Files
composable_kernel/dispatcher/codegen/codegen_common.py
Vidyasagar Ananthan 86591de476 [rocm-libraries] ROCm/rocm-libraries#5260 (commit a1834d2)
[CK] [CK_Tile] Add FMHA scaffolding to CK kernel dispatcher
 (#5260)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

## Motivation

The CK Tile dispatcher currently supports GEMM and Grouped Convolution
but has no support for Fused Multi-Head Attention (FMHA). The
example/ck_tile/01_fmha folder contains a comprehensive FMHA
implementation with forward, backward, split-KV, paged-KV, append-KV,
and batch-prefill kernels across multiple GPU architectures — but there
is no unified dispatch layer for it. This PR ports the FMHA stack into
the dispatcher, following the same architectural patterns established by
GEMM and Grouped Convolution, enabling runtime kernel selection, JIT
compilation from Python, and a declarative C++ example flow. Autotuning
heuristics to follow.

## Technical Details

This PR adds FMHA scaffolding to the CK dispatcher framework, mirroring
GEMM's layered architecture. Seven new C++ runtime headers provide type
definitions (coexisting with upstream headers via __has_include,
requiring zero modifications to example/ck_tile/01_fmha/), a problem
builder with 18+ setters, Signature + Algorithm kernel key matching, a
virtual kernel instance, a DECL_FMHA_KERNEL_SET macro with wildcard
support and named tile/wave/warp setters, arch-aware registry with JSON
export, and a dispatcher with seqtune-aware selection, configurable
timing, and multi-stage execution plans for split-KV (two-stage) and
backward (three-stage). The codegen pipeline is driven by a
fmha_arch_specs.json capturing per-arch tile tables and pipeline
constraints for five architectures (gfx90a/942/950/1100/1201), migrated
from hardcoded logic in 01_fmha/codegen/, with supporting modules for
C++ symbol mappings, validation rules, and named receipt profiles
(ck_default, flash, pytorch, aiter, fp32, fp8). Python integration
(fmha_utils.py) mirrors the C++ layer with JIT compilation, parallel
multi-kernel builds, HIP memory management via ctypes, tolerance-based
validation, and a NumPy CPU reference with GQA support. Twenty-seven C++
and thirty-two Python examples cover the full feature surface — forward,
split-KV, masks, bias, dropout, GQA, backward, append-KV, batch prefill,
fp8, logits soft cap, sink tokens, and parameter sweeps — all
JIT-compiled on the fly.

## Test Plan

Seven test files cover the runtime types, codegen, and end-to-end
correctness. C++ unit tests validate the problem builder, dispatcher
planning (single-stage for forward/paged-KV/append-KV; multi-stage for
split-KV and backward), registry operations, and the kernel-set
declaration macro. Python unit tests verify codegen emission, profile
filtering, and 15 validation rules for masks, hdim constraints, and
pipeline requirements. GPU execution validation in 01_basic_fmha
--validate reports zero errors across 65,536 elements with max absolute
error of 7.29e-05. A gold-standard parity suite (test_fmha_parity.py)
runs 14 configurations through both the upstream tile_example_fmha_fwd
and the dispatcher, comparing exit codes to confirm behavioral parity —
all 14 match.

## Test Result

The C++ smoke test builds and passes all 9 compiled examples, and a
Python JIT sweep (29_sweep_seqlen.py) passes 7/7 configurations reaching
up to 375 TFLOPS at seqlen 2048.

## Submission Checklist

- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
2026-05-17 07:30:33 +00:00

351 lines
11 KiB
Python

#!/usr/bin/env python3
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
"""
Shared codegen infrastructure for GEMM, grouped convolution, and FMHA code generators.
Extracted from unified_gemm_codegen.py + arch-aware expansion helpers from conv.
Both unified_gemm_codegen.py and unified_grouped_conv_codegen.py import from here
to eliminate duplication.
"""
import logging
import concurrent.futures
from dataclasses import dataclass
from typing import (
Callable,
ClassVar,
Dict,
FrozenSet,
List,
Optional,
Sequence,
Tuple,
TypeVar,
)
log = logging.getLogger(__name__)
T = TypeVar("T")
R = TypeVar("R")
ANY_INT = -1
# ============================================================================
# Tile and Trait Configuration (shared between GEMM and Conv)
# ============================================================================
@dataclass
class TileConfig:
"""Tile configuration parameters shared by GEMM and grouped conv."""
tile_m: int
tile_n: int
tile_k: int
warp_m: int
warp_n: int
warp_k: int
warp_tile_m: int
warp_tile_n: int
warp_tile_k: int
def is_valid(self) -> bool:
if self.tile_m <= 0 or self.tile_n <= 0 or self.tile_k <= 0:
return False
return (
self.tile_m % (self.warp_m * self.warp_tile_m) == 0
and self.tile_n % (self.warp_n * self.warp_tile_n) == 0
and self.tile_k % (self.warp_k * self.warp_tile_k) == 0
)
@dataclass
class TraitConfigBase:
"""
Base kernel trait configuration shared by GEMM and grouped conv.
GEMM extends this with ``persistent``; grouped conv extends with
``double_smem_buffer`` and ``num_groups_to_merge``.
"""
pipeline: str # mem, compv3, compv4, compv5, ...
epilogue: str # cshuffle, default
scheduler: str # intrawave, interwave
pad_m: bool
pad_n: bool
pad_k: bool
# Unsupported (pipeline, epilogue, scheduler) combinations.
# Only 'mem' and 'basic_v1' pipelines support interwave; all compute
# pipelines (compv3/v4/v5/v6/async) only support intrawave.
_UNSUPPORTED: ClassVar[FrozenSet] = frozenset(
{
("compv3", "cshuffle", "interwave"),
("compv3", "default", "interwave"),
("compv4", "cshuffle", "interwave"),
("compv4", "default", "interwave"),
("compv5", "cshuffle", "interwave"),
("compv5", "default", "interwave"),
("compv6", "cshuffle", "interwave"),
("compv6", "default", "interwave"),
("comp_async", "cshuffle", "interwave"),
("comp_async", "default", "interwave"),
("basic_async_v1", "cshuffle", "interwave"),
("basic_async_v1", "default", "interwave"),
}
)
def is_valid(self) -> bool:
return (self.pipeline, self.epilogue, self.scheduler) not in self._UNSUPPORTED
# ============================================================================
# Type Mappings (centralized for both GEMM and conv codegen)
# ============================================================================
class CommonTypeMappings:
"""Centralized type mappings shared by GEMM and grouped conv codegen."""
DTYPE_TO_CK = {
"fp16": "fp16_t",
"bf16": "bf16_t",
"fp32": "float",
"fp8": "fp8_t",
"bf8": "bf8_t",
"int8": "int8_t",
}
DTYPE_TO_CK_QUALIFIED = {
"fp16": "ck_tile::fp16_t",
"bf16": "ck_tile::bf16_t",
"fp32": "float",
"fp8": "ck_tile::fp8_t",
"bf8": "ck_tile::bf8_t",
"int8": "int8_t",
}
DTYPE_TO_DISPATCHER = {
"fp16": "DataType::FP16",
"bf16": "DataType::BF16",
"fp32": "DataType::FP32",
"fp8": "DataType::FP8",
"bf8": "DataType::BF8",
"int8": "DataType::INT8",
}
# GEMM-specific layout mappings ("r"/"c" for row/column major).
# Convolution layouts (NHWGC, GKYXC, etc.) are handled by
# unified_grouped_conv_codegen.py via GroupedConvLayout / GroupedConvTypeMappings.
GEMM_LAYOUT_TO_CK = {
"r": "tensor_layout::gemm::RowMajor",
"c": "tensor_layout::gemm::ColumnMajor",
}
LAYOUT_TO_CK = GEMM_LAYOUT_TO_CK # backward compat alias
GEMM_LAYOUT_TO_DISPATCHER = {
"r": "LayoutTag::RowMajor",
"c": "LayoutTag::ColMajor",
}
LAYOUT_TO_DISPATCHER = GEMM_LAYOUT_TO_DISPATCHER # backward compat alias
# GEMM-only pipeline mappings (used by unified_gemm_codegen.py).
# Convolution pipelines are in GroupedConvTypeMappings
# (unified_grouped_conv_codegen.py). CK Tile conv supports:
# BASIC_V1, Mem, CompV3, CompV4, CompV5, CompV6, ASYNC_V1, ASYNC_V4.
# The dispatcher currently generates: mem, compv3, compv4.
# preshufflev2 is GEMM-only (weight pre-shuffle for GEMM, not conv).
PIPELINE_TO_CK = {
"mem": "GemmPipelineAgBgCrMem",
"compv3": "GemmPipelineAgBgCrCompV3",
"compv4": "GemmPipelineAgBgCrCompV4",
"compv5": "GemmPipelineAgBgCrCompV5",
"preshufflev2": "WeightPreshufflePipelineAGmemBGmemCRegV2",
}
PIPELINE_TO_BASE = {
"mem": "BaseGemmPipelineAgBgCrMem",
"compv3": "BaseGemmPipelineAgBgCrCompV3",
"compv4": "BaseGemmPipelineAgBgCrCompV4",
"compv5": "BaseGemmPipelineAgBgCrCompV5",
"preshufflev2": "BaseWeightPreshufflePipelineAGmemBGmemCRegV2",
}
PIPELINE_TO_DISPATCHER = {
"mem": "Pipeline::Mem",
"compv3": "Pipeline::CompV3",
"compv4": "Pipeline::CompV4",
