Add API: Sparse Convolution3D (#41434)

paddle/phi/kernels/sparse/cpu/convolution_grad_kernel.cc

@@ -139,16 +139,16 @@ void Conv3dGradCPUKernel(const CPUContext& dev_ctx,
     T* tmp_in_ptr = in_features_ptr + offsets[i] * in_channels;
     T* tmp_out_grad_ptr = out_grad_features_ptr + offsets[i] * out_channels;
     const T* tmp_kernel_ptr = kernel_ptr + i * in_channels * out_channels;
-    T* tmp_d_x_ptr = d_x_features_ptr + offsets[i] * out_channels;
+    T* tmp_d_x_ptr = d_x_features_ptr + offsets[i] * in_channels;
     T* tmp_d_kernel_ptr = d_kernel_ptr + i * in_channels * out_channels;
 
     // call gemm: d_kernel = transpose(x) * out_grad
     // (in_channels, n) * (n, out_channels)
     blas.GEMM(CblasTrans,
               CblasNoTrans,
-              M,
-              N,
               K,
+              N,
+              M,
               static_cast<T>(1),
               tmp_in_ptr,
               tmp_out_grad_ptr,

paddle/phi/kernels/sparse/cpu/convolution_kernel.cc

@@ -50,16 +50,19 @@ void Conv3dCPUKernel(const CPUContext& dev_ctx,
     kernel_sizes[i] = kernel_dims[i];
   }
 
-  phi::funcs::sparse::GetOutShape(
-      x_dims, kernel_sizes, paddings, dilations, strides, &out_dims);
-  const int in_channels = kernel_dims[3];
-  const int out_channels = kernel_dims[4];
-
   std::vector<int> subm_paddings(paddings), subm_strides(strides);
   if (subm) {
+    // the out shape of subm_conv is same as input shape
+    // reset the padding=kernel_size/2 and strides=1
     phi::funcs::sparse::ResetSubmKernelSizeAndStrides(
         kernel.dims(), &subm_paddings, &subm_strides);
   }
+
+  phi::funcs::sparse::GetOutShape(
+      x_dims, kernel_sizes, subm_paddings, dilations, subm_strides, &out_dims);
+  const int in_channels = kernel_dims[3];
+  const int out_channels = kernel_dims[4];
+
   // Second algorithm:
   // https://pdfs.semanticscholar.org/5125/a16039cabc6320c908a4764f32596e018ad3.pdf
   // 1. product rulebook