Q8GEMM per-channel quant 32bit/16bit accumulation

[mortzur]

Micro-kernels implementation of the following:

• Q8GEMM with per-channel weights quantization parameters (.c) + unit tests + benchmarks

• Q8GEMM with per-channel weights quantization parameters for AARCH32 (.S) + unit tests + benchmarks

• Q8GEMM with per-channel weights quantization parameters with 16bit opportunistic accumulation (.c) + unit tests + benchmarks

• Q8GEMM with per-channel weights quantization parameters with 16bit opportunistic accumulation for AARCH32 (.S) + unit tests + benchmarks

[franksun007]

I believe the following patches are missing from the CMakeList.txt

diff --git a/CMakeLists.txt b/CMakeLists.txt
index a5ddc49..6320b1e 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -188,11 +188,15 @@ SET(QNNPACK_AARCH32_ASM_UKERNELS
   src/q8conv/4x8-aarch32-neon.S
   src/q8dwconv/up8x9-aarch32-neon.S
   src/q8gemm/4x8-aarch32-neon.S
-  src/q8gemm/4x8c2-xzp-aarch32-neon.S)
+  src/q8gemm/4x8c2-xzp-aarch32-neon.S
+  src/q8gemm/4x8-aarch32-neon-per-channel.S
+  src/q8gemm/4x8-aarch32-neon-per-channel-16bitAcc.S)
 
 SET(QNNPACK_AARCH64_ASM_UKERNELS
   src/q8conv/8x8-aarch64-neon.S
-  src/q8gemm/8x8-aarch64-neon.S)
+  src/q8gemm/8x8-aarch64-neon.S
+  src/q8gemm/4x8-neon_per_channel.c
+  src/q8gemm/4x8-neon_per_channel_16bitAcc.c)
 
 SET(QNNPACK_X86_SSE2_UKERNELS
   src/q8avgpool/mp8x9p8q-sse2.c

[franksun007]

Also, test and benchmarks failed to compile on ARM32 platform. This might be an easy fix with if guard.

Sorry, my bad. Only the benchmark is not compiling correctly.