use sgemm fp32 on arm platform,optimize conv1x1s2 (#1031)

src/layer/arm/convolution_arm.cpp

@@ -29,6 +29,7 @@ namespace ncnn {
 #include "convolution_4x4.h"
 #include "convolution_5x5.h"
 #include "convolution_7x7.h"
+#include "convolution_sgemm.h"
 #include "convolution_sgemm_int8.h"
 #include "convolution_1x1_int8.h"
 #include "convolution_3x3_int8.h"
@@ -79,7 +80,7 @@ int Convolution_arm::create_pipeline(const Option& opt)
     if (activation)
     {
         Option opt_cpu = opt;
-        opt_cpu.vulkan_compute = false;
+        opt_cpu.use_vulkan_compute = false;
         activation->create_pipeline(opt_cpu);
     }
 
@@ -155,6 +156,13 @@ int Convolution_arm::create_pipeline(const Option& opt)
         conv3x3s2_transform_kernel_neon(weight_data, weight_3x3s2_data, num_input, num_output);
     }
 
+    {
+        int kernel_size = kernel_w * kernel_h;
+        int num_input = weight_data_size / kernel_size / num_output;
+
+        conv_im2col_sgemm_transform_kernel_neon(weight_data, weight_sgemm_data, num_input, num_output, kernel_size);
+    }    
+
     return 0;
 }
 
@@ -163,7 +171,7 @@ int Convolution_arm::destroy_pipeline(const Option& opt)
     if (activation)
     {
         Option opt_cpu = opt;
-        opt_cpu.vulkan_compute = false;
+        opt_cpu.use_vulkan_compute = false;
         activation->destroy_pipeline(opt_cpu);
         delete activation;
         activation = 0;
@@ -581,10 +589,17 @@ int Convolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option
     {
         conv1x1s1_sgemm_neon(bottom_blob_bordered, top_blob, weight_1x1_sgemm_data, bias_data, opt);
     }
-    else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
+    else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
     {
-        conv3x3s2_packed_neon(bottom_blob_bordered, top_blob, weight_3x3s2_data, bias_data, opt);
+        conv_im2col_sgemm_neon(bottom_blob_bordered, top_blob, weight_sgemm_data, bias_data, kernel_w, kernel_h, stride_w, stride_h, opt);
     }
+    else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
+    {
+        if (outw >=8 && outh >=8)
+            conv3x3s2_packed_neon(bottom_blob_bordered, top_blob, weight_3x3s2_data, bias_data, opt);
+        else
+            conv_im2col_sgemm_neon(bottom_blob_bordered, top_blob, weight_sgemm_data, bias_data, kernel_w, kernel_h, stride_w, stride_h, opt);
+    }     
     else
         conv(bottom_blob_bordered, top_blob, weight_data, bias_data, opt);
 

src/layer/arm/convolution_arm.h

@@ -43,6 +43,7 @@ public:
     Mat weight_1x1s1_sgemm_int8_data;
     Mat weight_3x3_winograd23_data;
     Mat weight_sgemm_int8_data;
+    Mat weight_sgemm_data;
     std::vector<Mat> weight_3x3_winograd23_int8_data;
 };
 

src/layer/arm/convolutiondepthwise_arm.cpp

@@ -37,7 +37,7 @@ ConvolutionDepthWise_arm::ConvolutionDepthWise_arm()
 int ConvolutionDepthWise_arm::create_pipeline(const Option& opt)
 {
     Option opt_cpu = opt;
-    opt_cpu.vulkan_compute = false;
+    opt_cpu.use_vulkan_compute = false;
 
     if (activation_type == 1)
     {
@@ -175,7 +175,7 @@ int ConvolutionDepthWise_arm::create_pipeline(const Option& opt)
 int ConvolutionDepthWise_arm::destroy_pipeline(const Option& opt)
 {
     Option opt_cpu = opt;
-    opt_cpu.vulkan_compute = false;
+    opt_cpu.use_vulkan_compute = false;
 
     if (activation)
     {

src/layer/arm/deconvolution_arm.cpp

@@ -68,7 +68,7 @@ int Deconvolution_arm::create_pipeline(const Option& opt)
     if (activation)
     {
         Option opt_cpu = opt;
-        opt_cpu.vulkan_compute = false;
+        opt_cpu.use_vulkan_compute = false;
         activation->create_pipeline(opt_cpu);
     }
 