"compv5": "Pipeline::CompV5",
"preshufflev2": "Pipeline::PreShuffleV2",
}
SCHEDULER_TO_CK = {
"intrawave": "GemmPipelineScheduler::Intrawave",
"interwave": "GemmPipelineScheduler::Interwave",
"default": "GemmPipelineScheduler::Default",
}
SCHEDULER_TO_DISPATCHER = {
"intrawave": "Scheduler::Intrawave",
"interwave": "Scheduler::Interwave",
"default": "Scheduler::Auto",
}
EPILOGUE_TO_DISPATCHER = {
"cshuffle": "Epilogue::CShuffle",
"default": "Epilogue::Default",
}
@staticmethod
def get_output_dtype(dtype: str) -> str:
"""Get output datatype (fp8/bf8 -> fp16)."""
return "fp16" if dtype in ("fp8", "bf8") else dtype
# ============================================================================
# Code Generation Helpers
# ============================================================================
def generate_cpp_compilation_unit(kernel_name: str) -> str:
"""Generate a .cpp compilation unit that includes a kernel header.
This is the standard pattern: one .cpp per kernel that just includes
the generated .hpp header, causing template instantiation.
"""
return (
f"// Auto-generated compilation unit for {kernel_name}\n"
f'#include "{kernel_name}.hpp"\n'
)
def parallel_generate(
generate_fn: Callable[[T], R],
items: Sequence[T],
parallel: bool = True,
) -> List[R]:
"""Run ``generate_fn`` over ``items``, optionally in parallel.
Logs per-item progress (best-of-conv pattern).
Returns a flat list of results in completion order.
"""
results: List[R] = []
if not items:
return results
if parallel and len(items) > 1:
with concurrent.futures.ThreadPoolExecutor() as executor:
futures = {executor.submit(generate_fn, item): item for item in items}
for future in concurrent.futures.as_completed(futures):
result = future.result()
results.append(result)
log.info("Generated: %s", futures[future])
else:
for item in items:
result = generate_fn(item)
results.append(result)
log.info("Generated: %s", item)
return results
# ============================================================================
# Arch-Aware Expansion Helpers (adopted from conv kernel_decl.hpp)
# ============================================================================
# These load from arch_specs_generated when available, falling back to
# hardcoded defaults that match the most common arch (gfx942).
_arch_data_cache: Optional[Dict] = None
def _get_arch_data() -> Dict:
"""Load arch filter data, with caching."""
global _arch_data_cache
if _arch_data_cache is not None:
return _arch_data_cache
try:
from arch_specs_generated import (
WARP_SUPPORTED_COMBINATIONS,
WARP_TILE_SUPPORTED_COMBINATIONS,
TRAIT_UNSUPPORTED_COMBINATIONS,
get_supported_archs,
)
_arch_data_cache = {
"warp_combos": WARP_SUPPORTED_COMBINATIONS,
"warp_tile_combos": WARP_TILE_SUPPORTED_COMBINATIONS,
"trait_unsupported": TRAIT_UNSUPPORTED_COMBINATIONS,
"supported_archs": get_supported_archs(),
}
except ImportError:
_arch_data_cache = {
"warp_combos": {
"gfx942": [[1, 4, 1], [2, 2, 1], [4, 1, 1]],
"gfx90a": [[1, 4, 1], [2, 2, 1], [4, 1, 1]],
},
"warp_tile_combos": {
"gfx942": {"fp16_fp16_fp32": [[16, 16, 16], [32, 32, 16]]},
"gfx90a": {"fp16_fp16_fp32": [[16, 16, 16], [32, 32, 16]]},
},
"trait_unsupported": {
("compv3", "cshuffle", "interwave"),
("compv4", "cshuffle", "interwave"),
},
"supported_archs": ["gfx90a", "gfx942", "gfx950"],
}
return _arch_data_cache
def valid_wave_configs(arch: str) -> List[List[int]]:
"""Return valid [wave_m, wave_n, wave_k] combos for *arch*."""
data = _get_arch_data()
return data["warp_combos"].get(arch, [[2, 2, 1]])
def valid_warp_configs(arch: str, dtype: str) -> List[List[int]]:
"""Return valid [warp_tile_m, warp_tile_n, warp_tile_k] combos for *arch*/*dtype*.
The dtype key is constructed as ``{dtype}_{dtype}_{acc}`` where acc is
fp32 for float types and int32 for int8.
"""
data = _get_arch_data()
acc = "int32" if dtype == "int8" else "fp32"
dtype_key = f"{dtype}_{dtype}_{acc}"
arch_tiles = data["warp_tile_combos"].get(arch, {})
return arch_tiles.get(dtype_key, [[32, 32, 16]])
def valid_trait_configs() -> List[Tuple[str, str]]:
"""Return valid (pipeline, scheduler) pairs.
Compute pipelines only support intrawave; mem supports both.
"""
return [
("compv3", "intrawave"),
("compv4", "intrawave"),
("compv5", "intrawave"),
("mem", "intrawave"),
("mem", "interwave"),
]
def needs_wave_expansion(config: dict) -> bool:
"""True if wave_m or wave_n is a wildcard (ANY_INT = -1)."""
return config.get("wave_m", 2) == ANY_INT or config.get("wave_n", 2) == ANY_INT
def needs_warp_expansion(config: dict) -> bool:
"""True if warp_m or warp_n is a wildcard (ANY_INT = -1)."""
return config.get("warp_m", 32) == ANY_INT or config.get("warp_n", 32) == ANY_INT
def needs_pipeline_expansion(config: dict) -> bool:
"""True if pipeline is a wildcard (\"*\")."""
return config.get("pipeline", "compv4") == "*"