COMPMID-2789: Add support for QASYMM8_SIGNED in CLGEMMDeconvolutionLayer
Change-Id: I7e3bcb01025e827f6f62491749c691c205ee7481
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2844
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index a2a47d7..8e7db93 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -2166,7 +2166,7 @@
* @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor
- * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8
+ * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8/QASYMM8_SIGNED
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
diff --git a/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
index 14bda11..3298858 100644
--- a/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMDeconvolutionLayer.cpp
@@ -62,6 +62,33 @@
return { start, end };
}
+Status construct_gemmlowp_output_stage(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, GEMMLowpOutputStageInfo &output_stage_info)
+{
+ const auto data_type = input->data_type();
+
+ if(is_data_type_quantized_asymmetric(data_type))
+ {
+ const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
+ const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
+ const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
+
+ float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
+ int output_multiplier(0);
+ int output_shift(0);
+ ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
+
+ output_stage_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
+ output_stage_info.gemmlowp_multiplier = output_multiplier;
+ output_stage_info.gemmlowp_shift = output_shift;
+ output_stage_info.gemmlowp_offset = oq_info.offset;
+ const auto min_max_bound = get_min_max(data_type);
+ output_stage_info.gemmlowp_min_bound = (std::get<0>(min_max_bound)).get<int32_t>();
+ output_stage_info.gemmlowp_max_bound = (std::get<1>(min_max_bound)).get<int32_t>();
+ output_stage_info.output_data_type = data_type;
+ }
+ return Status{};
+}
+
} // namespace
CLGEMMDeconvolutionLayer::CLGEMMDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
@@ -93,7 +120,7 @@
Status CLGEMMDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &deconv_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QASYMM8);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights);
@@ -141,10 +168,14 @@
TensorInfo gemm_output_info = reshaped_t_info.clone()->set_tensor_shape(gemm_output_shape).set_is_resizable(true);
GEMMInfo gemm_info(false, false, true, input->dimension(idx_h), true);
+ GEMMLowpOutputStageInfo output_stage_info;
+
if(is_quantized)
{
ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyCore::validate(&input->clone()->set_tensor_shape(nhwc_input_shape), &reshaped_t_info, nullptr, &gemm_output_info.set_data_type(DataType::S32),
gemm_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(construct_gemmlowp_output_stage(input, weights, output, output_stage_info));
+
}
else
{
@@ -160,9 +191,8 @@
{
const auto start_end = compute_start_end_slice_coordinates(col2im_output_info, deconv_info, is_nchw);
ARM_COMPUTE_RETURN_ON_ERROR(CLDeconvolutionReshapeOutputKernel::validate(&gemm_output_info, bias, &col2im_output_info, input, weights, deconv_info));
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(&col2im_output_info, nullptr,
- &col2im_output_info.clone()->set_is_resizable(true).set_data_type(DataType::QASYMM8)));
- ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&col2im_output_info.clone()->set_is_resizable(true).set_data_type(DataType::QASYMM8), output, start_end.first, start_end.second));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOutputStage::validate(&col2im_output_info, nullptr, &col2im_output_info.clone()->set_is_resizable(true).set_data_type(input->data_type()), output_stage_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&col2im_output_info.clone()->set_is_resizable(true).set_data_type(input->data_type()), output, start_end.first, start_end.second));
}
else if(padded_input)
{
@@ -173,16 +203,7 @@
else if(is_quantized)
{
ARM_COMPUTE_RETURN_ON_ERROR(CLDeconvolutionReshapeOutputKernel::validate(&gemm_output_info, bias, &col2im_output_info, input, weights, deconv_info));
-
- const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
- const UniformQuantizationInfo wq_info = weights->quantization_info().uniform();
- const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
-
- float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
- int output_multiplier(0);
- int output_shift(0);
- ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(&col2im_output_info, nullptr, output));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOutputStage::validate(&col2im_output_info, nullptr, output, output_stage_info));
}
else
{
@@ -297,15 +318,9 @@
if(_is_quantized)
{
- const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform();
- const UniformQuantizationInfo wq_info = weights->info()->quantization_info().uniform();
- const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform();
-
- float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
- int output_multiplier(0);
- int output_shift(0);
- quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
- _gemmlowp_output_stage.configure(&_gemmlowp_final, nullptr, output_stage_output, output_multiplier, output_shift, oq_info.offset);
+ GEMMLowpOutputStageInfo output_stage_info;
+ construct_gemmlowp_output_stage(input->info(), weights->info(), output->info(), output_stage_info);
+ _gemmlowp_output_stage.configure(&_gemmlowp_final, nullptr, output_stage_output, output_stage_info);
_gemmlowp_final.allocator()->allocate();
}
diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index 54b63df..9346e93 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -142,6 +142,7 @@
}
// Pick up the GEMM configuration
+ // Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED doesn't matter, since it only affect the shape configuration
std::tie(lhs_info, rhs_info) = CLGEMMReshapedOnlyRHSKernelConfigurationFactory::create(gpu_target)->configure(m, n, k, batch_size, DataType::QASYMM8);
// Configure reshape RHS kernel
@@ -570,4 +571,4 @@
_is_prepared = true;
}
}
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
index 9551fc7..de00fd2 100644
--- a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
@@ -105,4 +105,48 @@
return CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(input, bias, output, min, max);
}
+void CLGEMMLowpOutputStage::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const GEMMLowpOutputStageInfo &info)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+ ARM_COMPUTE_ERROR_ON(info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT);
+
+ switch(info.output_data_type)
+ {
+ case DataType::QASYMM8:
+ {
+ auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel>();
+ k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+ _kernel = std::move(k);
+ break;
+ }
+ case DataType::QASYMM8_SIGNED:
+ {
+ auto k = arm_compute::support::cpp14::make_unique<CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel>();
+ k->configure(input, bias, output, info.gemmlowp_multiplier, info.gemmlowp_shift, info.gemmlowp_offset, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+ _kernel = std::move(k);
+ break;
+ }
+ default:
+ ARM_COMPUTE_ERROR("Unsupported output data type.");
+ }
+
+}
+
+Status CLGEMMLowpOutputStage::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo &info)
+{
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
+ ARM_COMPUTE_RETURN_ERROR_ON(info.type != GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT);
+
+ switch(output->data_type())
+ {
+ case DataType::QASYMM8:
+ return CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+ case DataType::QASYMM8_SIGNED:
+ return CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(input, bias, output, info.gemmlowp_min_bound, info.gemmlowp_max_bound);
+ default:
+ return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported output data type.");
+ }
+
+}
} // namespace arm_compute
\ No newline at end of file