Skip to content

GitLab

Commit 74d88bf8 authored by sangwz's avatar sangwz
Browse files

Merge branch 'dtk25.04' of http://developer.sourcefind.cn/codes/OpenDAS/dgl into 2.2.1

parents 2a1ac588 314cedc1
Hide whitespace changes
Inline Side-by-side
Showing with 517 additions and 362 deletions
src/array/cuda/array_index_select.cuh
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2021-2022 by Contributors
* @file array/cuda/array_index_select.cuh
......
src/array/cuda/array_index_select.cu → src/array/cuda/array_index_select.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2019 by Contributors
* @file array/cpu/array_index_select.cu
* @brief Array index select GPU implementation
*/
#include <dgl/array.h>
#include "../../../include/dgl/array.h"
#include "../../runtime/cuda/cuda_common.h"
#include "./array_index_select.cuh"
#include "./utils.h"
#include "array_index_select.cuh"
#include "utils.h"
namespace dgl {
using runtime::NDArray;
......@@ -31,9 +34,10 @@ NDArray IndexSelect(NDArray array, IdArray index) {
DType* ret_data = static_cast<DType*>(ret->data);
const DType* array_data = static_cast<DType*>(cuda::GetDevicePointer(array));
const IdType* idx_data = static_cast<IdType*>(index->data);
// const IdType* idx_data = static_cast<IdType*>(index->data);
const IdType* idx_data = static_cast<IdType*>(cuda::GetDevicePointer(index));
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
if (num_feat == 1) {
const int nt = cuda::FindNumThreads(len);
const int nb = (len + nt - 1) / nt;
......@@ -61,9 +65,9 @@ template NDArray IndexSelect<kDGLCUDA, int64_t, int64_t>(NDArray, IdArray);
template NDArray IndexSelect<kDGLCUDA, __half, int32_t>(NDArray, IdArray);
template NDArray IndexSelect<kDGLCUDA, __half, int64_t>(NDArray, IdArray);
#if BF16_ENABLED
template NDArray IndexSelect<kDGLCUDA, __nv_bfloat16, int32_t>(
template NDArray IndexSelect<kDGLCUDA, __hip_bfloat16, int32_t>(
NDArray, IdArray);
template NDArray IndexSelect<kDGLCUDA, __nv_bfloat16, int64_t>(
template NDArray IndexSelect<kDGLCUDA, __hip_bfloat16, int64_t>(
NDArray, IdArray);
#endif // BF16_ENABLED
template NDArray IndexSelect<kDGLCUDA, float, int32_t>(NDArray, IdArray);
......@@ -87,7 +91,7 @@ template uint32_t IndexSelect<kDGLCUDA, uint32_t>(NDArray array, int64_t index);
template uint64_t IndexSelect<kDGLCUDA, uint64_t>(NDArray array, int64_t index);
template __half IndexSelect<kDGLCUDA, __half>(NDArray array, int64_t index);
#if BF16_ENABLED
template __nv_bfloat16 IndexSelect<kDGLCUDA, __nv_bfloat16>(
template __hip_bfloat16 IndexSelect<kDGLCUDA, __hip_bfloat16>(
NDArray array, int64_t index);
#endif // BF16_ENABLED
template float IndexSelect<kDGLCUDA, float>(NDArray array, int64_t index);
......
src/array/cuda/array_nonzero.cu → src/array/cuda/array_nonzero.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020 by Contributors
* @file array/cpu/array_nonzero.cc
......@@ -5,11 +7,13 @@
*/
#include <dgl/array.h>
#include "../../../include/dgl/array.h"
#include <cub/cub.cuh>
#include <hipcub/hipcub.hpp>
#include "../../runtime/cuda/cuda_common.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
using runtime::NDArray;
......@@ -33,24 +37,24 @@ IdArray NonZero(IdArray array) {
const int64_t len = array->shape[0];
IdArray ret = NewIdArray(len, ctx, 64);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
const IdType* const in_data = static_cast<const IdType*>(array->data);
int64_t* const out_data = static_cast<int64_t*>(ret->data);
IsNonZeroIndex<IdType> comp(in_data);
cub::CountingInputIterator<int64_t> counter(0);
hipcub::CountingInputIterator<int64_t> counter(0);
// room for cub to output on GPU
int64_t* d_num_nonzeros =
static_cast<int64_t*>(device->AllocWorkspace(ctx, sizeof(int64_t)));
size_t temp_size = 0;
CUDA_CALL(cub::DeviceSelect::If(
CUDA_CALL(hipcub::DeviceSelect::If(
nullptr, temp_size, counter, out_data, d_num_nonzeros, len, comp,
stream));
void* temp = device->AllocWorkspace(ctx, temp_size);
CUDA_CALL(cub::DeviceSelect::If(
CUDA_CALL(hipcub::DeviceSelect::If(
temp, temp_size, counter, out_data, d_num_nonzeros, len, comp, stream));
device->FreeWorkspace(ctx, temp);
......
src/array/cuda/array_op_impl.cu → src/array/cuda/array_op_impl.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020-2021 by Contributors
* @file array/cuda/array_op_impl.cu
* @brief Array operator GPU implementation
*/
#include <dgl/array.h>
#include "../../../include/dgl/array.h"
#include "../../runtime/cuda/cuda_common.h"
#include "../../runtime/cuda/cuda_hashtable.cuh"
#include "../arith.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
using runtime::NDArray;
......@@ -36,7 +40,7 @@ IdArray BinaryElewise(IdArray lhs, IdArray rhs) {
const IdType* lhs_data = static_cast<IdType*>(lhs->data);
const IdType* rhs_data = static_cast<IdType*>(rhs->data);
IdType* ret_data = static_cast<IdType*>(ret->data);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
int nt = cuda::FindNumThreads(len);
int nb = (len + nt - 1) / nt;
CUDA_KERNEL_CALL(
......@@ -107,7 +111,7 @@ IdArray BinaryElewise(IdArray lhs, IdType rhs) {
IdArray ret = NewIdArray(lhs->shape[0], lhs->ctx, lhs->dtype.bits);
const IdType* lhs_data = static_cast<IdType*>(lhs->data);
IdType* ret_data = static_cast<IdType*>(ret->data);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
int nt = cuda::FindNumThreads(len);
int nb = (len + nt - 1) / nt;
CUDA_KERNEL_CALL(
......@@ -178,7 +182,7 @@ IdArray BinaryElewise(IdType lhs, IdArray rhs) {
IdArray ret = NewIdArray(rhs->shape[0], rhs->ctx, rhs->dtype.bits);
const IdType* rhs_data = static_cast<IdType*>(rhs->data);
IdType* ret_data = static_cast<IdType*>(ret->data);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
int nt = cuda::FindNumThreads(len);
int nb = (len + nt - 1) / nt;
CUDA_KERNEL_CALL(
......@@ -249,7 +253,7 @@ IdArray UnaryElewise(IdArray lhs) {
IdArray ret = NewIdArray(lhs->shape[0], lhs->ctx, lhs->dtype.bits);
const IdType* lhs_data = static_cast<IdType*>(lhs->data);
IdType* ret_data = static_cast<IdType*>(ret->data);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
int nt = cuda::FindNumThreads(len);
int nb = (len + nt - 1) / nt;
CUDA_KERNEL_CALL(
......@@ -277,7 +281,7 @@ template <DGLDeviceType XPU, typename DType>
NDArray Full(DType val, int64_t length, DGLContext ctx) {
NDArray ret = NDArray::Empty({length}, DGLDataTypeTraits<DType>::dtype, ctx);
DType* ret_data = static_cast<DType*>(ret->data);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
int nt = cuda::FindNumThreads(length);
int nb = (length + nt - 1) / nt;
CUDA_KERNEL_CALL(
......@@ -292,8 +296,8 @@ template IdArray Full<kDGLCUDA, int64_t>(
template IdArray Full<kDGLCUDA, __half>(
__half val, int64_t length, DGLContext ctx);
#if BF16_ENABLED
template IdArray Full<kDGLCUDA, __nv_bfloat16>(
__nv_bfloat16 val, int64_t length, DGLContext ctx);
template IdArray Full<kDGLCUDA, __hip_bfloat16>(
__hip_bfloat16 val, int64_t length, DGLContext ctx);
#endif // BF16_ENABLED
template IdArray Full<kDGLCUDA, float>(
float val, int64_t length, DGLContext ctx);
......@@ -319,7 +323,7 @@ IdArray Range(IdType low, IdType high, DGLContext ctx) {
IdArray ret = NewIdArray(length, ctx, sizeof(IdType) * 8);
if (length == 0) return ret;
IdType* ret_data = static_cast<IdType*>(ret->data);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
int nt = cuda::FindNumThreads(length);
int nb = (length + nt - 1) / nt;
CUDA_KERNEL_CALL(
......@@ -355,7 +359,7 @@ IdArray Relabel_(const std::vector<IdArray>& arrays) {
const auto& ctx = arrays[0]->ctx;
auto device = runtime::DeviceAPI::Get(ctx);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
// build node maps and get the induced nodes
OrderedHashTable<IdType> node_map(total_length, ctx, stream);
......@@ -364,7 +368,7 @@ IdArray Relabel_(const std::vector<IdArray>& arrays) {
static_cast<int64_t*>(device->AllocWorkspace(ctx, sizeof(int64_t)));
IdArray induced_nodes = NewIdArray(total_length, ctx, sizeof(IdType) * 8);
CUDA_CALL(cudaMemsetAsync(
CUDA_CALL(hipMemsetAsync(
num_induced_device, 0, sizeof(*num_induced_device), stream));
node_map.FillWithDuplicates(
......@@ -416,7 +420,7 @@ IdArray AsNumBits(IdArray arr, uint8_t bits) {
const std::vector<int64_t> shape(arr->shape, arr->shape + arr->ndim);
IdArray ret = IdArray::Empty(shape, DGLDataType{kDGLInt, bits, 1}, arr->ctx);
const int64_t length = ret.NumElements();
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
int nt = cuda::FindNumThreads(length);
int nb = (length + nt - 1) / nt;
if (bits == 32) {
......
