only floats

* use Ops.GROUP instead of Ops.NOOP for merging stores

* fs noop

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

@@ -41,6 +41,10 @@ inputs:
     description: "Install LLVM?"
     required: false
     default: 'false'
+  mesa:
+    description: "Install mesa"
+    required: false
+    default: 'false'
 runs:
   using: "composite"
   steps:
@@ -289,3 +293,13 @@ runs:
       if: inputs.llvm == 'true' && runner.os == 'macOS'
       shell: bash
       run: brew install llvm@20
+
+    # **** mesa ****
+    - name: Install mesa (linux)
+      if: inputs.mesa == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: sudo curl -L https://github.com/sirhcm/tinymesa/releases/download/tinymesa-32dc66c/libtinymesa_cpu-mesa-25.2.4-linux-amd64.so -o /usr/lib/libtinymesa_cpu.so
+    - name: Install mesa (macOS)
+      if: inputs.mesa == 'true' && runner.os == 'macOS'
+      shell: bash
+      run: brew install sirhcm/tinymesa/tinymesa_cpu

.github/workflows/autogen.yml

@@ -36,8 +36,9 @@ jobs:
         cuda: 'true'
         webgpu: 'true'
         llvm: 'true'
+        pydeps: 'pyyaml mako'
     - name: Install autogen support packages
-      run: sudo apt-get install -y --no-install-recommends llvm-14-dev libclang-14-dev
+      run: sudo apt-get install -y --no-install-recommends llvm-14-dev libclang-14-dev llvm-20-dev
     - name: Verify OpenCL autogen
       run: |
         cp tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
@@ -89,3 +90,8 @@ jobs:
         cp tinygrad/runtime/autogen/llvm.py /tmp/llvm.py.bak
         ./autogen_stubs.sh llvm
         diff /tmp/llvm.py.bak tinygrad/runtime/autogen/llvm.py
+    - name: Verify mesa autogen
+      run: |
+        cp tinygrad/runtime/autogen/mesa.py /tmp/mesa.py.bak
+        ./autogen_stubs.sh mesa
+        diff /tmp/mesa.py.bak tinygrad/runtime/autogen/mesa.py

.github/workflows/benchmark.yml

@@ -51,17 +51,18 @@ jobs:
         rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
     - name: reset process replay
       run: python3.11 test/external/process_replay/reset.py
+    - name: Print macOS version
+      run: sw_vers
     - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=800 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
     - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=900 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=800 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
     - name: Run Stable Diffusion v2
       # TODO: very slow step time
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=10000 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=4500 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
     # process replay can't capture this, the graph is too large
-    # TODO: too slow
-    # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+    - name: Run SDXL
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
     - name: Run model inference benchmark
       run: METAL=1 python3.11 test/external/external_model_benchmark.py
     - name: Test speed vs torch
@@ -130,7 +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)
@@ -318,9 +319,9 @@ 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
@@ -618,22 +619,24 @@ 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=4 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
@@ -641,19 +644,9 @@ jobs:
         ln -s /data/home/tiny/tinygrad/testsig-*.so .
         PYTHONPATH=. CC=clang-19 CPU=1 CPU_LLVM=0 QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
         # benchmark on DSP with NOOPT=1, the devectorizer has issues
-        PYTHONPATH=. CC=clang-19 DSP=1 DONT_REALIZE_EXPAND=1 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
+        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/mlperf.yml

@@ -12,7 +12,7 @@ jobs:
   run_script_job:
     runs-on: [self-hosted, Linux, tinybox]
     if: github.repository_owner == 'tinygrad'
-    timeout-minutes: 360
+    timeout-minutes: 720
 
     steps:
     - name: Checkout Code

.github/workflows/test.yml

@@ -89,64 +89,65 @@ jobs:
         clang -O2 recognize.c -lm -o recognize
         cat test/models/efficientnet/Chicken.jpg | ./recognize | grep cock
 
-  torchbackend:
-    name: Torch Backend Tests
-    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: torch-backend-pillow-torchvision-et-pt
-        deps: testing_minimal
-        pydeps: "pillow torchvision expecttest"
-        llvm: 'true'
-    - name: Install ninja
-      run: |
-        sudo apt update || true
-        sudo apt install -y --no-install-recommends ninja-build
-    - name: Lint with ruff
-      run: |
-        pip3 install --upgrade --force-reinstall ruff==0.11.0
-        python3 -m ruff check extra/torch_backend/backend.py
-    - name: Test one op
-      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add
-    - name: Test ResNet-18
-      run: DEBUG=2 python3 extra/torch_backend/example.py
-    - name: My (custom) tests
-      run: python3 extra/torch_backend/test.py
-    - name: Test one op in torch tests
-      run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
-    - name: Test Ops with TINY_BACKEND
-      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/test_ops.py --durations=20
-    - name: Test in-place operations on views
-      run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
-    - name: Test multi-gpu
-      run: CPU=1 CPU_LLVM=1 GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py
+  # TODO: fix the torch backend and reenable
+  # torchbackend:
+  #   name: Torch Backend Tests
+  #   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: torch-backend-pillow-torchvision-et-pt
+  #       deps: testing_minimal
+  #       pydeps: "pillow torchvision expecttest"
+  #       llvm: 'true'
+  #   - name: Install ninja
+  #     run: |
+  #       sudo apt update || true
+  #       sudo apt install -y --no-install-recommends ninja-build
+  #   - name: Lint with ruff
+  #     run: |
+  #       pip3 install --upgrade --force-reinstall ruff==0.11.0
+  #       python3 -m ruff check extra/torch_backend/backend.py
+  #   - name: Test one op
+  #     run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add
+  #   - name: Test ResNet-18
+  #     run: DEBUG=2 python3 extra/torch_backend/example.py
+  #   - name: My (custom) tests
+  #     run: python3 extra/torch_backend/test.py
+  #   - name: Test one op in torch tests
+  #     run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
+  #   - name: Test Ops with TINY_BACKEND
+  #     run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/test_ops.py --durations=20
+  #   - name: Test in-place operations on views
+  #     run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
+  #   - name: Test multi-gpu
+  #     run: CPU=1 CPU_LLVM=1 GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py
 
-  torchbackendmore:
-    name: Torch Backend Tests More
-    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: torch-backend-pillow-torchvision-et-pt
-        deps: testing_minimal
-        llvm: 'true'
-    - name: Install ninja
-      run: |
-        sudo apt update || true
-        sudo apt install -y --no-install-recommends ninja-build
-    - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: CPU=1 CPU_LLVM=1 TARGET_EVAL_ACC_PCT=96.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
-    - name: Test some torch tests (expect failure)
-      run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true
+  # torchbackendmore:
+  #   name: Torch Backend Tests More
+  #   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: torch-backend-pillow-torchvision-et-pt
+  #       deps: testing_minimal
+  #       llvm: 'true'
+  #   - name: Install ninja
+  #     run: |
+  #       sudo apt update || true
+  #       sudo apt install -y --no-install-recommends ninja-build
+  #   - name: Test beautiful_mnist in torch with TINY_BACKEND
+  #     run: STEPS=20 CPU=1 TARGET_EVAL_ACC_PCT=90.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
+  #   - name: Test some torch tests (expect failure)
+  #     run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true
 
   bepython:
     name: Python Backend
@@ -203,7 +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
@@ -263,8 +264,8 @@ 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: Check SPEC=2
+      run: SPEC=2 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
@@ -272,6 +273,8 @@ jobs:
     #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
     - name: Run Clip tests for SD MLPerf on NULL backend
       run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
+    - name: Run AMD emulated BERT training on NULL backend
+      run: EMULATE=AMD_RDNA4 NULL=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     # TODO: support fake weights
     #- name: Run LLaMA 7B on 4 fake devices
     #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
@@ -307,10 +310,6 @@ jobs:
       run: python test/external/fuzz_symbolic.py
     - name: Fuzz Test fast idiv
       run: python test/external/fuzz_fast_idiv.py
-    - name: Fuzz Test shapetracker
-      run: CNT=50 python test/external/fuzz_shapetracker.py
-    - name: Fuzz Test shapetracker math
-      run: CNT=200 python test/external/fuzz_shapetracker_math.py
     - name: Fuzz Test shape ops
       run: python test/external/fuzz_shape_ops.py
 
@@ -352,7 +351,7 @@ 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:
@@ -375,17 +374,13 @@ 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
-      - name: Test openpilot compile4
-        run: NOLOCALS=1 CL=1 IMAGE=2 FLOAT16=1 DEBUG=2 python3 examples/openpilot/compile4.py
+          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
 
@@ -527,11 +522,11 @@ 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)
@@ -675,7 +670,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        backend: [llvm, cpu, opencl]
+        backend: [llvm, cpu, opencl, lvp]
 
     name: Linux (${{ matrix.backend }})
     runs-on: ubuntu-22.04
@@ -689,9 +684,10 @@ jobs:
           key: ${{ matrix.backend }}-minimal
           deps: testing_minimal
           opencl: ${{ matrix.backend == 'opencl' && 'true' }}
-          llvm: ${{ matrix.backend == 'llvm' && 'true' }}
+          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
+          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
       - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'CL=1' }}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'CL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
       - name: Check Device.DEFAULT and print some source
         run: |
           python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
@@ -893,7 +889,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        backend: [metal, llvm, cpu]
+        backend: [metal, llvm, cpu, lvp]
     name: MacOS (${{ matrix.backend }})
     runs-on: macos-15
     timeout-minutes: 20
@@ -906,12 +902,13 @@ jobs:
           key: macos-${{ matrix.backend }}-minimal
           deps: testing_minimal
           pydeps: "capstone"
-          llvm: ${{ matrix.backend == 'llvm' && 'true' }}
+          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
+          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
       - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'metal' && 'METAL=1'}}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'metal' && 'METAL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
       - name: Check Device.DEFAULT and print some source
         run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
+          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
           DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
       - name: Run pytest (${{ matrix.backend }})
         run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20

.gitignore

@@ -38,6 +38,7 @@ extra/huggingface_onnx/models/*
 extra/huggingface_onnx/*.yaml
 extra/weights
 venv
+venv_sd_mlperf
 examples/**/net.*[js,json]
 examples/**/*.safetensors
 node_modules

autogen_stubs.sh

@@ -461,6 +461,80 @@ generate_libusb() {
   python3 -c "import tinygrad.runtime.autogen.libusb"
 }
 
+generate_mesa() {
+  MESA_TAG="mesa-25.2.4"
+  MESA_SRC=/tmp/mesa-$MESA_TAG
+  TINYMESA_TAG=tinymesa-32dc66c
+  TINYMESA_DIR=/tmp/tinymesa-$MESA_TAG-$TINYMESA_TAG/
+  TINYMESA_SO=$TINYMESA_DIR/libtinymesa_cpu.so
+  if [ ! -d "$MESA_SRC" ]; then
+    git clone --depth 1 --branch $MESA_TAG https://gitlab.freedesktop.org/mesa/mesa.git $MESA_SRC
+    pushd .
+    cd $MESA_SRC
+    git reset --hard $MESA_COMMIT_HASH
+    # clang 14 doesn't support packed enums
+    sed -i "s/enum \w\+ \(\w\+\);$/uint8_t \1;/" $MESA_SRC/src/nouveau/headers/nv_device_info.h
+    sed -i "s/enum \w\+ \(\w\+\);$/uint8_t \1;/" $MESA_SRC/src/nouveau/compiler/nak.h
+    sed -i "s/nir_instr_type \(\w\+\);/uint8_t \1;/" $MESA_SRC/src/compiler/nir/nir.h
+    mkdir -p gen/util/format
+    python3 src/util/format/u_format_table.py src/util/format/u_format.yaml --enums > gen/util/format/u_format_gen.h
+    python3 src/compiler/nir/nir_opcodes_h.py > gen/nir_opcodes.h
+    python3 src/compiler/nir/nir_intrinsics_h.py --outdir gen
+    python3 src/compiler/nir/nir_intrinsics_indices_h.py --outdir gen
+    python3 src/compiler/nir/nir_builder_opcodes_h.py > gen/nir_builder_opcodes.h
+    python3 src/compiler/nir/nir_intrinsics_h.py --outdir gen
+    python3 src/compiler/builtin_types_h.py gen/builtin_types.h
+    popd
+  fi
+
+  if [ ! -d "$TINYMESA_DIR" ]; then
+    mkdir $TINYMESA_DIR
+    curl -L https://github.com/sirhcm/tinymesa/releases/download/$TINYMESA_TAG/libtinymesa_cpu-$MESA_TAG-linux-amd64.so -o $TINYMESA_SO
+  fi
+
+  clang2py -k cdefstu \
+    $MESA_SRC/src/compiler/nir/nir.h \
+    $MESA_SRC/src/compiler/nir/nir_builder.h \
+    $MESA_SRC/src/compiler/nir/nir_shader_compiler_options.h \
+    $MESA_SRC/src/compiler/nir/nir_serialize.h \
+    $MESA_SRC/gen/nir_intrinsics.h \
+    $MESA_SRC/src/nouveau/headers/nv_device_info.h \
+    $MESA_SRC/src/nouveau/compiler/nak.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_passmgr.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_misc.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_type.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_init.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_nir.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_struct.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_jit_types.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_flow.h \
+    $MESA_SRC/src/gallium/auxiliary/gallivm/lp_bld_const.h \
+    $MESA_SRC/src/compiler/glsl_types.h \
+    $MESA_SRC/src/util/blob.h \
+    $MESA_SRC/src/util/ralloc.h \
+    --clang-args="-DHAVE_ENDIAN_H -DHAVE_STRUCT_TIMESPEC -DHAVE_PTHREAD -I$MESA_SRC/src -I$MESA_SRC/include -I$MESA_SRC/gen -I$MESA_SRC/src/compiler/nir -I$MESA_SRC/src/gallium/auxiliary -I$MESA_SRC/src/gallium/include -I$(llvm-config-20 --includedir)" \
+    -l $TINYMESA_SO \
+    -o $BASE/mesa.py
+
+  LVP_NIR_OPTIONS=$(./extra/mesa/lvp_nir_options.sh $MESA_SRC)
+
+  fixup $BASE/mesa.py
+  patch_dlopen $BASE/mesa.py tinymesa_cpu "(BASE:=os.getenv('MESA_PATH', f\"/usr{'/local/' if helpers.OSX else '/'}lib\"))+'/libtinymesa_cpu'+(EXT:='.dylib' if helpers.OSX else '.so')" "f'{BASE}/libtinymesa{EXT}'" "'/opt/homebrew/lib/libtinymesa_cpu.dylib'" "'/opt/homebrew/lib/libtinymesa.dylib'"
+  echo "lvp_nir_options = gzip.decompress(base64.b64decode('$LVP_NIR_OPTIONS'))" >> $BASE/mesa.py
+  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())/" $BASE/mesa.py
+  echo "def __getattr__(nm): raise AttributeError('LLVMpipe requires tinymesa_cpu' if 'tinymesa_cpu' not in dll._name else f'attribute {nm} not found') if dll else FileNotFoundError(f'libtinymesa not found (MESA_PATH={BASE}). See https://github.com/sirhcm/tinymesa ($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
+  sed -i "s/('\(\w\+\)', pipe_shader_type, 8)/('\1', ctypes.c_ubyte)/" $BASE/mesa.py
+  sed -i "s/\([0-9]\+\)()/\1/" $BASE/mesa.py
+  sed -i "s/\(struct_nir_builder._pack_\) = 1/\1 = 0/" $BASE/mesa.py
+  python3 -c "import tinygrad.runtime.autogen.mesa"
+}
+
 if [ "$1" == "opencl" ]; then generate_opencl
 elif [ "$1" == "hip" ]; then generate_hip
 elif [ "$1" == "comgr" ]; then generate_comgr
@@ -484,6 +558,7 @@ elif [ "$1" == "pci" ]; then generate_pci
 elif [ "$1" == "vfio" ]; then generate_vfio
 elif [ "$1" == "webgpu" ]; then generate_webgpu
 elif [ "$1" == "libusb" ]; then generate_libusb
-elif [ "$1" == "all" ]; then generate_opencl; generate_hip; generate_comgr; generate_cuda; generate_nvrtc; generate_hsa; generate_kfd; generate_nv; generate_amd; generate_io_uring; generate_libc; generate_am; generate_webgpu
+elif [ "$1" == "mesa" ]; then generate_mesa
+elif [ "$1" == "all" ]; then generate_opencl; generate_hip; generate_comgr; generate_cuda; generate_nvrtc; generate_hsa; generate_kfd; generate_nv; generate_amd; generate_io_uring; generate_libc; generate_am; generate_webgpu; generate_mesa
 else echo "usage: $0 <type>"
 fi

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`.

examples/gpt2.py

@@ -232,7 +232,7 @@ if __name__ == "__main__":
   gpt2 = GPT2.build_gguf(args.model_size) if args.model_size.startswith("gpt2_gguf_") else GPT2.build(args.model_size)
 
   if args.benchmark != -1:
-    gpt2.model(Tensor.rand(args.batch_size, args.benchmark), Variable("a", 0, MAX_CONTEXT).bind(0)).realize()
+    gpt2.model(Tensor.randint(args.batch_size, args.benchmark), Variable("a", 0, MAX_CONTEXT).bind(0)).realize()
   else:
     texts = gpt2.generate(args.prompt, args.count, args.temperature, timing=args.timing, batch_size=args.batch_size)
     if not args.noshow:

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/openpilot/compile4.py

@@ -1,47 +0,0 @@
-import sys
-from tinygrad import Tensor, fetch, GlobalCounters, dtypes
-from tinygrad.uop.ops import UOp
-from tinygrad.nn.onnx import OnnxRunner
-from tinygrad.schedule.rangeify import get_rangeify_map
-from tinygrad.engine.schedule import create_schedule_with_vars
-from tinygrad.engine.realize import run_schedule
-
-# NOLOCALS=1 CL=1 IMAGE=2 FLOAT16=1 VIZ=1 DEBUG=2 python3 examples/openpilot/compile4.py
-
-OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
-OUTPUT = sys.argv[2] if len(sys.argv) > 2 else "/tmp/openpilot.pkl"
-
-if __name__ == "__main__":
-  onnx_file = fetch(OPENPILOT_MODEL)
-  run_onnx = OnnxRunner(onnx_file)
-
-  inputs = run_onnx.get_empty_input_data("npy", dtypes.float32)
-  out: Tensor = next(iter(run_onnx({k:v.to(None) for k,v in inputs.items()}).values())).to('cpu')
-  root = out.uop
-  targets = [x.uop for x in inputs.values()]
-  print(targets)
-
-  # TODO: abstract this from gradient?
-
-  # compute the target path (top down)
-  in_target_path: dict[UOp, bool] = {}
-  for u in root.toposort(): in_target_path[u] = any(x in targets or in_target_path[x] for x in u.src)
-  independent_set = {}
-  for u in root.toposort():
-    if in_target_path[u]:
-      for s in u.src:
-        if not in_target_path[s]:
-          independent_set[s] = None
-  independent = UOp.sink(*independent_set.keys())
-  kernelized = get_rangeify_map(independent)
-  independent = independent.substitute(kernelized)
-  schedule, var_vals = create_schedule_with_vars(independent)
-  run_schedule(schedule)
-
-  print("**** real ****")
-  GlobalCounters.reset()
-  out.uop = root.substitute(kernelized)
-  out.kernelize()
-
-  # realize
-  out.realize()

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

@@ -263,13 +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):
-    load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['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():

examples/test_onnx_imagenet.py

@@ -19,8 +19,8 @@ from tinygrad.helpers import fetch, getenv
 
 # QUANT=1 python3 examples/test_onnx_imagenet.py
 #   https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx
-# DONT_REALIZE_EXPAND=1 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
-# VIZ=1 DONT_REALIZE_EXPAND=1 python3 examples/benchmark_onnx.py /tmp/model.quant.onnx
+# python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
+# VIZ=1 python3 examples/benchmark_onnx.py /tmp/model.quant.onnx
 
 def imagenet_dataloader(cnt=0):
   input_mean = Tensor([0.485, 0.456, 0.406]).reshape(1, -1, 1, 1)

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/mesa/lvp_nir_options.sh

@@ -0,0 +1,23 @@
+#!/bin/sh
+
+if [ "$#" -ne 1 ] || ! [ -d $1 ]; then
+  echo "usage: $0 MESA_PREFIX"
+  exit 1
+fi
+
+TMP=$(mktemp)
+trap 'rm -f "$TMP"' EXIT
+
+(
+  cat <<EOF
+#define HAVE_ENDIAN_H
+#define HAVE_STRUCT_TIMESPEC
+#define HAVE_PTHREAD
+#include <unistd.h>
+#include "nir_shader_compiler_options.h"
+#include "compiler/shader_enums.h"
+EOF
+  sed -n '/struct nir_shader_compiler_options/,/^}/{p;/^}/q}' $1/src/gallium/drivers/llvmpipe/lp_screen.c
+  echo "int main(void) { write(1, &gallivm_nir_options, sizeof(gallivm_nir_options)); }"
+) | cc -x c -o $TMP - -I$1/src/compiler/nir -I$1/src -I$1/include && $TMP | gzip | base64 -w0
+

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/onnx_helpers.py

@@ -3,8 +3,21 @@ from tinygrad.tensor import _to_np_dtype
 from tinygrad.nn.onnx import OnnxRunner, OnnxValue
 import numpy as np
 import onnxruntime as ort
+ort_options = ort.SessionOptions()
+ort_options.log_severity_level = 3
 
 def get_example_inputs(graph_inputs:dict[str, OnnxValue], config={}):
+  """
+  Generate example input tensors based on the provided ONNX graph input specifications.
+
+  NOTE: This is not guaranteed to be reliable. It's a best-effort helper
+  that uses heuristics to guess input shapes and values.
+
+  Example:
+    from tinygrad.nn.onnx import OnnxRunner
+    from extra.onnx_helpers import get_example_inputs
+    inputs = get_example_inputs(OnnxRunner(model_path).graph_inputs)
+  """
   def _get_shape(onnx_shape: tuple[str|int]):
     shape = []
     for onnx_dim in onnx_shape:
@@ -44,11 +57,9 @@ def get_example_inputs(graph_inputs:dict[str, OnnxValue], config={}):
     ret.update({name:value})
   return ret
 
-def validate(onnx_file, inputs, rtol=1e-5, atol=1e-5):
+def _get_tinygrad_and_ort_np_outputs(onnx_file, inputs):
   run_onnx = OnnxRunner(onnx_file)
 
-  ort_options = ort.SessionOptions()
-  ort_options.log_severity_level = 3
   ort_sess = ort.InferenceSession(onnx_file, ort_options, ["CPUExecutionProvider"])
   np_inputs = {k:v.numpy() if isinstance(v, Tensor) else v for k,v in inputs.items()}
   out_names = list(run_onnx.graph_outputs)
@@ -56,9 +67,101 @@ def validate(onnx_file, inputs, rtol=1e-5, atol=1e-5):
   ort_out = dict(zip(out_names, out_values))
 
   tinygrad_out = run_onnx(inputs)
+  Tensor.realize(*(x for x in tinygrad_out.values() if x is not None))
+  tinygrad_out = {k:v.numpy() if v is not None else None for k,v in tinygrad_out.items()}
+  return tinygrad_out, ort_out
+
+def validate(onnx_file, inputs, rtol=1e-5, atol=1e-5):
+  """
+  Compares the final output tensors of an onnx model run in tinygrad and onnxruntime.
+  """
+  tinygrad_out, ort_out = _get_tinygrad_and_ort_np_outputs(onnx_file, inputs)
 
   assert tinygrad_out.keys() == ort_out.keys()
   for k in tinygrad_out.keys():
     tiny_v, onnx_v = tinygrad_out[k], ort_out[k]
     if tiny_v is None: assert onnx_v is None, f"{k}: {tiny_v=}, {onnx_v=}"
-    else: np.testing.assert_allclose(tiny_v.numpy(), onnx_v, rtol=rtol, atol=atol, err_msg=f"For tensor '{k}' in {tinygrad_out.keys()}")
+    else: np.testing.assert_allclose(tiny_v, onnx_v, rtol=rtol, atol=atol, err_msg=f"For tensor '{k}' in {tinygrad_out.keys()}")
+
+def validate_all_intermediates(onnx_file, inputs, rtol=1e-5, atol=1e-5):
+  """
+  Compares all intermediate node output of an onnx model run in tinygrad and onnxruntime.
+  """
+  report = generate_node_output_report(onnx_file, inputs)
+  for i, node in enumerate(report):
+    node_name = node["node"]
+    op = node["op"]
+    outputs = node["outputs"]
+    for output in outputs:
+      output_name = output["name"]
+      tinygrad_out = output["tinygrad"]
+      ort_out = output["onnxruntime"]
+      try:
+        if tinygrad_out is None: assert ort_out is None, f"None outputs are not equal {tinygrad_out=} {ort_out=}"
+        else: np.testing.assert_allclose(tinygrad_out, ort_out, rtol=rtol, atol=atol)
+        print(f"Validated {i}: {op=} {node_name=} {output_name=}")
+      except AssertionError as e:
+        print(f"FAILED {i}: {op=} {node_name=} {output_name=}")
+        print(str(e).strip() + "\n")
+
+def generate_node_output_report(onnx_file, inputs):
+  """
+  Build a report of all ONNX node outputs from tinygrad and onnxruntime
+
+  Returns:
+    A list of dictionaries, where each entry corresponds to one
+    node in the ONNX graph. The structure is as follows:
+    [
+      {
+        "node": str,  # The name of the ONNX node.
+        "op": str,    # The operation type of the ONNX node.
+        "outputs": [
+          {
+            "name": str,                       # The name of the output tensor.
+            "tinygrad": np.ndarray | None,     # The output value from tinygrad.
+            "onnxruntime": np.ndarray | None,  # The output value from onnxruntime.
+          },
+          ...
+        ]
+      },
+      ...
+    ]
+  """
+  import onnx_graphsurgeon as gs
+  import onnx
+  import tempfile
+
+  # rewrite the model to output all the node outputs
+  # `infer_shapes` here tries to fill the shapes and dtypes of intermediate values which graphsurgeon requires when assigning them as outputs
+  inferred_model = onnx.shape_inference.infer_shapes(onnx.load(onnx_file))
+  model = gs.import_onnx(inferred_model)
+  model_nodes = model.nodes
+  node_outputs = [n.outputs for n in model.nodes]
+  model.outputs = [
+      each_output for outputs in node_outputs for each_output in outputs
+      if not (each_output.dtype is None and each_output.shape is None)  # output with None dtype and None shape is likely a `None` value
+  ]
+  rewritten_model = gs.export_onnx(model)
+
+  # TODO: remove this once ORT supports 1.18.0
+  if getattr(rewritten_model, "ir_version", 0) > 10:
+    rewritten_model.ir_version = 10
+
+  with tempfile.NamedTemporaryFile(suffix=".onnx") as f:
+    onnx.save(rewritten_model, f.name)
+    rewritten_model_path = f.name
+    tinygrad_out, ort_out = _get_tinygrad_and_ort_np_outputs(rewritten_model_path, inputs)
+
+  report = []
+  for node in model_nodes:
+    outputs = []
+    for each_output in node.outputs:
+      if each_output.dtype is None and each_output.shape is None:
+        continue
+      name = each_output.name
+      tinygrad_output = tinygrad_out[name]
+      ort_output = ort_out[name]
+      outputs.append({"name": name, "tinygrad": tinygrad_output, "onnxruntime": ort_output})
+    report.append({"node": node.name, "op": node.op, "outputs": outputs})
+
+  return report

