!5449 optimization for int8 packing

Merge pull request !5449 from lixian/master

mindspore/lite/nnacl/pack.c

@@ -487,9 +487,7 @@ void Im2ColPackUnitInt8Opt(const int8_t *input_data, int8_t *packed_input, int r
         for (int m = 0; m < ic4; m++) {
           int channel_block_stride = input_x_stride + m * C4NUM;
           int channel_block_offset = input_plane_offset + m * tile_num * C4NUM;
-          for (int k = 0; k < C4NUM; k++) {
-            (packed_input + channel_block_offset)[k] = (input_data + channel_block_stride)[k];
-          }
+          memcpy(packed_input + channel_block_offset, input_data + channel_block_stride, 4);
         }  // channel_block loop
       }    // kernel_w loop
     }      // kernel_h loop

mindspore/lite/src/runtime/kernel/arm/int8/convolution_1x1_int8.cc

@@ -68,7 +68,6 @@ void Convolution1x1Int8CPUKernel::CheckSupportOptimize() {
       matmul_func_ = nullptr;
     } else {
       support_optimize_ = true;
-      matmul_func_ = MatMulInt8_8x8_r;
     }
   } else {
     support_optimize_ = false;

mindspore/lite/src/runtime/kernel/arm/int8/convolution_int8.cc

@@ -400,9 +400,9 @@ kernel::LiteKernel *CpuConvInt8KernelCreator(const std::vector<lite::tensor::Ten
   kernel::LiteKernel *kernel;
   auto filter_quant_size = inputs[kWeightIndex]->GetQuantParams().size();
   if (kernel_h == 3 && kernel_w == 3 && stride_h == 1 && stride_w == 1 && dilation_h == 1 && dilation_w == 1) {
-    kernel = new (std::nothrow) kernel::Convolution3x3Int8CPUKernel(opParameter, inputs, outputs, ctx, primitive);
+    kernel = new (std::nothrow) kernel::ConvolutionInt8CPUKernel(opParameter, inputs, outputs, ctx, primitive);
   } else if (kernel_h == 1 && kernel_w == 1 && filter_quant_size == 1) {
-    kernel = new (std::nothrow) kernel::Convolution1x1Int8CPUKernel(opParameter, inputs, outputs, ctx, primitive);
+    kernel = new (std::nothrow) kernel::ConvolutionInt8CPUKernel(opParameter, inputs, outputs, ctx, primitive);
   } else {
     kernel = new (std::nothrow) kernel::ConvolutionInt8CPUKernel(opParameter, inputs, outputs, ctx, primitive);
   }