src/array/cuda/array_scatter.cu → src/array/cuda/array_scatter.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2019 by Contributors
* @file array/cuda/array_scatter.cu
* @brief Array scatter GPU implementation
*/
#include <dgl/array.h>
#include "../../../include/dgl/array.h"
#include "../../runtime/cuda/cuda_common.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
using runtime::NDArray;
......@@ -31,7 +35,7 @@ void Scatter_(IdArray index, NDArray value, NDArray out) {
const DType* val = value.Ptr<DType>();
DType* outd = out.Ptr<DType>();
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
const int nt = cuda::FindNumThreads(len);
const int nb = (len + nt - 1) / nt;
CUDA_KERNEL_CALL(_ScatterKernel, nb, nt, 0, stream, idx, val, len, outd);
......@@ -41,7 +45,7 @@ template void Scatter_<kDGLCUDA, int32_t, int32_t>(IdArray, NDArray, NDArray);
template void Scatter_<kDGLCUDA, int64_t, int32_t>(IdArray, NDArray, NDArray);
template void Scatter_<kDGLCUDA, __half, int32_t>(IdArray, NDArray, NDArray);
#if BF16_ENABLED
template void Scatter_<kDGLCUDA, __nv_bfloat16, int32_t>(
template void Scatter_<kDGLCUDA, __hip_bfloat16, int32_t>(
IdArray, NDArray, NDArray);
#endif // BF16_ENABLED
template void Scatter_<kDGLCUDA, float, int32_t>(IdArray, NDArray, NDArray);
......@@ -50,7 +54,7 @@ template void Scatter_<kDGLCUDA, int32_t, int64_t>(IdArray, NDArray, NDArray);
template void Scatter_<kDGLCUDA, int64_t, int64_t>(IdArray, NDArray, NDArray);
template void Scatter_<kDGLCUDA, __half, int64_t>(IdArray, NDArray, NDArray);
#if BF16_ENABLED
template void Scatter_<kDGLCUDA, __nv_bfloat16, int64_t>(
template void Scatter_<kDGLCUDA, __hip_bfloat16, int64_t>(
IdArray, NDArray, NDArray);
#endif // BF16_ENABLED
template void Scatter_<kDGLCUDA, float, int64_t>(IdArray, NDArray, NDArray);
......
src/array/cuda/array_sort.cu → src/array/cuda/array_sort.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020 by Contributors
* @file array/cpu/array_sort.cu
* @brief Array sort GPU implementation
*/
#include <dgl/array.h>
#include "../../../include/dgl/array.h"
#include <cub/cub.cuh>
#include <hipcub/hipcub.hpp>
#include "../../runtime/cuda/cuda_common.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
using runtime::NDArray;
......@@ -29,20 +33,20 @@ std::pair<IdArray, IdArray> Sort(IdArray array, int num_bits) {
IdType* keys_out = sorted_array.Ptr<IdType>();
int64_t* values_out = sorted_idx.Ptr<int64_t>();
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
if (num_bits == 0) {
num_bits = sizeof(IdType) * 8;
}
// Allocate workspace
size_t workspace_size = 0;
CUDA_CALL(cub::DeviceRadixSort::SortPairs(
CUDA_CALL(hipcub::DeviceRadixSort::SortPairs(
nullptr, workspace_size, keys_in, keys_out, values_in, values_out, nitems,
0, num_bits, stream));
void* workspace = device->AllocWorkspace(ctx, workspace_size);
// Compute
CUDA_CALL(cub::DeviceRadixSort::SortPairs(
CUDA_CALL(hipcub::DeviceRadixSort::SortPairs(
workspace, workspace_size, keys_in, keys_out, values_in, values_out,
nitems, 0, num_bits, stream));
......
src/array/cuda/atomic.cuh
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
/**
* Copyright (c) 2019 by Contributors
* @file array/cuda/atomic.cuh
......@@ -6,7 +7,7 @@
#ifndef DGL_ARRAY_CUDA_ATOMIC_CUH_
#define DGL_ARRAY_CUDA_ATOMIC_CUH_
#include <cuda_runtime.h>
#include <hip/hip_runtime.h>
#include <cassert>
#include <cstdint>
......@@ -15,8 +16,8 @@
#include "bf16.cuh"
#include "fp16.cuh"
#if __CUDA_ARCH__ >= 600
#include <cuda_fp16.h>
#if __HIPCC__
#include <hip/hip_fp16.h>
#endif
namespace dgl {
......@@ -56,39 +57,39 @@ struct Cast {
template <>
struct Cast<half> {
typedef Code<sizeof(half)>::Type Type;
static __device__ __forceinline__ Type Encode(half val) {
typedef half Type;
static __host__ __device__ __forceinline__ Type Encode(half val) {
return __half_as_ushort(val);
}
static __device__ __forceinline__ half Decode(Type code) {
static __host__ __device__ __forceinline__ half Decode(Type code) {
return __ushort_as_half(code);
}
};
#if BF16_ENABLED
template <>
struct Cast<__nv_bfloat16> {
typedef Code<sizeof(__nv_bfloat16)>::Type Type;
static __device__ __forceinline__ Type Encode(__nv_bfloat16 val) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
struct Cast<__hip_bfloat16> {
typedef __hip_bfloat16 Type;
static __host__ __device__ __forceinline__ Type Encode(__hip_bfloat16 val) {
#if defined(__HIP_DEVICE_COMPILE__)
return __bfloat16_as_ushort(val);
#else
printf(
"Atomic operations are not supported for bfloat16 (BF16) "
"on GPUs with compute capability less than 8.0.\n");
__trap();
// //__trap();
return static_cast<Type>(0);
#endif
}
static __device__ __forceinline__ __nv_bfloat16 Decode(Type code) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
static __host__ __device__ __forceinline__ __hip_bfloat16 Decode(Type code) {
#if defined(__HIP_DEVICE_COMPILE__)
return __ushort_as_bfloat16(code);
#else
printf(
"Atomic operations are not supported for bfloat16 (BF16) "
"on GPUs with compute capability less than 8.0.\n");
__trap();
return static_cast<__nv_bfloat16>(0.0f);
//__trap();
return static_cast<__hip_bfloat16>(0.0f);
#endif
}
};
......@@ -116,12 +117,12 @@ struct Cast<double> {
}
};
static __device__ __forceinline__ unsigned short int atomicCASshort( // NOLINT
static __host__ __device__ __forceinline__ unsigned short int atomicCASshort( // NOLINT
unsigned short int* address, // NOLINT
unsigned short int compare, // NOLINT
unsigned short int val) { // NOLINT
static_assert(CUDART_VERSION >= 10000, "Requires at least CUDA 10");
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__) >= 700)
static_assert(DTKRT_VERSION >= 10000, "Requires at least CUDA 10");
#if defined(__HIP_DEVICE_COMPILE__) && 0
return atomicCAS(address, compare, val);
#else
(void)address;
......@@ -130,9 +131,9 @@ static __device__ __forceinline__ unsigned short int atomicCASshort( // NOLINT
printf(
"Atomic operations are not supported for half precision (FP16) "
"on this GPU.\n");
__trap();
abort();
return val;
#endif // (defined(__CUDA_ARCH__) && (__CUDA_ARCH__) >= 700)
#endif // (defined(__HIP_DEVICE_COMPILE__)
}
#define DEFINE_ATOMIC(NAME) \
......@@ -168,19 +169,54 @@ static __device__ __forceinline__ unsigned short int atomicCASshort( // NOLINT
return Cast<dtype>::Decode(old); \
}
#define DEFINE_ATOMIC_16BIT_MAX(NAME, dtype) \
template <> \
__device__ __forceinline__ dtype Atomic##NAME<dtype>( \
dtype * addr, dtype val) { \
typedef uint16_t CT; \
CT* addr_as_ui = reinterpret_cast<CT*>(addr); \
CT old = *addr_as_ui; \
CT assumed = old; \
do { \
assumed = old; \
old = atomicCASshort( \
addr_as_ui, assumed, \
Cast<dtype>::Encode(dtype(max((float)val, (float)dtype(old))))); \
} while (assumed != old); \
return Cast<dtype>::Decode(old); \
}
#define DEFINE_ATOMIC_16BIT_MIN(NAME, dtype) \
template <> \
__device__ __forceinline__ dtype Atomic##NAME<dtype>( \
dtype * addr, dtype val) { \
typedef uint16_t CT; \
CT* addr_as_ui = reinterpret_cast<CT*>(addr); \
CT old = *addr_as_ui; \
CT assumed = old; \
do { \
assumed = old; \
old = atomicCASshort( \
addr_as_ui, assumed, \
Cast<dtype>::Encode(dtype(min((float)val,(float)old)))); \
} while (assumed != old); \
return Cast<dtype>::Decode(old); \
}
#define OP(a, b) max(a, b)
DEFINE_ATOMIC(Max)
DEFINE_ATOMIC_16BIT(Max, half)
DEFINE_ATOMIC_16BIT_MAX(Max, half)
#if BF16_ENABLED
DEFINE_ATOMIC_16BIT(Max, __nv_bfloat16)
#define OP_BF(a, b) max_bf((float)a, (float)b)
DEFINE_ATOMIC_16BIT_MAX(Max, __hip_bfloat16)
#endif // BF16_ENABLED
#undef OP
#define OP(a, b) min(a, b)
DEFINE_ATOMIC(Min)
DEFINE_ATOMIC_16BIT(Min, half)
DEFINE_ATOMIC_16BIT_MIN(Min, half)
#if BF16_ENABLED
DEFINE_ATOMIC_16BIT(Min, __nv_bfloat16)
DEFINE_ATOMIC_16BIT_MIN(Min, __hip_bfloat16)
#endif // BF16_ENABLED
#undef OP
......@@ -256,7 +292,7 @@ inline __device__ int32_t AtomicMax(int32_t* const address, const int32_t val) {
template <>
__device__ __forceinline__ float AtomicAdd<float>(float* addr, float val) {
#if __CUDA_ARCH__ >= 200
#if __HIP_DEVICE_COMPILE__
return atomicAdd(addr, val);
#else
typedef float T;
......@@ -270,12 +306,12 @@ __device__ __forceinline__ float AtomicAdd<float>(float* addr, float val) {
addr_as_ui, assumed, Cast<T>::Encode(Cast<T>::Decode(old) + val));
} while (assumed != old);
return Cast<T>::Decode(old);
#endif // __CUDA_ARCH__
#endif // __HIP_DEVICE_COMPILE__
}
template <>
__device__ __forceinline__ double AtomicAdd<double>(double* addr, double val) {
#if __CUDA_ARCH__ >= 600
#if __HIP_DEVICE_COMPILE__
return atomicAdd(addr, val);
#else
typedef double T;
......@@ -292,11 +328,11 @@ __device__ __forceinline__ double AtomicAdd<double>(double* addr, double val) {
#endif
}
#if defined(CUDART_VERSION) && CUDART_VERSION >= 10000
#if defined(DTKRT_VERSION) && DTKRT_VERSION >= 10000
template <>
__device__ __forceinline__ half AtomicAdd<half>(half* addr, half val) {
// make sure we have half support
#if __CUDA_ARCH__ >= 700
#if __HIP_DEVICE_COMPILE__
return atomicAdd(addr, val);
#else
(void)addr;
......@@ -304,18 +340,18 @@ __device__ __forceinline__ half AtomicAdd<half>(half* addr, half val) {
printf(
"Atomic operations are not supported for half precision (FP16) "
"on this GPU.\n");
__trap();
// //__trap();
return val;
#endif // __CUDA_ARCH__ >= 700
#endif // __HIP_DEVICE_COMPILE__
}
#endif // defined(CUDART_VERSION) && CUDART_VERSION >= 10000
#endif // defined(DTKRT_VERSION) && DTKRT_VERSION >= 10000
#if BF16_ENABLED
template <>
__device__ __forceinline__ __nv_bfloat16
AtomicAdd<__nv_bfloat16>(__nv_bfloat16* addr, __nv_bfloat16 val) {
__device__ __forceinline__ __hip_bfloat16
AtomicAdd<__hip_bfloat16>(__hip_bfloat16* addr, __hip_bfloat16 val) {
// make sure we have bfloat16 support
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
#if defined(__HIP_DEVICE_COMPILE__)
return atomicAdd(addr, val);
#else
(void)addr;
......@@ -323,9 +359,9 @@ AtomicAdd<__nv_bfloat16>(__nv_bfloat16* addr, __nv_bfloat16 val) {
printf(
"Atomic operations are not supported for bfloat16 (BF16) "
"on GPUs with compute capability less than 8.0.\n");
__trap();
//__trap();
return val;
#endif // defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
#endif // defined(__HIP_DEVICE_COMPILE__)
}
#endif // BF16_ENABLED
......
