You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Suggesting a change to the changelog for CUTLASS 3.3
Current ChangeLog
New [Mixed Precision Hopper GEMMs](https://github.com/NVIDIA/cutlass/blob/main/examples/55_hopper_mixed_dtype_gemm) support covering 16-bit x 8-bit input types with optimal performance.
New [Mixed Precision Ampere GEMMs](https://github.com/NVIDIA/cutlass/commit/7d8317a63e0a978a8dbb3c1fb7af4dbe4f286616) with support for canonical layouts (TN) and {fp16, bf16} x {s8/u8}. They also include fast numeric conversion recipes and warp level shuffles to achieve optimal performance.
Suggested change
New [Mixed-Input Hopper GEMMs](https://github.com/NVIDIA/cutlass/blob/main/examples/55_hopper_mixed_dtype_gemm) support covering 16-bit x 8-bit input types with optimal performance.
New [Mixed-Input Ampere GEMMs](https://github.com/NVIDIA/cutlass/pull/1084) with support for canonical layouts (TN) and {fp16, bf16} x {s8/u8}. They also include fast numeric conversion recipes and warp level shuffles to achieve optimal performance.
Summary and rationale for the suggested changes
Mixed-Precision to Mixed-Input. Mixed-Precision is taken by the GEMM data-type where inputs (DataType(operandA) == DataType(operandB) are mixed with a different accumulation data type (F16*F16+F32 and BF16*BF16+F32). The code uses cutlass::arch::OpMultiplyAddMixedInputUpcast tag to navigate and communicate that input data types are mixed. It would be good to set a nomenclature that is consistent and distinguishes between Mixed-Precision and Mixed-Input use-case.
Update the hyperlink for Mixed Precision Ampere GEMMs to the PR#1084 which has detailed description, steps to only compile Ampere mixed-input GEMMs, reproduce performance results, and a performance graph.
The text was updated successfully, but these errors were encountered:
This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.
Suggesting a change to the changelog for CUTLASS 3.3
Current ChangeLog
Suggested change
Summary and rationale for the suggested changes
Mixed-Precision
toMixed-Input
.Mixed-Precision
is taken by the GEMM data-type where inputs (DataType(operandA
) == DataType(operandB
) are mixed with a different accumulation data type (F16*F16+F32
andBF16*BF16+F32
). The code uses cutlass::arch::OpMultiplyAddMixedInputUpcast tag to navigate and communicate that input data types are mixed. It would be good to set a nomenclature that is consistent and distinguishes betweenMixed-Precision
andMixed-Input
use-case.Update the hyperlink for
Mixed Precision Ampere GEMMs
to the PR#1084 which has detailed description, steps to only compile Ampere mixed-input GEMMs, reproduce performance results, and a performance graph.The text was updated successfully, but these errors were encountered: