derek.holloway
7502bf64a4
CodeQL / Analyze (actions) (push) Canceled after 0s
CodeQL / Analyze (python) (push) Canceled after 0s
pre-commit / pre-commit (push) Canceled after 0s
Multi-Arch CI / setup (push) Canceled after 0s
Publish build_manylinux_x86_64 images / publish_build_manylinux_x86_64 (push) Canceled after 0s
Setup (Multi-Arch) / setup (push) Canceled after 0s
Multi-Arch CI / Linux::${{ fromJSON(needs.setup.outputs.linux_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubi10 image / publish_no_rocm_image_ubi10 (push) Canceled after 0s
Multi-Arch CI / Windows::${{ fromJSON(needs.setup.outputs.windows_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Multi-Arch CI / Manifest Diff (push) Canceled after 0s
TheRock Manifest Diff Report / Generate Manifest Diff Report (push) Canceled after 0s
Multi-Arch CI / CI Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_ocl_rt images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_rocgdb images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Unit Tests / Unit Tests :: ubuntu-24.04 (push) Canceled after 0s
Unit Tests / Unit Tests :: windows-2022 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_media (push) Canceled after 0s
Unit Tests / Unit Tests Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_openmpi (push) Canceled after 0s
622 lines
22 KiB
C++
622 lines
22 KiB
C++
/* ************************************************************************
|
|
* Copyright (C) 2016-2025 Advanced Micro Devices, Inc. All rights reserved.
|
|
*
|
|
* Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
* of this software and associated documentation files (the "Software"), to deal
|
|
* in the Software without restriction, including without limitation the rights
|
|
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
* copies of the Software, and to permit persons to whom the Software is
|
|
* furnished to do so, subject to the following conditions:
|
|
*
|
|
* The above copyright notice and this permission notice shall be included in
|
|
* all copies or substantial portions of the Software.
|
|
*
|
|
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
* SOFTWARE.
|
|
*
|
|
* ************************************************************************ */
|
|
|
|
#include <fstream>
|
|
#include <iostream>
|
|
#include <stdlib.h>
|
|
#include <vector>
|
|
|
|
#include "arg_check.h"
|
|
#include "testing_common.hpp"
|
|
#include <typeinfo>
|
|
|
|
/* ============================================================================================ */
|
|
|
|
using hipblasGemmBatchedModel = ArgumentModel<e_a_type,
|
|
e_transA,
|
|
e_transB,
|
|
e_M,
|
|
e_N,
|
|
e_K,
|
|
e_alpha,
|
|
e_lda,
|
|
e_ldb,
|
|
e_beta,
|
|
e_ldc,
|
|
e_batch_count>;
|
|
|
|
inline void testname_gemm_batched(const Arguments& arg, std::string& name)
|
|
{
|
|
hipblasGemmBatchedModel{}.test_name(arg, name);
|
|
}
|
|
|
|
template <typename T>
|
|
void testing_gemm_batched_bad_arg(const Arguments& arg)
|
|
{
|
|
using Ts = hipblas_internal_type<T>;
|
|
auto hipblasGemmBatchedFn
|
|
= arg.api == FORTRAN ? hipblasGemmBatched<T, true> : hipblasGemmBatched<T, false>;
|
|
auto hipblasGemmBatchedFn_64
|
|
