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
800 lines
32 KiB
C++
800 lines
32 KiB
C++
/* ************************************************************************
|
|
* Copyright (C) 2018-2025 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 cop-
|
|
* ies 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 IM-
|
|
* PLIED, 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 CONNE-
|
|
* CTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
|
|
*
|
|
* ************************************************************************ */
|
|
|
|
#pragma once
|
|
|
|
#include "testing_common.hpp"
|
|
|
|
#include "blas3/rocblas_trsm.hpp"
|
|
#include "src64/blas3/rocblas_trsm_64.hpp"
|
|
|
|
#define ERROR_EPS_MULTIPLIER 40
|
|
#define RESIDUAL_EPS_MULTIPLIER 40
|
|
|
|
template <typename T>
|
|
void testing_trsm_batched_bad_arg(const Arguments& arg)
|
|
{
|
|
auto rocblas_trsm_batched_fn
|
|
= arg.api & c_API_FORTRAN ? rocblas_trsm_batched<T, true> : rocblas_trsm_batched<T, false>;
|
|
auto rocblas_trsm_batched_fn_64 = arg.api & c_API_FORTRAN ? rocblas_trsm_batched_64<T, true>
|
|
: rocblas_trsm_batched_64<T, false>;
|
|
|
|
for(auto pointer_mode : {rocblas_pointer_mode_host, rocblas_pointer_mode_device})
|
|
{
|
|
rocblas_local_handle handle{arg};
|
|
CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, pointer_mode));
|
|
|
|
const int64_t M = 100;
|
|
const int64_t N = 100;
|
|
const int64_t lda = 100;
|
|
const int64_t ldb = 100;
|
|
const int64_t batch_count = 2;
|
|
|
|
DEVICE_MEMCHECK(device_vector<T>, alpha_d, (1));
|
|
DEVICE_MEMCHECK(device_vector<T>, zero_d, (1));
|
|
|
|
const T alpha_h(1), zero_h(0);
|
|
|
|
const T* alpha = &alpha_h;
|
|
const T* zero = &zero_h;
|
|
|
|
if(pointer_mode == rocblas_pointer_mode_device)
|
|
{
|
|
CHECK_HIP_ERROR(hipMemcpy(alpha_d, alpha, sizeof(*alpha), hipMemcpyHostToDevice));
|
|
alpha = alpha_d;
|
|
CHECK_HIP_ERROR(hipMemcpy(zero_d, zero, sizeof(*zero), hipMemcpyHostToDevice));
|
|
zero = zero_d;
|
|
}
|
|
|
|
const rocblas_side side = rocblas_side_left;
|
|
const rocblas_fill uplo = rocblas_fill_upper;
|
|
const rocblas_operation transA = rocblas_operation_none;
|
|
const rocblas_diagonal diag = rocblas_diagonal_non_unit;
|
|
|
|
int64_t K = side == rocblas_side_left ? M : N;
|
|
|
|
// Allocate device memory
|
|
DEVICE_MEMCHECK(device_batch_matrix<T>, dA, (K, K, lda, batch_count));
|
|
DEVICE_MEMCHECK(device_batch_matrix<T>, dB, (M, N, ldb, batch_count));
|
|
|
|
// check for invalid enum
|
|
DAPI_EXPECT(rocblas_status_invalid_value,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
rocblas_side_both,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
alpha,
|
|
dA,
|
|
lda,
|
|
dB,
|
|
ldb,
|
|
batch_count));
|
|
|
|
DAPI_EXPECT(rocblas_status_invalid_value,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
(rocblas_fill)rocblas_side_both,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
alpha,
|
|
dA,
|
|
lda,
|
|
dB,
|
|
ldb,
|
|
batch_count));
|
|
|
|
DAPI_EXPECT(rocblas_status_invalid_value,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
(rocblas_operation)rocblas_side_both,
|
|
diag,
|
|
M,
|
|
N,
|
|
alpha,
|
|
dA,
|
|
lda,
|
|
dB,
|
|
ldb,
|
|
batch_count));
|
|
|
|
DAPI_EXPECT(rocblas_status_invalid_value,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
(rocblas_diagonal)rocblas_side_both,
|
|
M,
|
|
N,
|
|
alpha,
|
|
dA,
|
|
lda,
|
|
dB,
|
|
ldb,
|
|
batch_count));
|
|
|
|
// check for invalid size
|
|
DAPI_EXPECT(
|
|
rocblas_status_invalid_size,
|
|
rocblas_trsm_batched_fn,
|
|
(handle, side, uplo, transA, diag, -1, N, alpha, dA, lda, dB, ldb, batch_count));
|
|
|
|
DAPI_EXPECT(
|
|
rocblas_status_invalid_size,
|
|
rocblas_trsm_batched_fn,
|
|
(handle, side, uplo, transA, diag, M, -1, alpha, dA, lda, dB, ldb, batch_count));
|
|
|
|
DAPI_EXPECT(rocblas_status_invalid_size,
