updates

Author: lidongzh

Ops/FWI/Src/Src_Rec.cu

@@ -64,8 +64,8 @@ Src_Rec::Src_Rec(Parameter &para, string survey_fname, const float *stf, int gro
 	float *d_win_start, *d_win_end;
 	float *d_weights;
 
-	dim3 threads(TX,TY);
-	dim3 blocks((para.nz()+TX-1)/TX, (para.nx()+TY-1)/TY);
+	// dim3 threads(TX,TY);
+	// dim3 blocks((para.nz()+TX-1)/TX, (para.nx()+TY-1)/TY);
 
 	if_res_ = para.if_res();
 	if_win_ = para.if_win();

Ops/FWI/Src/Torch_Fwi.cpp

@@ -60,7 +60,7 @@ std::vector<torch::Tensor> fwi_backward(const torch::Tensor &th_Lambda,
   auto th_grad_Lambda_sum = torch::zeros_like(th_Lambda);
   auto th_grad_Mu_sum = torch::zeros_like(th_Mu);
   auto th_grad_Den_sum = torch::zeros_like(th_Den);
-  auto th_misfit_sum = torch::zeros(1);
+  float misfit_sum = 0.0;
 
 #pragma omp parallel for num_threads(ngpu)
   for (int i = 0; i < ngpu; i++) {
@@ -86,14 +86,13 @@ std::vector<torch::Tensor> fwi_backward(const torch::Tensor &th_Lambda,
     // torch::Tensor th_misfit = torch::from_blob(&misfit, {1});
     vec_misfit.at(i) = misfit;
   }
-  float misfit_sum = 0.0;
   for (int i = 0; i < ngpu; i++) {
     th_grad_Lambda_sum += vec_grad_Lambda.at(i);
     th_grad_Mu_sum += vec_grad_Mu.at(i);
     th_grad_Den_sum += vec_grad_Den.at(i);
     misfit_sum += vec_misfit.at(i);
   }
-  return {torch::from_blob(&misfit_sum, {1}), th_grad_Lambda_sum, th_grad_Mu_sum,
+  return {torch::tensor({misfit_sum}), th_grad_Lambda_sum, th_grad_Mu_sum,
           th_grad_Den_sum, vec_grad_stf.at(0)};
 }
 

Ops/FWI/Src/libCUFD.cu

@@ -102,8 +102,8 @@ extern "C" void cufd(float *misfit, float *grad_Lambda, float *grad_Mu,
 
   dim3 threads(TX, TY);
   dim3 blocks((nz + TX - 1) / TX, (nx + TY - 1) / TY);
-  dim3 threads2(TX + 4, TY + 4);
-  dim3 blocks2((nz + TX + 3) / (TX + 4), (nx + TY + 3) / (TY + 4));
+  // dim3 threads2(TX + 4, TY + 4);
+  // dim3 blocks2((nz + TX + 3) / (TX + 4), (nx + TY + 3) / (TY + 4));
 
   float *d_vz, *d_vx, *d_szz, *d_sxx, *d_sxz, *d_vz_adj, *d_vx_adj, *d_szz_adj,
       *d_sxx_adj, *d_sxz_adj;

src/main.py

@@ -84,9 +84,9 @@
   cp_true_pad = np.ascontiguousarray(np.reshape(cp_true_pad, (nz_pad, -1), order='F'))
   cs_true_pad = np.zeros((nz_pad, nx_pad))
   print(f'cp_true_pad shape = {cp_true_pad.shape}')
-  plt.imshow(cp_true_pad, cmap='RdBu_r')
-  plt.colorbar()
-  plt.show()
+  # plt.imshow(cp_true_pad, cmap='RdBu_r')
+  # plt.colorbar()
+  # plt.show()
   den_true_pad = 2500.0 * np.ones((nz_pad, nx_pad))
   th_cp_pad = torch.tensor(cp_true_pad, dtype=torch.float32, requires_grad=False)
   th_cs_pad = torch.tensor(cs_true_pad, dtype=torch.float32, requires_grad=False)