= arg.api == FORTRAN_64 ? hipblasGemmBatched_64<T, true> : hipblasGemmBatched_64<T, false>;
|
|
|
|
hipblasLocalHandle handle(arg);
|
|
|
|
int64_t M = 101;
|
|
int64_t N = 100;
|
|
int64_t K = 102;
|
|
int64_t lda = 103;
|
|
int64_t ldb = 104;
|
|
int64_t ldc = 105;
|
|
int64_t batch_count = 2;
|
|
|
|
hipblasOperation_t transA = HIPBLAS_OP_N;
|
|
hipblasOperation_t transB = HIPBLAS_OP_N;
|
|
|
|
int64_t A_row = transA == HIPBLAS_OP_N ? M : std::max(K, int64_t(1));
|
|
int64_t A_col = transA == HIPBLAS_OP_N ? std::max(K, int64_t(1)) : M;
|
|
int64_t B_row = transB == HIPBLAS_OP_N ? std::max(K, int64_t(1)) : N;
|
|
int64_t B_col = transB == HIPBLAS_OP_N ? N : std::max(K, int64_t(1));
|
|
|
|
// Allocate device memory
|
|
device_batch_matrix<T> dA(A_row, A_col, lda, batch_count);
|
|
device_batch_matrix<T> dB(B_row, B_col, ldb, batch_count);
|
|
device_batch_matrix<T> dC(M, N, ldc, batch_count);
|
|
|
|
device_vector<T> d_alpha(1), d_beta(1), d_one(1), d_zero(1);
|
|
Ts h_alpha{1.0f}, h_beta{2.0f}, h_one{1.0f}, h_zero{0.0f};
|
|
|
|
if constexpr(std::is_same_v<T, hipblasHalf>)
|
|
h_one = float_to_half(1.0f);
|
|
else if constexpr(is_complex<T>)
|
|
h_one = {1, 0};
|
|
|
|
const Ts* alpha = &h_alpha;
|
|
const Ts* beta = &h_beta;
|
|
const Ts* one = &h_one;
|
|
const Ts* zero = &h_zero;
|
|
|
|
for(auto pointer_mode : {HIPBLAS_POINTER_MODE_HOST, HIPBLAS_POINTER_MODE_DEVICE})
|
|
{
|
|
CHECK_HIPBLAS_ERROR(hipblasSetPointerMode(handle, pointer_mode));
|
|
|
|
if(pointer_mode == HIPBLAS_POINTER_MODE_DEVICE)
|
|
{
|
|
CHECK_HIP_ERROR(hipMemcpy(d_alpha, alpha, sizeof(*alpha), hipMemcpyHostToDevice));
|
|
CHECK_HIP_ERROR(hipMemcpy(d_beta, beta, sizeof(*beta), hipMemcpyHostToDevice));
|
|
CHECK_HIP_ERROR(hipMemcpy(d_one, one, sizeof(*one), hipMemcpyHostToDevice));
|
|
CHECK_HIP_ERROR(hipMemcpy(d_zero, zero, sizeof(*zero), hipMemcpyHostToDevice));
|
|
alpha = d_alpha;
|
|
beta = d_beta;
|
|
one = d_one;
|
|
zero = d_zero;
|
|
}
|
|
|
|
DAPI_EXPECT(HIPBLAS_STATUS_NOT_INITIALIZED,
|
|
hipblasGemmBatchedFn,
|
|
(nullptr,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
alpha,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
beta,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
|
|
DAPI_EXPECT(HIPBLAS_STATUS_INVALID_ENUM,
|
|
hipblasGemmBatchedFn,
|
|
(handle,
|
|
(hipblasOperation_t)HIPBLAS_FILL_MODE_FULL,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
alpha,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
beta,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
DAPI_EXPECT(HIPBLAS_STATUS_INVALID_ENUM,
|
|
hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
(hipblasOperation_t)HIPBLAS_FILL_MODE_FULL,
|
|
M,
|
|
N,
|
|
K,
|
|
alpha,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
beta,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
|
|
if(arg.bad_arg_all)
|
|
{
|
|
DAPI_EXPECT(HIPBLAS_STATUS_INVALID_VALUE,
|
|
hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
alpha,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
nullptr,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
|
|
if(pointer_mode == HIPBLAS_POINTER_MODE_HOST)
|
|
{
|
|
// alpha check only for host mode. rocBLAS can handle this in device mode too but shouldn't assume in case this changes.