|
|
rocblas_trsm_batched_fn,
|
|
(handle, side, uplo, transA, diag, M, N, alpha, dA, lda, dB, ldb, -1));
|
|
|
|
// check for invalid leading dimension
|
|
DAPI_EXPECT(
|
|
rocblas_status_invalid_size,
|
|
rocblas_trsm_batched_fn,
|
|
(handle, side, uplo, transA, diag, M, N, alpha, dA, lda, dB, M - 1, batch_count));
|
|
|
|
DAPI_EXPECT(rocblas_status_invalid_size,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
rocblas_side_left,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
alpha,
|
|
dA,
|
|
M - 1,
|
|
dB,
|
|
ldb,
|
|
batch_count));
|
|
|
|
DAPI_EXPECT(rocblas_status_invalid_size,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
rocblas_side_right,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
alpha,
|
|
dA,
|
|
N - 1,
|
|
dB,
|
|
ldb,
|
|
batch_count));
|
|
|
|
// check that nullpointer gives rocblas_status_invalid_handle or rocblas_status_invalid_pointer
|
|
DAPI_EXPECT(
|
|
rocblas_status_invalid_handle,
|
|
rocblas_trsm_batched_fn,
|
|
(nullptr, side, uplo, transA, diag, M, N, alpha, dA, lda, dB, ldb, batch_count));
|
|
|
|
DAPI_EXPECT(
|
|
rocblas_status_invalid_pointer,
|
|
rocblas_trsm_batched_fn,
|
|
(handle, side, uplo, transA, diag, M, N, nullptr, dA, lda, dB, ldb, batch_count));
|
|
|
|
if(pointer_mode == rocblas_pointer_mode_host)
|
|
{
|
|
DAPI_EXPECT(rocblas_status_invalid_pointer,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
alpha,
|
|
nullptr,
|
|
lda,
|
|
dB,
|
|
ldb,
|
|
batch_count));
|
|
}
|
|
|
|
DAPI_EXPECT(
|
|
rocblas_status_invalid_pointer,
|
|
rocblas_trsm_batched_fn,
|
|
(handle, side, uplo, transA, diag, M, N, alpha, dA, lda, nullptr, ldb, batch_count));
|
|
|
|
// When batch_count==0, all pointers may be nullptr without error
|
|
DAPI_EXPECT(
|
|
rocblas_status_success,
|
|
rocblas_trsm_batched_fn,
|
|
(handle, side, uplo, transA, diag, M, N, nullptr, nullptr, lda, nullptr, ldb, 0));
|
|
|
|
// When M==0, all pointers may be nullptr without error
|
|
DAPI_EXPECT(rocblas_status_success,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
0,
|
|
N,
|
|
nullptr,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
batch_count));
|
|
|
|
// When N==0, all pointers may be nullptr without error
|
|
DAPI_EXPECT(rocblas_status_success,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
0,
|
|
nullptr,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
batch_count));
|
|
|
|
// If alpha==0, then A can be nullptr without error
|
|
DAPI_EXPECT(rocblas_status_success,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
zero,
|
|
nullptr,
|
|
lda,
|
|
dB.ptr_on_device(),
|
|
ldb,
|
|
batch_count));
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void testing_trsm_batched_internal_interfaces(const Arguments& arg)
|
|
{
|
|
// testing rocblas_internal_trsm_workspace_max_size to ensure that the sizes it gives
|
|
// is large enough for all/various sizes below the sizes given
|
|
|
|
int64_t M = arg.M;
|
|
int64_t N = arg.N;
|
|
int64_t batch_count = arg.batch_count;
|
|
|
|
rocblas_side side = char2rocblas_side(arg.side);
|
|
|
|
size_t w_x_tmp_size, w_x_tmp_arr_size, w_invA_size, w_invA_arr_size, w_x_tmp_size_backup;
|
|
|
|
CHECK_ROCBLAS_ERROR(
|
|
rocblas_internal_trsm_batched_workspace_max_size_64<T>(side,
|
|
M,
|
|
N,
|
|
batch_count,
|
|
&w_x_tmp_size,
|
|
&w_x_tmp_arr_size,
|
|
&w_invA_size,
|
|
&w_invA_arr_size,
|
|
&w_x_tmp_size_backup));
|
|
|
|
// test out below for various sizes below M and N
|
|
for(int64_t m_smaller = M; m_smaller > 0; m_smaller--)
|
|
{
|
|
for(int64_t n_smaller = N; n_smaller > 0; n_smaller--)
|
|
{
|
|
size_t w_x_tmp_size2, w_x_tmp_arr_size2, w_invA_size2, w_invA_arr_size2,
|
|
w_x_tmp_size_backup2;
|
|
|
|
// This is implementation-dependent, but currently we /may/ use less memory with "skinny"
|
|
// matrices when transA == non-transpose.
|
|
// Setting this to transpose will always allocate >= non-transpose invokations, so good
|
|
// for this test
|
|
rocblas_operation transA = rocblas_operation_transpose;
|
|
rocblas_status mem_status = rocblas_internal_trsm_workspace_size<T>(
|
|
side,
|
|
transA,
|
|
m_smaller,
|
|
n_smaller,
|
|
batch_count, // not bothering to test smaller batch_counts
|
|
0, // not supporting supplied invA for max_size fn
|
|
&w_x_tmp_size2,
|
|
&w_x_tmp_arr_size2,
|
|
&w_invA_size2,
|
|
&w_invA_arr_size2,
|
|
&w_x_tmp_size_backup2);
|
|
|
|
if(mem_status != rocblas_status_success && mem_status != rocblas_status_continue)
|
|
CHECK_ROCBLAS_ERROR(mem_status);
|
|
|
|
#ifdef GOOGLE_TEST
|
|
ASSERT_TRUE(w_x_tmp_size2 <= w_x_tmp_size && w_x_tmp_arr_size2 <= w_x_tmp_arr_size
|
|
&& w_invA_size2 <= w_invA_size && w_invA_arr_size2 <= w_invA_arr_size
|
|
&& w_x_tmp_size_backup2 <= w_x_tmp_size_backup);
|
|
#endif
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void testing_trsm_batched(const Arguments& arg)
|
|
{
|
|
auto rocblas_trsm_batched_fn
|
|
= arg.api & c_API_FORTRAN ? rocblas_trsm_batched<T, true> : rocblas_trsm_batched<T, false>;
|
|
auto rocblas_trsm_batched_fn_64 = arg.api & c_API_FORTRAN ? rocblas_trsm_batched_64<T, true>
|
|
: rocblas_trsm_batched_64<T, false>;
|
|
|
|
int64_t M = arg.M;
|
|
int64_t N = arg.N;
|
|
int64_t lda = arg.lda;
|
|
int64_t ldb = arg.ldb;
|
|
int64_t batch_count = arg.batch_count;
|
|
|
|
char char_side = arg.side;
|
|
char char_uplo = arg.uplo;
|
|
char char_transA = arg.transA;
|
|
char char_diag = arg.diag;
|
|
T alpha_h = arg.alpha;
|
|
|
|
rocblas_side side = char2rocblas_side(char_side);
|
|
rocblas_fill uplo = char2rocblas_fill(char_uplo);
|
|
rocblas_operation transA = char2rocblas_operation(char_transA);
|
|
rocblas_diagonal diag = char2rocblas_diagonal(char_diag);
|
|
|
|
// for internal interface testing: using ldc/ldd as offsets
|
|
const bool internal_api = arg.api == INTERNAL;
|
|
rocblas_stride offsetA = internal_api ? arg.ldc : 0;
|
|
rocblas_stride offsetB = internal_api ? arg.ldd : 0;
|
|
|
|
int64_t K = side == rocblas_side_left ? M : N;
|
|
|
|
rocblas_local_handle handle{arg};
|
|
// check here to prevent undefined memory allocation error
|
|
bool invalid_size = M < 0 || N < 0 || lda < K || ldb < M || batch_count < 0;
|
|
if(invalid_size || batch_count == 0)
|
|
{
|
|
CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_host));
|
|
DAPI_EXPECT(invalid_size ? rocblas_status_invalid_size : rocblas_status_success,
|
|
rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
nullptr,
|
|
nullptr,
|
|
lda,
|
|
nullptr,
|
|
ldb,
|
|
batch_count));
|
|
return;
|
|
}
|
|
|
|
// Naming: `h` is in CPU (host) memory(eg hA), `d` is in GPU (device) memory (eg dA).
|
|
// Allocate host memory
|
|
HOST_MEMCHECK(host_batch_matrix<T>, hA, (K, K, lda, batch_count));
|
|
HOST_MEMCHECK(host_batch_matrix<T>, hB, (M, N, M, batch_count)); // save memory when large ldb
|
|
HOST_MEMCHECK(host_batch_matrix<T>, hX, (M, N, ldb, batch_count));
|
|
HOST_MEMCHECK(host_batch_matrix<T>, hXorB_1, (M, N, ldb, batch_count));
|
|
HOST_MEMCHECK(host_vector<T>, halpha, (1));
|
|
halpha[0] = alpha_h;
|
|
|
|
// Allocate device memory
|
|
DEVICE_MEMCHECK(device_batch_matrix<T>, dA, (K, K, lda, batch_count, false, offsetA));
|
|
DEVICE_MEMCHECK(device_batch_matrix<T>, dXorB, (M, N, ldb, batch_count, false, offsetB));
|
|
DEVICE_MEMCHECK(device_vector<T>, alpha_d, (1));
|
|
|
|
// Initialize data on host memory
|
|
rocblas_init_matrix(hA,
|
|
arg,
|
|
rocblas_client_never_set_nan,
|
|
rocblas_client_diagonally_dominant_triangular_matrix,
|
|
true);
|
|
rocblas_init_matrix(
|
|
hX, arg, rocblas_client_never_set_nan, rocblas_client_general_matrix, false, true);
|
|
|
|
// make hA unit diagonal if diag == rocblas_diagonal_unit
|
|
if(diag == rocblas_diagonal_unit)
|
|
{
|
|
make_unit_diagonal(uplo, hA);
|
|
}
|
|
|
|
copy_matrix_with_different_leading_dimensions(hX, hB);
|
|
|
|
for(int b = 0; b < batch_count; b++)
|
|
{
|
|
// Calculate hB = hA*hX
|
|
ref_trmm<T>(side, uplo, transA, diag, M, N, 1.0 / alpha_h, hA[b], lda, hB[b], M);
|
|
}
|
|
|
|
copy_matrix_with_different_leading_dimensions(hB, hXorB_1);
|
|
|
|
CHECK_HIP_ERROR(dA.transfer_from(hA));
|
|
CHECK_HIP_ERROR(dXorB.transfer_from(hXorB_1));
|
|
|
|
double error_eps_multiplier = ERROR_EPS_MULTIPLIER;
|
|
double residual_eps_multiplier = RESIDUAL_EPS_MULTIPLIER;
|
|
double eps = std::numeric_limits<real_t<T>>::epsilon();
|
|
double err_host = 0.0;
|
|
double err_device = 0.0;
|
|
|
|
if(!ROCBLAS_REALLOC_ON_DEMAND)
|
|
{
|
|
// Compute size
|
|
CHECK_ROCBLAS_ERROR(rocblas_start_device_memory_size_query(handle));
|
|
DAPI_CHECK_ALLOC_QUERY(rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
&alpha_h,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dXorB.ptr_on_device(),
|
|
ldb,
|
|
batch_count));
|
|
size_t size;
|
|
CHECK_ROCBLAS_ERROR(rocblas_stop_device_memory_size_query(handle, &size));
|
|
|
|
// Allocate memory
|
|
CHECK_ROCBLAS_ERROR(rocblas_set_device_memory_size(handle, size));
|
|
}
|
|
|
|
if(arg.unit_check || arg.norm_check)
|
|
{
|
|
if(arg.pointer_mode_host)
|
|
{
|
|
// calculate dXorB <- A^(-1) B rocblas_device_pointer_host
|
|
CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_host));
|
|
CHECK_HIP_ERROR(dXorB.transfer_from(hXorB_1));
|
|
|
|
handle.pre_test(arg);
|
|
if(!internal_api)
|
|
{
|
|
DAPI_CHECK(rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
&alpha_h,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dXorB.ptr_on_device(),
|
|
ldb,
|
|
batch_count));
|
|
}
|
|
else
|
|
{
|
|
// internal function requires us to supply temporary memory ourselves
|
|
constexpr bool BATCHED = true;
|
|
bool optimal_mem = true;
|
|
int64_t supp_invA_size = 0; // used for trsm_ex
|
|
|
|
// first exported internal interface - calculate how much mem is needed
|
|
size_t w_x_tmp_size, w_x_tmp_arr_size, w_invA_size, w_invA_arr_size,
|
|
w_x_tmp_size_backup;
|
|
rocblas_status mem_status
|
|
= rocblas_internal_trsm_batched_workspace_size<T>(side,
|
|
transA,
|
|
M,
|
|
N,
|
|
batch_count,
|
|
supp_invA_size,
|
|
&w_x_tmp_size,
|
|
&w_x_tmp_arr_size,
|
|
&w_invA_size,
|
|
&w_invA_arr_size,
|
|
&w_x_tmp_size_backup);
|
|
|
|
if(mem_status != rocblas_status_success && mem_status != rocblas_status_continue)
|
|
CHECK_ROCBLAS_ERROR(mem_status);
|
|
|
|
// allocate memory ourselves
|
|
DEVICE_MEMCHECK(device_vector<T>, w_mem_x_tmp, (w_x_tmp_size / sizeof(T)));
|
|
DEVICE_MEMCHECK(
|
|
device_batch_vector<T>, w_mem_x_tmp_arr, (1, 1, w_x_tmp_arr_size / sizeof(T*)));
|
|
DEVICE_MEMCHECK(device_vector<T>, w_mem_invA, (w_invA_size / sizeof(T)));
|
|
DEVICE_MEMCHECK(
|
|
device_batch_vector<T>, w_mem_invA_arr, (1, 1, w_invA_arr_size / sizeof(T*)));
|
|
|
|
rocblas_stride strideA = 0, strideB = 0;
|
|
|
|
CHECK_ROCBLAS_ERROR(
|
|
rocblas_internal_trsm_batched_template(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
&alpha_h,
|
|
(const T* const*)dA.ptr_on_device(),
|
|
-offsetA,
|
|
lda,
|
|
strideA,
|
|
(T* const*)dXorB.ptr_on_device(),
|
|
-offsetB,
|
|
ldb,
|
|
strideB,
|
|
batch_count,
|
|
optimal_mem,
|
|
(void*)w_mem_x_tmp,
|
|
(void*)w_mem_x_tmp_arr.ptr_on_device(),
|
|
(void*)w_mem_invA,
|
|
(void*)w_mem_invA_arr.ptr_on_device()));
|
|
}
|
|
|
|
handle.post_test(arg);
|
|
CHECK_HIP_ERROR(hXorB_1.transfer_from(dXorB));
|
|
|
|
// doing unit tests here to save memory by having hB just use M as leading dimension,
|
|
// need to reuse hXorB for hipMemcpy later
|
|
if(alpha_h == 0)
|
|
{
|
|
// expecting 0 output, set hX == 0
|
|
for(int64_t b = 0; b < batch_count; b++)
|
|
rocblas_init_zero((T*)hX[b], M, N, ldb);
|
|
|
|
if(arg.unit_check)
|
|
unit_check_general<T>(M, N, ldb, hX, hXorB_1, batch_count);
|
|
if(arg.norm_check)
|
|
err_host
|
|
= std::abs(norm_check_general<T>('F', M, N, ldb, hX, hXorB_1, batch_count));
|
|
}
|
|
else
|
|
{
|
|
//computed result is in hx_or_b, so forward error is E = hx - hx_or_b
|
|
// calculate vector-induced-norm 1 of matrix E
|
|
for(int64_t b = 0; b < batch_count; b++)
|
|
{
|
|
double err_host_batch = matrix_norm_1<T>(M, N, ldb, hX[b], hXorB_1[b]);
|
|
|
|
if(arg.unit_check)
|
|
trsm_err_res_check<T>(err_host_batch, M, error_eps_multiplier, eps);
|
|
|
|
// hx_or_b contains A * (calculated X), so res = A * (calculated x) - b = hx_or_b - hb
|
|
ref_trmm<T>(
|
|
side, uplo, transA, diag, M, N, 1.0 / alpha_h, hA[b], lda, hXorB_1[b], ldb);
|
|
// calculate vector-induced-norm 1 of matrix res
|
|
double err_host_res_batch = matrix_norm_1<T>(M, N, hXorB_1[b], ldb, hB[b], M);
|
|
|
|
if(arg.unit_check)
|
|
trsm_err_res_check<T>(err_host_res_batch, M, residual_eps_multiplier, eps);
|
|
err_host = std::max(std::max(err_host_batch, err_host_res_batch), err_host);
|
|
}
|
|
}
|
|
}
|
|
|
|
if(arg.pointer_mode_device && !internal_api)
|
|
{
|
|
// calculate dXorB <- A^(-1) B rocblas_device_pointer_device
|
|
CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_device));
|
|
|
|
// copy hB to hXorB with correct leading dimension as hB still holds input
|
|
copy_matrix_with_different_leading_dimensions(hB, hXorB_1);
|
|
|
|
CHECK_HIP_ERROR(dXorB.transfer_from(hXorB_1));
|
|
CHECK_HIP_ERROR(alpha_d.transfer_from(halpha));
|
|
|
|
DAPI_CHECK(rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
alpha_d,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dXorB.ptr_on_device(),
|
|
ldb,
|
|
batch_count));
|
|
|
|
CHECK_HIP_ERROR(hXorB_1.transfer_from(dXorB));
|
|
|
|
if(arg.repeatability_check)
|
|
{
|
|
HOST_MEMCHECK(host_batch_matrix<T>, hXorB_copy, (M, N, ldb, batch_count));
|
|
|
|
// multi-GPU support
|
|
int device_id, device_count;
|
|
CHECK_HIP_ERROR(limit_device_count(device_count, (int)arg.devices));
|
|
|
|
for(int dev_id = 0; dev_id < device_count; dev_id++)
|
|
{
|
|
CHECK_HIP_ERROR(hipGetDevice(&device_id));
|
|
if(device_id != dev_id)
|
|
CHECK_HIP_ERROR(hipSetDevice(dev_id));
|
|
|
|
//New rocblas handle for new device
|
|
rocblas_local_handle handle_copy{arg};
|
|
|
|
//Allocate device memory in new device
|
|
DEVICE_MEMCHECK(
|
|
device_batch_matrix<T>, dA_copy, (K, K, lda, batch_count, false, offsetA));
|
|
DEVICE_MEMCHECK(device_batch_matrix<T>,
|
|
dXorB_copy,
|
|
(M, N, ldb, batch_count, false, offsetB));
|
|
DEVICE_MEMCHECK(device_vector<T>, alpha_d_copy, (1));
|
|
|
|
// copy data from CPU to device
|
|
CHECK_HIP_ERROR(dA_copy.transfer_from(hA));
|
|
CHECK_HIP_ERROR(alpha_d_copy.transfer_from(halpha));
|
|
|
|
CHECK_ROCBLAS_ERROR(
|
|
rocblas_set_pointer_mode(handle_copy, rocblas_pointer_mode_device));
|
|
|
|
for(int runs = 0; runs < arg.iters; runs++)
|
|
{
|
|
copy_matrix_with_different_leading_dimensions(hB, hXorB_copy);
|
|
CHECK_HIP_ERROR(dXorB_copy.transfer_from(hXorB_copy));
|
|
|
|
CHECK_ROCBLAS_ERROR(rocblas_trsm_batched_fn(handle_copy,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
alpha_d_copy,
|
|
dA_copy.ptr_on_device(),
|
|
lda,
|
|
dXorB_copy.ptr_on_device(),
|
|
ldb,
|
|
batch_count));
|
|
CHECK_HIP_ERROR(hXorB_copy.transfer_from(dXorB_copy));
|
|
unit_check_general<T>(M, N, ldb, hXorB_1, hXorB_copy, batch_count);
|
|
}
|
|
}
|
|
return;
|
|
}
|
|
|
|
if(alpha_h == 0)
|
|
{
|
|
// expecting 0 output, set hX == 0
|
|
for(int64_t b = 0; b < batch_count; b++)
|
|
rocblas_init_zero((T*)hX[b], M, N, ldb);
|
|
|
|
if(arg.unit_check)
|
|
unit_check_general<T>(M, N, ldb, hX, hXorB_1, batch_count);
|
|
if(arg.norm_check)
|
|
err_device
|
|
= std::abs(norm_check_general<T>('F', M, N, ldb, hX, hXorB_1, batch_count));
|
|
}
|
|
else
|
|
{
|
|
//computed result is in hx_or_b, so forward error is E = hx - hx_or_b
|
|
// calculate vector-induced-norm 1 of matrix E
|
|
for(int64_t b = 0; b < batch_count; b++)
|
|
{
|
|
double err_device_batch = matrix_norm_1<T>(M, N, ldb, hX[b], hXorB_1[b]);
|
|
|
|
if(arg.unit_check)
|
|
trsm_err_res_check<T>(err_device_batch, M, error_eps_multiplier, eps);
|
|
|
|
ref_trmm<T>(
|
|
side, uplo, transA, diag, M, N, 1.0 / alpha_h, hA[b], lda, hXorB_1[b], ldb);
|
|
double err_device_res_batch = matrix_norm_1<T>(M, N, hXorB_1[b], ldb, hB[b], M);
|
|
|
|
if(arg.unit_check)
|
|
trsm_err_res_check<T>(
|
|
err_device_res_batch, M, residual_eps_multiplier, eps);
|
|
err_device
|
|
= std::max(std::max(err_device_batch, err_device_res_batch), err_device);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
// can't use external interface with device memory as set up with
|
|
// internal tests/offsets, so ensuring not used when internal_api is set.
|
|
if(arg.timing && !internal_api)
|
|
{
|
|
double gpu_time_used, cpu_time_used;
|
|
int number_cold_calls = arg.cold_iters;
|
|
int total_calls = number_cold_calls + arg.iters;
|
|
|
|
// GPU rocBLAS
|
|
CHECK_HIP_ERROR(dXorB.transfer_from(hXorB_1));
|
|
|
|
CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_host));
|
|
|
|
hipStream_t stream;
|
|
CHECK_ROCBLAS_ERROR(rocblas_get_stream(handle, &stream));
|
|
|
|
for(int i = 0; i < total_calls; i++)
|
|
{
|
|
if(i == number_cold_calls)
|
|
gpu_time_used = get_time_us_sync(stream);
|
|
|
|
DAPI_DISPATCH(rocblas_trsm_batched_fn,
|
|
(handle,
|
|
side,
|
|
uplo,
|
|
transA,
|
|
diag,
|
|
M,
|
|
N,
|
|
&alpha_h,
|
|
dA.ptr_on_device(),
|
|
lda,
|
|
dXorB.ptr_on_device(),
|
|
ldb,
|
|
batch_count));
|
|
}
|
|
|
|
gpu_time_used = get_time_us_sync(stream) - gpu_time_used;
|
|
|
|
// CPU cblas
|
|
cpu_time_used = get_time_us_no_sync();
|
|
|
|
for(int64_t b = 0; b < batch_count; b++)
|
|
ref_trsm<T>(side, uplo, transA, diag, M, N, alpha_h, hA[b], lda, hXorB_1[b], ldb);
|
|
|
|
cpu_time_used = get_time_us_no_sync() - cpu_time_used;
|
|
|
|
ArgumentModel<e_side,
|
|
e_uplo,
|
|
e_transA,
|
|
e_diag,
|
|
e_M,
|
|
e_N,
|
|
e_alpha,
|
|
e_lda,
|
|
e_ldb,
|
|
e_batch_count>{}
|
|
.log_args<T>(rocblas_cout,
|
|
arg,
|
|
gpu_time_used,
|
|
trsm_gflop_count<T>(M, N, K),
|
|
ArgumentLogging::NA_value,
|
|
cpu_time_used,
|
|
err_host,
|
|
err_device);
|
|
}
|
|
}
|