@@ -80,7 +80,7 @@ int Deconvolution_arm::destroy_pipeline(const Option& opt)
     if (activation)
     {
         Option opt_cpu = opt;
-        opt_cpu.vulkan_compute = false;
+        opt_cpu.use_vulkan_compute = false;
         activation->destroy_pipeline(opt_cpu);
         delete activation;
         activation = 0;

src/layer/arm/deconvolutiondepthwise_arm.cpp

@@ -65,7 +65,7 @@ int DeconvolutionDepthWise_arm::create_pipeline(const Option& opt)
     if (activation)
     {
         Option opt_cpu = opt;
-        opt_cpu.vulkan_compute = false;
+        opt_cpu.use_vulkan_compute = false;
         activation->create_pipeline(opt_cpu);
     }
 
@@ -77,7 +77,7 @@ int DeconvolutionDepthWise_arm::destroy_pipeline(const Option& opt)
     if (activation)
     {
         Option opt_cpu = opt;
-        opt_cpu.vulkan_compute = false;
+        opt_cpu.use_vulkan_compute = false;
         activation->destroy_pipeline(opt_cpu);
         delete activation;
         activation = 0;

src/net.cpp

@@ -832,18 +832,38 @@ int Net::load_model(const unsigned char* _mem)
     return mem - _mem;
 }
 
-void Net::fuse_network()
+int Net::fuse_network()
 {
     // set the int8 op fusion:requantize
 #if NCNN_STRING && NCNN_REQUANT    
     // fprintf(stderr, "Test op fusion to int8 implement:\n");
+    // parse the network whether is a quantization model
+    bool net_quantized = false;
+    for (size_t i=0; i<layers.size(); i++)
+    {
+        Layer* layer = layers[i];
+        if (layer->type == "Convolution" || layer->type == "ConvolutionDepthWise")
+        {
+            if (layer->type == "Convolution" && (((Convolution*)layer)->use_int8_inference == false))
+                continue;
+            if (layer->type == "ConvolutionDepthWise" && (((ConvolutionDepthWise*)layer)->use_int8_inference == false))
+                continue;    
+            net_quantized = true;
+        }
+    }
+
+    if (net_quantized == false)
+        return 0;
+
     for (size_t i=0; i<layers.size(); i++)
     {
         Layer* layer = layers[i];
 
         if (layer->type == "Convolution" || layer->type == "ConvolutionDepthWise")
         {
-            if (((Convolution*)layer)->use_int8_inference == false)
+            if (layer->type == "Convolution" && (((Convolution*)layer)->use_int8_inference == false))
+                continue;
+            if (layer->type == "ConvolutionDepthWise" && (((ConvolutionDepthWise*)layer)->use_int8_inference == false))
                 continue;
 
             for (size_t n=0; n<blobs[layer->tops[0]].consumers.size(); n++)
@@ -858,6 +878,11 @@ void Net::fuse_network()
 
                     if (layer_next_2->type == "Convolution" || layer_next_2->type == "ConvolutionDepthWise")
                     {
+                        if (layer_next_2->type == "Convolution" && ((Convolution*)layer_next_2)->use_int8_inference == false)
+                            continue;
+                        if (layer_next_2->type == "ConvolutionDepthWise" && ((ConvolutionDepthWise*)layer_next_2)->use_int8_inference == false)
+                            continue;    
+
                         // fprintf(stderr, "%s, %s, %s\n", layer->name.c_str(), layer_next->name.c_str(), layer_next_2->name.c_str());
                         if (layer->type == "Convolution" && layer_next_2->type == "Convolution")
                         {
@@ -934,6 +959,7 @@ void Net::fuse_network()
         }
     }
 #endif
+    return 0;
 }
 
 void Net::clear()

src/net.h

@@ -96,7 +96,7 @@ public:
 protected:
     // parse the structure of network
     // fuse int8 op dequantize and quantize by requantize
-    void fuse_network();
+    int fuse_network();
 
 #if NCNN_VULKAN