|
|
DAPI_EXPECT(HIPBLAS_STATUS_INVALID_VALUE,
|
|
hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
nullptr,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
beta,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
|
|
// again, rocBLAS can handle this in device mode but shouldn't assume
|
|
DAPI_EXPECT(HIPBLAS_STATUS_INVALID_VALUE,
|
|
hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
alpha,
|
|
nullptr,
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
beta,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
DAPI_EXPECT(HIPBLAS_STATUS_INVALID_VALUE,
|
|
hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
alpha,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
beta,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
DAPI_EXPECT(HIPBLAS_STATUS_INVALID_VALUE,
|
|
hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
alpha,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
beta,
|
|
nullptr,
|
|
ldc,
|
|
batch_count));
|
|
}
|
|
|
|
// If alpha == 0 && beta == 1, can have A, B, C be nullptr
|
|
DAPI_CHECK(hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
zero,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
one,
|
|
nullptr,
|
|
ldc,
|
|
batch_count));
|
|
|
|
// If alpha == 0, A and B can be nullptr
|
|
DAPI_CHECK(hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
zero,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
beta,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
|
|
// If K == 0, A, and B can be nullptr
|
|
DAPI_CHECK(hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
0,
|
|
alpha,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
beta,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
|
|
// 64-bit interface test
|
|
DAPI_EXPECT((arg.api & c_API_64) ? HIPBLAS_STATUS_SUCCESS
|
|
: HIPBLAS_STATUS_INVALID_VALUE,
|
|
hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
c_i32_overflow,
|
|
c_i32_overflow,
|
|
c_i32_overflow,
|
|
zero,
|
|
nullptr,
|
|
c_i32_overflow,
|
|
nullptr,
|
|
c_i32_overflow,
|
|
one,
|
|
nullptr,
|
|
c_i32_overflow,
|
|
c_i32_overflow));
|
|
}
|
|
|
|
// If M == 0 || N == 0 || batch_count == 0, can have nullptrs
|
|
DAPI_CHECK(hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
0,
|
|
N,
|
|
K,
|
|
nullptr,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
nullptr,
|
|
nullptr,
|
|
ldc,
|
|
batch_count));
|
|
DAPI_CHECK(hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
0,
|
|
K,
|
|
nullptr,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
nullptr,
|
|
nullptr,
|
|
ldc,
|
|
batch_count));
|
|
DAPI_CHECK(hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
nullptr,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
nullptr,
|
|
nullptr,
|
|
ldc,
|
|
0));
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void testing_gemm_batched(const Arguments& arg)
|
|
{
|
|
using Ts = hipblas_internal_type<T>;
|
|
auto hipblasGemmBatchedFn
|
|
= arg.api == FORTRAN ? hipblasGemmBatched<T, true> : hipblasGemmBatched<T, false>;
|
|
auto hipblasGemmBatchedFn_64
|
|
= arg.api == FORTRAN_64 ? hipblasGemmBatched_64<T, true> : hipblasGemmBatched_64<T, false>;
|
|
|
|
hipblasOperation_t transA = char2hipblas_operation(arg.transA);
|
|
hipblasOperation_t transB = char2hipblas_operation(arg.transB);
|
|
int64_t M = arg.M;
|
|
int64_t N = arg.N;
|
|
int64_t K = arg.K;
|
|
int64_t lda = arg.lda;
|
|
int64_t ldb = arg.ldb;
|
|
int64_t ldc = arg.ldc;
|
|
int64_t batch_count = arg.batch_count;
|
|
|
|
hipblasLocalHandle handle(arg);
|
|
|
|
T h_alpha = arg.get_alpha<T>();
|
|
T h_beta = arg.get_beta<T>();
|
|
|
|
int64_t A_row = transA == HIPBLAS_OP_N ? M : std::max(K, int64_t(1));
|
|
int64_t A_col = transA == HIPBLAS_OP_N ? std::max(K, int64_t(1)) : M;
|
|
int64_t B_row = transB == HIPBLAS_OP_N ? std::max(K, int64_t(1)) : N;
|
|
int64_t B_col = transB == HIPBLAS_OP_N ? N : std::max(K, int64_t(1));
|
|
|
|
// check here to prevent undefined memory allocation error
|
|
bool invalid_size
|
|
= M < 0 || N < 0 || K < 0 || batch_count < 0 || lda < A_row || ldb < B_row || ldc < M;
|
|
if(invalid_size || !M || !N || !batch_count)
|
|
{
|
|
DAPI_EXPECT(invalid_size ? HIPBLAS_STATUS_INVALID_VALUE : HIPBLAS_STATUS_SUCCESS,
|
|
hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
nullptr,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
nullptr,
|
|
nullptr,
|
|
ldc,
|
|
batch_count));
|
|
return;
|
|
}
|
|
|
|
double gpu_time_used{0}, hipblas_error_host{0}, hipblas_error_device{0};
|
|
|
|
// Naming: `h` is in CPU (host) memory(eg hA), `d` is in GPU (device) memory (eg dA).
