[fix] fix conv

OpenPPL · Apr 3, 2024 · 1c97cf2 · 1c97cf2
1 parent a5c0ad8
commit 1c97cf2
Show file tree

Hide file tree

Showing 4 changed files with 28 additions and 8 deletions.
diff --git a/src/ppl/nn/engines/x86/kernels/onnx/conv1d_kernel.cc b/src/ppl/nn/engines/x86/kernels/onnx/conv1d_kernel.cc
@@ -84,13 +84,23 @@ ppl::common::RetCode Conv1dKernel::DoExecute(KernelExecContext* ctx) {
     }
 
 #ifdef DUMP_CONV
+    std::string conv_dump_name = GetName();
+    if (cur_executor->conv_param()->fuse_flag & kernel::x86::conv_fuse_flag::SUM) {
+        conv_dump_name += "-sum";
+    }
+    if (cur_executor->conv_param()->fuse_flag & kernel::x86::conv_fuse_flag::RELU) {
+        conv_dump_name += "-relu";
+    }
+    if (cur_executor->conv_param()->fuse_flag & kernel::x86::conv_fuse_flag::RELU6) {
+        conv_dump_name += "-relu6";
+    }
     fprintf(stderr, CASE_STRING_FMT() "\n", cur_executor->conv_param()->group, X_shape.GetDim(0),
             cur_executor->conv_param()->channels, X_shape.GetDim(2), X_shape.GetDim(3),
             cur_executor->conv_param()->num_output, Y_shape.GetDim(2), Y_shape.GetDim(3),
             cur_executor->conv_param()->kernel_h, cur_executor->conv_param()->kernel_w,
             cur_executor->conv_param()->stride_h, cur_executor->conv_param()->stride_w,
             cur_executor->conv_param()->pad_h, cur_executor->conv_param()->pad_w,
-            cur_executor->conv_param()->dilation_h - 1, cur_executor->conv_param()->dilation_w - 1, GetName().c_str());
+            cur_executor->conv_param()->dilation_h - 1, cur_executor->conv_param()->dilation_w - 1, conv_dump_name.c_str());
 #endif
 
     PPLNN_X86_REALLOC_TENSOR_BUFFER(Y);

diff --git a/src/ppl/nn/engines/x86/kernels/onnx/conv2d_kernel.cc b/src/ppl/nn/engines/x86/kernels/onnx/conv2d_kernel.cc
@@ -79,13 +79,23 @@ ppl::common::RetCode Conv2dKernel::DoExecute(KernelExecContext* ctx) {
     }
 
 #ifdef DUMP_CONV
+    std::string conv_dump_name = GetName();
+    if (cur_executor->conv_param()->fuse_flag & kernel::x86::conv_fuse_flag::SUM) {
+        conv_dump_name += "-sum";
+    }
+    if (cur_executor->conv_param()->fuse_flag & kernel::x86::conv_fuse_flag::RELU) {
+        conv_dump_name += "-relu";
+    }
+    if (cur_executor->conv_param()->fuse_flag & kernel::x86::conv_fuse_flag::RELU6) {
+        conv_dump_name += "-relu6";
+    }
     fprintf(stderr, CASE_STRING_FMT() "\n", cur_executor->conv_param()->group, X->GetShape()->GetDim(0),
             cur_executor->conv_param()->channels, X->GetShape()->GetDim(2), X->GetShape()->GetDim(3),
             cur_executor->conv_param()->num_output, Y->GetShape()->GetDim(2), Y->GetShape()->GetDim(3),
             cur_executor->conv_param()->kernel_h, cur_executor->conv_param()->kernel_w,
             cur_executor->conv_param()->stride_h, cur_executor->conv_param()->stride_w,
             cur_executor->conv_param()->pad_h, cur_executor->conv_param()->pad_w,
-            cur_executor->conv_param()->dilation_h - 1, cur_executor->conv_param()->dilation_w - 1, GetName().c_str());
+            cur_executor->conv_param()->dilation_h - 1, cur_executor->conv_param()->dilation_w - 1, conv_dump_name.c_str());
 #endif
 
     PPLNN_X86_REALLOC_TENSOR_BUFFER(Y);

diff --git a/src/ppl/nn/engines/x86/kernels/onnx/conv_kernel.cc b/src/ppl/nn/engines/x86/kernels/onnx/conv_kernel.cc
@@ -137,17 +137,17 @@ ppl::common::RetCode ConvKernel::DoExecute(KernelExecContext* ctx) {
     if (kernel_dims == 1) {
         return kernel::x86::conv1d_ndarray_fp32(
             GetISA(), X->GetShape(), sum_src_shape, Y->GetShape(),
-            X->GetBufferPtr<float>(), W->GetBufferPtr<float>(),
-            sum_src_data, b_data, param_->param->group, channels, num_output,
+            X->GetBufferPtr<float>(), sum_src_data, W->GetBufferPtr<float>(),
+            b_data, param_->param->group, channels, num_output,
             param_->param->kernel_shape[0], param_->param->strides[0],
             param_->param->pads[0], param_->param->dilations[0],
             param_->fuse_flag, tmp_buffer, Y->GetBufferPtr<float>());
     }
     if (kernel_dims == 2) {
         return kernel::x86::conv2d_ndarray_fp32(
             GetISA(), X->GetShape(), sum_src_shape, Y->GetShape(),
-            X->GetBufferPtr<float>(), W->GetBufferPtr<float>(),
-            sum_src_data, b_data, param_->param->group, channels, num_output,
+            X->GetBufferPtr<float>(), sum_src_data, W->GetBufferPtr<float>(),
+            b_data, param_->param->group, channels, num_output,
             param_->param->kernel_shape[0], param_->param->kernel_shape[1],
             param_->param->strides[0], param_->param->strides[1],
             param_->param->pads[0], param_->param->pads[1],

diff --git a/src/ppl/nn/engines/x86/optimizer/ops/onnx/conv_op.cc b/src/ppl/nn/engines/x86/optimizer/ops/onnx/conv_op.cc
@@ -66,9 +66,9 @@ RetCode ConvOp::DoInit(const OptKernelOptions& options) {
     infer_dims_func_ = [this](InputOutputInfo* info) -> RetCode {
         auto x = info->GetInput<TensorImpl>(0)->GetShape();
         auto w = info->GetInput<TensorImpl>(1)->GetShape();
-        if (x->GetDim(1) != w->GetDim(1)) {
+        if (x->GetDim(1) != w->GetDim(1) * param_->group) {
             LOG(ERROR) << "input tensor's channels(" << x->GetDim(1) <<
-                ") and weight's channels(" << w->GetDim(1) << ") not match.";
+                ") and weight's channels*group(" << w->GetDim(1) << "*"<< param_->group << ") not match.";
             return ppl::common::RC_INVALID_VALUE;
         }
         return onnx::ReshapeConv(info, param_.get());