[AMD] Enable scaled_dot(-, bf16) #5029
Merged
+16
−11
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
antiagainst
force-pushed
the
amd-scaled-dot-bf16
branch
from
b8346ba to
da6faa0
Compare
Luosuu
pushed a commit
to Luosuu/triton
that referenced
this pull request
Nov 13, 2024
guacamoleo
pushed a commit
to guacamoleo/triton
that referenced
this pull request
Nov 14, 2024
jataylo
pushed a commit
to jataylo/triton
that referenced
this pull request
Dec 13, 2024
(cherry picked from commit f062540)
jataylo
pushed a commit
to jataylo/triton
that referenced
this pull request
Dec 13, 2024
(cherry picked from commit f062540)
jataylo
added a commit
to ROCm/triton
that referenced
this pull request
Dec 13, 2024
* [BACKEND][NVIDIA] Add Lowering for Shared-to-MMAv3-DotOp Copy (triton-lang#5009) Allows for upcasting in DotOp encoding in RF. This lowering path is not currently in use; pending triton-lang#5003 (cherry picked from commit cfddb09) * [AMD] Add initial support for scaled_dot(mxfp8, fp8) (triton-lang#4994) This commit adds initial support for scaled_dot with mxfp8 LHS and fp8 RHS. It supports both mfma32 and mfma16 intrinsic variants. Right now we are missing software emulation for `Float8E4M3FN` type, so this only enables for `Float8E5M2`. (cherry picked from commit 3549db8) * [Frontend][Backend] Implement support for scale_dot(-, bf16) (triton-lang#4996) In the passing we also improve a few other things: - Now `scaled_dot` accepts both uint8/uint16 fp8/bf16 as inputs (before you had to cast it to uint8, which was weird when extending it to bf16). - Add `scaled_dot` to the docs and improve the docs overall (have not render them, might need a few further tweaks) (cherry picked from commit 23c9ec1) * [BACKEND] Improve detection of register to register conversion (triton-lang#4991) Specifically, it fixes problems when `srcLayout` and `dstLayout` have different number of registers but the same number of not free registers. We solved the problem by padding free registers to either `srcLayout` or `dstLayout`, but this can be improved by fixing the `invertAndCompose` function. (cherry picked from commit 15c5e55) * [BACKEND] Replace `isMmaToDotShortcut` with linear layout based logic (triton-lang#4951) This PR removes the legacy `isMmaToDotShortcut` and its associated shortcut conversion. (cherry picked from commit 1d5fdfe) * [BACKEND]Fix DotOperand(Ampere) LinearLayoutConversion (triton-lang#5038) We also clean a bit `TritonGPU/IR/Dialect.cpp` using some auxiliary functions to make the intentions a bit clearer. We add a few asserts in the `LinearLayoutConversion` to make sure it's clear why we do certain things here and there. We also kill `getCvtOrder`, as it was not used anywhere (cherry picked from commit 56584c4) * [BACKEND] Fix uses of getOrder(DotOperand(Nvidia) and MMA(Nvidia)) (triton-lang#5055) We use `getOrder` very liberally throughout the codebase, when we really meant to use `getThreadOrder`. This is an issue with the input layout is an `DotOperand(mma(opIdx=1))`, where the thread order and the matrix order are opposite. Found this to be an issue when a PR changed the `getOrder` of `DotOperand(Hopper)` to an incorrect one and CI still passed! The issue here is that the LLVM lowering for wgmma and the LinearLayout does not use `getOrder`, but there are many other subsystems do, and many heuristics would be getting an incorrect order, and potentially be disabled. This is particularly problematic for `DotOperand(opIdx=1)` in nvidia hardware, as `getThreadOrder` and `getOrder` are different! While doing so we: - Audit most (all?) the calls to `getOrder(dotOperand)`. It turns out that most of them really meant `getThreadOrder` - Fix the ordering methods of `SliceEncodingAttr` to be consistent - Move the implementation of `getWarpOrder` to the Attr classes, because of OOP The test strategy was to add `llvm::report_fatal_error("Testing");` within `getOrder(nvidiaMma)` and `getOrder(DotOperand(nvidiaMma))` and triaging all errors that were raised in CI. (cherry picked from commit 38a11b8) * [AMD] Reland instruction scheduling hint changes (triton-lang#4940) This commit relands triton-lang#4819 with the following fixes: * Changed to a better way to mark opIdx for loads * Replaced temlate-based `rewindUnaryOps` to use regular for-loops. The new way is more robust and can handle other unary ops automatically. * Replaced `instr.sched.barriers` using the ones from `rocdl` dialect from the MLIR upstream * Extended lit tests (cherry picked from commit ee5876c) * [AMD] Enable scaled_dot(-, bf16) (triton-lang#5029) (cherry picked from commit f062540) * [AMD] Add support for scaled_dot(mxfp4, -) (triton-lang#5034) This commit adds support for mxfp4 typed A tensor for sacled dot in the AMD backend. We moved the `convertMxfp4x2ToBf16x2` impl from NVIDIA side to a common path to reuse. (cherry picked from commit edc5c5c) * [BACKEND] Minor Bugfixes for SharedToDotOperand MMAv3 (triton-lang#5030) Two bugfixes following triton-lang#5009. - When `BLOCK_M=64` and `num_warps > 4`, the order of warps for DotOpEncoded tensor should be M-major instead of N-major, since WGMMA expects the 4 warps in each warp group to be stacked along the M dimension. - Should use `mmaBitwidth` instead of `bitwidth` when calculating `numRep` in `SharedToDotOperandMMAv2OrV3`. This was missed in a bad rebase. @lezcano I encountered these bugs when attempting to locally test the [DotOp hoisting PR](triton-lang#5003) after rebasing (they normally would be caught by `test_core.py` but that path was not yet enabled in the last PR). With these fixes added, I was able to successfully validate against pytorch. (cherry picked from commit e82dfd9) (cherry picked from commit 5287a68) * [BACKEND] Get rid of unpack/pack I32 (triton-lang#5044) - Removed functions related to unpacking and packing I32 values. - Updated utilities to handle conversion of mxfp4 values without packing/unpacking I32. - Move the register value ordering logic from the element-wise operation lowering to the dot operation lowering. - Use linear layout to handle conversions between almost all distributed layouts. - Clean up data loading and mma computation involving `repN`, `repK`, and `repM`. (cherry picked from commit 1cf7b1b) (cherry picked from commit 376fe7e) * Consolidate `getOrder` as "element order" and implement `getRepOrder` for general and NVIDIA layouts (triton-lang#5089) This partially reverts commit 38a11b8. Supersedes triton-lang#5085 It also documents that we are implicitly choosing a way to tile a full tensor depending on the layout. See triton-lang#5085 (comment) (cherry picked from commit 57643b3) (cherry picked from commit ffb2032) --------- Co-authored-by: Gary Geng <[email protected]> Co-authored-by: Lei Zhang <[email protected]> Co-authored-by: Mario Lezcano Casado <[email protected]> Co-authored-by: Keren Zhou <[email protected]> Co-authored-by: ravil-mobile <[email protected]>
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.
No description provided.