!4040 add int8 strided_slice kernel

Merge pull request !4040 from sunsuodong/strided_slice_int8

mindspore/lite/src/ops/strided_slice.cc

@@ -86,7 +86,7 @@ void StridedSlice::ApplyBeginMask() {
 
 void StridedSlice::ApplyEndMask() {
   for (int i = 0; i < ndim_; i++) {
-    if (ends_.at(i)) {
+    if (ends_mask_.at(i)) {
       ends_.at(i) = in_shape_.at(i);
     }
   }

mindspore/lite/src/populate_parameter.cc

@@ -62,7 +62,7 @@
 #include "src/runtime/kernel/arm/fp32/flatten.h"
 #include "src/runtime/kernel/arm/opclib/fp32/unsqueeze.h"
 #include "src/runtime/kernel/arm/opclib/fp32/one_hot.h"
-#include "src/runtime/kernel/arm/opclib/fp32/strided_slice.h"
+#include "src/runtime/kernel/arm/opclib/strided_slice.h"
 #include "src/runtime/kernel/arm/base/prior_box.h"
 #include "src/runtime/kernel/arm/opclib/fp32/space_to_depth.h"
 #include "src/runtime/kernel/arm/opclib/fp32/space_to_batch.h"

mindspore/lite/src/runtime/kernel/arm/base/strided_slice.cc

@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
-#include "src/runtime/kernel/arm/fp32/strided_slice.h"
+#include "src/runtime/kernel/arm/base/strided_slice.h"
 #include <vector>
-#include "src/runtime/kernel/arm/opclib/fp32/strided_slice.h"
+#include "src/runtime/kernel/arm/opclib/strided_slice.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
 #include "include/errorcode.h"
@@ -30,26 +30,24 @@ using mindspore::schema::PrimitiveType_StridedSlice;
 
 namespace mindspore::kernel {
 
-int StridedSliceCPUKernel::Init() { return RET_OK; }
+int StridedSliceCPUKernel::Init() {
-
+  auto input = inputs_.at(0);
-int StridedSliceCPUKernel::ReSize() { return 0; }
+  auto parameter = reinterpret_cast<StridedSliceParameter *>(opParameter);
-
+  MS_ASSERT(input);
-int StridedSliceCPUKernel::StridedSlice() {
+  MS_ASSERT(parameter);
-  StridedSliceParameter *param = reinterpret_cast<StridedSliceParameter *>(opParameter);
+  parameter->data_type = input->data_type() == kNumberTypeInt8 ? kDataTypeInt8 : kDataTypeFloat;
-  auto ret = DoStridedSlice(input_ptr_, output_ptr_, param);
-  if (ret != RET_OK) {
-    return RET_ERROR;
-  }
   return RET_OK;
 }
 
-int StridedSliceCPUKernel::Run() {
+int StridedSliceCPUKernel::ReSize() { return 0; }
-  auto input_tensor = inputs_.at(0);
-  auto output_tensor = outputs_.at(0);
-  input_ptr_ = reinterpret_cast<float *>(input_tensor->Data());
-  output_ptr_ = reinterpret_cast<float *>(output_tensor->Data());
 
-  auto ret = StridedSlice();
+int StridedSliceCPUKernel::Run() {
+  auto input = inputs_.at(0);
+  auto output = outputs_.at(0);
+  MS_ASSERT(input);
+  MS_ASSERT(output);
+
+  auto ret = DoStridedSlice(input->Data(), output->Data(), reinterpret_cast<StridedSliceParameter *>(opParameter));
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "StridedSlice error error_code[" << ret << "]";
     return RET_ERROR;
@@ -57,10 +55,10 @@ int StridedSliceCPUKernel::Run() {
   return RET_OK;
 }
 
