leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

Fixes for public issue #265

Browse Source
This commit is contained in:
Manikandan Ananth
2021-05-19 10:16:52 -07:00
parent b68113f5be
commit da2f110906
5 changed files with 93 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/cutlass/epilogue/thread/linear_combination_clamp.h
+7 -4
View File
@@ -186,11 +186,14 @@ public:
intermediate = mul_add_accumulator(alpha_, converted_accumulator, intermediate); // D = alpha * Accum + X
/// Clamping constant value
ElementCompute const kClamp =
ElementCompute((1U << (sizeof_bits<ElementOutput>::value - 1)) - 1);
ElementCompute const kClampMax =
ElementCompute(platform::numeric_limits<ElementOutput>::max());
intermediate = max_accumulator(intermediate, -kClamp - ElementCompute(1));
intermediate = min_accumulator(intermediate, kClamp);
ElementCompute const kClampMin =
ElementCompute(platform::numeric_limits<ElementOutput>::lowest());
intermediate = max_accumulator(intermediate, kClampMin);
intermediate = min_accumulator(intermediate, kClampMax);
// Convert to destination numeric type
NumericArrayConverter<ElementOutput, ElementCompute, kCount, Round> destination_converter;