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/tinyfs/fetch_file.py

@@ -0,0 +1,11 @@
+from tinygrad.tensor import Tensor
+import argparse
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser()
+  parser.add_argument("hash", type=str, required=True, help="file hash to fetch")
+  parser.add_argument("len", type=int, required=True, help="file length to fetch")
+  parser.add_argument("dest", type=str, required=True, help="destination path to save the file")
+  args = parser.parse_args()
+
+  Tensor(bytes.fromhex(args.hash), device="CPU").load(args.len).to(f"disk:{args.dest}").realize()

extra/tinyfs/fetch_raid.py

@@ -0,0 +1,39 @@
+import json, multiprocessing
+from pathlib import Path
+
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import tqdm, getenv
+
+raid_root = Path(getenv("RAID_ROOT", "/raid"))
+
+def fetch_file(item):
+  path, info = item
+  h, size = info["hash"], info["size"]
+
+  path = raid_root / Path(path)
+  path.parent.mkdir(parents=True, exist_ok=True)
+
+  try:
+    pt = Tensor(bytes.fromhex(h), device="CPU").load(size).to(f"disk:{path.as_posix()}").realize()
+  except Exception as e:
+    print(f"error fetching {path}, {h}, {size}: {e}")
+    raise
+
+  pt.uop.buffer.deallocate()
+
+def fetch_mapping():
+  mapping_tensor = Tensor(bytes.fromhex("d734f5e3be9f1e9d863bfaa4fc6c1ef2")).load(175866113).realize()
+  mapping = mapping_tensor.data().tobytes().decode()
+  mapping = json.loads(mapping)
+  mapped_files = mapping.items()
+  return list(mapped_files)
+
+if __name__ == "__main__":
+  with multiprocessing.Pool(processes=1) as pool:
+    mapped_files = pool.apply(fetch_mapping)
+
+  print(f"fetched mapping for {len(mapped_files)} files")
+
+  with multiprocessing.Pool(processes=multiprocessing.cpu_count()) as pool:
+    for _ in tqdm(pool.imap_unordered(fetch_file, mapped_files), total=len(mapped_files)):
+      pass

extra/tinyfs/upload_raid.py

@@ -0,0 +1,31 @@
+from pathlib import Path
+import multiprocessing, json
+
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import tqdm
+
+raid_root = Path("/raid")
+
+def upload_file(path: Path):
+  pt = Tensor(path).realize()
+  h = pt.store().realize()
+  pt.uop.realized.deallocate()
+  return h.data().hex(), path, pt.nbytes()
+
+if __name__ == "__main__":
+  raid_files = sorted([p for p in raid_root.rglob("*") if p.is_file()])
+  print(f"found {len(raid_files)} files in /raid")
+
+  mapping = {}
+  with multiprocessing.Pool(processes=multiprocessing.cpu_count()) as pool:
+    for h, p, s in tqdm(pool.imap_unordered(upload_file, raid_files), total=len(raid_files)):
+      mapping[p.relative_to(raid_root).as_posix()] = {"hash": h, "size": s}
+
+  # sort the mapping by key
+  mapping = dict(sorted(mapping.items()))
+
+  mapping = json.dumps(mapping).encode()
+  mapping_tensor = Tensor(mapping, device="CPU")
+  h = mapping_tensor.store().realize()
+
+  print(f"final hash: {h.data().hex()}, size: {len(mapping)}")

extra/torch_backend/backend.py

@@ -642,10 +642,11 @@ def get_real_tinygrad_buffers():
 torch.nn.modules.module.register_module_buffer_registration_hook(register_torch_buffer)
 
 from torch.nn.modules import Module
-def backward_hook(model:Module, _grad_input, _grad_out):
-  grads_to_realize = [unwrap(p.grad) for p in model.parameters() if p.grad is not None]
-  if len(grads_to_realize): Tensor.realize(*grads_to_realize)
-def module_hook(module:Module, _name, _submodule): module.register_backward_hook(backward_hook)
+def param_hook(_grad):
+  if _grad is not None and _grad.is_tiny: Tensor.realize(unwrap(_grad))
+def module_hook(module:Module, _name, _submodule):
+  for param in _submodule.parameters(recurse=False):
+    if param.requires_grad: param.register_hook(param_hook)
 torch.nn.modules.module.register_module_module_registration_hook(module_hook)
 
 def realize_optimizer_step(optimizer: torch.optim.Optimizer, *args, **kwargs):

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/.gitignore

@@ -0,0 +1,13 @@
+xcuserdata/
+
+**/*.xcodeproj/project.xcworkspace/*
+!**/*.xcodeproj/project.xcworkspace/xcshareddata
+
+**/*.xcodeproj/project.xcworkspace/xcshareddata/*
+!**/*.xcodeproj/project.xcworkspace/xcshareddata/WorkspaceSettings.xcsettings
+
+**/*.playground/playground.xcworkspace/*
+!**/*.playground/playground.xcworkspace/xcshareddata
+
+**/*.playground/playground.xcworkspace/xcshareddata/*
+!**/*.playground/playground.xcworkspace/xcshareddata/WorkspaceSettings.xcsettings

extra/usbgpu/tbgpu/installer/Shared/Assets.xcassets/AccentColor.colorset/Contents.json

@@ -0,0 +1,11 @@
+{
+  "colors" : [
+    {
+      "idiom" : "universal"
+    }
+  ],
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

extra/usbgpu/tbgpu/installer/Shared/Assets.xcassets/AppIcon.appiconset/Contents.json

@@ -0,0 +1,148 @@
+{
+  "images" : [
+    {
+      "idiom" : "iphone",
+      "scale" : "2x",
+      "size" : "20x20"
+    },
+    {
+      "idiom" : "iphone",
+      "scale" : "3x",
+      "size" : "20x20"
+    },
+    {
+      "idiom" : "iphone",
+      "scale" : "2x",
+      "size" : "29x29"
+    },
+    {
+      "idiom" : "iphone",
+      "scale" : "3x",
+      "size" : "29x29"
+    },
+    {
+      "idiom" : "iphone",
+      "scale" : "2x",
+      "size" : "40x40"
+    },
+    {
+      "idiom" : "iphone",
+      "scale" : "3x",
+      "size" : "40x40"
+    },
+    {
+      "idiom" : "iphone",
+      "scale" : "2x",
+      "size" : "60x60"
+    },
+    {
+      "idiom" : "iphone",
+      "scale" : "3x",
+      "size" : "60x60"
+    },
+    {
+      "idiom" : "ipad",
+      "scale" : "1x",
+      "size" : "20x20"
+    },
+    {
+      "idiom" : "ipad",
+      "scale" : "2x",
+      "size" : "20x20"
+    },
+    {
+      "idiom" : "ipad",
+      "scale" : "1x",
+      "size" : "29x29"
+    },
+    {
+      "idiom" : "ipad",
+      "scale" : "2x",
+      "size" : "29x29"
+    },
+    {
+      "idiom" : "ipad",
+      "scale" : "1x",
+      "size" : "40x40"
+    },
+    {
+      "idiom" : "ipad",
+      "scale" : "2x",
+      "size" : "40x40"
+    },
+    {
+      "idiom" : "ipad",
+      "scale" : "1x",
+      "size" : "76x76"
+    },
+    {
+      "idiom" : "ipad",
+      "scale" : "2x",
+      "size" : "76x76"
+    },
+    {
+      "idiom" : "ipad",
+      "scale" : "2x",
+      "size" : "83.5x83.5"
+    },
+    {
+      "idiom" : "ios-marketing",
+      "scale" : "1x",
+      "size" : "1024x1024"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "16x16"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "16x16"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "32x32"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "32x32"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "128x128"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "128x128"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "256x256"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "256x256"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "1x",
+      "size" : "512x512"
+    },
+    {
+      "idiom" : "mac",
+      "scale" : "2x",
+      "size" : "512x512"
+    }
+  ],
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

extra/usbgpu/tbgpu/installer/Shared/Assets.xcassets/Contents.json

@@ -0,0 +1,6 @@
+{
+  "info" : {
+    "author" : "xcode",
+    "version" : 1
+  }
+}

extra/usbgpu/tbgpu/installer/Shared/TinyGPUApp.swift

@@ -0,0 +1,19 @@
+import AppKit
+import SwiftUI
+
+final class AppDelegate: NSObject, NSApplicationDelegate {
+	func applicationShouldTerminateAfterLastWindowClosed(_ sender: NSApplication) -> Bool {
+		true
+	}
+}
+
+@main
+struct TinyGPUApp: App {
+	@NSApplicationDelegateAdaptor(AppDelegate.self) var appDelegate
+
+    var body: some Scene {
+        WindowGroup {
+			TinyGPUView()
+        }
+    }
+}

extra/usbgpu/tbgpu/installer/Shared/TinyGPUView.swift

@@ -0,0 +1,33 @@
+import SwiftUI
+
+struct TinyGPUView: View {
+	@ObservedObject var viewModel = TinyGPUViewModel()
+
+	var body: some View {
+#if os(macOS)
+		VStack(alignment: .center) {
+			Text("TinyGPU Intsaller")
+				.padding()
+				.font(.title)
+			Text(self.viewModel.dextLoadingState)
+				.multilineTextAlignment(.center)
+			HStack {
+				Button(
+					action: {
+						self.viewModel.activateMyDext()
+					}, label: {
+						Text("Install extension")
+					}
+				)
+			}
+		}
+		.frame(width: 500, height: 200, alignment: .center)
+#endif
+	}
+}
+
+struct TinyGPUView_Previews: PreviewProvider {
+    static var previews: some View {
+		TinyGPUView()
+    }
+}

extra/usbgpu/tbgpu/installer/Shared/TinyGPUViewModel.swift

@@ -0,0 +1,149 @@
+import Foundation
+import os.log
+import SystemExtensions
+
+class TinyGPUDriverLoadingStateMachine {
+  enum State { case unloaded, activating, needsApproval, activated, activationError }
+}
+
+class TinyGPUViewModel: NSObject {
+
+	@Published private var state: TinyGPUDriverLoadingStateMachine.State = .unloaded
+	
+	override init() {
+		super.init()
+		refreshInitialDextState()
+	}
+	
+	private func refreshInitialDextState() {
+#if os(macOS)
+		Task.detached { [dextIdentifier] in
+		  let newState = Self.queryDextState(bundleID: dextIdentifier)
+		  await MainActor.run { self.state = newState }
+		}
+#endif
+	}
+
+#if os(macOS)
+	private static func queryDextState(bundleID: String) -> TinyGPUDriverLoadingStateMachine.State {
+		let tool = "/usr/bin/systemextensionsctl"
+		let p = Process()
+		p.executableURL = URL(fileURLWithPath: tool)
+		p.arguments = ["list"]
+
+		let pipe = Pipe()
+		p.standardOutput = pipe
+		p.standardError = Pipe()
+
+		do {
+			try p.run()
+			p.waitUntilExit()
+			let data = pipe.fileHandleForReading.readDataToEndOfFile()
+			guard let output = String(data: data, encoding: .utf8) else { return .unloaded }
+
+			// Look for our bundle id line
+			if let line = output.split(separator: "\n").first(where: { $0.contains(bundleID) }) {
+				if line.contains("[activated enabled]") { return .activated }
+				if line.contains("[activated waiting for user]") { return .needsApproval }
+				if line.contains("terminated waiting to uninstall") { return .unloaded }
+				return .activating
+			} else {
+				return .unloaded
+			}
+		} catch {
+			return .unloaded
+		}
+	}
+#endif
+
+	private let dextIdentifier: String = "org.tinygrad.tinygpu.edriver"
+
+	public var dextLoadingState: String {
+		switch state {
+		case .unloaded:
+			return "TinyGPUDriver isn't loaded."
+		case .activating:
+			return "Activating TinyGPUDriver, please wait."
+		case .needsApproval:
+			return "Please follow the prompt to approve TinyGPUDriver."
+		case .activated:
+			return "TinyGPUDriver has been activated and is ready to use. You can close the installer."
+		case .activationError:
+			return "TinyGPUDriver has experienced an error during activation.\nPlease check the logs to find the error."
+		}
+	}
+}
+
+extension TinyGPUViewModel: ObservableObject {
+
+#if os(macOS)
+	func activateMyDext() {
+		activateExtension(dextIdentifier)
+	}
+	
+	func deactivateMyDext() {
+		deactivateExtension(dextIdentifier)
+	}
+
+	func activateExtension(_ dextIdentifier: String) {
+
+		let request = OSSystemExtensionRequest
+			.activationRequest(forExtensionWithIdentifier: dextIdentifier,
+							   queue: .main)
+		request.delegate = self
+		OSSystemExtensionManager.shared.submitRequest(request)
+
+		self.state = .activating
+	}
+	
+	func deactivateExtension(_ dextIdentifier: String) {
+
+		let request = OSSystemExtensionRequest.deactivationRequest(forExtensionWithIdentifier: dextIdentifier, queue: .main)
+		request.delegate = self
+		OSSystemExtensionManager.shared.submitRequest(request)
+
+		self.state = .unloaded
+	}
+#endif
+}
+
+#if os(macOS)
+extension TinyGPUViewModel: OSSystemExtensionRequestDelegate {
+
+	func request(
+		_ request: OSSystemExtensionRequest,
+		actionForReplacingExtension existing: OSSystemExtensionProperties,
+		withExtension ext: OSSystemExtensionProperties) -> OSSystemExtensionRequest.ReplacementAction {
+
+		var replacementAction: OSSystemExtensionRequest.ReplacementAction
+
+		os_log("sysex actionForReplacingExtension: %@ %@", existing, ext)
+
+		// Add appropriate logic here to determine whether to replace the extension
+		// with the new extension. Common things to check for include
+		// testing whether the new extension's version number is newer than
+		// the current version number, or whether the bundleIdentifier is different.
+		// For simplicity, this sample always replaces the current extension
+		// with the new one.
+		replacementAction = .replace
+
+		self.state = .activating
+		return replacementAction
+	}
+
+	func requestNeedsUserApproval(_ request: OSSystemExtensionRequest) {
+		os_log("sysex requestNeedsUserApproval")
+		self.state = .needsApproval
+	}
+
+	func request(_ request: OSSystemExtensionRequest, didFinishWithResult result: OSSystemExtensionRequest.Result) {
+		os_log("sysex didFinishWithResult: %d", result.rawValue)
+		self.state = .activated
+	}
+
+	func request(_ request: OSSystemExtensionRequest, didFailWithError error: Error) {
+		os_log("sysex didFailWithError: %@", error.localizedDescription)
+		self.state = .activationError
+	}
+}
+#endif

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

@@ -0,0 +1,589 @@
+// !$*UTF8*$!
+{
+	archiveVersion = 1;
+	classes = {
+	};
+	objectVersion = 54;
+	objects = {
+
+/* Begin PBXBuildFile section */
+		0ACB55392E9CB880007029EF /* PCIDriverKit.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 0ACB55382E9CB880007029EF /* PCIDriverKit.framework */; };
+		54798269286A3512009785F6 /* CoreAudio.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 54798268286A3512009785F6 /* CoreAudio.framework */; };
+		549EB121286A1A37009D38AB /* TinyGPUViewModel.swift in Sources */ = {isa = PBXBuildFile; fileRef = 549EB11F286A1A37009D38AB /* TinyGPUViewModel.swift */; };
+		549EB123286A1D48009D38AB /* org.tinygrad.tinygpu.edriver.dext in Embed System Extensions */ = {isa = PBXBuildFile; fileRef = C5B7D9BC26128AC50089B4C3 /* org.tinygrad.tinygpu.edriver.dext */; settings = {ATTRIBUTES = (RemoveHeadersOnCopy, ); }; };
+		549EB131286A2B98009D38AB /* IOKit.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549EB130286A2B98009D38AB /* IOKit.framework */; };
+		54E42BC8286A1697000E1E9A /* TinyGPUApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 54E42BB8286A1696000E1E9A /* TinyGPUApp.swift */; };
+		54E42BCA286A1697000E1E9A /* TinyGPUView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 54E42BB9286A1696000E1E9A /* TinyGPUView.swift */; };
+		54E42BCC286A1697000E1E9A /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 54E42BBA286A1697000E1E9A /* Assets.xcassets */; };
+		C5B7D9C326128AC50089B4C3 /* TinyGPUDriver.cpp in Sources */ = {isa = PBXBuildFile; fileRef = C5B7D9C226128AC50089B4C3 /* TinyGPUDriver.cpp */; };
+		C5B7D9C526128AC50089B4C3 /* TinyGPUDriver.iig in Sources */ = {isa = PBXBuildFile; fileRef = C5B7D9C426128AC50089B4C3 /* TinyGPUDriver.iig */; };
+		C5C3BBB32612ACDC003C7BFE /* AudioDriverKit.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = C5C3BBB12612ACD3003C7BFE /* AudioDriverKit.framework */; };
+		C5C3BBB52612ACEF003C7BFE /* DriverKit.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = C5C3BBB42612ACEF003C7BFE /* DriverKit.framework */; };
+		C5D787AC261667FC006047E5 /* TinyGPUDriverUserClient.iig in Sources */ = {isa = PBXBuildFile; fileRef = C5D787AB261667FC006047E5 /* TinyGPUDriverUserClient.iig */; };
+		C5D787AE26168E59006047E5 /* TinyGPUDriverUserClient.cpp in Sources */ = {isa = PBXBuildFile; fileRef = C5D787AD26168D1E006047E5 /* TinyGPUDriverUserClient.cpp */; };
+/* End PBXBuildFile section */
+
+/* Begin PBXContainerItemProxy section */
+		549EB126286A1D66009D38AB /* PBXContainerItemProxy */ = {
+			isa = PBXContainerItemProxy;
+			containerPortal = C5B7D9B326128AC50089B4C3 /* Project object */;
+			proxyType = 1;
+			remoteGlobalIDString = C5B7D9BB26128AC50089B4C3;
+			remoteInfo = SimpleAudioDriver;
+		};
+/* End PBXContainerItemProxy section */
+
+/* Begin PBXCopyFilesBuildPhase section */
+		549EB122286A1D3A009D38AB /* Embed System Extensions */ = {
+			isa = PBXCopyFilesBuildPhase;
+			buildActionMask = 2147483647;
+			dstPath = "$(SYSTEM_EXTENSIONS_FOLDER_PATH)";
+			dstSubfolderSpec = 16;
+			files = (
+				549EB123286A1D48009D38AB /* org.tinygrad.tinygpu.edriver.dext in Embed System Extensions */,
+			);
+			name = "Embed System Extensions";
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXCopyFilesBuildPhase section */
+
+/* Begin PBXFileReference section */
+		0ACB55382E9CB880007029EF /* PCIDriverKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = PCIDriverKit.framework; path = System/DriverKit/System/Library/Frameworks/PCIDriverKit.framework; sourceTree = SDKROOT; };
+		54798268286A3512009785F6 /* CoreAudio.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = CoreAudio.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX13.0.sdk/System/Library/Frameworks/CoreAudio.framework; sourceTree = DEVELOPER_DIR; };
+		549EB11F286A1A37009D38AB /* TinyGPUViewModel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TinyGPUViewModel.swift; sourceTree = "<group>"; usesTabs = 1; };
+		549EB130286A2B98009D38AB /* IOKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = IOKit.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX13.0.sdk/System/Library/Frameworks/IOKit.framework; sourceTree = DEVELOPER_DIR; };
+		549EB132286A2B9D009D38AB /* IOKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = IOKit.framework; path = Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS16.0.sdk/System/Library/Frameworks/IOKit.framework; sourceTree = DEVELOPER_DIR; };
+		54E42BB8286A1696000E1E9A /* TinyGPUApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TinyGPUApp.swift; sourceTree = "<group>"; };
+		54E42BB9286A1696000E1E9A /* TinyGPUView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TinyGPUView.swift; sourceTree = "<group>"; };
+		54E42BBA286A1697000E1E9A /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
+		54E42BC4286A1697000E1E9A /* TinyGPU.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = TinyGPU.app; sourceTree = BUILT_PRODUCTS_DIR; };
+		54E42BC6286A1697000E1E9A /* macOS.entitlements */ = {isa = PBXFileReference; lastKnownFileType = text.plist.entitlements; path = macOS.entitlements; sourceTree = "<group>"; };
+		C5B7D9BC26128AC50089B4C3 /* org.tinygrad.tinygpu.edriver.dext */ = {isa = PBXFileReference; explicitFileType = "wrapper.driver-extension"; includeInIndex = 0; path = org.tinygrad.tinygpu.edriver.dext; sourceTree = BUILT_PRODUCTS_DIR; };
+		C5B7D9BF26128AC50089B4C3 /* DriverKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = DriverKit.framework; path = Library/Frameworks/DriverKit.framework; sourceTree = DEVELOPER_DIR; };
+		C5B7D9C226128AC50089B4C3 /* TinyGPUDriver.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; path = TinyGPUDriver.cpp; sourceTree = "<group>"; usesTabs = 1; };
+		C5B7D9C426128AC50089B4C3 /* TinyGPUDriver.iig */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.iig; path = TinyGPUDriver.iig; sourceTree = "<group>"; };
+		C5B7D9C626128AC50089B4C3 /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = "<group>"; };
+		C5B7D9CC26128ADA0089B4C3 /* AudioDriverKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = AudioDriverKit.framework; path = System/DriverKit/System/Library/Frameworks/AudioDriverKit.framework; sourceTree = SDKROOT; };
+		C5B7D9CE26128B150089B4C3 /* TinyGPUDriver.entitlements */ = {isa = PBXFileReference; lastKnownFileType = text.plist.entitlements; path = TinyGPUDriver.entitlements; sourceTree = "<group>"; };
+		C5C0063326178F98003345D8 /* AppKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = AppKit.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX12.0.sdk/System/Library/Frameworks/AppKit.framework; sourceTree = DEVELOPER_DIR; };
+		C5C006352617ACB8003345D8 /* CoreAudio.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = CoreAudio.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX12.0.sdk/System/Library/Frameworks/CoreAudio.framework; sourceTree = DEVELOPER_DIR; };
+		C5C3BBB12612ACD3003C7BFE /* AudioDriverKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = AudioDriverKit.framework; path = Platforms/DriverKit.platform/Developer/SDKs/DriverKit.MacOSX21.0.Internal.sdk/System/DriverKit/System/Library/Frameworks/AudioDriverKit.framework; sourceTree = DEVELOPER_DIR; };
+		C5C3BBB42612ACEF003C7BFE /* DriverKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = DriverKit.framework; path = Platforms/DriverKit.platform/Developer/SDKs/DriverKit.MacOSX21.0.Internal.sdk/System/DriverKit/System/Library/Frameworks/DriverKit.framework; sourceTree = DEVELOPER_DIR; };
+		C5D787AB261667FC006047E5 /* TinyGPUDriverUserClient.iig */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.iig; path = TinyGPUDriverUserClient.iig; sourceTree = "<group>"; };
+		C5D787AD26168D1E006047E5 /* TinyGPUDriverUserClient.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; path = TinyGPUDriverUserClient.cpp; sourceTree = "<group>"; };
+		C5D787B026169723006047E5 /* IOKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = IOKit.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX12.0.sdk/System/Library/Frameworks/IOKit.framework; sourceTree = DEVELOPER_DIR; };
+		C5D787B22616973F006047E5 /* SystemExtensions.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = SystemExtensions.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX12.0.sdk/System/Library/Frameworks/SystemExtensions.framework; sourceTree = DEVELOPER_DIR; };
+		C5D787B426169747006047E5 /* Foundation.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Foundation.framework; path = Platforms/MacOSX.platform/Developer/SDKs/MacOSX12.0.sdk/System/Library/Frameworks/Foundation.framework; sourceTree = DEVELOPER_DIR; };
+/* End PBXFileReference section */
+
+/* Begin PBXFrameworksBuildPhase section */
+		54E42BC1286A1697000E1E9A /* Frameworks */ = {
+			isa = PBXFrameworksBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				54798269286A3512009785F6 /* CoreAudio.framework in Frameworks */,
+				549EB131286A2B98009D38AB /* IOKit.framework in Frameworks */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+		C5B7D9B926128AC50089B4C3 /* Frameworks */ = {
+			isa = PBXFrameworksBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				C5C3BBB32612ACDC003C7BFE /* AudioDriverKit.framework in Frameworks */,
+				C5C3BBB52612ACEF003C7BFE /* DriverKit.framework in Frameworks */,
+				0ACB55392E9CB880007029EF /* PCIDriverKit.framework in Frameworks */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXFrameworksBuildPhase section */
+
+/* Begin PBXGroup section */
+		54E42BB7286A1696000E1E9A /* Shared */ = {
+			isa = PBXGroup;
+			children = (
+				54E42BB8286A1696000E1E9A /* TinyGPUApp.swift */,
+				54E42BB9286A1696000E1E9A /* TinyGPUView.swift */,
+				549EB11F286A1A37009D38AB /* TinyGPUViewModel.swift */,
+				54E42BBA286A1697000E1E9A /* Assets.xcassets */,
+			);
+			path = Shared;
+			sourceTree = "<group>";
+		};
+		54E42BC5286A1697000E1E9A /* macOS */ = {
+			isa = PBXGroup;
+			children = (
+				54E42BC6286A1697000E1E9A /* macOS.entitlements */,
+			);
+			path = macOS;
+			sourceTree = "<group>";
+		};
+		C5B7D9B226128AC50089B4C3 = {
+			isa = PBXGroup;
+			children = (
+				C5B7D9C126128AC50089B4C3 /* TinyGPUDriverExtension */,
+				54E42BB7286A1696000E1E9A /* Shared */,
+				54E42BC5286A1697000E1E9A /* macOS */,
+				C5B7D9BE26128AC50089B4C3 /* Frameworks */,
+				C5B7D9BD26128AC50089B4C3 /* Products */,
+			);
+			sourceTree = "<group>";
+			usesTabs = 1;
+		};
+		C5B7D9BD26128AC50089B4C3 /* Products */ = {
+			isa = PBXGroup;
+			children = (
+				C5B7D9BC26128AC50089B4C3 /* org.tinygrad.tinygpu.edriver.dext */,
+				54E42BC4286A1697000E1E9A /* TinyGPU.app */,
+			);
+			name = Products;
+			sourceTree = "<group>";
+		};
+		C5B7D9BE26128AC50089B4C3 /* Frameworks */ = {
+			isa = PBXGroup;
+			children = (
+				0ACB55382E9CB880007029EF /* PCIDriverKit.framework */,
+				54798268286A3512009785F6 /* CoreAudio.framework */,
+				549EB130286A2B98009D38AB /* IOKit.framework */,
+				549EB132286A2B9D009D38AB /* IOKit.framework */,
+				C5C006352617ACB8003345D8 /* CoreAudio.framework */,
+				C5C0063326178F98003345D8 /* AppKit.framework */,
+				C5D787B426169747006047E5 /* Foundation.framework */,
+				C5D787B22616973F006047E5 /* SystemExtensions.framework */,
+				C5D787B026169723006047E5 /* IOKit.framework */,
+				C5C3BBB42612ACEF003C7BFE /* DriverKit.framework */,
+				C5B7D9CC26128ADA0089B4C3 /* AudioDriverKit.framework */,
+				C5C3BBB12612ACD3003C7BFE /* AudioDriverKit.framework */,
+				C5B7D9BF26128AC50089B4C3 /* DriverKit.framework */,
+			);
+			name = Frameworks;
+			sourceTree = "<group>";
+		};
+		C5B7D9C126128AC50089B4C3 /* TinyGPUDriverExtension */ = {
+			isa = PBXGroup;
+			children = (
+				C5B7D9C226128AC50089B4C3 /* TinyGPUDriver.cpp */,
+				C5B7D9C426128AC50089B4C3 /* TinyGPUDriver.iig */,
+				C5D787AD26168D1E006047E5 /* TinyGPUDriverUserClient.cpp */,
+				C5D787AB261667FC006047E5 /* TinyGPUDriverUserClient.iig */,
+				C5B7D9C626128AC50089B4C3 /* Info.plist */,
+				C5B7D9CE26128B150089B4C3 /* TinyGPUDriver.entitlements */,
+			);
+			path = TinyGPUDriverExtension;
+			sourceTree = "<group>";
+		};
+/* End PBXGroup section */
+
+/* Begin PBXHeadersBuildPhase section */
+		C5B7D9B726128AC50089B4C3 /* Headers */ = {
+			isa = PBXHeadersBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXHeadersBuildPhase section */
+
+/* Begin PBXNativeTarget section */
+		54E42BC3286A1697000E1E9A /* TinyGPU Installer (macOS) */ = {
+			isa = PBXNativeTarget;
+			buildConfigurationList = 54E42BD2286A1697000E1E9A /* Build configuration list for PBXNativeTarget "TinyGPU Installer (macOS)" */;
+			buildPhases = (
+				54E42BC0286A1697000E1E9A /* Sources */,
+				54E42BC1286A1697000E1E9A /* Frameworks */,
+				54E42BC2286A1697000E1E9A /* Resources */,
+				549EB122286A1D3A009D38AB /* Embed System Extensions */,
+			);
+			buildRules = (
+			);
+			dependencies = (
+				549EB127286A1D66009D38AB /* PBXTargetDependency */,
+			);
+			name = "TinyGPU Installer (macOS)";
+			productName = "SimpleAudioDriverExtension2 (macOS)";
+			productReference = 54E42BC4286A1697000E1E9A /* TinyGPU.app */;
+			productType = "com.apple.product-type.application";
+		};
+		C5B7D9BB26128AC50089B4C3 /* TinyGPUDriver */ = {
+			isa = PBXNativeTarget;
+			buildConfigurationList = C5B7D9C926128AC50089B4C3 /* Build configuration list for PBXNativeTarget "TinyGPUDriver" */;
+			buildPhases = (
+				C5B7D9B726128AC50089B4C3 /* Headers */,
+				C5B7D9B826128AC50089B4C3 /* Sources */,
+				C5B7D9B926128AC50089B4C3 /* Frameworks */,
+				C5B7D9BA26128AC50089B4C3 /* Resources */,
+			);
+			buildRules = (
+			);
+			dependencies = (
+			);
+			name = TinyGPUDriver;
+			productName = SimpleAudioDriverExtension;
+			productReference = C5B7D9BC26128AC50089B4C3 /* org.tinygrad.tinygpu.edriver.dext */;
+			productType = "com.apple.product-type.driver-extension";
+		};
+/* End PBXNativeTarget section */
+
+/* Begin PBXProject section */
+		C5B7D9B326128AC50089B4C3 /* Project object */ = {
+			isa = PBXProject;
+			attributes = {
+				BuildIndependentTargetsInParallel = YES;
+				DefaultBuildSystemTypeForWorkspace = Latest;
+				LastSwiftUpdateCheck = 1400;
+				LastUpgradeCheck = 1600;
+				ORGANIZATIONNAME = Apple;
+				TargetAttributes = {
+					54E42BC3286A1697000E1E9A = {
+						CreatedOnToolsVersion = 14.0;
+						LastSwiftMigration = 1400;
+					};
+					C5B7D9BB26128AC50089B4C3 = {
+						CreatedOnToolsVersion = 13.0;
+					};
+				};
+			};
+			buildConfigurationList = C5B7D9B626128AC50089B4C3 /* Build configuration list for PBXProject "TinyGPUDriverExtension" */;
+			compatibilityVersion = "Xcode 9.3";
+			developmentRegion = en;
+			hasScannedForEncodings = 0;
+			knownRegions = (
+				en,
+				Base,
+			);
+			mainGroup = C5B7D9B226128AC50089B4C3;
+			productRefGroup = C5B7D9BD26128AC50089B4C3 /* Products */;
+			projectDirPath = "";
+			projectRoot = "";
+			targets = (
+				C5B7D9BB26128AC50089B4C3 /* TinyGPUDriver */,
+				54E42BC3286A1697000E1E9A /* TinyGPU Installer (macOS) */,
+			);
+		};
+/* End PBXProject section */
+
+/* Begin PBXResourcesBuildPhase section */
+		54E42BC2286A1697000E1E9A /* Resources */ = {
+			isa = PBXResourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				54E42BCC286A1697000E1E9A /* Assets.xcassets in Resources */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+		C5B7D9BA26128AC50089B4C3 /* Resources */ = {
+			isa = PBXResourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXResourcesBuildPhase section */
+
+/* Begin PBXSourcesBuildPhase section */
+		54E42BC0286A1697000E1E9A /* Sources */ = {
+			isa = PBXSourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				549EB121286A1A37009D38AB /* TinyGPUViewModel.swift in Sources */,
+				54E42BCA286A1697000E1E9A /* TinyGPUView.swift in Sources */,
+				54E42BC8286A1697000E1E9A /* TinyGPUApp.swift in Sources */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+		C5B7D9B826128AC50089B4C3 /* Sources */ = {
+			isa = PBXSourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				C5B7D9C526128AC50089B4C3 /* TinyGPUDriver.iig in Sources */,
+				C5D787AE26168E59006047E5 /* TinyGPUDriverUserClient.cpp in Sources */,
+				C5D787AC261667FC006047E5 /* TinyGPUDriverUserClient.iig in Sources */,
+				C5B7D9C326128AC50089B4C3 /* TinyGPUDriver.cpp in Sources */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXSourcesBuildPhase section */
+
+/* Begin PBXTargetDependency section */
+		549EB127286A1D66009D38AB /* PBXTargetDependency */ = {
+			isa = PBXTargetDependency;
+			target = C5B7D9BB26128AC50089B4C3 /* TinyGPUDriver */;
+			targetProxy = 549EB126286A1D66009D38AB /* PBXContainerItemProxy */;
+		};
+/* End PBXTargetDependency section */
+
+/* Begin XCBuildConfiguration section */
+		54E42BCF286A1697000E1E9A /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+				CLANG_ENABLE_MODULES = YES;
+				CODE_SIGN_ENTITLEMENTS = macOS/macOS.entitlements;
+				CODE_SIGN_IDENTITY = "-";
+				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "-";
+				CODE_SIGN_STYLE = Automatic;
+				COMBINE_HIDPI_IMAGES = YES;
+				CURRENT_PROJECT_VERSION = 1;
+				DEAD_CODE_STRIPPING = YES;
+				DEVELOPMENT_TEAM = 9YG3G8543N;
+				ENABLE_HARDENED_RUNTIME = YES;
+				ENABLE_PREVIEWS = YES;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_KEY_NSHumanReadableCopyright = "";
+				LD_RUNPATH_SEARCH_PATHS = (
+					"$(inherited)",
+					"@executable_path/../Frameworks",
+				);
+				MACOSX_DEPLOYMENT_TARGET = 12.1;
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = org.tinygrad.tinygpu.installer;
+				PRODUCT_NAME = TinyGPU;
+				PROVISIONING_PROFILE_SPECIFIER = "";
+				SDKROOT = macosx;
+				SWIFT_ACTIVE_COMPILATION_CONDITIONS = DEBUG;
+				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
+				SWIFT_VERSION = 5.0;
+			};
+			name = Debug;
+		};
+		54E42BD0286A1697000E1E9A /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
+				CLANG_ENABLE_MODULES = YES;
+				CODE_SIGN_ENTITLEMENTS = macOS/macOS.entitlements;
+				CODE_SIGN_IDENTITY = "-";
+				"CODE_SIGN_IDENTITY[sdk=macosx*]" = "-";
+				CODE_SIGN_STYLE = Automatic;
+				COMBINE_HIDPI_IMAGES = YES;
+				CURRENT_PROJECT_VERSION = 1;
+				DEAD_CODE_STRIPPING = YES;
+				DEVELOPMENT_TEAM = 9YG3G8543N;
+				ENABLE_HARDENED_RUNTIME = YES;
+				ENABLE_PREVIEWS = YES;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_KEY_NSHumanReadableCopyright = "";
+				LD_RUNPATH_SEARCH_PATHS = (
+					"$(inherited)",
+					"@executable_path/../Frameworks",
+				);
+				MACOSX_DEPLOYMENT_TARGET = 12.1;
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = org.tinygrad.tinygpu.installer;
+				PRODUCT_NAME = TinyGPU;
+				PROVISIONING_PROFILE_SPECIFIER = "";
+				SDKROOT = macosx;
+				SWIFT_COMPILATION_MODE = wholemodule;
+				SWIFT_EMIT_LOC_STRINGS = YES;
+				SWIFT_OPTIMIZATION_LEVEL = "-O";
+				SWIFT_VERSION = 5.0;
+			};
+			name = Release;
+		};
+		C5B7D9C726128AC50089B4C3 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++17";
+				CLANG_CXX_LIBRARY = "libc++";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = dwarf;
+				DRIVERKIT_DEPLOYMENT_TARGET = 21.0;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				ENABLE_TESTABILITY = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu11;
+				GCC_DYNAMIC_NO_PIC = NO;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_OPTIMIZATION_LEVEL = 0;
+				GCC_PREPROCESSOR_DEFINITIONS = (
+					"DEBUG=1",
+					"$(inherited)",
+				);
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
+				MTL_FAST_MATH = YES;
+				ONLY_ACTIVE_ARCH = YES;
+				SDKROOT = driverkit;
+			};
+			name = Debug;
+		};
+		C5B7D9C826128AC50089B4C3 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++17";
+				CLANG_CXX_LIBRARY = "libc++";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
+				DRIVERKIT_DEPLOYMENT_TARGET = 21.0;
+				ENABLE_NS_ASSERTIONS = NO;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu11;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				MTL_ENABLE_DEBUG_INFO = NO;
+				MTL_FAST_MATH = YES;
+				SDKROOT = driverkit;
+				SWIFT_COMPILATION_MODE = wholemodule;
+			};
+			name = Release;
+		};
+		C5B7D9CA26128AC50089B4C3 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				AD_HOC_CODE_SIGNING_ALLOWED = YES;
+				CODE_SIGN_ENTITLEMENTS = TinyGPUDriverExtension/TinyGPUDriver.entitlements;
+				CODE_SIGN_IDENTITY = "-";
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_TEAM = 9YG3G8543N;
+				DRIVERKIT_DEPLOYMENT_TARGET = 21.0;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				FRAMEWORK_SEARCH_PATHS = (
+					"$(inherited)",
+					"$(SDKROOT)/System/DriverKit/System/Library/Frameworks",
+				);
+				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_FILE = TinyGPUDriverExtension/Info.plist;
+				INFOPLIST_KEY_OSBundleUsageDescription = "Sample Code Audio Driver Kit Extension";
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = org.tinygrad.tinygpu.edriver;
+				PRODUCT_NAME = "$(inherited)";
+				PROVISIONING_PROFILE_SPECIFIER = "";
+				RUN_CLANG_STATIC_ANALYZER = YES;
+				SDKROOT = driverkit;
+				SKIP_INSTALL = YES;
+			};
+			name = Debug;
+		};
+		C5B7D9CB26128AC50089B4C3 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				AD_HOC_CODE_SIGNING_ALLOWED = YES;
+				CODE_SIGN_ENTITLEMENTS = TinyGPUDriverExtension/TinyGPUDriver.entitlements;
+				CODE_SIGN_IDENTITY = "-";
+				CODE_SIGN_STYLE = Automatic;
+				CURRENT_PROJECT_VERSION = 1;
+				DEVELOPMENT_TEAM = 9YG3G8543N;
+				DRIVERKIT_DEPLOYMENT_TARGET = 21.0;
+				ENABLE_USER_SCRIPT_SANDBOXING = YES;
+				FRAMEWORK_SEARCH_PATHS = (
+					"$(inherited)",
+					"$(SDKROOT)/System/DriverKit/System/Library/Frameworks",
+				);
+				GENERATE_INFOPLIST_FILE = YES;
+				INFOPLIST_FILE = TinyGPUDriverExtension/Info.plist;
+				INFOPLIST_KEY_OSBundleUsageDescription = "Sample Code Audio Driver Kit Extension";
+				MARKETING_VERSION = 1.0;
+				PRODUCT_BUNDLE_IDENTIFIER = org.tinygrad.tinygpu.edriver;
+				PRODUCT_NAME = "$(inherited)";
+				PROVISIONING_PROFILE_SPECIFIER = "";
+				RUN_CLANG_STATIC_ANALYZER = YES;
+				SDKROOT = driverkit;
+				SKIP_INSTALL = YES;
+			};
+			name = Release;
+		};
+/* End XCBuildConfiguration section */
+
+/* Begin XCConfigurationList section */
+		54E42BD2286A1697000E1E9A /* Build configuration list for PBXNativeTarget "TinyGPU Installer (macOS)" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				54E42BCF286A1697000E1E9A /* Debug */,
+				54E42BD0286A1697000E1E9A /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+		C5B7D9B626128AC50089B4C3 /* Build configuration list for PBXProject "TinyGPUDriverExtension" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				C5B7D9C726128AC50089B4C3 /* Debug */,
+				C5B7D9C826128AC50089B4C3 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+		C5B7D9C926128AC50089B4C3 /* Build configuration list for PBXNativeTarget "TinyGPUDriver" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				C5B7D9CA26128AC50089B4C3 /* Debug */,
+				C5B7D9CB26128AC50089B4C3 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+/* End XCConfigurationList section */
+	};
+	rootObject = C5B7D9B326128AC50089B4C3 /* Project object */;
+}

extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension.xcodeproj/project.xcworkspace/xcshareddata/WorkspaceSettings.xcsettings

@@ -0,0 +1,10 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>BuildSystemType</key>
+	<string>Latest</string>
+	<key>DerivedDataLocationStyle</key>
+	<string>Default</string>
+</dict>
+</plist>

extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/Info.plist

@@ -0,0 +1,37 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>IOKitPersonalities</key>
+	<dict>
+		<key>TinyGPUDriver</key>
+		<dict>
+			<key>CFBundleIdentifier</key>
+			<string>$(PRODUCT_BUNDLE_IDENTIFIER)</string>
+			<key>IOClass</key>
+			<string>IOUserService</string>
+			<key>IOMatchCategory</key>
+			<string>TinyGPUDriver</string>
+			<key>IOPCIClassMatch</key>
+			<string>0x03000000</string>
+			<key>IOPCITunnelCompatible</key>
+			<true/>
+			<key>IOProviderClass</key>
+			<string>IOPCIDevice</string>
+			<key>IOResourceMatch</key>
+			<string>IOKit</string>
+			<key>IOUserClass</key>
+			<string>TinyGPUDriver</string>
+			<key>IOUserServerName</key>
+			<string>org.tinygrad.tinygpu.Driver</string>
+			<key>TinyGPUDriverUserClientProperties</key>
+			<dict>
+				<key>IOClass</key>
+				<string>IOUserUserClient</string>
+				<key>IOUserClass</key>
+				<string>TinyGPUDriverUserClient</string>
+			</dict>
+		</dict>
+	</dict>
+</dict>
+</plist>

extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriver.cpp

@@ -0,0 +1,259 @@
+#include "TinyGPUDriver.h"
+#include "TinyGPUDriverUserClient.h"
+#include <AudioDriverKit/AudioDriverKit.h>
+#include <DriverKit/IOUserServer.h>
+#include <DriverKit/IOLib.h>
+#include <DriverKit/OSString.h>
+#include <DriverKit/IOMemoryMap.h>
+#include <DriverKit/IODMACommand.h>
+#include <DriverKit/IODispatchQueue.h>
+#include <PCIDriverKit/PCIDriverKit.h>
+#include <DriverKit/OSAction.h>
+
+struct TinyGPUDriver_IVars
+{
+	IOPCIDevice *pci = nullptr;
+};
+
+bool TinyGPUDriver::init()
+{
+	os_log(OS_LOG_DEFAULT, "tinygpu: init");
+
+	auto answer = super::init();
+	if (!answer) {
+		return false;
+	}
+
+	ivars = new TinyGPUDriver_IVars();
+	if (ivars == nullptr) {
+		return false;
+	}
+
+	return true;
+}
+
+void TinyGPUDriver::free()
+{
+	if (ivars != nullptr) {
+		
+	}
+	IOSafeDeleteNULL(ivars, TinyGPUDriver_IVars, 1);
+	super::free();
+}
+
+kern_return_t TinyGPUDriver::Start_Impl(IOService* in_provider)
+{
+	IOServiceName service_name;
+	os_log(OS_LOG_DEFAULT, "tinygpu: on gpu detected");
+
+	kern_return_t err = Start(in_provider, SUPERDISPATCH);
+	if (err) return err;
+
+	ivars->pci = OSDynamicCast(IOPCIDevice, in_provider);
+	if (!ivars->pci) return kIOReturnNoDevice;
+
+	err = ivars->pci->Open(this, 0);
+	if (err) {
+		os_log(OS_LOG_DEFAULT, "tinygpu: Open() failed 0x%08x", err);
+		ivars->pci = nullptr;
+		return err;
+	}
+
+	uint16_t ven = 0, dev = 0;
+	ivars->pci->ConfigurationRead16(kIOPCIConfigurationOffsetVendorID, &ven);
+	ivars->pci->ConfigurationRead16(kIOPCIConfigurationOffsetDeviceID, &dev);
+	os_log(OS_LOG_DEFAULT, "tinygpu: opened device ven=0x%04x dev=0x%04x", ven, dev);
+
+#if 0
+	uint32_t off = 0x100;
+	while (off) {
+		uint32_t hdr = 0, next = 0, cap_id = 0;
+		ivars->pci->ConfigurationRead32(off, &hdr);
+		cap_id = hdr & 0xFFFFu;
+		next = (hdr >> 20) & 0xFFCu;
+		os_log(OS_LOG_DEFAULT, "tinygpu: cap: %u", cap_id);
+		if (cap_id == 0x15) {
+			uint32_t cap = 0, ctrl = 0;
+			ivars->pci->ConfigurationRead32(off+0x4, &cap);
+			ivars->pci->ConfigurationRead32(off+0x8, &ctrl);
+
+			uint32_t new_bar_size = 31 - __builtin_clz(cap >> 4);
+			uint32_t new_ctrl = (ctrl & ~0x1f00) | (new_bar_size << 8);
+			ivars->pci->ConfigurationWrite32(off+0x8, new_ctrl);
+
+			os_log(OS_LOG_DEFAULT, "tinygpu: rebar: cap=%u ctrl=%u new_bar_size=%u new_ctrl=%u", cap, ctrl, new_bar_size, new_ctrl);
+			ivars->pci->Reset(0);
+			break;
+		}
+		off = next;
+	}
+	ivars->pci->Reset(kIOPCIDeviceResetTypeHotReset);
+#endif
+
+	uint16_t commandRegister;
+	ivars->pci->ConfigurationRead16(kIOPCIConfigurationOffsetCommand, &commandRegister);
+	commandRegister |= (kIOPCICommandIOSpace | kIOPCICommandBusMaster | kIOPCICommandMemorySpace);
+	ivars->pci->ConfigurationWrite16(kIOPCIConfigurationOffsetCommand, commandRegister);
+
+	memcpy((void*)service_name, (void*)"tinygpu\0", 8);
+	SetName(service_name);
+
+	os_log(OS_LOG_DEFAULT, "tinygpu: will register service %s", service_name);
+	RegisterService();
+
+	os_log(OS_LOG_DEFAULT, "tinygpu: service started %s", service_name);
+	return 0;
+}
+
+kern_return_t TinyGPUDriver::Stop_Impl(IOService* in_provider)
+{
+	ivars->pci->Close(this, 0);
+	return 0;
+}
+
+kern_return_t TinyGPUDriver::NewUserClient_Impl(uint32_t in_type, IOUserClient** out_user_client)
+{
+	kern_return_t err = 0;
+
+	IOService* user_client_service = nullptr;
+	err = Create(this, "TinyGPUDriverUserClientProperties", &user_client_service);
+	if (err) {
+		os_log(OS_LOG_DEFAULT, "tinygpu: failed to create NewUserClient");
+		goto error;
+	}
+	*out_user_client = OSDynamicCast(IOUserClient, user_client_service);
+	os_log(OS_LOG_DEFAULT, "tinygpu: NewUserClient created");
+
+error:
+	return err;
+}
+
+kern_return_t TinyGPUDriver::MapBar(uint32_t bar, IOMemoryDescriptor** memory)
+{
+	kern_return_t err = 0;
+	uint8_t barMemoryIndex, barMemoryType;
+	uint64_t barMemorySize;
+	err = ivars->pci->GetBARInfo(bar, &barMemoryIndex, &barMemorySize, &barMemoryType);
+	if (err) return err;
+
+	os_log(OS_LOG_DEFAULT, "tinygpu: requested bar mapping %d, %d", bar, (uint32_t)barMemoryIndex);
+	err = ivars->pci->_CopyDeviceMemoryWithIndex(barMemoryIndex, memory, this);
+	return err;
+}
+
+kern_return_t TinyGPUDriver::CreateDMA(size_t size, TinyGPUCreateDMAResp* dmaDesc)
+{
+	kern_return_t err = 0;
+	IOMemoryMap* memoryMap = nullptr;
+	IOBufferMemoryDescriptor* sharedBuf = nullptr;
+	IODMACommand* dmaCmd = nullptr;
+	uint64_t flags = kIOMemoryDirectionInOut;
+	uint32_t segCount = 32;
+	IOAddressSegment segments[32];
+	IODMACommandSpecification dmaSpec = {
+		.options = 0,
+		.maxAddressBits = 40,
+	};
+
+	err = IOBufferMemoryDescriptor::Create(kIOMemoryDirectionInOut, size, IOVMPageSize, &sharedBuf);
+	if (err) {
+		os_log(OS_LOG_DEFAULT, "tinygpu: failed to alloc user buffer, err=%d", err);
+		goto error;
+	}
+	
+	err = IODMACommand::Create(ivars->pci, kIODMACommandCreateNoOptions, &dmaSpec, &dmaCmd);
+	if (err) {
+		os_log(OS_LOG_DEFAULT, "tinygpu: failed to create dma command, err=%d", err);
+		goto error;
+	}
+
+	err = dmaCmd->PrepareForDMA(kIODMACommandPrepareForDMANoOptions, sharedBuf, 0, size,
+								&flags, &segCount, segments);
+	if (err) {
+		os_log(OS_LOG_DEFAULT, "tinygpu: failed to prepare for dma, err=%d", err);
+		goto error;
+	}
+
+	// pass addresses to userland
+	{
+		// debug
+		for (int i = 0; i < segCount; i++) {
+			os_log(OS_LOG_DEFAULT, "tinygpu: new dma mapping (sz=0x%zx) %d 0x%llx 0x%llx", size, i, segments[i].address, segments[i].length);
+		}
+
+		err = sharedBuf->CreateMapping(0, 0, 0, IOVMPageSize, IOVMPageSize, &memoryMap); // one page should be fine
+		if (err) {
+			os_log(OS_LOG_DEFAULT, "tinygpu: failed to map memory, err=%d", err);
+			goto error;
+		}
+
+		// Send back gpu addresses
+		uint64_t* addr = (uint64_t*)memoryMap->GetAddress();
+		for (int i = 0; i < segCount; i++) {
+			addr[i * 2] = segments[i].address;
+			addr[i * 2 + 1] = segments[i].length;
+		}
+		addr[segCount * 2] = 0;
+		addr[segCount * 2 + 1] = 0;
+
+		// free memoryMap
+		memoryMap->release();
+		memoryMap = nullptr;
+	}
+
+	dmaDesc->sharedBuf = sharedBuf;
+	dmaDesc->dmaCmd = dmaCmd;
+	return 0;
+
+error:
+	if (memoryMap) {
+		memoryMap->release();
+		memoryMap = nullptr;
+	}
+	if (dmaCmd) {
+		dmaCmd->CompleteDMA(kIODMACommandCompleteDMANoOptions);
+		dmaCmd->release();
+		dmaCmd = nullptr;
+	}
+	if (sharedBuf) {
+		sharedBuf->release();
+		sharedBuf = nullptr;
+	}
+	return err;
+}
+
+kern_return_t TinyGPUDriver::CfgRead(uint32_t off, uint32_t size, uint32_t* outVal)
+{
+  if (!ivars->pci || !outVal) return kIOReturnNotReady;
+
+  if (size == 1) {
+	uint8_t v8 = 0;
+	ivars->pci->ConfigurationRead8(off, &v8);
+	*outVal = v8;
+  } else if (size == 2) {
+	uint16_t v16 = 0;
+	ivars->pci->ConfigurationRead16(off, &v16);
+	*outVal = v16;
+  } else if (size == 4) {
+	uint32_t v32 = 0;
+	ivars->pci->ConfigurationRead32(off, &v32);
+	*outVal = v32;
+  }
+  return 0;
+}
+
+kern_return_t TinyGPUDriver::CfgWrite(uint32_t off, uint32_t size, uint32_t val)
+{
+  if (!ivars->pci) return kIOReturnNotReady;
+  if (size == 1) ivars->pci->ConfigurationWrite8 (off, (uint8_t)val);
+  else if (size == 2) ivars->pci->ConfigurationWrite16(off, (uint16_t)val);
+  else if (size == 4) ivars->pci->ConfigurationWrite32(off, (uint32_t)val);
+  return 0;
+}
+
+kern_return_t TinyGPUDriver::ResetDevice()
+{
+	if (!ivars->pci) return kIOReturnNotReady;
+	ivars->pci->Reset(kIOPCIDeviceResetTypeFunctionReset);
+	return 0;
+}

extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriver.entitlements

@@ -0,0 +1,21 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>com.apple.developer.driverkit.transport.pci</key>
+	<array>
+		<dict>
+			<key>IOPCIMatch</key>
+			<string>0x00001002&amp;0x0000FFFF</string>
+		</dict>
+		<dict>
+			<key>IOPCIMatch</key>
+			<string>0x000010de&amp;0x0000FFFF</string>
+		</dict>
+	</array>
+	<key>com.apple.developer.driverkit.allow-any-userclient-access</key>
+	<true/>
+	<key>com.apple.developer.driverkit</key>
+	<true/>
+</dict>
+</plist>

extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriver.iig

@@ -0,0 +1,38 @@
+#ifndef TinyGPUDriver_h
+#define TinyGPUDriver_h
+
+#include <Availability.h>
+#include <DriverKit/IOService.iig>
+#include <PCIDriverKit/IOPCIDevice.iig>
+#include <DriverKit/IOMemoryMap.iig>
+#include <DriverKit/IODMACommand.iig>
+
+struct TinyGPUCreateDMAResp
+{
+	IOBufferMemoryDescriptor* sharedBuf;
+	IODMACommand* dmaCmd;
+};
+
+class TinyGPUDriver: public IOService
+{
+public:
+	virtual bool init() override;
+
+	virtual void free() override;
+
+	virtual kern_return_t Start(IOService * provider) override;
+
+	virtual kern_return_t Stop(IOService * provider) override;
+
+	virtual kern_return_t NewUserClient(uint32_t in_type, IOUserClient** out_user_client) override;
+
+	kern_return_t MapBar(uint32_t bar, IOMemoryDescriptor** memory) LOCALONLY;
+	kern_return_t CreateDMA(size_t size, TinyGPUCreateDMAResp* dmaDesc) LOCALONLY;
+
+	kern_return_t CfgRead(uint32_t off, uint32_t size, uint32_t* val) LOCALONLY;
+	kern_return_t CfgWrite(uint32_t off, uint32_t size, uint32_t val) LOCALONLY;
+	kern_return_t ResetDevice() LOCALONLY;
+	kern_return_t BarInfo() LOCALONLY;
+};
+
+#endif /* TinyGPUDriver_h */

extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriverUserClient.cpp

@@ -0,0 +1,127 @@
+#include "TinyGPUDriverUserClient.h"
+#include "TinyGPUDriver.h"
+#include <DriverKit/DriverKit.h>
+#include <DriverKit/OSSharedPtr.h>
+#include <PCIDriverKit/PCIDriverKit.h>
+
+struct TinyGPUDriverUserClient_IVars
+{
+	OSSharedPtr<TinyGPUDriver> provider = nullptr;
+};
+
+bool TinyGPUDriverUserClient::init()
+{
+	auto theAnswer = super::init();
+	if (!theAnswer) {
+		return false;
+	}
+
+	ivars = IONewZero(TinyGPUDriverUserClient_IVars, 1);
+	if (ivars == nullptr) {
+		return false;
+	}
+
+	return true;
+}
+
+void TinyGPUDriverUserClient::free()
+{
+	if (ivars != nullptr) {
+		ivars->provider.reset();
+	}
+
+	IOSafeDeleteNULL(ivars, TinyGPUDriverUserClient_IVars, 1);
+	super::free();
+}
+
+kern_return_t TinyGPUDriverUserClient::Start_Impl(IOService* in_provider)
+{
+	kern_return_t err = kIOReturnSuccess;
+	if (!in_provider) {
+		os_log(OS_LOG_DEFAULT, "tinygpu: provider is null");
+		err = kIOReturnBadArgument;
+		goto error;
+	}
+
+	err = Start(in_provider, SUPERDISPATCH);
+	if (err) {
+		os_log(OS_LOG_DEFAULT, "tinygpu: failed to start super (%d)", err);
+		goto error;
+	}
+
+	ivars->provider = OSSharedPtr(OSDynamicCast(TinyGPUDriver, in_provider), OSRetain);
+	return 0;
+
+error:
+	ivars->provider.reset();
+	return err;
+}
+
+kern_return_t TinyGPUDriverUserClient::Stop_Impl(IOService* in_provider)
+{
+	return Stop(in_provider, SUPERDISPATCH);
+}
+
+kern_return_t TinyGPUDriverUserClient::ExternalMethod(uint64_t selector, IOUserClientMethodArguments* args, const IOUserClientMethodDispatch* in_dispatch, OSObject* in_target, void* in_reference)
+{
+	kern_return_t err = 0;
+
+	os_log(OS_LOG_DEFAULT, "tinygpu: rpc (%llu) in:%d, out:%d", selector, args->scalarInputCount, args->scalarOutputCount);
+
+	if (selector == TinyGPURPC::ReadCfg) {
+		if (args->scalarInputCount != 2 or args->scalarOutputCount < 1) return kIOReturnBadArgument;
+
+		uint32_t off = uint32_t(args->scalarInput[0]);
+		uint32_t size = uint32_t(args->scalarInput[1]);
+
+		uint32_t val = 0;
+		err = ivars->provider->CfgRead(off, size, &val);
+		os_log(OS_LOG_DEFAULT, "tinygpu: read cfg off:%x sz:%d, val:%x", off, size, val);
+
+		if (!err) {
+			args->scalarOutput[0] = val;
+			args->scalarOutputCount = 1;
+		}
+		return err;
+	} else if (selector == TinyGPURPC::WriteCfg) {
+		if (args->scalarInputCount != 3) return kIOReturnBadArgument;
+
+		uint32_t off = uint32_t(args->scalarInput[0]);
+		uint32_t size = uint32_t(args->scalarInput[1]);
+		uint32_t val = uint32_t(args->scalarInput[2]);
+
+		os_log(OS_LOG_DEFAULT, "tinygpu: wr cfg off:%x sz:%d, val:%x", off, size, val);
+		return ivars->provider->CfgWrite(off, size, val);
+	} else if (selector == TinyGPURPC::Reset) {
+		os_log(OS_LOG_DEFAULT, "tinygpu: reset");
+		return ivars->provider->ResetDevice();
+	}
+
+	return kIOReturnUnsupported;
+}
+
+kern_return_t IMPL(TinyGPUDriverUserClient, CopyClientMemoryForType)
+{
+	if (!memory) {
+		return kIOReturnBadArgument;
+	}
+
+	if (ivars->provider.get() == nullptr) {
+		return kIOReturnNotAttached;
+	}
+
+	if (type < 6) {
+		uint32_t bar = (uint32_t)type;
+		return ivars->provider->MapBar(bar, memory);
+	}
+
+	// dma page buffer
+	TinyGPUCreateDMAResp buf;
+	kern_return_t err = ivars->provider->CreateDMA(type, &buf);
+	if (err) {
+		return err;
+	}
+
+	*memory = buf.sharedBuf;
+	return 0;
+}

extra/usbgpu/tbgpu/installer/TinyGPUDriverExtension/TinyGPUDriverUserClient.iig

@@ -0,0 +1,28 @@
+#ifndef TinyGPUDriverUserClient_h
+#define TinyGPUDriverUserClient_h
+
+#include <DriverKit/IOUserClient.iig>
+
+enum TinyGPURPC
+{
+	ReadCfg,
+	WriteCfg,
+	Reset
+};
+
+class TinyGPUDriverUserClient : public IOUserClient
+{
+public:
+	virtual bool init() final;
+	virtual void free() final;
+
+	virtual kern_return_t Start(IOService* in_provider) final;
+	virtual kern_return_t Stop(IOService* in_provider) final;
+
+	virtual kern_return_t ExternalMethod(uint64_t in_selector, IOUserClientMethodArguments* in_arguments, const IOUserClientMethodDispatch* in_dispatch, OSObject* in_target, void* in_reference) final;
+
+	virtual kern_return_t CopyClientMemoryForType(
+		uint64_t type, uint64_t *options, IOMemoryDescriptor **memory) final;
+};
+
+#endif /* TinyGPUDriverUserClient_h */

extra/usbgpu/tbgpu/installer/macOS/macOS.entitlements

@@ -0,0 +1,12 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>com.apple.security.app-sandbox</key>
+	<false/>
+	<key>com.apple.security.files.user-selected.read-only</key>
+	<true/>
+	<key>com.apple.developer.system-extension.install</key>
+	<true/>
+</dict>
+</plist>

extra/usbgpu/tbgpu/main.cpp

@@ -0,0 +1,39 @@
+#include <CoreFoundation/CoreFoundation.h>
+#include <IOKit/IOKitLib.h>
+#include <mach/mach.h>
+#include <stdio.h>
+#include <inttypes.h>
+
+static io_connect_t open_uc_by_name(const char *svc_name) {
+  io_connect_t conn = IO_OBJECT_NULL;
+  io_service_t service = IOServiceGetMatchingService(kIOMasterPortDefault, IOServiceNameMatching(svc_name));
+  if (!service) { fprintf(stderr, "service not found: %s\n", svc_name); return IO_OBJECT_NULL; }
+  kern_return_t kr = IOServiceOpen(service, mach_task_self(), /*type*/0, &conn);
+  IOObjectRelease(service);
+  if (kr) { fprintf(stderr, "IOServiceOpen 0x%x\n", kr); return IO_OBJECT_NULL; }
+  return conn;
+}
+
+int main(int argc, char **argv) {
+  uint32_t bar = (argc > 1) ? (uint32_t)strtoul(argv[1], NULL, 0) : 0;  // pick BAR index
+  io_connect_t conn = open_uc_by_name("tinygpu");
+  if (!conn) return 2;
+
+  mach_vm_address_t addr = 0;
+  mach_vm_size_t    size = 0;
+  kern_return_t kr = IOConnectMapMemory64(conn, bar, mach_task_self(), &addr, &size, kIOMapAnywhere);
+  if (kr) { fprintf(stderr, "Map BAR%u failed 0x%x\n", bar, kr); IOServiceClose(conn); return 3; }
+
+  printf("BAR%u mapped at 0x%llx, size 0x%llx\n", bar, (unsigned long long)addr, (unsigned long long)size);
+
+  // example: read a 32-bit register at offset 0x0 (make sure it’s safe!)
+  volatile uint32_t *mmio = (volatile uint32_t*)(uintptr_t)addr;
+  uint32_t v = mmio[0];
+  printf("mmio[0]=0x%08x\n", v);
+
+  kr = IOConnectUnmapMemory64(conn, bar, mach_task_self(), addr);
+  if (kr) fprintf(stderr, "Unmap failed 0x%x\n", kr);
+
+  IOServiceClose(conn);
+  return 0;
+}

extra/usbgpu/tbgpu/main.py

@@ -0,0 +1,82 @@
+import ctypes, ctypes.util, sys
+
+cf = ctypes.CDLL(ctypes.util.find_library("CoreFoundation"))
+iokit = ctypes.CDLL(ctypes.util.find_library("IOKit"))
+libsys = ctypes.CDLL(ctypes.util.find_library("System"))
+
+kern_return_t = ctypes.c_int
+mach_port_t = ctypes.c_uint
+io_object_t = mach_port_t
+io_service_t = io_object_t
+io_connect_t = mach_port_t
+CFMutableDictionaryRef = ctypes.c_void_p
+CFStringRef = ctypes.c_void_p
+
+kIOMasterPortDefault = mach_port_t(0)
+
+libsys.mach_task_self_.restype = mach_port_t
+
+iokit.IOServiceNameMatching.argtypes = [ctypes.c_char_p]
+iokit.IOServiceNameMatching.restype  = CFMutableDictionaryRef
+
+iokit.IOServiceGetMatchingService.argtypes = [mach_port_t, CFMutableDictionaryRef]
+iokit.IOServiceGetMatchingService.restype  = io_service_t
+
+iokit.IOObjectRelease.argtypes = [io_object_t]
+iokit.IOObjectRelease.restype  = kern_return_t
+
+iokit.IOServiceOpen.argtypes = [io_service_t, mach_port_t, ctypes.c_uint32, ctypes.POINTER(io_connect_t)]
+iokit.IOServiceOpen.restype  = kern_return_t
+
+iokit.IOConnectCallMethod.argtypes = [io_connect_t, ctypes.c_uint32, ctypes.POINTER(ctypes.c_uint64), ctypes.c_uint32, ctypes.c_void_p,
+  ctypes.c_size_t, ctypes.POINTER(ctypes.c_uint64), ctypes.POINTER(ctypes.c_uint32), ctypes.c_void_p, ctypes.POINTER(ctypes.c_size_t)]
+iokit.IOConnectCallMethod.restype = kern_return_t
+
+def open_userclient_by_name(name: str, uc_type: int = 0) -> io_connect_t:
+  mdict = iokit.IOServiceNameMatching(name.encode("utf-8"))
+  if not mdict: raise RuntimeError("IOServiceNameMatching returned NULL")
+
+  # Grab the first matching service
+  service = iokit.IOServiceGetMatchingService(kIOMasterPortDefault, mdict)
+  if not service: raise RuntimeError(f'service "{name}" not found')
+
+  # print("lol", service)
+  # print(libsys.mach_task_self_)
+  # cast libsys.mach_task_self_ to uint and print
+  # print("lol", ctypes.cast(libsys.mach_task_self_, ctypes.POINTER(ctypes.c_uint)).contents.value)
+
+  try:
+    # Open user client (type -> passed to NewUserClient_Impl)
+    conn = io_connect_t(0)
+    # print("lol", libsys.mach_task_self_)
+    kr = iokit.IOServiceOpen(service, ctypes.cast(libsys.mach_task_self_, ctypes.POINTER(ctypes.c_uint)).contents.value,
+                             ctypes.c_uint32(uc_type), ctypes.byref(conn))
+    if kr != 0: raise OSError(kr, f"IOServiceOpen failed (0x{kr:08x})")
+    return conn
+  finally: iokit.IOObjectRelease(service)
+
+def external_method(conn: io_connect_t, selector: int = 0) -> int:
+  # no scalars in/out, no struct in/out — just ping selector 0
+  in_scalars = ctypes.POINTER(ctypes.c_uint64)()  # NULL
+  out_scalars = (ctypes.c_uint64 * 1)()       # space if driver returns something
+  out_scalars_cnt = ctypes.c_uint32(0)      # driver can set this
+
+  return iokit.IOConnectCallMethod(conn, ctypes.c_uint32(selector), in_scalars, ctypes.c_uint32(0), None, ctypes.c_size_t(0),
+    out_scalars, ctypes.byref(out_scalars_cnt), None, ctypes.byref(ctypes.c_size_t(0)))
+
+def close_userclient(conn: io_connect_t) -> None:
+  # IOServiceClose is a macro; exported symbol is IOServiceClose in IOKit
+  iokit.IOServiceClose.argtypes = [io_connect_t]
+  iokit.IOServiceClose.restype  = kern_return_t
+  iokit.IOServiceClose(conn)
+
+if __name__ == "__main__":
+  try:
+    conn = open_userclient_by_name("tinygpu", uc_type=0)
+    kr = external_method(conn, selector=0)
+    print(f"ExternalMethod(0) -> 0x{kr:08x}")
+  except Exception as e:
+    print(e)
+    sys.exit(1)
+  finally:
+    if 'conn' in locals() and conn.value: close_userclient(conn)

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

@@ -9,7 +9,7 @@ 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",
@@ -42,7 +42,6 @@ setup(name='tinygrad',
         'tinygrad.runtime.support.am',
         'tinygrad.runtime.support.nv',
         'tinygrad.schedule',
-        'tinygrad.shape',
         'tinygrad.uop',
         'tinygrad.viz',
       ],

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_bert_softmax.py

@@ -1,4 +1,4 @@
-from tinygrad import Tensor, dtypes, Context, GlobalCounters
+from tinygrad import Tensor, dtypes, GlobalCounters
 dtypes.default_float = dtypes.float16
 from tinygrad.dtype import to_dtype
 from tinygrad.helpers import getenv
@@ -13,6 +13,5 @@ if __name__ == "__main__":
 
   # test single kernel softmax
   GlobalCounters.reset()
-  with Context(DONT_GROUP_REDUCES=1):
-    single_kernel_softmax(t, -1, acc_dtype).realize()
+  single_kernel_softmax(t, -1, acc_dtype).realize()
 

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.control_flow 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_debug_metal_sd_conv.py

@@ -1,46 +0,0 @@
-# ruff: noqa: E501
-from tinygrad.codegen.opt.kernel import Kernel, Opt, OptOps
-from tinygrad.dtype import dtypes
-from tinygrad.engine.realize import CompiledRunner, get_program
-from tinygrad.codegen.opt.search import bufs_from_lin
-from tinygrad.uop.ops import UOp, Ops
-from tinygrad.shape.shapetracker import ShapeTracker
-from tinygrad.shape.view import View
-
-ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-  UOp(Ops.STORE, dtypes.void, arg=None, src=(
-    UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(), arg=0, src=()),
-    UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 1, 1280, 8, 8, 1, 1, 1), strides=(81920, 0, 64, 8, 1, 0, 0, 0), offset=0, mask=None, contiguous=True),)), src=()),
-    UOp(Ops.ADD, dtypes.half, arg=None, src=(
-      UOp(Ops.ADD, dtypes.half, arg=None, src=(
-        UOp(Ops.CAST, dtypes.half, arg=None, src=(
-          UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (5, 6, 7)), src=(
-            UOp(Ops.CAST, dtypes.float, arg=None, src=(
-              UOp(Ops.MUL, dtypes.half, arg=None, src=(
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(), arg=1, src=()),
-                  UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1, 2, 1, 2560, 4, 10, 4, 10), strides=(0, 163840, 0, 64, 0, 8, 0, 1), offset=-9, mask=((0, 1), (0, 2), (0, 1), (0, 2560), (0, 4), (1, 9), (0, 4), (1, 9)), contiguous=False), View(shape=(2, 1, 1280, 8, 8, 2560, 3, 3), strides=(4096000, 0, 0, 40, 1, 1600, 440, 11), offset=0, mask=None, contiguous=False))), src=()),)),
-                UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(), arg=2, src=()),
-                  UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 1, 1280, 8, 8, 2560, 3, 3), strides=(0, 0, 23040, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),)),
-        UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(), arg=3, src=()),
-          x17:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(2, 1, 1280, 8, 8, 1, 1, 1), strides=(0, 0, 1, 0, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)), src=()),)),)),
-      UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-        UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(), arg=4, src=()),
-         x17,)),)),)),))
-opts = [Opt(op=OptOps.UPCAST, axis=3, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UNROLL, axis=2, arg=0), Opt(op=OptOps.UNROLL, axis=1, arg=0), Opt(op=OptOps.LOCAL, axis=1, arg=8), Opt(op=OptOps.LOCAL, axis=2, arg=8), Opt(op=OptOps.LOCAL, axis=2, arg=2)]
-
-k = Kernel(ast)
-k.apply_opts(opts)
-bufs = bufs_from_lin(k)
-
-prg = CompiledRunner(get_program(k.ast, k.opts, k.applied_opts))
-
-for i in range(10):
-  speed = prg(bufs, var_vals={}, wait=True)
-  print(f"kernel time: {speed*1e3:.2f} ms")
-
-# on M1 Max
-# 11ms before block 9b0859d71780fef5cf3831e317f74e53f2483229
-# 15ms after block cbcc1c20eb09a1342f6581cfbb99632bade982a8

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/external_test_train_gpt2.py

