ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-20 21:09:08 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
86591de476dbf8d16a3dce849fda0d2887220ffa
composable_kernel/dispatcher/examples/fmha/cpp/16_heuristics_fmha.cpp
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

429 lines
16 KiB
C++
Raw Blame History

// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
//
// Example 16: FMHA Heuristic-Based Kernel Selection
//
// Demonstrates:
// 1. Two kernels with different tile_m0 (128 vs 64) and selection_rank
// 2. Custom heuristic function that picks kernels based on seqlen
// 3. dispatcher.set_heuristic() + SelectionStrategy::Heuristic
// 4. Planning different problems to show which kernel is selected
// 5. GPU execution for at least one problem
//
// Usage:
// ./16_heuristics_fmha
// ./16_heuristics_fmha --arch gfx942
#include <hip/hip_runtime.h>
#include <cmath>
#include <iomanip>
#include <iostream>
#include <random>
#include <vector>
#include "ck_tile/dispatcher.hpp"
#include "ck_tile/dispatcher/example_args.hpp"
using namespace ck_tile::dispatcher;
using namespace ck_tile::dispatcher::utils;
using FmhaDataType = ck_tile::fp16_t;
DECL_FMHA_KERNEL_SET(heuristic_fmha_kernels,
// Kernel A: Large tile (128x128) -- better for long sequences
.add(FmhaSignature()
.family("fwd")
.dtype("fp16")
.mode("batch")
.vlayout("r")
.hdim(128)
.mask("no")
.bias("no")
.lse(false)
.dropout(false)
.qscale("no"),
FmhaAlgorithm()
// Stage 0 (Q*K^T): m0=seqlen_q, n0=seqlen_k, k0=hdim_q
.tile_m0(128)
.tile_n0(128)
.tile_k0(32)
// Stage 1 (Attn*V): n1=hdim_v, k1=seqlen_k, k0max=alignment
.tile_n1(128)
.tile_k1(32)
.tile_k0max(128)
.wave_m0(4)
.wave_n0(1)
.wave_k0(1)
.wave_m1(4)
.wave_n1(1)
.wave_k1(1)
.warp_m0(32)
.warp_n0(32)
.warp_k0(16)
.warp_m1(32)
.warp_n1(32)
.warp_k1(16)
.pipeline("qr_async")
.padding(true, true, true, true)
.alignments(128, 128)
.selection_rank(0),
"gfx950")
// Kernel B: Smaller tile_m0 (64x128) -- lower latency for short sequences
.add(FmhaSignature()
.family("fwd")
.dtype("fp16")
.mode("batch")
.vlayout("r")
.hdim(128)
.mask("no")
.bias("no")
.lse(false)
.dropout(false)
.qscale("no"),
FmhaAlgorithm()
.tile_m0(64)
.tile_n0(128)
.tile_k0(32)
.tile_n1(128)
.tile_k1(32)
.tile_k0max(128)
.wave_m0(4)
.wave_n0(1)
.wave_k0(1)
.wave_m1(4)
.wave_n1(1)
.wave_k1(1)
.warp_m0(32)
.warp_n0(32)
.warp_k0(16)
.warp_m1(32)
.warp_n1(32)
.warp_k1(16)
.pipeline("qr_async")
.padding(true, true, true, true)
.alignments(128, 128)
.selection_rank(1),
"gfx950"));
namespace {
void cpu_attention_fwd(const std::vector<float>& Q,
const std::vector<float>& K,
const std::vector<float>& V,
std::vector<float>& O,
int batch,
int nhead,
int seqlen_q,
int seqlen_k,
int hdim_q,
int hdim_v,
float scale)
{
for(int b = 0; b < batch; ++b)
{
for(int h = 0; h < nhead; ++h)
{
for(int sq = 0; sq < seqlen_q; ++sq)
{
