!21229 [LITE] fix magic numbers

Merge pull request !21229 from yefeng/140-fix_code-master

mindspore/lite/src/common/utils.cc

@@ -165,6 +165,5 @@ bool IsSupportSDot() {
 #endif
   return status;
 }
-
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/runtime/kernel/arm/fp16_grad/bn_fp16_grad.cc

@@ -34,6 +34,16 @@ using mindspore::lite::RET_OK;
 using mindspore::schema::PrimitiveType_BatchNormGrad;
 
 namespace mindspore::kernel {
+namespace {
+constexpr int kNumInputDim_0 = 0;
+constexpr int kNumInputDim_1 = 1;
+constexpr int kNumInputDim_2 = 2;
+constexpr int kNumInputDim_3 = 3;
+constexpr int kNumInputDim_4 = 4;
+constexpr int kNumInputDim_5 = 4;
+constexpr int kNumOutputDim_2 = 2;
+constexpr int kNumJobs = 4;
+}  // namespace
 int BNGradCPUKernelFp16::ReSize() {
   auto *input_x = in_tensors_.at(1);
   int channels = input_x->shape().at(kNHWC_C);
@@ -52,16 +62,16 @@ int BNGradCPUKernelFp16::Init() {
 }
 
 int BNGradCPUKernelFp16::Execute(int task_id) {
-  auto *input_yt = in_tensors_.at(0);
-  auto *input_x = in_tensors_.at(1);
-  auto *input_scale = in_tensors_.at(2);
-  auto *input_mean = in_tensors_.at(3);
-  auto *input_var = in_tensors_.at(4);
+  auto *input_yt = in_tensors_.at(kNumInputDim_0);
+  auto *input_x = in_tensors_.at(kNumInputDim_1);
+  auto *input_scale = in_tensors_.at(kNumInputDim_2);
+  auto *input_mean = in_tensors_.at(kNumInputDim_3);
+  auto *input_var = in_tensors_.at(kNumInputDim_4);
 
   auto kernel_name = this->name();
   if (kernel_name.find("FusedBatchNormGradCPU") != std::string::npos) {
-    input_mean = in_tensors_.at(4);
-    input_var = in_tensors_.at(5);
+    input_mean = in_tensors_.at(kNumInputDim_4);
+    input_var = in_tensors_.at(kNumInputDim_5);
   }
   auto bn_param = reinterpret_cast<BNGradParameter *>(op_parameter_);
   int stage = stage_;
@@ -71,7 +81,7 @@ int BNGradCPUKernelFp16::Execute(int task_id) {
 
   auto *output_dx = out_tensors_.at(0);
   auto *output_scale = out_tensors_.at(1);
-  auto *output_bias = out_tensors_.at(2);
+  auto *output_bias = out_tensors_.at(kNumOutputDim_2);
   int32_t batch = input_x->Batch();
   int32_t channels = input_x->Channel();
   int32_t spatial = input_x->Height() * input_x->Width();
@@ -91,7 +101,7 @@ int BNGradCPUKernelFp16::Execute(int task_id) {
   count = (count < 0) ? 0 : count;
   switch (stage) {
     case 0: {
-      for (int job = task_id; job < 4; job += thread_num) {
+      for (int job = task_id; job < kNumJobs; job += thread_num) {
         switch (job) {
           case 0:
             var2InvarFp16(save_var, input_var->ElementsNum(), bn_param->epsilon_);

mindspore/lite/src/runtime/kernel/arm/fp16_grad/dropout_fp16_grad.cc

@@ -41,7 +41,6 @@ int DropoutGradCPUKernelFp16::Init() {
     MS_LOG(ERROR) << "unsupported ratio value - Dropout ratio should be between zero to one";
     return RET_ERROR;
   }
-
   if (ratio >= 1.0f) {
     scale_ = 1.0f;
   } else {

mindspore/lite/src/runtime/kernel/arm/fp16_grad/layernorm_fp16_grad.cc

@@ -30,6 +30,16 @@ using mindspore::lite::RET_OK;
 using mindspore::schema::PrimitiveType_LayerNormGrad;
 
 namespace mindspore::kernel {
+namespace {
+constexpr int kNumInputDim_0 = 0;
+constexpr int kNumInputDim_1 = 1;
+constexpr int kNumInputDim_2 = 2;
+constexpr int kNumInputDim_3 = 3;
+constexpr int kNumInputDim_4 = 4;
+constexpr int kNumOutputDim_0 = 0;
+constexpr int kNumOutputDim_1 = 1;
+constexpr int kNumOutputDim_2 = 2;
+}  // namespace
 int LayerNormGradCPUKernelFp16::ReSize() { return RET_OK; }
 
 int LayerNormGradCPUKernelFp16::Init() {
@@ -63,14 +73,14 @@ int LayerNormGradCPUKernelFp16::Init() {
 }
 
 int LayerNormGradCPUKernelFp16::Execute(int task_id) {
-  auto input_x = in_tensors_.at(0);
-  auto input_dy = in_tensors_.at(1);
-  auto input_var = in_tensors_.at(2);
-  auto input_mean = in_tensors_.at(3);
-  auto input_gamma = in_tensors_.at(4);
-  auto output_dx = out_tensors_.at(0);
-  auto output_dg = out_tensors_.at(1);
-  auto output_db = out_tensors_.at(2);
+  auto input_x = in_tensors_.at(kNumInputDim_0);
+  auto input_dy = in_tensors_.at(kNumInputDim_1);
+  auto input_var = in_tensors_.at(kNumInputDim_2);
+  auto input_mean = in_tensors_.at(kNumInputDim_3);
+  auto input_gamma = in_tensors_.at(kNumInputDim_4);
+  auto output_dx = out_tensors_.at(kNumOutputDim_0);
+  auto output_dg = out_tensors_.at(kNumOutputDim_1);
+  auto output_db = out_tensors_.at(kNumOutputDim_2);
 
