COMPMID-1413 - Improve the performance of GEMMLowp with 8 bit dot product on OpenCL
COMPMID-1424 - Add dot product support for CLDepthwise QASYMM8 3x3 NHWC non-unit stride
With this patch we are able to improve the performance of MobileNet v1-qasymm8 by 37 %
Tried to use the dot product instruction in CLDepthwise QASYMM8 3x3 NHWC non-unit stride
but I have not seen any benefit (maybe because we have few arithemtic operation and we
do not have more load instructions). However Depthwise convolution has been improved by
30%
Change-Id: Id768a99c2e53a04276707e427af5d0ec93419ada
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/155082
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
index 60f6294..45e21b5 100644
--- a/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
+++ b/src/runtime/NEON/functions/NEFullyConnectedLayer.cpp
@@ -50,6 +50,7 @@
// Validate gemmlowp function
ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(&input.clone()->set_quantization_info(input_quantization_info),
&weights.clone()->set_quantization_info(weights_quantization_info),
+ nullptr,
&output));
}
else
@@ -93,7 +94,7 @@
weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
// Configure gemmlowp function
- _mm_gemmlowp.configure(input, weights, output);
+ _mm_gemmlowp.configure(input, weights, nullptr, output);
// Revert back QuantizatioInfo as input and weights could be used in other fully connected layers
input->info()->set_quantization_info(input_quantization_info);