std::vector<float> scores(seqlen_k, 0.0f);
float max_score = -1e30f;
for(int sk = 0; sk < seqlen_k; ++sk)
{
float dot = 0.0f;
for(int d = 0; d < hdim_q; ++d)
{
int q_idx = ((b * nhead + h) * seqlen_q + sq) * hdim_q + d;
int k_idx = ((b * nhead + h) * seqlen_k + sk) * hdim_q + d;
dot += Q[q_idx] * K[k_idx];
}
scores[sk] = dot * scale;
max_score = std::max(max_score, scores[sk]);
}
float sum_exp = 0.0f;
for(int sk = 0; sk < seqlen_k; ++sk)
{
scores[sk] = std::exp(scores[sk] - max_score);
sum_exp += scores[sk];
}
for(int sk = 0; sk < seqlen_k; ++sk)
{
scores[sk] /= sum_exp;
}
for(int dv = 0; dv < hdim_v; ++dv)
{
float acc = 0.0f;
for(int sk = 0; sk < seqlen_k; ++sk)
{
int v_idx = ((b * nhead + h) * seqlen_k + sk) * hdim_v + dv;
acc += scores[sk] * V[v_idx];
}
int o_idx = ((b * nhead + h) * seqlen_q + sq) * hdim_v + dv;
O[o_idx] = acc;
}
}
}
}
}
} // namespace
int main(int argc, char* argv[])
{
ExampleArgs args("Example 16: FMHA Heuristic Kernel Selection",
"Custom heuristic picks kernel based on seqlen");
args.add_option("--arch", "gfx950", "GPU architecture");
args.add_option("--nhead", "8", "Number of heads");
args.add_option("--hdim", "128", "Head dimension");
if(!args.parse(argc, argv))
return 0;
const std::string gfx_arch = args.get("--arch", "gfx950");
const int nhead = args.get_int("--nhead", 8);
const int hdim = args.get_int("--hdim", 128);
print_header("Example 16: FMHA Heuristic Kernel Selection");
// Step 1: Register kernels
std::cout << "\nStep 1: Register Kernels\n";
FmhaKernelSetRegistry::instance().print();
FmhaRegistry registry;
registry.set_name("heuristic_fmha");
REGISTER_GENERATED_KERNELS(registry, gfx_arch);
std::cout << " Registered " << registry.size() << " kernel(s)\n";
// Step 2: Set up heuristic
std::cout << "\nStep 2: Configure Heuristic\n";
std::cout << " Rule: seqlen >= 256 -> prefer large tile (128x128, rank=0)\n";
std::cout << " seqlen < 256 -> prefer small tile (64x128, rank=1)\n";
auto all_kernels = registry.all_kernels();
std::cout << " Available kernels:\n";
for(const auto& k : all_kernels)
{
std::cout << " - " << k->id() << "\n";
}
std::string kernel_a_id, kernel_b_id;
for(const auto& k : all_kernels)
{
auto kid = k->id();
if(kernel_a_id.empty())
kernel_a_id = kid;
else if(kernel_b_id.empty())
kernel_b_id = kid;
}
FmhaDispatcher dispatcher(&registry);
dispatcher.set_strategy(SelectionStrategy::Heuristic);
dispatcher.set_heuristic([&](const FmhaProblem& problem) -> std::vector<std::string> {
if(problem.seqlen_q >= 256)
return {kernel_a_id, kernel_b_id};
else
return {kernel_b_id, kernel_a_id};
});
dispatcher.set_benchmarking(true);
dispatcher.set_timing(1, 3);
// Step 3: Plan different problems to show kernel selection
std::cout << "\nStep 3: Plan Problems (show kernel selection)\n\n";
struct PlanCase
{
int batch;
int seqlen;
};
PlanCase plan_cases[] = {{1, 64}, {1, 128}, {2, 256}, {2, 512}, {4, 1024}};
std::cout << " " << std::setw(6) << "Batch" << " | " << std::setw(8) << "SeqLen" << " | "
<< std::setw(50) << "Selected Kernel" << "\n";
std::cout << " " << std::string(68, '-') << "\n";
for(const auto& pc : plan_cases)
{
auto problem = FmhaProblemBuilder()
.api_family(FmhaApiFamily::Fwd)
.kernel_family(FmhaKernelFamily::Fwd)
.gfx_arch(gfx_arch)
.data_type("fp16")
.dims(hdim, hdim, pc.batch, pc.seqlen, pc.seqlen)
.nheads(nhead, nhead)
.mask_type(0)
.bias_type(0)
.lse(false)
.dropout(false)
.build();
auto plan = dispatcher.plan(problem);
std::string selected = plan.is_valid() ? plan.stages[0].kernel_id : "(no match)";
std::cout << " " << std::setw(6) << pc.batch << " | " << std::setw(8) << pc.seqlen << " | "
<< std::setw(50) << selected << "\n";
}
// Step 4: GPU execution for a representative problem
std::cout << "\nStep 4: GPU Execution (batch=2, seqlen=256)\n";
const int batch = 2;
const int seqlen = 256;
const float scale = 1.0f / std::sqrt(static_cast<float>(hdim));
const int64_t elems = static_cast<int64_t>(batch) * nhead * seqlen * hdim;
GpuBuffer<FmhaDataType> q_dev(elems);
GpuBuffer<FmhaDataType> k_dev(elems);
GpuBuffer<FmhaDataType> v_dev(elems);
GpuBuffer<FmhaDataType> o_dev(elems);
std::mt19937 rng(42);
std::uniform_real_distribution<float> fdist(-0.5f, 0.5f);
std::vector<FmhaDataType> q_host(elems), k_host(elems), v_host(elems);
for(auto& x : q_host)
x = FmhaDataType(fdist(rng));
for(auto& x : k_host)
x = FmhaDataType(fdist(rng));
for(auto& x : v_host)
x = FmhaDataType(fdist(rng));
q_dev.copy_from_host(q_host.data());
k_dev.copy_from_host(k_host.data());
v_dev.copy_from_host(v_host.data());
o_dev.zero();
fmha_fwd_traits traits{};
traits.hdim_q = hdim;
traits.hdim_v = hdim;
traits.data_type = "fp16";
traits.is_group_mode = false;
traits.is_v_rowmajor = true;
traits.has_logits_soft_cap = false;
traits.mask_type = mask_enum::no_mask;
traits.bias_type = bias_enum::no_bias;
traits.has_lse = false;
traits.has_dropout = false;
traits.qscale_type = quant_scale_enum::no_scale;
fmha_fwd_args fmha_args{};
fmha_args.q_ptr = q_dev.get();
fmha_args.k_ptr = k_dev.get();
fmha_args.v_ptr = v_dev.get();
fmha_args.o_ptr = o_dev.get();
fmha_args.bias_ptr = nullptr;
fmha_args.q_descale_ptr = nullptr;
fmha_args.k_descale_ptr = nullptr;
fmha_args.v_descale_ptr = nullptr;
fmha_args.rand_val_ptr = nullptr;
fmha_args.lse_ptr = nullptr;
fmha_args.sink_ptr = nullptr;
fmha_args.block_scale_seqstart_q_ptr = nullptr;
fmha_args.block_scale_seqstart_k_ptr = nullptr;
fmha_args.seqlen_q = seqlen;
fmha_args.seqlen_k = seqlen;
fmha_args.batch = batch;
fmha_args.max_seqlen_q = seqlen;
fmha_args.hdim_q = hdim;
fmha_args.hdim_v = hdim;
fmha_args.nhead_q = nhead;
fmha_args.nhead_k = nhead;
fmha_args.scale_s = scale;
fmha_args.logits_soft_cap = 0.0f;
fmha_args.stride_q = hdim;
fmha_args.stride_k = hdim;
fmha_args.stride_v = hdim;