src/array/cuda/bf16.cuh
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
/**
* Copyright (c) 2022 by Contributors
*
......@@ -18,131 +19,142 @@
*/
#ifndef DGL_ARRAY_CUDA_BF16_CUH_
#define DGL_ARRAY_CUDA_BF16_CUH_
#include <hip/hip_runtime.h>
#if BF16_ENABLED
#include <cuda_bf16.h>
#include <hip/hip_bf16.h>
#include <algorithm>
static __device__ __forceinline__ __nv_bfloat16
max(__nv_bfloat16 a, __nv_bfloat16 b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
static __device__ __forceinline__ __hip_bfloat16
max(__hip_bfloat16 a, __hip_bfloat16 b) {
#if defined(__HIP_DEVICE_COMPILE__)
return __hmax(a, b);
#else
return __nv_bfloat16(max(float(a), float(b))); // NOLINT
return __hip_bfloat16(max(float(a), float(b))); // NOLINT
#endif
}
static __device__ __forceinline__ __nv_bfloat16
min(__nv_bfloat16 a, __nv_bfloat16 b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
static __device__ __forceinline__ __hip_bfloat16
min(__hip_bfloat16 a, __hip_bfloat16 b) {
#if defined(__HIP_DEVICE_COMPILE__)
return __hmin(a, b);
#else
return __nv_bfloat16(min(float(a), float(b))); // NOLINT
return __hip_bfloat16(min(float(a), float(b))); // NOLINT
#endif
}
#ifdef __CUDACC__
#if HIP_VERSION_MAJOR < 6
// Arithmetic BF16 operations for architecture >= 8.0 are already defined in
// cuda_bf16.h
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 800)
// CUDA 12.2 adds "emulated" support for older architectures.
#if defined(CUDART_VERSION) && (CUDART_VERSION < 12020)
__device__ __forceinline__ __nv_bfloat16
operator+(const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
return __nv_bfloat16(float(lh) + float(rh)); // NOLINT
}
__device__ __forceinline__ __nv_bfloat16
operator-(const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
return __nv_bfloat16(float(lh) - float(rh)); // NOLINT
}
__device__ __forceinline__ __nv_bfloat16
operator*(const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
return __nv_bfloat16(float(lh) * float(rh)); // NOLINT
}
__device__ __forceinline__ __nv_bfloat16
operator/(const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
return __nv_bfloat16(float(lh) / float(rh)); // NOLINT
// hip/__hip_bfloat16.h
// #if defined(__DTK_ARCH__) && (__DTK_ARCH__ < 800)
// // CUDA 12.2 adds "emulated" support for older architectures.
// #if defined(DTKRT_VERSION) && (DTKRT_VERSION < 12020)
__device__ __forceinline__ __hip_bfloat16
operator+(const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return __hip_bfloat16(float(lh) + float(rh)); // NOLINT
}
__device__ __forceinline__ __hip_bfloat16
operator-(const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return __hip_bfloat16(float(lh) - float(rh)); // NOLINT
}
__device__ __forceinline__ __hip_bfloat16
operator*(const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return __hip_bfloat16(float(lh) * float(rh)); // NOLINT
}
__device__ __forceinline__ __hip_bfloat16
operator/(const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return __hip_bfloat16(float(lh) / float(rh)); // NOLINT
}
__device__ __forceinline__ __nv_bfloat16& operator+=(
__nv_bfloat16& lh, const __nv_bfloat16& rh) { // NOLINT
lh = __nv_bfloat16(float(lh) + float(rh)); // NOLINT
__device__ __forceinline__ __hip_bfloat16& operator+=(
__hip_bfloat16& lh, const __hip_bfloat16& rh) { // NOLINT
lh = __hip_bfloat16(float(lh) + float(rh)); // NOLINT
return lh;
}
__device__ __forceinline__ __nv_bfloat16& operator-=(
__nv_bfloat16& lh, const __nv_bfloat16& rh) { // NOLINT
lh = __nv_bfloat16(float(lh) - float(rh)); // NOLINT
__device__ __forceinline__ __hip_bfloat16& operator-=(
__hip_bfloat16& lh, const __hip_bfloat16& rh) { // NOLINT
lh = __hip_bfloat16(float(lh) - float(rh)); // NOLINT
return lh;
}
__device__ __forceinline__ __nv_bfloat16& operator*=(
__nv_bfloat16& lh, const __nv_bfloat16& rh) { // NOLINT
lh = __nv_bfloat16(float(lh) * float(rh)); // NOLINT
__device__ __forceinline__ __hip_bfloat16& operator*=(
__hip_bfloat16& lh, const __hip_bfloat16& rh) { // NOLINT
lh = __hip_bfloat16(float(lh) * float(rh)); // NOLINT
return lh;
}
__device__ __forceinline__ __nv_bfloat16& operator/=(
__nv_bfloat16& lh, const __nv_bfloat16& rh) { // NOLINT
lh = __nv_bfloat16(float(lh) / float(rh)); // NOLINT
__device__ __forceinline__ __hip_bfloat16& operator/=(
__hip_bfloat16& lh, const __hip_bfloat16& rh) { // NOLINT
lh = __hip_bfloat16(float(lh) / float(rh)); // NOLINT
return lh;
}
__device__ __forceinline__ __nv_bfloat16& operator++(
__nv_bfloat16& h) { // NOLINT
h = __nv_bfloat16(float(h) + 1.0f); // NOLINT
__device__ __forceinline__ __hip_bfloat16& operator++(
__hip_bfloat16& h) { // NOLINT
h = __hip_bfloat16(float(h) + 1.0f); // NOLINT
return h;
}
__device__ __forceinline__ __nv_bfloat16& operator--(
__nv_bfloat16& h) { // NOLINT
h = __nv_bfloat16(float(h) - 1.0f); // NOLINT
__device__ __forceinline__ __hip_bfloat16& operator--(
__hip_bfloat16& h) { // NOLINT
h = __hip_bfloat16(float(h) - 1.0f); // NOLINT
return h;
}
__device__ __forceinline__ __nv_bfloat16
operator++(__nv_bfloat16& h, int) { // NOLINT
__nv_bfloat16 ret = h;
h = __nv_bfloat16(float(h) + 1.0f); // NOLINT
__device__ __forceinline__ __hip_bfloat16
operator++(__hip_bfloat16& h, int) { // NOLINT
__hip_bfloat16 ret = h;
h = __hip_bfloat16(float(h) + 1.0f); // NOLINT
return ret;
}
__device__ __forceinline__ __nv_bfloat16
operator--(__nv_bfloat16& h, int) { // NOLINT
__nv_bfloat16 ret = h;
h = __nv_bfloat16(float(h) - 1.0f); // NOLINT
__device__ __forceinline__ __hip_bfloat16
operator--(__hip_bfloat16& h, int) { // NOLINT
__hip_bfloat16 ret = h;
h = __hip_bfloat16(float(h) - 1.0f); // NOLINT
return ret;
}
__device__ __forceinline__ __nv_bfloat16 operator+(const __nv_bfloat16& h) {
__device__ __forceinline__ __hip_bfloat16 operator+(const __hip_bfloat16& h) {
return h;
}
__device__ __forceinline__ __nv_bfloat16 operator-(const __nv_bfloat16& h) {
return __nv_bfloat16(-float(h)); // NOLINT
__device__ __forceinline__ __hip_bfloat16 operator-(const __hip_bfloat16& h) {
return __hip_bfloat16(-float(h)); // NOLINT
}
__device__ __forceinline__ bool operator==(
const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return float(lh) == float(rh); // NOLINT
}
__device__ __forceinline__ bool operator!=(
const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return float(lh) != float(rh); // NOLINT
}
__device__ __forceinline__ bool operator>(
const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return float(lh) > float(rh); // NOLINT
}
__device__ __forceinline__ bool operator<(
const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return float(lh) < float(rh); // NOLINT
}
__device__ __forceinline__ bool operator>=(
const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return float(lh) >= float(rh); // NOLINT
}
__device__ __forceinline__ bool operator<=(
const __nv_bfloat16& lh, const __nv_bfloat16& rh) {
const __hip_bfloat16& lh, const __hip_bfloat16& rh) {
return float(lh) <= float(rh); // NOLINT
}
#endif // defined(CUDART_VERSION) && (CUDART_VERSION < 12020)
#endif // defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 800)
#endif // __CUDACC__
// #endif // defined(DTKRT_VERSION) && (DTKRT_VERSION < 12020)
// #endif // defined(__DTK_ARCH__) && (__DTK_ARCH__ < 800)
#endif
#if __HIPCC__
__device__
inline
__hip_bfloat16 __shfl_down(__hip_bfloat16 var, unsigned int lane_delta, int width = warpSize) {
union { unsigned short s; __hip_bfloat16 us; } tmp;
tmp.us = var;
tmp.s = __shfl_down(tmp.s, lane_delta, width);
return tmp.us;
}
#endif // __HIPCC__
#endif // BF16_ENABLED
......
src/array/cuda/coo2csr.cu → src/array/cuda/coo2csr.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020 by Contributors
* @file array/cuda/coo2csr.cc
* @brief COO2CSR
*/
#include <dgl/array.h>
#include "../../../include/dgl/array.h"
#include "../../runtime/cuda/cuda_common.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
......@@ -24,12 +28,12 @@ CSRMatrix COOToCSR(COOMatrix coo) {
template <>
CSRMatrix COOToCSR<kDGLCUDA, int32_t>(COOMatrix coo) {
auto* thr_entry = runtime::CUDAThreadEntry::ThreadLocal();
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
// allocate cusparse handle if needed
if (!thr_entry->cusparse_handle) {
CUSPARSE_CALL(cusparseCreate(&(thr_entry->cusparse_handle)));
CUSPARSE_CALL(hipsparseCreate(&(thr_entry->cusparse_handle)));
}
CUSPARSE_CALL(cusparseSetStream(thr_entry->cusparse_handle, stream));
CUSPARSE_CALL(hipsparseSetStream(thr_entry->cusparse_handle, stream));
bool row_sorted = coo.row_sorted;
bool col_sorted = coo.col_sorted;
......@@ -50,9 +54,9 @@ CSRMatrix COOToCSR<kDGLCUDA, int32_t>(COOMatrix coo) {
NDArray indptr =
aten::NewIdArray(coo.num_rows + 1, coo.row->ctx, coo.row->dtype.bits);
int32_t* indptr_ptr = static_cast<int32_t*>(indptr->data);
CUSPARSE_CALL(cusparseXcoo2csr(
CUSPARSE_CALL(hipsparseXcoo2csr(
thr_entry->cusparse_handle, coo.row.Ptr<int32_t>(), nnz, coo.num_rows,
indptr_ptr, CUSPARSE_INDEX_BASE_ZERO));
indptr_ptr, HIPSPARSE_INDEX_BASE_ZERO));
return CSRMatrix(
coo.num_rows, coo.num_cols, indptr, coo.col, coo.data, col_sorted);
......@@ -100,7 +104,7 @@ template <>
CSRMatrix COOToCSR<kDGLCUDA, int64_t>(COOMatrix coo) {
const auto& ctx = coo.row->ctx;
const auto nbits = coo.row->dtype.bits;
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
bool row_sorted = coo.row_sorted;
bool col_sorted = coo.col_sorted;
if (!row_sorted) {
......
