Files
composable_kernel/dispatcher/src/dispatcher.cpp
Vidyasagar Ananthan 40290297cd [CK] [CK_Tile] Add GroupConv to Kernel Dispatcher (#5168)
## Motivation

This PR adds CK Tile group convolution (forward, backward-data,
backward-weight) support to the kernel dispatcher, matching and unifying
with the existing dispatcher GEMM infrastructure in architecture and
usability. The dispatcher provides a unified kernel dispatch system with
both C++ and Python frontends, and until now only supported GEMM
operations. This PR enables framework integrators to use the same
declarative kernel workflow for convolutions as they do for GEMM:
declare kernels, build a registry JIT, select kernels within the
registry at runtime, and dispatch to GPU. Future PRs will include
runtime kernel selection heuristics for autotuning of kernel parameters
based on (problem, hardware arch).

## Technical Details

Grouped convolution support has been added to the CK Tile Dispatcher
with generated_conv_backend.hpp enabling dispatcher.run(in, wei, out,
problem) for all 6 conv variants (fwd/bwdd/bwdw x 2D/3D), runtime
heuristic kernel selection, and GroupedConvKernelKey with full
ConvConfigBase fields. Python side adds parallel JIT via
registry.build(max_workers) and heuristic registry.select(). Includes 7
C++ and 6 Python examples covering all directions with CPU reference
validation, and shared infrastructure improvements (BaseRegistry CRTP,
structured exceptions). As a sanity check, JIT compile times for a
single kernel remains the same and for multiple kernels there is better
parallelism:
Kernels | 1 worker | 8 workers
1 | 7.7 s | 7.7 s
2 | 15.9 s | 8.2 s
4 | 33.4 s | 9.7 s
6 | 52.3 s | 10.2 s

## Test Plan

145 ephemeral unit tests have been added to test basic functionality.
All 30 examples/integration tests run end-to-end on gfx950 (MI350): 7
C++ conv, 7 C++ GEMM, 6 Python conv, 10 Python GEMM. CPU reference
validation for forward, backward-data, and backward-weight (2D) in both
C++ and Python examples pass.

## Test Result

30 examples pass. Peak performance: 132 TFLOPS (Batch-32 forward 56x56),
53 TFLOPS (pointwise 1x1). CPU reference accuracy: max_abs_diff < 0.002
for all directions (fp16 vs fp32 reference).

## Submission Checklist

- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

---------

Co-authored-by: Yaswanth Raparti <113389104+yraparti@users.noreply.github.com>
2026-04-09 10:38:33 -07:00

154 lines
4.2 KiB
C++

// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "ck_tile/dispatcher/dispatcher.hpp"
#include "ck_tile/dispatcher/dispatcher_error.hpp"
#include <sstream>
#include <iostream>
namespace ck_tile {
namespace dispatcher {
Dispatcher::Dispatcher(Registry* registry, const std::string& gfx_arch)
: registry_(registry ? registry : &Registry::instance()),
heuristic_(nullptr),
strategy_(SelectionStrategy::FirstFit),
gfx_arch_(gfx_arch)
{
}
void Dispatcher::set_heuristic(HeuristicFunction heuristic)
{
heuristic_ = heuristic;
if(heuristic_)
{
strategy_ = SelectionStrategy::Heuristic;
}
}
void Dispatcher::set_strategy(SelectionStrategy strategy) { strategy_ = strategy; }
KernelInstancePtr Dispatcher::select_kernel(const Problem& problem) const
{
if(!problem.is_valid())
{
return nullptr;
}
switch(strategy_)
{
case SelectionStrategy::FirstFit: return select_first_fit(problem);
case SelectionStrategy::Heuristic: return select_heuristic(problem);
default: return nullptr;
}
}
float Dispatcher::run(
const void* a_ptr, const void* b_ptr, void* c_ptr, const Problem& problem, void* stream) const
{
return run_fused(a_ptr, b_ptr, c_ptr, nullptr, problem, stream);
}
float Dispatcher::run_fused(const void* a_ptr,
const void* b_ptr,
void* c_ptr,
const void** d_ptrs,
const Problem& problem,
void* stream) const
{
auto kernel = select_kernel(problem);
if(!kernel)
{
std::ostringstream oss;
oss << "No suitable kernel found for problem: M=" << problem.M << " N=" << problem.N
<< " K=" << problem.K;
throw NoKernelFound(oss.str());
}
return kernel->run(a_ptr, b_ptr, c_ptr, d_ptrs, problem, stream);
}
float Dispatcher::run_explicit(const std::string& kernel_id,
const void* a_ptr,
const void* b_ptr,
void* c_ptr,
const void** d_ptrs,
const Problem& problem,
void* stream) const
{
auto kernel = registry_->lookup(kernel_id);
if(!kernel)
{
throw NoKernelFound("Kernel not found: " + kernel_id);
}
if(!kernel->supports(problem))
{
std::ostringstream oss;
oss << "Kernel " << kernel_id << " does not support problem: M=" << problem.M
<< " N=" << problem.N << " K=" << problem.K;
throw UnsupportedProblem(oss.str());
}
return kernel->run(a_ptr, b_ptr, c_ptr, d_ptrs, problem, stream);
}
bool Dispatcher::validate(const void* a_ptr,
const void* b_ptr,
const void* c_ptr,
const void** d_ptrs,
const Problem& problem,
float tolerance) const
{
auto kernel = select_kernel(problem);
if(!kernel)
{
return false;
}
return kernel->validate(a_ptr, b_ptr, c_ptr, d_ptrs, problem, tolerance);
}
KernelInstancePtr Dispatcher::select_first_fit(const Problem& problem) const
{
auto all_kernels = registry_->get_all();
for(const auto& kernel : all_kernels)
{
if(kernel->supports(problem))
{
return kernel;
}
}
return nullptr;
}
KernelInstancePtr Dispatcher::select_heuristic(const Problem& problem) const
{
if(!heuristic_)
{
// Fall back to first-fit if no heuristic available
return select_first_fit(problem);
}
// Get ranked list of kernel identifiers from heuristic
auto candidates = heuristic_(problem);
// Try each candidate in order
for(const auto& kernel_id : candidates)
{
auto kernel = registry_->lookup(kernel_id);
if(kernel && kernel->supports(problem))
{
return kernel;
}
}
// If no heuristic candidate works, fall back to first-fit
return select_first_fit(problem);
}
} // namespace dispatcher
} // namespace ck_tile