!4135 [MS][LITE][Develop]opencl: image2d for pooling

Merge pull request !4135 from chenzhongming/new_master

mindspore/lite/src/runtime/kernel/opencl/cl/fp32/max_pool2d.cl

-__kernel void MaxPooling2d(__global float4 *input, __global float4 *output, const int4 input_shape,
-                           const int4 output_shape, const int2 stride, const int2 kernel_size, const int2 padding) {
+__kernel void MaxPooling2d_BUF(__global float4 *input, __global float4 *output, const int4 input_shape,
+                               const int4 output_shape, const int2 stride, const int2 kernel_size, const int2 padding) {
   // axis to dst tensor coordinate
   int X = get_global_id(0);
   int Y = get_global_id(1);
@@ -31,38 +31,37 @@ __kernel void MaxPooling2d(__global float4 *input, __global float4 *output, cons
   output[(output_shape.y * X + Y) * output_shape.w + Z] = maximum;
 }
 
-// __constant sampler_t sample_none = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;
+__constant sampler_t sample_none = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;
 
-//__kernel void MaxPooling2dImage2d(__read_only image2d_t input, __write_only image2d_t output, const int4 input_shape,
-//                                  const int4 output_shape, const int2 stride, const int2 kernel_size,
-//                                  const int2 padding) {
-//  // axis to dst tensor coordinate
-//  int X = get_global_id(0);
-//  int Y = get_global_id(1);
-//  int Z = get_global_id(2);
-//
-//  // boundary check
-//  if (X >= output_shape.x || Y >= output_shape.y || Z >= output_shape.w) {
-//    return;
-//  }
-//
-//  float4 maximum = (float4)(-10000.0f);
-//  int xs = X * stride.x + padding.x;
-//  int ys = Y * stride.y + padding.y;
-//
-//  for (int ky = 0; ky < kernel_size.y; ++ky) {
-//    int y_c = ys + ky;
-//    if (y_c < 0 || y_c >= input_shape.y) {
-//      continue;
-//    }
-//    for (int kx = 0; kx < kernel_size.x; ++kx) {
-//      int x_c = xs + kx;
-//      if (x_c < 0 || x_c >= input_shape.x) {
-//        continue;
-//      }
-//      float4 src = read_imagef(input, sample_none, (int2)(x_c, y_c * input_shape.w + Z));
-//      maximum = max(src, maximum);
-//    }
-//  }
-//  write_imagef(output, (int2)(X, Y * output_shape.w + Z), maximum);
-//}
\ No newline at end of file
+__kernel void MaxPooling2d_IMG(__read_only image2d_t input, __write_only image2d_t output, const int4 input_shape,
+                               const int4 output_shape, const int2 stride, const int2 kernel_size, const int2 padding) {
+  // axis to dst tensor coordinate
+  int X = get_global_id(0);
+  int Y = get_global_id(1);
+  int Z = get_global_id(2);
+
+  // boundary check
+  if (X >= output_shape.x || Y >= output_shape.y || Z >= output_shape.w) {
+    return;
+  }
+
+  float4 maximum = (float4)(-10000.0f);
+  int xs = X * stride.x + padding.x;
+  int ys = Y * stride.y + padding.y;
+
+  for (int kx = 0; kx < kernel_size.x; ++kx) {
+    int x_c = xs + kx;
+    if (x_c < 0 || x_c >= input_shape.x) {
+      continue;
+    }
+    for (int ky = 0; ky < kernel_size.y; ++ky) {
+      int y_c = ys + ky;
+      if (y_c < 0 || y_c >= input_shape.y) {
+        continue;
+      }
+      float4 src = read_imagef(input, sample_none, (int2)(x_c, y_c * input_shape.w + Z));
+      maximum = max(src, maximum);
+    }
+  }
+  write_imagef(output, (int2)(X, Y * output_shape.w + Z), maximum);
+}

