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

Add fixed_channel and few_channel mode to int8 in generator (#829)

Browse Source
This commit is contained in:
Shuai Shao 2023-02-21 18:15:39 -08:00 committed by GitHub
parent 95f673ecf7
commit 9cdbe33570
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 8 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
tools/library/scripts/generator.py
View File

@ -526,6 +526,8 @@ def CreateConv2dFixedChannelsOperator(manifest, layout, tile_descriptions, data_
manifest.append(new_operation) manifest.append(new_operation)
operations.append(new_operation) operations.append(new_operation)
return operations
# Convolution for 2D operations specialized for few channels # Convolution for 2D operations specialized for few channels
@ -570,6 +572,8 @@ def CreateConv2dFewChannelsOperator(manifest, layout, tile_descriptions, data_ty
manifest.append(new_operation) manifest.append(new_operation)
operations.append(new_operation) operations.append(new_operation)
return operations
# Convolution for 3D operations # Convolution for 3D operations
def CreateConv3dOperator(manifest, layout, tile_descriptions, data_type, alignment, \ def CreateConv3dOperator(manifest, layout, tile_descriptions, data_type, alignment, \
@ -1467,6 +1471,10 @@ def GenerateSM75_TensorOp_8816_TN(manifest, cuda_version):
operations += CreateConv2dOperator(manifest, conv_layout, tile_descriptions, operations += CreateConv2dOperator(manifest, conv_layout, tile_descriptions,
data_type_mixed, alignment_constraints, [ConvKind.Fprop], EpilogueFunctor.LinearCombinationClamp) data_type_mixed, alignment_constraints, [ConvKind.Fprop], EpilogueFunctor.LinearCombinationClamp)
operations += CreateConv2dFixedChannelsOperator(manifest, conv_layout, tile_descriptions,
data_type_mixed, [4, 8, 16], [ConvKind.Fprop], EpilogueFunctor.LinearCombinationClamp)
operations += CreateConv2dFewChannelsOperator(manifest, conv_layout, tile_descriptions,
data_type_mixed, [4, 8, 16], [ConvKind.Fprop], EpilogueFunctor.LinearCombinationClamp)
for op in operations: for op in operations:
if op.tile_description.threadblock_shape[1] >= 128: if op.tile_description.threadblock_shape[1] >= 128: