enable rangeify const folding

2026-08-14 21:18:26 +00:00 · 2025-09-15 11:50:13 +08:00
556 changed files with 9920 additions and 72093 deletions
@@ -41,10 +41,6 @@ inputs:
    description: "Install LLVM?"
    required: false
    default: 'false'
-  mesa:
-    description: "Install mesa"
-    required: false
-    default: 'false'
 runs:
  using: "composite"
  steps:
@@ -293,13 +289,3 @@ 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
@@ -2,7 +2,7 @@ name: Autogen
 env:
  # increment this when downloads substantially change to avoid the internet
  DOWNLOAD_CACHE_VERSION: '12'
-  PYTHON_CACHE_VERSION: '4'
+  PYTHON_CACHE_VERSION: '3'
  APT_CACHE_VERSION: '1'
  BUILD_CACHE_VERSION: '1'
  CAPTURE_PROCESS_REPLAY: 1
@@ -36,9 +36,8 @@ 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 llvm-20-dev
+      run: sudo apt-get install -y --no-install-recommends llvm-14-dev libclang-14-dev
    - name: Verify OpenCL autogen
      run: |
        cp tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
@@ -90,8 +89,3 @@ 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
@@ -28,7 +28,7 @@ jobs:
      # since sudo is required for usbgpu on macos, move the cache to a new location, as some of the files are owned by root
      PYTHONPYCACHEPREFIX: /tmp/tiny_python_pycache
    runs-on: [self-hosted, macOS]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -51,18 +51,15 @@ 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=720 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=500 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=800 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=700 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=4500 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=1600 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
    # process replay can't capture this, the graph is too large
    - 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
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3000 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
@@ -102,7 +99,7 @@ jobs:
    - name: Run GPT2
      run: |
        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=8 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
@@ -111,19 +108,14 @@ jobs:
      run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
    - name: Train MNIST
      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
-
-    # NOTE: this is failing in CI. it is not failing on my machine and I don't really have a way to debug it
-    # the error is "RuntimeError: Internal Error (0000000e:Internal Error)"
-    #- name: Run 10 CIFAR training steps
-    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
-    #- name: Run 10 CIFAR training steps w HALF
-    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
-
+    - name: Run 10 CIFAR training steps
+      run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=320 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+    - name: Run 10 CIFAR training steps w HALF
+      run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=385 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
    #- name: Run 10 CIFAR training steps w BF16
    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 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_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+    - name: Run 10 CIFAR training steps w winograd
+      run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: UsbGPU boot time
      run: sudo -E PYTHONPATH=. DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
    - name: UsbGPU tiny tests
@@ -131,7 +123,7 @@ 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 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 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
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (Mac)
@@ -168,7 +160,7 @@ jobs:
  testnvidiabenchmark:
    name: tinybox green Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 60
+    timeout-minutes: 30
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -211,7 +203,6 @@ jobs:
        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul.txt
        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_tf32.txt
-        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_fp8.txt
    - name: Run Tensor Core GEMM (PTX)
      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
    - name: Run Tensor Core GEMM (NV)
@@ -222,9 +213,8 @@ jobs:
      run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
    - name: Run Stable Diffusion
      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
-    # TODO: too slow
-    # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+    - name: Run SDXL
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA
      run: |
        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
@@ -239,8 +229,6 @@ jobs:
      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
-    - name: Run quantized LLaMA3
-      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8 | tee llama3_fp8.txt
    # - name: Run LLaMA-3 8B on 6 GPUs
    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
    # - name: Run LLaMA-2 70B
@@ -250,9 +238,9 @@ jobs:
    - name: Run GPT2
      run: |
        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
@@ -274,7 +262,6 @@ jobs:
          llama3_beam.txt
          llama3_four_gpu.txt
          llama3_six_gpu.txt
-          llama3_fp8.txt
          llama_2_70B.txt
          mixtral.txt
          gpt2_unjitted.txt
@@ -287,7 +274,7 @@ jobs:
  testmorenvidiabenchmark:
    name: tinybox green Training Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -312,27 +299,24 @@ 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
-    # TODO: too slow
-    # - name: Fuzz Padded Tensor Core GEMM (NV)
-    #   run: NV=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
-    # TODO: too slow
-    # - name: Fuzz Padded Tensor Core GEMM (PTX)
-    #   run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+    - name: Fuzz Padded Tensor Core GEMM (NV)
+      run: NV=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+    - name: Fuzz Padded Tensor Core GEMM (PTX)
+      run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
    - name: Train MNIST
      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - 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
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=85 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=240 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=68 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=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
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=75 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    - name: Run 10 CIFAR training steps w winograd
+      run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=35 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    - name: Run MLPerf resnet eval on training data
      run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
@@ -362,7 +346,7 @@ jobs:
  testamdbenchmark:
    name: tinybox red Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -431,10 +415,9 @@ jobs:
    - name: Test AM warm start time
      run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
-    # TODO: too slow
-    # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=450 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+    - name: Run SDXL
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=1400 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA 7B
      run: |
        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
@@ -493,7 +476,7 @@ jobs:
  testmoreamdbenchmark:
    name: tinybox red Training Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
+    timeout-minutes: 30
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -525,20 +508,19 @@ jobs:
    - name: Train MNIST
      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=330 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=85 AMD=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=350 AMD=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=288 AMD=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=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=188 AMD=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=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    - name: Run 10 CIFAR training steps w winograd
+      run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    #- name: Run full CIFAR training steps w 6 GPUS
-    #  run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+    #  run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    #- name: Run full CIFAR training steps w 6 GPUS (REMOTE)
-    #  run: time BENCHMARK_LOG=cifar_6gpu_remote REMOTE=1 REMOTEDEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu_remote.txt
+    #  run: time BENCHMARK_LOG=cifar_6gpu_remote REMOTE=1 REMOTEDEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu_remote.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AMD Training)
@@ -557,7 +539,7 @@ jobs:
  testmlperfamdbenchmark:
    name: tinybox red MLPerf Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
+    timeout-minutes: 30
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -623,24 +605,22 @@ 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 ASSERT_MIN_STEP_TIME=30 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 ASSERT_MIN_STEP_TIME=45 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 ASSERT_MIN_STEP_TIME=70 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: 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
+      run: PYTHONPATH="." 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: 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
+      run: PYTHONPATH="." 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: 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
+      run: PYTHONPATH="." 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="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/22aec22a10ce09384d4a4af2a0bbff08d54af7e0c888503508f356fae4ff0e29
+        PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/c824f68646a3b94f117f01c70dc8316fb466e05fbd42ccdba440b8a8dc86914b
    - name: benchmark MobileNetV2 on DSP
      run: |
        # generate quantized weights
@@ -648,14 +628,24 @@ 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 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
+        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
    - 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
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 20
+    timeout-minutes: 15
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -705,7 +695,7 @@ jobs:
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
    # TODO: enable
    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
@@ -726,7 +716,7 @@ jobs:
  testgreendriverbenchmark:
    name: NV Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 20
+    timeout-minutes: 15
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -768,7 +758,7 @@ jobs:
    - name: Test LLAMA-3
      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
    #  run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
    - name: Run 10 MLPerf Bert training steps (1 gpu)
@@ -12,7 +12,7 @@ jobs:
  run_script_job:
    runs-on: [self-hosted, Linux, tinybox]
    if: github.repository_owner == 'tinygrad'
-    timeout-minutes: 720
+    timeout-minutes: 360

    steps:
    - name: Checkout Code
@@ -2,7 +2,7 @@ name: Unit Tests
 env:
  # increment this when downloads substantially change to avoid the internet
  DOWNLOAD_CACHE_VERSION: '12'
-  PYTHON_CACHE_VERSION: '4'
+  PYTHON_CACHE_VERSION: '3'
  APT_CACHE_VERSION: '1'
  BUILD_CACHE_VERSION: '1'
  CAPTURE_PROCESS_REPLAY: 1
@@ -30,6 +30,8 @@ jobs:
        key: llvm-speed
        deps: testing_minimal
        llvm: 'true'
+    - name: External Benchmark Schedule
+      run: python3 test/external/external_benchmark_schedule.py
    - name: Speed Test
      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
    - name: Speed Test (BEAM=2)
@@ -46,7 +48,7 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        deps: docs
-        pydeps: "capstone torch"
+        pydeps: "capstone"
    - name: Build wheel and show size
      run: |
        pip install build
@@ -77,8 +79,6 @@ jobs:
      run: |
        python docs/abstractions2.py
        python docs/abstractions3.py
-    - name: Test README
-      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py &&  python README.py
    - name: Test Quickstart
      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' docs/quickstart.md > quickstart.py &&  python quickstart.py
    - name: Test DEBUG
@@ -89,65 +89,64 @@ jobs:
        clang -O2 recognize.c -lm -o recognize
        cat test/models/efficientnet/Chicken.jpg | ./recognize | grep cock

-  # 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
+  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: 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
+  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: SPLIT_REDUCEOP=0 FUSE_ARANGE=1 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

  bepython:
    name: Python Backend
@@ -204,7 +203,7 @@ 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_SM89 ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        DEBUG=2 EMULATE=CUDA 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
@@ -230,7 +229,7 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: linting-only
-        python-version: '3.11'
+        python-version: '3.10'
        deps: linting
    - name: Lint bad-indentation and trailing-whitespace with pylint
      run: python -m pylint --disable=all -e W0311 -e C0303 --jobs=0 --indent-string='  ' --recursive=y .
@@ -239,13 +238,14 @@ jobs:
        pip3 install --upgrade --force-reinstall ruff==0.11.0
        python3 -m ruff check .
        python3 -m ruff check examples/mlperf/ --ignore E501
+    - name: Lint tinygrad with pylint
+      run: python -m pylint tinygrad/
    - name: Run mypy
      run: |
        python -m mypy --strict-equality --lineprecision-report .
        cat lineprecision.txt
-    # broken because of UPatAny
-    #- name: Run TYPED=1
-    #  run: TYPED=1 python -c "import tinygrad"
+    - name: Run TYPED=1
+      run: TYPED=1 python -c "import tinygrad"

  unittest:
    name: Unit Tests
@@ -259,28 +259,21 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: unittest-12
-        pydeps: "pillow numpy ftfy regex"
+        pydeps: "pillow"
        deps: testing_unit
-    - name: Check Device.DEFAULT
-      run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
+    - name: Test README
+      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py &&  python README.py
    - name: Run unit tests
-      run: CPU=1 python -m pytest -n=auto test/unit/ --durations=20
+      run: python -m pytest -n=auto test/unit/ --durations=20
    - 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
-    # - name: Run SDXL on NULL backend
-    #   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
+      run: NULL=1 python3 test/test_multitensor.py TestMultiTensor.test_data_parallel_resnet_train_step
+    - name: Run SDXL on NULL backend
+      run: MAX_BUFFER_SIZE=0 NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
    # 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
    - name: Run GC tests
      run: python test/external/external_uop_gc.py
-    - name: External Benchmark Schedule
-      run: python3 test/external/external_benchmark_schedule.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
    - name: Regen dataset on test_tiny
@@ -293,25 +286,6 @@ jobs:
    - name: Repo line count < 18000 lines
      run: MAX_LINE_COUNT=18000 python sz.py

-  spec:
-    strategy:
-      fail-fast: false
-      matrix:
-        group: [1, 2]
-    name: SPEC=2 (${{ matrix.group }})
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: spec-unit
-        deps: testing_unit
-    - name: Test SPEC=2
-      run: IGNORE_OOB=0 SPEC=2 PYTHONPATH="." pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
-
  fuzzing:
    name: Fuzzing
    runs-on: ubuntu-latest
@@ -328,13 +302,17 @@ 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: |
+        python test/external/fuzz_shapetracker.py
+        python test/external/fuzz_shapetracker_math.py
    - name: Fuzz Test shape ops
      run: python test/external/fuzz_shape_ops.py

  testopenclimage:
-    name: CL IMAGE Tests
+    name: 'CL IMAGE Tests'
    runs-on: ubuntu-22.04
-    timeout-minutes: 15
+    timeout-minutes: 10
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -352,7 +330,7 @@ jobs:
        uses: ./.github/actions/process-replay

  testgpumisc:
-    name: CL Misc tests
+    name: 'CL Misc tests'
    runs-on: ubuntu-22.04
    timeout-minutes: 10
    steps:
@@ -369,7 +347,7 @@ 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 test/test_optim.py -k "not muon"
+        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py
      - name: Upload artifact
        uses: actions/upload-artifact@v4
        with:
@@ -377,7 +355,7 @@ jobs:
          path: /tmp/sops.gz

  testopenpilot:
-    name: openpilot Compile Tests
+    name: 'openpilot Compile Tests'
    runs-on: ubuntu-22.04
    timeout-minutes: 15
    steps:
@@ -392,20 +370,24 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          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
+          ALLOWED_KERNEL_COUNT=208 ALLOWED_READ_IMAGE=2175 ALLOWED_GATED_READ_IMAGE=16 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
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

 # ****** ONNX Tests ******

  testonnxcpu:
-    name: ONNX (CPU) Tests
+    name: 'ONNX (CPU) Tests'
    runs-on: ubuntu-22.04
    timeout-minutes: 20

@@ -433,7 +415,7 @@ jobs:
        uses: ./.github/actions/process-replay

  testopencl:
-    name: ONNX (CL)+Optimization Tests
+    name: 'ONNX (GPU)+Optimization Tests'
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    steps:
@@ -457,12 +439,8 @@ jobs:
        run: CL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
      - name: Test MLPerf stuff
        run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
-      - name: NULL=1 beautiful_mnist_multigpu
-        run: NULL=1 python examples/beautiful_mnist_multigpu.py
-      - name: Test Bert training
-        run: NULL=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
      - name: Test llama 3 training
-        run: NULL=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=8 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+        run: MAX_BUFFER_SIZE=0 DEV=NULL SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=8 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -525,6 +503,39 @@ jobs:

 # ****** Feature Tests ******

+  testrangeify:
+    name: Linux (rangeify)
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: rangeify-minimal-llvm
+        deps: testing_minimal
+        opencl: 'true'
+        llvm: "true"
+    - name: Test CPU=1 RANGEIFY=1
+      # TODO: add more passing tests here
+      # test_symbolic_arange_sym_step is passing now
+      # test_threefry_doesnt_use_long is because there's a contig after the long now
+      run: |
+        CPU=1 CPU_LLVM=0 RANGEIFY=1 python3 -m pytest -n auto --durations 20 \
+          -k "not test_symbolic_arange_sym_step and not test_threefry_doesnt_use_long" \
+          test/test_tiny.py test/test_rangeify.py test/test_ops.py test/test_tensor_variable.py \
+          test/test_outerworld_range.py test/test_sample.py test/test_randomness.py
+    - name: Test multitensor
+      run: RANGEIFY=1 PYTHONPATH="." python3 test/test_multitensor.py TestMultiTensor.test_matmul_shard_1_1 TestMultiTensor.test_simple_add_W
+    - name: Test GPU=1 RANGEIFY=1
+      run: GPU=1 RANGEIFY=1 pytest -n auto test/test_ops.py
+    - name: Test CPU=1 RANGEIFY=2
+      run: CPU=1 CPU_LLVM=0 RANGEIFY=2 python3 -m pytest -n auto test/test_tiny.py test/test_rangeify.py test/test_ops.py --durations 20
+    # slow (and still wrong on beautiful_mnist)
+    #- name: Test LLVM=1 RANGEIFY=1 (slow tests)
+    #  run: CPU=1 CPU_LLVM=1 RANGEIFY=1 python3 -m pytest -n auto test/models/test_mnist.py --durations 20
+
  testdevectorize:
    name: Linux (devectorize)
    runs-on: ubuntu-24.04
@@ -540,11 +551,11 @@ 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
+      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"
    - 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
+      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 FUSE_ARANGE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"

  testdsp:
    name: Linux (DSP)
@@ -645,8 +656,7 @@ jobs:
        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
      - name: Run TestOps.test_add with SQTT
        run: |
-          VIZ=1 PMC=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
-          VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
+          PROFILE=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
      - name: Run process replay tests
        uses: ./.github/actions/process-replay
@@ -689,7 +699,7 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
-        backend: [llvm, cpu, opencl, lvp]
+        backend: [llvm, cpu, opencl]

    name: Linux (${{ matrix.backend }})
    runs-on: ubuntu-22.04
@@ -703,10 +713,9 @@ jobs:
          key: ${{ matrix.backend }}-minimal
          deps: testing_minimal
          opencl: ${{ matrix.backend == 'opencl' && 'true' }}
-          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
-          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
+          llvm: ${{ matrix.backend == 'llvm' && 'true' }}
      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'CL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=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' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
@@ -908,7 +917,7 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
-        backend: [metal, llvm, cpu, lvp]
+        backend: [metal, llvm, cpu]
    name: MacOS (${{ matrix.backend }})
    runs-on: macos-15
    timeout-minutes: 20
@@ -921,13 +930,12 @@ jobs:
          key: macos-${{ matrix.backend }}-minimal
          deps: testing_minimal
          pydeps: "capstone"
-          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
-          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
+          llvm: ${{ matrix.backend == 'llvm' && 'true' }}
      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'metal' && 'METAL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=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'}}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
+          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'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
@@ -38,7 +38,6 @@ extra/huggingface_onnx/models/*
 extra/huggingface_onnx/*.yaml
 extra/weights
 venv
-venv_sd_mlperf
 examples/**/net.*[js,json]
 examples/**/*.safetensors
 node_modules
@@ -20,15 +20,21 @@ repos:
        language: system
        always_run: true
        pass_filenames: false
+      - id: tests
+        name: subset of tests
+        entry: env PYTHONPATH="." python3 -m pytest -n=4 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
+        language: system
+        always_run: true
+        pass_filenames: false
      - id: example
-        name: test all devices
+        name: multi device tests
        entry: python3 test/external/external_test_example.py
        language: system
        always_run: true
        pass_filenames: false
-      - id: tests
-        name: subset of tests
-        entry: env OMP_NUM_THREADS=1 PYTHONPATH="." python3 -m pytest -n=8 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
+      - id: pylint
+        name: pylint
+        entry: python3 -m pylint tinygrad/
        language: system
        always_run: true
-        pass_filenames: false
+        pass_filenames: false
@@ -30,6 +30,10 @@ persistent=yes
 # Specify a configuration file.
 #rcfile=

+# When enabled, pylint would attempt to guess common misconfiguration and emit
+# user-friendly hints instead of false-positive error messages
+suggestion-mode=yes
+
 # Allow loading of arbitrary C extensions. Extensions are imported into the
 # active Python interpreter and may run arbitrary code.
 unsafe-load-any-extension=no
@@ -414,29 +414,10 @@ generate_sqtt() {
  clang2py -k cdefstum \
    extra/sqtt/sqtt.h \
    -o $BASE/sqtt.py
+
  fixup $BASE/sqtt.py
  sed -i "s\import ctypes\import ctypes, os\g" $BASE/sqtt.py
  python3 -c "import tinygrad.runtime.autogen.sqtt"
-
-  ROCPROF_COMMIT_HASH=dd0485100971522cc4cd8ae136bdda431061a04d
-  ROCPROF_SRC=/tmp/rocprof-trace-decoder-$ROCPROF_COMMIT_HASH
-  if [ ! -d "$ROCPROF_SRC" ]; then
-    git clone https://github.com/ROCm/rocprof-trace-decoder $ROCPROF_SRC
-    pushd .
-    cd $ROCPROF_SRC
-    git reset --hard $ROCPROF_COMMIT_HASH
-    popd
-  fi
-
-  clang2py -k cdefstum \
-    $ROCPROF_SRC/include/rocprof_trace_decoder.h \
-    $ROCPROF_SRC/include/trace_decoder_instrument.h \
-    $ROCPROF_SRC/include/trace_decoder_types.h \
-    -o extra/sqtt/rocprof/rocprof.py
-  fixup extra/sqtt/rocprof/rocprof.py
-  sed -i '1s/^/# pylint: skip-file\n/' extra/sqtt/rocprof/rocprof.py
-  sed -i "s/import ctypes/import ctypes, ctypes.util/g" extra/sqtt/rocprof/rocprof.py
-  sed -i "s|FunctionFactoryStub()|ctypes.CDLL(ctypes.util.find_library('rocprof-trace-decoder'))|g" extra/sqtt/rocprof/rocprof.py
 }

 generate_webgpu() {
@@ -461,80 +442,6 @@ 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/" $BASE/mesa.py
-  sed -i '/struct_nir_builder._pack_ = 1 # source:False/d' "$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
@@ -558,7 +465,6 @@ 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" == "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
+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
 else echo "usage: $0 <type>"
 fi
@@ -42,6 +42,7 @@ import struct
 from tinygrad.dtype import dtypes
 from tinygrad.device import Buffer, Device
 from tinygrad.uop.ops import UOp, Ops
+from tinygrad.shape.shapetracker import ShapeTracker

 # allocate some buffers + load in values
 out = Buffer(DEVICE, 1, dtypes.int32).allocate()
@@ -50,12 +51,13 @@ b = Buffer(DEVICE, 1, dtypes.int32).allocate().copyin(memoryview(bytearray(struc
 # NOTE: a._buf is the same as the return from cpu.allocator.alloc

 # describe the computation
-idx = UOp.const(dtypes.index, 0)
 buf_1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 1)
 buf_2 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 2)
-alu = buf_1.index(idx) + buf_2.index(idx)
+ld_1 = UOp(Ops.LOAD, dtypes.int32, (buf_1.view(ShapeTracker.from_shape((1,))),))
+ld_2 = UOp(Ops.LOAD, dtypes.int32, (buf_2.view(ShapeTracker.from_shape((1,))),))
+alu = ld_1 + ld_2
 output_buf = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 0)
-st_0 = UOp(Ops.STORE, dtypes.void, (output_buf.index(idx), alu))
+st_0 = UOp(Ops.STORE, dtypes.void, (output_buf.view(ShapeTracker.from_shape((1,))), alu))
 s = UOp(Ops.SINK, dtypes.void, (st_0,))

 # convert the computation to a "linearized" format (print the format)
@@ -78,7 +80,7 @@ print("******** third, the UOp ***********")

 from tinygrad.engine.realize import run_schedule
 from tinygrad.engine.schedule import create_schedule_with_vars
-from tinygrad.schedule.rangeify import get_rangeify_map
+from tinygrad.schedule.kernelize import get_kernelize_map

 # allocate some values + load in values
 a = UOp.new_buffer(DEVICE, 1, dtypes.int32)
@@ -91,10 +93,10 @@ out = a + b
 s = UOp(Ops.SINK, dtypes.void, (out,))

 # group the computation into kernels
-becomes_map = get_rangeify_map(s)
+becomes_map = get_kernelize_map(s)

 # the compute maps to an assign
-assign = becomes_map[a+b].base
+assign = becomes_map[a+b]

 # the first source is the output buffer (data)
 assert assign.src[0].op is Ops.BUFFER
@@ -0,0 +1,109 @@
+# 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`.
@@ -10,7 +10,7 @@ Directories are listed in order of how they are processed.

 Group UOps into kernels.

-::: tinygrad.schedule.rangeify.get_rangeify_map
+::: tinygrad.schedule.kernelize.get_kernelize_map
    options:
        members: false
        show_labels: false
@@ -41,7 +41,9 @@ BEAM                | [#]        | number of beams in kernel beam search
 DEFAULT_FLOAT       | [HALF, ...]| specify the default float dtype (FLOAT32, HALF, BFLOAT16, FLOAT64, ...), default to FLOAT32
 IMAGE               | [1-2]      | enable 2d specific optimizations
 FLOAT16             | [1]        | use float16 for images instead of float32
-HCQ_VISIBLE_DEVICES | [list[int]]| restricts the HCQ devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
+PTX                 | [1]        | enable the specialized [PTX](https://docs.nvidia.com/cuda/parallel-thread-execution/) assembler for Nvidia GPUs. If not set, defaults to generic CUDA codegen backend.
+PROFILE             | [1]        | enable profiling. This feature is supported in NV, AMD, QCOM and METAL backends.
+VISIBLE_DEVICES     | [list[int]]| restricts the NV/AMD devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
 JIT                 | [0-2]      | 0=disabled, 1=[jit enabled](quickstart.md#jit) (default), 2=jit enabled, but graphs are disabled
 VIZ                 | [1]        | 0=disabled, 1=[viz enabled](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/viz)
 ALLOW_TF32          | [1]        | enable TensorFloat-32 tensor cores on Ampere or newer GPUs.
@@ -2,17 +2,17 @@

 tinygrad supports various runtimes, enabling your code to scale across a wide range of devices. The default runtime can be automatically selected based on the available hardware, or you can force a specific runtime to be default using environment variables (e.g., `CPU=1`).

-| Runtime | Description | Compiler Options | Requirements |
-|---------|-------------|------------------|--------------|
-| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | nvrtc (default)<br>PTX (`NV_PTX=1`) | Ampere/Ada/Blackwell series GPUs.<br>You can select an interface via `NV_IFACE=(NVK\|PCI)`. See [NV interfaces](#nv-interfaces) for details. |
-| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | LLVM (`AMD_LLVM=1`)<br>HIP/COMGR (`AMD_HIP=1`) | RDNA2 or newer GPUs.<br>You can select an interface via `AMD_IFACE=(KFD\|PCI\|USB)`. See [AMD interfaces](#amd-interfaces) for details. |
-| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | - | 6xx series GPUs |
-| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | - | M1+ Macs; Metal 3.0+ for `bfloat` support |
-| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | nvrtc (default)<br> PTX (`CUDA_PTX=1`) | NVIDIA GPU with CUDA support |
-| [CL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cl.py) | Accelerates computations using OpenCL on GPUs | - | OpenCL 2.0 compatible device |
-| [CPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang or llvm compiler | Clang JIT (default)<br>LLVM IR (`CPU_LLVM=1`) | `clang` compiler in system `PATH` |
-| [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | - | Dawn library installed and discoverable. Binaries: [pydawn v0.3.0](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0) |
-
+| Runtime | Description | Requirements |
+|---------|-------------|--------------|
+| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | Ampere/Ada series GPUs |
+| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | RDNA2/RDNA3/RDNA4 series GPUs. You can select one of the interfaces for communication by setting `AMD_IFACE=(KFD|PCI)`. See [AMD interfaces](#amd-interfaces) for more details. |
+| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | 6xx series GPUs |
+| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | M1+ Macs; Metal 3.0+ for `bfloat` support |
+| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | NVIDIA GPU with CUDA support |
+| [OpenCL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cl.py) | Accelerates computations using OpenCL on GPUs | OpenCL 2.0 compatible device |
+| [CPU (C Code)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang compiler | `clang` compiler in system `PATH` |
+| [LLVM (LLVM IR)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_llvm.py) | Runs on CPU using the LLVM compiler infrastructure | llvm libraries installed and findable |
+| [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | Dawn library installed and findable. Download binaries [here](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0). |

 ## Interoperability

@@ -70,12 +70,5 @@ AMD backend supports several interfaces for communicating with devices:

 * `KFD`: uses the amdgpu driver
 * `PCI`: uses the [AM driver](developer/am.md)
-* `USB`: USB3 interafce for asm24xx chips.

