sed -i '1s/^/# mypy: ignore-errors\n/' $1

sed -i 's/ *$//' $1

grep FIXME_STUB $1 || true

path=$1; shift

name=$1; shift

cat <<EOF | sed -i "/import ctypes.*/r /dev/stdin" $path

PATHS_TO_TRY = [

$(for p in "$@"; do echo " $p,"; done)

]

def _try_dlopen_$name():

library = ctypes.util.find_library("$name")

if library: return ctypes.CDLL(library)

for candidate in PATHS_TO_TRY:

try: return ctypes.CDLL(candidate)

except OSError: pass

raise RuntimeError("library $name not found")

EOF

clang2py /usr/include/CL/cl.h -o $BASE/opencl.py -l /usr/lib/x86_64-linux-gnu/libOpenCL.so.1 -k cdefstum

fixup $BASE/opencl.py

# hot patches

sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/opencl.py

sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libOpenCL.so.1')\ctypes.CDLL(ctypes.util.find_library('OpenCL'))\g" $BASE/opencl.py

python3 -c "import tinygrad.runtime.autogen.opencl"

clang2py /opt/rocm/include/hip/hip_ext.h /opt/rocm/include/hip/hiprtc.h \

/opt/rocm/include/hip/hip_runtime_api.h /opt/rocm/include/hip/driver_types.h \

--clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/hip.py -l /opt/rocm/lib/libamdhip64.so

echo "hipDeviceProp_t = hipDeviceProp_tR0600" >> $BASE/hip.py

echo "hipGetDeviceProperties = hipGetDevicePropertiesR0600" >> $BASE/hip.py

fixup $BASE/hip.py

# we can trust HIP is always at /opt/rocm/lib

#sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/hip.py

#sed -i "s\ctypes.CDLL('/opt/rocm/lib/libhiprtc.so')\ctypes.CDLL(ctypes.util.find_library('hiprtc'))\g" $BASE/hip.py

#sed -i "s\ctypes.CDLL('/opt/rocm/lib/libamdhip64.so')\ctypes.CDLL(ctypes.util.find_library('amdhip64'))\g" $BASE/hip.py

sed -i "s\import ctypes\import ctypes, os\g" $BASE/hip.py

sed -i "s\'/opt/rocm/\os.getenv('ROCM_PATH', '/opt/rocm/')+'/\g" $BASE/hip.py

python3 -c "import tinygrad.runtime.autogen.hip"

clang2py /opt/rocm/include/amd_comgr/amd_comgr.h \

--clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/comgr.py -l /opt/rocm/lib/libamd_comgr.so

fixup $BASE/comgr.py

sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/comgr.py

patch_dlopen $BASE/comgr.py amd_comgr "'/opt/rocm/lib/libamd_comgr.so'" "os.getenv('ROCM_PATH', '')+'/lib/libamd_comgr.so'"

sed -i "s\ctypes.CDLL('/opt/rocm/lib/libamd_comgr.so')\_try_dlopen_amd_comgr()\g" $BASE/comgr.py

python3 -c "import tinygrad.runtime.autogen.comgr"

clang2py /usr/include/linux/kfd_ioctl.h -o $BASE/kfd.py -k cdefstum

fixup $BASE/kfd.py

sed -i "s\import ctypes\import ctypes, os\g" $BASE/kfd.py

python3 -c "import tinygrad.runtime.autogen.kfd"

clang2py /usr/include/cuda.h -o $BASE/cuda.py -l /usr/lib/x86_64-linux-gnu/libcuda.so

sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/cuda.py

sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libcuda.so')\ctypes.CDLL(ctypes.util.find_library('cuda'))\g" $BASE/cuda.py

fixup $BASE/cuda.py

python3 -c "import tinygrad.runtime.autogen.cuda"

clang2py /usr/local/cuda/include/nvrtc.h /usr/local/cuda/include/nvJitLink.h -o $BASE/nvrtc.py -l /usr/local/cuda/lib64/libnvrtc.so -l /usr/local/cuda/lib64/libnvJitLink.so

sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/nvrtc.py

sed -i "s\ctypes.CDLL('/usr/local/cuda/lib64/libnvrtc.so')\ctypes.CDLL(ctypes.util.find_library('nvrtc'))\g" $BASE/nvrtc.py

sed -i "s\ctypes.CDLL('/usr/local/cuda/lib64/libnvJitLink.so')\ctypes.CDLL(ctypes.util.find_library('nvJitLink'))\g" $BASE/nvrtc.py

fixup $BASE/nvrtc.py

python3 -c "import tinygrad.runtime.autogen.nvrtc"