mindspore/lite/src/runtime/kernel/opencl/kernel/arithmetic.cc

@@ -13,13 +13,14 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
+
+#include "src/runtime/kernel/opencl/kernel/arithmetic.h"
 #include <set>
 #include <vector>
 #include <string>
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
 #include "src/runtime/kernel/opencl/utils.h"
-#include "src/runtime/kernel/opencl/kernel/arithmetic.h"
 #ifndef PROGRAM_WITH_IL
 #include "src/runtime/kernel/opencl/cl/fp32/arithmetic_buffer.cl.inc"
 #include "src/runtime/kernel/opencl/cl/fp32/arithmetic_image2d.cl.inc"
@@ -41,8 +42,8 @@ std::vector<size_t> ArithmeticOpenCLKernel::InitGlobalSize() const {
 void ArithmeticOpenCLKernel::Image2dGetWorkGroupSize() {
   global_size_ = InitGlobalSize();
   int max_work_group_size = runtime_->GetKernelMaxWorkGroupSize(kernel_(), (*runtime_->Device())());
-  local_size_ = GetLocalSize(global_size_, max_work_group_size);
-  global_size_ = GetGlobalSize(local_size_, global_size_);
+  local_size_ = GetCommonLocalSize(global_size_, max_work_group_size);
+  global_size_ = GetCommonGlobalSize(local_size_, global_size_);
 }
 
 void ArithmeticOpenCLKernel::BufferGetWorkGroupSize() {

mindspore/lite/src/runtime/kernel/opencl/kernel/depthwise_conv2d.cc

@@ -31,12 +31,11 @@
 
 #endif
 
-
-using mindspore::schema::PrimitiveType_DepthwiseConv2D;
 using mindspore::kernel::KERNEL_ARCH::kGPU;
 using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_DepthwiseConv2D;
 
 namespace mindspore::kernel {
 
@@ -117,11 +116,9 @@ int DepthwiseConv2dOpenCLKernel::InitBuffer() {
   return RET_OK;
 }
 
-int DepthwiseConv2dOpenCLKernel::ReSize() {
-  return RET_OK;
-}
+int DepthwiseConv2dOpenCLKernel::ReSize() { return RET_OK; }
 
-int DepthwiseConv2dOpenCLKernel::GetImageSize(size_t idx, std::vector<size_t>* img_size) {
+int DepthwiseConv2dOpenCLKernel::GetImageSize(size_t idx, std::vector<size_t> *img_size) {
   size_t CO4 = UP_DIV(outputs_[0]->Channel(), C4NUM);
   size_t im_dst_x, im_dst_y;
   if (inputs_[0]->GetFormat() == schema::Format_NHWC4) {
@@ -141,16 +138,18 @@ int DepthwiseConv2dOpenCLKernel::GetImageSize(size_t idx, std::vector<size_t>* i
   *img_size = vec;
   return RET_OK;
 }
-int DepthwiseConv2dOpenCLKernel::GetGlobalSize(size_t idx, std::vector<size_t>* global_size) {
+
+int DepthwiseConv2dOpenCLKernel::GetGlobalSize(size_t idx, std::vector<size_t> *global_size) {
   size_t CO4 = UP_DIV(outputs_[0]->Channel(), C4NUM);
-  std::vector <size_t> global = {(size_t) outputs_[0]->Width(), (size_t) outputs_[0]->Height(), CO4};
+  std::vector<size_t> global = {(size_t)outputs_[0]->Width(), (size_t)outputs_[0]->Height(), CO4};
   *global_size = std::move(global);
   return RET_OK;
 }
-int DepthwiseConv2dOpenCLKernel::GetLocalSize(size_t idx, const std::vector<size_t>& global_size,
-    std::vector<size_t>* local_size) {
+
+int DepthwiseConv2dOpenCLKernel::GetLocalSize(size_t idx, const std::vector<size_t> &global_size,
+                                              std::vector<size_t> *local_size) {
   size_t CO4 = UP_DIV(outputs_[0]->Channel(), C4NUM);
-  std::vector <size_t> local = {1, 1, CO4};
+  std::vector<size_t> local = {1, 1, CO4};
   *local_size = std::move(local);
   return RET_OK;
 }
@@ -161,8 +160,8 @@ int DepthwiseConv2dOpenCLKernel::Run() {
   auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
   size_t CO4 = UP_DIV(outputs_[0]->Channel(), C4NUM);
   size_t CI4 = UP_DIV(inputs_[0]->Channel(), C4NUM);
-  std::vector <size_t> global = {(size_t) outputs_[0]->Width(), (size_t) outputs_[0]->Height(), CO4};
-  std::vector <size_t> local;
+  std::vector<size_t> global = {(size_t)outputs_[0]->Width(), (size_t)outputs_[0]->Height(), CO4};
+  std::vector<size_t> local;
   GetLocalSize(0, global, &local);
 
