commit 01164235b4da4b5b7774b1336dd370261f866bce
author Oscar Andersson <oscar.andersson@arm.com> Mon Aug 02 16:19:24 2021 +0200
committer felix-johnny <48442848+felix-johnny@users.noreply.github.com> Fri Aug 13 13:07:23 2021 +0100
tree 5bfb07758eb832cf3204df357428f7b8dc7cb321
parent 061f747ca0b5f595e9d93f889fc76a5a481c0883

CMSIS-NN: Optimized convolution for 16x8 data type

Authors: Sebastian-Larsson, oscarandersson8218 (GitHub)

1. Optimized function is added for input data type of int16
and filter data type of int8. The main constraint is
that the kernel volume is < 512.

2. Unit tests are added as well.

Change-Id: I9f0f3067a275c8bf5f9263ebe76cd9a1033c7793

ARM.CMSIS.pdsc

diff --git a/ARM.CMSIS.pdsc b/ARM.CMSIS.pdsc
index 137f663..eef3f2e 100644
--- a/ARM.CMSIS.pdsc
+++ b/ARM.CMSIS.pdsc
@@ -11,8 +11,9 @@
     <release version="5.8.1">
       Active development ...
       CMSIS-DSP: 1.10.0 (see revision history for details)
-      CMSIS-NN: 3.0.1 (see revision history for details)
-       - Support for int16
+      CMSIS-NN: 3.1.0 (see revision history for details)
+       - Support for int16 convolution
+       - Support for DSP extension optimization for int16 convolution
     </release>
     <release version="5.8.0" date="2021-06-24">
       CMSIS-Core(M): 5.5.0 (see revision history for details)
@@ -2815,6 +2816,7 @@
         <file category="source" name="CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c"/>
         <file category="source" name="CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_s8_fast.c"/>
         <file category="source" name="CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8.c"/>
+        <file category="source" name="CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_fast_s16.c"/>
         <file category="source" name="CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c"/>
         <file category="source" name="CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_s8.c"/>
         <file category="source" name="CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_s16.c"/>
@@ -2853,6 +2855,7 @@
         <file category="source" name="CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_with_offset.c"/>
         <file category="source" name="CMSIS/NN/Source/NNSupportFunctions/arm_nn_accumulate_q7_to_q15.c"/>
         <file category="source" name="CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c"/>
+        <file category="source" name="CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_kernel_s16.c"/>
         <file category="source" name="CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_padded_s8.c"/>
         <file category="source" name="CMSIS/NN/Source/NNSupportFunctions/arm_nn_add_q7.c"/>
         <file category="source" name="CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_4x_s8.c"/>