Skip to content

fix: Preserve QASYMM8 output in GEMMLowp assembly fallback. #1307

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
morgolock wants to merge 1 commit into
base: main
Choose a base branch
from
+20 −11
Open

fix: Preserve QASYMM8 output in GEMMLowp assembly fallback. #1307

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 20 additions & 11 deletions src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -204,7 +204,9 @@ void CpuGemmLowpMatrixMultiplyCore::configure(
_mm_result_s32 = TensorInfo(dst->tensor_shape(), 1, DataType::S32);
}
#ifdef __aarch64__
ITensorInfo *dst_to_use = (_flip_signedness && _fuse_output_stage) ? &_signed_output : dst;
const bool use_signed_output_for_assembly =
_flip_signedness && (_fuse_output_stage || dst->data_type() == DataType::QASYMM8);
ITensorInfo *dst_to_use = use_signed_output_for_assembly ? &_signed_output : dst;

if (!(!b->are_values_constant() &&
b->tensor_shape().z() > 1)) // Disable batch matmul as optimized GeMM handles batching differently.
Expand All @@ -225,7 +227,7 @@ void CpuGemmLowpMatrixMultiplyCore::configure(
}
else
{
auto output_to_use = (_fuse_output_stage ? &_mm_result_s32 : dst);
auto output_to_use = (_fuse_output_stage ? &_mm_result_s32 : dst_to_use);
_asm_glue->configure(a_to_use, b, nullptr, output_to_use, asm_info);
}
_assembly_path = _asm_glue->is_configured();
Expand All @@ -239,6 +241,8 @@ void CpuGemmLowpMatrixMultiplyCore::configure(
}
}
#endif /* __aarch64__ */
const bool use_signed_output =
_flip_signedness && (_fuse_output_stage || (_assembly_path && dst->data_type() == DataType::QASYMM8));
if (!(_assembly_path || _run_vector_matrix_multiplication))
{
matrix_a = &_tmp_a;
Expand Down Expand Up @@ -295,7 +299,7 @@ void CpuGemmLowpMatrixMultiplyCore::configure(
std::make_unique<kernels::CpuGemmLowpOffsetContributionOutputStageKernel>();
_offset_contribution_output_stage_kernel->configure(
&_mm_result_s32, a_offset_kernel_needed ? &_vector_sum_col : nullptr,
b_offset_kernel_needed ? &_vector_sum_row : nullptr, c, _flip_signedness ? &_signed_output : dst,
b_offset_kernel_needed ? &_vector_sum_row : nullptr, c, use_signed_output ? &_signed_output : dst,
a->dimension(0), _a_offset, _b_offset, info.gemmlowp_output_stage());
}
else
Expand All @@ -318,7 +322,7 @@ void CpuGemmLowpMatrixMultiplyCore::configure(
a_to_use->dimension(0), _a_offset, _b_offset, dequantize_scale);
}
}
if (_fuse_output_stage && _flip_signedness)
if (use_signed_output)
{
_convert_from_signed_asymm = std::make_unique<kernels::CpuConvertQuantizedSignednessKernel>();
_convert_from_signed_asymm->configure(&_signed_output, dst);
Expand Down Expand Up @@ -477,7 +481,8 @@ Status CpuGemmLowpMatrixMultiplyCore::validate(const ITensorInfo *a,
// Update matrix a
matrix_a_info = &signed_a;
}
const ITensorInfo *output_to_use = (flip_signedness && fuse_output_stage) ? &signed_output : output;
const bool use_signed_output = flip_signedness && (fuse_output_stage || output->data_type() == DataType::QASYMM8);
const ITensorInfo *output_to_use = use_signed_output ? &signed_output : output;