NVKERN_COMMIT_HASH=d6b75a34094b0f56c2ccadf14e5d0bd515ed1ab6

NVKERN_SRC=/tmp/open-gpu-kernel-modules-$NVKERN_COMMIT_HASH

if [ ! -d "$NVKERN_SRC" ]; then

git clone https://github.com/tinygrad/open-gpu-kernel-modules $NVKERN_SRC

pushd .

cd $NVKERN_SRC

git reset --hard $NVKERN_COMMIT_HASH

popd

fi

clang2py -k cdefstum \

extra/nv_gpu_driver/clc6c0qmd.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl0080.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl2080_notification.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl83de.h \

$NVKERN_SRC/src/nvidia/generated/g_allclasses.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc6c0.h \

$NVKERN_SRC/kernel-open/nvidia-uvm/clc6b5.h \

$NVKERN_SRC/kernel-open/nvidia-uvm/uvm_ioctl.h \

$NVKERN_SRC/kernel-open/nvidia-uvm/uvm_linux_ioctl.h \

$NVKERN_SRC/kernel-open/nvidia-uvm/hwref/ampere/ga100/dev_fault.h \

$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv_escape.h \

$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl.h \

$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl-numbers.h \

$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl-numa.h \

$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-unix-nvos-params-wrappers.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/alloc/alloc_channel.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/nvos.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl0000/*.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl0080/*.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl2080/*.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl83de/*.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrlc36f.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrlcb33.h \

$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrla06c.h \

--clang-args="-include $NVKERN_SRC/src/common/sdk/nvidia/inc/nvtypes.h -I$NVKERN_SRC/src/common/inc -I$NVKERN_SRC/kernel-open/nvidia-uvm -I$NVKERN_SRC/kernel-open/common/inc -I$NVKERN_SRC/src/common/sdk/nvidia/inc -I$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include -I$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl" \

-o $BASE/nv_gpu.py

fixup $BASE/nv_gpu.py

sed -i "s\(0000000001)\1\g" $BASE/nv_gpu.py

sed -i "s\import ctypes\import ctypes, os\g" $BASE/nv_gpu.py

sed -i 's/#\?\s\([A-Za-z0-9_]\+\) = MW ( \([0-9]\+\) : \([0-9]\+\) )/\1 = (\2 , \3)/' $BASE/nv_gpu.py # NVC6C0_QMDV03_00 processing

sed -i 's/#\sdef NVC6C0_QMD\([A-Za-z0-9_()]\+\):/def NVC6C0_QMD\1:/' $BASE/nv_gpu.py

sed -i 's/#\s*return MW(\([0-9i()*+]\+\):\([0-9i()*+]\+\))/ return (\1 , \2)/' $BASE/nv_gpu.py

sed -i 's/#\?\s*\(.*\)\s*=\s*\(NV\)\?BIT\(32\)\?\s*(\s*\([0-9]\+\)\s*)/\1 = (1 << \4)/' $BASE/nv_gpu.py # name = BIT(x) -> name = (1 << x)

sed -i "s/UVM_\([A-Za-z0-9_]\+\) = \['i', '(', '\([0-9]\+\)', ')'\]/UVM_\1 = \2/" $BASE/nv_gpu.py # UVM_name = ['i', '(', '<num>', ')'] -> UVM_name = <num>

# Parse status codes

sed -n '1i\

nv_status_codes = {}

/^NV_STATUS_CODE/ { s/^NV_STATUS_CODE(\([^,]*\), *\([^,]*\), *"\([^"]*\)") *.*$/\1 = \2\nnv_status_codes[\1] = "\3"/; p }' $NVKERN_SRC/src/common/sdk/nvidia/inc/nvstatuscodes.h >> $BASE/nv_gpu.py

python3 -c "import tinygrad.runtime.autogen.nv_gpu"

