diff options
-rw-r--r-- | src/routines/levelx/xgemmbatched.cpp | 8 | ||||
-rw-r--r-- | src/routines/levelx/xgemmstridedbatched.cpp | 8 | ||||
-rw-r--r-- | src/utilities/buffer_test.hpp | 58 |
3 files changed, 64 insertions, 10 deletions
diff --git a/src/routines/levelx/xgemmbatched.cpp b/src/routines/levelx/xgemmbatched.cpp index b12b8734..0d7ae5ab 100644 --- a/src/routines/levelx/xgemmbatched.cpp +++ b/src/routines/levelx/xgemmbatched.cpp @@ -79,11 +79,9 @@ void XgemmBatched<T>::DoGemmBatched(const Layout layout, const Transpose a_trans gemm_kernel_id); // Tests the matrices for validity - for (auto batch = size_t{0}; batch < batch_count; ++batch) { - TestMatrixA(a_one, a_two, a_buffer, a_offsets[batch], a_ld, false); // don't test for invalid LD - TestMatrixB(b_one, b_two, b_buffer, b_offsets[batch], b_ld, false); // don't test for invalid LD - TestMatrixC(c_one, c_two, c_buffer, c_offsets[batch], c_ld); - } + TestBatchedMatrixA(a_one, a_two, a_buffer, a_offsets, a_ld, false); // don't test for invalid LD + TestBatchedMatrixB(b_one, b_two, b_buffer, b_offsets, b_ld, false); // don't test for invalid LD + TestBatchedMatrixC(c_one, c_two, c_buffer, c_offsets, c_ld); // Upload the scalar arguments to the device auto alphas_device = Buffer<T>(context_, BufferAccess::kReadWrite, batch_count); diff --git a/src/routines/levelx/xgemmstridedbatched.cpp b/src/routines/levelx/xgemmstridedbatched.cpp index 5d8945ce..8fb83141 100644 --- a/src/routines/levelx/xgemmstridedbatched.cpp +++ b/src/routines/levelx/xgemmstridedbatched.cpp @@ -78,11 +78,9 @@ void XgemmStridedBatched<T>::DoGemmStridedBatched(const Layout layout, const Tra gemm_kernel_id); // Tests the matrices for validity - for (auto batch = size_t{0}; batch < batch_count; ++batch) { - TestMatrixA(a_one, a_two, a_buffer, a_offset + a_stride * batch, a_ld); - TestMatrixB(b_one, b_two, b_buffer, b_offset + b_stride * batch, b_ld); - TestMatrixC(c_one, c_two, c_buffer, c_offset + c_stride * batch, c_ld); - } + TestStridedBatchedMatrixA(a_one, a_two, a_buffer, a_offset, a_stride, batch_count, a_ld); + TestStridedBatchedMatrixB(b_one, b_two, b_buffer, b_offset, b_stride, batch_count, b_ld); + TestStridedBatchedMatrixC(c_one, c_two, c_buffer, c_offset, c_stride, batch_count, c_ld); // Selects which version of the batched GEMM to run if (do_gemm_direct) { // single generic kernel diff --git a/src/utilities/buffer_test.hpp b/src/utilities/buffer_test.hpp index fd071434..4a2a2c95 100644 --- a/src/utilities/buffer_test.hpp +++ b/src/utilities/buffer_test.hpp @@ -15,6 +15,9 @@ #ifndef CLBLAST_BUFFER_TEST_H_ #define CLBLAST_BUFFER_TEST_H_ +#include <algorithm> +#include <vector> + #include "utilities/utilities.hpp" namespace clblast { @@ -105,6 +108,61 @@ void TestVectorIndex(const size_t n, const Buffer<T> &buffer, const size_t offse } // ================================================================================================= + +// Tests matrix 'A' for validity in a batched setting +template <typename T> +void TestBatchedMatrixA(const size_t one, const size_t two, const Buffer<T>& buffer, + const std::vector<size_t> &offsets, const size_t ld, const bool test_lead_dim = true) { + const auto max_offset = *std::max_element(offsets.begin(), offsets.end()); + TestMatrixA(one, two, buffer, max_offset, ld, test_lead_dim); +} + +// Tests matrix 'B' for validity in a batched setting +template <typename T> +void TestBatchedMatrixB(const size_t one, const size_t two, const Buffer<T>& buffer, + const std::vector<size_t>& offsets, const size_t ld, const bool test_lead_dim = true) { + const auto max_offset = *std::max_element(offsets.begin(), offsets.end()); + TestMatrixB(one, two, buffer, max_offset, ld, test_lead_dim); +} + +// Tests matrix 'C' for validity in a batched setting +template <typename T> +void TestBatchedMatrixC(const size_t one, const size_t two, const Buffer<T>& buffer, + const std::vector<size_t>& offsets, const size_t ld) { + const auto max_offset = *std::max_element(offsets.begin(), offsets.end()); + TestMatrixC(one, two, buffer, max_offset, ld); +} + +// ================================================================================================= + +// Tests matrix 'A' for validity in a strided batched setting +template <typename T> +void TestStridedBatchedMatrixA(const size_t one, const size_t two, const Buffer<T>& buffer, + const size_t offset, const size_t stride, const size_t batch_count, + const size_t ld, const bool test_lead_dim = true) { + const auto last_batch_offset = (batch_count - 1) * stride; + TestMatrixA(one, two, buffer, offset + last_batch_offset, ld, test_lead_dim); +} + +// Tests matrix 'B' for validity in a strided batched setting +template <typename T> +void TestStridedBatchedMatrixB(const size_t one, const size_t two, const Buffer<T>& buffer, + const size_t offset, const size_t stride, const size_t batch_count, + const size_t ld, const bool test_lead_dim = true) { + const auto last_batch_offset = (batch_count - 1) * stride; + TestMatrixB(one, two, buffer, offset + last_batch_offset, ld, test_lead_dim); +} + +// Tests matrix 'C' for validity in a strided batched setting +template <typename T> +void TestStridedBatchedMatrixC(const size_t one, const size_t two, const Buffer<T>& buffer, + const size_t offset, const size_t stride, const size_t batch_count, + const size_t ld) { + const auto last_batch_offset = (batch_count - 1) * stride; + TestMatrixC(one, two, buffer, offset + last_batch_offset, ld); +} + +// ================================================================================================= } // namespace clblast // CLBLAST_BUFFER_TEST_H_ |