simpler, but not inverse

fix bitcast
enforce a cast for bitcast
2026-06-25 10:24:01 -07:00 · 2026-06-25 10:11:13 -07:00 · 2026-06-25 10:00:14 -07:00 · 2026-06-25 09:49:43 -07:00 · 2026-06-25 09:29:32 -07:00 · 2026-06-25 08:56:34 -07:00
52 changed files with 1378 additions and 1290 deletions
@@ -81,85 +81,181 @@ jobs:
  #      source /tmp/tinygrad_pytest_ci/bin/activate
  #      pytest -nauto --durations=20

-  testmacbenchmark:
-    name: Mac Benchmark
-    runs-on: [self-hosted, macOS]
+  sharedbenchmarks:
+    name: Benchmark (DEV=${{ matrix.dev }})
+    runs-on: [self-hosted, "${{ matrix.dev == 'METAL' && 'macOS' || matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['METAL', 'AMD', 'NV']
    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
      uses: actions/checkout@v6
-    - name: Symlink models and datasets
+    - name: Setup (AMD)
+      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        mkdir -p weights
-        mkdir -p extra/disassemblers
-        ln -s ~/tinygrad/extra/disassemblers/applegpu extra/disassemblers/applegpu
-        ln -s ~/tinygrad/weights/sd-v1-4.ckpt weights/sd-v1-4.ckpt
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
+        ./extra/amdpci/setup_python_cap.sh
+        ./extra/hcq/hcq_smi.py amd rmmod
+        ./extra/hcq/hcq_smi.py amd kill_pids
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
        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
-    - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing
-    - 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
-    # 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
-    - name: Run model inference benchmark
-      run: DEV=METAL NOCLANG=1 python3.11 test/external/external_model_benchmark.py
-    - name: Test speed vs torch
-      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py
+      run: python3 test/external/process_replay/reset.py
    - name: Test tensor cores
-      run: DEV=METAL python3.11 test/opt/test_tensor_cores.py
-    - name: Run Tensor Core GEMM (float)
-      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (half)
-      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (bfloat16)
-      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py
-    - name: Fuzz Padded Tensor Core GEMM
-      run: DEV=METAL M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
+      run: |
+        if [[ "${{ matrix.dev }}" == "METAL" ]]; then
+          python3 test/opt/test_tensor_cores.py
+          DEBUG=2 SHOULD_USE_TC=1 python3 extra/gemm/simple_matmul.py
+          DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3 extra/gemm/simple_matmul.py
+          DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3 extra/gemm/simple_matmul.py
+          M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
+        elif [[ "${{ matrix.dev }}" == "NV" ]]; then
+          ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+          DEV=NV:PTX ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+          DEV=CUDA SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          DEV=CUDA SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          DEV=CUDA SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+          DEV=CUDA SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          DEV=NV:PTX SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          # TODO: too slow
+          # 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
+          # DEV=NV:PTX 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
+        else
+          python3 test/opt/test_tensor_cores.py
+          # TODO: this is flaky
+          # DEV=AMD:LLVM python3 test/opt/test_tensor_cores.py
+          SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+          # TODO: AMD compiler bug causes this to fail
+          # HSA=1 M_START=12 M_STOP=20 M_STEP=1 N_START=12 N_STOP=20 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
+        fi
+    - name: Run model inference benchmark
+      # TODO: unstable on AMD
+      if: ${{ matrix.dev != 'AMD' }}
+      run: CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
+    - name: Test speed vs torch
+      # TODO: unstable on AMD
+      if: ${{ matrix.dev != 'AMD' }}
+      env:
+        HALF: ${{ matrix.dev == 'NV' && '1' || '0' }}
+      run: CAPTURE_PROCESS_REPLAY=0 BIG=2 ${{ matrix.dev == 'METAL' && 'MPS=1' || 'TORCHCUDA=1' }} python3 test/speed/external_test_speed_v_torch.py
+    - name: Run Stable Diffusion
+      env:
+        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'METAL' && '720' || matrix.dev == 'AMD' && '550' || '0' }}
+      run: BENCHMARK_LOG=stable_diffusion python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+    - name: Run SDXL
+      if: ${{ matrix.dev != 'NV' }}
+      env:
+        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'METAL' && '5000' || matrix.dev == 'AMD' && '3200' || '2000' }}
+      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
    - name: Run llama3.2
-      run: BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
+      run: BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
+    - name: Run qwen3.5
+      # qwen3.5:35b-a3b doesn't fit on mac
+      if: ${{ matrix.dev != 'METAL' }}
+      run: BENCHMARK_LOG=qwen35_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m qwen3.5:35b-a3b --benchmark --warmup
    - name: Run olmoe
-      run: BENCHMARK_LOG=olmoe JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 -m tinygrad.llm -m olmoe --benchmark --warmup
+      # just metal for now
+      if: ${{ matrix.dev == 'METAL' }}
+      run: BENCHMARK_LOG=olmoe JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m olmoe --benchmark --warmup
    - name: Train MNIST
-      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py
-
-    # 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
-    #- 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
-
-    #- name: Run 10 CIFAR training steps w BF16
-    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py
-    # 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
+      run: time TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+    - name: Run 10 CIFAR training steps
+      env:
+        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'NV' && '130' || matrix.dev == 'AMD' && '200' || '3000' }}
+      run: BENCHMARK_LOG=cifar_10steps STEPS=10 python3 examples/hlb_cifar10.py
+    - name: Run 10 CIFAR training steps w HALF
+      env:
+        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'NV' && '120' || matrix.dev == 'AMD' && '230' || '3000' }}
+      run: BENCHMARK_LOG=cifar_10steps_half STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
    - uses: actions/upload-artifact@v7
+      if: ${{ matrix.dev != 'AMD' }}
      with:
-        name: Speed (Mac)
+        name: Speed (${{ matrix.dev }})
        path: |
          onnx_inference_speed.csv
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

+  tinyboxbenchmark:
+    name: Tinybox Benchmark (${{ matrix.dev }})
+    runs-on: [self-hosted, Linux, "${{ matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['AMD', 'NV']
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup (AMD)
+      if: ${{ matrix.dev == 'AMD' }}
+      run: |
+        ./extra/amdpci/setup_python_cap.sh
+        ./extra/hcq/hcq_smi.py amd rmmod
+        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        mkdir -p extra/datasets
+        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    - name: Test tiny
+      run: |
+        DEBUG=2 python -m pytest -rA test/test_tiny.py
+        if [[ "${{ matrix.dev }}" == "NV" ]]; then
+          DEBUG=2 DEV=CUDA python -m pytest -rA test/test_tiny.py
+        fi
+    - name: Test speed vs theoretical
+      run: IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+    - name: Test benchmark allreduce
+      if: ${{ matrix.dev == 'NV' }}
+      run: python test/external/external_benchmark_multitensor_allreduce.py
+    - name: HEVC Decode Benchmark
+      if: ${{ matrix.dev == 'NV' }}
+      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 PYTHONPATH=. python3 extra/hevc/decode.py
+    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
+      run: BENCHMARK_LOG=llama3_beam_4gpu 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
+    - name: Run full CIFAR training w 1 GPU
+      run: time BENCHMARK_LOG=cifar DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run full CIFAR training steps w 6 GPUS
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run MLPerf resnet eval on training data
+      run: time BENCHMARK_LOG=resnet_eval MODEL=resnet python3 examples/mlperf/model_eval.py
+    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+      run: BENCHMARK_LOG=resnet_10steps DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+      run: BENCHMARK_LOG=resnet_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf Bert training steps (6 gpu)
+      # TODO: remove BERT_LAYERS once scheduler is fast
+      run: BENCHMARK_LOG=bert_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
  testusbgpu:
    name: UsbGPU Benchmark
    runs-on: [self-hosted, macOS]
@@ -196,337 +292,6 @@ jobs:
    - name: UsbGPU (USB4/TB) tiny tests
      run: PYTHONPATH=. DEV=PCI+NV:NAK python3.11 test/test_tiny.py

-  testnvidiabenchmark:
-    name: tinybox green Benchmark
-    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Print nvidia-smi
-      run: nvidia-smi
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        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: Run model inference benchmark
-      run: DEV=NV CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
-    - name: Test speed vs torch
-      run: DEV=NV CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
-    - name: Test speed vs theoretical
-      run: DEV=NV IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
-    - name: Test benchmark allreduce
-      run: DEV=NV python test/external/external_benchmark_multitensor_allreduce.py
-    - name: Test tensor cores
-      run: |
-        DEV=NV ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-        DEV=NV:PTX ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-    - name: Run Tensor Core GEMM (CUDA)
-      run: |
-        DEV=CUDA SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        DEV=CUDA SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        DEV=CUDA SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-        DEV=CUDA SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (PTX)
-      run: DEV=NV:PTX SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (NV)
-      run: DEV=NV SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-    - name: Test DEV=NV
-      run: DEBUG=2 DEV=NV python -m pytest -rA test/test_tiny.py
-    - name: Test DEV=CUDA
-      run: DEBUG=2 DEV=CUDA python -m pytest -rA test/test_tiny.py
-    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion DEV=NV python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
-    # TODO: too slow
-    # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
-    - name: Run llama3.2
-      run: DEV=NV BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
-    - name: Run qwen3.5
-      run: DEV=NV BENCHMARK_LOG=qwen35_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 -m tinygrad.llm -m qwen3.5:35b-a3b --benchmark --warmup
-    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu DEV=NV 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
-    # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
-    # - name: Run LLaMA-2 70B
-    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
-    - uses: actions/upload-artifact@v7
-      with:
-        name: Speed (NVIDIA)
-        path: |
-          onnx_inference_speed.csv
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmorenvidiabenchmark:
-    name: tinybox green Training Benchmark
-    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        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: DEV=NV 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: DEV=NV:PTX 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: HEVC Decode Benchmark
-      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 DEV=NV PYTHONPATH=. python3 extra/hevc/decode.py
-    - name: Train MNIST
-      run: time PYTHONPATH=. DEV=NV TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
-    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=130 DEV=NV STEPS=10 python3 examples/hlb_cifar10.py
-    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=120 DEV=NV STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
-    - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 DEV=NV STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
-    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 DEV=NV WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
-    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar DEV=NV DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEV=NV DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run MLPerf resnet eval on training data
-      run: time BENCHMARK_LOG=resnet_eval DEV=NV MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps DEV=NV DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf Bert training steps (6 gpu)
-      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testamdbenchmark:
-    name: tinybox red Benchmark
-    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-    #- name: Insert amdgpu
-    #  run: sudo modprobe amdgpu
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        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: setup perflevel
-    #  run: |
-    #    examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
-    #    rocm-smi
-    #- name: Show off tinybox
-    #  run: /opt/rocm/bin/rocm-bandwidth-test
-    # TODO: unstable on AMD
-    #- name: Run model inference benchmark
-    #  run: LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 NOCLANG=1 python3 test/external/external_model_benchmark.py
-    # TODO: unstable on AMD
-    #- name: Test speed vs torch
-    #  run: |
-    #    python3 -c "import torch; print(torch.__version__)"
-    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
-    - name: Test speed vs theoretical
-      run: DEV=AMD IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
-    - name: Test tensor cores (no LLVM)
-      run: DEV=AMD python3 test/opt/test_tensor_cores.py
-    # TODO: this is flaky
-    # - name: Test tensor cores AMD:LLVM
-    #   run: DEV=AMD:LLVM python3 test/opt/test_tensor_cores.py
-    - name: Run Tensor Core GEMM (AMD)
-      run: |
-        DEV=AMD SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        DEV=AMD SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-    - name: Test DEV=AMD
-      run: DEBUG=2 DEV=AMD python -m pytest -rA test/test_tiny.py
-    #- name: Test HIP=1
-    #  run: DEBUG=2 HIP=1 python -m pytest -rA test/test_tiny.py
-    # TODO: AMD compiler bug causes this to fail
-    #- name: Fuzz Padded Tensor Core GEMM
-    #  run: HSA=1 M_START=12 M_STOP=20 M_STEP=1 N_START=12 N_STOP=20 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
-    #- name: Remove amdgpu
-    #  run: sleep 10 && sudo rmmod amdgpu # sleep a bit to let the driver unload the prev pid.
-    - name: Test AM cold start time
-      run: time DEV=AMD AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
-    - name: Test AM warm start time
-      run: time DEV=AMD python3 test/test_tiny.py TestTiny.test_plus
-    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 DEV=AMD python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
-    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 DEV=AMD python3 examples/sdxl.py --seed 0 --noshow --timing
-    - name: Run llama3.2
-      run: DEV=AMD BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
-    - name: Run qwen3.5
-      run: DEV=AMD BENCHMARK_LOG=qwen35_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 -m tinygrad.llm -m qwen3.5:35b-a3b --benchmark --warmup
-    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu DEV=AMD 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
-    # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
-    #- name: Restore amdgpu
-    #  run: sudo modprobe amdgpu
-    # - name: Run LLaMA-2 70B
-    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmoreamdbenchmark:
-    name: tinybox red Training Benchmark
-    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        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: Test GPU crash recovery
-      run: DEV=AMD python3 -m pytest -rA test/external/external_test_gpu_crash.py
-    - name: Train MNIST
-      run: time PYTHONPATH=. DEV=AMD TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
-    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 DEV=AMD STEPS=10 python3 examples/hlb_cifar10.py
-    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=230 DEV=AMD STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
-    # - name: Run 10 CIFAR training steps w BF16
-    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 DEV=AMD STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
-    # TODO: too slow
-    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 DEV=AMD WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
-    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar DEV=AMD DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu DEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    # TODO: broken on some of the machines
-    #- name: Test full tinyfs load
-    #  run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmlperfamdbenchmark:
-    name: tinybox red MLPerf Benchmark
-    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        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: Run MLPerf resnet eval
-      run: time BENCHMARK_LOG=resnet_eval DEV=AMD MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps DEV=AMD DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu DEV=AMD CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf Bert training steps (6 gpu)
-      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu DEV=AMD CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
  testcommalatest:
    name: comma Benchmark (0.11.0)
    runs-on: [self-hosted, Linux, comma]
@@ -648,31 +413,28 @@ jobs:
    - name: openpilot run_pickle 0.10.1 driving_vision
      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py

-  testreddriverbenchmark:
-    name: AM Benchmark
+  driverbenchmarks:
+    name: PCI Driver Benchmark (DEV=${{ matrix.dev }})
    runs-on: [self-hosted, Linux, tinyboxrandom]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['AMD', 'NV']
    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
      uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Symlink models and datasets
+    - name: Setup
      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
+        ./extra/amdpci/setup_python_cap.sh
+        ./extra/hcq/hcq_smi.py ${{ matrix.dev == 'AMD' && 'amd' || 'nv' }} rmmod
+        ./extra/hcq/hcq_smi.py ${{ matrix.dev == 'AMD' && 'amd' || 'nv' }} kill_pids
        mkdir -p extra/datasets
        ln -s /raid/datasets/imagenet extra/datasets/imagenet
    - name: setup staging db
@@ -683,104 +445,47 @@ jobs:
    - name: reset process replay
      run: test/external/process_replay/reset.py
    - name: Test driver cold start time
-      run: time DEBUG=3 DEV=AMD AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
+      run: time DEBUG=3 AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Test driver warm start time
-      run: time DEBUG=3 DEV=AMD python3 test/test_tiny.py TestTiny.test_plus
+      if: ${{ matrix.dev == 'AMD' }}
+      run: time DEBUG=3 python3 test/test_tiny.py TestTiny.test_plus
    - name: Test GPU crash recovery
-      run: DEV=AMD python3 -m pytest -rA test/external/external_test_gpu_crash.py
-    # Fails on 9070
-    # - name: Test tensor cores
-    #   run: |
-    #     DEV=AMD python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-    #     DEV=AMD:LLVM python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-    #     DEV=AMD SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (AMD)
-      run: DEV=AMD SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-    - name: Test DEV=AMD
-      run: DEBUG=2 DEV=AMD python -m pytest -rA test/test_tiny.py
+      if: ${{ matrix.dev == 'AMD' }}
+      run: python3 -m pytest -rA test/external/external_test_gpu_crash.py
+    - name: Test tensor cores
+      run: |
+        if [[ "${{ matrix.dev }}" == "AMD" ]]; then
+          # Fails on 9070
+          # python3 test/test_linearizer.py test/opt/test_tensor_cores.py
+          # DEV=AMD:LLVM python3 test/test_linearizer.py test/opt/test_tensor_cores.py
+          # SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+        else
+          ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+        fi
    - name: Test DISK copy time
-      run: DEV=AMD TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
+      run: TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
    - name: Test CPU copy time
      run: |
-        DEV=AMD GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
-        DEV=AMD GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
+        GRAPH_ONE_KERNEL=1 NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
+        GRAPH_ONE_KERNEL=1 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 DEV=AMD DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-    #   run: BENCHMARK_LOG=resnet_10steps DEV=AMD MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: time BENCHMARK_LOG=cifar DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+      if: ${{ matrix.dev == 'NV' }}
+      run: BENCHMARK_LOG=resnet_10steps MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
    - name: Run 10 MLPerf Bert training steps (1 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps DEV=AMD 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: BENCHMARK_LOG=bert_10steps CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
    - name: Remote
      run: |
        pkill -f 'extra/remote/serve.py' || true
        PYTHONPATH=. python3 extra/remote/serve.py 6482 &
        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 DEV=PCI+AMD python3 test/test_tiny.py
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 DEV=PCI+AMD AMD_AQL=1 python3 test/test_tiny.py
-        pkill -f 'extra/remote/serve.py' || true
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testgreendriverbenchmark:
-    name: NV Benchmark
-    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 20
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove nv modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        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: Test driver start time
-      run: time DEBUG=3 DEV=NV python3 test/test_tiny.py TestTiny.test_plus
-    - name: Test tensor cores
-      run: DEV=NV ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-    - name: Test DISK copy time
-      run: DEV=NV TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
-    - name: Test CPU copy time
-      run: |
-        DEV=NV GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
-        DEV=NV GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
-    - name: Test LLAMA-3
-      run: BENCHMARK_LOG=llama3_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0
-    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar DEV=NV DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps DEV=NV MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf Bert training steps (1 gpu)
-      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    - name: Remote
-      run: |
-        pkill -f 'extra/remote/serve.py' || true
-        PYTHONPATH=. python3 extra/remote/serve.py 6483 &
-        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6483 DEV=NV python3 test/test_tiny.py
+        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 python3 test/test_tiny.py
+        if [[ "${{ matrix.dev }}" == "AMD" ]]; then
+          DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD_AQL=1 python3 test/test_tiny.py
+        fi
        pkill -f 'extra/remote/serve.py' || true
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
@@ -327,7 +327,7 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1468 ALLOWED_GATED_READ_IMAGE=10 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=55 FLOAT16=1 DEV=CL IMAGE=1 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: |
          DEV=CL IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
@@ -370,7 +370,6 @@ jobs:
        with:
          key: optim
          deps: testing
-          pydeps: "tensorflow==2.19"
          opencl: 'true'
      #- name: Test Optimization Helpers
      #  run: DEBUG=1 python3 extra/optimization/test_helpers.py
@@ -379,7 +378,7 @@ jobs:
      - name: Test Beam Search
        run: DEV=CL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
      - name: Test MLPerf stuff
-        run: DEV=CL 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
+        run: DEV=CL python -m pytest -n=auto test/external/external_test_lr_schedule.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
      - name: DEV=NULL beautiful_mnist_multigpu
        run: DEV=NULL NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
      - name: Test Bert training
@@ -38,7 +38,7 @@ def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):

 def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_scale:Tensor|None=None,
           x_fp8:Tensor|None=None, x_new_amax:Tensor|None=None,
-           grad_amax_state:Tensor|None=None) -> tuple[Tensor,...]:
+           grad_amax_state:Tensor|None=None, x_prequant_mx:tuple|None=None) -> tuple[Tensor,...]:
  if not fp8:
    if ASM_GEMM:
      from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
@@ -47,12 +47,14 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
  assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
  if MXFP8:
    from extra.gemm.cdna_asm_gemm import asm_gemm, quantize_mxfp8, mx_pack, can_use_asm_gemm, _mx_block_scale
-    x_q, x_e8, x_si = quantize_mxfp8(x.reshape(-1, x.shape[-1]))
+    if x_prequant_mx is not None: x_q, x_e8, x_si = x_prequant_mx       # fused producer already quantized (2d)
+    else: x_q, x_e8, x_si = quantize_mxfp8(x.reshape(-1, x.shape[-1]))
+    l_shape = x.shape[:-1] if x is not None else x_q.shape[:-1]
    if can_use_asm_gemm(x_q, w.T):
      out = asm_gemm(x_q, w.T, mx=True, mx_scales=(x_si, x_e8, mx_pack(w_inv_scale), w_inv_scale),
-                     mx_w_stored=True).reshape(*x.shape[:-1], w.shape[0])
+                     mx_w_stored=True).reshape(*l_shape, w.shape[0])
    else:
-      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*x.shape[:-1], x.shape[-1])
+      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*l_shape, x_q.shape[-1])
      out = x_phys @ (w.cast(dtypes.bfloat16) * _mx_block_scale(w_inv_scale)).T
    return out, (amax_x.detach() if amax_x is not None else None), x_q
  if x_fp8 is None:
@@ -214,8 +216,15 @@ class FlatTransformer:
      x_w3, new_amax, *s = matmul(inp, kwargs["w3"], amax_x=kwargs["amax_x3"], w_inv_scale=kwargs["s_3"], grad_amax_state=kwargs["grad_amax_xw3"])
      amaxs.append(new_amax)
      saves.extend([*s, x_w3])
-      out, new_amax, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
-                                 grad_amax_state=kwargs["grad_amax_xout"])
+      if FUSED_SILU_W13 and MXFP8:
+        from extra.llama_kernels.fused_silu_mul_quantize_mxfp8 import fused_silu_mul_quantize_mxfp8
+        aq, ae8, asi = fused_silu_mul_quantize_mxfp8(x_w1.reshape(-1, x_w1.shape[-1]), x_w3.reshape(-1, x_w3.shape[-1]))
+        out, new_amax, *s = matmul(None, kwargs["w2"], x_prequant_mx=(aq, ae8, asi), amax_x=kwargs["amax_x2"],
+                                   w_inv_scale=kwargs["s_2"], grad_amax_state=kwargs["grad_amax_xout"])
+        out = out.reshape(*x_w1.shape[:-1], kwargs["w2"].shape[0])
+      else:
+        out, new_amax, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
+                                   grad_amax_state=kwargs["grad_amax_xout"])
      amaxs.append(new_amax)
      saves.extend([*s, out])
    else:
@@ -1,6 +1,6 @@
 #!/bin/bash
-export BENCHMARK=5
+export BENCHMARK=${BENCHMARK:-5}
 export EVAL_BS=0
 VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=${DEBUG:--0} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
 SRC="AMD"; [[ $DEV == NULL* ]] && SRC="NULL"
-python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"
+[ "$BENCHMARK" -le 3 ] || python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"
@@ -2899,7 +2899,7 @@ def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool, w_stored:bool=F

  g = Tensor(gradient, device=aq.device)[:aq.shape[0]].reshape(aq.shape[0]*aq.shape[1], bq.shape[0]).cast(dtypes.bfloat16)
  grad_a = asm_gemm(g, b_phys, mx=True)
-  grad_b = asm_gemm(g.T, a_phys, mx=True)
+  grad_b = asm_gemm(g.T, a_phys, mx=True, a_pretranspose=g)

  grad_a = (grad_a * _mx_block_scale(ae8)).reshape(aq.shape)
  if not w_stored: grad_b = grad_b * _mx_block_scale(be8)
@@ -2909,7 +2909,8 @@ def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool, w_stored:bool=F
 # ** main gemm function

 def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=None, grad_amax_state:Tensor|None=None,
-             w_post_scale:Tensor|None=None, mx:bool=False, mx_scales:tuple|None=None, mx_w_stored:bool=False, g_scale:Tensor|None=None) -> Tensor:
+             w_post_scale:Tensor|None=None, mx:bool=False, mx_scales:tuple|None=None, mx_w_stored:bool=False, g_scale:Tensor|None=None,
+             a_pretranspose:Tensor|None=None) -> Tensor:
  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
  counters["used"] += 1
  unfold_batch = a.ndim == 3 and isinstance(a.device, tuple) and a.uop.axis == 2 and b.uop.axis == 0
@@ -2946,6 +2947,11 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
      if mx_scales is not None:
        a_si, a_e8, b_si, b_e8 = mx_scales
        a_q, b_q = a.reshape(-1, a.shape[-1]), b.T
+      elif (a_pretranspose is not None and getenv("FUSED_GRAD_QUANTIZE", 0) and a_pretranspose.dtype == dtypes.bfloat16
+            and a_pretranspose.shape[0] % 32 == 0 and a_pretranspose.shape[1] % 256 == 0):
+        from extra.llama_kernels.transpose_quantize_mxfp8 import transpose_quantize_mxfp8
+        a_q, a_e8, a_si = transpose_quantize_mxfp8(a_pretranspose)
+        b_q, b_e8, b_si = quantize_mxfp8(b.T)
      else:
        a_q, a_e8, a_si = quantize_mxfp8(a.reshape(-1, a.shape[-1]))
        b_q, b_e8, b_si = quantize_mxfp8(b.T)
@@ -143,14 +143,17 @@ def make_getaddr(u, device=None):
 def make_ins(op, *srcs):
  return UOp(Ops.INS, dtypes.void, tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs), op)

+def make_patch(buf:UOp, off:sint, val:UOp, dtype=None) -> UOp:
+  dt = dtype or val.dtype
+  return UOp(Ops.SHRINK, buf.dtype.base, (buf, UOp.const(dtypes.int, off), UOp.const(dtypes.int, dt.itemsize))).bitcast(dt).store(val.cast(dt))
+
 def make_cmdbuf(lin, devs, tag):
  blob, patches = b'', []
  for s in (s for ins in lin.src for s in ins.src):
    if s.op is not Ops.CONST: patches.append((len(blob), s))
    blob += struct.pack(f'<{s.dtype.fmt}', s.arg if s.op is Ops.CONST else 0x0)
  buf = UOp.new_buffer(devs, len(blob), dtypes.uint8).rtag(tag)
-  stores = [buf.index(UOp.const(dtypes.int, off), dtype=buf.dtype.ptr()).cast(s.dtype.ptr()).store(s) for off, s in patches]
-  return buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=blob)), *stores)
+  return buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=blob)), *[make_patch(buf, off, s) for off, s in patches])

 def make_mstack(uops): return uops[0] if len(uops) == 1 else UOp(Ops.MSTACK, uops[0].dtype, tuple(uops))

@@ -211,15 +214,11 @@ def prep_program(call:UOp, prg:UOp) -> UOp|None:
  return prg.replace(src=(buf.after(buf.store(blob)),), arg=(data, prg.arg)).call(*call.src[1:], aux=HCQInfo.from_call(call))

 def prep_kernargs(call:UOp, prg:UOp) -> UOp:
-  data, info = prg.arg
-  patches = [(i*dtypes.uint64.itemsize, UOp(Ops.GETADDR, dtypes.uint64, src=(call.src[1+gi], UOp(Ops.DEVICE, arg=call.src[1+gi].device))),
-              dtypes.uint64) for i,gi in enumerate(info.globals)] \
-          + [(len(info.globals)*dtypes.uint64.itemsize + i*dtypes.uint32.itemsize, v, dtypes.uint32) for i,v in enumerate(info.vars)]
-
-  buf = UOp.new_buffer(call.src[1].device, data.kernargs_alloc_size, dtypes.uint8).rtag("kernargs")
-  kernargs = buf.after(*tuple(buf.index(UOp.const(dtypes.int, o), dtype=buf.dtype.ptr()).cast(dt.ptr()).store(val.cast(dt)) for o, val, dt in patches))
-
-  return call.replace(src=(prg.replace(src=prg.src + (kernargs,), arg=(data, info)),) + call.src[1:])
+  (data, info), dev_uop = prg.arg, UOp(Ops.DEVICE, arg=call.src[1].device)
+  buf = UOp.new_buffer(dev_uop.arg, data.kernargs_alloc_size, dtypes.uint8).rtag("kernargs")
+  patches = [make_patch(buf, i*8, UOp(Ops.GETADDR, dtypes.uint64, src=(call.src[1+gi], dev_uop))) for i,gi in enumerate(info.globals)] \
+          + [make_patch(buf, len(info.globals)*8 + i*4, v, dtypes.uint32) for i,v in enumerate(info.vars)]
+  return call.replace(src=(prg.replace(src=prg.src + (buf.after(*patches),), arg=(data, info)),) + call.src[1:])

 pm_prep_runtime = PatternMatcher([
  # bind generic PROGRAM device to the call's actual dev(s), then run device-specific lowering
@@ -532,9 +531,9 @@ pm_resolve_patches = PatternMatcher([
  (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat(Ops.CONST)], name="op"), push_stack),
  (UPat(Ops.CAST, src=(UPat(Ops.STACK, name="s"),), name="op"), push_stack),

-  # index on slice is index
-  (UPat(Ops.INDEX, src=(UPat(Ops.SLICE, name="bv"), UPat()), name="idx", allow_any_len=True),
-    lambda idx, bv: idx.replace(src=(bv.src[0], idx.src[1] + bv.src[1].cast(idx.src[1].dtype), *idx.src[2:]))),
+  # shrink on slice is shrink on base at offset
+  (UPat(Ops.SHRINK, src=(UPat(Ops.SLICE, name="bv"), UPat(), UPat()), name="shr"),
+    lambda shr, bv: shr.replace(src=(bv.src[0], shr.src[1] + bv.src[1].cast(shr.src[1].dtype), shr.src[2]))),

  # getaddr
  (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, name="bv"), UPat(Ops.DEVICE, name="dev"))), resolve_getaddr_slice), # getaddr(slice(x)) -> offset+getaddr(x)
@@ -542,8 +541,8 @@ pm_resolve_patches = PatternMatcher([

  # folders
  (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").store(UPat(Ops.BINARY, name="blob")), fold_blob_store),
-  (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").index(UPat.cvar("off")).or_casted().store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))),
-    fold_const_store),
+  (UPat(Ops.SHRINK, src=(UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf"), UPat.cvar("off"), UPat(Ops.CONST))).bitcast()
+    .store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))), fold_const_store),
 ]) + symbolic_simple

 # *****************
@@ -0,0 +1,104 @@
+import functools
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from extra.llama_kernels import FP8_MAX, THREADS_PER_WG, alloc_like
+
+BLK = 32
+PACK = 4
+LOG2E = 1.4426950408889634
+
+@functools.cache
+def _custom_silu_mul_quantize_mxfp8(fp8_out:UOp, e8_out:UOp, si_out:UOp, x_w1:UOp, x_w3:UOp) -> UOp:
+  rows, K = x_w1.shape
+  scale_K = K // BLK
+  n_elems = rows * K
+  n_super = n_elems // (BLK * PACK)
+  sk4 = scale_K // PACK
+  assert n_super % THREADS_PER_WG == 0, f"{n_super=} must divide over {THREADS_PER_WG=}"
+  nwg = n_super // THREADS_PER_WG
+
+  x_w1, x_w3 = x_w1.reshape(n_elems), x_w3.reshape(n_elems)
+  fp8_out = fp8_out.reshape(n_elems)
+  e8_out = e8_out.reshape(rows * scale_K)
+  si_out = si_out.reshape(sk4 * rows)
+
+  wg = UOp.range(nwg, 0, AxisType.GLOBAL)
+  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
+  sb = UOp.range(PACK, 2, AxisType.UNROLL)
+  lane = UOp.range(BLK, 3, AxisType.UNROLL)
+
+  super_idx = wg * THREADS_PER_WG + tid
+  idx = super_idx * (BLK * PACK) + sb * BLK + lane
+
+  w1 = x_w1[idx].cast(dtypes.float)
+  w3 = x_w3[idx].cast(dtypes.float)
+  sig = (1.0 + (w1 * -LOG2E).exp2()).reciprocal()
+  act = w1 * sig * w3
+  abs_a = (act < 0.0).where(-act, act)
+  blk_max = abs_a.reduce(lane, arg=Ops.MAX)
+  e8f = (blk_max.maximum(1e-38).log2().floor() + 127.0).maximum(0.0).minimum(254.0)
+  qscale = (127.0 - e8f).exp2()
+  scaled = (act * qscale).maximum(-FP8_MAX).minimum(FP8_MAX)
+  e8u8 = e8f.cast(dtypes.uint8)
+
+  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype.base)).end(lane)
+  e8_store = e8_out.after(fp8_store)[super_idx * PACK + sb].store(e8u8)
+  packed = (e8u8.cast(dtypes.uint32) << (sb.cast(dtypes.uint32) * 8)).reduce(sb, arg=Ops.ADD)
+  row, col4 = super_idx // sk4, super_idx % sk4
+  si_store = si_out.after(e8_store.end(sb))[col4 * rows + row].store(packed)
+  return si_store.end(tid, wg).sink(arg=KernelInfo(f"silu_mul_quantize_mxfp8_{n_elems}", opts_to_apply=()))
+
+@functools.cache
+def _custom_silu_mul_bwd_mxfp8(gx1_out:UOp, gx3_out:UOp, x_w1:UOp, x_w3:UOp, grad_aq:UOp, e8:UOp) -> UOp:
+  rows, K = x_w1.shape
+  scale_K = K // BLK
+  n_elems = rows * K
+  VEC = 8
+  assert n_elems % (THREADS_PER_WG * VEC) == 0, f"{n_elems=} must divide {THREADS_PER_WG*VEC=}"
+  nwg = n_elems // (THREADS_PER_WG * VEC)
+  x_w1, x_w3, grad_aq = x_w1.reshape(n_elems), x_w3.reshape(n_elems), grad_aq.reshape(n_elems)
+  gx1_out, gx3_out, e8 = gx1_out.reshape(n_elems), gx3_out.reshape(n_elems), e8.reshape(rows * scale_K)
+
+  wg = UOp.range(nwg, 0, AxisType.GLOBAL)
+  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
+  lane = UOp.range(VEC, 2, AxisType.UNROLL)
+  idx = (wg * THREADS_PER_WG + tid) * VEC + lane
+
+  e8v = e8[idx // BLK].cast(dtypes.float)
+  qscale = (127.0 - e8v).exp2()
+  ga = grad_aq[idx].cast(dtypes.float) * qscale
+  w1 = x_w1[idx].cast(dtypes.float)
+  w3 = x_w3[idx].cast(dtypes.float)
+  sig = (1.0 + (w1 * -LOG2E).exp2()).reciprocal()
+  s = w1 * sig
+  sprime = sig * (1.0 + w1 * (1.0 - sig))
+  gx1 = gx1_out[idx].store((ga * sprime * w3).cast(gx1_out.dtype.base))
+  gx3 = gx3_out.after(gx1)[idx].store((ga * s).cast(gx3_out.dtype.base))
+  return gx3.end(lane, tid, wg).sink(arg=KernelInfo(f"silu_mul_bwd_mxfp8_{n_elems}", opts_to_apply=()))
+
+def _silu_mul_quantize_mxfp8_bwd(gradient:UOp, kernel:UOp):
+  _, e8_out, _, x_w1, x_w3 = kernel.src[1:]
+  device = x_w1.device
+  rows, K = x_w1.shape
+  axis = x_w1.axis if isinstance(device, tuple) else None
+  gx1 = alloc_like((rows, K), dtypes.bfloat16, device, axis)
+  gx3 = alloc_like((rows, K), dtypes.bfloat16, device, axis)
+  gx1, gx3, *_ = Tensor.custom_kernel(gx1, gx3, Tensor(x_w1, device=device), Tensor(x_w3, device=device),
+                                      Tensor(gradient, device=device).cast(dtypes.bfloat16), Tensor(e8_out.after(kernel), device=device),
+                                      fxn=_custom_silu_mul_bwd_mxfp8)
+  return (None, None, None, gx1.uop, gx3.uop)
+
+def fused_silu_mul_quantize_mxfp8(x_w1:Tensor, x_w3:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  assert x_w1.shape == x_w3.shape, f"{x_w1.shape} != {x_w3.shape}"
+  assert x_w1.dtype == dtypes.bfloat16 and x_w3.dtype == dtypes.bfloat16
+  assert x_w1.ndim == 2, f"expected 2d, got {x_w1.shape}"
+  from extra.gemm.cdna_asm_gemm import FP8_DTYPE
+  rows, K = x_w1.shape
+  scale_K = K // BLK
+  axis = x_w1.uop.axis if isinstance(x_w1.device, tuple) else None
+  fp8_out = alloc_like((rows, K), FP8_DTYPE, x_w1.device, axis)
+  e8_out = alloc_like((rows, scale_K), dtypes.uint8, x_w1.device, axis)
+  si_out = alloc_like((scale_K // PACK, rows), dtypes.uint32, x_w1.device, None if axis is None else (1 if axis == 0 else 0))
+  fp8_out, e8_out, si_out, *_ = Tensor.custom_kernel(fp8_out, e8_out, si_out, x_w1, x_w3,
+                                                     fxn=_custom_silu_mul_quantize_mxfp8, grad_fxn=_silu_mul_quantize_mxfp8_bwd)
+  return fp8_out, e8_out, si_out
@@ -42,8 +42,8 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_partial:UOp, x:UOp, amax_st
  step = THREADS_PER_WG // 2
  while step:
    active = tid < step
-    other = lds[tid + step].load(UOp.const(dtypes.float, 0.0), active)
-    lds = lds.after(lds[tid].store(lds[tid].maximum(other), gate=active).barrier())
+    other = lds[(tid + step).valid(active)].load()
+    lds = lds.after(lds[tid.valid(active)].store(lds[tid].maximum(other)).barrier())
    step //= 2

  amax_store = amax_partial[tid.eq(0).where(wg, UOp.invalid())].store(lds[0])
@@ -0,0 +1,37 @@
+from __future__ import annotations
+import functools, pathlib
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.renderer import Estimates
+from extra.llama_kernels import THREADS_PER_WG, alloc_like, dname_of, compile_hip
+
+TILE_N = THREADS_PER_WG   # 256
+BLK = 32
+
+@functools.cache
+def _custom_transpose_quantize_mxfp8(q:UOp, e8:UOp, g:UOp, dname:str) -> UOp:
+  M, N = g.shape
+  num_wg = (M // BLK) * (N // TILE_N)
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(num_wg, "gidx0")
+  mem = M * N * 2 + M * N + (M // BLK) * N   # read bf16, write fp8 + e8
+  sink = UOp.sink(q.base, e8.base, g.base, threads, workgroups,
+                  arg=KernelInfo(f"transpose_quantize_mxfp8_{M}_{N}", estimates=Estimates(ops=M*N, mem=mem)))
+  src = (pathlib.Path(__file__).parent/"transpose_quantize_mxfp8.cpp").read_text()
+  defines = [f"-DM_DIM={M}", f"-DN_DIM={N}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+def transpose_quantize_mxfp8(g:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  # fused g.T quantize: returns (q, e8, si) == quantize_mxfp8(g.T) — q (N,M) fp8, e8 (N, M/32), si packed (M/128, N)
+  assert g.ndim == 2 and g.dtype == dtypes.bfloat16, f"{g.shape} {g.dtype}"
+  from extra.gemm.cdna_asm_gemm import FP8_DTYPE, mx_pack
+  M, N = g.shape
+  assert M % BLK == 0 and N % TILE_N == 0, f"M={M} must%{BLK}, N={N} must%{TILE_N}"
+  device = g.device
+  axis = g.uop.axis if isinstance(device, tuple) else None
+  out_axis = None if axis is None else (1 if axis == 0 else 0)
+  q = alloc_like((N, M), FP8_DTYPE, device, out_axis)
+  e8 = alloc_like((N, M // BLK), dtypes.uint8, device, out_axis)
+  fxn = functools.partial(_custom_transpose_quantize_mxfp8, dname=dname_of(device))
+  q, e8, *_ = Tensor.custom_kernel(q, e8, g, fxn=fxn)
+  return q, e8, mx_pack(e8)
@@ -0,0 +1,62 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_fp8.h>
+#include <hip/hip_bf16.h>
+
+#ifndef M_DIM
+#define M_DIM 8192
+#endif
+#ifndef N_DIM
+#define N_DIM 14336
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+
+constexpr int BLK = 32;
+constexpr int TILE_M = BLK;                 // one mxfp8 block along M per tile
+constexpr int TILE_N = THREADS_PER_WG;      // 256, one output column per thread
+constexpr int LDS_STRIDE = TILE_N + 1;      // +1 pad: stride 257 ≡ 1 (mod 32) -> conflict-free column reads
+constexpr int N_TILES_N = N_DIM / TILE_N;
+constexpr float FP8_MAX = 448.0f;
+
+static_assert(M_DIM % TILE_M == 0, "M_DIM must be a multiple of 32");
+static_assert(N_DIM % TILE_N == 0, "N_DIM must be a multiple of 256");
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+transpose_quantize_mxfp8(__hip_fp8_storage_t* __restrict__ q,        // (N_DIM, M_DIM)
+                         uint8_t* __restrict__ e8_out,               // (N_DIM, M_DIM/32)
+                         const __hip_bfloat16* __restrict__ g)       // (M_DIM, N_DIM)
+{
+  __shared__ __hip_bfloat16 lds[TILE_M * LDS_STRIDE];
+  const int tid    = threadIdx.x;
+  const int tile_m = blockIdx.x / N_TILES_N;     // which 32-block along M
+  const int tile_n = blockIdx.x % N_TILES_N;
+
+  #pragma unroll
+  for (int mm = 0; mm < TILE_M; mm++)
+    lds[mm * LDS_STRIDE + tid] = g[(long long)(tile_m * TILE_M + mm) * N_DIM + (tile_n * TILE_N + tid)];
+  __syncthreads();
+
+  float vals[TILE_M];
+  float amax = 0.0f;
+  #pragma unroll
+  for (int mm = 0; mm < TILE_M; mm++) {
+    float v = (float)lds[mm * LDS_STRIDE + tid];
+    vals[mm] = v;
+    amax = fmaxf(amax, fabsf(v));
+  }
+  int e8 = (int)floorf(log2f(fmaxf(amax, 1e-38f))) + 127;
+  e8 = max(0, min(254, e8));
+  float qscale = exp2f((float)(127 - e8));
+
+  const long long n = tile_n * TILE_N + tid;
+  __hip_fp8_storage_t out[TILE_M];
+  #pragma unroll
+  for (int mm = 0; mm < TILE_M; mm++)
+    out[mm] = __hip_cvt_float_to_fp8(fmaxf(-FP8_MAX, fminf(FP8_MAX, vals[mm] * qscale)), __HIP_SATFINITE, __HIP_E4M3);
+  // 32 contiguous fp8 along M -> two 16-byte vector stores
+  long long obase = n * M_DIM + (long long)(tile_m * TILE_M);
+  *reinterpret_cast<uint4*>(&q[obase])      = *reinterpret_cast<uint4*>(&out[0]);
+  *reinterpret_cast<uint4*>(&q[obase + 16]) = *reinterpret_cast<uint4*>(&out[16]);
+  e8_out[n * (M_DIM / BLK) + tile_m] = (uint8_t)e8;
+}
@@ -20,6 +20,7 @@ include-package-data = true
 packages = [
  'tinygrad',
  'tinygrad.codegen',
+  'tinygrad.codegen.decomp',
  'tinygrad.codegen.opt',
  'tinygrad.codegen.late',
  'tinygrad.engine',
@@ -140,7 +140,7 @@ class TestLinearizer(unittest.TestCase):
                       renderer=Device[Device.DEFAULT].renderer).src[2].src)
    num_loads = len([uop for uop in uops if uop.op is Ops.LOAD])
    assert num_loads <= 4, "more load uops than needed"
-    assert num_loads >= 4, "unexpected number of uops, maybe this test needs updating?"
+    assert num_loads >= 1, "expected at least one load uop"

  @unittest.skip("this is handled at higher level now")
  def test_upcast_cse(self):
@@ -0,0 +1,67 @@
+import unittest, math
+import numpy as np
+from tinygrad.tensor import Tensor
+from tinygrad.nn.optim import AdamW
+from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup, LambdaLR, LambdaLinearScheduler
+
+np.random.seed(1337)
+x_init = np.random.randn(1,4).astype(np.float32)
+W_init = np.random.randn(4,4).astype(np.float32)
+m_init = np.random.randn(1,4).astype(np.float32)
+
+class TinyNet:
+  def __init__(self):
+    self.x = Tensor(x_init.copy())
+    self.W = Tensor(W_init.copy())
+    self.m = Tensor(m_init.copy())
+
+  def forward(self):
+    out = self.x.matmul(self.W).relu()
+    out = out.log_softmax(1)
+    out = out.mul(self.m).add(self.m).sum()
+    return out
+
+class TestCosineAnnealingLRWithWarmup(unittest.TestCase):
+  # only tests the lr
+  def _test_lr(self, base_lr, end_lr, warmup_steps, decay_steps):
+    net = TinyNet()
+    optim = AdamW([net.W], lr=0.0)
+    tiny_lr = CosineAnnealingLRWithWarmup(optim, base_lr, end_lr, warmup_steps, decay_steps)
+    lr = []
+    for _ in range(warmup_steps+decay_steps):
+      lr.append(optim.lr.item())
+      tiny_lr.step()
+    # reimplemented in python
+    expected = []
+    for i in range(warmup_steps): expected.append((i+1)/warmup_steps*base_lr)
+    for i in range(decay_steps): expected.append(end_lr+(base_lr-end_lr)*(1+math.cos((i+1)/decay_steps*math.pi))/2)
+    np.testing.assert_allclose(lr, expected, rtol=1e-5)
+
+  def test_lr_0(self): self._test_lr(3e-4, 8e-5, 3, 5)
+  def test_lr_1(self): self._test_lr(3e-4, 8e-5, 10, 20)
+  def test_lr_llama3(self): self._test_lr(8e-5, 8e-7, 20, 100)
+
+class TestLambdaLRLinearWarmup(unittest.TestCase):
+  def test_linear_lr_warmup(self):
+    BS, BASE_LR = 304, 2.5e-7
+    lr = BS * BASE_LR
+    # Use a dummy Tensor parameter for optimizer because the lr_scheduler only needs the optimizer's device and lr, the params aren't touched.
+    optimizer = AdamW([Tensor([1.])])
+    lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
+    lr_scheduler = LambdaLR(optimizer, Tensor(lr, device=optimizer.device), lambda_lr_callback)
+    lrs = {}
+
+    # with above settings, optimizer.lr should warm up to lr over 1000 steps linearly
+    for i in range(1200):
+      lr_scheduler.step()
+      if i in {0, 499, 998, 999, 1000, 1199}:
+        lrs[i] = optimizer.lr.item()
+
+    np.testing.assert_allclose(lr, lrs[999], rtol=0, atol=1e-11)
+    np.testing.assert_equal(lrs[999], lrs[1000])
+    np.testing.assert_equal(lrs[999], lrs[1199])
+    np.testing.assert_allclose(lrs[999] / lrs[0], 1000, rtol=0, atol=1)
+    np.testing.assert_allclose(lrs[999] / lrs[499], 2, rtol=0, atol=1e-5)
+
+if __name__ == '__main__':
+  unittest.main()
@@ -1,5 +1,5 @@
 #!/usr/bin/env python
-import unittest, math
+import unittest
 import numpy as np
 import tensorflow as tf
 from tensorflow.keras.optimizers import Lamb
@@ -7,11 +7,11 @@ from tensorflow.python.ops import math_ops
 from extra.lr_scheduler import LRSchedulerGroup

 from tinygrad.tensor import Tensor
-from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, AdamW
+from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup

 from test.external.mlperf_resnet.lars_optimizer import LARSOptimizer

-from examples.mlperf.lr_schedulers import PolynomialDecayWithWarmup, CosineAnnealingLRWithWarmup, LambdaLR, LambdaLinearScheduler
+from examples.mlperf.lr_schedulers import PolynomialDecayWithWarmup
 from test.external.mlperf_resnet.lars_util import PolynomialDecayWithWarmup as PolynomialDecayWithWarmup_tf

 np.random.seed(1337)
@@ -173,48 +173,5 @@ class ExternalTestOptim(unittest.TestCase):
      'warmup': steps_per_epoch * warmup_epochs,
    }, 1e-5, 1e-5, do_optim=False)

-
-class TestCosineAnnealingLRWithWarmup(unittest.TestCase):
-  # only tests the lr
-  def _test_lr(self, base_lr, end_lr, warmup_steps, decay_steps):
-    net = TinyNet()
-    optim = AdamW([net.W], lr=0.0)
-    tiny_lr = CosineAnnealingLRWithWarmup(optim, base_lr, end_lr, warmup_steps, decay_steps)
-    lr = []
-    for _ in range(warmup_steps+decay_steps):
-      lr.append(optim.lr.item())
-      tiny_lr.step()
-    # reimplemented in python
-    expected = []
-    for i in range(warmup_steps): expected.append((i+1)/warmup_steps*base_lr)
-    for i in range(decay_steps): expected.append(end_lr+(base_lr-end_lr)*(1+math.cos((i+1)/decay_steps*math.pi))/2)
-    np.testing.assert_allclose(lr, expected, rtol=1e-5)
-
-  def test_lr_0(self): self._test_lr(3e-4, 8e-5, 3, 5)
-  def test_lr_1(self): self._test_lr(3e-4, 8e-5, 10, 20)
-  def test_lr_llama3(self): self._test_lr(8e-5, 8e-7, 20, 100)
-
-class TestLambdaLRLinearWarmup(unittest.TestCase):
-  def test_linear_lr_warmup(self):
-    BS, BASE_LR = 304, 2.5e-7
-    lr = BS * BASE_LR
-    # Use a dummy Tensor parameter for optimizer because the lr_scheduler only needs the optimizer's device and lr, the params aren't touched.
-    optimizer = AdamW([Tensor([1.])])
-    lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
-    lr_scheduler = LambdaLR(optimizer, Tensor(lr, device=optimizer.device), lambda_lr_callback)
-    lrs = {}
-
-    # with above settings, optimizer.lr should warm up to lr over 1000 steps linearly
-    for i in range(1200):
-      lr_scheduler.step()
-      if i in {0, 499, 998, 999, 1000, 1199}:
-        lrs[i] = optimizer.lr.item()
-
-    np.testing.assert_allclose(lr, lrs[999], rtol=0, atol=1e-11)
-    np.testing.assert_equal(lrs[999], lrs[1000])
-    np.testing.assert_equal(lrs[999], lrs[1199])
-    np.testing.assert_allclose(lrs[999] / lrs[0], 1000, rtol=0, atol=1)
-    np.testing.assert_allclose(lrs[999] / lrs[499], 2, rtol=0, atol=1e-5)
-
 if __name__ == '__main__':
  unittest.main()
@@ -3,7 +3,7 @@ from tinygrad import Tensor, UOp, Device, nn
 from tinygrad.schedule import schedule_cache
 from tinygrad.codegen import to_program, to_program_cache
 from tinygrad.schedule.indexing import apply_movement_op, _apply_reshape
-from tinygrad.uop.divandmod import fold_divmod_general
+from tinygrad.codegen.decomp.divandmod import fold_divmod_general
 from test.test_tiny import TestTiny

 def uops_allocated(): return sum([isinstance(x, UOp) for x in gc.get_objects()])
@@ -3,7 +3,7 @@ import z3
 from tinygrad import dtypes, Device
 from tinygrad.uop.validate import uops_to_z3, z3_cdiv
 from tinygrad.uop.ops import UOp
-from tinygrad.uop.decompositions import fast_idiv
+from tinygrad.codegen.decomp.op import fast_idiv
 random.seed(42)

 powers_of_two = [2**i for i in range(64)]
@@ -2,7 +2,7 @@
 from typing import Any, Callable
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import Ops, UOp
-from tinygrad.uop.decompositions import f2f
+from tinygrad.codegen.decomp.dtype import f2f

 # Type alias for vars dict: stores UOps and tuples for lambda definitions
 VarVal = UOp | tuple[str, list[str], str]
@@ -0,0 +1,45 @@
+import unittest
+from tinygrad import UOp, dtypes
+from tinygrad.uop.ops import shape_to_shape_arg, ParamArg, Ops, AddrSpace
+
+def placeholder(shape, dtype, slot):
+  return UOp(Ops.PARAM, dtype, (shape_to_shape_arg(shape),), arg=ParamArg(slot, AddrSpace.GLOBAL))
+
+class TestBitcastSpec(unittest.TestCase):
+  def test_bitcast_no_shape_change(self):
+    pl = placeholder((10,10), dtypes.int, 0)
+    out = pl.bitcast(dtypes.float)
+    self.assertEqual(out.shape, (10,10))
+
+  def test_bitcast_increase_shape(self):
+    pl = placeholder((10,10), dtypes.int, 0)
+    out = pl.bitcast(dtypes.short)
+    self.assertEqual(out.shape, (10,20))
+
+  def test_bitcast_decrease_shape(self):
+    pl = placeholder((10,10), dtypes.int, 0)
+    out = pl.bitcast(dtypes.long)
+    self.assertEqual(out.shape, (10,5))
+
+  def test_bitcast_remove_ones(self):
+    pl = placeholder((10,2), dtypes.int, 0)
+    out = pl.bitcast(dtypes.long)
+    self.assertEqual(out.shape, (10,1))
+
+  def test_bitcast_remove_ones_full(self):
+    pl = placeholder((2,), dtypes.int, 0)
+    out = pl.bitcast(dtypes.long)
+    self.assertEqual(out.shape, (1,))
+
+  def test_bitcast_add_ones_full(self):
+    pl = placeholder((), dtypes.long, 0)
+    out = pl.bitcast(dtypes.int)
+    self.assertEqual(out.shape, (2,))
+
+  def test_bitcast_add_ones_full_uchar(self):
+    pl = placeholder((), dtypes.long, 0)
+    out = pl.bitcast(dtypes.uchar)
+    self.assertEqual(out.shape, (8,))
+
+if __name__ == '__main__':
+  unittest.main()
@@ -16,41 +16,17 @@ def simplify_image_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move
 def get_gated_load_uop(valid:UOp, idx:UOp):
  return UOp(Ops.LOAD, dtypes.float, (
    UOp.param(0, dtypes.float.ptr()).index(idx.valid(valid), ptr=True),
-    UOp.const(dtypes.float, 0.0)
  ))

 def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UOp]):
  return UOp(Ops.LOAD, dtypes.float.vec(4), (
    UOp.param(0, dtypes.imagef(image_shape)).index(idx[1].valid(valid), idx[0].valid(valid), ptr=True),
-    UOp(Ops.STACK, dtypes.float.vec(4), src=(UOp.const(dtypes.float, 0.0),) * 4)
  ))

 def Special(expr, nmax): return UOp(Ops.SPECIAL, dtypes.weakint, (UOp.const(dtypes.weakint, nmax),), expr)
 def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax)
 def Range(n, nmax): return UOp.range(nmax, n)

-class TestHelpers(unittest.TestCase):
-  def test_is_increasing(self):
-    idx1 = Special("idx1", 32)
-    idx2 = Special("idx2", 64)
-    ridx0 = Variable("ridx0", 0, 5)
-    ridx1 = Variable("ridx1", 0, 2)
-    ridx2 = Variable("ridx2", 0, 2)
-    # (ridx0+(idx1*48)+(ridx2*6)+(-6)),((idx2*2)+ridx1+(-1)))
-    f0 = ((idx1*24)+(ridx2*3)+ridx0+765)%768
-    f1 = ridx0+(idx1*48)+(ridx2*6)+(-6)
-    f2 = (idx2*2)+ridx1+((idx1+((ridx2+7)//8)+31)//32)+(-2)
-    f3 = (idx2*2)+ridx1+(-1)
-
-    self.assertFalse(f0.is_increasing())
-    self.assertTrue(f1.is_increasing())
-    self.assertTrue(f2.is_increasing())
-    self.assertTrue(f3.is_increasing())
-
-    rng = UOp.range(5, 2)
-    self.assertTrue(rng.is_increasing())
-    self.assertTrue((rng+2).is_increasing())
-
 class TestValidIdxSimplification(unittest.TestCase):
  def check(self, load, sidx, svalid, extra=()):
    load = simplify_valid_idx(UOp.sink(load, *extra)).src[0]
@@ -506,6 +482,16 @@ class TestImageSimplification(unittest.TestCase):
      self.check(load, "(((lidx1<1)!=True)&(((lidx0+r0)<3)!=True)&((lidx0+r0)<11))",
                       "(lidx2+gidx0*4+lidx1*256+(lidx0*1024+r0*1024)+-3264)", "0")

+  def test_drop_non_monotonic_window(self):
+    # two-sided window valid (645 <= gidx0 < 653) on a non-monotonic index (lane split via %4 and //4):
+    # gidx0 outside the window pushes idx_x out of the (1, 48) image, so the gate is dropped
+    gidx0 = Special("gidx0", 1064)
+    r12 = Range(12, 3)
+    valid = ((gidx0 < 645).ne(True)) & (gidx0 < 653)
+    idx = (r12*4 + (gidx0+3)%4 + (gidx0+3)//4*24 - 3888, UOp.const(dtypes.weakint, 0))
+    load = get_load_image_uop((1, 48, 4), valid, idx)
+    self.check(load, None, "(r12*4+(gidx0+3)%4+(gidx0+3)//4*24+-3888)", "0")
+
 class TestDropTrueGate(unittest.TestCase):
  def test_drop_true_gate_on_index(self):
    # test that INDEX with a constant True valid gets simplified to drop the valid
@@ -2,8 +2,8 @@ import unittest, math
 import numpy as np
 from tinygrad import dtypes
 from tinygrad.uop.ops import UOp
-from tinygrad.uop.decompositions import TRANSCENDENTAL_DTYPES, payne_hanek_reduction, cody_waite_reduction
-from tinygrad.uop.decompositions import frexp, rintk, xpow, xexp2, xlog2, trig_poly, pow2if
+from tinygrad.codegen.decomp.transcendental import TRANSCENDENTAL_DTYPES, payne_hanek_reduction, cody_waite_reduction
+from tinygrad.codegen.decomp.transcendental import frexp, rintk, xpow, xexp2, xlog2, trig_poly, pow2if
 from test.helpers import eval_uop

 class TestTranscendentalFunctions(unittest.TestCase):
@@ -7,7 +7,7 @@ from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, TrackedPatternMatch
 from tinygrad.uop.symbolic import sym
 from tinygrad.dtype import dtypes, AddrSpace
 from tinygrad.helpers import colored, ansistrip, flatten, TracingKey, ProfileRangeEvent, ProfileEvent, Context, cpu_events, profile_marker
-from tinygrad.helpers import cpu_profile, ProfilePointEvent, unwrap, VIZ
+from tinygrad.helpers import cpu_profile, ProfilePointEvent, unwrap, VIZ, BEAM
 from tinygrad.device import Buffer

 from tinygrad.uop.ops import tracked_keys, tracked_ctxs, uop_fields, active_rewrites, active_group, _name_cnt, RewriteTrace
@@ -341,41 +341,28 @@ class TestVizGC(unittest.TestCase):
 from tinygrad import Tensor, Device, TinyJit, Variable, function

 class TestVizIntegration(unittest.TestCase):
-  # codegen supports rendering of code blocks
-  def test_codegen_tracing(self):
-    with save_viz() as viz:
-      ast = (Tensor.empty(4)+Tensor.empty(4)).schedule_linear().src[0].src[0]
-      prg = do_to_program(ast, Device[Device.DEFAULT].renderer)
-    lst = viz.list_items()
-    self.assertEqual(len(lst), 3)
-    self.assertEqual(lst[0]["name"], "Callify 1 Buffer n1")
-    self.assertEqual(lst[1]["name"], "Schedule 1 Kernel n1")
-    self.assertEqual(lst[2]["name"], prg.arg.name)
-    input_ast = next(viz.get_details(2, 0))["graph"].values()
-    for u in input_ast:
-      if u["label"].startswith("PARAM\n"): self.assertEqual(u["addrspace"], addrspace_colors[AddrSpace.GLOBAL])
-
-  # schedule graph CALL nodes have a link to jump to codegen
  def test_link_sched_codegen(self):
+    c1 = Tensor.empty(4, device="NULL")
+    c2 = Tensor.empty(8, device="NULL")
+    # uniquely named A = B + 1 kernel
+    kernel_name = f"custom_add1_link_sched_codegen_{BEAM.value}"
+    def custom_add1(A:UOp, B:UOp): return A[0].store(B[0]+1).sink(arg=KernelInfo(kernel_name))
    with save_viz() as viz:
-      c1 = Tensor.empty(4, device="NULL").add(1)
-      c2 = Tensor.empty(8, device="NULL").add(1)
-      with Context(SCACHE=0):
-        sched = c1.schedule_linear(c2)
-      from tinygrad.engine.realize import compile_linear
-      sched = compile_linear(sched)
-      with Context(NO_COLOR=0):
-        prgs = [do_to_program(si.src[0], Device[c1.device].renderer).arg.name for si in sched.src]
+      c1 = Tensor.custom_kernel(c1, c2, fxn=custom_add1)[0]
+      c1.realize()
    lst = viz.list_items()
+    # schedule graph CALL nodes have a link to jump to codegen
    sched_idx = next(i for i,l in enumerate(lst) if l["name"].startswith("Schedule"))
    viz_kernel = next(i for i,s in enumerate(lst[sched_idx]["steps"]) if s["name"] == "View Kernel Graph")
-    with Context(NO_COLOR=1):
-      graph = next(viz.get_details(sched_idx, viz_kernel))["graph"]
+    graph = next(viz.get_details(sched_idx, viz_kernel))["graph"]
    call_nodes = [n for n in graph.values() if n["label"].startswith("CALL")]
    for i,n in enumerate(call_nodes):
      assert n["ref"] is not None
-      self.assertEqual(lst[n["ref"]]["name"], prgs[i])
-      assert ansistrip(prgs[i]) in n["label"], f"CALL must contain kernel name, got {n['label']}"
+      self.assertEqual(lst[n["ref"]]["name"], kernel_name)
+      assert kernel_name[i] in n["label"], f"CALL must contain kernel name, got {n['label']}"
+    # UOp addrspace is colored
+    for u in graph.values():
+      if u["label"].startswith("PARAM\n"): self.assertEqual(u["addrspace"], addrspace_colors[AddrSpace.GLOBAL])

  def test_link_sched_codegen_beam(self):
    with Context(BEAM=2):
@@ -493,6 +480,24 @@ class TestVizIntegration(unittest.TestCase):
    bin_render = get_render(viz.data, steps[bin_idx]["query"])["src"]
    self.assertIn(type(e.exception).__name__, bin_render)

+  def test_view_source_alt(self):
+    src = "void E_3(float* data0_3) {}"
+    binary = Device["CPU"].renderer.compiler.compile(src)
+    def custom_binary(X:UOp):
+      sink = UOp.sink(X, arg=KernelInfo("custom_binary"))
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="CPU"), UOp(Ops.LINEAR, src=sink.src+(sink,)), UOp(Ops.SOURCE, arg=src),
+                                   UOp(Ops.BINARY, arg=binary)))
+    x = Tensor.custom_kernel(Tensor.empty(1, device="CPU"), fxn=custom_binary)[0]
+    with save_viz() as viz:
+      x.realize()
+    lst = viz.list_items()
+    codegen_idx = len(lst)-1
+    steps = lst[codegen_idx]["steps"]
+    src_idx = next((i for i,s in enumerate(steps) if s["name"] == "View Source"), None)
+    assert src_idx is not None, "must have source rendering in list"
+    src_render = get_render(viz.data, steps[src_idx]["query"])["src"]
+    self.assertEqual(src, src_render)
+
 from tinygrad.device import ProfileDeviceEvent, ProfileGraphEvent, ProfileGraphEntry
 from tinygrad.viz.serve import get_profile
 from tinygrad.viz.cli import decode_profile
@@ -1,4 +1,3 @@
-from typing import cast
 from dataclasses import replace
 import itertools
 from tinygrad.helpers import DISABLE_FAST_IDIV, TRANSCENDENTAL, SPEC, DEBUG, VIZ, IMAGE, NOOPT, EMULATED_DTYPES, NOLOCALS, USE_TC
@@ -13,7 +12,9 @@ from tinygrad.dtype import dtypes, PtrDType, ImageDType
 # import all pattern matchers here
 from tinygrad.codegen.gpudims import pm_add_gpudims
 from tinygrad.uop.symbolic import sym, symbolic_simple, gep_pushing, symbolic, pm_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
-from tinygrad.uop.decompositions import get_late_rewrite_patterns, get_transcendental_patterns, pm_dtype_decomps
+from tinygrad.codegen.decomp.dtype import pm_dtype_decomps
+from tinygrad.codegen.decomp.op import get_late_rewrite_patterns, get_simplifying_rewrite_patterns
+from tinygrad.codegen.decomp.transcendental import get_transcendental_patterns
 from tinygrad.codegen.late.expander import expander, pm_pre_expander, pm_group_for_reduce
 from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize_buf_and_index, devectorize_alu, pm_reduce, \
  ReduceContext, correct_load_store, pm_render, pm_add_loads, pm_make_images
@@ -23,6 +24,7 @@ from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_s
 from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen, pm_mops, pm_syntactic_sugar, pm_store_ranges
 from tinygrad.codegen.late.linearizer import CFGContext, pm_split_ends, pm_add_control_flow, linearize
 from tinygrad.codegen.late.regalloc import LinearScanRegallocContext, pm_regalloc_rewrite
+from tinygrad.codegen.late.coalese import memory_coalesing

 pm_index_is_shrink = PatternMatcher([
  # rewrite non-image INDEX to SHRINK
@@ -52,6 +54,10 @@ pm_number_params = PatternMatcher([
  (UPat(Ops.PARAM, name="x"), do_number_param),
 ])

+pm_no_weakints = PatternMatcher([
+  (UPat(GroupOp.All, dtype=dtypes.weakint, name="x"), lambda x: x.replace(dtype=dtypes.int))
+])
+
 def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
  if VIZ: graph_rewrite(ast, PatternMatcher([]), name="View Base AST")
  if DEBUG >= 5: print(pyrender(ast))
@@ -113,27 +119,30 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
  # optional pre matcher
  if ren.pre_matcher is not None: sink = graph_rewrite(sink, ren.pre_matcher, name="pre_matcher")

-  # decompositions
+  # floordiv+mod / dtype decomp (early)
  supported_ops = tuple(ren.code_for_op.keys())
-  pm_decomp = symbolic_simple+get_late_rewrite_patterns(supported_ops, bool(DISABLE_FAST_IDIV))
-  pm_transcendental = symbolic_simple+get_transcendental_patterns(supported_ops, TRANSCENDENTAL>=2)
-  sink = graph_rewrite(sink, pm_decomp, ctx=ren, name="decompositions")
-  sink = graph_rewrite(sink, pm_dtype_decomps, ctx=(set(), ren), name="decomp dtypes")
-  sink = graph_rewrite(sink, pm_transcendental, name="transcendental")
+  pm_decomp = symbolic_simple+get_simplifying_rewrite_patterns(supported_ops)
+  sink = graph_rewrite(sink, pm_decomp, name="early decompositions")

-  # GEP/STACK stuff
+  # do memory coalesing (late)
+  sink = memory_coalesing(sink, ren)
+
+  # this is new style (TODO: this should all be removed)
  sink = graph_rewrite(sink, pm_render, name="pm_render gep/stack")
-
-  # this is new style
  sink = graph_rewrite(sink, pm_index_is_shrink, name="index is shrink")
  sink = graph_rewrite(sink, pm_remove_vec_dtypes, name="transform to new style")

-  # move gates from unrenderable INVALID where
+  # late decomps + move gates from unrenderable INVALID where
+  sink = graph_rewrite(sink, pm_dtype_decomps, ctx=(set(), ren), name="decomp dtypes")
+  pm_decomp = pm_decomp+\
+    get_late_rewrite_patterns(supported_ops, bool(DISABLE_FAST_IDIV))+\
+    get_transcendental_patterns(supported_ops, TRANSCENDENTAL>=2)
+  sink = graph_rewrite(sink, pm_decomp, ctx=ren, name="late decompositions")
  sink = graph_rewrite(sink, pm_move_gates_from_index, name="move gates from index")

  # final rules for the renderer (without sym)
  extra_matcher = ren.extra_matcher if ren.extra_matcher is not None else PatternMatcher([])
-  pm_final_rewrite = pm_decomp+extra_matcher+pm_split_ends
+  pm_final_rewrite = pm_decomp+extra_matcher+pm_split_ends+pm_no_weakints
  sink = graph_rewrite(sink, pm_final_rewrite, ctx=ren, name="final rewrite")

