Merge pull request #246 from mengchihe/master

support unalignment input for conv2d fprop stage=2 Fix for issue #242
NVIDIA · Sep 8, 2021 · 9ac2558 · 9ac2558
2 parents 6c2f8f2 + 59e2aa5
commit 9ac2558
Show file tree

Hide file tree

Showing 45 changed files with 1,922 additions and 533 deletions.
diff --git a/include/cutlass/conv/device/implicit_gemm_convolution.h b/include/cutlass/conv/device/implicit_gemm_convolution.h
@@ -105,6 +105,18 @@ class ImplicitGemmConvolution {
       return status;
     }
 
+    static int const kAlignmentC = ImplicitGemmKernel::Epilogue::OutputTileIterator::kElementsPerAccess;
+    if (kConvolutionalOperator == conv::Operator::kFprop) {
+      if (args.problem_size.K % kAlignmentC)
+        return Status::kErrorMisalignedOperand;
+    } else if (kConvolutionalOperator == conv::Operator::kDgrad) {
+       if (args.problem_size.C % kAlignmentC)
+        return Status::kErrorMisalignedOperand;
+    } else if (kConvolutionalOperator == conv::Operator::kWgrad) {
+       if (args.problem_size.C % kAlignmentC)
+        return Status::kErrorMisalignedOperand;
+    }
+
     // check for unsupported problem sizes for strided dgrad implementation
     if (kConvolutionalOperator == conv::Operator::kDgrad && 
       kStrideSupport == conv::StrideSupport::kStrided) {