Docs: Add precision support reference page #2111
Open
+369
−0
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
adeljo-amd
requested review from
babakpst,
yoichiyoshida,
bragadeesh and
AlexBrownAMD
as code owners
neon60
approved these changes
Mar 18, 2025
SwRaw
reviewed
Mar 26, 2025
| .. _precision-support: | ||
|
|
||
| ******************************** | ||
| Precision Support |
There was a problem hiding this comment.
Suggested change
| Precision Support | |
| Precision support |
| ******************************** | ||
|
|
||
| Tensile supports a rich variety of data types for matrix multiplication operations, enabling optimized performance | ||
| across different precision requirements. This document outlines the supported data types and precision formats |
There was a problem hiding this comment.
Suggested change
| across different precision requirements. This document outlines the supported data types and precision formats | |
| across different precision requirements. This topic outlines the supported data types and precision formats |
| across different precision requirements. This document outlines the supported data types and precision formats | ||
| used in Tensile's GEMM implementations. | ||
|
|
||
| Data Types |
There was a problem hiding this comment.
Suggested change
| Data Types | |
| Data types |
| - ``__hip_fp8_e5m2`` / ``__hip_fp8_e5m2_fnuz`` | ||
| - 8-bit | ||
| - | Brain float8 format with 5 exponent bits, 2 mantissa bits, and 1 sign bit. Provides greater dynamic range than | ||
| | F8 at the cost of reduced precision. |
There was a problem hiding this comment.
Suggested change
| | F8 at the cost of reduced precision. | |
| F8 at the cost of reduced precision. |
| * - X | ||
| - N/A | ||
| - 32-bit | ||
| - | Tensorfloat equivalent with custom bit distribution for enhanced precision in specific computation patterns |
There was a problem hiding this comment.
Suggested change
| - | Tensorfloat equivalent with custom bit distribution for enhanced precision in specific computation patterns | |
| - | Tensorfloat equivalent to custom bit distribution. Used for enhanced precision in specific computation patterns |
| - # SGEMM | ||
| - {M: 5504, N: 5504, K: 5504, transposeA: false, transposeB: true, dataType: S} | ||
|
|
||
| **Half-Precision with Single-Precision Accumulation** |
There was a problem hiding this comment.
Suggested change
| **Half-Precision with Single-Precision Accumulation** | |
| **Half-precision with single-precision accumulation** |
| - # GEMM_EX (HHS) | ||
| - {M: 5504, N: 5504, K: 5504, transposeA: false, transposeB: true, dataType: H, destDataType: H, computeDataType: S} | ||
|
|
||
| **BFloat16 Input with Float32 Output** |
There was a problem hiding this comment.
Suggested change
| **BFloat16 Input with Float32 Output** | |
| **BFloat16 input with Float32 output** |
| - # GEMM_EX (BSS) | ||
| - {M: 4096, N: 4096, K: 4096, transposeA: false, transposeB: true, dataType: B, destDataType: S, computeDataType: S} | ||
|
|
||
| **8-bit Integer Operations** |
There was a problem hiding this comment.
Suggested change
| **8-bit Integer Operations** | |
| **8-bit integer operations** |
| - # GEMM_EX (I8II) | ||
| - {M: 4096, N: 4096, K: 4096, transposeA: false, transposeB: true, dataType: I8, destDataType: I, computeDataType: I} | ||
|
|
||
| **Mixed F8/B8 Input with Half Precision Output** |
There was a problem hiding this comment.
Suggested change
| **Mixed F8/B8 Input with Half Precision Output** | |
| **Mixed F8/B8 input with half precision output** |
| - # GEMM_EX | ||
| - {M: 5504, N: 5504, K: 5504, transposeA: false, transposeB: true, dataType: F8B8, destDataType: H, computeDataType: S} | ||
|
|
||
| Library Logic File Naming |
There was a problem hiding this comment.
Suggested change
| Library Logic File Naming | |
| Library logic file naming |
SwRaw
reviewed
Mar 26, 2025
| - S | ||
| - Matrix A is bfloat8, Matrix B is float8, with half precision output | ||
|
|
||
| Configuration in Tensile |
There was a problem hiding this comment.
Is this heading apt? How about "Data types in configuration files"?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
resolves #___
Summary:
Add precison support reference page which details supported data types in Tensile.
Outcomes:
Only affects documentation side of this project.
Notable changes:
Only addition of one RST file, and modification of two files to add a link.