src/array/cuda/coo_sort.cu → src/array/cuda/coo_sort.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020 by Contributors
* @file array/cuda/coo_sort.cc
* @brief Sort COO index
*/
#include <dgl/array.h>
#include "../../../include/dgl/array.h"
#include "../../c_api_common.h"
#include "../../runtime/cuda/cuda_common.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
......@@ -65,7 +69,7 @@ __global__ void _COODecodeEdgesKernel(
template <DGLDeviceType XPU, typename IdType>
void COOSort_(COOMatrix* coo, bool sort_column) {
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
const int row_bits = cuda::_NumberOfBits(coo->num_rows);
const int64_t nnz = coo->row->shape[0];
......@@ -138,7 +142,7 @@ template <DGLDeviceType XPU, typename IdType>
std::pair<bool, bool> COOIsSorted(COOMatrix coo) {
const int64_t nnz = coo.row->shape[0];
const auto& ctx = coo.row->ctx;
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
auto device = runtime::DeviceAPI::Get(ctx);
// We allocate a workspace of 2*nnz bytes. It wastes a little bit memory but
// should be fine.
......
src/array/cuda/csr2coo.cu → src/array/cuda/csr2coo.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020 by Contributors
* @file array/cuda/csr2coo.cc
......@@ -7,11 +9,12 @@
#include <thrust/iterator/constant_iterator.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/iterator/transform_iterator.h>
#include <hipcub/backend/rocprim/device/device_copy.hpp>
#include <cub/cub.cuh>
#include <hipcub/hipcub.hpp>
#include "../../runtime/cuda/cuda_common.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
......@@ -29,12 +32,12 @@ COOMatrix CSRToCOO(CSRMatrix csr) {
template <>
COOMatrix CSRToCOO<kDGLCUDA, int32_t>(CSRMatrix csr) {
auto* thr_entry = runtime::CUDAThreadEntry::ThreadLocal();
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
// allocate cusparse handle if needed
if (!thr_entry->cusparse_handle) {
CUSPARSE_CALL(cusparseCreate(&(thr_entry->cusparse_handle)));
CUSPARSE_CALL(hipsparseCreate(&(thr_entry->cusparse_handle)));
}
CUSPARSE_CALL(cusparseSetStream(thr_entry->cusparse_handle, stream));
CUSPARSE_CALL(hipsparseSetStream(thr_entry->cusparse_handle, stream));
NDArray indptr = csr.indptr, indices = csr.indices, data = csr.data;
const int32_t* indptr_ptr = static_cast<int32_t*>(indptr->data);
......@@ -42,9 +45,9 @@ COOMatrix CSRToCOO<kDGLCUDA, int32_t>(CSRMatrix csr) {
aten::NewIdArray(indices->shape[0], indptr->ctx, indptr->dtype.bits);
int32_t* row_ptr = static_cast<int32_t*>(row->data);
CUSPARSE_CALL(cusparseXcsr2coo(
CUSPARSE_CALL(hipsparseXcsr2coo(
thr_entry->cusparse_handle, indptr_ptr, indices->shape[0], csr.num_rows,
row_ptr, CUSPARSE_INDEX_BASE_ZERO));
row_ptr, HIPSPARSE_INDEX_BASE_ZERO));
return COOMatrix(
csr.num_rows, csr.num_cols, row, indices, data, true, csr.sorted);
......@@ -72,10 +75,40 @@ struct AdjacentDifference {
}
};
/*!
* \brief Repeat elements
* \param val Value to repeat
* \param repeats Number of repeats for each value
* \param pos The position of the output buffer to write the value.
* \param out Output buffer.
* \param length Number of values
*
* For example:
* val = [3, 0, 1]
* repeats = [1, 0, 2]
* pos = [0, 1, 1] # write to output buffer position 0, 1, 1
* then,
* out = [3, 1, 1]
*/
template <typename DType, typename IdType>
__global__ void _RepeatKernel(
const DType* val, const IdType* pos,
DType* out, int64_t n_row, int64_t length) {
IdType tx = static_cast<IdType>(blockIdx.x) * blockDim.x + threadIdx.x;
const int stride_x = gridDim.x * blockDim.x;
while (tx < length) {
IdType i = dgl::cuda::_UpperBound(pos, n_row, tx) - 1;
out[tx] = val[i];
tx += stride_x;
}
}
#if 1
template <>
COOMatrix CSRToCOO<kDGLCUDA, int64_t>(CSRMatrix csr) {
const auto& ctx = csr.indptr->ctx;
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
const int64_t nnz = csr.indices->shape[0];
const auto nbits = csr.indptr->dtype.bits;
......@@ -94,16 +127,16 @@ COOMatrix CSRToCOO<kDGLCUDA, int64_t>(CSRMatrix csr) {
constexpr int64_t max_copy_at_once = std::numeric_limits<int32_t>::max();
for (int64_t i = 0; i < csr.num_rows; i += max_copy_at_once) {
std::size_t temp_storage_bytes = 0;
CUDA_CALL(cub::DeviceCopy::Batched(
CUDA_CALL(hipcub::DeviceCopy::Batched(
nullptr, temp_storage_bytes, input_buffer + i, output_buffer + i,
buffer_sizes + i, std::min(csr.num_rows - i, max_copy_at_once),
buffer_sizes + i, ::min(csr.num_rows - i, max_copy_at_once),
stream));
auto temp = allocator.alloc_unique<char>(temp_storage_bytes);
CUDA_CALL(cub::DeviceCopy::Batched(
CUDA_CALL(hipcub::DeviceCopy::Batched(
temp.get(), temp_storage_bytes, input_buffer + i, output_buffer + i,
buffer_sizes + i, std::min(csr.num_rows - i, max_copy_at_once),
buffer_sizes + i, ::min(csr.num_rows - i, max_copy_at_once),
stream));
}
......@@ -111,6 +144,30 @@ COOMatrix CSRToCOO<kDGLCUDA, int64_t>(CSRMatrix csr) {
csr.num_rows, csr.num_cols, ret_row, csr.indices, csr.data, true,
csr.sorted);
}
#else
template <>
COOMatrix CSRToCOO<kDGLCUDA, int64_t>(CSRMatrix csr) {
const auto& ctx = csr.indptr->ctx;
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
const int64_t nnz = csr.indices->shape[0];
const auto nbits = csr.indptr->dtype.bits;
IdArray rowids = Range(0, csr.num_rows, nbits, ctx);
IdArray ret_row = NewIdArray(nnz, ctx, nbits);
const int nt = 256;
const int nb = (nnz + nt - 1) / nt;
CUDA_KERNEL_CALL(_RepeatKernel,
nb, nt, 0, stream,
rowids.Ptr<int64_t>(),
csr.indptr.Ptr<int64_t>(), ret_row.Ptr<int64_t>(),
csr.num_rows, nnz);
return COOMatrix(csr.num_rows, csr.num_cols,
ret_row, csr.indices, csr.data,
true, csr.sorted);
}
#endif
template COOMatrix CSRToCOO<kDGLCUDA, int32_t>(CSRMatrix csr);
template COOMatrix CSRToCOO<kDGLCUDA, int64_t>(CSRMatrix csr);
......@@ -128,12 +185,12 @@ COOMatrix CSRToCOODataAsOrder<kDGLCUDA, int32_t>(CSRMatrix csr) {
auto* thr_entry = runtime::CUDAThreadEntry::ThreadLocal();
auto device = runtime::DeviceAPI::Get(coo.row->ctx);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
// allocate cusparse handle if needed
if (!thr_entry->cusparse_handle) {
CUSPARSE_CALL(cusparseCreate(&(thr_entry->cusparse_handle)));
CUSPARSE_CALL(hipsparseCreate(&(thr_entry->cusparse_handle)));
}
CUSPARSE_CALL(cusparseSetStream(thr_entry->cusparse_handle, stream));
CUSPARSE_CALL(hipsparseSetStream(thr_entry->cusparse_handle, stream));
NDArray row = coo.row, col = coo.col, data = coo.data;
int32_t* row_ptr = static_cast<int32_t*>(row->data);
......@@ -141,11 +198,11 @@ COOMatrix CSRToCOODataAsOrder<kDGLCUDA, int32_t>(CSRMatrix csr) {
int32_t* data_ptr = static_cast<int32_t*>(data->data);
size_t workspace_size = 0;
CUSPARSE_CALL(cusparseXcoosort_bufferSizeExt(
CUSPARSE_CALL(hipsparseXcoosort_bufferSizeExt(
thr_entry->cusparse_handle, coo.num_rows, coo.num_cols, row->shape[0],
data_ptr, row_ptr, &workspace_size));
void* workspace = device->AllocWorkspace(row->ctx, workspace_size);
CUSPARSE_CALL(cusparseXcoosortByRow(
CUSPARSE_CALL(hipsparseXcoosortByRow(
thr_entry->cusparse_handle, coo.num_rows, coo.num_cols, row->shape[0],
data_ptr, row_ptr, col_ptr, workspace));
device->FreeWorkspace(row->ctx, workspace);
......
src/array/cuda/csr_get_data.cu → src/array/cuda/csr_get_data.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2021 by Contributors
* @file array/cuda/csr_get_data.cu
* @brief Retrieve entries of a CSR matrix
*/
#include <dgl/array.h>
#include "../../../include/dgl/array.h"
#include <numeric>
#include <unordered_set>
#include <vector>
#include "../../runtime/cuda/cuda_common.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
......@@ -32,11 +36,11 @@ NDArray CSRGetData(
const int64_t row_stride = (rowlen == 1 && collen != 1) ? 0 : 1;
const int64_t col_stride = (collen == 1 && rowlen != 1) ? 0 : 1;
const int64_t rstlen = std::max(rowlen, collen);
const int64_t rstlen = ::max(rowlen, collen);
IdArray rst = NDArray::Empty({rstlen}, weights->dtype, rows->ctx);
if (rstlen == 0) return rst;
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
const int nt = cuda::FindNumThreads(rstlen);
const int nb = (rstlen + nt - 1) / nt;
if (return_eids)
......@@ -67,12 +71,12 @@ template NDArray CSRGetData<kDGLCUDA, int64_t, __half>(
CSRMatrix csr, NDArray rows, NDArray cols, bool return_eids,
NDArray weights, __half filler);
#if BF16_ENABLED
template NDArray CSRGetData<kDGLCUDA, int32_t, __nv_bfloat16>(
template NDArray CSRGetData<kDGLCUDA, int32_t, __hip_bfloat16>(
CSRMatrix csr, NDArray rows, NDArray cols, bool return_eids,
NDArray weights, __nv_bfloat16 filler);
template NDArray CSRGetData<kDGLCUDA, int64_t, __nv_bfloat16>(
NDArray weights, __hip_bfloat16 filler);
template NDArray CSRGetData<kDGLCUDA, int64_t, __hip_bfloat16>(
CSRMatrix csr, NDArray rows, NDArray cols, bool return_eids,
NDArray weights, __nv_bfloat16 filler);
NDArray weights, __hip_bfloat16 filler);
#endif // BF16_ENABLED
template NDArray CSRGetData<kDGLCUDA, int32_t, float>(
CSRMatrix csr, NDArray rows, NDArray cols, bool return_eids,
......
src/array/cuda/csr_mm.cu → src/array/cuda/csr_mm.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020 by Contributors
* @file array/cuda/csr_mm.cu
* @brief SpSpMM/SpGEMM C APIs and definitions.