  # this was the linearizer
@@ -164,7 +173,7 @@ def line_rewrite(lst:list[UOp], pm:PatternMatcher, ctx=None) -> list[UOp]:
  replaced: dict[UOp, UOp] = {}
  for u in lst:
    nu = u.replace(src=tuple([replaced.get(x, x) for x in u.src]))
-    ret: tuple[UOp, list[UOp]] = cast(tuple[UOp, list[UOp]]|None, pm.rewrite(nu, ctx)) or (nu, [nu])
+    ret: tuple[UOp, list[UOp]] = pm.rewrite(nu, ctx) or (nu, [nu])
    replaced[u] = ret[0]
    newlst.extend(ret[1])
  return newlst
@@ -0,0 +1,185 @@
+from tinygrad.dtype import dtypes, DType, truncate
+from tinygrad.helpers import flatten, DEBUG, EMULATED_DTYPES
+from tinygrad.uop import GroupOp
+from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, graph_rewrite
+from tinygrad.renderer import Renderer
+from tinygrad.codegen.decomp.transcendental import exponent_bias, shl, shr
+
+# ***** long as 2 ints *****
+
+l2i_dt = {dtypes.long: dtypes.int, dtypes.ulong: dtypes.uint}
+def unpack32(v:UOp) -> tuple[UOp, UOp]: return v.bitcast(dtypes.uint) & 0xFFFF, shr(v.bitcast(dtypes.uint), 16)
+def reindex(idx:UOp, off:int, mul=2) -> UOp:
+  if idx.op is Ops.SHRINK:
+    assert mul == 1, "can't reindex SHRINK with mul != 1"
+    return idx.replace(op=Ops.INDEX, src=(idx.src[0], idx.src[1]+off))
+  return idx.replace(src=(idx.src[0], idx.src[1]*mul+off, *idx.src[2:]))
+
+# 4.3.1 is the relevant section in TAOCP
+def l2i(op: Ops, dt: DType, *uops:UOp):
+  zero = UOp.const(dt, 0)
+  if len(uops) == 2: a0, a1 = uops
+  elif len(uops) == 4: a0, a1, b0, b1 = uops
+  match op:
+    case Ops.NEG: return l2i(Ops.SUB, dt, zero, zero, *uops)
+    case Ops.CAST if dt in (dtypes.long, dtypes.ulong) and uops[0].dtype not in dtypes.floats:
+      return uops[0].cast(l2i_dt[dt]), (uops[0] < 0).where(UOp.const(l2i_dt[dt], -1), UOp.const(l2i_dt[dt], 0))
+    case Ops.CAST if dt in (dtypes.long, dtypes.ulong):
+      return (lo:=uops[0].cast(l2i_dt[dt])), (uops[0] / 2**32).cast(l2i_dt[dt]) - ((uops[0] < 0) & lo.ne(0)).cast(l2i_dt[dt])
+    case Ops.CAST if dt in dtypes.floats:
+      small = (a1.eq(0) & (a0 >= 0)) | (a1.eq(-1) & (a0 < 0))
+      return small.where(a0.cast(dt), ((a1.cast(dtypes.float32) * (2**32)) + a0.bitcast(dtypes.uint).cast(dtypes.float32)).cast(dt))
+    case Ops.CAST: return a0.bitcast(dtypes.uint).cast(dt)
+    case Ops.BITCAST: return a0.bitcast(dt), a1.bitcast(dt)
+    case Ops.SHL:
+      lo, hi = shl(a0, b0_mod:=b0 & 31), shl(a1, b0_mod) | shr(shr(a0, 1), 31 - b0_mod)
+      return (b0 >= 32).where(zero, lo), (b0 >= 32).where(lo, hi)
+    case Ops.SHR:
+      lo, hi = shr(a0, b0_mod:=b0 & 31) | shl(shl(a1, 1), 31 - b0_mod), shr(a1, b0_mod)
+      return (b0 >= 32).where(hi, lo), (b0 >= 32).where(zero, hi)
+    case Ops.ADD: return (low:=a0+b0), (a1 + b1).replace(dtype=dt) + (low.bitcast(dtypes.uint) < a0.bitcast(dtypes.uint)).cast(dt)
+    case Ops.SUB: return a0 - b0, a1 - b1 - (a0.bitcast(dtypes.uint) < b0.bitcast(dtypes.uint)).cast(dt)
+    case Ops.MUL:
+      (a00, a01), (b00, b01) = unpack32(a0), unpack32(b0)
+      mid = l2i(Ops.ADD, dt, shl(a00*b01, 16).bitcast(dt), shr(a00*b01, 16).bitcast(dt), shl(a01*b00, 16).bitcast(dt), shr(a01*b00, 16).bitcast(dt))
+      return l2i(Ops.ADD, dt, *mid, (a00*b00).bitcast(dt), (a01*b01).bitcast(dt) + a0*b1 + a1*b0)
+    case Ops.CDIV | Ops.CMOD:
+      # TAOCP Algorithm 4.3.1D could be faster here, but must be parameterized over the width of b
+      if dt == dtypes.int:
+        ua0, ua1, ub0, ub1 = a0.bitcast(dtypes.uint), a1.bitcast(dtypes.uint), b0.bitcast(dtypes.uint), b1.bitcast(dtypes.uint)
+        a0, a1 = (a_neg:=a1 < zero).where((n:=l2i(Ops.NEG, dtypes.uint, ua0, ua1))[0], ua0), a_neg.where(n[1], ua1)
+        b0, b1 = (b_neg:=b1 < zero).where((n:=l2i(Ops.NEG, dtypes.uint, ub0, ub1))[0], ub0), b_neg.where(n[1], ub1)
+      q, r = (z:=UOp.const(dtypes.uint, 0), z), (z, z)
+      for i in range(63, -1, -1):
+        r = l2i(Ops.SHL, dtypes.uint, *r, UOp.const(dtypes.uint, 1), z)
+        r = (r[0] | l2i(Ops.SHR, dtypes.uint, a0, a1, UOp.const(dtypes.uint, i), z)[0] & 1), r[1]
+        cond = l2i(Ops.CMPLT, dtypes.uint, *r, b0, b1).logical_not()
+        diff = l2i(Ops.SUB, dtypes.uint, *r, b0, b1)
+        q = ((q[0] | shl(cond.cast(dtypes.uint), i % 32), q[1]) if i < 32 else (q[0], q[1] | shl(cond.cast(dtypes.uint), i % 32)))
+        r = l2i(Ops.WHERE, dtypes.uint, cond, *diff, *r)
+      if dt == dtypes.int:
+        (nq0, nq1), (nr0, nr1) = l2i(Ops.BITCAST, dt, *l2i(Ops.NEG, dtypes.uint, *q)), l2i(Ops.BITCAST, dt, *l2i(Ops.NEG, dtypes.uint, *r))
+        (q0, q1), (r0, r1) = l2i(Ops.BITCAST, dt, *q), l2i(Ops.BITCAST, dt, *r)
+        return (a_neg.where(nr0, r0), a_neg.where(nr1, r1)) if op == Ops.CMOD else ((a_neg^b_neg).where(nq0, q0), (a_neg^b_neg).where(nq1, q1))
+      return (r[0].bitcast(dt), r[1].bitcast(dt)) if op == Ops.CMOD else (q[0].bitcast(dt), q[1].bitcast(dt))
+    case Ops.CMPLT: return (a1 < b1) | ((a1.eq(b1)) & (a0.bitcast(dtypes.uint) < b0.bitcast(dtypes.uint)))
+    case Ops.CMPEQ: return a0.eq(b0) & a1.eq(b1)
+    case Ops.CMPNE: return a0.ne(b0) | a1.ne(b1)
+    case Ops.XOR | Ops.OR | Ops.AND: return UOp(op, dt, src=(a0, b0)), UOp(op, dt, src=(a1, b1))
+    case Ops.WHERE: return uops[0].where(uops[1], uops[3]), uops[0].where(uops[2], uops[4])
+    case Ops.MAX: return l2i(Ops.WHERE, dt, l2i(Ops.CMPLT, dt, *uops), b0, b1, a0, a1)
+    case _: raise NotImplementedError(f"long decomposition of {op} unsupported")
+
+# ***** floats *****
+f2f_dt = { f:getattr(dtypes, f"uint{f.bitsize}") for f in dtypes.floats }
+
+def rne(v: UOp, s) -> UOp: return shr(v, s) + ((shr(v, s - 1) & 1) & ((v & ((1 << (s - 1)) - 1)).ne(0).cast(v.dtype) | (shr(v, s) & 1)))
+
+def f2f(v, fr:DType, to:DType, sat=True):
+  fs, fb, (fe, fm), ts, tb, (te, tm) = fr.bitsize, exponent_bias(fr), dtypes.finfo(fr), to.bitsize, exponent_bias(to), dtypes.finfo(to)
+  # NB: denormals are zero!
+  if fe <= te and fm < tm:
+    sign, nosign = shl((v & shl(1, fs-1)).cast(f2f_dt[to]), ts - fs), (v & (shl(1, fs-1) - 1)).cast(f2f_dt[to])
+    exp, norm = shr(nosign, fm), shl(nosign, tm - fm) + shl(tb - fb, tm)
+    nan = shl(nosign, tm - fm) | shl((shl(1, te) - 1), tm)
+    if fr in dtypes.fp8_fnuz:
+      fnuz_nan = sign.ne(0) & nosign.eq(0)
+      qnan = shl(shl(1, te) - 1, tm) | shl(1, tm - 1)
+      return fnuz_nan.where(qnan, sign | exp.eq(0).where(0, norm)).bitcast(to)
+    # fp8e4m3 has only one nan
+    is_nan = (nosign.eq(shl(1, fm + fe) - 1) if fr == dtypes.fp8e4m3 else exp.eq(shl(1, fe) - 1))
+    return (sign | exp.eq(0).where(0, is_nan.where(nan, norm))).bitcast(to)
+  elif fe >= te and fm > tm:
+    v = f2f_clamp(v.bitcast(fr), to, sat).bitcast(f2f_dt[fr])
+    sign, nosign = shr(v, fs - ts) & shl(1, ts - 1), v & (shl(1, fs - 1) - 1)
+    norm = (rne(nosign, fm - tm) - shl(fb - tb, tm)).cast(f2f_dt[to])
+    underflow = (shr(v, fm) & (shl(1, fe) - 1)) < (1 + fb - tb)
+    nan_mantissa = (shl(1, tm) - 1) if to == dtypes.fp8e4m3 else (shr(nosign, fm - tm) & (shl(1, tm) - 1))
+    nan = (sign | nan_mantissa | shl(shl(1, te) - 1, tm)).cast(f2f_dt[to])
+    is_nan = (shr(v, fm) & (shl(1, fe) - 1)).eq(shl(1, fe) - 1)
+    if to in dtypes.fp8_fnuz: return is_nan.where(shl(1, ts - 1), underflow.where(0, sign.cast(f2f_dt[to]) | norm))
+    return is_nan.where(nan, sign.cast(f2f_dt[to]) | underflow.where(0, norm))
+  else: raise NotImplementedError(f"unsupported decomp {fr} -> {to}")
+
+def f2f_clamp(val:UOp, dt:DType, sat=True) -> UOp:
+  e, m = dtypes.finfo(dt)
+  if dt in dtypes.fp8_fnuz: max_exp, max_man = (1 << e) - 1, (1 << m) - 1
+  else: max_exp, max_man = ((1 << e) - 1, (1 << m) - 2) if dt == dtypes.fp8e4m3 else ((1 << e) - 2, (1 << m) - 1)
+  mx = val.const_like(2.0**(max_exp - exponent_bias(dt)) * (1.0 + max_man / (1 << m)))
+  sat = mx if dt in dtypes.fp8s and sat else val.const_like(float('inf'))
+  # FIXME: CMPLT of nan is undefined
+  return val.ne(val).where(val, (val < -mx).where(-sat, (mx < val).where(sat, val)))
+
+def f2f_load(x: UOp, fr:DType, to:DType) -> UOp:
+  if (n:=x.max_numel()) == 1: return f2f(x.replace(dtype=f2f_dt[fr]), fr, to)
+  return UOp(Ops.STACK, to, tuple(f2f(x.replace(dtype=f2f_dt[fr], src=(reindex(x.src[0], i, 1),)), fr, to) for i in range(n)))
+
+def f2f_store(st, idx, val, fr:DType, to:DType):
+  if (n:=val.max_numel()) == 1: return st.replace(src=(idx, f2f(val.bitcast(f2f_dt[to]), to, fr)))
+  return UOp.group(*(st.replace(src=(reindex(idx, i, 1), f2f(val.gep(i).bitcast(f2f_dt[to]), to, fr))) for i in range(n)))
+
+pm_long_decomp = PatternMatcher([
+  (UPat(GroupOp.Defines, src=(UPat.var("sz"),), name="x"), lambda x,sz:
+   x.replace(dtype=l2i_dt[x.dtype.base], src=(sz*2,)) if x.dtype.base in l2i_dt else None),
+  (UPat(Ops.INDEX, tuple(l2i_dt.keys()), name='x'), lambda x: reindex(x, x.tag).replace(dtype=l2i_dt[x.dtype]) if x.tag is not None else None),
+  (UPat(Ops.STORE, src=(UPat.var('idx'), UPat.var('val', tuple(l2i_dt.keys()))), name='st'), lambda st,idx,val:
+   st.replace(src=(idx.rtag(0), val.rtag(0))).group(st.replace(src=(idx.rtag(1), val.rtag(1)))) if val.tag is None else None),
+  (UPat(GroupOp.Comparison, src=(UPat.var('a', tuple(l2i_dt.keys())), UPat.var('b', tuple(l2i_dt.keys()))), name="x"), lambda a,b,x:
+   l2i(x.op, dt:=l2i_dt[a.dtype], a.rtag(0).cast(dt), a.rtag(1).cast(dt), b.rtag(0).cast(dt), b.rtag(1).cast(dt))),
+  (UPat(Ops.CAST, tuple(l2i_dt.keys()), src=(UPat.var('a'),), name="x"), lambda a,x:
+   l2i(x.op, x.dtype, a)[x.tag] if x.tag is not None and a.dtype not in l2i_dt else None),
+  (UPat(Ops.CAST, tuple(l2i_dt.keys()), src=(UPat.var('a', tuple(l2i_dt.keys())),), name="x"), lambda a,x:
+   (a.rtag(0).cast(dt:=l2i_dt[a.dtype]).bitcast(xdt:=l2i_dt[x.dtype]), a.rtag(1).cast(dt).bitcast(xdt))[x.tag]),
+  (UPat(Ops.CAST, src=(UPat.var('a', tuple(l2i_dt.keys())),), name="x"), lambda a,x:
+   l2i(x.op, x.dtype, a.rtag(0).cast(dt:=l2i_dt[a.dtype]), a.rtag(1).cast(dt)) if x.dtype not in l2i_dt and a.tag is None else None),
+  (UPat((*(GroupOp.ALU - GroupOp.Comparison), Ops.BITCAST), tuple(l2i_dt.keys()), name="x"), lambda x:
+   l2i(x.op, l2i_dt[x.dtype], *flatten((a.rtag(0).cast(dt:=l2i_dt[x.src[-1].dtype]), a.rtag(1).cast(dt))
+                                       if a.dtype in l2i_dt else (a,) for a in x.src))[x.tag] if x.tag is not None else None),
+  (UPat(Ops.LOAD, tuple(l2i_dt.keys()), src=(UPat.var('idx'),), name='x'), lambda x,idx:
+   x.replace(dtype=l2i_dt[x.dtype], src=(reindex(idx, x.tag).replace(dtype=l2i_dt[x.dtype]),))),
+  (UPat(Ops.CONST, tuple(l2i_dt.keys()), name='x'), lambda x:
+   UOp.const(dt:=l2i_dt[x.dtype], truncate[dt]((x.arg >> 32) if x.tag == 1 else (x.arg & 0xFFFFFFFF))))
+])
+
+# float decomposition patterns - ctx is (fr, to) tuple
+pm_float_decomp = PatternMatcher([
+  (UPat((*GroupOp.Defines, Ops.INDEX, Ops.SHRINK), name="x"), lambda ctx,x:
+   x.replace(dtype=f2f_dt[ctx[0]], tag=ctx[0])
+   if x.dtype.base == ctx[0] and (x.op is not Ops.INDEX or x.src[0].op not in {Ops.LOAD, Ops.STACK}) else None),
+  (UPat(Ops.LOAD, dtypes.floats, name="x"), lambda ctx,x: f2f_load(x, *ctx) if x.dtype.scalar() == ctx[0] else None),
+  # bitcasted load should just replace load
+  (UPat(Ops.BITCAST, src=(UPat(Ops.LOAD, name="ld"),), name="bc"), lambda ctx,bc,ld:
+   ld.replace(dtype=f2f_dt[ctx[0]]).bitcast(bc.dtype) if ld.dtype == ctx[0] else None),
+  # bitcast from
+  (UPat(Ops.BITCAST, src=(UPat.var("x", dtypes.floats),), name="bc"), lambda ctx,bc,x:
+   bc.replace(src=(f2f(x.bitcast(f2f_dt[ctx[1]]), ctx[1], ctx[0]),)) if x.dtype == ctx[1] and bc.dtype.bitsize == ctx[0].bitsize else None),
+  # bitcast to
+  (UPat(Ops.BITCAST, src=(UPat.var("x"),), name="bc"), lambda ctx,bc,x:
+   f2f(x.bitcast(f2f_dt[ctx[0]]), ctx[0], ctx[1]) if bc.dtype == ctx[0] else None),
+  (UPat(Ops.CAST, dtypes.floats, src=(UPat.var("val"),), name="x"), lambda ctx,x,val:
+   f2f_clamp(val.cast(ctx[1]), ctx[0]) if x.dtype.scalar() == ctx[0] else None),
+  (UPat(GroupOp.All-{Ops.BITCAST}, dtypes.floats, name="x"), lambda ctx,x:
+   x.replace(dtype=ctx[1].vec(x.dtype.count), src=tuple(s.cast(ctx[1]) if s.dtype == ctx[0] else s for s in x.src))
+   if x.dtype.scalar() == ctx[0] else None),
+  (UPat(Ops.STORE, src=(UPat.var("idx"), UPat(Ops.BITCAST, dtypes.floats, name="val")), name='st'), lambda ctx,st,idx,val:
+   st.replace(src=(idx, val.replace(dtype=f2f_dt[ctx[0]]))) if val.dtype == ctx[0] and idx.tag == ctx[0] else None),
+  (UPat(Ops.STORE, src=(UPat.var("idx"), UPat.var("val", dtypes.floats)), name='st'), lambda ctx,st,idx,val:
+   f2f_store(st, idx, val, *ctx) if val.dtype.scalar() == ctx[1] and (idx:=idx.src[0] if idx.op == Ops.CAST else idx).tag == ctx[0] else None),
+])
+
+def do_dtype_decomps(sink:UOp, ctx:tuple[set[DType], Renderer]) -> UOp:
+  def _should_emulate(dt): return dt in EMULATED_DTYPES.tolist(dtypes) or dt not in ctx[1].supported_dtypes()
+  for fr in sorted(filter(_should_emulate, ctx[0])):
+    to = dtypes.int if fr == dtypes.long else dtypes.half if not _should_emulate(dtypes.half) and fr in dtypes.fp8s else dtypes.float
+    if DEBUG >= 2: print(f"emulating {fr} as {to}")
+    sink = graph_rewrite(sink, pm_float_decomp if fr in dtypes.floats else pm_long_decomp, name=f"decomp {fr} -> {to}", ctx=(fr, to), bottom_up=True)
+  ctx[0].clear()
+  return sink
+
+pm_dtype_decomps = PatternMatcher([
+  # detect dtypes to decompose
+  (UPat(GroupOp.All, (*dtypes.fp8s, dtypes.bfloat16, dtypes.half, dtypes.long, dtypes.ulong), name="x"), lambda x,ctx:
+   ctx[0].add({dtypes.ulong:dtypes.long}.get(dt:=x.dtype.base.scalar(), dt))),
+  # do the rewrites
+  (UPat(Ops.SINK, name="sink"), do_dtype_decomps),
+])
@@ -0,0 +1,131 @@
+from typing import Callable
+import functools
+from tinygrad.dtype import dtypes, promo_lattice
+from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher
+from tinygrad.renderer import Renderer
+
+# *** integer division ***
+
+@functools.lru_cache(None)
+def magicgu(vmax:int, d:int) -> tuple[int,int]:
+  # calculate m,s such that x//d == (x*m) >> s for all 0 <= x <= vmax, d>0; adapted from Hacker's Delight, Chapter 10
+  nc = (vmax+1)//(d) * d - 1
+  nbits = vmax.bit_length()
+  for s in range(0, 2*nbits + 1):
+    if 2**s > nc*(d - 1 - (2**s - 1) % d):
+      m = (2**s + d - 1 - (2**s - 1) % d)//d
+      return m, s
+  assert False
+
+def fast_idiv(ren: Renderer, x: UOp, d: int, dont_cast=False) -> UOp|None:
+  from tinygrad.renderer.cstyle import MetalRenderer
+  # NOTE: disable for METAL due to compiler bug. keccak with -O0 works but not with optimization
+  if isinstance(ren, MetalRenderer): return None
+  # If d is a power of two this is not valid for signed ints!
+  is_unsigned = x.vmin>=0 or x.dtype in dtypes.uints
+  assert d>0, "Sign should have been taken out of divisor"
+  vmin,vmax = max(x.vmin, x.dtype.min), min(x.vmax, x.dtype.max)
+  if vmin > -d and vmax < d: return x.const_like(0)
+  m,s = magicgu(max(vmax, abs(vmin)), d)
+  if m*vmin >= x.dtype.min and m*vmax <= x.dtype.max:
+    return ((x*m) >> s) if is_unsigned else ((x*m) >> s) + (x<0).where(x.ufix(1), 0)
+  # before we try casting to a larger dtype (slow), we see if there are powers of two in d we can shift to make x smaller
+  # use explicit Ops.CDIV (trunc) since the recursion assumes trunc semantics throughout
+  if (largest_factor_of_two_in_d := (d & -d)) > 1:
+    if (ret:=fast_idiv(ren, x.alu(Ops.CDIV, x.const_like(largest_factor_of_two_in_d)),
+                       d//largest_factor_of_two_in_d, dont_cast=True)) is not None: return ret
+  if dont_cast: return None
+  # promo_lattice needs to return an unsigned type if the type is unsigned
+  if dtypes.is_int(next_dtype := promo_lattice[x.dtype.scalar()][-1]) and next_dtype in ren.supported_dtypes():
+    if m*vmin >= next_dtype.min and m*vmax <= next_dtype.max:
+      return ((x.cast(next_dtype)*m) >> s).cast(x.dtype) if is_unsigned else ((x.cast(next_dtype)*m) >> s).cast(x.dtype) + (x<0).where(x.ufix(1), 0)
+  return None
+
+# ***** threefry *****
+
+def threefry2x32(x: UOp, key: UOp):
+  # split x and key from uint64 to two uint32
+  x0, x1 = (x & 0xffffffff).cast(dtypes.uint32), ((x // 2**32) & 0xffffffff).cast(dtypes.uint32)
+  key0, key1 = (key & 0xffffffff).cast(dtypes.uint32), ((key // 2**32) & 0xffffffff).cast(dtypes.uint32)
+
+  rotations = [[13, 15, 26, 6], [17, 29, 16, 24]]
+  ks = [key1, key0 ^ key1 ^ 0x1BD11BDA, key0]
+  xr:list[UOp] = [x0 + ks[-1], x1 + ks[0]]
+  for i in range(5):
+    for r in rotations[i % 2]: xr[0], xr[1] = (x0 := xr[0] + xr[1]), x0 ^ ((xr[1] * 2**r) + (xr[1] // 2**(32 - r)))
+    xr = [(xr[0] + ks[i % 3]), (xr[1] + ks[(i + 1) % 3] + i + 1)]
+
+  return xr[1].cast(dtypes.uint64) * 2**32 | xr[0].cast(dtypes.uint64)
+
+# ***** decomposition patterns *****
+
+def floordiv_to_idiv(a:UOp, b:UOp) -> UOp:
+  if (a.vmin >= 0 and b.vmin > 0) or (a.vmax <= 0 and b.vmax < 0): return a.alu(Ops.CDIV, b)
+  return a.alu(Ops.CDIV, b) - (a.alu(Ops.CMOD, b).ne(0) & (a<0).ne(b<0)).cast(a.dtype)
+
+def floormod_to_mod(a:UOp, b:UOp) -> UOp:
+  if (a.vmin >= 0 and b.vmin > 0) or (a.vmax <= 0 and b.vmax < 0): return a.alu(Ops.CMOD, b)
+  r = a.alu(Ops.CMOD, b)
+  # use where instead of mul to avoid being fused into MULACC (which int64 long-decomp doesn't handle)
+  return r + (r.ne(0) & (a<0).ne(b<0)).where(b, b.const_like(0))
+
+powers_of_two: dict[int, int] = {2**i:i for i in range(64)}
+@functools.cache
+def get_simplifying_rewrite_patterns(ops:tuple[Ops, ...]) -> PatternMatcher:
+  # these are rewrites that make things simpler
+  pat: list[tuple[UPat, Callable]] = [(UPat.var("a")//UPat.var("b"), floordiv_to_idiv)]
+  # FLOORMOD by 2**y -> x & (2**y-1) (correct floor mod for any sign in two's complement); fires before floormod_to_mod
+  if Ops.AND in ops: pat.append((UPat.var("x", dtypes.ints)%UPat.cvar("c"), lambda x,c: x & (c.arg-1) if c.arg in powers_of_two else None))
+  pat.append((UPat.var("a")%UPat.var("b"), floormod_to_mod))
+  # no real hardware supports THREEFRY, but NullRenderer does
+  if Ops.THREEFRY not in ops: pat.append((UPat(Ops.THREEFRY, dtype=dtypes.uint64, src=(UPat.var("x"), UPat.var("key"))), threefry2x32))
+  # MAX can be rewritten as CMPLT + WHERE (max function is annoying on many cstyle backends)
+  if Ops.MAX not in ops and Ops.CMPLT in ops: pat.append((UPat(Ops.MAX, name="m"), lambda m: (m.src[0] < m.src[1]).where(m.src[1], m.src[0])))
+  return PatternMatcher(pat)
+
+@functools.cache
+def get_late_rewrite_patterns(ops:tuple[Ops, ...], disable_fast_idiv:bool) -> PatternMatcher:
+  pat: list[tuple[UPat, Callable]] = []
+  if Ops.OR in ops: pat += [(UPat.var("x", dtypes.bool).logical_not()&UPat.var("y", dtypes.bool).logical_not(),
+    lambda x,y: (x | y).logical_not())]
+  # rewrite MUL/CDIV to SHL+SHR: x*(2**y) -> shl(x,y) and x//(2**y) -> shr(x,y)
+  if Ops.SHL in ops: pat += [(UPat.var("x", dtypes.ints)*UPat.cvar("c"), lambda c,x: x << v if (v:=powers_of_two.get(c.arg, 0)) else None)]
+  if Ops.SHR in ops:
+    # uint CDIV by 2**v -> x >> v (FLOORDIV is lowered to CDIV by the rule above before reaching here)
+    pat += [(UPat(Ops.CDIV, src=(UPat.var("x", dtypes.uints), UPat.cvar("c"))),
+      lambda x,c: x >> v if (v:=powers_of_two.get(c.arg, 0)) else None)]
+    # signed CDIV (trunc) by 2**v -> (x + (x<0 ? c-1 : 0)) >> v
+    pat += [(UPat(Ops.CDIV, src=(UPat.var("x", dtypes.ints), UPat.cvar("c"))),
+      lambda x,c: (x+(l.const_like(l.vmin) if (l:=(x<0)).vmin==l.vmax else l).where(c-1, 0)) >> v
+        if (v:=powers_of_two.get(c.arg, 0)) else None)]
+    if not disable_fast_idiv:
+      # fast_idiv handles non-pow2: only fire on non-negative inputs (signed magic-mul is unreliable for x<0)
+      pat += [(UPat(Ops.CDIV, src=(UPat.var("x", dtypes.ints), UPat.cvar("d"))),
+        lambda ctx, x, d: fast_idiv(ctx, x, d.arg) if x.vmin >= 0 or x.dtype in dtypes.uints else None)]
+      # rewrite raw CMOD -> x - d*CDIV(x,d) so fast_idiv can pick up the CDIV. only on non-negative inputs;
+      # avoids disturbing floormod_to_mod's general-path output (which uses a trunc Ops.CMOD as an implementation detail)
+      pat += [(UPat(Ops.CMOD, src=(UPat.var("x", dtypes.ints), UPat.var("d"))),
+        lambda x, d: x - d * x.alu(Ops.CDIV, d) if x.vmin >= 0 or x.dtype in dtypes.uints else None)]
+  if Ops.NEG in ops:
+    pat += [(UPat.var('x')*-1, lambda ctx,x: x.alu(Ops.NEG))]
+    if Ops.SUB in ops: pat += [(UPat.var('x')+UPat.var('y').alu(Ops.NEG), lambda ctx,x,y: x.alu(Ops.SUB, y))]
+  if Ops.CMPLT in ops:
+    # These are late rewrites because simplex expects equalities to be a certain format
+    pat += [
+      ((UPat.var("x", dtypes.sints) < UPat.cvar("c", dtypes.sints)).logical_not(), lambda x,c: c-1<x),
+      ((UPat.cvar("c", dtypes.sints) < UPat.var("x", dtypes.sints)).logical_not(), lambda x,c: x<c+1),
+      (UPat.var("x", dtypes.sints)*-1 < UPat.var("y", dtypes.sints)*UPat.cvar("c"), lambda x,y,c: y*(-c)<x),
+      (UPat.var("x", dtypes.sints)*-1 < UPat.cvar("c"), lambda x,c:-c<x),
+      ((UPat.cvar("c1")<UPat.var("x", dtypes.sints)) & (UPat.var("x", dtypes.sints)<UPat.cvar("c2")),
+        lambda x,c1,c2: x.eq(c1+1) if c1.arg+1==c2.arg-1 else None),  # (c-1)<x & x<(c+1) -> x==c
+    ]
+  if Ops.CMPEQ in ops: pat += [(UPat.var('x').ne(UPat.var('y')).logical_not(), lambda x,y: x.alu(Ops.CMPEQ, y))]
+  if Ops.MULACC in ops:
+    pat += [(UPat.var('a')*UPat.var('b')+UPat.var('c'), lambda a,b,c: a.alu(Ops.MULACC, b, c))]
+    # also fuse (x << n) + c → MULACC(x, 2^n, c) since MUL→SHL may run first
+    if Ops.SHL in ops: pat += [(UPat.var('x').alu(Ops.SHL, UPat.cvar('n'))+UPat.var('c'), lambda x,n,c: x.alu(Ops.MULACC, x.const_like(1<<n.arg), c))]
+  # some backends emit FDIV for RECIP, in that case: a*(1/b) -> a/b
+  if Ops.FDIV in ops:
+    pat += [(UPat.var("x").reciprocal(), lambda x: x.const_like(1).alu(Ops.FDIV, x))]
+    pat += [(UPat.var("a", dtypes.floats) * UPat.const(dtypes.floats, 1).alu(Ops.FDIV, UPat.var("b")), lambda a,b: a.alu(Ops.FDIV, b))]
+  return PatternMatcher(pat)
@@ -0,0 +1,277 @@
+from typing import Callable
+import math, functools
+from tinygrad.dtype import dtypes, DType
+from tinygrad.helpers import polyN
+from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher
+
+TRANSCENDENTAL_DTYPES = (dtypes.float16, dtypes.float32, dtypes.float64)
+
+def _lazy_map_numbers(x:UOp, inf:UOp, _inf:UOp, nan:UOp, ratio:UOp):
+  """replace inf -> inf, -inf -> _inf, nan -> nan, otherwise -> ratio"""
+  return x.ne(math.inf).where(x.ne(x).where(nan, x.ne(-math.inf).where(ratio, _inf)), inf)
+
+# *** helper functions for bit manipulation ***
+def mantissa_bits(d:DType) -> int: return dtypes.finfo(d.scalar())[1]
+def exponent_bias(d:DType) -> int: return (1 << (dtypes.finfo(d.scalar())[0] - 1)) - (0 if d.scalar() in dtypes.fp8_fnuz else 1)
+def exponent_mask(d:DType) -> int: return (1 << dtypes.finfo(d.scalar())[0]) - 1
+
+# **** utils ****
+def shr(x:UOp|int, y:UOp|int) -> UOp: return x // (2**(y.simplify().arg) if isinstance(y, UOp) else 2**y)
+def shl(x:UOp|int, y:UOp|int) -> UOp: return x * (2**(y.simplify().arg) if isinstance(y, UOp) else 2**y)
+
+def rintk(d:UOp) -> UOp:
+  """round d:float to int away from 0"""
+  out_dtype = {dtypes.float64: dtypes.int64, dtypes.float32: dtypes.int32, dtypes.float16: dtypes.int16}[d.dtype.scalar()].vec(d.dtype.vcount)
+  return (d + (d<0.0).where(d.const_like(-0.5), d.const_like(0.5))).cast(out_dtype)
+
+def pow2if(q:UOp, float_dtype:DType):
+  """cast(2^q, float_dtype) where q is any integer in the range of [-126, 127]"""
+  out_dtype = {dtypes.int64: dtypes.float64, dtypes.int32: dtypes.float32, dtypes.int16: float_dtype.scalar()}[q.dtype.scalar()].vec(q.dtype.vcount)
+  return shl(q + exponent_bias(out_dtype), mantissa_bits(out_dtype)).bitcast(out_dtype)
+
+def ilogb2k(d:UOp) -> UOp:
+  """calculate the integer part of log2(d), where d is normalized fp value in the range of [0, +inf)."""
+  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
+  dint = d.bitcast({dtypes.float64: dtypes.int64, dtypes.float32: dtypes.int32, dtypes.float16: dtypes.int16}[d.dtype.scalar()].vec(d.dtype.vcount))
+  # -1 <= ilog2bk(d) <= 128
+  return (shr(dint, mantissa_bits(d.dtype)) & exponent_mask(d.dtype)) - exponent_bias(d.dtype)
+
+def ldexp3k(d:UOp, e:UOp) -> UOp:
+  """d*2^e. e is a number obtained by casting an integer in the range [-127, 127] to a float. d is any float number."""
+  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES and e.dtype.scalar() in TRANSCENDENTAL_DTYPES
+  dtype = {dtypes.float64: dtypes.int64, dtypes.float32: dtypes.int32, dtypes.float16: dtypes.int16}[d.dtype.scalar()].vec(d.dtype.count)
+  m1 = d.bitcast(dtype)
+  m2 = shl(e.cast(dtype), mantissa_bits(d.dtype))
+  return (m1 + m2).bitcast(d.dtype)
+
+def ldexp2k(d:UOp, e:UOp) -> UOp:
+  """d*2^e. much faster than ldexp3k but risky. d > 0 and d is not denormal."""
+  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES and e.dtype.scalar() in (dtypes.int16, dtypes.int32, dtypes.int64)
+  return (d * pow2if(shr(e, 1), d.dtype)) * pow2if(e - shr(e, 1), d.dtype)
+
+def frexp(v:UOp) -> tuple[UOp, UOp]:
+  """frexp(v) -> (mantissa, exponent) assuming v != 0"""
+  assert v.dtype.scalar() in TRANSCENDENTAL_DTYPES
+  # m1 = masks for mantissa, m2 = masks to normalize the mantissa.
+  m1 = {dtypes.float64: 0x000FFFFFFFFFFFFF, dtypes.float32: 0x807FFFFF, dtypes.float16: 0x83FF}[v.dtype.scalar()]
+  m2 = {dtypes.float64: 0x3FE0000000000000, dtypes.float32: 0x3F000000, dtypes.float16: 0x3800}[v.dtype.scalar()]
+  bits = v.bitcast({dtypes.float64: dtypes.uint64, dtypes.float32: dtypes.uint32, dtypes.float16: dtypes.uint16}[v.dtype.scalar()].vec(v.dtype.count))
+  exponent = shr(bits, mantissa_bits(v.dtype)) & exponent_mask(v.dtype)
+  # Set the exponent bits appropriately to normalize the mantissa into the range of [0.5, 1.0).
+  mantissa = ((bits & m1) | m2).bitcast(v.dtype)
+  exp = exponent - exponent_bias(v.dtype) + 1
+  return mantissa, exp
+
+# *** reduction algorithms for sine ***
+def payne_hanek_reduction(d:UOp) -> tuple[UOp, UOp]:
+  """
+  Performs Payne-Hanek Reduction: computes the remainder of `d` modulo pi/2 for the values `d` where
+    39800.0 <= d <= +Inf
+  Returns a tuple of `(r, q)`:
+  - `r`[d.dtype] is the reminder value corresponding to `round_to_nearest(x % pi/2)`.
+  - `q`[int32] is an integer, and q % 4 is corresponding to the quadrant of the original angle `d`.
+  """
+  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
+  # https://stackoverflow.com/questions/30463616/payne-hanek-algorithm-implementation-in-c/30465751#30465751
+  # 190 bits of 2/pi for Payne-Hanek style argument reduction
+  two_over_pi_f = [0x00000000, 0x28be60db, 0x9391054a, 0x7f09d5f4, 0x7d4d3770, 0x36d8a566, 0x4f10e410]
+
+  intermediate_dtype = dtypes.float32.vec(d.dtype.count) if d.dtype.base.scalar() == dtypes.float16 else d.dtype
+
+  f, e = frexp(d)
+  ia = (f.cast(intermediate_dtype) * 4.294967296e9).cast(dtypes.uint64)
+  # extract 96 relevant bits of 2/pi based on magnitude of argument
+  i = shr(e.cast(dtypes.uint64), 5)
+  e = e.cast(dtypes.int32) & 31
+  offset = 32 - e
+
+  def _take(an:UOp, offset:int, count:int=0) -> UOp:
+    """an = two_over_pi_f[i+offset]"""
+    if count+offset < len(two_over_pi_f) - 1:
+      an = i.ne(count).where(_take(an, offset, count=count+1), an.const_like(two_over_pi_f[count+offset]))
+    return an
+  def _shl_lazy(x:UOp, y:UOp): return (x.cast(dtypes.uint64) * pow2if(y, d.dtype).cast(dtypes.uint64)).cast(dtypes.uint32)
+  def _shr_lazy(x:UOp, y:UOp): return (x.cast(dtypes.uint64) // pow2if(y, d.dtype).cast(dtypes.uint64)).cast(dtypes.uint32)
+
+  a = [_take(UOp.const(dtypes.uint32.vec(d.dtype.count), 0), i) for i in range(4)]
+  #  (two_over_pi_f[Int(i) + n] << e) | (two_over_pi_f[Int(i) + n+1] >> (nbits - e))
+  # Note: e >= 1 for all numbers d >= 1.0. assume e != 0
+  hi = _shl_lazy(a[0], e) | _shr_lazy(a[1], offset)
+  mi = _shl_lazy(a[1], e) | _shr_lazy(a[2], offset)
+  lo = _shl_lazy(a[2], e) | _shr_lazy(a[3], offset)
+
+  def _hp_mul(x:UOp, y:UOp) -> UOp: return x.cast(dtypes.uint64) * y.cast(dtypes.uint64)
+  # compute x * 2/pi
+  p = shl(_hp_mul(ia, hi), 32) + _hp_mul(ia, mi) + shr(_hp_mul(ia, lo), 32)
+
+  # round quotient to nearest
+  q = shr(p, 62).cast(dtypes.int32)
+  p = p & 0x3fffffffffffffff
+  r = (p.cast(intermediate_dtype) * (3.4061215800865545e-19)).cast(d.dtype)
+
+  # if fraction >= 0.5, r -= pi/2, q += 1
+  return (f<0.5).where(r, r - math.pi/2), (f<0.5).where(q, q + 1)
+
+def cody_waite_reduction(d:UOp) -> tuple[UOp, UOp]:
+  """
+  Performs Cody-Waite Reduction: computes the reminder of `d` modulo pi/2 for the values `d` where
+      0 <= abs(d) <= 39800.0
+  Returns a tuple of `(r, q)`, where the output format is the same as that of `payne_hanek_reduction`.
+  """
+  def _reduce_d(x:UOp, q:UOp):
+    # https://github.com/shibatch/sleef/blob/4e08851f59fc2b545f9c393c6a23dfd311a26308/src/libm/sleefdp.c#L789-L823
+    if x.dtype.scalar() == dtypes.float64:
+      # https://github.com/shibatch/sleef/blob/f6d8a841fbfddd26ce712834d4da220cd76048fb/src/common/misc.h#L77
+      PI_A, PI_B, PI_C, PI_D = 3.1415926218032836914, 3.1786509424591713469e-08, 1.2246467864107188502e-16, 1.2736634327021899816e-24
+      d = qdh * -PI_A + x
+      d = q * -PI_A + d
+      d = qdh * -PI_B + d
+      d = q * -PI_B + d
+      d = qdh * -PI_C + d
+      d = q * -PI_C + d
+      d = (qdh + q) * -PI_D + d
+    elif x.dtype.scalar() == dtypes.float16:
+      # [FIXME] when reducing `d`, FP16 needs FP32 precision to achieve 1.0 ULP precision.
+      d = _reduce_d(x.cast(dtypes.float32), q.cast(dtypes.float32)).cast(dtypes.float16)
+    else:
+      # https://github.com/shibatch/sleef/blob/4e08851f59fc2b545f9c393c6a23dfd311a26308/src/libm/sleefsp.c#L464-L503
+      d = q * -3.1414794921875 + x
+      d = q * -0.00011315941810607910156 + d
+      d = q * -1.9841872589410058936e-09 + d
+      d = q * -1.2154201256553420762e-10 + d
+    return d
+
+  m_1_pi = 0.318309886183790671537767526745028724
+  qdh = (d * (m_1_pi / 2.0**24)).cast(dtypes.int64).cast(d.dtype) * (2.0**24)
+  quadrant = rintk(d * m_1_pi -qdh) if d.dtype.base.scalar() == dtypes.float64 else rintk(d * m_1_pi)
+  return _reduce_d(d, quadrant.cast(d.dtype)), quadrant.cast(dtypes.int32)
+
+# *** approximate sine on small angle. ***
+def trig_poly(d:UOp, coeff32, coeff64): return d * (polyN(d*d, coeff64) if d.dtype.scalar() == dtypes.float64 else polyN(d*d, coeff32))
+# approximate sine on [-pi/2, pi/2]
+def sin_poly(d:UOp) -> UOp:
+  return trig_poly(d, [2.6083159809786593541503e-06, -0.0001981069071916863322258, 0.00833307858556509017944336, -0.166666597127914428710938, 1.0],
+                      [-7.97255955009037868891952e-18, 2.81009972710863200091251e-15, -7.64712219118158833288484e-13, 1.60590430605664501629054e-10,
+                       -2.50521083763502045810755e-08, 2.75573192239198747630416e-06, -0.000198412698412696162806809, 0.00833333333333332974823815,
+                       -0.166666666666666657414808,    1.0])
+
+def _ifand(q:UOp, n:int): return (q & n).ne(0)
+
+def sin_poly_small(d:UOp, q:UOp) -> UOp:
+  r = sin_poly(d)
+  return r * _ifand(q, 1).where(r.const_like(-1), r.const_like(1))
+
+def sin_poly_large(d:UOp, q:UOp) -> UOp:
+  r = sin_poly(d + _ifand(q, 1).where(d.const_like(math.pi / 2), d.const_like(0)))
+  return r * _ifand(q, 2).where(r.const_like(-1), r.const_like(1))
+
+# *** toplevel functions for xsin/xlog2/xexp2 ***
+
+def xsin(d:UOp, fast:bool=False, switch_over:float=30.0) -> UOp:
+  """
+  Implements a 1.0 ULP approximation for Ops.SIN.
+  - fast=True assumes x <= switch_over.
+  - switch_over is the threshold for switching to payne_hanek_reduction.
+  """
+  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
+  # mask +-inf/nan as zero
+  x = _lazy_map_numbers(d, d.const_like(0.0), d.const_like(0.0), d.const_like(0.0), d)
+  # x_sign = sign(x)
+  x_sign = x.ne(0).where((x<0).where(x.const_like(-1), x.const_like(1)), x.const_like(0))
+  x_abs = x * x_sign
+  r, q = (cody_waite_reduction if fast else payne_hanek_reduction)(x_abs)
+  if fast: result = sin_poly_small(r, q)
+  else:
+    # Payne Hanek Reduction assumes abs(x) >= pi/4, so for smaller values, use cody_waite_reduction.
+    r_small, q_small = cody_waite_reduction(x_abs)
+    result = (x_abs<switch_over).where(sin_poly_small(r_small, q_small), sin_poly_large(r, q))
+  # adjusts the sign for abs(x)
+  result = result * x_sign
+  # sin(Inf) = NaN, sin(-Inf) = NaN, sin(NaN) = NaN
+  return _lazy_map_numbers(d, d.const_like(math.nan), d.const_like(math.nan), d.const_like(math.nan), result)
+
+def xexp2(d:UOp) -> UOp:
+  """
+  Implements a 1.0 ULP approximation for Ops.EXP2
+  - Paper: https://arxiv.org/pdf/2001.09258
+  """
+  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
+  # mask +=inf/nan as zero.
+  x = _lazy_map_numbers(d, d.const_like(0.0), d.const_like(0.0), d.const_like(0.0), d)
+  q = rintk(x)
+  # s = d - round(d)
+  s = x - q.cast(x.dtype)
+  # a polynomial approximation with 13 non-zero terms in the range of [−(log 2)/2,(log 2)/2].
+  if d.dtype.scalar() == dtypes.float64:
+    u = polyN(s, [0.4434359082926529454e-9, 0.7073164598085707425e-8, 0.1017819260921760451e-6, 0.1321543872511327615e-5, 0.1525273353517584730e-4,
+                  0.1540353045101147808e-3, 0.1333355814670499073e-2, 0.9618129107597600536e-2, 0.5550410866482046596e-1, 0.2402265069591012214e+0,
+                  0.6931471805599452862e+0, 0.1000000000000000000e+1])
+  else: u = polyN(s, [0.1535920892e-3, 0.1339262701e-2, 0.9618384764e-2, 0.5550347269e-1, 0.2402264476e+0, 0.6931471825e+0, 1.0])
+  u = ldexp2k(u, q) # u*2^q
+  upper, lower = {dtypes.float64: (1024, -2000), dtypes.float32: (128, -150), dtypes.float16: (23, -22)}[d.dtype.scalar()]
+  # Replace x >= upper with +inf
+  u = (d >= upper).where(d.const_like(math.inf), u)
+  # Replace x < lower with zero.
+  u = (d<lower).where(d.const_like(0.0), u)
+  # exp2(NaN) = NaN
+  return d.ne(d).where(d.const_like(math.nan), u)
+
+def xlog2(d:UOp) -> UOp:
+  """
+  Implements a 1.0 ULP approximation for Ops.LOG2
+  Paper: https://arxiv.org/pdf/2001.09258 5.5
+  """
+  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
+  # float16 uses 2^10 for denormal scaling (2^64 overflows), float32/64 use 2^64
+  denormal_exp = 10 if d.dtype.scalar() == dtypes.float16 else 64
+  FLT_MIN = d.const_like({dtypes.float16: 6.1e-5, dtypes.float32: 1e-4, dtypes.float64: 1e-4}[d.dtype.scalar()])
+  is_denormal = d<FLT_MIN
+  a = is_denormal.where(d * (2 ** denormal_exp), d)
+
+  e = ilogb2k(a * (1.0 / 0.75)).cast(a.dtype)
+  m = ldexp3k(a, -e)
+  e = is_denormal.where(e - denormal_exp, e)
+
+  x = (m - 1.0) / (m + 1.0)
+  x2 = x * x
+  if d.dtype.scalar() == dtypes.float64:
+    t = polyN(x2, [0.2211941750456081490e+0, 0.2200768693152277689e+0, 0.2623708057488514656e+0, 0.3205977477944495502e+0,
+                   0.4121985945485324709e+0, 0.5770780162997058982e+0, 0.96179669392608091449])
+    r = t * (x * x2) + e + x * 2.885390081777926774
+  else:
+    t = polyN(x2, [0.4374550283e+0, 0.5764790177e+0, 0.9618012905120])
+    # s_lo term (x*3.27e-08) only for float32 - underflows in float16
+    r = t * (x * x2) + e + x * 2.8853900432586669922 + (x * 3.2734474483568488616e-08 if d.dtype.scalar() == dtypes.float32 else 0)
+
+  # log2(Inf) = Inf
+  r = d.ne(math.inf).where(r, r.const_like(math.inf))
+  # log2(0) = -Inf (handle both +0.0 and -0.0)
+  r = d.ne(0.0).where(r, r.const_like(-math.inf))
+  # log2(x) = NaN for x < 0
+  r = (d<-0.0).where(r.const_like(math.nan), r)
+  # log2(NaN) = NaN
+  r = d.ne(d).where(r.const_like(math.nan), r)
+  # log2(-0.0) = -Inf. In certain devices like PTX, x == -0.0 won't be true. so making reciprocal.
+  return d.reciprocal().ne(-math.inf).where(r, r.const_like(-math.inf))
+
+def xpow(base:UOp, exponent:UOp) -> UOp:
+  # start with b ** e = exp2(e * log2(b))
+  ret = (base < 0).where(-base, base).log2().mul(exponent).exp2()
+  # negative base: nan for non-integer exponent, negate for odd integer exponent
+  non_int = exponent != exponent.cast(dtypes.int32).cast(exponent.dtype)
+  is_odd = (exponent < 0).where(-exponent, exponent).cast(dtypes.int32).mod(2).cast(dtypes.bool)
+  neg_base = non_int.where(ret.const_like(math.nan), is_odd.where(-ret, ret))
+  # fix 0 ** 0 = 1
+  return (base.eq(0) & exponent.eq(0)).where(ret.const_like(1), (base < 0).where(neg_base, ret))
+
+@functools.cache
+def get_transcendental_patterns(ops:tuple[Ops, ...], force_transcendental:bool) -> PatternMatcher:
+  pat: list[tuple[UPat, Callable]] = []
+  for op,f in ((Ops.EXP2, xexp2), (Ops.LOG2, xlog2), (Ops.SIN, xsin)):
+    if op not in ops or force_transcendental:
+      pat += [(UPat(op, dtype=TRANSCENDENTAL_DTYPES, src=(UPat.var("d"),)), f),
+              (UPat(op, dtype=tuple(dt for dt in dtypes.floats if dt not in TRANSCENDENTAL_DTYPES), src=(UPat.var("d"),), name="x"),
+                lambda x,d: d.cast(dtypes.float32).alu(x.op).cast(x.dtype))]
+  # rewrite SQRT to xpow 0.5
+  if Ops.SQRT not in ops or force_transcendental: pat.append((UPat(Ops.SQRT, src=UPat.var("d")), lambda d: xpow(d, d.const_like(0.5))))
+  return PatternMatcher(pat)
@@ -0,0 +1,74 @@
+from typing import Any
+import itertools
+from collections import defaultdict
+from tinygrad.dtype import dtypes, AddrSpace, Invalid, ImageDType
+from tinygrad.uop.ops import UOp, Ops
+from tinygrad.helpers import getenv
+from tinygrad.renderer import Renderer
+
+def memory_coalesing(sink:UOp, ctx:Renderer) -> UOp:
+  if getenv("DMC"): return sink
+
+  # collect
+  memory: defaultdict[tuple[Ops, UOp, Any, Any], dict[int, list[UOp]]]  = defaultdict(dict)
+  for u in sink.toposort():
+    # TODO: this should handle images too, it's just memory coalesing
+    if u.op in {Ops.LOAD, Ops.STORE} and not isinstance(u.src[0].src[0].dtype, ImageDType):
+      assert len(u.src) == (2 if u.op is Ops.STORE else 1), "memory coalesing does not support gated loads/stores"
+      if u.src[0].op is not Ops.INDEX: continue
+      buf, idx_u = u.src[0].src
+      if buf.addrspace == AddrSpace.REG: continue
+      idx: Any = idx_u.src[1] if idx_u.op is Ops.WHERE and idx_u.src[2].arg is Invalid else idx_u
+      valid: Any = idx_u.src[0] if idx_u.op is Ops.WHERE and idx_u.src[2].arg is Invalid else None
+      if idx.op is Ops.ADD and idx.src[1].op is Ops.CONST: root_src, arg = idx.src[0], idx.src[1].arg
+      elif idx.op is Ops.ADD and idx.src[0].op is Ops.CONST: root_src, arg = idx.src[1], idx.src[0].arg
+      elif idx.op is Ops.CONST and idx.arg is Invalid: root_src, arg = "INVALID", 0
+      elif idx.op is Ops.CONST: root_src, arg = "CONST", idx.arg
+      else: root_src, arg = idx, 0
+      memory[(u.op, buf, root_src, valid)].setdefault(arg, []).append(u)
+
+  # build replacements
+  replacements = {}
+  for (op,buf,base,valid),offsets in memory.items():
+    # allowed lengths (copied in)
+    lengths = []
+    must_divide = True
+    if ctx is not None and ctx.target.device == "DSP":
+      lengths = [128,64,32,16,8,4]
+      must_divide = False
+    elif buf.dtype.base not in (dtypes.float, dtypes.half, *dtypes.fp8s) and not isinstance(buf.dtype, ImageDType):
+      pass
+    elif buf.addrspace == AddrSpace.REG:
+      pass
+    elif isinstance(buf.dtype, ImageDType):
+      lengths = [4]
+    elif ctx is not None and ctx.supports_float4:
+      # TODO: a better way to get this than ctx
+      lengths = [8,4,2] if buf.dtype.base == dtypes.half and getenv("ALLOW_HALF8") else [4,2]
+    lengths.append(1)  # worst case, it's not folded
+    # do the grouping
+    grouped_offsets = [[x for _,x in group] for _,group in itertools.groupby(enumerate(sorted(offsets.keys())), lambda x: x[1]-x[0])]
+    for full_grp in grouped_offsets:
+      while len(full_grp):
+        offset = (base+full_grp[0]) if isinstance(base, UOp) else UOp.const(dtypes.int, full_grp[0])
+        length = [l for l in lengths if l <= len(full_grp) and (not must_divide or offset.divides(l) is not None)][0]
+        grp = full_grp[:length]
+        # NOTE: we apply the valid again after we determine the length
+        offset = valid.where(offset, UOp(Ops.CONST, offset.dtype, arg=Invalid)) if valid is not None else offset
+        idx = UOp(Ops.SHRINK, dtype=buf.dtype, src=(buf, offset, UOp.const(dtypes.int, len(grp)))) if len(grp) > 1 else buf.index(offset)
+        if op == Ops.STORE:
+          datas = []
+          for i,g in enumerate(grp):
+            assert len(offsets[g]) == 1, f"attempting multiple stores: {len(offsets[g])}"
+            datas.append(offsets[g][0].src[1])
+          store = idx.store(UOp.vectorize(*datas) if len(datas) > 1 else datas[0])
+          for i,g in enumerate(grp): replacements[offsets[g][0]] = store
+        else:
+          ld = idx.load()
+          for i,g in enumerate(grp):
+            for oo in offsets[g]:
+              replacements[oo] = ld.index(UOp.const(dtypes.int, i)) if len(grp) > 1 else ld
+        full_grp = full_grp[length:]
+
+  # apply
+  return sink.substitute(replacements, name="memory coalesing")
@@ -5,7 +5,7 @@ from dataclasses import dataclass
 from tinygrad.dtype import dtypes, ImageDType, DType, AddrSpace, Invalid, PtrDType
 from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, GroupOp, identity_element
 from tinygrad.uop.symbolic import uop_given_valid, parse_valid, invalid_gate
-from tinygrad.helpers import getenv, flatten, prod
+from tinygrad.helpers import getenv, flatten, prod, OSX, ceildiv
 from tinygrad.renderer import Renderer

 # ***** image load valid simplification *****
@@ -14,7 +14,7 @@ from tinygrad.renderer import Renderer
 def _drop_valid_stmts(valid:UOp, idx:UOp, height:int, width:int) -> list[UOp]:
  # can drop valid if idx is out of bound when valid is False
  drop_stmt = []
-  for stmt in valid.split_uop(Ops.AND):
+  for i,stmt in enumerate(valid.split_uop(Ops.AND)):
    if (res:=parse_valid(stmt)) is None: continue
    X, is_upper_bound, c = res

@@ -25,12 +25,12 @@ def _drop_valid_stmts(valid:UOp, idx:UOp, height:int, width:int) -> list[UOp]:
        drop_stmt.append(stmt)
        continue

-    # if X <= c, check if it's out of bound when X = c+1
-    # if X >= c, check if it's out of bound when X = c-1
-    test_value = c + 1 if is_upper_bound else c - 1
-    for i,b in zip(idx.src, (width, height)):
-      if i.is_increasing():
-        rw = i.substitute({X:X.const_like(test_value)})
+    # check if idx is out of bound when X is on the wrong side of the bound: X in [c+1, vmax] or [vmin, c-1]
+    lo, hi = (c + 1, X.vmax) if is_upper_bound else (X.vmin, c - 1)
+    if lo <= hi:
+      fake = UOp.variable(f"fake{i}", lo, hi, X.dtype)
+      for coord,b in zip(idx.src, (width, height)):
+        rw = coord.substitute({X:fake}).simplify()
        if rw.vmin >= b or rw.vmax < 0:
          drop_stmt.append(stmt)
          break
@@ -60,6 +60,15 @@ load_store_indexing = PatternMatcher([

 # ***** load/store grouping *****

+# get list of (height, width) that do not require pitch padding
+def image_valid_dims(base:DType, size:int, arch:str) -> list[tuple[int,int]]:
+  if (ALIGN:=next((int(p.split('=')[1]) for p in arch.split(',') if p.startswith("IMAGE_PITCH_ALIGNMENT=")), 0)) == 0: return []
+  MAXW, pxls = 16384, size // 4
+  if base not in (dtypes.half, dtypes.float) or size > 4*MAXW*MAXW: return []
+  # height=1 images just need to abide by alignment requirements in bytes, not pixels!
+  if size % (ALIGN * 4) != 0: return [] if (base.itemsize * size) % (64 if OSX else ALIGN) != 0 or pxls > MAXW else [(1, pxls)]
+  return [(pxls//ALIGN//k, ALIGN*k) for k in range(ceildiv(pxls//ALIGN, MAXW), min(pxls//ALIGN, MAXW//ALIGN)+1) if (pxls//ALIGN)%k == 0]
+
 def expand_index(ctx, buf:UOp, vec:UOp):
  # determine optimal image shapes
  if isinstance(dt:=buf.dtype, ImageDType):
@@ -68,7 +77,7 @@ def expand_index(ctx, buf:UOp, vec:UOp):
    x, valid = idxs.gep(lane), valids.gep(lane)
    # search for dims that drop the most valid statements
    best_drop, cands = -1, []
-    for ch, cw in ImageDType.valid_dims(dt, ctx.target.arch):
+    for ch, cw in image_valid_dims(dt.base, dt.size, ctx.target.arch):
      if (dropped:=len(_drop_valid_stmts(valid, cidx:=uop_given_valid(valid, UOp.vectorize((x//4)%cw, x//(4*cw))), ch, cw))) > best_drop:
        best_drop, cands = dropped, [(ch, cw, cidx)]
      elif dropped == best_drop: cands.append((ch, cw, cidx))
@@ -162,18 +171,8 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
  # determine fold lengths
  lengths = []
  must_divide = True
-  if ctx is not None and ctx.target.device == "DSP":
-    lengths = [128,64,32,16,8,4]
-    must_divide = False
-  elif buf.dtype.base not in (dtypes.float, dtypes.half, *dtypes.fp8s) and not isinstance(buf.dtype, ImageDType):
-    pass
-  elif buf.addrspace == AddrSpace.REG:
-    pass
-  elif isinstance(buf.dtype, ImageDType):
-    lengths = [4]
-  elif ctx is not None and ctx.supports_float4:
-    # TODO: a better way to get this than ctx
-    lengths = [8,4,2] if buf.dtype.base == dtypes.half and getenv("ALLOW_HALF8") else [4,2]
+  # TODO: this belongs in coalese
+  if isinstance(buf.dtype, ImageDType): lengths = [4]
  lengths.append(1)  # worst case, it's not folded

  # filter fold lengths that don't divide
@@ -384,7 +383,7 @@ pm_imageh_store = PatternMatcher([

 def make_image(ctx, ls, buf, off):
  if (vcount:=buf.dtype.vcount) != 1: buf = buf.src[0]
-  if buf.op == Ops.PARAM and not isinstance(dt:=buf.dtype, ImageDType) and (dims:=ImageDType.valid_dims(dt, ctx)):
+  if buf.op == Ops.PARAM and not isinstance(dt:=buf.dtype, ImageDType) and (dims:=image_valid_dims(dt.base, dt.size, ctx)):
    buf = buf.replace(dtype=(dtypes.imageh if dt.base == dtypes.half else dtypes.imagef)((*dims[0], 4))).flatten()
    if vcount != 1: buf = UOp.vectorize(*([buf] * vcount))
    if ls.op is Ops.LOAD: return ls.replace(src=(buf.index(off, ptr=True),), dtype=dtypes.float.vec(ls.dtype.vcount)).cast(dt.base)
@@ -1,8 +1,9 @@
 import itertools
 from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
 from tinygrad.helpers import getenv, DEBUG, prod, NOLOCALS, TC_OPT, TC_SELECT, USE_TC, IMAGE
-from tinygrad.dtype import PtrDType, ImageDType
+from tinygrad.dtype import PtrDType
 from tinygrad.uop.ops import Ops, resolve, AxisType
+from tinygrad.codegen.late.devectorizer import image_valid_dims
 from tinygrad.codegen.opt.postrange import Scheduler

 def hand_coded_optimizations(k:Scheduler) -> Scheduler:
@@ -50,7 +51,7 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
  # upcast float4 images, this must be early so we don't accidentally add locals before the upcast
  if IMAGE:
    for buf_index,buf in enumerate(k.bufs):
-      if isinstance(buf.src[0].dtype, PtrDType) and ImageDType.valid_dims(buf.src[0].dtype, k.ren.target.arch):
+      if isinstance(buf.src[0].dtype, PtrDType) and image_valid_dims(buf.src[0].dtype.base, buf.src[0].dtype.size, k.ren.target.arch):
        # part of is_expanded
        unit_stride_axes_mul_4 = [k.rngs.index(c) for c in k.bufs[buf_index].src[1].get_idx().split_uop(Ops.ADD) if
          c.op is Ops.RANGE and (c.vmax+1)%4 == 0]
@@ -101,6 +102,12 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
    # for Schedule, we check if the range is used in INDEX gates or WHERE gates
    is_masked = k.rngs[axis] in where_gate_rngs
    if k.full_shape[axis] <= 7 and is_masked and prod(k.full_shape[j] for j in to_upcast) * k.full_shape[axis] <= 7 * 7:
+      # upcasting a masked global axis moves that range out of the launch grid into each work-item
+      # under IMAGE, skip the upcast unless enough global work-items remain after it to hide memory latency
+      if IMAGE and k.axis_types[axis] is AxisType.GLOBAL:
+        global_upcast = prod(k.full_shape[i] for i in to_upcast if k.axis_types[i] is AxisType.GLOBAL) * k.full_shape[axis]
+        global_items_after = prod(k.full_shape[i] for i in k.axes_of(AxisType.GLOBAL)) // global_upcast
+        if resolve(global_items_after < getenv("OCCUPANCY_FLOOR", 4096), False): continue
      if DEBUG >= 4: print(f"upcasting masked axis : {axis}")
      to_upcast.append(axis)
  for axis in to_upcast[::-1]: k.apply_opt(Opt(OptOps.UPCAST, axis, 0))
@@ -2,7 +2,7 @@ from __future__ import annotations
 from typing import Final, ClassVar, Callable, Literal
 import math, struct, ctypes, functools
 from dataclasses import dataclass, fields
-from tinygrad.helpers import ceildiv, getenv, prod, round_up, OSX
+from tinygrad.helpers import getenv, prod, round_up, OSX
 from enum import IntEnum, auto

 class ConstFloat(float):
@@ -136,15 +136,6 @@ class ImageDType(PtrDType):
  @property
  def pitch(self): return (round_up(self.shape[1], 256) if OSX else self.shape[1]) * 4 * self.itemsize

-  # get list of (height, width) that do not require pitch padding
-  @staticmethod
-  def valid_dims(ptr:PtrDType, arch:str) -> list[tuple[int,int]]:
-    if (ALIGN:=next((int(p.split('=')[1]) for p in arch.split(',') if p.startswith("IMAGE_PITCH_ALIGNMENT=")), 0)) == 0: return []
-    MAXW, pxls = 16384, ptr.size // 4
-    if ptr.base not in (dtypes.half, dtypes.float) or ptr.size > 4*MAXW*MAXW: return []
-    # height=1 images just need to abide by alignment requirements in bytes, not pixels!
-    if ptr.size % (ALIGN * 4) != 0: return [] if ptr.nbytes() % (64 if OSX else ALIGN) != 0 or pxls > MAXW else [(1, pxls)]
-    return [(pxls//ALIGN//k, ALIGN*k) for k in range(ceildiv(pxls//ALIGN, MAXW), min(pxls//ALIGN, MAXW//ALIGN)+1) if (pxls//ALIGN)%k == 0]

 class dtypes:
  @staticmethod
@@ -281,4 +281,4 @@ def time_call(call:UOp, var_vals:dict[str, int]|None=None, timeout:int|None=None
      from tinygrad.tensor import Tensor
      with Context(DEBUG=0, BEAM=0, CAPTURING=0, TRACK_MATCH_STATS=0): Tensor.ones(1024, 1024).contiguous().realize(do_update_stats=False)
  call = compile_linear(UOp(Ops.LINEAR, src=(call,)), beam=0).src[0]
-  return cast(float, pm_exec.rewrite(call, ExecContext(var_vals or {}, update_stats=False, wait=True, timeout=timeout, cache=False)))
+  return pm_exec.rewrite(call, ExecContext(var_vals or {}, update_stats=False, wait=True, timeout=timeout, cache=False))
@@ -1,12 +1,12 @@
 from __future__ import annotations
 import functools, itertools, string
-from typing import TYPE_CHECKING, Callable, Self, Sequence, Literal, get_args, cast
+from typing import TYPE_CHECKING, Callable, Self, Sequence, Literal, get_args
 from tinygrad.mixin.elementwise import ElementwiseMixin
 from tinygrad.mixin.movement import MovementMixin
 from tinygrad.mixin.reduce import ReduceMixin
 from tinygrad.uop import Ops
 from tinygrad.uop.ops import _broadcast_shape, resolve, smax, smin, identity_element
-from tinygrad.dtype import ConstType, DTypeLike, PtrDType, PyConst, dtypes, least_upper_dtype, sum_acc_dtype, to_dtype
+from tinygrad.dtype import ConstType, DTypeLike, Invalid, PtrDType, PyConst, dtypes, least_upper_dtype, sum_acc_dtype, to_dtype
 from tinygrad.helpers import all_int, argfix, argsort, ceildiv, flatten, flat_to_grouped, fully_flatten, get_shape, make_tuple, merge_dicts, prod
 from tinygrad.helpers import resolve_pool_pads, round_up

@@ -285,7 +285,7 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
    pads = tuple((smax(pB,0), smax(pA,0)) for pB,pA in pX) if has_neg else pX
    base = MovementMixin.pad(X, pads)
    if value == 0: return base
-    base = base.cast(least_upper_dtype(base.dtype, dtypes.from_py(value)))
+    if value is not Invalid: base = base.cast(least_upper_dtype(base.dtype, dtypes.from_py(value)))
    return MovementMixin.pad(X.const_like(1).cast(dtypes.bool), pads).where(base, base.const_like(value))

  def _pad_circular(self, pX:tuple[tuple[sint, sint], ...]) -> Self:
@@ -441,7 +441,7 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
    for i, (s, x) in enumerate(zip(inputs, xs)):
      for c in set(s):
        while s.count(c) > 1:
-          j, k, n = s.index(c), s.index(c, s.index(c)+1), cast(int, x.shape[s.index(c)])
+          j, k, n = s.index(c), s.index(c, s.index(c)+1), x.shape[s.index(c)]
          perm = [d for d in range(x.ndim) if d not in (j,k)]+[j,k]
          x = x.permute(perm).flatten(-2).pad(((0,0),)*(x.ndim-2)+((0,n),)).unflatten(-1,(n,n+1))[...,0] if x.ndim > 2 else x.diagonal()
          s = s[:k] + s[k+1:]
@@ -360,7 +360,7 @@ required_input_python_consts: dict[str, tuple[int, ...]] = {
 }

 def _to_python_const(t:Tensor) -> list[ConstType]|ConstType|bytes:
-  return t.data().tobytes() if t.dtype == dtypes.uint8 else cast(list[ConstType]|ConstType, t.tolist())
+  return t.data().tobytes() if t.dtype == dtypes.uint8 else t.tolist()

 # ***** runner ******
 debug = int(getenv("DEBUGONNX", "0"))
@@ -1188,7 +1188,9 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
  def ScatterElements(x: Tensor, indices: Tensor, updates: Tensor, axis=0, reduction:Literal["none", "add", "mul", "min", "max"]="none"):
    indices = (indices < 0).where(x.shape[axis], 0) + indices
    if reduction == "none": return x.scatter(axis, indices, updates)
-    reduction_ = cast(Literal["sum", "prod", "amin", "amax"], {"add": "sum", "mul": "prod", "min": "amin", "max": "amax"}[reduction])
+    reduction_map: dict[Literal["add", "mul", "min", "max"], Literal["sum", "prod", "amin", "amax"]] = \
+      {"add": "sum", "mul": "prod", "min": "amin", "max": "amax"}
+    reduction_ = reduction_map[reduction]
    return x.scatter_reduce(axis, indices, updates, reduction_)
  def GatherElements(x:Tensor, indices:Tensor, axis:int):
    indices = (indices < 0).where(x.shape[axis], 0) + indices
@@ -1,4 +1,4 @@
-from typing import Literal, Callable, cast
+from typing import Literal, Callable
 import math, sys, struct
 from collections import defaultdict, Counter
 from tinygrad.codegen.opt import tc
@@ -22,6 +22,7 @@ base_rewrite = PatternMatcher([
  (UPat(Ops.CAST, name="x"), lambda ctx,x: f"__builtin_convertvector({ctx[x.src[0]]}, {ctx.render_type(x)})" \
    if x.max_numel() > 1 and x.addrspace is AddrSpace.REG else None),
  (UPat(Ops.CAST, name="x"), lambda ctx,x: f"({ctx.render_cast(x, ctx[x.src[0]])})"),
+  (UPat(Ops.BITCAST, name="x"), lambda ctx,x: ctx[x.src[0]] if x.addrspace in (AddrSpace.GLOBAL, AddrSpace.LOCAL) else None),
  (UPat(Ops.BITCAST, name="x"), lambda ctx,x: f"__builtin_bit_cast({ctx.render_type(x)}, ({ctx.render_type(x.src[0])})({ctx[x.src[0]]}))"),

  # GPU stuff
@@ -222,7 +223,7 @@ class CStyleLanguage(Renderer):
                  Ops.INDEX: "bidx", Ops.LOAD: "val"}.get(u.op, "alu")
        r[u] = f"{prefix}{c[prefix]}"

-      l = cast(str, self.string_rewrite.rewrite(u, ctx=self))
+      l: str|None = self.string_rewrite.rewrite(u, ctx=self)
      assert l is not None, f"failed to render {u.op} {u.dtype} {[(x.op,x.dtype) for x in u.src]} {u.arg}"

      if u.op in {Ops.ENDIF, Ops.END}: depth -= 1
@@ -1,9 +1,8 @@
-from typing import cast
 import math, struct, sys
 from tinygrad.codegen.opt import tc
 from tinygrad.renderer import Renderer
 from tinygrad.renderer.cstyle import HIPRenderer, create_non_native_float_pats, pm_manual_bf16_cast
-from tinygrad.uop.decompositions import xexp2, xlog2
+from tinygrad.codegen.decomp.transcendental import xexp2, xlog2
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, GroupOp, range_str
 from tinygrad.dtype import dtypes, float_to_fp8, DType, PtrDType, truncate, AddrSpace
 from tinygrad.helpers import prod, Target, CPU_COUNT, getenv, OSX
@@ -170,9 +169,10 @@ class LLVMRenderer(Renderer):
          r[u] = f"%v{vc}"

        # do the rendering of the llvm ir code
-        if (l:=self.string_rewrite.rewrite(u, ctx=r)) is None:
+        l: str|None = self.string_rewrite.rewrite(u, ctx=r)
+        if l is None:
          raise RuntimeError(f"failed to render {u.op} with {u.dtype} srcs {[x.dtype for x in u.src]}")
-        kernel.append(cast(str, l))
+        kernel.append(l)
    return tuple(local_args), self._render_fn(name, args, kernel, prefix)

 class CPULLVMRenderer(LLVMRenderer):
@@ -1,4 +1,4 @@
-from typing import Callable, cast, Any
+from typing import Callable, Any
 from tinygrad.dtype import AddrSpace, DType, ImageDType, dtypes, truncate
 from tinygrad.helpers import DEBUG, OSX, unwrap, fromimport, Target
 from tinygrad.renderer import Renderer
@@ -216,8 +216,9 @@ class NIRRenderer(Renderer):
        nstore(self.b, AddrSpace.REG, ranges.pop(), next_i),
        mesa.nir_pop_loop(self.b, None)
      else:
-        if (d:=self.def_rewrite.rewrite(u, ctx=self)) is None: raise RuntimeError(f"failed to render {u.op} srcs {[x.dtype for x in u.src]}")
-        self.r[u] = cast(mesa.nir_def, d)
+        d: mesa.nir_def|None = self.def_rewrite.rewrite(u, ctx=self)
+        if d is None: raise RuntimeError(f"failed to render {u.op} srcs {[x.dtype for x in u.src]}")
+        self.r[u] = d
    self.postrender(uops)

    mesa.nir_validate_shader(self.b.shader, b"after render")
@@ -213,7 +213,8 @@ class PTXRenderer(Renderer):
        Ops.PARAM: ("dat", "u64" if u.addrspace is AddrSpace.GLOBAL else None), **{op: ("alu", None) for op in GroupOp.ALU}}.get(u.op, (None, None))
      if prefix: r[u] = ssa(prefix, u, dtype)

-      if (l:=cast(str|list[str], string_rewrite.rewrite(u, ctx=self))) is None:
+      l: str|list[str]|None = string_rewrite.rewrite(u, ctx=self)
+      if l is None:
        raise RuntimeError(f"failed to render {u.op} with {u.dtype} srcs {[x.dtype for x in u.src]}")
      kernel.extend([l] if isinstance(l, str) else l)

@@ -21,7 +21,7 @@ class CUDAGraph(MultiGraphRunner):

        c_deps, new_node = self.new_node([b.base for b in bufs], ast.arg.outs)
        c_args, vargs = encode_args([b._buf for b in bufs], [device_vars.get(x.expr, 0) for x in ast.arg.vars])
-        kern_params = cuda.CUDA_KERNEL_NODE_PARAMS_v1(runtime.prg, *global_size, *local_size, 0,
+        kern_params = cuda.CUDA_KERNEL_NODE_PARAMS_v1(runtime.prg, *global_size, *local_size, runtime.smem,
                                                      ctypes.cast(0, ctypes.POINTER(ctypes.c_void_p)), vargs)
        check(cuda.cuGraphAddKernelNode(ctypes.byref(new_node), self.graph, c_deps, len(c_deps or []), ctypes.byref(kern_params)))

@@ -243,7 +243,7 @@ class NVArgsState(CLikeArgsState):
    if isinstance(prg.dev.iface, MOCKIface): prg.cbuf_0[80:82] = [len(bufs), len(vals)]
    super().__init__(buf, prg, bufs, vals=vals, prefix=prg.cbuf_0 or None)

-class NVProgram(HCQProgram):
+class NVProgram(HCQProgram['NVDevice']):
  def __init__(self, dev:NVDevice, name:str, lib:bytes, **kwargs):
    self.dev, self.name, self.lib = dev, name, lib
    self.constbufs: dict[int, tuple[int, int]] = {0: (0, 0x160)} # dict[constbuf index, tuple[va_addr, size]]
@@ -325,7 +325,7 @@ class NVProgram(HCQProgram):

  def __call__(self, *bufs, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1), vals:tuple[int|None, ...]=(),
               wait=False, timeout:int|None=None):
-    if prod(local_size) > 1024 or self.max_threads < prod(local_size) or self.lcmem_usage > cast(NVDevice, self.dev).slm_per_thread:
+    if prod(local_size) > 1024 or self.max_threads < prod(local_size) or self.lcmem_usage > self.dev.slm_per_thread:
      raise RuntimeError(f"Too many resources requested for launch, {prod(local_size)=}, {self.max_threads=}")
    if any(cur > mx for cur,mx in zip(global_size, [2147483647, 65535, 65535])) or any(cur > mx for cur,mx in zip(local_size, [1024, 1024, 64])):
      raise RuntimeError(f"Invalid global/local dims {global_size=}, {local_size=}")
@@ -105,7 +105,7 @@ class PythonProgram:
          if u.src[0].addrspace == AddrSpace.ALU:
            ret = [src_values[0][i][t] for t,i in enumerate(src_values[1])]
          elif isinstance(src_dtypes[0], ImageDType):
-            assert len(src_values) == 3, "image index must be 3 srcs"
+            assert len(src_values) == 3, f"image index must be 3 srcs, not {len(src_values)}"
            for m,oy,ox in zip(*src_values):
              if ox < 0 or ox >= src_dtypes[0].shape[1] or oy < 0 or oy >= src_dtypes[0].shape[0]: ret.append((m, None))
              else: ret.append((m, ox*4 + oy*src_dtypes[0].shape[1]*4))
@@ -215,7 +215,7 @@ class HWQueue(Generic[SignalType, HCQDeviceType, ProgramType, ArgsStateType]):
      else: self.mv_sints.append((mv, i, self._new_sym(val), mask))

  def _apply_var_vals(self, var_vals:dict[str, int]):
-    resolved_syms = [sym_infer(sym, var_vals) for sym in self.syms]
+    resolved_syms: list[int|None] = [sym_infer(sym, var_vals) for sym in self.syms]

    for off, sym_idx in self.q_sints:
      if self._prev_resolved_syms[sym_idx] == resolved_syms[sym_idx]: continue
@@ -225,7 +225,7 @@ class HWQueue(Generic[SignalType, HCQDeviceType, ProgramType, ArgsStateType]):
      if self._prev_resolved_syms[sym_idx] == resolved_syms[sym_idx]: continue
      mv[off] = resolved_syms[sym_idx] if mask is None else ((mv[off] & ~mask) | resolved_syms[sym_idx])

-    self._prev_resolved_syms = cast(list[int|None], resolved_syms)
+    self._prev_resolved_syms = resolved_syms

  def submit(self, dev:HCQDeviceType, var_vals:dict[str, int]|None=None):
    """
@@ -141,7 +141,7 @@ def apply_movement_op(op:Ops, in_shape:tuple[sint,...], arg:tuple, rngs:tuple[UO
        symbolic+pm_simplify_valid, name="pad").where(r-off, UOp.invalid()) for r,sh,(off,sz) in zip(rngs, in_shape, arg))
    case Ops.RESHAPE:
      sink = UOp.sink(*rngs).simplify() # NOTE: this applies any commutative flips to the rngs early
-      sub_array = {r:UOp.range(r.src[0], i, AxisType.PLACEHOLDER) for i,r in enumerate(sink.ranges)}
+      sub_array = {r:UOp.range(r.src[0], i, AxisType.PLACEHOLDER, dtype=r.dtype) for i,r in enumerate(sink.ranges)}
      rngs = _apply_reshape(in_shape, arg, sink.substitute(sub_array)).substitute({v:k for k,v in sub_array.items()}).src
    case _: raise RuntimeError(f"{op} is not a MovementOp")
  return rngs
@@ -18,7 +18,7 @@ import sys
 sys.setrecursionlimit(10000)

 def add_ranges_to_store(ctx, x):
-  if x.src[0]._shape is None or x.src[1]._shape is None or x.src[0].shape == (): return None
+  if x.src[0]._shape is None or x.src[1]._shape is None or x.src[0].shape == () or x.src[0].max_numel() == x.src[1].max_numel() == 1: return None
  assert x.src[0].shape == x.src[1].shape, "bad store shape"
  idxs = [UOp.range(r, next(ctx), AxisType.LOOP) for r in x.src[0].shape]
  return UOp.store(x.src[0].index(*idxs), x.src[1].index(*idxs)).end(*idxs)
@@ -859,8 +859,7 @@ class Tensor(RandMixin, metaclass=TensorMeta):
    def is_pow2(v): return v > 0 and v & (v - 1) == 0
    # pad dimension i to amt with invalids
    def ipad(t, i, amt):
-      shape = (None,)*i + (amt,) + (None,)*(t.ndim-i-1)
-      return Tensor(True, device=t.device).expand(t.shape).pad_to(shape).where(t.pad_to(shape), Invalid) if amt != t.shape[i] else t
+      return t.pad(tuple(None if d != i else (0, amt-s) for d,s in enumerate(t.shape)), value=Invalid) if amt != t.shape[i] else t
    # align a dimension, use at to specify the dimension to pad in, defaults to first
    def pad_align(t, dim, at=None, force=False):
      # align to 64 pixels when height is real, otherwise 64 bytes is sufficient
@@ -895,13 +894,10 @@ class Tensor(RandMixin, metaclass=TensorMeta):
    # prepare weights
    w = w.permute(0,4,2,5,1,3).reshape((1, 1, 1, *group_shape, *rcout_expand, rcin_hi, rcin_lo, H, W))

-    added_ox = round_up(ox, math.lcm(cout, 64 // dtsz) // cout) - ox
-    if added_ox: x = x.pad_to(None, None, ox + added_ox, None, None, None, None, None, None, None, None)
-
    # the conv!
    ret = (x*w).cast(dtypes.float32).sum((-4, -3, -2, -1), dtype=dtype)

-    ret = ret.reshape(bs, oy, ox + added_ox, groups, rcout)[:, :, :ox, :, :]
+    ret = ret.reshape(bs, oy, ox, groups, rcout)
    # undo hack for non multiples of 4 on C.rcout
    if added_output_channels: ret = ret[:, :, :, :, :-added_output_channels]
    # NCHW output
@@ -1,572 +0,0 @@
-from typing import Callable
-import math, functools
-from tinygrad.dtype import dtypes, DType, promo_lattice, truncate
-from tinygrad.helpers import flatten, polyN, DEBUG, EMULATED_DTYPES
-from tinygrad.uop import GroupOp
-from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, graph_rewrite
-from tinygrad.renderer import Renderer
-
-TRANSCENDENTAL_DTYPES = (dtypes.float16, dtypes.float32, dtypes.float64)
-
-def _lazy_map_numbers(x:UOp, inf:UOp, _inf:UOp, nan:UOp, ratio:UOp):
-  """replace inf -> inf, -inf -> _inf, nan -> nan, otherwise -> ratio"""
-  return x.ne(math.inf).where(x.ne(x).where(nan, x.ne(-math.inf).where(ratio, _inf)), inf)
-
-# *** helper functions for bit manipulation ***
-def mantissa_bits(d:DType) -> int: return dtypes.finfo(d.scalar())[1]
-def exponent_bias(d:DType) -> int: return (1 << (dtypes.finfo(d.scalar())[0] - 1)) - (0 if d.scalar() in dtypes.fp8_fnuz else 1)
-def exponent_mask(d:DType) -> int: return (1 << dtypes.finfo(d.scalar())[0]) - 1
-
-# **** utils ****
-def shr(x:UOp|int, y:UOp|int) -> UOp: return x // (2**(y.simplify().arg) if isinstance(y, UOp) else 2**y)
-def shl(x:UOp|int, y:UOp|int) -> UOp: return x * (2**(y.simplify().arg) if isinstance(y, UOp) else 2**y)
-
-def rintk(d:UOp) -> UOp:
-  """round d:float to int away from 0"""
-  out_dtype = {dtypes.float64: dtypes.int64, dtypes.float32: dtypes.int32, dtypes.float16: dtypes.int16}[d.dtype.scalar()].vec(d.dtype.vcount)
-  return (d + (d<0.0).where(d.const_like(-0.5), d.const_like(0.5))).cast(out_dtype)
-
-def pow2if(q:UOp, float_dtype:DType):
-  """cast(2^q, float_dtype) where q is any integer in the range of [-126, 127]"""
-  out_dtype = {dtypes.int64: dtypes.float64, dtypes.int32: dtypes.float32, dtypes.int16: float_dtype.scalar()}[q.dtype.scalar()].vec(q.dtype.vcount)
-  return shl(q + exponent_bias(out_dtype), mantissa_bits(out_dtype)).bitcast(out_dtype)
-
-def ilogb2k(d:UOp) -> UOp:
-  """calculate the integer part of log2(d), where d is normalized fp value in the range of [0, +inf)."""
-  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
-  dint = d.bitcast({dtypes.float64: dtypes.int64, dtypes.float32: dtypes.int32, dtypes.float16: dtypes.int16}[d.dtype.scalar()].vec(d.dtype.vcount))
-  # -1 <= ilog2bk(d) <= 128
-  return (shr(dint, mantissa_bits(d.dtype)) & exponent_mask(d.dtype)) - exponent_bias(d.dtype)
-
-def ldexp3k(d:UOp, e:UOp) -> UOp:
-  """d*2^e. e is a number obtained by casting an integer in the range [-127, 127] to a float. d is any float number."""
-  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES and e.dtype.scalar() in TRANSCENDENTAL_DTYPES
-  dtype = {dtypes.float64: dtypes.int64, dtypes.float32: dtypes.int32, dtypes.float16: dtypes.int16}[d.dtype.scalar()].vec(d.dtype.count)
-  m1 = d.bitcast(dtype)
-  m2 = shl(e.cast(dtype), mantissa_bits(d.dtype))
-  return (m1 + m2).bitcast(d.dtype)
-
-def ldexp2k(d:UOp, e:UOp) -> UOp:
-  """d*2^e. much faster than ldexp3k but risky. d > 0 and d is not denormal."""
-  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES and e.dtype.scalar() in (dtypes.int16, dtypes.int32, dtypes.int64)
-  return (d * pow2if(shr(e, 1), d.dtype)) * pow2if(e - shr(e, 1), d.dtype)
-
-def frexp(v:UOp) -> tuple[UOp, UOp]:
-  """frexp(v) -> (mantissa, exponent) assuming v != 0"""
-  assert v.dtype.scalar() in TRANSCENDENTAL_DTYPES
-  # m1 = masks for mantissa, m2 = masks to normalize the mantissa.
-  m1 = {dtypes.float64: 0x000FFFFFFFFFFFFF, dtypes.float32: 0x807FFFFF, dtypes.float16: 0x83FF}[v.dtype.scalar()]
-  m2 = {dtypes.float64: 0x3FE0000000000000, dtypes.float32: 0x3F000000, dtypes.float16: 0x3800}[v.dtype.scalar()]
-  bits = v.bitcast({dtypes.float64: dtypes.uint64, dtypes.float32: dtypes.uint32, dtypes.float16: dtypes.uint16}[v.dtype.scalar()].vec(v.dtype.count))
-  exponent = shr(bits, mantissa_bits(v.dtype)) & exponent_mask(v.dtype)
-  # Set the exponent bits appropriately to normalize the mantissa into the range of [0.5, 1.0).
-  mantissa = ((bits & m1) | m2).bitcast(v.dtype)
-  exp = exponent - exponent_bias(v.dtype) + 1
-  return mantissa, exp
-
-# *** reduction algorithms for sine ***
-def payne_hanek_reduction(d:UOp) -> tuple[UOp, UOp]:
-  """
-  Performs Payne-Hanek Reduction: computes the remainder of `d` modulo pi/2 for the values `d` where
-    39800.0 <= d <= +Inf
-  Returns a tuple of `(r, q)`:
-  - `r`[d.dtype] is the reminder value corresponding to `round_to_nearest(x % pi/2)`.
-  - `q`[int32] is an integer, and q % 4 is corresponding to the quadrant of the original angle `d`.
-  """
-  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
-  # https://stackoverflow.com/questions/30463616/payne-hanek-algorithm-implementation-in-c/30465751#30465751
-  # 190 bits of 2/pi for Payne-Hanek style argument reduction
-  two_over_pi_f = [0x00000000, 0x28be60db, 0x9391054a, 0x7f09d5f4, 0x7d4d3770, 0x36d8a566, 0x4f10e410]
-
-  intermediate_dtype = dtypes.float32.vec(d.dtype.count) if d.dtype.base.scalar() == dtypes.float16 else d.dtype
-
-  f, e = frexp(d)
-  ia = (f.cast(intermediate_dtype) * 4.294967296e9).cast(dtypes.uint64)
-  # extract 96 relevant bits of 2/pi based on magnitude of argument
-  i = shr(e.cast(dtypes.uint64), 5)
-  e = e.cast(dtypes.int32) & 31
-  offset = 32 - e
-
-  def _take(an:UOp, offset:int, count:int=0) -> UOp:
-    """an = two_over_pi_f[i+offset]"""
-    if count+offset < len(two_over_pi_f) - 1:
-      an = i.ne(count).where(_take(an, offset, count=count+1), an.const_like(two_over_pi_f[count+offset]))
-    return an
-  def _shl_lazy(x:UOp, y:UOp): return (x.cast(dtypes.uint64) * pow2if(y, d.dtype).cast(dtypes.uint64)).cast(dtypes.uint32)
-  def _shr_lazy(x:UOp, y:UOp): return (x.cast(dtypes.uint64) // pow2if(y, d.dtype).cast(dtypes.uint64)).cast(dtypes.uint32)
-
-  a = [_take(UOp.const(dtypes.uint32.vec(d.dtype.count), 0), i) for i in range(4)]
-  #  (two_over_pi_f[Int(i) + n] << e) | (two_over_pi_f[Int(i) + n+1] >> (nbits - e))
-  # Note: e >= 1 for all numbers d >= 1.0. assume e != 0
-  hi = _shl_lazy(a[0], e) | _shr_lazy(a[1], offset)
-  mi = _shl_lazy(a[1], e) | _shr_lazy(a[2], offset)
-  lo = _shl_lazy(a[2], e) | _shr_lazy(a[3], offset)
-
-  def _hp_mul(x:UOp, y:UOp) -> UOp: return x.cast(dtypes.uint64) * y.cast(dtypes.uint64)
-  # compute x * 2/pi
-  p = shl(_hp_mul(ia, hi), 32) + _hp_mul(ia, mi) + shr(_hp_mul(ia, lo), 32)
-
-  # round quotient to nearest
-  q = shr(p, 62).cast(dtypes.int32)
-  p = p & 0x3fffffffffffffff
-  r = (p.cast(intermediate_dtype) * (3.4061215800865545e-19)).cast(d.dtype)
-
-  # if fraction >= 0.5, r -= pi/2, q += 1
-  return (f<0.5).where(r, r - math.pi/2), (f<0.5).where(q, q + 1)
-
-def cody_waite_reduction(d:UOp) -> tuple[UOp, UOp]:
-  """
-  Performs Cody-Waite Reduction: computes the reminder of `d` modulo pi/2 for the values `d` where
-      0 <= abs(d) <= 39800.0
-  Returns a tuple of `(r, q)`, where the output format is the same as that of `payne_hanek_reduction`.
-  """
-  def _reduce_d(x:UOp, q:UOp):
-    # https://github.com/shibatch/sleef/blob/4e08851f59fc2b545f9c393c6a23dfd311a26308/src/libm/sleefdp.c#L789-L823
-    if x.dtype.scalar() == dtypes.float64:
-      # https://github.com/shibatch/sleef/blob/f6d8a841fbfddd26ce712834d4da220cd76048fb/src/common/misc.h#L77
-      PI_A, PI_B, PI_C, PI_D = 3.1415926218032836914, 3.1786509424591713469e-08, 1.2246467864107188502e-16, 1.2736634327021899816e-24
-      d = qdh * -PI_A + x
-      d = q * -PI_A + d
-      d = qdh * -PI_B + d
-      d = q * -PI_B + d
-      d = qdh * -PI_C + d
-      d = q * -PI_C + d
-      d = (qdh + q) * -PI_D + d
-    elif x.dtype.scalar() == dtypes.float16:
-      # [FIXME] when reducing `d`, FP16 needs FP32 precision to achieve 1.0 ULP precision.
-      d = _reduce_d(x.cast(dtypes.float32), q.cast(dtypes.float32)).cast(dtypes.float16)
-    else:
-      # https://github.com/shibatch/sleef/blob/4e08851f59fc2b545f9c393c6a23dfd311a26308/src/libm/sleefsp.c#L464-L503
-      d = q * -3.1414794921875 + x
-      d = q * -0.00011315941810607910156 + d
-      d = q * -1.9841872589410058936e-09 + d
-      d = q * -1.2154201256553420762e-10 + d
-    return d
-
-  m_1_pi = 0.318309886183790671537767526745028724
-  qdh = (d * (m_1_pi / 2.0**24)).cast(dtypes.int64).cast(d.dtype) * (2.0**24)
-  quadrant = rintk(d * m_1_pi -qdh) if d.dtype.base.scalar() == dtypes.float64 else rintk(d * m_1_pi)
-  return _reduce_d(d, quadrant.cast(d.dtype)), quadrant.cast(dtypes.int32)
-
-# *** approximate sine on small angle. ***
-def trig_poly(d:UOp, coeff32, coeff64): return d * (polyN(d*d, coeff64) if d.dtype.scalar() == dtypes.float64 else polyN(d*d, coeff32))
-# approximate sine on [-pi/2, pi/2]
-def sin_poly(d:UOp) -> UOp:
-  return trig_poly(d, [2.6083159809786593541503e-06, -0.0001981069071916863322258, 0.00833307858556509017944336, -0.166666597127914428710938, 1.0],
-                      [-7.97255955009037868891952e-18, 2.81009972710863200091251e-15, -7.64712219118158833288484e-13, 1.60590430605664501629054e-10,
-                       -2.50521083763502045810755e-08, 2.75573192239198747630416e-06, -0.000198412698412696162806809, 0.00833333333333332974823815,
-                       -0.166666666666666657414808,    1.0])
-
-def _ifand(q:UOp, n:int): return (q & n).ne(0)
-
-def sin_poly_small(d:UOp, q:UOp) -> UOp:
-  r = sin_poly(d)
-  return r * _ifand(q, 1).where(r.const_like(-1), r.const_like(1))
-
-def sin_poly_large(d:UOp, q:UOp) -> UOp:
-  r = sin_poly(d + _ifand(q, 1).where(d.const_like(math.pi / 2), d.const_like(0)))
-  return r * _ifand(q, 2).where(r.const_like(-1), r.const_like(1))
-
-# *** toplevel functions for xsin/xlog2/xexp2 ***
-
-def xsin(d:UOp, fast:bool=False, switch_over:float=30.0) -> UOp:
-  """
-  Implements a 1.0 ULP approximation for Ops.SIN.
-  - fast=True assumes x <= switch_over.
-  - switch_over is the threshold for switching to payne_hanek_reduction.
-  """
-  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
-  # mask +-inf/nan as zero
-  x = _lazy_map_numbers(d, d.const_like(0.0), d.const_like(0.0), d.const_like(0.0), d)
-  # x_sign = sign(x)
-  x_sign = x.ne(0).where((x<0).where(x.const_like(-1), x.const_like(1)), x.const_like(0))
-  x_abs = x * x_sign
-  r, q = (cody_waite_reduction if fast else payne_hanek_reduction)(x_abs)
-  if fast: result = sin_poly_small(r, q)
-  else:
-    # Payne Hanek Reduction assumes abs(x) >= pi/4, so for smaller values, use cody_waite_reduction.
-    r_small, q_small = cody_waite_reduction(x_abs)
-    result = (x_abs<switch_over).where(sin_poly_small(r_small, q_small), sin_poly_large(r, q))
-  # adjusts the sign for abs(x)
-  result = result * x_sign
-  # sin(Inf) = NaN, sin(-Inf) = NaN, sin(NaN) = NaN
-  return _lazy_map_numbers(d, d.const_like(math.nan), d.const_like(math.nan), d.const_like(math.nan), result)
-
-def xexp2(d:UOp) -> UOp:
-  """
-  Implements a 1.0 ULP approximation for Ops.EXP2
-  - Paper: https://arxiv.org/pdf/2001.09258
-  """
-  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
-  # mask +=inf/nan as zero.
-  x = _lazy_map_numbers(d, d.const_like(0.0), d.const_like(0.0), d.const_like(0.0), d)
-  q = rintk(x)
-  # s = d - round(d)
-  s = x - q.cast(x.dtype)
-  # a polynomial approximation with 13 non-zero terms in the range of [−(log 2)/2,(log 2)/2].
-  if d.dtype.scalar() == dtypes.float64:
-    u = polyN(s, [0.4434359082926529454e-9, 0.7073164598085707425e-8, 0.1017819260921760451e-6, 0.1321543872511327615e-5, 0.1525273353517584730e-4,
-                  0.1540353045101147808e-3, 0.1333355814670499073e-2, 0.9618129107597600536e-2, 0.5550410866482046596e-1, 0.2402265069591012214e+0,
-                  0.6931471805599452862e+0, 0.1000000000000000000e+1])
-  else: u = polyN(s, [0.1535920892e-3, 0.1339262701e-2, 0.9618384764e-2, 0.5550347269e-1, 0.2402264476e+0, 0.6931471825e+0, 1.0])
-  u = ldexp2k(u, q) # u*2^q
-  upper, lower = {dtypes.float64: (1024, -2000), dtypes.float32: (128, -150), dtypes.float16: (23, -22)}[d.dtype.scalar()]
-  # Replace x >= upper with +inf
-  u = (d >= upper).where(d.const_like(math.inf), u)
-  # Replace x < lower with zero.
-  u = (d<lower).where(d.const_like(0.0), u)
-  # exp2(NaN) = NaN
-  return d.ne(d).where(d.const_like(math.nan), u)
-
-def xlog2(d:UOp) -> UOp:
-  """
-  Implements a 1.0 ULP approximation for Ops.LOG2
-  Paper: https://arxiv.org/pdf/2001.09258 5.5
-  """
-  assert d.dtype.scalar() in TRANSCENDENTAL_DTYPES
-  # float16 uses 2^10 for denormal scaling (2^64 overflows), float32/64 use 2^64
-  denormal_exp = 10 if d.dtype.scalar() == dtypes.float16 else 64
-  FLT_MIN = d.const_like({dtypes.float16: 6.1e-5, dtypes.float32: 1e-4, dtypes.float64: 1e-4}[d.dtype.scalar()])
-  is_denormal = d<FLT_MIN
-  a = is_denormal.where(d * (2 ** denormal_exp), d)
-
-  e = ilogb2k(a * (1.0 / 0.75)).cast(a.dtype)
-  m = ldexp3k(a, -e)
-  e = is_denormal.where(e - denormal_exp, e)
-
-  x = (m - 1.0) / (m + 1.0)
-  x2 = x * x
-  if d.dtype.scalar() == dtypes.float64:
-    t = polyN(x2, [0.2211941750456081490e+0, 0.2200768693152277689e+0, 0.2623708057488514656e+0, 0.3205977477944495502e+0,
-                   0.4121985945485324709e+0, 0.5770780162997058982e+0, 0.96179669392608091449])
-    r = t * (x * x2) + e + x * 2.885390081777926774
-  else:
-    t = polyN(x2, [0.4374550283e+0, 0.5764790177e+0, 0.9618012905120])
-    # s_lo term (x*3.27e-08) only for float32 - underflows in float16
-    r = t * (x * x2) + e + x * 2.8853900432586669922 + (x * 3.2734474483568488616e-08 if d.dtype.scalar() == dtypes.float32 else 0)
-
-  # log2(Inf) = Inf
-  r = d.ne(math.inf).where(r, r.const_like(math.inf))
-  # log2(0) = -Inf (handle both +0.0 and -0.0)
-  r = d.ne(0.0).where(r, r.const_like(-math.inf))
-  # log2(x) = NaN for x < 0
-  r = (d<-0.0).where(r.const_like(math.nan), r)
-  # log2(NaN) = NaN
-  r = d.ne(d).where(r.const_like(math.nan), r)
-  # log2(-0.0) = -Inf. In certain devices like PTX, x == -0.0 won't be true. so making reciprocal.
-  return d.reciprocal().ne(-math.inf).where(r, r.const_like(-math.inf))
-
-def xpow(base:UOp, exponent:UOp) -> UOp:
-  # start with b ** e = exp2(e * log2(b))
-  ret = (base < 0).where(-base, base).log2().mul(exponent).exp2()
-  # negative base: nan for non-integer exponent, negate for odd integer exponent
-  non_int = exponent != exponent.cast(dtypes.int32).cast(exponent.dtype)
-  is_odd = (exponent < 0).where(-exponent, exponent).cast(dtypes.int32).mod(2).cast(dtypes.bool)
-  neg_base = non_int.where(ret.const_like(math.nan), is_odd.where(-ret, ret))
-  # fix 0 ** 0 = 1
-  return (base.eq(0) & exponent.eq(0)).where(ret.const_like(1), (base < 0).where(neg_base, ret))
-
-# *** integer division ***
-
-@functools.lru_cache(None)
-def magicgu(vmax:int, d:int) -> tuple[int,int]:
-  # calculate m,s such that x//d == (x*m) >> s for all 0 <= x <= vmax, d>0; adapted from Hacker's Delight, Chapter 10
-  nc = (vmax+1)//(d) * d - 1
-  nbits = vmax.bit_length()
-  for s in range(0, 2*nbits + 1):
-    if 2**s > nc*(d - 1 - (2**s - 1) % d):
-      m = (2**s + d - 1 - (2**s - 1) % d)//d
-      return m, s
-  assert False
-
-def fast_idiv(ren: Renderer, x: UOp, d: int, dont_cast=False) -> UOp|None:
-  from tinygrad.renderer.cstyle import MetalRenderer
-  # NOTE: disable for METAL due to compiler bug. keccak with -O0 works but not with optimization
-  if isinstance(ren, MetalRenderer): return None
-  # If d is a power of two this is not valid for signed ints!
-  is_unsigned = x.vmin>=0 or x.dtype in dtypes.uints
-  assert d>0, "Sign should have been taken out of divisor"
-  vmin,vmax = max(x.vmin, x.dtype.min), min(x.vmax, x.dtype.max)
-  if vmin > -d and vmax < d: return x.const_like(0)
-  m,s = magicgu(max(vmax, abs(vmin)), d)
-  if m*vmin >= x.dtype.min and m*vmax <= x.dtype.max:
-    return ((x*m) >> s) if is_unsigned else ((x*m) >> s) + (x<0).where(x.ufix(1), 0)
-  # before we try casting to a larger dtype (slow), we see if there are powers of two in d we can shift to make x smaller
-  # use explicit Ops.CDIV (trunc) since the recursion assumes trunc semantics throughout
-  if (largest_factor_of_two_in_d := (d & -d)) > 1:
-    if (ret:=fast_idiv(ren, x.alu(Ops.CDIV, x.const_like(largest_factor_of_two_in_d)),
-                       d//largest_factor_of_two_in_d, dont_cast=True)) is not None: return ret
-  if dont_cast: return None
-  # promo_lattice needs to return an unsigned type if the type is unsigned
-  if dtypes.is_int(next_dtype := promo_lattice[x.dtype.scalar()][-1]) and next_dtype in ren.supported_dtypes():
-    if m*vmin >= next_dtype.min and m*vmax <= next_dtype.max:
-      return ((x.cast(next_dtype)*m) >> s).cast(x.dtype) if is_unsigned else ((x.cast(next_dtype)*m) >> s).cast(x.dtype) + (x<0).where(x.ufix(1), 0)
-  return None
-
-# ***** threefry *****
-
-def threefry2x32(x: UOp, key: UOp):
-  # split x and key from uint64 to two uint32
-  x0, x1 = (x & 0xffffffff).cast(dtypes.uint32), ((x // 2**32) & 0xffffffff).cast(dtypes.uint32)
-  key0, key1 = (key & 0xffffffff).cast(dtypes.uint32), ((key // 2**32) & 0xffffffff).cast(dtypes.uint32)
-
-  rotations = [[13, 15, 26, 6], [17, 29, 16, 24]]
-  ks = [key1, key0 ^ key1 ^ 0x1BD11BDA, key0]
-  xr:list[UOp] = [x0 + ks[-1], x1 + ks[0]]
-  for i in range(5):
-    for r in rotations[i % 2]: xr[0], xr[1] = (x0 := xr[0] + xr[1]), x0 ^ ((xr[1] * 2**r) + (xr[1] // 2**(32 - r)))
-    xr = [(xr[0] + ks[i % 3]), (xr[1] + ks[(i + 1) % 3] + i + 1)]
-
-  return xr[1].cast(dtypes.uint64) * 2**32 | xr[0].cast(dtypes.uint64)
-
-# ***** long as 2 ints *****
-
-l2i_dt = {dtypes.long: dtypes.int, dtypes.ulong: dtypes.uint}
-def unpack32(v:UOp) -> tuple[UOp, UOp]: return v.bitcast(dtypes.uint) & 0xFFFF, shr(v.bitcast(dtypes.uint), 16)
-def reindex(idx:UOp, off:int, mul=2) -> UOp: return idx.replace(src=(idx.src[0], idx.src[1]*mul+off, *idx.src[2:]))
-
-# 4.3.1 is the relevant section in TAOCP
-def l2i(op: Ops, dt: DType, *uops:UOp):
-  zero = UOp.const(dt, 0)
-  if len(uops) == 2: a0, a1 = uops
-  elif len(uops) == 4: a0, a1, b0, b1 = uops
-  match op:
-    case Ops.NEG: return l2i(Ops.SUB, dt, zero, zero, *uops)
-    case Ops.CAST if dt in (dtypes.long, dtypes.ulong) and uops[0].dtype not in dtypes.floats:
-      return uops[0].cast(l2i_dt[dt]), (uops[0] < 0).where(UOp.const(l2i_dt[dt], -1), UOp.const(l2i_dt[dt], 0))
-    case Ops.CAST if dt in (dtypes.long, dtypes.ulong):
-      return (lo:=uops[0].cast(l2i_dt[dt])), (uops[0] / 2**32).cast(l2i_dt[dt]) - ((uops[0] < 0) & lo.ne(0)).cast(l2i_dt[dt])
-    case Ops.CAST if dt in dtypes.floats:
-      small = (a1.eq(0) & (a0 >= 0)) | (a1.eq(-1) & (a0 < 0))
-      return small.where(a0.cast(dt), ((a1.cast(dtypes.float32) * (2**32)) + a0.bitcast(dtypes.uint).cast(dtypes.float32)).cast(dt))
-    case Ops.CAST: return a0.bitcast(dtypes.uint).cast(dt)
-    case Ops.BITCAST: return a0.bitcast(dt), a1.bitcast(dt)
-    case Ops.SHL:
-      lo, hi = shl(a0, b0_mod:=b0 & 31), shl(a1, b0_mod) | shr(shr(a0, 1), 31 - b0_mod)
-      return (b0 >= 32).where(zero, lo), (b0 >= 32).where(lo, hi)
-    case Ops.SHR:
-      lo, hi = shr(a0, b0_mod:=b0 & 31) | shl(shl(a1, 1), 31 - b0_mod), shr(a1, b0_mod)
-      return (b0 >= 32).where(hi, lo), (b0 >= 32).where(zero, hi)
-    case Ops.ADD: return (low:=a0+b0), (a1 + b1).replace(dtype=dt) + (low.bitcast(dtypes.uint) < a0.bitcast(dtypes.uint)).cast(dt)
-    case Ops.SUB: return a0 - b0, a1 - b1 - (a0.bitcast(dtypes.uint) < b0.bitcast(dtypes.uint)).cast(dt)
-    case Ops.MUL:
-      (a00, a01), (b00, b01) = unpack32(a0), unpack32(b0)
-      mid = l2i(Ops.ADD, dt, shl(a00*b01, 16).bitcast(dt), shr(a00*b01, 16).bitcast(dt), shl(a01*b00, 16).bitcast(dt), shr(a01*b00, 16).bitcast(dt))
-      return l2i(Ops.ADD, dt, *mid, (a00*b00).bitcast(dt), (a01*b01).bitcast(dt) + a0*b1 + a1*b0)
-    case Ops.CDIV | Ops.CMOD:
-      # TAOCP Algorithm 4.3.1D could be faster here, but must be parameterized over the width of b
-      if dt == dtypes.int:
-        ua0, ua1, ub0, ub1 = a0.bitcast(dtypes.uint), a1.bitcast(dtypes.uint), b0.bitcast(dtypes.uint), b1.bitcast(dtypes.uint)
-        a0, a1 = (a_neg:=a1 < zero).where((n:=l2i(Ops.NEG, dtypes.uint, ua0, ua1))[0], ua0), a_neg.where(n[1], ua1)
-        b0, b1 = (b_neg:=b1 < zero).where((n:=l2i(Ops.NEG, dtypes.uint, ub0, ub1))[0], ub0), b_neg.where(n[1], ub1)
-      q, r = (z:=UOp.const(dtypes.uint, 0), z), (z, z)
-      for i in range(63, -1, -1):
-        r = l2i(Ops.SHL, dtypes.uint, *r, UOp.const(dtypes.uint, 1), z)
-        r = (r[0] | l2i(Ops.SHR, dtypes.uint, a0, a1, UOp.const(dtypes.uint, i), z)[0] & 1), r[1]
-        cond = l2i(Ops.CMPLT, dtypes.uint, *r, b0, b1).logical_not()
-        diff = l2i(Ops.SUB, dtypes.uint, *r, b0, b1)
-        q = ((q[0] | shl(cond.cast(dtypes.uint), i % 32), q[1]) if i < 32 else (q[0], q[1] | shl(cond.cast(dtypes.uint), i % 32)))
-        r = l2i(Ops.WHERE, dtypes.uint, cond, *diff, *r)
-      if dt == dtypes.int:
-        (nq0, nq1), (nr0, nr1) = l2i(Ops.BITCAST, dt, *l2i(Ops.NEG, dtypes.uint, *q)), l2i(Ops.BITCAST, dt, *l2i(Ops.NEG, dtypes.uint, *r))
-        (q0, q1), (r0, r1) = l2i(Ops.BITCAST, dt, *q), l2i(Ops.BITCAST, dt, *r)
-        return (a_neg.where(nr0, r0), a_neg.where(nr1, r1)) if op == Ops.CMOD else ((a_neg^b_neg).where(nq0, q0), (a_neg^b_neg).where(nq1, q1))
-      return (r[0].bitcast(dt), r[1].bitcast(dt)) if op == Ops.CMOD else (q[0].bitcast(dt), q[1].bitcast(dt))
-    case Ops.CMPLT: return (a1 < b1) | ((a1.eq(b1)) & (a0.bitcast(dtypes.uint) < b0.bitcast(dtypes.uint)))
-    case Ops.CMPEQ: return a0.eq(b0) & a1.eq(b1)
-    case Ops.CMPNE: return a0.ne(b0) | a1.ne(b1)
-    case Ops.XOR | Ops.OR | Ops.AND: return UOp(op, dt, src=(a0, b0)), UOp(op, dt, src=(a1, b1))
-    case Ops.WHERE: return uops[0].where(uops[1], uops[3]), uops[0].where(uops[2], uops[4])
-    case Ops.MAX: return l2i(Ops.WHERE, dt, l2i(Ops.CMPLT, dt, *uops), b0, b1, a0, a1)
-    case _: raise NotImplementedError(f"long decomposition of {op} unsupported")
-
-# ***** floats *****
-f2f_dt = { f:getattr(dtypes, f"uint{f.bitsize}") for f in dtypes.floats }
-
-def rne(v: UOp, s) -> UOp: return shr(v, s) + ((shr(v, s - 1) & 1) & ((v & ((1 << (s - 1)) - 1)).ne(0).cast(v.dtype) | (shr(v, s) & 1)))
-
-def f2f(v, fr:DType, to:DType, sat=True):
-  fs, fb, (fe, fm), ts, tb, (te, tm) = fr.bitsize, exponent_bias(fr), dtypes.finfo(fr), to.bitsize, exponent_bias(to), dtypes.finfo(to)
-  # NB: denormals are zero!
-  if fe <= te and fm < tm:
-    sign, nosign = shl((v & shl(1, fs-1)).cast(f2f_dt[to]), ts - fs), (v & (shl(1, fs-1) - 1)).cast(f2f_dt[to])
-    exp, norm = shr(nosign, fm), shl(nosign, tm - fm) + shl(tb - fb, tm)
-    nan = shl(nosign, tm - fm) | shl((shl(1, te) - 1), tm)
-    if fr in dtypes.fp8_fnuz:
-      fnuz_nan = sign.ne(0) & nosign.eq(0)
-      qnan = shl(shl(1, te) - 1, tm) | shl(1, tm - 1)
-      return fnuz_nan.where(qnan, sign | exp.eq(0).where(0, norm)).bitcast(to)
-    # fp8e4m3 has only one nan
-    is_nan = (nosign.eq(shl(1, fm + fe) - 1) if fr == dtypes.fp8e4m3 else exp.eq(shl(1, fe) - 1))
-    return (sign | exp.eq(0).where(0, is_nan.where(nan, norm))).bitcast(to)
-  elif fe >= te and fm > tm:
-    v = f2f_clamp(v.bitcast(fr), to, sat).bitcast(f2f_dt[fr])
-    sign, nosign = shr(v, fs - ts) & shl(1, ts - 1), v & (shl(1, fs - 1) - 1)
-    norm = (rne(nosign, fm - tm) - shl(fb - tb, tm)).cast(f2f_dt[to])
-    underflow = (shr(v, fm) & (shl(1, fe) - 1)) < (1 + fb - tb)
-    nan_mantissa = (shl(1, tm) - 1) if to == dtypes.fp8e4m3 else (shr(nosign, fm - tm) & (shl(1, tm) - 1))
-    nan = (sign | nan_mantissa | shl(shl(1, te) - 1, tm)).cast(f2f_dt[to])
-    is_nan = (shr(v, fm) & (shl(1, fe) - 1)).eq(shl(1, fe) - 1)
-    if to in dtypes.fp8_fnuz: return is_nan.where(shl(1, ts - 1), underflow.where(0, sign.cast(f2f_dt[to]) | norm))
-    return is_nan.where(nan, sign.cast(f2f_dt[to]) | underflow.where(0, norm))
-  else: raise NotImplementedError(f"unsupported decomp {fr} -> {to}")
-
-def f2f_clamp(val:UOp, dt:DType, sat=True) -> UOp:
-  e, m = dtypes.finfo(dt)
-  if dt in dtypes.fp8_fnuz: max_exp, max_man = (1 << e) - 1, (1 << m) - 1
-  else: max_exp, max_man = ((1 << e) - 1, (1 << m) - 2) if dt == dtypes.fp8e4m3 else ((1 << e) - 2, (1 << m) - 1)
-  mx = val.const_like(2.0**(max_exp - exponent_bias(dt)) * (1.0 + max_man / (1 << m)))
-  sat = mx if dt in dtypes.fp8s and sat else val.const_like(float('inf'))
-  # FIXME: CMPLT of nan is undefined
-  return val.ne(val).where(val, (val < -mx).where(-sat, (mx < val).where(sat, val)))
-
-def f2f_load(x: UOp, fr:DType, to:DType) -> UOp:
-  if (n:=x.dtype.count) == 1: return f2f(x.replace(dtype=f2f_dt[fr]), fr, to)
-  return UOp.vectorize(*(f2f(x.replace(dtype=f2f_dt[fr], src=(reindex(x.src[0].src[0], i, 1),)), fr, to) for i in range(n)))
-
-def f2f_store(st, idx, val, fr:DType, to:DType):
-  if (n:=val.dtype.count) == 1: return st.replace(src=(idx, f2f(val.bitcast(f2f_dt[to]), to, fr)))
-  return UOp.group(*(st.replace(src=(reindex(idx, i, 1), f2f(val.gep(i).bitcast(f2f_dt[to]), to, fr))) for i in range(n)))
-
-# ***** decomposition patterns *****
-
-@functools.cache
-def get_transcendental_patterns(ops:tuple[Ops, ...], force_transcendental:bool) -> PatternMatcher:
-  pat: list[tuple[UPat, Callable]] = []
-  for op,f in ((Ops.EXP2, xexp2), (Ops.LOG2, xlog2), (Ops.SIN, xsin)):
-    if op not in ops or force_transcendental:
-      pat += [(UPat(op, dtype=TRANSCENDENTAL_DTYPES, src=(UPat.var("d"),)), f),
-              (UPat(op, dtype=tuple(dt for dt in dtypes.floats if dt not in TRANSCENDENTAL_DTYPES), src=(UPat.var("d"),), name="x"),
-                lambda x,d: d.cast(dtypes.float32).alu(x.op).cast(x.dtype))]
-  # rewrite SQRT to xpow 0.5
-  if Ops.SQRT not in ops or force_transcendental: pat.append((UPat(Ops.SQRT, src=UPat.var("d")), lambda d: xpow(d, d.const_like(0.5))))
-  return PatternMatcher(pat)
-
-def floordiv_to_idiv(a:UOp, b:UOp) -> UOp:
-  if (a.vmin >= 0 and b.vmin > 0) or (a.vmax <= 0 and b.vmax < 0): return a.alu(Ops.CDIV, b)
-  return a.alu(Ops.CDIV, b) - (a.alu(Ops.CMOD, b).ne(0) & (a<0).ne(b<0)).cast(a.dtype)
-
-def floormod_to_mod(a:UOp, b:UOp) -> UOp:
-  if (a.vmin >= 0 and b.vmin > 0) or (a.vmax <= 0 and b.vmax < 0): return a.alu(Ops.CMOD, b)
-  r = a.alu(Ops.CMOD, b)
-  # use where instead of mul to avoid being fused into MULACC (which int64 long-decomp doesn't handle)
-  return r + (r.ne(0) & (a<0).ne(b<0)).where(b, b.const_like(0))
-
-powers_of_two: dict[int, int] = {2**i:i for i in range(64)}
-@functools.cache
-def get_late_rewrite_patterns(ops:tuple[Ops, ...], disable_fast_idiv:bool) -> PatternMatcher:
-  pat: list[tuple[UPat, Callable]] = [(UPat.var("a")//UPat.var("b"), floordiv_to_idiv)]
-  # FLOORMOD by 2**y -> x & (2**y-1) (correct floor mod for any sign in two's complement); fires before floormod_to_mod
-  if Ops.AND in ops: pat.append((UPat.var("x", dtypes.ints)%UPat.cvar("c"), lambda x,c: x & (c.arg-1) if c.arg in powers_of_two else None))
-  pat.append((UPat.var("a")%UPat.var("b"), floormod_to_mod))
-  # no real hardware supports THREEFRY, but NullRenderer does
-  if Ops.THREEFRY not in ops: pat.append((UPat(Ops.THREEFRY, dtype=dtypes.uint64, src=(UPat.var("x"), UPat.var("key"))), threefry2x32))
-  # MAX can be rewritten as CMPLT + WHERE (max function is annoying on many cstyle backends)
-  if Ops.MAX not in ops and Ops.CMPLT in ops: pat.append((UPat(Ops.MAX, name="m"), lambda m: (m.src[0] < m.src[1]).where(m.src[1], m.src[0])))
-  if Ops.OR in ops: pat += [(UPat.var("x", dtypes.bool).logical_not()&UPat.var("y", dtypes.bool).logical_not(),
-    lambda x,y: (x | y).logical_not())]
-  # rewrite MUL/CDIV to SHL+SHR: x*(2**y) -> shl(x,y) and x//(2**y) -> shr(x,y)
-  if Ops.SHL in ops: pat += [(UPat.var("x", dtypes.ints)*UPat.cvar("c"), lambda c,x: x << v if (v:=powers_of_two.get(c.arg, 0)) else None)]
-  if Ops.SHR in ops:
-    # uint CDIV by 2**v -> x >> v (FLOORDIV is lowered to CDIV by the rule above before reaching here)
-    pat += [(UPat(Ops.CDIV, src=(UPat.var("x", dtypes.uints), UPat.cvar("c"))),
-      lambda x,c: x >> v if (v:=powers_of_two.get(c.arg, 0)) else None)]
-    # signed CDIV (trunc) by 2**v -> (x + (x<0 ? c-1 : 0)) >> v
-    pat += [(UPat(Ops.CDIV, src=(UPat.var("x", dtypes.ints), UPat.cvar("c"))),
-      lambda x,c: (x+(l.const_like(l.vmin) if (l:=(x<0)).vmin==l.vmax else l).where(c-1, 0)) >> v
-        if (v:=powers_of_two.get(c.arg, 0)) else None)]
-    if not disable_fast_idiv:
-      # fast_idiv handles non-pow2: only fire on non-negative inputs (signed magic-mul is unreliable for x<0)
-      pat += [(UPat(Ops.CDIV, src=(UPat.var("x", dtypes.ints), UPat.cvar("d"))),
-        lambda ctx, x, d: fast_idiv(ctx, x, d.arg) if x.vmin >= 0 or x.dtype in dtypes.uints else None)]
-      # rewrite raw CMOD -> x - d*CDIV(x,d) so fast_idiv can pick up the CDIV. only on non-negative inputs;
-      # avoids disturbing floormod_to_mod's general-path output (which uses a trunc Ops.CMOD as an implementation detail)
-      pat += [(UPat(Ops.CMOD, src=(UPat.var("x", dtypes.ints), UPat.var("d"))),
-        lambda x, d: x - d * x.alu(Ops.CDIV, d) if x.vmin >= 0 or x.dtype in dtypes.uints else None)]
-  if Ops.NEG in ops:
-    pat += [(UPat.var('x')*-1, lambda ctx,x: x.alu(Ops.NEG))]
-    if Ops.SUB in ops: pat += [(UPat.var('x')+UPat.var('y').alu(Ops.NEG), lambda ctx,x,y: x.alu(Ops.SUB, y))]
-  if Ops.CMPLT in ops:
-    # These are late rewrites because simplex expects equalities to be a certain format
-    pat += [
-      ((UPat.var("x", dtypes.sints) < UPat.cvar("c", dtypes.sints)).logical_not(), lambda x,c: c-1<x),
-      ((UPat.cvar("c", dtypes.sints) < UPat.var("x", dtypes.sints)).logical_not(), lambda x,c: x<c+1),
-      (UPat.var("x", dtypes.sints)*-1 < UPat.var("y", dtypes.sints)*UPat.cvar("c"), lambda x,y,c: y*(-c)<x),
-      (UPat.var("x", dtypes.sints)*-1 < UPat.cvar("c"), lambda x,c:-c<x),
-      ((UPat.cvar("c1")<UPat.var("x", dtypes.sints)) & (UPat.var("x", dtypes.sints)<UPat.cvar("c2")),
-        lambda x,c1,c2: x.eq(c1+1) if c1.arg+1==c2.arg-1 else None),  # (c-1)<x & x<(c+1) -> x==c
-    ]
-  if Ops.CMPEQ in ops: pat += [(UPat.var('x').ne(UPat.var('y')).logical_not(), lambda x,y: x.alu(Ops.CMPEQ, y))]
-  if Ops.MULACC in ops:
-    pat += [(UPat.var('a')*UPat.var('b')+UPat.var('c'), lambda a,b,c: a.alu(Ops.MULACC, b, c))]
-    # also fuse (x << n) + c → MULACC(x, 2^n, c) since MUL→SHL may run first
-    if Ops.SHL in ops: pat += [(UPat.var('x').alu(Ops.SHL, UPat.cvar('n'))+UPat.var('c'), lambda x,n,c: x.alu(Ops.MULACC, x.const_like(1<<n.arg), c))]
-  # some backends emit FDIV for RECIP, in that case: a*(1/b) -> a/b
-  if Ops.FDIV in ops:
-    pat += [(UPat.var("x").reciprocal(), lambda x: x.const_like(1).alu(Ops.FDIV, x))]
-    pat += [(UPat.var("a", dtypes.floats) * UPat.const(dtypes.floats, 1).alu(Ops.FDIV, UPat.var("b")), lambda a,b: a.alu(Ops.FDIV, b))]
-  return PatternMatcher(pat)
-
-pm_long_decomp = PatternMatcher([
-  (UPat((*GroupOp.Defines, Ops.BUFFER, Ops.INDEX), name="x"), lambda x:
-   x.replace(dtype=l2i_dt[x.dtype.base].ptr(x.dtype.size * 2)) if hasattr(x.dtype, 'size') and x.dtype.base in l2i_dt else None),
-  (UPat(Ops.INDEX, tuple(l2i_dt.keys()), name='x'), lambda x: reindex(x, x.tag).replace(dtype=l2i_dt[x.dtype])),
-  (UPat(Ops.STORE, src=(UPat.var('idx'), UPat.var('val', tuple(l2i_dt.keys()))), name='st'), lambda st,idx,val:
-   st.replace(src=(reindex(idx, 0), val.rtag(0))).group(st.replace(src=(reindex(idx, 1), val.rtag(1)))) if val.tag is None else None),
-  (UPat(GroupOp.Comparison, src=(UPat.var('a', tuple(l2i_dt.keys())), UPat.var('b', tuple(l2i_dt.keys()))), name="x"), lambda a,b,x:
-   l2i(x.op, dt:=l2i_dt[a.dtype], a.rtag(0).cast(dt), a.rtag(1).cast(dt), b.rtag(0).cast(dt), b.rtag(1).cast(dt))),
-  (UPat(Ops.CAST, tuple(l2i_dt.keys()), src=(UPat.var('a'),), name="x"), lambda a,x:
-   l2i(x.op, x.dtype, a)[x.tag] if x.tag is not None and a.dtype not in l2i_dt else None),
-  (UPat(Ops.CAST, tuple(l2i_dt.keys()), src=(UPat.var('a', tuple(l2i_dt.keys())),), name="x"), lambda a,x:
-   (a.rtag(0).cast(dt:=l2i_dt[a.dtype]).bitcast(xdt:=l2i_dt[x.dtype]), a.rtag(1).cast(dt).bitcast(xdt))[x.tag]),
-  (UPat(Ops.CAST, src=(UPat.var('a', tuple(l2i_dt.keys())),), name="x"), lambda a,x:
-   l2i(x.op, x.dtype, a.rtag(0).cast(dt:=l2i_dt[a.dtype]), a.rtag(1).cast(dt)) if x.dtype not in l2i_dt and a.tag is None else None),
-  (UPat((*(GroupOp.ALU - GroupOp.Comparison), Ops.BITCAST), tuple(l2i_dt.keys()), name="x"), lambda x:
-   l2i(x.op, l2i_dt[x.dtype], *flatten((a.rtag(0).cast(dt:=l2i_dt[x.src[-1].dtype]), a.rtag(1).cast(dt))
-                                       if a.dtype in l2i_dt else (a,) for a in x.src))[x.tag] if x.tag is not None else None),
-  (UPat(Ops.LOAD, tuple(l2i_dt.keys()), src=(UPat.var('idx'),), name='x'), lambda x,idx: x.replace(dtype=l2i_dt[x.dtype],src=(reindex(idx, x.tag),))),
-  (UPat(Ops.CONST, tuple(l2i_dt.keys()), name='x'), lambda x:
-   UOp.const(dt:=l2i_dt[x.dtype], truncate[dt]((x.arg >> 32) if x.tag == 1 else (x.arg & 0xFFFFFFFF))))
-])
-
-# float decomposition patterns - ctx is (fr, to) tuple
-pm_float_decomp = PatternMatcher([
-  (UPat((*GroupOp.Defines, Ops.BUFFER, Ops.INDEX), name="x"), lambda ctx,x:
-   x.replace(dtype=f2f_dt[ctx[0]].ptr(x.dtype.size), tag=ctx[0]) if x.dtype.base == ctx[0] else None),
-  (UPat(Ops.LOAD, dtypes.floats, name="x"), lambda ctx,x: f2f_load(x, *ctx) if x.dtype.scalar() == ctx[0] else None),
-  # bitcasted load should just replace load
-  (UPat(Ops.BITCAST, src=(UPat(Ops.LOAD, name="ld"),), name="bc"), lambda ctx,bc,ld:
-   ld.replace(dtype=f2f_dt[ctx[0]]).bitcast(bc.dtype) if ld.dtype == ctx[0] else None),
-  # bitcast from
-  (UPat(Ops.BITCAST, src=(UPat.var("x", dtypes.floats),), name="bc"), lambda ctx,bc,x:
-   bc.replace(src=(f2f(x.bitcast(f2f_dt[ctx[1]]), ctx[1], ctx[0]),)) if x.dtype == ctx[1] and bc.dtype.bitsize == ctx[0].bitsize else None),
-  # bitcast to
-  (UPat(Ops.BITCAST, src=(UPat.var("x"),), name="bc"), lambda ctx,bc,x:
-   f2f(x.bitcast(f2f_dt[ctx[0]]), ctx[0], ctx[1]) if bc.dtype == ctx[0] else None),
-  (UPat(Ops.CAST, dtypes.floats, src=(UPat.var("val"),), name="x"), lambda ctx,x,val:
-   f2f_clamp(val.cast(ctx[1]), ctx[0]) if x.dtype.scalar() == ctx[0] else None),
-  (UPat(GroupOp.All-{Ops.BITCAST}, dtypes.floats, name="x"), lambda ctx,x:
-   x.replace(dtype=ctx[1].vec(x.dtype.count), src=tuple(s.cast(ctx[1]) if s.dtype == ctx[0] else s for s in x.src))
-   if x.dtype.scalar() == ctx[0] else None),
-  (UPat(Ops.STORE, src=(UPat.var("idx"), UPat(Ops.BITCAST, dtypes.floats, name="val")), name='st'), lambda ctx,st,idx,val:
-   st.replace(src=(idx, val.replace(dtype=f2f_dt[ctx[0]]))) if val.dtype == ctx[0] and idx.tag == ctx[0] else None),
-  (UPat(Ops.STORE, src=(UPat.var("idx"), UPat.var("val", dtypes.floats)), name='st'), lambda ctx,st,idx,val:
-   f2f_store(st, idx, val, *ctx) if val.dtype.scalar() == ctx[1] and (idx:=idx.src[0] if idx.op == Ops.CAST else idx).tag == ctx[0] else None),
-])
-
-def do_dtype_decomps(sink:UOp, ctx:tuple[set[DType], Renderer]) -> UOp:
-  def _should_emulate(dt): return dt in EMULATED_DTYPES.tolist(dtypes) or dt not in ctx[1].supported_dtypes()
-  for fr in sorted(filter(_should_emulate, ctx[0])):
-    to = dtypes.int if fr == dtypes.long else dtypes.half if not _should_emulate(dtypes.half) and fr in dtypes.fp8s else dtypes.float
-    if DEBUG >= 2: print(f"emulating {fr} as {to}")
-    sink = graph_rewrite(sink, pm_float_decomp if fr in dtypes.floats else pm_long_decomp, name=f"decomp {fr} -> {to}", ctx=(fr, to), bottom_up=True)
-  ctx[0].clear()
-  return sink
-
-pm_dtype_decomps = PatternMatcher([
-  # detect dtypes to decompose
-  (UPat(GroupOp.All, (*dtypes.fp8s, dtypes.bfloat16, dtypes.half, dtypes.long, dtypes.ulong), name="x"), lambda x,ctx:
-   ctx[0].add({dtypes.ulong:dtypes.long}.get(dt:=x.dtype.base.scalar(), dt))),
-  # do the rewrites
-  (UPat(Ops.SINK, name="sink"), do_dtype_decomps),
-])
@@ -312,10 +312,12 @@ class UOp(RandMixin, metaclass=UOpMetaClass):

      # TODO: disallow shape changing bitcast
      case Ops.BITCAST:
-        ps = self.src[0]._shape
-        if ps is None: return None
+        ps = self.src[0].shape
        if (output_sz:=self.dtype.itemsize) != (input_sz:=self.src[0].dtype.itemsize):
-          return ps[:-1]+(ssimplify((ps[-1]*input_sz) // output_sz),) if len(ps) > 0 else ps
+          if ps == ():
+            if output_sz > input_sz: raise RuntimeError(f"shape {ps} must be an expanding bitcast {output_sz=} {input_sz=}")
+            ps = (1,)
+          return ps[:-1]+(ssimplify((ps[-1]*input_sz) // output_sz),)
        return ps

      # MULTI marker has no shape
@@ -919,12 +921,6 @@ class UOp(RandMixin, metaclass=UOpMetaClass):

  # *** uop symbolic stuff ***

-  def is_increasing(self:UOp) -> bool:
-    # is f a monotonically increasing function regards its input
-    if self.op in GroupOp.Irreducible: return True
-    if self.op is Ops.ADD: return self.src[0].is_increasing() and self.src[1].is_increasing()
-    if self.op in (Ops.MUL, Ops.CDIV, Ops.FLOORDIV) and self.src[1].op is Ops.CONST and self.src[1].arg >= 0: return self.src[0].is_increasing()
-    return False  # False if not sure
  def const_factor(self) -> int:
    """largest known int that divides self"""
    # TODO: for negatives it's not the largest
@@ -1,4 +1,3 @@
-from typing import cast
 from tinygrad.dtype import dtypes
 from tinygrad.uop import Ops, GroupOp
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, multirange_str, range_str, consumer_map_from_toposort
@@ -120,7 +119,7 @@ def _render_with_splits(lst:list[UOp], pm:PatternMatcher, to_render:set[UOp], sp
    op_depth = 1 + max([depth.get(s, 0) for s in u.src], default=0)
    if op_depth > split_depth: to_render.add(u)
    depth[u] = 0 if u in to_render else op_depth
-    ren = cast(str, pm.rewrite(u, ctx=r))
+    ren = pm.rewrite(u, ctx=r)
    assert isinstance(ren, str)
    if u.tag is not None: ren += f".rtag({repr(u.tag)})"
    if u not in to_render: r[u] = ren
@@ -1,5 +1,5 @@
 import math
-from typing import cast, Any
+from typing import Any
 from tinygrad.uop.ops import PatternMatcher, UPat, GroupOp, Ops, UOp, AxisType, KernelInfo, ParamArg
 from tinygrad.uop.render import print_uops, pyrender
 from tinygrad.dtype import DType, ImageDType, dtypes, PtrDType, AddrSpace, Invalid, ConstFloat
@@ -38,8 +38,8 @@ def type_verify(ast:UOp|list[UOp], check_spec:PatternMatcher):

  with Context(TRACK_MATCH_STATS=0):
    for i,u in enumerate(lst):
-      ret = check_spec.rewrite(u)
-      if cast(bool|None, ret) is not True:
+      ret: bool|None = check_spec.rewrite(u)
+      if ret is not True:
        if DEBUG >= 3: print_uops(lst)
        raise RuntimeError(f"UOp verification failed at {i} on {u.op} {u.dtype} {len(u.src)} {[(x.op, x.dtype, x.arg) for x in u.src]} {u.arg}")

@@ -4,8 +4,11 @@ from collections import defaultdict
 from tinygrad.uop.ops import Ops, PatternMatcher, UPat, UOp, GroupOp, exec_alu
 from tinygrad.dtype import PyConst, ConstType, dtypes, PtrDType, can_lossless_cast, Invalid
 from tinygrad.helpers import partition, all_same, prod, flatten, get_single_element, unwrap, IMAGE, dedup
-from tinygrad.uop.decompositions import threefry2x32, xpow
-from tinygrad.uop.divandmod import div_and_mod_symbolic
+
+# TODO: symbolic shouldn't be importing from codegen
+from tinygrad.codegen.decomp.op import threefry2x32
+from tinygrad.codegen.decomp.transcendental import xpow
+from tinygrad.codegen.decomp.divandmod import div_and_mod_symbolic

 # ******** phase 1 of symbolic used to live in ops, it's the most generic folding rules ********

@@ -121,6 +124,8 @@ symbolic_simple = propagate_invalid + PatternMatcher([
  # TODO: combine this with "# rules for threefry" below
  ((UPat.var("x") & UPat.cvar("mask")) >> UPat.cvar("k"),
   lambda x,mask,k: x >> k.arg if mask.arg | ((1 << k.arg) - 1) == -1 else None),
+  ((UPat.var("x") & UPat.cvar("mask")) // UPat.cvar("c"),
+   lambda x,mask,c: x // c.arg if c.arg > 0 and c.arg & (c.arg-1) == 0 and mask.arg | (c.arg-1) == -1 else None),
  (UPat.var("x", dtype=dtypes.ints+(dtypes.bool, dtypes.weakint)) != UPat.var("x"),
   lambda x: x.const_like(False).cast(dtypes.bool.vec(x.dtype.count))), # x != x -> False (only ints)
  # ** constant folding **
@@ -160,6 +165,7 @@ symbolic_simple = propagate_invalid + PatternMatcher([
  (((UPat.var(None, dtypes.uint64)*(1<<32)) | UPat.var('y',  dtypes.uint32).cast(dtypes.uint64)).cast(dtypes.uint32), lambda y: y),
  (((UPat.var('x',  dtypes.uint64)*(1<<32)) | UPat.var(None, dtypes.uint32).cast(dtypes.uint64))//(1<<32), lambda x: x),
  (((UPat.var(None, dtypes.uint64)<<32) | UPat.var('y',  dtypes.uint32).cast(dtypes.uint64)).cast(dtypes.uint32), lambda y: y),
+  (((UPat.var('x',  dtypes.uint64)<<32) | UPat.var(None, dtypes.uint32).cast(dtypes.uint64))//(1<<32), lambda x: x),
  (((UPat.var('x',  dtypes.uint64)<<32) | UPat.var(None, dtypes.uint32).cast(dtypes.uint64))>>32, lambda x: x),
  # ** simple where folding **
  # a conditional with the same results either way is a noop, also fold const conditionals
@@ -167,6 +173,9 @@ symbolic_simple = propagate_invalid + PatternMatcher([
  (UPat.cvar("gate").where(UPat.var("c0"), UPat.var("c1")), lambda gate, c0, c1: c0 if gate.arg else c1),
  # a.where(b.where(c, d), d) -> (a & b).where(c, d)
  (UPat.var("a").where(UPat.var("b").where(UPat.var("c"), UPat.var("d")), UPat.var("d")), lambda a,b,c,d: (a&b).where(c,d)),
+  # STACK on INDEX CONST (TODO: remove all the GEP crap)
+  (UPat(Ops.STACK, src=UPat(Ops.INDEX, src=(UPat.var("src"), UPat(Ops.CONST))), name="stk"),
+   lambda src,stk: src if stk.shape == src.shape and list(range(len(stk.src))) == [x.src[1].arg for x in stk.src] else None),
 ])

 # ******** phase 2 builds on phase 1, it includes the old "symbolic", rules that match deeper ********
@@ -1,4 +1,4 @@
-from typing import Callable, cast
+from typing import Callable
 from tinygrad.uop.ops import PatternMatcher, UPat, GroupOp, Ops, UOp, python_alu
 from tinygrad.dtype import dtypes, Invalid
 from tinygrad.helpers import cpu_profile
@@ -58,9 +58,9 @@ def uops_to_z3(solver:z3.Solver, *uops: UOp) -> list[z3.ExprRef]:
  for u in lst:
    # NOTE: we skip STACK here, it can't actually be accessed
    if u.op is Ops.STACK: continue
-    z3_rewritten = z3_renderer.rewrite(u, ctx=(solver.ctx, z3map))
+    z3_rewritten: tuple[z3.ExprRef, z3.BoolRef|None]|None = z3_renderer.rewrite(u, ctx=(solver.ctx, z3map))
    if z3_rewritten is None: raise NotImplementedError(f"{u.op} is not supported by z3")
-    new_u, constraint = cast(tuple[z3.ArithRef, z3.BoolRef|None], z3_rewritten)
+    new_u, constraint = z3_rewritten
    if constraint is not None: solver.add(constraint)
    z3map[u] = new_u
  assert all(u in z3map for u in uops), "UOp failed to rewrite to z3!"
@@ -193,7 +193,7 @@ def get_full_rewrite(data:VizData, ctx:TrackedGraphRewrite, depth:int|None=None)

 def get_sink_at(upats:tuple[str, ...], viz_data:VizData, ctx:TrackedGraphRewrite, depth:int|None=None) -> UOp|None:
  for s in get_full_rewrite(viz_data, ctx, depth=depth):
-    if s["upat"] is not None and any(n in s["upat"][1] for n in upats): return s["_sink"]
+    if (s["upat"] is not None and any(n in s["upat"][1] for n in upats)) or len(ctx.matches) == 0: return s["_sink"]
  return None

 # encoder helpers
Author	SHA1	Message	Date
geohot	a8d13380f9	simpler, but not inverse	2026-06-25 10:24:01 -07:00
geohot	afb0463f10	fix bitcast	2026-06-25 10:11:13 -07:00
geohot	947d6a3c69	enforce a cast for bitcast	2026-06-25 10:00:14 -07:00
George HotzandGitHub	dc57efc3a3	move down dtype decomps (try 2) (#16746 ) * move down dtype decomps (try 2) * gpt fixes * fixes * that did need reindex * valid.where * upd to master * reindex	2026-06-25 09:49:43 -07:00
George HotzandGitHub	b5e0bca764	move the load/store gate back to gater (#16747 ) * move the load/store gate back to gater * order	2026-06-25 09:29:32 -07:00
geohot	fa60fa96a6	hotfix: decomp in pyproject	2026-06-25 08:56:34 -07:00
George HotzandGitHub	e72ecd101f	split decomp into files (#16745 ) * split decomp into files * that's in op	2026-06-25 08:40:02 -07:00
George HotzandGitHub	54887cbbdf	move decomp files into codegen (#16744 )	2026-06-25 08:19:02 -07:00
chenyuandGitHub	96c9a68205	imrpove some type casts (#16743 )	2026-06-25 10:48:28 -04:00
chenyuandGitHub	db2257ad09	don't run tinyboxbenchmark on fork (#16741 )	2026-06-25 09:12:56 -04:00
qazalandGitHub	daf8c8f509	viz: simpler test_link_sched_codegen, merge with the other codegen test (#16739 ) * simpler test_link_sched_codegen * merge test_codegen_tracing in the other test * include BEAM.value	2026-06-25 15:02:00 +09:00
qazalandGitHub	695ffb3a08	viz: fix rendering custom Ops.PROGRAM source (#16738 ) * failing test * fix * simplify test	2026-06-25 14:21:31 +09:00
George HotzandGitHub	2f91a401a1	move late decomp after new style (#16734 )	2026-06-24 22:04:59 -07:00
qazalandGitHub	cb9df8bd33	fix profile.sh BENCHMARK=3 exit status (#16736 )	2026-06-25 13:41:03 +09:00
sirhcmandGitHub	63f0e08593	start benchmarks cleanup (#16729 )	2026-06-24 23:33:46 -04:00
wozeparrotandGitHub	e0c69d7a12	llama: fused grad quantize (#16731 )	2026-06-24 23:25:39 -04:00
chenyuandGitHub	65dd099b63	support Invalid in _pad_constant [PR] (#16733 ) and simplified image_conv2d	2026-06-24 21:03:36 -04:00
George HotzandGitHub	9d5d2253b3	refactor to image_valid_dims (#16730 )	2026-06-24 16:53:09 -07:00
nimlgenandGitHub	b99b9e1875	do not assert index in mem coalesing (#16723 )	2026-06-24 14:28:49 +03:00
Nikolay Murzin GitHub Claude Opus 4.8	f38836d14d	cuda: pass kernel dynamic shared memory in the graph runner (#16699 ) The CUDA graph runner hardcoded sharedMemBytes=0 in the kernel node params, so a custom_kernel using dynamic shared memory (extern __shared__) failed with CUDA error 700 on graph replay while the eager path worked. Pass runtime.smem (0 for normal kernels, which use static shared). Co-authored-by: Claude Opus 4.8 (1M context) <[email protected]>	2026-06-24 14:20:10 +03:00
qazalandGitHub	535c806c9f	llama: profile.sh BENCHMARK=3 (#16722 )	2026-06-24 19:20:51 +09:00
chenyuandGitHub	857a435af2	remove dead added_ox (#16721 )	2026-06-23 23:59:17 -04:00
chenyuandGitHub	687ade119e	IMAGE hand_coded_optimizations update (#16720 )	2026-06-23 21:55:28 -04:00
George HotzandGitHub	0a8e61d0c5	switch to the new memory coaleser [pr] (#16716 ) * switch to the new memory coalese * move that stuff * copy in allowed length logic * mulitple buffers * new coalese is better * fine * earlier * fixes * work * work * valid * stack on index const	2026-06-23 18:03:48 -07:00
wozeparrotandGitHub	dfea9e7994	llama: fused silu mul quantize mxfp8 (#16704 )	2026-06-23 16:59:50 -07:00
chenyuandGitHub	ce87d80911	better _drop_valid_stmts [pr] (#16719 ) also dropped the unused is_increasing	2026-06-23 19:35:01 -04:00
George HotzandGitHub	5a2b3b7b06	early dtype decomp (#16718 ) * early dtype decomp * simplify * cleanup * that goes there * doing too much * stupid symbolic rules	2026-06-23 16:07:20 -07:00
sirhcmandGitHub	116045cc8e	ci: remove tensorflow from testoptim (#16717 )	2026-06-23 18:11:48 -04:00
nimlgenandGitHub	7c1d0b6d9a	hcq2: use shrink(bitcast) (#16713 ) * hcq2: use shrink(bitcast) * x	2026-06-23 18:11:39 +03:00