   float relu_clip1 = 6.0;

mindspore/lite/src/runtime/kernel/opencl/kernel/depthwise_conv2d.h

@@ -28,11 +28,10 @@ namespace mindspore::kernel {
 class DepthwiseConv2dOpenCLKernel : public OpenCLKernel {
  public:
   explicit DepthwiseConv2dOpenCLKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
-      const std::vector<lite::tensor::Tensor *> &outputs)
-      : OpenCLKernel(parameter, inputs, outputs),
-        packed_weight_(nullptr), bias_data_(nullptr), kernel_(nullptr) {}
+                                       const std::vector<lite::tensor::Tensor *> &outputs)
+      : OpenCLKernel(parameter, inputs, outputs), packed_weight_(nullptr), bias_data_(nullptr), kernel_(nullptr) {}
 
-  ~DepthwiseConv2dOpenCLKernel() override {};
+  ~DepthwiseConv2dOpenCLKernel() override{};
 
   int Init() override;
 
@@ -42,20 +41,16 @@ class DepthwiseConv2dOpenCLKernel : public OpenCLKernel {
 
   int InitBuffer();
 
-  int GetImageSize(size_t idx, std::vector<size_t>* img_size) override;
-  int GetGlobalSize(size_t idx, std::vector<size_t>* global_size) override;
-  int GetLocalSize(size_t idx, const std::vector<size_t>& global_size,
-                   std::vector<size_t>* local_size) override;
+  int GetImageSize(size_t idx, std::vector<size_t> *img_size) override;
+  int GetGlobalSize(size_t idx, std::vector<size_t> *global_size) override;
+  int GetLocalSize(size_t idx, const std::vector<size_t> &global_size, std::vector<size_t> *local_size) override;
 
  private:
   FLOAT_t *packed_weight_;
   FLOAT_t *bias_data_;
   cl::Kernel kernel_;
-  enum class MEM_TYPE {
-    BUF, IMG
-  } mem_type_{MEM_TYPE::IMG};
+  enum class MEM_TYPE { BUF, IMG } mem_type_{MEM_TYPE::IMG};
 };
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_BACKEND_OPENCL_DEPTHWISE_H_
-

mindspore/lite/src/runtime/kernel/opencl/kernel/pooling2d.cc

@@ -64,12 +64,18 @@ int PoolingOpenCLKernel::Init() {
 #ifdef PROGRAM_WITH_IL
   ocl_runtime->CreateKernelFromIL(kernel_(), kernel_name);
 #else
+  if (mem_type_ == MEM_TYPE::BUF) {
+    kernel_name += "_BUF";
+  } else {
+    kernel_name += "_IMG";
+  }
   std::set<std::string> build_options;
   ocl_runtime->LoadSource(program_name, source);
   ocl_runtime->BuildKernel(kernel_, program_name, kernel_name, build_options);
 #endif
   outputs_[0]->SetFormat(schema::Format_NHWC4);
   MS_LOG(DEBUG) << kernel_name << " Init Done!";
+
   return RET_OK;
 }
 
@@ -81,8 +87,30 @@ std::vector<size_t> PoolingOpenCLKernel::InitGlobalSize() const {
   return global;
 }
 