@@ -1,55 +0,0 @@
-# ruff: noqa: E501
-import unittest
-
-from tinygrad.uop.ops import UOp, Ops
-from .search import Opt, OptOps
-from tinygrad.dtype import dtypes
-from tinygrad.shape.shapetracker import ShapeTracker
-from tinygrad.shape.view import View
-from tinygrad.codegen.opt.kernel import Kernel
-
-from test.external.fuzz_linearizer import run_linearizer
-
-class TestTrainGpt2Kernel(unittest.TestCase):
-  def test_1(self):
-    # kernel 244
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(206045184), arg=0, src=()),
-        UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4, 1024, 50304, 1), strides=(51511296, 50304, 1, 0), offset=0, mask=None, contiguous=True),)), src=()),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (3,)), src=(
-        UOp(Ops.MUL, dtypes.float, arg=None, src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(3145728), arg=1, src=()),
-          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4, 1024, 50304, 768), strides=(786432, 768, 0, 1), offset=0, mask=None, contiguous=False),)), src=()),)),
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(38633472), arg=2, src=()),
-          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4, 1024, 50304, 768), strides=(0, 0, 768, 1), offset=0, mask=None, contiguous=False),)), src=()),)),)),)),)),))
-
-    opts = [Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.UPCAST, axis=1, arg=3), Opt(op=OptOps.LOCAL, axis=0, arg=2)]
-    kernel = Kernel(ast)
-    kernel.apply_opts(opts)
-    run_linearizer(kernel)
-
-  def test_2(self):
-    # kernel 254
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(3145728), arg=0, src=()),
-        UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4, 1024, 1, 768), strides=(786432, 768, 0, 1), offset=0, mask=None, contiguous=True),)), src=()),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (2,)), src=(
-        UOp(Ops.MUL, dtypes.float, arg=None, src=(
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(38633472), arg=1, src=()),
-          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4, 1024, 50304, 768), strides=(0, 0, 768, 1), offset=0, mask=None, contiguous=False),)), src=()),)),
-          UOp(Ops.LOAD, dtypes.float, arg=None, src=(
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(205852672), arg=2, src=()),
-          UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4, 1024, 50304, 768), strides=(51463168, 50257, 1, 0), offset=0, mask=((0, 4), (0, 1024), (0, 50257), (0, 768)), contiguous=False),)), src=()),)),)),)),)),))
-
-    opts = [Opt(op=OptOps.LOCAL, axis=1, arg=16), Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.UPCAST, axis=2, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.LOCAL, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=3, arg=4)]
-    kernel = Kernel(ast)
-    kernel.apply_opts(opts)
-    run_linearizer(kernel)
-
-if __name__ == "__main__":
-  unittest.main()

test/external/external_uop_gc.py

@@ -1,6 +1,5 @@
 import gc
 from tinygrad import Tensor, UOp, Device, nn
-from tinygrad.shape.shapetracker import views_to_valid_uop
 from tinygrad.engine.realize import method_cache, get_program
 from tinygrad.schedule.indexing import apply_movement_op
 from test.test_tiny import TestTiny
@@ -69,7 +68,6 @@ if __name__ == "__main__":
 
     # these caches will keep uops alive
     method_cache.clear()
-    views_to_valid_uop.cache_clear()
     apply_movement_op.cache_clear()
     Tensor._device_seeds.clear()
     Tensor._device_rng_counters.clear()

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/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/opt/test_gen_float4.py

@@ -2,7 +2,6 @@ import unittest
 from tinygrad import Device, Tensor, dtypes
 from tinygrad.uop.ops import UOp, Ops
 from tinygrad.codegen.opt import Opt, OptOps
-from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.engine.realize import get_program
 from tinygrad.helpers import AMX
 
@@ -149,33 +148,5 @@ class TestFloat4(unittest.TestCase):
 
     assert TestFloat4.count_float4(uops) == (1, 1)
 
-  @unittest.skip("Ops.VIEW no longer exists")
-  def test_half4_load_unrolled(self):
-    # from llama 7B shard 4 gpus
-    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
-      UOp(Ops.STORE, dtypes.void, arg=None, src=(
-        UOp(Ops.VIEW, dtypes.float.ptr(96000), arg=ShapeTracker(views=(View(shape=(1, 3, 32000, 1), strides=(0, 32000, 1, 0), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
-          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96000), arg=0, src=()),)),
-        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (3,)), src=(
-          UOp(Ops.CAST, dtypes.float, arg=None, src=(
-            UOp(Ops.MUL, dtypes.half, arg=None, src=(
-              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.half.ptr(9216), arg=ShapeTracker(views=(View(shape=(1, 3, 32000, 1024), strides=(0, 4096, 0, 1), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(9216), arg=1, src=()),)),)),
-              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
-                UOp(Ops.VIEW, dtypes.half.ptr(32768000), arg=ShapeTracker(views=(View(shape=(1, 3, 32000, 1024), strides=(0, 0, 1024, 1), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
-                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(32768000), arg=2, src=()),)),)),)),)),)),)),))
-
-    # TODO: fix this, expected might change but should be positive
-    for expected, opts in [
-      ((7, 0), [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=3), Opt(op=OptOps.UNROLL, axis=0, arg=4)]),
-      ((5, 0), [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=4)]),
-      ((2, 0), [Opt(op=OptOps.UNROLL, axis=0, arg=4)]),
-    ]:
-      program = get_program(ast, Device[Device.DEFAULT].renderer, opts=opts)
-
-      count = TestFloat4.count_half4(program.uops)
-      assert count == expected, f"{count=}, {expected=}"
-
 if __name__ == '__main__':
   unittest.main()

test/opt/test_kernel_opts.py

@@ -1,6 +1,5 @@
 import unittest
 from tinygrad import Device, Tensor, dtypes
-from tinygrad.helpers import CI
 from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
 
 # TODO: write a clean version of this
@@ -177,9 +176,7 @@ class TestKernelOpts(unittest.TestCase):
     ], apply_tc=True, atol=atol, rtol=rtol)
 
   def test_padto_matmul(self):
-    if (CI and Device.DEFAULT in ["AMD", "NV", "CUDA"]):
-      self.skipTest("super slow on CUDA and AMD because of the big grid dims")
-    N = 17 * 17
+    N = 17
     Tensor.manual_seed(289)
     a = Tensor.rand(N, N)
     b = Tensor.rand(N, N)
@@ -213,7 +210,7 @@ class TestKernelOpts(unittest.TestCase):
       helper_linearizer_opt(a@b, [[Opt(OptOps.UNROLL, 0, 0), Opt(OptOps.PADTO, 2, 8)]])
 
   def test_padto_sum_ok(self):
-    N = 18 * 18
+    N = 18
     # NOTE: this setup prevents 17 * 17 contiguous merged into one dimension
     a = Tensor.rand(N, N).realize().shrink(((0, 17), (0, 17))) * 100
     b = (Tensor.rand(N, N) < 0.5).realize().shrink(((0, 17), (0, 17)))
@@ -244,7 +241,7 @@ class TestKernelOpts(unittest.TestCase):
     helper_linearizer_opt(a.sum(0).exp(), [[Opt(OptOps.PADTO, 1, 32)],])
 
   def test_padto_sum_not_ok(self):
-    N = 18 * 18
+    N = 18
     # NOTE: this setup prevents 17 * 17 contiguous merged into one dimension
     a = Tensor.rand(N, N).shrink(((0, 17), (0, 17))).exp()
     # exp is not safe to pad
@@ -261,7 +258,7 @@ class TestKernelOpts(unittest.TestCase):
       helper_linearizer_opt(b.sum(0), [[Opt(OptOps.PADTO, 1, 32)],])
 
   def test_padto_max(self):
-    N = 18 * 18
+    N = 18
     # NOTE: this setup prevents 17 * 17 contiguous merged into one axis
     a = -Tensor.rand(N, N).shrink(((0, 17), (0, 17))) * 100
 
@@ -282,7 +279,7 @@ class TestKernelOpts(unittest.TestCase):
 
   def test_padto_where(self):
     Tensor.manual_seed(0)
-    N = 17 * 17
+    N = 17
     a = (Tensor.randn(N, N).realize().max(axis=0, keepdim=True) > 1).where(1, 0)
     helper_linearizer_opt(a.max(0), [
       [Opt(OptOps.PADTO, 0, 32)],
@@ -291,7 +288,7 @@ class TestKernelOpts(unittest.TestCase):
 
   def test_padto_where_multioutput(self):
     Tensor.manual_seed(0)
-    N = 17 * 17
+    N = 17
     r = Tensor.randn(N, N).realize().max(axis=0, keepdim=True) > 1
     a0 = r.where(1, 0)
     a1 = r.where(2, 0)

test/test_arange.py

@@ -131,8 +131,7 @@ class TestIndexing(unittest.TestCase):
     # llama3 is 128256
     vocab_size, embed_size = (10, 3) if CI else (32000, 4096)
     emb = nn.Embedding(vocab_size, embed_size)
-    # TODO: why is a new realize needed here
-    emb_w = emb.weight.realize().numpy()
+    emb_w = emb.weight.numpy()
     x = Tensor([1,2,3,4])
     with Context(NOOPT=noopt):
       GlobalCounters.reset()

test/test_compile_failures.py

@@ -1,7 +1,7 @@
 import unittest, io
 from contextlib import redirect_stdout
 from tinygrad import Tensor, dtypes, Device
-from tinygrad.helpers import OSX, CPU_LLVM
+from tinygrad.helpers import OSX, CPU_LLVM, CPU_LVP
 from tinygrad.engine.realize import lower_schedule
 from tinygrad.device import is_dtype_supported
 from tinygrad.engine.realize import get_program
@@ -19,7 +19,7 @@ class TestCompileFailures(unittest.TestCase):
 
 class TestDisassembly(unittest.TestCase):
   # TODO: fails on llvm. llvm.LLVMGetHostCPUName() returns "generic"
-  @unittest.skipUnless(Device.DEFAULT in ("CPU",) and not CPU_LLVM and OSX, "m series cpus support fp16 arithmetic")
+  @unittest.skipUnless(Device.DEFAULT in ("CPU",) and not (CPU_LLVM or CPU_LVP) and OSX, "m series cpus support fp16 arithmetic")
   def test_float16_alu(self):
     c = Tensor([1], dtype=dtypes.float16) + Tensor([1], dtype=dtypes.float16)
     s = c.schedule()[-1]

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.py

@@ -6,6 +6,7 @@ from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import getenv, DEBUG, CI
 from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float, float_to_fp8, _to_np_dtype, _to_torch_dtype, truncate
 from tinygrad.renderer.ptx import PTXRenderer
+from tinygrad.renderer.nir import NIRRenderer
 from tinygrad import Device, Tensor, dtypes
 from hypothesis import given, settings, strategies as strat
 from test.helpers import rand_for_dtype
@@ -102,7 +103,7 @@ class TestDType(unittest.TestCase):
     ))
 
   @unittest.skipIf(Device.DEFAULT == "PYTHON", "skip for now")
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "skip for now")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)), "skip for now")
   def test_uint_overflow(self):
     if not dtypes.is_unsigned(self.DTYPE): raise unittest.SkipTest("only for unsigned")
     v = dtypes.max(self.DTYPE)
@@ -261,7 +262,7 @@ class TestFloatDType(TestDType):
 class TestDoubleDType(TestDType):
   DTYPE = dtypes.double
   @unittest.skipIf((CI and Device.DEFAULT in {"CUDA", "NV"}) or \
-                    isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "conversion not supported on CI CUDA and PTX")  # TODO: why not?
+   isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)), "conversion not supported on CI CUDA, PTX, and NIR")  # TODO: why not?
   def test_float64_increased_precision(self):
     for func in [
       lambda t: t.exp(),

test/test_dtype_alu.py

@@ -6,6 +6,7 @@ from tinygrad.tensor import _to_np_dtype
 from tinygrad.device import is_dtype_supported
 from tinygrad.runtime.ops_python import from_storage_scalar
 from tinygrad.renderer.ptx import PTXRenderer
+from tinygrad.renderer.nir import NIRRenderer
 import numpy as np
 import pytest
 from hypothesis import assume, given, strategies as strat, settings, HealthCheck
@@ -29,8 +30,8 @@ unary_operations = [(Tensor.exp, np.exp), (Tensor.log, np.log), (Tensor.sin, np.
 # TODO: enable this (this is a dtype issue)
 #binary_operations.append(operator.truediv)
 
-# TODO: CI CUDA segfaults on sin, WEBGPU sin is not precise enough for large numbers
-if (getenv("MOCKGPU") and Device.DEFAULT in {"NV", "CUDA"}) or Device.DEFAULT == "WEBGPU":
+# TODO: CI CUDA segfaults on sin, WEBGPU and NIR sines are not precise enough for large numbers
+if (getenv("MOCKGPU") and Device.DEFAULT in {"NV", "CUDA"}) or Device.DEFAULT == "WEBGPU" or isinstance(Device[Device.DEFAULT].renderer, NIRRenderer):
   unary_operations.remove((Tensor.sin, np.sin))
   unary_operations.remove((Tensor.cos, np.cos))
 
@@ -74,7 +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))
@@ -184,8 +188,8 @@ class TestDTypeALU(unittest.TestCase):
   @given(ht.int32, ht.int32, ht.float32, strat.sampled_from(integer_binary_operations), strat.sampled_from(binary_operations))
   def test_int32_midcast_float(self, a, b, c, op1, op2): universal_test_midcast(a, b, c, op1, op2, dtypes.int32, dtypes.float32)
 
-  # Metal and CUDA and HIP behave differently than numpy in CI for overflows
-  skip_overflow = CI and Device.DEFAULT in {"AMD", "NV", "CUDA"}
+  # Metal and CUDA and HIP and NIR behave differently than numpy in CI for overflows
+  skip_overflow = (CI and Device.DEFAULT in {"AMD", "NV", "CUDA"}) or isinstance(Device[Device.DEFAULT].renderer, NIRRenderer)
   @given(strat.floats(width=32, min_value=0, max_value=10.0) if skip_overflow else ht.float32,
          strat.floats(width=32, min_value=0, max_value=10.0) if skip_overflow else ht.float32,
          ht.int32, strat.sampled_from(binary_operations), strat.sampled_from(integer_binary_operations))

test/test_edgecases.py

@@ -26,8 +26,9 @@ import unittest
 import numpy as np
 import torch
 from tinygrad import Tensor, dtypes, nn
-from tinygrad.device import is_dtype_supported
+from tinygrad.device import Device, is_dtype_supported
 from tinygrad.helpers import getenv
+from tinygrad.renderer.nir import NIRRenderer
 
 MOCKGPU = getenv("MOCKGPU")
 
@@ -206,7 +207,8 @@ class TestUOpValidationIssue(unittest.TestCase):
   # these fail with UOp verification error.
   # we want more of these with diverse errors!
 
-  @unittest.skipIf((not is_dtype_supported(dtypes.long)) or MOCKGPU, "hangs gpuocelot")
+  @unittest.skipIf((not is_dtype_supported(dtypes.long)) or MOCKGPU or isinstance(Device[Device.DEFAULT].renderer, NIRRenderer),
+                   "hangs gpuocelot, NIR cannot render")
   def test_tensor_index_overflow(self):
     val = Tensor([1])
     big = val.expand(2**31 + 3)

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}}"
@@ -205,7 +205,7 @@ 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 +214,8 @@ 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 +393,15 @@ 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

@@ -4,12 +4,9 @@
 
 import unittest
 from tinygrad import Device, dtypes
-from tinygrad.device import is_dtype_supported
 from tinygrad.uop.ops import UOp, Ops, AxisType, KernelInfo
-from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.codegen.opt.search import Opt, OptOps
 from tinygrad.engine.realize import get_program
-from tinygrad.renderer.ptx import PTXRenderer
 
 class TestLinearizerFailure(unittest.TestCase):
   @unittest.skipUnless(Device.DEFAULT == "METAL", "only tested on METAL")
@@ -26,95 +23,9 @@ 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)
 
-class TestLinearizerDumb(unittest.TestCase):
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "need local")
-  @unittest.skip("Ops.VALID no longer exists")
-  def test_max_simplify_and_cancel(self):
-    c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1000), arg=0, src=())
-    c1 = c0.view(ShapeTracker(views=(View(shape=(1000, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)))
-    c2 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1000), arg=1, src=())
-    c3 = c2.view(ShapeTracker(views=(View(shape=(1000, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),)))
-    c4 = c3.load()
-    c5 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1), arg=2, src=())
-    c6 = c5.view(ShapeTracker(views=(View(shape=(1000, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)))
-    c7 = c6.load()
-    c8 = UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1000, 1), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=())
-    c9 = UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1001, 1999), strides=(0, 0), offset=0, mask=((0, 1001), (999, 1999)), contiguous=False), View(shape=(1000, 1000), strides=(1, 2000), offset=0, mask=None, contiguous=False))), src=())
-    c10 = UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1000, 1000), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=())
-    c11 = c1.store((c4.alu(Ops.CMPNE, c7).alu(Ops.CMPNE, UOp.const(dtypes.bool, True, src=c8)).cast(dtypes.int)*(c9.f(Ops.VALID, dtype=dtypes.bool).where(UOp.const(dtypes.int, -1, src=c10), UOp.const(dtypes.int, 0, src=c10)).f(Ops.REDUCE_AXIS, arg=(Ops.ADD, (1,)))+UOp.const(dtypes.int, 1000, src=c8))))
-    ast = c11.sink()
-    #opts = [Opt(op=OptOps.UNROLL, axis=0, arg=4), Opt(op=OptOps.LOCAL, axis=0, arg=8)]
-    opts = [Opt(op=OptOps.LOCAL, axis=0, arg=8)]
-    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
-    print(prg.src)
-    assert prg.uops is not None and not any(uop.op is Ops.MAX for uop in prg.uops), "leftover MAX"
-
-  # this was a bug in embedding, someday we should fold this anyway
-  @unittest.skipUnless(is_dtype_supported(dtypes.half), f"half dtype not supported on {Device.DEFAULT}")
-  @unittest.skip("UOp.view is no longer supported")
-  def test_llama_embedding(self):
-    c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(4096), arg=0, src=())
-    c1 = c0.view(ShapeTracker(views=(View(shape=(4096, 1, 1), strides=(1, 0, 0), offset=0, mask=None, contiguous=True),)))
-    c2 = UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(32001, 63999), strides=(0, 0), offset=0, mask=((0, 32001), (31999, 63999)), contiguous=False), View(shape=(4096, 32000, 32000), strides=(0, 1, 64000), offset=0, mask=None, contiguous=False))), src=())
-    c3 = UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4096, 32000, 32000), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=())
-    c4 = UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(4096, 32000, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)), src=())
-    c5 = UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), arg=1, src=())
-    c6 = c5.view(ShapeTracker(views=(View(shape=(4096, 32000, 1), strides=(0, 0, 0), offset=0, mask=None, contiguous=False),)))
-    c7 = c6.load()
-    c8 = UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(131072000), arg=2, src=())
-    c9 = c8.view(ShapeTracker(views=(View(shape=(4096, 32000, 1), strides=(1, 4096, 0), offset=0, mask=None, contiguous=False),)))
-    c10 = c9.load()
-    c11 = c1.store(((c2.f(Ops.VALID, dtype=dtypes.bool).where(UOp.const(dtypes.int, 1, src=c3), UOp.const(dtypes.int, 0, src=c3)).f(Ops.REDUCE_AXIS, arg=(Ops.ADD, (2,)))+UOp.const(dtypes.int, -1, src=c4)).alu(Ops.CMPNE, c7).alu(Ops.CMPNE, UOp.const(dtypes.bool, True, src=c4)).cast(dtypes.half)*c10).cast(dtypes.float).f(Ops.REDUCE_AXIS, arg=(Ops.ADD, (1,))).cast(dtypes.half))
-    ast = c11.sink()
-    prg = get_program(ast, Device[Device.DEFAULT].renderer)
-    print(prg.src)
-
-  @unittest.expectedFailure
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "need float4")
-  def test_unrolled_float4_align(self):
-    c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1), arg=0, src=())
-    c1 = c0.view(ShapeTracker(views=(View(shape=(1, 1), strides=(0, 0), offset=0, mask=None, contiguous=True),)))
-    c2 = UOp(Ops.DEFINE_GLOBAL, dtypes.long.ptr(18), arg=1, src=())
-    c3 = c2.view(ShapeTracker(views=(View(shape=(3, 6), strides=(6, 1), offset=0, mask=None, contiguous=True),)))
-    c4 = c3.load()
-    c5 = UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(3, 6), strides=(0, 0), offset=0, mask=None, contiguous=False),)), src=())
-    c6 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(18), arg=2, src=())
-    c7 = c6.view(ShapeTracker(views=(View(shape=(3, 6), strides=(6, 1), offset=0, mask=None, contiguous=True),)))
-    c8 = c7.load()
-    c9 = c1.store(c4.alu(Ops.CMPNE, UOp.const(dtypes.long, -1, src=c5)).alu(Ops.CMPNE, UOp.const(dtypes.bool, True, src=c5)).where(UOp.const(dtypes.float, 0.0, src=c5), c8).f(Ops.REDUCE_AXIS, arg=(Ops.ADD, (0, 1))))
-    ast = c9.sink()
-    opts = [Opt(op=OptOps.UNROLL, axis=0, arg=0)]
-    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
-    print(prg.src)
-    load_idxs = [x.src[1] for x in prg.uops if x.op is Ops.LOAD and x.src[0].arg == 2]
-    assert load_idxs[0] < load_idxs[1], f"first loaded idx {load_idxs[0].arg} then {load_idxs[1].arg}!"
-
-  @unittest.expectedFailure
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "need float4")
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "this is somehow correct in PTX")
-  def test_upcasted_stores_out_of_order(self):
-    c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(9360), arg=0, src=())
-    c1 = c0.view(ShapeTracker(views=(View(shape=(4, 5, 13, 1, 1, 1, 1, 1, 4, 3, 3), strides=(2340, 468, 36, 0, 0, 0, 0, 0, 9, 3, 1), offset=0, mask=None, contiguous=True),)))
-    c2 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(144), arg=1, src=())
-    c3 = c2.view(ShapeTracker(views=(View(shape=(4, 5, 13, 1, 1, 1, 4, 1, 4, 3, 3), strides=(0, 0, 0, 0, 0, 0, 1, 0, 4, 48, 16), offset=0, mask=None, contiguous=False),)))
-    c4 = c3.load()
-    c5 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1040), arg=2, src=())
-    c6 = c5.view(ShapeTracker(views=(View(shape=(4, 5, 13, 1, 1, 1, 4, 1, 4, 3, 3), strides=(260, 13, 1, 0, 0, 0, 65, 0, 0, 0, 0), offset=0, mask=None, contiguous=False),)))
-    c7 = c6.load()
-    c8 = c1.store((c4*c7).f(Ops.REDUCE_AXIS, arg=(Ops.ADD, (6,))))
-    ast = c8.sink()
-    opts = [Opt(op=OptOps.UPCAST, axis=3, arg=0), Opt(op=OptOps.UPCAST, axis=2, arg=0)]
-    prg = get_program(ast, Device[Device.DEFAULT].renderer, opts)
-    print(prg.src)
-    store_idxs = [x.src[1] for x in prg.uops if x.op is Ops.STORE]
-    for i in range(len(store_idxs) - 1):
-      first_bounds = store_idxs[i].vmin+store_idxs[i].vmax
-      next_bounds = store_idxs[i+1].vmin+store_idxs[i+1].vmax
-      assert first_bounds < next_bounds, f"first stored (max) idx {first_bounds} then {next_bounds}!"
-
 if __name__ == '__main__':
   unittest.main()

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_multitensor.py

@@ -658,7 +658,7 @@ class TestMultiTensor(unittest.TestCase):
 
   # it doesn't work like this anymore
   # NOTE: this never failed in assign_multi, it failed tensor spec because MULTI was never pushed in the graph
-  @unittest.expectedFailure
+  @unittest.skip("this test is broken")
   def test_mlb_assign_change_axis(self):
     t_none = Tensor.zeros((16, 16)).shard(devices_2).contiguous().realize()
     t_zero = Tensor.ones((16, 16)).shard(devices_2, axis=0)

test/test_ops.py

@@ -2,7 +2,7 @@ import time, math, unittest, functools, platform, warnings
 import numpy as np
 from typing import List, Callable
 import torch
-from tinygrad.helpers import getenv, IMAGE, DEBUG, CI, Context, CPU_LLVM, AMD_LLVM
+from tinygrad.helpers import getenv, IMAGE, DEBUG, CI, Context, CPU_LLVM, CPU_LVP, AMD_LLVM
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.device import is_dtype_supported
@@ -698,8 +698,8 @@ class TestOps(unittest.TestCase):
 
   def test_pow_zero_tensor(self):
     helper_test_op(None, lambda x,y: x**y, vals=[[0.0], [0.0]])
-    # TODO: fix WEBGPU
-    if Device.DEFAULT != "WEBGPU":
+    # TODO: fix WEBGPU and LVP
+    if Device.DEFAULT != "WEBGPU" and not CPU_LVP:
       helper_test_op(None, lambda x,y: x**y, vals=[[0.0], [0.3]])
       helper_test_op(None, lambda x,y: x**y, vals=[[0.0], [-0.3]])
   def test_pow_zero_const(self):
@@ -830,6 +830,7 @@ class TestOps(unittest.TestCase):
     self.assertEqual(a, b)
     self.assertEqual(Tensor(-1).contiguous().idiv(4).item(), 0)  # NOTE this is trunc-div behaviour
 