// Offset kernel is need if offset is non-zero or it may change (i.e. dynamic).
bool a_offset_kernel_needed = a_offset != 0 || a->quantization_info().is_dynamic();
Expand All @@ -499,7 +504,7 @@ Status CpuGemmLowpMatrixMultiplyCore::validate(const ITensorInfo *a,
else
{
run_optimised = bool(CpuGemmAssemblyDispatch::validate(
a_to_use, b, nullptr, fuse_output_stage ? &mm_result_s32_info : output, asm_info));
a_to_use, b, nullptr, fuse_output_stage ? &mm_result_s32_info : output_to_use, asm_info));
}
}
auto validate_lowp_reductions = [&](const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *a_to_use,
Expand Down Expand Up @@ -622,7 +627,7 @@ Status CpuGemmLowpMatrixMultiplyCore::validate(const ITensorInfo *a,
// Validate offset contribution kernel
ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuGemmLowpOffsetContributionOutputStageKernel::validate(
&mm_result_s32_info, a_offset_kernel_needed ? &info_vector_sum_col : nullptr,
b_offset_kernel_needed ? &info_vector_sum_row : nullptr, c, flip_signedness ? &signed_output : output,
b_offset_kernel_needed ? &info_vector_sum_row : nullptr, c, use_signed_output ? &signed_output : output,
a_offset, b_offset, info.gemmlowp_output_stage()));
}
else
Expand Down Expand Up @@ -697,6 +702,9 @@ void CpuGemmLowpMatrixMultiplyCore::run(ITensorPack &tensors)
a_to_use = signed_a.get();
matrix_a = signed_a.get();
}
const bool use_signed_output =
_flip_signedness &&
(_fuse_output_stage || (_asm_glue->is_configured() && dst->info()->data_type() == DataType::QASYMM8));

// Run GEMM
if (_asm_glue->is_configured())
Expand All @@ -705,15 +713,16 @@ void CpuGemmLowpMatrixMultiplyCore::run(ITensorPack &tensors)
if (is_data_type_quantized_asymmetric(a_to_use->info()->data_type()) &&
_gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
{
auto output_to_use = (_flip_signedness && _fuse_output_stage) ? signed_output.get() : dst;
auto output_to_use = use_signed_output ? signed_output.get() : dst;
asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_0, a_to_use);
asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_1, b);
asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_2, c);
asm_glue_tensors.add_tensor(TensorType::ACL_DST, output_to_use);
}
else
{
auto output_to_use = (_fuse_output_stage ? mm_result_s32.get() : dst);
auto output_to_use =
(_fuse_output_stage ? mm_result_s32.get() : (use_signed_output ? signed_output.get() : dst));
asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_0, a_to_use);
asm_glue_tensors.add_const_tensor(TensorType::ACL_SRC_1, b);
asm_glue_tensors.add_tensor(TensorType::ACL_DST, output_to_use);
Expand Down Expand Up @@ -781,7 +790,7 @@ void CpuGemmLowpMatrixMultiplyCore::run(ITensorPack &tensors)
pack.add_tensor(TensorType::ACL_SRC_1, _a_offset == 0 ? nullptr : vector_sum_col.get());
pack.add_tensor(TensorType::ACL_SRC_2, _b_offset == 0 ? nullptr : vector_sum_row.get());
pack.add_tensor(TensorType::ACL_SRC_3, c);
pack.add_tensor(TensorType::ACL_DST, _flip_signedness ? signed_output.get() : dst);
pack.add_tensor(TensorType::ACL_DST, use_signed_output ? signed_output.get() : dst);

// Run offset contribution kernel
NEScheduler::get().schedule_op(_offset_contribution_output_stage_kernel.get(), Window::DimY,
Expand Down Expand Up @@ -811,7 +820,7 @@ void CpuGemmLowpMatrixMultiplyCore::run(ITensorPack &tensors)
}

// Convert QASYMM8_SIGNED->QASYMM8
if (_fuse_output_stage && _flip_signedness)
if (use_signed_output)
{
ITensorPack pack = {{TensorType::ACL_SRC, signed_output.get()}, {TensorType::ACL_DST, dst}};
NEScheduler::get().schedule_op(_convert_from_signed_asymm.get(), Window::DimY,
Expand Down