-int PoolingOpenCLKernel::InitBuffer() { return 0; }
-int PoolingOpenCLKernel::ReSize() { return 0; }
+int PoolingOpenCLKernel::GetImageSize(size_t idx, std::vector<size_t> *img_size) {
+  size_t CO4 = UP_DIV(outputs_[0]->Channel(), C4NUM);
+  size_t im_dst_x, im_dst_y;
+  if (inputs_[0]->GetFormat() == schema::Format_NHWC4) {
+    im_dst_x = outputs_[0]->Height();
+    im_dst_y = outputs_[0]->Width() * CO4;
+  } else {
+    im_dst_y = outputs_[0]->Width();
+    im_dst_x = outputs_[0]->Height() * CO4;
+  }
+#ifdef ENABLE_FP16
+  size_t img_dtype = CL_HALF_FLOAT;
+#else
+  size_t img_dtype = CL_FLOAT;
+#endif
+  img_size->clear();
+  std::vector<size_t> vec{im_dst_x, im_dst_y, img_dtype};
+  *img_size = vec;
+  return RET_OK;
+}
+
+int PoolingOpenCLKernel::InitBuffer() { return RET_OK; }
+
+int PoolingOpenCLKernel::ReSize() { return RET_OK; }
 
 int PoolingOpenCLKernel::Run() {
   MS_LOG(DEBUG) << this->Name() << " Running!";
@@ -110,12 +138,11 @@ int PoolingOpenCLKernel::Run() {
   std::vector<size_t> local_size;
   std::vector<size_t> global_size = InitGlobalSize();
   int max_work_group_size = ocl_runtime->GetKernelMaxWorkGroupSize(kernel_(), (*ocl_runtime->Device())());
-  local_size = GetLocalSize(global_size, max_work_group_size);
-  global_size = GetGlobalSize(local_size, global_size);
+  local_size = GetCommonLocalSize(global_size, max_work_group_size);
+  global_size = GetCommonGlobalSize(local_size, global_size);
 
   // run opengl kernel
   ocl_runtime->RunKernel(kernel_, global_size, local_size, nullptr);
-
   return RET_OK;
 }
 

mindspore/lite/src/runtime/kernel/opencl/kernel/pooling2d.h

@@ -25,11 +25,11 @@
 
 namespace mindspore::kernel {
 
-class PoolingOpenCLKernel : public LiteKernel {
+class PoolingOpenCLKernel : public OpenCLKernel {
  public:
   explicit PoolingOpenCLKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
                                const std::vector<lite::tensor::Tensor *> &outputs)
-      : LiteKernel(parameter, inputs, outputs) {
+      : OpenCLKernel(parameter, inputs, outputs) {
     parameter_ = reinterpret_cast<PoolingParameter *>(parameter);
   }
   ~PoolingOpenCLKernel() override{};
@@ -38,10 +38,11 @@ class PoolingOpenCLKernel : public LiteKernel {
   int ReSize() override;
   int Run() override;
   int InitBuffer();
+  int GetImageSize(size_t idx, std::vector<size_t> *img_size) override;
 
  private:
   std::vector<size_t> InitGlobalSize() const;
-
+  enum class MEM_TYPE { BUF, IMG } mem_type_{MEM_TYPE::IMG};
   PoolingParameter *parameter_;
   cl::Kernel kernel_;
 };
@@ -49,4 +50,3 @@ class PoolingOpenCLKernel : public LiteKernel {
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_BACKEND_OPENCL_POOLING_H_
-

mindspore/lite/src/runtime/kernel/opencl/utils.cc

@@ -22,7 +22,7 @@
 namespace mindspore {
 namespace kernel {
 
-std::vector<size_t> GetGlobalSize(const std::vector<size_t> &local, const std::vector<size_t> &global) {
+std::vector<size_t> GetCommonGlobalSize(const std::vector<size_t> &local, const std::vector<size_t> &global) {
   std::vector<size_t> result(3, 1);
   for (int i = 0; i < 3; ++i) {
     result[i] = AlignByN(global[i], local[i]);
@@ -30,7 +30,7 @@ std::vector<size_t> GetGlobalSize(const std::vector<size_t> &local, const std::v
   return result;
 }
 
-std::vector<size_t> GetLocalSize(const std::vector<size_t> &global, int max_size) {
+std::vector<size_t> GetCommonLocalSize(const std::vector<size_t> &global, int max_size) {
   size_t wg_z = GetBiggestDividerWithPriority(global[2], 8);
   size_t wg_xy_size = max_size / wg_z;
   size_t wg_x = std::min(DivideRoundUp(global[0], 2), wg_xy_size);

mindspore/lite/src/runtime/kernel/opencl/utils.h

@@ -75,10 +75,10 @@ T AlignByN(T number, N n) {
 }
 
 // GetGlobalSize
-std::vector<size_t> GetGlobalSize(const std::vector<size_t> &local, const std::vector<size_t> &global);
+std::vector<size_t> GetCommonGlobalSize(const std::vector<size_t> &local, const std::vector<size_t> &global);
 