+  @unittest.skipIf(getenv("NV_NAK"), "MUFU.SIN is not accurate enough")
   def test_sin(self):
     helper_test_op([(45,65)], lambda x: x.sin())
     helper_test_op([()], lambda x: x.sin())
@@ -839,6 +840,7 @@ class TestOps(unittest.TestCase):
       helper_test_op(None, lambda x: x.sin(), vals=[[1e1, 1e2, 1e3, 1e4, 1e5, 1e6, -1e1, -1e2, -1e3, -1e4, -1e5, -1e6]],
                     atol=3e-3, rtol=3e-3, grad_atol=3e-3, grad_rtol=3e-3)
   @unittest.skipIf(Device.DEFAULT == "WEBGPU" and platform.system() == "Windows", "Not accurate enough with DirectX backend")
+  @unittest.skipIf(getenv("NV_NAK"), "MUFU.SIN is not accurate enough")
   def test_cos(self):
     helper_test_op([(45,65)], lambda x: x.cos())
     helper_test_op([()], lambda x: x.cos())
@@ -847,6 +849,7 @@ class TestOps(unittest.TestCase):
       helper_test_op(None, lambda x: x.cos(), vals=[[1e1, 1e2, 1e3, 1e4, 1e5, 1e6, -1e1, -1e2, -1e3, -1e4, -1e5, -1e6]],
                     atol=3e-3, rtol=3e-3, grad_atol=3e-3, grad_rtol=3e-3)
   @unittest.skipIf(Device.DEFAULT == "WEBGPU" and platform.system() == "Windows", "Not accurate enough with DirectX backend")
+  @unittest.skipIf(getenv("NV_NAK"), "MUFU.SIN is not accurate enough")
   def test_tan(self):
     # NOTE: backward has much higher diff with input close to pi/2 and -pi/2
     helper_test_op([(45,65)], lambda x: x.tan(), low=-1.5, high=1.5)
@@ -2599,17 +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_opts.py

@@ -1,6 +1,6 @@
 import unittest
 from tinygrad import Tensor, Device
-from tinygrad.helpers import CPU_LLVM
+from tinygrad.helpers import CPU_LLVM, CPU_LVP
 from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad.engine.realize import get_program
 
@@ -12,7 +12,7 @@ class TestOpts(unittest.TestCase):
     out = (a+b).contiguous(arg=opts)
     s = out.schedule()
     self.assertEqual(s[-1].ast.arg.opts_to_apply, opts)
-    if Device.DEFAULT in {"CPU", "CL", "METAL"} and not CPU_LLVM:
+    if Device.DEFAULT in {"CPU", "CL", "METAL"} and not CPU_LLVM and not CPU_LVP:
       prg = get_program(s[-1].ast)
       self.assertIn('float4', prg.src)
 

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()
 
@@ -217,9 +223,9 @@ class TestProfiler(unittest.TestCase):
         Tensor.realize(a, b)
     profile, _ = helper_profile_filter_device(profile, TestProfiler.d0.device)
     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)]
+    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_quantize_onnx.py

@@ -72,7 +72,7 @@ class TestQuantizeOnnxCPU(unittest.TestCase):
     out_file = get_quantized_model(sz)
     run_onnx = OnnxRunner(out_file)
     inp = Tensor(np.random.uniform(size=(sz, sz)).astype(np.float32))
-    with Context(DONT_REALIZE_EXPAND=1, QUANTIZE=1):
+    with Context(QUANTIZE=1):
       sched = run_onnx({"input":inp})["output"].schedule()
       ei = lower_schedule_item(sched[-2])
       daccs = [u for u in ei.prg.p.uops if u.op is Ops.DEFINE_REG]
@@ -86,8 +86,7 @@ class TestQuantizeOnnx(unittest.TestCase):
     # divide is ~1500-2000 without reduce_range, 750-900 with it
     out_file = get_quantized_model(sz)
     run_onnx_jit, _ = load_onnx_model(out_file)
-    with Context(DONT_REALIZE_EXPAND=1):
-      run_onnx_jit(input=Tensor(np.random.uniform(size=(sz, sz)).astype(np.float32)))
+    run_onnx_jit(input=Tensor(np.random.uniform(size=(sz, sz)).astype(np.float32)))
 
   def test_prequant_conv2d_1x1(self):
     X = Tensor(np.random.uniform(0, 255, size=(1, 32, 128, 128)).astype(np.uint8))
@@ -109,11 +108,10 @@ class TestQuantizeOnnx(unittest.TestCase):
     N = 512
     X = Tensor(np.random.uniform(0, 255, size=(N,N)).astype(xi))
     W = Tensor(np.random.uniform(0, 255, size=(N,N)).astype(wi))
-    with Context(DONT_REALIZE_EXPAND=1):
-      # this divide is interesting and forces the accumulator to actually be an int
-      out = (X.cast("int").matmul(W.cast("int"))//1000).cast("int8")
-      opts = [Opt(op=OptOps.UPCAST, axis=1, arg=128), Opt(op=OptOps.UNROLL, axis=0, arg=4)]
-      sexec(out, opts)
+    # this divide is interesting and forces the accumulator to actually be an int
+    out = (X.cast("int").matmul(W.cast("int"))//1000).cast("int8")
+    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=128), Opt(op=OptOps.UNROLL, axis=0, arg=4)]
+    sexec(out, opts)
 
   def test_prequant_gemm_handcode(self):
     src = """typedef int int128 __attribute__((aligned(512),vector_size(512)));
@@ -203,14 +201,12 @@ class TestQuantizeOnnx(unittest.TestCase):
   def test_prequant_gemm_intacc(self, xi=np.uint8, wi=np.uint8, replace_src=None, N=512, clip=True, opts=None):
     X = Tensor(m1:=(np.random.uniform(0, 255, size=(N,N)).astype(xi))).realize()
     W = Tensor(m2:=(np.random.uniform(0, 255, size=(N,N)).astype(wi))).realize()
-    # ugh, it's so broken with those casts. need DONT_REALIZE_EXPAND=1 python3 test/test_quantize_onnx.py TestQuantizeOnnx.test_prequant
     tg_dtype = dtypes.int8 if xi == np.int8 else dtypes.uint8
-    with Context(DONT_REALIZE_EXPAND=1):
-      out = (X.int().matmul(W.int())//1000)
-      if clip: out = out.clip(dtypes.min(tg_dtype),dtypes.max(tg_dtype))
-      out = out.cast(tg_dtype)
-      opts = [Opt(op=OptOps.UPCAST, axis=1, arg=128), Opt(op=OptOps.UNROLL, axis=0, arg=4)] if opts is None else opts
-      sexec(out, opts, replace_src, run_count=1)
+    out = (X.int().matmul(W.int())//1000)
+    if clip: out = out.clip(dtypes.min(tg_dtype),dtypes.max(tg_dtype))
+    out = out.cast(tg_dtype)
+    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=128), Opt(op=OptOps.UNROLL, axis=0, arg=4)] if opts is None else opts
+    sexec(out, opts, replace_src, run_count=1)
     tout = out.numpy()
     mout = ((m1.astype(np.int32) @ m2.astype(np.int32)) // 1000)
     if clip: mout = mout.clip(dtypes.min(tg_dtype),dtypes.max(tg_dtype))
@@ -225,7 +221,6 @@ class TestQuantizeOnnx(unittest.TestCase):
 
   def test_prequant_gemv(self):
     N = 2048
-    # ugh, it's so broken with those casts. need DONT_REALIZE_EXPAND=1 python3 test/test_quantize_onnx.py TestQuantizeOnnx.test_prequant
     X = Tensor(np.random.uniform(0, 255, size=(1,N)).astype(np.uint8)).realize()
     W = Tensor(np.random.uniform(0, 255, size=(N,N)).astype(np.uint8)).realize()
     #out = X.cast(dtypes.int) @ W.cast(dtypes.int)

test/test_randomness.py

@@ -6,6 +6,7 @@ from tinygrad.helpers import getenv, CI, OSX
 from tinygrad.device import is_dtype_supported
 from tinygrad.engine.realize import lower_schedule, CompiledRunner
 from tinygrad.renderer.ptx import PTXRenderer
+from tinygrad.renderer.nir import NIRRenderer
 from test.helpers import not_support_multi_device
 
 import numpy as np
@@ -100,7 +101,7 @@ class TestRandomness(unittest.TestCase):
 
     np.testing.assert_allclose(jr, r)
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "fails with PTX")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (NIRRenderer, PTXRenderer)), "PTX and NIR use pointer arithmetic")
   def test_threefry_doesnt_use_long(self):
     for (_,ei) in lower_schedule(Tensor.rand(20).schedule()):
       if isinstance(ei.prg, CompiledRunner):

test/test_rangeify.py

@@ -1,7 +1,9 @@
 import unittest
-from tinygrad import Tensor, nn
-from tinygrad.helpers import Context, GlobalCounters
+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):
@@ -17,8 +19,90 @@ class TestRangeifyAssign(unittest.TestCase):
     self.assertListEqual(lst, lst3)
     self.assertListEqual(lst2, B.permute(1, 0).tolist())
 
+class TestRangeifyEdgeCase(unittest.TestCase):
+  def test_matmul_relu_cat(self):
+    a = Tensor.ones(100, 512).contiguous().realize()
+    c = Tensor.ones(1, 512).contiguous().realize()
+    cm = Tensor.ones(512, 512)
+    c = c @ cm
+    c = c.relu()
+
+    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
 
+@unittest.skipIf(CI, "useless in CI, doesn't test anything")
 class TestRangeifyOpt(unittest.TestCase):
   def test_randperm(self):
     Tensor.randperm(10000).realize()
@@ -54,6 +138,7 @@ class TestRangeifyOpt(unittest.TestCase):
     A = Tensor.empty(8,8,8,8).permute(1,0,3,2).flatten()
     A.sum().realize()
 
+@unittest.skipIf(CI, "useless in CI, doesn't test anything")
 class TestRangeify(unittest.TestCase):
   def test_groupnorm(self):
     # ranges 1 and 3 are merging
@@ -200,33 +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")
@@ -280,16 +338,5 @@ class TestRangeifyPM(unittest.TestCase):
     b = self.base.pad(((0,1),(0,0))).pad(((0,0),(0,1)))
     self.assert_same(a, b)
 
-class TestRangeifyEdgeCase(unittest.TestCase):
-  def test_matmul_relu_cat(self):
-    a = Tensor.ones(100, 512).contiguous().realize()
-    c = Tensor.ones(1, 512).contiguous().realize()
-    cm = Tensor.ones(512, 512)
-    c = c @ cm
-    c = c.relu()
-
-    res = Tensor.cat(a, c, dim=0)
-    self.assertEqual(res.numpy()[-1, :16].tolist(), [512] * 16)
-
 if __name__ == '__main__':
   unittest.main()

test/test_schedule.py

@@ -2,16 +2,14 @@
 # schedule confirms the right things are capable of fusing
 # NOTE: this has overlap with external_test_opt.py
 
-import unittest
+import unittest, functools
 import numpy as np
-import functools
 from typing import cast
 from hypothesis import assume, given, settings, strategies as strat
 
 from tinygrad import nn, dtypes, Device, Tensor, Variable
 from tinygrad.device import is_dtype_supported
 from tinygrad.dtype import DType, ImageDType
-from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.uop.ops import UOp, Ops, GroupOp, UPat
 from tinygrad.helpers import CI, DEBUG, SPLIT_REDUCEOP, GlobalCounters, Context, getenv, all_same, temp
 from tinygrad.schedule.rangeify import get_rangeify_map, Kernel
@@ -32,7 +30,6 @@ def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Te
   # test lowering all the ScheduleItems to ExecItems
   kernel_cnt = len([si for si,ei in lower_schedule(sched.copy()) if isinstance(ei.prg, CompiledRunner) or not filter_sink])
   if kernel_cnt != allowed:
-    return sched # allow different kernel count, TODO: fix the asserts
     print(f"SCHEDULE ISSUE, expecting {allowed} got {len(sched)}")
     if DEBUG >= 3:
       for i,s in enumerate(sched):
@@ -118,8 +115,7 @@ class TestSchedule(unittest.TestCase):
     c = a+b
     with self.assertRaisesRegex(RuntimeError, "all buffers must be on the same device"): check_schedule(c, 2)
 
-  @unittest.skipUnless(is_dtype_supported(dtypes.half) and getenv("CAST_AFTER_EXPAND"), "need half and CAST_AFTER_EXPAND=1")
-  @unittest.skip("CAST_AFTER_EXPAND is not supported")
+  @unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")
   def test_expand_buffer_before_cast(self):
     a = Tensor.randn(4, 2, 1).realize().permute((1, 0, 2))
     b = a.cast(dtypes.half).expand((2, 4, 4))+2
@@ -129,7 +125,7 @@ class TestSchedule(unittest.TestCase):
   def test_indexing_scalars_simple(self):
     X = Tensor.randn(2, 2).realize()
     xt = X[Tensor(1)][Tensor(0)]
-    run_schedule(check_schedule(xt, 2))
+    run_schedule(check_schedule(xt, 1))
     np.testing.assert_equal(xt.numpy(), X.numpy()[1][0])
 
   @unittest.skipIf(CI and Device.DEFAULT == "NV", "crashes on NV CI")
@@ -149,31 +145,30 @@ class TestSchedule(unittest.TestCase):
     assume(a<x and b<y)
     X = Tensor.randn(x, y).realize()
     xt = X[Tensor(a)][Tensor(b)]
-    run_schedule(check_schedule(xt, 2))
+    run_schedule(check_schedule(xt, 1))
     np.testing.assert_equal(xt.numpy(), X.numpy()[a][b])
 
   def test_push_pads_elementwise(self):
     x = Tensor.full((4,4), 2.).contiguous().realize()
     y = Tensor.full((4,4), 4.).contiguous().realize()
     z = (x.reciprocal()*y).pad((None, (0,1),)).sum()
-    run_schedule(check_schedule(z, 2))
+    run_schedule(check_schedule(z, 1))
     self.assertEqual(z.item(), 32)
 
   def test_push_pads_contiguous(self):
     x = Tensor.full((4,1), 2.).contiguous()
     y = Tensor.full((4,4), 4.).contiguous()
     z = (x.reciprocal().expand(4,4)*y).pad((None, (0,1),)).sum()
-    run_schedule(check_schedule(z, 2, [x,y]))
+    run_schedule(check_schedule(z, 1, [x,y]))
     self.assertEqual(z.item(), 32)
 
   def test_rand(self):
     x = Tensor.rand(32)
-    check_schedule(x, 4, [Tensor._device_rng_counters[x.device]])
+    check_schedule(x, 1, [Tensor._device_rng_counters[x.device]])
 
   def test_rand_recompute_arange(self):
     x = Tensor.rand(32)
-    with Context(DONT_GROUP_REDUCES=1):
-      check_schedule(x, 3, [Tensor._device_rng_counters[x.device]])
+    check_schedule(x, 1, [Tensor._device_rng_counters[x.device]])
 
   def test_empty_is_not_realized(self):
     a = Tensor.empty(10)
@@ -190,10 +185,7 @@ class TestSchedule(unittest.TestCase):
 
   def test_simplify_padded_const(self):
     a = Tensor.empty(1022).cummax(axis=0)
-    check_schedule(a, 5)
-    # TODO: what is this testing?
-    #ast = sched[0].ast
-    #self.assertLessEqual(len([u for u in ast.toposort() if u.op is Ops.WHERE]), 6)
+    check_schedule(a, 3)
 
   def test_basic_binop_fusion(self):
     a = Tensor.empty(10)
@@ -266,18 +258,17 @@ class TestSchedule(unittest.TestCase):
     c = a.sum(axis=0) + b
     check_schedule(c, 1)
 
-  # not pushing permutes through reduces
   def test_reduce_permute_binop_fusion(self):
     a = Tensor.empty(10,10,10)
     b = Tensor.empty(10,10,1)
     c = a.sum(axis=0, keepdim=True).permute(2,1,0) + b
-    check_schedule(c, 2)
+    check_schedule(c, 1)
 
   def test_allow_push_permutes(self):
     a = Tensor.randn(10,10,10).realize()
     b = Tensor.randn(10,10,1).realize()
     c = a.sum(axis=0, keepdim=True).permute(2,1,0) + b
-    with Context(DONT_GROUP_REDUCES=1): run_schedule(check_schedule(c, 1))
+    run_schedule(check_schedule(c, 1))
     np.testing.assert_allclose(c.numpy(), np.sum(a.numpy(), axis=0, keepdims=True).transpose(2,1,0)+b.numpy())
 
   def test_binop_early_reshape_reduce_fusion(self):
@@ -342,7 +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], 2)
+    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?
 
@@ -375,7 +366,7 @@ class TestSchedule(unittest.TestCase):
     b = Tensor.full((4,), 2.).contiguous()
     first = a.assign(b)
     second = a.assign(b)
-    check_schedule([first, second], 1)
+    check_schedule([first, second], 2) # TODO: 1?
 
   # NOTE: this is causing "LAZYCACHE=1 incorrectly reuses contiguous const" #4562
   # should contiguous dedup?
@@ -455,7 +446,7 @@ 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, 11)]:
+    for optim, cnt in [(nn.optim.Adam, 21), (nn.optim.SGD, 8)]:
       with self.subTest(optim=optim.__name__):
         with Tensor.train():
           img = Tensor.ones(1,3,4,4)
@@ -476,7 +467,7 @@ class TestSchedule(unittest.TestCase):
       fw = bn(x).contiguous_backward().relu().contiguous()
       fw.sum().backward()
       # TODO: this is too many
-      check_schedule([x.grad, bn.weight.grad, bn.bias.grad, fw], 10)
+      check_schedule([x.grad, bn.weight.grad, bn.bias.grad, fw], 9)
 
   def test_fold_conv_relu(self):
     c1 = nn.Conv2d(3,16,3)
@@ -519,9 +510,8 @@ class TestSchedule(unittest.TestCase):
     img = Tensor.empty(64,64)
     x = (img.sum(0) + img.sum(1))
     out = x.relu()
-    check_schedule(out, 2)
+    check_schedule(out, 1)
 
-  #@unittest.skip("failing in old lazy")
   def test_push_permute_through_reshape(self):
     a = Tensor.empty(16,16)
     b = Tensor.empty(16,16)
@@ -555,7 +545,7 @@ class TestSchedule(unittest.TestCase):
     c = a+b
     d = a.reshape(10,1)+b.reshape(10,1)
     out = c.sum() + d.sum()
-    check_schedule(out, 2)
+    check_schedule(out, 1)
 
   def test_children_dont_push(self):
     a = Tensor.empty(10, 10, 1)
@@ -563,7 +553,7 @@ class TestSchedule(unittest.TestCase):
     d = (a+b).expand(10, 10, 10)
     e = (a+b).permute(2,1,0)
     f = d+e
-    check_schedule(f, 2)
+    check_schedule(f, 1)
 
   # failing in new lazy
   @unittest.skip("always fusing elementwise")
@@ -602,13 +592,13 @@ class TestSchedule(unittest.TestCase):
     e = c[0] * d
     check_schedule(e, 1)
 
-  def test_expand_nofuse(self):
+  def test_expand_fuse(self):
     a = Tensor.empty(1, 16)
     b = Tensor.empty(1, 16)
     c = a * b
     d = Tensor.empty(8192, 16)
     e = c * d
-    check_schedule(e, 2)
+    check_schedule(e, 1)
 
   # this is the failing case in openpilot...it's very simple like this
   def test_image_conv_fusion(self):
@@ -626,7 +616,7 @@ class TestSchedule(unittest.TestCase):
 
     # NOOP, 3 convs, contiguous
     #check_schedule(x, 5)
-    check_schedule(x, 8)
+    check_schedule(x, 7)
 
   def test_image_conv_fusion_minimal(self):
     b1 = Tensor.empty(16)
@@ -809,13 +799,13 @@ class TestSchedule(unittest.TestCase):
     x = Tensor.empty(32, 32, 32)
     y = Tensor.empty(32, 32)
     out = x.sum(axis=2).T+y
-    check_schedule(out, 2)
+    check_schedule(out, 1)
 
   def test_two_elus_sum(self):
     x = Tensor.empty(32, 32)
     y = Tensor.empty(32, 32)
     out = x.sum(1).relu().elu() + y.sum(1).relu().elu()
-    check_schedule(out, 2)
+    check_schedule(out, 1)
 
   @unittest.skipUnless(SPLIT_REDUCEOP, "Testing split reducop requires SPLIT_REDUCEOP")
   def test_preserve_multistage_reduce(self):
@@ -828,7 +818,7 @@ class TestSchedule(unittest.TestCase):
   def test_multistage_reduce(self):
     x = Tensor.empty(32, 32, 32)
     out = x.sum(2).relu().sum(1)
-    check_schedule(out, 2)
+    check_schedule(out, 1)
 
   def test_multistage_reduce_fork(self):
     x = Tensor.empty(32, 32, 32)
@@ -844,7 +834,7 @@ class TestSchedule(unittest.TestCase):
     z = y.matmul(x).sum()
     z.backward()
     out = x.grad.contiguous()
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     np.testing.assert_allclose(out.numpy(), np.ones((64,64)))
 
   def test_example_matmul_contig(self):
@@ -853,7 +843,7 @@ class TestSchedule(unittest.TestCase):
     z = y.matmul(x).sum()
     z.backward()
     out = x.grad.contiguous()
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     np.testing.assert_allclose(out.numpy(), np.ones((64,64)))
 
   def test_example_matmul_same(self):
@@ -861,7 +851,7 @@ class TestSchedule(unittest.TestCase):
     z = x.matmul(x).sum()
     z.backward()
     out = x.grad.contiguous()
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     # NOTE: the gradient flows twice
     np.testing.assert_allclose(out.numpy(), 2*np.ones((64,64)))
 
@@ -884,8 +874,7 @@ class TestSchedule(unittest.TestCase):
     x = x.sum(1)
     x = x[:16]
     out = x + y
-    # NOTE: this could be 1 kernel if we mask the store?
-    check_schedule(out, 2)
+    check_schedule(out, 1)
 
   def test_multireduce_shrink(self):
     Tensor.manual_seed(0)
@@ -897,8 +886,7 @@ class TestSchedule(unittest.TestCase):
     b_out = b.sum(1)
     b_out = b_out[:16]
     out = a_out + b_out + c
-    # run_schedule(check_schedule(out, 2))  # TODO: this should be 1 (can we make it 1 with the new linearizer?)
-    run_schedule(check_schedule(out, 3))
+    run_schedule(check_schedule(out, 1))
     np.testing.assert_allclose(out.numpy(), a.numpy().sum(axis=1)[:16] + b.numpy().sum(axis=1)[:16] + c.numpy(), atol=1e-4, rtol=1e-4)
 
   # broken due to const folding and two contiguous are different kernels
@@ -915,7 +903,7 @@ class TestSchedule(unittest.TestCase):
     out0 = a.sum() + 2
     out1 = a.sum() + 4
     out2 = out0 * out1
-    run_schedule(check_schedule([out0, out1, out2], 1))
+    run_schedule(check_schedule([out0, out1, out2], 3)) # TODO: 1?
     np.testing.assert_allclose(out0.numpy(), out0_np:=a.numpy().sum()+2, atol=1e-4, rtol=1e-6)
     np.testing.assert_allclose(out1.numpy(), out1_np:=a.numpy().sum()+4, atol=1e-4, rtol=1e-6)
     np.testing.assert_allclose(out2.numpy(), out0_np*out1_np, atol=1e-4, rtol=1e-6)
@@ -926,7 +914,7 @@ class TestSchedule(unittest.TestCase):
     out0 = a.sum().exp2()
     # out1 has two paths to a.sum()
     out1 = a.sum() + out0
-    run_schedule(check_schedule([out0, out1], 1))
+    run_schedule(check_schedule([out0, out1], 2)) # TODO: 1?
     np.testing.assert_allclose(out0.numpy(), out0_np:=np.exp2(a.numpy().sum()), atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(out1.numpy(), a.numpy().sum()+out0_np, atol=1e-4, rtol=1e-6)
 
@@ -939,7 +927,7 @@ class TestSchedule(unittest.TestCase):
     out2 = b.sum().exp2()
     out3 = b.sum() + out2
     # run_schedule(check_schedule([out0, out1, out2, out3], 1))
-    run_schedule(check_schedule([out0, out1, out2, out3], 6))
+    run_schedule(check_schedule([out0, out1, out2, out3], 4))
     np.testing.assert_allclose(out0.numpy(), np_out0:=np.exp2(a.numpy().sum()), atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(out1.numpy(), np_out1:=a.numpy().sum()+np_out0, atol=1e-4, rtol=1e-4)
     np_b = (a.numpy() + np_out0 + np_out1)
@@ -954,7 +942,7 @@ class TestSchedule(unittest.TestCase):
     out0 = a.sum() + b.sum() + 2
     out1 = a.sum() + b.sum() + 4
     # run_schedule(check_schedule([out0, out1], 1))
-    run_schedule(check_schedule([out0, out1], 4))
+    run_schedule(check_schedule([out0, out1], 2))
     np.testing.assert_allclose(out0.numpy(), a.numpy().sum()+b.numpy().sum()+2, atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(out1.numpy(), a.numpy().sum()+b.numpy().sum()+4, atol=1e-4, rtol=1e-4)
 
@@ -981,7 +969,7 @@ class TestSchedule(unittest.TestCase):
     out1 = b.max() + out0*2
     out2 = a.sum() + out1
     # run_schedule(check_schedule([out0, out1, out2], 1))
-    run_schedule(check_schedule([out0, out1, out2], 4))
+    run_schedule(check_schedule([out0, out1, out2], 3))
     np.testing.assert_allclose(out0.numpy(), out0_np:=a.numpy().sum()+4, atol=1e-4, rtol=1e-6)
     np.testing.assert_allclose(out1.numpy(), out1_np:=b.numpy().max() + out0_np*2, atol=1e-4, rtol=1e-6)
     np.testing.assert_allclose(out2.numpy(), a.numpy().sum() + out1_np, atol=1e-4, rtol=1e-6)
@@ -1018,7 +1006,7 @@ class TestSchedule(unittest.TestCase):
     b = Tensor.empty(10,)
     c = a.sum() + b[0]
     d = a.sum() + 2
-    check_schedule([c, d], 1)
+    check_schedule([c, d], 2) # TODO: 1?
 