-kernel::LiteKernel *CpuStridedSliceFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+kernel::LiteKernel *CpuStridedSliceKernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                                     const std::vector<lite::tensor::Tensor *> &outputs,
+                                                 const std::vector<lite::tensor::Tensor *> &outputs,
-                                                     OpParameter *opParameter, const lite::Context *ctx,
+                                                 OpParameter *opParameter, const lite::Context *ctx,
-                                                     const kernel::KernelKey &desc) {
+                                                 const kernel::KernelKey &desc) {
   MS_ASSERT(desc.type == schema::PrimitiveType_StridedSlice);
   if (opParameter == nullptr) {
     MS_LOG(ERROR) << "opParameter null pointer dereferencing.";
@@ -82,5 +80,6 @@ kernel::LiteKernel *CpuStridedSliceFp32KernelCreator(const std::vector<lite::ten
   return kernel;
 }
 
-REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_StridedSlice, CpuStridedSliceFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_StridedSlice, CpuStridedSliceKernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_StridedSlice, CpuStridedSliceKernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/base/strided_slice.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_SRC_BACKEND_ARM_FP32_STRIDED_SLICE_H_
+#ifndef MINDSPORE_LITE_SRC_BACKEND_ARM_BASE_STRIDED_SLICE_H_
-#define MINDSPORE_LITE_SRC_BACKEND_ARM_FP32_STRIDED_SLICE_H_
+#define MINDSPORE_LITE_SRC_BACKEND_ARM_BASE_STRIDED_SLICE_H_
 
 #include <vector>
 #include "ir/anf.h"
@@ -32,13 +32,10 @@ class StridedSliceCPUKernel : public LiteKernel {
   int Init() override;
   int ReSize() override;
   int Run() override;
-  int StridedSlice();
 
  private:
   int thread_num_;
-  float *input_ptr_;
-  float *output_ptr_;
 };
 }  // namespace mindspore::kernel
 
-#endif  // MINDSPORE_LITE_SRC_BACKEND_ARM_FP32_STRIDED_SLICE_H_
+#endif  // MINDSPORE_LITE_SRC_BACKEND_ARM_BASE_STRIDED_SLICE_H_

mindspore/lite/src/runtime/kernel/arm/opclib/op_base.h

@@ -48,6 +48,11 @@
 #define kInputSize1 2
 #define kInputSize2 3
 
+enum LiteDataType {
+  kDataTypeFloat,
+  kDataTypeInt8,
+};
+
 struct OpParameter {
   char name_[100];
   int type_;
@@ -55,4 +60,3 @@ struct OpParameter {
 };
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_OPCLIB_OP_BASE_H_
-

mindspore/lite/src/runtime/kernel/arm/opclib/strided_slice.cc

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "src/runtime/kernel/arm/opclib/fp32/strided_slice.h"
+#include "src/runtime/kernel/arm/opclib/strided_slice.h"
 #include "src/runtime/kernel/arm/opclib/errorcode.h"
 
 void PadStridedSliceParameterTo4D(StridedSliceParameter *param) {
@@ -45,7 +45,7 @@ void PadStridedSliceParameterTo4D(StridedSliceParameter *param) {
   param->num_axes_ = DIMENSION_4D;
 }
 
-int DoStridedSlice(const float *in_data, float *out_data, StridedSliceParameter *param) {
+int DoStridedSlice(const void *in_data, void *out_data, StridedSliceParameter *param) {
   if (in_data == nullptr || out_data == nullptr || param == nullptr) {
     return OPCLIB_NULL_PTR;
   }
@@ -66,17 +66,21 @@ int DoStridedSlice(const float *in_data, float *out_data, StridedSliceParameter
   dim_offset[2] = in_shape[3];
   dim_offset[1] = dim_offset[2] * in_shape[2];
   dim_offset[0] = dim_offset[1] * in_shape[1];
-  size_t output_index = 0;
+  size_t out_offset = 0;
   for (int32_t dim0 = begins[0]; dim0 < ends[0]; dim0 += strides[0]) {
     for (int32_t dim1 = begins[1]; dim1 < ends[1]; dim1 += strides[1]) {
       for (int32_t dim2 = begins[2]; dim2 < ends[2]; dim2 += strides[2]) {
         for (int32_t dim3 = begins[3]; dim3 < ends[3]; dim3 += strides[3]) {
-          out_data[output_index++] =
+          int32_t in_offset = dim0 * dim_offset[0] + dim1 * dim_offset[1] + dim2 * dim_offset[2] + dim3;
-            *(in_data + dim0 * dim_offset[0] + dim1 * dim_offset[1] + dim2 * dim_offset[2] + dim3);
+          if (param->data_type == kDataTypeFloat) {
+            *((float *)out_data + out_offset) = *((float *)in_data + in_offset);
+          } else {
+            *((int8_t *)out_data + out_offset) = *((int8_t *)in_data + in_offset);
+          }
+          out_offset++;
         }
       }
     }
   }
-
   return OPCLIB_OK;
 }

mindspore/lite/src/runtime/kernel/arm/opclib/strided_slice.h

@@ -13,20 +13,21 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#ifndef MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_FP32_STRIDED_SLICE_H_
+#ifndef MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_STRIDED_SLICE_H_
-#define MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_FP32_STRIDED_SLICE_H_
+#define MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_STRIDED_SLICE_H_
 
 #include "src/runtime/kernel/arm/opclib/op_base.h"
 
 struct StridedSliceParameter {
-    OpParameter op_parameter_;
+  OpParameter op_parameter_;
-    int begins_[8] = {0};
+  int begins_[8] = {0};
-    int ends_[8] = {0};
+  int ends_[8] = {0};
-    int strides_[8] = {1};
+  int strides_[8] = {1};
-    int isScale;
+  int isScale;
-    int num_axes_;
+  int num_axes_;
-    int in_shape_[8];
+  int in_shape_[8];
+  LiteDataType data_type;
 };
 
-int DoStridedSlice(const float *inputs, float *output, StridedSliceParameter * param);
+int DoStridedSlice(const void *inputs, void *output, StridedSliceParameter *param);
-#endif  // MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_FP32_STRIDED_SLICE_H_
+#endif  // MINDSPORE_LITE_SRC_BACKEND_ARM_OPCLIB_STRIDED_SLICE_H_

mindspore/lite/test/CMakeLists.txt

@@ -252,6 +252,7 @@ else()
 endif()
 ### test src
 file(GLOB_RECURSE TEST_CASE_KERNEL_SRC
+    ${TEST_DIR}/ut/src/runtime/kernel/arm/common/*.cc
     ${TEST_DIR}/ut/src/runtime/kernel/arm/fp32/*.cc
     ${TEST_DIR}/ut/src/runtime/kernel/arm/int8/*.cc
 )

