File size: 10,671 Bytes
c1af2fa |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 |
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
#pragma once
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/tunable/TunableOp.h>
#include <ATen/cuda/tunable/GemmCommon.h>
#include <c10/util/StringUtil.h>
#include <fmt/printf.h>
#define ROCBLAS_BETA_FEATURES_API
#include <rocblas/rocblas.h>
#define TORCH_ROCBLAS_CHECK(EXPR) \
do { \
rocblas_status __err = EXPR; \
TORCH_CHECK(__err == rocblas_status_success, \
"rocblas error: ", \
rocblas_status_to_string(__err), \
" when calling `" #EXPR "`"); \
} while (0)
namespace at::cuda::tunable {
template <typename T>
constexpr rocblas_datatype RocBlasDataTypeFor();
template <>
constexpr rocblas_datatype RocBlasDataTypeFor<float>() {
return rocblas_datatype_f32_r;
}
template <>
constexpr rocblas_datatype RocBlasDataTypeFor<double>() {
return rocblas_datatype_f64_r;
}
template <>
constexpr rocblas_datatype RocBlasDataTypeFor<Half>() {
return rocblas_datatype_f16_r;
}
template <>
constexpr rocblas_datatype RocBlasDataTypeFor<BFloat16>() {
return rocblas_datatype_bf16_r;
}
template <>
constexpr rocblas_datatype RocBlasDataTypeFor<c10::complex<float>>() {
return rocblas_datatype_f32_c;
}
template <>
constexpr rocblas_datatype RocBlasDataTypeFor<c10::complex<double>>() {
return rocblas_datatype_f64_c;
}
template <typename T>
constexpr rocblas_datatype RocBlasComputeTypeFor();
template <>
constexpr rocblas_datatype RocBlasComputeTypeFor<float>() {
return rocblas_datatype_f32_r;
}
template <>
constexpr rocblas_datatype RocBlasComputeTypeFor<double>() {
return rocblas_datatype_f64_r;
}
template <>
constexpr rocblas_datatype RocBlasComputeTypeFor<Half>() {
// Note that we're returning the _compute_ type for a given datatype.
// As of 12/2022, using compute type FP16 for 16-bit floats was much
// slower than using compute type FP32. So we use FP32 compute even for
// FP16 datatypes. This is how GEMM is implemented even in the function
// rocblasGemmHelper (see fpgeneric.h)
return rocblas_datatype_f32_r;
}
template <>
constexpr rocblas_datatype RocBlasComputeTypeFor<BFloat16>() {
// Note that we're returning the _compute_ type for a given datatype.
// As of 12/2022, using compute type FP16 for 16-bit floats was much
// slower than using compute type FP32. So we use FP32 compute even for
// BF16 datatypes. This is how GEMM is implemented even in the function
// rocblasGemmHelper (see fpgeneric.h)
return rocblas_datatype_f32_r;
}
template <>
constexpr rocblas_datatype RocBlasComputeTypeFor<c10::complex<float>>() {
return rocblas_datatype_f32_c;
}
template <>
constexpr rocblas_datatype RocBlasComputeTypeFor<c10::complex<double>>() {
return rocblas_datatype_f64_c;
}
template <typename T>
auto DoCastForHalfOrBfloat16(const T fp) {
return fp;
}
template <>
inline auto DoCastForHalfOrBfloat16<Half>(const Half fp) {
// alpha and beta should be the same as compute_type, in Half case it is float.
float h = fp;
return h;
}
template <>
inline auto DoCastForHalfOrBfloat16<BFloat16>(const BFloat16 fp) {
// alpha and beta should be the same as compute_type, in bfloat16 case it is float.
float h = fp;
return h;
}
static rocblas_operation _rocblasOpFromChar(char op) {
switch (op) {
case 'n':
case 'N':
return rocblas_operation_none;
case 't':
case 'T':
return rocblas_operation_transpose;
case 'c':
case 'C':
return rocblas_operation_conjugate_transpose;
}
TORCH_CHECK(false,
"_rocblasOpFromChar input should be 't', 'n' or 'c' but got `", op, "`");
}
template <typename T>
class RocblasGemmOp : public Callable<GemmParams<T>> {
public:
RocblasGemmOp(int solution) : solution_{solution} {}
TuningStatus Call(const GemmParams<T>* params) override {
auto input_output_type = RocBlasDataTypeFor<T>();
if (at::globalContext().allowTF32CuBLAS() && input_output_type == rocblas_datatype_f32_r)
return FAIL; // no support for TF32 in rocBLAS
auto compute_type = RocBlasComputeTypeFor<T>();
auto h_a = DoCastForHalfOrBfloat16(params->alpha);
auto h_b = DoCastForHalfOrBfloat16(params->beta);
auto status = rocblas_gemm_ex(
(rocblas_handle)at::cuda::getCurrentCUDABlasHandle(),
_rocblasOpFromChar(params->transa),
_rocblasOpFromChar(params->transb),
params->m, params->n, params->k,
&h_a,
params->a, input_output_type, params->lda,
params->b, input_output_type, params->ldb,
&h_b,
params->c, input_output_type, params->ldc,
params->c, input_output_type, params->ldc,
compute_type,
rocblas_gemm_algo_solution_index,
solution_,
rocblas_gemm_flags_none);
if (status != rocblas_status_success) {
return FAIL;
}
return OK;
}
private:
int solution_;
};
template <typename T>
auto GetRocBlasGemmTypeStringAndOps() {
rocblas_handle handle = (rocblas_handle)at::cuda::getCurrentCUDABlasHandle();
int solution_size;
auto input_output_type = RocBlasDataTypeFor<T>();
auto compute_type = RocBlasComputeTypeFor<T>();
// Get the number of available solutions
TORCH_ROCBLAS_CHECK(rocblas_gemm_ex_get_solutions_by_type(handle,
input_output_type,
input_output_type,
compute_type,
rocblas_gemm_flags_none,
nullptr,
&solution_size));
std::vector<int> solutions(solution_size);
// Get the list of available solutions
TORCH_ROCBLAS_CHECK(rocblas_gemm_ex_get_solutions_by_type(handle,
input_output_type,
input_output_type,
compute_type,
rocblas_gemm_flags_none,
solutions.data(),
&solution_size));
std::vector<std::pair<std::string, std::unique_ptr<Callable<GemmParams<T>>>>> ret;
for (size_t i = 0; i < solutions.size(); ++i) {
auto callable = std::make_unique<RocblasGemmOp<T>>(solutions[i]);
ret.emplace_back(std::make_pair(fmt::sprintf("Gemm_Rocblas_%d", solutions[i]), std::move(callable)));
}
return ret;
}
template <typename T>
class RocblasGemmStridedBatchedOp : public Callable<GemmStridedBatchedParams<T>> {
public:
RocblasGemmStridedBatchedOp(int solution) : solution_{solution} {}
TuningStatus Call(const GemmStridedBatchedParams<T>* params) override {
auto input_output_type = RocBlasDataTypeFor<T>();
if (at::globalContext().allowTF32CuBLAS() && input_output_type == rocblas_datatype_f32_r)
return FAIL; // no support for TF32 in rocBLAS
auto compute_type = RocBlasComputeTypeFor<T>();
auto h_a = DoCastForHalfOrBfloat16(params->alpha);
auto h_b = DoCastForHalfOrBfloat16(params->beta);
auto status = rocblas_gemm_strided_batched_ex(
(rocblas_handle)at::cuda::getCurrentCUDABlasHandle(),
_rocblasOpFromChar(params->transa),
_rocblasOpFromChar(params->transb),
params->m, params->n, params->k,
&h_a,
params->a, input_output_type, params->lda, params->stride_a,
params->b, input_output_type, params->ldb, params->stride_b,
&h_b,
params->c, input_output_type, params->ldc, params->stride_c,
params->c, input_output_type, params->ldc, params->stride_c,
params->batch,
compute_type,
rocblas_gemm_algo_solution_index,
solution_,
rocblas_gemm_flags_none);
if (status != rocblas_status_success) {
return FAIL;
}
return OK;
}
private:
int solution_;
};
template <typename T>
auto GetRocBlasGemmStridedBatchedTypeStringAndOps() {
rocblas_handle handle = (rocblas_handle)at::cuda::getCurrentCUDABlasHandle();
int solution_size;
auto input_output_type = RocBlasDataTypeFor<T>();
auto compute_type = RocBlasComputeTypeFor<T>();
// Get the number of available solutions
TORCH_ROCBLAS_CHECK(rocblas_gemm_ex_get_solutions_by_type(handle,
input_output_type,
input_output_type,
compute_type,
rocblas_gemm_flags_none,
nullptr,
&solution_size));
std::vector<int> solutions(solution_size);
// Get the list of available solutions
TORCH_ROCBLAS_CHECK(rocblas_gemm_ex_get_solutions_by_type(handle,
input_output_type,
input_output_type,
compute_type,
rocblas_gemm_flags_none,
solutions.data(),
&solution_size));
// Sort the solutions in ascending order to make the solution vector deterministic across runs
std::sort(solutions.begin(), solutions.end());
std::vector<std::pair<std::string, std::unique_ptr<Callable<GemmStridedBatchedParams<T>>>>> ret;
for (size_t i = 0; i < solutions.size(); ++i) {
auto callable = std::make_unique<RocblasGemmStridedBatchedOp<T>>(solutions[i]);
ret.emplace_back(std::make_pair(c10::str("Gemm_Rocblas_", solutions[i]), std::move(callable)));
}
return ret;
}
} // namespace at::cuda::tunable
|