Merge pull request #518 from notoraptor/fix-sger

nouiz · web-flow · commit 342016c39813 · 2017-09-08T19:40:53.000-04:00
Try to fix theano error related to sgerBatch.
diff --git a/src/gpuarray_blas_cuda_cublas.c b/src/gpuarray_blas_cuda_cublas.c
@@ -1636,9 +1636,9 @@ static int sgerBatch(cb_order order, size_t M, size_t N, float alpha,
 
 
   for (i = 0; i < batchCount; i++) {
-    GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(A[i], CUDA_WAIT_READ));
+    GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(A[i], CUDA_WAIT_ALL));
     GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(x[i], CUDA_WAIT_READ));
-    GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(y[i], CUDA_WAIT_ALL));
+    GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(y[i], CUDA_WAIT_READ));
   }
 
   cuda_exit(ctx);
@@ -1766,9 +1766,9 @@ static int dgerBatch(cb_order order, size_t M, size_t N, double alpha,
   }
 
   for (i = 0; i < batchCount; i++) {
-    GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(A[i], CUDA_WAIT_READ));
+    GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(A[i], CUDA_WAIT_ALL));
     GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(x[i], CUDA_WAIT_READ));
-    GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(y[i], CUDA_WAIT_ALL));
+    GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(y[i], CUDA_WAIT_READ));
   }
 
   cuda_exit(ctx);

Original file line number	Diff line number	Diff line change
`@@ -1636,9 +1636,9 @@ static int sgerBatch(cb_order order, size_t M, size_t N, float alpha,`
`1636`	`1636`
`1637`	`1637`
`1638`	`1638`	`for (i = 0; i < batchCount; i++) {`
`1639`		`- GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(A[i], CUDA_WAIT_READ));`
	`1639`	`+ GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(A[i], CUDA_WAIT_ALL));`
`1640`	`1640`	`GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(x[i], CUDA_WAIT_READ));`
`1641`		`- GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(y[i], CUDA_WAIT_ALL));`
	`1641`	`+ GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(y[i], CUDA_WAIT_READ));`
`1642`	`1642`	`}`
`1643`	`1643`
`1644`	`1644`	`cuda_exit(ctx);`
`@@ -1766,9 +1766,9 @@ static int dgerBatch(cb_order order, size_t M, size_t N, double alpha,`
`1766`	`1766`	`}`
`1767`	`1767`
`1768`	`1768`	`for (i = 0; i < batchCount; i++) {`
`1769`		`- GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(A[i], CUDA_WAIT_READ));`
	`1769`	`+ GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(A[i], CUDA_WAIT_ALL));`
`1770`	`1770`	`GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(x[i], CUDA_WAIT_READ));`
`1771`		`- GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(y[i], CUDA_WAIT_ALL));`
	`1771`	`+ GA_CUDA_EXIT_ON_ERROR(ctx, cuda_record(y[i], CUDA_WAIT_READ));`
`1772`	`1772`	`}`
`1773`	`1773`
`1774`	`1774`	`cuda_exit(ctx);`