 You can force an interface by setting `AMD_IFACE` to one of these values. In the case of `AMD_IFACE=PCI`, this may unbind your GPU from the amdgpu driver.
-
-## NV Interfaces
-NV backend supports several interfaces for communicating with devices:
-
-* `NVK`: uses the nvidia driver
-* `PCI`: uses the [NV driver](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/support/nv/nvdev.py)
@@ -10,7 +10,7 @@ GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]

 # override tinygrad defaults
 dtypes.default_float = dtypes.half
-Context(FUSE_OPTIM=1).__enter__()
+Context(FUSE_ARANGE=1, FUSE_OPTIM=1).__enter__()

 # from https://github.com/tysam-code/hlb-CIFAR10/blob/main/main.py
 batchsize = getenv("BS", 1024)
@@ -1,6 +1,6 @@
 import sys, time
 from tinygrad import TinyJit, GlobalCounters, fetch, getenv
-from tinygrad.nn.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner
 from extra.onnx_helpers import get_example_inputs, validate

 def load_onnx_model(onnx_file):
@@ -8,7 +8,7 @@ import numpy as np
 import subprocess
 import tensorflow as tf
 import tf2onnx
-from tinygrad.nn.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import to_mv
 from extra.export_model import export_model_clang, compile_net, jit_model
@@ -26,8 +26,8 @@ class Attention:
      start_pos = start_pos.val

    if HALF: x = x.half()
-    xqkv = self.c_attn(x).reshape(None, None, 3, self.n_heads, self.head_dim)
-    xq, xk, xv = [xqkv[:, :, i, :, :] for i in range(3)]
+    xqkv = self.c_attn(x)
+    xq, xk, xv = [xqkv.shrink((None, None, (i*self.dim, (i+1)*self.dim))).reshape(None, None, self.n_heads, self.head_dim) for i in range(3)]
    bsz, seqlen, _, _ = xq.shape

    # create kv cache
@@ -35,11 +35,11 @@ class Attention:
      self.cache_kv = Tensor.zeros(2, bsz, MAX_CONTEXT, self.n_heads, self.head_dim, dtype=x.dtype).contiguous().realize()

    # update the cache
-    self.cache_kv[:, :, start_pos:start_pos+seqlen, :, :].assign(Tensor.stack(xk, xv)).realize()
+    self.cache_kv.shrink((None, None,(start_pos,start_pos+seqlen),None,None)).assign(Tensor.stack(xk, xv)).realize()

    if start_pos > 0:
-      keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
-      values = self.cache_kv[1][:, :start_pos+seqlen, :, :]
+      keys = self.cache_kv[0].shrink((None, (0, start_pos+seqlen), None, None))
+      values = self.cache_kv[1].shrink((None, (0, start_pos+seqlen), None, None))
    else:
      keys = xk
      values = xv
@@ -64,7 +64,7 @@ class TransformerBlock:

  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]):
    h = x + self.attn(self.ln_1(x), start_pos, mask).float()
-    return (h + self.mlp(self.ln_2(h))).contiguous()
+    return (h + self.mlp(self.ln_2(h)))

 class Transformer:
  def __init__(self, dim, n_heads, n_layers, norm_eps, vocab_size, max_seq_len=1024):
@@ -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.randint(args.batch_size, args.benchmark), Variable("a", 0, MAX_CONTEXT).bind(0)).realize()
+    gpt2.model(Tensor.rand(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:
@@ -145,6 +145,7 @@ hyp = {
  },
 }

+@Context(FUSE_ARANGE=getenv("FUSE_ARANGE", 1))
 def train_cifar():

  def set_seed(seed):
@@ -228,8 +229,7 @@ def train_cifar():
    if getenv("RANDOM_CROP", 1):
      X = random_crop(X, crop_size=32)
    if getenv("RANDOM_FLIP", 1):
-      # NOTE: RANGEIFY=1 needs this contiguous or the X[perms] is very slow
-      X = (Tensor.rand(X.shape[0],1,1,1) < 0.5).where(X.flip(-1), X).contiguous() # flip LR
+      X = (Tensor.rand(X.shape[0],1,1,1) < 0.5).where(X.flip(-1), X) # flip LR
    X, Y = X[perms], Y[perms]
    return X, Y, *cutmix(X, Y, perms, mask_size=hyp['net']['cutmix_size'])

@@ -145,41 +145,6 @@ def NF4Linear(block_size):
      return new_state_dict
  return _NF4Linear

-def quantize_to_fp8(x: Tensor, dtype=dtypes.fp8e4m3):
-  fp8_min = -448.0 if dtype == dtypes.fp8e4m3 else -57344.0
-  fp8_max = 448.0 if dtype == dtypes.fp8e4m3 else 57344.0
-  scale = fp8_max / x.abs().max()
-  x_scl_sat = (x * scale).clamp(fp8_min, fp8_max)
-  return x_scl_sat.cast(dtype), scale.float().reciprocal()
-
-class FP8Linear:
-  def __init__(self, in_features, out_features, bias=True):
-    self.weight = Tensor.empty(out_features, in_features, dtype=dtypes.fp8e4m3)
-    self.bias = Tensor.empty(out_features, dtype=dtypes.float16) if bias else None
-    self.weight_scale = Tensor.empty((), dtype=dtypes.float16)
-
-  def __call__(self, x:Tensor):
-    y = x.dot(self.weight.T.cast(dtypes.float32)) * self.weight_scale
-    if self.bias is not None: y = y + self.bias.cast(y.dtype)
-    return y.cast(x.dtype)
-
-  @staticmethod
-  def quantize(tensors, device, scale_dtype=dtypes.float16, quantize_embeds=False):
-    assert not quantize_embeds
-    new_tensors = {}
-    for name,v in tensors.items():
-      if "feed_forward" in name or "attention.w" in name:
-        assert "weight" in name, name
-        fp8_weight, scale = quantize_to_fp8(v)
-        new_tensors[name] = fp8_weight
-        new_tensors[name.replace('weight', 'weight_scale')] = scale.cast(scale_dtype)
-        if isinstance(device, tuple):
-          new_tensors[name].shard_(device, axis=-1)
-          new_tensors[name.replace('weight', 'weight_scale')].shard_(device, axis=None)
-      else:
-        new_tensors[name] = v
-    return new_tensors
-
 MODEL_PARAMS = {
  "1B": {
    "args": {"dim": 2048, "n_heads": 32, "n_kv_heads": 8, "n_layers": 16, "norm_eps": 1e-5, "rope_theta": 500000, "vocab_size": 128256, "hidden_dim": 8192},
@@ -202,7 +167,6 @@ def build_transformer(model_path: Path, model_size="8B", quantize=None, scale_dt
  # build model
  if quantize == "int8": linear, embedding, quantize_embeds = Int8Linear, Int8Embedding, True
  elif quantize == "nf4": linear, embedding, quantize_embeds = NF4Linear(64), nn.Embedding, False
-  elif quantize == "fp8": linear, embedding, quantize_embeds = FP8Linear, nn.Embedding, False
  else: linear, embedding, quantize_embeds = nn.Linear, nn.Embedding, False
  model = Transformer(**MODEL_PARAMS[model_size]["args"], linear=linear, embedding=embedding, max_context=max_context, jit=True)

@@ -278,7 +242,7 @@ if __name__ == "__main__":
  parser.add_argument("--model", type=Path, help="Model path")
  parser.add_argument("--size", choices=["1B", "8B", "70B", "405B"], default="1B", help="Model size")
  parser.add_argument("--shard", type=int, default=1, help="Shard the model across multiple devices")
-  parser.add_argument("--quantize", choices=["int8", "nf4", "float16", "fp8"], help="Quantization method")
+  parser.add_argument("--quantize", choices=["int8", "nf4", "float16"], help="Quantization method")
  parser.add_argument("--no_api", action="store_true", help="Disable the api and run a cli test interface")
  parser.add_argument("--host", type=str, default="0.0.0.0", help="Web server bind address")
  parser.add_argument("--port", type=int, default=7776, help="Web server port")
@@ -511,33 +511,6 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
      # happens with BENCHMARK set
      pass

-# stable diffusion callbacks to match mlperf ref; declared here because they're pickled
-def filter_dataset(sample:dict): return {k:v for k,v in sample.items() if k in {'npy', 'txt'}}
-def collate(batch:list[dict]):
-  ret = {"npy": [], "txt": [], "__key__": []}
-  for sample in batch:
-    for k,v in sample.items():
-      ret[k].append(v)
-  return ret
-def collate_fn(batch): return batch
-
-# Reference (code): https://github.com/mlcommons/training/blob/2f4a93fb4888180755a8ef55f4b977ef8f60a89e/stable_diffusion/ldm/data/webdatasets.py, Line 55
-# Reference (params): https://github.com/mlcommons/training/blob/ab4ae1ca718d7fe62c369710a316dff18768d04b/stable_diffusion/configs/train_01x08x08.yaml, Line 107
-def batch_load_train_stable_diffusion(urls:str, BS:int):
-  import webdataset
-  dataset = webdataset.WebDataset(urls=urls, resampled=True, cache_size=-1, cache_dir=None)
-  dataset = dataset.shuffle(size=1000)
-  dataset = dataset.decode()
-  dataset = dataset.map(filter_dataset)
-  dataset = dataset.batched(BS, partial=False, collation_fn=collate)
-  dataset = webdataset.WebLoader(dataset, batch_size=None, shuffle=False, num_workers=1, persistent_workers=True, collate_fn=collate_fn)
-
-  for x in dataset:
-    assert isinstance(x, dict) and all(isinstance(k, str) for k in x.keys()) and all(isinstance(v, list) for v in x.values())
-    assert all(isinstance(moment_mean_logvar, np.ndarray) and moment_mean_logvar.shape==(1,8,64,64) for moment_mean_logvar in x["npy"])
-    assert all(isinstance(caption, str) for caption in x["txt"])
-    yield x
-
 # llama3

 class BinIdxDataset:
@@ -2,9 +2,7 @@ import math
 from typing import Union

 from tinygrad import Tensor, nn, dtypes
-from tinygrad.helpers import prod, argfix, Context
-from tinygrad.nn.state import get_parameters
-from extra.models.unet import UNetModel
+from tinygrad.helpers import prod, argfix

 # rejection sampling truncated randn
 def rand_truncn(*shape, dtype=None, truncstds=2, **kwargs) -> Tensor:
@@ -19,10 +17,6 @@ def he_normal(*shape, a: float = 0.00, **kwargs) -> Tensor:
  std = math.sqrt(2.0 / (1 + a ** 2)) / math.sqrt(prod(argfix(*shape)[1:])) / 0.87962566103423978
  return std * rand_truncn(*shape, **kwargs)

-# Stable Diffusion v2 training uses default torch gelu, which doesn't use tanh approximation
-def gelu_erf(x:Tensor) -> Tensor:
-  return 0.5 * x * (1.0 + (x / 1.4142135623730951).erf())
-
 class Conv2dHeNormal(nn.Conv2d):
  def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
    super().__init__(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
@@ -133,59 +127,3 @@ class Conv2dRetinaNet(nn.Conv2d):
  def __call__(self, x:Tensor) -> Tensor:
    return x.conv2d(self.weight.cast(dtypes.default_float), self.bias.cast(dtypes.default_float) if self.bias is not None else None,
                    groups=self.groups, stride=self.stride, dilation=self.dilation, padding=self.padding)
-
-# copy torch AMP: isolate mixed precision to just the below autocast ops, instead of using dtypes.default_float which affects all new Tensors
-class AutocastLinear(nn.Linear):
-  cast_dtype=dtypes.bfloat16 # enable monkeypatching of the mixed precision dtype
-  def __call__(self, x:Tensor) -> Tensor:
-    dtype = type(self).cast_dtype
-    return x.cast(dtype).linear(self.weight.cast(dtype).transpose(), self.bias.cast(dtype) if self.bias is not None else None)
-
-class AutocastConv2d(nn.Conv2d):
-  cast_dtype=dtypes.bfloat16
-  def __call__(self, x:Tensor) -> Tensor:
-    dtype = type(self).cast_dtype
-    return x.cast(dtype).conv2d(self.weight.cast(dtype), self.bias.cast(dtype), self.groups, self.stride, self.dilation, self.padding)
-
-# copy torch AMP: upcast to float32 before GroupNorm and LayerNorm
-class AutocastGroupNorm(nn.GroupNorm):
-  def __call__(self, x:Tensor) -> Tensor:
-    return super().__call__(x.cast(dtypes.float32))
-
-class AutocastLayerNorm(nn.LayerNorm):
-  def __call__(self, x:Tensor) -> Tensor:
-    return super().__call__(x.cast(dtypes.float32))
-
-def zero_module(module):
-  for p in get_parameters(module): p.assign(Tensor.zeros_like(p).contiguous())
-
-# Stable Diffusion mlperf reference doesn't call scaled_dot_product_attention
-# copy torch AMP: upcast to float32 before softmax on CUDA
-def attn_f32_softmax(q:Tensor, k:Tensor, v:Tensor) -> Tensor:
-  return (q.matmul(k.transpose(-2,-1), dtype=dtypes.float32) / math.sqrt(q.shape[-1])).softmax(-1).cast(q.dtype) @ v
-
-def init_stable_diffusion(version:str, pretrained:str, devices:list[str]):
-  from examples.stable_diffusion import StableDiffusion
-  from tinygrad.nn.state import safe_load, safe_save, load_state_dict, get_state_dict
-  from tempfile import TemporaryDirectory
-  model = StableDiffusion(version=version, pretrained=pretrained)
-  unet:UNetModel = model.model.diffusion_model
-
-  # this prevents extra consumption of memory, enabling much larger BS
-  Tensor.realize(*get_parameters(unet))
-  with TemporaryDirectory(prefix="unet_init") as tmp:
-    safe_save(get_state_dict(unet), init_fn:=f"{tmp}/init_model.safetensors")
-    load_state_dict(unet, safe_load(init_fn))
-
-  sqrt_alphas_cumprod = model.alphas_cumprod.sqrt().realize()
-  sqrt_one_minus_alphas_cumprod = (1 - model.alphas_cumprod).sqrt().realize()
-
-  if len(devices) > 1:
-    to_move = [sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod]
-    if version == "v2-mlperf-train": to_move += get_parameters(unet) + get_parameters(model.cond_stage_model)
-    for p in to_move:
-      p.to_(devices)
-    with Context(BEAM=0):
-      Tensor.realize(*to_move)
-
-  return model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod
@@ -1,9 +1,8 @@
 import math
-from tinygrad import dtypes, Tensor
+from tinygrad import dtypes
 from tinygrad.nn.optim import Optimizer

 from extra.lr_scheduler import LR_Scheduler
-from typing import Callable

 # https://github.com/mlcommons/training/blob/e237206991d10449d9675d95606459a3cb6c21ad/image_classification/tensorflow2/lars_util.py
 class PolynomialDecayWithWarmup(LR_Scheduler):
@@ -37,24 +36,4 @@ class CosineAnnealingLRWithWarmup(LR_Scheduler):
  def get_lr(self):
    warmup_lr = ((self.epoch_counter+1) / self.warmup_steps) * self.base_lr
    decay_lr = self.end_lr + 0.5 * (self.base_lr-self.end_lr) * (1 + (((self.epoch_counter+1-self.warmup_steps)/self.decay_steps) * math.pi).cos())
-    return (self.epoch_counter < self.warmup_steps).where(warmup_lr, decay_lr).cast(self.optimizer.lr.dtype)
-
-# Reference: https://github.com/mlcommons/training/blob/64b14a9abc74e08779a175abca7d291f8c957632/stable_diffusion/ldm/lr_scheduler.py, Lines 36-97
-class LambdaLinearScheduler:
-  def __init__(self, warm_up_steps:int, f_min:float, f_max:float, f_start:float, cycle_lengths:int):
-    self.lr_warm_up_steps, self.f_min, self.f_max, self.f_start, self.cycle_lengths = warm_up_steps, f_min, f_max, f_start, cycle_lengths
-
-  def schedule(self, n:Tensor) -> Tensor:
-    warm_up = (n < self.lr_warm_up_steps)
-    f_warm_up = (self.f_max - self.f_start) / self.lr_warm_up_steps * n + self.f_start
-    return warm_up.where(f_warm_up, self.f_min + (self.f_max - self.f_min) * (self.cycle_lengths - n) / (self.cycle_lengths))
-
-# based on torch.optim.lr_scheduler.LambdaLR
-class LambdaLR(LR_Scheduler):
-  def __init__(self, optimizer:Optimizer, base_lr:Tensor, lr_lambda:Callable):
-    super().__init__(optimizer)
-    self.base_lr, self.lr_lambda = base_lr, lr_lambda
-    self.step()
-
-  def get_lr(self):
-    return self.base_lr * self.lr_lambda(self.epoch_counter - 1)
+    return (self.epoch_counter < self.warmup_steps).where(warmup_lr, decay_lr).cast(self.optimizer.lr.dtype)
@@ -1,10 +1,10 @@
-import time, math, os
+import time, math
 start = time.perf_counter()
 from pathlib import Path
 import numpy as np
 from tinygrad import Tensor, Device, dtypes, GlobalCounters, TinyJit
 from tinygrad.nn.state import get_parameters, load_state_dict, safe_load
-from tinygrad.helpers import getenv, Context, prod
+from tinygrad.helpers import getenv
 from extra.bench_log import BenchEvent, WallTimeEvent
 def tlog(x): print(f"{x:25s}  @ {time.perf_counter()-start:5.2f}s")

@@ -287,256 +287,6 @@ def eval_llama3():
  log_perplexity = np.mean(losses)
  print(f"Log Perplexity: {log_perplexity}")

-# NOTE: BEAM hangs on 8xmi300x with DECODE_BS=384 in final realize below; function is declared here for external testing
-@TinyJit
-def vae_decode(x:Tensor, vae, disable_beam=False) -> Tensor:
-  from examples.stable_diffusion import AutoencoderKL
-  assert isinstance(vae, AutoencoderKL)
-  x = vae.post_quant_conv(1./0.18215 * x)
-
-  x = vae.decoder.conv_in(x)
-  x = vae.decoder.mid(x)
-  for i, l in enumerate(vae.decoder.up[::-1]):
-    print("decode", x.shape)
-    for b in l['block']: x = b(x)
-    if 'upsample' in l:
-      bs,c,py,px = x.shape
-      x = x.reshape(bs, c, py, 1, px, 1).expand(bs, c, py, 2, px, 2).reshape(bs, c, py*2, px*2)
-      x = l['upsample']['conv'](x)
-    if i == len(vae.decoder.up) - 1 and disable_beam:
-      with Context(BEAM=0): x.realize()
-    else: x.realize()
-  x = vae.decoder.conv_out(vae.decoder.norm_out(x).swish())
-
-  x = ((x + 1.0) / 2.0).clip(0.0, 1.0)
-  return x
-
-def eval_stable_diffusion():
-  import csv, PIL, sys
-  from tqdm import tqdm
-  from examples.mlperf.initializers import init_stable_diffusion, gelu_erf
-  from examples.stable_diffusion import AutoencoderKL
-  from extra.models.unet import UNetModel
-  from tinygrad.nn.state import load_state_dict, torch_load
-  from tinygrad.helpers import BEAM
-  from extra.models import clip
-  from extra.models.clip import FrozenOpenClipEmbedder
-  from extra.models.clip import OpenClipEncoder
-  from extra.models.inception import FidInceptionV3
-
-  config = {}
-  GPUS               = config["GPUS"]                   = [f"{Device.DEFAULT}:{i}" for i in range(getenv("GPUS", 1))]
-  for x in GPUS: Device[x]
-  print(f"running eval on {GPUS}")
-  seed               = config["seed"]                   = getenv("SEED", 12345)
-  CKPTDIR            = config["CKPTDIR"]                = Path(getenv("CKPTDIR", "./checkpoints"))
-  DATADIR            = config["DATADIR"]                = Path(getenv("DATADIR", "./datasets"))
-  CONTEXT_BS         = config["CONTEXT_BS"]             = getenv("CONTEXT_BS", 1 * len(GPUS))
-  DENOISE_BS         = config["DENOISE_BS"]             = getenv("DENOISE_BS", 1 * len(GPUS))
-  DECODE_BS          = config["DECODE_BS"]              = getenv("DECODE_BS", 1 * len(GPUS))
-  INCEPTION_BS       = config["INCEPTION_BS"]           = getenv("INCEPTION_BS", 1 * len(GPUS))
-  CLIP_BS            = config["CLIP_BS"]                = getenv("CLIP_BS", 1 * len(GPUS))
-  EVAL_CKPT_DIR      = config["EVAL_CKPT_DIR"]          = getenv("EVAL_CKPT_DIR", "")
-  STOP_IF_CONVERGED  = config["STOP_IF_CONVERGED"]      = getenv("STOP_IF_CONVERGED", 0)
-
-  if (WANDB := getenv("WANDB", "")):
-    import wandb
-    wandb.init(config=config, project="MLPerf-Stable-Diffusion")
-
-  assert EVAL_CKPT_DIR != "", "provide a directory with checkpoints to be evaluated"
-  print(f"running eval on checkpoints in {EVAL_CKPT_DIR}\nSEED={seed}")
-  eval_queue:list[tuple[int, Path]] = []
-  for p in Path(EVAL_CKPT_DIR).iterdir():
-    if p.name.endswith(".safetensors"):
-      ckpt_iteration = p.name.split(".safetensors")[0]
-      assert ckpt_iteration.isdigit(), f"invalid checkpoint name: {p.name}, expected <digits>.safetensors"
-      eval_queue.append((int(ckpt_iteration), p))
-  assert len(eval_queue), f'no files ending with ".safetensors" were found in {EVAL_CKPT_DIR}'
-  print(sorted(eval_queue, reverse=True))
-
-  Tensor.manual_seed(seed)  # seed for weight initialization
-  model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod = init_stable_diffusion("v2-mlperf-eval", CKPTDIR / "sd" / "512-base-ema.ckpt", GPUS)
-
-  # load prompts for generating images for validation; 2 MB of data total
-  with open(DATADIR / "coco2014" / "val2014_30k.tsv") as f:
-    reader = csv.DictReader(f, delimiter="\t")
-    eval_inputs:list[dict] = [{"image_id": int(row["image_id"]), "id": int(row["id"]), "caption": row["caption"]} for row in reader]
-  assert len(eval_inputs) == 30_000
-  # NOTE: the clip weights are the same between model.cond_stage_model and clip_encoder
-  eval_timesteps = list(reversed(range(1, 1000, 20)))
-
-  original_device, Device.DEFAULT = Device.DEFAULT, "CPU"
-  # The choice of alphas_prev[0] = alphas_cumprod[0] seems arbitrary, but it's how the mlperf ref does it:
-  #   alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
-  eval_alphas_prev = model.alphas_cumprod[0:1].cat(model.alphas_cumprod[list(range(1, 1000, 20))[:-1]]).to(GPUS).realize()
-  inception = FidInceptionV3().load_from_pretrained(CKPTDIR / "inception" / "pt_inception-2015-12-05-6726825d.pth")
-  vision_cfg = {'width': 1280, 'layers': 32, 'd_head': 80, 'image_size': 224, 'patch_size': 14}
-  text_cfg = {'width': 1024, 'n_heads': 16, 'layers': 24, 'vocab_size': 49408, 'ctx_length': 77}
-  clip.gelu = gelu_erf
-  clip_encoder = OpenClipEncoder(1024, text_cfg, vision_cfg)
-  loaded = torch_load(CKPTDIR / "clip" / "open_clip_pytorch_model.bin")
-  loaded.update({"attn_mask": clip_encoder.attn_mask, "mean": clip_encoder.mean, "std": clip_encoder.std})
-  load_state_dict(clip_encoder, loaded)
-  Device.DEFAULT=original_device
-
-  @TinyJit
-  def denoise_step(x:Tensor, x_x:Tensor, t_t:Tensor, uc_c:Tensor, sqrt_alphas_cumprod_t:Tensor, sqrt_one_minus_alphas_cumprod_t:Tensor,
-                    alpha_prev:Tensor, unet:UNetModel, GPUS) -> Tensor:
-    out_uncond, out = unet(x_x, t_t, uc_c).to("CPU").reshape(-1, 2, 4, 64, 64).chunk(2, dim=1)
-    out_uncond = out_uncond.squeeze(1).shard(GPUS,axis=0)
-    out = out.squeeze(1).shard(GPUS,axis=0)
-    v_t = out_uncond + 8.0 * (out - out_uncond)
-    e_t = sqrt_alphas_cumprod_t * v_t + sqrt_one_minus_alphas_cumprod_t * x
-    pred_x0 = sqrt_alphas_cumprod_t * x - sqrt_one_minus_alphas_cumprod_t * v_t
-    dir_xt = (1. - alpha_prev).sqrt() * e_t
-    x_prev = alpha_prev.sqrt() * pred_x0 + dir_xt
-    return x_prev.realize()
-
-  def shard_tensor(t:Tensor) -> Tensor: return t.shard(GPUS, axis=0) if len(GPUS) > 1 else t.to(GPUS[0])
-  def get_batch(whole:Tensor, i:int, bs:int) -> tuple[Tensor, int]:
-    batch = whole[i: i + bs].to("CPU")
-    if (unpadded_bs:=batch.shape[0]) < bs:
-      batch = batch.cat(batch[-1:].expand(bs - unpadded_bs, *batch[-1].shape))
-    return batch, unpadded_bs 
-
-  @Tensor.train(mode=False)
-  def eval_unet(eval_inputs:list[dict], unet:UNetModel, cond_stage:FrozenOpenClipEmbedder, first_stage:AutoencoderKL,
-                inception:FidInceptionV3, clip:OpenClipEncoder) -> tuple[float, float]:
-    # Eval is divided into 5 jits, one per model
-    # It doesn't make sense to merge these jits, e.g. unet repeats 50 times in isolation; images fork to separate inception/clip
-    # We're generating and scoring 30,000 images per eval, and all the data can flow through one jit at a time
-    # To maximize throughput for each jit, we have only one model/jit on the GPU at a time, and pool outputs from each jit off-GPU
-    for model in (unet, first_stage, inception, clip):
-      Tensor.realize(*[p.to_("CPU") for p in get_parameters(model)])
-
-    uc_written = False
-    models = (cond_stage, unet, first_stage, inception, clip)
-    jits = (jit_context:=TinyJit(cond_stage.embed_tokens), denoise_step, vae_decode, jit_inception:=TinyJit(inception),
-            jit_clip:=TinyJit(clip.get_clip_score))
-    all_bs = (CONTEXT_BS, DENOISE_BS, DECODE_BS, INCEPTION_BS, CLIP_BS)
-    if (EVAL_SAMPLES:=getenv("EVAL_SAMPLES", 0)) and EVAL_SAMPLES > 0:
-      eval_inputs = eval_inputs[0:EVAL_SAMPLES]
-    output_shapes = [(ns:=len(eval_inputs),77), (ns,77,1024), (ns,4,64,64), (ns,3,512,512), (ns,2048), (ns,)]
-    # Writing progress to disk lets us resume eval if we crash
-    stages = ["tokens", "embeds", "latents", "imgs", "inception", "clip"]
-    disk_tensor_names, disk_tensor_shapes = stages + ["end", "uc"], output_shapes + [(6,), (1,77,1024)]
-    if not all(os.path.exists(f"{EVAL_CKPT_DIR}/{name}.bytes") for name in disk_tensor_names):
-      for name, shape in zip(disk_tensor_names, disk_tensor_shapes):
-        file = Path(f"{EVAL_CKPT_DIR}/{name}.bytes")
-        file.unlink(missing_ok=True)
-        with file.open("wb") as f: f.truncate(prod(shape) * 4)
-    progress = {name: Tensor.empty(*shape, device=f"disk:{EVAL_CKPT_DIR}/{name}.bytes", dtype=dtypes.int if name in {"tokens", "end"} else dtypes.float)
-                for name, shape in zip(disk_tensor_names, disk_tensor_shapes)}
-
-    def embed_tokens(tokens:Tensor) -> Tensor:
-      nonlocal uc_written
-      if not uc_written:
-        with Context(BEAM=0): progress["uc"].assign(cond_stage.embed_tokens(cond_stage.tokenize("").to(GPUS)).to("CPU").realize()).realize()
-        uc_written = True
-      return jit_context(shard_tensor(tokens))
-
-    def generate_latents(embeds:Tensor) -> Tensor:
-      uc_c = Tensor.stack(progress["uc"].to("CPU").expand(bs, 77, 1024), embeds, dim=1).reshape(-1, 77, 1024)
-      uc_c = shard_tensor(uc_c)
-      x = shard_tensor(Tensor.randn(bs,4,64,64))
-      for step_idx, timestep in enumerate(tqdm(eval_timesteps)):
-        reversed_idx = Tensor([50 - step_idx - 1], device=GPUS)
-        alpha_prev = eval_alphas_prev[reversed_idx]
-        ts = Tensor.full(bs, fill_value=timestep, dtype=dtypes.int, device="CPU")
-        ts_ts = shard_tensor(ts.cat(ts))
-        ts = shard_tensor(ts)
-        sqrt_alphas_cumprod_t = sqrt_alphas_cumprod[ts].reshape(bs, 1, 1, 1)
-        sqrt_one_minus_alphas_cumprod_t = sqrt_one_minus_alphas_cumprod[ts].reshape(bs, 1, 1, 1)
-        x_x = shard_tensor(Tensor.stack(x.to("CPU"), x.to("CPU"), dim=1).reshape(-1, 4, 64, 64))
-        x.assign(denoise_step(x, x_x, ts_ts, uc_c, sqrt_alphas_cumprod_t, sqrt_one_minus_alphas_cumprod_t, alpha_prev, unet, GPUS)).realize()
-      return x
-
-    def decode_latents(latents:Tensor) -> Tensor: return vae_decode(shard_tensor(latents), first_stage, disable_beam=True)
-    def generate_inception(imgs:Tensor) -> Tensor: return jit_inception(shard_tensor(imgs))[:,:,0,0]
-
-    def calc_clip_scores(batch:Tensor, batch_tokens:Tensor) -> Tensor:
-      # Tensor.interpolate does not yet support bicubic, so we use PIL
-      batch = (batch.to(GPUS[0]).permute(0,2,3,1) * 255).clip(0, 255).cast(dtypes.uint8).numpy()
-      batch = [np.array(PIL.Image.fromarray(batch[i]).resize((224,224), PIL.Image.BICUBIC)) for i in range(bs)]
-      batch = shard_tensor(Tensor(np.stack(batch, axis=0).transpose(0,3,1,2), device="CPU").realize())
-      batch = batch.cast(dtypes.float) / 255
-      batch = (batch - model.mean) / model.std
-      batch = jit_clip(shard_tensor(batch_tokens), batch)
-      return batch
-
-    callbacks = (embed_tokens, generate_latents, decode_latents, generate_inception, calc_clip_scores)
-
-    # save every forward pass output to disk; NOTE: this needs ~100 GB disk space because 30k images are large
-    def stage_progress(stage_idx:int) -> int: return progress["end"].to("CPU")[stage_idx].item()
-    if stage_progress(0) < len(eval_inputs):
-      tokens = []
-      for i in tqdm(range(0, len(eval_inputs), CONTEXT_BS)):
-        subset = [cond_stage.tokenize(row["caption"], device="CPU") for row in eval_inputs[i: i+CONTEXT_BS]]
-        tokens.append(Tensor.cat(*subset, dim=0).realize())
-      progress["tokens"].assign(Tensor.cat(*tokens, dim=0).realize()).realize()
-      progress["end"][0:1].assign(Tensor([len(eval_inputs)], dtype=dtypes.int)).realize()
-    prev_stage = "tokens"
-    tokens = progress["tokens"]
-
-    # wrapper code for every model
-    for stage_idx, model, jit, bs, callback in zip(range(1,6), models, jits, all_bs, callbacks):
-      stage = stages[stage_idx]
-      if stage_progress(stage_idx) >= len(eval_inputs):
-        prev_stage = stage
-        continue # use cache
-      t0 = time.perf_counter()
-      print(f"starting eval with model: {model}")
-      if stage_idx == 1: inputs = tokens
-      elif stage_idx == 5: inputs = progress["imgs"]
-      else: inputs = progress[prev_stage]
-
-      Tensor.realize(*[p.to_(GPUS) for p in get_parameters(model)])
-      for batch_idx in tqdm(range(stage_progress(stage_idx), inputs.shape[0], bs)):
-        t1 = time.perf_counter()
-        batch, unpadded_bs = get_batch(inputs, batch_idx, bs)
-        if isinstance(model, OpenClipEncoder): batch = callback(batch, get_batch(tokens, batch_idx, bs)[0].realize())
-        else: batch = callback(batch)
-        # to(GPUS[0]) is necessary for this to work, without that the result is still on GPUS, probably due to a bug
-        batch = batch.to(GPUS[0]).to("CPU")[0:unpadded_bs].realize()
-        progress[stage][batch_idx: batch_idx + bs].assign(batch).realize()
-        # keep track of what our last output was, so we can resume from there if we crash in this loop
-        progress["end"][stage_idx: stage_idx + 1].assign(Tensor([batch_idx + bs], dtype=dtypes.int)).realize()
-        print(f"model: {model}, batch_idx: {batch_idx}, elapsed: {(time.perf_counter() - t1):.2f}")
-      del batch
-        
-      jit.reset()
-      Tensor.realize(*[p.to_("CPU") for p in get_parameters(model)])
-      print(f"done with model: {model}, elapsed: {(time.perf_counter() - t0):.2f}")
-      prev_stage = stage
-
-    inception_stats_fn = str(DATADIR / "coco2014" / "val2014_30k_stats.npz")
-    fid_score = inception.compute_score(progress["inception"].to("CPU"), inception_stats_fn)
-    clip_score = progress["clip"].to(GPUS[0]).mean().item()
-    for name in disk_tensor_names:
-      Path(f"{EVAL_CKPT_DIR}/{name}.bytes").unlink(missing_ok=True)
-    
-    if EVAL_SAMPLES and BEAM:
-      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
-      sys.exit() # Don't eval additional models; we don't care about clip/fid scores when running BEAM on eval sample subset
-
-    return clip_score, fid_score
-
-  # evaluate checkpoints in reverse chronological order
-  for ckpt_iteration, p in sorted(eval_queue, reverse=True):
-    unet_ckpt = safe_load(p)
-    load_state_dict(unet, unet_ckpt)
-    clip_score, fid_score = eval_unet(eval_inputs, unet, model.cond_stage_model, model.first_stage_model, inception, clip_encoder)
-    converged = True if clip_score >= 0.15 and fid_score <= 90 else False
-    print(f"eval results for {EVAL_CKPT_DIR}/{p.name}: clip={clip_score}, fid={fid_score}, converged={converged}")
-    if WANDB:
-      wandb.log({"eval/ckpt_iteration": ckpt_iteration, "eval/clip_score": clip_score, "eval/fid_score": fid_score})
-    if converged and STOP_IF_CONVERGED:
-      print(f"Convergence detected, exiting early before evaluating other checkpoints due to STOP_IF_CONVERGED={STOP_IF_CONVERGED}")
-      sys.exit()
-
-  # for testing
-  return clip_score, fid_score, ckpt_iteration
-
 if __name__ == "__main__":
  # inference only
  Tensor.training = False
@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing

 from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
-from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling
+from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, FUSE_CONV_BW, Profiling
 from tinygrad.nn.state import get_parameters, get_state_dict, load_state_dict, safe_load, safe_save
 from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam, AdamW

@@ -707,7 +707,7 @@ def train_unet3d():
  ```BASEDIR=<folder_path> ./examples/mlperf/scripts/setup_kits19_dataset.sh```

  2) To start training the model, run the following:
-  ```time PYTHONPATH=. WANDB=1 TRAIN_BEAM=3 GPUS=6 BS=6 MODEL=unet3d python3 examples/mlperf/model_train.py```
+  ```time PYTHONPATH=. WANDB=1 TRAIN_BEAM=3 FUSE_CONV_BW=1 GPUS=6 BS=6 MODEL=unet3d python3 examples/mlperf/model_train.py```
  """
  from examples.mlperf.losses import dice_ce_loss
  from examples.mlperf.metrics import dice_score
@@ -749,6 +749,7 @@ def train_unet3d():
    "train_beam": TRAIN_BEAM,
    "eval_beam": EVAL_BEAM,
    "wino": WINO.value,
+    "fuse_conv_bw": FUSE_CONV_BW.value,
    "gpus": GPUS,
    "default_float": dtypes.default_float.name
  }
@@ -1188,9 +1189,7 @@ def train_bert():
      if MLLOGGER and RUNMLPERF:
        MLLOGGER.start(key=mllog_constants.EVAL_START, value=None, metadata={"epoch_num": i*GBS, "step_num": i})
      if getenv("RESET_STEP"): train_step_bert.reset()
-      elif getenv("FREE_INTERMEDIATE", 0) and train_step_bert.captured is not None:
-        # TODO: FREE_INTERMEDIATE nan'ed after jit step 2
-        train_step_bert.captured.free_intermediates()
+      elif getenv("FREE_INTERMEDIATE", 1) and train_step_bert.captured is not None: train_step_bert.captured.free_intermediates()
      eval_lm_losses = []
      eval_clsf_losses = []
      eval_lm_accs = []
@@ -1224,7 +1223,7 @@ def train_bert():
          return

      if getenv("RESET_STEP"): eval_step_bert.reset()
-      elif getenv("FREE_INTERMEDIATE", 0) and eval_step_bert.captured is not None: eval_step_bert.captured.free_intermediates()
+      elif getenv("FREE_INTERMEDIATE", 1) and eval_step_bert.captured is not None: eval_step_bert.captured.free_intermediates()

      del eval_data
      avg_lm_loss = sum(eval_lm_losses) / len(eval_lm_losses)
@@ -1310,7 +1309,7 @@ def train_llama3():
  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
  EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 5.6)

-  # LR=1e-4 TRAIN_ON_VAL=1 DEFAULT_FLOAT=bfloat16 JITBEAM=2 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B WARMUP_STEPS=36 DECAY_STEPS=360 SEQLEN=512 PYTHONPATH=. AMD=1 AMD_LLVM=0 MODEL=llama3 python3 examples/mlperf/model_train.py
+  # LR=1e-4 TRAIN_ON_VAL=1 DEFAULT_FLOAT=bfloat16 FUSE_ARANGE=1 JITBEAM=2 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B WARMUP_STEPS=36 DECAY_STEPS=360 SEQLEN=512 PYTHONPATH=. AMD=1 AMD_LLVM=0 MODEL=llama3 python3 examples/mlperf/model_train.py
  # trains to 7

  opt_adamw_beta_1 = 0.9
@@ -1494,144 +1493,6 @@ def train_llama3():
          safe_save(get_state_dict(model), fn)
        break

-def train_stable_diffusion():
-  from extra.models.unet import UNetModel
-  from examples.mlperf.dataloader import batch_load_train_stable_diffusion
-  from examples.mlperf.lr_schedulers import LambdaLR, LambdaLinearScheduler
-  from examples.mlperf.initializers import init_stable_diffusion
-  from examples.mlperf.helpers import get_training_state
-  import numpy as np
-
-  config = {}
-  GPUS               = config["GPUS"]                   = [f"{Device.DEFAULT}:{i}" for i in range(getenv("GPUS", 1))]
-  seed               = config["seed"]                   = getenv("SEED", 12345)
-  # ** hyperparameters **
-  BS                 = config["BS"]                     = getenv("BS", 1 * len(GPUS))
-  BASE_LR            = config["LEARNING_RATE"]          = getenv("LEARNING_RATE", 2.5e-7)
-  # https://github.com/mlcommons/training_policies/blob/cfa99da479b8d5931f7a3c67612d021dfb47510a/training_rules.adoc#benchmark_specific_rules
-  # "Checkpoint must be collected every 512,000 images. CEIL(512000 / global_batch_size) if 512000 is not divisible by GBS."
-  # NOTE: It's inferred that "steps" is the unit for the output of the CEIL formula, based on all other cases of CEIL in the rules
-  CKPT_STEP_INTERVAL = config["CKPT_STEP_INTERVAL"]     = getenv("CKPT_STEP_INTERVAL", math.ceil(512_000 / BS))
-  CKPTDIR            = config["CKPTDIR"]                = Path(getenv("CKPTDIR", "./checkpoints"))
-  DATADIR            = config["DATADIR"]                = Path(getenv("DATADIR", "./datasets"))
-  UNET_CKPTDIR       = config["UNET_CKPTDIR"]           = Path(getenv("UNET_CKPTDIR", "./checkpoints"))
-  TOTAL_CKPTS        = config["TOTAL_CKPTS"]            = getenv("TOTAL_CKPTS", 0)
-
-  print(f"training on {GPUS}")
-  lr = BS * BASE_LR
-  print(f"BS={BS}, BASE_LR={BASE_LR}, lr={lr}")
-  print(f"CKPT_STEP_INTERVAL = {CKPT_STEP_INTERVAL}")
-  for x in GPUS: Device[x]
-  if (WANDB := getenv("WANDB", "")):
-    import wandb
-    wandb.init(config=config, project="MLPerf-Stable-Diffusion")
-
-  Tensor.manual_seed(seed)  # seed for weight initialization
-  model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod = init_stable_diffusion("v2-mlperf-train", CKPTDIR / "sd" / "512-base-ema.ckpt", GPUS)
-
-  optimizer = AdamW(get_parameters(unet))
-  lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
-  lr_scheduler = LambdaLR(optimizer, Tensor(lr, dtype=dtypes.float, device=optimizer.device), lambda_lr_callback)
-
-  @TinyJit
-  def train_step(mean:Tensor, logvar:Tensor, tokens:Tensor, unet:UNetModel, optimizer:LAMB, lr_scheduler:LambdaLR) -> Tensor:
-    optimizer.zero_grad()
-
-    timestep = Tensor.randint(BS, low=0, high=model.alphas_cumprod.shape[0], dtype=dtypes.int, device=GPUS[0])
-    latent_randn = Tensor.randn(*mean.shape, device=GPUS[0])
-    noise = Tensor.randn(*mean.shape, device=GPUS[0])
-    for t in (mean, logvar, tokens, timestep, latent_randn, noise):
-      t.shard_(GPUS, axis=0)
-
-    std = Tensor.exp(0.5 * logvar.clamp(-30.0, 20.0))
-    latent = (mean + std * latent_randn) * 0.18215
-
-    sqrt_alphas_cumprod_t = sqrt_alphas_cumprod[timestep].reshape(timestep.shape[0], 1, 1, 1)
-    sqrt_one_minus_alphas_cumprod_t = sqrt_one_minus_alphas_cumprod[timestep].reshape(timestep.shape[0], 1, 1, 1)
-    latent_with_noise = sqrt_alphas_cumprod_t * latent + sqrt_one_minus_alphas_cumprod_t * noise
-    v_true = sqrt_alphas_cumprod_t * noise - sqrt_one_minus_alphas_cumprod_t * latent
-
-    context = model.cond_stage_model.embed_tokens(tokens)
-
-    out = unet(latent_with_noise, timestep, context)
-    loss = ((out - v_true) ** 2).mean()
-    del mean, logvar, std, latent, noise, sqrt_alphas_cumprod_t, sqrt_one_minus_alphas_cumprod_t
-    del out, v_true, context, latent_randn, tokens, timestep
-    loss.backward()
-
-    optimizer.step()
-    lr_scheduler.step()
-    loss, out_lr = loss.detach().to("CPU"), optimizer.lr.to("CPU")
-    Tensor.realize(loss, out_lr)
-    return loss, out_lr
-    
-  # checkpointing takes ~9 minutes without this, and ~1 minute with this
-  @TinyJit
-  def ckpt_to_cpu():
-    ckpt = get_training_state(unet, optimizer, lr_scheduler)
-    # move to CPU first so more GPU bufs aren't created (can trigger OOM)
-    for k,v in ckpt.items(): ckpt[k] = v.detach().to("CPU")
-    Tensor.realize(*[v for v in ckpt.values()])
-    for k,v in ckpt.items(): ckpt[k] = v.cast(v.dtype.base).contiguous()
-    Tensor.realize(*[v for v in ckpt.values()])
-    return ckpt
-
-  # training loop
-  dl = batch_load_train_stable_diffusion(f'{DATADIR}/laion-400m/webdataset-moments-filtered/{{00000..00831}}.tar', BS)
-  # for tests
-  saved_checkpoints = []
-
-  train_start_time = time.perf_counter()
-  t0 = t6 = time.perf_counter()
-  for i, batch in enumerate(dl, start=1):
-    loop_time = time.perf_counter() - t0
-    t0 = time.perf_counter()
-    dl_time = t0 - t6
-    GlobalCounters.reset()
-
-    mean, logvar = np.split(np.concatenate(batch["npy"], axis=0), 2, axis=1)
-    mean, logvar = Tensor(mean, dtype=dtypes.float32, device="CPU"), Tensor(logvar, dtype=dtypes.float32, device="CPU")
-    tokens = []
-    for text in batch['txt']: tokens += model.cond_stage_model.tokenizer.encode(text, pad_with_zeros=True)
-    tokens = Tensor(tokens, dtype=dtypes.int32, device="CPU").reshape(-1, 77)
-
-    t1 = time.perf_counter()
-    loss, lr = train_step(mean, logvar, tokens, unet, optimizer, lr_scheduler)
-    loss_item, lr_item = loss.item(), lr.item()
-    t2 = time.perf_counter()
-
-    if i == 3:
-      for _ in range(3): ckpt_to_cpu() # do this at the beginning of run to prevent OOM surprises when checkpointing
-      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
-      
-    total_train_time = time.perf_counter() - train_start_time
-    if WANDB:
-      wandb.log({"train/loss": loss_item, "train/lr": lr_item, "train/loop_time_prev": loop_time, "train/dl_time": dl_time, "train/step": i,
-                    "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (t2-t1), "train/input_prep_time": t1-t0,
-                    "train/train_step_time": t2-t1, "train/total_time": total_train_time})
-
-      if i == 1 and wandb.run is not None:
-        with open(f"{UNET_CKPTDIR}/wandb_run_id_{wandb.run.id}", "w") as f:
-          f.write(f"wandb.run.id = {wandb.run.id}")
-
-    if i % CKPT_STEP_INTERVAL == 0:
-      # https://github.com/mlcommons/training_policies/blob/cfa99da479b8d5931f7a3c67612d021dfb47510a/training_rules.adoc#benchmark_specific_rules
-      # "evaluation is done offline, the time is not counted towards the submission time."
-      fn = f"{UNET_CKPTDIR}/{i}.safetensors"
-      print(f"saving unet checkpoint at {fn}")
-      saved_checkpoints.append(fn)
-      safe_save({k.replace("model.", ""):v for k,v in ckpt_to_cpu().items() if k.startswith("model.")}, fn)
-      if TOTAL_CKPTS and i == TOTAL_CKPTS * CKPT_STEP_INTERVAL:
-        print(f"ending run after {i} steps ({TOTAL_CKPTS} checkpoints collected)")
-        return saved_checkpoints
-
-    t3 = time.perf_counter()
-    print(f"""step {i}: {GlobalCounters.global_ops * 1e-9 / (t2-t1):9.2f} GFLOPS, mem_used: {GlobalCounters.mem_used / 1e9:.2f} GB,
-  loop_time_prev: {loop_time:.2f}, dl_time: {dl_time:.2f}, input_prep_time: {t1-t0:.2f}, train_step_time: {t2-t1:.2f},
-  t3-t2: {t3-t2:.4f}, loss:{loss_item:.5f}, lr:{lr_item:.3e}, total_train_time:{total_train_time:.2f}
-  """)
-    t6 = time.perf_counter()
-
 if __name__ == "__main__":
  multiprocessing.set_start_method('spawn')

@@ -1640,7 +1501,7 @@ if __name__ == "__main__":
  else: bench_log_manager = contextlib.nullcontext()

  with Tensor.train():
-    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn,stable_diffusion").split(","):
+    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn").split(","):
      nm = f"train_{m}"
      if nm in globals():
        print(f"training {m}")
@@ -1,57 +0,0 @@
-#!/usr/bin/env bash
-# adapted from https://github.com/mlcommons/training/blob/4bdf5c8ed218ad76565a2ba1ac27c919ccc6d689/stable_diffusion/README.md
-
-# setup dirs
-
-DATA=/raid/datasets/stable_diffusion
-
-LAION=$DATA/laion-400m/webdataset-moments-filtered 
-COCO=$DATA/coco2014
-mkdir -p $LAION $COCO
-
-CKPT=/raid/weights/stable_diffusion
-mkdir -p $CKPT/clip $CKPT/sd $CKPT/inception
-
-# download data
-
-# if rclone isn't installed system-wide / in your PATH, put the executable path in quotes below
-#RCLONE=""
-RCLONE="rclone"
-
-## VAE-encoded image latents, from 6.1M image subset of laion-400m
-## about 1 TB for whole download
-$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/laion-400m/moments-webdataset-filtered/ ${LAION} --include="*.tar" -P
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/laion-400m/moments-webdataset-filtered/sha512sums.txt ${LAION} -P
-cd $LAION && grep -E '\.tar$' sha512sums.txt | sha512sum -c --quiet - && \
-  echo "All .tar files verified" || { echo "Checksum failure when validating downloaded Laion moments"; exit 1; }
-
-## prompts and FID statistics from 30k image subset of coco2014
-## 33 MB
-$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/coco2014/val2014_30k.tsv ${COCO} -P
-
-$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/coco2014/val2014_30k_stats.npz ${COCO} -P
-
-# download checkpoints
-
-## clip (needed for text and vision encoders for validation)
-CLIP_WEIGHTS_URL="https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/resolve/main/open_clip_pytorch_model.bin"
-CLIP_WEIGHTS_SHA256="9a78ef8e8c73fd0df621682e7a8e8eb36c6916cb3c16b291a082ecd52ab79cc4"
-CLIP_CONFIG_URL="https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/raw/main/open_clip_config.json"
-wget -N -P ${CKPT}/clip ${CLIP_WEIGHTS_URL}
-wget -N -P ${CKPT}/clip ${CLIP_CONFIG_URL}
-echo "${CLIP_WEIGHTS_SHA256}  ${CKPT}/clip/open_clip_pytorch_model.bin"                    | sha256sum -c
-
-## sd (needed for latent->image decoder for validation, also has clip text encoder for training)
-SD_WEIGHTS_URL='https://huggingface.co/stabilityai/stable-diffusion-2-base/resolve/main/512-base-ema.ckpt'
-SD_WEIGHTS_SHA256="d635794c1fedfdfa261e065370bea59c651fc9bfa65dc6d67ad29e11869a1824"
-wget -N -P ${CKPT}/sd ${SD_WEIGHTS_URL}
-echo "${SD_WEIGHTS_SHA256}  ${CKPT}/sd/512-base-ema.ckpt"                    | sha256sum -c
-
-## inception (needed for validation)
-FID_WEIGHTS_URL='https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth'
-FID_WEIGHTS_SHA1="bd836944fd6db519dfd8d924aa457f5b3c8357ff"
-wget -N -P ${CKPT}/inception ${FID_WEIGHTS_URL}
-echo "${FID_WEIGHTS_SHA1}  ${CKPT}/inception/pt_inception-2015-12-05-6726825d.pth"                    | sha1sum -c
@@ -1,72 +0,0 @@
-#!/usr/bin/env bash
-
-DATETIME=${2:-$(date "+%m%d%H%M")}
-LOGFILE="${HOME}/logs/sd_mi300x_${DATETIME}.log"
-# UNET_CKPTDIR must be set: training saves checkpoints to this path, then a separate eval process scans this path to know which checkpoints to eval
-export UNET_CKPTDIR="${HOME}/stable_diffusion/training_checkpoints/${DATETIME}"
-mkdir -p "${HOME}/logs" "$UNET_CKPTDIR"
-
-# run this script in isolation when using the --bg flag
-if [[ "${1:-}" == "--bg" ]]; then
-  echo "logging output to $LOGFILE"
-  echo "saving UNet checkpoints to $UNET_CKPTDIR"
-  script_path="$(readlink -f "${BASH_SOURCE[0]}")"
-  nohup bash "$script_path" run "$DATETIME" >"$LOGFILE" 2>&1 & disown $!
-  exit 0
-fi
-
-# venv management
-if [[ -d .venv-sd-mlperf ]]; then
-  . .venv-sd-mlperf/bin/activate
-else
-  python3 -m venv .venv-sd-mlperf && . .venv-sd-mlperf/bin/activate
-  pip install --index-url https://download.pytorch.org/whl/cpu torch && pip install tqdm numpy ftfy regex pillow scipy wandb webdataset
-fi
-pip list
-apt list --installed | grep amdgpu
-rocm-smi --version
-modinfo amdgpu | grep version
-
-export BEAM=2 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 IGNORE_JIT_FIRST_BEAM=1 HCQDEV_WAIT_TIMEOUT_MS=300000
-export AMD_LLVM=0 # bf16 seems to require this
-export DATADIR="/raid/datasets/stable_diffusion"
-export CKPTDIR="/raid/weights/stable_diffusion"
-export EVAL_CKPT_DIR=$UNET_CKPTDIR
-export MODEL="stable_diffusion" PYTHONPATH="."
-export GPUS=8 BS=304
-export CONTEXT_BS=816 DENOISE_BS=600 DECODE_BS=384 INCEPTION_BS=560 CLIP_BS=240
-export WANDB=1
-export PARALLEL=4
-export PYTHONUNBUFFERED=1
-sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setperfdeterminism 1500 || exit 1
-
-# Retry BEAM search if script fails before BEAM COMPLETE is printed, but don't retry after that
-run_retry(){ local try=0 max=5 code tmp py pgid kids
-  while :; do
-    tmp=$(mktemp)
-    setsid bash -c 'exec env "$@"' _ "$@" > >(tee -a "$LOGFILE" | tee "$tmp") 2>&1 &
-    py=$!; pgid=$(ps -o pgid= -p "$py" | tr -d ' ')
-    wait "$py"; code=$?
-    [[ -n "$pgid" ]] && { kill -TERM -"$pgid" 2>/dev/null; sleep 1; kill -KILL -"$pgid" 2>/dev/null; }
-    kids=$(pgrep -P "$py" || true)
-    while [[ -n "$kids" ]]; do
-      kill -TERM $kids 2>/dev/null; sleep 0.5
-      kids=$(for k in $kids; do pgrep -P "$k" || true; done)
-    done
-    grep -q 'BEAM COMPLETE' "$tmp" && { rm -f "$tmp"; return 1; }
-    rm -f "$tmp"
-    ((code==0)) && return 0
-    ((try>=max)) && return 2
-    ((try++)); sleep 90; echo "try = ${try}"
-  done
-}
-
-# Power limiting to 400W is only needed if GPUs fall out of sync (causing 2.2x increased train time) at higher power, which has been observed at 450W
-sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setpoweroverdrive 750 && \
-run_retry TOTAL_CKPTS=7 python3 examples/mlperf/model_train.py; (( $? == 2 )) && { echo "training failed before BEAM completion"; exit 2; }
-sleep 90
-
-run_retry EVAL_SAMPLES=600 python3 examples/mlperf/model_eval.py; (( $? == 2 )) && { echo "eval failed before BEAM completion"; exit 2; }
-# Checkpoints will be evaluated in reverse chronological order, even if above training crashed early
-# STOP_IF_CONVERGED=1: Stop the eval after the first time convergence is detected; no more checkpoints will be evaluated after that.
-STOP_IF_CONVERGED=1 python3 examples/mlperf/model_eval.py
@@ -1,17 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128
-
-export IGNORE_OOB=1
-
-export BEAM=3 BEAM_UOPS_MAX=4000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-# export BEAM_LOG_SURPASS_MAX=1
-# export BASEDIR="/raid/datasets/wiki"
-
-export RESET_STEP=1
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,17 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-export BENCHMARK=10 BERT_LAYERS=2
-
-python3 examples/mlperf/model_train.py
@@ -1,20 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,31 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_8xMI300X"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_8xMI300x_${DATETIME}_${SEED}.log"
-
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,17 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BEAM_LOG_SURPASS_MAX=1
-export BASEDIR="/raid/datasets/wiki"
-
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,16 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,28 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." NV=1
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,18 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BEAM_LOG_SURPASS_MAX=1
-export BASEDIR="/raid/datasets/wiki"
-
-export RESET_STEP=1
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,16 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,31 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_red"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_red_${DATETIME}_${SEED}.log"
-
-export HCQDEV_WAIT_TIMEOUT_MS=100000  # prevents hang?
-
-# init
-sleep 5 && sudo rmmod amdgpu || true
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,50 +0,0 @@
-# 1. Problem
-
-This problem uses the ResNet-50 CNN to do image classification.
-
-## Requirements
-
-Install tinygrad and mlperf-logging from master.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-### tinybox_red
-Disable cwsr
-This is the default on production tinybox red.
-```
-sudo vi /etc/modprobe.d/amdgpu.conf
-cat <<EOF > /etc/modprobe.d/amdgpu.conf
-options amdgpu cwsr_enable=0
-EOF
-sudo update-initramfs -u
-sudo reboot
-
-# validate
-sudo cat /sys/module/amdgpu/parameters/cwsr_enable #= 0
-```
-
-# 2. Directions
-
-## Steps to download and verify data
-
-```
-IMGNET_TRAIN=1 python3 extra/datasets/imagenet_download.py
-```
-
-## Steps for one time setup
-
-### tinybox_red
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/setup.sh
-```
-
-## Steps to run benchmark
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-```
@@ -1,13 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
-
-export BENCHMARK=10 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
-
-export EVAL_START_EPOCH=3 EVAL_FREQ=4
-
-export WANDB=1 PARALLEL=0
-
-python3 examples/mlperf/model_train.py
@@ -1,25 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." NV=1
-export MODEL="resnet"
-export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="resnet_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 EVAL_START_EPOCH=3 EVAL_FREQ=4 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,50 +0,0 @@
-# 1. Problem
-
-This problem uses the ResNet-50 CNN to do image classification.
-
-## Requirements
-
-Install tinygrad and mlperf-logging from master.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-### tinybox_red
-Disable cwsr
-This is the default on production tinybox red.
-```
-sudo vi /etc/modprobe.d/amdgpu.conf
-cat <<EOF > /etc/modprobe.d/amdgpu.conf
-options amdgpu cwsr_enable=0
-EOF
-sudo update-initramfs -u
-sudo reboot
-
-# validate
-sudo cat /sys/module/amdgpu/parameters/cwsr_enable #= 0
-```
-
-# 2. Directions
-
-## Steps to download and verify data
-
-```
-IMGNET_TRAIN=1 python3 extra/datasets/imagenet_download.py
-```
-
-## Steps for one time setup
-
-### tinybox_red
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/setup.sh
-```
-
-## Steps to run benchmark
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-```
@@ -1,13 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=10 DEBUG=${DEBUG:-2}
-
-python3 examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export EVAL_START_EPOCH=3 EVAL_FREQ=4
-
-export WANDB=1 PARALLEL=0
-
-python3 examples/mlperf/model_train.py
@@ -1,26 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." AMD=1
-export MODEL="resnet"
-export SUBMISSION_PLATFORM="tinybox_red"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="resnet_red_${DATETIME}_${SEED}.log"
-
-# init
-sleep 5 && sudo rmmod amdgpu || true
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 EVAL_START_EPOCH=3 EVAL_FREQ=4 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,8 +0,0 @@
-#!/bin/bash
-
-rocm-smi --setprofile compute
-rocm-smi --setmclk 3
-rocm-smi --setperflevel high
-
-# power cap to 350W
-echo "350000000" | sudo tee /sys/class/drm/card{1..6}/device/hwmon/hwmon*/power1_cap
@@ -1,38 +0,0 @@
-# 1. Problem
-
-This problem uses RetinaNet for SSD.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-Also install the following dependencies:
-```
-pip install tqdm numpy pycocotools boto3 pandas torch torchvision
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download data
-
-Run the following:
-```
-BASEDIR=/raid/datasets/openimages python3 extra/datasets/openimages.py
-```
-
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/run_and_time.sh
-```
@@ -1,14 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=5 DEBUG=2
-
-python examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export WANDB=1 PARALLEL=0
-export RUNMLPERF=1
-
-python examples/mlperf/model_train.py
@@ -1,25 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." NV=1
-export MODEL="retinanet"
-export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export TRAIN_BEAM=2 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/openimages"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="retinanet_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,14 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=5 DEBUG=2
-
-python examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export WANDB=1 PARALLEL=0
-export RUNMLPERF=1
-
-python examples/mlperf/model_train.py
@@ -1,38 +0,0 @@
-{
-    "submitter": "tinycorp",
-    "division": "closed",
-    "status": "Available on-premise",
-    "system_name": "tinybox 8xMI300X",
-    "number_of_nodes": "1",
-    "host_processors_per_node": "2",
-    "host_processor_model_name": "AMD EPYC 9354",
-    "host_processor_core_count": "32",
-    "host_processor_vcpu_count": "64",
-    "host_processor_frequency": "",
-    "host_processor_caches": "",
-    "host_processor_interconnect": "",
-    "host_memory_capacity": "2304GB",
-    "host_storage_type": "NVMe SSD",
-    "host_storage_capacity": "3x 4TB raid array",
-    "host_networking": "",
-    "host_networking_topology": "",
-    "host_memory_configuration": "24x 96GB DDR5",
-    "accelerators_per_node": "8",
-    "accelerator_model_name": "AMD Instinct MI300X 192GB HBM3",
-    "accelerator_host_interconnect": "PCIe 5.0 x16",
-    "accelerator_frequency": "",
-    "accelerator_on-chip_memories": "",
-    "accelerator_memory_configuration": "HBM3",
-    "accelerator_memory_capacity": "192GB",
-    "accelerator_interconnect": "",
-    "accelerator_interconnect_topology": "",
-    "cooling": "air",
-    "hw_notes": "",
-    "framework": "tinygrad, branch mlperf_training_v5.0",
-    "other_software_stack": {
-        "python": "3.10.16",
-        "ROCm": "3.0.0+94441cb"
-    },
-    "operating_system": "Ubuntu 24.04.1 LTS",
-    "sw_notes": ""
-  }
@@ -1,38 +0,0 @@
-{
-  "submitter": "tinycorp",
-  "division": "closed",
-  "status": "Available on-premise",
-  "system_name": "tinybox green",
-  "number_of_nodes": "1",
-  "host_processors_per_node": "1",
-  "host_processor_model_name": "AMD EPYC 7532",
-  "host_processor_core_count": "32",
-  "host_processor_vcpu_count": "64",
-  "host_processor_frequency": "",
-  "host_processor_caches": "",
-  "host_processor_interconnect": "",
-  "host_memory_capacity": "128GB",
-  "host_storage_type": "NVMe SSD",
-  "host_storage_capacity": "4 TB raid array + 1 TB boot",
-  "host_networking": "",
-  "host_networking_topology": "",
-  "host_memory_configuration": "8x 16GB DDR4",
-  "accelerators_per_node": "6",
-  "accelerator_model_name": "NVIDIA GeForce RTX 4090",
-  "accelerator_host_interconnect": "PCIe 4.0 x16",
-  "accelerator_frequency": "",
-  "accelerator_on-chip_memories": "",
-  "accelerator_memory_configuration": "GDDR6X",
-  "accelerator_memory_capacity": "24GB",
-  "accelerator_interconnect": "",
-  "accelerator_interconnect_topology": "",
-  "cooling": "air",
-  "hw_notes": "",
-  "framework": "tinygrad, branch mlperf_training_v5.0",
-  "other_software_stack": {
-    "python": "3.10.12",
-    "CUDA": "12.4"
-  },
-  "operating_system": "Ubuntu 22.04.4",
-  "sw_notes": ""
-}
@@ -1,37 +0,0 @@
-{
-  "submitter": "tinycorp",
-  "division": "closed",
-  "status": "Available on-premise",
-  "system_name": "tinybox red",
-  "number_of_nodes": "1",
-  "host_processors_per_node": "1",
-  "host_processor_model_name": "AMD EPYC 7532",
-  "host_processor_core_count": "32",
-  "host_processor_vcpu_count": "64",
-  "host_processor_frequency": "",
-  "host_processor_caches": "",
-  "host_processor_interconnect": "",
-  "host_memory_capacity": "128GB",
-  "host_storage_type": "NVMe SSD",
-  "host_storage_capacity": "4 TB raid array + 1 TB boot",
-  "host_networking": "",
-  "host_networking_topology": "",
-  "host_memory_configuration": "8x 16GB DDR4",
-  "accelerators_per_node": "6",
-  "accelerator_model_name": "AMD Radeon RX 7900 XTX",
-  "accelerator_host_interconnect": "PCIe 4.0 x16",
-  "accelerator_frequency": "",
-  "accelerator_on-chip_memories": "",
-  "accelerator_memory_configuration": "GDDR6",
-  "accelerator_memory_capacity": "24GB",
-  "accelerator_interconnect": "",
-  "accelerator_interconnect_topology": "",
-  "cooling": "air",
-  "hw_notes": "",
-  "framework": "tinygrad, branch mlperf_training_v5.0",
-  "other_software_stack": {
-    "python": "3.10.12"
-  },
-  "operating_system": "Ubuntu 22.04.4",
-  "sw_notes": ""
-}
@@ -1,12 +1,16 @@
-import os, sys, pickle, time, re
+import os, sys, pickle, time
 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
 from tinygrad.engine.realize import CompiledRunner

 import onnx
-from tinygrad.nn.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner

 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"
@@ -17,14 +21,11 @@ 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)
-  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()}
+  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()}
  print("created tensors")

  run_onnx_jit = TinyJit(lambda **kwargs:
@@ -32,6 +33,8 @@ 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
@@ -49,8 +52,6 @@ def compile(onnx_file):
      kernel_count += 1
      read_image_count += ei.prg.p.src.count("read_image")
      gated_read_image_count += ei.prg.p.src.count("?read_image")
-      for v in [m.group(1) for m in re.finditer(r'(val\d+)\s*=\s*read_imagef\(', ei.prg.p.src)]:
-        if len(re.findall(fr'[\?\:]{v}\.[xyzw]', ei.prg.p.src)) > 0: gated_read_image_count += 1
  print(f"{kernel_count=},  {read_image_count=}, {gated_read_image_count=}")
  if (allowed_kernel_count:=getenv("ALLOWED_KERNEL_COUNT", -1)) != -1:
    assert kernel_count == allowed_kernel_count, f"different kernels! {kernel_count=}, {allowed_kernel_count=}"
@@ -66,76 +67,85 @@ 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 inputs, test_val
+  return test_val

-def test_vs_compile(run, inputs, test_val=None):
+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}}

  # run 20 times
-  step_times = []
  for _ in range(20):
    st = time.perf_counter()
    out = run(**inputs)
    mt = time.perf_counter()
    val = out.numpy()
    et = time.perf_counter()
-    step_times.append((et-st)*1e3)
-    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {step_times[-1]:6.2f} ms")
-
-  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
-    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(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {(et-st)*1e3:6.2f} 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
-  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)
+  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)
  return val

-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()}
+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()}
  onnx_model = onnx.load(onnx_file)

-  ORT_TO_NP_DTYPES: dict[str, np.dtype] = {
-    'tensor(float)': np.dtype('float32'),
-    'tensor(float16)': np.dtype('float16'),
-    'tensor(uint8)': np.dtype('uint8'),
-  }
-
  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}
+  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()

-  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")
+  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")
  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)
-  inputs, outputs = compile(onnx_file)
+  test_val = compile(onnx_file) if not getenv("RUN") else None

  with open(OUTPUT, "rb") as f: pickle_loaded = pickle.load(f)

-  test_vs_compile(pickle_loaded, inputs, outputs)
-  if getenv("SELFTEST"):
-    test_vs_onnx(inputs, outputs, onnx_file, 1e-4)
+  # 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"))

-  if getenv("BENCHMARK_LOG", ""):
-    bench(pickle_loaded, inputs)
@@ -0,0 +1,47 @@
+import sys
+from tinygrad import Tensor, fetch, GlobalCounters, dtypes
+from tinygrad.uop.ops import UOp
+from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.schedule.kernelize import get_kernelize_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_kernelize_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()
@@ -27,7 +27,7 @@ class Model(nn.Module):

 if __name__ == "__main__":
  if getenv("TINY_BACKEND"):
-    import tinygrad.nn.torch  # noqa: F401
+    import tinygrad.frontend.torch  # noqa: F401
    device = torch.device("tiny")
  else:
    device = torch.device({"METAL":"mps","NV":"cuda"}.get(Device.DEFAULT, "cpu"))
@@ -99,7 +99,6 @@ 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
@@ -113,22 +112,19 @@ if __name__ == "__main__":

  model = StableDiffusionV2(**params)

-  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    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)))
+  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)
+  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
+    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))
-
-    Tensor.realize(*get_state_dict(model).values())
+          v.replace(v.cast(dtypes.float16).realize())

  c  = { "crossattn": model.cond_stage_model(args.prompt) }
  uc = { "crossattn": model.cond_stage_model("") }
@@ -9,13 +9,11 @@ from typing import Dict, Any
 from PIL import Image
 import numpy as np
 from tinygrad import Device, GlobalCounters, dtypes, Tensor, TinyJit
-from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten
+from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm
 from tinygrad.nn import Conv2d, GroupNorm
 from tinygrad.nn.state import torch_load, load_state_dict, get_state_dict
-from extra.models.clip import Closed, Tokenizer, FrozenOpenClipEmbedder
-from extra.models import unet, clip
+from extra.models.clip import Closed, Tokenizer
 from extra.models.unet import UNetModel
-from examples.mlperf.initializers import AutocastLinear, AutocastConv2d, AutocastGroupNorm, AutocastLayerNorm, zero_module, attn_f32_softmax, gelu_erf
 from extra.bench_log import BenchEvent, WallTimeEvent

 class AttnBlock:
@@ -156,46 +154,12 @@ unet_params: Dict[str,Any] = {
  "use_linear": False,
 }

-mlperf_params: Dict[str,Any] = {"adm_in_ch": None, "in_ch": 4, "out_ch": 4, "model_ch": 320, "attention_resolutions": [4, 2, 1], "num_res_blocks": 2,
-                                "channel_mult": [1, 2, 4, 4], "d_head": 64, "transformer_depth": [1, 1, 1, 1], "ctx_dim": 1024, "use_linear": True,
-                                "num_groups":16, "st_norm_eps":1e-6}
-
 class StableDiffusion:
-  def __init__(self, version:str|None=None, pretrained:str|None=None):
+  def __init__(self):
    self.alphas_cumprod = get_alphas_cumprod()
-    if version != "v2-mlperf-train":
-      self.first_stage_model = AutoencoderKL() # only needed for decoding generated latents to images; not needed in mlperf training from preprocessed moments
-
-    if not version:
-      self.cond_stage_model = namedtuple("CondStageModel", ["transformer"])(transformer = namedtuple("Transformer", ["text_model"])(text_model = Closed.ClipTextTransformer()))
-      unet_init_params = unet_params
-    elif version in {"v2-mlperf-train", "v2-mlperf-eval"}:
-      unet_init_params = mlperf_params
-      clip.gelu = gelu_erf
-      self.cond_stage_model = FrozenOpenClipEmbedder(**{"dims": 1024, "n_heads": 16, "layers": 24, "return_pooled": False, "ln_penultimate": True,
-                                                        "clip_tokenizer_version": "sd_mlperf_v5_0"})
-      unet.Linear, unet.Conv2d, unet.GroupNorm, unet.LayerNorm = AutocastLinear, AutocastConv2d, AutocastGroupNorm, AutocastLayerNorm
-      unet.attention, unet.gelu, unet.mixed_precision_dtype = attn_f32_softmax, gelu_erf, dtypes.bfloat16
-      if pretrained:
-        print("loading text encoder")
-        weights: dict[str,Tensor] = {k.replace("cond_stage_model.", "", 1):v for k,v in torch_load(pretrained)["state_dict"].items() if k.startswith("cond_stage_model.")}
-        weights["model.attn_mask"] = Tensor.full((77, 77), fill_value=float("-inf")).triu(1)
-        load_state_dict(self.cond_stage_model, weights)
-        # only the eval model needs the decoder
-        if version == "v2-mlperf-eval":
-          print("loading image latent encoder")
-          weights = {k.replace("first_stage_model.", "", 1):v for k,v in torch_load(pretrained)["state_dict"].items() if k.startswith("first_stage_model.")}
-          load_state_dict(self.first_stage_model, weights)
-
-    self.model = namedtuple("DiffusionModel", ["diffusion_model"])(diffusion_model = UNetModel(**unet_init_params))
-    if version == "v2-mlperf-train":
-      # the mlperf reference inits certain weights as zeroes
-      for bb in flatten(self.model.diffusion_model.input_blocks) + self.model.diffusion_model.middle_block + flatten(self.model.diffusion_model.output_blocks):
-        if isinstance(bb, unet.ResBlock):
-          zero_module(bb.out_layers[3])
-        elif isinstance(bb, unet.SpatialTransformer):
-          zero_module(bb.proj_out)
-      zero_module(self.model.diffusion_model.out[2])
+    self.model = namedtuple("DiffusionModel", ["diffusion_model"])(diffusion_model = UNetModel(**unet_params))
+    self.first_stage_model = AutoencoderKL()
+    self.cond_stage_model = namedtuple("CondStageModel", ["transformer"])(transformer = namedtuple("Transformer", ["text_model"])(text_model = Closed.ClipTextTransformer()))

  def get_x_prev_and_pred_x0(self, x, e_t, a_t, a_prev):
    temperature = 1
@@ -263,23 +227,18 @@ 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):
-    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)
+    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'], strict=False)

    if args.fp16:
      for k,v in get_state_dict(model).items():
        if k.startswith("model"):
-          v.replace(v.cast(dtypes.float16))
-
-    Tensor.realize(*get_state_dict(model).values())
+          v.replace(v.cast(dtypes.float16).realize())

  # run through CLIP to get context
  tokenizer = Tokenizer.ClipTokenizer()
@@ -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
-# python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
-# VIZ=1 python3 examples/benchmark_onnx.py /tmp/model.quant.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

 def imagenet_dataloader(cnt=0):
  input_mean = Tensor([0.485, 0.456, 0.406]).reshape(1, -1, 1, 1)
@@ -32,7 +32,7 @@ if __name__ == "__main__":

  lr = 5e-3
  transform = ComposeTransforms([
-    lambda x: [Image.fromarray(xx).resize((64, 64)) for xx in x],
+    lambda x: [Image.fromarray(xx, mode='L').resize((64, 64)) for xx in x],
    lambda x: np.stack([np.asarray(xx) for xx in x], 0),
    lambda x: x / 255.0,
    lambda x: np.tile(np.expand_dims(x, 1), (1, 3, 1, 1)).astype(np.float32),
@@ -109,7 +109,7 @@ class TextDecoder:

  def forward(self, x:Tensor, pos:Union[Variable, Literal[0]], encoded_audio:Tensor):
    seqlen = x.shape[-1]
-    x = self.token_embedding(x) + self.positional_embedding.shrink(((pos, pos+seqlen), None))
+    x = self.token_embedding(x) + self.positional_embedding.shrink(((pos, pos+seqlen), None, None))
    for block in self.blocks: x = block(x, xa=encoded_audio, mask=self.mask, len=pos)
    return self.output_tok(x)

@@ -2,7 +2,7 @@
 import os
 from ultralytics import YOLO
 from pathlib import Path
-from tinygrad.nn.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner
 from extra.onnx_helpers import get_example_inputs

 os.chdir("/tmp")
@@ -1,4 +0,0 @@
-# source extra/cl_android.sh
-export LD_LIBRARY_PATH=/data/data/com.termux/files/usr/lib:/system/vendor/lib64
-export LD_PRELOAD=/system/vendor/lib64/libOpenCL.so
-
@@ -1,180 +1,355 @@
 from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
-from tinygrad.uop.ops import UOp, KernelInfo
-from tinygrad.engine.realize import ExecItem, get_runner
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, graph_rewrite, AxisType, PatternMatcher, UPat
+from tinygrad.engine.realize import CompiledRunner, ExecItem, get_program
 from tinygrad.dtype import AddrSpace
-from tinygrad.helpers import getenv
+from tinygrad.helpers import getenv, colored, prod, unwrap
+from tinygrad.shape.shapetracker import ShapeTracker, View
+from tinygrad.shape.view import strides_for_shape
+from tinygrad.codegen.opt.kernel import axis_colors, Opt, OptOps
+from tinygrad.codegen.opt.swizzler import merge_views, view_left
+
+def to_colored(full_shape, axis_types): return '_'.join([colored(str(s), axis_colors[at]) for s,at in zip(full_shape, axis_types)])

 N = 4096
 run_count = 5

-# ---------------------------
-# launch/config constants
-# ---------------------------
+BN = 128
+BM = 128
+BK = 8

-WARP_SIZE = 32
+TN = 4
+TM = 4

-# Threadblock tile sizes (block-level tile of C that a block computes)
-BLOCK_N = 128   # columns of C (N-dim) per block
-BLOCK_M = 128   # rows of C (M-dim) per block
-BLOCK_K = 8     # K-slice per block iteration
+# NOTE: this is from testgrad
+# change reduceop axes and input ShapeTrackers, view gets replaced with a reshape.
+# src->r->view  -->   src->view->r
+def swizzle_reduceop(src:UOp, r:UOp, view:UOp):
+  if r.tag is not None: return None
+  # confirm the input is in order
+  # TODO: replace this with a UOp that allows for nothing else then remove this
+  permute = tuple(i for i in range(len(src.shape)) if i not in r.axis_arg)+r.axis_arg
+  assert permute == tuple(range(len(permute))), f"reduce axis must already be in order, {permute} isn't"

-# Register tile sizes (per-thread accumulator tile of C)
-TN = 4     # columns per thread
-TM = 4     # rows per thread
+  # append the reduce shape to each of the views
+  prshape = prod(rshape:=src.shape[-len(r.axis_arg):])
+  rstrides = strides_for_shape(rshape)
+  nv = [View.create(v.shape+rshape, tuple(x*prshape for x in v.strides)+rstrides, v.offset*prshape,
+                    v.mask+tuple((0,s) for s in rshape) if v.mask is not None else None) for v in unwrap(view.st).views]

-is_kernel5 = getenv("K5", 0)
-THREADS_PER_BLOCK = 128 if is_kernel5 else 256
-assert THREADS_PER_BLOCK % BLOCK_N == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_N"
-assert THREADS_PER_BLOCK % BLOCK_K == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_K"
-assert (BLOCK_N * BLOCK_K) % THREADS_PER_BLOCK == 0
-assert (BLOCK_M * BLOCK_K) % THREADS_PER_BLOCK == 0
+  # no reshape required with shrinking REDUCE_AXIS
+  return UOp(Ops.REDUCE_AXIS, r.dtype, (src.view(ShapeTracker(tuple(nv))),),
+             (r.arg[0], tuple(range(len(view.shape), len(view.shape) + len(r.axis_arg)))))

-WARPS_PER_BLOCK = THREADS_PER_BLOCK // WARP_SIZE
-WAVE_TILE_N = 128 if is_kernel5 else 64
-WAVE_TILE_M = BLOCK_N * BLOCK_M // WARPS_PER_BLOCK // WAVE_TILE_N
-assert BLOCK_N % WAVE_TILE_N == 0, "BN must be a multiple of WN"
-assert BLOCK_M % WAVE_TILE_M == 0, "BM must be a multiple of WM"
-WAVES_IN_BLOCK_X = BLOCK_N // WAVE_TILE_N
-WAVES_IN_BLOCK_Y = BLOCK_M // WAVE_TILE_M
-assert WAVES_IN_BLOCK_X * WAVES_IN_BLOCK_Y == WARPS_PER_BLOCK, "wave grid must match warps/block"
+pm = PatternMatcher([
+  (UPat(Ops.VIEW, src=(UPat(Ops.REDUCE_AXIS, src=(UPat.var("src"),), name="r"),), name="view"), swizzle_reduceop),
+])

-LANES_PER_WAVE_X = 8
-LANES_PER_WAVE_Y = 4
-ITERS_PER_WAVE_N = WAVE_TILE_N // (LANES_PER_WAVE_X * TN)
-ITERS_PER_WAVE_M = WAVE_TILE_M // (LANES_PER_WAVE_Y * TM)
-N_PER_ITER = WAVE_TILE_N // ITERS_PER_WAVE_N
-M_PER_ITER = WAVE_TILE_M // ITERS_PER_WAVE_M
-assert WAVE_TILE_N % (LANES_PER_WAVE_X * TN) == 0, "WAVE_TILE_N must be divisible by LANES_PER_WAVE_X*TN"
-assert WAVE_TILE_M % (LANES_PER_WAVE_Y * TM) == 0, "WAVE_TILE_M must be divisible by LANES_PER_WAVE_Y*TM"
+def rangeify_kernel3():
+  a = Tensor.empty(N,N)
+  b = Tensor.empty(N,N)
+  c = a@b
+  #c = c.reshape((32,2,16,4,32,2,16,4)).contiguous()
+  with Context(RANGEIFY=1):
+    sink = c.schedule()[-1].ast
+  #print(sink)

-def hand_spec_kernel3():
-  # ---------------------------
-  # per-thread read mapping
-  # ---------------------------
-  # A: read BK x BN tiles; B: read BN x BK tiles
-  tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
+  opts  = [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.UPCAST, 0, 2)]
+  opts += [Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.LOCAL, 1, 16), Opt(OptOps.UPCAST, 1, 2)]
+  opts += [Opt(OptOps.UNROLL, 0, 8)]

-  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
-  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
-  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y
+  return sink.replace(arg=KernelInfo(opts_to_apply=tuple(opts)))

-  idxInWave = (tid % WARP_SIZE) % LANES_PER_WAVE_X
-  idyInWave = (tid % WARP_SIZE) // LANES_PER_WAVE_X
-  assert idyInWave.vmax+1 == LANES_PER_WAVE_Y
+def top_spec_kernel3():
+  a = Tensor.empty(N,N)
+  b = Tensor.empty(N,N)
+  c = a@b
+  sink = c.schedule()[-1].ast
+  L = 16
+  sink = sink.reshape((N//L, L, N//L, L)) #.lift({0:UOp.range(N//BM, 0), 2:UOp.range(N//BN, 1)})
+  sink = graph_rewrite(sink, view_left+pm)
+  axis_types = (AxisType.GLOBAL, AxisType.LOCAL, AxisType.GLOBAL, AxisType.LOCAL, AxisType.REDUCE)
+  return sink.replace(arg=KernelInfo(name="top_"+to_colored(sink.full_shape, axis_types), axis_types=axis_types))

-  # ---------------------------
-  # block indices & placeholders
-  # ---------------------------
-  blockIdx_x = UOp.special(N // BLOCK_N, "gidx0")
-  blockIdx_y = UOp.special(N // BLOCK_M, "gidx1")
+def hl_spec_kernel3():
+  nbIterWaveM = 2
+  nbIterWaveN = 2

-  a = UOp.placeholder((N, N), dtypes.float, slot=1)
-  b = UOp.placeholder((N, N), dtypes.float, slot=2)
-  c = UOp.placeholder((N, N), dtypes.float, slot=0)
+  # define buffers
+  # TODO: remove these views once the defines have a shape
+  a = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=1).view(ShapeTracker.from_shape((N,N)))
+  b = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=2).view(ShapeTracker.from_shape((N,N))).permute((1,0))
+  c = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=0).view(ShapeTracker.from_shape((N,N)))
+  As = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(BK*BM, AddrSpace.LOCAL), arg=0).view(ShapeTracker.from_shape((BK, BM))).permute((1,0))
+  Bs = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(BK*BN, AddrSpace.LOCAL), arg=1).view(ShapeTracker.from_shape((BK, BN))).permute((1,0))
+  A_col = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbIterWaveM * TM, AddrSpace.REG), arg=0).view(ShapeTracker.from_shape((nbIterWaveM * TM,)))
+  B_row = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbIterWaveN * TN, AddrSpace.REG), arg=1).view(ShapeTracker.from_shape((nbIterWaveN * TN,)))

-  BM_As_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
-  As = UOp.placeholder((BLOCK_K, BM_As_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
-  Bs = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
+  # shape buffers. TODO: permutes
+  full_shape = (N//BM, nbIterWaveM, BM//(nbIterWaveM * TM), TM, N//BN, nbIterWaveN, BN//(nbIterWaveN * TN), TN, N//BK, BK)
+  a = a.reshape((N//BM, nbIterWaveM, BM//(nbIterWaveM * TM), TM, 1, 1, 1, 1, N//BK, BK)).expand(full_shape)
+  b = b.reshape((1, 1, 1, 1, N//BN, nbIterWaveN, BN//(nbIterWaveN * TN), TN, N//BK, BK)).expand(full_shape)
+  c = c.reshape((N//BM, nbIterWaveM, BM//(nbIterWaveM * TM), TM, N//BN, nbIterWaveN, BN//(nbIterWaveN * TN), TN, 1, 1))
+  As = As.reshape((1, nbIterWaveM, BM//(nbIterWaveM * TM), TM, 1, 1, 1, 1, 1, BK)).expand(full_shape)
+  Bs = Bs.reshape((1, 1, 1, 1, 1, nbIterWaveN, BN//(nbIterWaveN * TN), TN, 1, BK)).expand(full_shape)
+  A_col = A_col.reshape((1, nbIterWaveM, 1, TM, 1, 1, 1, 1, 1, 1)).expand(full_shape)
+  B_row = B_row.reshape((1, 1, 1, 1, 1, nbIterWaveN, 1, TN, 1, 1)).expand(full_shape)

-  A_col = UOp.placeholder((ITERS_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-  B_row = UOp.placeholder((ITERS_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-  c_regs = UOp.placeholder((ITERS_PER_WAVE_M, TM, ITERS_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
+  #                     U1   L2 L3 L4 L5   U6 U7      U9   L10 L11 L12 L13   U14 U15      U17  U18  U19
+  expanded_shape = (32, 2,   2, 2, 2, 2,   2, 2,  32, 2,   2,  2,  2,  2,    2,  2,  512, 2,   2,   2)
+  assert len(expanded_shape) == 20
+  permute_a = list(range(len(expanded_shape)))
+  permute_b = permute_a[:]

-  i = UOp.range(c_regs.size, 16)
-  c_regs = c_regs[i].set(0.0, end=i)
+  # this makes all the global loads match
+  # this can also be more simply done by rebinding the RANGEs
+  # but sadly, rebinding the RANGEs doesn't work to change the order of the local axes
+  permute_a[17:20] = [11,12,13]
+  permute_a[11:14] = [17,18,19]
+  permute_a[7], permute_a[10] = permute_a[10], permute_a[7]
+  permute_a[2:7] = [3,4,5,6,2]

-  k_tile_range = UOp.range(N // BLOCK_K, 0)
+  permute_b[2:16] = [19,9,10,11,17,18,8,2,12,13,14,15,3,4]
+  permute_b[17:20] = [5,6,7]

-  # ---------------------------
-  # GLOBAL -> LOCAL (As, Bs)
-  # ---------------------------
-  b = b.reshape(N // BLOCK_K, BLOCK_K,
-                N // BLOCK_N, BLOCK_N)
-  i = UOp.range(BLOCK_N * BLOCK_K // THREADS_PER_BLOCK, 1)
-  index_x = tid % BLOCK_N
-  index_y = (tid // BLOCK_N) + (THREADS_PER_BLOCK // BLOCK_N) * i
-  Bs_store = Bs[index_y, index_x].store(b[k_tile_range, index_y, blockIdx_x, index_x]).end(i)
+  a_permute   = a.reshape(expanded_shape).permute(tuple(permute_a)).reshape(full_shape)
+  As_permute = As.reshape(expanded_shape).permute(tuple(permute_a)).reshape(full_shape)

-  a = a.reshape(N // BLOCK_M, BLOCK_M,
-                N // BLOCK_K, BLOCK_K)
-  i = UOp.range(BLOCK_M * BLOCK_K // THREADS_PER_BLOCK, 2)
-  index_x = tid % BLOCK_K
-  index_y = (tid // BLOCK_K) + (THREADS_PER_BLOCK // BLOCK_K) * i
-  As_store = As[index_x, index_y].store(a[blockIdx_y, index_y, k_tile_range, index_x]).end(i)
+  b_permute   = b.reshape(expanded_shape).permute(tuple(permute_b)).reshape(full_shape)
+  Bs_permute = Bs.reshape(expanded_shape).permute(tuple(permute_b)).reshape(full_shape)

-  # TODO: can we automate barrier?
-  barrier = UOp.barrier(As_store, Bs_store)
-  Bs = Bs.after(barrier)
-  As = As.after(barrier)
+  #out = (a.load() * b.load()).r(Ops.ADD, (8, 9))
+  out = (As.load(As_permute.store(a_permute.load())) * Bs.load(Bs_permute.store(b_permute.load()))).r(Ops.ADD, (8, 9))
+  #out = (A_col.load(A_col.store(As.load(As.store(a.load())))) * B_row.load(B_row.store(Bs.load(Bs.store(b.load()))))).r(Ops.ADD, (8, 9))

-  # open inner k range
-  k = UOp.range(BLOCK_K, 3)
+  axis_types = (
+    AxisType.GLOBAL, AxisType.UPCAST, AxisType.LOCAL, AxisType.UPCAST,
+    AxisType.GLOBAL, AxisType.UPCAST, AxisType.LOCAL, AxisType.UPCAST,
+    AxisType.REDUCE, AxisType.REDUCE)

-  # ---------------------------
-  # LOCAL -> REG (per-wave tiles)
-  # ---------------------------
-  Bs_view = Bs.reshape(BLOCK_K, WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
-  iterWaveN = UOp.range(ITERS_PER_WAVE_N, 4)
-  i = UOp.range(TN, 5)
-  B_row = B_row[iterWaveN, i].set(Bs_view[k, waveIdx, iterWaveN, idxInWave, i], end=(iterWaveN, i))
+  sink = c.store(out).sink(arg=KernelInfo(name="tg_"+to_colored(full_shape, axis_types), axis_types=axis_types))
+  sink = graph_rewrite(sink, merge_views)
+  return sink

-  As_view = As.reshape(BLOCK_K, WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM)
-  iterWaveM = UOp.range(ITERS_PER_WAVE_M, 6)
-  i = UOp.range(TM, 7)
-  A_col = A_col[iterWaveM, i].set(As_view[k, waveIdy, iterWaveM, idyInWave, i], end=(iterWaveM, i))
+def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
+  BLOCK_SIZE = 128 if kernel5 else 256

-  # ---------------------------
-  # FMA: c_regs += A_col * B_row
-  # ---------------------------
-  iterWaveM = UOp.range(ITERS_PER_WAVE_M, 8)
-  yt = UOp.range(TM, 9)
-  iterWaveN = UOp.range(ITERS_PER_WAVE_N, 10)
-  xt = UOp.range(TN, 12)
-  c_idx = c_regs.after(k, k_tile_range)[iterWaveM, yt, iterWaveN, xt]
-  sink = c_idx.store(c_idx + A_col[iterWaveM, yt] * B_row[iterWaveN, xt]).end(iterWaveM, iterWaveN, yt, xt)
+  nbWaves = BLOCK_SIZE // 32
+  WN = 128 if kernel5 else 64
+  WM = BN * BM // nbWaves // WN

-  # Close k, sync, and close K tiles
-  sink = sink.end(k).barrier().end(k_tile_range)
+  nbWaveX = BN // WN
+  nbWaveY = BM // WM

-  # ---------------------------
-  # REG -> GLOBAL (epilogue)
-  # ---------------------------
-  c = c.reshape(N//BLOCK_M, WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM,
-                N//BLOCK_N, WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
-  iterWaveM = UOp.range(ITERS_PER_WAVE_M, 1000)
+  threadIdx_x = UOp(Ops.SPECIAL, dtypes.int, arg=("lidx0", BLOCK_SIZE))
+  waveIndex = threadIdx_x // 32
+  waveIdx = waveIndex % nbWaveX
+  waveIdy = waveIndex // nbWaveX
+  indexInWave = threadIdx_x % 32
+
+  nbThreadXPerWave = 8
+  nbThreadYPerWave = 4
+
+  idxInWave = indexInWave % nbThreadXPerWave
+  idyInWave = indexInWave // nbThreadXPerWave
+
+  nbIterWaveN = WN // (nbThreadXPerWave * TN)
+  nbIterWaveM = WM // (nbThreadYPerWave * TM)
+
+  SUBWN = WN // nbIterWaveN
+  SUBWM = WM // nbIterWaveM
+
+  # Thread mapping to read BKxBN block from A
+  rAIdx = threadIdx_x % BK
+  rAIdy = threadIdx_x // BK
+  # Thread mapping to read BNxBK block from B
+  rBIdx = threadIdx_x % BN
+  rBIdy = threadIdx_x // BN
+
+  strideReadB = BLOCK_SIZE // BN
+  strideReadA = BLOCK_SIZE // BK
+  nbReadsB = BN * BK // BLOCK_SIZE
+  nbReadsA = BM * BK // BLOCK_SIZE
+
+  blockIdx_x = UOp(Ops.SPECIAL, dtypes.int, arg=("gidx0", N//BN))
+  blockIdx_y = UOp(Ops.SPECIAL, dtypes.int, arg=("gidx1", N//BM))
+
+  a = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=1)
+  b = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=2)
+  c = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(N*N), arg=0)
+
+  A_col = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbIterWaveM * TM, AddrSpace.REG), arg=0)
+  B_row = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbIterWaveN * TN, AddrSpace.REG), arg=1)
+
+  BM_As_stride = (BM+4) if kernel5 else BM
+  As = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(BK*BM_As_stride, AddrSpace.LOCAL), arg=0)
+  Bs = UOp(Ops.DEFINE_LOCAL, dtypes.float.ptr(BK*BN, AddrSpace.LOCAL), arg=1)
+
+  c_regs = UOp(Ops.DEFINE_REG, dtypes.float.ptr(TM * nbIterWaveM * TN * nbIterWaveN), arg=2)
+
+  i = UOp.range(c_regs.dtype.size, 16)
+  init_store = c_regs[i].store(UOp.const(dtypes.float, 0.0), i)
+
+  if kernel4:
+    regA = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbReadsA, AddrSpace.REG), arg=3)
+    regB = UOp(Ops.DEFINE_REG, dtypes.float.ptr(nbReadsB, AddrSpace.REG), arg=4)
+
+    # initial load from globals into locals (0)
+    kId = 0
+
+    # load from globals into locals
+    i = UOp.range(nbReadsB, 0)
+    index_x = BN * blockIdx_x + rBIdx
+    index_y = rBIdy + i * strideReadB + kId
+    Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)
+
+    i = UOp.range(nbReadsA, 1)
+    index_x = rAIdx + kId
+    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
+    As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)
+
+    # iterate over the middle chunk
+    kId_range = UOp.range(N//BK-1, 2)
+    kId = kId_range*BK
+
+    barrier = UOp.barrier(As_store, Bs_store)
+
+    # load from globals into registers (next round)
+    i = UOp.range(nbReadsB, 3)
+    index_x = BN * blockIdx_x + rBIdx
+    index_y = rBIdy + i * strideReadB + kId + BK
+    regB_store = regB[i].store(b[N * index_y + index_x].load(), i)
+
+    i = UOp.range(nbReadsA, 4)
+    index_x = rAIdx + kId + BK
+    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
+    regA_store = regA[i].store(a[N * index_y + index_x].load(), i)
+
+    def inner_loop(first_range, inp_dep=()):
+      # inner unroll
+      k = UOp.range(BK, first_range+0)
+
+      # load from locals into registers
+      iterWave = UOp.range(nbIterWaveN, first_range+1)
+      i = UOp.range(TN, first_range+2)
+      index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
+      B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(*inp_dep), iterWave, i)
+
+      iterWave = UOp.range(nbIterWaveM, first_range+3)
+      i = UOp.range(TM, first_range+4)
+      index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
+      A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(*inp_dep), iterWave, i)
+
+      # do the GEMM math
+      iterWaveM = UOp.range(nbIterWaveM, first_range+5)
+      yt = UOp.range(TM, first_range+6)
+      iterWaveN = UOp.range(nbIterWaveN, first_range+7)
+      xt = UOp.range(TN, first_range+8)
+      x = iterWaveN * TN + xt
+      y = iterWaveM * TM + yt
+      c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
+      # sketchy, this should end the kId_range but it doesn't
+      sink = c_regs_idx.store(c_regs_idx.load(init_store) + A_col[y].load(A_col_store) * B_row[x].load(B_row_store),
+                              iterWaveM, iterWaveN, yt, xt, k)
+      return sink
+
+    # TODO: kId_range should endrange after a barrier
+    sink = inner_loop(5, (barrier, regB_store, regA_store)).barrier()
+
+    # load from registers into locals
+    i = UOp.range(nbReadsB, 14)
+    index_x = BN * blockIdx_x + rBIdx
+    index_y = rBIdy + i * strideReadB + kId + BK
+    Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(regB[i].load(sink), i, kId_range)
+
+    i = UOp.range(nbReadsA, 15)
+    index_x = rAIdx + kId + BK
+    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
+    As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(regA[i].load(sink), i, kId_range)
+
+    # final iteration without the copy
+    sink = inner_loop(16, (UOp.barrier(Bs_store, As_store),))
+  else:
+    kId_range = UOp.range(N//BK, 0)
+    kId = kId_range*BK
+
+    # load from globals into locals
+    i = UOp.range(nbReadsB, 1)
+    index_x = BN * blockIdx_x + rBIdx
+    index_y = rBIdy + i * strideReadB + kId
+    Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)
+
+    i = UOp.range(nbReadsA, 2)
+    index_x = rAIdx + kId
+    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
+    As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)
+
+    barrier = UOp.barrier(As_store, Bs_store)
+
+    k = UOp.range(BK, 3)
+
+    # load from locals into registers
+    iterWave = UOp.range(nbIterWaveN, 4)
+    i = UOp.range(TN, 5)
+    index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
+    B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(barrier), iterWave, i)
+
+    iterWave = UOp.range(nbIterWaveM, 6)
+    i = UOp.range(TM, 7)
+    index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
+    A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(barrier), iterWave, i)
+
+    # do the GEMM math
+    iterWaveM = UOp.range(nbIterWaveM, 8)
+    yt = UOp.range(TM, 9)
+    iterWaveN = UOp.range(nbIterWaveN, 10)
+    xt = UOp.range(TN, 12)
+    x = iterWaveN * TN + xt
+    y = iterWaveM * TM + yt
+    c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
+    sink = c_regs_idx.store(c_regs_idx.load(init_store) + A_col[y].load(A_col_store) * B_row[x].load(B_row_store),
+                            iterWaveM, iterWaveN, yt, xt, k, kId_range)
+
+  # store c_regs into c
+  iterWaveM = UOp.range(nbIterWaveM, 1000)
  yt = UOp.range(TM, 1001)
-  iterWaveN = UOp.range(ITERS_PER_WAVE_N, 1002)
+  iterWaveN = UOp.range(nbIterWaveN, 1002)
  xt = UOp.range(TN, 1003)
-  c_glbl_idx = c[blockIdx_y, waveIdy, iterWaveM, idyInWave, yt, blockIdx_x, waveIdx, iterWaveN, idxInWave, xt]
-  sink = c_glbl_idx.store(c_regs.after(sink)[iterWaveM, yt, iterWaveN, xt])
-  sink = sink.end(iterWaveM, iterWaveN, yt, xt)
+  xOut = blockIdx_x * BN + waveIdx * WN + iterWaveN * SUBWN + TN * idxInWave
+  yOut = blockIdx_y * BM + waveIdy * WM + iterWaveM * SUBWM + TM * idyInWave
+  indexC = N * (yOut + yt) + xOut + xt
+  sink = c[indexC].store(c_regs[TN * nbIterWaveN * (iterWaveM * TM + yt) + (iterWaveN * TN + xt)].load(sink),
+                         iterWaveM, iterWaveN, yt, xt)

-  return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify()
-
-def test_matmul(sink:UOp, N=N):
-  with Context(DEBUG=0):
-    a = Tensor.randn(N, N)
-    b = Tensor.randn(N, N)
-    hc = Tensor.empty(N, N)
-    Tensor.realize(a, b, hc)
-
-  ei = ExecItem(get_runner(Device.DEFAULT, sink), [t.uop.buffer for t in [hc, a, b]])
-
-  GlobalCounters.reset()
-  ets = []
-  with Context(DEBUG=2):
-    for _ in range(run_count):
-      ets.append(ei.run(wait=True))
-  print(f"REAL TFLOPS {N * N * N * 2 / min(ets) * 1e-12:.2f}")
-
-  GlobalCounters.reset()
-  with Context(DEBUG=2):
-    tc = (a @ b).realize()
-  with Context(DEBUG=0):
-    err = (hc - tc).square().mean().item()
-  print(f"mean squared error {err}")
-  if err > 1e-06:
-    raise RuntimeError("matmul is wrong!")
+  return sink.sink(arg=KernelInfo(name="tinygemm"))

 if __name__ == "__main__":
-  test_matmul(hand_spec_kernel3(), N=N)
+  HL = getenv("HL")
+  if HL == 3: hprg = rangeify_kernel3()
+  elif HL == 2: hprg = top_spec_kernel3()
+  elif HL == 1: hprg = hl_spec_kernel3()
+  else: hprg = hand_spec_kernel3()
+  if HL == 3:
+    with Context(RANGEIFY=1, BLOCK_REORDER=0):
+      prg = get_program(hprg, Device.default.renderer)
+  else:
+    prg = get_program(hprg, Device.default.renderer)
+  print(prg.src)
+  if getenv("SRC"): exit(0)
+  hrunner = CompiledRunner(prg)
+
+  a = Tensor.randn(N, N).realize()
+  b = Tensor.randn(N, N).realize()
+  hc = Tensor.zeros(N, N).contiguous().realize()
+
+  GlobalCounters.reset()
+  with Context(DEBUG=2):
+    for _ in range(run_count): tc = (a@b).realize()
+
+  GlobalCounters.reset()
+  buffers = [hc.uop.buffer, a.uop.buffer, b.uop.buffer]
+  ei = ExecItem(hrunner, buffers)
+  with Context(DEBUG=2):
+    for _ in range(run_count): ei.run(wait=True)
+  err = (hc-tc).square().mean().item()
+  print(f"hrunner {err}")
+  if err > 1e-06: raise RuntimeError("matmul is wrong!")
@@ -1,42 +0,0 @@
-from tinygrad import UOp, dtypes
-from tinygrad.uop.ops import AxisType, Ops, KernelInfo, AddrSpace
-from extra.gemm.amd_uop_matmul import test_matmul
-
-N = 2048
-
-# metal has an 8x8 tensor core. this is the indexing
-def mat_idx(buf, g0, g1, warp, u):
-  l = [(warp//2**i)%2 for i in range(5)]
-  return buf[g0, l[4]*4 + l[2]*2 + l[1], g1, l[3]*4 + l[0]*2 + u]
-
-def hand_spec_tc_cores():
-  gx = UOp.special(N // 8, "gidx0")
-  gy = UOp.special(N // 8, "gidx1")
-  warp = UOp.special(32, "lidx0")
-
-  c = UOp.placeholder((N, N), dtypes.float, slot=0).reshape((N//8, 8, N//8, 8))
-  a = UOp.placeholder((N, N), dtypes.float, slot=1).reshape((N//8, 8, N//8, 8))
-  b = UOp.placeholder((N, N), dtypes.float, slot=2).reshape((N//8, 8, N//8, 8))
-
-  gk = UOp.range(N // 8, 0, AxisType.REDUCE)
-
-  a_tc = UOp.vectorize(*[mat_idx(a, gx, gk, warp, i) for i in range(2)])
-  b_tc = UOp.vectorize(*[mat_idx(b, gk, gy, warp, i) for i in range(2)])
-
-  acc = UOp.placeholder((2,), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-  acc = acc[0].set(0.0)
-  acc = acc[1].set(0.0)
-
-  # TODO: make this simple
-  wmma_arg = ('WMMA_8_8_8_float_float', (8, 8, 8), dtypes.float, dtypes.float, 'METAL', 32, (((3, 2),), ((3, 2),), ((3, 2),)), ())
-
-  acc_load = UOp.vectorize(acc.after(gk)[0], acc.after(gk)[1])
-  out = UOp(Ops.WMMA, dtypes.float.vec(2), (a_tc, b_tc, acc_load), arg=wmma_arg)
-
-  end_loop = UOp.group(*[acc[i].store(out.gep(i)) for i in range(2)]).end(gk)
-
-  sink = UOp.group(*[mat_idx(c.after(end_loop), gx, gy, warp, i).store(acc[i]) for i in range(2)])
-  return sink.sink(arg=KernelInfo(name="custom_metal_matmul", opts_to_apply=())).simplify()
-
-if __name__ == "__main__":
-  test_matmul(hand_spec_tc_cores(), N=N)
@@ -1,229 +0,0 @@
-import os
-import numpy as np
-np.set_printoptions(linewidth=1000000)
-os.environ["AMD_LLVM"] = "0"
-
-from tinygrad import Tensor, Context, dtypes, UOp, GlobalCounters
-from tinygrad.helpers import DEBUG, getenv
-from tinygrad.dtype import AddrSpace
-from tinygrad.uop.ops import AxisType, KernelInfo, Ops
-
-WARP_SIZE = 64
-
-# Reg tile sizes (tensor cores)
-TC_M = 16
-TC_N = 16
-TC_K = 32
-
-# 1024 matrix cores
-# 16 cycle mfma
-# 2.2 GHz
-# 16x16x32x2 FLOPS/mma = 16384
-# 2.2*1e9*16384*1024/16*1e-12 TFLOPS = 2306 TFLOPS
-
-#N,M,K = 256,256,64
-N,M,K = 4096,4096,4096
-
-# Threadblock tile sizes (block-level tile of C that a block computes)
-#BLOCK_M = 128   # rows of C (M-dim) per block
-#BLOCK_N = 128   # columns of C (N-dim) per block
-#BLOCK_K = 128   # K-slice per block iteration
-
-BLOCK_M = 64
-BLOCK_N = 64
-BLOCK_K = 128
-
-WARPGROUP_SIZE = 1
-BLOCK_M = BLOCK_M * WARPGROUP_SIZE
-
-# TODO: improve the syntax of this. better syntax, faster iteration
-#  -- add working slice a[gx, :, i] -> shape of the : (aka (16,16,32) becomes (16,))
-#  -- add argfix to movement (traits shared with Tensor)
-#  -- fix WMMA to not require all the junk
-#  -- improve syntax for vectorized loads/stores (both with DEVECTORIZE and without)
-#  -- be able to use CONTRACT on a range
-#  -- fix upcasted RANGE on an already vectorized buffer
-#  -- improve "all ranges not ended error" / fix the bug with after on ended ranges (if you are after end of range, range is closed)
-
-CUS_PER_GPU = 256
-assert ((M//BLOCK_M) * (N//BLOCK_N)) >= CUS_PER_GPU, "not enough globals"
-
-def custom_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
-  # A = (M x K)
-  # B = (K x N)
-  # C = (M x N)
-
-  # check it's proper matmul
-  assert C.shape[0] == A.shape[0]
-  assert C.shape[1] == B.shape[1]
-  assert A.shape[1] == B.shape[0]
-
-  gx, gy = UOp.special(M//BLOCK_M, "gidx0"), UOp.special(N//BLOCK_N, "gidx1")
-  warp = UOp.special(WARP_SIZE, "lidx0")
-  warpgroup = UOp.special(WARPGROUP_SIZE, "lidx1")
-
-  # generic copy logic (not good)
-  def generic_copy(glbl, gargs, lcl, rng):
-    # Fully coalesced 128-bit loads/stores.
-    INNER_SIZE = 8
-    cp_i = UOp.range(lcl.size//(WARPGROUP_SIZE*WARP_SIZE*INNER_SIZE), rng)
-    cp_inner = UOp.range(INNER_SIZE, rng+1, AxisType.UPCAST)
-    idx_i = cp_i*WARPGROUP_SIZE*WARP_SIZE*INNER_SIZE + warpgroup*WARP_SIZE*INNER_SIZE + warp*INNER_SIZE + cp_inner
-    return lcl[idx_i].store(glbl[*gargs, idx_i]).end(cp_i, cp_inner)
-
-  # split out the globals into blocks
-  C = C.reshape((M//BLOCK_M, BLOCK_M, N//BLOCK_N, BLOCK_N))
-  A = A.reshape((M//BLOCK_M, BLOCK_M, K//BLOCK_K, BLOCK_K))
-  B = B.reshape((K//BLOCK_K, BLOCK_K, N//BLOCK_N, BLOCK_N))
-
-  # this is the big accumulator
-  acc = UOp.placeholder((BLOCK_N//TC_N, BLOCK_M//TC_M//WARPGROUP_SIZE), dtypes.float.vec(4), 0, AddrSpace.REG)
-  assert acc.size*WARP_SIZE*WARPGROUP_SIZE*4 == BLOCK_M*BLOCK_N
-  acc = acc[init_l:=UOp.range(acc.size, 500)].set(UOp.const(dtypes.float.vec(4), 0.0), end=init_l)
-
-  # create locals (note A is permuted, and the stride is changed to avoid bank conflicts)
-  def make_locals(slot) -> tuple[UOp, UOp]:
-    BM_As_stride = (BLOCK_M + 1)
-    BN_Bs_stride = (BLOCK_N + 0)
-    INNER_SLICE = 8
-    As = UOp.placeholder((BLOCK_K//INNER_SLICE, BM_As_stride, INNER_SLICE), dtypes.half, slot=slot, addrspace=AddrSpace.LOCAL)
-    INNER_SLICE = 1
-    Bs = UOp.placeholder((BLOCK_K//INNER_SLICE, BN_Bs_stride, INNER_SLICE), dtypes.half, slot=slot+1, addrspace=AddrSpace.LOCAL)
-    As = As.permute((0,2,1)).reshape((BLOCK_K, BM_As_stride)).shrink_to((BLOCK_K, BLOCK_M))
-    Bs = Bs.permute((0,2,1)).reshape((BLOCK_K, BN_Bs_stride)).shrink_to((BLOCK_K, BLOCK_N))
-    return As, Bs
-
-  # load from globals into locals (TODO: use the warpgroup)
-
-  def load_to_locals(l_K_outer_loop:UOp, Asl:UOp, Bsl:UOp, rng:int, barrier=True) -> tuple[UOp, UOp]:
-    if getenv("FAKE"):
-      return Asl[0].set(0), Bsl[0].set(0)
-    else:
-      pA = A.permute((0,2,1,3)).reshape((M//BLOCK_M, K//BLOCK_K, BLOCK_M*BLOCK_K))
-      pas = Asl.permute((1,0)).reshape((BLOCK_M*BLOCK_K,))
-      As_store = generic_copy(pA, (gx, l_K_outer_loop), pas, rng)
-
-      pB = B.permute((0,2,1,3)).reshape((K//BLOCK_K, N//BLOCK_N, BLOCK_K*BLOCK_N))
-      pbs = Bsl.reshape((BLOCK_K*BLOCK_N,))
-      Bs_store = generic_copy(pB, (l_K_outer_loop, gy), pbs, rng+2)
-
-      barrier = UOp.barrier(As_store, Bs_store) if barrier else UOp.group(As_store, Bs_store)
-      return Asl.after(barrier), Bsl.after(barrier)
-
-  def compute_on_locals(acc:UOp, Asl:UOp, Bsl:UOp, rng:int, afters:tuple[UOp, ...]=()) -> UOp:
-    K_inner_loop = UOp.range(BLOCK_K//TC_K, rng, AxisType.REDUCE)
-
-    # load from locals into registers
-    Ar = UOp.placeholder((BLOCK_M//TC_M//WARPGROUP_SIZE,), dtypes.half.vec(8), slot=1, addrspace=AddrSpace.REG)
-    Br = UOp.placeholder((BLOCK_N//TC_N,), dtypes.half.vec(8), slot=2, addrspace=AddrSpace.REG)
-
-    M_load_loop = UOp.range(BLOCK_M//TC_M//WARPGROUP_SIZE, rng+10)
-    Asl = Asl.reshape((BLOCK_K//TC_K, TC_K, BLOCK_M//TC_M//WARPGROUP_SIZE, WARPGROUP_SIZE, TC_M))
-    load_rng = UOp.range(8, rng+11, axis_type=AxisType.UPCAST)
-    A_in = Asl[K_inner_loop, (warp//16)*8+load_rng, M_load_loop, warpgroup, warp%16].contract(load_rng)
-    Ar = Ar[M_load_loop].set(A_in, end=M_load_loop)
-
-    N_load_loop = UOp.range(BLOCK_N//TC_N, rng+20)
-    Bsl = Bsl.reshape((BLOCK_K//TC_K, TC_K, BLOCK_N//TC_N, TC_N))
-    load_rng = UOp.range(8, rng+21, axis_type=AxisType.UPCAST)
-    B_in = Bsl[K_inner_loop, (warp//16)*8+load_rng, N_load_loop, warp%16].contract(load_rng)
-    Br = Br[N_load_loop].set(B_in, end=N_load_loop)
-
-    M_inner_loop = UOp.range(BLOCK_M//TC_M//WARPGROUP_SIZE, rng+30)
-    N_inner_loop = UOp.range(BLOCK_N//TC_N, rng+31)
-
-    # load values
-    acc_after = acc.after(*afters, M_inner_loop, N_inner_loop, K_inner_loop)
-    acc_load = acc_after[N_inner_loop, M_inner_loop]
-
-    # do WMMA
-    wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float, 'AMD', 64, ((), (), ((3, 2), (2, 2))), ())
-    out = UOp(Ops.WMMA, dtypes.float.vec(4), (Ar[M_inner_loop], Br[N_inner_loop], acc_load), arg=wmma_arg)
-
-    # store back the acc
-    acc_store = acc[N_inner_loop, M_inner_loop].store(out)
-    return acc_store.end(M_inner_loop, N_inner_loop, K_inner_loop)
-
-  # **** START INNER LOOP *****
-  # inner loop -- locals -> regs
-
-  # no pipeline
-  if not getenv("PIPELINE"):
-    As, Bs = make_locals(slot=0)
-
-    K_outer_loop = UOp.range(K//BLOCK_K, 0, AxisType.REDUCE)
-    As, Bs = load_to_locals(K_outer_loop, As, Bs, 1000, barrier=True)
-    acc_store = compute_on_locals(acc, As, Bs, 1500, afters=(K_outer_loop,))
-    acc = acc.after(acc_store.barrier().end(K_outer_loop))
-  else:
-    # this doesn't work
-    As0, Bs0 = make_locals(slot=0)
-    As1, Bs1 = make_locals(slot=2)
-    As0, Bs0 = load_to_locals(0, As0, Bs0, 1000)
-
-    K_outer_loop = UOp.range((K//BLOCK_K-2)//2, 0, AxisType.REDUCE)
-    As1, Bs1 = load_to_locals(K_outer_loop+1, As1, Bs1, 2000, barrier=False)
-    acc_store = compute_on_locals(acc, As0, Bs0, 1500, afters=(K_outer_loop,))
-    As0, Bs0 = load_to_locals(K_outer_loop+2, As0, Bs0, 3000, barrier=False)
-    acc_store = compute_on_locals(acc, As1, Bs1, 2500, afters=(acc_store, As0, Bs0))
-    acc = acc.after(acc_store.barrier().end(K_outer_loop))
-
-    #acc_store = compute_on_locals(acc, As0, Bs0, 3500, afters=(acc_store.barrier().end(K_outer_loop)))
-    """
-    As1, Bs1 = load_to_locals(K//BLOCK_K-1, As1, Bs1, 4000)
-    acc_store = compute_on_locals(acc, As1, Bs1, 4500, afters=(acc_store))
-    """
-    #acc = acc.after(acc_store)
-
-  # **** END LOOPS *****
-
-  # store the acc into gmem
-  cp_i, cp_j = UOp.range(BLOCK_M//TC_M//WARPGROUP_SIZE, 10004), UOp.range(BLOCK_N//TC_N, 10005)
-  c_load = lambda i: C[gx, cp_i*TC_M*WARPGROUP_SIZE + warpgroup*TC_M + (warp//16)*4+i, gy, cp_j*TC_N + warp%16]
-  store = UOp.group(*[c_load(i).store(acc[cp_j, cp_i].gep(i)) for i in range(4)])
-  store = store.end(cp_i, cp_j)
-
-  return store.sink(arg=KernelInfo(name="custom_gemm", opts_to_apply=())).simplify()
-
-# simplest WMMA
-"""
-# init the acc
-acc = UOp.placeholder((4,), dtypes.float, 0, AddrSpace.REG)
-acc = acc[init_l:=UOp.range(4, 1)].set(0.0, end=init_l)
-
-# do the wmma
-acc_load = UOp.vectorize(*[acc.after(K_loop)[i] for i in range(4)])
-wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float, 'AMD', 64, ((), (), ((3, 2), (2, 2))), ())
-out = UOp(Ops.WMMA, dtypes.float.vec(4), (A_in, B_in, acc_load), arg=wmma_arg)
-
-# store back the acc
-acc = acc.after(UOp.group(*[acc[i].store(out.gep(i)) for i in range(4)]).end(K_loop))
-
-# store the acc into gmem
-store = UOp.group(*[C[gx, (warp//16)*4+i, gy, warp%16].store(acc[i]) for i in range(4)])
-"""
-
-if __name__ == "__main__":
-  a = Tensor.randn(M, K, dtype=dtypes.half)
-  b = Tensor.randn(K, N, dtype=dtypes.half)
-
-  #a = Tensor.zeros(M, K, dtype=dtypes.half).contiguous()
-  #a[0,16] = 1
-  #b = Tensor.ones(K, N, dtype=dtypes.half).contiguous()
-
-  c = Tensor.empty(M, N, dtype=dtypes.float)
-  with Context(DEBUG=0): Tensor.realize(a,b)
-
-  ref = a.dot(b, dtype=dtypes.float)
-  ref.realize()
-
-  GlobalCounters.reset()
-  with Context(DEBUG=max(2, DEBUG.value), DEVECTORIZE=2):
-    tst = Tensor.custom_kernel(c, a, b, fxn=custom_gemm)[0]
-    tst.realize()
-  print(f"{(N*M*K*2 / GlobalCounters.time_sum_s)*1e-12:.2f} REAL TFLOPS")
-
-  with Context(DEBUG=0):
-    #print(ref.numpy())
-    #print(tst.numpy())
-    assert Tensor.isclose(ref, tst, atol=1e-2).all().item(), "matrix not close"
@@ -5,10 +5,8 @@ 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.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)
+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
 if getenv("INT"):  dtype_in, acc_dtype = dtypes.int8, dtypes.int32
 if getenv("UINT"): dtype_in, acc_dtype = dtypes.uint8, dtypes.int32

@@ -16,10 +14,8 @@ N = getenv("N", 4096)
 M = getenv("M", N)
 K = getenv("K", N)
 CNT = getenv("CNT", 10)
-
-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)
-
+ATOL = getenv("ATOL", 1e-4)
+RTOL = getenv("RTOL", 3e-2)
 INT_LOW = getenv("INT_LOW", 0)
 INT_HIGH = getenv("INT_HIGH", 10)

@@ -8,23 +8,19 @@ import torch
 torch.set_num_threads(1)
 from tinygrad.helpers import getenv
 CUDA = getenv("CUDA", 1)
-MPS = getenv("MPS", 0)
-if getenv("FP16_ACC"): torch.backends.cuda.matmul.allow_fp16_accumulation = True

-for dtype in [torch.float32, torch.float16, torch.bfloat16]:
+for dtype in [torch.float32, torch.float16]:
  for N in [256, 512, 1024, 2048, 4096]:
    FLOPS = N*N*N*2

    b = torch.rand((N,N), dtype=dtype)
    c = torch.rand((N,N), dtype=dtype)
    if CUDA: b,c = b.cuda(),c.cuda()
-    if MPS: b,c = b.to('mps'),c.to('mps')

    def torch_prog(b, c):
      st = time.perf_counter()
      a = b@c
      if CUDA: torch.cuda.synchronize()
-      if MPS: torch.mps.synchronize()
      return time.perf_counter() - st
    tm = min([torch_prog(b, c) for _ in range(20)])
    print(f"{N*N:10d} {tm*1e6:9.2f} us, would be {FLOPS*1e-9/tm:9.2f} GFLOPS {N:4d}x{N:4d}x{N:4d} matmul in {dtype}")
@@ -1,6 +1,7 @@
 #!/usr/bin/env python3

-import argparse, glob, os, time, subprocess, sys
+from tinygrad.runtime.support.system import System
+import argparse, glob, os, re, time, subprocess, sys

 def scan_devs_based_on_lock(prefix:str, args) -> list[str]:
  target_dev = args.pci_bus if 'pci_bus' in args.__dir__() else ""
@@ -11,7 +12,7 @@ def scan_devs_based_on_lock(prefix:str, args) -> list[str]:
    if os.path.exists(f"/sys/bus/pci/devices/{dev_id}") and dev_id.startswith(target_dev): devs.append(dev_id)
  return devs

-def _do_reset_device(pci_bus): os.system(f"sudo sh -c 'echo 1 > /sys/bus/pci/devices/{pci_bus}/reset'")
+def _do_reset_device(pci_bus): System.pci_reset(pci_bus)
 def _is_module_loaded(name: str) -> bool: return os.path.isdir(f"/sys/module/{name}")

 def cmd_remove_module(args):
@@ -7,6 +7,7 @@ bert_train_params = {
  "GPUS": 6,
  "BS": 96,
  "EVAL_BS": 96,
+  "FUSE_ARANGE": 1,
  "BASEDIR": "/raid/datasets/wiki",
 }

@@ -50,7 +50,7 @@ def ioctls_from_header():
  hdr = (pathlib.Path(__file__).parent / "kfd_ioctl.h").read_text().replace("\\\n", "")
  pattern = r'#define\s+(AMDKFD_IOC_[A-Z0-9_]+)\s+AMDKFD_IOW?R?\((0x[0-9a-fA-F]+),\s+struct\s([A-Za-z0-9_]+)\)'
  matches = re.findall(pattern, hdr, re.MULTILINE)
-  return {int(nr, 0x10):(name, getattr(kfd_ioctl, "struct_"+sname, None)) for name, nr, sname in matches}
+  return {int(nr, 0x10):(name, getattr(kfd_ioctl, "struct_"+sname)) for name, nr, sname in matches}
 nrs = ioctls_from_header()

@ctypes.CFUNCTYPE(ctypes.c_int, ctypes.c_int, ctypes.c_ulong, ctypes.c_void_p)
@@ -1,7 +1,7 @@
 import onnx, yaml, tempfile, time, argparse, json
 from pathlib import Path
 from typing import Any
-from tinygrad.nn.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner
 from extra.onnx_helpers import validate, get_example_inputs
 from extra.huggingface_onnx.huggingface_manager import DOWNLOADS_DIR, snapshot_download_with_retry

@@ -1,23 +0,0 @@
-#!/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
-
@@ -7,34 +7,31 @@ import os
 NUM_WORKGROUPS = 96
 WAVE_SIZE = 32
 NUM_WAVES = 2
-FLOPS_PER_MATMUL = 16*16*16*2
-INTERNAL_LOOP = 1_000_00
-INSTRUCTIONS_PER_LOOP = 200
-DIRECTIVE = ".amdhsa_wavefront_size32 1"
+FLOPS_PER_MATMUL =  16*16*16*2
+INTERNAL_LOOP  = 1_000_000
+INSTRUCTIONS_PER_LOOP = 1_000

 assemblyTemplate = (pathlib.Path(__file__).parent / "template.s").read_text()

-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:
+def launchBenchmark(instruction, vgprIndices, dense = True):
+  if dense:
    instructions = "{} v[0:{}], v[{}:{}], v[{}:{}], 1\n".format(instruction, vgprIndices[0],
                                                                vgprIndices[1], vgprIndices[2],
-                                                                vgprIndices[1], vgprIndices[2])
+                                                                vgprIndices[1], vgprIndices[2]) * INSTRUCTIONS_PER_LOOP
  else:
    instructions = "{} v[0:{}], v[{}:{}], v[{}:{}], v{}\n".format(instruction, vgprIndices[0],
-                                                                  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)
+                                                                vgprIndices[1], vgprIndices[2],
+                                                                vgprIndices[3], vgprIndices[4],
+                                                                vgprIndices[5]) * INSTRUCTIONS_PER_LOOP
+  src = assemblyTemplate.replace("INSTRUCTION", instructions)
  lib = COMPILER.compile(src)
  fxn = AMDProgram(DEV, "matmul", lib)
-  elapsed = fxn(global_size=(NUM_WORKGROUPS,1,1), local_size=(WAVE_SIZE*NUM_WAVES,1,1), wait=True)
+  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
  FLOPs = FLOPS_PER_MATMUL * NUM_WAVES * NUM_WORKGROUPS * INTERNAL_LOOP * INSTRUCTIONS_PER_LOOP
-  print(f"{instruction:<29} : {FLOPs/elapsed/10**12:.2f} T(FL)OPS")
+  print("{:<29} : {} T(FL)OPS".format(instruction, round(FLOPs/elapsed/10**12, 2)))

 if __name__=="__main__":
  DEVICENUM = os.getenv("DEVICENUM", "0")
@@ -43,17 +40,18 @@ if __name__=="__main__":
  except:
    raise RuntimeError("Error while initiating AMD device")

-  COMPILER = HIPCompiler(DEV.arch)
-  if DEV.arch in {'gfx1100', 'gfx1103'}:
-    if DEV.arch == 'gfx1103':
-      NUM_WORKGROUPS = 8
+  if (ARCH := DEV.arch) not in ['gfx1100', 'gfx1201']:
+    raise RuntimeError("only gfx1100 and gfx1201 supported")
+  COMPILER = HIPCompiler(ARCH)
+
+  if ARCH == 'gfx1100':
    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))
-  elif DEV.arch == 'gfx1201':
+  if ARCH == 'gfx1201':
    NUM_WORKGROUPS = 64
    launchBenchmark("v_wmma_bf16_16x16x16_bf16", (3,4,7))
    launchBenchmark("v_wmma_f16_16x16x16_f16", (3,4,7))
@@ -78,20 +76,4 @@ 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)
-  elif DEV.arch == 'gfx950':
-    DIRECTIVE = ".amdhsa_accum_offset 4"
-    NUM_WORKGROUPS = 256
-    WAVE_SIZE = 64
-    NUM_WAVES = 4
-    launchBenchmark("v_mfma_f32_16x16x16_f16", (3,0,1), accum=True)
-    launchBenchmark("v_mfma_f32_16x16x16_bf16", (3,0,1), accum=True)
-    FLOPS_PER_MATMUL = 16*16*32*2
-    launchBenchmark("v_mfma_f32_16x16x32_f16", (3,0,3), accum=True)
-    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.")
+    launchBenchmark("v_swmmac_i32_16x16x64_iu4", (7,8,9,10,13,14), False)
@@ -1,9 +1,9 @@
    .text
    .globl matmul
-    .p2align 8
+    .p2align 8 
    .type matmul,@function
 matmul:
-    s_mov_b32 s1, INTERNAL_LOOP
+    s_mov_b32 s1, 1000000
    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
-  DIRECTIVE
+  .amdhsa_wavefront_size32 1
 .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
@@ -36,5 +36,6 @@ amdhsa.kernels:
    .sgpr_count: 8
    .vgpr_count: 32
    .max_flat_workgroup_size: 1024
+    .args:
 ...
 .end_amdgpu_metadata
@@ -9,9 +9,6 @@ from PIL import Image
 import numpy as np
 import re, gzip

-# Allow for monkeypatching for mlperf.
-gelu = Tensor.gelu
-
@lru_cache()
 def default_bpe():
  # Clip tokenizer, taken from https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py (MIT license)
@@ -56,8 +53,8 @@ class Tokenizer:
    cs = [chr(n) for n in cs]
    return dict(zip(bs, cs))
  class ClipTokenizer:
-    def __init__(self, version=None):
-      self.byte_encoder, self.version = Tokenizer.bytes_to_unicode(), version
+    def __init__(self):
+      self.byte_encoder = Tokenizer.bytes_to_unicode()
      merges = gzip.open(default_bpe()).read().decode("utf-8").split('\n')
      merges = merges[1:49152-256-2+1]
      merges = [tuple(merge.split()) for merge in merges]
@@ -65,17 +62,11 @@ class Tokenizer:
      vocab = vocab + [v+'</w>' for v in vocab]
      for merge in merges:
        vocab.append(''.join(merge))
-      if self.version == "sd_mlperf_v5_0":
-        import regex
-        vocab.extend(['<start_of_text>', '<end_of_text>'])
-        self.cache = {'<start_of_text>': '<start_of_text>', '<end_of_text>': '<end_of_text>'}
-        self.pat = regex.compile(r"""<start_of_text>|<end_of_text>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", regex.IGNORECASE)
-      else:
-        vocab.extend(['<|startoftext|>', '<|endoftext|>'])
-        self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
-        self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[^\s]+""", re.IGNORECASE)
+      vocab.extend(['<|startoftext|>', '<|endoftext|>'])
      self.encoder = dict(zip(vocab, range(len(vocab))))
      self.bpe_ranks = dict(zip(merges, range(len(merges))))
+      self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
+      self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[^\s]+""", re.IGNORECASE)

    def bpe(self, token):
      if token in self.cache:
@@ -119,17 +110,8 @@ class Tokenizer:

    def encode(self, text:str, pad_with_zeros:bool=False) -> List[int]:
      bpe_tokens: List[int] = []
-      if self.version == "sd_mlperf_v5_0":
-        import regex, ftfy, html
-        text = ftfy.fix_text(text)
-        text = html.unescape(html.unescape(text)).strip()
-        text = Tokenizer.whitespace_clean(text).lower()
-        re_module = regex
-      else:
-        text = Tokenizer.whitespace_clean(text.strip()).lower()
-        re_module = re
-
-      for token in re_module.findall(self.pat, text):
+      text = Tokenizer.whitespace_clean(text.strip()).lower()
+      for token in re.findall(self.pat, text):
        token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
        bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
      # Truncation, keeping two slots for start and end tokens.
@@ -270,8 +252,10 @@ class Open:
      q,k,v = [y.reshape(T, B*self.n_heads, self.d_head).transpose(0, 1).reshape(B, self.n_heads, T, self.d_head) for y in proj.chunk(3)]

      attn_output = Tensor.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
-      attn_output = attn_output.permute(2, 0, 1, 3).reshape(T, B, C)
+      attn_output = attn_output.permute(2, 0, 1, 3).reshape(T*B, C)
+
      attn_output = self.out_proj(attn_output)
+      attn_output = attn_output.reshape(T, B, C)

      return attn_output

@@ -279,10 +263,9 @@ class Open:
    def __init__(self, dims, hidden_dims):
      self.c_fc   = Linear(dims, hidden_dims)
      self.c_proj = Linear(hidden_dims, dims)
-      self.gelu = gelu

    def __call__(self, x:Tensor) -> Tensor:
-      return x.sequential([self.c_fc, self.gelu, self.c_proj])
+      return x.sequential([self.c_fc, Tensor.gelu, self.c_proj])

  # https://github.com/mlfoundations/open_clip/blob/58e4e39aaabc6040839b0d2a7e8bf20979e4558a/src/open_clip/transformer.py#L210
  class ResidualAttentionBlock:
@@ -367,15 +350,15 @@ class Open:
 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/encoders/modules.py#L396
 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/encoders/modules.py#L498
 class FrozenOpenClipEmbedder(Embedder):
-  def __init__(self, dims:int, n_heads:int, layers:int, return_pooled:bool, ln_penultimate:bool=False, clip_tokenizer_version=None):
-    self.tokenizer = Tokenizer.ClipTokenizer(version=clip_tokenizer_version)
+  def __init__(self, dims:int, n_heads:int, layers:int, return_pooled:bool, ln_penultimate:bool=False):
+    self.tokenizer = Tokenizer.ClipTokenizer()
    self.model = Open.ClipTextTransformer(dims, n_heads, layers)
    self.return_pooled = return_pooled
    self.input_key = "txt"
    self.ln_penultimate = ln_penultimate

  def tokenize(self, text:str, device:Optional[str]=None) -> Tensor:
-    return Tensor(self.tokenizer.encode(text, pad_with_zeros=True), dtype=dtypes.int32, device=device).reshape(1,-1)
+    return Tensor(self.tokenizer.encode(text, pad_with_zeros=True), dtype=dtypes.int64, device=device).reshape(1,-1)

  def text_transformer_forward(self, x:Tensor, attn_mask:Optional[Tensor]=None):
    for r in self.model.transformer.resblocks:
@@ -466,7 +449,7 @@ class OpenClipEncoder:
    x = x + self.positional_embedding
    x = self.transformer(x, attn_mask=self.attn_mask)
    x = self.ln_final(x)
-    x = x[Tensor.arange(x.shape[0], device=x.device), tokens.argmax(axis=-1)]
+    x = x[:, tokens.argmax(axis=-1)]
    x = x @ self.text_projection
    return x

@@ -270,10 +270,8 @@ class FidInceptionV3:
    self.Mixed_7b = inception.Mixed_7b
    self.Mixed_7c = inception.Mixed_7c

-  def load_from_pretrained(self, path=None):
-    if path is None:
-      path = fetch("https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth", "pt_inception-2015-12-05-6726825d.pth")
-    state_dict = torch_load(str(path))
+  def load_from_pretrained(self):
+    state_dict = torch_load(str(fetch("https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth", "pt_inception-2015-12-05-6726825d.pth")))
    for k,v in state_dict.items():
      if k.endswith(".num_batches_tracked"):
        state_dict[k] = v.reshape(1)
@@ -1,24 +1,21 @@
-from tinygrad import Tensor, dtypes, nn
+from tinygrad import Tensor, dtypes
+from tinygrad.nn import Linear, Conv2d, GroupNorm, LayerNorm
 from tinygrad.device import is_dtype_supported
-from typing import Optional, Union, List, Any, Tuple, Callable
+from typing import Optional, Union, List, Any, Tuple
 import math

-# allow for monkeypatching
-Linear, Conv2d, GroupNorm, LayerNorm = nn.Linear, nn.Conv2d, nn.GroupNorm, nn.LayerNorm
-attention, gelu, mixed_precision_dtype = Tensor.scaled_dot_product_attention, Tensor.gelu, dtypes.float16
-
 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/diffusionmodules/util.py#L207
 def timestep_embedding(timesteps:Tensor, dim:int, max_period=10000):
  half = dim // 2
  freqs = (-math.log(max_period) * Tensor.arange(half, device=timesteps.device) / half).exp()
  args = timesteps.unsqueeze(1) * freqs.unsqueeze(0)
  out = Tensor.cat(args.cos(), args.sin(), dim=-1)
-  return out.cast(mixed_precision_dtype) if is_dtype_supported(mixed_precision_dtype) else out
+  return out.cast(dtypes.float16) if is_dtype_supported(dtypes.float16) else out

 class ResBlock:
-  def __init__(self, channels:int, emb_channels:int, out_channels:int, num_groups:int=32):
+  def __init__(self, channels:int, emb_channels:int, out_channels:int):
    self.in_layers = [
-      GroupNorm(num_groups, channels),
+      GroupNorm(32, channels),
      Tensor.silu,
      Conv2d(channels, out_channels, 3, padding=1),
    ]
@@ -27,7 +24,7 @@ class ResBlock:
      Linear(emb_channels, out_channels),
    ]
    self.out_layers = [
-      GroupNorm(num_groups, out_channels),
+      GroupNorm(32, out_channels),
      Tensor.silu,
      lambda x: x,  # needed for weights loading code to work
      Conv2d(out_channels, out_channels, 3, padding=1),
@@ -48,37 +45,35 @@ class CrossAttention:
    self.to_v = Linear(ctx_dim,   n_heads*d_head, bias=False)
    self.num_heads = n_heads
    self.head_size = d_head
-    self.attn = attention
    self.to_out = [Linear(n_heads*d_head, query_dim)]

  def __call__(self, x:Tensor, ctx:Optional[Tensor]=None) -> Tensor:
    ctx = x if ctx is None else ctx
    q,k,v = self.to_q(x), self.to_k(ctx), self.to_v(ctx)
    q,k,v = [y.reshape(x.shape[0], -1, self.num_heads, self.head_size).transpose(1,2) for y in (q,k,v)]
-    attention = self.attn(q, k, v).transpose(1,2)
+    attention = Tensor.scaled_dot_product_attention(q, k, v).transpose(1,2)
    h_ = attention.reshape(x.shape[0], -1, self.num_heads * self.head_size)
    return h_.sequential(self.to_out)

 class GEGLU:
  def __init__(self, dim_in:int, dim_out:int):
    self.proj = Linear(dim_in, dim_out * 2)
-    self.gelu = gelu
    self.dim_out = dim_out

  def __call__(self, x:Tensor) -> Tensor:
    x, gate = self.proj(x).chunk(2, dim=-1)
-    return x * self.gelu(gate)
+    return x * gate.gelu()

 class FeedForward:
  def __init__(self, dim:int, mult:int=4):
-    self.net: tuple[GEGLU, Callable, nn.Linear] = (
+    self.net = [
      GEGLU(dim, dim*mult),
      lambda x: x,  # needed for weights loading code to work
      Linear(dim*mult, dim)
-    )
+    ]

  def __call__(self, x:Tensor) -> Tensor:
-    return x.sequential(list(self.net))
+    return x.sequential(self.net)

 class BasicTransformerBlock:
  def __init__(self, dim:int, ctx_dim:int, n_heads:int, d_head:int):
@@ -97,13 +92,12 @@ class BasicTransformerBlock:

 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/attention.py#L619
 class SpatialTransformer:
-  def __init__(self, channels:int, n_heads:int, d_head:int, ctx_dim:Union[int,List[int]], use_linear:bool, depth:int=1,
-               norm_eps:float=1e-5):
+  def __init__(self, channels:int, n_heads:int, d_head:int, ctx_dim:Union[int,List[int]], use_linear:bool, depth:int=1):
    if isinstance(ctx_dim, int):
      ctx_dim = [ctx_dim]*depth
    else:
      assert isinstance(ctx_dim, list) and depth == len(ctx_dim)
-    self.norm = GroupNorm(32, channels, eps=norm_eps)
+    self.norm = GroupNorm(32, channels)
    assert channels == n_heads * d_head
    self.proj_in  = Linear(channels, channels) if use_linear else Conv2d(channels, channels, 1)
    self.transformer_blocks = [BasicTransformerBlock(channels, ctx_dim[d], n_heads, d_head) for d in range(depth)]
@@ -140,9 +134,7 @@ class Upsample:

 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/diffusionmodules/openaimodel.py#L472
 class UNetModel:
-  def __init__(self, adm_in_ch:Optional[int], in_ch:int, out_ch:int, model_ch:int, attention_resolutions:List[int], num_res_blocks:int,
-               channel_mult:List[int], transformer_depth:List[int], ctx_dim:Union[int,List[int]], use_linear:bool=False, d_head:Optional[int]=None,
-               n_heads:Optional[int]=None, num_groups:int=32, st_norm_eps:float=1e-5):
+  def __init__(self, adm_in_ch:Optional[int], in_ch:int, out_ch:int, model_ch:int, attention_resolutions:List[int], num_res_blocks:int, channel_mult:List[int], transformer_depth:List[int], ctx_dim:Union[int,List[int]], use_linear:bool=False, d_head:Optional[int]=None, n_heads:Optional[int]=None):
    self.model_ch = model_ch
    self.num_res_blocks = [num_res_blocks] * len(channel_mult)

@@ -182,12 +174,12 @@ class UNetModel:
    for idx, mult in enumerate(channel_mult):
      for _ in range(self.num_res_blocks[idx]):
        layers: List[Any] = [
-          ResBlock(ch, time_embed_dim, model_ch*mult, num_groups),
+          ResBlock(ch, time_embed_dim, model_ch*mult),
        ]
        ch = mult * model_ch
        if ds in attention_resolutions:
          d_head, n_heads = get_d_and_n_heads(ch)
-          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx], norm_eps=st_norm_eps))
+          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx]))

        self.input_blocks.append(layers)
        input_block_channels.append(ch)
@@ -201,9 +193,9 @@ class UNetModel:

    d_head, n_heads = get_d_and_n_heads(ch)
    self.middle_block: List = [
-      ResBlock(ch, time_embed_dim, ch, num_groups),
-      SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[-1], norm_eps=st_norm_eps),
-      ResBlock(ch, time_embed_dim, ch, num_groups),
+      ResBlock(ch, time_embed_dim, ch),
+      SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[-1]),
+      ResBlock(ch, time_embed_dim, ch),
    ]

    self.output_blocks = []
@@ -211,13 +203,13 @@ class UNetModel:
      for i in range(self.num_res_blocks[idx] + 1):
        ich = input_block_channels.pop()
        layers = [
-          ResBlock(ch + ich, time_embed_dim, model_ch*mult, num_groups),
+          ResBlock(ch + ich, time_embed_dim, model_ch*mult),
        ]
        ch = model_ch * mult

        if ds in attention_resolutions:
          d_head, n_heads = get_d_and_n_heads(ch)
-          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx], norm_eps=st_norm_eps))
+          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx]))

        if idx > 0 and i == self.num_res_blocks[idx]:
          layers.append(Upsample(ch))
@@ -225,7 +217,7 @@ class UNetModel:
        self.output_blocks.append(layers)

    self.out = [
-      GroupNorm(num_groups, ch),
+      GroupNorm(32, ch),
      Tensor.silu,
      Conv2d(model_ch, out_ch, 3, padding=1),
    ]
@@ -238,10 +230,10 @@ class UNetModel:
      assert y.shape[0] == x.shape[0]
      emb = emb + y.sequential(self.label_emb[0])

-    if is_dtype_supported(mixed_precision_dtype):
-      emb = emb.cast(mixed_precision_dtype)
-      ctx = ctx.cast(mixed_precision_dtype)
-      x   = x  .cast(mixed_precision_dtype)
+    if is_dtype_supported(dtypes.float16):
+      emb = emb.cast(dtypes.float16)
+      ctx = ctx.cast(dtypes.float16)
+      x   = x  .cast(dtypes.float16)

    def run(x:Tensor, bb) -> Tensor:
      if isinstance(bb, ResBlock): x = bb(x, emb)
@@ -1,23 +1,10 @@
 from tinygrad import Tensor
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.nn.onnx import OnnxRunner, OnnxValue
+from tinygrad.frontend.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:
@@ -57,9 +44,11 @@ def get_example_inputs(graph_inputs:dict[str, OnnxValue], config={}):
    ret.update({name:value})
  return ret

-def _get_tinygrad_and_ort_np_outputs(onnx_file, inputs):
+def validate(onnx_file, inputs, rtol=1e-5, atol=1e-5):
  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)
@@ -67,101 +56,9 @@ def _get_tinygrad_and_ort_np_outputs(onnx_file, inputs):
  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, 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
+    else: np.testing.assert_allclose(tiny_v.numpy(), onnx_v, rtol=rtol, atol=atol, err_msg=f"For tensor '{k}' in {tinygrad_out.keys()}")
@@ -81,7 +81,7 @@ def lin_to_feats(lin:Kernel, use_sts=True):
  ret = [float(x) for x in ret]

  if use_sts:
-    my_sts = dedup([(x.shape == lin.full_shape, x.is_expanded(), any(v.mask is not None for v in x.views), len(x.views)) for x in lin.sts])
+    my_sts = dedup([(x.shape == lin.full_shape, x.real_strides(), any(v.mask is not None for v in x.views), len(x.views)) for x in lin.sts])
    assert len(my_sts) < MAX_BUFS
    sts_len = 3 + 5*MAX_DIMS
    for s in my_sts:
@@ -50,7 +50,7 @@ class TestBeamSearch(unittest.TestCase):
  def test_variable_shrink_prime_number(self):
    v = Variable("v", 1, 400).bind(367)
    a = rand(400, 367)
-    b = (a.shrink(((0,v), None))+1)[:367,:367].realize()
+    b = (a.shrink(((0,v), None))+1).reshape(367,367).realize()
    np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)

  def test_no_mutate_rawbuffers(self):
@@ -1,8 +1,7 @@
 <?xml version="1.0" encoding="UTF-8"?>
 <database xmlns="http://nouveau.freedesktop.org/"
 xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
-xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.xsd">
-<import file="freedreno_copyright.xml"/>
+xsi:schemaLocation="http://nouveau.freedesktop.org/ rules-ng.xsd">
 <import file="adreno/adreno_common.xml"/>

 <enum name="vgt_event_type" varset="chip">
@@ -21,9 +20,9 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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"/>
@@ -31,8 +30,8 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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_COUNTER_CNTL.copy is true, writes OQ Z passed
-		sample counts to RB_SAMPLE_COUNTER_BASE.  This writes to main
+		If A6XX_RB_SAMPLE_COUNT_CONTROL.copy is true, writes OQ Z passed
+		sample counts to RB_SAMPLE_COUNT_ADDR.  This writes to main
 		memory, skipping UCHE.
 	</doc>
 	<value name="ZPASS_DONE" value="21"/>
@@ -97,13 +96,6 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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.
@@ -120,12 +112,11 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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="LRZ_CACHE_INVALIDATE" value="40" variants="A7XX-"/>
-	<value name="LRZ_Q_CACHE_INVALIDATE" value="41" variants="A7XX-"/>
+	<value name="UNK_40" value="40" variants="A7XX"/>
 	<value name="BLIT_OP_SCALE_2D" value="42" variants="A5XX-"/>
 	<value name="CONTEXT_DONE_2D" value="43" variants="A5XX-"/>
-	<value name="VSC_BINNING_START" value="44" variants="A5XX-"/>
-	<value name="VSC_BINNING_END" value="45" variants="A5XX-"/>
+	<value name="UNK_2C" value="44" variants="A5XX-"/>
+	<value name="UNK_2D" value="45" variants="A5XX-"/>

 	<!-- a6xx events -->
 	<doc>
@@ -138,22 +129,21 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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="DEPTH_BUFFER_FLIP" value="0x3d" variants="A8XX-"/>
+	<value name="CACHE_FLUSH7" value="50" variants="A7XX"/>
+	<value name="CACHE_INVALIDATE7" value="51" variants="A7XX"/>
 </enum>

 <enum name="pc_di_primtype">
@@ -334,7 +324,7 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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" variants="A2XX"/>
+	<value name="CP_SET_CONSTANT" value="0x2d"/>
 	<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>
@@ -381,7 +371,7 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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" variants="A3XX-A5XX"/>
+	<value name="CP_COND_INDIRECT_BUFFER_PFE" value="0x3a"/>
 	<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>
@@ -524,7 +514,7 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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
 	 -->
@@ -547,7 +537,7 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.xsd">
 	<value name="CP_LOAD_STATE6_GEOM" value="0x32" variants="A6XX-"/>
 	<value name="CP_LOAD_STATE6_FRAG" value="0x34" variants="A6XX-"/>
 	<!--
-	Note: For UAV state (Image/SSBOs) which have shared state across
+	Note: For IBO 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.
@@ -576,21 +566,20 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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"/>
-	<value name="CP_SET_AMBLE" value="0x55" variants="A6XX-"/>
+	<!-- Note, kgsl calls this CP_SET_AMBLE: -->
+	<value name="CP_SET_CTXSWITCH_IB" value="0x55" variants="A6XX-"/>

 	<!--
 	Seems to always have the payload:
@@ -641,7 +630,8 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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-"/>
-	<value name="CP_NON_CONTEXT_REG_BUNCH" value="0x5d" variants="A7XX-"/>
+	<!-- similar to CP_CONTEXT_REG_BUNCH, but discards first two dwords?? -->
+	<value name="CP_CONTEXT_REG_BUNCH2" 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.
@@ -658,11 +648,6 @@ xsi:schemaLocation="https://gitlab.freedesktop.org/freedreno/ rules-fd.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>


@@ -805,14 +790,14 @@ 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_UAV"       value="0xe"/>
-		<value name="SB6_CS_UAV"    value="0xf"/>
+		<value name="SB6_IBO"       value="0xe"/>
+		<value name="SB6_CS_IBO"    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_UAV" value="3"/>
+		<value name="ST6_IBO" value="3"/>
 	</enum>
 	<enum name="a6xx_state_src">
 		<value name="SS6_DIRECT" value="0"/>
@@ -918,6 +903,12 @@ 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 -->
@@ -1093,10 +1084,8 @@ 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"/>
-			<!-- high bit is 28 until a750: -->
-			<bitfield name="GROUP_ID" low="24" high="29" type="uint"/>
+			<bitfield name="GROUP_ID" low="24" high="28" 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>
@@ -1130,63 +1119,39 @@ 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>
-	<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>
+	<!-- 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"/>
 </domain>

 <domain name="CP_SET_BIN_DATA5_OFFSET" width="32">
@@ -1197,42 +1162,23 @@ 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>
-	<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>
+	<!-- 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>
 </domain>

 <domain name="CP_REG_RMW" width="32">
@@ -1250,9 +1196,6 @@ 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"/>
@@ -1266,7 +1209,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	</reg32>
 </domain>

-<domain name="CP_REG_TO_MEM" width="32" prefix="chip">
+<domain name="CP_REG_TO_MEM" width="32">
 	<reg32 offset="0" name="0">
 		<bitfield name="REG" low="0" high="17" type="hex"/>
 		<!-- number of registers/dwords copied is max(CNT, 1). -->
@@ -1274,12 +1217,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="64B" pos="30" type="boolean"/>
 		<bitfield name="ACCUMULATE" pos="31" type="boolean"/>
 	</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>
+	<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>
 </domain>

 <domain name="CP_REG_TO_MEM_OFFSET_REG" width="32">
@@ -1295,7 +1238,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="64B" pos="30" type="boolean"/>
 		<bitfield name="ACCUMULATE" pos="31" type="boolean"/>
 	</reg32>
-	<reg64 offset="1" name="DEST" type="waddress"/>
+	<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="OFFSET0" low="0" high="17" type="hex"/>
 		<bitfield name="OFFSET0_SCRATCH" pos="19" type="boolean"/>
@@ -1315,8 +1263,18 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<bitfield name="64B" pos="30" type="boolean"/>
 		<bitfield name="ACCUMULATE" pos="31" type="boolean"/>
 	</reg32>
-	<reg64 offset="1" name="DEST" type="waddress"/>
-	<reg64 offset="3" name="OFFSET" type="waddress"/>
+	<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>
 </domain>

 <domain name="CP_MEM_TO_REG" width="32">
@@ -1329,12 +1287,12 @@ 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>
-	<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>
+	<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>
 </domain>

 <domain name="CP_MEM_TO_MEM" width="32">
@@ -1354,10 +1312,6 @@ 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
@@ -1392,8 +1346,6 @@ 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>

@@ -1416,12 +1368,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 </domain>

 <domain name="CP_MEM_WRITE" width="32">
-	<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>
+	<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>
 	<!-- followed by the DWORDs to write -->
 </domain>

@@ -1473,14 +1425,24 @@ 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>
-	<reg64 offset="1" name="POLL_ADDR" type="address"/>
+	<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>
 	<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>
-	<reg64 offset="5" name="WRITE_ADDR" type="waddress"/>
+	<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>
 	<reg32 offset="7" name="7">
 		<bitfield name="WRITE_DATA" low="0" high="31"/>
 	</reg32>
@@ -1495,7 +1457,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		<!-- Reserved for flags, presumably? Unused in FW -->
 		<bitfield name="RESERVED" low="0" high="31" type="hex"/>
 	</reg32>
-	<reg64 offset="1" name="POLL_ADDR" type="address"/>
+	<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>
 	<reg32 offset="3" name="3">
 		<bitfield name="REF" low="0" high="31"/>
 	</reg32>
@@ -1513,7 +1480,12 @@ 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>
-	<reg64 offset="1" name="POLL_ADDR" type="address"/>
+	<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>
 	<reg32 offset="3" name="3">
 		<bitfield name="REF" low="0" high="31"/>
 	</reg32>
@@ -1647,7 +1619,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	TODO what is gpuaddr for, seems to be all 0's.. maybe needed for
 	context switch?
 	 -->
-	<reg64 offset="1" name="ADDR" type="waddress"/>
+	<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>
 	<reg32 offset="3" name="3">
 		<!-- ??? -->
 	</reg32>
@@ -1676,8 +1653,8 @@ 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_ACCUM_SAMPLE_COUNT_DIFF" pos="14" type="boolean"/>
+		<!-- *(iova + 8) = *(iova + 16) - *iova -->
+		<bitfield name="WRITE_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 -->
@@ -1691,11 +1668,15 @@ 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">
-		<reg64 offset="1" name="1" type="waddress"/>
+		<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>
 		<reg32 offset="3" name="3">
 			<bitfield name="PAYLOAD_0" low="0" high="31"/>
 		</reg32>
@@ -1762,7 +1743,9 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	<reg32 offset="0" name="0">
 	</reg32>
 	<stripe varset="chip" variants="A4XX">
-		<reg32 offset="1" name="ADDR" type="address"/>
+		<reg32 offset="1" name="1">
+			<bitfield name="ADDR" low="0" high="31"/>
+		</reg32>
 		<reg32 offset="2" name="2">
 			<!-- localsize is value minus one: -->
 			<bitfield name="LOCALSIZEX" low="2" high="11" type="uint"/>
@@ -1771,7 +1754,12 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 		</reg32>
 	</stripe>
 	<stripe varset="chip" variants="A5XX-">
-		<reg64 offset="1" name="ADDR" type="address"/>
+		<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>
 		<reg32 offset="3" name="3">
 			<!-- localsize is value minus one: -->
 			<bitfield name="LOCALSIZEX" low="2" high="11" type="uint"/>
@@ -1783,88 +1771,40 @@ 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_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="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="8" name="RM6_COMPUTE"/>
-		<value value="12" name="RM6_BLIT2DSCALE"/>  <!-- no-op (at least on current sqe fw) -->
+		<value value="0xc" 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="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"/>
+		<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"/>
 	</enum>
-	<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>
+	<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>
 </domain>

 <domain name="CP_SET_PSEUDO_REG" width="32" varset="chip" prefix="chip" variants="A6XX-">
@@ -1890,9 +1830,9 @@ 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="VSC_PIPE_DATA_DRAW_BASE"/>
-		<value value="9" name="VSC_SIZE_BASE"/>
-		<value value="10" name="VSC_PIPE_DATA_PRIM_BASE"/>
+		<value value="8" name="DRAW_STRM_ADDRESS"/>
+		<value value="9" name="DRAW_STRM_SIZE_ADDRESS"/>
+		<value value="10" name="PRIM_STRM_ADDRESS"/>
 		<value value="11" name="UNK_STRM_ADDRESS"/>
 		<value value="12" name="UNK_STRM_SIZE_ADDRESS"/>

@@ -1993,11 +1933,11 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 			a bitmask of which modes pass the test.
 		-->

-		<!-- RM6_BIN_VISIBILITY -->
+		<!-- RM6_BINNING -->
 		<bitfield name="BINNING" pos="25" variants="RENDER_MODE" type="boolean"/>
 		<!-- all others -->
 		<bitfield name="GMEM" pos="26" variants="RENDER_MODE" type="boolean"/>
-		<!-- RM6_DIRECT_RENDER -->
+		<!-- RM6_BYPASS -->
 		<bitfield name="SYSMEM" pos="27" variants="RENDER_MODE" type="boolean"/>

 		<bitfield name="BV" pos="25" variants="THREAD_MODE" type="boolean"/>
@@ -2070,45 +2010,54 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
 	</reg32>
 </domain>

-<domain name="CP_SET_AMBLE" width="32">
+<domain name="CP_SET_CTXSWITCH_IB" width="32">
 	<doc>
-		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.
+                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.
 	</doc>
-	<enum name="amble_type">
-		<value name="PREAMBLE_AMBLE_TYPE" value="0">
+	<enum name="ctxswitch_ib">
+		<value name="RESTORE_IB" value="0">
 			<doc>Executed unconditionally when switching back to the context.</doc>
 		</value>
-		<value name="BIN_PREAMBLE_AMBLE_TYPE" value="1">
+		<value name="YIELD_RESTORE_IB" value="1">
                        <doc>
 				Executed when switching back after switching
 				away during execution of
-				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.
+				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.
 			</doc>
 		</value>
-		<value name="POSTAMBLE_AMBLE_TYPE" value="2">
+		<value name="SAVE_IB" value="2">
                        <doc>
 				Executed when switching away from the context,
 				except for context switches initiated via
 				CP_YIELD.
                        </doc>
 		</value>
-		<value name="KMD_AMBLE_TYPE" value="3">
+		<value name="RB_SAVE_IB" 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 POSTAMBLE_AMBLE_TYPE.
+				is otherwise similar to SAVE_IB.
+
+				Note, kgsl calls this CP_KMD_AMBLE_TYPE
 			</doc>
 		</value>
 	</enum>
-	<reg64 offset="0" name="ADDR" type="address"/>
+	<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>
 	<reg32 offset="2" name="2">
 		<bitfield name="DWORDS" low="0" high="19" type="uint"/>
-		<bitfield name="TYPE" low="20" high="21" type="amble_type"/>
+		<bitfield name="TYPE" low="20" high="21" type="ctxswitch_ib"/>
 	</reg32>
 </domain>

@@ -2140,12 +2089,12 @@ 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_VIEW_INFO with previous values, and if one of
+			GRAS_LRZ_DEPTH_VIEW 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_VIEW_INFO has changed
+			- GRAS_LRZ_DEPTH_VIEW has changed
 			then it does a LRZ_FLUSH with GRAS_LRZ_CNTL::ENABLE
 			forced to 1.
 			Only exists in a650_sqe.fw.
@@ -2260,7 +2209,7 @@ 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_CHANNEL_VISIBILITY registers
+		Best guess is that it is a faster way to fetch all the VSC_STATE registers
 		and keep them in a local scratch memory instead of fetching every time
 		when skipping IBs.
 	</doc>
@@ -2308,25 +2257,7 @@ 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_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"/>
+		<bitfield name="CLEAR_GLOBAL_LOCAL_TS" pos="2" type="boolean"/>
 	</reg32>
 </domain>

@@ -10,8 +10,6 @@ import ctypes


 class AsDictMixin:
-    import sys
-    if sys.version_info >= (3, 14): _layout_ = 'ms'
    @classmethod
    def as_dict(cls, self):
        result = {}
@@ -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 > 3:
+          if state_type == ST6_SHADER and IOCTL > 2:
            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)
@@ -106,30 +106,25 @@ def parse_cmd_buf(dat):
              print('constants')
              hexdump(x)
          if state_type == ST6_IBO:
-            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)
+            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:
-            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)
+            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:
-            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)
+            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)
@@ -157,10 +152,6 @@ 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]))
@@ -180,8 +171,8 @@ 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":
-      mmaped[s.gpuaddr] = mmap.mmap(fd, s.size, offset=s.id*0x1000)
+    if name == "IOCTL_KGSL_GPUOBJ_INFO": pass
+      # 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)
@@ -882,11 +882,6 @@ 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 {
@@ -935,7 +930,7 @@ impl<'a> Thread<'a> {

            let op = ((instr >> 16) & 0x3ff) as u32;
            match op {
-                764 | 765 | 288 | 289 | 290 | 766 | 767 | 768 | 769 | 770 => {
+                764 | 765 | 288 | 289 | 290 | 766 | 768 | 769 => {
                    let vdst = (instr & 0xff) as usize;
                    let sdst = ((instr >> 8) & 0x7f) as usize;
                    let f = |i: u32| -> usize { ((instr >> i) & 0x1ff) as usize };
@@ -949,16 +944,6 @@ impl<'a> Thread<'a> {
                    assert_eq!(clmp, 0);

                    let vcc = match op {
-                        767 => {
-                            let (s0, s1, s2): (u32, u32, u64) = (self.val(s0), self.val(s1), self.val(s2));
-                            let (mul_result, overflow_mul) = (s0 as i64).overflowing_mul(s1 as i64);
-                            let (ret, overflow_add) = mul_result.overflowing_add(s2 as i64);
-                            let overflowed = overflow_mul || overflow_add;
-                            if self.exec.read() {
-                                self.vec_reg.write64(vdst, ret as u64);
-                            }
-                            overflowed
-                        },
                        766 => {
                            let (s0, s1, s2): (u32, u32, u64) = (self.val(s0), self.val(s1), self.val(s2));
                            let (mul_result, overflow_mul) = (s0 as u64).overflowing_mul(s1 as u64);
@@ -1001,10 +986,6 @@ 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() {
@@ -1,32 +1,98 @@
 import numpy as np
 import unittest
 import subprocess, struct, math
-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
+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

-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()
+@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 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]
@@ -39,57 +105,54 @@ class TestHW(unittest.TestCase):

  def test_simple(self):
    out = get_output("""
-    v_mov_b32_e32 %1 42
-    v_mov_b32_e32 %2 %1
-    """)[0]
+    v_mov_b32_e32 v10 42
+    v_mov_b32_e32 v1 v10
+    """, n_threads=2)
    np.testing.assert_equal(out, 42)

  def test_exec_mov(self):
    out = get_output("""
-    v_mov_b32_e32 %1 42
+    v_mov_b32_e32 v10 42
    s_mov_b32_e32 exec_lo 0b10
-    v_mov_b32_e32 %1 10
+    v_mov_b32_e32 v10 10
    s_mov_b32_e32 exec_lo 0b11
-    v_mov_b32_e32 %2 %1
+    v_mov_b32_e32 v1 v10
    """, 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 %1 42
-    v_mov_b32_e32 %2 10
+    v_mov_b32_e32 v10 42
+    v_mov_b32_e32 v11 10
    s_mov_b32_e32 exec_lo 0b01
-    v_cmp_ne_u32 %1 %2
+    v_cmp_ne_u32 v10 v11
    s_mov_b32_e32 exec_lo 0b11
-    v_mov_b32_e32 %2 vcc_lo
+    v_mov_b32_e32 v1 vcc_lo
    """, n_threads=2)
    np.testing.assert_equal(out, 0b01)

  def test_exec_cmpx_vop3(self):
    out = get_output("""
-    s_mov_b32_e32 exec_lo 0b11
-    v_mov_b32_e32 %1 42
-    v_mov_b32_e32 %2 10
+    v_mov_b32_e32 v10 42
+    v_mov_b32_e32 v11 10
    s_mov_b32_e32 exec_lo 0b01
-    v_cmpx_ne_u32 %1 %2
+    v_cmpx_ne_u32 v10 v11
    s_mov_b32_e32 s10 exec_lo
    s_mov_b32_e32 exec_lo 0b11
-    v_mov_b32_e32 %2 s10
-    """, n_threads=2)[0]
-    np.testing.assert_equal(out & 0b11, 0b01)
+    v_mov_b32_e32 v1 s10
+    """, n_threads=2)
+    np.testing.assert_equal(out, 0b01)

  def test_fmac_vop3_modifier(self):
    init_state = f"""
-    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));
+    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)}
    """
-    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.))
+    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.))

  def test_s_abs_i32(self):
    def s_abs_i32(x, y, dst="s10", scc=0):
@@ -97,7 +160,7 @@ class TestHW(unittest.TestCase):
        self.assertEqual(get_output(f"""
        s_mov_b32_e32 {dst} {x}
        s_abs_i32 {dst} {dst}
-        v_mov_b32_e32 %2 {reg}
+        v_mov_b32_e32 v1 {reg}
        """)[0], val)
    s_abs_i32(0x00000001, 0x00000001, scc=1)
    s_abs_i32(0x7fffffff, 0x7fffffff, scc=1)
@@ -110,8 +173,8 @@ 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 %2 {f32_to_bits(x)}
-      v_rcp_f32_e64 %2, -%2
+      v_mov_b32_e32 v1 {f32_to_bits(x)}
+      v_rcp_f32_e64 v1, -v1
      """)[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)
@@ -123,11 +186,10 @@ 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 %2 {f32_to_bits(x)}
-      s_mov_b32_e32 s10 1
-      v_cndmask_b32 %2, %2, -%2 s10
+      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
      """)[0]
      assert out == f32_to_bits(y), f"{f32_from_bits(out)} != {y} / {out} != {f32_to_bits(y)}"
    v_neg(-0.0, 0.0)
@@ -136,12 +198,5 @@ 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()
@@ -2,7 +2,9 @@

 ## Getting SQ Thread Trace

-SQTT is implemented on top of normal tinygrad profiling, `VIZ=1 SQTT=1` to get profile pickle with sqtt data embedded in it.
+Only supported on 7900XTX, requires either AM (`rmmod amdgpu`) or disabling power gating on AMD (`ppfeaturemask=0xffff3fff`, don't forget to rebuild initramfs)
+
+SQTT is implemented on top of normal tinygrad PROFILE=1, `PROFILE=1 SQTT=1` to get profile pickle with sqtt data embedded in it.

 `SQTT_BUFFER_SIZE=X` to change size of SQTT buffer (per shader engine, 6 SEs on 7900xtx) in megabytes, default 256.

@@ -1,68 +0,0 @@
-import ctypes
-from dataclasses import dataclass
-import tinygrad.runtime.autogen.comgr as comgr
-from tinygrad.runtime.support.compiler_amd import check
-
-@dataclass
-class InstrCtx:
-  pc:int=0
-  inst:str=""
-
-@comgr.amd_comgr_create_disassembly_info.argtypes[2]
-def instr_cb(text, user_data):
-  c = ctypes.cast(user_data, ctypes.POINTER(ctypes.py_object)).contents.value
-  c.inst = ctypes.string_at(text).decode("utf-8","replace").strip()
-  return comgr.AMD_COMGR_STATUS_SUCCESS
-
-# nop callback
-@comgr.amd_comgr_create_disassembly_info.argtypes[3]
-def addr_cb(*args): return comgr.AMD_COMGR_STATUS_SUCCESS
-
-def comgr_get_address_table(lib:bytes) -> dict[int, tuple[str, int]]:
-  check(comgr.amd_comgr_create_data(comgr.AMD_COMGR_DATA_KIND_EXECUTABLE, ctypes.byref(data_src:=comgr.amd_comgr_data_t())))
-  lib_buf = ctypes.create_string_buffer(lib, len(lib))
-  check(comgr.amd_comgr_set_data(data_src, len(lib), lib_buf))
-  check(comgr.amd_comgr_get_data_isa_name(data_src, isa_sz:=ctypes.c_size_t(128), isa:=(ctypes.c_char*isa_sz.value)()))
-
-  @comgr.amd_comgr_create_disassembly_info.argtypes[1]
-  def memory_cb(from_addr, to, size, _):
-    base, buf_len = ctypes.addressof(lib_buf), len(lib_buf)
-    start = int(from_addr) - base
-    if start < 0 or start >= buf_len: return 0
-    ctypes.memmove(to, base + start, n:=min(int(size), buf_len - start))
-    return n
-
-  info_src = comgr.amd_comgr_disassembly_info_t()
-  check(comgr.amd_comgr_create_disassembly_info(ctypes.cast(isa, ctypes.POINTER(ctypes.c_char)), memory_cb, instr_cb, addr_cb, info_src))
-
-  @comgr.amd_comgr_iterate_symbols.argtypes[1]
-  def sym_callback(sym, udata):
-    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_TYPE, ctypes.byref(sym_type:=ctypes.c_int())))
-    if sym_type.value != comgr.AMD_COMGR_SYMBOL_TYPE_FUNC: return comgr.AMD_COMGR_STATUS_SUCCESS
-    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_VALUE, ctypes.byref(vaddr:=ctypes.c_uint64())))
-    check(comgr.amd_comgr_symbol_get_info(sym, comgr.AMD_COMGR_SYMBOL_INFO_SIZE, ctypes.byref(size:=ctypes.c_uint64())))
-    check(comgr.amd_comgr_map_elf_virtual_address_to_code_object_offset(data_src, vaddr.value, ctypes.byref(offset:=ctypes.c_uint64()),
-                                                                        ctypes.byref(ctypes.c_uint64()), ctypes.byref(nobits:=ctypes.c_bool())))
-    check(nobits.value)
-    base = ctypes.addressof(lib_buf)
-    pc = base + offset.value
-    end = pc + size.value
-    addr_table = ctypes.cast(udata, ctypes.POINTER(ctypes.py_object)).contents.value
-    instr_ref = ctypes.py_object(ctx:=InstrCtx())
-    instr_ptr = ctypes.cast(ctypes.pointer(instr_ref), ctypes.c_void_p)
-    while pc < end:
-      size_read = ctypes.c_uint64(0)
-      ctx.pc = pc
-      st = comgr.amd_comgr_disassemble_instruction(info_src, ctypes.c_uint64(pc), instr_ptr, ctypes.byref(size_read))
-      if st == comgr.AMD_COMGR_STATUS_SUCCESS and size_read.value:
-        rel = (pc - base) - offset.value
-        addr_table[vaddr.value + rel] = (ctx.inst, int(size_read.value))
-        pc += size_read.value
-      else: # don't inf loop if comgr fails
-        b = ctypes.c_ubyte.from_buffer(lib_buf, pc - base).value
-        addr_table[vaddr.value + (pc - base - offset.value)] = (f"DISASSEMBLER ISSUE 0x{b:02x}", 1)
-        pc += 1
-    return comgr.AMD_COMGR_STATUS_SUCCESS
-  addr_table:dict[int, tuple[str, int]] = {}
-  check(comgr.amd_comgr_iterate_symbols(data_src, sym_callback, ctypes.cast(ctypes.pointer(ctypes.py_object(addr_table)), ctypes.c_void_p)))
-  return addr_table
@@ -155,10 +155,6 @@ class RGP:
      device_event = device_events[device]
    sqtt_events = [x for x in profile if isinstance(x, ProfileSQTTEvent) and x.device == device_event.device]
    if len(sqtt_events) == 0: raise RuntimeError(f"Device {device_event.device} doesn't contain SQTT data")
-    device_props = sqtt_events[0].props
-    gfx_ver = device_props['gfx_target_version'] // 10000
-    gfx_iplvl = getattr(sqtt, f"SQTT_GFXIP_LEVEL_GFXIP_{device_props['gfx_target_version']//10000}_{(device_props['gfx_target_version']//100)%100}",
-                        getattr(sqtt, f"SQTT_GFXIP_LEVEL_GFXIP_{device_props['gfx_target_version']//10000}", None))
    sqtt_itrace_enabled = any([event.itrace for event in sqtt_events])
    sqtt_itrace_masked = not all_same([event.itrace for event in sqtt_events])
    sqtt_itrace_se_mask = functools.reduce(lambda a,b: a|b, [int(event.itrace) << event.se for event in sqtt_events], 0) if sqtt_itrace_masked else 0
@@ -196,21 +192,21 @@ class RGP:
        flags=0,
        trace_shader_core_clock=0x93f05080,
        trace_memory_clock=0x4a723a40,
-        device_id={110000: 0x744c, 110003: 0x7480, 120001: 0x7550}[device_props['gfx_target_version']],
+        device_id=0x744c,
        device_revision_id=0xc8,
        vgprs_per_simd=1536,
        sgprs_per_simd=128*16,
-        shader_engines=device_props['array_count'] // device_props['simd_arrays_per_engine'],
-        compute_unit_per_shader_engine=device_props['simd_count'] // device_props['simd_per_cu'] // (device_props['array_count'] // device_props['simd_arrays_per_engine']),
-        simd_per_compute_unit=device_props['simd_per_cu'],
-        wavefronts_per_simd=device_props['max_waves_per_simd'],
+        shader_engines=6,
+        compute_unit_per_shader_engine=16,
+        simd_per_compute_unit=2,
+        wavefronts_per_simd=16,
        minimum_vgpr_alloc=4,
        vgpr_alloc_granularity=8,
        minimum_sgpr_alloc=128,
        sgpr_alloc_granularity=128,
        hardware_contexts=8,
        gpu_type=sqtt.SQTT_GPU_TYPE_DISCRETE,
-        gfxip_level=gfx_iplvl,
+        gfxip_level=sqtt.SQTT_GFXIP_LEVEL_GFXIP_11_0,
        gpu_index=0,
        gds_size=0,
        gds_per_shader_engine=0,
@@ -222,7 +218,7 @@ class RGP:
        vram_bus_width=384, # 384-bit
        l2_cache_size=6 * 1024 * 1024, # 6 MB
        l1_cache_size=32 * 1024, # 32 KB per SIMD (?)
-        lds_size=device_props['lds_size_in_kb'] * 1024,
+        lds_size=65536, # 64 KB per CU
        gpu_name=b'NAVI31',
        alu_per_clock=0,
        texture_per_clock=0,
@@ -261,7 +257,7 @@ class RGP:
            major_version=0, minor_version=2,
          ),
          shader_engine_index=sqtt_event.se,
-          sqtt_version={11: sqtt.SQTT_VERSION_3_2, 12: sqtt.SQTT_VERSION_3_3}.get(gfx_ver),
+          sqtt_version=sqtt.SQTT_VERSION_3_2,
          _0=sqtt.union_sqtt_file_chunk_sqtt_desc_0(
            v1=sqtt.struct_sqtt_file_chunk_sqtt_desc_0_v1(
              instrumentation_spec_version=1,
@@ -1,114 +0,0 @@
-import ctypes, pathlib, argparse, pickle, re, functools, dataclasses, itertools
-from extra.sqtt.rocprof import rocprof
-from extra.sqtt.disasm import comgr_get_address_table
-from tinygrad.helpers import temp, DEBUG
-from tinygrad.device import ProfileEvent, ProfileProgramEvent
-from tinygrad.runtime.ops_amd import ProfileSQTTEvent, ProfilePMCEvent
-
-@dataclasses.dataclass
-class InstInfo:
-  typ:str=""
-  inst:str=""
-  hit:int=0
-  lat:int=0
-  stall:int=0
-  def __str__(self): return f"{self.inst:>20} hits:{self.typ:>6} hits:{self.hit:>6} latency:{self.lat:>6} stall:{self.stall:>6}"
-
-  def on_ev(self, ev):
-    self.hit, self.lat, self.stall = self.hit + 1, self.lat + ev.duration, self.stall + ev.stall
-
-class _ROCParseCtx:
-  def __init__(self, sqtt_evs:list[ProfileSQTTEvent], prog_evs:list[ProfileProgramEvent]):
-    self.sqtt_evs, self.prog_evs = iter(sqtt_evs), prog_evs
-    self.wave_events, self.disasms, self.addr2prg = {}, {}, {}
-
-    for prog in prog_evs:
-      for addr, info in comgr_get_address_table(prog.lib).items():
-        self.disasms[prog.base + addr] = info
-        self.addr2prg[prog.base + addr] = prog
-
-  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, 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, self.active_se, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
-
-    asm = {}
-    for j in range(ev.instructions_size):
-      inst_ev = ev.instructions_array[j]
-      inst_typ = rocprof.rocprofiler_thread_trace_decoder_inst_category_t__enumvalues[inst_ev.category]
-      asm.setdefault(inst_ev.pc.address, InstInfo(typ=inst_typ, inst=self.disasms[inst_ev.pc.address][0]))
-      asm[inst_ev.pc.address].on_ev(inst_ev)
-
-    self.wave_events[(self.find_program(ev.instructions_array[0].pc.address).name, ev.wave_id, ev.cu, ev.simd)] = asm
-
-if __name__ == "__main__":
-  parser = argparse.ArgumentParser()
-  parser.add_argument('--profile', type=pathlib.Path, help='Path to profile', default=pathlib.Path(temp("profile.pkl", append_user=True)))
-  args = parser.parse_args()
-
-  with args.profile.open("rb") as f: profile = pickle.load(f)
-  sqtt_events:list[ProfileSQTTEvent] = []
-  pmc_events:list[ProfilePMCEvent] = []
-  prog_events:list[ProfileProgramEvent] = []
-  for e in profile:
-    if isinstance(e, ProfileSQTTEvent): sqtt_events.append(e)
-    if isinstance(e, ProfilePMCEvent): pmc_events.append(e)
-    if isinstance(e, ProfileProgramEvent) and e.device.startswith("AMD"): prog_events.append(e)
-
-  ROCParseCtx = _ROCParseCtx(sqtt_events, prog_events)
-
-  @rocprof.rocprof_trace_decoder_se_data_callback_t
-  def copy_cb(buf, buf_size, data_ptr):
-    if (prof:=ROCParseCtx.next_sqtt()) is None: return 0
-    buf[0] = ctypes.cast((ctypes.c_ubyte * len(prof.blob)).from_buffer_copy(prof.blob), ctypes.POINTER(ctypes.c_ubyte))
-    buf_size[0] = len(prof.blob)
-    return len(prof.blob)
-
-  @rocprof.rocprof_trace_decoder_trace_callback_t
-  def trace_cb(record_type, events_ptr, n, data_ptr):
-    match record_type:
-      case rocprof.ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY:
-        for ev in (rocprof.rocprofiler_thread_trace_decoder_occupancy_t * n).from_address(events_ptr): ROCParseCtx.on_occupancy_ev(ev)
-      case rocprof.ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE:
-        for ev in (rocprof.rocprofiler_thread_trace_decoder_wave_t * n).from_address(events_ptr): ROCParseCtx.on_wave_ev(ev)
-      case _:
-        if DEBUG >= 2: print(rocprof.rocprofiler_thread_trace_decoder_record_type_t__enumvalues[record_type], events_ptr, n)
-    return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS
-
-  @rocprof.rocprof_trace_decoder_isa_callback_t
-  def isa_cb(instr_ptr, mem_size_ptr, size_ptr, pc, data_ptr):
-    instr, mem_size_ptr[0] = ROCParseCtx.disasms[pc.address]
-
-    # this is the number of bytes to next instruction, set to 0 for end_pgm
-    if instr == "s_endpgm": mem_size_ptr[0] = 0
-    if (max_sz:=size_ptr[0]) == 0: return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES
-
-    # truncate the instr if it doesn't fit
-    if (str_sz:=len(instr_bytes:=instr.encode()))+1 > max_sz: str_sz = max_sz
-    ctypes.memmove(instr_ptr, instr_bytes, str_sz)
-    size_ptr[0] = str_sz
-
-    return rocprof.ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS
-
-  try:
-    rocprof.rocprof_trace_decoder_parse_data(copy_cb, trace_cb, isa_cb, None)
-    print('SQTT:', ROCParseCtx.wave_events.keys())
-  except Exception as e: print("Error in sqtt decoder:", e)
-
-  for ev in pmc_events:
-    print(f"PMC Event: dev={ev.device} kern={ev.kern}")
-    ptr = 0
-    for s in ev.sched:
-      view = memoryview(ev.blob).cast('Q')
-      print(f"\t{s.name}")
-      for xcc, inst, se_idx, sa_idx, wgp_idx in itertools.product(range(s.xcc), range(s.inst), range(s.se), range(s.sa), range(s.wgp)):
-        print(f"\t\tXCC {xcc} Inst {inst} SE {se_idx} SA {sa_idx} WGP {wgp_idx}: {view[ptr]:#x}")
-        ptr += 1
@@ -1,18 +0,0 @@
-#!/usr/bin/env python3
-import os, shutil
-from pathlib import Path
-from tinygrad.helpers import fetch, OSX
-
-DEST = Path("/usr/local/lib")
-DEST.mkdir(exist_ok=True)
-
-if __name__ == "__main__":
-  if OSX:
-    fp = fetch("https://github.com/ROCm/rocprof-trace-decoder/releases/download/0.1.4/rocprof-trace-decoder-macos-arm64-0.1.4-Darwin.sh")
-    lib = fp.parent/"rocprof-trace-decoder-macos-arm64-0.1.4-Darwin"/"lib"/"librocprof-trace-decoder.dylib"
-    os.chmod(fp, 0o755)
-    os.system(f"sudo {fp} --prefix={fp.parent} --include-subdir")
-  else:
-    lib = fetch("https://github.com/ROCm/rocprof-trace-decoder/raw/5420409ad0963b2d76450add067b9058493ccbd0/releases/linux_glibc_2_28_x86_64/librocprof-trace-decoder.so", name="librocprof-trace-decoder.so")
-  shutil.copy2(lib, DEST)
-  print(f"Installed {lib.name} to", DEST)
@@ -1,658 +0,0 @@
-# pylint: skip-file
-# mypy: ignore-errors
-# -*- coding: utf-8 -*-
-#
-# TARGET arch is: []
-# WORD_SIZE is: 8
-# POINTER_SIZE is: 8
-# LONGDOUBLE_SIZE is: 16
-#
-import ctypes, ctypes.util
-
-
-class AsDictMixin:
-    import sys
-    if sys.version_info >= (3, 14): _layout_ = 'ms'
-    @classmethod
-    def as_dict(cls, self):
-        result = {}
-        if not isinstance(self, AsDictMixin):
-            # not a structure, assume it's already a python object
-            return self
-        if not hasattr(cls, "_fields_"):
-            return result
-        # sys.version_info >= (3, 5)
-        # for (field, *_) in cls._fields_:  # noqa
-        for field_tuple in cls._fields_:  # noqa
-            field = field_tuple[0]
-            if field.startswith('PADDING_'):
-                continue
-            value = getattr(self, field)
-            type_ = type(value)
-            if hasattr(value, "_length_") and hasattr(value, "_type_"):
-                # array
-                if not hasattr(type_, "as_dict"):
-                    value = [v for v in value]
-                else:
-                    type_ = type_._type_
-                    value = [type_.as_dict(v) for v in value]
-            elif hasattr(value, "contents") and hasattr(value, "_type_"):
-                # pointer
-                try:
-                    if not hasattr(type_, "as_dict"):
-                        value = value.contents
-                    else:
-                        type_ = type_._type_
-                        value = type_.as_dict(value.contents)
-                except ValueError:
-                    # nullptr
-                    value = None
-            elif isinstance(value, AsDictMixin):
-                # other structure
-                value = type_.as_dict(value)
-            result[field] = value
-        return result
-
-
-class Structure(ctypes.Structure, AsDictMixin):
-
-    def __init__(self, *args, **kwds):
-        # We don't want to use positional arguments fill PADDING_* fields
-
-        args = dict(zip(self.__class__._field_names_(), args))
-        args.update(kwds)
-        super(Structure, self).__init__(**args)
-
-    @classmethod
-    def _field_names_(cls):
-        if hasattr(cls, '_fields_'):
-            return (f[0] for f in cls._fields_ if not f[0].startswith('PADDING'))
-        else:
-            return ()
-
-    @classmethod
-    def get_type(cls, field):
-        for f in cls._fields_:
-            if f[0] == field:
-                return f[1]
-        return None
-
-    @classmethod
-    def bind(cls, bound_fields):
-        fields = {}
-        for name, type_ in cls._fields_:
-            if hasattr(type_, "restype"):
-                if name in bound_fields:
-                    if bound_fields[name] is None:
-                        fields[name] = type_()
-                    else:
-                        # use a closure to capture the callback from the loop scope
-                        fields[name] = (
-                            type_((lambda callback: lambda *args: callback(*args))(
-                                bound_fields[name]))
-                        )
-                    del bound_fields[name]
-                else:
-                    # default callback implementation (does nothing)
-                    try:
-                        default_ = type_(0).restype().value
-                    except TypeError:
-                        default_ = None
-                    fields[name] = type_((
-                        lambda default_: lambda *args: default_)(default_))
-            else:
-                # not a callback function, use default initialization
-                if name in bound_fields:
-                    fields[name] = bound_fields[name]
-                    del bound_fields[name]
-                else:
-                    fields[name] = type_()
-        if len(bound_fields) != 0:
-            raise ValueError(
-                "Cannot bind the following unknown callback(s) {}.{}".format(
-                    cls.__name__, bound_fields.keys()
-            ))
-        return cls(**fields)
-
-
-class Union(ctypes.Union, AsDictMixin):
-    pass
-
-
-
-c_int128 = ctypes.c_ubyte*16
-c_uint128 = c_int128
-void = None
-if ctypes.sizeof(ctypes.c_longdouble) == 16:
-    c_long_double_t = ctypes.c_longdouble
-else:
-    c_long_double_t = ctypes.c_ubyte*16
-
-def string_cast(char_pointer, encoding='utf-8', errors='strict'):
-    value = ctypes.cast(char_pointer, ctypes.c_char_p).value
-    if value is not None and encoding is not None:
-        value = value.decode(encoding, errors=errors)
-    return value
-
-
-def char_pointer_cast(string, encoding='utf-8'):
-    if encoding is not None:
-        try:
-            string = string.encode(encoding)
-        except AttributeError:
-            # In Python3, bytes has no encode attribute
-            pass
-    string = ctypes.c_char_p(string)
-    return ctypes.cast(string, ctypes.POINTER(ctypes.c_char))
-
-
-
-class FunctionFactoryStub:
-    def __getattr__(self, _):
-      return ctypes.CFUNCTYPE(lambda y:y)
-
-# libraries['FIXME_STUB'] explanation
-# As you did not list (-l libraryname.so) a library that exports this function
-# This is a non-working stub instead.
-# You can either re-run clan2py with -l /path/to/library.so
-# Or manually fix this by comment the ctypes.CDLL loading
-_libraries = {}
-_libraries['FIXME_STUB'] = ctypes.CDLL(ctypes.util.find_library('rocprof-trace-decoder')) #  ctypes.CDLL('FIXME_STUB')
-
-
-
-# values for enumeration 'rocprofiler_thread_trace_decoder_info_t'
-rocprofiler_thread_trace_decoder_info_t__enumvalues = {
-    0: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_NONE',
-    1: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_DATA_LOST',
-    2: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_STITCH_INCOMPLETE',
-    3: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_WAVE_INCOMPLETE',
-    4: 'ROCPROFILER_THREAD_TRACE_DECODER_INFO_LAST',
-}
-ROCPROFILER_THREAD_TRACE_DECODER_INFO_NONE = 0
-ROCPROFILER_THREAD_TRACE_DECODER_INFO_DATA_LOST = 1
-ROCPROFILER_THREAD_TRACE_DECODER_INFO_STITCH_INCOMPLETE = 2
-ROCPROFILER_THREAD_TRACE_DECODER_INFO_WAVE_INCOMPLETE = 3
-ROCPROFILER_THREAD_TRACE_DECODER_INFO_LAST = 4
-rocprofiler_thread_trace_decoder_info_t = ctypes.c_uint32 # enum
-class struct_rocprofiler_thread_trace_decoder_pc_t(Structure):
-    pass
-
-struct_rocprofiler_thread_trace_decoder_pc_t._pack_ = 1 # source:False
-struct_rocprofiler_thread_trace_decoder_pc_t._fields_ = [
