Reduce TestMatrix calls for xgemmstridedbatched.
Replace the looped test by a single one with the offset of the last batch.pull/372/head
parent
bf50c4e53e
commit
21b66ca761
|
@ -78,11 +78,9 @@ void XgemmStridedBatched<T>::DoGemmStridedBatched(const Layout layout, const Tra
|
||||||
gemm_kernel_id);
|
gemm_kernel_id);
|
||||||
|
|
||||||
// Tests the matrices for validity
|
// Tests the matrices for validity
|
||||||
for (auto batch = size_t{0}; batch < batch_count; ++batch) {
|
TestStridedBatchedMatrixA(a_one, a_two, a_buffer, a_offset, a_stride, batch_count, a_ld);
|
||||||
TestMatrixA(a_one, a_two, a_buffer, a_offset + a_stride * batch, a_ld);
|
TestStridedBatchedMatrixB(b_one, b_two, b_buffer, b_offset, b_stride, batch_count, b_ld);
|
||||||
TestMatrixB(b_one, b_two, b_buffer, b_offset + b_stride * batch, b_ld);
|
TestStridedBatchedMatrixC(c_one, c_two, c_buffer, c_offset, c_stride, batch_count, c_ld);
|
||||||
TestMatrixC(c_one, c_two, c_buffer, c_offset + c_stride * batch, c_ld);
|
|
||||||
}
|
|
||||||
|
|
||||||
// Selects which version of the batched GEMM to run
|
// Selects which version of the batched GEMM to run
|
||||||
if (do_gemm_direct) { // single generic kernel
|
if (do_gemm_direct) { // single generic kernel
|
||||||
|
|
|
@ -133,6 +133,35 @@ void TestBatchedMatrixC(const size_t one, const size_t two, const Buffer<T>& buf
|
||||||
TestMatrixC(one, two, buffer, max_offset, ld);
|
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
|
} // namespace clblast
|
||||||
|
|
||||||
|
|
Loading…
Reference in New Issue