Neon depthwise conv with filterSize = 4 and stride = 2.

paddle/function/neon/NeonDepthwiseConv.cpp

@@ -364,6 +364,116 @@ struct DepthwiseConvKernel<4, 1> {
   }
 };
 
+/**
+ * Each step calculates four elements of the output.
+ */
+template <>
+struct DepthwiseConvKernel<4, 2> {
+  static void run(const float* inputData,
+                  const float* filterData,
+                  int inputHeight,
+                  int inputWidth,
+                  int outputChannels,
+                  int outputHeight,
+                  int outputWidth,
+                  int filterMultiplier,
+                  float* outputData) {
+    const int steps = outputWidth >> 2;
+    const int remain = outputWidth & 3;
+    for (int c = 0; c < outputChannels; c++, filterData += 16) {
+      // Load the filters
+      float32x4_t k[4];
+      k[0] = vld1q_f32(filterData);
+      k[1] = vld1q_f32(filterData + 4);
+      k[2] = vld1q_f32(filterData + 8);
+      k[3] = vld1q_f32(filterData + 12);
+
+      const float* start =
+          inputData + (c / filterMultiplier) * (inputHeight * inputWidth);
+      float32x4_t input[4][4];
+      for (int h = 0; h < outputHeight; h++) {
+        const float* r0 = start + 2 * h * inputWidth;
+        const float* r1 = start + (2 * h + 1) * inputWidth;
+        const float* r2 = start + (2 * h + 2) * inputWidth;
+        const float* r3 = start + (2 * h + 3) * inputWidth;
+        for (int s = 0; s < steps; s++) {
+          // Load the inputs
+          float32x4x2_t data1;
+          float32x4x2_t data2;
+
+          data1 = vld2q_f32(r0);
+          data2 = vld2q_f32(r0 + 8);
+          input[0][0] = data1.val[0];
+          input[0][1] = data1.val[1];
+          input[0][2] = vextq_f32(data1.val[0], data2.val[0], 1);
+          input[0][3] = vextq_f32(data1.val[1], data2.val[1], 1);
+
+          data1 = vld2q_f32(r1);
+          data2 = vld2q_f32(r1 + 8);
+          input[1][0] = data1.val[0];
+          input[1][1] = data1.val[1];
+          input[1][2] = vextq_f32(data1.val[0], data2.val[0], 1);
+          input[1][3] = vextq_f32(data1.val[1], data2.val[1], 1);
+
+          data1 = vld2q_f32(r2);
+          data2 = vld2q_f32(r2 + 8);
+          input[2][0] = data1.val[0];
+          input[2][1] = data1.val[1];
+          input[2][2] = vextq_f32(data1.val[0], data2.val[0], 1);
+          input[2][3] = vextq_f32(data1.val[1], data2.val[1], 1);
+
+          data1 = vld2q_f32(r3);
+          data2 = vld2q_f32(r3 + 8);
+          input[3][0] = data1.val[0];
+          input[3][1] = data1.val[1];
+          input[3][2] = vextq_f32(data1.val[0], data2.val[0], 1);
+          input[3][3] = vextq_f32(data1.val[1], data2.val[1], 1);
+
+          float32x4_t tmp1 = vdupq_n_f32(0.f);
+          float32x4_t tmp2 = vdupq_n_f32(0.f);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1);
+          tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2);
+          tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3);
+          tmp1 = vaddq_f32(tmp1, tmp2);
+
+          vst1q_f32(outputData, tmp1);
+          r0 += 8;
+          r1 += 8;
+          r2 += 8;
+          r3 += 8;
+          outputData += 4;
+        }
+
+        for (int r = 0; r < remain; r++) {
+          float32x4_t i0 = vld1q_f32(r0);
+          float32x4_t i1 = vld1q_f32(r1);
+          float32x4_t i2 = vld1q_f32(r2);
+          float32x4_t i3 = vld1q_f32(r3);
+          *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]);
+          r0 += 2;
+          r1 += 2;
+          r2 += 2;
+          r3 += 2;
+          outputData++;
+        }
+      }
+    }
+  }
+};
+
 template <DeviceType Device>
 class NeonDepthwiseConvFunction : public ConvFunctionBase {
 public:
@@ -449,7 +559,7 @@ public:
                                        outputWidth,
                                        filterMultiplier,
                                        outputData);
-      } else if (filterWidth == 4) {
+      } else if (filterWidth == 4 && strideH() == 1) {
         DepthwiseConvKernel<4, 1>::run(inputPadding,
                                        filterData,
                                        inputHeight,
@@ -459,6 +569,16 @@ public:
                                        outputWidth,
                                        filterMultiplier,
                                        outputData);
+      } else if (filterWidth == 4 && strideH() == 2) {
+        DepthwiseConvKernel<4, 2>::run(inputPadding,
+                                       filterData,
+                                       inputHeight,
+                                       inputWidth,
+                                       outputChannels,
+                                       outputHeight,
+                                       outputWidth,
+                                       filterMultiplier,
+                                       outputData);
       }
 
       inputPadding += inputChannels * inputHeight * inputWidth;