|
|
// Allocate host memory
|
|
host_batch_matrix<T> hA(A_row, A_col, lda, batch_count);
|
|
host_batch_matrix<T> hB(B_row, B_col, ldb, batch_count);
|
|
host_batch_matrix<T> hC_host(M, N, ldc, batch_count);
|
|
host_batch_matrix<T> hC_device(M, N, ldc, batch_count);
|
|
host_batch_matrix<T> hC_cpu(M, N, ldc, batch_count);
|
|
|
|
// Check host memory allocation
|
|
CHECK_HIP_ERROR(hA.memcheck());
|
|
CHECK_HIP_ERROR(hB.memcheck());
|
|
CHECK_HIP_ERROR(hC_host.memcheck());
|
|
CHECK_HIP_ERROR(hC_device.memcheck());
|
|
CHECK_HIP_ERROR(hC_cpu.memcheck());
|
|
|
|
// Allocate device memory
|
|
device_batch_matrix<T> dA(A_row, A_col, lda, batch_count);
|
|
device_batch_matrix<T> dB(B_row, B_col, ldb, batch_count);
|
|
device_batch_matrix<T> dC(M, N, ldc, batch_count);
|
|
device_vector<T> d_alpha(1);
|
|
device_vector<T> d_beta(1);
|
|
|
|
// Check device memory allocation
|
|
CHECK_DEVICE_ALLOCATION(dA.memcheck());
|
|
CHECK_DEVICE_ALLOCATION(dB.memcheck());
|
|
CHECK_DEVICE_ALLOCATION(dC.memcheck());
|
|
CHECK_DEVICE_ALLOCATION(d_alpha.memcheck());
|
|
CHECK_DEVICE_ALLOCATION(d_beta.memcheck());
|
|
|
|
// Initial Data on CPU
|
|
hipblas_init_matrix(hA, arg, hipblas_client_alpha_sets_nan, hipblas_general_matrix, true);
|
|
hipblas_init_matrix(
|
|
hB, arg, hipblas_client_alpha_sets_nan, hipblas_general_matrix, false, true);
|
|
hipblas_init_matrix(hC_host, arg, hipblas_client_beta_sets_nan, hipblas_general_matrix);
|
|
|
|
// copy vector
|
|
hC_device.copy_from(hC_host);
|
|
hC_cpu.copy_from(hC_host);
|
|
|
|
// copy data from CPU to device, does not work for lda != A_row
|
|
CHECK_HIP_ERROR(dA.transfer_from(hA));
|
|
CHECK_HIP_ERROR(dB.transfer_from(hB));
|
|
CHECK_HIP_ERROR(dC.transfer_from(hC_host));
|
|
CHECK_HIP_ERROR(hipMemcpy(d_alpha, &h_alpha, sizeof(T), hipMemcpyHostToDevice));
|
|
CHECK_HIP_ERROR(hipMemcpy(d_beta, &h_beta, sizeof(T), hipMemcpyHostToDevice));
|
|
|
|
if(arg.unit_check || arg.norm_check)
|
|
{
|
|
// calculate "golden" result on CPU
|
|
for(int64_t i = 0; i < batch_count; i++)
|
|
{
|
|
ref_gemm<T>(transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
h_alpha,
|
|
(T*)hA[i],
|
|
lda,
|
|
(T*)hB[i],
|
|
ldb,
|
|
h_beta,
|
|
(T*)hC_cpu[i],
|
|
ldc);
|
|
}
|
|
|
|
// test hipBLAS batched gemm with alpha and beta pointers on device
|
|
CHECK_HIPBLAS_ERROR(hipblasSetPointerMode(handle, HIPBLAS_POINTER_MODE_DEVICE));
|
|
DAPI_CHECK(hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
d_alpha,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
d_beta,
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
|
|
CHECK_HIP_ERROR(hC_device.transfer_from(dC));
|
|
|
|
// test hipBLAS batched gemm with alpha and beta pointers on host
|
|
CHECK_HIP_ERROR(dC.transfer_from(hC_host));
|
|
CHECK_HIPBLAS_ERROR(hipblasSetPointerMode(handle, HIPBLAS_POINTER_MODE_HOST));
|
|
DAPI_CHECK(hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
reinterpret_cast<Ts*>(&h_alpha),
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
reinterpret_cast<Ts*>(&h_beta),
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
|
|
CHECK_HIP_ERROR(hC_host.transfer_from(dC));
|
|
|
|
if(arg.unit_check)
|
|
{
|
|
if(std::is_same_v<T, hipblasHalf> && (getArchMajor() == 11))
|
|
{
|
|
const double tol = K * sum_error_tolerance_for_gfx11<T, T, T>;
|
|
near_check_general<T>(M, N, batch_count, ldc, hC_cpu, hC_host, tol);
|
|
near_check_general<T>(M, N, batch_count, ldc, hC_cpu, hC_device, tol);
|
|
}
|
|
else
|
|
{
|
|
unit_check_general<T>(M, N, batch_count, ldc, hC_cpu, hC_host);
|
|
unit_check_general<T>(M, N, batch_count, ldc, hC_cpu, hC_device);
|
|
}
|
|
}
|
|
|
|
if(arg.norm_check)
|
|
{
|
|
hipblas_error_host
|
|
= norm_check_general<T>('F', M, N, ldc, hC_cpu, hC_host, batch_count);
|
|
hipblas_error_device
|
|
= norm_check_general<T>('F', M, N, ldc, hC_cpu, hC_device, batch_count);
|
|
}
|
|
}
|
|
|
|
if(arg.timing)
|
|
{
|
|
hipStream_t stream;
|
|
CHECK_HIPBLAS_ERROR(hipblasGetStream(handle, &stream));
|
|
|
|
// gemm has better performance in host mode. In rocBLAS in device mode
|
|
// we need to copy alpha and beta to the host.
|
|
CHECK_HIPBLAS_ERROR(hipblasSetPointerMode(handle, HIPBLAS_POINTER_MODE_HOST));
|
|
|
|
int runs = arg.cold_iters + arg.iters;
|
|
for(int iter = 0; iter < runs; iter++)
|
|
{
|
|
if(iter == arg.cold_iters)
|
|
gpu_time_used = get_time_us_sync(stream);
|
|
|
|
DAPI_DISPATCH(hipblasGemmBatchedFn,
|
|
(handle,
|
|
transA,
|
|
transB,
|
|
M,
|
|
N,
|
|
K,
|
|
reinterpret_cast<Ts*>(&h_alpha),
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
reinterpret_cast<Ts*>(&h_beta),
|
|
dC.ptr_on_device(),
|
|
ldc,
|
|
batch_count));
|
|
}
|
|
|
|
gpu_time_used = get_time_us_sync(stream) - gpu_time_used;
|
|
|
|
hipblasGemmBatchedModel{}.log_args<T>(std::cout,
|
|
arg,
|
|
gpu_time_used,
|
|
gemm_gflop_count<T>(M, N, K),
|
|
gemm_gbyte_count<T>(M, N, K),
|
|
hipblas_error_host,
|
|
hipblas_error_device);
|
|
}
|
|
}
|