 // GetLocalSize
-std::vector<size_t> GetLocalSize(const std::vector<size_t> &global, int max_size);
+std::vector<size_t> GetCommonLocalSize(const std::vector<size_t> &global, int max_size);
 
 std::string CLErrorCode(cl_int error_code);
 

mindspore/lite/test/ut/src/runtime/kernel/opencl/max_pooling_tests.cc

@@ -43,33 +43,42 @@ TEST_F(TestMaxPoolingOpenCL, MaxPool_1_32_512_96) {
   MS_LOG(INFO) << "ocl runtime";
   auto ocl_runtime = lite::opencl::OpenCLRuntime::GetInstance();
   ocl_runtime->Init();
+  auto allocator = ocl_runtime->GetAllocator();
 
   MS_LOG(INFO) << "PoolingParameter";
   auto param = new PoolingParameter;
   InitParameter(param);
 
   // define tensor
-  MS_LOG(INFO) << "define tensor";
+  MS_LOG(INFO) << "define tensor1";
   std::vector<int> input_shape = {1, 16, 256, 192};
   std::vector<int> output_shape = {1, 8, 128, 192};
   auto data_type = kNumberTypeFloat32;
   auto tensorType = schema::NodeType_ValueNode;
+  MS_LOG(INFO) << "define tensor2";
   auto input_tensor = new lite::tensor::Tensor(data_type, input_shape, schema::Format_NHWC4, tensorType);
   auto output_tensor = new lite::tensor::Tensor(data_type, output_shape, schema::Format_NHWC4, tensorType);
+  MS_LOG(INFO) << "define input";
   std::vector<lite::tensor::Tensor *> inputs{input_tensor};
   std::vector<lite::tensor::Tensor *> outputs{output_tensor};
 
   // run
+  MS_LOG(INFO) << "pooling_kernel";
   auto *pooling_kernel = new kernel::PoolingOpenCLKernel(reinterpret_cast<OpParameter *>(param), inputs, outputs);
+  MS_LOG(INFO) << "pooling_kernel init";
   pooling_kernel->Init();
+
   std::vector<kernel::LiteKernel *> kernels{pooling_kernel};
+  inputs[0]->MallocData(allocator);
   auto *pGraph = new kernel::SubGraphOpenCLKernel(inputs, outputs, kernels, kernels, kernels);
+  MS_LOG(INFO) << "pGraph init";
   pGraph->Init();
 
   // load data
-  MS_LOG(INFO) << "load data";
+  MS_LOG(INFO) << "load data1";
   std::string input_file = "maxpool_in.bin";
   std::string expect_file = "maxpool_out.bin";
+  MS_LOG(INFO) << "load data2";
   LoadTestData(input_tensor->Data(), input_tensor->Size(), input_file);
   auto *input_data = reinterpret_cast<float *>(input_tensor->Data());
   printf("input[0:10]:");
@@ -81,6 +90,7 @@ TEST_F(TestMaxPoolingOpenCL, MaxPool_1_32_512_96) {
   pGraph->Run();
 
   MS_LOG(INFO) << "compare result";
+  std::cout << "compare result" << std::endl;
   CompareOutput(output_tensor, expect_file);
 }
 

mindspore/lite/test/ut/src/runtime/kernel/opencl/utils_tests.cc

@@ -24,9 +24,14 @@ namespace mindspore {
 
 void LoadTestData(void *dst, size_t dst_size, const std::string &file_path) {
   if (file_path.empty()) {
-    memset(dst, dst_size, dst_size);
+    memset(dst, 0x00, dst_size);
   } else {
-    memcpy(dst, reinterpret_cast<const void *>(dst_size), dst_size);
+    auto src_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(file_path.c_str(), &dst_size));
+    if (src_data != nullptr) {
+      memcpy(dst, src_data, dst_size);
+    } else {
+      MS_LOG(ERROR) << "read file empty.";
+    }
   }
 }