Optimal OPENCL conv_trans_3x3s2 Performance (#2015)

* add opencl  depthwise_conv_trans_op

* test=develop

* Optimal OPENCL conv_3x3S2 Performance

* test=develop

* Optimal OPENCL conv_trans_3x3s2 Performance , test=develop

mobile/src/operators/kernel/cl/cl-kernel-func/conv_func.cpp

@@ -638,5 +638,123 @@ void ConvTransposeAddBnRelu(framework::CLHelper *cl_helper,
   clEnqueueNDRangeKernel(cl_helper->CLCommandQueue(), kernel, 3, NULL,
                          work_size, NULL, 0, NULL, NULL);
 }
+
+void ConvTranspose3x3s2AddBnRelu(framework::CLHelper *cl_helper,
+                                 const ConvTransposeParam<GPU_CL> &param,
+                                 bool ifRelu, const framework::CLImage *biase,
+                                 const framework::CLImage *new_scale,
+                                 const framework::CLImage *new_bias) {
+  auto kernel = cl_helper->KernelAt(0);
+  auto default_work_size = cl_helper->DefaultWorkSize(*param.Output());
+  int c_block = default_work_size[0];
+  int w = default_work_size[1];
+  int nh = default_work_size[2];
+
+  int w_blk_size = 5;
+  int w_blk = (w + w_blk_size - 1 + 5) / w_blk_size / 2 * 2;
+  default_work_size[1] = w_blk;
+
+  int h_blk_size = 1;
+  int h_blk = (nh + h_blk_size - 1) / h_blk_size;
+  default_work_size[2] = h_blk;
+
+  auto input = param.Input()->GetCLImage();
+  auto filter = param.Filter()->GetCLImage();
+
+  auto output = param.Output()->GetCLImage();
+  int stride = param.Strides()[0];
+  int pad = param.Paddings()[0];
+  int dilation = param.Dilations()[0];
+
+  int input_channel = param.Input()->dims()[1];
+  int input_height = param.Input()->dims()[2];
+  int input_width = param.Input()->dims()[3];
+
+  int output_height = param.Output()->dims()[2];
+  int output_width = param.Output()->dims()[3];
+
+  int filter_height = param.Filter()->dims()[2];
+  int filter_width = param.Filter()->dims()[3];
+
+  cl_int status;
+  int index = 0;
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &c_block);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &w_blk);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &h_blk);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &input);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &filter);
+  CL_CHECK_ERRORS(status);
+
+  if (biase) {
+    auto bias_mem = biase->GetCLImage();
+    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &bias_mem);
+    CL_CHECK_ERRORS(status);
+  }
+
+  if (new_scale && new_bias) {
+    auto new_scale_mem = new_scale->GetCLImage();
+    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_scale_mem);
+    CL_CHECK_ERRORS(status);
+
+    auto new_bias_mem = new_bias->GetCLImage();
+    status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &new_bias_mem);
+    CL_CHECK_ERRORS(status);
+  }
+
+  status = clSetKernelArg(kernel, index++, sizeof(cl_mem), &output);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &stride);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &pad);
+  CL_CHECK_ERRORS(status);
+  status = clSetKernelArg(kernel, index++, sizeof(int), &dilation);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &input_channel);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &input_width);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &input_height);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &output_width);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &output_height);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &filter_width);
+  CL_CHECK_ERRORS(status);
+
+  status = clSetKernelArg(kernel, index++, sizeof(int), &filter_height);
+  CL_CHECK_ERRORS(status);
+
+  if (default_work_size.data()[1] % 60 == 0 && use_lws) {
+    const size_t local_work_size[3] = {static_cast<const uint32_t>(1),
+                                       static_cast<const uint32_t>(60),
+                                       static_cast<const uint32_t>(1)};
+    status = clEnqueueNDRangeKernel(
+        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
+        default_work_size.data(), local_work_size, 0, NULL, NULL);
+  } else {
+    status = clEnqueueNDRangeKernel(
+        cl_helper->CLCommandQueue(), kernel, default_work_size.size(), NULL,
+        default_work_size.data(), NULL, 0, NULL, NULL);
+  }
+  CL_CHECK_ERRORS(status);
+}
 }  // namespace operators
 }  // namespace paddle_mobile

mobile/src/operators/kernel/cl/cl-kernel-func/conv_func.h

@@ -64,6 +64,12 @@ void ConvTransposeAddBnRelu(framework::CLHelper *cl_helper,
                             const framework::CLImage *biase = nullptr,
                             const framework::CLImage *new_scale = nullptr,
                             const framework::CLImage *new_bias = nullptr);
+void ConvTranspose3x3s2AddBnRelu(framework::CLHelper *cl_helper,
+                                 const ConvTransposeParam<GPU_CL> &param,
+                                 bool ifRelu = false,
+                                 const framework::CLImage *biase = nullptr,
+                                 const framework::CLImage *new_scale = nullptr,
+                                 const framework::CLImage *new_bias = nullptr);
 
