diff options
author | morgolock <pablo.tello@arm.com> | 2020-10-07 10:36:28 +0100 |
---|---|---|
committer | Pablo Marquez <pablo.tello@arm.com> | 2020-10-07 15:07:04 +0000 |
commit | f316e9ae24d12e98dd2669316d6dfb97fcccc24d (patch) | |
tree | f63601d2e70e0fe051b3e53c1943358fcfac9848 /src/runtime/NEON/functions | |
parent | d304adbb1c6a2f66144c9cac1104f6e3f30d255a (diff) | |
download | ComputeLibrary-f316e9ae24d12e98dd2669316d6dfb97fcccc24d.tar.gz |
MLCE-229: enabled gemmlowp asm kernel for neg shifts
Change-Id: I9354f7f1cb2583677441cc7b1ac857a5e950e42e
Signed-off-by: morgolock <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4100
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/runtime/NEON/functions')
-rw-r--r-- | src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp | 29 |
1 files changed, 4 insertions, 25 deletions
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp index dada6d16da..83db146a8a 100644 --- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp +++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp @@ -117,18 +117,8 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, { if(is_data_type_quantized_asymmetric(a_to_use->info()->data_type()) && info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) { - // Result shifts < 0 are not supported by asm kernels - const std::vector<int32_t> &shifts = info.gemmlowp_output_stage().gemmlowp_shifts; - const bool is_asm_supported = info.gemmlowp_output_stage().gemmlowp_shift >= 0 - && std::all_of(shifts.cbegin(), shifts.cend(), [](int32_t val) - { - return val >= 0; - }); - if(is_asm_supported) - { - _asm_glue.configure(a_to_use, b, c, output, gemm_info); - _fused_assembly_path = _asm_glue.is_configured(); - } + _asm_glue.configure(a_to_use, b, c, output, gemm_info); + _fused_assembly_path = _asm_glue.is_configured(); } else { @@ -339,19 +329,8 @@ Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITenso bool run_optimised_requantized = false; if(is_data_type_quantized_asymmetric(a_to_use->data_type()) && info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) { - // Result shifts < 0 are not supported by asm kernels - const std::vector<int32_t> &shifts = info.gemmlowp_output_stage().gemmlowp_shifts; - const bool is_asm_supported = info.gemmlowp_output_stage().gemmlowp_shift >= 0 - && std::all_of(shifts.cbegin(), shifts.cend(), [](int32_t val) - { - return val >= 0; - }); - - if(is_asm_supported) - { - run_optimised = bool(NEGEMMAssemblyDispatch::validate(a_to_use, b, c, output, gemm_info)); - run_optimised_requantized = run_optimised; - } + run_optimised = bool(NEGEMMAssemblyDispatch::validate(a_to_use, b, c, output, gemm_info)); + run_optimised_requantized = run_optimised; } else { |