182 lines
5.0 KiB
C++

// Prelude for PyTorch cpp binding.
// clang-format off
#include <torch/extension.h>
// clang-format on
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/CUDAUtils.h>
#include <stdexcept>
#include <string>
#include <vector>
#ifdef SLANG_LLVM
#include "slang-llvm.h"
#else // SLANG_LLVM
#if SLANG_GCC_FAMILY && __GNUC__ < 6
#include <cmath>
#define SLANG_PRELUDE_STD std::
#else
#include <math.h>
#define SLANG_PRELUDE_STD
#endif
#include <assert.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#endif // SLANG_LLVM
#include "../source/core/slang-string.h"
#if defined(_MSC_VER)
#define SLANG_PRELUDE_SHARED_LIB_EXPORT __declspec(dllexport)
#else
#define SLANG_PRELUDE_SHARED_LIB_EXPORT __attribute__((__visibility__("default")))
// # define SLANG_PRELUDE_SHARED_LIB_EXPORT __attribute__ ((dllexport))
// __attribute__((__visibility__("default")))
#endif
#ifdef __cplusplus
#define SLANG_PRELUDE_EXTERN_C extern "C"
#define SLANG_PRELUDE_EXTERN_C_START \
extern "C" \
{
#define SLANG_PRELUDE_EXTERN_C_END }
#else
#define SLANG_PRELUDE_EXTERN_C
#define SLANG_PRELUDE_EXTERN_C_START
#define SLANG_PRELUDE_EXTERN_C_END
#endif
#define SLANG_PRELUDE_NAMESPACE
#ifndef SLANG_NO_THROW
#define SLANG_NO_THROW
#endif
#ifndef SLANG_STDCALL
#define SLANG_STDCALL
#endif
#ifndef SLANG_MCALL
#define SLANG_MCALL SLANG_STDCALL
#endif
#ifndef SLANG_FORCE_INLINE
#define SLANG_FORCE_INLINE inline
#endif
#include "slang-cpp-scalar-intrinsics.h"
#include "slang-cpp-types-core.h"
static const int kSlangTorchTensorMaxDim = 5;
struct TensorView
{
uint8_t* data;
uint32_t strides[kSlangTorchTensorMaxDim];
uint32_t sizes[kSlangTorchTensorMaxDim];
uint32_t dimensionCount;
};
TensorView make_tensor_view(
torch::Tensor val,
const char* name,
torch::ScalarType targetScalarType,
bool requireContiguous)
{
// We're currently not trying to implicitly cast or transfer to device for two reasons:
// 1. There appears to be a bug with .to() where successive calls after the first one fail.
// 2. Silent casts like this can cause large memory allocations & unexpected overheads.
// It's better to be explicit.
// Expect tensors to be on CUDA device
if (!val.device().is_cuda())
throw std::runtime_error(
std::string(name).append(": tensor is not on CUDA device.").c_str());
// Expect tensors to be the right type.
if (val.dtype() != targetScalarType)
throw std::runtime_error(
std::string(name).append(": tensor is not of the expected type.").c_str());
// Check that the tensor is contiguous
if (requireContiguous && !val.is_contiguous())
throw std::runtime_error(std::string(name).append(": tensor is not contiguous.").c_str());
TensorView res = {};
res.dimensionCount = val.dim();
res.data = nullptr;
size_t elementSize = 4;
switch (val.scalar_type())
{
case torch::kInt8:
case torch::kUInt8:
elementSize = 1;
res.data = (uint8_t*)val.data_ptr<uint8_t>();
break;
case torch::kBFloat16:
elementSize = 2;
res.data = (uint8_t*)val.data_ptr<torch::BFloat16>();
break;
case torch::kFloat16:
elementSize = 2;
res.data = (uint8_t*)val.data_ptr<at::Half>();
break;
case torch::kInt16:
elementSize = 2;
res.data = (uint8_t*)val.data_ptr<int16_t>();
break;
case torch::kFloat32:
elementSize = 4;
res.data = (uint8_t*)val.data_ptr<float>();
break;
case torch::kInt32:
elementSize = 4;
res.data = (uint8_t*)val.data_ptr<int32_t>();
break;
case torch::kFloat64:
elementSize = 8;
res.data = (uint8_t*)val.data_ptr<double>();
break;
case torch::kInt64:
elementSize = 8;
res.data = (uint8_t*)val.data_ptr<int64_t>();
break;
case torch::kBool:
elementSize = 1;
res.data = (uint8_t*)val.data_ptr<bool>();
break;
}
if (val.dim() > kSlangTorchTensorMaxDim)
throw std::runtime_error(std::string(name)
.append(": number of dimensions exceeds limit (")
.append(std::to_string(kSlangTorchTensorMaxDim))
.append(")")
.c_str());
bool isEmpty = true;
for (int i = 0; i < val.dim(); ++i)
{
res.strides[i] = val.stride(i) * elementSize;
if (res.strides[i] == 0)
throw std::runtime_error(
std::string(name)
.append(": tensors with broadcasted dimensions are not supported (use "
"tensor.contiguous() to make tensor whole)")
.c_str());
res.sizes[i] = val.size(i);
if (res.sizes[i] > 0)
isEmpty = false;
}
if (!res.data && !isEmpty)
throw std::runtime_error(std::string(name).append(": data pointer is invalid.").c_str());
return res;
}
#define SLANG_PRELUDE_EXPORT