   float16_t *x = reinterpret_cast<float16_t *>(input_x->data_c());
   float16_t *dy = reinterpret_cast<float16_t *>(input_dy->data_c());

mindspore/lite/src/runtime/kernel/arm/fp16_grad/pooling_fp16_grad.cc

@@ -29,24 +29,23 @@ using mindspore::schema::PrimitiveType_AvgPoolGrad;
 using mindspore::schema::PrimitiveType_MaxPoolGrad;
 
 namespace mindspore::kernel {
+namespace {
+constexpr int kNumInputDim_2 = 2;
+constexpr int kNumShapeDim_2 = 2;
+}  // namespace
 int PoolingGradCPUKernelFp16::ReSize() {
   PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *>(op_parameter_);
-
   auto in_shape = in_tensors_.at(0)->shape();
   auto out_shape = in_tensors_.at(1)->shape();
-
   if (pool_param->pool_mode_ == PoolMode_AvgPool) {
-    out_shape = in_tensors_.at(2)->shape();
+    out_shape = in_tensors_.at(kNumInputDim_2)->shape();
   }
-
   int input_h = in_shape.at(1);
-  int input_w = in_shape.at(2);
-
+  int input_w = in_shape.at(kNumShapeDim_2);
   if (pool_param->global_) {
     pool_param->window_w_ = input_w;
     pool_param->window_h_ = input_h;
   }
-
   pool_param->input_h_ = in_shape[kNHWC_H];
   pool_param->input_w_ = in_shape[kNHWC_W];
   pool_param->input_batch_ = in_shape[kNHWC_N];
@@ -55,7 +54,6 @@ int PoolingGradCPUKernelFp16::ReSize() {
   pool_param->output_w_ = out_shape[kNHWC_W];
   pool_param->output_batch_ = out_shape[kNHWC_N];
   pool_param->output_channel_ = out_shape[kNHWC_C];
-
   return RET_OK;
 }
 
@@ -73,11 +71,11 @@ int PoolingGradCPUKernelFp16::Execute(int task_id) {
     std::fill(output_ptr + task_id * stride * in_batch_size, output_ptr + ((task_id * stride) + count) * in_batch_size,
               0.f);
     if (pool_param->pool_mode_ == PoolMode_MaxPool) {
-      auto dy_ptr = reinterpret_cast<float16_t *>(in_tensors_.at(2)->data_c());
+      auto dy_ptr = reinterpret_cast<float16_t *>(in_tensors_.at(kNumInputDim_2)->data_c());
       MaxPoolingFp16Grad(input_ptr + task_id * stride * in_batch_size, dy_ptr + task_id * stride * out_batch_size,
                          output_ptr + task_id * stride * in_batch_size, count, pool_param);
     } else {
-      input_ptr = reinterpret_cast<float16_t *>(in_tensors_.at(2)->data_c());
+      input_ptr = reinterpret_cast<float16_t *>(in_tensors_.at(kNumInputDim_2)->data_c());
       AvgPoolingFp16Grad(input_ptr + task_id * stride * out_batch_size, output_ptr + task_id * stride * in_batch_size,
                          count, pool_param);
     }

mindspore/lite/src/runtime/kernel/arm/fp16_grad/resize_fp16_grad.cc

@@ -46,7 +46,6 @@ int ResizeGradCPUKernelFp16::ReSize() {
   param->out_width_ = static_cast<size_t>(out_tensors_.at(0)->Width());
   param->height_scale_ = ScalingFp16(param->out_height_, param->in_height_, align_corners);
   param->width_scale_ = ScalingFp16(param->out_width_, param->in_width_, align_corners);
-
   return RET_OK;
 }
 
@@ -67,7 +66,6 @@ int ResizeGradCPUKernelFp16::Execute(int task_id) {
   }
   auto batch_size = in_tensors_.at(0)->Batch();
   auto channel = in_tensors_.at(0)->Channel();
-
   if (param->method == static_cast<int>(schema::ResizeMethod_NEAREST)) {
     ResizeNearestNeighborFp16Grad(in_addr, out_addr, batch_size, channel, in_tensors_.at(0)->format(), param);
   } else {

mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.cc

@@ -23,6 +23,9 @@ using mindspore::lite::RET_OK;
 using mindspore::schema::PrimitiveType_FusedBatchNorm;
 
 namespace mindspore::kernel {
+namespace {
+constexpr int kNumInputSize = 5;
+}  // namespace
 int FusedBatchnormCPUKernel::ReSize() {
   FreeMeanAndVariance();
   FreeScaleAndOffset();
@@ -66,7 +69,7 @@ int FusedBatchnormCPUKernel::InitConstTensor() {
 
 int FusedBatchnormCPUKernel::Run() {
   auto param = reinterpret_cast<BatchNormParameter *>(op_parameter_);
-  if (IsTrain() && IsTrainable() && in_tensors_.size() >= 5) {
+  if (IsTrain() && IsTrainable() && in_tensors_.size() >= kNumInputSize) {
     float *in = static_cast<float *>(in_tensors_[0]->MutableData());
     float *scale = static_cast<float *>(in_tensors_[1]->MutableData());
     float *offset = static_cast<float *>(in_tensors_[2]->MutableData());