Browse Source

[MSLITE] depthwise fp16 sw bug

tags/v1.1.0
ling 5 years ago
parent
commit
ef2de7a494
2 changed files with 2 additions and 1 deletions
  1. +1
    -1
      mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_slidewindow_fp16.cc
  2. +1
    -0
      mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_fp16.cc

+ 1
- 1
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_slidewindow_fp16.cc View File

@@ -41,7 +41,7 @@ ConvolutionDepthwiseSWFp16CPUKernel::~ConvolutionDepthwiseSWFp16CPUKernel() {
} }


int ConvolutionDepthwiseSWFp16CPUKernel::InitBuffer() { int ConvolutionDepthwiseSWFp16CPUKernel::InitBuffer() {
if (conv_param_->input_channel_ % C4NUM != 0) {
if (conv_param_->input_channel_ % C8NUM != 0) {
need_align_ = true; need_align_ = true;
int C8 = UP_DIV(conv_param_->input_channel_, C8NUM); int C8 = UP_DIV(conv_param_->input_channel_, C8NUM);
int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C8NUM * C8; int pack_input_size = conv_param_->input_batch_ * conv_param_->input_h_ * conv_param_->input_w_ * C8NUM * C8;


+ 1
- 0
mindspore/lite/src/runtime/kernel/arm/fp16/deconvolution_fp16.cc View File

@@ -264,4 +264,5 @@ kernel::LiteKernel *CpuDeConvFp16KernelCreator(const std::vector<lite::Tensor *>
} }
return kernel; return kernel;
} }
REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_DeConv2D, CpuDeConvFp16KernelCreator)
} // namespace mindspore::kernel } // namespace mindspore::kernel

Loading…
Cancel
Save