*/
#include <dgl/array.h>
#include "../../../include/dgl/array.h"
#include <dgl/runtime/device_api.h>
#include <limits>
#include "../../runtime/cuda/cuda_common.h"
#include "./cusparse_dispatcher.cuh"
#include "./functor.cuh"
#include "cusparse_dispatcher.cuh"
#include "functor.cuh"
namespace dgl {
using namespace dgl::runtime;
......@@ -18,7 +22,7 @@ using namespace dgl::runtime;
namespace aten {
namespace cusparse {
#if CUDART_VERSION >= 12000
#if DTKRT_VERSION >= 12000
/** @brief Cusparse implementation of SpGEMM on Csr format for CUDA 12.0+ */
template <typename DType, typename IdType>
......@@ -31,74 +35,74 @@ std::pair<CSRMatrix, NDArray> CusparseSpgemm(
const int nnzB = B.indices->shape[0];
const DType alpha = 1.0;
const DType beta = 0.0;
auto transA = CUSPARSE_OPERATION_NON_TRANSPOSE;
auto transB = CUSPARSE_OPERATION_NON_TRANSPOSE;
auto transA = HIPSPARSE_OPERATION_NON_TRANSPOSE;
auto transB = HIPSPARSE_OPERATION_NON_TRANSPOSE;
// device
auto ctx = A.indptr->ctx;
auto device = runtime::DeviceAPI::Get(ctx);
auto* thr_entry = runtime::CUDAThreadEntry::ThreadLocal();
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
const DType* A_weights = A_weights_array.Ptr<DType>();
const DType* B_weights = B_weights_array.Ptr<DType>();
// allocate cusparse handle if needed
if (!thr_entry->cusparse_handle) {
CUSPARSE_CALL(cusparseCreate(&(thr_entry->cusparse_handle)));
CUSPARSE_CALL(hipsparseCreate(&(thr_entry->cusparse_handle)));
}
CUSPARSE_CALL(cusparseSetStream(thr_entry->cusparse_handle, stream));
CUSPARSE_CALL(hipsparseSetStream(thr_entry->cusparse_handle, stream));
// all one data array
cusparseSpMatDescr_t matA, matB, matC;
hipsparseSpMatDescr_t matA, matB, matC;
IdArray dC_csrOffsets =
IdArray::Empty({A.num_rows + 1}, A.indptr->dtype, A.indptr->ctx);
IdType* dC_csrOffsets_data = dC_csrOffsets.Ptr<IdType>();
constexpr auto idtype = cusparse_idtype<IdType>::value;
constexpr auto dtype = cuda_dtype<DType>::value;
// Create sparse matrix A, B and C in CSR format
CUSPARSE_CALL(cusparseCreateCsr(
CUSPARSE_CALL(hipsparseCreateCsr(
&matA, A.num_rows, A.num_cols, nnzA, A.indptr.Ptr<IdType>(),
A.indices.Ptr<IdType>(),
// cusparseCreateCsr only accepts non-const pointers.
const_cast<DType*>(A_weights), idtype, idtype, CUSPARSE_INDEX_BASE_ZERO,
// hipsparseCreateCsr only accepts non-const pointers.
const_cast<DType*>(A_weights), idtype, idtype, HIPSPARSE_INDEX_BASE_ZERO,
dtype));
CUSPARSE_CALL(cusparseCreateCsr(
CUSPARSE_CALL(hipsparseCreateCsr(
&matB, B.num_rows, B.num_cols, nnzB, B.indptr.Ptr<IdType>(),
B.indices.Ptr<IdType>(),
// cusparseCreateCsr only accepts non-const pointers.
const_cast<DType*>(B_weights), idtype, idtype, CUSPARSE_INDEX_BASE_ZERO,
// hipsparseCreateCsr only accepts non-const pointers.
const_cast<DType*>(B_weights), idtype, idtype, HIPSPARSE_INDEX_BASE_ZERO,
dtype));
CUSPARSE_CALL(cusparseCreateCsr(
CUSPARSE_CALL(hipsparseCreateCsr(
&matC, A.num_rows, B.num_cols, 0, dC_csrOffsets_data, nullptr, nullptr,
idtype, idtype, CUSPARSE_INDEX_BASE_ZERO, dtype));
idtype, idtype, HIPSPARSE_INDEX_BASE_ZERO, dtype));
// SpGEMM Computation
cusparseSpGEMMDescr_t spgemmDesc;
cusparseSpGEMMAlg_t alg = CUSPARSE_SPGEMM_DEFAULT;
hipsparseSpGEMMDescr_t spgemmDesc;
cusparseSpGEMMAlg_t alg = HIPSPARSE_SPGEMM_DEFAULT;
CUSPARSE_CALL(cusparseSpGEMM_createDescr(&spgemmDesc));
CUSPARSE_CALL(hipsparseSpGEMM_createDescr(&spgemmDesc));
size_t workspace_size1 = 0, workspace_size2 = 0, workspace_size3 = 0;
// ask bufferSize1 bytes for external memory
CUSPARSE_CALL(cusparseSpGEMM_workEstimation(
CUSPARSE_CALL(hipsparseSpGEMM_workEstimation(
thr_entry->cusparse_handle, transA, transB, &alpha, matA, matB, &beta,
matC, dtype, alg, spgemmDesc, &workspace_size1, NULL));
void* workspace1 = (device->AllocWorkspace(ctx, workspace_size1));
// inspect the matrices A and B to understand the memory requiremnent
cusparseStatus_t e = cusparseSpGEMM_workEstimation(
hipsparseStatus_t e = hipsparseSpGEMM_workEstimation(
thr_entry->cusparse_handle, transA, transB, &alpha, matA, matB, &beta,
matC, dtype, alg, spgemmDesc, &workspace_size1, workspace1);
// CUSPARSE_SPGEMM_DEFAULT not support getting num_prods > 2^31 -1
// HIPSPARSE_SPGEMM_DEFAULT not support getting num_prods > 2^31 -1
// and throws insufficient memory error within workEstimation call
if (e == CUSPARSE_STATUS_INSUFFICIENT_RESOURCES) {
// fall back to ALG2 to estimate num_prods
alg = CUSPARSE_SPGEMM_ALG2;
device->FreeWorkspace(ctx, workspace1);
// rerun cusparseSpGEMM_workEstimation
CUSPARSE_CALL(cusparseSpGEMM_workEstimation(
// rerun hipsparseSpGEMM_workEstimation
CUSPARSE_CALL(hipsparseSpGEMM_workEstimation(
thr_entry->cusparse_handle, transA, transB, &alpha, matA, matB, &beta,
matC, dtype, alg, spgemmDesc, &workspace_size1, NULL));
workspace1 = (device->AllocWorkspace(ctx, workspace_size1));
CUSPARSE_CALL(cusparseSpGEMM_workEstimation(
CUSPARSE_CALL(hipsparseSpGEMM_workEstimation(
thr_entry->cusparse_handle, transA, transB, &alpha, matA, matB, &beta,
matC, dtype, alg, spgemmDesc, &workspace_size1, workspace1));
} else {
CHECK(e == CUSPARSE_STATUS_SUCCESS) << "CUSPARSE ERROR in SpGEMM: " << e;
CHECK(e == HIPSPARSE_STATUS_SUCCESS) << "CUSPARSE ERROR in SpGEMM: " << e;
}
// get the number of intermediate products required for SpGEMM compute
......@@ -113,22 +117,22 @@ std::pair<CSRMatrix, NDArray> CusparseSpgemm(
int64_t LARGE_NUM_PRODUCTS = 800000000; // 800*1000*1000;
// switch to ALG2/ALG3 for medium & large problem size
if (alg == CUSPARSE_SPGEMM_DEFAULT && num_prods > MEDIUM_NUM_PRODUCTS) {
if (alg == HIPSPARSE_SPGEMM_DEFAULT && num_prods > MEDIUM_NUM_PRODUCTS) {
// use ALG3 for very large problem
alg = num_prods > LARGE_NUM_PRODUCTS ? CUSPARSE_SPGEMM_ALG3
: CUSPARSE_SPGEMM_ALG2;
device->FreeWorkspace(ctx, workspace1);
// rerun cusparseSpGEMM_workEstimation
CUSPARSE_CALL(cusparseSpGEMM_workEstimation(
// rerun hipsparseSpGEMM_workEstimation
CUSPARSE_CALL(hipsparseSpGEMM_workEstimation(
thr_entry->cusparse_handle, transA, transB, &alpha, matA, matB, &beta,
matC, dtype, alg, spgemmDesc, &workspace_size1, NULL));
workspace1 = (device->AllocWorkspace(ctx, workspace_size1));
CUSPARSE_CALL(cusparseSpGEMM_workEstimation(
CUSPARSE_CALL(hipsparseSpGEMM_workEstimation(
thr_entry->cusparse_handle, transA, transB, &alpha, matA, matB, &beta,
matC, dtype, alg, spgemmDesc, &workspace_size1, workspace1));
} else if (alg == CUSPARSE_SPGEMM_ALG2 && num_prods > LARGE_NUM_PRODUCTS) {
// no need to rerun cusparseSpGEMM_workEstimation between ALG2 and ALG3
// no need to rerun hipsparseSpGEMM_workEstimation between ALG2 and ALG3
alg = CUSPARSE_SPGEMM_ALG3;
}
......@@ -147,40 +151,40 @@ std::pair<CSRMatrix, NDArray> CusparseSpgemm(
workspace3, &workspace_size2));
device->FreeWorkspace(ctx, workspace3);
} else {
CUSPARSE_CALL(cusparseSpGEMM_compute(
CUSPARSE_CALL(hipsparseSpGEMM_compute(
thr_entry->cusparse_handle, transA, transB, &alpha, matA, matB, &beta,
matC, dtype, alg, spgemmDesc, &workspace_size2, NULL));
}
// ask bufferSize2 bytes for external memory
void* workspace2 = device->AllocWorkspace(ctx, workspace_size2);
// compute the intermediate product of A * B
CUSPARSE_CALL(cusparseSpGEMM_compute(
CUSPARSE_CALL(hipsparseSpGEMM_compute(
thr_entry->cusparse_handle, transA, transB, &alpha, matA, matB, &beta,
matC, dtype, alg, spgemmDesc, &workspace_size2, workspace2));
// get matrix C non-zero entries C_nnz1
int64_t C_num_rows1, C_num_cols1, C_nnz1;
CUSPARSE_CALL(
cusparseSpMatGetSize(matC, &C_num_rows1, &C_num_cols1, &C_nnz1));
hipsparseSpMatGetSize(matC, &C_num_rows1, &C_num_cols1, &C_nnz1));
IdArray dC_columns = IdArray::Empty({C_nnz1}, A.indptr->dtype, A.indptr->ctx);
NDArray dC_weights =
NDArray::Empty({C_nnz1}, A_weights_array->dtype, A.indptr->ctx);
IdType* dC_columns_data = dC_columns.Ptr<IdType>();
DType* dC_weights_data = dC_weights.Ptr<DType>();
// update matC with the new pointers
CUSPARSE_CALL(cusparseCsrSetPointers(
CUSPARSE_CALL(hipsparseCsrSetPointers(
matC, dC_csrOffsets_data, dC_columns_data, dC_weights_data));
