Fix int32 overflow in lstm, lstsq, qr_grad, and spectral_norm_grad impl

Co-authored-by: zrr1999 <46243324+zrr1999@users.noreply.github.com>

Author: Copilot

paddle/phi/kernels/impl/lstm_kernel_impl.h

@@ -60,7 +60,7 @@ void LSTMKernel(const Context& dev_ctx,
   to_batch(dev_ctx, input, batch_gate_new, true, is_reverse);
 
   auto in_dims = input.dims();
-  int frame_size = static_cast<int>(in_dims[1] / 4);
+  int64_t frame_size = in_dims[1] / 4;
   phi::DDim dims({in_dims[0], frame_size});
 
   if (bias.initialized()) {
@@ -254,7 +254,7 @@ void LSTMGradKernel(const Context& dev_ctx,
 
   auto in_dims = input->dims();
   auto out_dims = hidden_g->dims();
-  int frame_size = static_cast<int>(in_dims[1] / 4);
+  int64_t frame_size = in_dims[1] / 4;
   PADDLE_ENFORCE_EQ(frame_size,
                     out_dims[1],
                     common::errors::InvalidArgument(

paddle/phi/kernels/impl/lstsq_kernel_impl.h

@@ -66,8 +66,8 @@ inline void GetResidualsTensor(const DeviceContext& dev_ctx,
                                DenseTensor* rank) {
   auto x_dims = x.dims();
   int dim_size = x_dims.size();
-  int m = x_dims[dim_size - 2];
-  int n = x_dims[dim_size - 1];
+  int64_t m = x_dims[dim_size - 2];
+  int64_t n = x_dims[dim_size - 1];
 
   if (m > n && driver != "gelsy") {
     bool compute_residuals = true;

paddle/phi/kernels/impl/qr_grad_kernel_impl.h

@@ -76,8 +76,8 @@ void QrGradKernel(const Context& dev_ctx,
 
   auto a_dims = A.dims();
   int a_rank = a_dims.size();
-  int m = a_dims[a_rank - 2];
-  int n = a_dims[a_rank - 1];
+  int64_t m = a_dims[a_rank - 2];
+  int64_t n = a_dims[a_rank - 1];
 
   if ((m > n) && (!reduced)) {
     PADDLE_THROW(errors::InvalidArgument(

paddle/phi/kernels/impl/spectral_norm_grad_kernel_impl.h

@@ -31,8 +31,8 @@ void SpectralNormGradKernel(const Context& dev_ctx,
   auto& place = *dev_ctx.eigen_device();
   auto blas = phi::funcs::GetBlas<Context, T>(dev_ctx);
 
-  const int h = u.dims()[0];
-  const int w = v.dims()[0];
+  const int64_t h = u.dims()[0];
+  const int64_t w = v.dims()[0];
 
   DenseTensor weight_mat, out_grad_mat;
   auto dims = weight.dims();