 }  // namespace operators
 }  // namespace paddle_mobile

mobile/src/operators/kernel/cl/cl_kernel/conv_transpose_kernel.cl

@@ -242,7 +242,202 @@ __read_only image2d_t new_scale,
 }
 
 
+/* batch == 1 pad(output) == 1 out_w % 2 == 0 */
+__kernel void conv_transpose3x3s2(__private const int item_ch,
+                                  __private const int item_w,
+                                  __private const int item_h,
+                                  __read_only image2d_t input_image,
+                                  __read_only image2d_t filter_image,
+#if defined(BIASE_CH) || defined(BIASE_ELE)
+        __read_only image2d_t bias,
+#endif
+#ifdef BATCH_NORM
+__read_only image2d_t new_scale,
+                                              __read_only image2d_t new_biase,
+#endif
+                                  __write_only image2d_t output_image,
+                                  __private const int stride,
+                                  __private const int pad,
+                                  __private const int dilation,
+                                  __private const int in_ch,
+                                  __private const int in_w,
+                                  __private const int in_h,
+                                  __private const int out_w,
+                                  __private const int out_h,
+                                  __private const int filter_w,
+                                  __private const int filter_h) {
+
+    const sampler_t sampler = CLK_NORMALIZED_COORDS_TRUE |
+                              CLK_ADDRESS_CLAMP          |
+                              CLK_FILTER_NEAREST;
+    // item_id
+    const int item_ch_id = get_global_id(0);
+    const int item_w_id = get_global_id(1);
+    const int item_h_id = get_global_id(2);
+
+    // out_id
+    int out_w_id_per_ch_blk = item_w_id / 2 * 10 + item_w_id % 2;
+    int out_h_id = item_h_id;
+    int out_w_id0 = item_ch_id * out_w + out_w_id_per_ch_blk;
+    int out_w_id1 = out_w_id0 + 2;
+    int out_w_id2 = out_w_id1 + 2;
+    int out_w_id3 = out_w_id2 + 2;
+    int out_w_id4 = out_w_id3 + 2;
+
+    // in_id
+    int in_w_id_per_ch_blk = (out_w_id_per_ch_blk) / 2;
+    in_w_id_per_ch_blk = in_w_id_per_ch_blk > 0 ? in_w_id_per_ch_blk : 0;
+    int in_h_id_per_batch = (out_h_id) / 2;
+    in_h_id_per_batch = in_h_id_per_batch > 0 ? in_h_id_per_batch : 0;
+
+    // filter_id
+    int align_w_i = out_w_id_per_ch_blk - 1;
+    int align_w = align_w_i % 2 > 0 ?
+                  align_w_i % 2 - 2 : align_w_i % 2;
+    int filter_w_id_per_ch_blk = out_w_id_per_ch_blk + 1 < 3 ? out_w_id_per_ch_blk + 1 : 2 + align_w;
+
+    int align_h_i = out_h_id - 1;
+    int align_h = align_h_i % 2 > 0 ?
+                  align_h_i % 2 - 2 : align_h_i % 2;
+    int filter_h_id_per_out_ch = out_h_id + 1 < 3 ? out_h_id + 1 : 2 + align_h;
 
+#ifdef BIASE_CH
+    half4 output[5];
+    output[0] = read_imageh(bias, sampler, (int2)(item_ch_id, 0));
+    output[1] = output[0];
+    output[2] = output[0];
+    output[3] = output[0];
+    output[4] = output[0];
+
+#elif defined(BIASE_ELE)
+    half4 output[5];
+    output[0] = read_imageh(bias, sampler, (int2)(out_w_id0, item_h_id));
+    if (out_w_id_per_ch_blk + 2 < out_w) {
+        output[1] = read_imageh(bias, sampler, (int2)(out_w_id1, item_h_id));
+    }
+    if (out_w_id_per_ch_blk + 4 < out_w) {
+        output[2] = read_imageh(bias, sampler, (int2)(out_w_id2, item_h_id));
+    }
+    if (out_w_id_per_ch_blk + 6 < out_w) {
+        output[3] = read_imageh(bias, sampler, (int2)(out_w_id3, item_h_id));
+    }
+    if (out_w_id_per_ch_blk + 8 < out_w) {
+        output[4] = read_imageh(bias, sampler, (int2)(out_w_id4, item_h_id));
+    }
 