// copy the final products to the matrix C
CUSPARSE_CALL(cusparseSpGEMM_copy(
CUSPARSE_CALL(hipsparseSpGEMM_copy(
thr_entry->cusparse_handle, transA, transB, &alpha, matA, matB, &beta,
matC, dtype, alg, spgemmDesc));
device->FreeWorkspace(ctx, workspace1);
device->FreeWorkspace(ctx, workspace2);
// destroy matrix/vector descriptors
CUSPARSE_CALL(cusparseSpGEMM_destroyDescr(spgemmDesc));
CUSPARSE_CALL(cusparseDestroySpMat(matA));
CUSPARSE_CALL(cusparseDestroySpMat(matB));
CUSPARSE_CALL(cusparseDestroySpMat(matC));
CUSPARSE_CALL(hipsparseSpGEMM_destroyDescr(spgemmDesc));
CUSPARSE_CALL(hipsparseDestroySpMat(matA));
CUSPARSE_CALL(hipsparseDestroySpMat(matB));
CUSPARSE_CALL(hipsparseDestroySpMat(matC));
return {
CSRMatrix(
A.num_rows, B.num_cols, dC_csrOffsets, dC_columns,
......@@ -188,7 +192,7 @@ std::pair<CSRMatrix, NDArray> CusparseSpgemm(
dC_weights};
}
#else // CUDART_VERSION < 12000
#else // DTKRT_VERSION < 12000
/** @brief Cusparse implementation of SpGEMM on Csr format for older CUDA
* versions */
......@@ -208,25 +212,25 @@ std::pair<CSRMatrix, NDArray> CusparseSpgemm(
auto ctx = A.indptr->ctx;
auto device = runtime::DeviceAPI::Get(ctx);
auto* thr_entry = runtime::CUDAThreadEntry::ThreadLocal();
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
auto idtype = A.indptr->dtype;
auto dtype = A_weights_array->dtype;
const DType* A_weights = A_weights_array.Ptr<DType>();
const DType* B_weights = B_weights_array.Ptr<DType>();
if (!thr_entry->cusparse_handle) {
CUSPARSE_CALL(cusparseCreate(&(thr_entry->cusparse_handle)));
CUSPARSE_CALL(hipsparseCreate(&(thr_entry->cusparse_handle)));
}
CUSPARSE_CALL(cusparseSetStream(thr_entry->cusparse_handle, stream));
CUSPARSE_CALL(cusparseSetPointerMode(
thr_entry->cusparse_handle, CUSPARSE_POINTER_MODE_HOST));
CUSPARSE_CALL(hipsparseSetStream(thr_entry->cusparse_handle, stream));
CUSPARSE_CALL(hipsparseSetPointerMode(
thr_entry->cusparse_handle, HIPSPARSE_POINTER_MODE_HOST));
CUSPARSE_CALL(cusparseCreateCsrgemm2Info(&info));
CUSPARSE_CALL(hipsparseCreateCsrgemm2Info(&info));
cusparseMatDescr_t matA, matB, matC, matD;
CUSPARSE_CALL(cusparseCreateMatDescr(&matA));
CUSPARSE_CALL(cusparseCreateMatDescr(&matB));
CUSPARSE_CALL(cusparseCreateMatDescr(&matC));
CUSPARSE_CALL(cusparseCreateMatDescr(&matD)); // needed even if D is null
hipsparseMatDescr_t matA, matB, matC, matD;
CUSPARSE_CALL(hipsparseCreateMatDescr(&matA));
CUSPARSE_CALL(hipsparseCreateMatDescr(&matB));
CUSPARSE_CALL(hipsparseCreateMatDescr(&matC));
CUSPARSE_CALL(hipsparseCreateMatDescr(&matD)); // needed even if D is null
CUSPARSE_CALL(CSRGEMM<DType>::bufferSizeExt(
thr_entry->cusparse_handle, m, n, k, &alpha, matA, nnzA,
......@@ -252,11 +256,11 @@ std::pair<CSRMatrix, NDArray> CusparseSpgemm(
C_indptr.Ptr<IdType>(), C_indices.Ptr<IdType>(), info, workspace));
device->FreeWorkspace(ctx, workspace);
CUSPARSE_CALL(cusparseDestroyCsrgemm2Info(info));
CUSPARSE_CALL(cusparseDestroyMatDescr(matA));
CUSPARSE_CALL(cusparseDestroyMatDescr(matB));
CUSPARSE_CALL(cusparseDestroyMatDescr(matC));
CUSPARSE_CALL(cusparseDestroyMatDescr(matD));
CUSPARSE_CALL(hipsparseDestroyCsrgemm2Info(info));
CUSPARSE_CALL(hipsparseDestroyMatDescr(matA));
CUSPARSE_CALL(hipsparseDestroyMatDescr(matB));
CUSPARSE_CALL(hipsparseDestroyMatDescr(matC));
CUSPARSE_CALL(hipsparseDestroyMatDescr(matD));
return {
CSRMatrix(
......@@ -264,7 +268,7 @@ std::pair<CSRMatrix, NDArray> CusparseSpgemm(
C_weights};
}
#endif // CUDART_VERSION >= 12000
#endif // DTKRT_VERSION >= 12000
} // namespace cusparse
template <int XPU, typename IdType, typename DType>
......@@ -314,9 +318,9 @@ template std::pair<CSRMatrix, NDArray> CSRMM<kDGLCUDA, int32_t, __half>(
template std::pair<CSRMatrix, NDArray> CSRMM<kDGLCUDA, int64_t, __half>(
const CSRMatrix&, NDArray, const CSRMatrix&, NDArray);
#if BF16_ENABLED
template std::pair<CSRMatrix, NDArray> CSRMM<kDGLCUDA, int32_t, __nv_bfloat16>(
template std::pair<CSRMatrix, NDArray> CSRMM<kDGLCUDA, int32_t, __hip_bfloat16>(
const CSRMatrix&, NDArray, const CSRMatrix&, NDArray);
template std::pair<CSRMatrix, NDArray> CSRMM<kDGLCUDA, int64_t, __nv_bfloat16>(
template std::pair<CSRMatrix, NDArray> CSRMM<kDGLCUDA, int64_t, __hip_bfloat16>(
const CSRMatrix&, NDArray, const CSRMatrix&, NDArray);
#endif // BF16_ENABLED
template std::pair<CSRMatrix, NDArray> CSRMM<kDGLCUDA, int32_t, float>(
......
src/array/cuda/csr_sort.cu → src/array/cuda/csr_sort.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020 by Contributors
* @file array/cuda/csr_sort.cc
......@@ -5,10 +7,10 @@
*/
#include <dgl/array.h>
#include <cub/cub.cuh>
#include <hipcub/hipcub.hpp>
#include "../../runtime/cuda/cuda_common.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
......@@ -39,7 +41,7 @@ __global__ void _SegmentIsSorted(
template <DGLDeviceType XPU, typename IdType>
bool CSRIsSorted(CSRMatrix csr) {
const auto& ctx = csr.indptr->ctx;
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
auto device = runtime::DeviceAPI::Get(ctx);
// We allocate a workspace of num_rows bytes. It wastes a little bit memory
// but should be fine.
......@@ -67,12 +69,12 @@ template <>
void CSRSort_<kDGLCUDA, int32_t>(CSRMatrix* csr) {
auto* thr_entry = runtime::CUDAThreadEntry::ThreadLocal();
auto device = runtime::DeviceAPI::Get(csr->indptr->ctx);
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
// allocate cusparse handle if needed
if (!thr_entry->cusparse_handle) {
CUSPARSE_CALL(cusparseCreate(&(thr_entry->cusparse_handle)));
CUSPARSE_CALL(hipsparseCreate(&(thr_entry->cusparse_handle)));
}
CUSPARSE_CALL(cusparseSetStream(thr_entry->cusparse_handle, stream));
CUSPARSE_CALL(hipsparseSetStream(thr_entry->cusparse_handle, stream));
NDArray indptr = csr->indptr;
NDArray indices = csr->indices;
......@@ -83,16 +85,16 @@ void CSRSort_<kDGLCUDA, int32_t>(CSRMatrix* csr) {
NDArray data = csr->data;
size_t workspace_size = 0;
CUSPARSE_CALL(cusparseXcsrsort_bufferSizeExt(
CUSPARSE_CALL(hipsparseXcsrsort_bufferSizeExt(
thr_entry->cusparse_handle, csr->num_rows, csr->num_cols, nnz,
indptr.Ptr<int32_t>(), indices.Ptr<int32_t>(), &workspace_size));
void* workspace = device->AllocWorkspace(ctx, workspace_size);
cusparseMatDescr_t descr;
CUSPARSE_CALL(cusparseCreateMatDescr(&descr));
CUSPARSE_CALL(cusparseSetMatType(descr, CUSPARSE_MATRIX_TYPE_GENERAL));
CUSPARSE_CALL(cusparseSetMatIndexBase(descr, CUSPARSE_INDEX_BASE_ZERO));
CUSPARSE_CALL(cusparseXcsrsort(
hipsparseMatDescr_t descr;
CUSPARSE_CALL(hipsparseCreateMatDescr(&descr));
CUSPARSE_CALL(hipsparseSetMatType(descr, HIPSPARSE_MATRIX_TYPE_GENERAL));
CUSPARSE_CALL(hipsparseSetMatIndexBase(descr, HIPSPARSE_INDEX_BASE_ZERO));
CUSPARSE_CALL(hipsparseXcsrsort(
thr_entry->cusparse_handle, csr->num_rows, csr->num_cols, nnz, descr,
indptr.Ptr<int32_t>(), indices.Ptr<int32_t>(), data.Ptr<int32_t>(),
workspace));
......@@ -100,13 +102,13 @@ void CSRSort_<kDGLCUDA, int32_t>(CSRMatrix* csr) {
csr->sorted = true;
// free resources
CUSPARSE_CALL(cusparseDestroyMatDescr(descr));
CUSPARSE_CALL(hipsparseDestroyMatDescr(descr));
device->FreeWorkspace(ctx, workspace);
}
template <>
void CSRSort_<kDGLCUDA, int64_t>(CSRMatrix* csr) {
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
auto device = runtime::DeviceAPI::Get(csr->indptr->ctx);
const auto& ctx = csr->indptr->ctx;
......@@ -125,13 +127,13 @@ void CSRSort_<kDGLCUDA, int64_t>(CSRMatrix* csr) {
// Allocate workspace
size_t workspace_size = 0;
CUDA_CALL(cub::DeviceSegmentedRadixSort::SortPairs(
CUDA_CALL(hipcub::DeviceSegmentedRadixSort::SortPairs(
nullptr, workspace_size, key_in, key_out, value_in, value_out, nnz,
csr->num_rows, offsets, offsets + 1, 0, sizeof(int64_t) * 8, stream));
void* workspace = device->AllocWorkspace(ctx, workspace_size);
// Compute
CUDA_CALL(cub::DeviceSegmentedRadixSort::SortPairs(
CUDA_CALL(hipcub::DeviceSegmentedRadixSort::SortPairs(
workspace, workspace_size, key_in, key_out, value_in, value_out, nnz,
csr->num_rows, offsets, offsets + 1, 0, sizeof(int64_t) * 8, stream));
......