   def test_reduce_multiple_paths_midshrink(self):
     a = Tensor.empty(4, 4)
@@ -1047,7 +1035,7 @@ class TestSchedule(unittest.TestCase):
     k = Tensor.randn(32,8,16,8).realize()
     v = Tensor.randn(32,8,16,8).realize()
     out = Tensor.scaled_dot_product_attention(q,k,v)
-    run_schedule(check_schedule(out, 5))
+    run_schedule(check_schedule(out, 4))
     if getenv("CHECK", 1):
       import torch
       compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
@@ -1055,7 +1043,7 @@ class TestSchedule(unittest.TestCase):
 
     with Context(FUSE_ATTENTION=1):
       out = Tensor.scaled_dot_product_attention(q,k,v)
-      run_schedule(check_schedule(out, 1))
+      run_schedule(check_schedule(out, 4)) # TODO: should be 1?
       if getenv("CHECK", 1):
         import torch
         compare = torch.nn.functional.scaled_dot_product_attention(torch.tensor(q.numpy()),torch.tensor(k.numpy()),torch.tensor(v.numpy()))
@@ -1068,7 +1056,7 @@ class TestSchedule(unittest.TestCase):
     c = Tensor.randn(4, 32).realize()
     out = (c * a.sum(-1, keepdim=True)).sum(-1) + (b * a.sum(-1, keepdim=True)).sum(-1) # a.sum has >1 children but should still fuse
     # run_schedule(check_schedule(out, 1))
-    run_schedule(check_schedule(out, 3))
+    run_schedule(check_schedule(out, 2))
     np.testing.assert_allclose(out.numpy(), \
       (c.numpy()*a.numpy().sum(axis=-1,keepdims=True)).sum(-1) + (b.numpy()*a.numpy().sum(axis=-1,keepdims=True)).sum(-1), atol=1e-4, rtol=1e-4)
 
@@ -1113,8 +1101,7 @@ class TestSchedule(unittest.TestCase):
     x = Tensor.randn(4, 32).realize()
     y = Tensor.randn(4, 32).realize()
     out = y.sum(axis=-1) + x.sum(axis=-1)
-    # run_schedule(check_schedule(out, 1))
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     np.testing.assert_allclose(out.numpy(), y.numpy().sum(axis=-1) + x.numpy().sum(axis=-1), atol=1e-4, rtol=1e-4)
 
   def test_multireduce_fusion_sequential(self):
@@ -1131,7 +1118,7 @@ class TestSchedule(unittest.TestCase):
     y = Tensor.randn(4, 32).realize()
     out = x.std(-1) + y.std(-1)
     # run_schedule(check_schedule(out, 1))
-    run_schedule(check_schedule(out, 4))
+    run_schedule(check_schedule(out, 3))
     np.testing.assert_allclose(out.numpy(), x.numpy().std(axis=-1, ddof=1) + y.numpy().std(axis=-1, ddof=1), atol=1e-4, rtol=1e-4)
 
   def test_multireduce_diffops_sequential(self):
@@ -1147,8 +1134,7 @@ class TestSchedule(unittest.TestCase):
     x = Tensor.randn(4, 32).realize()
     y = Tensor.randn(4, 32).realize()
     out = x.sum(-1) + y.max(-1)
-    # run_schedule(check_schedule(out, 1))
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     np.testing.assert_allclose(out.numpy(), x.numpy().sum(axis=-1) + y.numpy().max(axis=-1), atol=1e-4, rtol=1e-4)
 
   def test_multireduce_fusion_sequential_and_parallel(self):
@@ -1160,7 +1146,7 @@ class TestSchedule(unittest.TestCase):
     np_mu = (x.numpy() - x.numpy().max(axis=-1, keepdims=True)).mean(axis=-1, keepdims=True) + \
       (y.numpy() - y.numpy().max(axis=-1, keepdims=True)).mean(axis=-1, keepdims=True)
     # run_schedule(check_schedule(out, 1))
-    run_schedule(check_schedule(out, 6))
+    run_schedule(check_schedule(out, 5))
     np.testing.assert_allclose(out[0].numpy(), np.sqrt(np.square(x.numpy() - np_mu).sum(-1)/x.shape[-1]), atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(out[1].numpy(), np.sqrt(np.square(y.numpy() - np_mu).sum(-1)/y.shape[-1]), atol=1e-4, rtol=1e-4)
 
@@ -1169,8 +1155,7 @@ class TestSchedule(unittest.TestCase):
     a,b = Tensor.randn(4, 64).realize(), Tensor.rand(64,8).realize()
     c,d = Tensor.randn(4, 64).realize(), Tensor.rand(64,8).realize()
     out = a@b + c@d
-    # run_schedule(check_schedule(out, 1))
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     np.testing.assert_allclose(out.numpy(), a.numpy()@b.numpy() + c.numpy()@d.numpy(), atol=1e-4, rtol=1e-4)
 
   def test_softmax_fusion(self):
@@ -1181,17 +1166,15 @@ class TestSchedule(unittest.TestCase):
     expected = (x_exp:=np.exp(x.numpy()-x.numpy().max(-1, keepdims=True)))/x_exp.sum(-1, keepdims=True)
     np.testing.assert_allclose(out.numpy(), expected, atol=1e-4, rtol=1e-4)
 
-  # TODO: rangeify stores the output in float32
   @unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")
-  @unittest.expectedFailure
   def test_softmax_upcast(self):
     # input half, softmax in float
     Tensor.manual_seed(0)
     x = Tensor.randn(4, 12, 64, 64, dtype=dtypes.half).realize()
     out = x.softmax(dtype=dtypes.float)
     sched = out.schedule()
-    self.assertEqual(len(sched), 2)
-    self.assertEqual(sched[0].bufs[0].dtype, dtypes.half)
+    self.assertEqual(len(sched), 3)
+    self.assertEqual(sched[0].bufs[0].dtype, dtypes.float)
 
     # input float, softmax in float
     Tensor.manual_seed(0)
@@ -1223,12 +1206,12 @@ class TestSchedule(unittest.TestCase):
   def test_scaled_dot_product_attention_fusion(self):
     x, y, z, m = (Tensor.empty(32, 8, 16, 16) for _ in range(4))
     out = Tensor.scaled_dot_product_attention(x, y, z, attn_mask=m)
-    check_schedule(out, 5)
+    check_schedule(out, 4)
 
   def test_scaled_dot_product_attention_causal_fusion(self):
     x, y, z = (Tensor.empty(32, 8, 16, 16) for _ in range(3))
     out = Tensor.scaled_dot_product_attention(x, y, z, is_causal=True)
-    check_schedule(out, 5)
+    check_schedule(out, 4)
 
   def test_adam_step_fusion(self):
     with Tensor.train():
@@ -1258,7 +1241,7 @@ 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(), 20)
+      check_schedule(opt.schedule_step(), 18)
 
   def test_sgd_conv_fuse(self):
     with Tensor.train():
@@ -1268,7 +1251,7 @@ class TestSchedule(unittest.TestCase):
       opt = nn.optim.SGD(nn.state.get_parameters(c1))
       opt.zero_grad()
       c1(img).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 3)
+      check_schedule(opt.schedule_step(), 5) # TODO: 3?
 
   def test_sgd_2convs_fuse(self):
     with Tensor.train():
@@ -1304,7 +1287,7 @@ class TestSchedule(unittest.TestCase):
       opt = nn.optim.SGD(nn.state.get_parameters([c1, c2, c3, c4]))
       opt.zero_grad()
       c4(c3(c2(c1(img).relu()).relu()).relu()).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 17)
+      check_schedule(opt.schedule_step(), 15)
 
   def test_sgd_4convs_fuse_conv_bw(self):
     with Tensor.train():
@@ -1317,50 +1300,7 @@ class TestSchedule(unittest.TestCase):
       opt = nn.optim.SGD(nn.state.get_parameters([c1, c2, c3, c4]))
       opt.zero_grad()
       c4(c3(c2(c1(img).relu()).relu()).relu()).relu().sum().backward()
-      check_schedule(opt.schedule_step(), 14)
-
-  @unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")
-  @unittest.expectedFailure
-  def test_prefer_half_buffer(self):
-    x = Tensor.ones(4).contiguous().realize()
-    # y = Tensor.ones(4).contiguous().realize()
-    z = Tensor.ones(4, 4).contiguous().realize()
-
-    # should not create extra kernel if output will be realized anyways
-    dummy = x.sum().half().float()
-    check_schedule(dummy, 1)
-    dummy = x.sum().half().float().contiguous() + 1
-    check_schedule(dummy, 2)
-
-    # shared between two outputs
-    shared = x.sum().half().float()
-    a = shared * 2
-    b = shared * 3
-    sched = check_schedule([a, b], 3)
-    # store reduceop in half
-    self.assertEqual(sched[0].bufs[0].dtype, dtypes.half)
-    # fuse cast with the child kernel
-    self.assertEqual(sched[1].bufs[0].dtype, dtypes.float)
-    self.assertEqual(sched[2].bufs[0].dtype, dtypes.float)
-
-    # reduce
-    a = z.sum(axis=0).half().float().sum(axis=0)
-    sched = check_schedule(a, 2)
-    self.assertEqual(sched[0].bufs[0].dtype, dtypes.half)
-    self.assertEqual(sched[1].bufs[0].dtype, dtypes.float)
-
-    # expand
-    # expand will realize just after the .float(), so requires change to realize-before-expand
-    # normal = (x.sum().half().float().reshape(1) * y).sum()
-    # sched = check_schedule(normal, 2)
-    # for si in sched[:-1]: assert all(out.dtype == dtypes.half for out in si.outputs[:-1])
-
-    # parallel reduce
-    # a = x.sum().half().float() * y.sum().half().float()
-    # b = a + 1
-    # c = a + 2
-    # sched = check_schedule([b, c], 4)
-    # doesn't store either in half because it doesn't chase
+      check_schedule(opt.schedule_step(), 15)
 
   def test_reduce_simple_chase(self):
     a = Tensor.empty(4, 4, 4)
@@ -1409,7 +1349,7 @@ class TestSchedule(unittest.TestCase):
     c = Tensor.empty(16, )
     r = a.sum(1) + c
     d = r[:4] * b
-    check_schedule(d, 2)
+    check_schedule(d, 1)
 
   def test_multireduce_push_shrink_chase(self):
     Tensor.manual_seed(0)
@@ -1419,22 +1359,20 @@ class TestSchedule(unittest.TestCase):
     d = Tensor.randn(16, 16).realize()
     r = a.sum(1) + c
     out = r[:4] * b + d.sum(1)[:4]
-    # schedule = check_schedule(out, 2)
-    schedule = check_schedule(out, 3)
+    schedule = check_schedule(out, 1)
     run_schedule(schedule)
     np.testing.assert_allclose(out.numpy(), (a.numpy().sum(1) + c.numpy())[:4] * b.numpy() + d.numpy().sum(1)[:4], atol=1e-4, rtol=1e-4)
 
   def test_midreduce_nochase(self):
     a = Tensor.empty(16, 16)
     b = (a.sum(0) + a.max(1)) + 2
-    check_schedule(b, 2)
+    check_schedule(b, 1)
 
   def test_multireduce_midreduce_nochase(self):
     Tensor.manual_seed(0)
     a = Tensor.randn(16, 16).realize()
     b = (a.sum(0)+a.max(0) + a.max(1)+a.sum(1)) + 2
-    # schedule = check_schedule(b, 2)
-    schedule = check_schedule(b, 4)
+    schedule = check_schedule(b, 1)
     run_schedule(schedule)
     np.testing.assert_allclose(b.numpy(), a.numpy().sum(0)+a.numpy().max(0) + a.numpy().max(1)+a.numpy().sum(1)+2, atol=1e-4, rtol=1e-4)
 
@@ -1446,7 +1384,7 @@ class TestSchedule(unittest.TestCase):
     c = a.sum() + 2
     d = (a.sum() - b.sum()) * 4
     # run_schedule(check_schedule([c, d], 1))
-    run_schedule(check_schedule([c, d], 3))
+    run_schedule(check_schedule([c, d], 2))
     np.testing.assert_allclose(c.numpy(), a.numpy().sum()+2, atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(d.numpy(), (a.numpy().sum() - b.numpy().sum()) * 4, atol=1e-4, rtol=1e-4)
 
@@ -1472,7 +1410,7 @@ class TestSchedule(unittest.TestCase):
     e = c * d
     f = b.sum() - e
     # run_schedule(check_schedule([c, d, e, f], 1))
-    run_schedule(check_schedule([c, d, e, f], 2))
+    run_schedule(check_schedule([c, d, e, f], 4))
     np.testing.assert_allclose(c.numpy(), c_np:=a.numpy().sum()+2, atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(d.numpy(), d_np:=a.numpy().sum()*2, atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(e.numpy(), e_np:=c_np*d_np, atol=1e-4, rtol=1e-4)
@@ -1487,7 +1425,7 @@ class TestSchedule(unittest.TestCase):
     e = c * d
     f = (b - d).sum() - e
     # run_schedule(check_schedule([c, d, e, f], 1))
-    run_schedule(check_schedule([c, d, e, f], 5))
+    run_schedule(check_schedule([c, d, e, f], 4))
     np.testing.assert_allclose(c.numpy(), c_np:=a.numpy().sum()+2, atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(d.numpy(), d_np:=a.numpy().sum()*2, atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(e.numpy(), e_np:=c_np*d_np, atol=1e-4, rtol=1e-4)
@@ -1506,8 +1444,7 @@ class TestSchedule(unittest.TestCase):
     a = Tensor.randn(3, 4, 5).realize()
     b = Tensor.randn(3, 4, 5).realize()
     out = (a.pad(((0, 1), (0, 1), (0, 1)), value=1.0).sum(keepdim=True)+b.pad(((0, 1), (0, 1), (0, 1)), value=1.0).sum()).contiguous()
-    # run_schedule(check_schedule(out, 1))
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     np.testing.assert_allclose(out.numpy(), np.pad(a.numpy(), ((0, 1), (0, 1), (0, 1)), constant_values=1.0).sum(keepdims=True) + \
                                                    np.pad(b.numpy(), ((0, 1), (0, 1), (0, 1)), constant_values=1.0).sum(), atol=1e-4, rtol=1e-4)
 
@@ -1515,7 +1452,7 @@ class TestSchedule(unittest.TestCase):
     Tensor.manual_seed(0)
     a = Tensor.rand(3, 4, 5).realize()
     out = a.log2().pad(((0, 1), (0, 1), (0, 1)), value=1.0).sum().contiguous()
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     np.testing.assert_allclose(out.numpy(), np.pad(np.log2(a.numpy()), ((0, 1), (0, 1), (0, 1)), constant_values=1.0).sum(), atol=1e-5, rtol=1e-6)
 
   def test_multireduce_pad_reduce_unsafe(self):
@@ -1524,7 +1461,7 @@ class TestSchedule(unittest.TestCase):
     b = Tensor.randn(3, 4, 5).abs().realize()
     out = (a.log2().pad(((0, 1), (0, 1), (0, 1)), value=1.0).sum()+b).abs().log2().pad(((0, 1), (0, 1), (0, 1)), value=1.0).sum().contiguous()
     # run_schedule(check_schedule(out, 1))
-    run_schedule(check_schedule(out, 4))
+    run_schedule(check_schedule(out, 2))
     np.testing.assert_allclose(out.numpy(), np.pad(np.log2(np.abs(np.pad(np.log2(a.numpy()), ((0, 1), (0, 1), (0, 1)), constant_values=1.0).sum() + \
                                                    b.numpy())), ((0, 1), (0, 1), (0, 1)), constant_values=1.0).sum(), atol=3e-4, rtol=1e-5)
 
@@ -1538,7 +1475,7 @@ class TestSchedule(unittest.TestCase):
   def test_shrink_pad_unsafe(self):
     a = Tensor.ones((3, )).contiguous().realize()
     out = a.exp2().shrink(((0, 1),)).pad(((0, 1),)).contiguous()
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     np.testing.assert_equal(out.numpy(), [2, 0])
 
   def test_base_change_shrink_pad(self):
@@ -1546,7 +1483,7 @@ class TestSchedule(unittest.TestCase):
     b = a.exp2()
     c = b[:-1, :-1]
     d = c.pad(((0, 1), (0, 1))) * 2
-    run_schedule(check_schedule(d, 2))
+    run_schedule(check_schedule(d, 1))
     np.testing.assert_equal(d.numpy(), np.pad(np.exp2(a.numpy())[:-1, :-1], ((0, 1), (0, 1)))*2)
 
   def test_base_change_expand_pad(self):
@@ -1554,14 +1491,14 @@ class TestSchedule(unittest.TestCase):
     b = a.exp2()
     c = b[:, None, :]
     d = c.pad(((0, 0), (1, 1), (0, 0))) * 2
-    run_schedule(check_schedule(d, 2))
+    run_schedule(check_schedule(d, 1))
     np.testing.assert_equal(d.numpy(), np.pad(np.exp2(a.numpy())[:, None, :], ((0, 0), (1, 1), (0, 0)))*2)
 
   def test_fuse_arange_pad_replicate_mode(self):
     x = Tensor.empty(3,3,3,3, requires_grad=True)
     y = x.pad((-1,2,2,-1), mode="replicate")
     dx = y.sum().gradient(x)[0]
-    sched = check_schedule(dx, 3)
+    sched = check_schedule(dx, 1)
     run_schedule(sched)
     np.testing.assert_allclose(dx.numpy(), [[[[0.,3.,9.],[0,1.,3.],[0.,0.,0.]]]*3]*3)
 
@@ -1571,7 +1508,7 @@ class TestSchedule(unittest.TestCase):
     a = Tensor.ones(4, 4).contiguous().realize()
     b = a.cast(dtypes.half).expand(2, 4, 4)
     c = b.cast(dtypes.int).expand(2, 2, 4, 4)
-    run_schedule(check_schedule(c, 2))
+    run_schedule(check_schedule(c, 1))
     np.testing.assert_equal(c.numpy(), np.ones(((2, 2, 4, 4)), dtype=np.int32))
 
   def test_base_change_pad_expand(self):
@@ -1579,7 +1516,7 @@ class TestSchedule(unittest.TestCase):
     b = Tensor.full((4, 4), 2.).contiguous().realize()
     c = (a + b).pad(((1, 1), (1, 1)))
     d = c.cast(dtypes.int).expand((2, 6, 6)) * 4
-    run_schedule(check_schedule(d, 2))
+    run_schedule(check_schedule(d, 1))
     c_np = np.pad((np.full((4, 4), 2., dtype=np.float32) + np.full((4, 4), 1., dtype=np.float32)), ((1, 1), (1, 1)), constant_values=0.0)
     np.testing.assert_equal(d.numpy(), np.broadcast_to(c_np.astype(np.half), (2, *c_np.shape)) * 4)
 
@@ -1678,7 +1615,7 @@ class TestSchedule(unittest.TestCase):
     self._test_fusion([(4, 4), (1, 4)], lambda a,b:a.sum(1).reshape(b.shape)+b, 1)
 
   def test_late_fusion_post_permute(self):
-    self._test_fusion([(4, 6, 4), (4, 4, 1)], lambda a,b:a.sum(1, keepdim=True).permute((2, 0, 1))+b, 2)
+    self._test_fusion([(4, 6, 4), (4, 4, 1)], lambda a,b:a.sum(1, keepdim=True).permute((2, 0, 1))+b, 1)
 
   def test_late_fusion_double_transpose(self):
     self._test_fusion([(32, 16, 1)],
@@ -1716,6 +1653,7 @@ class TestSchedule(unittest.TestCase):
     self.assertListEqual(realized_const_view.tolist(), [[1, 1, 1, 1], [1, 1, 1, 1], [1, 1, 1, 1], [1, 1, 1, 1]])
 
   @given(strat.sampled_from(dtypes.all), strat.sampled_from(dtypes.all))
+  @unittest.skip("kernel count depends on input")
   def test_cast_padded_const(self, dt1, dt2):
     assume(is_dtype_supported(dt1) and is_dtype_supported(dt2))
     a = Tensor(1, dtype=dt1).reshape(1, 1).pad(((1, 1), None))
@@ -1729,7 +1667,7 @@ class TestSchedule(unittest.TestCase):
     X = Tensor.randn(10, 10).realize()
     idxs = Tensor([0, 2]).realize()
     xt = X[idxs]
-    run_schedule(check_schedule(xt, 2))
+    run_schedule(check_schedule(xt, 1))
     np.testing.assert_equal(xt.numpy(), X.numpy()[idxs.numpy()])
 
   def test_simple_indexing_alt(self):
@@ -1747,7 +1685,7 @@ class TestSchedule(unittest.TestCase):
   def test_advanced_indexing_alt(self):
     X = Tensor.arange(6).reshape(3, 2)+1
     xt = X[[Tensor([2]), Tensor([1])]]
-    run_schedule(check_schedule(xt, 3))
+    run_schedule(check_schedule(xt, 1))
     np.testing.assert_equal(xt.numpy(), 6)
 
   def test_advanced_simple_indexing_combined(self):
@@ -1795,7 +1733,7 @@ class TestSchedule(unittest.TestCase):
     x = Tensor.full((2,2), 16)
     y = x.idiv(Tensor.linspace(2, 8, steps=4, dtype=dtypes.int).reshape(2,2)).pad(((1,1), (1,1)))
     out = y.sum(axis=1)
-    run_schedule(check_schedule(out, 2))
+    run_schedule(check_schedule(out, 1))
     self.assertListEqual(out.tolist(), [0, 12, 4, 0])
 
   def test_arange_transposed_descendants(self):
@@ -1828,7 +1766,7 @@ class TestSchedule(unittest.TestCase):
     x = Tensor.randn(5, 2).realize()
     a = Tensor.arange(10).contiguous()
     out = (x + a[2]).sum()
-    run_schedule(check_schedule(out, 3))
+    run_schedule(check_schedule(out, 2))
     np.testing.assert_allclose(out.numpy(), (x.numpy()+np.arange(10)[2]).sum(), atol=1e-5, rtol=1e-6)
 
   def test_arange_index_child(self):
@@ -1844,7 +1782,7 @@ class TestSchedule(unittest.TestCase):
     x = Tensor.randn(5, 2).realize()
     a = (Tensor.arange(10)+1).contiguous()
     out = (x + a[2]).sum()
-    run_schedule(check_schedule(out, 3))
+    run_schedule(check_schedule(out, 2))
     np.testing.assert_allclose(out.numpy(), (x.numpy()+(np.arange(10)+1)[2]).sum(), atol=1e-5, rtol=1e-6)
 
   @unittest.skip("BUFFER_VIEW no longer supported on non-disk devices")
@@ -1859,10 +1797,10 @@ class TestSchedule(unittest.TestCase):
     from extra.models.llama import precompute_freqs_cis
     args = {"dim":32 if CI else 128, "end":2048 if CI else 8192, "theta":10000}
     fused = precompute_freqs_cis(**args)
-    run_schedule(check_schedule(fused, 3))
+    run_schedule(check_schedule(fused, 1))
     if getenv("CHECK", 1):
       ref = precompute_freqs_cis(**args)
-      run_schedule(check_schedule(ref, 3))
+      run_schedule(check_schedule(ref, 1))
       np.testing.assert_equal(fused.numpy(), ref.numpy())
 
   def test_fuse_assign_contiguous(self):
@@ -1904,7 +1842,7 @@ class TestSchedule(unittest.TestCase):
     X = Tensor([[0, 2, 3], [1, 2, 3]]).realize()
     Y = Tensor([1, 2]).realize()
     loss = X.sparse_categorical_crossentropy(Y)
-    run_schedule(check_schedule(loss, 4))
+    run_schedule(check_schedule(loss, 3))
     np.testing.assert_allclose(loss.item(), 0.878309, atol=1e-5, rtol=1e-6)
 
   def test_const_folding_alt(self):
@@ -1925,7 +1863,7 @@ 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, 6))
+      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)
@@ -1935,7 +1873,7 @@ class TestSchedule(unittest.TestCase):
     r = (X+Tensor.arange(16).reshape(4, 4)).sum()
     out0 = r+2
     out1 = r+3
-    run_schedule(check_schedule([out0, out1], 1))
+    run_schedule(check_schedule([out0, out1], 2)) # TODO: 1?
     r_ref = (X.numpy()+np.arange(16).reshape(4, 4)).sum()
     np.testing.assert_allclose(out0.numpy(), r_ref+2, rtol=2e-7)
     np.testing.assert_allclose(out1.numpy(), r_ref+3, rtol=2e-7)
@@ -1977,8 +1915,7 @@ class TestSwizzle(unittest.TestCase):
       a = Tensor.randint(32, 32).realize()
     r = (a+a).sum(1).sum(0)
     # double reduce collapses to a single reduce
-    with Context(DONT_GROUP_REDUCES=1):
-      run_schedule(check_schedule(r, 1))
+    run_schedule(check_schedule(r, 1))
     self.assertEqual(r.numpy(), (a.numpy()+a.numpy()).sum(1).sum(0))
 
   def test_single_swizzle(self):
@@ -1998,33 +1935,29 @@ class TestSwizzle(unittest.TestCase):
       b = Tensor.randint(4,).realize()
     # parallel reduce!
     add = a.sum(0)+b.sum(0)
-    with Context(DONT_GROUP_REDUCES=1):
-      run_schedule(check_schedule(add, 1))
+    run_schedule(check_schedule(add, 1))
     self.assertEqual(add.numpy(), a.numpy().sum(0)+b.numpy().sum(0))
 
