Add MXDOTP operation group with FP4/FP6/FP8 source format support#17
Open
gamzeisl wants to merge 16 commits intopulp-platform:pulpfrom
Open
Add MXDOTP operation group with FP4/FP6/FP8 source format support#17gamzeisl wants to merge 16 commits intopulp-platform:pulpfrom
gamzeisl wants to merge 16 commits intopulp-platform:pulpfrom
Conversation
Extended fpnew_pkg.sv with new floating-point formats and MXDOTP operation group for MX dot product operations: - New formats: FP6(E3M2), FP6ALT(E2M3), FP4(E2M1) - Increased NUM_FP_FORMATS from 6 to 9 - Added MXDOTP operation group (6th group) - New operations: MXDOTPF (FP), MXDOTPI (INT) - Updated all format masks from 6-bit to 9-bit - Added bias_constant() helper function for MXDOTP - Updated FPU configurations (DEFAULT_NOREGS, DEFAULT_SNITCH)
Introduces fpnew_mxdotp_multi_pkg.sv with parameterized configuration for MXDOTP operations supporting mixed-precision arithmetic with low precision formats. Configuration: - Source formats: FP4, FP6, FP6ALT, FP8, FP8ALT, INT8 - Destination formats: FP32, FP16ALT
Add core MXDOTP implementation supporting very low-precision floating-point formats (FP4, FP6, FP8) and INT8. New files: - fpnew_mxdotp_multi_modules.sv: 14 modules implementing the MXDOTP datapath (classification, multiplication, shifting, accumulation, normalization, rounding) - fpnew_mxdotp_multi.sv: Top-level MXDOTP unit integrating all modules
New file: - fpnew_mxdotp_multi_wrapper.sv: Wrapper handling operand unpacking, FP6 extended operand processing (3-step with unroll factor), NaN-boxing, and scale extraction Changes to core module: - Add NumPipeRegs and PipeConfig as module parameters - Compute NUM_INP_REGS, NUM_MID_REGS, NUM_OUT_REGS from parameters
Add MX parameter and format-specific classification logic to support low-precision formats used in MXDOTP operations. Changes: - Add MX parameter (default 1) to enable MX-specific classification - FP8ALT (E4M3): No infinity, NaN when exp=all1s and man=all1s - FP6/FP6ALT/FP4 (E3M2/E2M3/E2M1): No infinity or NaN - Other formats: Standard IEEE-754 classification
- Add elaboration-time checks: fatal for Width!=64, missing FP32, missing FP8/INT8; warnings for inactive FP6/FP6ALT/FP4 - Add NUM_MX_LANES localparam and lane generation for MXDOTP - Instantiate fpnew_mxdotp_multi_wrapper with FpFmtConfig and IntFmtConfig
- Widen FpSrcFmtConfig bitmasks from 6b to 9b to match the extended NUM_FP_FORMATS (FP6, FP6ALT, FP4 added but masked off for SDOTP)
Relax format validation in fpnew_opgroup_multifmt_slice to require only FP8 and FP8ALT as mandatory base formats, allowing INT8 to be disabled.
There was a problem hiding this comment.
Pull request overview
This pull request introduces the MXDOTP (Microscaling Dot Product) operation group to FPnew, implementing support for the OCP MX specification's low-precision floating-point formats. The implementation adds three new FP formats (FP6, FP6ALT, FP4) and enables scaled dot-product-accumulate operations over these formats plus existing FP8/FP8ALT/INT8 formats.
Changes:
- Add FP6 (E3M2), FP6ALT (E2M3), and FP4 (E2M1) floating-point formats, expanding NUM_FP_FORMATS from 6 to 9
- Introduce MXDOTP operation group with MXDOTPF (FP sources) and MXDOTPI (INT8 sources) operations, producing FP32/FP16ALT results
- Extend fpnew_classifier with MX parameter to handle MX-specific special cases (no infinity/NaN for FP6/FP6ALT/FP4)
Reviewed changes
Copilot reviewed 14 out of 14 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
| src/fpnew_pkg.sv | Add new formats, operations, opgroup; expand format masks to 9 bits; add MxFpFmtMask/MxIntFmtMask configuration fields |
| src/fpnew_classifier.sv | Add MX parameter for MX-specific special case handling (FP8ALT: no inf; FP6/FP6ALT/FP4: no inf/NaN) |
| src/mxdotp/fpnew_mxdotp_multi_pkg.sv | Define MXDOTP-specific parameters, types, and helper functions |
| src/mxdotp/fpnew_mxdotp_multi_modules.sv | Implement 14 datapath modules for classification, multiplication, shifting, accumulation, normalization, and rounding |
| src/fpnew_mxdotp_multi.sv | Top-level MXDOTP unit with pipeline stages and datapath integration |
| src/fpnew_mxdotp_multi_wrapper.sv | Handle operand unpacking and FP6 3-step lane extraction with stateful counter |
| src/fpnew_opgroup_multifmt_slice.sv | Add MXDOTP lane generation with elaboration-time format validation checks |
| src/fpnew_opgroup_block.sv | Propagate MxFpFmtMask and MxIntFmtMask parameters |
| src/fpnew_top.sv | Pass MX format masks to opgroup blocks |
| src/fpnew_sdotp_multi_wrapper.sv | Update FpSrcFmtConfig masks from 6 to 9 bits |
| Bender.yml, src_files.yml | Add new MXDOTP source files to build system |
| docs/README.md | Document new operations, formats, and operation group |
| docs/CHANGELOG-PULP.md | Record additions and changes for MXDOTP feature |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
gamzeisl
force-pushed
the
feature/mxdotp_multi
branch
from
a78adf2 to
669ba20
Compare
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 14 out of 14 changed files in this pull request and generated no new comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Author
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Summary
Introduces the MXDOTP (Microscaling Dot Product) operation group to CVFPU, implementing scaled dot-product-accumulate over low-precision MX element formats as described in the OCP MX specification.
New formats
Three new floating-point formats are added to
fpnew_pkg, expandingNUM_FP_FORMATSfrom 6 to 9:New operation group
A sixth operation group (
MXDOTP) is added. It supports two operations:MXDOTPF– Scaled dot-product and accumulate for FP source elements (FP4, FP6, FP6ALT, FP8, FP8ALT)MXDOTPI– Scaled dot-product and accumulate for INT source elements (INT8)Both produce results in FP32 or FP16ALT and accept two 8-bit shared exponent scale factors (one per operand vector).
New source files
src/mxdotp/fpnew_mxdotp_multi_pkg.svsrc/mxdotp/fpnew_mxdotp_multi_modules.svsrc/fpnew_mxdotp_multi.svsrc/fpnew_mxdotp_multi_wrapper.svChanges to existing files
fpnew_pkg.sv: New formats, new opgroup, new operations, updated format masks (6→9 bits), updated default FPU configurationsfpnew_classifier.sv: MX-specific special cases (FP8ALT: no Inf; FP6/FP6ALT/FP4: no Inf, no NaN)fpnew_opgroup_multifmt_slice.sv: MXDOTP lane generation and wrapper instantiation; elaboration-time checks for mandatory formatsfpnew_opgroup_block.sv,fpnew_top.sv: Parameter propagation for the new opgroupfpnew_sdotp_multi_wrapper.sv: WidenedFpSrcFmtConfigmasks from 6 to 9 bitsBender.yml,src_files.yml: New source files addedKnown limitations
TODOcomments in the source files.