src/array/cuda/csr_sum.cu → src/array/cuda/csr_sum.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020 by Contributors
* @file array/cuda/spmm.cu
......@@ -7,8 +9,8 @@
#include <dgl/runtime/device_api.h>
#include "../../runtime/cuda/cuda_common.h"
#include "./cusparse_dispatcher.cuh"
#include "./functor.cuh"
#include "cusparse_dispatcher.cuh"
#include "functor.cuh"
namespace dgl {
......@@ -32,21 +34,21 @@ std::pair<CSRMatrix, NDArray> CusparseCsrgeam2(
auto ctx = A.indptr->ctx;
auto device = runtime::DeviceAPI::Get(ctx);
auto* thr_entry = runtime::CUDAThreadEntry::ThreadLocal();
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
const DType* A_weights = A_weights_array.Ptr<DType>();
const DType* B_weights = B_weights_array.Ptr<DType>();
// allocate cusparse handle if needed
if (!thr_entry->cusparse_handle)
CUSPARSE_CALL(cusparseCreate(&(thr_entry->cusparse_handle)));
CUSPARSE_CALL(cusparseSetStream(thr_entry->cusparse_handle, stream));
CUSPARSE_CALL(hipsparseCreate(&(thr_entry->cusparse_handle)));
CUSPARSE_CALL(hipsparseSetStream(thr_entry->cusparse_handle, stream));
cusparseMatDescr_t matA, matB, matC;
CUSPARSE_CALL(cusparseCreateMatDescr(&matA));
CUSPARSE_CALL(cusparseCreateMatDescr(&matB));
CUSPARSE_CALL(cusparseCreateMatDescr(&matC));
hipsparseMatDescr_t matA, matB, matC;
CUSPARSE_CALL(hipsparseCreateMatDescr(&matA));
CUSPARSE_CALL(hipsparseCreateMatDescr(&matB));
CUSPARSE_CALL(hipsparseCreateMatDescr(&matC));
cusparseSetPointerMode(
thr_entry->cusparse_handle, CUSPARSE_POINTER_MODE_HOST);
hipsparseSetPointerMode(
thr_entry->cusparse_handle, HIPSPARSE_POINTER_MODE_HOST);
size_t workspace_size = 0;
/* prepare output C */
IdArray dC_csrOffsets = IdArray::Empty({m + 1}, A.indptr->dtype, ctx);
......@@ -81,9 +83,9 @@ std::pair<CSRMatrix, NDArray> CusparseCsrgeam2(
device->FreeWorkspace(ctx, workspace);
// destroy matrix/vector descriptors
CUSPARSE_CALL(cusparseDestroyMatDescr(matA));
CUSPARSE_CALL(cusparseDestroyMatDescr(matB));
CUSPARSE_CALL(cusparseDestroyMatDescr(matC));
CUSPARSE_CALL(hipsparseDestroyMatDescr(matA));
CUSPARSE_CALL(hipsparseDestroyMatDescr(matB));
CUSPARSE_CALL(hipsparseDestroyMatDescr(matC));
return {
CSRMatrix(
A.num_rows, A.num_cols, dC_csrOffsets, dC_columns,
......@@ -159,9 +161,9 @@ template std::pair<CSRMatrix, NDArray> CSRSum<kDGLCUDA, int32_t, __half>(
template std::pair<CSRMatrix, NDArray> CSRSum<kDGLCUDA, int64_t, __half>(
const std::vector<CSRMatrix>&, const std::vector<NDArray>&);
#if BF16_ENABLED
template std::pair<CSRMatrix, NDArray> CSRSum<kDGLCUDA, int32_t, __nv_bfloat16>(
template std::pair<CSRMatrix, NDArray> CSRSum<kDGLCUDA, int32_t, __hip_bfloat16>(
const std::vector<CSRMatrix>&, const std::vector<NDArray>&);
template std::pair<CSRMatrix, NDArray> CSRSum<kDGLCUDA, int64_t, __nv_bfloat16>(
template std::pair<CSRMatrix, NDArray> CSRSum<kDGLCUDA, int64_t, __hip_bfloat16>(
const std::vector<CSRMatrix>&, const std::vector<NDArray>&);
#endif // BF16_ENABLED
template std::pair<CSRMatrix, NDArray> CSRSum<kDGLCUDA, int32_t, float>(
......
src/array/cuda/csr_transpose.cc
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2020 by Contributors
* @file array/cuda/csr_transpose.cc
......@@ -24,12 +26,12 @@ template <>
CSRMatrix CSRTranspose<kDGLCUDA, int32_t>(CSRMatrix csr) {
#if CUDART_VERSION < 12000
auto* thr_entry = runtime::CUDAThreadEntry::ThreadLocal();
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
// allocate cusparse handle if needed
if (!thr_entry->cusparse_handle) {
CUSPARSE_CALL(cusparseCreate(&(thr_entry->cusparse_handle)));
CUSPARSE_CALL(hipsparseCreate(&(thr_entry->cusparse_handle)));
}
CUSPARSE_CALL(cusparseSetStream(thr_entry->cusparse_handle, stream));
CUSPARSE_CALL(hipsparseSetStream(thr_entry->cusparse_handle, stream));
NDArray indptr = csr.indptr, indices = csr.indices, data = csr.data;
const int64_t nnz = indices->shape[0];
......@@ -49,30 +51,30 @@ CSRMatrix CSRTranspose<kDGLCUDA, int32_t>(CSRMatrix csr) {
int32_t* t_indices_ptr = static_cast<int32_t*>(t_indices->data);
void* t_data_ptr = t_data->data;
#if CUDART_VERSION >= 10010
#if DTKRT_VERSION >= 10010
auto device = runtime::DeviceAPI::Get(csr.indptr->ctx);
// workspace
size_t workspace_size;
CUSPARSE_CALL(cusparseCsr2cscEx2_bufferSize(
CUSPARSE_CALL(hipsparseCsr2cscEx2_bufferSize(
thr_entry->cusparse_handle, csr.num_rows, csr.num_cols, nnz, data_ptr,
indptr_ptr, indices_ptr, t_data_ptr, t_indptr_ptr, t_indices_ptr,
CUDA_R_32F, CUSPARSE_ACTION_NUMERIC, CUSPARSE_INDEX_BASE_ZERO,
CUSPARSE_CSR2CSC_ALG1, // see cusparse doc for reference
HIP_R_32F, HIPSPARSE_ACTION_NUMERIC, HIPSPARSE_INDEX_BASE_ZERO,
HIPSPARSE_CSR2CSC_ALG1, // see cusparse doc for reference
&workspace_size));
void* workspace = device->AllocWorkspace(ctx, workspace_size);
CUSPARSE_CALL(cusparseCsr2cscEx2(
CUSPARSE_CALL(hipsparseCsr2cscEx2(
thr_entry->cusparse_handle, csr.num_rows, csr.num_cols, nnz, data_ptr,
indptr_ptr, indices_ptr, t_data_ptr, t_indptr_ptr, t_indices_ptr,
CUDA_R_32F, CUSPARSE_ACTION_NUMERIC, CUSPARSE_INDEX_BASE_ZERO,
CUSPARSE_CSR2CSC_ALG1, // see cusparse doc for reference
HIP_R_32F, HIPSPARSE_ACTION_NUMERIC, HIPSPARSE_INDEX_BASE_ZERO,
HIPSPARSE_CSR2CSC_ALG1, // see cusparse doc for reference
workspace));
device->FreeWorkspace(ctx, workspace);
#else
CUSPARSE_CALL(cusparseScsr2csc(
CUSPARSE_CALL(hipsparseScsr2csc(
thr_entry->cusparse_handle, csr.num_rows, csr.num_cols, nnz,
static_cast<const float*>(data_ptr), indptr_ptr, indices_ptr,
static_cast<float*>(t_data_ptr), t_indices_ptr, t_indptr_ptr,
CUSPARSE_ACTION_NUMERIC, CUSPARSE_INDEX_BASE_ZERO));
HIPSPARSE_ACTION_NUMERIC, HIPSPARSE_INDEX_BASE_ZERO));
#endif
return CSRMatrix(
......
