Merge branch 'master' into replace_if_with_range

* zero len ranges fail

* fix Python backend

* fix llvm

* fix ptx

* yolo fix nir

* this works...

* always store...

* always store...

* Revert "always store..."

This reverts commit 0816cf344d.

.github/actions/setup-tinygrad/action.yml

@@ -302,4 +302,4 @@ runs:
     - name: Install mesa (macOS)
       if: inputs.mesa == 'true' && runner.os == 'macOS'
       shell: bash
-      run: brew install sirhcm/tinymesa/tinymesa
+      run: brew install sirhcm/tinymesa/tinymesa_cpu

.github/workflows/benchmark.yml

@@ -131,7 +131,7 @@ jobs:
     - name: UsbGPU copy speeds
       run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
     #- name: UsbGPU openpilot test
-    #  run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB NOLOCALS=0 IMAGE=0 GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+    #  run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
     - uses: actions/upload-artifact@v4
       with:
         name: Speed (Mac)
@@ -238,6 +238,8 @@ jobs:
       run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
     - name: Run LLaMA-3 8B on 4 GPUs with BEAM
       run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
+    - name: Run quantized LLaMA3
+      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8 | tee llama3_fp8.txt
     # - name: Run LLaMA-3 8B on 6 GPUs
     #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
     # - name: Run LLaMA-2 70B
@@ -271,6 +273,7 @@ jobs:
           llama3_beam.txt
           llama3_four_gpu.txt
           llama3_six_gpu.txt
+          llama3_fp8.txt
           llama_2_70B.txt
           mixtral.txt
           gpt2_unjitted.txt
@@ -319,9 +322,9 @@ jobs:
     - name: Run 10 CIFAR training steps
       run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=270 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
     - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=310 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=240 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
     - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=310 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=270 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
     # TODO: too slow
     # - name: Run 10 CIFAR training steps w winograd
     #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
@@ -619,22 +622,24 @@ jobs:
         rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
     - name: reset process replay
       run: test/external/process_replay/reset.py
-    - name: benchmark openpilot 0.9.9 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_9_9_vision PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
-    - name: benchmark openpilot 0.9.9 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_9_9_policy PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
-    - name: benchmark openpilot 0.9.9 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: openpilot compile3 0.9.9 driving_vision
-      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=18 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+      run: BENCHMARK_LOG=openpilot_0_9_9_vision PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
     - name: openpilot compile3 0.9.9 driving_policy
-      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=7 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_9_9_policy PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
     - name: openpilot compile3 0.9.9 dmonitoring
-      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=12 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
-    - name: openpilot compile3 Space Lab policy + vision
-      run: |
-        PYTHONPATH="." ASSERT_MIN_STEP_TIME=5 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/22aec22a10ce09384d4a4af2a0bbff08d54af7e0c888503508f356fae4ff0e29
-        PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/c824f68646a3b94f117f01c70dc8316fb466e05fbd42ccdba440b8a8dc86914b
+      run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: openpilot compile3 0.10.0 driving_policy
+      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=5 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
+    - name: openpilot compile3 0.10.0 dmonitoring
+      run: BENCHMARK_LOG=openpilot_0_10_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=13 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: openpilot compile3 0.10.1 driving_vision
+      # TODO: ASSERT_MIN_STEP_TIME=17
+      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=25 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot compile3 0.10.1 driving_policy
+      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=5 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
+    - name: openpilot compile3 0.10.1 dmonitoring
+      # TODO: ASSERT_MIN_STEP_TIME=10
+      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=13 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: benchmark MobileNetV2 on DSP
       run: |
         # generate quantized weights
@@ -645,16 +650,6 @@ jobs:
         PYTHONPATH=. CC=clang-19 DSP=1 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
     - name: Run process replay tests
       run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
-    - uses: actions/upload-artifact@v4
-      with:
-        name: Speed (comma)
-        path: |
-          openpilot_compile_0_9_4.txt
-          openpilot_compile_0_9_7.txt
-          openpilot_0_9_4.txt
-          openpilot_0_9_7.txt
-          openpilot_image_0_9_4.txt
-          openpilot_image_0_9_7.txt
 
   testreddriverbenchmark:
     name: AM Benchmark

.github/workflows/test.yml

@@ -204,7 +204,7 @@ jobs:
         DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
         DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
         DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        DEBUG=2 EMULATE=CUDA_SM89 ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
     - name: Test emulated INTEL OpenCL tensor cores
       run: DEBUG=2 EMULATE=INTEL FORWARD_ONLY=1 PYTHON=1 HALF=1 N=64 python3 ./extra/gemm/simple_matmul.py
     - name: Test emulated AMX tensor cores
@@ -264,8 +264,6 @@ jobs:
       run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
     - name: Run unit tests
       run: CPU=1 python -m pytest -n=auto test/unit/ --durations=20
-    - name: Check SPEC=1
-      run: SPEC=1 python3 test/test_tiny.py
     - name: Run targetted tests on NULL backend
       run: NULL=1 python3 -m unittest test.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step test/device/test_null.py
     # TODO: too slow
@@ -294,6 +292,25 @@ jobs:
     - name: Repo line count < 18000 lines
       run: MAX_LINE_COUNT=18000 python sz.py
 
+  spec:
+    strategy:
+      fail-fast: false
+      matrix:
+        group: [1, 2]
+    name: SPEC=2 (${{ matrix.group }})
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: spec-unit
+        deps: testing_unit
+    - name: Test SPEC=2
+      run: IGNORE_OOB=0 SPEC=2 PYTHONPATH="." pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
+
   fuzzing:
     name: Fuzzing
     runs-on: ubuntu-latest
@@ -351,7 +368,7 @@ jobs:
       - name: Run Kernel Count Test
         run: CL=1 python -m pytest -n=auto test/external/external_test_opt.py
       - name: Run fused optimizer tests
-        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py
+        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/test_optim.py -k "not muon"
       - name: Upload artifact
         uses: actions/upload-artifact@v4
         with:
@@ -374,15 +391,13 @@ jobs:
           llvm: 'true'
       - name: Test openpilot model kernel count and gate usage
         run: |
-          ALLOWED_KERNEL_COUNT=190 ALLOWED_READ_IMAGE=2081 ALLOWED_GATED_READ_IMAGE=28 FLOAT16=0 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx
-      - name: Test openpilot alt model correctness (float32)
-        run: FLOAT16=0 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/3799fe46b3a629e491d4b8498b8ae83e4c88c304/selfdrive/modeld/models/supercombo.onnx
-      - name: Test openpilot fastvits model correctness (float32)
-        run: FLOAT16=0 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
-      # - name: Test openpilot simple_plan vision model correctness (float32)
-      #   run: FLOAT16=0 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/35ff4f4577002f2685e50c8346addae33fe8da27a41dd4d6a0f14d1f4b1af81b
-      - name: Test openpilot LLVM compile
-        run: CPU=1 CPU_LLVM=1 LLVMOPT=1 JIT=2 BEAM=0 IMAGE=0 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1452 ALLOWED_GATED_READ_IMAGE=122 FLOAT16=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+      - name: Test openpilot CL compile fp16
+        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+      - name: Test openpilot CL compile fp32 (test correctness)
+        run: DEBUGCL=1 CL=1 IMAGE=2 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
+      - name: Test openpilot LLVM compile fp16
+        run: FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -524,11 +539,11 @@ jobs:
         pydeps: "pillow"
         llvm: "true"
     - name: Test LLVM=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
+      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py
     - name: Test LLVM=1 DEVECTORIZE=0 for model
       run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
     - name: Test CPU=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
+      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py
 
   testdsp:
     name: Linux (DSP)

.pre-commit-config.yaml

@@ -28,7 +28,7 @@ repos:
         pass_filenames: false
       - id: tests
         name: subset of tests
-        entry: env PYTHONPATH="." python3 -m pytest -n=8 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
+        entry: env OMP_NUM_THREADS=1 PYTHONPATH="." python3 -m pytest -n=8 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
         language: system
         always_run: true
-        pass_filenames: false
+        pass_filenames: false

autogen_stubs.sh

@@ -520,17 +520,12 @@ generate_mesa() {
   LVP_NIR_OPTIONS=$(./extra/mesa/lvp_nir_options.sh $MESA_SRC)
 
   fixup $BASE/mesa.py
-  patch_dlopen $BASE/mesa.py tinymesa_cpu "(BASE:=os.getenv('MESA_PATH', f\"/usr{'/local/' if helpers.OSX else '/'}lib\"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so')" "f'{BASE}/libtinymesa{EXT}'" "f'{brew_prefix()}/lib/libtinymesa_cpu.dylib'"
+  patch_dlopen $BASE/mesa.py tinymesa_cpu "(BASE:=os.getenv('MESA_PATH', f\"/usr{'/local/' if helpers.OSX else '/'}lib\"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so')" "f'{BASE}/libtinymesa{EXT}'" "'/opt/homebrew/lib/libtinymesa_cpu.dylib'" "'/opt/homebrew/lib/libtinymesa.dylib'"
   echo "lvp_nir_options = gzip.decompress(base64.b64decode('$LVP_NIR_OPTIONS'))" >> $BASE/mesa.py
-  cat <<EOF | sed -i "/import ctypes.*/r /dev/stdin" $BASE/mesa.py
-def brew_prefix():
-  try: return subprocess.check_output(['brew', '--prefix', 'tinymesa']).decode().strip()
-  except Exception: return ''
-EOF
-  sed -i "/in_dll/s/.*/try: &\nexcept AttributeError: pass/" $BASE/mesa.py
+  sed -i "/in_dll/s/.*/try: &\nexcept (AttributeError, ValueError): pass/" $BASE/mesa.py
   sed -i "s/import ctypes/import ctypes, ctypes.util, os, gzip, base64, subprocess, tinygrad.helpers as helpers/" $BASE/mesa.py
   sed -i "s/ctypes.CDLL('.\+')/(dll := _try_dlopen_tinymesa_cpu())/" $BASE/mesa.py
-  echo "def __getattr__(nm): raise AttributeError() if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa ($TINYMESA_TAG, $MESA_TAG)')" >> $BASE/mesa.py
+  echo "def __getattr__(nm): raise AttributeError('LLVMpipe requires tinymesa_cpu' if 'tinymesa_cpu' not in dll._name else f'attribute {nm} not found') if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa ($TINYMESA_TAG, $MESA_TAG)')" >> $BASE/mesa.py
   sed -i "s/ctypes.glsl_base_type/glsl_base_type/" $BASE/mesa.py
   # bitfield bug in clang2py
   sed -i "s/('fp_fast_math', ctypes.c_bool, 9)/('fp_fast_math', ctypes.c_uint32, 9)/" $BASE/mesa.py

docs/developer/kernelize.md

@@ -1,109 +0,0 @@
-# Kernel Creation
-
-Tinygrad lazily builds up a graph of Tensor operations. The Tensor graph includes a mix of:
-
-- Buffer and Assignment Ops: `BUFFER`, `BUFFER_VIEW`, `COPY`, `ASSIGN`
-- Movement Ops: `RESHAPE`, `EXPAND`, `PERMUTE`, `PAD`, `SHRINK`, `FLIP`
-- Compute Ops: `ADD`, `MUL`, `REDUCE_AXIS`, ...
-
-`Tensor.kernelize` creates the kernels and buffers needed to realize the output Tensor(s).
-
-## Kernelize flow
-
-Let's see how a multiply add Tensor graph becomes a fused elementwise kernel.
-
-```py
-# initialize 3 input buffers on the device
-a = Tensor([1]).realize()
-b = Tensor([2]).realize()
-c = Tensor([3]).realize()
-
-# create the Tensor graph
-mul = a*b
-out = mul+c
-
-print(mul) # <Tensor <UOp METAL (1,) int (<Ops.MUL: 48>, None)> on METAL with grad None>
-print(out) # <Tensor <UOp METAL (1,) int (<Ops.ADD: 52>, None)> on METAL with grad None>
-
-out.kernelize()
-
-print(mul) # <Tensor <UOp METAL (1,) int (<Ops.MUL: 48>, None)> on METAL with grad None>
-print(out) # <Tensor <UOp METAL (1,) int (<Ops.ASSIGN: 66>, None)> on METAL with grad None>
-```
-
-The multiply Tensor stays the same because it is fused. The output Tensor's UOp becomes a new ASSIGN UOp:
-
-```py
-print(out.uop)
-```
-
-The first source is the output BUFFER:
-
-```
-UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
-  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),
-  UOp(Ops.UNIQUE, dtypes.void, arg=6, src=()),))
-```
-
-And the second source is the KERNEL and its 4 buffer edges (output_buffer, a, b, c):
-
-```
-UOp(Ops.KERNEL, dtypes.void, arg=<Kernel 12 SINK(<Ops.STORE: 45>,) (__add__, __mul__)>, src=(
-  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
-    x1:=UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),
-    UOp(Ops.UNIQUE, dtypes.void, arg=6, src=()),)),
-  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
-     x1,
-    UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),)),
-  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
-     x1,
-    UOp(Ops.UNIQUE, dtypes.void, arg=3, src=()),)),
-  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
-     x1,
-    UOp(Ops.UNIQUE, dtypes.void, arg=5, src=()),)),))
-```
-
-KERNEL describes the compute AST, metadata and memory dependencies.
-
-BUFFER holds a reference to the device memory where the output will be stored.
-
-Once a Tensor is kernelized, all children will LOAD its BUFFER, instead of fusing it:
-
-```py
-child = out+2
-child.kernelize()
-print(child.uop.src[1].arg.ast)
-```
-
-```
-UOp(Ops.SINK, dtypes.void, arg=None, src=(
-  UOp(Ops.STORE, dtypes.void, arg=None, src=(
-    UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), arg=0, src=()),
-    x2:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1,), strides=(0,), offset=0, mask=None, contiguous=True),)), src=()),
-    UOp(Ops.ADD, dtypes.int, arg=None, src=(
-      UOp(Ops.LOAD, dtypes.int, arg=None, src=(
-        UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), arg=1, src=()),
-         x2,)),
-      UOp(Ops.CONST, dtypes.int, arg=2, src=(
-         x2,)),)),)),))
-```
-
-`Tensor.realize` will execute the kernels and write outputs to memory:
-
-```py
-Tensor.realize(out)
-print(out)        # <Tensor <UOp METAL (1,) int (<Ops.BUFFER: 23>, <buf real:True device:METAL size:1 dtype:dtypes.int offset:0>)> on METAL with grad None>
-print(out.item()) # 5
-```
-
-<hr />
-
-**Summary**
-
-- The large Tensor graph is built from a mix of data, compute and movement Ops.
-
-- `Tensor.kernelize` splits the Tensor graph into data (BUFFER), compute (KERNEL) and links dependencies with ASSIGN.
-
-- `Tensor.realize` executes KERNELs on device and replaces the Tensor graph with just a BUFFER.
-
-- Kernelize can be called multiple times on a Tensor. This allows for incrementally building the kernel fusion layout of a large Tensor graph, without having to call `realize` or `schedule`.

docs/env_vars.md

@@ -41,7 +41,7 @@ BEAM                | [#]        | number of beams in kernel beam search
 DEFAULT_FLOAT       | [HALF, ...]| specify the default float dtype (FLOAT32, HALF, BFLOAT16, FLOAT64, ...), default to FLOAT32
 IMAGE               | [1-2]      | enable 2d specific optimizations
 FLOAT16             | [1]        | use float16 for images instead of float32
-VISIBLE_DEVICES     | [list[int]]| restricts the NV/AMD devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
+HCQ_VISIBLE_DEVICES | [list[int]]| restricts the HCQ devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
 JIT                 | [0-2]      | 0=disabled, 1=[jit enabled](quickstart.md#jit) (default), 2=jit enabled, but graphs are disabled
 VIZ                 | [1]        | 0=disabled, 1=[viz enabled](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/viz)
 ALLOW_TF32          | [1]        | enable TensorFloat-32 tensor cores on Ampere or newer GPUs.

examples/llama3.py

@@ -145,6 +145,41 @@ def NF4Linear(block_size):
       return new_state_dict
   return _NF4Linear
 
+def quantize_to_fp8(x: Tensor, dtype=dtypes.fp8e4m3):
+  fp8_min = -448.0 if dtype == dtypes.fp8e4m3 else -57344.0
+  fp8_max = 448.0 if dtype == dtypes.fp8e4m3 else 57344.0
+  scale = fp8_max / x.abs().max()
+  x_scl_sat = (x * scale).clamp(fp8_min, fp8_max)
+  return x_scl_sat.cast(dtype), scale.float().reciprocal()
+
+class FP8Linear:
+  def __init__(self, in_features, out_features, bias=True):
+    self.weight = Tensor.empty(out_features, in_features, dtype=dtypes.fp8e4m3)
+    self.bias = Tensor.empty(out_features, dtype=dtypes.float16) if bias else None
+    self.weight_scale = Tensor.empty((), dtype=dtypes.float16)
+
+  def __call__(self, x:Tensor):
+    y = x.dot(self.weight.T.cast(dtypes.float32)) * self.weight_scale
+    if self.bias is not None: y = y + self.bias.cast(y.dtype)
+    return y.cast(x.dtype)
+
+  @staticmethod
+  def quantize(tensors, device, scale_dtype=dtypes.float16, quantize_embeds=False):
+    assert not quantize_embeds
+    new_tensors = {}
+    for name,v in tensors.items():
+      if "feed_forward" in name or "attention.w" in name:
+        assert "weight" in name, name
+        fp8_weight, scale = quantize_to_fp8(v)
+        new_tensors[name] = fp8_weight
+        new_tensors[name.replace('weight', 'weight_scale')] = scale.cast(scale_dtype)
+        if isinstance(device, tuple):
+          new_tensors[name].shard_(device, axis=-1)
+          new_tensors[name.replace('weight', 'weight_scale')].shard_(device, axis=None)
+      else:
+        new_tensors[name] = v
+    return new_tensors
+
 MODEL_PARAMS = {
   "1B": {
     "args": {"dim": 2048, "n_heads": 32, "n_kv_heads": 8, "n_layers": 16, "norm_eps": 1e-5, "rope_theta": 500000, "vocab_size": 128256, "hidden_dim": 8192},
@@ -167,6 +202,7 @@ def build_transformer(model_path: Path, model_size="8B", quantize=None, scale_dt
   # build model
   if quantize == "int8": linear, embedding, quantize_embeds = Int8Linear, Int8Embedding, True
   elif quantize == "nf4": linear, embedding, quantize_embeds = NF4Linear(64), nn.Embedding, False
+  elif quantize == "fp8": linear, embedding, quantize_embeds = FP8Linear, nn.Embedding, False
   else: linear, embedding, quantize_embeds = nn.Linear, nn.Embedding, False
   model = Transformer(**MODEL_PARAMS[model_size]["args"], linear=linear, embedding=embedding, max_context=max_context, jit=True)
 
@@ -242,7 +278,7 @@ if __name__ == "__main__":
   parser.add_argument("--model", type=Path, help="Model path")
   parser.add_argument("--size", choices=["1B", "8B", "70B", "405B"], default="1B", help="Model size")
   parser.add_argument("--shard", type=int, default=1, help="Shard the model across multiple devices")
-  parser.add_argument("--quantize", choices=["int8", "nf4", "float16"], help="Quantization method")
+  parser.add_argument("--quantize", choices=["int8", "nf4", "float16", "fp8"], help="Quantization method")
   parser.add_argument("--no_api", action="store_true", help="Disable the api and run a cli test interface")
   parser.add_argument("--host", type=str, default="0.0.0.0", help="Web server bind address")
   parser.add_argument("--port", type=int, default=7776, help="Web server port")

examples/openpilot/compile3.py

@@ -1,9 +1,5 @@
 import os, sys, pickle, time, re
 import numpy as np
-if "FLOAT16" not in os.environ: os.environ["FLOAT16"] = "1"
-if "IMAGE" not in os.environ: os.environ["IMAGE"] = "2"
-if "NOLOCALS" not in os.environ: os.environ["NOLOCALS"] = "1"
-if "JIT_BATCH_SIZE" not in os.environ: os.environ["JIT_BATCH_SIZE"] = "0"
 
 from tinygrad import fetch, Tensor, TinyJit, Context, GlobalCounters, Device, dtypes
 from tinygrad.helpers import DEBUG, getenv
@@ -21,11 +17,14 @@ def compile(onnx_file):
 
   input_shapes = {name: spec.shape for name, spec in run_onnx.graph_inputs.items()}
   input_types = {name: spec.dtype for name, spec in run_onnx.graph_inputs.items()}
+
   # Float inputs and outputs to tinyjits for openpilot are always float32
+  # TODO this seems dumb
   input_types = {k:(dtypes.float32 if v is dtypes.float16 else v) for k,v in input_types.items()}
   Tensor.manual_seed(100)
-  new_inputs = {k:Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize() for k,shp in sorted(input_shapes.items())}
-  new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
+  inputs = {k:Tensor(Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize().numpy(), device='NPY') for k,shp in sorted(input_shapes.items())}
+  if not getenv("NPY_IMG"):
+    inputs = {k:Tensor(v.numpy(), device=Device.DEFAULT).realize() if 'img' in k else v for k,v in inputs.items()}
   print("created tensors")
 
   run_onnx_jit = TinyJit(lambda **kwargs:
@@ -33,8 +32,6 @@ def compile(onnx_file):
   for i in range(3):
     GlobalCounters.reset()
     print(f"run {i}")
-    inputs = {**{k:v.clone() for k,v in new_inputs.items() if 'img' in k},
-              **{k:Tensor(v, device="NPY").realize() for k,v in new_inputs_numpy.items() if 'img' not in k}}
     with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1)):
       ret = run_onnx_jit(**inputs).numpy()
     # copy i == 1 so use of JITBEAM is okay
@@ -69,14 +66,9 @@ def compile(onnx_file):
   print(f"mdl size is {mdl_sz/1e6:.2f}M")
   print(f"pkl size is {pkl_sz/1e6:.2f}M")
   print("**** compile done ****")
-  return test_val
+  return inputs, test_val
 
-def test_vs_compile(run, new_inputs, test_val=None):
-  new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
-
-  # create fake "from_blob" tensors for the inputs, and wrapped NPY tensors for the numpy inputs (these have the same underlying memory)
-  inputs = {**{k:v for k,v in new_inputs.items() if 'img' in k},
-            **{k:Tensor(v, device="NPY").realize() for k,v in new_inputs_numpy.items() if 'img' not in k}}
+def test_vs_compile(run, inputs, test_val=None):
 
   # run 20 times
   step_times = []
@@ -93,68 +85,57 @@ def test_vs_compile(run, new_inputs, test_val=None):
     min_time = min(step_times)
     assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
 
-  print(out, val.shape, val.dtype)
   if test_val is not None: np.testing.assert_equal(test_val, val)
   print("**** test done ****")
 
   # test that changing the numpy changes the model outputs
-  if any([x.device == 'NPY' for x in inputs.values()]):
-    for v in new_inputs_numpy.values(): v *= 2
-    out = run(**inputs)
-    changed_val = out.numpy()
-    np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, val, changed_val)
+  inputs_2x = {k: Tensor(v.numpy()*2, device=v.device) for k,v in inputs.items()}
+  out = run(**inputs_2x)
+  changed_val = out.numpy()
+  np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, val, changed_val)
   return val
 
-def test_vs_onnx(new_inputs, test_val, onnx_file, ort=False):
-  new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
+def test_vs_onnx(new_inputs, test_val, onnx_file, tol):
+  import onnxruntime as ort
+  
+  onnx_inputs = {k:v.numpy() for k,v in new_inputs.items()}
   onnx_model = onnx.load(onnx_file)
 
-  timings = []
-  if ort:
-    # test with onnxruntime
-    import onnxruntime as ort
-    onnx_session = ort.InferenceSession(onnx_file)
-    for _ in range(1 if test_val is not None else 5):
-      st = time.perf_counter()
-      onnx_output = onnx_session.run([onnx_model.graph.output[0].name], {k:v.astype(np.float16) for k,v in new_inputs_numpy.items()})
-      timings.append(time.perf_counter() - st)
-    new_torch_out = onnx_output[0]
-  else:
-    # test with torch
-    import torch
-    from onnx2torch import convert
-    inputs = {k.name:new_inputs_numpy[k.name] for k in onnx_model.graph.input}
-    torch_model = convert(onnx_model).float()
-    with torch.no_grad():
-      for _ in range(1 if test_val is not None else 5):
-        st = time.perf_counter()
-        torch_out = torch_model(*[torch.tensor(x) for x in inputs.values()])
-        timings.append(time.perf_counter() - st)
-      new_torch_out = torch_out.numpy()
+  ORT_TO_NP_DTYPES: dict[str, np.dtype] = {
+    'tensor(float)': np.dtype('float32'),
+    'tensor(float16)': np.dtype('float16'),
+    'tensor(uint8)': np.dtype('uint8'),
+  }
 
-  if test_val is not None:
-    np.testing.assert_allclose(new_torch_out.reshape(test_val.shape), test_val, atol=1e-4, rtol=1e-2)
-    print("test vs onnx passed")
+  timings = []
+  onnx_session = ort.InferenceSession(onnx_file)
+  onnx_types = {x.name: ORT_TO_NP_DTYPES[x.type] for x in onnx_session.get_inputs()}
+  onnx_inputs = {k:onnx_inputs[k].astype(onnx_types[k]) for k in onnx_inputs}
+
+  for _ in range(1 if test_val is not None else 5):
+    st = time.perf_counter()
+    onnx_output = onnx_session.run([onnx_model.graph.output[0].name], onnx_inputs)
+    timings.append(time.perf_counter() - st)
+
+  np.testing.assert_allclose(onnx_output[0].reshape(test_val.shape), test_val, atol=tol, rtol=tol)
+  print("test vs onnx passed")
   return timings
 
+def bench(run, inputs):
+  from extra.bench_log import WallTimeEvent, BenchEvent
+  for _ in range(10):
+    with WallTimeEvent(BenchEvent.STEP):
+      run(**inputs).numpy()
+
 if __name__ == "__main__":
   onnx_file = fetch(OPENPILOT_MODEL)
-  test_val = compile(onnx_file) if not getenv("RUN") else None
+  inputs, outputs = compile(onnx_file)
 
   with open(OUTPUT, "rb") as f: pickle_loaded = pickle.load(f)
 
-  # same randomness as compile
-  Tensor.manual_seed(100)
-  new_inputs = {nm:Tensor.randn(*st.shape, dtype=dtype).mul(8).realize() for nm, (st, _, dtype, _) in
-                sorted(zip(pickle_loaded.captured.expected_names, pickle_loaded.captured.expected_st_vars_dtype_device))}
-
-  test_val = test_vs_compile(pickle_loaded, new_inputs, test_val)
-  if getenv("BENCHMARK"):
-    for be in ["torch", "ort"]:
-      try:
-        timings = test_vs_onnx(new_inputs, None, onnx_file, be=="ort")
-        print(f"timing {be}: {min(timings)*1000:.2f} ms")
-      except Exception as e:
-        print(f"{be} fail with {e}")
-  if not getenv("FLOAT16"): test_vs_onnx(new_inputs, test_val, onnx_file, getenv("ORT"))
+  test_vs_compile(pickle_loaded, inputs, outputs)
+  if getenv("SELFTEST"):
+    test_vs_onnx(inputs, outputs, onnx_file, 1e-4)
 
+  if getenv("BENCHMARK_LOG", ""):
+    bench(pickle_loaded, inputs)

examples/sdv2.py

@@ -99,6 +99,7 @@ if __name__ == "__main__":
   parser.add_argument('--timing',      action='store_true', help="Print timing per step")
   parser.add_argument('--noshow',      action='store_true', help="Don't show the image")
   parser.add_argument('--fp16',        action='store_true', help="Cast the weights to float16")
+  parser.add_argument('--fakeweights', action='store_true', help="Skip loading checkpoints and use fake weights")
   args = parser.parse_args()
 
   N = 1
@@ -112,19 +113,22 @@ if __name__ == "__main__":
 
   model = StableDiffusionV2(**params)
 
-  default_weights_url = 'https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors'
-  weights_fn = args.weights_fn
-  if not weights_fn:
-    weights_url = args.weights_url if args.weights_url else default_weights_url
-    weights_fn  = fetch(weights_url, os.path.basename(str(weights_url)))
-
   with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    load_state_dict(model, safe_load(weights_fn), strict=False)
+    if not args.fakeweights:
+      default_weights_url = 'https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors'
+      weights_fn = args.weights_fn
+      if not weights_fn:
+        weights_url = args.weights_url if args.weights_url else default_weights_url
+        weights_fn  = fetch(weights_url, os.path.basename(str(weights_url)))
+
+        load_state_dict(model, safe_load(weights_fn), strict=False)
 
     if args.fp16:
       for k,v in get_state_dict(model).items():
         if k.startswith("model"):
-          v.replace(v.cast(dtypes.float16).realize())
+          v.replace(v.cast(dtypes.float16))
+
+    Tensor.realize(*get_state_dict(model).values())
 
   c  = { "crossattn": model.cond_stage_model(args.prompt) }
   uc = { "crossattn": model.cond_stage_model("") }

examples/stable_diffusion.py

@@ -263,14 +263,16 @@ if __name__ == "__main__":
   parser.add_argument('--timing', action='store_true', help="Print timing per step")
   parser.add_argument('--seed', type=int, help="Set the random latent seed")
   parser.add_argument('--guidance', type=float, default=7.5, help="Prompt strength")
+  parser.add_argument('--fakeweights', action='store_true', help="Skip loading checkpoints and use fake weights")
   args = parser.parse_args()
 
   model = StableDiffusion()
 
   # load in weights
   with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    model_bin = fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt')
-    load_state_dict(model, torch_load(model_bin)['state_dict'], verbose=False, strict=False, realize=False)
+    if not args.fakeweights:
+      model_bin = fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt')
+      load_state_dict(model, torch_load(model_bin)['state_dict'], verbose=False, strict=False, realize=False)
 
     if args.fp16:
       for k,v in get_state_dict(model).items():

extra/cl_android.sh

@@ -0,0 +1,4 @@
+# source extra/cl_android.sh
+export LD_LIBRARY_PATH=/data/data/com.termux/files/usr/lib:/system/vendor/lib64
+export LD_PRELOAD=/system/vendor/lib64/libOpenCL.so
+

extra/gemm/amd_uop_matmul.py

@@ -328,8 +328,7 @@ if __name__ == "__main__":
   elif HL == 1: hprg = hl_spec_kernel3()
   else: hprg = hand_spec_kernel3()
   if HL == 3:
-    with Context(BLOCK_REORDER=0):
-      prg = get_program(hprg, Device.default.renderer)
+    prg = get_program(hprg, Device.default.renderer)
   else:
     prg = get_program(hprg, Device.default.renderer)
   print(prg.src)

extra/gemm/simple_matmul.py

@@ -5,8 +5,10 @@ from tinygrad.dtype import _to_np_dtype
 from tinygrad.codegen.opt import OptOps
 from tinygrad.engine.realize import lower_schedule
 
-dtype_in = dtypes.half if getenv("HALF") else dtypes.bfloat16 if getenv("BFLOAT16") else dtypes.float
-acc_dtype = dtypes.half if getenv("ACC_HALF") else dtypes.bfloat16 if getenv("ACC_BFLOAT16") else None
+dtype_in = (dtypes.half if getenv("HALF") else dtypes.bfloat16 if getenv("BFLOAT16") else
+            dtypes.fp8e4m3 if getenv("FP8E4M3") else dtypes.fp8e5m2 if getenv("FP8E5M2") else dtypes.float)
+acc_dtype = (dtypes.half if getenv("ACC_HALF") else dtypes.bfloat16 if getenv("ACC_BFLOAT16") else
+            dtypes.fp8e4m3 if getenv("ACC_FP8E4M3") else dtypes.fp8e5m2 if getenv("ACC_FP8E5M2") else None)
 if getenv("INT"):  dtype_in, acc_dtype = dtypes.int8, dtypes.int32
 if getenv("UINT"): dtype_in, acc_dtype = dtypes.uint8, dtypes.int32
 
@@ -14,8 +16,10 @@ N = getenv("N", 4096)
 M = getenv("M", N)
 K = getenv("K", N)
 CNT = getenv("CNT", 10)
-ATOL = getenv("ATOL", 1e-4)
-RTOL = getenv("RTOL", 3e-2)
+
+atol, rtol = {dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1)}.get(dtype_in, (1e-4, 3e-2))
+ATOL, RTOL = getenv("ATOL", atol), getenv("RTOL", rtol)
+
 INT_LOW = getenv("INT_LOW", 0)
 INT_HIGH = getenv("INT_HIGH", 10)
 

extra/gemm/torch_gemm.py

@@ -8,19 +8,22 @@ import torch
 torch.set_num_threads(1)
 from tinygrad.helpers import getenv
 CUDA = getenv("CUDA", 1)
+MPS = getenv("MPS", 0)
 
-for dtype in [torch.float32, torch.float16]:
+for dtype in [torch.float32, torch.float16, torch.bfloat16]:
   for N in [256, 512, 1024, 2048, 4096]:
     FLOPS = N*N*N*2
 
     b = torch.rand((N,N), dtype=dtype)
     c = torch.rand((N,N), dtype=dtype)
     if CUDA: b,c = b.cuda(),c.cuda()
+    if MPS: b,c = b.to('mps'),c.to('mps')
 
     def torch_prog(b, c):
       st = time.perf_counter()
       a = b@c
       if CUDA: torch.cuda.synchronize()
+      if MPS: torch.mps.synchronize()
       return time.perf_counter() - st
     tm = min([torch_prog(b, c) for _ in range(20)])
     print(f"{N*N:10d} {tm*1e6:9.2f} us, would be {FLOPS*1e-9/tm:9.2f} GFLOPS {N:4d}x{N:4d}x{N:4d} matmul in {dtype}")

extra/hcq/hcq_smi.py

@@ -1,7 +1,6 @@
 #!/usr/bin/env python3
 
-from tinygrad.runtime.support.system import System
-import argparse, glob, os, re, time, subprocess, sys
+import argparse, glob, os, time, subprocess, sys
 
 def scan_devs_based_on_lock(prefix:str, args) -> list[str]:
   target_dev = args.pci_bus if 'pci_bus' in args.__dir__() else ""
@@ -12,7 +11,7 @@ def scan_devs_based_on_lock(prefix:str, args) -> list[str]:
     if os.path.exists(f"/sys/bus/pci/devices/{dev_id}") and dev_id.startswith(target_dev): devs.append(dev_id)
   return devs
 
-def _do_reset_device(pci_bus): System.pci_reset(pci_bus)
+def _do_reset_device(pci_bus): os.system(f"sudo sh -c 'echo 1 > /sys/bus/pci/devices/{pci_bus}/reset'")
 def _is_module_loaded(name: str) -> bool: return os.path.isdir(f"/sys/module/{name}")
 
 def cmd_remove_module(args):

extra/mmapeak/mmapeak.py

@@ -7,31 +7,34 @@ import os
 NUM_WORKGROUPS = 96
 WAVE_SIZE = 32
 NUM_WAVES = 2
-FLOPS_PER_MATMUL =  16*16*16*2
-INTERNAL_LOOP  = 1_000_000
-INSTRUCTIONS_PER_LOOP = 1_000
+FLOPS_PER_MATMUL = 16*16*16*2
+INTERNAL_LOOP = 1_000_00
+INSTRUCTIONS_PER_LOOP = 200
+DIRECTIVE = ".amdhsa_wavefront_size32 1"
 
 assemblyTemplate = (pathlib.Path(__file__).parent / "template.s").read_text()
 
-def launchBenchmark(instruction, vgprIndices, dense = True):
-  if dense:
+def launchBenchmark(instruction, vgprIndices, dense=True, accum=False, extra=""):
+  if accum:
+    instructions = "{} a[0:{}], v[{}:{}], v[{}:{}], 1{}\n".format(instruction, vgprIndices[0],
+                                                                vgprIndices[1], vgprIndices[2],
+                                                                vgprIndices[1], vgprIndices[2], extra)
+  elif dense:
     instructions = "{} v[0:{}], v[{}:{}], v[{}:{}], 1\n".format(instruction, vgprIndices[0],
                                                                 vgprIndices[1], vgprIndices[2],
-                                                                vgprIndices[1], vgprIndices[2]) * INSTRUCTIONS_PER_LOOP
+                                                                vgprIndices[1], vgprIndices[2])
   else:
     instructions = "{} v[0:{}], v[{}:{}], v[{}:{}], v{}\n".format(instruction, vgprIndices[0],
-                                                                vgprIndices[1], vgprIndices[2],
-                                                                vgprIndices[3], vgprIndices[4],
-                                                                vgprIndices[5]) * INSTRUCTIONS_PER_LOOP
-  src = assemblyTemplate.replace("INSTRUCTION", instructions)
+                                                                  vgprIndices[1], vgprIndices[2],
+                                                                  vgprIndices[3], vgprIndices[4],
+                                                                  vgprIndices[5])
+  src = assemblyTemplate.replace("INTERNAL_LOOP", str(INTERNAL_LOOP)).replace("INSTRUCTION", instructions*INSTRUCTIONS_PER_LOOP)
+  src = src.replace("DIRECTIVE", DIRECTIVE)
   lib = COMPILER.compile(src)
   fxn = AMDProgram(DEV, "matmul", lib)
-  start = time.perf_counter()
-  fxn(global_size=(NUM_WORKGROUPS,1,1), local_size=(WAVE_SIZE*NUM_WAVES,1,1), wait=True) #For some reason the returned time is very small after the first kernel execution
-  end = time.perf_counter()
-  elapsed = end-start
+  elapsed = fxn(global_size=(NUM_WORKGROUPS,1,1), local_size=(WAVE_SIZE*NUM_WAVES,1,1), wait=True)
   FLOPs = FLOPS_PER_MATMUL * NUM_WAVES * NUM_WORKGROUPS * INTERNAL_LOOP * INSTRUCTIONS_PER_LOOP
-  print("{:<29} : {} T(FL)OPS".format(instruction, round(FLOPs/elapsed/10**12, 2)))
+  print(f"{instruction:<29} : {FLOPs/elapsed/10**12:.2f} T(FL)OPS")
 
 if __name__=="__main__":
   DEVICENUM = os.getenv("DEVICENUM", "0")
@@ -40,18 +43,17 @@ if __name__=="__main__":
   except:
     raise RuntimeError("Error while initiating AMD device")
 
-  if (ARCH := DEV.arch) not in ['gfx1100', 'gfx1201']:
-    raise RuntimeError("only gfx1100 and gfx1201 supported")
-  COMPILER = HIPCompiler(ARCH)
-
-  if ARCH == 'gfx1100':
+  COMPILER = HIPCompiler(DEV.arch)
+  if DEV.arch in {'gfx1100', 'gfx1103'}:
+    if DEV.arch == 'gfx1103':
+      NUM_WORKGROUPS = 8
     launchBenchmark("v_wmma_bf16_16x16x16_bf16", (7,8,15))
     launchBenchmark("v_wmma_f16_16x16x16_f16", (7,8,15))
     launchBenchmark("v_wmma_f32_16x16x16_bf16", (7,8,15))
     launchBenchmark("v_wmma_f32_16x16x16_f16", (7,8,15))
     launchBenchmark("v_wmma_i32_16x16x16_iu4", (7,8,9))
     launchBenchmark("v_wmma_i32_16x16x16_iu8", (7,8,11))
-  if ARCH == 'gfx1201':
+  elif DEV.arch == 'gfx1201':
     NUM_WORKGROUPS = 64
     launchBenchmark("v_wmma_bf16_16x16x16_bf16", (3,4,7))
     launchBenchmark("v_wmma_f16_16x16x16_f16", (3,4,7))
@@ -76,4 +78,18 @@ if __name__=="__main__":
     launchBenchmark("v_swmmac_f32_16x16x32_bf8_fp8", (7,8,9,10,13,14), False)
     launchBenchmark("v_swmmac_f32_16x16x32_bf8_bf8", (7,8,9,10,13,14), False)
     FLOPS_PER_MATMUL = 16*16*64*2
-    launchBenchmark("v_swmmac_i32_16x16x64_iu4", (7,8,9,10,13,14), False)
+    launchBenchmark("v_swmmac_i32_16x16x64_iu4", (7,8,9,10,13,14), False)
+  elif DEV.arch == 'gfx950':
+    DIRECTIVE = ".amdhsa_accum_offset 4"
+    NUM_WORKGROUPS = 256
+    WAVE_SIZE = 64
+    NUM_WAVES = 4
+    launchBenchmark("v_mfma_f32_16x16x16_bf16", (3,0,1), accum=True)
+    FLOPS_PER_MATMUL = 16*16*32*2
+    launchBenchmark("v_mfma_f32_16x16x32_bf16", (3,0,3), accum=True)
+    FLOPS_PER_MATMUL = 16*16*128*2
+    launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,7), accum=True) # fp8
+    launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,5), accum=True, extra=", cbsz:2 blgp:2") # fp6
+    launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,3), accum=True, extra=", cbsz:4 blgp:4") # fp4
+  else:
+    raise RuntimeError(f"arch {DEV.arch} not supported.")

extra/mmapeak/template.s

@@ -1,9 +1,9 @@
     .text
     .globl matmul
-    .p2align 8 
+    .p2align 8
     .type matmul,@function
 matmul:
-    s_mov_b32 s1, 1000000
+    s_mov_b32 s1, INTERNAL_LOOP
     s_mov_b32 s2, 0
     inner_loop:
         INSTRUCTION
@@ -17,7 +17,7 @@ matmul:
 .amdhsa_kernel matmul
   .amdhsa_next_free_vgpr .amdgcn.next_free_vgpr
   .amdhsa_next_free_sgpr .amdgcn.next_free_sgpr
-  .amdhsa_wavefront_size32 1
+  DIRECTIVE
 .end_amdhsa_kernel
 
 .amdgpu_metadata
@@ -28,7 +28,7 @@ amdhsa.version:
 amdhsa.kernels:
   - .name: matmul
     .symbol: matmul.kd
-    .kernarg_segment_size:  0
+    .kernarg_segment_size: 0
     .group_segment_fixed_size: 0
     .private_segment_fixed_size: 0
     .kernarg_segment_align: 4
@@ -36,6 +36,5 @@ amdhsa.kernels:
     .sgpr_count: 8
     .vgpr_count: 32
     .max_flat_workgroup_size: 1024
-    .args:
 ...
 .end_amdgpu_metadata

extra/qcom_gpu_driver/adreno_pm4.xml

@@ -1,7 +1,8 @@
 <?xml version="1.0" encoding="UTF-8"?>
 <database xmlns="http://nouveau.freedesktop.org/"
 xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
-xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
+xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.xsd">
+<import file="freedreno_copyright.xml"/>
 <import file="adreno/adreno_common.xml"/>
 
 <enum name="vgt_event_type" varset="chip">
@@ -20,9 +21,9 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<value name="HLSQ_FLUSH" value="7" variants="A3XX-A4XX"/>
 	<value name="VIZQUERY_END" value="8" variants="A2XX"/>
 	<value name="SC_WAIT_WC" value="9" variants="A2XX"/>
-	<value name="WRITE_PRIMITIVE_COUNTS" value="9" variants="A6XX"/>
-	<value name="START_PRIMITIVE_CTRS" value="11" variants="A6XX"/>
-	<value name="STOP_PRIMITIVE_CTRS" value="12" variants="A6XX"/>
+	<value name="WRITE_PRIMITIVE_COUNTS" value="9" variants="A6XX-"/>
+	<value name="START_PRIMITIVE_CTRS" value="11" variants="A6XX-"/>
+	<value name="STOP_PRIMITIVE_CTRS" value="12" variants="A6XX-"/>
 	<!-- Not sure that these 4 events don't have the same meaning as on A5XX+ -->
 	<value name="RST_PIX_CNT" value="13" variants="A2XX-A4XX"/>
 	<value name="RST_VTX_CNT" value="14" variants="A2XX-A4XX"/>
@@ -30,8 +31,8 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<value name="STAT_EVENT" value="16" variants="A2XX-A4XX"/>
 	<value name="CACHE_FLUSH_AND_INV_TS_EVENT" value="20" variants="A2XX-A4XX"/>
 	<doc>
-		If A6XX_RB_SAMPLE_COUNT_CONTROL.copy is true, writes OQ Z passed
-		sample counts to RB_SAMPLE_COUNT_ADDR.  This writes to main
+		If A6XX_RB_SAMPLE_COUNTER_CNTL.copy is true, writes OQ Z passed
+		sample counts to RB_SAMPLE_COUNTER_BASE.  This writes to main
 		memory, skipping UCHE.
 	</doc>
 	<value name="ZPASS_DONE" value="21"/>
@@ -96,6 +97,13 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	</doc>
 	<value name="BLIT" value="30" variants="A5XX-"/>
 
+	<doc>
+	Flip between the primary and secondary LRZ buffers. This is used
+	for concurrent binning, so that BV can write to one buffer while
+	BR reads from the other.
+	</doc>
+	<value name="LRZ_FLIP_BUFFER" value="36" variants="A7XX-"/>
+
 	<doc>
 		Clears based on GRAS_LRZ_CNTL configuration, could clear
 		fast-clear buffer or LRZ direction.
@@ -112,11 +120,12 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<value name="LRZ_FLUSH" value="38" variants="A5XX-"/>
 	<value name="BLIT_OP_FILL_2D" value="39" variants="A5XX-"/>
 	<value name="BLIT_OP_COPY_2D" value="40" variants="A5XX-A6XX"/>
-	<value name="UNK_40" value="40" variants="A7XX"/>
+	<value name="LRZ_CACHE_INVALIDATE" value="40" variants="A7XX-"/>
+	<value name="LRZ_Q_CACHE_INVALIDATE" value="41" variants="A7XX-"/>
 	<value name="BLIT_OP_SCALE_2D" value="42" variants="A5XX-"/>
 	<value name="CONTEXT_DONE_2D" value="43" variants="A5XX-"/>
-	<value name="UNK_2C" value="44" variants="A5XX-"/>
-	<value name="UNK_2D" value="45" variants="A5XX-"/>
+	<value name="VSC_BINNING_START" value="44" variants="A5XX-"/>
+	<value name="VSC_BINNING_END" value="45" variants="A5XX-"/>
 
 	<!-- a6xx events -->
 	<doc>
@@ -129,21 +138,22 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<!-- note, some of these are the same as a6xx, just named differently -->
 
 	<doc> Doesn't seem to do anything </doc>
-	<value name="DUMMY_EVENT" value="1" variants="A7XX"/>
-	<value name="CCU_INVALIDATE_DEPTH" value="24" variants="A7XX"/>
-	<value name="CCU_INVALIDATE_COLOR" value="25" variants="A7XX"/>
-	<value name="CCU_RESOLVE_CLEAN" value="26" variants="A7XX"/>
-	<value name="CCU_FLUSH_DEPTH" value="28" variants="A7XX"/>
-	<value name="CCU_FLUSH_COLOR" value="29" variants="A7XX"/>
-	<value name="CCU_RESOLVE" value="30" variants="A7XX"/>
-	<value name="CCU_END_RESOLVE_GROUP" value="31" variants="A7XX"/>
-	<value name="CCU_CLEAN_DEPTH" value="32" variants="A7XX"/>
-	<value name="CCU_CLEAN_COLOR" value="33" variants="A7XX"/>
-	<value name="CACHE_RESET" value="48" variants="A7XX"/>
-	<value name="CACHE_CLEAN" value="49" variants="A7XX"/>
+	<value name="DUMMY_EVENT" value="1" variants="A7XX-"/>
+	<value name="CCU_INVALIDATE_DEPTH" value="24" variants="A7XX-"/>
+	<value name="CCU_INVALIDATE_COLOR" value="25" variants="A7XX-"/>
+	<value name="CCU_RESOLVE_CLEAN" value="26" variants="A7XX-"/>
+	<value name="CCU_FLUSH_DEPTH" value="28" variants="A7XX-"/>
+	<value name="CCU_FLUSH_COLOR" value="29" variants="A7XX-"/>
+	<value name="CCU_RESOLVE" value="30" variants="A7XX-"/>
+	<value name="CCU_END_RESOLVE_GROUP" value="31" variants="A7XX-"/>
+	<value name="CCU_CLEAN_DEPTH" value="32" variants="A7XX-"/>
+	<value name="CCU_CLEAN_COLOR" value="33" variants="A7XX-"/>
+	<value name="CACHE_RESET" value="48" variants="A7XX-"/>
+	<value name="CACHE_CLEAN" value="49" variants="A7XX-"/>
 	<!-- TODO: deal with name conflicts with other gens -->
-	<value name="CACHE_FLUSH7" value="50" variants="A7XX"/>
-	<value name="CACHE_INVALIDATE7" value="51" variants="A7XX"/>
+	<value name="CACHE_FLUSH7" value="50" variants="A7XX-"/>
+	<value name="CACHE_INVALIDATE7" value="51" variants="A7XX-"/>
+	<value name="DEPTH_BUFFER_FLIP" value="0x3d" variants="A8XX-"/>
 </enum>
 
 <enum name="pc_di_primtype">
@@ -324,7 +334,7 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<doc>fetch state sub-blocks and initiate shader code DMAs</doc>
 	<value name="CP_SET_STATE" value="0x25"/>
 	<doc>load constant into chip and to memory</doc>
-	<value name="CP_SET_CONSTANT" value="0x2d"/>
+	<value name="CP_SET_CONSTANT" value="0x2d" variants="A2XX"/>
 	<doc>load sequencer instruction memory (pointer-based)</doc>
 	<value name="CP_IM_LOAD" value="0x27"/>
 	<doc>load sequencer instruction memory (code embedded in packet)</doc>
@@ -371,7 +381,7 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<value name="CP_LOAD_STATE" value="0x30" variants="A3XX"/>
 	<value name="CP_LOAD_STATE4" value="0x30" variants="A4XX-A5XX"/>
 	<doc>Conditionally load a IB based on a flag, prefetch enabled</doc>
-	<value name="CP_COND_INDIRECT_BUFFER_PFE" value="0x3a"/>
+	<value name="CP_COND_INDIRECT_BUFFER_PFE" value="0x3a" variants="A3XX-A5XX"/>
 	<doc>Conditionally load a IB based on a flag, prefetch disabled</doc>
 	<value name="CP_COND_INDIRECT_BUFFER_PFD" value="0x32" variants="A3XX"/>
 	<doc>Load a buffer with pre-fetch enabled</doc>
@@ -514,7 +524,7 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<!--
 	Seems to set the mode flags which control which CP_SET_DRAW_STATE
 	packets are executed, based on their ENABLE_MASK values
-	
+
 	CP_SET_MODE w/ payload of 0x1 seems to cause CP_SET_DRAW_STATE
 	packets w/ ENABLE_MASK & 0x6 to execute immediately
 	 -->
@@ -537,7 +547,7 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<value name="CP_LOAD_STATE6_GEOM" value="0x32" variants="A6XX-"/>
 	<value name="CP_LOAD_STATE6_FRAG" value="0x34" variants="A6XX-"/>
 	<!--
-	Note: For IBO state (Image/SSBOs) which have shared state across
+	Note: For UAV state (Image/SSBOs) which have shared state across
 	shader stages, for 3d pipeline CP_LOAD_STATE6 is used.  But for
 	compute shaders, CP_LOAD_STATE6_FRAG is used.  Possibly they are
 	interchangable.
@@ -566,20 +576,21 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<value name="IN_PREEMPT" value="0x0f" variants="A6XX-"/>
 
 	<!-- TODO do these exist on A5xx? -->
-	<value name="CP_SCRATCH_WRITE" value="0x4c" variants="A6XX"/>
+	<value name="CP_SCRATCH_WRITE" value="0x4c" variants="A6XX-"/>
 	<value name="CP_REG_TO_MEM_OFFSET_MEM" value="0x74" variants="A6XX-"/>
 	<value name="CP_REG_TO_MEM_OFFSET_REG" value="0x72" variants="A6XX-"/>
 	<value name="CP_WAIT_MEM_GTE" value="0x14" variants="A6XX"/>
 	<value name="CP_WAIT_TWO_REGS" value="0x70" variants="A6XX"/>
 	<value name="CP_MEMCPY" value="0x75" variants="A6XX-"/>
 	<value name="CP_SET_BIN_DATA5_OFFSET" value="0x2e" variants="A6XX-"/>
+	<!-- A750+, set in place of CP_SET_BIN_DATA5_OFFSET but has different values -->
+	<value name="CP_SET_UNK_BIN_DATA" value="0x2d" variants="A7XX-"/>
 	<doc>
                 Write CP_CONTEXT_SWITCH_*_INFO from CP to the following dwords,
                 and forcibly switch to the indicated context.
 	</doc>
 	<value name="CP_CONTEXT_SWITCH" value="0x54" variants="A6XX"/>
-	<!-- Note, kgsl calls this CP_SET_AMBLE: -->
-	<value name="CP_SET_CTXSWITCH_IB" value="0x55" variants="A6XX-"/>
+	<value name="CP_SET_AMBLE" value="0x55" variants="A6XX-"/>
 
 	<!--
 	Seems to always have the payload:
@@ -630,8 +641,7 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 	<value name="CP_BV_BR_COUNT_OPS" value="0x1b" variants="A7XX-"/>
 	<doc> Clears, adds to local, or adds to global timestamp </doc>
 	<value name="CP_MODIFY_TIMESTAMP" value="0x1c" variants="A7XX-"/>
-	<!-- similar to CP_CONTEXT_REG_BUNCH, but discards first two dwords?? -->
-	<value name="CP_CONTEXT_REG_BUNCH2" value="0x5d" variants="A7XX-"/>
+	<value name="CP_NON_CONTEXT_REG_BUNCH" value="0x5d" variants="A7XX-"/>
 	<doc>
 		Write to a scratch memory that is read by CP_REG_TEST with
 		SOURCE_SCRATCH_MEM set. It's not the same scratch as scratch registers.
@@ -648,6 +658,11 @@ xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 
 	<doc>Reset various on-chip state used for synchronization</doc>
 	<value name="CP_RESET_CONTEXT_STATE" value="0x1f" variants="A7XX-"/>
+
+	<doc>Invalidates the "CCHE" introduced on a740</doc>
+	<value name="CP_CCHE_INVALIDATE" value="0x3a" variants="A7XX-"/>
+
+	<value name="CP_SCOPE_CNTL" value="0x6c" variants="A7XX-"/>
 </enum>
 
 
@@ -790,14 +805,14 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<value name="SB6_GS_SHADER" value="0xb"/>
 		<value name="SB6_FS_SHADER" value="0xc"/>
 		<value name="SB6_CS_SHADER" value="0xd"/>
-		<value name="SB6_IBO"       value="0xe"/>
-		<value name="SB6_CS_IBO"    value="0xf"/>
+		<value name="SB6_UAV"       value="0xe"/>
+		<value name="SB6_CS_UAV"    value="0xf"/>
 	</enum>
 	<enum name="a6xx_state_type">
 		<value name="ST6_SHADER" value="0"/>
 		<value name="ST6_CONSTANTS" value="1"/>
 		<value name="ST6_UBO" value="2"/>
-		<value name="ST6_IBO" value="3"/>
+		<value name="ST6_UAV" value="3"/>
 	</enum>
 	<enum name="a6xx_state_src">
 		<value name="SS6_DIRECT" value="0"/>
@@ -903,12 +918,6 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	</reg32>
 
 	<stripe varset="chip" variants="A5XX-">
-		<reg32 offset="4" name="4">
-			<bitfield name="INDX_BASE_LO" low="0" high="31"/>
-		</reg32>
-		<reg32 offset="5" name="5">
-			<bitfield name="INDX_BASE_HI" low="0" high="31"/>
-		</reg32>
 		<reg64 offset="4" name="INDX_BASE" type="address"/>
 		<reg32 offset="6" name="6">
 			<!-- max # of elements in index buffer -->
@@ -1084,8 +1093,10 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 			<bitfield name="BINNING" pos="20" varset="chip" variants="A6XX-" type="boolean"/>
 			<bitfield name="GMEM" pos="21" varset="chip" variants="A6XX-" type="boolean"/>
 			<bitfield name="SYSMEM" pos="22" varset="chip" variants="A6XX-" type="boolean"/>
-			<bitfield name="GROUP_ID" low="24" high="28" type="uint"/>
+			<!-- high bit is 28 until a750: -->
+			<bitfield name="GROUP_ID" low="24" high="29" type="uint"/>
 		</reg32>
+		<reg64 offset="1" name="ADDR" type="address"/>
 		<reg32 offset="1" name="1">
 			<bitfield name="ADDR_LO" low="0" high="31" type="hex"/>
 		</reg32>
@@ -1119,39 +1130,63 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	</reg32>
 </domain>
 
+<enum name="a7xx_abs_mask_mode">
+	<value name="ABS_MASK" value="0x1"/>
+	<value name="NO_ABS_MASK" value="0x0"/>
+</enum>
+
 <domain name="CP_SET_BIN_DATA5" width="32">
 	<reg32 offset="0" name="0">
+		<bitfield name="VSC_MASK" low="0" high="15" type="hex">
+			<doc>
+				A mask of bins, starting at VSC_N, whose
+				visibility is OR'd together. A value of 0 is
+				interpreted as 1 (i.e. just use VSC_N for
+				visbility) for backwards compatibility. Only
+				exists on a7xx.
+			</doc>
+		</bitfield>
 		<!-- equiv to PC_VSTREAM_CONTROL.SIZE on a3xx/a4xx: -->
 		<bitfield name="VSC_SIZE" low="16" high="21" type="uint"/>
 		<!-- equiv to PC_VSTREAM_CONTROL.N on a3xx/a4xx: -->
 		<bitfield name="VSC_N" low="22" high="26" type="uint"/>
+		<bitfield name="ABS_MASK" pos="28" type="a7xx_abs_mask_mode" addvariant="yes">
+			<doc>
+				If this field is 1, VSC_MASK and VSC_N are
+				ignored and instead a new ordinal immediately
+				after specifies the full 32-bit mask of bins
+				to use. The mask is "absolute" instead of
+				relative to VSC_N.
+			</doc>
+		</bitfield>
 	</reg32>
-	<!-- BIN_DATA_ADDR -> VSC_PIPE[p].DATA_ADDRESS -->
-	<reg32 offset="1" name="1">
-		<bitfield name="BIN_DATA_ADDR_LO" low="0" high="31" type="hex"/>
-	</reg32>
-	<reg32 offset="2" name="2">
-		<bitfield name="BIN_DATA_ADDR_HI" low="0" high="31" type="hex"/>
-	</reg32>
-	<!-- BIN_SIZE_ADDRESS -> VSC_SIZE_ADDRESS + (p * 4)-->
-	<reg32 offset="3" name="3">
-		<bitfield name="BIN_SIZE_ADDRESS_LO" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="4" name="4">
-		<bitfield name="BIN_SIZE_ADDRESS_HI" low="0" high="31"/>
-	</reg32>
-	<!-- new on a6xx, where BIN_DATA_ADDR is the DRAW_STRM: -->
-	<reg32 offset="5" name="5">
-		<bitfield name="BIN_PRIM_STRM_LO" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="6" name="6">
-		<bitfield name="BIN_PRIM_STRM_HI" low="0" high="31"/>
-	</reg32>
-	<!--
-		a7xx adds a few more addresses to the end of the pkt
-	 -->
-	<reg64 offset="7" name="7"/>
-	<reg64 offset="9" name="9"/>
+	<stripe varset="a7xx_abs_mask_mode" variants="NO_ABS_MASK">
+		<!-- BIN_DATA_ADDR -> VSC_PIPE[p].DATA_ADDRESS -->
+		<reg64 offset="1" name="BIN_DATA_ADDR" type="address"/>
+		<!-- BIN_SIZE_ADDRESS -> VSC_SIZE_ADDRESS + (p * 4)-->
+		<reg64 offset="3" name="BIN_SIZE_ADDR" type="address"/>
+		<!-- new on a6xx, where BIN_DATA_ADDR is the DRAW_STRM: -->
+		<reg64 offset="5" name="BIN_PRIM_STRM" type="address"/>
+		<!--
+			a7xx adds a few more addresses to the end of the pkt
+		 -->
+		<reg64 offset="7" name="7"/>
+		<reg64 offset="9" name="9"/>
+	</stripe>
+	<stripe varset="a7xx_abs_mask_mode" variants="ABS_MASK">
+		<reg32 offset="1" name="ABS_MASK"/>
+		<!-- BIN_DATA_ADDR -> VSC_PIPE[p].DATA_ADDRESS -->
+		<reg64 offset="2" name="BIN_DATA_ADDR" type="address"/>
+		<!-- BIN_SIZE_ADDRESS -> VSC_SIZE_ADDRESS + (p * 4)-->
+		<reg64 offset="4" name="BIN_SIZE_ADDR" type="address"/>
+		<!-- new on a6xx, where BIN_DATA_ADDR is the DRAW_STRM: -->
+		<reg64 offset="6" name="BIN_PRIM_STRM" type="address"/>
+		<!--
+			a7xx adds a few more addresses to the end of the pkt
+		 -->
+		<reg64 offset="8" name="8"/>
+		<reg64 offset="10" name="10"/>
+	</stripe>
 </domain>
 
 <domain name="CP_SET_BIN_DATA5_OFFSET" width="32">
@@ -1162,23 +1197,42 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
                 stream is recorded.
 	</doc>
 	<reg32 offset="0" name="0">
+		<bitfield name="VSC_MASK" low="0" high="15" type="hex"/>
 		<!-- equiv to PC_VSTREAM_CONTROL.SIZE on a3xx/a4xx: -->
 		<bitfield name="VSC_SIZE" low="16" high="21" type="uint"/>
 		<!-- equiv to PC_VSTREAM_CONTROL.N on a3xx/a4xx: -->
 		<bitfield name="VSC_N" low="22" high="26" type="uint"/>
+		<bitfield name="ABS_MASK" pos="28" type="a7xx_abs_mask_mode" addvariant="yes"/>
 	</reg32>
-	<!-- BIN_DATA_ADDR -> VSC_PIPE[p].DATA_ADDRESS -->
-	<reg32 offset="1" name="1">
-		<bitfield name="BIN_DATA_OFFSET" low="0" high="31" type="uint"/>
-	</reg32>
-	<!-- BIN_SIZE_ADDRESS -> VSC_SIZE_ADDRESS + (p * 4)-->
-	<reg32 offset="2" name="2">
-		<bitfield name="BIN_SIZE_OFFSET" low="0" high="31" type="uint"/>
-	</reg32>
-	<!-- BIN_DATA2_ADDR -> VSC_PIPE[p].DATA2_ADDRESS -->
-	<reg32 offset="3" name="3">
-		<bitfield name="BIN_DATA2_OFFSET" low="0" high="31" type="uint"/>
-	</reg32>
+	<stripe varset="a7xx_abs_mask_mode" variants="NO_ABS_MASK">
+		<!-- BIN_DATA_ADDR -> VSC_PIPE[p].DATA_ADDRESS -->
+		<reg32 offset="1" name="1">
+			<bitfield name="BIN_DATA_OFFSET" low="0" high="31" type="uint"/>
+		</reg32>
+		<!-- BIN_SIZE_ADDRESS -> VSC_SIZE_ADDRESS + (p * 4)-->
+		<reg32 offset="2" name="2">
+			<bitfield name="BIN_SIZE_OFFSET" low="0" high="31" type="uint"/>
+		</reg32>
+		<!-- BIN_DATA2_ADDR -> VSC_PIPE[p].DATA2_ADDRESS -->
+		<reg32 offset="3" name="3">
+			<bitfield name="BIN_DATA2_OFFSET" low="0" high="31" type="uint"/>
+		</reg32>
+	</stripe>
+	<stripe varset="a7xx_abs_mask_mode" variants="ABS_MASK">
+		<reg32 offset="1" name="ABS_MASK"/>
+		<!-- BIN_DATA_ADDR -> VSC_PIPE[p].DATA_ADDRESS -->
+		<reg32 offset="2" name="2">
+			<bitfield name="BIN_DATA_OFFSET" low="0" high="31" type="uint"/>
+		</reg32>
+		<!-- BIN_SIZE_ADDRESS -> VSC_SIZE_ADDRESS + (p * 4)-->
+		<reg32 offset="3" name="3">
+			<bitfield name="BIN_SIZE_OFFSET" low="0" high="31" type="uint"/>
+		</reg32>
+		<!-- BIN_DATA2_ADDR -> VSC_PIPE[p].DATA2_ADDRESS -->
+		<reg32 offset="4" name="4">
+			<bitfield name="BIN_DATA2_OFFSET" low="0" high="31" type="uint"/>
+		</reg32>
+	</stripe>
 </domain>
 
 <domain name="CP_REG_RMW" width="32">
@@ -1196,6 +1250,9 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	</doc>
 	<reg32 offset="0" name="0">
 		<bitfield name="DST_REG" low="0" high="17" type="hex"/>
+		<bitfield name="DST_SCRATCH" pos="19" type="boolean" varset="chip" variants="A7XX-"/>
+		<!-- skip implied CP_WAIT_FOR_IDLE + CP_WAIT_FOR_ME -->
+		<bitfield name="SKIP_WAIT_FOR_ME" pos="23" type="boolean" varset="chip" variants="A7XX-"/>
 		<bitfield name="ROTATE" low="24" high="28" type="uint"/>
 		<bitfield name="SRC1_ADD" pos="29" type="boolean"/>
 		<bitfield name="SRC1_IS_REG" pos="30" type="boolean"/>
@@ -1209,7 +1266,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	</reg32>
 </domain>
 
-<domain name="CP_REG_TO_MEM" width="32">
+<domain name="CP_REG_TO_MEM" width="32" prefix="chip">
 	<reg32 offset="0" name="0">
 		<bitfield name="REG" low="0" high="17" type="hex"/>
 		<!-- number of registers/dwords copied is max(CNT, 1). -->
@@ -1217,12 +1274,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="64B" pos="30" type="boolean"/>
 		<bitfield name="ACCUMULATE" pos="31" type="boolean"/>
 	</reg32>
-	<reg32 offset="1" name="1">
-		<bitfield name="DEST" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="2" name="2" varset="chip" variants="A5XX-">
-		<bitfield name="DEST_HI" low="0" high="31"/>
-	</reg32>
+	<stripe varset="chip" variants="A2XX-A4XX">
+		<reg32 offset="1" name="DEST" type="address"/>
+	</stripe>
+	<stripe varset="chip" variants="A5XX-">
+		<reg64 offset="1" name="DEST" type="address"/>
+	</stripe>
 </domain>
 
 <domain name="CP_REG_TO_MEM_OFFSET_REG" width="32">
@@ -1238,12 +1295,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="64B" pos="30" type="boolean"/>
 		<bitfield name="ACCUMULATE" pos="31" type="boolean"/>
 	</reg32>
-	<reg32 offset="1" name="1">
-		<bitfield name="DEST" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="2" name="2" varset="chip" variants="A5XX-">
-		<bitfield name="DEST_HI" low="0" high="31"/>
-	</reg32>
+	<reg64 offset="1" name="DEST" type="waddress"/>
 	<reg32 offset="3" name="3">
 		<bitfield name="OFFSET0" low="0" high="17" type="hex"/>
 		<bitfield name="OFFSET0_SCRATCH" pos="19" type="boolean"/>
@@ -1263,18 +1315,8 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="64B" pos="30" type="boolean"/>
 		<bitfield name="ACCUMULATE" pos="31" type="boolean"/>
 	</reg32>
-	<reg32 offset="1" name="1">
-		<bitfield name="DEST" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="2" name="2" varset="chip" variants="A5XX-">
-		<bitfield name="DEST_HI" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="3" name="3">
-		<bitfield name="OFFSET_LO" low="0" high="31" type="hex"/>
-	</reg32>
-	<reg32 offset="4" name="4">
-		<bitfield name="OFFSET_HI" low="0" high="31" type="hex"/>
-	</reg32>
+	<reg64 offset="1" name="DEST" type="waddress"/>
+	<reg64 offset="3" name="OFFSET" type="waddress"/>
 </domain>
 
 <domain name="CP_MEM_TO_REG" width="32">
@@ -1287,12 +1329,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<!-- does the same thing as CP_MEM_TO_MEM::UNK31 -->
 		<bitfield name="UNK31" pos="31" type="boolean"/>
 	</reg32>
-	<reg32 offset="1" name="1">
-		<bitfield name="SRC" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="2" name="2" varset="chip" variants="A5XX-">
-		<bitfield name="SRC_HI" low="0" high="31"/>
-	</reg32>
+	<stripe varset="chip" variants="A2XX-A4XX">
+		<reg32 offset="1" name="SRC" type="address"/>
+	</stripe>
+	<stripe varset="chip" variants="A5XX-">
+		<reg64 offset="1" name="SRC" type="address"/>
+	</stripe>
 </domain>
 
 <domain name="CP_MEM_TO_MEM" width="32">
@@ -1312,6 +1354,10 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<!-- some other kind of wait -->
 		<bitfield name="UNK31" pos="31" type="boolean"/>
 	</reg32>
+	<reg64 offset="1" name="DST" type="waddress"/>
+	<reg64 offset="3" name="SRC_A" type="address"/>
+	<reg64 offset="5" name="SRC_B" type="address"/>
+	<reg64 offset="7" name="SRC_C" type="address"/>
 	<!--
 	followed by sequence of addresses.. the first is the
 	destination and the rest are N src addresses which are
@@ -1346,6 +1392,8 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="SCRATCH" low="20" high="22" type="uint"/>
 		<!-- number of registers/dwords copied is CNT + 1. -->
 		<bitfield name="CNT" low="24" high="26" type="uint"/>
+		<!-- skip implied CP_WAIT_FOR_IDLE + CP_WAIT_FOR_ME -->
+		<bitfield name="SKIP_WAIT_FOR_ME" pos="27" type="boolean" varset="chip" variants="A7XX-"/>
 	</reg32>
 </domain>
 
@@ -1368,12 +1416,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 </domain>
 
 <domain name="CP_MEM_WRITE" width="32">
-	<reg32 offset="0" name="0">
-		<bitfield name="ADDR_LO" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="1" name="1">
-		<bitfield name="ADDR_HI" low="0" high="31"/>
-	</reg32>
+	<stripe varset="chip" variants="A2XX-A4XX">
+		<reg32 offset="0" name="ADDR" type="address"/>
+	</stripe>
+	<stripe varset="chip" variants="A5XX-">
+		<reg64 offset="0" name="ADDR" type="address"/>
+	</stripe>
 	<!-- followed by the DWORDs to write -->
 </domain>
 
@@ -1425,24 +1473,14 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="POLL" low="4" high="5" type="poll_memory_type"/>
 		<bitfield name="WRITE_MEMORY" pos="8" type="boolean"/>
 	</reg32>
-	<reg32 offset="1" name="1">
-		<bitfield name="POLL_ADDR_LO" low="0" high="31" type="hex"/>
-	</reg32>
-	<reg32 offset="2" name="2">
-		<bitfield name="POLL_ADDR_HI" low="0" high="31" type="hex"/>
-	</reg32>
+	<reg64 offset="1" name="POLL_ADDR" type="address"/>
 	<reg32 offset="3" name="3">
 		<bitfield name="REF" low="0" high="31"/>
 	</reg32>
 	<reg32 offset="4" name="4">
 		<bitfield name="MASK" low="0" high="31"/>
 	</reg32>
-	<reg32 offset="5" name="5">
-		<bitfield name="WRITE_ADDR_LO" low="0" high="31" type="hex"/>
-	</reg32>
-	<reg32 offset="6" name="6">
-		<bitfield name="WRITE_ADDR_HI" low="0" high="31" type="hex"/>
-	</reg32>
+	<reg64 offset="5" name="WRITE_ADDR" type="waddress"/>
 	<reg32 offset="7" name="7">
 		<bitfield name="WRITE_DATA" low="0" high="31"/>
 	</reg32>
@@ -1457,12 +1495,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<!-- Reserved for flags, presumably? Unused in FW -->
 		<bitfield name="RESERVED" low="0" high="31" type="hex"/>
 	</reg32>
-	<reg32 offset="1" name="1">
-		<bitfield name="POLL_ADDR_LO" low="0" high="31" type="hex"/>
-	</reg32>
-	<reg32 offset="2" name="2">
-		<bitfield name="POLL_ADDR_HI" low="0" high="31" type="hex"/>
-	</reg32>
+	<reg64 offset="1" name="POLL_ADDR" type="address"/>
 	<reg32 offset="3" name="3">
 		<bitfield name="REF" low="0" high="31"/>
 	</reg32>
@@ -1480,12 +1513,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="POLL" low="4" high="5" type="poll_memory_type"/>
 		<bitfield name="WRITE_MEMORY" pos="8" type="boolean"/>
 	</reg32>
-	<reg32 offset="1" name="1">
-		<bitfield name="POLL_ADDR_LO" low="0" high="31" type="hex"/>
-	</reg32>
-	<reg32 offset="2" name="2">
-		<bitfield name="POLL_ADDR_HI" low="0" high="31" type="hex"/>
-	</reg32>
+	<reg64 offset="1" name="POLL_ADDR" type="address"/>
 	<reg32 offset="3" name="3">
 		<bitfield name="REF" low="0" high="31"/>
 	</reg32>
@@ -1619,12 +1647,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	TODO what is gpuaddr for, seems to be all 0's.. maybe needed for
 	context switch?
 	 -->
-	<reg32 offset="1" name="1">
-		<bitfield name="ADDR_0_LO" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="2" name="2">
-		<bitfield name="ADDR_0_HI" low="0" high="31"/>
-	</reg32>
+	<reg64 offset="1" name="ADDR" type="waddress"/>
 	<reg32 offset="3" name="3">
 		<!-- ??? -->
 	</reg32>
@@ -1653,8 +1676,8 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="WRITE_SAMPLE_COUNT" pos="12" type="boolean"/>
 		<!-- Write sample count at (iova + 16) -->
 		<bitfield name="SAMPLE_COUNT_END_OFFSET" pos="13" type="boolean"/>
-		<!-- *(iova + 8) = *(iova + 16) - *iova -->
-		<bitfield name="WRITE_SAMPLE_COUNT_DIFF" pos="14" type="boolean"/>
+		<!-- *(iova + 8) += *(iova + 16) - *iova -->
+		<bitfield name="WRITE_ACCUM_SAMPLE_COUNT_DIFF" pos="14" type="boolean"/>
 
 		<!-- Next 4 flags are valid to set only when concurrent binning is enabled -->
 		<!-- Increment 16b BV counter. Valid only in BV pipe -->
@@ -1668,15 +1691,11 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="WRITE_DST" pos="24" type="event_write_dst" addvariant="yes"/>
 		<!-- Writes into WRITE_DST from WRITE_SRC. RB_DONE_TS requires WRITE_ENABLED. -->
 		<bitfield name="WRITE_ENABLED" pos="27" type="boolean"/>
+		<bitfield name="IRQ" pos="31" type="boolean"/>
 	</reg32>
 
 	<stripe varset="event_write_dst" variants="EV_DST_RAM">
-		<reg32 offset="1" name="1">
-			<bitfield name="ADDR_0_LO" low="0" high="31"/>
-		</reg32>
-		<reg32 offset="2" name="2">
-			<bitfield name="ADDR_0_HI" low="0" high="31"/>
-		</reg32>
+		<reg64 offset="1" name="1" type="waddress"/>
 		<reg32 offset="3" name="3">
 			<bitfield name="PAYLOAD_0" low="0" high="31"/>
 		</reg32>
@@ -1743,9 +1762,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	<reg32 offset="0" name="0">
 	</reg32>
 	<stripe varset="chip" variants="A4XX">
-		<reg32 offset="1" name="1">
-			<bitfield name="ADDR" low="0" high="31"/>
-		</reg32>
+		<reg32 offset="1" name="ADDR" type="address"/>
 		<reg32 offset="2" name="2">
 			<!-- localsize is value minus one: -->
 			<bitfield name="LOCALSIZEX" low="2" high="11" type="uint"/>
@@ -1754,12 +1771,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		</reg32>
 	</stripe>
 	<stripe varset="chip" variants="A5XX-">
-		<reg32 offset="1" name="1">
-			<bitfield name="ADDR_LO" low="0" high="31"/>
-		</reg32>
-		<reg32 offset="2" name="2">
-			<bitfield name="ADDR_HI" low="0" high="31"/>
-		</reg32>
+		<reg64 offset="1" name="ADDR" type="address"/>
 		<reg32 offset="3" name="3">
 			<!-- localsize is value minus one: -->
 			<bitfield name="LOCALSIZEX" low="2" high="11" type="uint"/>
@@ -1771,40 +1783,88 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 
 <domain name="CP_SET_MARKER" width="32" varset="chip" prefix="chip" variants="A6XX-">
 	<doc>Tell CP the current operation mode, indicates save and restore procedure</doc>
+	<enum name="set_marker_mode">
+		<value value="0" name="SET_RENDER_MODE"/>
+		<!-- IFPC - inter-frame power collapse -->
+		<value value="1" name="SET_IFPC_MODE"/>
+	</enum>
+	<enum name="a6xx_ifpc_mode">
+		<value value="0" name="IFPC_ENABLE"/>
+		<value value="1" name="IFPC_DISABLE"/>
+	</enum>
 	<enum name="a6xx_marker">
-		<value value="1" name="RM6_BYPASS"/>
-		<value value="2" name="RM6_BINNING"/>
-		<value value="4" name="RM6_GMEM"/>
-		<value value="5" name="RM6_ENDVIS"/>
-		<value value="6" name="RM6_RESOLVE"/>
-		<value value="7" name="RM6_YIELD"/>
+		<value value="1" name="RM6_DIRECT_RENDER"/>
+		<value value="2" name="RM6_BIN_VISIBILITY"/>
+		<value value="3" name="RM6_BIN_DIRECT"/>
+		<value value="4" name="RM6_BIN_RENDER_START"/>
+		<value value="5" name="RM6_BIN_END_OF_DRAWS"/>
+		<value value="6" name="RM6_BIN_RESOLVE"/>
+		<value value="7" name="RM6_BIN_RENDER_END"/>
 		<value value="8" name="RM6_COMPUTE"/>
-		<value value="0xc" name="RM6_BLIT2DSCALE"/>  <!-- no-op (at least on current sqe fw) -->
+		<value value="12" name="RM6_BLIT2DSCALE"/>  <!-- no-op (at least on current sqe fw) -->
 
 		<!--
 			These values come from a6xx_set_marker() in the
 			downstream kernel, and they can only be set by the kernel
 		-->
-		<value value="0xd" name="RM6_IB1LIST_START"/>
-		<value value="0xe" name="RM6_IB1LIST_END"/>
-		<!-- IFPC - inter-frame power collapse -->
-		<value value="0x100" name="RM6_IFPC_ENABLE"/>
-		<value value="0x101" name="RM6_IFPC_DISABLE"/>
+		<value value="13" name="RM6_IB1LIST_START"/>
+		<value value="14" name="RM6_IB1LIST_END"/>
+		<value value="15" name="RM7_BIN_VISIBILITY_END"/>
+
+		<!-- new in a8xx: -->
+		<value value="32" name="RM8_DEPTH_PASS_START"/>
+		<value value="33" name="RM8_DEPTH_PASS_END"/>
 	</enum>
-	<reg32 offset="0" name="0">
-		<!--
-			NOTE: blob driver and some versions of freedreno/turnip set
-			b4, which is unused (at least by current sqe fw), but interferes
-			with parsing if we extend the size of the bitfield to include
-			b8 (only sent by kernel mode driver).  Really, the way the
-			parsing works in the firmware, only b0-b3 are considered, but
-			if b8 is set, the low bits are interpreted differently.  To
-			model this, without getting confused by spurious b4, this is
-			described as two overlapping bitfields:
-		 -->
-		<bitfield name="MODE" low="0" high="8" type="a6xx_marker"/>
-		<bitfield name="MARKER" low="0" high="3" type="a6xx_marker"/>
-	</reg32>
+	<stripe varset="chip" variants="A6XX-A7XX">
+		<reg32 offset="0" name="0">
+			<!-- if b8 is set, the low bits are interpreted differently (and b4 ignored) -->
+			<bitfield name="MARKER_MODE" pos="8" type="set_marker_mode" addvariant="yes"/>
+
+
+			<bitfield name="MODE" low="0" high="3" type="a6xx_marker" varset="set_marker_mode" variants="SET_RENDER_MODE"/>
+			<!-- used by preemption to determine if GMEM needs to be saved or not -->
+			<bitfield name="USES_GMEM" pos="4" type="boolean" varset="set_marker_mode" variants="SET_RENDER_MODE"/>
+
+
+			<bitfield name="IFPC_MODE" pos="0" type="a6xx_ifpc_mode" varset="set_marker_mode" variants="SET_IFPC_MODE"/>
+
+
+			<!--
+				CP_SET_MARKER is used with these bits to create a
+				critical section around a workaround for ray tracing.
+				The workaround happens after BVH building, and appears
+				to invalidate the RTU's BVH node cache. It makes sure
+				that only one of BR/BV/LPAC is executing the
+				workaround at a time, and no draws using RT on BV/LPAC
+				are executing while the workaround is executed on BR (or
+				vice versa, that no draws on BV/BR using RT are executed
+				while the workaround executes on LPAC), by
+				hooking subsequent CP_EVENT_WRITE/CP_DRAW_*/CP_EXEC_CS.
+				The blob usage is:
+
+
+				CP_SET_MARKER(RT_WA_START)
+				... workaround here ...
+				CP_SET_MARKER(RT_WA_END)
+				...
+				CP_SET_MARKER(SHADER_USES_RT)
+				CP_DRAW_INDX(...) or CP_EXEC_CS(...)
+			-->
+			<bitfield name="SHADER_USES_RT" pos="9" type="boolean" variants="A7XX-"/>
+			<bitfield name="RT_WA_START" pos="10" type="boolean" variants="A7XX-"/>
+			<bitfield name="RT_WA_END" pos="11" type="boolean" variants="A7XX-"/>
+		</reg32>
+	</stripe>
+	<stripe varset="chip" variants="A8XX-">
+		<reg32 offset="0" name="0">
+			<!-- if b8 is set, the low bits are interpreted differently (and b4 ignored) -->
+			<bitfield name="MARKER_MODE" pos="8" type="set_marker_mode" addvariant="yes"/>
+			<bitfield name="USES_GMEM" pos="7" type="boolean" varset="set_marker_mode" variants="SET_RENDER_MODE"/>
+			<bitfield name="MODE" low="0" high="6" type="a6xx_marker" varset="set_marker_mode" variants="SET_RENDER_MODE"/>
+			<bitfield name="IFPC_MODE" pos="0" type="a6xx_ifpc_mode" varset="set_marker_mode" variants="SET_IFPC_MODE"/>
+			<!-- idk if the RT w/a fields apply to a8xx as well -->
+		</reg32>
+	</stripe>
 </domain>
 
 <domain name="CP_SET_PSEUDO_REG" width="32" varset="chip" prefix="chip" variants="A6XX-">
@@ -1830,9 +1890,9 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 			If concurrent binning is disabled then BR also does binning so it will also
 			write the "real" registers in BR.
 		-->
-		<value value="8" name="DRAW_STRM_ADDRESS"/>
-		<value value="9" name="DRAW_STRM_SIZE_ADDRESS"/>
-		<value value="10" name="PRIM_STRM_ADDRESS"/>
+		<value value="8" name="VSC_PIPE_DATA_DRAW_BASE"/>
+		<value value="9" name="VSC_SIZE_BASE"/>
+		<value value="10" name="VSC_PIPE_DATA_PRIM_BASE"/>
 		<value value="11" name="UNK_STRM_ADDRESS"/>
 		<value value="12" name="UNK_STRM_SIZE_ADDRESS"/>
 
@@ -1933,11 +1993,11 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 			a bitmask of which modes pass the test.
 		-->
 
-		<!-- RM6_BINNING -->
+		<!-- RM6_BIN_VISIBILITY -->
 		<bitfield name="BINNING" pos="25" variants="RENDER_MODE" type="boolean"/>
 		<!-- all others -->
 		<bitfield name="GMEM" pos="26" variants="RENDER_MODE" type="boolean"/>
-		<!-- RM6_BYPASS -->
+		<!-- RM6_DIRECT_RENDER -->
 		<bitfield name="SYSMEM" pos="27" variants="RENDER_MODE" type="boolean"/>
 
 		<bitfield name="BV" pos="25" variants="THREAD_MODE" type="boolean"/>
@@ -2010,54 +2070,45 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	</reg32>
 </domain>
 
-<domain name="CP_SET_CTXSWITCH_IB" width="32">
+<domain name="CP_SET_AMBLE" width="32">
 	<doc>
-                Used by the userspace driver to set various IB's which are
-                executed during context save/restore for handling
-                state that isn't restored by the
-                context switch routine itself.
+		Used by the userspace and kernel drivers to set various IB's
+		which are executed during context save/restore for handling
+		state that isn't restored by the context switch routine itself.
 	</doc>
-	<enum name="ctxswitch_ib">
-		<value name="RESTORE_IB" value="0">
+	<enum name="amble_type">
+		<value name="PREAMBLE_AMBLE_TYPE" value="0">
 			<doc>Executed unconditionally when switching back to the context.</doc>
 		</value>
-		<value name="YIELD_RESTORE_IB" value="1">
+		<value name="BIN_PREAMBLE_AMBLE_TYPE" value="1">
                         <doc>
 				Executed when switching back after switching
 				away during execution of
-				a CP_SET_MARKER packet with RM6_YIELD as the
-				payload *and* the normal save routine was
-				bypassed for a shorter one. I think this is
-				connected to the "skipsaverestore" bit set by
-				the kernel when preempting.
+				a CP_SET_MARKER packet with RM6_BIN_RENDER_END as the
+				payload *and* skipsaverestore is set. This is
+				expected to restore static register values not
+				saved when skipsaverestore is set.
 			</doc>
 		</value>
-		<value name="SAVE_IB" value="2">
+		<value name="POSTAMBLE_AMBLE_TYPE" value="2">
                         <doc>
 				Executed when switching away from the context,
 				except for context switches initiated via
 				CP_YIELD.
                         </doc>
 		</value>
-		<value name="RB_SAVE_IB" value="3">
+		<value name="KMD_AMBLE_TYPE" value="3">
 			<doc>
 				This can only be set by the RB (i.e. the kernel)
 				and executes with protected mode off, but
-				is otherwise similar to SAVE_IB.
-
-				Note, kgsl calls this CP_KMD_AMBLE_TYPE
+				is otherwise similar to POSTAMBLE_AMBLE_TYPE.
 			</doc>
 		</value>
 	</enum>
-	<reg32 offset="0" name="0">
-		<bitfield name="ADDR_LO" low="0" high="31"/>
-	</reg32>
-	<reg32 offset="1" name="1">
-		<bitfield name="ADDR_HI" low="0" high="31"/>
-	</reg32>
+	<reg64 offset="0" name="ADDR" type="address"/>
 	<reg32 offset="2" name="2">
 		<bitfield name="DWORDS" low="0" high="19" type="uint"/>
-		<bitfield name="TYPE" low="20" high="21" type="ctxswitch_ib"/>
+		<bitfield name="TYPE" low="20" high="21" type="amble_type"/>
 	</reg32>
 </domain>
 
@@ -2089,12 +2140,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<value name="UNK_EVENT_WRITE" value="0x4"/>
 		<doc>
 			Tracks GRAS_LRZ_CNTL::GREATER, GRAS_LRZ_CNTL::DIR, and
-			GRAS_LRZ_DEPTH_VIEW with previous values, and if one of
+			GRAS_LRZ_VIEW_INFO with previous values, and if one of
 			the following is true:
 			- GRAS_LRZ_CNTL::GREATER has changed
 			- GRAS_LRZ_CNTL::DIR has changed, the old value is not
 			  CUR_DIR_GE, and the new value is not CUR_DIR_DISABLED
-			- GRAS_LRZ_DEPTH_VIEW has changed
+			- GRAS_LRZ_VIEW_INFO has changed
 			then it does a LRZ_FLUSH with GRAS_LRZ_CNTL::ENABLE
 			forced to 1.
 			Only exists in a650_sqe.fw.
@@ -2209,7 +2260,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 
 <domain name="CP_MEM_TO_SCRATCH_MEM" width="32">
 	<doc>
-		Best guess is that it is a faster way to fetch all the VSC_STATE registers
+		Best guess is that it is a faster way to fetch all the VSC_CHANNEL_VISIBILITY registers
 		and keep them in a local scratch memory instead of fetching every time
 		when skipping IBs.
 	</doc>
@@ -2257,7 +2308,25 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	<reg32 offset="0" name="0">
 		<bitfield name="CLEAR_ON_CHIP_TS" pos="0" type="boolean"/>
 		<bitfield name="CLEAR_RESOURCE_TABLE" pos="1" type="boolean"/>
-		<bitfield name="CLEAR_GLOBAL_LOCAL_TS" pos="2" type="boolean"/>
+		<bitfield name="CLEAR_BV_BR_COUNTER" pos="2" type="boolean"/>
+		<bitfield name="RESET_GLOBAL_LOCAL_TS" pos="3" type="boolean"/>
+	</reg32>
+</domain>
+
+<domain name="CP_SCOPE_CNTL" width="32">
+	<enum name="cp_scope">
+		<value value="0" name="INTERRUPTS"/>
+	</enum>
+	<reg32 offset="0" name="0">
+		<bitfield name="DISABLE_PREEMPTION" pos="0" type="boolean"/>
+		<bitfield low="28" high="31" name="SCOPE" type="cp_scope"/>
+	</reg32>
+</domain>
+
+<domain name="CP_INDIRECT_BUFFER" width="32" varset="chip" prefix="chip" variants="A5XX-">
+	<reg64 offset="0" name="IB_BASE" type="address"/>
+	<reg32 offset="2" name="2">
+		<bitfield name="IB_SIZE" low="0" high="19"/>
 	</reg32>
 </domain>
 

extra/qcom_gpu_driver/opencl_ioctl.py

@@ -97,7 +97,7 @@ def parse_cmd_buf(dat):
 
         if state_block == SB6_CS_SHADER:
           from extra.disassemblers.adreno import disasm_raw
-          if state_type == ST6_SHADER and IOCTL > 2:
+          if state_type == ST6_SHADER and IOCTL > 3:
             disasm_raw(get_mem(((vals[2] << 32) | vals[1]), num_unit * 128))
           if state_type == ST6_CONSTANTS:
             x = get_mem(((vals[2] << 32) | vals[1]), num_unit*4)
@@ -106,25 +106,30 @@ def parse_cmd_buf(dat):
               print('constants')
               hexdump(x)
           if state_type == ST6_IBO:
-            ibos_bytes = get_mem((vals[2] << 32) | vals[1], num_unit * 16 * 4)
+            if state_src == 0x1:
+              ibos_bytes = get_mem(CAPTURED_STATE['bindless_base'] + ((vals[2] << 32) | vals[1]) * 4, num_unit * 64)
+            else: ibos_bytes = get_mem((vals[2] << 32) | vals[1], num_unit * 16 * 4)
             CAPTURED_STATE['ibos'] = ibos_bytes[:]
             if IOCTL > 1:
               print('texture ibos')
               hexdump(ibos_bytes)
         elif state_block == SB6_CS_TEX:
           if state_type == ST6_SHADER:
-            samplers_bytes = get_mem((vals[2] << 32) | vals[1], num_unit * 4 * 4)
+            if state_src == 0x1:
+              samplers_bytes = get_mem(CAPTURED_STATE['bindless_base'] + ((vals[2] << 32) | vals[1]) * 4, num_unit * 64)
+            else: samplers_bytes = get_mem((vals[2] << 32) | vals[1], num_unit * 4 * 4)
             CAPTURED_STATE['samplers'] = samplers_bytes[:]
             if IOCTL > 1:
               print('texture samplers')
               hexdump(samplers_bytes)
           if state_type == ST6_CONSTANTS:
-            descriptors_bytes = get_mem((vals[2] << 32) | vals[1], 1600)
+            if state_src == 0x1:
+              descriptors_bytes = get_mem(CAPTURED_STATE['bindless_base'] + ((vals[2] << 32) | vals[1]) * 4, num_unit * 64)
+            else: descriptors_bytes = get_mem((vals[2] << 32) | vals[1], 1600)
             CAPTURED_STATE['descriptors'] = descriptors_bytes[:]
             if IOCTL > 1:
               print('texture descriptors')
               hexdump(descriptors_bytes)
-
       elif ops[opcode] == "CP_REG_TO_MEM":
         reg, cnt, b64, accum = vals[0] & 0x3FFFF, (vals[0] >> 18) & 0xFFF, (vals[0] >> 30) & 0x1, (vals[0] >> 31) & 0x1
         dest = vals[1] | (vals[2] << 32)
@@ -152,6 +157,10 @@ def parse_cmd_buf(dat):
         if IOCTL > 0:
           print(f'THREADSIZE-{(vals[0] >> 20)&0x1}\nEARLYPREAMBLE-{(vals[0] >> 23) & 0x1}\nMERGEDREGS-{(vals[0] >> 3) & 0x1}\nTHREADMODE-{vals[0] & 0x1}\nHALFREGFOOTPRINT-{(vals[0] >> 1) & 0x3f}\nFULLREGFOOTPRINT-{(vals[0] >> 7) & 0x3f}\nBRANCHSTACK-{(vals[0] >> 14) & 0x3f}\n')
           print(f'SP_CS_UNKNOWN_A9B1-{vals[1]}\nSP_CS_BRANCH_COND-{vals[2]}\nSP_CS_OBJ_FIRST_EXEC_OFFSET-{vals[3]}\nSP_CS_OBJ_START-{vals[4] | (vals[5] << 32)}\nSP_CS_PVT_MEM_PARAM-{vals[6]}\nSP_CS_PVT_MEM_ADDR-{vals[7] | (vals[8] << 32)}\nSP_CS_PVT_MEM_SIZE-{vals[9]}')
+      if offset == 0xa9e8:
+        CAPTURED_STATE['bindless_base'] = (vals[0] | (vals[1] << 32)) & ~0b11
+        # print(hex(CAPTURED_STATE['bindless_base']))
+        # hexdump(get_mem(CAPTURED_STATE['bindless_base'], 0x200))
       if offset == 0xb180:
         if IOCTL > 0:
           print('border color offset', hex(vals[1] << 32 | vals[0]))
@@ -171,8 +180,8 @@ def ioctl(fd, request, argp):
     name, stype = nrs[nr]
     s = get_struct(argp, stype)
     if IOCTL > 0: print(f"{ret:2d} = {name:40s}", ' '.join(format_struct(s)))
-    if name == "IOCTL_KGSL_GPUOBJ_INFO": pass
-      # mmaped[s.gpuaddr] = mmap.mmap(fd, s.size, offset=s.id*0x1000)
+    if name == "IOCTL_KGSL_GPUOBJ_INFO":
+      mmaped[s.gpuaddr] = mmap.mmap(fd, s.size, offset=s.id*0x1000)
     if name == "IOCTL_KGSL_GPU_COMMAND":
       for i in range(s.numcmds):
         cmd = get_struct(s.cmdlist+ctypes.sizeof(msm_kgsl.struct_kgsl_command_object)*i, msm_kgsl.struct_kgsl_command_object)

extra/remu/src/thread.rs

@@ -882,6 +882,11 @@ impl<'a> Thread<'a> {
                             let s1 = sign_ext((s1 & 0xffffff) as u64, 24) as i32;
                             (s0 * s1) as u32
                         }
+                        10 => {
+                            let s0 = sign_ext((s0 & 0xffffff) as u64, 24) as i64;
+                            let s1 = sign_ext((s1 & 0xffffff) as u64, 24) as i64;
+                            ((s0 * s1) >> 32) as u32
+                        }
                         17 | 18 | 26 => {
                             let (s0, s1) = (s0 as i32, s1 as i32);
                             (match op {
@@ -930,7 +935,7 @@ impl<'a> Thread<'a> {
 
             let op = ((instr >> 16) & 0x3ff) as u32;
             match op {
-                764 | 765 | 288 | 289 | 290 | 766 | 767 | 768 | 769 => {
+                764 | 765 | 288 | 289 | 290 | 766 | 767 | 768 | 769 | 770 => {
                     let vdst = (instr & 0xff) as usize;
                     let sdst = ((instr >> 8) & 0x7f) as usize;
                     let f = |i: u32| -> usize { ((instr >> i) & 0x1ff) as usize };
@@ -996,6 +1001,10 @@ impl<'a> Thread<'a> {
                                     let ret = s0.wrapping_sub(s1);
                                     (ret as u32, s1 > s0)
                                 }
+                                770 => {
+                                    let ret = s1.wrapping_sub(s0);
+                                    (ret as u32, s0 > s1)
+                                }
                                 _ => todo_instr!(instruction)?,
                             };
                             if self.exec.read() {

extra/remu/test/hwtest.py

@@ -1,98 +1,32 @@
 import numpy as np
 import unittest
 import subprocess, struct, math
-from typing import cast
-from tinygrad.runtime.ops_amd import AMDProgram, AMDDevice
-from tinygrad import Tensor, dtypes, Device
-from tinygrad.helpers import diskcache, OSX, getenv
+from tinygrad import Tensor, dtypes, Device, UOp
+from tinygrad.helpers import getenv
+from tinygrad.runtime.support.compiler_amd import amdgpu_disassemble
+from tinygrad.renderer import ProgramSpec
+from tinygrad.engine.realize import CompiledRunner
 
-@diskcache
-def assemble(code:str) -> bytes:
-  try:
-    LLVM_MC = "llvm-mc" if OSX else "/opt/rocm/llvm/bin/llvm-mc"
-    return subprocess.run([LLVM_MC, "--arch=amdgcn", "--mcpu=gfx1100", "--triple=amdgcn-amd-amdhsa", "-filetype=obj", "-o", "-"],
-                           input=code.encode("utf-8"), stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True).stdout
-  except subprocess.CalledProcessError as e:
-    print("stderr:")
-    print(e.stderr.decode())
-    raise
-
-# copied from extra/rdna
-def get_prg(code:str, v_cnt:int, s_cnt:int):
-  function_name = "test"
-  metadata = f"""
-amdhsa.kernels:
-- .args:
-  - .address_space: global
-    .name: buf_0
-    .offset: 0
-    .size: 8
-    .type_name: unsigned int*
-    .value_kind: global_buffer
-  .group_segment_fixed_size: 0
-  .kernarg_segment_align: 8
-  .kernarg_segment_size: 8
-  .language: OpenCL C
-  .language_version:
-  - 1
-  - 2
-  .max_flat_workgroup_size: 256
-  .name: test
-  .private_segment_fixed_size: 0
-  .sgpr_count: {s_cnt}
-  .sgpr_spill_count: 0
-  .symbol: test.kd
-  .uses_dynamic_stack: false
-  .vgpr_count: {v_cnt}
-  .vgpr_spill_count: 0
-  .wavefront_size: 32
-amdhsa.target: amdgcn-amd-amdhsa--gfx1100
-amdhsa.version:
-- 1
-- 2
-  """
-  boilerplate_start = f"""
-  .rodata
-  .global {function_name}.kd
-  .type {function_name}.kd,STT_OBJECT
-  .align 0x10
-  .amdhsa_kernel {function_name}"""
-  kernel_desc = {
-    '.amdhsa_group_segment_fixed_size': 0, '.amdhsa_private_segment_fixed_size': 0, '.amdhsa_kernarg_size': 0,
-    '.amdhsa_next_free_vgpr': v_cnt,   # this matters!
-    '.amdhsa_reserve_vcc': 0, '.amdhsa_reserve_xnack_mask': 0,
-    '.amdhsa_next_free_sgpr': s_cnt,
-    '.amdhsa_float_round_mode_32': 0, '.amdhsa_float_round_mode_16_64': 0, '.amdhsa_float_denorm_mode_32': 3, '.amdhsa_float_denorm_mode_16_64': 3,
-    '.amdhsa_dx10_clamp': 1, '.amdhsa_ieee_mode': 1, '.amdhsa_fp16_overflow': 0,
-    '.amdhsa_workgroup_processor_mode': 1, '.amdhsa_memory_ordered': 1, '.amdhsa_forward_progress': 0, '.amdhsa_enable_private_segment': 0,
-    '.amdhsa_system_sgpr_workgroup_id_x': 1, '.amdhsa_system_sgpr_workgroup_id_y': 1, '.amdhsa_system_sgpr_workgroup_id_z': 1,
-    '.amdhsa_system_sgpr_workgroup_info': 0, '.amdhsa_system_vgpr_workitem_id': 2, # is amdhsa_system_vgpr_workitem_id real?
-    '.amdhsa_exception_fp_ieee_invalid_op': 0, '.amdhsa_exception_fp_denorm_src': 0,
-    '.amdhsa_exception_fp_ieee_div_zero': 0, '.amdhsa_exception_fp_ieee_overflow': 0, '.amdhsa_exception_fp_ieee_underflow': 0,
-    '.amdhsa_exception_fp_ieee_inexact': 0, '.amdhsa_exception_int_div_zero': 0,
-    '.amdhsa_user_sgpr_dispatch_ptr': 0, '.amdhsa_user_sgpr_queue_ptr': 0, '.amdhsa_user_sgpr_kernarg_segment_ptr': 1,
-    '.amdhsa_user_sgpr_dispatch_id': 0, '.amdhsa_user_sgpr_private_segment_size': 0, '.amdhsa_wavefront_size32': 1, '.amdhsa_uses_dynamic_stack': 0}
-  code_start = f""".end_amdhsa_kernel
-  .text
-  .global {function_name}
-  .type {function_name},@function
-  .p2align 8
-  {function_name}:
-  """
-  ret = ".amdgpu_metadata\n" + metadata + ".end_amdgpu_metadata" + boilerplate_start + "\n" + '\n'.join("%s %d" % x for x in kernel_desc.items()) \
-      + "\n" + code_start + code + f"\n.size {function_name}, .-{function_name}"
-  return AMDProgram(cast(AMDDevice, Device["AMD"]), function_name, assemble(ret))
-
-def get_output(s:str, n_threads:int=1):
-  assert n_threads <= 32
-  code = "\n".join(["s_load_b64 s[0:1], s[0:1], null", "v_lshlrev_b32_e32 v0, 2, v0", s,
-    "s_waitcnt 0",
-    "global_store_b32 v0, v1, s[0:1]",
-    "s_nop 0", "s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)", "s_endpgm"])
-  test = Tensor.zeros((n_threads,), dtype=dtypes.uint32).contiguous().realize().uop.buffer
-  prg = get_prg(code, 32, 32)
-  prg(test._buf, global_size=(1, 1, 1), local_size=(n_threads, 1, 1), wait=True)
-  return test.numpy()
+def get_output(asm:str, n_threads:int=1):
+  input_asm = "\n".join([ln if ln.strip().startswith('asm volatile') else f'asm volatile("{ln.strip().lstrip()}" : "+v"(a), "+v"(b));'
+                         for ln in asm.strip().splitlines() if ln.strip()])
+  src = f"""
+  typedef long unsigned int size_t;
+  extern "C" __attribute__((device, const)) size_t __ockl_get_local_id(unsigned int);
+  extern "C" __attribute__((global)) void __attribute__((amdgpu_flat_work_group_size(1, {n_threads}))) test(unsigned int* data0_1) {{
+    int l = __ockl_get_local_id(0);
+    unsigned a = 0, b = 0, c = 0;
+    {input_asm}
+    unsigned res;
+    asm volatile("v_mov_b32 %0, %1" : "=v"(res) : "v"(a));
+    *(data0_1+l) = res;
+  }}"""
+  t = Tensor.zeros(n_threads, dtype=dtypes.uint32).contiguous().realize()
+  prg = ProgramSpec("test", src, Device.DEFAULT, UOp.sink(t), global_size=[1, 1, 1], local_size=[n_threads, 1, 1])
+  car = CompiledRunner(prg)
+  if getenv("PRINT_ASM"): amdgpu_disassemble(car.lib)
+  car([t.uop.buffer], {}, wait=True)
+  return t.numpy()
 
 def f16_to_bits(x:float) -> int: return struct.unpack('<H', struct.pack('<e', x))[0]
 def f32_from_bits(x:int) -> float: return struct.unpack('<f', struct.pack('<I', x))[0]
@@ -105,54 +39,57 @@ class TestHW(unittest.TestCase):
 
   def test_simple(self):
     out = get_output("""
-    v_mov_b32_e32 v10 42
-    v_mov_b32_e32 v1 v10
-    """, n_threads=2)
+    v_mov_b32_e32 %1 42
+    v_mov_b32_e32 %2 %1
+    """)[0]
     np.testing.assert_equal(out, 42)
 
   def test_exec_mov(self):
     out = get_output("""
-    v_mov_b32_e32 v10 42
+    v_mov_b32_e32 %1 42
     s_mov_b32_e32 exec_lo 0b10
-    v_mov_b32_e32 v10 10
+    v_mov_b32_e32 %1 10
     s_mov_b32_e32 exec_lo 0b11
-    v_mov_b32_e32 v1 v10
+    v_mov_b32_e32 %2 %1
     """, n_threads=2)
     np.testing.assert_equal(out, [42, 10])
 
   def test_exec_cmp_vopc(self):
     out = get_output("""
     s_mov_b32 vcc_lo 0 // reset vcc
-    v_mov_b32_e32 v10 42
-    v_mov_b32_e32 v11 10
+    v_mov_b32_e32 %1 42
+    v_mov_b32_e32 %2 10
     s_mov_b32_e32 exec_lo 0b01
-    v_cmp_ne_u32 v10 v11
+    v_cmp_ne_u32 %1 %2
     s_mov_b32_e32 exec_lo 0b11
-    v_mov_b32_e32 v1 vcc_lo
+    v_mov_b32_e32 %2 vcc_lo
     """, n_threads=2)
     np.testing.assert_equal(out, 0b01)
 
   def test_exec_cmpx_vop3(self):
     out = get_output("""
-    v_mov_b32_e32 v10 42
-    v_mov_b32_e32 v11 10
+    s_mov_b32_e32 exec_lo 0b11
+    v_mov_b32_e32 %1 42
+    v_mov_b32_e32 %2 10
     s_mov_b32_e32 exec_lo 0b01
-    v_cmpx_ne_u32 v10 v11
+    v_cmpx_ne_u32 %1 %2
     s_mov_b32_e32 s10 exec_lo
     s_mov_b32_e32 exec_lo 0b11
-    v_mov_b32_e32 v1 s10
-    """, n_threads=2)
-    np.testing.assert_equal(out, 0b01)
+    v_mov_b32_e32 %2 s10
+    """, n_threads=2)[0]
+    np.testing.assert_equal(out & 0b11, 0b01)
 
   def test_fmac_vop3_modifier(self):
     init_state = f"""
-    v_mov_b32_e32 v10 {f16_to_bits(4.0)}
-    v_mov_b32_e32 v11 {f16_to_bits(3.0)}
-    v_mov_b32_e32 v1 {f16_to_bits(2.0)}
+    asm volatile("v_mov_b32_e32 %1, {f16_to_bits(4.0)}" : "+v"(a));
+    asm volatile("v_mov_b32_e32 %1, {f16_to_bits(3.0)}" : "+v"(b));
+    asm volatile("v_mov_b32_e32 %1, {f16_to_bits(2.0)}" : "+v"(c));
     """
-    self.assertEqual(get_output(init_state+"\n"+"v_fmac_f16_e64 v1 v11 v10"), f16_to_bits(14.))
-    self.assertEqual(get_output(init_state+"\n"+"v_fmac_f16_e64 v1 -v11 v10"), f16_to_bits(-10.))
-    self.assertEqual(get_output(init_state+"\n"+"v_fmac_f16_e64 v1 -v11 -v10"), f16_to_bits(14.))
+    mov = """asm volatile("v_mov_b32_e32 %1, %2" : "+v"(c), "+v"(a));"""
+    def fmac(a, b, c): return f"""asm volatile("v_fmac_f16_e64 {c}, {a}, {b}" : "+v"(c) : "v"(a), "v"(b));"""+"\n"+mov
+    self.assertEqual(get_output(init_state+"\n"+fmac("%1", "%2", "%3")), f16_to_bits(14.))
+    self.assertEqual(get_output(init_state+"\n"+fmac("%1", "-%2", "%3")), f16_to_bits(-10.))
+    self.assertEqual(get_output(init_state+"\n"+fmac("-%1", "-%2", "%3")), f16_to_bits(14.))
 
   def test_s_abs_i32(self):
     def s_abs_i32(x, y, dst="s10", scc=0):
@@ -160,7 +97,7 @@ class TestHW(unittest.TestCase):
         self.assertEqual(get_output(f"""
         s_mov_b32_e32 {dst} {x}
         s_abs_i32 {dst} {dst}
-        v_mov_b32_e32 v1 {reg}
+        v_mov_b32_e32 %2 {reg}
         """)[0], val)
     s_abs_i32(0x00000001, 0x00000001, scc=1)
     s_abs_i32(0x7fffffff, 0x7fffffff, scc=1)
@@ -173,8 +110,8 @@ class TestHW(unittest.TestCase):
   def test_v_rcp_f32_neg_vop3(self):
     def v_neg_rcp_f32(x:float, y:float):
       out = get_output(f"""
-      v_mov_b32_e32 v1 {f32_to_bits(x)}
-      v_rcp_f32_e64 v1, -v1
+      v_mov_b32_e32 %2 {f32_to_bits(x)}
+      v_rcp_f32_e64 %2, -%2
       """)[0]
       assert out == f32_to_bits(y), f"{f32_from_bits(out)} != {y} / {out} != {f32_to_bits(y)}"
     v_neg_rcp_f32(math.inf, -0.0)
@@ -186,10 +123,11 @@ class TestHW(unittest.TestCase):
 
   def test_v_cndmask_b32_neg(self):
     def v_neg(x:int|float, y:float):
+      # always pick -v1
       out = get_output(f"""
-      v_mov_b32_e32 v1 {f32_to_bits(x)}
-      s_mov_b32_e32 s10 1 // always pick -v1
-      v_cndmask_b32 v1, v1, -v1 s10
+      v_mov_b32_e32 %2 {f32_to_bits(x)}
+      s_mov_b32_e32 s10 1
+      v_cndmask_b32 %2, %2, -%2 s10
       """)[0]
       assert out == f32_to_bits(y), f"{f32_from_bits(out)} != {y} / {out} != {f32_to_bits(y)}"
     v_neg(-0.0, 0.0)
@@ -198,5 +136,12 @@ class TestHW(unittest.TestCase):
     v_neg(math.inf, -math.inf)
     v_neg(-math.inf, math.inf)
 
+  def test_v_subrev_wrap(self):
+    out = get_output("""
+    v_dual_mov_b32 %1, 0xffffffff :: v_dual_mov_b32 %2, 0x0
+    v_subrev_co_u32 %2, vcc_lo, %2, %1
+    """)[0]
+    self.assertEqual(out, 0xffff_ffff)
+
 if __name__ == "__main__":
   unittest.main()

extra/sqtt/roc.py

@@ -27,14 +27,18 @@ class _ROCParseCtx:
         self.disasms[prog.base + addr] = info
         self.addr2prg[prog.base + addr] = prog
 
-  def next_sqtt(self): return next(self.sqtt_evs, None)
+  def next_sqtt(self):
+    x = next(self.sqtt_evs, None)
+    self.active_se = x.se if x is not None else None
+    return x
+
   def find_program(self, addr): return self.addr2prg[addr]
 
   def on_occupancy_ev(self, ev):
-    if DEBUG >= 4: print("OCC", ev.time, ev.cu, ev.simd, ev.wave_id, ev.start)
+    if DEBUG >= 4: print("OCC", ev.time, self.active_se, ev.cu, ev.simd, ev.wave_id, ev.start)
 
   def on_wave_ev(self, ev):
-    if DEBUG >= 4: print("WAVE", ev.wave_id, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
+    if DEBUG >= 4: print("WAVE", ev.wave_id, self.active_se, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
 
     asm = {}
     for j in range(ev.instructions_size):

extra/thunder/cuda/include/ops/group/group.cuh

@@ -46,9 +46,9 @@ __device__ static inline void arrive(int id) {
 #include "memory/memory.cuh"
 #include "shared/shared.cuh"
 #include "register/register.cuh"
+#include "mma/mma.cuh"
 
 #ifdef KITTENS_HOPPER
-#include "mma/mma.cuh"
 
 template<int n_reg> __device__ static inline void increase_registers() {
     static_assert(n_reg % 8 == 0, "n_reg must be a multiple of 8");
@@ -93,4 +93,4 @@ __device__ static inline void sync() {
 using warp = group<1>;      // scope used by most pre-Hopper GPUs, and also for most register operations.
 using warpgroup = group<4>; // special scope commonly used by Hopper and later.
 
-}
+}

extra/thunder/cuda/include/types/global/gl.cuh

@@ -65,8 +65,8 @@ template<typename _T, int _axis=-9999, bool _swizzle_flag=true> struct descripto
 namespace detail {
 template<typename... Args>
 struct descriptor_dict {
-    __host__ descriptor_dict() {}
-    template<typename T> __host__ descriptor_dict(T _, int b, int d, int r, int c) {}
+    __host__ __device__ descriptor_dict() {}
+    template<typename T> __host__ __device__ descriptor_dict(T _, int b, int d, int r, int c) {}
     __host__ __device__ descriptor_dict(const descriptor_dict &other) {}
 #ifdef KITTENS_HOPPER
     template<typename T, int U> __device__ const CUtensorMap* get() const {
@@ -85,8 +85,8 @@ struct descriptor_dict<_T, Args...> {
     using DESC = kittens::tma::descriptor<_T>; // copy or initialize with a default value
     CUtensorMap tma_desc;
     descriptor_dict<Args...> other_descs;
-    __host__ descriptor_dict() {}
-    __host__ descriptor_dict(typename DESC::T::dtype *data, int b, int d, int r, int c): other_descs(data, b, d, r, c) {
+    __host__ __device__ descriptor_dict() {}
+    __host__ __device__ descriptor_dict(typename DESC::T::dtype *data, int b, int d, int r, int c): other_descs(data, b, d, r, c) {
         kittens::detail::tma::create_tensor_map<typename DESC::T, DESC::axis, DESC::swizzle_flag>(&tma_desc, data, b, d, r, c);
     }
     __host__ __device__ inline descriptor_dict(const descriptor_dict &other) :
@@ -135,7 +135,7 @@ struct gl {
 
     detail::descriptor_dict<TMA_Types...> tma_descs;
 
-    __host__ inline gl(T *_data,
+    __host__ __device__ inline gl(T *_data,
                         ducks::gl::make_arg_t<b> _batch,
                         ducks::gl::make_arg_t<d> _depth,
                         ducks::gl::make_arg_t<r> _rows,

extra/thunder/cuda/include/types/global/tma.cuh

@@ -425,4 +425,4 @@ __host__ static inline CUtensorMap* allocate_and_create_tensor_map(const typenam
 
 } // namespace tma
 } // namespace detail
-} // namespace kittens
+} // namespace kittens

extra/thunder/cuda/matmul.cu

@@ -0,0 +1,45 @@
+// https://github.com/HazyResearch/ThunderKittens/blob/main/kernels/matmul/educational/level_04.cu
+#include "kittens.cuh"
+using namespace kittens;
+
+constexpr int g_N = 8192;
+constexpr int BLOCK_SIZE = 32;
+#define NUM_WORKERS (1)
+#define NUM_THREADS (NUM_WORKERS*kittens::WARP_THREADS)
+
+using sub_tile = st_bf<BLOCK_SIZE,BLOCK_SIZE>;
+using tile_gl =  gl<bf16,  1, 1, g_N, g_N, sub_tile>;
+
+__global__ void kernel(bf16 *c_ptr, bf16 *a_ptr, bf16 *b_ptr) {
+  tile_gl g_C{c_ptr, nullptr, nullptr, nullptr, nullptr};
+  tile_gl g_A{a_ptr, nullptr, nullptr, nullptr, nullptr};
+  tile_gl g_B{b_ptr, nullptr, nullptr, nullptr, nullptr};
+
+  extern __shared__ alignment_dummy __shm[];
+  shared_allocator al((int*)&__shm[0]);
+  st_bf<BLOCK_SIZE,BLOCK_SIZE> &As = al.allocate<st_bf<BLOCK_SIZE,BLOCK_SIZE>>();
+  st_bf<BLOCK_SIZE,BLOCK_SIZE> &Bs = al.allocate<st_bf<BLOCK_SIZE,BLOCK_SIZE>>();
+
+  rt_bf<BLOCK_SIZE,BLOCK_SIZE> A_reg;
+  rt_bf<BLOCK_SIZE,BLOCK_SIZE> B_reg;
+  rt_bf<BLOCK_SIZE,BLOCK_SIZE, ducks::rt_layout::col> B_reg_col;
+  rt_fl<BLOCK_SIZE,BLOCK_SIZE> C_accum;
+
+  int col = blockIdx.x;
+  int row = blockIdx.y;
+
+  warp::zero(C_accum);
+  int num_tiles = (g_N + BLOCK_SIZE - 1) / BLOCK_SIZE;
+  for (int tile = 0; tile < num_tiles; ++tile) {
+    warp::load(As, g_A, {0, 0, row, tile});
+    warp::load(Bs, g_B, {0, 0, tile, col});
+    __syncthreads();
+    warp::load(A_reg, As);
+    warp::load(B_reg, Bs);
+    warp::swap_layout(B_reg_col, B_reg);
+    __syncthreads();
+    warp::mma_AB(C_accum, A_reg, B_reg_col, C_accum);
+    __syncthreads();
+  }
+  warp::store(g_C, C_accum, {0, 0, row, col});
+}

extra/thunder/cuda/matmul.py

@@ -0,0 +1,37 @@
+import pathlib
+from tinygrad import Device, Tensor
+from tinygrad.helpers import Context
+from tinygrad.runtime.support.compiler_cuda import pretty_ptx, NVCCCompiler
+
+if __name__ == "__main__":
+  code = (pathlib.Path(__file__).parent / "matmul.cu").read_text()
+  device = Device["CUDA"]
+  kitten_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "--expt-relaxed-constexpr"]
+  lib = NVCCCompiler(device.compiler.arch, kitten_args).compile(code)
+  kernel_name = lib.decode().split(".globl\t")[1].split("\n")[0]
+  print("kernel name", kernel_name)
+  print(pretty_ptx(lib.decode()))
+
+  prg = device.runtime(kernel_name, lib)
+  prg.smem = 10000
+
+  N = 8192
+  a = Tensor.randn(N, N, device='CUDA', dtype="bfloat16")
+  b = Tensor.randn(N, N, device='CUDA', dtype="bfloat16")
+  c = Tensor.empty(N, N, device='CUDA', dtype="bfloat16")
+  Tensor.realize(a, b, c)
+
+  BLOCK_SIZE = 32
+
+  gsz = (N // BLOCK_SIZE, N // BLOCK_SIZE, 1)
+  for _ in range(5):
+    et = prg(c.uop.buffer.ensure_allocated()._buf, a.uop.buffer._buf, b.uop.buffer._buf,
+             global_size=gsz, local_size=(32,1,1), wait=True)
+    print(f"{N*N*N*2/(et*1e9):2f} GFLOPS")
+
+  for _ in range(5):
+    with Context(DEBUG=2):
+      ref = (a@b).realize()
+
+  ref, c = ref.float(), c.float()
+  print((ref-c).mean().item(), (ref-c).max().item())

extra/torch_muon.py

@@ -1,75 +0,0 @@
-import torch
-
-#credit to KellerJordan at https://github.com/KellerJordan/Muon/tree/master
-#some changes: classic momentum instead of weighting gradient
-#added ns_steps, ns_params, nesterov as hyperparams
-def zeropower_via_newtonschulz5(G:torch.tensor, steps:int, params:tuple[int, ...]):
-  """
-  Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
-  quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
-  of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
-  zero even beyond the point where the iteration no longer converges all the way to one everywhere
-  on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
-  where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
-  performance at all relative to UV^T, where USV^T = G is the SVD.
-  """
-  assert G.ndim >= 2 # batched Muon implementation by @scottjmaddox, and put into practice in the record by @YouJiacheng
-
-  a, b, c = params
-  X = G
-  if G.size(-2) > G.size(-1):
-    X = X.mT
-
-  # Ensure spectral norm is at most 1
-  X = X / (X.norm(dim=(-2, -1), keepdim=True) + 1e-7)
-  # Perform the NS iterations
-  for _ in range(steps):
-    A = X @ X.mT
-    B = b * A + c * A @ A # quintic computation strategy adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
-    X = a * X + B @ X
-
-  if G.size(-2) > G.size(-1):
-    X = X.mT
-
-  return X
-
-def muon_update(grad, momentum, beta=0.95, ns_steps=5, ns_params=(3.4445, -4.7750,  2.0315), nesterov=True):
-  if beta:
-    momentum.mul_(beta).add_(grad)
-    update = grad.add(momentum,alpha=beta) if nesterov else momentum
-  else: update = grad
-  if update.ndim == 4: # for the case of conv filters
-    update = update.view(len(update), -1)
-  update = zeropower_via_newtonschulz5(update, steps=ns_steps, params=ns_params)
-  return update
-
-class SingleDeviceMuon(torch.optim.Optimizer):
-  """
-  Muon variant for usage in non-distributed settings.
-  """
-  def __init__(self, params, lr=0.02, weight_decay=0.0, momentum=0.95, ns_steps=5, ns_params=(3.4445, -4.7750,  2.0315), nesterov=True):
-    defaults = dict(lr=lr, weight_decay=weight_decay, momentum=momentum, ns_steps=ns_steps, ns_params=ns_params, nesterov=nesterov)
-    super().__init__(params, defaults)
-
-  @torch.no_grad()
-  def step(self, closure=None):
-
-    loss = None
-    if closure is not None:
-      with torch.enable_grad():
-        loss = closure()
-
-    for group in self.param_groups:
-      for p in group["params"]:
-        if p.grad is None:
-          p.grad = torch.zeros_like(p)  # Force synchronization
-        state = self.state[p]
-        if len(state) == 0:
-          state["momentum_buffer"] = torch.zeros_like(p)
-        update = muon_update(p.grad, state["momentum_buffer"], beta=group["momentum"], ns_steps=group["ns_steps"],
-                             ns_params=group["ns_params"], nesterov=group["nesterov"])
-        p.mul_(1.0 - group["lr"] * group["weight_decay"])
-
-        p.add_(update.reshape(p.shape), alpha=-group["lr"])
-
-    return loss

extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension.xcodeproj/project.pbxproj

@@ -321,7 +321,7 @@
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_ENTITLEMENTS = macOS/macOS.entitlements;
 				CODE_SIGN_IDENTITY = "-";
-				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "Apple Development";
+				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "-";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
 				CURRENT_PROJECT_VERSION = 1;
@@ -357,7 +357,7 @@
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_ENTITLEMENTS = macOS/macOS.entitlements;
 				CODE_SIGN_IDENTITY = "-";
-				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "Apple Development";
+				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "-";
 				CODE_SIGN_STYLE = Automatic;
 				COMBINE_HIDPI_IMAGES = YES;
 				CURRENT_PROJECT_VERSION = 1;
@@ -502,7 +502,7 @@
 			buildSettings = {
 				AD_HOC_CODE_SIGNING_ALLOWED = YES;
 				CODE_SIGN_ENTITLEMENTS = TinyGPUDriverExtension/TinyGPUDriver.entitlements;
-				CODE_SIGN_IDENTITY = "Apple Development";
+				CODE_SIGN_IDENTITY = "-";
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
 				DEVELOPMENT_TEAM = 9YG3G8543N;
@@ -530,7 +530,7 @@
 			buildSettings = {
 				AD_HOC_CODE_SIGNING_ALLOWED = YES;
 				CODE_SIGN_ENTITLEMENTS = TinyGPUDriverExtension/TinyGPUDriver.entitlements;
-				CODE_SIGN_IDENTITY = "Apple Development";
+				CODE_SIGN_IDENTITY = "-";
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
 				DEVELOPMENT_TEAM = 9YG3G8543N;

mkdocs.yml

@@ -25,8 +25,6 @@ nav:
     - Layout: developer/layout.md
     - Speed: developer/speed.md
     - UOp: developer/uop.md
-    - Grouper:
-      - developer/kernelize.md
     - Runtime:
       - developer/runtime.md
       - HCQ: developer/hcq.md

setup.py

@@ -9,10 +9,11 @@ with open(directory / 'README.md', encoding='utf-8') as f:
 
 testing_minimal = [
   "numpy",
-  "torch==2.8.0",
+  "torch==2.9.0",
   "pytest",
   "pytest-xdist",
   "pytest-timeout",
+  "pytest-split",
   "hypothesis",
   "z3-solver",
 ]

sz.py

@@ -54,6 +54,8 @@ def gen_diff(table_old, table_new):
 
 def display_diff(diff): return "+"+str(diff) if diff > 0 else str(diff)
 
+NONCORE_DIRS = {"tinygrad/apps", "tinygrad/nn", "tinygrad/renderer", "tinygrad/runtime", "tinygrad/viz"}
+
 if __name__ == "__main__":
   if len(sys.argv) == 3:
     headers = ["Name", "Lines", "Diff", "Tokens/Line", "Diff"]
@@ -76,9 +78,12 @@ if __name__ == "__main__":
     else:
       print(tabulate([headers] + sorted(table, key=lambda x: -x[1]), headers="firstrow", floatfmt=".1f")+"\n")
       groups = sorted([('/'.join(x[0].rsplit("/", 1)[0].split("/")[0:2]), x[1], x[2]) for x in table])
+      dir_sizes = {}
       for dir_name, group in itertools.groupby(groups, key=lambda x:x[0]):
-        print(f"{dir_name:30s} : {sum([x[1] for x in group]):6d}")
+        dir_sizes[dir_name] = sum([x[1] for x in group])
+        print(f"{dir_name:30s} : {dir_sizes[dir_name]:6d}")
+      print(f"\n core line count: {sum([v for k,v in dir_sizes.items() if k not in NONCORE_DIRS])}")
       total_lines = sum([x[1] for x in table])
-      print(f"\ntotal line count: {total_lines}")
+      print(f"total line count: {total_lines}")
       max_line_count = int(os.getenv("MAX_LINE_COUNT", "-1"))
       assert max_line_count == -1 or total_lines <= max_line_count, f"OVER {max_line_count} LINES"

test/external/external_benchmark_openpilot.py

@@ -1,63 +0,0 @@
-import time, sys, hashlib
-from pathlib import Path
-from tinygrad.nn.onnx import OnnxRunner
-from tinygrad import Tensor, dtypes, TinyJit
-from tinygrad.helpers import IMAGE, GlobalCounters, fetch, colored, getenv, trange
-import numpy as np
-from extra.bench_log import BenchEvent, WallTimeEvent
-
-OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx"
-
-if __name__ == "__main__":
-  run_onnx = OnnxRunner(fetch(OPENPILOT_MODEL))
-
-  Tensor.manual_seed(100)
-  input_shapes = {name: spec.shape for name, spec in run_onnx.graph_inputs.items()}
-  input_types = {name: spec.dtype for name, spec in run_onnx.graph_inputs.items()}
-  new_inputs = {k:Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize() for k,shp in input_shapes.items()}
-  new_inputs_junk = {k:Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize() for k,shp in input_shapes.items()}
-  new_inputs_junk_numpy = {k:v.numpy() for k,v in new_inputs_junk.items()}
-
-  # benchmark
-  for _ in range(5):
-    GlobalCounters.reset()
-    st = time.perf_counter_ns()
-    ret = next(iter(run_onnx(new_inputs_junk).values())).cast(dtypes.float32).numpy()
-    print(f"unjitted: {(time.perf_counter_ns() - st)*1e-6:7.4f} ms")
-
-  # NOTE: the inputs to a JIT must be first level arguments
-  run_onnx_jit = TinyJit(lambda **kwargs: run_onnx(kwargs), prune=True)
-  step_times = []
-  for _ in range(20):
-    GlobalCounters.reset()
-    st = time.perf_counter_ns()
-    with WallTimeEvent(BenchEvent.STEP):
-      # Need to cast non-image inputs from numpy, this is only realistic way to run model
-      inputs = {**{k:v for k,v in new_inputs_junk.items() if 'img' in k},
-                **{k:Tensor(v) for k,v in new_inputs_junk_numpy.items() if 'img' not in k}}
-      ret = next(iter(run_onnx_jit(**inputs).values())).cast(dtypes.float32).numpy()
-      step_times.append(t:=(time.perf_counter_ns() - st)*1e-6)
-    print(f"jitted:  {t:7.4f} ms")
-
-  suffix = ""
-  if IMAGE.value < 2: suffix += f"_image{IMAGE.value}" # image=2 has no suffix for compatibility
-  if getenv("FLOAT16") == 1: suffix += "_float16"
-  path = Path(__file__).parent / "openpilot" / f"{hashlib.md5(OPENPILOT_MODEL.encode()).hexdigest()}{suffix}.npy"
-
-  # validate if we have records
-  tinygrad_out = next(iter(run_onnx_jit(**new_inputs).values())).cast(dtypes.float32).numpy()
-  if getenv("SAVE_OUTPUT"):
-    np.save(path, tinygrad_out)
-    print(f"saved output to {path}!")
-  elif getenv("FUZZ") and path.exists():
-    known_good_out = np.load(path)
-    for _ in trange(1000):
-      ret = next(iter(run_onnx_jit(**new_inputs).values())).cast(dtypes.float32).numpy()
-      np.testing.assert_allclose(known_good_out, ret, atol=1e-2, rtol=1e-2)
-    print(colored("fuzz validated!", "green"))
-  elif path.exists():
-    known_good_out = np.load(path)
-    np.testing.assert_allclose(known_good_out, tinygrad_out, atol=1e-2, rtol=1e-2)
-    print(colored("outputs validated!", "green"))
-  else:
-    print(colored("skipping validation", "yellow"))

test/external/external_benchmark_schedule.py

@@ -2,8 +2,9 @@ from extra.models.resnet import ResNet50
 from tinygrad import Tensor, nn, Device
 from tinygrad.helpers import Profiling, Timing, getenv
 from tinygrad.uop.ops import Ops
-from tinygrad.codegen import get_rewrites_for_renderer, apply_rewrites, rewrites_for_linearizer
-from tinygrad.uop.spec import type_verify
+from tinygrad.codegen import full_rewrite_to_sink
+from tinygrad.codegen.late.linearizer import linearize
+from tinygrad.uop.spec import type_verify, program_spec
 
 if __name__ == "__main__":
   mdl = ResNet50()
@@ -28,18 +29,17 @@ if __name__ == "__main__":
         asts = list({x.ast.key:x.ast for x in sched if x.ast.op is Ops.SINK}.values())
         if (restrict_kernel := getenv("RESTRICT_KERNEL", -1)) != -1: asts = asts[restrict_kernel:restrict_kernel+1]
 
-        rewrites = get_rewrites_for_renderer(Device.default.renderer, linearizer=False)
         with Profiling(PROFILE, fn="/tmp/rewrite.prof"):
           with Timing("***** model rewrite in   "):
             rewritten_uops = []
             for u in asts:
-              rewritten_uops.append(apply_rewrites(u, rewrites))
+              rewritten_uops.append(full_rewrite_to_sink(u, ren=Device.default.renderer))
 
         if LINEARIZE:
           with Timing("***** model linearize in "):
             uops_line = []
             for u in rewritten_uops:
-              uops_line.append(apply_rewrites(u, rewrites_for_linearizer))
+              uops_line.append(linearize(u))
           with Timing("***** model verify in    "):
-            for u in uops_line: type_verify(u.arg.lst)
-          print(sum(len(u.arg.lst) for u in uops_line))
+            for u in uops_line: type_verify(u, program_spec)
+          print(sum(len(u) for u in uops_line))

test/external/external_nhwc_conv.py

@@ -0,0 +1,38 @@
+from tinygrad import Tensor, nn, Context, GlobalCounters
+
+if __name__ == "__main__":
+  conv = nn.Conv2d(64, 128, 3)
+  img = Tensor.randn((1,64,128,128))
+  with Context(DEBUG=0, BEAM=0):
+    Tensor.realize(img, conv.weight, conv.bias)
+
+  tst = conv(img).permute(0,2,3,1).realize()
+  print(tst.shape)
+
+  print("NEW")
+  img_perm = img.permute(0,2,3,1).contiguous()
+  print(img_perm.shape)
+  pp = img_perm.permute(0,3,1,2)._pool((3,3)).permute(0,2,3,4,5,1)
+
+  def hwio(pp, conv):
+    pp = pp.unsqueeze(-1)
+    weight = conv.weight.permute(2,3,1,0).contiguous()
+    print(pp.shape, weight.shape, (pp*weight).shape)
+    return (pp * weight).sum([-4,-3, -2])
+
+  def ohwi(pp, conv):
+    pp = pp.unsqueeze(-4)
+    weight = conv.weight.permute(0,2,3,1).contiguous()
+    print(pp.shape, weight.shape, (pp*weight).shape)
+    return (pp * weight).sum([-3,-2,-1])
+
+  for f in [hwio, ohwi]:
+    GlobalCounters.reset()
+    print("\n**************", f.__name__, "**************")
+    out = f(pp, conv)
+    out.realize()
+    print(out.shape)
+
+    with Context(DEBUG=0, BEAM=0):
+      err = (tst-out).square()
+      print(err.mean().item(), err.max().item())

test/external/external_test_onnx_ops.py

@@ -272,6 +272,10 @@ class TestMainOnnxOps(TestOnnxOps):
   def test_qlinearmatmul_2D_int8_float32(self): self._run_qlinearmatmul_test(np.int8, np.float32, 2)
   def test_qlinearmatmul_3D_int8_float32(self): self._run_qlinearmatmul_test(np.int8, np.float32, 3)
 
+  def test_reduce_l2_half(self):
+    inputs = {"data": np.random.randn(1, 1, 32, 32, 32).astype(np.half)*100}
+    self.helper_test_single_op("ReduceL2", inputs, {}, ["reduced"])
+
 class TestTrainingOnnxOps(TestOnnxOps):
   # NOTE: ORT doesn't actually support training ops on cpu so we test using functions provided by onnx
   DOMAIN = AI_ONNX_PREVIEW_TRAINING_DOMAIN
@@ -282,11 +286,11 @@ class TestTrainingOnnxOps(TestOnnxOps):
     tiny_out = runner(inps)
     onnx_out = onnx_fxn(**inps, **opts)
     for (nm, t_out), o_out in  zip(tiny_out.items(), onnx_out):
-      np.testing.assert_allclose(t_out.numpy(), o_out, rtol=1e-3, atol=1e-6, err_msg=f"{nm} failed")
+      np.testing.assert_allclose(t_out.numpy(), o_out, rtol=1e-6, atol=1e-6, err_msg=f"{nm} failed")
 
-  def test_adagrad_t_greater_than_zero(self):
+  def test_adagrad_t(self):
     from onnx.backend.test.case.node.adagrad import apply_adagrad
-    for t in [1, 3, 100]:
+    for t in [0, 1, 3, 100]:
       inputs = {
         "r": np.array(0.01, dtype=np.float32),
         "t": np.array(t, dtype=np.int32),
@@ -298,10 +302,10 @@ class TestTrainingOnnxOps(TestOnnxOps):
       outputs = ["X_out", "H_out"]
       self._validate_training("Adagrad", apply_adagrad, inputs, attributes, outputs)
 
-  def test_momentum_t_greater_than_zero(self):
+  def test_momentum(self):
     from onnx.backend.test.case.node.momentum import apply_momentum, apply_nesterov
     for onnx_fxn, mode in ((apply_momentum, "standard"), (apply_nesterov, "nesterov")):
-      for t in [1, 3, 100]:
+      for t in [0, 1, 3, 100]:
         inputs = {
           "r": np.array(0.01, dtype=np.float32),
           "t": np.array(t, dtype=np.int32),
@@ -313,9 +317,9 @@ class TestTrainingOnnxOps(TestOnnxOps):
         outputs = ["X_out", "V_out"]
         self._validate_training("Momentum", onnx_fxn, inputs, attributes, outputs)
 
-  def test_adam_t_greater_than_zero(self):
+  def test_adam(self):
     from onnx.backend.test.case.node.adam import apply_adam
-    for t in [1, 3, 100]:
+    for t in [0, 1, 3, 100]:
       inputs = {
         "r": np.array(0.01, dtype=np.float32),
         "t": np.array(t, dtype=np.int32),
@@ -422,6 +426,7 @@ class TestContribOnnxOps(TestOnnxOps):
         outputs = ["C"]
         self.helper_test_single_op("QLinearAdd", inputs, attributes, outputs, atol=1) # TODO: look into why this is inaccurate
 
+  def test_qlinear_add_round_half_to_even(self):
     with self.subTest(test_case="round_half_to_even"):
       inputs = {
         "A": np.array([1, 1, 1, 1], dtype=np.int8),
@@ -435,7 +440,7 @@ class TestContribOnnxOps(TestOnnxOps):
       }
       attributes = {}
       outputs = ["C"]
-      self.helper_test_single_op("QLinearAdd", inputs, attributes, outputs)
+      self.helper_test_single_op("QLinearAdd", inputs, attributes, outputs, atol=1) # TODO: look into why this is inaccurate
 
   def test_qlinear_mul(self):
     for dtype, zero_point in [(np.uint8, 128), (np.int8, 0)]:
@@ -486,4 +491,4 @@ class TestContribOnnxOps(TestOnnxOps):
         self.helper_test_single_op("QLinearGlobalAveragePool", inputs, attributes, outputs)
 
 if __name__ == "__main__":
-  unittest.main()
+  unittest.main()

test/external/fuzz_fast_idiv.py

@@ -1,7 +1,7 @@
 import random
 import z3
 from tinygrad import dtypes
-from tinygrad.uop.spec import uops_to_z3, z3_cdiv
+from tinygrad.uop.validate import uops_to_z3, z3_cdiv
 from tinygrad.uop.ops import UOp
 from tinygrad.uop.decompositions import fast_idiv
 random.seed(42)

test/external/fuzz_linearizer.py

@@ -207,7 +207,7 @@ def fuzz_linearizer(lin: Kernel, rtol=1e-2, atol=1e-2, opts_list=None):
         if not FUZZ_ALL_ACTIONS and test_lin.applied_opts: print(f"applied opts: {test_lin.applied_opts}")
 
         # stop if kernel uops repeat
-        try: tuops = tuplize_uops(get_program(test_lin.get_optimized_ast(), test_lin.opts).uops)
+        try: tuops = tuplize_uops(get_program(test_lin.get_optimized_ast(), test_lin.ren).uops)
         except KeyboardInterrupt: raise
         except BaseException as e:
           print(test_lin.ast)
@@ -224,7 +224,7 @@ def fuzz_linearizer(lin: Kernel, rtol=1e-2, atol=1e-2, opts_list=None):
         (msg, rawbufs, var_vals, ground_truth, state1) = compare_linearizer(test_lin, rawbufs, var_vals, ground_truth, rtol=rtol, atol=atol)
         if state1 is not None and validate_device is not None:
           validate_lin = test_lin.copy()
-          validate_lin.opts = validate_device.renderer
+          validate_lin.ren = validate_device.renderer
           if validate_rawbufs is None:
             validate_rawbufs = [get_fuzz_rawbuf_like(x, copy=True, force_device=validate_device.device) for x in rawbufs]
           (_msg, _, _, _, state2) = compare_linearizer(validate_lin, validate_rawbufs, var_vals, ground_truth, rtol=rtol, atol=atol)

test/external/fuzz_symbolic.py

@@ -2,7 +2,7 @@ import random, operator
 import z3
 from tinygrad import Variable, dtypes
 from tinygrad.uop.ops import UOp
-from tinygrad.uop.spec import uops_to_z3
+from tinygrad.uop.validate import uops_to_z3
 from tinygrad.helpers import DEBUG, Context
 
 seed = random.randint(0, 100)

test/external/process_replay/process_replay.py

@@ -13,7 +13,7 @@ try:
   from tinygrad.engine.realize import get_program
   from tinygrad.uop.ops import UOp, Ops, KernelInfo
   from tinygrad.codegen.opt import Opt
-  from tinygrad.helpers import VERSION, Context, ContextVar, colored, db_connection, getenv, tqdm
+  from tinygrad.helpers import VERSION, Context, ContextVar, colored, db_connection, getenv, tqdm, BEAM
   from tinygrad.device import Device
 except ImportError as e:
   print(repr(e))
@@ -51,8 +51,8 @@ def replay_get_rangeify_map(ret:dict[UOp, UOp], big_sink:UOp) -> tuple[str, str,
   return to_str(new_sink), to_str(big_sink.substitute(ret)), (big_sink,)
 
 def replay_get_program(p:ProgramSpec, ast:UOp, renderer:Renderer|None=None, opts:list[Opt]|None=None) -> tuple[str, str, tuple[Any, ...]]:
-  # NOTE: this always uses the opts_to_apply path
-  sink_arg = ast.arg or KernelInfo(opts_to_apply=p.applied_opts)
+  # the ast.arg is non None if we are inside of search.py
+  sink_arg = ast.arg or KernelInfo(opts_to_apply=tuple(opts) if opts is not None else p.applied_opts if BEAM>=1 else None)
   input_ast = ast.replace(arg=replace(sink_arg, name=p.name))
   # if no renderer was provided, open the device to get it
   if renderer is None: renderer = Device[p.device].renderer

test/external/speed_v_theoretical.py

@@ -91,11 +91,11 @@ class TestKernelSpeed(unittest.TestCase):
 
   # theoretical is nv_tflops=165, amd_tflops=123
   def test_gemm_4096(self): self._test_matmul(4096, nv_tflops=115, amd_tflops=65)
-  def test_gemm_8192(self): self._test_matmul(8192, nv_tflops=125, amd_tflops=60)
+  def test_gemm_8192(self): self._test_matmul(8192, nv_tflops=115, amd_tflops=60)
 
   # theoretical is nv_gbs=1008, amd_gbs=960
   def test_gemv_16384_4096(self): self._test_matmul(16384, 4096, 1, nv_gbs=840, amd_gbs=750)
-  def test_gemv_4096_16384(self): self._test_matmul(4096, 16384, 1, nv_gbs=830, amd_gbs=750)
+  def test_gemv_4096_16384(self): self._test_matmul(4096, 16384, 1, nv_gbs=820, amd_gbs=750)
 
 if __name__ == '__main__':
   unittest.main()

test/models/test_real_world.py

@@ -112,7 +112,7 @@ class TestRealWorld(unittest.TestCase):
         loss.backward()
         optimizer.step()
 
-      helper_test("train_mnist", lambda: (Tensor.randn(BS, 1, 28, 28),), train, 0.07, 102)
+      helper_test("train_mnist", lambda: (Tensor.randn(BS, 1, 28, 28),), train, 0.07, 103)
 
   @unittest.skipIf(CI and Device.DEFAULT in {"CPU", "CL"}, "slow")
   def test_forward_cifar(self):
@@ -176,7 +176,7 @@ class TestRealWorld(unittest.TestCase):
       for v in data.values(): v.to_(Device.DEFAULT)
 
       helper_test("train_bert", lambda: (data["input_ids"], data["segment_ids"], data["input_mask"], data["masked_lm_positions"], \
-          data["masked_lm_ids"], data["masked_lm_weights"], data["next_sentence_labels"]), train, 0.31, 358)
+          data["masked_lm_ids"], data["masked_lm_weights"], data["next_sentence_labels"]), train, 0.31, 427)
 
 if __name__ == '__main__':
   unittest.main()

test/opt/test_tensor_cores.py

@@ -14,6 +14,8 @@ from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
 # TODO: write a clean version of this
 from test.test_linearizer import helper_realized_ast, helper_linearizer_opt
 
+# NOTE: get_program always passes in Device[Device.DEFAULT].renderer explicitly for process_replay!!!
+
 def helper_tc_ensure_uops_and_opts_count(N: int, M:int, K:int, dtype_in:DType, dtype_out:DType, axis:int=0, tc_select:int=-1, tc_opt:int=0,
                                          ensure_triggered:bool=True):
   a, b = Tensor.rand(M, K, dtype=dtype_in), Tensor.rand(K, N, dtype=dtype_in)
@@ -41,7 +43,7 @@ def helper_tc_allclose(N:int, M:int, K:int, dtype_in:DType, dtype_out:DType, axi
   if dtype_in == dtypes.bfloat16: r = r.float()
   realized_ast, bufs = helper_realized_ast(r)
   opts = [Opt(op=OptOps.TC, axis=axis, arg=(tc_select, tc_opt, use_tensor_cores))]
-  prg = CompiledRunner(replace(get_program(realized_ast, opts=opts), device=Device.DEFAULT))
+  prg = CompiledRunner(replace(get_program(realized_ast, Device[Device.DEFAULT].renderer, opts=opts), device=Device.DEFAULT))
   if use_tensor_cores == 1: assert len([uop for uop in prg.p.uops if uop.op is Ops.WMMA]) > 0, "wmma not triggered"
   assert len([x for x in prg.p.uops[-1].arg.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
   prg.exec(bufs)
@@ -68,7 +70,7 @@ class TestTensorCores(unittest.TestCase):
       n, m, k = tc.dims[0], tc.dims[1], 2 if AMX else tc.dims[2]
       a, b = Tensor.rand(m, k, dtype=tc.dtype_in), Tensor.rand(k, n, dtype=tc.dtype_in)
       r = a.matmul(b, dtype=tc.dtype_out)
-      prg = get_program(r.schedule()[-1].ast, opts=[Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))])
+      prg = get_program(r.schedule()[-1].ast, Device[Device.DEFAULT].renderer, opts=[Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))])
       if Device.DEFAULT == "CPU" and CPU_LLVM:
         assert "0x201000" in prg.src
       elif Device.DEFAULT == "AMD" and AMD_LLVM:
@@ -154,7 +156,7 @@ class TestTensorCores(unittest.TestCase):
     r = x.matmul(y, dtype=tc.dtype_out)
     opts = [Opt(OptOps.UNROLL, 0, 4)]
     ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
-    for u in get_program(ast, opts=opts).uops:
+    for u in get_program(ast, Device[Device.DEFAULT].renderer, opts=opts).uops:
       if u.op is Ops.WMMA:
         assert u.src[-1].src[0].op != Ops.STORE
 
@@ -167,7 +169,7 @@ class TestTensorCores(unittest.TestCase):
     r = x.matmul(y, dtype=tc.dtype_out)
     opts = [Opt(OptOps.UNROLL, 0, 4)]
     ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
-    for u in get_program(ast, opts=opts).uops:
+    for u in get_program(ast, Device[Device.DEFAULT].renderer, opts=opts).uops:
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
         assert u.src[-1].src[0].op != Ops.STORE
@@ -182,7 +184,7 @@ class TestTensorCores(unittest.TestCase):
     r = x.matmul(y, dtype=tc.dtype_out).relu()
     opts = [Opt(OptOps.UNROLL, 0, 4)]
     ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
-    for u in get_program(ast, opts=opts).uops:
+    for u in get_program(ast, Device[Device.DEFAULT].renderer, opts=opts).uops:
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
         assert u.src[-1].src[0].op != Ops.STORE

test/test_const_folding.py

@@ -24,7 +24,6 @@ class TestUnaryOpsConstFolding(unittest.TestCase):
     _check_ast_count(0, Tensor.ones(4).cast(dtypes.int16))
     _check_ast_count(0, Tensor.full(4, fill_value=-1).cast(dtypes.uint16))
 
-  @unittest.expectedFailure  # no two level fold
   def test_neg_folding(self):
     _check_ast_count(0, Tensor([1, 2, 3]).mul(-1).neg())
     _check_ast_count(0, Tensor([1, 2, 3]).neg().mul(-1))
@@ -68,12 +67,9 @@ class TestBinaryOpsConstFolding(unittest.TestCase):
   def test_tensor_one_mul(self):
     _check_ast_count(0, Tensor.ones(4) * Tensor([1.0, 2, 3, 4]))
 
-  # TODO: these will be fixed with better folding
-  @unittest.expectedFailure
   def test_bool_tensor_mul_bool(self):
     _check_ast_count(0, Tensor([True, False]) * True)
     _check_ast_count(0, Tensor([True, False]) * False)
-  @unittest.expectedFailure
   def test_bool_mul_bool_tensor(self):
     _check_ast_count(0, True * Tensor([True, False]))
     _check_ast_count(0, False * Tensor([True, False]))
@@ -83,10 +79,8 @@ class TestBinaryOpsConstFolding(unittest.TestCase):
   def test_div_tensor_one(self):
     _check_ast_count(0, Tensor([1.0, 2, 3, 4]) / Tensor.ones(4))
 
-  @unittest.expectedFailure  # TODO: fix
   def test_idiv_literal_one(self):
     _check_ast_count(0, Tensor([1, 2, 3, 4]) // 1)
-  @unittest.expectedFailure  # TODO: fix
   def test_idiv_tensor_one(self):
     _check_ast_count(0, Tensor([1, 2, 3, 4]) // Tensor.ones(4, dtype=dtypes.int32))
 
@@ -188,7 +182,7 @@ class TestReduceOpsConstFolding(unittest.TestCase):
     np.testing.assert_equal(Tensor(4).sum().numpy(), 4)
 
   def test_padded_const_sum(self):
-    _check_ast_count(1, Tensor.ones(4).pad(((1, 1),)).sum())
+    _check_ast_count(0, Tensor.ones(4).pad(((1, 1),)).sum())
     np.testing.assert_equal(Tensor.ones(4).pad(((1, 1),)).sum().numpy(), 4)
 
     # NOTE: cannot just count the non-padded area because some Ops f do not have f(0) = 0.

test/test_dtype_alu.py

@@ -75,7 +75,10 @@ def universal_test_unary(a, dtype, op):
   out: Tensor = op[0](ta)
   tensor_value = out.numpy()
   numpy_value = op[1](ta.numpy())
-  if dtype in dtypes.fp8s: numpy_value = truncate[dtype](numpy_value)
+  if dtype in dtypes.fp8s:
+    # cuda cast f32 inf to f8 MAX, amd cast it to nan(E4M3)/inf(E5M2)
+    if math.isinf(numpy_value): return
+    numpy_value = truncate[dtype](numpy_value)
   if dtype in dtypes.floats:
     atol, rtol = { dtypes.float16:(1e-3, 1e-2), dtypes.bfloat16:(1e-3, 2e-2),
       dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2: (1.0, 5e-1)}.get(dtype, (1e-6, 1e-5))

test/test_fusion_op.py

@@ -51,7 +51,7 @@ class TestFusionOp(unittest.TestCase):
     a = Tensor(val)
     for _ in range(24): a = Tensor.stack(a, a)[0]
     sched = a.schedule()
-    self.assertEqual(len(sched), 1)
+    self.assertEqual(len(sched), 0)
     self.assertLess(time.perf_counter()-st, 2.0)
 
   def test_recursive_reshape(self):

test/test_image_dtype.py

@@ -52,7 +52,6 @@ class TestImageDType(unittest.TestCase):
     assert isinstance(it.uop.base.realized.dtype, ImageDType)
     np.testing.assert_equal(tst, it.numpy())
 
-  @unittest.expectedFailure # this isn't supported anymore, CAST to ImageDType stays ImageDType
   def test_image_cast_and_back_collapses(self):
     data = Tensor.randn(9*27*4).realize()
     tst = data.numpy()

test/test_linearizer.py

@@ -41,7 +41,7 @@ class TestLinearizer(unittest.TestCase):
   def _test_no_nested_ranges(self, lins, skip=None):
     for l in lins:
       range_in_acc = flatten([[x for x in u.src if x.op is Ops.RANGE] for u in l.uops if u.op is Ops.DEFINE_REG])
-      ranges = [u.op for u in l.uops if (u.op is Ops.RANGE and u in range_in_acc) or (u.op is Ops.ENDRANGE and u.src[0] in range_in_acc)]
+      ranges = [u.op for u in l.uops if (u.op is Ops.RANGE and u in range_in_acc) or (u.op is Ops.END and u.src[0] in range_in_acc)]
       for i,u in enumerate(ranges):
         if skip and i in skip: continue
         assert ranges[i-1] != u, f"multireduce nested the ranges! {ranges[i-1], {u}}"
@@ -155,6 +155,7 @@ class TestLinearizer(unittest.TestCase):
     assert stores[1].src[1].dtype == dtypes.float
     assert any(x.op is Ops.DEFINE_GLOBAL for x in stores[1].toposort())
 
+  @unittest.skipIf(Device.DEFAULT=="CPU", "CPU splits the cat so cant upcast")
   def test_zero_fold(self):
     a, b = Tensor.randn(1).realize(), Tensor.randn(1).realize()
     r = Tensor.stack(a, b)
@@ -205,7 +206,7 @@ class TestLinearizer(unittest.TestCase):
     # the uops graph is DEFINE_REG -> 4x STORE 0.0 -> RANGE -> 4x ALU -> 4x STORE -> ENDRANGE
     uops = get_program(ast, opts=opt).uops
     begin_range = [i for i, x in enumerate(uops) if x.op is Ops.RANGE][-1]
-    end_range = [i for i, x in enumerate(uops) if x.op is Ops.ENDRANGE][0]
+    end_range = [i for i, x in enumerate(uops) if x.op is Ops.END][0]
     for i,u in enumerate(uops): print(i, u.op, [uops.index(s) for s in u.src], u.arg, u.dtype)
     for u in uops:
       if u.op is Ops.STORE and isinstance(dt:=u.src[0].dtype, PtrDType) and dt.addrspace is AddrSpace.REG:
@@ -214,8 +215,8 @@ class TestLinearizer(unittest.TestCase):
         else:
           assert u.src[1].op in GroupOp.ALU
           assert begin_range < uops.index(u) < end_range
-      # children of STORE are placed after ENDRANGE
-      if any(x.op is Ops.STORE and x.src[1].op in GroupOp.ALU for x in u.src):
+      # children of END are placed after ENDRANGE
+      if any(x.op is Ops.END and x.src[1].op in GroupOp.ALU for x in u.src):
         assert end_range < uops.index(u)
 
   def test_grouped_dims(self):
@@ -393,14 +394,15 @@ class TestLinearizer(unittest.TestCase):
     uops = get_program(ast, opts=opt).uops
     local_stores = [u for u in uops if u.op is Ops.STORE and any(x.op is Ops.DEFINE_LOCAL for x in get_recursive(u.src[0]))]
     global_stores = [u for u in uops if u.op is Ops.STORE and any(x.op is Ops.DEFINE_GLOBAL for x in get_recursive(u.src[0]))]
-    barrier = [u for u in uops if u.op is Ops.BARRIER][0]
+    barrier = [u for u in uops if u.op is Ops.BARRIER]
+    assert len(barrier) == 1
     # check that the float4 cast collapses for all stores
     for store in local_stores+global_stores:
       assert store.src[1].dtype.count > 1 # and store.src[2].op is not Ops.VECTORIZE
     # # check the children's vins
     # TODO: src ALU are not the same, should it?
     # assert barrier.src == tuple(local_stores)
-    assert len([u for u in uops if u.op is Ops.IF and u.src[-1] == barrier]) == 1
+    #assert len([u for u in uops if u.op is Ops.IF])
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")

test/test_linearizer_dumb.py

@@ -23,7 +23,7 @@ class TestLinearizerFailure(unittest.TestCase):
     c9 = UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(47040000), arg=2, src=())
     c10 = c9.index((((c3*UOp.const(dtypes.index, 4704000))+c2)+(c6*UOp.const(dtypes.index, 784))).valid(UOp.const(dtypes.bool, True))).load()
     c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(dtypes.index, 6000))+c6)+((c7*UOp.const(dtypes.index, 16))+c8)).alu(Ops.CMPLT, UOp.const(dtypes.index, 59999)).where(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1)).reduce(c7, c8, arg=Ops.ADD)+UOp.const(dtypes.int, -1))).where(UOp.const(dtypes.uchar, 0), c10).reduce(c6, arg=Ops.ADD)
-    c12 = c0.index((((c1*UOp.const(dtypes.index, 7840))+(c2*UOp.const(dtypes.index, 10)))+c3).valid(UOp.const(dtypes.bool, True))).store(c11, c1, c2, c3)
+    c12 = c0.index((((c1*UOp.const(dtypes.index, 7840))+(c2*UOp.const(dtypes.index, 10)))+c3).valid(UOp.const(dtypes.bool, True))).store(c11).end(c1, c2, c3)
     ast = c12.sink(arg=KernelInfo(name='test', axis_types=(), dont_use_locals=False, applied_opts=(Opt(op=OptOps.GROUP, axis=1, arg=16),), opts_to_apply=None))
     _ = get_program(ast, Device["METAL"].renderer)
 

test/test_linearizer_failures.py

@@ -16,7 +16,7 @@ class TestLinearizerFailures(unittest.TestCase):
     c7 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(64), arg=2, src=())
     c8 = c7.index(c3).load()
     c9 = ((((c6+(c8*UOp.const(dtypes.float, -1.0)))*(c6+(c8*UOp.const(dtypes.float, -1.0)))).reduce(c5, arg=Ops.ADD)*UOp.const(dtypes.float, 0.000390625))+UOp.const(dtypes.float, 1e-05)).sqrt().reciprocal()
-    c10 = c0.index(c3).store(c9, c1, c2)
+    c10 = c0.index(c3).store(c9).end(c1, c2)
     ast = c10.sink()
     get_program(ast)
 

test/test_ops.py

@@ -2602,17 +2602,13 @@ class TestOps(unittest.TestCase):
       lambda x: torch.nn.functional.avg_pool2d(x, kernel_size=(111,28)),
       lambda x: Tensor.avg_pool2d(x, kernel_size=(111,28)), rtol=1e-5)
 
-  def test_avg_pool3d_failure(self):
-    with Context(NOOPT=0):
-      helper_test_op([(1,1,16,16,16)],
-        lambda x: torch.nn.functional.avg_pool3d(x, kernel_size=(8,8,8), stride=5, padding=1, count_include_pad=False),
-        lambda x: Tensor.avg_pool2d(x, kernel_size=(8,8,8), stride=5, padding=1, count_include_pad=False), rtol=1e-5, forward_only=True)
-
-  def test_avg_pool3d_noopt(self):
-    with Context(NOOPT=1):
-      helper_test_op([(1,1,16,16,16)],
-        lambda x: torch.nn.functional.avg_pool3d(x, kernel_size=(8,8,8), stride=5, padding=1, count_include_pad=False),
-        lambda x: Tensor.avg_pool2d(x, kernel_size=(8,8,8), stride=5, padding=1, count_include_pad=False), rtol=1e-5, forward_only=True)
+  def test_avg_pool3d(self):
+    # TODO: AMD_LLVM has larger atol
+    # TODO: PYTHON=1 backward hangs?
+    atol = 1e-2 if AMD_LLVM else 1e-6
+    helper_test_op([(1,1,16,16,16)],
+      lambda x: torch.nn.functional.avg_pool3d(x, kernel_size=(8,8,8), stride=5, padding=1, count_include_pad=False),
+      lambda x: Tensor.avg_pool2d(x, kernel_size=(8,8,8), stride=5, padding=1, count_include_pad=False), atol=atol, rtol=1e-5, forward_only=True)
 
   def test_interpolate_linear(self):
     for in_sz, out_sz in [((52,),(29,)), ((29,),(52,))]:

test/test_optim.py

@@ -5,7 +5,6 @@ from tinygrad import Tensor, Device, dtypes
 from tinygrad.nn.optim import Adam, SGD, AdamW, Muon
 from tinygrad.helpers import CI
 from tinygrad.device import is_dtype_supported
-from extra.torch_muon import SingleDeviceMuon as TorchMuon
 
 np.random.seed(1337)
 x_init = np.random.randn(1,4).astype(np.float32)
@@ -58,12 +57,11 @@ class TestOptim(unittest.TestCase):
   def _test_sgd(self, steps, opts, atol, rtol): self._test_optim(SGD, torch.optim.SGD, steps, opts, atol, rtol)
   def _test_adam(self, steps, opts, atol, rtol): self._test_optim(Adam, torch.optim.Adam, steps, opts, atol, rtol)
   def _test_adamw(self, steps, opts, atol, rtol): self._test_optim(AdamW, torch.optim.AdamW, steps, opts, atol, rtol)
-  #TODO: use torch.muon when it comes out
-  def _test_muon(self, steps, opts, atol, rtol): self._test_optim(Muon, TorchMuon, steps, opts, atol, rtol)
+  def _test_muon(self, steps, opts, atol, rtol): self._test_optim(Muon, torch.optim.Muon, steps, opts, atol, rtol)
 
   def test_multistep_sgd_high_lr_teeny(self): self._test_sgd(2, {'lr': 1.1, 'teeny': True}, 1e-6, 1e-5)
   def test_multistep_adam_high_lr_teeny(self): self._test_adam(2, {'lr': 1.1, 'teeny': True}, 2e-4, 5e-4)
-  def test_multistep_muon_high_lr_teeny(self): self._test_muon(2, {'lr': 1.1, 'teeny': True}, 2e-4, 5e-4)
+  def test_multistep_muon_high_lr_teeny(self): self._test_muon(2, {'lr': 1.1, 'teeny': True}, 1e-2, 5e-4)
 
   def test_sgd(self): self._test_sgd(1, {'lr': 0.001}, 1e-6, 0)
   def test_sgd_high_lr(self): self._test_sgd(1, {'lr': 10}, 1e-6, 1e-5)
@@ -87,27 +85,34 @@ class TestOptim(unittest.TestCase):
   def test_multistep_sgd_high_lr_nesterov_momentum_wd(self):
     self._test_sgd(10, {'lr': 9, 'momentum': 0.9, 'nesterov': True, 'weight_decay': 0.1}, 1e-5, 3e-4)
 
-  def test_muon(self): self._test_muon(1, {'lr': 0.001}, 1e-6, 0)
-  def test_muon_high_lr(self): self._test_muon(1, {'lr': 10}, 1e-6, 3e-4)
-  def test_muon_wd(self): self._test_muon(1, {'lr': 0.001, 'weight_decay': 0.01}, 1e-6, 0)
-  def test_muon_high_lr_wd(self): self._test_muon(1, {'lr': 10, 'weight_decay': 0.01}, 1e-6, 5e-4)
+  def test_muon(self): self._test_muon(1, {'lr': 0.001}, 1e-3, 0)
+  # TODO: disabled due to big atol
+  # def test_muon_high_lr(self): self._test_muon(1, {'lr': 10}, 1e-6, 3e-4)
+  def test_muon_wd(self): self._test_muon(1, {'lr': 0.001, 'weight_decay': 0.01}, 1e-3, 3e-4)
+  # TODO: disabled due to big atol
+  # def test_muon_high_lr_wd(self): self._test_muon(1, {'lr': 10, 'weight_decay': 0.01}, 1e-6, 5e-4)
 
   # NOTE: momentum set to 0.95 by default, nesterov set to True by default
-  def test_multistep_muon_momentum_wd(self): self._test_muon(10, {'lr': 0.001, 'weight_decay': 0.01}, 1e-5, 0)
+  def test_multistep_muon_momentum_wd(self): self._test_muon(10, {'lr': 0.001, 'weight_decay': 0.01}, 3e-3, 0)
   # ns defaults are numerically unstable, but it is tolerable in real training (see nsteps/nparam tests)
-  def test_multistep_muon_high_lr_momentum_wd(self): self._test_muon(10, {'lr': 10, 'weight_decay': 0.01}, 1e-1, 3e-4)
-  def test_multistep_muon_no_nesterov_momentum(self): self._test_muon(10, {'lr': 0.001, 'nesterov': False}, 1e-5, 0)
-  def test_multistep_muon_high_lr_no_nesterov_momentum(self): self._test_muon(10, {'lr': 10, 'nesterov': False}, 0.5e-1, 1e-1)
+  # TODO: disabled due to big atol
+  # def test_multistep_muon_high_lr_momentum_wd(self): self._test_muon(10, {'lr': 10, 'weight_decay': 0.01}, 1e-1, 3e-4)
+  def test_multistep_muon_no_nesterov_momentum(self): self._test_muon(10, {'lr': 0.001, 'nesterov': False}, 1e-3, 0)
+  # TODO: disabled due to big atol
+  # def test_multistep_muon_high_lr_no_nesterov_momentum(self): self._test_muon(10, {'lr': 10, 'nesterov': False}, 5e-2, 1e-1)
 
-  def test_muon_ns_steps(self): self._test_muon(1, {'lr': 0.001, 'ns_steps': 3}, 1e-6, 0)
-  def test_muon_high_lr_ns_steps(self): self._test_muon(1, {'lr': 10, 'ns_steps': 3}, 1e-5, 3e-4)
-  def test_muon_ns_params(self): self._test_muon(1, {'lr': 0.001,'ns_params': (2.0,-1.5,0.5)}, 1e-6, 0)
-  def test_muon_high_lr_ns_params(self): self._test_muon(1, {'lr': 10,'ns_params': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
+  def test_muon_ns_steps(self): self._test_muon(1, {'lr': 0.001, 'ns_steps': 3}, 1e-4, 0)
+  # TODO: disabled due to big atol
+  # def test_muon_high_lr_ns_steps(self): self._test_muon(1, {'lr': 10, 'ns_steps': 3}, 1e-5, 3e-4)
+  def test_muon_ns_coefficients(self): self._test_muon(1, {'lr': 0.001,'ns_coefficients': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
+  # TODO: disabled due to big atol
+  # def test_muon_high_lr_ns_coefficients(self): self._test_muon(1, {'lr': 10,'ns_coefficients': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
 
-  def test_muon_momentum_wd_ns_steps_ns_params(self):
-    self._test_muon(10, {'lr': 0.001, 'momentum': 0.90, 'weight_decay': 0.01, 'ns_steps': 3, 'ns_params': (2.0,-1.5,0.5)}, 1e-5, 0)
-  def test_multistep_muon_high_lr_momentum_wd_ns_steps_ns_params(self):
-    self._test_muon(10, {'lr': 10, 'momentum': 0.90, 'weight_decay': 0.01, 'ns_steps': 3, 'ns_params': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
+  def test_muon_momentum_wd_ns_steps_ns_coefficients(self):
+    self._test_muon(10, {'lr': 0.001, 'momentum': 0.90, 'weight_decay': 0.01, 'ns_steps': 3, 'ns_coefficients': (2.0,-1.5,0.5)}, 1e-4, 0)
+  # TODO: disabled due to big atol
+  # def test_multistep_muon_high_lr_momentum_wd_ns_steps_ns_coefficients(self):
+  #   self._test_muon(10, {'lr': 10, 'momentum': 0.90, 'weight_decay': 0.01, 'ns_steps': 3, 'ns_coefficients': (2.0,-1.5,0.5)}, 1e-5, 3e-4)
 
   def test_adam(self): self._test_adam(1, {'lr': 0.001}, 1e-5, 0)
   def test_adam_high_lr(self): self._test_adam(1, {'lr': 10}, 1e-4, 1e-4)

test/test_profiler.py

@@ -92,7 +92,9 @@ class TestProfiler(unittest.TestCase):
     #  assert evs[i].st > evs[i-1].en, "timestamp not aranged"
 
   def test_profile_multidev(self):
-    d1 = Device[f"{Device.DEFAULT}:1"]
+    try: d1 = Device[f"{Device.DEFAULT}:1"]
+    except Exception as e: self.skipTest(f"second device not available {e}")
+
     buf1 = Buffer(Device.DEFAULT, 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
     buf2 = Buffer(f"{Device.DEFAULT}:1", 2, dtypes.float, options=BufferSpec(nolru=True)).ensure_allocated()
 
@@ -109,7 +111,8 @@ class TestProfiler(unittest.TestCase):
       assert evs[0].is_copy, "kernel should be copy"
 
   def test_profile_multidev_transfer(self):
-    d1 = Device[f"{Device.DEFAULT}:1"]
+    try: d1 = Device[f"{Device.DEFAULT}:1"]
+    except Exception as e: self.skipTest(f"second device not available {e}")
 
     buf1 = Tensor.randn(10, 10, device=f"{Device.DEFAULT}:0").realize()
     with helper_collect_profile(TestProfiler.d0, d1) as profile:
@@ -122,7 +125,8 @@ class TestProfiler(unittest.TestCase):
 
   @unittest.skipIf(Device.DEFAULT in "METAL" or (MOCKGPU and Device.DEFAULT == "AMD"), "AMD mockgpu does not support queue wait interrupts")
   def test_profile_graph(self):
-    d1 = Device[f"{Device.DEFAULT}:1"]
+    try: d1 = Device[f"{Device.DEFAULT}:1"]
+    except Exception as e: self.skipTest(f"second device not available {e}")
 
     def f(a):
       x = (a + 1).realize()
@@ -145,7 +149,9 @@ class TestProfiler(unittest.TestCase):
   @unittest.skipIf(CI or not issubclass(type(Device[Device.DEFAULT]), HCQCompiled), "skip CI")
   def test_dev_jitter_matrix(self):
     dev_cnt = 6
-    devs = [Device[f"{Device.DEFAULT}:{i}"] for i in range(dev_cnt)]
+    try: devs = [Device[f"{Device.DEFAULT}:{i}"] for i in range(dev_cnt)]
+    except Exception as e: self.skipTest(f"multiple devices not available {e}")
+
     for dev in devs: dev.synchronize()
     for dev in devs: dev._at_profile_finalize()
 
@@ -219,7 +225,7 @@ class TestProfiler(unittest.TestCase):
     exec_points = [e for e in profile if isinstance(e, ProfilePointEvent) and e.name == "exec"]
     range_events = [e for e in profile if isinstance(e, ProfileRangeEvent) and not e.is_copy]
     self.assertEqual(len(exec_points), len(range_events), 2)
-    self.assertEqual(len(dedup(e.key for e in exec_points)), 1)
+    self.assertEqual(len(dedup(e.arg['name'] for e in exec_points)), 1)
     self.assertEqual(len(dedup(e.arg['metadata'] for e in exec_points)), 1)
 
 if __name__ == "__main__":

test/test_rangeify.py

@@ -1,7 +1,9 @@
 import unittest
-from tinygrad import Tensor, nn
-from tinygrad.helpers import Context, GlobalCounters, CI
+from tinygrad import Tensor, nn, Device
+from tinygrad.helpers import Context, GlobalCounters, CI, getenv, PCONTIG
 from tinygrad.uop.ops import graph_rewrite, PatternMatcher, UPat, Ops
+from tinygrad.renderer.ptx import PTXRenderer
+from tinygrad.renderer.nir import NIRRenderer
 
 class TestRangeifyAssign(unittest.TestCase):
   def test_assign_permuted(self):
@@ -28,6 +30,74 @@ class TestRangeifyEdgeCase(unittest.TestCase):
     res = Tensor.cat(a, c, dim=0)
     self.assertEqual(res.numpy()[-1, :16].tolist(), [512] * 16)
 
+if getenv("BIG") > 2:
+  # llama 8B (8192)
+  BS, HEADS, SEQLEN, EMB = 4, 32, 8192, 128
+elif getenv("BIG") > 1:
+  # llama 8B
+  BS, HEADS, SEQLEN, EMB = 4, 32, 2048, 128
+elif getenv("BIG") > 0:
+  # bigger
+  BS, HEADS, SEQLEN, EMB = 4, 32, 1024, 64
+else:
+  BS, HEADS, SEQLEN, EMB = 4, 2, 16, 8
+
+def fa():
+  Tensor.manual_seed(1337)
+  with Context(DEBUG=0): q,k,v = [Tensor.rand(BS, HEADS, SEQLEN, EMB).contiguous().realize() for _ in range(3)]
+  GlobalCounters.reset()
+  return q.scaled_dot_product_attention(k, v)
+
+def fa_bw():
+  Tensor.manual_seed(1337)
+  with Context(DEBUG=0):
+    q,k,v = [Tensor.rand(BS, HEADS, SEQLEN, EMB).contiguous().realize().requires_grad_() for _ in range(3)]
+    attn_output = nn.Linear(HEADS*EMB, HEADS*EMB, bias=False)
+    attn_output.weight.requires_grad_().realize()
+    target = Tensor.rand(BS, SEQLEN, HEADS*EMB).contiguous().realize()
+
+  GlobalCounters.reset()
+  attn = q.scaled_dot_product_attention(k, v).contiguous().contiguous_backward()
+  attn = attn.transpose(1, 2).reshape(BS, SEQLEN, -1)
+  out = attn_output(attn)
+  loss = (out - target).square().mean()
+  loss.backward()
+  #ret = [out, Tensor.stack(q.grad, k.grad, v.grad, dim=-1)]
+  #ret = [out, Tensor.stack(q.grad, k.grad, dim=-1), v.grad]
+  ret = [out, q.grad, k.grad, v.grad]
+  Tensor.realize(*ret)
+  return ret
+
+@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (NIRRenderer, PTXRenderer)), "broken in LVP and PTX")
+class TestPcontig(unittest.TestCase):
+  def test_flash_attention_bw(self):
+    with Context(PCONTIG=max(2, PCONTIG.value), DEBUG=2):
+      grads = fa_bw()
+      print(f"{GlobalCounters.global_ops/1e9:.2f} GFLOPS")
+
+    with Context(PCONTIG=0, DEBUG=2):
+      cmp_grads = fa_bw()
+      print(f"{GlobalCounters.global_ops/1e9:.2f} GFLOPS")
+
+    with Context(DEBUG=0):
+      mses = [((x-y)**2).sum().item() for x,y in zip(grads, cmp_grads)]
+    mse = sum(mses)
+    print(f"mse: {mse}")
+    self.assertLessEqual(mse, 1e-6)
+
+  def test_flash_attention(self):
+    with Context(PCONTIG=2, DEBUG=2):
+      ret = fa().realize()
+      print(f"{GlobalCounters.global_ops/1e9:.2f} GFLOPS")
+    with Context(DEBUG=2):
+      cmp = fa().realize()
+      print(f"{GlobalCounters.global_ops/1e9:.2f} GFLOPS")
+    with Context(DEBUG=0):
+      mse = ((cmp-ret)**2).sum().item()
+    print(f"mse: {mse}")
+    self.assertLessEqual(mse, 1e-6)
+
+
 # *** non CI rangeify tests below this line ***
 
 N = 256
@@ -215,33 +285,6 @@ class TestRangeify(unittest.TestCase):
     out = blk._feed_forward(x)
     out.realize()
 
-  @unittest.skip("RANGEIFY=0 does nothing")
-  def test_flash_attention(self):
-    BS, HEADS, SEQLEN, EMB = 4, 2, 16, 8
-
-    # bigger
-    #BS, HEADS, SEQLEN, EMB = 4, 32, 1024, 64
-
-    # llama 8B
-    #BS, HEADS, SEQLEN, EMB = 4, 32, 2048, 128
-
-    def fa():
-      Tensor.manual_seed(1337)
-      with Context(DEBUG=0): q,k,v = [Tensor.rand(BS, HEADS, SEQLEN, EMB).contiguous().realize() for _ in range(3)]
-      return q.scaled_dot_product_attention(k, v).realize()
-
-    with Context(DEBUG=4):
-      GlobalCounters.reset()
-      ret = fa()
-    with Context(RANGEIFY=0):
-      with Context(DEBUG=2):
-        GlobalCounters.reset()
-        cmp = fa()
-      with Context(DEBUG=0):
-        mse = ((cmp-ret)**2).sum().item()
-    print(f"mse: {mse}")
-    self.assertLessEqual(mse, 1e-6)
-
 # contiguous + reduce can support ranges?
 
 @unittest.skip("pm_rangeify no longer exists. test this in a different way")

test/test_schedule.py

@@ -333,7 +333,7 @@ class TestSchedule(unittest.TestCase):
     r1 = (x - r0).sum(axis=0).div(2)
     out0 = r0 + y
     out1 = r1 + y
-    schedule = check_schedule([out0, out1], 4)
+    schedule = check_schedule([out0, out1], 3)
     reduceops = [x for si in schedule for x in si.ast.toposort() if x.op in {Ops.REDUCE_AXIS, Ops.REDUCE}]
     self.assertEqual(len(reduceops), 2) # why is RANGEIFY different?
 
@@ -370,6 +370,7 @@ class TestSchedule(unittest.TestCase):
 
   # NOTE: this is causing "LAZYCACHE=1 incorrectly reuses contiguous const" #4562
   # should contiguous dedup?
+  @unittest.skip("we do the exact opposite now")
   def test_dedup_contiguous(self):
     a = Tensor.ones(4).contiguous()
     b = Tensor.ones(4).contiguous()
@@ -446,7 +447,7 @@ class TestSchedule(unittest.TestCase):
   @unittest.skipUnless(is_dtype_supported(dtypes.ulong), "Needs ulong")
   def test_fold_conv_batchnorm_optim(self):
     # this is too high
-    for optim, cnt in [(nn.optim.Adam, 30), (nn.optim.SGD, 13)]:
+    for optim, cnt in [(nn.optim.Adam, 28), (nn.optim.SGD, 8)]:
       with self.subTest(optim=optim.__name__):
         with Tensor.train():
           img = Tensor.ones(1,3,4,4)
@@ -759,7 +760,7 @@ class TestSchedule(unittest.TestCase):
 
   def test_pow_neg_05_is_rsqrt(self):
     t = Tensor([1.0, 2.0, 3.0]) ** -0.5
-    self.assertEqual(self._alu_from_tensor(t), [Ops.RECIP, Ops.SQRT])
+    self.assertEqual(self._alu_from_tensor(t), [Ops.RECIPROCAL, Ops.SQRT])
 
   def test_pow_2_has_1_mul(self):
     t = Tensor([1.0, 2.0, 3.0]) ** Tensor(2.0)
@@ -1220,7 +1221,7 @@ class TestSchedule(unittest.TestCase):
       _realize_weights(layer)
       opt = nn.optim.Adam(nn.state.get_parameters(layer), lr=1e-4)
       layer(x).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 16)
+      check_schedule(opt.schedule_step(), 19)
 
   def test_adam_conv_fuse(self):
     with Tensor.train():
@@ -1230,7 +1231,7 @@ class TestSchedule(unittest.TestCase):
       opt = nn.optim.Adam(nn.state.get_parameters(c1), lr=1e-4)
       opt.zero_grad()
       c1(img).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 16)
+      check_schedule(opt.schedule_step(), 19)
 
   def test_adam_2convs_fuse(self):
     with Tensor.train():
@@ -1241,7 +1242,7 @@ class TestSchedule(unittest.TestCase):
       opt = nn.optim.Adam(nn.state.get_parameters([c1, c2]), lr=1e-4)
       opt.zero_grad()
       c2(c1(img).relu()).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 18)
+      check_schedule(opt.schedule_step(), 21)
 
   def test_sgd_conv_fuse(self):
     with Tensor.train():
@@ -1274,7 +1275,7 @@ class TestSchedule(unittest.TestCase):
       opt = nn.optim.SGD(nn.state.get_parameters([c1, c2]), nesterov=True, momentum=0.9, weight_decay=0.1)
       opt.zero_grad()
       c2(c1(img).relu()).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 15)
+      check_schedule(opt.schedule_step(), 13)
 
   def test_sgd_4convs_fuse(self):
     with Tensor.train():
@@ -1863,7 +1864,7 @@ class TestSchedule(unittest.TestCase):
     yt = Tensor.randn(BS, 10).realize()
     with Context(SPLIT_REDUCEOP=0):
       loss = yt.sparse_categorical_crossentropy(Y_train[samples])
-      run_schedule(check_schedule(loss, 5))
+      run_schedule(check_schedule(loss, 4))
       loss_fused = loss.numpy()
     loss_ref = torch.nn.CrossEntropyLoss()(torch.tensor(yt.numpy()), torch.tensor(Y_train.numpy())[torch.tensor(samples.numpy())])
     np.testing.assert_allclose(loss_fused, loss_ref.numpy(), atol=1e-6, rtol=1e-6)
@@ -2076,6 +2077,11 @@ class TestCopyFolding(unittest.TestCase):
     check_schedule(b, 0, filter_sink=False)
     assert b.item() == 1
 
+  def test_one_hot_with_copy(self):
+    y = Tensor([1, 2, 3]).to("CPU")
+    x = y.one_hot(10)
+    check_schedule(x, 3, filter_sink=False)
+
   def test_const_copy_multi(self):
     x = Tensor.ones(1, device="CPU").to_(["CPU", "CPU:1"])
     check_schedule(x, 0, filter_sink=False)
@@ -2085,7 +2091,7 @@ class TestCopyFolding(unittest.TestCase):
     a = Tensor.arange(3).realize()
     zeros = Tensor.zeros(3).realize()
     b = (a*zeros).to("CPU")
-    run_schedule(check_schedule(b, 2, filter_sink=False)) # TODO: 0?
+    run_schedule(check_schedule(b, 0, filter_sink=False))
     self.assertListEqual(b.tolist(), [0, 0, 0])
     self.assertEqual(b.device, "CPU")
 

test/test_symbolic_ops.py

@@ -287,7 +287,6 @@ class TestSymbolicOps(unittest.TestCase):
         symbolic = symbolic_result[:].numpy()
       np.testing.assert_allclose(symbolic, expected, atol=1e-6, rtol=0)
 
-  @unittest.expectedFailure
   def test_conv2d_ceildiv_edge_case(self):
     v = Variable('v', 11, 50_000)
     val = 39601
@@ -295,9 +294,10 @@ class TestSymbolicOps(unittest.TestCase):
     weight = Tensor.randn(256, 22, 12)
 
     result = x.conv2d(weight=weight, groups=1, stride=6, dilation=1, padding=(3, 3))
-    var_val = {v: val}
+    var_val = {v.expr: val}
     shape = tuple(sym_infer(s, var_val) for s in result.shape)
-    self.assertEqual(shape, (1, 256, 6600))  # TODO: fails if ceildiv is incorrect
+    with self.assertRaises(AssertionError):
+      self.assertEqual(shape, (1, 256, 6600))  # TODO: fails if ceildiv is incorrect
     # TODO: test output is correct
 
 if __name__ == '__main__':

test/test_tensor.py

@@ -839,12 +839,11 @@ class TestTensorMetadata(unittest.TestCase):
     self.assertEqual(y.grad.uop.metadata[0].name, "sigmoid")
     self.assertTrue(y.grad.uop.metadata[0].backward)
     si = Tensor.schedule(out, x.grad, y.grad)[-1]
-    self.assertEqual(len(si.metadata), 4, f"failed with {si.metadata}")
-    self.assertSetEqual(set(m.name for m in si.metadata), {"__mul__", "sigmoid", "relu"})
+    self.assertEqual(len(si.metadata), 3, f"failed with {si.metadata}")
+    self.assertSetEqual(set(m.name for m in si.metadata), {"sigmoid", "relu"})
     bw = [m for m in si.metadata if m.backward]
-    self.assertEqual(len(bw), 2)
-    self.assertEqual(bw[0].name, "__mul__")
-    self.assertEqual(bw[1].name, "sigmoid")
+    self.assertEqual(len(bw), 1)
+    self.assertEqual(bw[0].name, "sigmoid")
 
 class TestIdxUpcast(unittest.TestCase):
   def _find_op(self, ast: UOp, op: Ops):
@@ -920,5 +919,38 @@ class TestIdxUpcast(unittest.TestCase):
     a = Tensor.empty(2**11, 2**11, 1, dtype=dtypes.int8).permute((2, 0, 1)).expand((2**9+10, -1, -1)).contiguous()
     a.realize()
 
+class TestTensorUnique(unittest.TestCase):
+  def test_empty_bufs_unique(self):
+    a = Tensor.empty(10, 10).contiguous()
+    b = Tensor.empty(10, 10).contiguous()
+    Tensor.realize(a,b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_zeros_bufs_unique_sep(self):
+    a = Tensor.zeros(10, 10).contiguous()
+    Tensor.realize(a)
+    b = Tensor.zeros(10, 10).contiguous()
+    Tensor.realize(b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_zeros_bufs_unique(self):
+    a = Tensor.zeros(10, 10).contiguous()
+    b = Tensor.zeros(10, 10).contiguous()
+    Tensor.realize(a,b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_eye_bufs_unique(self):
+    a = Tensor.eye(10).contiguous()
+    b = Tensor.eye(10).contiguous()
+    Tensor.realize(a,b)
+    self.assertIsNot(a.uop.buffer, b.uop.buffer)
+
+  def test_times_2_not_unique(self):
+    a = Tensor.zeros(10, 10).contiguous()
+    b = a * 2
+    c = a * 2
+    Tensor.realize(b,c)
+    self.assertIs(b.uop.buffer, c.uop.buffer)
+
 if __name__ == '__main__':
   unittest.main()

test/test_tensor_uop.py

@@ -3,7 +3,7 @@ import numpy as np
 import unittest
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.engine.realize import run_schedule
-from tinygrad.uop.ops import Ops, UOp, UPat
+from tinygrad.uop.ops import UOp
 from tinygrad.helpers import SPLIT_REDUCEOP
 
 class TestTensorUOp(unittest.TestCase):
@@ -93,7 +93,6 @@ class TestTensorUOp(unittest.TestCase):
     out.realize()
     self.assertEqual(out.tolist(), Tensor.zeros(4, 8).tolist())
 
-reduce_kernel = UPat(Ops.SINK, src=(UPat(Ops.STORE, allow_any_len=True, src=(UPat(), UPat((Ops.REDUCE_AXIS, Ops.REDUCE))))))
 @unittest.skipUnless(SPLIT_REDUCEOP, "only for SPLIT_REDUCEOP")
 class TestReduceOp(unittest.TestCase):
   def test_no_split_reduce_kernel(self):
@@ -101,23 +100,18 @@ class TestReduceOp(unittest.TestCase):
     a = a.sum()
     sched = a.schedule()
     assert len(sched) == 1
-    assert reduce_kernel.match(sched[0].ast, {})
 
   def test_split_reduce_kernel_dim0(self):
     a = Tensor.rand(256, 255).realize()
     a = a.sum()
     sched = a.schedule()
     assert len(sched) == 2
-    for s in sched:
-      assert reduce_kernel.match(s.ast, {})
 
   def test_split_reduce_kernel_dim1(self):
     a = Tensor.rand(255, 256).realize()
     a = a.sum()
     sched = a.schedule()
     assert len(sched) == 2
-    for s in sched:
-      assert reduce_kernel.match(s.ast, {})
 
 if __name__ == "__main__":
   unittest.main()

test/test_uop_graph.py

@@ -1,12 +1,11 @@
-from typing import List
 import unittest, pytest
 from tinygrad import dtypes, Variable
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import DEBUG, Context
-from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, track_rewrites, graph_rewrite, GroupOp, KernelInfo
+from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, track_rewrites, graph_rewrite, GroupOp, AxisType
 from tinygrad.uop.symbolic import sym
-from tinygrad.codegen import full_rewrite, full_rewrite_to_sink
 from tinygrad.codegen.late.expander import expander
+from test.test_uops import to_uops_list
 
 simple_pm = PatternMatcher([
   (UPat.cvar('x', dtypes.int), lambda x: UOp.const(dtypes.float, 1.0) + UOp.const(dtypes.float, 2.0)),
@@ -15,12 +14,6 @@ simple_pm = PatternMatcher([
   ((UPat.var('x') + UPat.cvar('c1')) + UPat.cvar('c2'), lambda x,c1,c2: x + (c1.arg+c2.arg)),
 ])
 
-def to_uops_list(u:List[UOp]) -> List[UOp]:
-  # we strip the SINK here for legacy reasons
-  ret = full_rewrite(UOp.sink(*u, arg=KernelInfo(opts_to_apply=())))
-  assert ret[-1].op is Ops.SINK
-  return ret[:-1]
-
 class TestGraphRewriteConst(unittest.TestCase):
   def test_gep_const(self):
     v1 = UOp.const(dtypes.int.vec(3), (0,1,2))
@@ -270,6 +263,7 @@ class TestUOpGraph(unittest.TestCase):
     uops = to_uops_list([out])
     self.assertEqual(len([x for x in uops if x.op is Ops.VECTORIZE]), 0)
 
+  @unittest.skip("this test isn't valid uops")
   def test_gep_vec_fold(self):
     d0 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
     d1 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 1)
@@ -459,30 +453,29 @@ class TestUOpGraph(unittest.TestCase):
     idx = d0.index(ridx0)
     ld = idx.load()
     val = (ridx0<50).where(5, ld)
-    st = idx.store(val, ridx0)
+    st = idx.store(val).end(ridx0)
     uops = to_uops_list([st])
     for u in uops:
       assert u.op is not Ops.WHERE
       if u.op is Ops.STORE: assert u.src[1].arg==5
 
   def test_load_idx_becomes_int(self):
-    # These loads wont overflow int since we know from the gate that the value is bounded
-    r0 = UOp.range(10, 0)
-    d0 = UOp(Ops.DEFINE_GLOBAL, dtypes.long.ptr(), (), 0)
-    d1 = UOp(Ops.DEFINE_GLOBAL, dtypes.long.ptr(), (), 1)
-    l0 = UOp(Ops.LOAD, dtypes.long, (d0.index(UOp.const(dtypes.int, 0)),)).cast(dtypes.index)
-    idx = l0 * 600
-    valid = (l0<-1).ne(True)&(l0<3000)
-    l1 = UOp(Ops.LOAD, dtypes.long, (d1.index(idx.valid(valid)),))
-    uops = to_uops_list([l1])
+    # mnist indexing with split reduceop
+    # Make sure we are not doign math on the loaded index, which would promote it to long
+    c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(128000), arg=0, src=())
+    c1 = UOp.range(UOp.const(dtypes.index, 512), 1, AxisType.LOOP)
+    c2 = UOp.range(UOp.const(dtypes.index, 250), 2, AxisType.LOOP)
+    c3 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(512), arg=1, src=())
+    c4 = c3.index(c1).load()
+    c5 = UOp.range(UOp.const(dtypes.index, 240), 0, AxisType.REDUCE)
+    c6 = ((c2*UOp.const(dtypes.index, 240))+c5)
+    c7 = UOp(Ops.DEFINE_GLOBAL, dtypes.uchar.ptr(60000), arg=2, src=())
+    c8 = c7.index(c6).load()
+    c9 = ((c4<0).where((c4+60000), c4)!=c6.cast(dtypes.int)).where(0, c8.cast(dtypes.uint).cast(dtypes.uchar)).reduce(c5, arg=Ops.ADD)
+    c10 = c0.index(((c1*UOp.const(dtypes.index, 250))+c2)).store(c9).end(c1, c2)
+    uops = to_uops_list([c10])
     for u in uops:
-      if u.op is Ops.INDEX: self.assertEqual(u.src[1].dtype, dtypes.int)
-
-    valid = (10*r0<5-l0).ne(True)&(l0<3000)
-    l2 = UOp(Ops.LOAD, dtypes.long, (d1.index(idx.valid(valid)),))
-    uops = to_uops_list([l2])
-    for u in uops:
-      if u.op is Ops.INDEX: self.assertEqual(u.src[1].dtype, dtypes.int)
+      self.assertNotEqual(u.dtype, dtypes.long)
 
   def test_in_out_of_bounds_access(self):
     with Context(IGNORE_OOB=0):
@@ -518,6 +511,7 @@ class TestUOpGraph(unittest.TestCase):
       st1 = UOp(Ops.STORE, dtypes.void, (glbl0.index(v), v, v<20))
       with self.assertRaises(RuntimeError): to_uops_list([st1])
 
+  @unittest.skip("if not allowed in graph")
   def test_in_bounds_access_gated_local(self):
     with Context(IGNORE_OOB=0):
       # Define buffers
@@ -579,7 +573,7 @@ class TestUOpGraph(unittest.TestCase):
   def test_in_out_bounds_access_with_mask(self):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
-      gidx0 = UOp(Ops.SPECIAL, dtypes.index, (UOp.const(dtypes.index, 42),), "gidx0")
+      gidx0 = UOp.range(42, 0, AxisType.GLOBAL)
       ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0, (5<gidx0)&(gidx0<16)),))
       ld1 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0, gidx0<16),))
       to_uops_list([ld0, ld1])
@@ -603,7 +597,7 @@ class TestUOpGraph(unittest.TestCase):
     with Context(IGNORE_OOB=0):
       glbl0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(16), (), 0)
       glbl1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(64), (), 0)
-      gidx0 = UOp(Ops.SPECIAL, dtypes.index, (UOp.const(dtypes.index, 42),), "gidx0")
+      gidx0 = UOp.range(42, 0, AxisType.GLOBAL)
       ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(gidx0, gidx0<8),)).cast(dtypes.index)
       ld1 = UOp(Ops.LOAD, dtypes.int, (glbl1.index(ld0*2, (ld0>=0)&(ld0<32)),))
       to_uops_list([ld1])
@@ -638,13 +632,13 @@ class TestUOpGraph(unittest.TestCase):
     lidx = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 16),), "lidx0")
     st = UOp(Ops.STORE, dtypes.void, (smem.index(lidx), UOp.load(glbl0.index(lidx), dtype=dtypes.int)))
     barrier = UOp(Ops.BARRIER, dtypes.void, (st, ))
-    ld0 = UOp(Ops.LOAD, dtypes.int, (smem.index(UOp.invalid()), barrier))
-    ld1 = UOp(Ops.LOAD, dtypes.int, (smem.index(lidx+2, UOp.const(dtypes.bool, True)), barrier))
+    ld0 = UOp(Ops.LOAD, dtypes.int, (smem.after(barrier).index(UOp.invalid()),))
+    ld1 = UOp(Ops.LOAD, dtypes.int, (smem.after(barrier).index(lidx+2, UOp.const(dtypes.bool, True)),))
     uops = to_uops_list([UOp(Ops.STORE, dtypes.void, (glbl0.index(lidx), ld1+ld0))])
 
     ld0 = uops[-1].src[-1]
     # the gate and invalid value are deleted from ld1
-    self.assertEqual(ld0.src[0], smem.index(lidx+2))
+    self.assertEqual(ld0.src[0], smem.after(barrier).index(lidx+2))
 
   def test_fold_gated_store(self):
     glbl = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
@@ -665,19 +659,6 @@ class TestUOpGraph(unittest.TestCase):
     bad_gate = UOp.const(dtypes.int, 1)
     with self.assertRaises(AssertionError): to_uops_list([UOp(Ops.STORE, dtypes.void, (glbl0, idx, UOp.const(dtypes.int, 42), bad_gate))])
 
-  def test_switched_range_order(self):
-    glbl = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(), (), 0)
-    cf = UOp.const(dtypes.float, 0.0)
-    r1 = UOp.range(2, 0)
-    r2 = UOp.range(2, 1)
-    alu = UOp(Ops.MUL, dtypes.int, (r2, r1))
-    store = UOp(Ops.STORE, dtypes.void, (glbl.index(alu), cf))
-    uops = to_uops_list([store])
-    ranges = [x for x in uops if x.op is Ops.RANGE]
-    endranges = [x for x in uops if x.op is Ops.ENDRANGE]
-    # ranges are closed in the right order
-    self.assertEqual(endranges[-1].src[0], ranges[0])
-
 @track_rewrites()
 def expander_rewrite(sink): return graph_rewrite(sink, sym + expander)
 
@@ -739,7 +720,7 @@ class TestExpander(unittest.TestCase):
     self.assertTupleEqual(sink.src[0].arg, (0,2,1,3,4,6,5,7))
 
   def test_contract_no_expand(self):
-    e1 = UOp(Ops.DEFINE_VAR, dtypes.int)
+    e1 = UOp.variable("i", 0, 10, dtype=dtypes.int)
     con = UOp(Ops.CONTRACT, dtypes.int.vec(2), (e1,), ((2,2),))
     sink = expander_rewrite(con)
     assert sink.op is Ops.VECTORIZE and len(sink.src) == 2
@@ -828,60 +809,6 @@ class TestExpander(unittest.TestCase):
     sink = expander_rewrite(sink)
     print(sink)
 
-class TestIFUOps(unittest.TestCase):
-  def test_create_ifs(self):
-    gbuf = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
-    sbuf = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(size=4, addrspace=AddrSpace.LOCAL), (), "smem")
-    valid = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 10),), "gidx0")<5
-    lidx = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), "lidx0")
-    gate = valid&(lidx.ne(2))
-    idx = UOp.const(dtypes.int, 0)
-    st = UOp(Ops.STORE, dtypes.void, (sbuf.index(idx), UOp.const(dtypes.float, 42)))
-    barrier = UOp(Ops.BARRIER, dtypes.void, (st,))
-    lbuf = UOp(Ops.LOAD, dtypes.float, (sbuf.index(UOp.const(dtypes.int, 0)), barrier))
-    store = UOp(Ops.STORE, dtypes.void, (gbuf.index(UOp.const(dtypes.int, 0), gate), lbuf))
-    sink = UOp(Ops.SINK, dtypes.void, (store,))
-    sink = full_rewrite_to_sink(sink)
-    if_uops = [u for u in sink.toposort() if u.op is Ops.IF]
-    self.assertEqual(len(if_uops), 1)
-    self.assertEqual(if_uops[0].src[0], gate)
-    for st in sink.src:
-      self.assertEqual(len(st.src), 2)
-
-  def test_expand_ifs_one_gate(self):
-    gbuf = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
-    sbuf = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(size=16, addrspace=AddrSpace.LOCAL), (), "smem")
-    valid = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), "gidx0")<1
-    lidx = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 16),), "lidx0")
-    gate = valid&(lidx.ne(2))
-    st = UOp(Ops.STORE, dtypes.void, (sbuf, lidx, UOp.const(dtypes.float, 42)))
-    barrier = UOp(Ops.BARRIER, dtypes.void, (st,))
-    lbufs = [UOp(Ops.LOAD, dtypes.float, (sbuf.index(UOp.const(dtypes.int, i)), barrier)) for i in range(4)]
-    stores = [UOp(Ops.STORE, dtypes.void, (gbuf.index(UOp.const(dtypes.int, i), gate), lbufs[i])) for i in range(4)]
-    sink = UOp(Ops.SINK, dtypes.void, tuple(stores))
-    sink = full_rewrite_to_sink(sink)
-    if_uops = [u for u in sink.toposort() if u.op is Ops.IF]
-    self.assertEqual(len(if_uops), 1)
-    self.assertEqual(if_uops[0].src[0], gate)
-    for st in sink.src:
-      self.assertEqual(len(st.src), 2)
-
-  # this will be fixed with the merge gated stores bounty
-  @unittest.expectedFailure
-  def test_expand_ifs_dumb(self):
-    buf = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
-    valid = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 10),), "gidx0")<5
-    lidx = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), "lidx0")
-    gate = valid&(lidx.ne(2))
-    stores = [UOp(Ops.STORE, dtypes.void, (buf, UOp.const(dtypes.int, i), UOp.const(dtypes.float, i), gate)) for i in range(4)]
-    sink = UOp(Ops.SINK, dtypes.void, tuple(stores))
-    sink = full_rewrite_to_sink(sink)
-    if_uops = [u for u in sink.toposort() if u.op is Ops.IF]
-    self.assertEqual(len(if_uops), 1)
-    self.assertEqual(if_uops[0].src[0], gate)
-    for st in sink.src:
-      self.assertEqual(len(st.src), 2)
-
 class TestUOpTags(unittest.TestCase):
   def test_inc_by_one(self):
     g = UOp.const(dtypes.int, 1) + UOp.const(dtypes.int, 1)

test/test_uops.py

@@ -6,7 +6,7 @@ from tinygrad.helpers import CI, DEBUG, getenv, Timing
 from tinygrad.dtype import dtypes, DType, AddrSpace
 from tinygrad.device import Buffer, Device
 from tinygrad.uop.ops import Ops, UOp, UPat, KernelInfo, exec_alu # noqa F401
-from tinygrad.uop.spec import spec
+from tinygrad.uop.spec import shared_spec
 from tinygrad.renderer import ProgramSpec
 from tinygrad.engine.realize import CompiledRunner, get_program
 from tinygrad.codegen import full_rewrite
@@ -15,10 +15,16 @@ from tinygrad.device import is_dtype_supported
 from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad.renderer.ptx import PTXRenderer
 
-def to_uops_list(u:list[UOp], opts=None, skip_check=False) -> list[UOp]: return full_rewrite(UOp.sink(*u), opts)
+def to_uops_list(u:list[UOp], ren=None) -> list[UOp]:
+  sink = UOp.group(*u)
+  for r in sink.ranges: sink = sink.end(r)
+  # we strip the SINK here for legacy reasons
+  ret = full_rewrite(sink.sink(arg=KernelInfo(opts_to_apply=())), ren)
+  assert ret[-1].op is Ops.SINK
+  return ret[:-1]
 
 def _uops_to_prg(uops_list):
-  uops = full_rewrite(ast:=UOp.sink(*uops_list), opts=Device[Device.DEFAULT].renderer)
+  uops = full_rewrite(ast:=UOp.sink(*uops_list), ren=Device[Device.DEFAULT].renderer)
   src = Device[Device.DEFAULT].renderer.render(uops)
   has_local = Device[Device.DEFAULT].renderer.has_local
   return CompiledRunner(ProgramSpec(uops[-1].arg.name if uops[-1].arg is not None else "test", src, Device.DEFAULT, ast, uops=uops,
@@ -109,7 +115,7 @@ class TestFloatUOps(TestUOps):
   def test_log2(self): self._test_uop_fxn(Ops.LOG2, lambda a: math.log2(a) if a > 0 else float('-inf' if a==0 else 'nan'))
   @unittest.skipIf(Device.DEFAULT == "CPU", 'not supported as uop')
   def test_sin(self): self._test_uop_fxn(Ops.SIN, lambda a: math.sin(a))
-  def test_recip(self): self._test_uop_fxn(Ops.RECIP, lambda a: 1/a if a != 0 else float('inf'))
+  def test_recip(self): self._test_uop_fxn(Ops.RECIPROCAL, lambda a: 1/a if a != 0 else float('inf'))
   def test_sqrt(self): self._test_uop_fxn(Ops.SQRT, lambda a: math.sqrt(a) if a >= 0 else float('nan'))
 
   def test_add(self): self._test_bop_fxn(Ops.ADD, lambda a,b: a+b)
@@ -212,18 +218,18 @@ class TestExecALU(TestUOps):
     self.assertEqual(exec_alu(Ops.IDIV, dtypes.int8, (7, -3)), -2)
     self.assertEqual(exec_alu(Ops.IDIV, dtypes.int8, (-50, 6)), -8)
 
-    np.testing.assert_allclose(exec_alu(Ops.MUL, dtypes.float32, (7.0, exec_alu(Ops.RECIP, dtypes.float32, (3.0,)))), 2+(1.0/3.0))
-    np.testing.assert_allclose(exec_alu(Ops.MUL, dtypes.float32, (7.0, exec_alu(Ops.RECIP, dtypes.float32, (-3.0,)))), -2-(1.0/3.0))
+    np.testing.assert_allclose(exec_alu(Ops.MUL, dtypes.float32, (7.0, exec_alu(Ops.RECIPROCAL, dtypes.float32, (3.0,)))), 2+(1.0/3.0))
+    np.testing.assert_allclose(exec_alu(Ops.MUL, dtypes.float32, (7.0, exec_alu(Ops.RECIPROCAL, dtypes.float32, (-3.0,)))), -2-(1.0/3.0))
 
   def test_recip(self):
-    np.testing.assert_allclose(exec_alu(Ops.RECIP, dtypes.float32, (8,)), 1/8)
-    np.testing.assert_allclose(exec_alu(Ops.RECIP, dtypes.float32, (7,)), 1/7)
-    np.testing.assert_allclose(exec_alu(Ops.RECIP, dtypes.float32, (-3,)), 1/-3)
-    np.testing.assert_allclose(exec_alu(Ops.RECIP, dtypes.float32, (-50,)), 1/-50)
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (8,)), 1/8)
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (7,)), 1/7)
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (-3,)), 1/-3)
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (-50,)), 1/-50)
 
-    np.testing.assert_allclose(exec_alu(Ops.RECIP, dtypes.float32, ((32+521+3),)), 1/(32+521+3))
-    np.testing.assert_allclose(exec_alu(Ops.RECIP, dtypes.float32, ((34**2),)), 1/(34**2))
-    np.testing.assert_allclose(exec_alu(Ops.RECIP, dtypes.float32, (10,)), 1/10)
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, ((32+521+3),)), 1/(32+521+3))
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, ((34**2),)), 1/(34**2))
+    np.testing.assert_allclose(exec_alu(Ops.RECIPROCAL, dtypes.float32, (10,)), 1/10)
 
   def test_bool_cmplt(self):
     self.assertEqual(exec_alu(Ops.CMPLT, dtypes.bool, (False, False)), False)
@@ -266,6 +272,7 @@ class TestConstantFolding(unittest.TestCase):
     si = t.schedule()
     assert len(si) == 0
 
+@unittest.skip("no more if statements")
 class TestGatedStoreRewrite(unittest.TestCase):
   def test_tiny_gate_store(self):
     gmem = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
@@ -302,6 +309,7 @@ class TestGatedStoreRewrite(unittest.TestCase):
     self.assertIs(gated_uops[-1].op, Ops.STORE)
 
   # scaled down version of TestLinearizerDumb.test_unmerged_ifs
+  @unittest.skip("we don't merge ifs anymore")
   def test_merge_ifs_alt(self):
     gmem0 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 0)
     gmem1 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(), (), 1)
@@ -331,7 +339,7 @@ class TestLocalAccess(unittest.TestCase):
     smem = uop(uops, Ops.DEFINE_LOCAL, dtypes.float32.ptr(size=16, addrspace=AddrSpace.LOCAL), (), 'smem')
     st = uop(uops, Ops.STORE, dtypes.void, (smem.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), uop(uops, Ops.CONST, dtypes.float32, (), 42.0)))
     barr = uop(uops, Ops.BARRIER, dtypes.void, (st,))
-    sres = uop(uops, Ops.LOAD, dtypes.float32, (smem.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), barr))
+    sres = uop(uops, Ops.LOAD, dtypes.float32, (smem.after(barr).index(uop(uops, Ops.CONST, dtypes.int32, (), 0)),))
     self.assertEqual(_test_uops_result(dtypes.float32, uops, sres), 42)
 
   # NOTE: webgpu specific, since only webgpu performs bitpacking
@@ -341,7 +349,7 @@ class TestLocalAccess(unittest.TestCase):
     smem = uop(uops, Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=16, addrspace=AddrSpace.LOCAL), (), 'smem')
     st = uop(uops, Ops.STORE, dtypes.void, (smem.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), uop(uops, Ops.CONST, dtypes.uint8, (), 42)))
     barr = uop(uops, Ops.BARRIER, dtypes.void, (st,))
-    sres = uop(uops, Ops.LOAD, dtypes.uint8, (smem.index(uop(uops, Ops.CONST, dtypes.int32, (), 0)), barr))
+    sres = uop(uops, Ops.LOAD, dtypes.uint8, (smem.after(barr).index(uop(uops, Ops.CONST, dtypes.int32, (), 0)),))
     self.assertEqual(_test_uops_result(dtypes.uint8, uops, sres), 42)
 
   # NOTE: webgpu specific, since only webgpu performs bitpacking
@@ -351,7 +359,7 @@ class TestLocalAccess(unittest.TestCase):
     size = 16
     for dtype in _dtypes:
       temp = UOp(Ops.DEFINE_LOCAL, dtype.ptr(size=size, addrspace=AddrSpace.LOCAL), (), 'smem')
-      uops = to_uops_list([temp], opts=Device[Device.DEFAULT].renderer)
+      uops = to_uops_list([temp], ren=Device[Device.DEFAULT].renderer)
       out = Device[Device.DEFAULT].renderer.render(uops)
       # half is supported in wgsl, so it doesn't have to be packed
       corrected_size = size//(4//dtype.itemsize) if dtype != dtypes.half else size
@@ -378,7 +386,7 @@ class TestAssembly(unittest.TestCase):
     l1 = UOp(Ops.LOAD, dtypes.int, (g1.index(c1),))
     a1 = UOp(Ops.MUL, dtypes.int, (l1, c1))
     a2 = UOp(Ops.MUL, dtypes.int, (l1, c2))
-    uops = to_uops_list([a1,a2], opts=Device[Device.DEFAULT].renderer)
+    uops = to_uops_list([a1,a2], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
     ops = [x.op for x in uops]
     self.assertIn(Ops.SHL, ops)
@@ -390,7 +398,7 @@ class TestAssembly(unittest.TestCase):
       c = UOp(Ops.CONST, dt, (), 2)
       l = UOp(Ops.LOAD, dt, (g.index(c),))
       a = UOp(Ops.IDIV, dt, (l, c))
-      uops = to_uops_list([a], opts=Device[Device.DEFAULT].renderer)
+      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
       Device[Device.DEFAULT].renderer.render(uops)
       ops = [x.op for x in uops]
       self.assertIn(Ops.SHR, ops, f"For dtype={dt} divison by power of two did not simplify to shift")
@@ -401,14 +409,14 @@ class TestAssembly(unittest.TestCase):
     c = UOp(Ops.CONST, dtypes.uint, (), 3)
     l = UOp(Ops.LOAD, dtypes.uint, (g.index(c),))
     a = UOp(Ops.IDIV, dtypes.uint, (l, c))
-    uops = to_uops_list([a], opts=Device[Device.DEFAULT].renderer)
+    uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
     ops = [x.op for x in uops]
     self.assertIn(Ops.SHR, ops)
     self.assertNotIn(Ops.IDIV, ops)
 
     b = UOp(Ops.MOD, dtypes.uint, (l, c))
-    uops = to_uops_list([b], opts=Device[Device.DEFAULT].renderer)
+    uops = to_uops_list([b], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
     ops = [x.op for x in uops]
     self.assertIn(Ops.SHR, ops)
@@ -421,7 +429,7 @@ class TestAssembly(unittest.TestCase):
     c = UOp(Ops.CONST, dtypes.uint, (), 7)
     l = UOp(Ops.LOAD, dtypes.uint, (g.index(c),))
     a = UOp(Ops.IDIV, dtypes.uint, (l, c))
-    uops = to_uops_list([a], opts=Device[Device.DEFAULT].renderer)
+    uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
     ops = [x.op for x in uops]
     self.assertIn(Ops.SHR, ops)
@@ -429,7 +437,7 @@ class TestAssembly(unittest.TestCase):
 
   def test_fast_idiv_remove_powers_of_two(self):
     ridx = UOp.range(2**20, 0)
-    uops = to_uops_list([ridx//(7*64)], opts=Device[Device.DEFAULT].renderer)
+    uops = to_uops_list([ridx//(7*64)], ren=Device[Device.DEFAULT].renderer)
     ops = [x.op for x in uops]
     # this requires shifting out the powers of two before doing fast_idiv
     # (((ridx0>>6)*18725)>>17) instead of (int)((((long)(ridx0)*1198373)>>29))
@@ -453,7 +461,7 @@ class TestAssembly(unittest.TestCase):
     c = UOp(Ops.CONST, dtypes.uint, (), 7)
     l = UOp(Ops.LOAD, dtypes.uint, (g.index(c),))
     comp = l.ne(c).ne(True)
-    uops = to_uops_list([comp], opts=Device[Device.DEFAULT].renderer)
+    uops = to_uops_list([comp], ren=Device[Device.DEFAULT].renderer)
     Device[Device.DEFAULT].renderer.render(uops)
     ops = [x.op for x in uops]
     self.assertIn(Ops.CMPEQ, ops)
@@ -512,7 +520,7 @@ class TestUOpStr(unittest.TestCase):
 class TestUPatHelpers(unittest.TestCase):
   def test_location(self):
     self.assertEqual(sym.patterns[-1][0].location[0].replace("\\", "/").split("/")[-1], "symbolic.py")
-    self.assertEqual(spec.patterns[0][0].location[0].replace("\\", "/").split("/")[-1], "spec.py")
+    self.assertEqual(shared_spec.patterns[0][0].location[0].replace("\\", "/").split("/")[-1], "spec.py")
     test_upat = UPat(Ops.CONST, dtypes.bool)
     self.assertEqual(test_upat.location[0].split("/")[-1], __file__.replace("\\", "/").split("/")[-1])
     test_upat_named = test_upat.named("test_name")
@@ -540,11 +548,24 @@ class TestUopsObject(unittest.TestCase):
 
 class TestUOpRender(unittest.TestCase):
   def test_render_vectorize_same(self):
-    u = UOp(Ops.VECTORIZE, src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0)))
-    self.assertEqual(u.render(), "{0, ...}")
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(3), src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0)))
+    self.assertEqual(u.render(simplify=False), "{0, ...}")
   def test_render_vectorize_different(self):
-    u = UOp(Ops.VECTORIZE, src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1), UOp.const(dtypes.int, 2)))
-    self.assertEqual(u.render(), "{0,1,2}")
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(3), src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1), UOp.const(dtypes.int, 2)))
+    self.assertEqual(u.render(simplify=False), "{0,1,2}")
+  def test_render_vectorize_same_simplified(self):
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(3), src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 0)))
+    self.assertEqual(u.render(), "0")
+  def test_render_vectorize_different_simplified(self):
+    u = UOp(Ops.VECTORIZE, dtype=dtypes.int.vec(3), src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1), UOp.const(dtypes.int, 2)))
+    self.assertEqual(u.render(), "(0, 1, 2)")
+
+class TestZeroRange(unittest.TestCase):
+  def test_reduce_variable(self):
+    for i in range(3,-1,-1):
+      v = UOp.variable("i", 0, 5).bind(i)
+      out = Tensor.ones(10, dtype=dtypes.int).contiguous().shrink(((0,v),)).sum()
+      self.assertEqual(out.item(), i)
 
 if __name__ == '__main__':
   unittest.main(verbosity=2)

test/unit/test_block_reorder.py

@@ -1,76 +0,0 @@
-import unittest, random
-from tinygrad.dtype import dtypes
-from tinygrad.uop.ops import print_uops, UOp, Ops
-from tinygrad.codegen.late.linearize import block_reorder
-from tinygrad.renderer.cstyle import OpenCLRenderer
-
-def is_toposorted(lst:list[UOp]):
-  seen = set()
-  for u in lst:
-    if any(p not in seen for p in u.src): return False
-    seen.add(u)
-  return True
-
-class TestBlockReorder(unittest.TestCase):
-  def _test_randomize(self, golden:list[UOp]):
-    # test random order is always same
-    for _ in range(50):
-      # shuffle and form a valid toposort
-      lst = golden[:]
-      random.shuffle(lst)
-      topolst = []
-      for u in lst:
-        for p in u.toposort():
-          if p not in topolst: topolst.append(p)
-      assert is_toposorted(topolst)
-
-      for x,y in zip(golden, this_order:=block_reorder(topolst)):
-        if x is not y:
-          print_uops(golden)
-          print_uops(this_order)
-        self.assertIs(x, y)
-
-  def _test_render(self, golden:list[UOp]):
-    return OpenCLRenderer().render(golden)
-
-  def test_loads(self):
-    a = UOp(Ops.DEFINE_GLOBAL, dtype=dtypes.float.ptr(), arg=0)
-    b = UOp(Ops.DEFINE_GLOBAL, dtype=dtypes.float.ptr(), arg=1)
-    c = UOp(Ops.DEFINE_GLOBAL, dtype=dtypes.float.ptr(), arg=2)
-    v1 = UOp(Ops.SPECIAL, dtype=dtypes.int, src=(UOp.const(dtypes.int, 4),), arg="gidx0")
-    v2 = UOp(Ops.SPECIAL, dtype=dtypes.int, src=(UOp.const(dtypes.int, 4),), arg="gidx1")
-    v1 = v1*27
-    v2 = v2*4
-    loads = [
-      a.index(v1).load(dtype=dtypes.float),
-      a.index(v1+1).load(dtype=dtypes.float),
-      a.index(v1+2).load(dtype=dtypes.float),
-      a.index(v1+3).load(dtype=dtypes.float),
-      b.index(v2).load(dtype=dtypes.float),
-      b.index(v2+1).load(dtype=dtypes.float),
-      b.index(v2+2).load(dtype=dtypes.float),
-      b.index(v2+3).load(dtype=dtypes.float)]
-    #random.shuffle(loads)
-    sink = c.store(sum(loads)).sink()
-
-    # determine golden order
-    golden = block_reorder(list(sink.toposort()))
-
-    # render for test
-    print(self._test_render(golden))
-    #print_uops(golden)
-
-    # assert the loads are in this order
-    self.assertListEqual([g.src[0].src[1].render() for g in golden if g.op is Ops.LOAD],
-                         ['(gidx1*4)', '((gidx1*4)+1)', '((gidx1*4)+2)', '((gidx1*4)+3)',
-                          '(gidx0*27)', '((gidx0*27)+1)', '((gidx0*27)+2)', '((gidx0*27)+3)'])
-
-    # assert math is after loads
-    first_math = [i for i,g in enumerate(golden) if g.op is Ops.ADD and g.dtype == dtypes.float][0]
-    assert not any(x.op is Ops.LOAD for x in golden[first_math:])
-
-    # confirm the sort is stable
-    self._test_randomize(golden)
-
-if __name__ == '__main__':
-  unittest.main()

test/unit/test_disk_tensor.py

@@ -433,17 +433,15 @@ class TestDiskTensorMovement(unittest.TestCase):
     t = Tensor(self.fn)
     self.assertListEqual(t[16:18].tolist(), [16,17])
 
-  # TODO: fix this! at least assert on it
-  @unittest.expectedFailure
   def test_slice_read_cat(self):
     t = Tensor(self.fn)
-    self.assertListEqual(Tensor.cat(t[16:18], t[20:22]).tolist(), [16,17,20,21])
+    with self.assertRaises(AssertionError):
+      self.assertListEqual(Tensor.cat(t[16:18], t[20:22]).tolist(), [16,17,20,21])
 
-  # TODO: fix this! at least assert on it
-  @unittest.expectedFailure
   def test_slice_sum(self):
     t = Tensor(self.fn)
-    self.assertListEqual((t[16:18]+t[20:22]).tolist(), [16+20,17+21])
+    with self.assertRaises(AssertionError):
+      self.assertListEqual((t[16:18]+t[20:22]).tolist(), [16+20,17+21])
 
 if __name__ == "__main__":
   unittest.main()

test/unit/test_graph_rewrite.py

@@ -1,11 +1,12 @@
 import unittest, math
 from tinygrad import dtypes
-from tinygrad.helpers import all_same
+from tinygrad.helpers import all_same, Context
 from tinygrad.uop.ops import GroupOp, UOp, Ops, exec_alu, PatternMatcher, TrackedPatternMatcher, UPat
 from tinygrad.codegen import full_rewrite_to_sink
 from hypothesis import given, strategies as strat
 
 # Helper function to apply the graph rewrite
+@Context(SPEC=0)
 def apply_rewrite(expr):
   return full_rewrite_to_sink(expr.sink()).src[0]
 
@@ -305,19 +306,19 @@ class TestRecurse(unittest.TestCase):
     graph_rewrite(a, pm, bottom_up=True)
 
   def test_inf_loop(self):
-    a = UOp.variable('a', 0, 10)
+    a = UOp.const(dtypes.int, 3)
     pm = PatternMatcher([
-      (UPat(Ops.DEFINE_VAR, name="x"), lambda x: x.replace(op=Ops.CONST)),
-      (UPat(Ops.CONST, name="x"), lambda x: x.replace(op=Ops.DEFINE_VAR)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
     ])
     with self.assertRaises(RuntimeError):
       graph_rewrite(a, pm)
 
   def test_inf_loop_bottom_up(self):
-    a = UOp.variable('a', 0, 10)
+    a = UOp.const(dtypes.int, 3)
     pm = PatternMatcher([
-      (UPat(Ops.DEFINE_VAR, name="x"), lambda x: x.replace(op=Ops.CONST)),
-      (UPat(Ops.CONST, name="x"), lambda x: x.replace(op=Ops.DEFINE_VAR)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
     ])
     with self.assertRaises(RuntimeError):
       graph_rewrite(a, pm, bottom_up=True)

test/unit/test_hashing.py

@@ -3,6 +3,8 @@ import hashlib, random, unittest
 from tinygrad import Tensor, Device, getenv, dtypes
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import CI
+from tinygrad.uop.ops import UOp
+from tinygrad.engine.jit import TinyJit
 
 @unittest.skipUnless(is_dtype_supported(dtypes.uint8) and is_dtype_supported(dtypes.uint64), "Device must support uint8 and uint64")
 @unittest.skipIf(getenv("MOCKGPU") and Device.DEFAULT == "NV", "crashes in NV CI")
@@ -72,5 +74,52 @@ class TestKeccak(unittest.TestCase):
     data = b"\x00" * 1000
     self.assertEqual(bytes(Tensor(data).keccak("shake_128").tolist()), hashlib.shake_128(data).digest(16))
 
+  def test_variable_bs(self):
+    data = Tensor([b"abc", b"abc", b"abc"], dtype=dtypes.uint8).repeat(2048, 1)
+
+    bs = UOp.variable("bs", 1, 4096).bind(1)
+    out = data.shrink_to(bs, data.shape[-1]).keccak().shrink_to(1, 32)
+    self.assertEqual(bytes(out[0].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+
+    bs = UOp.variable("bs", 1, 4096).bind(2)
+    out = data.shrink_to(bs, data.shape[-1]).keccak().shrink_to(2, 32)
+    self.assertEqual(bytes(out[0].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+    self.assertEqual(bytes(out[1].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+
+    bs = UOp.variable("bs", 1, 4096).bind(3)
+    data = Tensor([b"abc", b"abc", b"def"], dtype=dtypes.uint8).repeat(2048, 1)
+    out = data.shrink_to(bs, data.shape[-1]).keccak().shrink_to(3, 32)
+    self.assertEqual(bytes(out[0].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+    self.assertEqual(bytes(out[1].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+    self.assertEqual(bytes(out[2].tolist()), bytearray.fromhex("8e0d8f672252acb0 ffc5093db8653b18 1513bf9a2097e737 b4f73533dcaf46df"))
+
+  def test_variable_bs_jit(self):
+    def f(data):
+      return data.keccak()
+    jit_f = TinyJit(f)
+
+    data = Tensor([b"abc", b"abc", b"abc"], dtype=dtypes.uint8).repeat(2048, 1)
+
+    # initialize jit
+    for _ in range(3):
+      bs = UOp.variable("bs", 1, 4096).bind(4096)
+      _ = jit_f(data.shrink_to(bs, data.shape[-1]))
+
+    bs = UOp.variable("bs", 1, 4096).bind(1)
+    out = jit_f(data.shrink_to(bs, data.shape[-1])).shrink_to(1, 32)
+    self.assertEqual(bytes(out[0].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+
+    bs = UOp.variable("bs", 1, 4096).bind(2)
+    out = jit_f(data.shrink_to(bs, data.shape[-1])).shrink_to(2, 32)
+    self.assertEqual(bytes(out[0].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+    self.assertEqual(bytes(out[1].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+
+    bs = UOp.variable("bs", 1, 4096).bind(3)
+    data = Tensor([b"abc", b"abc", b"def"], dtype=dtypes.uint8).repeat(2048, 1)
+    out = jit_f(data.shrink_to(bs, data.shape[-1])).shrink_to(3, 32)
+    self.assertEqual(bytes(out[0].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+    self.assertEqual(bytes(out[1].tolist()), bytearray.fromhex("3a985da74fe225b2 045c172d6bd390bd 855f086e3e9d525b 46bfe24511431532"))
+    self.assertEqual(bytes(out[2].tolist()), bytearray.fromhex("8e0d8f672252acb0 ffc5093db8653b18 1513bf9a2097e737 b4f73533dcaf46df"))
+
 if __name__ == "__main__":
   unittest.main()

test/unit/test_kernelize.py

@@ -20,8 +20,8 @@ class TestKernelize(unittest.TestCase):
     self.assertEqual(len([s for s in a0.uop.toposort() if s.op is Ops.KERNEL]), 2)
     self.assertIs(a1.uop.base.op, Ops.REDUCE_AXIS)
     # input Tensor and user contiguous kernelize
-    self.assertIs(a0.uop.base.op, Ops.ASSIGN)
-    self.assertIs(a.uop.base.op, Ops.ASSIGN)
+    self.assertIs(a0.uop.base.op, Ops.AFTER)
+    self.assertIs(a.uop.base.op, Ops.AFTER)
 
   def test_two_reduce_w_add(self):
     a = Tensor.ones(16,16).contiguous()
@@ -31,7 +31,7 @@ class TestKernelize(unittest.TestCase):
     # NOTE: the +1 is fused with a1, so a1 is not kernelized
     self.assertIs(a1.uop.base.op, Ops.REDUCE_AXIS)
     # the input to the REDUCE_AXIS is an ASSIGN though
-    self.assertIs(a1.uop.base.src[0].base.op, Ops.ASSIGN)
+    self.assertIs(a1.uop.base.src[0].base.op, Ops.AFTER)
 
 if __name__ == '__main__':
   unittest.main()

test/unit/test_pattern_matcher.py

@@ -50,7 +50,7 @@ class TestPatternMatcher(unittest.TestCase):
     def fxn(ctx, x):
       ctx.append(True)
       assert len(x.src) == 0
-      return UOp(Ops.CONST, src=(UOp(Ops.CONST),))
+      return x.replace(src=(UOp(Ops.DEVICE, arg="blah"),))
     matcher = PatternMatcher([(UPat(Ops.CONST, src=(), name="x"), fxn)])
     c1 = UOp(Ops.CONST, dtypes.float, arg=1.0)
     # second rewrite shouldn't match anything

test/unit/test_simplify_valid_idx.py

@@ -5,7 +5,7 @@ from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import UOp, Ops
 from tinygrad.uop.symbolic import simplify_valid
 from tinygrad.helpers import Context
-from .test_uop_symbolic import check_uop_against_string
+from test.unit.test_uop_symbolic import check_uop_against_string
 
 def get_gated_load_uop(valid:UOp, idx:UOp):
   return UOp(Ops.LOAD, dtypes.float, (
@@ -41,13 +41,13 @@ class TestHelpers(unittest.TestCase):
     self.assertTrue(f2.is_increasing())
     self.assertTrue(f3.is_increasing())
 
-    rng = UOp(Ops.RANGE, dtypes.int, arg=(2, True), src=(UOp(Ops.CONST, dtypes.int, arg=5, src=()),))
+    rng = UOp.range(5, 2)
     self.assertTrue(rng.is_increasing())
     self.assertTrue((rng+2).is_increasing())
 
 class TestValidIdxSimplification(unittest.TestCase):
   def check(self, load, sidx, svalid):
-    with Context(NOOPT=1):
+    with Context(NOOPT=1, SPEC=0):
       load = full_rewrite_to_sink(load.sink()).src[0]
     idx, valid = load.src[0].src[1], load.src[0].src[2]
     check_uop_against_string(self, idx, sidx)
@@ -213,7 +213,7 @@ class TestValidIdxSimplification(unittest.TestCase):
 
 class TestImageSimplification(unittest.TestCase):
   def check(self, load, svalid, sidx0, sidx1):
-    with Context(NOOPT=1):
+    with Context(NOOPT=1, SPEC=0):
       load = full_rewrite_to_sink(load.sink()).src[0]
     idx = load.src[0].src[1]
     self.assertEqual(idx.op, Ops.VECTORIZE)
@@ -283,7 +283,8 @@ class TestImageSimplification(unittest.TestCase):
 
     # empty -> invalid
     load = get_load_image_uop(shape, (gidx0<8) & (gidx0<8).ne(True), idx)
-    load = full_rewrite_to_sink(load.sink()).src[0]
+    with Context(NOOPT=1, SPEC=0):
+      load = full_rewrite_to_sink(load.sink()).src[0]
     self.assertEqual(load.op, Ops.VECTORIZE)
     self.assertEqual(load.dtype.count, 4)
 

test/unit/test_uop_symbolic.py

@@ -7,7 +7,7 @@ from tinygrad.codegen import full_rewrite
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, sym_infer
 from tinygrad.uop.symbolic import sym, commutative
-from tinygrad.uop.spec import uops_to_z3
+from tinygrad.uop.validate import uops_to_z3
 
 def check_uop_against_string(self, v:UOp, s:str):
   sym_vars = {v.render():v for v in v.toposort() if v.op in (Ops.DEFINE_VAR, Ops.RANGE, Ops.SPECIAL)}

test/unit/test_uop_vmin_vmax.py

@@ -40,15 +40,14 @@ class TestVminVmaxProperties(unittest.TestCase):
     self.assertEqual(uop.vmin, 0)
     self.assertEqual(uop.vmax, 5)
 
-    # this can be improved
     uop = x & 15
     self.assertEqual(uop.vmin, 0)
     self.assertEqual(uop.vmax, 15)
 
-    # this can be improved
+    # TODO: this can be improved
     uop = x & 32
     self.assertEqual(uop.vmin, 0)
-    self.assertEqual(uop.vmax, 20)
+    self.assertEqual(uop.vmax, 20) # shoud be 0
 
   def test_vmin_vmax_multiplication_with_variable(self):
     # vmin and vmax for multiplication with a variable

test/unit/test_upat_compile.py

@@ -1,5 +1,5 @@
 import unittest
-from tinygrad.helpers import DEBUG
+from tinygrad.helpers import DEBUG, Context
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import UPat, track_rewrites, GroupOp, Ops
 from tinygrad.uop.upat import _get_code, upat_compile
@@ -14,6 +14,7 @@ def do_compile(up):
   if DEBUG >= 2: dis.dis(match)
   return match_code[0]
 
+@Context(SPEC=0)
 class TestUPatCompile(unittest.TestCase):
   def test_double(self):
     up = UPat.var("x") * UPat.cvar("c0") + UPat.var("x") * UPat.cvar("c1")

test/unit/test_viz.py

@@ -2,7 +2,7 @@ import unittest, decimal, json, struct
 from dataclasses import dataclass
 from typing import Generator
 
-from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, TrackedPatternMatcher, graph_rewrite, track_rewrites, TRACK_MATCH_STATS
+from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, TrackedPatternMatcher, graph_rewrite, track_rewrites, TRACK_MATCH_STATS, profile_matches
 from tinygrad.uop.symbolic import sym
 from tinygrad.dtype import dtypes
 from tinygrad.helpers import PROFILE, colored, ansistrip, flatten, TracingKey, ProfileRangeEvent, ProfileEvent, Context, cpu_events, profile_marker
@@ -117,6 +117,28 @@ class TestViz(BaseTestViz):
     # NOTE: names from TracingKey do not get deduped
     self.assertEqual(lst[0]["name"], "custom_name")
 
+  def test_profile_matches(self):
+    @profile_matches
+    def nested_function(u:UOp):
+      for i in range(2): graph_rewrite(u, PatternMatcher([]), name=f"step {i+1}")
+
+    @track_rewrites()
+    def main_rewrite(u:UOp):
+      graph_rewrite(u, PatternMatcher([]), name="init")
+      nested_function(u)
+
+    main_rewrite(UOp.variable("a", 1, 10)+UOp.variable("b", 1, 10))
+    steps = get_viz_list()[0]["steps"]
+    self.assertEqual(steps[0]["name"], "init")
+    self.assertEqual(steps[1]["name"], "nested_function")
+    self.assertEqual(len(steps), 4)
+
+  def test_profile_matches_invalid_arg(self):
+    @profile_matches
+    def invalid_fxn(arg:str): return graph_rewrite(UOp(Ops.SINK), PatternMatcher([]))
+    with self.assertRaisesRegex(AssertionError, "invalid match tracing input"):
+      invalid_fxn("test")
+
   def test_colored_label(self):
     # NOTE: dataclass repr prints literal escape codes instead of unicode chars
     @dataclass(frozen=True)
@@ -126,12 +148,20 @@ class TestViz(BaseTestViz):
     a2 = uop_to_json(a)[id(a)]
     self.assertEqual(ansistrip(a2["label"]), f"CUSTOM\n{TestStruct.__qualname__}(colored_field='xyz12345')")
 
+  def test_colored_label_multiline(self):
+    arg = colored("x", "green")+"\n"+colored("y", "red")+colored("z", "yellow")+colored("ww\nw", "magenta")
+    src = [Tensor.empty(1).uop for _ in range(10)]
+    a = UOp(Ops.CUSTOM, src=tuple(src), arg=arg)
+    exec_rewrite(a, [PatternMatcher([])])
+    a2 = next(get_viz_details(0, 0))["graph"][id(a)]
+    self.assertEqual(ansistrip(a2["label"]), "CUSTOM\nx\nyzww\nw")
+
   def test_inf_loop(self):
-    a = UOp.variable('a', 0, 10)
-    b = a.replace(op=Ops.CONST)
+    a = UOp.const(dtypes.int, 3)
+    b = UOp.const(dtypes.int, 4)
     pm = PatternMatcher([
-      (UPat(Ops.DEFINE_VAR, name="x"), lambda x: x.replace(op=Ops.CONST)),
-      (UPat(Ops.CONST, name="x"), lambda x: x.replace(op=Ops.DEFINE_VAR)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
     ])
     with self.assertRaises(RuntimeError): exec_rewrite(a, [pm])
     graphs = flatten(x["graph"].values() for x in get_viz_details(0, 0))
@@ -142,8 +172,8 @@ class TestViz(BaseTestViz):
     self.assertEqual(graphs[2], uop_to_json(nop)[id(nop)])
 
   def test_const_node_visibility(self):
-    a = UOp.variable("a", 0, 10)
-    z = UOp.const(dtypes.index, 0)
+    a = UOp.variable("a", 0, 10, dtype=dtypes.int)
+    z = UOp.const(a.dtype, 0)
     alu = a*z
     exec_rewrite(alu, [sym])
     lst = get_viz_list()
@@ -326,14 +356,14 @@ def load_profile(lst:list[ProfileEvent]) -> dict:
     event_type, event_count = u("<BI")
     if event_type == 0:
       for _ in range(event_count):
-        name, ref, st, dur, _ = u("<IIIfI")
-        v["events"].append({"name":strings[name], "ref":option(ref), "st":st, "dur":dur})
+        name, ref, key, st, dur, _ = u("<IIIIfI")
+        v["events"].append({"name":strings[name], "ref":option(ref), "key":option(key), "st":st, "dur":dur})
     else:
       v["peak"] = u("<Q")[0]
       for _ in range(event_count):
         alloc, ts, key = u("<BII")
         if alloc: v["events"].append({"event":"alloc", "ts":ts, "key":key, "arg": {"dtype":strings[u("<I")[0]], "sz":u("<Q")[0]}})
-        else: v["events"].append({"event":"free", "ts":ts, "key":key, "arg":{"users":[u("<I")[0] for _ in range(u("<I")[0])]}})
+        else: v["events"].append({"event":"free", "ts":ts, "key":key, "arg": {"users":[u("<IIBB") for _ in range(u("<I")[0])]}})
   return {"dur":total_dur, "peak":global_peak, "layout":layout, "markers":markers}
 
 class TestVizProfiler(unittest.TestCase):
@@ -498,6 +528,27 @@ class TestVizMemoryLayout(BaseTestViz):
     buffers = profile["layout"]["NULL Memory"]["events"]
     user_cnt = [len(b["arg"]["users"]) for b in buffers if b["arg"].get("users")]
     self.assertEqual(max(user_cnt), n)
+    input_buf = buffers.pop()
+    assert all(u[3] == 0 for u in input_buf["arg"]["users"])
+
+  def test_annotate_read_write(self):
+    a = Tensor.ones(4, device="NULL").contiguous().realize()
+    b = a.assign(a+2)
+    c = a+1
+    Tensor.realize(b, c)
+    buf_events = load_profile(cpu_events+Buffer.profile_events)["layout"]["NULL Memory"]["events"]
+    users = next((b["arg"]["users"] for b in buf_events if len(b["arg"].get("users",[])) == 3))
+    self.assertEqual(users[0][3], 1) # write Tensor.ones
+    self.assertEqual(users[1][3], 2) # read+write Tensor.assign
+    self.assertEqual(users[2][3], 0) # readonly
+
+  def test_dedup_users(self):
+    a = Tensor.empty(1, device="NULL")
+    for _ in range(n:=4): a.add(1).realize()
+    profile = load_profile(cpu_events+Buffer.profile_events)
+    programs = profile["layout"][a.device]["events"]
+    users = profile["layout"][f"{a.device} Memory"]["events"].pop()["arg"]["users"]
+    self.assertEqual(len(programs), len(set(users)), n)
 
 if __name__ == "__main__":
   unittest.main()

tinygrad/codegen/__init__.py

@@ -1,9 +1,7 @@
-from typing import Any, Callable
-import functools
-from dataclasses import dataclass
-from tinygrad.helpers import QUANTIZE, DEVECTORIZE, TRANSCENDENTAL
+import itertools
+from tinygrad.helpers import QUANTIZE, DEVECTORIZE, TRANSCENDENTAL, SPEC
 from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, pm_lower_index_dtype
-from tinygrad.uop.spec import type_verify
+from tinygrad.uop.spec import type_verify, program_spec, kernel_spec
 from tinygrad.renderer import Renderer
 
 # import all pattern matchers here
@@ -14,111 +12,103 @@ from tinygrad.uop.decompositions import get_late_rewrite_patterns
 from tinygrad.codegen.late.expander import migrate_indexing, expander, pm_pre_expander, pm_group_for_reduce
 from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize, pm_reduce, \
   ReduceContext, correct_load_store, pm_render
-from tinygrad.codegen.late.linearize import block_create, pm_blockend_merge, block_merge, pm_finalize, BlockContext
-from tinygrad.codegen.opt.postrange import pm_postrange_opt
-from tinygrad.codegen.simplify import pm_simplify_ranges, pm_reduce_simplify, pm_flatten_range, pm_split_ranges
-from tinygrad.schedule.rangeify import pm_add_buffers, rangeify_codegen
+from tinygrad.codegen.opt.postrange import apply_opts
+from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse, pm_split_store
+from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen
+from tinygrad.codegen.late.linearizer import CFGContext, pm_prepare_control_flow, pm_add_control_flow, linearize
 
-@dataclass
-class RewriteStep:
-  pm: PatternMatcher
-  ctx: Callable[[UOp], Any]|None = None
-  name: str|None = None
-  bottom_up: bool = False
-  def __call__(self, sink:UOp):
-    return graph_rewrite(sink, self.pm, ctx=self.ctx(sink) if self.ctx is not None else None, name=self.name, bottom_up=self.bottom_up)
+def full_rewrite_to_sink(sink:UOp, ren:Renderer|None=None, optimize:bool=True) -> UOp:
+  if ren is None: ren = Renderer()
 
-def apply_rewrites(sink:UOp, rewrites:list[RewriteStep]): return functools.reduce(lambda x,f: f(x), rewrites, sink)
-
-rewrites_for_linearizer = [
-  RewriteStep(block_create, ctx=BlockContext.from_sink, name="Linearizer: Create Blocks", bottom_up=True),
-  RewriteStep(pm_blockend_merge, name="Linearizer: Merge Blockends"),
-  RewriteStep(block_merge, name="Linearizer: Merge Blocks"),
-  RewriteStep(pm_finalize, name="Linearizer: Finalize")]
-
-def get_rewrites_for_renderer(opts:Renderer, optimize:bool=True, linearizer:bool=True) -> list[RewriteStep]:
-  # cache with the values of the context vars
-  return _get_rewrites_for_renderer(opts, optimize, linearizer, QUANTIZE.value, DEVECTORIZE.value, TRANSCENDENTAL.value)
-
-@functools.cache
-def _get_rewrites_for_renderer(opts:Renderer, optimize:bool, linearizer:bool, _QUANTIZE, _DEVECTORIZE, _TRANSCENDENTAL) -> list[RewriteStep]:
-  # ** lowerer **
-  ret: list[RewriteStep] = []
+  if SPEC: type_verify(sink, kernel_spec)
 
+  # first we optimize
   if optimize:
+    if QUANTIZE and ren.device in {"CPU", "DSP"}: sink = graph_rewrite(sink, pm_quant, name="quantize")
 
-    # lowerer first
-    if _QUANTIZE and opts.device in {"CPU", "DSP"}: ret.append(RewriteStep(pm_quant, name="quantize"))
+    # TODO: fix expander and remove this
+    sink = graph_rewrite(sink, pm_add_buffers_local, name="add locals early")
+
+    # collapse loads reduce (indexing by a tensor)
+    sink = graph_rewrite(sink, pm_load_collapse, name="load collapse")
 
     # split ranges
-    ret.append(RewriteStep(pm_split_ranges+pm_flatten_range, ctx=lambda _: {}, name="split ranges"))
+    sink = graph_rewrite(sink, pm_split_ranges+pm_flatten_range, ctx={}, name="split ranges")
 
     # symbolic (NOTE: this is a requirement for pm_simplify_ranges to be correct)
-    ret.append(RewriteStep(sym+pm_flatten_range, name="initial symbolic"))
+    sink = graph_rewrite(sink, sym+pm_flatten_range, name="initial symbolic")
 
     # optimize (schedule) the AST
-    ret.append(RewriteStep(pm_simplify_ranges, name="simplify ranges"))
-    ret.append(RewriteStep(pm_reduce_simplify, name="simplify reduces"))
-    ret.append(RewriteStep(pm_postrange_opt, ctx=lambda _: opts, name="post optimize ast"))
+    sink = graph_rewrite(sink, pm_simplify_ranges, name="simplify ranges")
+
+    # split store range (only on CPU for now)
+    sink = graph_rewrite(sink, pm_split_store, ctx=ren.device, name="cut store ranges")
+
+    # do postrange optimization, BEAM or hand_coded_optimizations
+    sink = apply_opts(sink, ren)
 
   # ** expander (expand_rewrite) **
-  ret.append(RewriteStep(sym+migrate_indexing+pm_move_where_on_load, name="postopt symbolic"))
+  sink = graph_rewrite(sink, sym+migrate_indexing+pm_move_where_on_load, name="postopt symbolic")
 
   # expand
-  ret.append(RewriteStep(sym+pm_pre_expander+pm_group_for_reduce+expander, name="expander"))
+  sink = graph_rewrite(sink, sym+pm_pre_expander+pm_group_for_reduce+expander, name="expander")
 
   # add locals
-  ret.append(RewriteStep(pm_add_buffers+rangeify_codegen, name="add local buffers"))
+  sink = graph_rewrite(sink, pm_add_buffers_local+rangeify_codegen, name="add local buffers")
 
   # ** devectorizer (full_graph_rewrite) **
   # remove reduce
-  ret.append(RewriteStep(pm_reduce+gep_pushing, lambda _: ReduceContext(), name="remove_reduce"))
+  sink = graph_rewrite(sink, pm_reduce+gep_pushing, ctx=ReduceContext(), name="remove_reduce")
 
   # add gpu dims (late). this works after devectorize, but it's faster here
-  ret.append(RewriteStep(pm_add_gpudims, lambda _: opts, name="add gpudims"))
+  sink = graph_rewrite(sink, pm_add_gpudims, ctx=ren, name="add gpudims")
 
   # devectorize (TODO: does this need opts?)
-  if _DEVECTORIZE >= 2: pm_devectorize = sym+load_store_folding+load_store_indexing
-  elif _DEVECTORIZE: pm_devectorize = sym+devectorize+load_store_folding+correct_load_store+load_store_indexing
+  if DEVECTORIZE >= 2: pm_devectorize = sym+load_store_folding+load_store_indexing
+  elif DEVECTORIZE: pm_devectorize = sym+devectorize+load_store_folding+correct_load_store+load_store_indexing
   else: pm_devectorize = sym+load_store_folding+correct_load_store+load_store_indexing
-  ret.append(RewriteStep(pm_devectorize, lambda _: opts, name="devectorize"))
-
-  supported_ops = tuple(opts.code_for_op.keys())
-  extra_matcher = opts.extra_matcher if opts.extra_matcher is not None else PatternMatcher([])
+  sink = graph_rewrite(sink, pm_devectorize, ctx=ren, name="devectorize")
 
   # lower the index dtype to a concrete int
-  ret.append(RewriteStep(pm_lower_index_dtype+load_store_indexing, lambda _: opts.device, name="lower all index dtypes"))
-  ret.append(RewriteStep(symbolic, name="post index symbolic"))
+  sink = graph_rewrite(sink, pm_lower_index_dtype+load_store_indexing, ctx=ren.device, name="lower all index dtypes")
+  sink = graph_rewrite(sink, symbolic, name="post index symbolic")
 
   # optional pre matcher
-  if opts.pre_matcher is not None: ret.append(RewriteStep(opts.pre_matcher, name="pre_matcher"))
+  if ren.pre_matcher is not None: sink = graph_rewrite(sink, ren.pre_matcher, name="pre_matcher")
 
   # decompositions
-  pm_decomp = symbolic_simple+get_late_rewrite_patterns(supported_ops, _TRANSCENDENTAL>=2)
-  ret.append(RewriteStep(pm_decomp, lambda _: opts.device, name="decompositions"))
+  supported_ops = tuple(ren.code_for_op.keys())
+  pm_decomp = symbolic_simple+get_late_rewrite_patterns(supported_ops, TRANSCENDENTAL>=2)
+  sink = graph_rewrite(sink, pm_decomp, ctx=ren.device, name="decompositions")
 
   # final rules for the renderer (without sym)
+  extra_matcher = ren.extra_matcher if ren.extra_matcher is not None else PatternMatcher([])
   pm_final_rewrite = pm_decomp+pm_render+extra_matcher
-  ret.append(RewriteStep(pm_final_rewrite, lambda _: opts.device, name="final rewrite"))
+  sink = graph_rewrite(sink, pm_final_rewrite, ctx=ren.device, name="final rewrite")
 
-  # return the list (with optional linearizer)
-  return ret + (rewrites_for_linearizer if linearizer else [])
+  # prepare for control flow
+  sink = graph_rewrite(sink, pm_prepare_control_flow, ctx=itertools.count(10000), name="split ends + add if ranges")
 
-def full_rewrite_to_sink(sink:UOp, opts:Renderer|None=None, optimize:bool=True, linearizer:bool=False) -> UOp:
-  return apply_rewrites(sink, get_rewrites_for_renderer(opts if opts is not None else Renderer(), optimize, linearizer))
+  # this was the linearizer
+  sink = graph_rewrite(sink, pm_add_control_flow, ctx=CFGContext(sink), name="add control flow", bottom_up=True)
 
-def full_rewrite(sink:UOp, opts:Renderer|None=None) -> list[UOp]:
+  # return the rewritten sink
+  return sink
+
+def full_rewrite(sink:UOp, ren:Renderer|None=None) -> list[UOp]:
   """
   Function to transform the Kernel UOp graph into a linearized program.
 
   Args:
     sink: The Ops.SINK rooting the Kernel graph.
-    opts: The Renderer (can change how things are processed, fix this).
+    ren: The Renderer (can change how things are processed, fix this).
 
   Returns:
     Linear program in UOps.
   """
 
-  lst = list(full_rewrite_to_sink(sink, opts, optimize=sink.tag is None, linearizer=True).arg.lst)
-  if __debug__: type_verify(lst)
+  full_sink = full_rewrite_to_sink(sink, ren, optimize=sink.tag is None)
+  assert len(full_sink.ranges) == 0, "all ranges must end by the sink"
+  lst = linearize(full_sink)
+  if SPEC: type_verify(lst, program_spec)
   return lst

tinygrad/codegen/gpudims.py

@@ -1,7 +1,7 @@
-import math
+import math, functools, operator
 from tinygrad.uop.ops import UOp, Ops, sint, PatternMatcher, UPat, KernelInfo, ssimplify, AxisType, sint_to_uop
 from tinygrad.helpers import all_int, dedup, get_contraction
-from tinygrad.dtype import dtypes
+from tinygrad.dtype import dtypes, AddrSpace, Invalid
 from tinygrad.renderer import Renderer
 
 def _group_dims(dims:tuple[sint, ...], max_sizes:tuple[int, ...]):
@@ -79,6 +79,14 @@ def add_gpudims(ctx:Renderer, s:UOp):
   # apply to multiple ranges
   subs = {}
   for r in s_topo:
+    # look for local INDEXes that are not used in the GLOBAL store, then add them as an INVALID
+    if r.op is Ops.STORE and r.src[0].ptrdtype.addrspace == AddrSpace.GLOBAL:
+      idx = r.src[0]
+      missing_locals = [all_ranges[rng] for rng in local_dims if all_ranges[rng] not in idx.ranges]
+      if len(missing_locals):
+        assert len(idx.src) == 2, "index has 2 sources"
+        mask: UOp = functools.reduce(operator.and_, [x.eq(0) for x in missing_locals])
+        subs[idx] = idx.replace(src=(idx.src[0], mask.broadcast(idx.src[1].dtype.count).where(idx.src[1], Invalid)))
     if r.op is not Ops.RANGE: continue
     try:
       ii = (global_dims+local_dims).index(r.arg[0:-1])

tinygrad/codegen/late/devectorizer.py

@@ -50,7 +50,6 @@ def delete_redundant_gates(store:UOp, buf:UOp, idx:UOp, val:UOp, store_gate:UOp,
   # remove the gate from the index
   return UOp.store(buf.index(idx).cast(cast.dtype) if cast is not None else buf.index(idx), val, *store.src[2:])
 
-def no_load(u:UOp) -> bool: return not any(x.op is Ops.LOAD for x in u.backward_slice_with_self)
 load_store_indexing = PatternMatcher([
   # image load valid idx simplification
   (UPat(Ops.INDEX, src=(UPat.var("buf"), invalid_gate)), lambda buf,x,i,cond: simplify_valid_load(buf, x, cond)),
@@ -61,8 +60,6 @@ load_store_indexing = PatternMatcher([
   # delete_redundant_gates (after expand)
   (UPat(Ops.STORE, src=(UPat.any(stidx:=UPat.var("buf").index(UPat.var("idx"), UPat.var("store_gate")), stidx.cast().named("cast")),
                                   UPat.var("val")), name="store", allow_any_len=True), delete_redundant_gates),
-  # we want to make sure we dont do math on a loaded index since that can cause overflow, this undoes a pattern in reduce_collapse
-  (UPat.var("c")<(UPat.var("x", dtypes.index)+UPat.var("y")), lambda x,y,c: (-x < -(c-y)) if no_load(y) and no_load(c) and not no_load(x) else None),
 ])
 
 # ***** load/store grouping *****
@@ -112,7 +109,7 @@ def cat_after_store(cat:UOp, data:UOp, sto:UOp):
   for s in cat.src:
     ret.append(s.store(data.gep(tuple(range(offset, offset+s.dtype.count))), *sto.src[2:]))
     offset += s.dtype.count
-  return UOp(Ops.NOOP, src=tuple(ret))
+  return UOp.group(*ret)
 
 def gep_on_store(gep:UOp, st:UOp, sto:UOp):
   # NOTE: we need to invert the gep here, but it may be an expanding gep
@@ -123,7 +120,7 @@ def gep_on_store(gep:UOp, st:UOp, sto:UOp):
   return gep.src[0].store(st.gep(new_arg), *sto.src[2:])
 
 load_store_folding = PatternMatcher([
-  (UPat(Ops.INDEX, src=(UPat(Ops.VECTORIZE, src=UPat(GroupOp.Defines, name="buf")), UPat.var("vec"))), expand_index),
+  (UPat(Ops.INDEX, src=(UPat(Ops.VECTORIZE, src=UPat(GroupOp.Defines).or_after(name="buf")), UPat.var("vec"))), expand_index),
   # GEP after LOAD
   (UPat(Ops.LOAD, src=(UPat(Ops.GEP, name="gep"),), name="ld", allow_any_len=True),
    lambda gep, ld: ld.replace(dtype=ld.dtype.scalar().vec(gep.dtype.count), src=(gep.src[0],)+ld.src[1:]).gep(gep.arg)),
@@ -182,7 +179,7 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
 
   # if it wasn't split, we return None. otherwise we CAT them
   if len(ret) <= 1: return None
-  return UOp(Ops.CAT, ls.dtype, tuple(ret)) if ls.op is Ops.LOAD else UOp(Ops.NOOP, src=tuple(ret))
+  return UOp(Ops.CAT, ls.dtype, tuple(ret)) if ls.op is Ops.LOAD else UOp.group(*ret)
 
 def image_fixup(ls:UOp):
   # normal image load or store, with the CAST from expand_index
@@ -242,11 +239,13 @@ def no_vectorized_index(buf:UOp, cast:UOp, idx:UOp):
   return buf.broadcast(cnt).index(idx.broadcast(cnt)*cnt+UOp.const(dtypes.index.vec(cnt), tuple(range(cnt))))
 
 devectorize = PatternMatcher([
+  # CAST after AFTER
+  (UPat(Ops.CAST, name="c").f(Ops.AFTER, allow_any_len=True, name="a"), lambda c,a: c.src[0].after(*a.src[1:]).cast(c.dtype)),
   # no ALU on vectorized dtypes
   (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST), name="alu"), no_vectorized_alu),
   (UPat(Ops.WMMA, name="wmma"), no_vectorized_wmma),
   (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf"), no_vectorized_buf),
-  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG), name="buf").cast(name="cast").index(UPat.var("idx")), no_vectorized_index),
+  (UPat((Ops.DEFINE_LOCAL, Ops.DEFINE_REG)).or_after(name="buf").cast(name="cast").index(UPat.var("idx")), no_vectorized_index),
 ])
 
 pm_render = PatternMatcher([
@@ -266,10 +265,6 @@ pm_render = PatternMatcher([
     UPat.var("a")), lambda c,idx,l,a: l.replace(src=(l.src[0], a.cast(l.dtype))+l.src[2:]).cast(a.dtype)),
   (UPat.var("c").where(UPat.var("a"), UPat(Ops.LOAD, src=(UPat().index(UPat.var("idx"), UPat.var("c").logical_not()).or_casted(),),
     allow_any_len=True, name="l").or_casted()), lambda c,idx,l,a: l.replace(src=(l.src[0], a.cast(l.dtype))+l.src[2:]).cast(a.dtype)),
-  # gate any stores that aren't gated with ifs
-  (UPat(Ops.STORE, src=(UPat(src=(UPat(), UPat(), UPat(dtype=dtypes.bool)), name="idx").or_casted(), UPat()), name="store", allow_any_len=True),
-    lambda store,idx: UOp(Ops.STORE, dtype=store.dtype, src=store.src[:2]+(UOp(Ops.IF, src=(idx.src[2],)),)+store.src[2:]) if \
-      len(store.src) <= 2 or store.src[2].op != Ops.IF else None),
 ])
 
 # *** Ops.REDUCE -> Ops.DEFINE_ACC ***
@@ -293,15 +288,17 @@ def reduce_to_acc(ctx:ReduceContext, red:UOp):
   # if we have a range
   if len(reduce_range) != 0:
     topo = inp.toposort()
-    stored_ranges = flatten([x.src[2:] for x in topo if x.op is Ops.STORE])
-    input_ranges = tuple([x for x in topo if x.op is Ops.RANGE and x not in reduce_range and x not in stored_ranges])
+    ended_ranges = flatten([x.ended_ranges for x in topo if x.op is Ops.END])
+    input_ranges = tuple([x for x in topo if x.op is Ops.RANGE and x not in reduce_range and x not in ended_ranges])
     identity = red.const(red.dtype, identity_element(red.arg, red.dtype.scalar()))
-    acc = UOp(Ops.DEFINE_REG, red.dtype.ptr(size=1, addrspace=AddrSpace.REG), arg=(ctx.acc_num,)).index(UOp.const(dtypes.int, 0))
-    do_store = acc.store(identity, UOp(Ops.NOOP, src=input_ranges)) if len(input_ranges) else acc.store(identity)
-    lst = [acc.load(do_store, *reduce_range)] + lst  # put acc as the first element
+    acc = UOp(Ops.DEFINE_REG, red.dtype.ptr(size=1, addrspace=AddrSpace.REG), arg=(ctx.acc_num,))
+    acc_init = acc.after(*input_ranges).index(UOp.const(dtypes.int, 0)).store(identity) if len(input_ranges) else \
+               acc.index(UOp.const(dtypes.int, 0)).store(identity)
+    lst = [acc.after(acc_init, *reduce_range).index(UOp.const(dtypes.int, 0)).load()] + lst  # put acc as the first element
     ctx.acc_num += 1
   ret = functools.reduce(lambda x,y: x.alu(red.arg, y), lst)
-  return acc.load(acc.store(ret, *reduce_range)) if len(reduce_range) != 0 else ret
+  if len(reduce_range) == 0: return ret
+  return acc.after(acc.index(UOp.const(dtypes.int, 0)).store(ret).end(*reduce_range)).index(UOp.const(dtypes.int, 0)).load()
 
 pm_reduce = PatternMatcher([
   # REDUCE -> DEFINE_ACC+ASSIGN

tinygrad/codegen/late/expander.py

@@ -87,7 +87,7 @@ expander = PatternMatcher([
    lambda outer, inner: UOp(Ops.UNROLL, outer.dtype, (inner.src[0],), inner.arg+outer.arg)),
   # do expansion
   (UPat((*GroupOp.ALU, Ops.CAST, Ops.BITCAST, Ops.GEP, Ops.WMMA, Ops.LOAD, Ops.STORE, Ops.INDEX, Ops.BUFFERIZE,
-         Ops.VECTORIZE, Ops.IF, Ops.REDUCE), name="root", custom_early_reject=set([Ops.UNROLL])), do_expand),
+         Ops.VECTORIZE, Ops.IF, Ops.REDUCE, Ops.END), name="root", custom_early_reject=set([Ops.UNROLL])), do_expand),
   (UPat(Ops.CONTRACT, name="con"), do_contract),
   # BARRIERs aren't actually expanded
   (UPat(Ops.BARRIER, src=(UPat(Ops.UNROLL, name="ex"),)),
@@ -145,8 +145,7 @@ def fix_group_for_reduce(x:UOp):
   reduce_loop = [x.replace(arg=(x.arg[0]+100, AxisType.REDUCE)) for x in reduce_gfr]
   buf = ret.bufferize(*upstream_locals, *reduce_gfr, arg=BufferizeOpts(reduce_gfr[0].arg[0], AddrSpace.LOCAL)).index(*upstream_locals, *reduce_loop)
 
-  # gate with an if on the store + do the final reduce
-  buf = UOp(Ops.IF, dtype=buf.dtype, src=(functools.reduce(operator.and_, [x.eq(0) for x in reduce_gfr]), buf))
+  # do the final reduce (if/barrier are added in gpudims step)
   return buf.reduce(*reduce_loop, arg=x.arg)
 
 pm_pre_expander = PatternMatcher([

tinygrad/codegen/late/linearize.py

@@ -1,243 +0,0 @@
-from __future__ import annotations
-import heapq
-from collections import defaultdict
-from dataclasses import dataclass, replace
-from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp, BottomUpGate
-from tinygrad.helpers import dedup, all_same, flatten, BLOCK_REORDER
-
-# NOTE: any toposort should be valid here, unlike last time this isn't required, it's just for speed
-def block_reorder(lst:list[UOp]) -> list[UOp]:
-  in_this_block = set(lst)
-  local_children: defaultdict[UOp, list[UOp]] = defaultdict(list)
-  in_degree:dict[UOp, int] = {}
-  priorities:dict[UOp, int] = {}
-
-  # get local children and assign priorities
-  # NOTE: this requires the lst be locally toposorted
-  for u in reversed(lst):
-    in_degree[u] = 0
-    for s in u.src:
-      if s in in_this_block:
-        local_children[s].append(u)
-        in_degree[u] += 1
-    # put loads in the beginning of the block and prevent priority inversion. hack for BARRIER grouping too
-    priority = [0] + [priorities[x] for x in local_children[u]]
-    if u.op is Ops.LOAD: priority.append(-1000)
-    if u.op is Ops.BARRIER: priority.append(-1500)
-    priorities[u] = min(priority)
-
-  # number the uops in "ideal" order
-  nkey = {u:i for i,u in enumerate(sorted(lst, key=lambda x: (priorities[x],)+x.tuplize))}
-
-  # then force then to be toposorted in as close to the ideal order as possible
-  heapq.heapify(heap:=[(nkey[u],u) for u in lst if in_degree[u] == 0])
-  newlst = []
-  while heap:
-    newlst.append(u:=heapq.heappop(heap)[1])
-    for v in local_children[u]:
-      in_degree[v] -= 1
-      if in_degree[v] == 0: heapq.heappush(heap, (nkey[v],v))
-
-  assert len(newlst) == len(lst), f"len mismatch {len(newlst)} != {len(lst)}"
-  return newlst
-
-# ***** basic block *****
-
-def disp(y:UOp) -> str:
-  if y.op is Ops.IF: return f'IF{id(y)}'
-  if y.op is Ops.RANGE: return str(y.arg)
-  return "<NONE>"
-
-@dataclass(frozen=True, eq=False)
-class BasicBlock:
-  lst: tuple[UOp, ...]
-  ctx: tuple[UOp, ...] = ()
-  end: UOp|None = None
-  cnt: int = 0
-  child_ctx: tuple[UOp, ...]|None = None
-  def __lt__(self, _:BasicBlock): raise RuntimeError("no comparing basic blocks")
-  def __repr__(self):
-    return f"{(str(disp(self.end))+' ') if self.end is not None else ''}"+f'f{self.cnt} '+\
-           f"{[disp(y) for y in self.ctx]} {[disp(y) for y in self.child_ctx] if self.child_ctx is not None else '-'} "+\
-           f"{len(self.lst)}" + "\n" + '\n'.join([str(x.op) for x in self.lst])
-  def last_ctx(self): return self.child_ctx if self.child_ctx is not None else self.ctx
-
-def _sort_ctx(inp): return tuple(sorted(dedup(inp), key=lambda x: x.tuplize))
-
-# ***** block context *****
-
-@dataclass
-class BlockContext:
-  child_count: dict[UOp, int]
-  block_ctxs: dict[UOp, tuple[UOp, ...]]
-  child_ctxs: dict[UOp, tuple[UOp, ...]]
-  def last_ctx(self, u): return self.child_ctxs.get(u, self.block_ctxs[u])
-  @staticmethod
-  def from_sink(sink:UOp) -> BlockContext:
-    # get children and all block contexts
-    ctx = BlockContext({}, {}, {})
-    for u in sink.toposort(gate=lambda u:u.op is not Ops.SPECIAL):
-      this_block_ctx: list[UOp] = []
-      ctx.child_count[u] = 0
-
-      # get children and accumulate the last_ctx
-      for s in u.src:
-        if s.op is Ops.SPECIAL: continue
-        # NOTE: if a parent appears multiple times in the src, it counts multiple times as a child
-        ctx.child_count[s] += 1
-        this_block_ctx += ctx.last_ctx(s)
-
-      # save the block ctx. SINK never has anything
-      ctx.block_ctxs[u] = _sort_ctx(this_block_ctx) if u.op is not Ops.SINK else ()
-
-      # RANGE/IF add to the next ctx
-      # STORE/ASSIGN subtract from the next ctx
-      if u.op in {Ops.RANGE, Ops.IF}: ctx.child_ctxs[u] = _sort_ctx(ctx.block_ctxs[u] + (u,))
-      elif u.op is Ops.STORE: ctx.child_ctxs[u] = tuple([y for y in ctx.block_ctxs[u] if y not in u.src])
-    return ctx
-
-# ***** make blocks *****
-
-DONT_PLACE_IN_BLOCK = {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG, Ops.DEFINE_VAR, Ops.SPECIAL, Ops.CONST}
-
-def add_blockends(base_block:UOp, new_ctx:tuple[UOp, ...], current_ctx:tuple[UOp, ...], cnt:int=1) -> UOp:
-  ends_to_add = [z for z in new_ctx if z not in current_ctx]
-  while len(ends_to_add):
-    r:UOp = ends_to_add.pop(-1)
-    new_ctx = tuple([z for z in new_ctx if z is not r])
-    end_uop = UOp(Ops.ENDIF if r.op is Ops.IF else Ops.ENDRANGE, src=(r,))
-    base_block = UOp(Ops.BLOCKEND, src=(base_block,)*cnt, arg=BasicBlock((end_uop,), tuple(new_ctx), end=r, cnt=cnt))
-  return base_block
-
-def make_block_bottom_up(ctx:BlockContext, x:UOp):
-  if x.op is Ops.BLOCKSTART:
-    current_ctx, child_ctx = x.arg
-    lst = list(x.src)
-    child_count = 1
-  else:
-    current_ctx, child_count, child_ctx = ctx.block_ctxs[x], ctx.child_count[x], ctx.child_ctxs.get(x, None)
-    lst = [x]
-
-  # count of times we've seen this block, or a seed for a new block if we can't merge it
-  unmergable: defaultdict[UOp, int] = defaultdict(int)
-  blockseeds = defaultdict(list)
-
-  # add the srcs of this to the frontier
-  # NOTE: things may be in here multiple times, that's okay
-  frontier_nodes = list(flatten(y.src[::-1] for y in lst))
-  while len(frontier_nodes):
-    u = frontier_nodes.pop(0)
-    if u.op not in DONT_PLACE_IN_BLOCK and ctx.child_count[u] == unmergable[u]+1:
-      # count is correct
-      if (newctx:=ctx.block_ctxs[u]) == current_ctx:
-        # block has same context, merge it, and put the srcs on the frontier
-        lst.append(u)
-        frontier_nodes.extend(u.src[::-1])
-      else:
-        # block has different context, add it to blockseeds
-        blockseeds[(newctx, ctx.child_ctxs.get(u, None))].append(u)
-      del unmergable[u]
-    else:
-      # count is incorrect (or it's DONT_PLACE_IN_BLOCK), add it to unmergable
-      unmergable[u] += 1
-
-  # add unmergables to sources
-  srcs = []
-  for u,cnt in unmergable.items(): srcs += [add_blockends(u, ctx.block_ctxs.get(u,()), current_ctx, cnt=cnt)]*cnt
-
-  # add blockseeds, with blockends as needed
-  for (new_ctx, new_child_ctx), v in blockseeds.items():
-    base_block = UOp(Ops.BLOCKSTART, src=tuple(v), arg=(new_ctx, new_child_ctx))
-    srcs.append(add_blockends(base_block, new_ctx, current_ctx))
-
-  lst = lst[::-1]
-  if BLOCK_REORDER: lst = block_reorder(lst)
-  bb = BasicBlock(tuple(lst), ctx=current_ctx, cnt=child_count, child_ctx=child_ctx)
-  return UOp(Ops.BLOCK, src=tuple(srcs), arg=bb)
-
-# we prevent the source of the SPECIAL from being linearized since its not part of the kernel
-def raise_bottom_up_gate(): raise BottomUpGate()
-
-block_create = PatternMatcher([
-  (UPat(GroupOp.All-DONT_PLACE_IN_BLOCK.union({Ops.BLOCK, Ops.BLOCKEND}), name="x"), make_block_bottom_up),
-  (UPat(Ops.SPECIAL), raise_bottom_up_gate)
-])
-
-# ***** blockend merging ****
-
-def merge_blockends(sink:UOp) -> UOp|None:
-  # only run on the final BLOCK with the SINK in it
-  if sink.arg.lst[-1].op is not Ops.SINK: return None
-  # combine matching BLOCKENDS, the keys of this dictionary are the RANGE UOps, values are the BLOCKENDs
-  blockends_to_arg: dict[UOp, list[UOp]] = {}
-  for be in sink.toposort():
-    if be.op is Ops.BLOCKEND: blockends_to_arg.setdefault(be.arg.end, []).append(be)
-  new_forks = {}
-  for k,v in blockends_to_arg.items():
-    # NOTE: if any BLOCKEND is the parent of any other with the same arg, this algo fails
-    if len(v) > 1:
-      bb = BasicBlock(v[0].arg.lst, _sort_ctx(flatten([y.arg.ctx for y in v])), k, cnt=sum(y.arg.cnt for y in v))
-      out = UOp(Ops.BLOCKEND, src=tuple(flatten([x.src for x in v])), arg=bb)
-      # NOTE: bb.ctx != u.arg.ctx can cause problems here
-      for u in v: new_forks[u] = out
-  if len(new_forks) == 0: return None
-  return sink.substitute(new_forks)
-
-pm_blockend_merge = PatternMatcher([(UPat(Ops.BLOCK, name="sink"), merge_blockends)])
-
-# ***** block merging ****
-
-def merge_block(x:UOp):
-  unmergable_blocks, mergable_blocks = [], []
-  mergable_dict: defaultdict[UOp, int] = defaultdict(int)
-  for y in x.src:
-    if y.op is Ops.BLOCK and x.op is Ops.BLOCK and x.arg.ctx == y.arg.ctx: mergable_dict[y] += 1
-    elif y.op is Ops.BLOCK and x.op is Ops.BLOCKEND and x.arg.end in y.arg.ctx: mergable_dict[y] += 1
-    else: unmergable_blocks.append(y)
-  for k,v in mergable_dict.items():
-    if v == k.arg.cnt: mergable_blocks.append(k)
-    else: unmergable_blocks.extend([k]*v)
-  if len(mergable_blocks) == 0: return None
-  del mergable_dict
-
-  # create the block
-  arg = replace(x.arg, lst=tuple(flatten([y.arg.lst for y in mergable_blocks]))+x.arg.lst)
-  return UOp(x.op, src=tuple(flatten([y.src for y in mergable_blocks])+unmergable_blocks), arg=arg)
-
-def remove_blockend(x:UOp):
-  # if there's any remaining blocks that need to go in this BLOCKEND, we don't remove it
-  if any(x.arg.end in y.arg.ctx for y in x.src if y.op in {Ops.BLOCK, Ops.BLOCKEND}): return None
-
-  if (parent_blocks := [y for y in x.src if y.op is Ops.BLOCK and y.arg.child_ctx is not None and x.arg.end in y.arg.child_ctx]):
-    assert all_same(parent_blocks), f"should never have two parent blocks (has {len(parent_blocks)})"
-    parent_block = parent_blocks[0]
-    assert len(parent_blocks) == parent_block.arg.cnt
-    # NOTE: DEFINE_ACC doesn't have to be handled in any special way
-    late_ops = list(x.arg.lst)
-    # NOTE: we have to add a barrier at the start if barrier is used in the range
-    if x.op is Ops.BLOCKEND and any(y.op is Ops.BARRIER for y in late_ops) and late_ops[-1].op is Ops.ENDRANGE:
-      late_ops = [UOp(Ops.BARRIER)] + late_ops
-    # peephole opt, remove any BARRIERs next to each other
-    for i in range(len(late_ops)-1):
-      if late_ops[i].op is Ops.BARRIER and late_ops[i+1].op is Ops.BARRIER: late_ops[i+1] = UOp(Ops.NOOP)
-    arg = BasicBlock(parent_block.arg.lst+tuple(late_ops), tuple([y for y in x.arg.ctx if y is not x.arg.end]), cnt=x.arg.cnt)
-    return UOp(Ops.BLOCK, src=tuple(y for y in x.src if y is not parent_block)+parent_block.src, arg=arg)
-  # else the whole context ended by the blockend is already in this block and we can safely turn it into a block
-  return UOp(Ops.BLOCK, src=x.src, arg=BasicBlock(x.arg.lst, tuple([y for y in x.arg.ctx if y is not x.arg.end]), cnt=x.arg.cnt))
-
-block_merge = PatternMatcher([
-  (UPat((Ops.BLOCK, Ops.BLOCKEND), name="x"), merge_block),
-  (UPat(Ops.BLOCKEND, name="x"), remove_blockend),
-])
-
-# ****** finalize ******
-
-def finalize(sink:UOp) -> UOp:
-  if sink.op is not Ops.BLOCK or not all(x.op in DONT_PLACE_IN_BLOCK for x in sink.src):
-    raise RuntimeError(f"linearize failure {sink.op} {[x.op for x in sink.src if x.op not in DONT_PLACE_IN_BLOCK]}")
-
-  # place the early things
-  lst = sorted(dedup(sink.src), key=lambda x: x.tuplize) + list(sink.arg.lst)
-  return UOp(Ops.BLOCKFINAL, arg=BasicBlock(tuple(lst)))
-
-pm_finalize = PatternMatcher([(UPat(Ops.BLOCK, name="sink"), finalize)])

tinygrad/codegen/late/linearizer.py

@@ -0,0 +1,86 @@
+import heapq
+from collections import defaultdict
+from tinygrad.dtype import dtypes
+from tinygrad.uop.ops import PatternMatcher, UOp, Ops, UPat, AxisType, GroupOp
+
+def linearize(u:UOp) -> list[UOp]:
+  # this is a toposort with priority
+  lst = list(u.toposort())
+  consumers: defaultdict[UOp, list[UOp]] = defaultdict(list)
+  in_degree:dict[UOp, int] = {}
+  priorities:dict[UOp, int] = {}
+
+  # get consumers and assign priorities
+  # NOTE: this requires the lst be locally toposorted
+  for u in reversed(lst):
+    for s in u.src: consumers[s].append(u)
+    in_degree[u] = len(u.src)
+    # put loads in the beginning of the block and prevent priority inversion. hack for BARRIER grouping too
+    priority = [0] + [priorities[x] for x in consumers[u]]
+    if u.op is Ops.LOAD: priority.append(-1000)
+    if u.op is Ops.BARRIER: priority.append(-1500)
+    # ranges are scheduled as late as possible so anything that can be outside is
+    # if u.op is Ops.RANGE: priority = [2000]
+    if u.op is Ops.END: priority = [-1000]
+    # move defines and consts to the top
+    if u.op in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG, Ops.DEFINE_VAR, Ops.SPECIAL, Ops.CONST}: priority.append(-2000)
+    priorities[u] = min(priority)
+
+  # number the uops in "ideal" order
+  nkey = {u:i for i,u in enumerate(sorted(lst, key=lambda x: (priorities[x],)+x.tuplize))}
+
+  # then force then to be toposorted in as close to the ideal order as possible
+  heapq.heapify(heap:=[(nkey[u],u) for u in lst if in_degree[u] == 0])
+  newlst = []
+  while heap:
+    newlst.append(u:=heapq.heappop(heap)[1])
+    for v in consumers[u]:
+      in_degree[v] -= 1
+      if in_degree[v] == 0: heapq.heappush(heap, (nkey[v],v))
+  assert len(newlst) == len(lst), f"len mismatch {len(newlst)} != {len(lst)}"
+  return newlst
+
+class CFGContext:
+  def __init__(self, sink:UOp):
+    # there are 3 relationships between ranges:
+    # nested, meaning endrange y is a dependency of endrange x and range x is a dependency of endrange y
+    # dependent, meaning endrange y is a dependency of endrange x and range x is not a dependency of endrange y
+    # independent, endrange y is not a dependency of endrange x
+    # everything is nested inside the sink
+    deps: dict[UOp, dict[UOp, None]] = {}
+    nesting: dict[UOp, UOp] = {}
+    for u in sink.toposort():
+      # get the deps from the src
+      deps[u] = {}
+      for s in u.src: deps[u] |= deps[s]
+
+      if u.op in (Ops.END, Ops.SINK):
+        nesting |= {x:u for x in deps[u] if x.op is Ops.END and (u.op is Ops.SINK or u.src[1] in deps[x]) and x not in nesting}
+      if u.op in (Ops.RANGE, Ops.END): deps[u][u] = None
+
+    self.edges: dict[UOp, UOp] = {}
+    siblings: dict[UOp, list[UOp]] = {}
+    for k,vv in nesting.items(): siblings.setdefault(vv, []).append(k)
+    for k,v in siblings.items():
+      # ranges that have dependencies on other siblings need to be scheduled after them
+      order = sorted(v, key=lambda x: len([u for u in v if u in deps[x]]))
+      zipped = zip(order, order[1:]) if k.op is Ops.SINK else zip([k.src[1]] + order, order)
+      for x,y in zipped: self.edges[y.src[1]] = x
+
+pm_add_control_flow = PatternMatcher([
+  (UPat(Ops.RANGE, name="x"), lambda ctx,x: x.replace(src=x.src+(y,)) if (y:=ctx.edges.get(x)) is not None else None),
+])
+
+def do_split_ends(e:UOp):
+  ret = e.src[0]
+  for r in list(UOp.sink(*e.src[1:]).ranges)[::-1]: ret = ret.end(r)
+  return ret
+
+pm_prepare_control_flow = PatternMatcher([
+  # split the ends
+  (UPat(Ops.END, name="e"), do_split_ends),
+  # add if ranges
+  (UPat(GroupOp.Defines, name="buf").index(UPat.var("idx"), UPat(name="gate", dtype=dtypes.bool)).or_casted("cast").store(UPat.var("val")),
+    lambda ctx,buf,idx,gate,cast,val:
+      buf.after(r:=UOp.range(gate.cast(dtypes.int), next(ctx), AxisType.IF, dtype=dtypes.int)).index(idx, gate).cast(cast.dtype).store(val).end(r)),
+])

tinygrad/codegen/opt/__init__.py

@@ -2,7 +2,6 @@
 from __future__ import annotations
 from enum import Enum, auto
 from dataclasses import dataclass
-from tinygrad.uop.ops import AxisType
 
 class OptOps(Enum):
   TC = auto(); UPCAST = auto(); UNROLL = auto(); LOCAL = auto(); THREAD = auto() # noqa: E702
@@ -16,11 +15,6 @@ class Opt:
   arg: int|tuple|None = None
   def __repr__(self): return f"Opt(op={self.op}, axis={self.axis}, arg={self.arg})"
 
-axis_letters = {AxisType.GLOBAL: "g", AxisType.THREAD: "t", AxisType.LOCAL: "l", AxisType.WARP: "w", AxisType.LOOP: "L", AxisType.UPCAST: "u",
-                AxisType.GROUP_REDUCE: "G", AxisType.REDUCE: "R", AxisType.UNROLL: "r"}
-axis_colors = {AxisType.GLOBAL: "blue", AxisType.THREAD: "BLUE", AxisType.LOCAL: "cyan", AxisType.WARP: "CYAN", AxisType.LOOP: "WHITE",
-               AxisType.UPCAST: "yellow", AxisType.GROUP_REDUCE: "RED", AxisType.REDUCE: "red", AxisType.UNROLL: "magenta"}
-
 class KernelOptError(Exception): pass
 def check(cond:bool, msg:str=""):
   if not cond: raise KernelOptError(msg)

tinygrad/codegen/opt/heuristic.py

@@ -27,15 +27,15 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
   # NOTE: unless TC_OPT is > 0, we only trigger tensor cores if there's only one reduce axis
   if USE_TC > 0 and (len(k.axes_of(AxisType.GROUP_REDUCE, AxisType.REDUCE)) == 1 or (TC_OPT.value >= 1)):
     good_tc_opt = False
+    tk = k.copy()
     try: # check TC first and apply hand-coded opts if successful
-      tk = k.copy()
       rngs = tk.apply_opt(Opt(OptOps.TC, 0, (TC_SELECT.value, TC_OPT.value, USE_TC.value)))
       good_tc_opt = True
     except KernelOptError:
       pass
-    if good_tc_opt:
-      # skip hand-coded TC opts if AMX, upcasting will make kernel slower
-      if rngs is not None and not AMX:
+    # skip hand-coded TC opts if AMX, upcasting will make kernel slower
+    if good_tc_opt and not AMX:
+      if rngs is not None:
         for tc_dim in [1,0]: # attempt to upcast M and N
           szs = [sz for sz in [5,4,3,2] if rngs[tc_dim].src[0].divides(sz) is not None]
           if szs:
@@ -62,8 +62,8 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
 
   # should use matvec - TODO: adjust/tune based on the wide vs tall/large vs small mat
   MV_BLOCKSIZE, MV_THREADS_PER_ROW, MV_ROWS_PER_THREAD = getenv("MV_BLOCKSIZE", 4), getenv("MV_THREADS_PER_ROW", 8), getenv("MV_ROWS_PER_THREAD", 4)
-  if k.opts.has_local and getenv("MV",1) != 0 and (MV_BLOCKSIZE > 1 or MV_THREADS_PER_ROW > 1 or MV_ROWS_PER_THREAD > 1) and  \
-    k.reduceop is not None and k.reduceop.arg[0] is Ops.ADD and len(k.full_shape) >= 2 and k.opts.has_shared and \
+  if k.ren.has_local and getenv("MV",1) != 0 and (MV_BLOCKSIZE > 1 or MV_THREADS_PER_ROW > 1 or MV_ROWS_PER_THREAD > 1) and  \
+    k.reduceop is not None and k.reduceop.arg[0] is Ops.ADD and len(k.full_shape) >= 2 and k.ren.has_shared and \
     (mulop:=k.reduceop.src[0]).op is Ops.MUL and mulop.src[0].op is Ops.LOAD and mulop.src[1].op is Ops.LOAD:
     idx0, idx1 = mulop.src[0].src[0].src[1].get_idx(), mulop.src[1].src[0].src[1].get_idx()
     if k.ranges_of(AxisType.REDUCE):
@@ -103,7 +103,7 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
   for axis in to_upcast[::-1]: k.apply_opt(Opt(OptOps.UPCAST, axis, 0))
 
   # potentially do more upcasts of non reduce axes based on a heuristic
-  is_dsp = k.opts is not None and k.opts.device == "DSP"
+  is_dsp = k.ren is not None and k.ren.device == "DSP"
   upcasted_axis: set[int] = set()
   while resolve(prod(k.output_shape[i] for i in k.upcastable_dims) >= 1024):
     xb_choices = []
@@ -149,13 +149,12 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
 
   # if nothing at all is upcasted and it's easy to, do an upcast
   for splits in [4]:
-    # TODO: somehow this never hits a reduce
     if not k.upcasted and k.upcastable_dims and k.full_shape[k.upcastable_dims[-1]] % splits == 0:
       k.apply_opt(Opt(OptOps.UPCAST, k.upcastable_dims[-1], splits))
 
   # **** local groups ****
 
-  if k.opts.has_local:
+  if k.ren.has_local:
     if NOLOCALS:
       k.apply_opt(Opt(OptOps.NOLOCALS))
     else:
@@ -176,13 +175,14 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
 
   # **** threading ****
 
-  if k.opts.has_threads and k.opts.global_max is not None:
+  if k.ren.has_threads and k.ren.global_max is not None:
     for threads in [32,16,12,8,6,5,4,3,2]:
       # Skip if too many threads. Heuristic: use about 128K ops per thread
-      if threads > k.opts.global_max[0] or resolve(prod(k.full_shape) // (128 << 10) < threads): continue
+      if threads > k.ren.global_max[0] or resolve(prod(k.full_shape) // (128 << 10) < threads): continue
       for axis in k.axes_of(AxisType.LOOP):
         if k.full_shape[axis] % threads == 0:
-          k.apply_opt(Opt(OptOps.THREAD, axis, threads))
+          try: k.apply_opt(Opt(OptOps.THREAD, axis, threads))
+          except KernelOptError: pass
           break
       if k.applied_opts and k.applied_opts[-1].op is OptOps.THREAD: break
 

tinygrad/codegen/opt/postrange.py

@@ -2,11 +2,11 @@ from __future__ import annotations
 import math, itertools
 from collections import defaultdict
 from typing import cast, Final
-from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo, graph_rewrite, AxisType, ssimplify, GroupOp
+from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, KernelInfo, graph_rewrite, AxisType, ssimplify, GroupOp, axis_letters, axis_colors
 from tinygrad.device import Buffer
-from tinygrad.dtype import AddrSpace, dtypes, ImageDType
-from tinygrad.helpers import colored, BEAM, getenv, DEBUG, to_function_name, NOOPT, argsort, round_up, prod, merge_dicts, get_single_element
-from tinygrad.codegen.opt import axis_colors, Opt, OptOps, KernelOptError, check, axis_letters
+from tinygrad.dtype import dtypes, ImageDType
+from tinygrad.helpers import colored, BEAM, getenv, DEBUG, to_function_name, NOOPT, argsort, round_up, prod, merge_dicts, get_single_element, flatten
+from tinygrad.codegen.opt import Opt, OptOps, KernelOptError, check
 from tinygrad.codegen.simplify import pm_flatten_range
 from tinygrad.renderer import Renderer
 
@@ -17,8 +17,8 @@ axis_to_pos = {AxisType.LOOP: -1, AxisType.THREAD: 0, AxisType.GLOBAL: 0, AxisTy
                AxisType.GROUP_REDUCE: 2, AxisType.REDUCE: 4, AxisType.UNROLL: 5}
 
 class Scheduler:
-  def __init__(self, ast:UOp, opts:Renderer):
-    self.ast, self.opts = ast, opts
+  def __init__(self, ast:UOp, ren:Renderer):
+    self.ast, self.ren = ast, ren
     self.dont_use_locals = self.ast.arg.dont_use_locals if self.ast.arg is not None else False
     self.applied_opts = list(self.ast.arg.applied_opts) if self.ast.arg is not None else []
 
@@ -46,7 +46,7 @@ class Scheduler:
   def shape_str_to_axis(self, nms:list[str]) -> tuple[int, ...]: return tuple([self.shape_str().index(x) for x in nms])
 
   def copy(self):
-    ret = Scheduler(self.ast, self.opts)
+    ret = Scheduler(self.ast, self.ren)
     ret.dont_use_locals = self.dont_use_locals
     ret.applied_opts = self.applied_opts[:]
     return ret
@@ -64,31 +64,24 @@ class Scheduler:
     return self.ast.replace(arg=KernelInfo(name=name, applied_opts=tuple(self.applied_opts), dont_use_locals=self.dont_use_locals), tag=1)
 
   def _globalizable_rngs(self) -> list[UOp]:
-    store_rngs = self.ast.src[0].src[2:]
-
-    # filter any not in local stores
-    local_store_rngs = [x.ranges for x in self.ast.toposort() if (x.op is Ops.STORE and x.src[0].ptrdtype.addrspace == AddrSpace.LOCAL) \
-                        or (x.op is Ops.BUFFERIZE and x.arg == AddrSpace.LOCAL)]
-    for ls in local_store_rngs: store_rngs = tuple([x for x in store_rngs if x in ls])
-
-    # filter any not in reduces
-    # TODO: enable this
-    """
-    reduce_rngs = [x.ranges for x in self.ast.toposort() if x.op is Ops.REDUCE]
-    for ls in reduce_rngs: store_rngs = tuple([x for x in store_rngs if x in ls])
-    """
-
-    return [x for x in UOp.sink(*store_rngs).toposort() if x.op is Ops.RANGE and x.arg[-1] == AxisType.LOOP] if store_rngs else []
+    return flatten([list(UOp.sink(*s.src[1:]).ranges) for s in self.ast.src if s.op is Ops.END])
 
   def convert_loop_to_global(self):
-    if not self.opts.has_local: return None
+    if not self.ren.has_local: return None
 
     globalizible_rngs = self._globalizable_rngs()
     rng = [x.replace(arg=x.arg[0:-1]+(AxisType.GLOBAL,)) if x in globalizible_rngs else x for x in self.rngs]
 
     self.ast = self.ast.substitute(dict(zip(self.rngs, rng)))
 
-  def colors(self) -> list[str]: return [axis_colors[x] if not self.dont_use_locals or not x == AxisType.GLOBAL else "BLUE" for x in self.axis_types]
+  def colors(self) -> list[str]:
+    output_rngs = self._globalizable_rngs()
+    ret = []
+    for x,r in zip(self.axis_types, self.rngs):
+      if self.dont_use_locals and x == AxisType.GLOBAL: ret.append("BLUE")
+      elif r not in output_rngs and x == AxisType.LOOP: ret.append("BLACK")
+      else: ret.append(axis_colors[x])
+    return ret
   def colored_shape(self) -> str: return ' '.join([colored(f'{x.src[0].render():>4s}', color) for x,color in zip(self.rngs, self.colors())])
 
   def shift_to(self, rng:UOp, amount:int, new_type:AxisType, top:bool=False, input_new_rng=None):
@@ -128,7 +121,7 @@ class Scheduler:
       return
 
     if opt.op in {OptOps.LOCAL, OptOps.GROUP, OptOps.GROUPTOP}:
-      check(self.opts.has_local, "locals needed for opt")
+      check(self.ren.has_local, "locals needed for opt")
 
     rng = self.rngs[real_axis] if (real_axis:=self.real_axis(opt.op, opt.axis)) >= 0 else UOp(Ops.NOOP)
 
@@ -146,7 +139,7 @@ class Scheduler:
                                         (self.group_for_reduces and opt.op not in {OptOps.NOLOCALS, OptOps.PADTO})):
         upcast_local_sz = prod([self.full_shape[a] for a in self.axes_of(AxisType.UPCAST, AxisType.WARP, AxisType.LOCAL, AxisType.GROUP_REDUCE)])
         smem_sz = amt*upcast_local_sz*self.reduceop.dtype.itemsize
-        check(smem_sz <= self.opts.shared_max, f"exceeds maximum shared memory size: needs {smem_sz}, max {self.opts.shared_max}")
+        check(smem_sz <= self.ren.shared_max, f"exceeds maximum shared memory size: needs {smem_sz}, max {self.ren.shared_max}")
       if self.reduceop is not None and (opt.op in {OptOps.GROUP, OptOps.GROUPTOP}):
         # We currently dont support a group within another rudece, TODO: fix if-contexts
         reduce = [u for u in self.ast.backward_slice if u.op is Ops.REDUCE and rng in merge_dicts([r.ranges for r in u.src[1:]])][0]
@@ -157,14 +150,14 @@ class Scheduler:
         check(amt <= 32, "don't unroll more than 32")
         check(rng.arg[-1] in {AxisType.GROUP_REDUCE, AxisType.REDUCE}, "unroll is for GROUP_REDUCE/REDUCE")
       if opt.op is OptOps.UPCAST:
-        check((self.opts is not None and self.opts.device == "DSP") or amt <= 16, "don't upcast more than 16")
+        check((self.ren is not None and self.ren.device == "DSP") or amt <= 16, "don't upcast more than 16")
         check(rng.arg[-1] in {AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP}, f"upcast is for GLOBAL/LOCAL/LOOP, not {rng.arg[-1]}")
       if opt.op is OptOps.LOCAL:
         check(not self.dont_use_locals, "can't use locals")
         check(rng.arg[-1] in {AxisType.GLOBAL, AxisType.LOOP}, "local is for globals")
       if opt.op is OptOps.THREAD:
-        check(self.opts is not None and self.opts.has_threads, "target does not support threads")
-        check(self.opts is not None and self.opts.global_max is not None and amt <= self.opts.global_max[0], "too many threads")
+        check(self.ren is not None and self.ren.has_threads, "target does not support threads")
+        check(self.ren is not None and self.ren.global_max is not None and amt <= self.ren.global_max[0], "too many threads")
         check(all(x is not AxisType.THREAD for x in self.axis_types), "already threaded")
         check(rng in self._globalizable_rngs(), "can't apply range to this dim")
       if opt.op in {OptOps.GROUP, OptOps.GROUPTOP}:
@@ -176,7 +169,7 @@ class Scheduler:
       check(len(self.applied_opts) == 0, "tensor core opts must be first") # TODO: remove the need for this by having warps
       check(opt.axis is not None, "tensor core opts must have an axis")
       check(opt.arg is not None and isinstance(opt.arg, tuple) and len(opt.arg) == 3, "tensor core opts must have valid arg")
-      check(-1 <= (tc_select:=cast(tuple, opt.arg)[0]) < len(self.opts.tensor_cores), "tensor core opts must have valid tc_select")
+      check(-1 <= (tc_select:=cast(tuple, opt.arg)[0]) < len(self.ren.tensor_cores), "tensor core opts must have valid tc_select")
       check(0 <= (tc_opt:=cast(tuple, opt.arg)[1]) <= 2, "tensor core opts must have valid tc_opt")
       check(0 < (use_tensor_cores:=cast(tuple, opt.arg)[2]) <= 2, "use_tensor_cores value is not valid")
       try: ret = self._apply_tc_opt(use_tensor_cores, cast(int, opt.axis), tc_select, tc_opt)
@@ -223,7 +216,7 @@ class Scheduler:
       if mul.op is not Ops.MUL: return None
       in0, in1 = mul.src
       try:
-        tensor_cores = self.opts.tensor_cores if tc_select == -1 else [self.opts.tensor_cores[tc_select]]
+        tensor_cores = self.ren.tensor_cores if tc_select == -1 else [self.ren.tensor_cores[tc_select]]
       except IndexError:
         raise KernelOptError(f"invalid tensor core choice {tc_select}")
       for tc in tensor_cores:
@@ -294,7 +287,7 @@ class Scheduler:
             # TODO: remove tc_upcast_axes from the arg
             # do the reduce_axes always disappear? i think they don't
             # they need to be moved into the WMMA srcs
-            wmma_arg = (str(tc), tc.dims, tc.dtype_in, tc.dtype_out, self.opts.device, tc.threads, tc_upcast_axes, ()) #, tc_reduce_axes)
+            wmma_arg = (str(tc), tc.dims, tc.dtype_in, tc.dtype_out, self.ren.device, tc.threads, tc_upcast_axes, ()) #, tc_reduce_axes)
             wmma = UOp(Ops.WMMA, dtype=tc.dtype_out.vec(tc.elements_per_thread[2]), src=(
               UOp(Ops.CONTRACT, dtype=srcs[0].dtype.vec(tc.elements_per_thread[0]), src=(srcs[0],), arg=tc_upcast_axes[0], tag=1),
               UOp(Ops.CONTRACT, dtype=srcs[1].dtype.vec(tc.elements_per_thread[1]), src=(srcs[1],), arg=tc_upcast_axes[1], tag=1),
@@ -328,23 +321,19 @@ def bufs_from_ast(ast:UOp, dname:str) -> list[Buffer]:
   glbls = sorted([x for x in ast.backward_slice if x.op is Ops.DEFINE_GLOBAL], key=lambda x: x.arg)
   return [Buffer(dname, x.ptrdtype.size, x.dtype.base if not isinstance(x.dtype, ImageDType) else x.dtype) for x in glbls]
 
-def apply_opts(ctx:Renderer, ast:UOp):
-  if ast.tag is not None: return None
-  k = Scheduler(ast, ctx)
+def apply_opts(ast:UOp, ren:Renderer) -> UOp:
+  if ast.tag is not None: return ast
+  k = Scheduler(ast, ren)
   k.convert_loop_to_global()
   if ast.arg is not None and ast.arg.opts_to_apply is not None:
     for opt in ast.arg.opts_to_apply: k.apply_opt(opt)
   elif BEAM >= 1:
     from tinygrad.codegen.opt.search import beam_search
-    rawbufs = bufs_from_ast(ast, ctx.device)
+    rawbufs = bufs_from_ast(ast, ren.device)
     k = beam_search(k, rawbufs, BEAM.value, bool(getenv("BEAM_ESTIMATE", 1)))
   elif not NOOPT and (ast.arg is None or ast.arg.applied_opts == ()):
     from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
     # NOTE: hand_coded_optimizations doesn't support multiblock opts yet
-    if all(len(u.src) == 1 for u in ast.backward_slice if u.op is Ops.LOAD):
+    if not any(u.op is Ops.AFTER and u.src[0].op is Ops.DEFINE_LOCAL for u in ast.backward_slice):
       k = hand_coded_optimizations(k)
   return k.get_optimized_ast(name_override=ast.arg.name if ast.arg is not None and ast.arg.name != "test" else None)
-
-pm_postrange_opt = PatternMatcher([
-  (UPat(Ops.SINK, name="ast"), apply_opts),
-])

tinygrad/codegen/opt/search.py

@@ -66,7 +66,7 @@ def _try_compile_linearized_w_idx(x:tuple[int,Scheduler], compiler:Compiler) ->
     signal.alarm(getenv("BEAM_TIMEOUT_SEC", 10))
   ret = None
   try:
-    p = get_program(x[1].copy().get_optimized_ast(name_override="test"), x[1].opts)
+    p = get_program(x[1].copy().get_optimized_ast(name_override="test"), x[1].ren)
     assert p.uops is not None, "uop list wasn't generated?"
     if len(p.uops) >= (uops_max:=getenv("BEAM_UOPS_MAX", 3000)) > 0:
       if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too many uops. {len(p.uops)=}, {uops_max=}")
@@ -119,7 +119,7 @@ def get_kernel_actions(lin:Scheduler, include_0=True, candidates:list[Opt]|None=
 beam_pool, BEAM_DEBUG = None, getenv("BEAM_DEBUG")
 def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=True, disable_cache=IGNORE_BEAM_CACHE.value):
   global beam_pool
-  key = {"ast": lin.ast.key, "amt": amt, "allow_test_size": allow_test_size, "device": lin.opts.device, "suffix": lin.opts.suffix}
+  key = {"ast": lin.ast.key, "amt": amt, "allow_test_size": allow_test_size, "device": lin.ren.device, "suffix": lin.ren.suffix}
   if not disable_cache and CACHELEVEL >= 1 and (val:=diskcache_get("beam_search", key)) is not None:
     ret = lin.copy()
     for o in val[len(lin.applied_opts):]: ret.apply_opt(o)
@@ -128,7 +128,7 @@ def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=Tr
   beam: list[tuple[Scheduler, float]] = [(lin, float("inf"))]
   seen_libs = set()
 
-  default_parallel = multiprocessing.cpu_count() if lin.opts.device in {"CUDA", "AMD", "NV", "METAL", "HIP"} else 0
+  default_parallel = multiprocessing.cpu_count() if lin.ren.device in {"CUDA", "AMD", "NV", "METAL", "HIP"} else 0
   if beam_pool is None and (workers := getenv("PARALLEL", default_parallel)):
     beam_pool = multiprocessing.get_context("spawn").Pool(workers, _init_worker, (), getenv("BEAM_MAX_TASKS_PER_CHILD", 16))
     @atexit.register
@@ -137,14 +137,14 @@ def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=Tr
   min_progress = getenv("BEAM_MIN_PROGRESS", 0.01)/1e6
   if BEAM_DEBUG:
     print("BEAM_SEARCH:")
-    print('\n'.join(pyrender(lin.ast.replace(arg=None))))
+    print(pyrender(lin.ast.replace(arg=None)))
   if DEBUG >= 2: print(f"   0.00s:                from   1 ->   1 actions {lin.colored_shape()}")
 
   try:
     rawbufs = _ensure_buffer_alloc(rawbufs)
     var_vals: dict[str, int] = {k.expr:int(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
     exiting, st = False, time.perf_counter()
-    dev = Device[lin.opts.device]
+    dev = Device[lin.ren.device]
     while not exiting:
       acted_lins: list[Scheduler] = flatten([get_kernel_actions(lin, include_0=False).values() for lin,_ in beam])
       timed_lins: list[tuple[Scheduler, float]] = []
@@ -156,7 +156,9 @@ def beam_search(lin:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=Tr
         if lib in seen_libs: continue
         # filter out kernels that use 1000x more compute than the smallest
         least_compute_ops = min(this_compute_ops:=sym_infer(p.estimates.ops, var_vals), least_compute_ops)
-        if least_compute_ops*1000 < this_compute_ops: continue
+        if least_compute_ops*1000 < this_compute_ops:
+          if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too much compute. {this_compute_ops} when least is {least_compute_ops}")
+          continue
         seen_libs.add(lib)
         try: tms = _time_program(p, lib, var_vals, rawbufs, early_stop=beam[0][1]*3 if len(beam) else 1.0,
                                  allow_test_size=allow_test_size, clear_l2=hasattr(dev, 'invalidate_caches'))

tinygrad/codegen/opt/tc.py

@@ -80,6 +80,10 @@ cuda_81616 = [TensorCore(dims=(8,16,16), threads=32, elements_per_thread=(8,4,4)
   swizzle=((('r1', 'r2', 'l2', 'l3', 'l4'), ('u1', 'r3'), ('l0', 'l1', 'u0', 'r0')),
            (('r1', 'r2', 'u0', 'l0', 'l1'), ('r0', 'r3'), ('l2', 'l3', 'l4', 'u1'))))
   for di,do in [(dtypes.half,dtypes.float), (dtypes.bfloat16,dtypes.float), (dtypes.half,dtypes.half)]]
+cuda_81632_f8 = [TensorCore(dims=(8,16,32), threads=32, elements_per_thread=(16,8,4), dtype_in=di, dtype_out=do, opts=cuda_tc_opts,
+  swizzle=((('r2', 'r3', 'l2', 'l3', 'l4'), ('u1', 'r4'), ('l0', 'l1', 'u0', 'r0', 'r1')),
+           (('r2', 'r3', 'u0', 'l0', 'l1'), ('r1', 'r4'), ('l2', 'l3', 'l4', 'u1', 'r0'))))
+  for di,do in [(dtypes.fp8e4m3,dtypes.float),(dtypes.fp8e5m2,dtypes.float)]]
 cuda_8168_f16 = [TensorCore(dims=(8,16,8), threads=32, elements_per_thread=(4,2,4), dtype_in=di, dtype_out=do, opts=cuda_tc_opts,
   swizzle=((('r1', 'r2', 'l2', 'l3', 'l4'), ('r0', 'u1'), ('l0', 'l1', 'u0')),
            (('r1', 'r2', 'u0', 'l0', 'l1'), ('u1', 'r0'), ('l2', 'l3', 'l4'))))
@@ -87,9 +91,10 @@ cuda_8168_f16 = [TensorCore(dims=(8,16,8), threads=32, elements_per_thread=(4,2,
 cuda_8168_tf32 = [TensorCore(dims=(8,16,8), threads=32, elements_per_thread=(4,2,4), dtype_in=dtypes.float, dtype_out=dtypes.float, opts=cuda_tc_opts,
   swizzle=((('r0', 'r1', 'l2', 'l3', 'l4'), ('u1', 'r2'), ('l0', 'l1', 'u0')),
            (('r0', 'r1', 'u0', 'l0', 'l1'), ('u1', 'r2'), ('l2', 'l3', 'l4'))))]
+cuda_sm75: list[TensorCore] = cuda_8168_f16
 cuda_sm80: list[TensorCore] = cuda_81616 + cuda_8168_f16
 if getenv("ALLOW_TF32", 0): cuda_sm80 += cuda_8168_tf32
-cuda_sm75: list[TensorCore] = cuda_8168_f16
+cuda_sm89: list[TensorCore] = cuda_sm80 + cuda_81632_f8
 
 # ***** AMD *****
 
@@ -112,6 +117,14 @@ amd_cdna = [TensorCore(dims=(16,16,16), threads=64, elements_per_thread=(4,4,4),
            (('l0', 'l1', 'l2', 'l3', 'r2', 'r3'), ('r0', 'r1'), ('l4', 'l5', 'u0', 'u1'))))
   for di,do in [(dtypes.half,dtypes.float),(dtypes.bfloat16,dtypes.float)]]
 
+amd_cdna_161632 = [TensorCore(dims=(16,16,32), threads=64, elements_per_thread=(8,8,4), dtype_in=di, dtype_out=do,
+  opts=("l0","l0","l0","l0","u1","u1","l1","l1"),
+  swizzle=((('u0','u1','l4','l5','r3','r4'), ('r0','r1'), ('l0','l1','l2','l3','r2')),
+           (('l0','l1','l2','l3','r3','r4'), ('r0','r1'), ('l4','l5','u0','u1','r2'))))
+  for di,do in [(dtypes.half,dtypes.float),(dtypes.bfloat16,dtypes.float)]]
+
+amd_cdna4 = amd_cdna_161632 + amd_cdna
+
 # ***** Apple Metal *****
 
 metal = [TensorCore(dims=(8,8,8), threads=32, elements_per_thread=(2,2,2), dtype_in=di, dtype_out=do,

tinygrad/codegen/simplify.py

@@ -1,6 +1,6 @@
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, graph_rewrite, _substitute, range_start, ImageDType
-from tinygrad.uop.symbolic import symbolic_flat, sym
-from tinygrad.helpers import partition
+from tinygrad.uop.symbolic import symbolic_flat
+from tinygrad.helpers import partition, dedup
 from tinygrad.dtype import dtypes
 
 def flatten_range(r:UOp):
@@ -12,15 +12,15 @@ def flatten_range(r:UOp):
 
 pm_flatten_range = PatternMatcher([
   # real ranges only
-  (UPat((Ops.REDUCE, Ops.STORE), name="r"), flatten_range),
+  (UPat((Ops.REDUCE, Ops.STORE, Ops.END), name="r"), flatten_range),
 ])
 
 def count_divmod(x:UOp): return len([u for u in x.toposort() if u.op in {Ops.IDIV, Ops.MOD}])
 def simplify_merge_adjacent(u:UOp) -> UOp|None:
   reduce_ranges = [x.ranges for x in u.backward_slice_with_self if x.op is Ops.REDUCE]
-  i = range_start[u.op]
-  while i < len(u.src)-1:
-    r0, r1 = u.src[i], u.src[i+1]
+  i = 0
+  while i < len(u.ended_ranges)-1:
+    r0, r1 = u.ended_ranges[i], u.ended_ranges[i+1]
     # check same type
     if r0.arg[-1] == r1.arg[-1]:
       # check if the ranges to merge are in the same reduces
@@ -39,7 +39,7 @@ def simplify_merge_adjacent(u:UOp) -> UOp|None:
   return u
 
 pm_simplify_ranges = PatternMatcher([
-  (UPat((Ops.STORE, Ops.REDUCE), name="u"), simplify_merge_adjacent),
+  (UPat((Ops.END, Ops.REDUCE), name="u"), simplify_merge_adjacent),
 ])
 
 def mark_range_mod(ctx, r:UOp, c:UOp):
@@ -57,7 +57,7 @@ def do_substitute(ctx, x: UOp):
 
 def dont_sub_ranges_for_image(ctx, x:UOp):
   if isinstance(x.src[0].dtype, ImageDType):
-    for s in x.src[1:]: ctx[s] = None
+    for s in x.src[0].ranges: ctx[s] = None
 
 pm_split_ranges = PatternMatcher([
   (UPat(Ops.RANGE, name="r")%UPat.cvar("c"), mark_range_mod),
@@ -69,59 +69,6 @@ pm_split_ranges = PatternMatcher([
 
 def no_range(u:UOp) -> bool: return not any(x.op is Ops.RANGE for x in u.backward_slice_with_self)
 
-def reduce_rangeless(red:UOp):
-  # TODO: share code with reduce_unparented
-  if red.arg not in {Ops.ADD, Ops.MAX}: return None
-  if red.src[0].dtype != red.dtype: return None
-  if not no_range(red.src[0]): return None
-  ret = red.src[0]
-  if red.arg is Ops.ADD:
-    for r in red.src[1:]:
-      ret = ret * r.src[0].cast(ret.dtype.scalar()).broadcast(ret.dtype.count)
-  return ret
-
-pm_reduce_collapse = PatternMatcher([
-  # lift x+y out of reduce on lt
-  ((UPat.var("x")+UPat.var("y")).or_casted() < UPat.var("c"), lambda x,y,c: (x < (c.cast(y.dtype)-y)) if no_range(y) and no_range(c) else None),
-  # lift x*y out of reduce
-  ((UPat.var("x")*UPat.var("y")) < UPat.var("c"),
-   lambda x,y,c: (x < ((c+y-1) // y)) if no_range(y) and no_range(c) and y.vmin > 0 else None),
-  # lift x+y out of reduce on ne
-  ((UPat.var("x")+UPat.var("y")).or_casted() != UPat.var("c"), lambda x,y,c: (x != (c.cast(y.dtype)-y)) if no_range(y) and no_range(c) else None),
-  # fold the range
-  ((UPat(Ops.RANGE, name="r") < UPat.var("cut")).where(0, UPat.cvar("val")).reduce(UPat.var("r"), arg=Ops.ADD),
-   lambda r,cut,val: (r.src[0]-cut).maximum(0).minimum(r.src[0]).cast(val.dtype) * val),
-  (((UPat.var("r")<UPat.var("lower")).logical_not()&(UPat(Ops.RANGE, name="r")<UPat.var("upper"))).where(UPat.cvar("val"), 0).reduce(UPat.var("r"),
-    arg=Ops.ADD), lambda r,lower,upper,val: (upper.minimum(r.src[0])-lower.maximum(0)).maximum(0).minimum(r.src[0]).cast(val.dtype) * val),
-  ((UPat(Ops.RANGE, name="r") < UPat.var("cut")).where(UPat.cvar("val"), 0).reduce(UPat.var("r"), arg=Ops.ADD),
-   lambda r,cut,val: cut.maximum(0).minimum(r.src[0]).cast(val.dtype) * val),
-  # REDUCE on ADD
-  ((UPat.var("x")+UPat.var("y")).reduce(arg=Ops.ADD, allow_any_len=True, name="r"),
-   lambda x,y,r: x.reduce(*r.src[1:], arg=Ops.ADD) + y.reduce(*r.src[1:],arg=Ops.ADD)),
-  # MUL casted bool
-  ((UPat.var("x") * UPat.var("gate", dtype=dtypes.bool).cast()), lambda x,gate: gate.where(x, 0)),
-  # reduce on gated load becomes can substitute the range and remove the reduce
-  ((UPat.var("idx")!=(UPat(Ops.RANGE, name="r").or_casted())).where(0, UPat.var("expr")).reduce(UPat.var("r"), arg=Ops.ADD),
-   lambda r,idx,expr: (v:=(idx.cast(r.dtype) >= 0) & (idx.cast(r.dtype) < r.src[0])).where(expr.substitute({r:idx.cast(r.dtype).valid(v)}),0)),
-  # AND on WHERE
-  ((UPat(Ops.DEFINE_VAR, name="x") & UPat.var("y")).where(UPat.cvar("c"), 0).reduce(arg=Ops.ADD, allow_any_len=True, name="r"),
-    lambda x,y,c,r: y.where(c, 0).reduce(*r.src[1:], arg=Ops.ADD)*x.cast(c.dtype)),
-  # remove REDUCEs that no longer have a RANGE in the src
-  (UPat(Ops.REDUCE, name="red"), reduce_rangeless),
-])+sym
-
-def reduce_collapse(red:UOp):
-  included, not_included = partition(red.backward_slice, lambda x: any(y in x.backward_slice_with_self for y in red.src[1:]))
-  if any(x.op in {Ops.STORE, Ops.REDUCE} for x in included): return None
-  replaces: dict[UOp, UOp] = {}
-  for u in included:
-    for s in u.src:
-      if s in not_included and s not in replaces and s.op not in {Ops.CONST, Ops.VCONST, Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_VAR}:
-        replaces[s] = UOp(Ops.DEFINE_VAR, dtype=s.dtype, arg=(f'in{len(replaces)}', s.vmin, s.vmax))
-  collapse_fxn = red.substitute(replaces)
-  sink = graph_rewrite(collapse_fxn, pm_reduce_collapse, name="reduce_collapse")
-  return sink.substitute({v:k for k,v in replaces.items()}) if no_range(sink) else None
-
 def reduce_unparented(red:UOp):
   if red.arg not in {Ops.ADD, Ops.MAX, Ops.MUL}: return None
   assert all(x.op is Ops.RANGE for x in red.src[1:]), "some reduce srcs aren't ranges"
@@ -139,7 +86,66 @@ pm_reduce_unparented = PatternMatcher([
   (UPat(Ops.REDUCE, name="red"), reduce_unparented),
 ])
 
-pm_reduce_simplify = pm_reduce_unparented + PatternMatcher([
-  # remove REDUCE without loads (generic arange opt / indexing). TODO: support multi range
-  (UPat(Ops.REDUCE, src=(UPat(), UPat()), name="red"), reduce_collapse),
+pm_reduce_collapse = pm_reduce_unparented + PatternMatcher([
+  # lift x+y out of reduce on lt
+  ((UPat.var("x")+UPat.var("y")).or_casted() < UPat.var("c"), lambda x,y,c: (x < (c.cast(y.dtype)-y)) if no_range(y) and no_range(c) else None),
+  # lift x*y out of reduce
+  ((UPat.var("x")*UPat.var("y")) < UPat.var("c"), lambda x,y,c: (x < ((c+y-1) // y)) if no_range(y) and no_range(c) and y.vmin > 0 else None),
+  # fold the range
+  ((UPat(Ops.RANGE, name="r") < UPat.var("cut")).where(0, UPat.cvar("val")).reduce(UPat.var("r"), arg=Ops.ADD),
+   lambda r,cut,val: (r.src[0]-cut).maximum(0).minimum(r.src[0]).cast(val.dtype) * val),
+  (((UPat.var("r")<UPat.var("lower")).logical_not()&(UPat(Ops.RANGE, name="r")<UPat.var("upper"))).where(UPat.cvar("val"), 0).reduce(UPat.var("r"),
+    arg=Ops.ADD), lambda r,lower,upper,val: (upper.minimum(r.src[0])-lower.maximum(0)).maximum(0).minimum(r.src[0]).cast(val.dtype) * val),
+  ((UPat(Ops.RANGE, name="r") < UPat.var("cut")).where(UPat.cvar("val"), 0).reduce(UPat.var("r"), arg=Ops.ADD),
+   lambda r,cut,val: cut.maximum(0).minimum(r.src[0]).cast(val.dtype) * val),
+  # REDUCE on ADD
+  ((UPat.var("x")+UPat.var("y")).reduce(arg=Ops.ADD, allow_any_len=True, name="r"),
+   lambda x,y,r: x.reduce(*r.src[1:], arg=Ops.ADD) + y.reduce(*r.src[1:],arg=Ops.ADD)),
+])+symbolic_flat
+
+pm_reduce_load_collapse = PatternMatcher([
+  # MUL casted bool
+  ((UPat.var("x") * UPat.var("gate", dtype=dtypes.bool).cast()), lambda x,gate: gate.where(x, 0)),
+  # lift x+y out of reduce on ne
+  ((UPat.var("x")+UPat.var("y")).or_casted() != UPat.var("c"), lambda x,y,c: (x != (c.cast(y.dtype)-y)) if no_range(y) and no_range(c) else None),
+  # reduce on gated load becomes can substitute the range and remove the reduce
+  ((UPat.var("idx")!=(UPat(Ops.RANGE, name="r").or_casted())).where(0, UPat.var("expr")).reduce(UPat.var("r"), arg=Ops.ADD),
+   lambda r,idx,expr: (v:=(idx.cast(r.dtype) >= 0) & (idx.cast(r.dtype) < r.src[0])).where(expr.substitute({r:idx.cast(r.dtype).valid(v)}),0)),
+])+symbolic_flat
+
+def reduce_collapse(red:UOp, pm=pm_reduce_collapse):
+  included = red.src[0].toposort(gate=lambda x: any(y in x.ranges for y in red.src[1:]))
+  if any(x.op in {Ops.STORE, Ops.REDUCE} for x in included): return None
+  replaces: dict[UOp, UOp] = {}
+  for u in included:
+    for s in u.src:
+      if s in included or s in replaces or s.op in {Ops.CONST, Ops.VCONST, Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_VAR}: continue
+      replaces[s] = UOp(Ops.DEFINE_VAR, dtype=s.dtype, arg=(f'in{len(replaces)}', s.vmin, s.vmax))
+  collapse_fxn = red.substitute(replaces)
+  sink = graph_rewrite(collapse_fxn, pm, name="reduce_collapse")
+  return sink.substitute({v:k for k,v in replaces.items()}) if no_range(sink) else None
+
+def reduce_load_collapse(red:UOp): return reduce_collapse(red, pm=pm_reduce_load_collapse)
+
+# remove REDUCE without loads (generic arange opt / indexing). TODO: support multi range
+pm_reduce_simplify = pm_reduce_unparented + PatternMatcher([(UPat(Ops.REDUCE, src=(UPat(), UPat()), name="red"), reduce_collapse),])
+# remove REDUCE on load, comes from indexing a tensor with another tensor
+def no_load(u:UOp) -> bool: return not any(x.op is Ops.LOAD for x in u.backward_slice_with_self)
+pm_load_collapse = PatternMatcher([
+  (UPat(Ops.REDUCE, src=(UPat(), UPat()), name="red"), reduce_load_collapse),
+  # we want to make sure we dont do math on a loaded index since that can cause overflow, this undoes the rule in pm_reduce_load_collapse
+  ((UPat.var("x", dtypes.index)+UPat.var("y"))<UPat.var("c"), lambda x,y,c: x < c-y if no_load(y) and no_load(c) and not no_load(x) else None),
+])
+
+def cut_store_range(ctx, store:UOp, r:UOp):
+  # only cut ranges on CPU for now
+  if r.src[0].op is not Ops.CONST or ctx!="CPU": return None
+  if not (cuts:=[c.src[1].arg for c in store.get_consumer_map()[r] if c.op is Ops.CMPLT and r is c.src[0] and c.src[1].op is Ops.CONST]): return None
+  cuts = sorted(dedup([0] + cuts + [r.src[0].arg]))
+  ranges = [UOp.range((end-start), *(r.arg[0:-1]+(i,r.arg[-1]))) for i,(start,end) in enumerate(zip(cuts[:-1], cuts[1:]))]
+
+  return UOp.group(*[store.substitute({r: new_r+start}).end(new_r) for new_r, start in zip(ranges, cuts[:-1])])
+
+pm_split_store = pm_flatten_range+PatternMatcher([
+  (UPat(Ops.END, src=(UPat(Ops.STORE, name="store"), UPat.var("r"))), cut_store_range),
 ])

tinygrad/device.py

@@ -5,7 +5,7 @@ from typing import Any, Generic, TypeVar, Iterator, Sequence, cast, Generator
 import importlib, inspect, functools, pathlib, os, platform, contextlib, sys, re, atexit, pickle, decimal
 from tinygrad.helpers import CI, OSX, LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, flat_mv, PROFILE, temp, colored, CPU_LLVM
 from tinygrad.helpers import Context, DISABLE_COMPILER_CACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, dedup
-from tinygrad.helpers import unwrap_class_type
+from tinygrad.helpers import unwrap_class_type, suppress_finalizing
 from tinygrad.dtype import DType, ImageDType, PtrDType, dtypes, _to_np_dtype
 from tinygrad.renderer import Renderer
 
@@ -163,6 +163,7 @@ class Buffer:
     return self._trace_num
   @property
   def nbytes(self): return self.size*self.dtype.itemsize
+  @suppress_finalizing
   def __del__(self): (not hasattr(self, '_buf')) or self.deallocate()
   def __repr__(self):
     return f"<buf real:{self.is_allocated()} device:{self.device} size:{self.size} dtype:{self.dtype}" + \
@@ -330,7 +331,9 @@ def is_dtype_supported(dtype:DType, device:str|None=None) -> bool:
     if device in {"CUDA", "NV"}: return not CI and not getenv(f"{device}_PTX") and not getenv("NV_NAK")
     if device in {"CPU"}: return not CI and platform.machine() in {"arm", "arm64", "aarch64", "x86_64", "amd64"} and not getenv("CPU_LVP")
     return device in {"AMD", "PYTHON", "NULL"}
-  if dtype in dtypes.fp8s: return device in {"PYTHON", "NULL"}
+  if dtype in dtypes.fp8s:
+    if device in {"CUDA", "NV"}: return not CI and not getenv(f"{device}_PTX") and not getenv("NV_NAK")
+    return device in {"PYTHON", "NULL"}
   if device == "WEBGPU": return dtype in [dtypes.bool, dtypes.char, dtypes.uchar, dtypes.short,
                                           dtypes.ushort, dtypes.float, dtypes.int32, dtypes.uint32, dtypes.half]
   # for CI GPU and OSX, cl_khr_fp16 isn't supported

tinygrad/dtype.py

@@ -7,13 +7,15 @@ from enum import Enum, auto
 
 class InvalidTypeMetaClass(type):
   instance:None|InvalidType = None
-  def __call__(cls, *args, **kwargs):
+  def __call__(cls):
     if (ret:=InvalidTypeMetaClass.instance) is not None: return ret
     InvalidTypeMetaClass.instance = ret = super().__call__()
     return ret
 
 class InvalidType(metaclass=InvalidTypeMetaClass):
   def __eq__(self, other): return self is other
+  def __lt__(self, other): return self is not other
+  def __gt__(self, other): return self is not other
   def __hash__(self): return id(self)
   def __repr__(self): return "Invalid"
   def __reduce__(self): return (InvalidType, ())  # Return the global Invalid instance
@@ -47,7 +49,7 @@ class DType(metaclass=DTypeMetaClass):
   @staticmethod
   def new(priority:int, itemsize:int, name:str, fmt:FmtStr|None): return DType(priority, itemsize, name, fmt, 1, None)
   def __reduce__(self): return type(self), tuple(getattr(self, f.name) for f in fields(self))
-  def __repr__(self): return f"dtypes.{INVERSE_DTYPES_DICT[self.scalar().name]}"+(f".vec({self.count})" if self.count > 1 else "")
+  def __repr__(self): return f"dtypes.{INVERSE_DTYPES_DICT[self.scalar().name]}"+(f".vec({self.count})" if self.count != 1 else "")
   def __lt__(self, o:DType): return (self.priority, self.itemsize, self.name, self.fmt, self.count) < (o.priority, o.itemsize, o.name, o.fmt, o.count)
   @property
   def base(self): return self
@@ -61,7 +63,7 @@ class DType(metaclass=DTypeMetaClass):
   def ptr(self, size=-1, addrspace=AddrSpace.GLOBAL) -> PtrDType:
     return PtrDType(self.priority, self.itemsize, self.name, self.fmt, self.count, None, self, addrspace, 1, size)
   def scalar(self) -> DType: return self._scalar if self._scalar is not None else self
-  def nbytes(self): raise RuntimeError("only ptr types have nbytes")
+  def nbytes(self) -> int: raise RuntimeError("only ptr types have nbytes")
   @property
   def min(self): return dtypes.min(self)
   @property
@@ -82,7 +84,7 @@ class PtrDType(DType):
     if isinstance(self, ImageDType):
       return ImageDType(self.priority, self.itemsize, self.name, self.fmt, self.count, self, self._base, self.addrspace, sz, self.size, self.shape)
     return type(self)(self.priority, self.itemsize, self.name, self.fmt, self.count, self, self._base, self.addrspace, sz, self.size)
-  def ptr(self, size=-1, addrspace=AddrSpace.GLOBAL): raise RuntimeError("can't make a pointer from a pointer")
+  def ptr(self, size=-1, addrspace=AddrSpace.GLOBAL) -> PtrDType: raise RuntimeError("can't make a pointer from a pointer")
   def nbytes(self) -> int:
     if self.size == -1: raise RuntimeError("can't get nbytes of a pointer with unlimited size")
     return self.size*self.itemsize

tinygrad/engine/realize.py

@@ -26,7 +26,7 @@ def get_program(ast:UOp, renderer:Renderer|None=None, opts:list[Opt]|None=None)
   """
 
   if getenv("VIZ"): graph_rewrite(ast, PatternMatcher([]), name="View Base AST")
-  if DEBUG >= 5: print('\n'.join(pyrender(ast)))
+  if DEBUG >= 5: print(pyrender(ast))
 
   # linearize
   if renderer is None: renderer = Device.default.renderer
@@ -38,7 +38,7 @@ def get_program(ast:UOp, renderer:Renderer|None=None, opts:list[Opt]|None=None)
   except RuntimeError as e:
     print("***** LINEARIZE FAILURE *****")
     print(e)
-    print('\n'.join(pyrender(ast)))
+    print(pyrender(ast))
     raise
   assert uops[-1].op is Ops.SINK, "last uop must be sink"
 
