fix WinogradConvolution3 allocation size (LeelaChessZero#864)

Author: borg323

src/neural/blas/network_blas.cc

@@ -170,8 +170,7 @@ void BlasComputation::ComputeBlocking() {
   std::vector<float> res_buffer3(largest_batch_size * output_channels *
                                  kSquares);
 
-  WinogradConvolution3 convolve3(largest_batch_size, max_channels,
-                                 output_channels);
+  WinogradConvolution3 convolve3(largest_batch_size, max_channels);
 
   std::vector<float> policy_buffer(largest_batch_size *
                                    num_policy_input_planes * kSquares);

src/neural/blas/winograd_convolution3.cc

@@ -44,10 +44,9 @@ using ConstEigenMatrixMap =
 #endif
 
 WinogradConvolution3::WinogradConvolution3(const size_t max_batch_size,
-                                           const size_t max_input_layers,
-                                           const size_t max_output_layers)
-    : V_(max_batch_size * kWinogradTile * max_input_layers * kTiles),
-      M_(max_batch_size * kWinogradTile * max_output_layers * kTiles) {}
+                                           const size_t max_channels)
+    : V_(max_batch_size * kWinogradTile * max_channels * kTiles),
+      M_(max_batch_size * kWinogradTile * max_channels * kTiles) {}
 
 void WinogradConvolution3::Forward(const size_t batch_size,
                                    const size_t input_channels,

src/neural/blas/winograd_convolution3.h

@@ -39,9 +39,7 @@ class WinogradConvolution3 {
   // The instance will allocate memory resources for the
   // largest batch size, and the largest input and output
   // layers.
-  WinogradConvolution3(const size_t max_batch_size,
-                       const size_t max_input_layers,
-                       const size_t max_output_layers);
+  WinogradConvolution3(const size_t max_batch_size, const size_t max_channels);
 
   // Forward inference, batched.
   void Forward(const size_t batch_size, const size_t input_channels,