# clang2py broken when pass -x c++ to prev headers

clang2py -k cdefstum \

extra/hip_gpu_driver/sdma_registers.h \

extra/hip_gpu_driver/nvd.h \

extra/hip_gpu_driver/sdma_v6_0_0_pkt_open.h \

extra/hip_gpu_driver/gc_11_0_0_offset.h \

extra/hip_gpu_driver/gc_10_3_0_offset.h \

--clang-args="-I/opt/rocm/include -x c++" \

-o $BASE/amd_gpu.py

fixup $BASE/amd_gpu.py

sed -i "s\import ctypes\import ctypes, os\g" $BASE/amd_gpu.py

python3 -c "import tinygrad.runtime.autogen.amd_gpu"

clang2py \

/opt/rocm/include/hsa/hsa.h \

/opt/rocm/include/hsa/hsa_ext_amd.h \

/opt/rocm/include/hsa/amd_hsa_signal.h \

/opt/rocm/include/hsa/amd_hsa_queue.h \

/opt/rocm/include/hsa/amd_hsa_kernel_code.h \

/opt/rocm/include/hsa/hsa_ext_finalize.h /opt/rocm/include/hsa/hsa_ext_image.h \

/opt/rocm/include/hsa/hsa_ven_amd_aqlprofile.h \

--clang-args="-I/opt/rocm/include" \

-o $BASE/hsa.py -l /opt/rocm/lib/libhsa-runtime64.so

fixup $BASE/hsa.py

sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/hsa.py

sed -i "s\ctypes.CDLL('/opt/rocm/lib/libhsa-runtime64.so')\ctypes.CDLL(os.getenv('ROCM_PATH')+'/lib/libhsa-runtime64.so' if os.getenv('ROCM_PATH') else ctypes.util.find_library('hsa-runtime64'))\g" $BASE/hsa.py

python3 -c "import tinygrad.runtime.autogen.hsa"

clang2py -k cdefstum \

/usr/include/liburing.h \

/usr/include/linux/io_uring.h \

-o $BASE/io_uring.py

sed -r '/^#define __NR_io_uring/ s/^#define __(NR_io_uring[^ ]+) (.*)$/\1 = \2/; t; d' /usr/include/asm-generic/unistd.h >> $BASE/io_uring.py # io_uring syscalls numbers

fixup $BASE/io_uring.py

clang2py -k cdefstum \

$(dpkg -L libc6-dev | grep sys/mman.h) \

$(dpkg -L libc6-dev | grep sys/syscall.h) \

/usr/include/elf.h \

/usr/include/unistd.h \

-o $BASE/libc.py

sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/libc.py

sed -i "s\FIXME_STUB\libc\g" $BASE/libc.py

sed -i "s\FunctionFactoryStub()\None if (libc_path := ctypes.util.find_library('c')) is None else ctypes.CDLL(libc_path)\g" $BASE/libc.py

fixup $BASE/libc.py

clang2py extra/qcom_gpu_driver/msm_kgsl.h -o $BASE/kgsl.py -k cdefstum

fixup $BASE/kgsl.py

sed -i "s\import ctypes\import ctypes, os\g" $BASE/kgsl.py

sed -nE 's/#define ([A-Za-z0-9_]+)_SHIFT\s*[^\S\r\n]*[0-9]*$/def \1(val): return (val << \1_SHIFT) \& \1_MASK/p' extra/qcom_gpu_driver/msm_kgsl.h >> $BASE/kgsl.py

python3 -c "import tinygrad.runtime.autogen.kgsl"

clang2py extra/qcom_gpu_driver/a6xx.xml.h -o $BASE/adreno.py -k cestum

sed -nE 's/#define ([A-Za-z0-9_]+)__SHIFT\s*[^\S\r\n]*[0-9]*$/def \1(val): return (val << \1__SHIFT) \& \1__MASK/p' extra/qcom_gpu_driver/a6xx.xml.h >> $BASE/adreno.py

fixup $BASE/adreno.py

sed -i "s\import ctypes\import ctypes, os\g" $BASE/adreno.py

python3 -c "import tinygrad.runtime.autogen.adreno"

clang2py -k cdefstum \

extra/dsp/include/ion.h \

extra/dsp/include/msm_ion.h \

extra/dsp/include/adsprpc_shared.h \

extra/dsp/include/remote_default.h \

extra/dsp/include/apps_std.h \

-o $BASE/qcom_dsp.py

fixup $BASE/qcom_dsp.py

python3 -c "import tinygrad.runtime.autogen.qcom_dsp"