@@ -166,8 +166,9 @@ class ExecItem:
     var_vals = self.fixedvars if _var_vals is None else (_var_vals|self.fixedvars)
     bufs = [cast(Buffer, x) for x in self.bufs] if jit else [cast(Buffer, x).ensure_allocated() for x in self.bufs]
     if PROFILE:
-      payload = {"metadata":self.metadata, "var_vals":var_vals, "bufs":[b.trace_num for b in bufs]}
-      cpu_events.append(ProfilePointEvent(self.prg.device, "exec", self.prg.display_name, payload))
+      payload = {"metadata":self.metadata, "var_vals":var_vals, "bufs":[b.trace_num for b in bufs], "name":self.prg.display_name}
+      payload["outputs"], payload["inputs"] = (self.prg.p.outs, self.prg.p.ins) if isinstance(self.prg, CompiledRunner) else ([0], [1])
+      cpu_events.append(ProfilePointEvent(self.prg.device, "exec", len(cpu_events), payload))
     et = self.prg(bufs, var_vals, wait=wait or DEBUG >= 2)
     if do_update_stats:
       GlobalCounters.kernel_count += 1
@@ -180,10 +181,11 @@ class ExecItem:
         header_color = 'magenta' if jit else ('green' if self.prg.first_run else None)
         ptm = colored(time_to_str(et, w=9), "yellow" if et > 0.01 else None) if et is not None else ""
         flops, membw, ldsbw = op_est/(et or 1e-20), mem_est/(et or 1e-20), lds_est/(et or 1e-20)
-        flops_str = f"{flops*1e-9:9.2f} GFLOPS" if flops < 1e14 else colored(f"{flops*1e-12:9.2f} TFLOPS", 'green')
-        mem_str = f"{membw*1e-9:6.1f}|{ldsbw*1e-9:<7.1f} GB/s" if membw < 1e13 else colored(f"{membw*1e-12:6.1f}|{ldsbw*1e-12:<7.1f} TB/s", 'green')
+        flops_str = f"{flops*1e-9:7.0f} GFLOPS" if flops < 1e14 else colored(f"{flops*1e-12:7.0f} TFLOPS", 'green')
+        mem_str = f"{membw*1e-9:4.0f}|{ldsbw*1e-9:<6.0f} GB/s" if membw < 1e13 and ldsbw < 1e15 else \
+          colored(f"{membw*1e-12:4.0f}|{ldsbw*1e-12:<6.0f} TB/s", 'green')
         print(f"{colored(f'*** {self.prg.device[:7]:7s} {GlobalCounters.kernel_count:4d}', header_color)}"+
-          f" {self.prg.display_name+' '*(44-ansilen(self.prg.display_name))} arg {len(bufs):2d} mem {GlobalCounters.mem_used/1e9:5.2f} GB"+
+          f" {self.prg.display_name+' '*(46-ansilen(self.prg.display_name))} arg {len(bufs):2d} mem {GlobalCounters.mem_used/1e9:6.2f} GB"+
           ("" if et is None else f" tm {ptm}/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({flops_str} {mem_str})")+
           f" {[repr(m) if TRACEMETA >= 2 else str(m) for m in self.metadata] if self.metadata else ''}")
       self.prg.first_run = False

tinygrad/engine/schedule.py

@@ -22,18 +22,18 @@ def create_schedule_with_vars(sched_sink:UOp) -> tuple[list[ScheduleItem], dict[
   in_degree: dict[UOp, int] = {}
   var_vals: dict[str, int] = {}
   for u in sched_sink.toposort():
-    if u.op is not Ops.ASSIGN: continue  # anything that's not an ASSIGN doesn't write a kernel, so we can skip
+    if u.op is not Ops.AFTER: continue  # anything that's not an ASSIGN doesn't write a kernel, so we can skip
     k = u.src[1]
     in_degree.setdefault(k, 0)
     for s in k.src:
-      if s.op is Ops.ASSIGN:
+      if s.op is Ops.AFTER:
         children[s.src[1]].append(k)
         in_degree[k] += 1
       elif s.op in {Ops.MSELECT, Ops.MSTACK}:
         for ss in s.src:
           if ss.op is Ops.MSELECT: ss = ss.src[0]
           if ss.op is not Ops.BUFFER:
-            assert ss.op is Ops.ASSIGN, f"ss.op is not ASSIGN, it's {ss.op}"
+            assert ss.op is Ops.AFTER, f"ss.op is not AFTER, it's {ss.op}"
             children[ss.src[1]].append(k)
             in_degree[k] += 1
       elif s.op is Ops.BUFFER:
@@ -43,7 +43,7 @@ def create_schedule_with_vars(sched_sink:UOp) -> tuple[list[ScheduleItem], dict[
         assert var.expr not in var_vals or var_vals[var.expr] == val, f"bind mismatch on {var}, {var_vals[var.expr]} != {val}"
         var_vals[var.expr] = val
       else:
-        raise RuntimeError(f"input to kernel must be ASSIGN or BUFFER, not {s.op}")
+        raise RuntimeError(f"input to kernel must be AFTER or BUFFER, not {s.op}")
 
   # linearize KERNEL UOps into ScheduleItems in BFS order
 

tinygrad/gradient.py

@@ -15,7 +15,7 @@ def reduce_gradient(ctx:UOp, ret:UOp):
 # ctx is grad_output
 pm_gradient = PatternMatcher([
   (UPat(Ops.CAST, name="ret"), lambda ctx, ret: (ctx.cast(ret.src[0].dtype),)),
-  (UPat(Ops.RECIP, name="ret"), lambda ctx, ret: (-ctx * ret * ret,)),
+  (UPat(Ops.RECIPROCAL, name="ret"), lambda ctx, ret: (-ctx * ret * ret,)),
   (UPat(Ops.SIN, name="ret"), lambda ctx, ret: ((math.pi/2 - ret.src[0]).sin() * ctx,)),
   (UPat(Ops.LOG2, name="ret"), lambda ctx, ret: (ctx / (ret.src[0] * math.log(2)),)),
   (UPat(Ops.EXP2, name="ret"), lambda ctx, ret: (ret * ctx * math.log(2),)),
@@ -24,7 +24,7 @@ pm_gradient = PatternMatcher([
   (UPat(Ops.ADD), lambda ctx: (ctx, ctx)),
   (UPat(Ops.POW, name="ret", src=(UPat.var("b"), UPat.var("e"))), lambda ctx, ret, b, e:
     (ctx * (b.eq(0)&e.eq(0)).where(e, e*b.pow(e-1)), ctx * b.eq(0).where((e<0).where(ret.const_like(-math.inf), 0), ret*b.log2()*math.log(2.0)))),
-  (UPat(Ops.MAX, name="ret", src=(UPat.var("x"), UPat.var("y"))), lambda ctx, ret, x, y:
+  (UPat(Ops.MAX, src=(UPat.var("x"), UPat.var("y"))), lambda ctx, x, y:
     ((x>y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)), (x<y).where(ctx, (x.eq(y)).where(ctx * 0.5, 0)))),
   (UPat(Ops.MUL, name="ret"), lambda ctx, ret: (ret.src[1]*ctx, ret.src[0]*ctx)),
   (UPat(Ops.WHERE, name="ret"), lambda ctx, ret: (None, ret.src[0].where(ctx, ctx.const_like(0)), ret.src[0].where(ctx.const_like(0), ctx))),

tinygrad/helpers.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 import os, functools, platform, time, re, contextlib, operator, hashlib, pickle, sqlite3, tempfile, pathlib, string, ctypes, sys, gzip, getpass
 import urllib.request, subprocess, shutil, math, types, copyreg, inspect, importlib, decimal, itertools
 from dataclasses import dataclass, field
-from typing import ClassVar, Iterable, Any, TypeVar, Callable, Sequence, TypeGuard, Iterator, Generic, Generator, cast
+from typing import ClassVar, Iterable, Any, TypeVar, Callable, Sequence, TypeGuard, Iterator, Generic, Generator, cast, overload
 
 T = TypeVar("T")
 U = TypeVar("U")
@@ -85,6 +85,9 @@ def word_wrap(x, wrap=80):
   while len(ansistrip(x[:i])) < wrap and i < len(x): i += 1
   return x[:i] + "\n" + word_wrap(x[i:], wrap)
 def pad_bytes(b:bytes, align:int) -> bytes: return b + b'\x00' * ((align - (len(b) % align)) % align)
+def panic(e:Exception|None=None):
+  if e is None: raise RuntimeError("PANIC!")
+  raise e
 
 @functools.cache
 def canonicalize_strides(shape:tuple[T, ...], strides:tuple[T, ...]) -> tuple[T, ...]:
@@ -123,8 +126,13 @@ def polyN(x:T, p:list[float]) -> T: return functools.reduce(lambda acc,c: acc*x+
 
 @functools.cache
 def to_function_name(s:str): return ''.join([c if c in (string.ascii_letters+string.digits+'_') else f'{ord(c):02X}' for c in ansistrip(s)])
+@overload
+def getenv(key:str) -> int: ...
+@overload
+def getenv(key:str, default:T) -> T: ...
 @functools.cache
-def getenv(key:str, default=0): return type(default)(os.getenv(key, default))
+def getenv(key:str, default:Any=0): return type(default)(os.getenv(key, default))
+
 def temp(x:str, append_user:bool=False) -> str:
   return (pathlib.Path(tempfile.gettempdir()) / (f"{x}.{getpass.getuser()}" if append_user else x)).as_posix()
 
@@ -157,7 +165,7 @@ TRANSCENDENTAL, NOLOCALS = ContextVar("TRANSCENDENTAL", 1), ContextVar("NOLOCALS
 SPLIT_REDUCEOP, NO_MEMORY_PLANNER, RING = ContextVar("SPLIT_REDUCEOP", 1), ContextVar("NO_MEMORY_PLANNER", 0), ContextVar("RING", 1)
 PICKLE_BUFFERS, LRU = ContextVar("PICKLE_BUFFERS", 1), ContextVar("LRU", 1)
 CACHELEVEL, IGNORE_BEAM_CACHE, DEVECTORIZE = ContextVar("CACHELEVEL", 2), ContextVar("IGNORE_BEAM_CACHE", 0), ContextVar("DEVECTORIZE", 1)
-DISABLE_COMPILER_CACHE, BLOCK_REORDER = ContextVar("DISABLE_COMPILER_CACHE", 0), ContextVar("BLOCK_REORDER", 1)
+DISABLE_COMPILER_CACHE = ContextVar("DISABLE_COMPILER_CACHE", 0)
 QUANTIZE, VALIDATE_WITH_CPU, DISABLE_FAST_IDIV = ContextVar("QUANTIZE", 0), ContextVar("VALIDATE_WITH_CPU", 0), ContextVar("DISABLE_FAST_IDIV", 0)
 CORRECT_DIVMOD_FOLDING, FUSE_OPTIM = ContextVar("CORRECT_DIVMOD_FOLDING", 0), ContextVar("FUSE_OPTIM", 0)
 ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE = ContextVar("ALLOW_DEVICE_USAGE", 1), ContextVar("MAX_BUFFER_SIZE", 0)
@@ -166,9 +174,11 @@ EMULATE = ContextVar("EMULATE", "")
 CPU_COUNT = ContextVar("CPU_COUNT", max(1, len(os.sched_getaffinity(0)) if hasattr(os, "sched_getaffinity") else (os.cpu_count() or 1)))
 CPU_LLVM, CPU_LVP, AMD_LLVM = ContextVar("CPU_LLVM", 0), ContextVar("CPU_LVP", 0), ContextVar("AMD_LLVM", 1)
 VIZ = PROFILE = ContextVar("VIZ", 0)
-SPEC = ContextVar("SPEC", 0)
+SPEC = ContextVar("SPEC", 1)
 # TODO: disable by default due to speed
 IGNORE_OOB = ContextVar("IGNORE_OOB", 1)
+PCONTIG = ContextVar("PCONTIG", 0)  # partial contiguous in rangeify
+DEBUG_RANGEIFY = ContextVar("DEBUG_RANGEIFY", 0)
 
 @dataclass(frozen=True)
 class Metadata:
@@ -176,7 +186,6 @@ class Metadata:
   caller: str
   backward: bool = False
   def __hash__(self): return hash(self.name)
-  def __repr__(self): return str(self) + (f" - {self.caller}" if self.caller else "")
   def __str__(self): return self.name + (" bw" if self.backward else "")
 
 # **************** global state Counters ****************

tinygrad/nn/__init__.py

@@ -36,7 +36,7 @@ class BatchNorm:
     self.weight: Tensor|None = Tensor.ones(sz) if affine else None
     self.bias: Tensor|None = Tensor.zeros(sz) if affine else None
 
-    self.num_batches_tracked = Tensor.zeros(1, dtype='long' if is_dtype_supported(dtypes.long) else 'int', requires_grad=False)
+    self.num_batches_tracked = Tensor.zeros(dtype='long' if is_dtype_supported(dtypes.long) else 'int', requires_grad=False)
     if track_running_stats: self.running_mean, self.running_var = Tensor.zeros(sz, requires_grad=False), Tensor.ones(sz, requires_grad=False)
 
   def calc_stats(self, x:Tensor) -> tuple[Tensor, Tensor]:

tinygrad/nn/onnx.py

@@ -5,7 +5,7 @@ from io import BufferedReader
 from tinygrad.nn.state import TensorIO
 from tinygrad.tensor import Tensor, _broadcast_shape, ReductionStr
 from tinygrad.helpers import getenv, DEBUG, all_same, prod, flatten, make_tuple, argsort, is_numpy_ndarray, get_single_element, polyN
-from tinygrad.dtype import DType, ConstType, dtypes, _from_np_dtype, truncate
+from tinygrad.dtype import DType, ConstType, dtypes, _from_np_dtype, truncate, least_upper_dtype
 from tinygrad.device import is_dtype_supported, Device
 
 # ***** protobuf definitions ******
@@ -670,7 +670,8 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
   def ReduceL1(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
     return ReduceSum(data.abs(), axes, keepdims, noop_with_empty_axes)
   def ReduceL2(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
-    return ReduceSumSquare(data, axes, keepdims, noop_with_empty_axes).sqrt()
+    dtype = dtypes.float if data.dtype in (dtypes.float16, dtypes.bfloat16) else data.dtype
+    return ReduceSum(data.cast(dtype).square(), axes, keepdims, noop_with_empty_axes).sqrt().cast(data.dtype)
   def ReduceLogSum(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
     return ReduceSum(data, axes, keepdims, noop_with_empty_axes).log()
   def ReduceLogSumExp(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
@@ -897,7 +898,7 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
   def BatchNormalization(X:Tensor, scale:Tensor, B:Tensor, input_mean:Tensor, input_var:Tensor, epsilon:float=1e-05, momentum:float=0.9,
                         training_mode:int=0, spatial=1, is_test=0):
     if training_mode:
-      x_detached = X.detach()
+      x_detached = X.detach().cast(least_upper_dtype(X.dtype, dtypes.float32))
       current_mean = x_detached.mean(axis=(0,2,3))
       y = (x_detached - current_mean.reshape(shape=[1, -1, 1, 1]))
       current_var = (y*y).mean(axis=(0,2,3))
@@ -906,18 +907,20 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
       running_mean = input_mean * momentum + current_mean * (1 - momentum)
       running_var = input_var * momentum + current_var * (1 - momentum)
 
-      return X.batchnorm(scale, B, current_mean, current_invstd), running_mean, running_var
+      return X.batchnorm(scale, B, current_mean, current_invstd).cast(X.dtype),running_mean.cast(input_mean.dtype),running_var.cast(input_var.dtype)
     return X.batchnorm(scale, B, input_mean, (input_var + epsilon).rsqrt())
-  def GroupNormalization(x:Tensor, scale:Tensor, bias:Tensor, num_groups:int, epsilon:float=1e-05):
-    x = x.reshape(x.shape[0], num_groups, -1).layernorm(eps=epsilon).reshape(x.shape)
+  def GroupNormalization(x:Tensor, scale:Tensor, bias:Tensor, num_groups:int, epsilon:float=1e-05, stash_type:int=1):
+    assert stash_type == 1, "only float32 is supported"
+    x = x.reshape(x.shape[0], num_groups, -1).cast(dtypes.float).layernorm(eps=epsilon).cast(x.dtype).reshape(x.shape)
     return x * scale.reshape(1, -1, *[1] * (x.ndim-2)) + bias.reshape(1, -1, *[1] * (x.ndim-2))
   def InstanceNormalization(x:Tensor, scale:Tensor, bias:Tensor, epsilon:float=1e-05):
     return GroupNormalization(x, scale, bias, num_groups=cast(int, x.shape[1]), epsilon=epsilon)
   def LayerNormalization(x:Tensor, scale:Tensor, bias:Tensor, axis:int=-1, epsilon:float=1e-05, stash_type:int=1):
     assert stash_type == 1, "only float32 is supported"
     axes = tuple(i for i in range(axis if axis >= 0 else x.ndim + axis, x.ndim))
-    mean = x.mean(axis=axes, keepdim=True)
-    return x.layernorm(axes, epsilon).mul(scale).add(bias), mean, (x.sub(mean)).square().mean(axis=axes, keepdim=True).add(epsilon).rsqrt()
+    mean = (x32:=x.cast(dtypes.float)).mean(axis=axes, keepdim=True)
+    inv_std_dev = (x32.sub(mean)).square().mean(axis=axes, keepdim=True).add(epsilon).rsqrt()
+    return (x32.sub(mean)*inv_std_dev).cast(x.dtype).mul(scale).add(bias), mean, inv_std_dev
   def SkipLayerNormalization(x:Tensor, skip:Tensor, gamma:Tensor, beta:Tensor|None=None, bias:Tensor|None=None, epsilon:float=1e-12):
     x = x + skip
     if bias is not None: x = x + bias
@@ -1089,9 +1092,10 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
     return output, present_key, present_value, qk_matmul_return_val
   Attention = {OpSetId(Domain.ONNX, 1): attention_onnx, OpSetId(Domain.MICROSOFT_CONTRIB_OPS, 1): attention_contrib}
 
-  def RMSNormalization(X:Tensor, scale:Tensor, axis:int=-1, epsilon:float=1e-5):
-    norm = X.square().mean(axis=tuple(range(axis + X.ndim if axis < 0 else axis, X.ndim)), keepdim=True).add(epsilon).rsqrt()
-    return X * norm * scale
+  def RMSNormalization(X:Tensor, scale:Tensor, axis:int=-1, epsilon:float=1e-5, stash_type:int=1):
+    assert stash_type == 1, "only float32 is supported"
+    norm = X.cast(dtypes.float).square().mean(axis=tuple(range(axis + X.ndim if axis < 0 else axis, X.ndim)), keepdim=True).add(epsilon).rsqrt()
+    return X.cast(X.dtype) * norm * scale
 
   def RotaryEmbedding(X:Tensor, cos_cache:Tensor, sin_cache:Tensor, position_ids:Tensor|None=None, interleaved:int=0, num_heads:int|None=None,
                       rotary_embedding_dim:int=0):
@@ -1242,7 +1246,8 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
     G, V, H = G.detach(), V.detach(), H.detach()
     X.grad = norm_coefficient * X.detach() + G
     opt = TinyAdam([X], b1=alpha, b2=beta, eps=epsilon)
-    opt.m, opt.v, opt.lr = [V], [H], R
+    # NOTE: FUSE_OPTIM can change shapes of m and v
+    opt.m, opt.v, opt.lr = [V.reshape(opt.m[0].shape)], [H.reshape(opt.v[0].shape)], R
     # need no-op for m_hat and v_hat if T == 0
     if T == 0: opt.b1_t, opt.b2_t = opt.b1_t.zeros_like(), opt.b2_t.zeros_like()
     else:

tinygrad/nn/optim.py

@@ -80,7 +80,7 @@ def SGD(params: list[Tensor], lr=0.001, momentum=0.0, weight_decay=0.0, nesterov
   return LARS(params, lr, momentum, weight_decay, 0, None, nesterov, classic=classic, pre_wd=True, tcoef=0.0, fused=fused)
 
 # Muon applies the newton schulz algorithm on gradient. also can include momentum, nesterov, and weight decay
-def Muon(params: list[Tensor], lr=0.02, momentum=0.95, weight_decay=0.0, ns_steps=5, ns_params=(3.4445, -4.775, 2.0315),
+def Muon(params: list[Tensor], lr=0.001, momentum=0.95, weight_decay=0.1, ns_steps=5, ns_coefficients=(3.4445, -4.775, 2.0315),
          nesterov=True, fused=FUSE_OPTIM):
   """
   SGD with newton-schulz iteration and post momentum weight decay.
@@ -89,7 +89,7 @@ def Muon(params: list[Tensor], lr=0.02, momentum=0.95, weight_decay=0.0, ns_step
   - Paper: https://arxiv.org/pdf/2502.16982
   """
   assert not fused, "FUSE_OPTIM not allowed for Muon optimizer"
-  return LARS(params, lr, momentum, weight_decay, ns_steps, ns_params, nesterov, classic=False, pre_wd=False, tcoef=0.0, fused=fused)
+  return LARS(params, lr, momentum, weight_decay, ns_steps, ns_coefficients, nesterov, classic=False, pre_wd=False, tcoef=0.0, fused=fused)
 
 class LARS(Optimizer):
   """
@@ -97,10 +97,10 @@ class LARS(Optimizer):
 
   - Paper: https://arxiv.org/abs/1708.03888v3
   """
-  def __init__(self, params:list[Tensor], lr=0.001, momentum=0.9, weight_decay=1e-4, ns_steps=0, ns_params=None,
+  def __init__(self, params:list[Tensor], lr=0.001, momentum=0.9, weight_decay=1e-4, ns_steps=0, ns_coefficients=None,
                nesterov=False, classic=True, pre_wd=True, tcoef=0.001, fused=FUSE_OPTIM):
     super().__init__(params, lr, fused)
-    self.momentum, self.wd, self.ns_steps, self.ns_params  = momentum, weight_decay, ns_steps, ns_params
+    self.momentum, self.wd, self.ns_steps, self.ns_coefficients  = momentum, weight_decay, ns_steps, ns_coefficients
     self.nesterov, self.classic, self.pre_wd, self.tcoef = nesterov, classic, pre_wd, tcoef
     self.b = self._new_optim_param() if self.momentum else []
 
@@ -116,11 +116,9 @@ class LARS(Optimizer):
       # classic momentum does post learning rate update
       if self.classic: g = g * r * self.lr
       if self.momentum:
-        # TODO: this contiguous is required for correctness because self.b[i] becomes a non contiguous view
-        # the scheduler should detect this and just insert contiguous
-        self.b[i].assign(self.momentum * self.b[i].contiguous() + g)  # NOTE: self.b[i] is zero on the first run, no if required
+        self.b[i].assign(self.momentum * self.b[i] + g)  # NOTE: self.b[i] is zero on the first run, no if required
         g = (g + self.momentum * self.b[i]) if self.nesterov else self.b[i]
-      if self.ns_params: g = g.reshape(g.shape[0], -1).newton_schulz(self.ns_steps, self.ns_params).reshape(g.shape)
+      if self.ns_coefficients: g = g.reshape(g.shape[0], -1).newton_schulz(self.ns_steps, self.ns_coefficients).reshape(g.shape)
       # muon does post momentum weight decay
       if not self.pre_wd and self.wd > 0: t = t.detach() * (1.0 - self.wd * self.lr)
       # popular momentum does pre learning rate update

tinygrad/renderer/__init__.py

@@ -28,12 +28,17 @@ class Estimates:
     mult_stack: list[sint] = []
     dont_count: set[UOp] = set()
     if ignore_indexing:
+      def range_gate(x): return x.op is not Ops.RANGE
       for u in uops:
         if u.op in {Ops.LOAD, Ops.STORE} and (not isinstance(u.src[0].dtype, PtrDType) or u.src[0].dtype.addrspace != AddrSpace.REG):
-          dont_count = dont_count.union(u.src[0].toposort())
+          # if u.src[0] is INDEX, we have to include the buffer since it might be an AFTER
+          dont_count = dont_count.union((UOp.sink(*u.src[0].src[1:]) if u.src[0].op is Ops.INDEX else u.src[0]).toposort(range_gate))
+          # TODO: is this correct? this all needs to be cleaned up
           if len(u.src) > 2: dont_count = dont_count.union(u.src[2].toposort())
         elif u.op is Ops.IF:
           dont_count = dont_count.union(u.src[0].toposort())
+        elif u.op is Ops.RANGE:
+          dont_count = dont_count.union(u.src[0].toposort())
     for u in uops:
       if u.op in {Ops.LOAD, Ops.STORE}:
         buf = u
@@ -42,18 +47,18 @@ class Estimates:
           mem[(buf, u.op)] = buf.ptrdtype.size * buf.dtype.itemsize
       if u.op is Ops.RANGE:
         mult_stack.append(mults)
-        mults *= cast(sint, u.src[0].ssimplify())
+        mults = cast(sint, (mults*u.src[0]).ssimplify())
         # SPECIAL are already counted in mults
         mults = mults.substitute({x:x.const_like(0) for x in mults.toposort() if x.op is Ops.SPECIAL}) if isinstance(mults, UOp) else mults
-      elif u.op is Ops.ENDRANGE: mults = mult_stack.pop(-1)
-      elif u.op is Ops.SPECIAL: mults *= cast(sint, u.src[0].ssimplify()) # NOTE: we don't push to the mult_stack here, you can't end these
+      elif u.op is Ops.END: mults = mult_stack.pop(-1)
+      elif u.op is Ops.SPECIAL: mults = cast(sint, (mults*u.src[0]).ssimplify()) # NOTE: we don't push to the mult_stack here, you can't end these
       elif u.op is Ops.LOAD and (not isinstance(u.src[0].dtype, PtrDType) or u.src[0].dtype.addrspace != AddrSpace.REG):
         lds += u.dtype.itemsize * mults
       elif u.op is Ops.STORE and (not isinstance(u.src[0].dtype, PtrDType) or u.src[0].dtype.addrspace != AddrSpace.REG):
         lds += u.src[1].dtype.itemsize * mults
       elif u.op in GroupOp.ALU and u not in dont_count: flops += (mults * (2 if u.op is Ops.MULACC else 1)) * u.dtype.count
       elif u.op is Ops.WMMA and u not in dont_count: flops += 2 * prod(u.arg[1]) // u.arg[5] * mults
-    return Estimates(flops, lds, sum(mem.values()))
+    return Estimates(ssimplify(flops), ssimplify(lds), sum(mem.values()))
 
 @dataclass
 class ProgramSpec:
@@ -78,8 +83,12 @@ class ProgramSpec:
       for u in self.uops:
         if u.op is Ops.DEFINE_VAR: self.vars.append(u)
         if u.op is Ops.DEFINE_GLOBAL: self.globals.append(u.arg)
-        if u.op is Ops.STORE: self.outs.extend([x.arg for x in u.src[0].toposort() if x.op is Ops.DEFINE_GLOBAL])
-        if u.op is Ops.LOAD: self.ins.extend([x.arg for x in u.src[0].toposort() if x.op is Ops.DEFINE_GLOBAL])
+        if u.op is Ops.STORE and (u.src[0].op is Ops.INDEX or (u.src[0].op is Ops.CAST and u.src[0].src[0].op is Ops.INDEX)):
+          idx = u.src[0] if u.src[0].op is Ops.INDEX else u.src[0].src[0]
+          if (buf:=idx.src[0]).op is Ops.DEFINE_GLOBAL: self.outs.append(buf.arg)
+        if u.op is Ops.LOAD and (u.src[0].op is Ops.INDEX or (u.src[0].op is Ops.CAST and u.src[0].src[0].op is Ops.INDEX)):
+          idx = u.src[0] if u.src[0].op is Ops.INDEX else u.src[0].src[0]
+          if (buf:=idx.src[0]).op is Ops.DEFINE_GLOBAL: self.ins.append(buf.arg)
         if u.op is Ops.SPECIAL:
           # NOTE: you have to set local_size and global_size to the base [1,1,1] outside this
           if u.arg[0] == 'i': self.local_size = None

tinygrad/renderer/cstyle.py

@@ -2,7 +2,7 @@ from typing import Literal, Callable, cast
 import os, math, sys
 from collections import defaultdict, Counter
 from tinygrad.codegen.opt import tc
-from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat, range_str
+from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat, range_str, axis_letters
 from tinygrad.helpers import strip_parens, getenv, prod, dedup, AMX, CPU_COUNT
 from tinygrad.dtype import ImageDType, dtypes, DType, PtrDType, AddrSpace, truncate
 from tinygrad.renderer import Renderer
@@ -11,7 +11,7 @@ from tinygrad.codegen.late.devectorizer import no_vectorized_alu
 base_rewrite = PatternMatcher([
   (UPat(Ops.DEFINE_REG, name="x"), lambda ctx,x: f"{ctx.render_dtype(x.dtype.base)} {ctx[x]}[{x.dtype.size}];"),
   (UPat(Ops.IF, name="x"), lambda ctx,x: f"if ({ctx[x.src[0]]}) {{"),
-  (UPat((Ops.ENDIF, Ops.ENDRANGE)), lambda ctx: "}"),
+  (UPat((Ops.ENDIF, Ops.END)), lambda ctx: "}"),
   (UPat(Ops.WMMA, name="x"), lambda ctx,x: f"__{x.arg[0]}({ctx[x.src[0]]}, {ctx[x.src[1]]}, {ctx[x.src[2]]})"),
   # r method accesses
   (UPat(Ops.RANGE, name="x"),
@@ -37,7 +37,7 @@ base_rewrite = PatternMatcher([
   (UPat(Ops.CONST, dtype=dtypes.uint32, name="x"), lambda ctx,x: f"{truncate[x.dtype](x.arg)}u"),
   (UPat(Ops.CONST, dtype=dtypes.bool, name="x"), lambda ctx,x: "1" if x.arg else "0"),
   # consts are rendered to larger type and casted
-  (UPat(Ops.CONST, (dtypes.bfloat16, dtypes.half), name="x"), lambda ctx,x: f"({ctx.render_cast(x.dtype, f'{x.arg}f')})"),
+  (UPat(Ops.CONST, (*dtypes.fp8s, dtypes.bfloat16, dtypes.half), name="x"), lambda ctx,x: f"({ctx.render_cast(x.dtype, f'{x.arg}f')})"),
   (UPat(Ops.CONST, (dtypes.uint8, dtypes.uint16), name="x"), lambda ctx,x: f"({ctx.render_cast(x.dtype, f'{x.arg}u')})"),
   (UPat(Ops.CONST, (dtypes.int8, dtypes.int16), name="x"), lambda ctx,x: f"({ctx.render_cast(x.dtype, str(x.arg))})"),
   # default const render
@@ -95,7 +95,7 @@ class CStyleLanguage(Renderer):
   infinity: str = "INFINITY"
   nan: str = "NAN"
   code_for_op: dict = {
-    Ops.SQRT: lambda x,dtype: f"sqrt({x})", Ops.RECIP: lambda x,dtype: f"(1/{x})", Ops.NEG: lambda x,dtype: f"-{x}",
+    Ops.SQRT: lambda x,dtype: f"sqrt({x})", Ops.RECIPROCAL: lambda x,dtype: f"(1/{x})", Ops.NEG: lambda x,dtype: f"-{x}",
     Ops.EXP2: lambda x,dtype: f"exp2({x})", Ops.LOG2: lambda x,dtype: f"log2({x})", Ops.SIN: lambda x,dtype: f"sin({x})",
     Ops.TRUNC: lambda x,dtype: f"trunc({x})",
     Ops.AND: lambda a,b,dtype: f"({a}&{b})", Ops.XOR: lambda a,b,dtype: f"({a}^{b})", Ops.OR: lambda a,b,dtype: f"({a}|{b})",
@@ -143,7 +143,10 @@ class CStyleLanguage(Renderer):
     c: defaultdict[str, int] = defaultdict(int)
     name = "test"
     for u in uops:
-      if u.op is Ops.NOOP: continue
+      if u.op in {Ops.NOOP, Ops.GROUP}: continue
+      if u.op is Ops.AFTER:
+        r[u] = r[u.src[0]]
+        continue
       if u.op is Ops.SINK:
         if u.arg is not None: name = u.arg.function_name
         continue
@@ -154,13 +157,14 @@ class CStyleLanguage(Renderer):
 
       # mark buffers that we store to writable
       if u.op is Ops.STORE:
-        for up in u.src[0].toposort():
+        # NOTE: we gate on RANGE to not follow it back
+        for up in u.src[0].toposort(lambda x: x.op is not Ops.RANGE):
           if up.op is Ops.DEFINE_GLOBAL: bufs[up] = (bufs[up][0], (bufs[up][1][0], True))
 
       # naming
       prefix = None
       if u.op is Ops.SPECIAL: r[u] = u.arg
-      elif u.op is Ops.RANGE: r[u] = "ridx"+range_str(u)
+      elif u.op is Ops.RANGE: r[u] = f"{axis_letters[u.arg[-1]]}idx"+range_str(u)
       else:
         prefix = {Ops.WMMA: "wmma", Ops.DEFINE_LOCAL: "temp", Ops.CONST: "const",
                   Ops.CAST: "cast", Ops.BITCAST: "cast", Ops.GEP: "gep", Ops.VECTORIZE: "cast", Ops.PRECAST: "precast",
@@ -170,7 +174,7 @@ class CStyleLanguage(Renderer):
       l = cast(str, self.string_rewrite.rewrite(u, ctx=self))
       assert l is not None, f"failed to render {u.op} {u.dtype} {[(x.op,x.dtype) for x in u.src]} {u.arg}"
 
-      if u.op in {Ops.ENDIF, Ops.ENDRANGE}: depth -= 1
+      if u.op in {Ops.ENDIF, Ops.END}: depth -= 1
       if (u.op is not Ops.CAST or u.dtype.vcount == 1) and (u.op in {Ops.CONST, Ops.GEP, Ops.INDEX, Ops.CUSTOMI} or \
         (u.op is Ops.LOAD and u.src[0].ptrdtype.addrspace == AddrSpace.REG) or \
         (u.op is Ops.CAST and isinstance(u.dtype, PtrDType)) or \
@@ -205,7 +209,7 @@ class ClangRenderer(CStyleLanguage):
   # language options
   buffer_suffix = " restrict"
   type_map = {dtypes.bool:"_Bool", dtypes.half:"__fp16"}
-  code_for_op = {**({k:v for k,v in CStyleLanguage.code_for_op.items() if k not in [Ops.EXP2, Ops.SIN, Ops.LOG2, Ops.TRUNC, Ops.RECIP]}),
+  code_for_op = {**({k:v for k,v in CStyleLanguage.code_for_op.items() if k not in [Ops.EXP2, Ops.SIN, Ops.LOG2, Ops.TRUNC, Ops.RECIPROCAL]}),
                  Ops.SQRT: lambda x,dtype: f"__builtin_sqrt({x})" if dtype == dtypes.float64 else f"__builtin_sqrtf({x})",
                  Ops.TRUNC: lambda x,dtype: f"__builtin_trunc({x})" if dtype == dtypes.float64 else f"__builtin_truncf({x})",
                  Ops.FDIV: lambda a,b,dtype: f"({a}/{b})"}
@@ -266,7 +270,8 @@ class OpenCLRenderer(CStyleLanguage):
       lambda ctx,buf,idx,var,gate: f"({ctx[gate]}?read_imagef({ctx[buf]}, smp, {ctx[idx]}):{ctx[var]})"),
     (UPat(Ops.LOAD, dtype=dtypes.float.vec(4), src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2))),)),
       lambda ctx,buf,idx: f"read_imagef({ctx[buf]}, smp, {ctx[idx]})"),
-    (UPat(Ops.STORE, src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2))), UPat.var("var", dtypes.float.vec(4))), allow_any_len=True),
+    (UPat(Ops.STORE, src=(UPat.var('buf').index(UPat.var('idx', dtypes.int.vec(2)), allow_any_len=True),
+                          UPat.var("var", dtypes.float.vec(4))), allow_any_len=True),
       lambda ctx,buf,idx,var: f"write_imagef({ctx[buf]}, {ctx[idx]}, {ctx[var]});"),
   ]) + base_rewrite
 
@@ -342,7 +347,8 @@ class CUDARenderer(CStyleLanguage):
   shared_max = 49152
 
   def __init__(self, arch:str):
-    self.tensor_cores, self.arch = tc.cuda_sm80 if int(arch[3:]) >= 80 else tc.cuda_sm75 if int(arch[3:]) >= 75 else [], arch
+    self.arch = arch
+    self.tensor_cores = tc.cuda_sm89 if int(arch[3:]) >= 89 else tc.cuda_sm80 if int(arch[3:]) >= 80 else tc.cuda_sm75 if int(arch[3:]) >= 75 else []
   def __reduce__(self): return self.__class__, (self.arch,)
 
   # language options
@@ -360,9 +366,15 @@ class CUDARenderer(CStyleLanguage):
     Ops.LOG2: lambda x,dtype: f"hlog2({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"log2({x})",
     Ops.EXP2: lambda x,dtype: f"hexp2({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"exp2({x})",
     Ops.SQRT: lambda x,dtype: f"hsqrt({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"sqrt({x})",
-    Ops.RECIP: lambda x,dtype: f"hrcp({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"(1/{x})" }
-  type_map = {dtypes.bfloat16: "nv_bfloat16"}
-
+    Ops.RECIPROCAL: lambda x,dtype: f"hrcp({x})" if dtype in (dtypes.half, dtypes.bfloat16) else f"(1/{x})" }
+  type_map = {dtypes.bfloat16: "nv_bfloat16", dtypes.fp8e4m3: "__nv_fp8_e4m3", dtypes.fp8e5m2: "__nv_fp8_e5m2"}
+  extra_matcher = PatternMatcher([
+    (UPat(Ops.CAST, dtypes.fp8s, UPat.var("x", dtypes.fp8s), name='y'), lambda x,y: x.cast(dtypes.float).cast(y.dtype) if x.dtype!=y.dtype else None),
+    (UPat(GroupOp.ALU, dtype=dtypes.fp8s, name="x"),
+     lambda x: UOp(x.op, dtypes.float, tuple(vv.cast(dtypes.float) for vv in x.src), x.arg).cast(x.dtype)),
+    (UPat(GroupOp.ALU, dtypes.bool, name="alu", src=(UPat.var("x", dtype=dtypes.fp8s), UPat.var("y", dtype=dtypes.fp8s))),
+     lambda alu,x,y: UOp(alu.op, dtypes.bool, (x.cast(dtypes.float), y.cast(dtypes.float)), alu.arg)),
+  ]) + extra_pm
   def render_vector_prefix(self, dt:DType) -> str:
     vec, scal = self.render_dtype(dt), self.render_dtype(dt.scalar()),
     elems, header = ', '.join(_nms[:dt.count]), ', '.join([f"{scal} {x}" for x in _nms[:dt.count]])
@@ -373,11 +385,12 @@ class CUDARenderer(CStyleLanguage):
     prefix = ["#define INFINITY (__int_as_float(0x7f800000))","#define NAN (__int_as_float(0x7fffffff))"]
 
     used_dtypes = uops_to_dtypes(uops)
+    if any(dt.scalar() in dtypes.fp8s for dt in used_dtypes): prefix.append("#include <cuda_fp8.h>")
     if any(dt.scalar() == dtypes.half for dt in used_dtypes): prefix.append("#include <cuda_fp16.h>")
     if any(dt.scalar() == dtypes.bfloat16 for dt in used_dtypes): prefix.append("#include <cuda_bf16.h>")
-    prefix += [self.render_vector_prefix(dt) for dt in used_dtypes if dt.count in (4,8) and dt.scalar() in {dtypes.half, dtypes.bfloat16}]
-
-    dt_map_in = { dtypes.float: "tf32", dtypes.half: "f16", dtypes.bfloat16: "bf16" }
+    prefix += [self.render_vector_prefix(dt) for dt in used_dtypes if (dt.count in (4,8) and dt.scalar() in {dtypes.half, dtypes.bfloat16})
+      or (dt.count in (8,16) and dt.scalar() in dtypes.fp8s)]
+    dt_map_in = { dtypes.float: "tf32", dtypes.half: "f16", dtypes.bfloat16: "bf16", dtypes.fp8e4m3: "e4m3", dtypes.fp8e5m2: "e5m2" }
     dt_map_out = { dtypes.float: "f32", dtypes.half: "f16" }
     for name, (N, M, K), dtype_in, dtype_out, _, _, upcast_axes, _ in wmma_args(uops):
       upcast_sizes = [prod(size for _, size in upcast) for upcast in upcast_axes]
@@ -411,7 +424,7 @@ class AMDRenderer(CStyleLanguage):
 
   @staticmethod
   def get_tensor_cores(arch):
-    return {"gfx942": tc.amd_cdna, "gfx950": tc.amd_cdna, "gfx1200": tc.amd_rdna4, "gfx1201": tc.amd_rdna4}.get(arch.split(":")[0], tc.amd_rdna3)
+    return {"gfx942": tc.amd_cdna, "gfx950": tc.amd_cdna4, "gfx1200": tc.amd_rdna4, "gfx1201": tc.amd_rdna4}.get(arch.split(":")[0], tc.amd_rdna3)
   def __init__(self, arch:str): # gfx942 => MI300, gfx1100 => RX 7900, gfx1201 => RX 9700
     self.arch = arch
     self.tensor_cores = self.get_tensor_cores(arch)

tinygrad/renderer/llvmir.py

@@ -49,9 +49,11 @@ def render_wmma_amx(ctx, wmma: UOp) -> str:
 def render_wmma_amd(ctx, wmma: UOp, cdna=False) -> str:
   dt_map = {dtypes.half: "f16", dtypes.float: "f32", dtypes.ushort: "bf16.1k" if cdna else "bf16", dtypes.bfloat16: "bf16.1k" if cdna else "bf16"}
   # https://github.com/llvm/llvm-project/blob/main/clang/test/CodeGenOpenCL/builtins-amdgcn-mfma.cl
+  N,M,K = wmma.arg[1]
   if cdna:
+    if K == 32: dt_map.update({dtypes.half: ".f16", dtypes.bfloat16: ".bf16"})
     return f"  {ctx[wmma]} = call {ldt(wmma.dtype)} @llvm.amdgcn.mfma.{dt_map[wmma.src[-1].dtype.scalar()]}" + \
-           f".16x16x16{dt_map[wmma.src[0].dtype.scalar()]}(" + ", ".join([f"{ldt(w.dtype)} {ctx[w]}" for w in wmma.src]) + ", i32 0, i32 0, i32 0)"
+           f".{N}x{M}x{K}{dt_map[wmma.src[0].dtype.scalar()]}(" + ", ".join([f"{ldt(w.dtype)} {ctx[w]}" for w in wmma.src]) + ", i32 0, i32 0, i32 0)"
   # https://github.com/llvm/llvm-project/blob/main/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.wmma_32.ll
   # example: %wmma0 = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.f16(<16 x half> %v99,<16 x half> %v100,<8 x float> %v101)
   return f"  {ctx[wmma]} = call {ldt(wmma.dtype)} @llvm.amdgcn.wmma.{dt_map[wmma.src[-1].dtype.scalar()]}.16x16x16." + \
@@ -104,15 +106,21 @@ base_rewrite = PatternMatcher([
    f"  {ctx[x]} = select {ldt(x.src[0].dtype)} {ctx[x.src[0]]}, {ldt(x.src[1].dtype)} {ctx[x.src[1]]}, {ldt(x.src[2].dtype)} {ctx[x.src[2]]}"),
 
   # range
-  (UPat(Ops.RANGE, name="x"), lambda ctx,x:
-   f"  br label %loop_entry_{range_str(x)}\nloop_entry_{range_str(x)}:\n"
-   f"  br label %loop_body_{range_str(x)}\nloop_body_{range_str(x)}:\n"
-   f"  {ctx[x]} = phi {ldt(x.dtype)} [ 0, %loop_entry_{range_str(x)} ], [ {ctx[x]}phi, %loop_latch_{range_str(x)} ]"),
-  (UPat(Ops.ENDRANGE, name="x"), lambda ctx,x:
-   f"  br label %loop_latch_{range_str(x.src[0])}\nloop_latch_{range_str(x.src[0])}:\n"
-   f"  {ctx[x.src[0]]}phi = add {ldt(x.src[0].dtype)} {ctx[x.src[0]]}, 1\n"
-   f"  {ctx[x]} = icmp ult {ldt(x.src[0].dtype)} {ctx[x.src[0]]}phi, {ctx[x.src[0].src[0]]}\n"
-   f"  br i1 {ctx[x]}, label %loop_body_{range_str(x.src[0])}, label %loop_exit_{range_str(x.src[0])}\nloop_exit_{range_str(x.src[0])}:"),
+  (UPat(Ops.RANGE, name="r"), lambda ctx,r:
+   f"  br label %loop_entry_{range_str(r)}\n"
+   f"loop_entry_{range_str(r)}:\n"
+   f"  br label %loop_latch_{range_str(r)}\n"
+   f"loop_latch_{range_str(r)}:\n"
+   f"  {ctx[r]} = phi {ldt(r.dtype)} [ 0, %loop_entry_{range_str(r)} ], [ {ctx[r]}phi, %loop_footer_{range_str(r)} ]\n"
+   f"  {ctx[r]}phi = add {ldt(r.dtype)} {ctx[r]}, 1\n"
+   f"  {ctx[r]}cmp = icmp ult {ldt(r.dtype)} {ctx[r]}, {ctx[r.src[0]]}\n"
+   f"  br i1 {ctx[r]}cmp, label %loop_body_{range_str(r)}, label %loop_exit_{range_str(r)}\n"
+   f"loop_body_{range_str(r)}:"),
+  (UPat(Ops.END, src=(UPat(), UPat(Ops.RANGE, name="r"))), lambda r:
+   f"  br label %loop_footer_{range_str(r)}\n"
+   f"loop_footer_{range_str(r)}:\n"
+   f"  br label %loop_latch_{range_str(r)}\n"
+   f"loop_exit_{range_str(r)}:"),
 
   # if
   (UPat(Ops.IF, name="x"), lambda ctx,x: f"  br i1 {ctx[x.src[0]]}, label %ifbody_{ctx[x][1:]}, label %ifskip_{ctx[x][1:]}\nifbody_{ctx[x][1:]}:"),
@@ -166,7 +174,10 @@ class LLVMRenderer(Renderer):
 
     name = "test"
     for u in uops:
-      if u.op is Ops.NOOP: continue
+      if u.op in {Ops.NOOP, Ops.GROUP}: continue
+      if u.op is Ops.AFTER:
+        r[u] = r[u.src[0]]
+        continue
       if u.op is Ops.SINK:
         if u.arg is not None: name = u.arg.function_name
         continue

tinygrad/renderer/nir.py

@@ -1,9 +1,9 @@
-from typing import Callable, cast
+from typing import Callable, cast, Any
 from tinygrad.dtype import AddrSpace, DType, PtrDType, dtypes
 from tinygrad.helpers import DEBUG, OSX, unwrap
 from tinygrad.renderer import Renderer
 from tinygrad.renderer.cstyle import CUDARenderer
-from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat
+from tinygrad.uop.ops import GroupOp, Ops, UOp, PatternMatcher, UPat, range_str
 import tinygrad.runtime.autogen.mesa as mesa
 import base64, ctypes, ctypes.util, struct, functools, inspect
 
@@ -21,7 +21,7 @@ def glsl_type(t:DType) -> mesa.struct_glsl_type:
 u_aop = { Ops.ADD: "iadd", Ops.MUL: "imul", Ops.IDIV: "udiv", Ops.MOD: "umod", Ops.CMPLT: "ult", Ops.CMPNE: "ine", Ops.CMPEQ: "ieq", Ops.OR: "ior",
           Ops.AND: "iand", Ops.XOR: "ixor", Ops.WHERE: "bcsel", Ops.MAX: "umax"}
 s_aop = {**u_aop, Ops.CMPLT: "ilt", Ops.IDIV: "idiv", Ops.MOD: "irem", Ops.MAX: "imax"}
-f_aop = { Ops.ADD: "fadd", Ops.MUL: "fmul", Ops.CMPLT: "flt", Ops.CMPNE: "fneu", Ops.CMPEQ: "feq", Ops.FDIV: "fdiv", Ops.RECIP: "frcp",
+f_aop = { Ops.ADD: "fadd", Ops.MUL: "fmul", Ops.CMPLT: "flt", Ops.CMPNE: "fneu", Ops.CMPEQ: "feq", Ops.FDIV: "fdiv", Ops.RECIPROCAL: "frcp",
           Ops.MAX: "fmax", Ops.TRUNC: "ftrunc", Ops.SIN: "fsin", Ops.EXP2: "fexp2", Ops.LOG2: "flog2"}
 aop = {**{x:u_aop for x in (dtypes.bool,)+dtypes.uints}, **{x:s_aop for x in dtypes.sints}, **{x:f_aop for x in dtypes.floats}}
 
@@ -169,23 +169,30 @@ class NIRRenderer(Renderer):
   def render(self, uops:list[UOp]):
     self.prerender(uops)
     for u in [u for u in uops if u.op is Ops.SPECIAL and u.arg[0] == "l"]: self.b.shader.contents.info.workgroup_size[int(u.arg[-1])] = u.src[0].arg
-    self.r, self.param_idx, ranges = {}, 0, []
+    self.r: dict[UOp, Any] = {}
+    self.param_idx, ranges = 0, []
 
     for u in uops:
-      if u.op == Ops.NOOP or u.op == Ops.INDEX: pass
+      if u.op in {Ops.NOOP, Ops.GROUP, Ops.INDEX}: pass
+      elif u.op is Ops.AFTER:
+        self.r[u] = self.r[u.src[0]]
       elif u.op == Ops.SINK:
         if u.arg is not None: self.b.shader.contents.info.name = mesa.char_pointer_cast(u.arg.function_name)
       elif u.op == Ops.DEFINE_LOCAL:
         self.r[u] = nimm(self.b, self.b.shader.contents.info.shared_size, dtypes.long)
         self.b.shader.contents.info.shared_size += u.dtype.nbytes()
       elif u.op == Ops.RANGE:
-        ranges.append(i:=deref_var(self.b, mesa.nir_local_variable_create(self.b.impl, glsl_type(u.dtype), f"idx{u.arg[0]}".encode()).contents))
+        ranges.append(i:=deref_var(self.b, mesa.nir_local_variable_create(self.b.impl, glsl_type(u.dtype), f"idx{range_str(u)}".encode()).contents))
         nstore(self.b, AddrSpace.REG, i, nimm(self.b, 0, u.dtype), u.dtype)
         mesa.nir_push_loop(self.b)
         self.r[u] = nload(self.b, AddrSpace.REG, i, u.dtype)
-      elif u.op == Ops.ENDRANGE:
-        nif(self.b, nalu(self.b, "ilt", x:=nalu(self.b, "iadd", self.r[u.src[0]], nimm(self.b, 1, u.src[0].dtype)), self.r[u.src[0].src[0]]),
-            functools.partial(nstore, self.b, AddrSpace.REG, ranges.pop(), x, u.src[0].dtype), lambda: njump(self.b, mesa.nir_jump_break))
+        nif(self.b, nalu(self.b, "ilt", self.r[u], self.r[u.src[0]]), lambda: None, lambda: njump(self.b, mesa.nir_jump_break))
+      elif u.op == Ops.END:
+        r = u.src[1]
+        next_i = nalu(self.b, "iadd", self.r[r], nimm(self.b, 1, r.dtype))
+        # TODO: this nif should be removable ... but TestMultiTensor.test_double_matmul_shard_W_0 segfaults with it gone
+        nif(self.b, nalu(self.b, "ilt", next_i, self.r[r.src[0]]), lambda: None, lambda: njump(self.b, mesa.nir_jump_break))
+        nstore(self.b, AddrSpace.REG, ranges.pop(), next_i, r.dtype),
         mesa.nir_pop_loop(self.b, None)
       else:
         if (d:=self.def_rewrite.rewrite(u, ctx=self)) is None: raise RuntimeError(f"failed to render {u.op} srcs {[x.dtype for x in u.src]}")

tinygrad/renderer/ptx.py

@@ -16,7 +16,7 @@ def render_val(x, dtype):
   return str(int(x)) + ("U" if dtypes.is_unsigned(dtype) else "")
 
 asm_for_op: dict[Ops, Callable] = {
-  Ops.RECIP: lambda d,a,dt,name: f"rcp{'.approx' if dtypes.is_float(dt) else ''}.{name} {d}, {a};",
+  Ops.RECIPROCAL: lambda d,a,dt,name: f"rcp{'.approx' if dtypes.is_float(dt) else ''}.{name} {d}, {a};",
   Ops.EXP2: lambda d,a,dt,name: f"ex2.approx.{name} {d}, {a};", Ops.LOG2: lambda d,a,dt,name: f"lg2.approx.{name} {d}, {a};",
   Ops.SIN: lambda d,a,dt,name: f"sin.approx.{name} {d}, {a};", Ops.SQRT: lambda d,a,dt,name: f"sqrt.approx.{name} {d}, {a};",
   Ops.TRUNC: lambda d,a,dt,name: f"cvt.rzi.{name}.{name} {d}, {a};",
@@ -49,21 +49,23 @@ ptx_matcher = PatternMatcher([
    lambda x: UOp(x.op, dtypes.uint8, x.src[0:1] + ((x.src[1].cast(dtypes.uint8),) if len(x.src) >= 2 else ()) + x.src[2:]).cast(dtypes.bool)),
   (UPat(Ops.STORE, src=(UPat(dtype=dtypes.int64), UPat(dtype=dtypes.bool)), name="x", allow_any_len=True),
    lambda x: UOp(x.op, dtypes.void, x.src[0:1] + (x.src[1].cast(dtypes.uint8),) + x.src[2:])),
+  # indexing on PTX is in uint64, we do the math while it's still in the graph
+  (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("idx")), name="op", allow_any_len=True), lambda buf,idx,op:
+    UOp(Ops.INDEX, dtype=dtypes.int64, src=(buf, buf.cast(dtypes.int64)+idx.cast(dtypes.int64)*buf.dtype.itemsize)+op.src[2:]) \
+      if op.dtype != dtypes.int64 and buf.dtype.addrspace != AddrSpace.REG else None),
   # load/store use pointer arithmetic, and the cast does nothing
-  (UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("idx"))),
-   lambda buf,idx: (buf.cast(dtypes.int64) + idx.cast(dtypes.int64)*buf.dtype.itemsize) if buf.dtype.addrspace != AddrSpace.REG else None),
   (UPat(Ops.CAST, name="x"), lambda x: x.src[0] if isinstance(x.dtype, PtrDType) or x.src[0].dtype == dtypes.void else None),
-  # move mask from INDEX to the load/store to enable pointer arithmetic
-  (UPat(Ops.LOAD, src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("idx"), UPat.var("gate"))), UPat.var("alt")), allow_any_len=True, name="l"),
-   lambda buf,idx,gate,alt,l: UOp(Ops.LOAD, alt.dtype, (buf.index(idx), alt, gate, *l.src[2:]))),
-  (UPat(Ops.STORE, src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("idx"), UPat())), UPat.var("val"), UPat.var("gate")), allow_any_len=True),
-   lambda buf,idx,val,gate: UOp.store(buf.index(idx), val, gate)),
   # ptx shr and shl instructions require y to be uint
   (UPat.var("x") << UPat.var("y"), lambda x,y: UOp(Ops.SHL, x.dtype, (x,y.cast(dtypes.uint))) if y.dtype != dtypes.uint else None),
   (UPat.var("x") >> UPat.var("y"), lambda x,y: UOp(Ops.SHR, x.dtype, (x,y.cast(dtypes.uint))) if y.dtype != dtypes.uint else None),
 ])
 
-def mem_type(x: UOp): return 'shared' if any(_x.op is Ops.DEFINE_LOCAL for _x in x.src[0].toposort()) else 'global'
+def mem_type(x:UOp) -> str:
+  match x.op:
+    case Ops.AFTER: return mem_type(x.src[0])
+    case Ops.DEFINE_LOCAL: return 'shared'
+    case Ops.DEFINE_GLOBAL: return 'global'
+    case _: raise RuntimeError(f"{x.op} needs to be memory")
 
 def render_wmma(ctx: "PTXRenderer", wmma: UOp):
   assert ctx.wmma_r, "registry values for wmma must be populated"
@@ -88,9 +90,6 @@ def modifier(a: DType, b: DType): return '.rzi' if dtypes.is_int(a) and dtypes.i
 string_rewrite = PatternMatcher([
   (UPat.cvar("x", dtypes.bool), lambda ctx, x: f"setp.ne.s16 {ctx.r[x]}, {render_val(x.arg, x.dtype)}, 0;"),
   (UPat.cvar("x"), lambda ctx, x: f"mov.b{ctx.types[x.dtype][1:]} {ctx.r[x]}, {render_val(x.arg, x.dtype)};"),
-  (UPat(Ops.STORE, name="x", src=(UPat.var('bidx'), UPat.var("var")), allow_any_len=True), lambda ctx, x, bidx, var: f"st.{mem_type(bidx)}" + \
-    f"{f'.v{cnt}' if ((cnt:=var.dtype.count)>1) else ''}.{ctx.mem_types[var.dtype.scalar()]} " + \
-    f"[{ctx.r[bidx]}+0], {('{' + ', '.join(ctx.r[var]) + '}') if var.dtype.count > 1 else ctx.r[var]};"),
   (UPat(Ops.SPECIAL, name="x"), lambda ctx,x: f"mov.u32 %{x.arg}, %{'ctaid' if x.arg[0] == 'g' else 'tid'}.{chr(120+int(x.arg[-1]))};"),
   (UPat(Ops.DEFINE_GLOBAL, name="x"), lambda ctx, x: f"ld.param.{ctx.types[dtypes.ulong]} {ctx.r[x]}, [data{x.arg}+0];"),
   (UPat((Ops.CMPLT, Ops.CMPNE, Ops.CMPEQ), name="x", allow_any_len=True, src=(UPat.var("src0"),)),
@@ -103,22 +102,32 @@ string_rewrite = PatternMatcher([
    lambda ctx, x, a: f"setp.ne.b{ctx.types[a.dtype][1:]} {ctx.r[x]}, {ctx.r[a]}, {render_val(0, a.dtype)};"),
   (UPat(Ops.CAST, name="x", src=(UPat.var("a"),)),
    lambda ctx, x, a: f"cvt{modifier(x.dtype, a.dtype)}.{ctx.cast_types[x.dtype]}.{ctx.cast_types[a.dtype]} {ctx.r[x]}, {ctx.r[a]};"),
-  (UPat(Ops.LOAD, name="x", src=(UPat.var('loc'), UPat(name='alt'), UPat(name="gate", op=GroupOp.ALU)), allow_any_len=True),
-    lambda ctx, x, loc, alt, gate: flatten([
+  # store / gated load / load
+  (UPat(Ops.STORE, src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("loc")), allow_any_len=True), UPat.var("var"))),
+   lambda ctx, loc, var, buf: f"st.{mem_type(buf)}" + \
+    f"{f'.v{cnt}' if ((cnt:=var.dtype.count)>1) else ''}.{ctx.mem_types[var.dtype.scalar()]} " + \
+    f"[{ctx.r[loc]}+0], {('{' + ', '.join(ctx.r[var]) + '}') if var.dtype.count > 1 else ctx.r[var]};"),
+  (UPat(Ops.LOAD, name="x", src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("loc"), UPat.var("gate"))), UPat.var("alt")), allow_any_len=True),
+    lambda ctx, x, loc, alt, gate, buf: flatten([
     [f"mov.{ctx.mem_types[x.dtype.scalar()]} {v}, {render_val(0, x.dtype.scalar())};" for v in ctx.r[x]],
-    [f"@{ctx.r[gate]} ld.{mem_type(x)}.v{x.dtype.count}.{ctx.mem_types[x.dtype.scalar()]} {{{', '.join(ctx.r[x])}}}, [{ctx.r[loc]}+0];"]
+    [f"@{ctx.r[gate]} ld.{mem_type(buf)}.v{x.dtype.count}.{ctx.mem_types[x.dtype.scalar()]} {{{', '.join(ctx.r[x])}}}, [{ctx.r[loc]}+0];"]
   ]) if alt.dtype.count > 1 else [
-    f"@{ctx.r[gate]} ld.{mem_type(x)}.{ctx.mem_types[x.dtype.scalar()]} {ctx.r[x]}, [{ctx.r[loc]}+0];",
+    f"@{ctx.r[gate]} ld.{mem_type(buf)}.{ctx.mem_types[x.dtype.scalar()]} {ctx.r[x]}, [{ctx.r[loc]}+0];",
     f"@!{ctx.r[gate]} mov.b{ctx.types[x.dtype.scalar()][1:]} {ctx.r[x]}, {ctx.r[alt]};"]),
-  (UPat(Ops.LOAD, name="x", src=(UPat.var('loc'),), allow_any_len=True),
-   lambda ctx, x, loc: f"ld.{mem_type(x)}.v{x.dtype.count}.{ctx.mem_types[x.dtype.scalar()]} {{{', '.join(ctx.r[x])}}}, [{ctx.r[loc]}+0];" \
-     if x.dtype.count > 1 else f"ld.{mem_type(x)}.{ctx.mem_types[x.dtype]} {ctx.r[x]}, [{ctx.r[loc]}+0];"),
+  (UPat(Ops.LOAD, name="x", src=(UPat(Ops.INDEX, src=(UPat.var("buf"), UPat.var("loc"))),), allow_any_len=True),
+    lambda ctx, x, loc, buf: f"ld.{mem_type(buf)}.v{x.dtype.count}.{ctx.mem_types[x.dtype.scalar()]} {{{', '.join(ctx.r[x])}}}, [{ctx.r[loc]}+0];" \
+     if x.dtype.count > 1 else f"ld.{mem_type(buf)}.{ctx.mem_types[x.dtype]} {ctx.r[x]}, [{ctx.r[loc]}+0];"),
+  # simple
   (UPat(Ops.DEFINE_REG, src=()), lambda ctx: []),
-  (UPat(Ops.RANGE, name="x"), lambda ctx, x: [f"mov.u32 {ctx.r[x]}, 0;", "LOOP_" + f"{ctx.r[x][1:]}:"]),
-  (UPat(Ops.ENDRANGE, name="x", src=(UPat.var("src0"),)), lambda ctx, x, src0: [
-    ctx.code_for_op[Ops.ADD](ctx.r[src0], ctx.r[src0], "1", dtypes.int, ctx.types[dtypes.int]),
-    ctx.code_for_op[Ops.CMPLT](ctx.r[x], ctx.r[x.src[0]], ctx.r[src0.src[0]], dtypes.int, ctx.types[dtypes.int]),
-    f"@{ctx.r[x]} bra LOOP_{ctx.r[src0][1:]};"]),
+  (UPat(Ops.RANGE, name="r"), lambda ctx, r: [
+    f"mov.u32 {ctx.r[r]}, -1;",
+    f"bra END_{ctx.r[r][1:]};",
+    "LOOP_" + f"{ctx.r[r][1:]}:"]),
+  (UPat(Ops.END, name="x", src=(UPat(), UPat(Ops.RANGE, name="r"))), lambda ctx, x, r: [
+    "END_" + f"{ctx.r[r][1:]}:",
+    ctx.code_for_op[Ops.ADD](ctx.r[r], ctx.r[r], "1", dtypes.int, ctx.types[dtypes.int]),
+    ctx.code_for_op[Ops.CMPLT](ctx.r[x], ctx.r[r], ctx.r[r.src[0]], dtypes.int, ctx.types[dtypes.int]),
+    f"@{ctx.r[x]} bra LOOP_{ctx.r[r][1:]};"]),
   (UPat(Ops.DEFINE_LOCAL, name="x"),
    lambda ctx, x: [f".shared .align 16 .b8 local{x.arg}[{x.dtype.size*x.dtype.itemsize}];", f"mov.u64 {ctx.r[x]}, local{x.arg}[0];"]),
   (UPat(Ops.IF, name="x"), lambda ctx, x: f"@!{ctx.r[x.src[0]]} bra IF_{ctx.r[x.src[0]][1:]}_{ctx.uops.index(x)};"),
@@ -178,7 +187,10 @@ class PTXRenderer(Renderer):
 
     name = "test"
     for u in uops:
-      if u.op is Ops.NOOP: continue
+      if u.op in {Ops.NOOP, Ops.GROUP}: continue
+      if u.op is Ops.AFTER:
+        self.r[u] = self.r[u.src[0]]
+        continue
       if u.op is Ops.SINK:
         if u.arg is not None: name = u.arg.function_name
         continue
@@ -204,6 +216,7 @@ class PTXRenderer(Renderer):
             typ = "pred" if u.src[1].dtype == dtypes.bool else ("b"+self.types[u.src[1].dtype][1:])
             kernel.append(f"mov.{typ} {self.r[u.src[0]]}, {self.r[u.src[1]]};")
         continue
+      if u.op is Ops.INDEX: continue  # other index we can skip
       if u.op is Ops.SPECIAL: r[u] = "%" + u.arg
       elif u.op is Ops.DEFINE_VAR: bufs.append((u.arg[0], u.dtype))
       elif u.op is Ops.LOAD:
@@ -216,7 +229,7 @@ class PTXRenderer(Renderer):
                        [ssa("wmma_in", dtype="b32") for _ in range(0, len(r[u.src[1]]), 4 // u.src[0].dtype.scalar().itemsize)],
                        [ssa("wmma_acc", dtype="b32") for _ in range(0, len(r[u.src[2]]), 4 // u.dtype.scalar().itemsize)]]
         r[u] = [ssa("wmma", dtype=self.types[u.dtype.scalar()]) for _ in range(u.dtype.count)]
-      prefix, dtype = {Ops.CAST: ("cast", None), Ops.BITCAST: ("cast", None), Ops.ENDRANGE: ("pred", "pred"), Ops.RANGE: ("ridx", None),
+      prefix, dtype = {Ops.CAST: ("cast", None), Ops.BITCAST: ("cast", None), Ops.END: ("pred", "pred"), Ops.RANGE: ("ridx", None),
         Ops.DEFINE_VAR: ("dat", None), Ops.CONST: ("const", None), Ops.DEFINE_LOCAL: ("local",self.types[dtypes.ulong]),
         Ops.DEFINE_GLOBAL: ("dat", self.types[dtypes.ulong]), **{op: ("alu", None) for op in GroupOp.ALU}}.get(u.op, (None, None))
       if prefix: r[u] = ssa(prefix, u, dtype)

tinygrad/runtime/autogen/mesa.py

@@ -7,13 +7,11 @@
 # LONGDOUBLE_SIZE is: 16
 #
 import ctypes, ctypes.util, os, gzip, base64, subprocess, tinygrad.helpers as helpers
-def brew_prefix():
-  try: return subprocess.check_output(['brew', '--prefix', 'tinymesa']).decode().strip()
-  except Exception: return ''
 PATHS_TO_TRY = [
   (BASE:=os.getenv('MESA_PATH', f"/usr{'/local/' if helpers.OSX else '/'}lib"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so'),
   f'{BASE}/libtinymesa{EXT}',
-  f'{brew_prefix()}/lib/libtinymesa_cpu.dylib',
+  '/opt/homebrew/lib/libtinymesa_cpu.dylib',
+  '/opt/homebrew/lib/libtinymesa.dylib',
 ]
 def _try_dlopen_tinymesa_cpu():
   library = ctypes.util.find_library("tinymesa_cpu")
@@ -6087,7 +6085,7 @@ struct_nir_op_info._fields_ = [
 
 nir_op_info = struct_nir_op_info
 try: nir_op_infos = (struct_nir_op_info * 489).in_dll(_libraries['libtinymesa_cpu.so'], 'nir_op_infos')
-except AttributeError: pass
+except (AttributeError, ValueError): pass
 try:
     nir_op_is_selection = _libraries['FIXME_STUB'].nir_op_is_selection
     nir_op_is_selection.restype = ctypes.c_bool
@@ -8118,7 +8116,7 @@ c__EA_nir_intrinsic_index_flag = ctypes.c_uint32 # enum
 nir_intrinsic_index_flag = c__EA_nir_intrinsic_index_flag
 nir_intrinsic_index_flag__enumvalues = c__EA_nir_intrinsic_index_flag__enumvalues
 try: nir_intrinsic_index_names = (ctypes.POINTER(ctypes.c_char) * 75).in_dll(_libraries['libtinymesa_cpu.so'], 'nir_intrinsic_index_names')
-except AttributeError: pass
+except (AttributeError, ValueError): pass
 class struct_nir_intrinsic_instr(Structure):
     pass
 
@@ -8242,7 +8240,7 @@ struct_nir_intrinsic_info._fields_ = [
 
 nir_intrinsic_info = struct_nir_intrinsic_info
 try: nir_intrinsic_infos = (struct_nir_intrinsic_info * 732).in_dll(_libraries['libtinymesa_cpu.so'], 'nir_intrinsic_infos')
-except AttributeError: pass
+except (AttributeError, ValueError): pass
 try:
     nir_intrinsic_src_components = _libraries['libtinymesa_cpu.so'].nir_intrinsic_src_components
     nir_intrinsic_src_components.restype = ctypes.c_uint32
@@ -19877,4 +19875,4 @@ __all__ = \
     'union_util_format_description_0', 'util_format_colorspace',
     'util_format_layout', 'va_list']
 lvp_nir_options = gzip.decompress(base64.b64decode('H4sIAAAAAAAAA2NgZGRkYGAAkYxgCsQFsxigwgwQBoxmhCqFq2WEKwIrAEGIkQxoAEMALwCqVsCiGUwLMHA0QPn29nBJkswHANb8YpH4AAAA'))
-def __getattr__(nm): raise AttributeError() if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa (tinymesa-32dc66c, mesa-25.2.4)')
+def __getattr__(nm): raise AttributeError('LLVMpipe requires tinymesa_cpu' if 'tinymesa_cpu' not in dll._name else f'attribute {nm} not found') if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa (tinymesa-32dc66c, mesa-25.2.4)')

tinygrad/runtime/ops_amd.py

@@ -4,7 +4,7 @@ import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, co
 assert sys.platform != 'win32'
 from dataclasses import dataclass
 from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, HWQueue, CLikeArgsState, HCQSignal, HCQProgram, FileIOInterface
-from tinygrad.runtime.support.hcq import MMIOInterface, BumpAllocator
+from tinygrad.runtime.support.hcq import MMIOInterface, BumpAllocator, hcq_filter_visible_devices
 from tinygrad.uop.ops import sint
 from tinygrad.device import Compiled, DMAFdRef, BufferSpec, CompilerPairT
 from tinygrad.helpers import getenv, round_up, data64_le, DEBUG, PROFILE, ProfileEvent, suppress_finalizing, lo32, hi32, colored
@@ -15,11 +15,11 @@ from tinygrad.runtime.autogen.am import am
 from tinygrad.runtime.support.compiler_amd import HIPCompiler, AMDLLVMCompiler
 from tinygrad.runtime.support.elf import elf_loader
 from tinygrad.runtime.support.am.amdev import AMDev, AMMemoryManager
-from tinygrad.runtime.support.amd import AMDReg, AMDIP, import_module, import_soc, setup_pci_bars
-from tinygrad.runtime.support.system import System, PCIIfaceBase, PCIAllocationMeta, MAP_FIXED, MAP_NORESERVE
-from tinygrad.runtime.support.usb import ASM24Controller, USBMMIOInterface
+from tinygrad.runtime.support.amd import AMDReg, AMDIP, import_module, import_soc, import_ip_offsets
+from tinygrad.runtime.support.system import System, PCIIfaceBase, PCIAllocationMeta, PCIDevice, USBPCIDevice, MAP_FIXED, MAP_NORESERVE
 if getenv("IOCTL"): import extra.hip_gpu_driver.hip_ioctl  # noqa: F401 # pylint: disable=unused-import
 
+SQTT = getenv("SQTT", 0)
 EVENT_INDEX_PARTIAL_FLUSH = 4 # based on a comment in nvd.h
 WAIT_REG_MEM_FUNCTION_EQ  = 3 # ==
 WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=
@@ -66,11 +66,15 @@ class AMDComputeQueue(HWQueue):
     if self.dev.xccs > 1:
       self._q[prev_len-1] |= (len(self._q) - prev_len)
 
-  def wait_reg_mem(self, value, mask=0xffffffff, mem=None, reg_req=None, reg_done=None):
-    wrm_info_dw = self.pm4.WAIT_REG_MEM_MEM_SPACE(int(mem is not None)) | self.pm4.WAIT_REG_MEM_OPERATION(int(mem is None)) \
-                | self.pm4.WAIT_REG_MEM_FUNCTION(WAIT_REG_MEM_FUNCTION_GEQ) | self.pm4.WAIT_REG_MEM_ENGINE(0)
+  def set_grbm_broadcast(self):
+    self.wreg(self.gc.regGRBM_GFX_INDEX, **{f'{f}_broadcast_writes': 1 for f in ['se', 'sh' if self.dev.target[0] == 9 else 'sa', 'instance']})
+  def set_grbm_se(self, se): self.wreg(self.gc.regGRBM_GFX_INDEX, se_index=se, instance_broadcast_writes=1)
 
-    self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, wrm_info_dw, *(data64_le(mem) if mem is not None else (reg_req, reg_done)), value, mask, 4)
+  def wait_reg_mem(self, value, mask=0xffffffff, mem=None, reg=None, reg_done=0, op=WAIT_REG_MEM_FUNCTION_GEQ):
+    wrm_info_dw = self.pm4.WAIT_REG_MEM_MEM_SPACE(int(mem is not None)) | self.pm4.WAIT_REG_MEM_OPERATION(int(mem is None and reg_done > 0)) \
+                | self.pm4.WAIT_REG_MEM_FUNCTION(op) | self.pm4.WAIT_REG_MEM_ENGINE(0)
+
+    self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, wrm_info_dw, *(data64_le(mem) if mem is not None else (reg, reg_done)), value, mask, 4)
 
   def acquire_mem(self, addr=0x0, sz=(1 << 64)-1, gli=1, glm=1, glk=1, glv=1, gl1=1, gl2=1):
     if self.dev.target >= (10,0,0):
@@ -113,7 +117,7 @@ class AMDComputeQueue(HWQueue):
 
   def memory_barrier(self):
     pf = '' if self.nbio.version[0] == 2 else '0' if self.nbio.version[:2] != (7, 11) else '1'
-    self.wait_reg_mem(reg_req=getattr(self.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_REQ').addr[0],
+    self.wait_reg_mem(reg=getattr(self.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_REQ').addr[0],
                       reg_done=getattr(self.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_DONE').addr[0], value=0xffffffff)
     self.acquire_mem()
     return self
@@ -124,6 +128,22 @@ class AMDComputeQueue(HWQueue):
 
   ### SQTT ###
 
+  def sqtt_setup_exec(self, prg, global_size):
+    self.sqtt_userdata(sqtt.struct_rgp_sqtt_marker_pipeline_bind(
+      _0=sqtt.union_rgp_sqtt_marker_pipeline_bind_0(_0=sqtt.struct_rgp_sqtt_marker_pipeline_bind_0_0(
+        identifier=sqtt.RGP_SQTT_MARKER_IDENTIFIER_BIND_PIPELINE, bind_point=(__BIND_POINT_COMPUTE:=1))),
+      _1=sqtt.union_rgp_sqtt_marker_pipeline_bind_1(api_pso_hash=data64_le(prg.libhash[0]))))
+
+    self.sqtt_userdata(sqtt.struct_rgp_sqtt_marker_event(
+      _0=sqtt.union_rgp_sqtt_marker_event_0(_0=sqtt.struct_rgp_sqtt_marker_event_0_0(has_thread_dims=1)),
+      _2=sqtt.union_rgp_sqtt_marker_event_2(cmd_id=next(prg.dev.sqtt_next_cmd_id))), *global_size)
+
+    for xcc in range(self.dev.xccs):
+      with self.pred_exec(xcc_mask=1 << xcc):
+        for i in range(8 if prg.dev.target >= (11,0,0) else 4):
+          self.wreg(getattr(self.gc, f'regCOMPUTE_STATIC_THREAD_MGMT_SE{i}'),
+            ((prg.dev.sqtt_itrace_se_mask >> ((self.dev.se_cnt // self.dev.xccs) * xcc + i)) & 0b1) if SQTT >= 2 else 0xffffffff)
+
   def sqtt_userdata(self, data, *extra_dwords):
     data_ints = [x[0] for x in struct.iter_unpack('<I', bytes(data))] + list(extra_dwords)
     for i in range(0, len(data_ints), 2):
@@ -134,87 +154,97 @@ class AMDComputeQueue(HWQueue):
     self.wreg(self.gc.regSQ_THREAD_TRACE_CTRL, draw_event_en=1, spi_stall_en=1, sq_stall_en=1, reg_at_hwm=2, hiwater=1, util_timer=1,
       mode=int(tracing), **trace_ctrl)
 
-  # Magic values from mesa/src/amd/vulkan/radv_sqtt.c:radv_emit_spi_config_cntl and src/amd/common/ac_sqtt.c:ac_sqtt_emit_start
   def sqtt_start(self, buf0s:list[HCQBuffer], se_mask:int):
     self.memory_barrier()
-    self.spi_config(tracing=True)
-    # One buffer for one SE, mesa does it with a single buffer and ac_sqtt_get_data_offset, but this is simpler and should work just as well
-    for se in range(len(buf0s)):
-      self.wreg(self.gc.regGRBM_GFX_INDEX, se_index=se, instance_broadcast_writes=1)
-      buf0_lo, buf0_hi = data64_le(buf0s[se].va_addr >> 12)
-      if self.dev.target >= (12,0,0):
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_SIZE, size=buf0s[se].size >> 12)
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE_LO, base_lo=buf0_lo)
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE_HI, base_hi=buf0_hi)
-      else:
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_SIZE, base_hi=buf0_hi, size=buf0s[se].size >> 12)
-        self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE, base_lo=buf0_lo)
-      # NOTE: SQTT can only trace instructions on one simd per se, this selects first simd in first wgp in first sa.
-      # For RGP to display instruction trace it has to see it on first SE. Howerver ACE/MEC/whatever does the dispatching starting with second se,
-      # and on amdgpu/non-AM it also does weird things with dispatch order inside se: around 7 times out of 10 it starts from the last cu, but
-      # sometimes not, especially if the kernel has more than one wavefront which means that kernels with small global size might get unlucky and
-      # be dispatched on something else and not be seen in instruction tracing tab. You can force the wavefronts of a kernel to be dispatched on the
-      # CUs you want to by disabling other CUs via bits in regCOMPUTE_STATIC_THREAD_MGMT_SE<x> and trace even kernels that only have one wavefront.
-      cs_wtype = (1 << 6) if self.dev.target >= (12,0,0) else self.soc.SQ_TT_WTYPE_INCLUDE_CS_BIT
-      self.wreg(self.gc.regSQ_THREAD_TRACE_MASK, wtype_include=cs_wtype, simd_sel=0, wgp_sel=0, sa_sel=0)
-      reg_include = self.soc.SQ_TT_TOKEN_MASK_SQDEC_BIT | self.soc.SQ_TT_TOKEN_MASK_SHDEC_BIT | self.soc.SQ_TT_TOKEN_MASK_GFXUDEC_BIT | \
-                    self.soc.SQ_TT_TOKEN_MASK_COMP_BIT | self.soc.SQ_TT_TOKEN_MASK_CONTEXT_BIT
-      token_exclude = (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_PERF_SHIFT) if self.dev.target < (12,0,0) else 0
+    if self.dev.target[0] == 9:
+      self.set_grbm_broadcast()
+      self.wreg(self.gc.regSQ_THREAD_TRACE_MASK, simd_en=0xf, cu_sel=0, sq_stall_en=1, spi_stall_en=1, reg_stall_en=1, vm_id_mask=0)
+      for se in range(len(buf0s)):
+        mask = (__SQTT_MISC:=1<<0) | (__SQTT_TIME:=1<<1) | (__SQTT_REG:=1<<2) | (__SQTT_WAVE_START:=1<<3) | (__SQTT_WAVE_END:=1<<6) \
+             | (__SQTT_USERDATA:=1<<12) | (__SQTT_REG_CS:=1<<5) | (__SQTT_REG_CS_PRIV:=1<<15)
+        if (se_mask >> se) & 0b1: mask |= (__SQTTINST:=1<<10) | (__SQTT_INST_PC:=1<<11) | (__SQTT_ISSUE:=1<<13)
 
-      # disable tracing
-      if not (se_mask >> se) & 0b1:
-        # gfx12 doesn't have enums with all fields, so it's hardcoded, but it's the same as gfx11.
-        token_exclude |= (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VMEMEXEC_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_ALUEXEC_SHIFT | \
-                         1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VALUINST_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_IMMEDIATE_SHIFT | \
-                         1 << self.soc.SQ_TT_TOKEN_EXCLUDE_INST_SHIFT) if self.dev.target < (12,0,0) else 0x927
+        with self.pred_exec(xcc_mask=1<<(se // (ses_per_xcc:=(self.dev.se_cnt // self.dev.xccs)))):
+          self.set_grbm_se(se % ses_per_xcc)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK, reg_mask=0xf, token_mask=mask)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK2, inst_mask=0xffffffff)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BASE, addr=lo32(buf0s[se].va_addr >> 12))
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BASE2, addr_hi=hi32(buf0s[se].va_addr >> 12))
+          self.wreg(self.gc.regSQ_THREAD_TRACE_SIZE, size=buf0s[se].size >> 12)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_CTRL, reset_buffer=1)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_MODE, mask_cs=1, autoflush_en=1, mode=1)
+    else:
+      self.spi_config(tracing=True)
+      # One buffer for one SE, mesa does it with a single buffer and ac_sqtt_get_data_offset, but this is simpler and should work just as well
+      for se in range(len(buf0s)):
+        self.set_grbm_se(se)
 
-      token_mask = {} if self.dev.target < (12,0,0) else {'exclude_barrier_wait': 1}
-      self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK, reg_include=reg_include, token_exclude=token_exclude, bop_events_token_include=1, **token_mask)
-      # Enable SQTT
-      self.sqtt_config(tracing=True)
-    # Restore global broadcasting
-    self.wreg(self.gc.regGRBM_GFX_INDEX, se_broadcast_writes=1, sa_broadcast_writes=1, instance_broadcast_writes=1)
-    self.wreg(self.gc.regCOMPUTE_THREAD_TRACE_ENABLE, 1)
+        buf0_lo, buf0_hi = data64_le(buf0s[se].va_addr >> 12)
+        if self.dev.target >= (12,0,0):
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_SIZE, size=buf0s[se].size >> 12)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE_LO, base_lo=buf0_lo)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE_HI, base_hi=buf0_hi)
+        else:
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_SIZE, base_hi=buf0_hi, size=buf0s[se].size >> 12)
+          self.wreg(self.gc.regSQ_THREAD_TRACE_BUF0_BASE, base_lo=buf0_lo)
+        # NOTE: SQTT can only trace instructions on one simd per se, this selects first simd in first wgp in first sa.
+        # For RGP to display instruction trace it has to see it on first SE. Howerver ACE/MEC/whatever does the dispatching starting with second se,
+        # and on amdgpu/non-AM it also does weird things with dispatch order inside se: around 7 times out of 10 it starts from the last cu, but
+        # sometimes not, especially if the kernel has more than one wavefront which means that kernels with small global size might get unlucky and
+        # be dispatched on something else and not be seen in instruction tracing tab. You can force the wavefronts of a kernel to be dispatched on the
+        # CUs you want to by disabling other CUs via bits in regCOMPUTE_STATIC_THREAD_MGMT_SE<x> and trace even kernels that only have one wavefront.
+        cs_wtype = (1 << 6) if self.dev.target >= (12,0,0) else self.soc.SQ_TT_WTYPE_INCLUDE_CS_BIT
+        self.wreg(self.gc.regSQ_THREAD_TRACE_MASK, wtype_include=cs_wtype, simd_sel=0, wgp_sel=0, sa_sel=0)
+        reg_include = self.soc.SQ_TT_TOKEN_MASK_SQDEC_BIT | self.soc.SQ_TT_TOKEN_MASK_SHDEC_BIT | self.soc.SQ_TT_TOKEN_MASK_GFXUDEC_BIT | \
+                      self.soc.SQ_TT_TOKEN_MASK_COMP_BIT | self.soc.SQ_TT_TOKEN_MASK_CONTEXT_BIT
+        token_exclude = (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_PERF_SHIFT) if self.dev.target < (12,0,0) else 0
+
+        # disable instr tracing
+        if not (se_mask >> se) & 0b1:
+          # gfx12 doesn't have enums with all fields, so it's hardcoded, but it's the same as gfx11.
+          token_exclude |= (1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VMEMEXEC_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_ALUEXEC_SHIFT | \
+                            1 << self.soc.SQ_TT_TOKEN_EXCLUDE_VALUINST_SHIFT | 1 << self.soc.SQ_TT_TOKEN_EXCLUDE_IMMEDIATE_SHIFT | \
+                            1 << self.soc.SQ_TT_TOKEN_EXCLUDE_INST_SHIFT) if self.dev.target < (12,0,0) else 0x927
+
+        self.wreg(self.gc.regSQ_THREAD_TRACE_TOKEN_MASK, reg_include=reg_include, token_exclude=token_exclude, bop_events_token_include=1,
+                  **({} if self.dev.target < (12,0,0) else {'exclude_barrier_wait': 1}))
+        self.sqtt_config(tracing=True)
+
+    self.set_grbm_broadcast()
+    if self.dev.target[0] > 9: self.wreg(self.gc.regCOMPUTE_THREAD_TRACE_ENABLE, 1)
     self.memory_barrier()
     return self
 
   # Magic values from src/amd/common/ac_sqtt.c:ac_sqtt_emit_stop and src/amd/common/ac_sqtt.c:ac_sqtt_emit_wait
-  def sqtt_stop(self, ses: int, wptrs: HCQBuffer):
+  def sqtt_stop(self, ses:int, wptrs:HCQBuffer):
     self.memory_barrier()
+    self.set_grbm_broadcast()
+
     # Start shutting everything down
-    self.wreg(self.gc.regCOMPUTE_THREAD_TRACE_ENABLE, 0)
-    self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.THREAD_TRACE_FINISH) | self.pm4.EVENT_INDEX(0))
+    if self.dev.target[0] == 9: self.wreg(self.gc.regSQ_THREAD_TRACE_MODE, mask_cs=1, autoflush_en=1, mode=0)
+    else:
+      self.wreg(self.gc.regCOMPUTE_THREAD_TRACE_ENABLE, 0)
+      self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.THREAD_TRACE_FINISH) | self.pm4.EVENT_INDEX(0))
+
     # For each SE wait for finish to complete and copy regSQ_THREAD_TRACE_WPTR to know where in the buffer trace data ends
     for se in range(ses):
-      self.wreg(self.gc.regGRBM_GFX_INDEX, se_index=se, instance_broadcast_writes=1)
-      # Wait for FINISH_PENDING==0
-      self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, self.pm4.WAIT_REG_MEM_FUNCTION(WAIT_REG_MEM_FUNCTION_EQ),
-                self.gc.regSQ_THREAD_TRACE_STATUS.addr[0], 0, 0, self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('finish_pending'), 4)
-      # Disable SQTT
-      self.sqtt_config(tracing=False)
-      # Wait for BUSY==0
-      self.pkt3(self.pm4.PACKET3_WAIT_REG_MEM, self.pm4.WAIT_REG_MEM_FUNCTION(WAIT_REG_MEM_FUNCTION_EQ),
-                self.gc.regSQ_THREAD_TRACE_STATUS.addr[0], 0, 0, self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('busy'), 4)
+      self.set_grbm_se(se)
+
+      status_reg = self.gc.regSQ_THREAD_TRACE_STATUS.addr[0] - (self.pm4.PACKET3_SET_UCONFIG_REG_START if self.dev.target[0] == 9 else 0)
+      if self.dev.target >= (10, 0, 0):
+        self.wait_reg_mem(reg=status_reg, mask=self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('finish_pending'), op=WAIT_REG_MEM_FUNCTION_EQ, value=0)
+        self.sqtt_config(tracing=False)
+      self.wait_reg_mem(reg=status_reg, mask=self.gc.regSQ_THREAD_TRACE_STATUS.fields_mask('busy'), op=WAIT_REG_MEM_FUNCTION_EQ, value=0)
+      self.pkt3(self.pm4.PACKET3_EVENT_WRITE, self.pm4.EVENT_TYPE(self.soc.CS_PARTIAL_FLUSH) | self.pm4.EVENT_INDEX(EVENT_INDEX_PARTIAL_FLUSH))
+
       # Copy WPTR to memory (src_sel = perf, dst_sel = tc_l2, wr_confirm = True)
       self.pkt3(self.pm4.PACKET3_COPY_DATA, 1 << 20 | 2 << 8 | 4, self.gc.regSQ_THREAD_TRACE_WPTR.addr[0], 0, *data64_le(wptrs.va_addr+(se*4)))
-    # Restore global broadcasting
-    self.wreg(self.gc.regGRBM_GFX_INDEX, se_broadcast_writes=1, sa_broadcast_writes=1, instance_broadcast_writes=1)
-    self.spi_config(tracing=False)
+
+    self.set_grbm_broadcast()
+    if self.dev.target[0] > 9: self.spi_config(tracing=False)
     self.memory_barrier()
     return self
 
-  def sqtt_prg_marker(self, prg:AMDProgram, global_size:tuple[sint, ...]):
-    BIND_POINT_COMPUTE = 1
-
-    self.sqtt_userdata(sqtt.struct_rgp_sqtt_marker_pipeline_bind(
-      _0=sqtt.union_rgp_sqtt_marker_pipeline_bind_0(_0=sqtt.struct_rgp_sqtt_marker_pipeline_bind_0_0(
-        identifier=sqtt.RGP_SQTT_MARKER_IDENTIFIER_BIND_PIPELINE, bind_point=BIND_POINT_COMPUTE)),
-      _1=sqtt.union_rgp_sqtt_marker_pipeline_bind_1(api_pso_hash=data64_le(prg.libhash[0]))))
-
-    self.sqtt_userdata(sqtt.struct_rgp_sqtt_marker_event(
-      _0=sqtt.union_rgp_sqtt_marker_event_0(_0=sqtt.struct_rgp_sqtt_marker_event_0_0(has_thread_dims=1)),
-      _2=sqtt.union_rgp_sqtt_marker_event_2(cmd_id=next(prg.dev.sqtt_next_cmd_id))), *global_size)
-
   def exec(self, prg:AMDProgram, args_state:CLikeArgsState, global_size:tuple[sint, ...], local_size:tuple[sint, ...]):
     self.bind_args_state(args_state)
 
@@ -238,7 +268,7 @@ class AMDComputeQueue(HWQueue):
 
     user_regs += [*data64_le(args_state.buf.va_addr)]
 
-    if prg.dev.sqtt_enabled: self.sqtt_prg_marker(prg, global_size)
+    if prg.dev.sqtt_enabled: self.sqtt_setup_exec(prg, global_size)
 
     self.wreg(self.gc.regCOMPUTE_PGM_LO, *data64_le(prg.prog_addr >> 8))
     self.wreg(self.gc.regCOMPUTE_PGM_RSRC1, prg.rsrc1, prg.rsrc2)
@@ -254,13 +284,8 @@ class AMDComputeQueue(HWQueue):
     if (10,0,0) <= prg.dev.target < (11,0,0): self.wreg(self.gc.mmCP_COHER_START_DELAY, 0x20)
 
     self.wreg(self.gc.regCOMPUTE_RESTART_X, 0, 0, 0)
-    self.wreg(self.gc.regCOMPUTE_STATIC_THREAD_MGMT_SE0, 0xFFFFFFFF, 0xFFFFFFFF)
-    self.wreg(self.gc.regCOMPUTE_STATIC_THREAD_MGMT_SE2, 0xFFFFFFFF, 0xFFFFFFFF)
-    if prg.dev.target >= (11,0,0): self.wreg(self.gc.regCOMPUTE_STATIC_THREAD_MGMT_SE4, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF)
-
     self.wreg(self.gc.regCOMPUTE_USER_DATA_0, *user_regs)
     self.wreg(self.gc.regCOMPUTE_RESOURCE_LIMITS, 0)
-
     self.wreg(self.gc.regCOMPUTE_START_X, 0, 0, 0, *local_size, 0, 0)
 
     gfx10p = {'cs_w32_en': int(prg.wave32)} if prg.dev.target >= (10,0,0) else {}
@@ -278,6 +303,7 @@ class AMDComputeQueue(HWQueue):
   def timestamp(self, signal:AMDSignal):
     with self.pred_exec(xcc_mask=0b1):
       self.release_mem(signal.timestamp_addr, 0, self.pm4.data_sel__mec_release_mem__send_gpu_clock_counter, self.pm4.int_sel__mec_release_mem__none)
+      self.acquire_mem() # ensure timestamp is written
     return self
 
   def signal(self, signal:AMDSignal, value:sint=0):
@@ -320,6 +346,7 @@ class AMDComputeQueue(HWQueue):
 class AMDComputeAQLQueue(AMDComputeQueue):
   def exec(self, prg:AMDProgram, args_state:CLikeArgsState, global_size:tuple[sint, ...], local_size:tuple[sint, ...]):
     self.bind_args_state(args_state)
+    if prg.dev.sqtt_enabled: self.sqtt_setup_exec(prg, global_size)
     self._q.append(pkt:=hsa.hsa_kernel_dispatch_packet_t(header=AQL_HDR | (hsa.HSA_PACKET_TYPE_KERNEL_DISPATCH << hsa.HSA_PACKET_HEADER_TYPE),
       setup=3<<hsa.HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS, private_segment_size=prg.private_segment_size,
       group_segment_size=prg.group_segment_size, kernel_object=prg.aql_prog_addr, kernarg_address=args_state.buf.va_addr))
@@ -334,7 +361,7 @@ class AMDComputeAQLQueue(AMDComputeQueue):
     def flush_pm4_batch():
       nonlocal pm4_batch
       if not pm4_batch: return bytes()
-      dev.pm4_ibs.cpu_view().view(off:=dev.pm4_ib_alloc.alloc(len(pm4_batch) * 4), fmt='I')[:len(pm4_batch)] = array.array('I', pm4_batch)
+      dev.pm4_ibs.cpu_view().view(off:=dev.pm4_ib_alloc.alloc(len(pm4_batch) * 4, 16), fmt='I')[:len(pm4_batch)] = array.array('I', pm4_batch)
       pkt = [AQL_HDR | (hsa.HSA_PACKET_TYPE_VENDOR_SPECIFIC << hsa.HSA_PACKET_HEADER_TYPE) | (1 << 16),
         self.pm4.PACKET3(self.pm4.PACKET3_INDIRECT_BUFFER, 2), *data64_le(dev.pm4_ibs.va_addr+off), len(pm4_batch)|self.pm4.INDIRECT_BUFFER_VALID, 10]
       pm4_batch.clear()
@@ -548,9 +575,7 @@ class KFDIface:
     if KFDIface.kfd is None:
       KFDIface.kfd = FileIOInterface("/dev/kfd", os.O_RDWR)
       gpus = [g for g in FileIOInterface(kfd_topo_path).listdir() if self._is_usable_gpu(FileIOInterface(f"{kfd_topo_path}/{g}/gpu_id"))]
-      gpus = sorted(gpus, key=lambda x: int(x.split('/')[-1]))
-      visible_devices = [int(x) for x in (getenv('VISIBLE_DEVICES', getenv('HIP_VISIBLE_DEVICES', ''))).split(',') if x.strip()]
-      KFDIface.gpus = [gpus[x] for x in visible_devices] if visible_devices else gpus
+      KFDIface.gpus = hcq_filter_visible_devices(sorted(gpus, key=lambda x: int(x.split('/')[-1])))
 
     if device_id >= len(KFDIface.gpus): raise RuntimeError(f"No device found for {device_id}. Requesting more devices than the system has?")
 
@@ -560,8 +585,6 @@ class KFDIface:
     id2ip = {am.GC_HWID: am.GC_HWIP, am.SDMA0_HWID: am.SDMA0_HWIP, am.NBIF_HWID: am.NBIF_HWIP}
     ip_hw = [(id2ip[int(hwid)], int(hwid)) for hwid in FileIOInterface(ip_base).listdir() if hwid.isnumeric() and int(hwid) in id2ip]
     self.ip_versions = {ip:tuple(int(FileIOInterface(f'{ip_base}/{hw}/0/{part}').read()) for part in ['major','minor','revision']) for ip,hw in ip_hw}
-    self.ip_offsets = {ip:{int(i):tuple(int(x, 16) for x in FileIOInterface(f'{ip_base}/{hw}/{i}/base_addr').read().splitlines())
-      for i in FileIOInterface(f'{ip_base}/{hw}').listdir()} for ip,hw in ip_hw }
     self.drm_fd = FileIOInterface(f"/dev/dri/renderD{self.props['drm_render_minor']}", os.O_RDWR)
 
     self.kfd_ver = ((ver_st:=kfd.AMDKFD_IOC_GET_VERSION(KFDIface.kfd)).major_version, ver_st.minor_version)
@@ -672,12 +695,12 @@ class PCIIface(PCIIfaceBase):
   def __init__(self, dev, dev_id):
     super().__init__(dev, dev_id, vendor=0x1002, devices=[0x744c, 0x7480, 0x7550, 0x7590], bars=[0, 2, 5], vram_bar=0,
       va_start=AMMemoryManager.va_allocator.base, va_size=AMMemoryManager.va_allocator.size)
-    self._setup_adev(self.pci_dev.pcibus, self.pci_dev.map_bar(0), self.pci_dev.map_bar(2, fmt='Q'), self.pci_dev.map_bar(5, fmt='I'))
+    self._setup_adev(self.pci_dev)
     self.pci_dev.write_config(pci.PCI_COMMAND, self.pci_dev.read_config(pci.PCI_COMMAND, 2) | pci.PCI_COMMAND_MASTER, 2)
 
-  def _setup_adev(self, name, vram:MMIOInterface, doorbell:MMIOInterface, mmio:MMIOInterface, dma_regions:list[tuple[int, MMIOInterface]]|None=None):
-    self.dev_impl:AMDev = AMDev(name, vram, doorbell, mmio, dma_regions)
-    self.ip_offsets, self.ip_versions = self.dev_impl.regs_offset, self.dev_impl.ip_ver
+  def _setup_adev(self, pci_dev:PCIDevice, dma_regions:list[tuple[int, MMIOInterface]]|None=None):
+    self.dev_impl:AMDev = AMDev(pci_dev, dma_regions)
+    self.ip_versions = self.dev_impl.ip_ver
 
     gfxver = int(f"{self.dev_impl.ip_ver[am.GC_HWIP][0]:02d}{self.dev_impl.ip_ver[am.GC_HWIP][1]:02d}{self.dev_impl.ip_ver[am.GC_HWIP][2]:02d}")
     array_count = self.dev_impl.gc_info.gc_num_sa_per_se * self.dev_impl.gc_info.gc_num_se
@@ -701,7 +724,7 @@ class PCIIface(PCIIfaceBase):
       read_ptrs=[gart.cpu_view().view(offset=rptr, size=8, fmt='Q')], write_ptrs=[gart.cpu_view().view(offset=wptr, size=8, fmt='Q')])
 
   def sleep(self, timeout):
-    if self.pci_dev.irq_poller is not None and (events_cnt:=len(self.pci_dev.irq_poller.poll(timeout))):
+    if hasattr(self.pci_dev, 'irq_poller') and self.pci_dev.irq_poller is not None and (events_cnt:=len(self.pci_dev.irq_poller.poll(timeout))):
       self.pci_dev.irq_fd.read(8 * events_cnt)
       self.dev_impl.ih.interrupt_handler()
 
@@ -714,34 +737,29 @@ class PCIIface(PCIIfaceBase):
 
 class USBIface(PCIIface):
   def __init__(self, dev, dev_id): # pylint: disable=super-init-not-called
-    self.dev = dev
-    self.usb = ASM24Controller()
-    self.bars = setup_pci_bars(self.usb, gpu_bus=4, mem_base=0x10000000, pref_mem_base=(32 << 30))
-
-    self._setup_adev(f"usb:{dev_id}", USBMMIOInterface(self.usb, *self.bars[0], fmt='B'), USBMMIOInterface(self.usb, *self.bars[2], fmt='Q'),
-      USBMMIOInterface(self.usb, *self.bars[5], fmt='I'), dma_regions=[(0x200000, self._dma_view(0xf000, 0x80000))])
-    self.usb._pci_cacheable += [self.bars[2]] # doorbell region is cacheable
+    self.dev, self.pci_dev = dev, USBPCIDevice(f"usb:{dev_id}", bars=[0, 2, 5])
+    self._setup_adev(self.pci_dev, dma_regions=[(0x200000, self.pci_dev.dma_view(0xf000, 0x80000))])
+    self.pci_dev.usb._pci_cacheable += [(self.pci_dev.bar_info[2].addr, self.pci_dev.bar_info[2].size)] # doorbell region is cacheable
 
     # special regions
     self.copy_bufs = [self._dma_region(ctrl_addr=0xf000, sys_addr=0x200000, size=0x80000)]
     self.sys_buf, self.sys_next_off = self._dma_region(ctrl_addr=0xa000, sys_addr=0x820000, size=0x1000), 0x800
 
-  def _dma_view(self, ctrl_addr, size): return USBMMIOInterface(self.usb, ctrl_addr, size, fmt='B', pcimem=False)
   def _dma_region(self, ctrl_addr, sys_addr, size):
     region = self.dev_impl.mm.map_range(vaddr:=self.dev_impl.mm.alloc_vaddr(size=size), size, [(sys_addr, size)], system=True, uncached=True)
-    return HCQBuffer(vaddr, size, meta=PCIAllocationMeta(region, has_cpu_mapping=False), view=self._dma_view(ctrl_addr, size), owner=self.dev)
+    return HCQBuffer(vaddr, size, meta=PCIAllocationMeta(region, has_cpu_mapping=False), view=self.pci_dev.dma_view(ctrl_addr, size), owner=self.dev)
 
   def alloc(self, size:int, host=False, uncached=False, cpu_access=False, contiguous=False, **kwargs) -> HCQBuffer:
     if (host or (uncached and cpu_access)) and self.sys_next_off + size < self.sys_buf.size:
       self.sys_next_off += size
       return self.sys_buf.offset(self.sys_next_off - size, size)
 
-    am_mapping = self.dev_impl.mm.valloc(size:=round_up(size, 4 << 10), uncached=uncached, contiguous=cpu_access)
-    return HCQBuffer(am_mapping.va_addr, size, meta=PCIAllocationMeta(am_mapping, has_cpu_mapping=False),
-      view=USBMMIOInterface(self.usb, self.bars[0][0] + am_mapping.paddrs[0][0], size, fmt='B') if cpu_access else None, owner=self.dev)
+    mapping = self.dev_impl.mm.valloc(size:=round_up(size, 4 << 10), uncached=uncached, contiguous=cpu_access)
+    barview = self.pci_dev.map_bar(bar=0, off=mapping.paddrs[0][0], size=mapping.size) if cpu_access else None
+    return HCQBuffer(mapping.va_addr, size, meta=PCIAllocationMeta(mapping, has_cpu_mapping=False), view=barview, owner=self.dev)
 
   def create_queue(self, queue_type, ring, gart, rptr, wptr, eop_buffer=None, cwsr_buffer=None, ctl_stack_size=0, ctx_save_restore_size=0, xcc_id=0):
-    if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE: self.usb._pci_cacheable += [(ring.cpu_view().addr, ring.size)]
+    if queue_type == kfd.KFD_IOC_QUEUE_TYPE_COMPUTE: self.pci_dev.usb._pci_cacheable += [(ring.cpu_view().addr, ring.size)]
     return super().create_queue(queue_type, ring, gart, rptr, wptr, eop_buffer, cwsr_buffer, ctl_stack_size, ctx_save_restore_size, xcc_id)
 
   def sleep(self, timeout): pass
@@ -777,14 +795,15 @@ class AMDDevice(HCQCompiled):
     debug_memory_size = round_up((self.max_cu_id + 1 if self.target >= (10,1,0) else 1) * (self.max_wave_id + 1) * 32, 64)
     if self.target[0] == 10: ctl_stack_size = min(ctl_stack_size, 0x7000)
 
+    self.ip_off = import_ip_offsets(self.target)
     self.soc = import_soc(self.target)
     self.pm4 = importlib.import_module(f"tinygrad.runtime.autogen.am.pm4_{'nv' if self.target[0] >= 10 else 'soc15'}")
     self.sdma = import_module('sdma', min(self.iface.ip_versions[am.SDMA0_HWIP], (6, 0, 0)))
-    self.gc = AMDIP('gc', self.iface.ip_versions[am.GC_HWIP], self.iface.ip_offsets[am.GC_HWIP])
+    self.gc = AMDIP('gc', self.iface.ip_versions[am.GC_HWIP],
+                    bases={i: tuple(getattr(self.ip_off, f'GC_BASE__INST{i}_SEG{s}', 0) for s in range(6)) for i in range(6)})
 
-    nbio_name = 'nbio' if self.target[0] < 12 else 'nbif'
-    nbio_pad = (0,) if self.target[0] == 9 else ()
-    self.nbio = AMDIP(nbio_name, self.iface.ip_versions[am.NBIF_HWIP], {i:nbio_pad+x for i,x in self.iface.ip_offsets[am.NBIF_HWIP].items()})
+    self.nbio = AMDIP('nbio' if self.target[0] < 12 else 'nbif', self.iface.ip_versions[am.NBIF_HWIP],
+                      bases={i: tuple(getattr(self.ip_off, f'NBIO_BASE__INST{i}_SEG{s}', 0) for s in range(9)) for i in range(6)})
 
     self.is_aql = getenv("AMD_AQL", int(self.xccs > 1))
     if self.is_aql:
@@ -811,16 +830,16 @@ class AMDDevice(HCQCompiled):
     self._ensure_has_local_memory(128) # set default scratch size to 128 bytes per thread
 
     # SQTT is disabled by default because of runtime overhead and big file sizes (~200mb to Tensor.full() two 4096x4096 tensors and matmul them)
-    self.sqtt_enabled = PROFILE and bool(getenv("SQTT", 0))
+    self.sqtt_enabled = PROFILE and SQTT > 0
     if self.sqtt_enabled:
-      if self.target[0] < 11: raise RuntimeError(f'SQ Thread Tracing is not supported on gc:{self.target}')
+      if self.target[0] not in {9, 11, 12}: raise RuntimeError(f'SQ Thread Tracing is not supported on gc:{self.target}')
       if not self.is_am() and (ppfeaturemask:=int(FileIOInterface('/sys/module/amdgpu/parameters/ppfeaturemask', os.O_RDONLY).read(), 16))&0x8000:
         raise RuntimeError("SQTT can't be enabled because of hardware bug, to workaround either use AMD_IFACE=PCI or add "
                            f"ppfeaturemask={(ppfeaturemask&~0x8000):#x} (current {ppfeaturemask=:#x} & ~PP_GFXOFF_MASK) to amdgpu module parameters\n"
                            "For more information read https://github.com/tinygrad/tinygrad/blob/master/extra/sqtt/README.md")
       SQTT_BUFFER_SIZE = getenv("SQTT_BUFFER_SIZE", 256) # in mb, per shader engine
-      self.sqtt_buffers = [self.allocator.alloc(SQTT_BUFFER_SIZE*1024*1024, BufferSpec(nolru=True)) for _ in range(self.se_cnt)]
-      self.sqtt_itrace_se_mask = getenv("SQTT_ITRACE_SE_MASK", 2) # -1 enable all, 0 disable all, >0 bitmask for where to enable instruction tracing
+      self.sqtt_buffers = [self.allocator.alloc(SQTT_BUFFER_SIZE*1024*1024, BufferSpec(nolru=True, uncached=True)) for _ in range(self.se_cnt)]
+      self.sqtt_itrace_se_mask = getenv("SQTT_ITRACE_SE_MASK", -1 if SQTT >= 2 else (1 << 1)) # se bitmask: -1 enable all, 0 disable all
       self.sqtt_next_cmd_id = itertools.count(0)
       cast(AMDComputeQueue, self.hw_compute_queue_t()).sqtt_start(self.sqtt_buffers, self.sqtt_itrace_se_mask).submit(self)
 
@@ -833,7 +852,7 @@ class AMDDevice(HCQCompiled):
         read_dispatch_id_field_base_byte_offset=getattr(hsa.amd_queue_t, 'read_dispatch_id').offset,
         max_cu_id=self.max_cu_id, max_wave_id=self.max_wave_id)
       gart.cpu_view().view(fmt='B')[:ctypes.sizeof(aql_desc)] = bytes(aql_desc)
-      self.aql_desc = hsa.amd_queue_t.from_address(gart.va_addr)
+      self.aql_desc = hsa.amd_queue_t.from_address(gart.cpu_view().addr)
 
     cwsr_buffer_size = round_up((ctx_save_restore_size + debug_memory_size) * self.iface.props.get('num_xcc', 1), mmap.PAGESIZE)
     cwsr_buffer = self.iface.alloc(cwsr_buffer_size) if ctx_save_restore_size else None
@@ -880,12 +899,13 @@ class AMDDevice(HCQCompiled):
       self.synchronize()
       if DEBUG >= 2: print(f'{self.device}: Saving SQTT in profile...')
       for i,buf0 in enumerate(self.sqtt_buffers):
-        wptr = ((wptrs_buf.cpu_view().view(fmt='I')[i] & 0x1FFFFFFF) - (((buf0.va_addr//32) & 0x1FFFFFFF) if self.target < (12,0,0) else 0)) * 32
+        wptr = ((wptrs_buf.cpu_view().view(fmt='I')[i] & 0x1FFFFFFF) - (((buf0.va_addr//32) & 0x1FFFFFFF) if self.target[0] == 11 else 0)) * 32
         if DEBUG >= 2: print(f'\t{self.device}: SE {i} blob size {wptr:#x}')
         assert wptr >= 0 and wptr <= buf0.size, f"{wptr} > {buf0.size}, should never happen"
         # When sqtt buffer overflows, wptr stops at the last dword
         if wptr >= buf0.size - 32:
           print(colored(f"{self.device}: Warning: SQTT buffer is full (SE {i})! Increase SQTT buffer with SQTT_BUFFER_SIZE=X (in MB)", "yellow"))
         self.allocator._copyout(sqtt_buf:=memoryview(bytearray(wptr)), buf0)
+        if self.target[0] == 9: sqtt_buf = memoryview(struct.pack('<Q', 0x11 | (4 << 13) | (0xf << 16) | (i << 24)) + sqtt_buf)
         Compiled.profile_events += [ProfileSQTTEvent(self.device, i, self.iface.props, bytes(sqtt_buf), bool((self.sqtt_itrace_se_mask >> i) & 0b1))]
     super()._at_profile_finalize()