diff options
Diffstat (limited to 'src/runtime/CL/functions/CLGEMM.cpp')
| -rw-r--r-- | src/runtime/CL/functions/CLGEMM.cpp | 30 |
1 files changed, 8 insertions, 22 deletions
diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp index 07b19421d6..6d22825694 100644 --- a/src/runtime/CL/functions/CLGEMM.cpp +++ b/src/runtime/CL/functions/CLGEMM.cpp @@ -45,13 +45,11 @@ CLGEMM::CLGEMM() void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::F32, DataType::F16); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32, DataType::F16); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QS8, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output); if(c != nullptr) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c); ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A"); ARM_COMPUTE_ERROR_ON_MSG(b->info()->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix C"); @@ -59,7 +57,6 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor * ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != output->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix"); } - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output); ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); // Check if the first input tensor is a vector. If so, all the kernels for reshaping the tensors can be skipped @@ -73,25 +70,14 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor * shape_tmp_a.set(0, a->info()->dimension(0) * 4); shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f)); - if(DataType::F32 == a->info()->data_type()) - { - shape_tmp_b.set(0, b->info()->dimension(1) * 4); - shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 4.0f)); - } - else if(DataType::F16 == a->info()->data_type()) - { - shape_tmp_b.set(0, b->info()->dimension(1) * 8); - shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 8.0f)); - } - else - { - ARM_COMPUTE_ERROR("DataType not supported"); - } - - TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type()); + const unsigned int transpose_w = max_cl_vector_width / data_size_from_type(b->info()->data_type()); + shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w); + shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast<float>(transpose_w))); + + TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position()); _tmp_a.allocator()->init(info_a); - TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type()); + TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), b->info()->fixed_point_position()); _tmp_b.allocator()->init(info_b); // Configure interleave kernel |