diff options
author | Pablo Tello <pablo.tello@arm.com> | 2017-11-13 16:44:08 +0000 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:35:24 +0000 |
commit | 6681d24ccc084a0d98d84edadc8aeb5416159261 (patch) | |
tree | e90b7a771df5daac6bc2ac38d70d906fc9908e77 /src/runtime/NEON/functions | |
parent | 2f8e077378a757128540428b0a1318f80b6dea75 (diff) | |
download | ComputeLibrary-6681d24ccc084a0d98d84edadc8aeb5416159261.tar.gz |
COMPMID-675 - Fixed mismatches in GEMMLowpMatrixMultiplyKernel dotproduct path
Change-Id: I791a08c1e333ce6fc5d537f50ab731fbe066e9c9
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/95737
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Diffstat (limited to 'src/runtime/NEON/functions')
-rw-r--r-- | src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp | 13 |
1 files changed, 4 insertions, 9 deletions
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp index 929ee41220..0fff6c9ca1 100644 --- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp +++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp @@ -41,7 +41,7 @@ namespace arm_compute { #include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp" -#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_s8_12x8.hpp" +#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_12x8.hpp" } // namespace arm_compute using namespace arm_compute; @@ -75,20 +75,15 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, { dot_product_path = true; - // If the DOT product instruction is available, the computation will be performed in int8_t - // In order to take into account this, we need to subtract -128 from a_offset and b_offset - _a_offset -= 128; - _b_offset -= 128; - // Configure matrix multiply kernel struct CPUInfo ci = NEScheduler::get().cpu_info(); const int M = output->info()->tensor_shape().y(); const int N = output->info()->tensor_shape().x(); const int K = a->info()->tensor_shape().x(); - GemmInterleaved<gemm_s8_12x8, int8_t, int32_t> gemm(&ci, M, N, K, false, false); - constexpr size_t alignment = 4096; - _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8)); + const size_t workbench_size = GemmInterleaved<gemm_u8_12x8, gemm_u8_12x8::operand_type, gemm_u8_12x8::result_type>(&ci, M, N, K, false, false).get_working_size(); + constexpr size_t alignment = 4096; + _workspace.allocator()->init(TensorInfo(TensorShape{ (workbench_size + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8)); _memory_group.manage(&_workspace); // Configure matrix multiplication kernel |