-    ('address', ctypes.c_uint64),
-    ('code_object_id', ctypes.c_uint64),
-]
-
-rocprofiler_thread_trace_decoder_pc_t = struct_rocprofiler_thread_trace_decoder_pc_t
-class struct_rocprofiler_thread_trace_decoder_perfevent_t(Structure):
-    pass
-
-struct_rocprofiler_thread_trace_decoder_perfevent_t._pack_ = 1 # source:False
-struct_rocprofiler_thread_trace_decoder_perfevent_t._fields_ = [
-    ('time', ctypes.c_int64),
-    ('events0', ctypes.c_uint16),
-    ('events1', ctypes.c_uint16),
-    ('events2', ctypes.c_uint16),
-    ('events3', ctypes.c_uint16),
-    ('CU', ctypes.c_ubyte),
-    ('bank', ctypes.c_ubyte),
-    ('PADDING_0', ctypes.c_ubyte * 6),
-]
-
-rocprofiler_thread_trace_decoder_perfevent_t = struct_rocprofiler_thread_trace_decoder_perfevent_t
-class struct_rocprofiler_thread_trace_decoder_occupancy_t(Structure):
-    pass
-
-struct_rocprofiler_thread_trace_decoder_occupancy_t._pack_ = 1 # source:False
-struct_rocprofiler_thread_trace_decoder_occupancy_t._fields_ = [
-    ('pc', rocprofiler_thread_trace_decoder_pc_t),
-    ('time', ctypes.c_uint64),
-    ('reserved', ctypes.c_ubyte),
-    ('cu', ctypes.c_ubyte),
-    ('simd', ctypes.c_ubyte),
-    ('wave_id', ctypes.c_ubyte),
-    ('start', ctypes.c_uint32, 1),
-    ('_rsvd', ctypes.c_uint32, 31),
-]
-
-rocprofiler_thread_trace_decoder_occupancy_t = struct_rocprofiler_thread_trace_decoder_occupancy_t
-
-# values for enumeration 'rocprofiler_thread_trace_decoder_wstate_type_t'
-rocprofiler_thread_trace_decoder_wstate_type_t__enumvalues = {
-    0: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EMPTY',
-    1: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_IDLE',
-    2: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EXEC',
-    3: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_WAIT',
-    4: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_STALL',
-    5: 'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_LAST',
-}
-ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EMPTY = 0
-ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_IDLE = 1
-ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EXEC = 2
-ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_WAIT = 3
-ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_STALL = 4
-ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_LAST = 5
-rocprofiler_thread_trace_decoder_wstate_type_t = ctypes.c_uint32 # enum
-class struct_rocprofiler_thread_trace_decoder_wave_state_t(Structure):
-    pass
-
-struct_rocprofiler_thread_trace_decoder_wave_state_t._pack_ = 1 # source:False
-struct_rocprofiler_thread_trace_decoder_wave_state_t._fields_ = [
-    ('type', ctypes.c_int32),
-    ('duration', ctypes.c_int32),
-]
-
-rocprofiler_thread_trace_decoder_wave_state_t = struct_rocprofiler_thread_trace_decoder_wave_state_t
-
-# values for enumeration 'rocprofiler_thread_trace_decoder_inst_category_t'
-rocprofiler_thread_trace_decoder_inst_category_t__enumvalues = {
-    0: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_NONE',
-    1: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_SMEM',
-    2: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_SALU',
-    3: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_VMEM',
-    4: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_FLAT',
-    5: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_LDS',
-    6: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_VALU',
-    7: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_JUMP',
-    8: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_NEXT',
-    9: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_IMMED',
-    10: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_CONTEXT',
-    11: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_MESSAGE',
-    12: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_BVH',
-    13: 'ROCPROFILER_THREAD_TRACE_DECODER_INST_LAST',
-}
-ROCPROFILER_THREAD_TRACE_DECODER_INST_NONE = 0
-ROCPROFILER_THREAD_TRACE_DECODER_INST_SMEM = 1
-ROCPROFILER_THREAD_TRACE_DECODER_INST_SALU = 2
-ROCPROFILER_THREAD_TRACE_DECODER_INST_VMEM = 3
-ROCPROFILER_THREAD_TRACE_DECODER_INST_FLAT = 4
-ROCPROFILER_THREAD_TRACE_DECODER_INST_LDS = 5
-ROCPROFILER_THREAD_TRACE_DECODER_INST_VALU = 6
-ROCPROFILER_THREAD_TRACE_DECODER_INST_JUMP = 7
-ROCPROFILER_THREAD_TRACE_DECODER_INST_NEXT = 8
-ROCPROFILER_THREAD_TRACE_DECODER_INST_IMMED = 9
-ROCPROFILER_THREAD_TRACE_DECODER_INST_CONTEXT = 10
-ROCPROFILER_THREAD_TRACE_DECODER_INST_MESSAGE = 11
-ROCPROFILER_THREAD_TRACE_DECODER_INST_BVH = 12
-ROCPROFILER_THREAD_TRACE_DECODER_INST_LAST = 13
-rocprofiler_thread_trace_decoder_inst_category_t = ctypes.c_uint32 # enum
-class struct_rocprofiler_thread_trace_decoder_inst_t(Structure):
-    pass
-
-struct_rocprofiler_thread_trace_decoder_inst_t._pack_ = 1 # source:False
-struct_rocprofiler_thread_trace_decoder_inst_t._fields_ = [
-    ('category', ctypes.c_uint32, 8),
-    ('stall', ctypes.c_uint32, 24),
-    ('duration', ctypes.c_int32),
-    ('time', ctypes.c_int64),
-    ('pc', rocprofiler_thread_trace_decoder_pc_t),
-]
-
-rocprofiler_thread_trace_decoder_inst_t = struct_rocprofiler_thread_trace_decoder_inst_t
-class struct_rocprofiler_thread_trace_decoder_wave_t(Structure):
-    pass
-
-struct_rocprofiler_thread_trace_decoder_wave_t._pack_ = 1 # source:False
-struct_rocprofiler_thread_trace_decoder_wave_t._fields_ = [
-    ('cu', ctypes.c_ubyte),
-    ('simd', ctypes.c_ubyte),
-    ('wave_id', ctypes.c_ubyte),
-    ('contexts', ctypes.c_ubyte),
-    ('_rsvd1', ctypes.c_uint32),
-    ('_rsvd2', ctypes.c_uint32),
-    ('_rsvd3', ctypes.c_uint32),
-    ('begin_time', ctypes.c_int64),
-    ('end_time', ctypes.c_int64),
-    ('timeline_size', ctypes.c_uint64),
-    ('instructions_size', ctypes.c_uint64),
-    ('timeline_array', ctypes.POINTER(struct_rocprofiler_thread_trace_decoder_wave_state_t)),
-    ('instructions_array', ctypes.POINTER(struct_rocprofiler_thread_trace_decoder_inst_t)),
-]
-
-rocprofiler_thread_trace_decoder_wave_t = struct_rocprofiler_thread_trace_decoder_wave_t
-class struct_rocprofiler_thread_trace_decoder_realtime_t(Structure):
-    pass
-
-struct_rocprofiler_thread_trace_decoder_realtime_t._pack_ = 1 # source:False
-struct_rocprofiler_thread_trace_decoder_realtime_t._fields_ = [
-    ('shader_clock', ctypes.c_int64),
-    ('realtime_clock', ctypes.c_uint64),
-    ('reserved', ctypes.c_uint64),
-]
-
-rocprofiler_thread_trace_decoder_realtime_t = struct_rocprofiler_thread_trace_decoder_realtime_t
-
-# values for enumeration 'rocprofiler_thread_trace_decoder_shaderdata_flags_t'
-rocprofiler_thread_trace_decoder_shaderdata_flags_t__enumvalues = {
-    0: 'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_IMM',
-    1: 'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_PRIV',
-}
-ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_IMM = 0
-ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_PRIV = 1
-rocprofiler_thread_trace_decoder_shaderdata_flags_t = ctypes.c_uint32 # enum
-class struct_rocprofiler_thread_trace_decoder_shaderdata_t(Structure):
-    pass
-
-struct_rocprofiler_thread_trace_decoder_shaderdata_t._pack_ = 1 # source:False
-struct_rocprofiler_thread_trace_decoder_shaderdata_t._fields_ = [
-    ('time', ctypes.c_int64),
-    ('value', ctypes.c_uint64),
-    ('cu', ctypes.c_ubyte),
-    ('simd', ctypes.c_ubyte),
-    ('wave_id', ctypes.c_ubyte),
-    ('flags', ctypes.c_ubyte),
-    ('reserved', ctypes.c_uint32),
-]
-
-rocprofiler_thread_trace_decoder_shaderdata_t = struct_rocprofiler_thread_trace_decoder_shaderdata_t
-
-# values for enumeration 'rocprofiler_thread_trace_decoder_record_type_t'
-rocprofiler_thread_trace_decoder_record_type_t__enumvalues = {
-    0: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_GFXIP',
-    1: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY',
-    2: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_PERFEVENT',
-    3: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE',
-    4: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_INFO',
-    5: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_DEBUG',
-    6: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_SHADERDATA',
-    7: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_REALTIME',
-    8: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_RT_FREQUENCY',
-    9: 'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_LAST',
-}
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_GFXIP = 0
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY = 1
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_PERFEVENT = 2
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE = 3
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_INFO = 4
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_DEBUG = 5
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_SHADERDATA = 6
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_REALTIME = 7
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_RT_FREQUENCY = 8
-ROCPROFILER_THREAD_TRACE_DECODER_RECORD_LAST = 9
-rocprofiler_thread_trace_decoder_record_type_t = ctypes.c_uint32 # enum
-
-# values for enumeration 'c__EA_rocprofiler_thread_trace_decoder_status_t'
-c__EA_rocprofiler_thread_trace_decoder_status_t__enumvalues = {
-    0: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS',
-    1: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR',
-    2: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES',
-    3: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_ARGUMENT',
-    4: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_SHADER_DATA',
-    5: 'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_LAST',
-}
-ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS = 0
-ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR = 1
-ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES = 2
-ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_ARGUMENT = 3
-ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_SHADER_DATA = 4
-ROCPROFILER_THREAD_TRACE_DECODER_STATUS_LAST = 5
-c__EA_rocprofiler_thread_trace_decoder_status_t = ctypes.c_uint32 # enum
-rocprofiler_thread_trace_decoder_status_t = c__EA_rocprofiler_thread_trace_decoder_status_t
-rocprofiler_thread_trace_decoder_status_t__enumvalues = c__EA_rocprofiler_thread_trace_decoder_status_t__enumvalues
-rocprof_trace_decoder_trace_callback_t = ctypes.CFUNCTYPE(c__EA_rocprofiler_thread_trace_decoder_status_t, rocprofiler_thread_trace_decoder_record_type_t, ctypes.POINTER(None), ctypes.c_uint64, ctypes.POINTER(None))
-rocprof_trace_decoder_isa_callback_t = ctypes.CFUNCTYPE(c__EA_rocprofiler_thread_trace_decoder_status_t, ctypes.POINTER(ctypes.c_char), ctypes.POINTER(ctypes.c_uint64), ctypes.POINTER(ctypes.c_uint64), struct_rocprofiler_thread_trace_decoder_pc_t, ctypes.POINTER(None))
-rocprof_trace_decoder_se_data_callback_t = ctypes.CFUNCTYPE(ctypes.c_uint64, ctypes.POINTER(ctypes.POINTER(ctypes.c_ubyte)), ctypes.POINTER(ctypes.c_uint64), ctypes.POINTER(None))
-try:
-    rocprof_trace_decoder_parse_data = _libraries['FIXME_STUB'].rocprof_trace_decoder_parse_data
-    rocprof_trace_decoder_parse_data.restype = rocprofiler_thread_trace_decoder_status_t
-    rocprof_trace_decoder_parse_data.argtypes = [rocprof_trace_decoder_se_data_callback_t, rocprof_trace_decoder_trace_callback_t, rocprof_trace_decoder_isa_callback_t, ctypes.POINTER(None)]
-except AttributeError:
-    pass
-try:
-    rocprof_trace_decoder_get_info_string = _libraries['FIXME_STUB'].rocprof_trace_decoder_get_info_string
-    rocprof_trace_decoder_get_info_string.restype = ctypes.POINTER(ctypes.c_char)
-    rocprof_trace_decoder_get_info_string.argtypes = [rocprofiler_thread_trace_decoder_info_t]
-except AttributeError:
-    pass
-try:
-    rocprof_trace_decoder_get_status_string = _libraries['FIXME_STUB'].rocprof_trace_decoder_get_status_string
-    rocprof_trace_decoder_get_status_string.restype = ctypes.POINTER(ctypes.c_char)
-    rocprof_trace_decoder_get_status_string.argtypes = [rocprofiler_thread_trace_decoder_status_t]
-except AttributeError:
-    pass
-rocprofiler_thread_trace_decoder_debug_callback_t = ctypes.CFUNCTYPE(None, ctypes.c_int64, ctypes.POINTER(ctypes.c_char), ctypes.POINTER(ctypes.c_char), ctypes.POINTER(None))
-uint64_t = ctypes.c_uint64
-try:
-    rocprof_trace_decoder_dump_data = _libraries['FIXME_STUB'].rocprof_trace_decoder_dump_data
-    rocprof_trace_decoder_dump_data.restype = rocprofiler_thread_trace_decoder_status_t
-    rocprof_trace_decoder_dump_data.argtypes = [ctypes.POINTER(ctypes.c_char), uint64_t, rocprofiler_thread_trace_decoder_debug_callback_t, ctypes.POINTER(None)]
-except AttributeError:
-    pass
-class union_rocprof_trace_decoder_gfx9_header_t(Union):
-    pass
-
-class struct_rocprof_trace_decoder_gfx9_header_t_0(Structure):
-    pass
-
-struct_rocprof_trace_decoder_gfx9_header_t_0._pack_ = 1 # source:False
-struct_rocprof_trace_decoder_gfx9_header_t_0._fields_ = [
-    ('legacy_version', ctypes.c_uint64, 13),
-    ('gfx9_version2', ctypes.c_uint64, 3),
-    ('DSIMDM', ctypes.c_uint64, 4),
-    ('DCU', ctypes.c_uint64, 5),
-    ('reserved1', ctypes.c_uint64, 1),
-    ('SEID', ctypes.c_uint64, 6),
-    ('reserved2', ctypes.c_uint64, 32),
-]
-
-union_rocprof_trace_decoder_gfx9_header_t._pack_ = 1 # source:False
-union_rocprof_trace_decoder_gfx9_header_t._anonymous_ = ('_0',)
-union_rocprof_trace_decoder_gfx9_header_t._fields_ = [
-    ('_0', struct_rocprof_trace_decoder_gfx9_header_t_0),
-    ('raw', ctypes.c_uint64),
-]
-
-rocprof_trace_decoder_gfx9_header_t = union_rocprof_trace_decoder_gfx9_header_t
-class union_rocprof_trace_decoder_instrument_enable_t(Union):
-    pass
-
-class struct_rocprof_trace_decoder_instrument_enable_t_0(Structure):
-    pass
-
-struct_rocprof_trace_decoder_instrument_enable_t_0._pack_ = 1 # source:False
-struct_rocprof_trace_decoder_instrument_enable_t_0._fields_ = [
-    ('char1', ctypes.c_uint32, 8),
-    ('char2', ctypes.c_uint32, 8),
-    ('char3', ctypes.c_uint32, 8),
-    ('char4', ctypes.c_uint32, 8),
-]
-
-union_rocprof_trace_decoder_instrument_enable_t._pack_ = 1 # source:False
-union_rocprof_trace_decoder_instrument_enable_t._anonymous_ = ('_0',)
-union_rocprof_trace_decoder_instrument_enable_t._fields_ = [
-    ('_0', struct_rocprof_trace_decoder_instrument_enable_t_0),
-    ('u32All', ctypes.c_uint32),
-]
-
-rocprof_trace_decoder_instrument_enable_t = union_rocprof_trace_decoder_instrument_enable_t
-class union_rocprof_trace_decoder_packet_header_t(Union):
-    pass
-
-class struct_rocprof_trace_decoder_packet_header_t_0(Structure):
-    pass
-
-struct_rocprof_trace_decoder_packet_header_t_0._pack_ = 1 # source:False
-struct_rocprof_trace_decoder_packet_header_t_0._fields_ = [
-    ('opcode', ctypes.c_uint32, 8),
-    ('type', ctypes.c_uint32, 4),
-    ('data20', ctypes.c_uint32, 20),
-]
-
-union_rocprof_trace_decoder_packet_header_t._pack_ = 1 # source:False
-union_rocprof_trace_decoder_packet_header_t._anonymous_ = ('_0',)
-union_rocprof_trace_decoder_packet_header_t._fields_ = [
-    ('_0', struct_rocprof_trace_decoder_packet_header_t_0),
-    ('u32All', ctypes.c_uint32),
-]
-
-rocprof_trace_decoder_packet_header_t = union_rocprof_trace_decoder_packet_header_t
-
-# values for enumeration 'rocprof_trace_decoder_packet_opcode_t'
-rocprof_trace_decoder_packet_opcode_t__enumvalues = {
-    4: 'ROCPROF_TRACE_DECODER_PACKET_OPCODE_CODEOBJ',
-    5: 'ROCPROF_TRACE_DECODER_PACKET_OPCODE_RT_TIMESTAMP',
-    6: 'ROCPROF_TRACE_DECODER_PACKET_OPCODE_AGENT_INFO',
-}
-ROCPROF_TRACE_DECODER_PACKET_OPCODE_CODEOBJ = 4
-ROCPROF_TRACE_DECODER_PACKET_OPCODE_RT_TIMESTAMP = 5
-ROCPROF_TRACE_DECODER_PACKET_OPCODE_AGENT_INFO = 6
-rocprof_trace_decoder_packet_opcode_t = ctypes.c_uint32 # enum
-
-# values for enumeration 'rocprof_trace_decoder_agent_info_type_t'
-rocprof_trace_decoder_agent_info_type_t__enumvalues = {
-    0: 'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_RT_FREQUENCY_KHZ',
-    1: 'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_COUNTER_INTERVAL',
-    2: 'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_LAST',
-}
-ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_RT_FREQUENCY_KHZ = 0
-ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_COUNTER_INTERVAL = 1
-ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_LAST = 2
-rocprof_trace_decoder_agent_info_type_t = ctypes.c_uint32 # enum
-class union_rocprof_trace_decoder_codeobj_marker_tail_t(Union):
-    pass
-
-class struct_rocprof_trace_decoder_codeobj_marker_tail_t_0(Structure):
-    pass
-
-struct_rocprof_trace_decoder_codeobj_marker_tail_t_0._pack_ = 1 # source:False
-struct_rocprof_trace_decoder_codeobj_marker_tail_t_0._fields_ = [
-    ('isUnload', ctypes.c_uint32, 1),
-    ('bFromStart', ctypes.c_uint32, 1),
-    ('legacy_id', ctypes.c_uint32, 30),
-]
-
-union_rocprof_trace_decoder_codeobj_marker_tail_t._pack_ = 1 # source:False
-union_rocprof_trace_decoder_codeobj_marker_tail_t._anonymous_ = ('_0',)
-union_rocprof_trace_decoder_codeobj_marker_tail_t._fields_ = [
-    ('_0', struct_rocprof_trace_decoder_codeobj_marker_tail_t_0),
-    ('raw', ctypes.c_uint32),
-]
-
-rocprof_trace_decoder_codeobj_marker_tail_t = union_rocprof_trace_decoder_codeobj_marker_tail_t
-
-# values for enumeration 'rocprof_trace_decoder_codeobj_marker_type_t'
-rocprof_trace_decoder_codeobj_marker_type_t__enumvalues = {
-    0: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_TAIL',
-    1: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_LO',
-    2: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_LO',
-    3: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_HI',
-    4: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_HI',
-    5: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_LO',
-    6: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_HI',
-    7: 'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_LAST',
-}
-ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_TAIL = 0
-ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_LO = 1
-ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_LO = 2
-ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_HI = 3
-ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_HI = 4
-ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_LO = 5
-ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_HI = 6
-ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_LAST = 7
-rocprof_trace_decoder_codeobj_marker_type_t = ctypes.c_uint32 # enum
-__all__ = \
-    ['ROCPROFILER_THREAD_TRACE_DECODER_INFO_DATA_LOST',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_LAST',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_NONE',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_STITCH_INCOMPLETE',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INFO_WAVE_INCOMPLETE',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_BVH',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_CONTEXT',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_FLAT',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_IMMED',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_JUMP',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_LAST',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_LDS',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_MESSAGE',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_NEXT',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_NONE',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_SALU',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_SMEM',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_VALU',
-    'ROCPROFILER_THREAD_TRACE_DECODER_INST_VMEM',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_DEBUG',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_GFXIP',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_INFO',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_LAST',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_OCCUPANCY',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_PERFEVENT',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_REALTIME',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_RT_FREQUENCY',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_SHADERDATA',
-    'ROCPROFILER_THREAD_TRACE_DECODER_RECORD_WAVE',
-    'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_IMM',
-    'ROCPROFILER_THREAD_TRACE_DECODER_SHADERDATA_FLAGS_PRIV',
-    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR',
-    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_ARGUMENT',
-    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_INVALID_SHADER_DATA',
-    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_ERROR_OUT_OF_RESOURCES',
-    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_LAST',
-    'ROCPROFILER_THREAD_TRACE_DECODER_STATUS_SUCCESS',
-    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EMPTY',
-    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_EXEC',
-    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_IDLE',
-    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_LAST',
-    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_STALL',
-    'ROCPROFILER_THREAD_TRACE_DECODER_WSTATE_WAIT',
-    'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_COUNTER_INTERVAL',
-    'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_LAST',
-    'ROCPROF_TRACE_DECODER_AGENT_INFO_TYPE_RT_FREQUENCY_KHZ',
-    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_HI',
-    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ADDR_LO',
-    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_HI',
-    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_ID_LO',
-    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_LAST',
-    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_HI',
-    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_SIZE_LO',
-    'ROCPROF_TRACE_DECODER_CODEOBJ_MARKER_TYPE_TAIL',
-    'ROCPROF_TRACE_DECODER_PACKET_OPCODE_AGENT_INFO',
-    'ROCPROF_TRACE_DECODER_PACKET_OPCODE_CODEOBJ',
-    'ROCPROF_TRACE_DECODER_PACKET_OPCODE_RT_TIMESTAMP',
-    'c__EA_rocprofiler_thread_trace_decoder_status_t',
-    'rocprof_trace_decoder_agent_info_type_t',
-    'rocprof_trace_decoder_codeobj_marker_tail_t',
-    'rocprof_trace_decoder_codeobj_marker_type_t',
-    'rocprof_trace_decoder_dump_data',
-    'rocprof_trace_decoder_get_info_string',
-    'rocprof_trace_decoder_get_status_string',
-    'rocprof_trace_decoder_gfx9_header_t',
-    'rocprof_trace_decoder_instrument_enable_t',
-    'rocprof_trace_decoder_isa_callback_t',
-    'rocprof_trace_decoder_packet_header_t',
-    'rocprof_trace_decoder_packet_opcode_t',
-    'rocprof_trace_decoder_parse_data',
-    'rocprof_trace_decoder_se_data_callback_t',
-    'rocprof_trace_decoder_trace_callback_t',
-    'rocprofiler_thread_trace_decoder_debug_callback_t',
-    'rocprofiler_thread_trace_decoder_info_t',
-    'rocprofiler_thread_trace_decoder_inst_category_t',
-    'rocprofiler_thread_trace_decoder_inst_t',
-    'rocprofiler_thread_trace_decoder_occupancy_t',
-    'rocprofiler_thread_trace_decoder_pc_t',
-    'rocprofiler_thread_trace_decoder_perfevent_t',
-    'rocprofiler_thread_trace_decoder_realtime_t',
-    'rocprofiler_thread_trace_decoder_record_type_t',
-    'rocprofiler_thread_trace_decoder_shaderdata_flags_t',
-    'rocprofiler_thread_trace_decoder_shaderdata_t',
-    'rocprofiler_thread_trace_decoder_status_t',
-    'rocprofiler_thread_trace_decoder_status_t__enumvalues',
-    'rocprofiler_thread_trace_decoder_wave_state_t',
-    'rocprofiler_thread_trace_decoder_wave_t',
-    'rocprofiler_thread_trace_decoder_wstate_type_t',
-    'struct_rocprof_trace_decoder_codeobj_marker_tail_t_0',
-    'struct_rocprof_trace_decoder_gfx9_header_t_0',
-    'struct_rocprof_trace_decoder_instrument_enable_t_0',
-    'struct_rocprof_trace_decoder_packet_header_t_0',
-    'struct_rocprofiler_thread_trace_decoder_inst_t',
-    'struct_rocprofiler_thread_trace_decoder_occupancy_t',
-    'struct_rocprofiler_thread_trace_decoder_pc_t',
-    'struct_rocprofiler_thread_trace_decoder_perfevent_t',
-    'struct_rocprofiler_thread_trace_decoder_realtime_t',
-    'struct_rocprofiler_thread_trace_decoder_shaderdata_t',
-    'struct_rocprofiler_thread_trace_decoder_wave_state_t',
-    'struct_rocprofiler_thread_trace_decoder_wave_t', 'uint64_t',
-    'union_rocprof_trace_decoder_codeobj_marker_tail_t',
-    'union_rocprof_trace_decoder_gfx9_header_t',
-    'union_rocprof_trace_decoder_instrument_enable_t',
-    'union_rocprof_trace_decoder_packet_header_t']
@@ -43,7 +43,6 @@ enum sqtt_version
   SQTT_VERSION_2_3 = 0x6, /* GFX9 */
   SQTT_VERSION_2_4 = 0x7, /* GFX10+ */
   SQTT_VERSION_3_2 = 0xb, /* GFX11+ */
-   SQTT_VERSION_3_3 = 0xc, /* GFX12+ */
 };

 enum sqtt_file_chunk_type
@@ -145,8 +144,6 @@ enum sqtt_gfxip_level
   SQTT_GFXIP_LEVEL_GFXIP_10_1 = 0x7,
   SQTT_GFXIP_LEVEL_GFXIP_10_3 = 0x9,
   SQTT_GFXIP_LEVEL_GFXIP_11_0 = 0xc,
-   SQTT_GFXIP_LEVEL_GFXIP_11_5 = 0xd,
-   SQTT_GFXIP_LEVEL_GFXIP_12 = 0x10,
 };

 enum sqtt_memory_type
@@ -430,8 +427,6 @@ enum elf_gfxip_level
   EF_AMDGPU_MACH_AMDGCN_GFX1010 = 0x033,
   EF_AMDGPU_MACH_AMDGCN_GFX1030 = 0x036,
   EF_AMDGPU_MACH_AMDGCN_GFX1100 = 0x041,
-   EF_AMDGPU_MACH_AMDGCN_GFX1150 = 0x043,
-   EF_AMDGPU_MACH_AMDGCN_GFX1200 = 0x04e,
 };

 struct sqtt_file_chunk_spm_db {
@@ -0,0 +1,40 @@
+import time
+from extra.optimization.helpers import load_worlds, ast_str_to_ast
+from tinygrad import Device
+from tinygrad.codegen.lowerer import pm_lowerer, get_index
+from tinygrad.uop.ops import graph_rewrite
+from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.codegen.opt.postrange import Scheduler
+from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.helpers import getenv
+
+if __name__ == "__main__":
+  renderer = Device.default.renderer
+  ast_strs = load_worlds()
+  if (n:=getenv("N", -1)) != -1: ast_strs = ast_strs[n:n+1]
+  good = 0
+  for i, ast_str in enumerate(ast_strs):
+    ast = ast_str_to_ast(ast_str)
+
+    st = time.perf_counter()
+    lin = Kernel(ast, renderer)
+    opt1 = hand_coded_optimizations(lin)
+    et_lin = time.perf_counter() - st
+
+    lowered = graph_rewrite(ast, pm_lowerer, ctx=get_index(ast), bottom_up=True)
+    st = time.perf_counter()
+    sch = Scheduler(lowered, renderer)
+    sch.convert_loop_to_global()
+    sch.simplify_merge_adjacent()
+    opt2 = hand_coded_optimizations(sch)
+    et_sch = time.perf_counter() - st
+
+    if opt1 != opt2:
+      print(f"******* {i:6d}")
+      print("Kernel:    ", lin.colored_shape(), "->", lin.apply_opts(opt1).colored_shape())
+      print("Scheduler: ", sch.colored_shape(), "->", sch.apply_opts(opt2).colored_shape())
+      print(opt1)
+      print(opt2)
+    else:
+      good += 1
+      print(f"******* {i:6d} MATCH {good/(i+1)*100:.2f}% -- {et_lin/et_sch:4.2f}x speedup")
--- a/Show More
+++ b/Show More