mindspore/lite/test/ut/src/runtime/kernel/arm/common/strided_slice_tests.cc

@@ -0,0 +1,126 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <memory>
+#include "common/common_test.h"
+#include "mindspore/lite/src/runtime/kernel/arm/opclib/strided_slice.h"
+#include "mindspore/lite/src/kernel_registry.h"
+
+namespace mindspore {
+class TestStridedSlice : public mindspore::Common {
+ public:
+  TestStridedSlice() {}
+};
+
+void InitStridedSliceParam(StridedSliceParameter *strided_slice_param) {
+  strided_slice_param->begins_[0] = 0;
+  strided_slice_param->begins_[1] = 0;
+  strided_slice_param->begins_[2] = 0;
+
+  strided_slice_param->ends_[0] = 1;
+  strided_slice_param->ends_[1] = 2;
+  strided_slice_param->ends_[2] = 4;
+
+  strided_slice_param->strides_[0] = 1;
+  strided_slice_param->strides_[1] = 2;
+  strided_slice_param->strides_[2] = 2;
+
+  strided_slice_param->in_shape_[0] = 1;
+  strided_slice_param->in_shape_[1] = 2;
+  strided_slice_param->in_shape_[2] = 4;
+  strided_slice_param->num_axes_ = 3;
+}
+
+TEST_F(TestStridedSlice, StridedSlice) {
+  lite::tensor::Tensor in_tensor(kNumberTypeFloat32, {1, 2, 4});
+  lite::tensor::Tensor out_tensor(kNumberTypeFloat32, {1, 1, 2});
+  float input_data[] = {0.2390374, 0.92039955, 0.05051243, 0.49574447, 0.8355223, 0.02647042, 0.08811307, 0.4566604};
+  float output_data[2] = {0};
+  in_tensor.SetData(input_data);
+  out_tensor.SetData(output_data);
+  std::vector<lite::tensor::Tensor *> inputs = {&in_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&out_tensor};
+
+  StridedSliceParameter parameter = {0};
+  InitStridedSliceParam(&parameter);
+  parameter.op_parameter_.type_ = schema::PrimitiveType_StridedSlice;
+  kernel::KernelKey desc = {kernel::KERNEL_ARCH::kCPU, kNumberTypeFloat32, schema::PrimitiveType_StridedSlice};
+
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
+
+  auto ctx = std::make_shared<lite::Context>();
+  auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(&parameter), ctx.get(), desc);
+  ASSERT_NE(kernel, nullptr);
+
+  auto ret = kernel->Run();
+  EXPECT_EQ(0, ret);
+
+  float expect[2] = {0.2390374, 0.05051243};
+  CompareOutputData(output_data, expect, 2, 0.000001);
+
+  in_tensor.SetData(nullptr);
+  out_tensor.SetData(nullptr);
+}
+
+TEST_F(TestStridedSlice, StridedSliceInt8) {
+  lite::tensor::Tensor in_tensor(kNumberTypeInt8, {2, 3, 4});
+  lite::tensor::Tensor out_tensor(kNumberTypeInt8, {2, 3, 4});
+  int8_t input_data[] = {-12, -11, -10, -9,   -8, -7, -6, -5,   -4, -3, -2, -1,
+                          1,  2,  3, 4,       5, 6, 7, 8,       9, 10, 11, 12};
+  int8_t output_data[4] = {0};
+  in_tensor.SetData(input_data);
+  out_tensor.SetData(output_data);
+  std::vector<lite::tensor::Tensor *> inputs = {&in_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&out_tensor};
+
+  StridedSliceParameter parameter = {0};
+  parameter.begins_[0] = 0;
+  parameter.begins_[1] = 1;
+  parameter.begins_[2] = 2;
+  parameter.ends_[0] = 2;
+  parameter.ends_[1] = 3;
+  parameter.ends_[2] = 4;
+  parameter.strides_[0] = 1;
+  parameter.strides_[1] = 2;
+  parameter.strides_[2] = 1;
+  parameter.in_shape_[0] = 2;
+  parameter.in_shape_[1] = 3;
+  parameter.in_shape_[2] = 4;
+  parameter.num_axes_ = 3;
+
+  parameter.op_parameter_.type_ = schema::PrimitiveType_StridedSlice;
+  kernel::KernelKey desc = {kernel::KERNEL_ARCH::kCPU, kNumberTypeInt8, schema::PrimitiveType_StridedSlice};
+
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
+
+  auto ctx = std::make_shared<lite::Context>();
+  auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(&parameter), ctx.get(), desc);
+  ASSERT_NE(kernel, nullptr);
+
+  auto ret = kernel->Run();
+  EXPECT_EQ(0, ret);
+
+  int8_t expect[4] = {-6, -5, 7, 8};
+  for (int i = 0; i < sizeof(expect); ++i) {
+    EXPECT_EQ(output_data[i], expect[i]);
+  }
+
+  in_tensor.SetData(nullptr);
+  out_tensor.SetData(nullptr);
+}
+}  // namespace mindspore