-  @unittest.skip("TODO: how do we express the norm")
   def test_softmax_one_kernel(self):
     Tensor.manual_seed(0)
     with Context(DEBUG=0, TRACK_MATCH_STATS=0):
       a = Tensor.randn(32, 32).realize()
     t = a.softmax()
-    with Context(DONT_GROUP_REDUCES=1, DONT_REALIZE_EXPAND=1):
-      check_schedule(t, 1)
+    check_schedule(t, 3) # TODO: 1?
 
   def test_argmax_one_kernel(self):
     Tensor.manual_seed(0)
     with Context(DEBUG=0, TRACK_MATCH_STATS=0):
       a = Tensor.randn(10, 20).realize()
     t = a.argmax(0)
-    with Context(DONT_GROUP_REDUCES=1, DONT_REALIZE_EXPAND=1): t.realize()
+    check_schedule(t, 2) # TODO: 1?
 
   def test_swizzle_reduceop(self):
     Tensor.manual_seed(0)
     x = Tensor.randn(4,4).realize()
     y = Tensor.randn(4,4,4).realize()
     out = x.reshape(4,4,1).expand(4,4,4).sum(axis=(1,))+y
-    with Context(DONT_REALIZE_EXPAND=1, DONT_GROUP_REDUCES=1):
-      run_schedule(check_schedule(out, 1))
+    run_schedule(check_schedule(out, 2)) # TODO: 1?
     np.testing.assert_allclose(out.numpy(), np.tile(x.numpy().reshape(4,4,1), (1,1,4)).sum(axis=1)+y.numpy())
 
   def test_permute_rewrite(self):
@@ -2032,7 +1965,7 @@ class TestSwizzle(unittest.TestCase):
     y = Tensor.randn(4, 1, 16).realize()
     z = Tensor.randn(4, 4, 1).realize()
     t = (x*y).sum(axis=(0, 2)).reshape(1, 4, 1).permute(0, 2, 1)+z
-    with Context(DONT_GROUP_REDUCES=1, DONT_REALIZE_EXPAND=1): run_schedule(check_schedule(t, 1))
+    run_schedule(check_schedule(t, 2)) # TODO: 1?
     t_np = (x.numpy()*y.numpy()).sum(axis=(0, 2)).reshape(1, 4, 1).transpose(0, 2, 1)+z.numpy()
     np.testing.assert_allclose(t.numpy(), t_np, atol=1e-6, rtol=1e-3)
 
@@ -2043,14 +1976,14 @@ class TestSwizzle(unittest.TestCase):
     a_reduce = a.sum(axis=(2,), keepdim=True).sum(axis=(1,))
     b_reduce = b.sum(axis=(0,))
     t = a_reduce+b_reduce
-    with Context(DONT_GROUP_REDUCES=1, DONT_REALIZE_EXPAND=1): run_schedule(check_schedule(t, 1))
+    run_schedule(check_schedule(t, 1))
 
   def test_parallel_reduce_possible(self):
     Tensor.manual_seed(0)
     x = Tensor.randn(4, 2, 2).realize()
     y = Tensor.randn(4, 2, 2).realize()
     t = x.sum(axis=1)+y.sum(axis=1)
-    with Context(DONT_GROUP_REDUCES=1): run_schedule(check_schedule(t, 1))
+    run_schedule(check_schedule(t, 1))
     np.testing.assert_allclose(t.numpy(), x.numpy().sum(axis=1)+y.numpy().sum(axis=1), atol=1e-6, rtol=1e-3)
 
   # kernels can only have 1 or n in each dim
@@ -2059,7 +1992,7 @@ class TestSwizzle(unittest.TestCase):
     x = Tensor.randn(4, 2, 2).realize()
     y = Tensor.randn(4, 3, 2).realize()
     t = x.sum(axis=1)+y.sum(axis=1)
-    with Context(DONT_GROUP_REDUCES=1): run_schedule(check_schedule(t, 1))
+    run_schedule(check_schedule(t, 1))
     np.testing.assert_allclose(t.numpy(), x.numpy().sum(axis=1)+y.numpy().sum(axis=1), atol=1e-6, rtol=1e-3)
 
   def test_unsafe_pad(self):
@@ -2143,6 +2076,11 @@ 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)
@@ -2160,8 +2098,8 @@ class TestCopyFolding(unittest.TestCase):
     a = Tensor.ones((4,)).to("CPU")
     b = Tensor.empty(4, device="CPU")
     add = a+b
-    add.kernelize()
     assert all_same([x.device for x in add.uop.src]), f"ALU has different devices! {[x.device for x in add.src]}"
+    add.kernelize()
 
   def test_alu_before_copy(self):
     buf = Tensor.ones(1).contiguous().realize()
@@ -2172,12 +2110,12 @@ class TestCopyFolding(unittest.TestCase):
   def test_copy_to_same_device(self):
     a = Tensor.empty(4).uop
     b = a.copy_to_device(a.device)
-    check_schedule(b, 0, filter_sink=False)
+    check_schedule(b, 1, filter_sink=False) # TODO: 0?
 
   def test_copy_to_same_device_alt(self):
     a = Tensor.empty(4, 4).uop
     b = a.copy_to_device(a.device)
-    check_schedule(b, 0, filter_sink=False)
+    check_schedule(b, 1, filter_sink=False) # TODO: 0?
 
   def test_copy_to_same_device_sched(self):
     a = Tensor.ones(4).contiguous().realize().uop.as_buf()
@@ -2192,13 +2130,11 @@ class TestCopyFolding(unittest.TestCase):
     a = Tensor.empty(4)
     check_schedule(a.clone(), 1, filter_sink=False)
 
-  # NOTE: moving copy before view might change this
   def test_shrink_copy(self):
     a = Tensor.arange(4)
     view = a.shrink(((0, 2),))
     b = view.clone()
-    # NOTE: this was sort of a bug making this 2
-    run_schedule(check_schedule(b, 2, filter_sink=False))
+    run_schedule(check_schedule(b, 1, filter_sink=False))
     self.assertEqual(b.uop.base.buffer.size, 2)
     self.assertEqual(b.uop.size, 2)
     self.assertListEqual(b.tolist(), [0, 1])
@@ -2207,7 +2143,7 @@ class TestCopyFolding(unittest.TestCase):
     a = Tensor.arange(2)
     view = a.reshape(2, 1).expand(2, 2)
     b = view.clone()
-    run_schedule(check_schedule(b, 2, filter_sink=False))
+    run_schedule(check_schedule(b, 1, filter_sink=False))
     self.assertEqual(b.uop.base.buffer.size, 4)
     self.assertEqual(b.uop.size, 4)
     self.assertListEqual(b.tolist(), [[0, 0], [1, 1]])
@@ -2251,7 +2187,7 @@ class TestBufferUOp(unittest.TestCase):
   def test_buffer_has_buffer(self):
     buf = Tensor.empty(10)
     self.assertIsNotNone(buf.uop.buffer)
-    self.assertEqual(buf.uop.st, ShapeTracker.from_shape((10,)))
+    self.assertEqual(buf.uop.shape, (10,))
     # the device Buffer remains unallocated until it's we run the schedule
     self.assertFalse(buf.uop.buffer.is_allocated())
     add = buf+1
@@ -2330,7 +2266,7 @@ class TestContiguous(unittest.TestCase):
   def test_double_contiguous_realizes_once(self):
     a = Tensor.empty(4, 1)
     b = a.expand((4, 4)).contiguous().contiguous()
-    check_schedule(b, 1)
+    check_schedule(b, 2) # TODO: should be 1?
 
   def test_view_does_not_realize(self):
     a = Tensor.empty(4)
@@ -2466,10 +2402,6 @@ class TestUOpBecome(unittest.TestCase):
     c = (a.reshape(1, 1, 4, 4)+0).shrink(((0, 1), (0, 1), (0, 3), (0, 3)))+0
     check_schedule([b, c], 0)
     assert all_same([x.uop.base.realized for x in [a,b,c]])
-    # these movement ops result in the same ShapeTracker
-    assert b.uop.st == c.uop.st
-    assert b.uop is c.uop
-    assert UPat(Ops.VIEW, src=(UPat(Ops.BUFFER),)).match(c.uop, {})
 
   def test_setitem_becomes_subbuffer(self):
     a = Tensor.full((4,), 2.).contiguous().realize()

test/test_setitem.py

@@ -52,7 +52,6 @@ class TestSetitem(unittest.TestCase):
 
   def test_setitem_into_noncontiguous(self):
     t = Tensor.ones(4)
-    self.assertFalse(t.uop.st.contiguous)
     with self.assertRaises(RuntimeError): t[1] = 5
 
   @unittest.skip("TODO: flaky")

test/test_softmax_fusion.py

@@ -165,8 +165,7 @@ class TestSoftmaxFusion(unittest.TestCase):
       sout.realize()
 
     print("*** single kernel softmax ***")
-    # NOTE: DONT_GROUP_REDUCES is required here
-    with Context(NOOPT=1, DEBUG=max(DEBUG.value, 2), DONT_GROUP_REDUCES=1):
+    with Context(NOOPT=1, DEBUG=max(DEBUG.value, 2)):
       out = single_kernel_softmax(self.test)
       out.realize()
 
@@ -186,7 +185,6 @@ class TestSoftmaxFusion(unittest.TestCase):
 
     np.testing.assert_allclose(sout.numpy(), out.numpy(), atol=3e-7)
 
-  @unittest.skip("recursion error no longer raised")
   def test_softmax_bw(self):
     print("*** softmax bw ***")
     self.test.requires_grad_()
@@ -197,14 +195,11 @@ class TestSoftmaxFusion(unittest.TestCase):
     self.test.grad = None
 
     print("*** single kernel softmax bw ***")
-    # NOTE: DONT_GROUP_REDUCES is required here
-    # TODO: fix RecursionError with DONT_GROUP_REDUCES
-    with self.assertRaises(RecursionError):
-      with Context(NOOPT=1, DEBUG=max(DEBUG.value, 2), DONT_GROUP_REDUCES=1):
-        single_kernel_softmax(self.test).sum().backward()
-        g = self.test.grad.realize()
+    with Context(NOOPT=1, DEBUG=max(DEBUG.value, 2)):
+      single_kernel_softmax(self.test).sum().backward()
+      g = self.test.grad.realize()
 
-      np.testing.assert_allclose(sg.numpy(), g.numpy(), atol=1e-7)
+    np.testing.assert_allclose(sg.numpy(), g.numpy(), atol=1e-7)
 
 if __name__ == '__main__':
   unittest.main()

test/test_symbolic_ops.py

@@ -1,6 +1,5 @@
 import unittest
 from tinygrad import Tensor, Variable, GlobalCounters
-from tinygrad.shape.shapetracker import View
 from tinygrad.uop.ops import sym_infer
 from tinygrad.dtype import dtypes
 from tinygrad.device import is_dtype_supported
@@ -64,14 +63,6 @@ class TestSymbolicOps(unittest.TestCase):
     self.test_attention(imin=4, imax=5, use_symbolic=False)
     self.test_attention(imin=4, imax=5, use_symbolic=True)
 
-  # until this works, symbolic single kernel softmax won't
-  @unittest.expectedFailure
-  def test_attention_simple_view(self):
-    i = Variable("i", 2, 10)
-    v1 = View.create((2,4,1,i,i), ((i*4),i,0,0,1))
-    v2 = View.create((2,4,1,i,i,i), (((i*i)*4),(i*i),0,0,i,1))
-    self.assertIsNotNone(v1+v2)
-
   def test_attention_training(self):
     with Tensor.train():
       self.test_attention(dropout_p=0.0)
@@ -296,7 +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
@@ -304,9 +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

@@ -9,6 +9,7 @@ from hypothesis import given, settings, strategies as strat
 from tinygrad.device import is_dtype_supported
 from tinygrad.uop.ops import Ops, UOp
 from tinygrad.renderer.ptx import PTXRenderer
+from tinygrad.renderer.nir import NIRRenderer
 from tinygrad.codegen import full_rewrite
 from tinygrad.dtype import DType
 
@@ -526,6 +527,7 @@ class TestTinygrad(unittest.TestCase):
     self.assertListEqual(t.shrink_to(16).tolist(), list(range(16)))
     t = t.reshape(4, 8).contiguous().realize()
     self.assertListEqual(t.shrink_to(2, 2).tolist(), [[0, 1], [8, 9]])
+    self.assertListEqual(t.shrink_to(None, 2).tolist(), t.shrink_to(4, 2).tolist())
     with self.assertRaises(ValueError): t.shrink_to(2)
     with self.assertRaises(ValueError): t.shrink_to(2, 2, 2)
 
@@ -568,22 +570,6 @@ class TestMoveTensor(unittest.TestCase):
     np.testing.assert_equal(x.grad.numpy(), [[2,2,2],[0,0,0],[-2,-2,-2]])
 
 class TestZeroShapeTensor(unittest.TestCase):
-  def test_shape_is_expanded(self):
-    t = Tensor.empty(3, 2, 0)
-    assert t.shape == (3, 2, 0)
-    # numpy has stride 0, 0, 0; torch has stride 2, 1, 1
-    assert t.uop.st.is_expanded() == (True, True, True)
-
-    t = Tensor.empty(3, 0, 2)
-    assert t.shape == (3, 0, 2)
-    # numpy has stride 0, 0, 0; torch has stride 2, 2, 1
-    assert t.uop.st.is_expanded() == (True, True, True)
-
-    t = Tensor.empty(0, 0, 0)
-    assert t.shape == (0, 0, 0)
-    # numpy has stride 0, 0, 0; torch has stride 1, 1, 1
-    assert t.uop.st.is_expanded() == (True, True, True)
-
   def test_rand(self):
     t = Tensor.rand(3, 2, 0)
     assert t.shape == (3, 2, 0)
@@ -635,8 +621,10 @@ class TestZeroShapeTensor(unittest.TestCase):
 
     np.testing.assert_equal(Tensor([1, 2]).pad_to(4).numpy(), [1, 2, 0, 0])
     np.testing.assert_equal(Tensor([[1, 2]]).pad_to(2, 3).numpy(), [[1, 2, 0], [0, 0, 0]])
-    with self.assertRaises(TypeError): Tensor([1, 2]).pad_to(2, 3)
-    with self.assertRaises(TypeError): Tensor([[1, 2]]).pad_to(3)
+    np.testing.assert_equal(Tensor([[1, 2]]).pad_to(1, 3).numpy(), [[1, 2, 0]])
+    np.testing.assert_equal(Tensor([[1, 2]]).pad_to(None, 3).numpy(), [[1, 2, 0]])
+    with self.assertRaises(ValueError): Tensor([1, 2]).pad_to(2, 3)
+    with self.assertRaises(ValueError): Tensor([[1, 2]]).pad_to(3)
 
   def test_shrink_into_zero(self):
     t = Tensor.rand(3, 4).realize()
@@ -822,6 +810,13 @@ class TestTensorMetadata(unittest.TestCase):
     self.assertEqual(len(si.metadata), 1)
     self.assertEqual(si.metadata[0].name, "relu")
 
+  def test_assign(self):
+    x = Tensor.empty(10, 10).realize()
+    x.assign(Tensor.ones(10, 10).contiguous())
+    si = x.schedule()[-1]
+    self.assertEqual(len(si.metadata), 1)
+    self.assertEqual(si.metadata[0].name, "assign")
+
   def test_complex(self):
     x = Tensor.rand(3, requires_grad=True)
     y = Tensor.rand(3, requires_grad=True)
@@ -833,7 +828,6 @@ class TestTensorMetadata(unittest.TestCase):
     self.assertEqual(len(si.metadata), 3)
     self.assertEqual(set(m.name for m in si.metadata), {"relu", "sigmoid", "__mul__"})
 
-  @unittest.skip("not accurate")
   def test_complex_backward(self):
     x = Tensor.rand(3, requires_grad=True).realize()
     y = Tensor.rand(3, requires_grad=True).realize()
@@ -871,7 +865,8 @@ class TestIdxUpcast(unittest.TestCase):
     store = next(uop for uop in uops if uop.op is Ops.STORE)
     assert store.op is Ops.STORE
     idx = self._find_op(store, Ops.INDEX)
-    if idx is not None: # PTX turns Ops.INDEX into pointer arithmetic earlier than cstyle, plus it's already cast to int64
+    # PTX and NIR turn Ops.INDEX into pointer arithmetic earlier than cstyle, plus it's already cast to int64
+    if not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)):
       assert idx.op is Ops.INDEX
       idx_val = idx.src[1]
       assert idx_val.dtype is dtype
@@ -895,7 +890,7 @@ class TestIdxUpcast(unittest.TestCase):
   def test_regular_sym(self):
     self.do_op_then_assert(dtypes.int, 2048, 2048, UOp.variable("dim3", 1, 64).bind(32))
 
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX always convert Ops.INDEX to int64")
+  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)), "PTX and NIR always converts Ops.INDEX to int64")
   def test_symfold(self):
     # This would cause an overflow, but after sym fold it's within int32
     a = Tensor.arange(65535)

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):
@@ -11,8 +11,7 @@ class TestTensorUOp(unittest.TestCase):
     def helper(a: np.ndarray):
       print(a.shape, a.strides, a.flags.c_contiguous)
       b = Tensor(a).uop
-      #assert b.st.contiguous == a.flags.c_contiguous
-      assert b.st.shape == a.shape
+      assert b.shape == a.shape
       np.testing.assert_equal(a, Tensor(b).numpy())
 
     for ndims in range(1, 4):
@@ -94,7 +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):
@@ -102,23 +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_tensor_variable.py

@@ -93,7 +93,7 @@ class TestTensorVariable(unittest.TestCase):
     vb = v.bind(3)
     t = Tensor.empty(3, vb)
     assert t.uop.base.buffer.size == 30
-    assert t.uop.st.shape == (3, vb)
+    assert t.uop.shape == (3, vb)
 
 
 if __name__ == '__main__':

test/test_transcendental.py

@@ -149,6 +149,7 @@ class TestTranscendentalVectorized(unittest.TestCase):
     for vec_size in [1,2,3,4,5,127,128]: self._test_vectorized_op(Tensor.log2, np.log2, (0.001, 200), vec_size)
 
   @unittest.skipIf(getenv("DSP"), "requires int division")
+  @unittest.skipIf(getenv("NV_NAK"), "MUFU.SIN is not accurate enough")
   def test_sin_vectorized(self):
     for vec_size in [1,2,3,4,5,127,128]: self._test_vectorized_op(Tensor.sin, np.sin, (-100, 100), vec_size)
 

test/test_uop_graph.py

@@ -1,12 +1,12 @@
-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 import 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 +15,6 @@ 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))
@@ -459,30 +453,30 @@ 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)
+    ast = c10.sink()
+    uops = to_uops_list([ast])
     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 +512,7 @@ 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 +574,7 @@ 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 +598,7 @@ 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 +633,13 @@ 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 +660,6 @@ 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)
 
@@ -845,8 +827,6 @@ class TestIFUOps(unittest.TestCase):
     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)
@@ -854,7 +834,7 @@ class TestIFUOps(unittest.TestCase):
     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)))
+    st = UOp(Ops.STORE, dtypes.void, (sbuf.index(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)]
@@ -863,8 +843,6 @@ class TestIFUOps(unittest.TestCase):
     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
@@ -879,8 +857,6 @@ class TestIFUOps(unittest.TestCase):
     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):

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,
@@ -302,6 +308,7 @@ 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 +338,7 @@ 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 +348,7 @@ 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 +358,7 @@ 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 +385,7 @@ 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 +397,7 @@ 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 +408,14 @@ 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 +428,7 @@ 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 +436,7 @@ 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 +460,7 @@ 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 +519,7 @@ 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")
@@ -540,10 +547,10 @@ 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)))
+    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(self):
-    u = UOp(Ops.VECTORIZE, src=(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1), UOp.const(dtypes.int, 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(), "{0,1,2}")
 
 if __name__ == '__main__':

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

@@ -418,5 +418,30 @@ class TestPathTensor(unittest.TestCase):
       Tensor(pathlib.Path(test_file)).tolist()
     os.chmod(test_file, 0o644)
     assert Tensor(pathlib.Path(test_file)).tolist(), list(range(10))
+
+class TestDiskTensorMovement(unittest.TestCase):
+  def setUp(self):
+    self.fn = pathlib.Path(temp("custom_disk_range"))
+    self.fn.unlink(missing_ok=True)
+    Tensor.arange(100, dtype=dtypes.uint8).to(f"disk:{str(self.fn)}").realize()
+
+  def test_simple_read(self):
+    t = Tensor(self.fn)
+    self.assertTrue(Tensor.all(t.to(None) == Tensor.arange(100, dtype=dtypes.uint8)).item())
+
+  def test_slice_read(self):
+    t = Tensor(self.fn)
+    self.assertListEqual(t[16:18].tolist(), [16,17])
+
+  def test_slice_read_cat(self):
+    t = Tensor(self.fn)
+    with self.assertRaises(AssertionError):
+      self.assertListEqual(Tensor.cat(t[16:18], t[20:22]).tolist(), [16,17,20,21])
+
+  def test_slice_sum(self):
+    t = Tensor(self.fn)
+    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_elf.py

@@ -24,6 +24,15 @@ class TestElfLoader(unittest.TestCase):
     '''
     with self.assertRaisesRegex(RuntimeError, 'evil_external_function'):
       ClangJITCompiler().compile(src)
+  def test_link(self):
+    src = '''
+      float powf(float, float); // from libm
+      float test(float x, float y) { return powf(x, y); }
+    '''
+    args = ('-x', 'c', '-c', '-target', f'{platform.machine()}-none-unknown-elf', '-march=native', '-fPIC', '-O2', '-ffreestanding', '-nostdlib')
+    obj = subprocess.check_output(('clang',) + args + ('-', '-o', '-'), input=src.encode())
+    with self.assertRaisesRegex(RuntimeError, 'powf'): elf_loader(obj)
+    elf_loader(obj, link_libs=['m'])
 
 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)
@@ -334,5 +335,19 @@ class TestBidirectional(unittest.TestCase):
     graph_rewrite(c, pm, ctx=ctx_list, bpm=bpm)
     self.assertListEqual(ctx_list, [('+', True), (1, True), (1, False), (2, True), (2, False), ('+', False)])
 
+class TestStopEarly(unittest.TestCase):
+  def test_stop_early(self):
+    a = UOp.const(dtypes.int, 3)
+    b = UOp.const(dtypes.int, 4)
+    c = a+b
+    cn = UOp.const(dtypes.int, 7)
+    d = UOp.const(dtypes.int, 2)
+    def visit_const(c:UOp):
+      print(f"visit {c.arg}")
+      assert c.arg not in (3,4)
+    pm_cvisit = PatternMatcher([(UPat(Ops.CONST, name="c"), visit_const),])
+    ret = (c+d).substitute({c:cn}, extra_pm=pm_cvisit)
+    assert ret == cn+d
+
 if __name__ == '__main__':
   unittest.main()

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_indexing.py

@@ -5,8 +5,6 @@ import numpy as np
 
 from tinygrad import Tensor, dtypes, Device, TinyJit
 from tinygrad.device import is_dtype_supported
-from tinygrad.shape.shapetracker import ShapeTracker
-from tinygrad.shape.view import View
 from tinygrad.helpers import CI, all_same, prod
 
 random.seed(42)
@@ -22,11 +20,13 @@ def consec(shape, start=1):
 # creates strided tensor with base set to reference tensor's base, equivalent to torch.set_()
 def set_(reference: Tensor, shape, strides, offset):
   raise NotImplementedError("need to implement without calling uop.view")
+  """
   if reference.uop.base.realized is None: reference.realize()
   assert reference.uop.base.realized, "base has to be realized before setting it to strided's base"
   strided = Tensor(reference.uop.view(ShapeTracker((View.create(shape=shape, strides=strides, offset=offset),))))
   assert strided.uop.st.real_strides() == strides, "real_strides should equal strides for strided"
   return strided
+  """
 
 def clone(original:Tensor): return original.clone()
 def copy_(src:Tensor, other:Tensor) -> Tensor: return src.clone()
@@ -501,10 +501,6 @@ class TestIndexing(unittest.TestCase):
     y = x[:, :, :, 1]
     z = y[:, 1:1, :]
     numpy_testing_assert_equal_helper((2, 0, 4), z.shape)
-    # this isn't technically necessary, but matches NumPy stride calculations.
-    # NOTE: this is empty and shouldn't have strides
-    numpy_testing_assert_equal_helper((True, True, True), z.uop.st.is_expanded())
-    self.assertTrue(z.uop.st.contiguous)
 
   @unittest.skip("bool indexing not supported")
   def test_index_getitem_copy_bools_slices(self):

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()