+#else
+    half4 output[5] = {0.0f};
+#endif
+    half4 filter[4] = {0.0f};
+    half4 filter_trans[4] = {0.0f};
+
+    half4 input[5] = {0.0f};
+    for (int ch = 0; ch < (in_ch + 3) / 4; ch++) {
+        int filter_w_id = ch * 3;
+        int h_idx = 0;
+        for (int h = filter_h_id_per_out_ch; h >= 0; h -= 2) {
+            int in_h_id = select(in_h_id_per_batch + h_idx, -1,
+                                 in_h_id_per_batch + h_idx < 0 || in_h_id_per_batch + h_idx >= in_h);
+            int filter_h_id = item_ch_id * 12 + h;
+            int w_idx = 0;
+            for (int w = filter_w_id_per_ch_blk; w >= 0; w -= 2) {
+                int in_w_id0 = select(ch * in_w + in_w_id_per_ch_blk + w_idx, -1,
+                                     in_w_id_per_ch_blk + w_idx < 0 || in_w_id_per_ch_blk + w_idx >= in_w);
+                int in_w_id1 = select(ch * in_w + in_w_id_per_ch_blk + 1 + w_idx, -1,
+                                     in_w_id_per_ch_blk + 1 + w_idx < 0 || in_w_id_per_ch_blk + 1 + w_idx >= in_w);
+                int in_w_id2 = select(ch * in_w + in_w_id_per_ch_blk + 2 + w_idx, -1,
+                                     in_w_id_per_ch_blk + 2 + w_idx < 0 || in_w_id_per_ch_blk + 2 + w_idx >= in_w);
+                int in_w_id3 = select(ch * in_w + in_w_id_per_ch_blk + 3 + w_idx, -1,
+                                     in_w_id_per_ch_blk + 3 + w_idx < 0 || in_w_id_per_ch_blk + 3 + w_idx >= in_w);
+                int in_w_id4 = select(ch * in_w + in_w_id_per_ch_blk + 4 + w_idx, -1,
+                                     in_w_id_per_ch_blk + 4 + w_idx < 0 || in_w_id_per_ch_blk + 4 + w_idx >= in_w);
+
+                input[0] = read_imageh(input_image, sampler, (int2)(in_w_id0, in_h_id));
+                input[1] = read_imageh(input_image, sampler, (int2)(in_w_id1, in_h_id));
+                input[2] = read_imageh(input_image, sampler, (int2)(in_w_id2, in_h_id));
+                input[3] = read_imageh(input_image, sampler, (int2)(in_w_id3, in_h_id));
+                input[4] = read_imageh(input_image, sampler, (int2)(in_w_id4, in_h_id));
+
+                filter[0] = read_imageh(filter_image, sampler, (int2)(filter_w_id + w, filter_h_id));      // in_ch:0-3,out_ch:0
+                filter[1] = read_imageh(filter_image, sampler, (int2)(filter_w_id + w, filter_h_id + 3));  // in_ch:0-3,out_ch:1
+                filter[2] = read_imageh(filter_image, sampler, (int2)(filter_w_id + w, filter_h_id + 6));  // in_ch:0-3,out_ch:2
+                filter[3] = read_imageh(filter_image, sampler, (int2)(filter_w_id + w, filter_h_id + 9));  // in_ch:0-3,out_ch:3
+
+                filter_trans[0] = (half4)(filter[0].x, filter[1].x, filter[2].x, filter[3].x);             // in_ch:0,out_ch:0-3
+                filter_trans[1] = (half4)(filter[0].y, filter[1].y, filter[2].y, filter[3].y);             // in_ch:1,out_ch:0-3
+                filter_trans[2] = (half4)(filter[0].z, filter[1].z, filter[2].z, filter[3].z);             // in_ch:2,out_ch:0-3
+                filter_trans[3] = (half4)(filter[0].w, filter[1].w, filter[2].w, filter[3].w);             // in_ch:3,out_ch:0-3
+
+                output[0] = mad(input[0].x, filter_trans[0], output[0]);
+                output[0] = mad(input[0].y, filter_trans[1], output[0]);
+                output[0] = mad(input[0].z, filter_trans[2], output[0]);
+                output[0] = mad(input[0].w, filter_trans[3], output[0]);
+
+                output[1] = mad(input[1].x, filter_trans[0], output[1]);
+                output[1] = mad(input[1].y, filter_trans[1], output[1]);
+                output[1] = mad(input[1].z, filter_trans[2], output[1]);
+                output[1] = mad(input[1].w, filter_trans[3], output[1]);
+
+                output[2] = mad(input[2].x, filter_trans[0], output[2]);
+                output[2] = mad(input[2].y, filter_trans[1], output[2]);
+                output[2] = mad(input[2].z, filter_trans[2], output[2]);
+                output[2] = mad(input[2].w, filter_trans[3], output[2]);
+
+                output[3] = mad(input[3].x, filter_trans[0], output[3]);
+                output[3] = mad(input[3].y, filter_trans[1], output[3]);
+                output[3] = mad(input[3].z, filter_trans[2], output[3]);
+                output[3] = mad(input[3].w, filter_trans[3], output[3]);
+
+                output[4] = mad(input[4].x, filter_trans[0], output[4]);
+                output[4] = mad(input[4].y, filter_trans[1], output[4]);
+                output[4] = mad(input[4].z, filter_trans[2], output[4]);
+                output[4] = mad(input[4].w, filter_trans[3], output[4]);
+                w_idx++;
+            }
+            h_idx++;
+        }
+    }
+#ifdef BATCH_NORM
+    half4 scale = read_imageh(new_scale, sampler, (int2)(item_ch_id, 0));
+    half4 biase = read_imageh(new_biase, sampler, (int2)(item_ch_id, 0));
+    output[0] = mad(scale, output[0], biase);
+    if (out_w_id_per_ch_blk + 2 < out_w) {
+        output[1] = mad(scale, output[1], biase);
+    }
+    if (out_w_id_per_ch_blk + 4 < out_w) {
+        output[2] = mad(scale, output[2], biase);
+    }
+    if (out_w_id_per_ch_blk + 6 < out_w) {
+        output[3] = mad(scale, output[3], biase);
+    }
+    if (out_w_id_per_ch_blk + 8 < out_w) {
+        output[4] = mad(scale, output[4], biase);
+    }
+#endif
+
+#ifdef RELU
+    output[0] = activation(output[0]);
+    output[1] = activation(output[1]);
+    output[2] = activation(output[2]);
+    output[3] = activation(output[3]);
+    output[4] = activation(output[4]);
+
+#endif
+
+    write_imageh(output_image, (int2)(out_w_id0, item_h_id), output[0]);
+
+    if (out_w_id_per_ch_blk + 2 < out_w) {
+        write_imageh(output_image, (int2)(out_w_id1, item_h_id), output[1]);
+    }
+    if (out_w_id_per_ch_blk + 4 < out_w) {
+        write_imageh(output_image, (int2)(out_w_id2, item_h_id), output[2]);
+    }
+    if (out_w_id_per_ch_blk + 6 < out_w) {
+        write_imageh(output_image, (int2)(out_w_id3, item_h_id), output[3]);
+    }
+    if (out_w_id_per_ch_blk + 8 < out_w) {
+        write_imageh(output_image, (int2)(out_w_id4, item_h_id), output[4]);
+    }
+}
 