src/array/cuda/cuda_filter.cu → src/array/cuda/cuda_filter.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2021 by Contributors
* @file array/cuda/cuda_filter.cc
......@@ -6,7 +8,7 @@
#include <dgl/runtime/device_api.h>
#include <cub/cub.cuh>
#include <hipcub/hipcub.hpp>
#include "../../runtime/cuda/cuda_common.h"
#include "../../runtime/cuda/cuda_hashtable.cuh"
......@@ -45,7 +47,7 @@ IdArray _PerformFilter(const OrderedHashTable<IdType>& table, IdArray test) {
const auto& ctx = test->ctx;
auto device = runtime::DeviceAPI::Get(ctx);
const int64_t size = test->shape[0];
cudaStream_t cudaStream = runtime::getCurrentCUDAStream();
hipStream_t cudaStream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
if (size == 0) {
return test;
......@@ -74,12 +76,12 @@ IdArray _PerformFilter(const OrderedHashTable<IdType>& table, IdArray test) {
// generate prefix-sum
{
size_t workspace_bytes;
CUDA_CALL(cub::DeviceScan::ExclusiveSum(
CUDA_CALL(hipcub::DeviceScan::ExclusiveSum(
nullptr, workspace_bytes, static_cast<IdType*>(nullptr),
static_cast<IdType*>(nullptr), size + 1, cudaStream));
void* workspace = device->AllocWorkspace(ctx, workspace_bytes);
CUDA_CALL(cub::DeviceScan::ExclusiveSum(
CUDA_CALL(hipcub::DeviceScan::ExclusiveSum(
workspace, workspace_bytes, prefix, prefix, size + 1, cudaStream));
device->FreeWorkspace(ctx, workspace);
}
......@@ -108,8 +110,8 @@ template <typename IdType>
class CudaFilterSet : public Filter {
public:
explicit CudaFilterSet(IdArray array)
: table_(array->shape[0], array->ctx, runtime::getCurrentCUDAStream()) {
cudaStream_t cudaStream = runtime::getCurrentCUDAStream();
: table_(array->shape[0], array->ctx, runtime::getCurrentHIPStreamMasqueradingAsCUDA()) {
hipStream_t cudaStream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
table_.FillWithUnique(
static_cast<const IdType*>(array->data), array->shape[0], cudaStream);
}
......
src/array/cuda/cusparse_dispatcher.cuh
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
/**
* Copyright (c) 2020 by Contributors
* @file array/cuda/dispatcher.cuh
......@@ -7,7 +8,7 @@
#ifndef DGL_ARRAY_CUDA_CUSPARSE_DISPATCHER_CUH_
#define DGL_ARRAY_CUDA_CUSPARSE_DISPATCHER_CUH_
#include <cusparse.h>
#include <hipsparse/hipsparse.h>
#include <dgl/runtime/c_runtime_api.h>
#include "bf16.cuh"
......@@ -20,70 +21,70 @@ namespace aten {
template <typename DType>
struct CSRGEMM {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
BUG_IF_FAIL(false) << "This piece of code should not be reached.";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgemm2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgemm2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
static inline hipsparseStatus_t compute(Args... args) {
BUG_IF_FAIL(false) << "This piece of code should not be reached.";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
};
template <>
struct CSRGEMM<__half> {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
// TODO(ndickson): There is no cusparseHcsrgemm2_bufferSizeExt, so a
// different implementation would be required.
LOG(FATAL) << "CSRGEMM::bufferSizeExt does not support dtype half (FP16).";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgemm2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgemm2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
static inline hipsparseStatus_t compute(Args... args) {
// TODO(ndickson): There is no cusparseHcsrgemm2, so a different
// implementation would be required.
LOG(FATAL) << "CSRGEMM::compute does not support dtype half (FP16).";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
};
#if BF16_ENABLED
template <>
struct CSRGEMM<__nv_bfloat16> {
struct CSRGEMM<__hip_bfloat16> {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
// TODO(ndickson): There is no cusparseHcsrgemm2_bufferSizeExt, so a
// different implementation would be required.
LOG(FATAL)
<< "CSRGEMM::bufferSizeExt does not support dtype bfloat16 (BF16).";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgemm2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgemm2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
static inline hipsparseStatus_t compute(Args... args) {
// TODO(ndickson): There is no cusparseHcsrgemm2, so a different
// implementation would be required.
LOG(FATAL) << "CSRGEMM::compute does not support dtype bfloat16 (BF16).";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
};
#endif // BF16_ENABLED
......@@ -91,36 +92,36 @@ struct CSRGEMM<__nv_bfloat16> {
template <>
struct CSRGEMM<float> {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
return cusparseScsrgemm2_bufferSizeExt(args...);
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
return hipsparseScsrgemm2_bufferSizeExt(args...);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgemm2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgemm2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
return cusparseScsrgemm2(args...);
static inline hipsparseStatus_t compute(Args... args) {
return hipsparseScsrgemm2(args...);
}
};
template <>
struct CSRGEMM<double> {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
return cusparseDcsrgemm2_bufferSizeExt(args...);
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
return hipsparseDcsrgemm2_bufferSizeExt(args...);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgemm2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgemm2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
return cusparseDcsrgemm2(args...);
static inline hipsparseStatus_t compute(Args... args) {
return hipsparseDcsrgemm2(args...);
}
};
......@@ -128,70 +129,70 @@ struct CSRGEMM<double> {
template <typename DType>
struct CSRGEAM {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
BUG_IF_FAIL(false) << "This piece of code should not be reached.";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgeam2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgeam2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
static inline hipsparseStatus_t compute(Args... args) {
BUG_IF_FAIL(false) << "This piece of code should not be reached.";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
};
template <>
struct CSRGEAM<__half> {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
// TODO(ndickson): There is no cusparseHcsrgeam2_bufferSizeExt, so a
// different implementation would be required.
LOG(FATAL) << "CSRGEAM::bufferSizeExt does not support dtype half (FP16).";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgeam2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgeam2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
static inline hipsparseStatus_t compute(Args... args) {
// TODO(ndickson): There is no cusparseHcsrgeam2, so a different
// implementation would be required.
LOG(FATAL) << "CSRGEAM::compute does not support dtype half (FP16).";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
};
#if BF16_ENABLED
template <>
struct CSRGEAM<__nv_bfloat16> {
struct CSRGEAM<__hip_bfloat16> {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
// TODO(ndickson): There is no cusparseHcsrgeam2_bufferSizeExt, so a
// different implementation would be required.
LOG(FATAL)
<< "CSRGEAM::bufferSizeExt does not support dtype bfloat16 (BF16).";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgeam2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgeam2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
static inline hipsparseStatus_t compute(Args... args) {
// TODO(ndickson): There is no cusparseHcsrgeam2, so a different
// implementation would be required.
LOG(FATAL) << "CSRGEAM::compute does not support dtype bfloat16 (BF16).";
return static_cast<cusparseStatus_t>(0);
return static_cast<hipsparseStatus_t>(0);
}
};
#endif // BF16_ENABLED
......@@ -199,36 +200,36 @@ struct CSRGEAM<__nv_bfloat16> {
template <>
struct CSRGEAM<float> {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
return cusparseScsrgeam2_bufferSizeExt(args...);
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
return hipsparseScsrgeam2_bufferSizeExt(args...);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgeam2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgeam2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
return cusparseScsrgeam2(args...);
static inline hipsparseStatus_t compute(Args... args) {
return hipsparseScsrgeam2(args...);
}
};
template <>
struct CSRGEAM<double> {
template <typename... Args>
static inline cusparseStatus_t bufferSizeExt(Args... args) {
return cusparseDcsrgeam2_bufferSizeExt(args...);
static inline hipsparseStatus_t bufferSizeExt(Args... args) {
return hipsparseDcsrgeam2_bufferSizeExt(args...);
}
template <typename... Args>
static inline cusparseStatus_t nnz(Args... args) {
return cusparseXcsrgeam2Nnz(args...);
static inline hipsparseStatus_t nnz(Args... args) {
return hipsparseXcsrgeam2Nnz(args...);
}
template <typename... Args>
static inline cusparseStatus_t compute(Args... args) {
return cusparseDcsrgeam2(args...);
static inline hipsparseStatus_t compute(Args... args) {
return hipsparseDcsrgeam2(args...);
}
};
......
src/array/cuda/disjoint_union.cu → src/array/cuda/disjoint_union.hip
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
#include "hip/hip_runtime.h"
/**
* Copyright (c) 2022, NVIDIA CORPORATION.
*
......@@ -24,7 +26,7 @@
#include <vector>
#include "../../runtime/cuda/cuda_common.h"
#include "./utils.h"
#include "utils.h"
namespace dgl {
using runtime::NDArray;
......@@ -78,7 +80,7 @@ std::tuple<IdArray, IdArray, IdArray> _ComputePrefixSums(
template <DGLDeviceType XPU, typename IdType>
void _Merge(
IdType** arrs, IdType* prefix, IdType* offset, IdType* out, int64_t n_arrs,
int n_elms, DGLContext ctx, DGLDataType dtype, cudaStream_t stream) {
int n_elms, DGLContext ctx, DGLDataType dtype, hipStream_t stream) {
auto device = runtime::DeviceAPI::Get(ctx);
int nt = 256;
int nb = (n_elms + nt - 1) / nt;
......@@ -99,7 +101,7 @@ void _Merge(
template <DGLDeviceType XPU, typename IdType>
COOMatrix DisjointUnionCoo(const std::vector<COOMatrix>& coos) {
cudaStream_t stream = runtime::getCurrentCUDAStream();
hipStream_t stream = runtime::getCurrentHIPStreamMasqueradingAsCUDA();
auto device = runtime::DeviceAPI::Get(coos[0].row->ctx);
uint64_t src_offset = 0, dst_offset = 0;
bool has_data = false;
......
src/array/cuda/fp16.cuh
View file @ 74d88bf8
// !!! This is a file automatically generated by hipify!!!
/**
* Copyright (c) 2020-2022 by Contributors
*
......@@ -21,12 +22,12 @@
#ifndef DGL_ARRAY_CUDA_FP16_CUH_
#define DGL_ARRAY_CUDA_FP16_CUH_
#include <cuda_fp16.h>
#include <hip/hip_fp16.h>
#include <algorithm>
static __device__ __forceinline__ half max(half a, half b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
#if defined(__HIP_DEVICE_COMPILE__)
return __hgt(__half(a), __half(b)) ? a : b;
#else
return __half(max(float(a), float(b))); // NOLINT
......@@ -34,19 +35,19 @@ static __device__ __forceinline__ half max(half a, half b) {
}
static __device__ __forceinline__ half min(half a, half b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
#if defined(__HIP_DEVICE_COMPILE__)
return __hlt(__half(a), __half(b)) ? a : b;
#else
return __half(min(float(a), float(b))); // NOLINT
#endif
}
#ifdef __CUDACC__
#if 0
#ifdef __HIPCC__
// Arithmetic FP16 operations for architecture >= 5.3 are already defined in
// cuda_fp16.h
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 530)
// hip/hip_fp16.h
#if defined(__HIP_DEVICE_COMPILE__)
// CUDA 12.2 adds "emulated" support for older architectures.
#if defined(CUDART_VERSION) && (CUDART_VERSION < 12020)
#if defined(DTKRT_VERSION) && (DTKRT_VERSION < 12020)
__device__ __forceinline__ __half
operator+(const __half& lh, const __half& rh) {
return __half(float(lh) + float(rh)); // NOLINT
......@@ -127,8 +128,8 @@ __device__ __forceinline__ bool operator>=(const __half& lh, const __half& rh) {
__device__ __forceinline__ bool operator<=(const __half& lh, const __half& rh) {
return float(lh) <= float(rh); // NOLINT
}
#endif // defined(CUDART_VERSION) && (CUDART_VERSION < 12020)
#endif // defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 530)
#endif // __CUDACC__
#endif // defined(DTKRT_VERSION) && (DTKRT_VERSION < 12020)
#endif // defined(__HIP_DEVICE_COMPILE__)
#endif // __HIPCC__
#endif
#endif // DGL_ARRAY_CUDA_FP16_CUH_
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment