From ec4dee8c68a3d0f6d63db184bfb2f4589429778e Mon Sep 17 00:00:00 2001 From: morgolock Date: Tue, 1 Sep 2020 11:14:27 +0100 Subject: COMPMID-3750: Disable asm kernels when shifts are negative. Change-Id: I65a738221a6c6fc3527ececda42f7a7e547755c1 Signed-off-by: morgolock Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3896 Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- .../functions/NEGEMMLowpMatrixMultiplyCore.cpp | 29 +++++++++++++++++++--- 1 file changed, 25 insertions(+), 4 deletions(-) (limited to 'src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp') diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp index 83db146a8a..dada6d16da 100644 --- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp +++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp @@ -117,8 +117,18 @@ 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) { - _asm_glue.configure(a_to_use, b, c, output, gemm_info); - _fused_assembly_path = _asm_glue.is_configured(); + // Result shifts < 0 are not supported by asm kernels + const std::vector &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(); + } } else { @@ -329,8 +339,19 @@ 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) { - run_optimised = bool(NEGEMMAssemblyDispatch::validate(a_to_use, b, c, output, gemm_info)); - run_optimised_requantized = run_optimised; + // Result shifts < 0 are not supported by asm kernels + const std::vector &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; + } } else { -- cgit v1.2.1