 

mobile/src/operators/kernel/cl/conv_transpose_kernel.cpp

@@ -40,7 +40,8 @@ bool ConvTransposeKernel<GPU_CL, float>::Init(
     param->ExecMode() = ConvTransposeParam<GPU_CL>::EXEC_CONVTRANS3x3s2_FLOAT;
     param->Filter()->InitConv2dTransposeFilterCLImage(
         cl_helper_.CLContext(), cl_helper_.CLCommandQueue());
-    this->cl_helper_.AddKernel("conv_transpose", "conv_transpose_kernel.cl");
+    this->cl_helper_.AddKernel("conv_transpose3x3s2",
+                               "conv_transpose_kernel.cl");
   } else {
     PADDLE_MOBILE_THROW_EXCEPTION(" not support ");
   }
@@ -55,7 +56,7 @@ void ConvTransposeKernel<GPU_CL, float>::Compute(
       DWConvTransposeAddBnRelu(&this->cl_helper_, param);
       break;
     case ConvTransposeParam<GPU_CL>::EXEC_CONVTRANS3x3s2_FLOAT:
-      ConvTransposeAddBnRelu(&this->cl_helper_, param);
+      ConvTranspose3x3s2AddBnRelu(&this->cl_helper_, param);
       break;
     default:
       PADDLE_MOBILE_THROW_EXCEPTION(