fmha_args.stride_bias = 0;
fmha_args.stride_randval = 0;
fmha_args.stride_o = hdim;
fmha_args.nhead_stride_q = seqlen * hdim;
fmha_args.nhead_stride_k = seqlen * hdim;
fmha_args.nhead_stride_v = seqlen * hdim;
fmha_args.nhead_stride_bias = 0;
fmha_args.nhead_stride_randval = 0;
fmha_args.nhead_stride_lse = 0;
fmha_args.nhead_stride_o = seqlen * hdim;
fmha_args.nhead_stride_q_descale = 0;
fmha_args.nhead_stride_k_descale = 0;
fmha_args.nhead_stride_v_descale = 0;
fmha_args.batch_stride_q = nhead * seqlen * hdim;
fmha_args.batch_stride_k = nhead * seqlen * hdim;
fmha_args.batch_stride_v = nhead * seqlen * hdim;
fmha_args.batch_stride_bias = 0;
fmha_args.batch_stride_randval = 0;
fmha_args.batch_stride_lse = 0;
fmha_args.batch_stride_o = nhead * seqlen * hdim;
fmha_args.batch_stride_q_descale = 0;
fmha_args.batch_stride_k_descale = 0;
fmha_args.batch_stride_v_descale = 0;
fmha_args.window_size_left = -1;
fmha_args.window_size_right = -1;
fmha_args.sink_size = 0;
fmha_args.mask_type = 0;
fmha_args.min_seqlen_q = 0;
fmha_args.p_drop = 0.0f;
fmha_args.s_randval = false;
fmha_args.drop_seed_offset = std::make_pair(uint64_t(0), uint64_t(0));
fmha_args.block_scale_size_q = 0;
fmha_args.block_scale_size_kv = 0;
float time_ms = 0.0f;
bool passed = false;
try
{
time_ms = dispatcher.run_fwd(traits, fmha_args, nullptr);
auto problem =
FmhaProblem::from_invocation(FmhaInvocation::make(traits, fmha_args), gfx_arch);
double tflops = static_cast<double>(problem.num_ops()) / (time_ms * 1e-3) / 1e12;
std::cout << " Time: " << std::fixed << std::setprecision(4) << time_ms << " ms\n";
std::cout << " TFLOPS: " << std::setprecision(2) << tflops << "\n";
// Validate against CPU reference
std::vector<FmhaDataType> o_host(elems);
o_dev.copy_to_host(o_host.data());
std::vector<float> q_f32(elems), k_f32(elems), v_f32(elems), o_ref(elems, 0.0f);
for(int64_t i = 0; i < elems; ++i)
q_f32[i] = static_cast<float>(q_host[i]);
for(int64_t i = 0; i < elems; ++i)
k_f32[i] = static_cast<float>(k_host[i]);
for(int64_t i = 0; i < elems; ++i)
v_f32[i] = static_cast<float>(v_host[i]);
cpu_attention_fwd(
q_f32, k_f32, v_f32, o_ref, batch, nhead, seqlen, seqlen, hdim, hdim, scale);
double max_abs_err = 0.0;
int errors = 0;
for(int64_t i = 0; i < elems; ++i)
{
double abs_err = std::abs(static_cast<float>(o_host[i]) - o_ref[i]);
max_abs_err = std::max(max_abs_err, abs_err);
if(abs_err > 1e-2 + 1e-2 * std::abs(o_ref[i]))
++errors;
}
std::cout << " Max abs error: " << std::scientific << max_abs_err << "\n";
std::cout << " Errors: " << errors << " / " << elems << "\n";
passed = (errors == 0);
}
catch(const std::exception& e)
{
std::cerr << " ERROR: " << e.what() << "\n";
}
print_separator();
std::cout << "Status: " << (passed ? "PASS" : "FAIL") << "\n";
print_separator();
return passed ? 0 : 1;
}
Reference in New Issue View Git Blame Copy Permalink