ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK TILE] Support fp8/fp16 with pk_int4_t as data types for tensors A and B (#2805)

Browse Source 
- Add support for tensor A/B in both fp16+pk_int4_t and fp8+pk_int4_t formats
- Implement A(bf8) B(i4) support in universal GEMM
- Use new implementation for i4 to fp8 conversion in Block Scale
This commit is contained in:
Cong Ma
2025-09-09 17:40:52 -06:00
committed by GitHub
parent 91178b4011
commit 82890192dd
15 changed files with 320 additions and 135 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
example/ck_tile/03_gemm/gemm_utils.hpp
View File

@@ -344,6 +344,24 @@ struct GemmTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t>
using CDataType = ck_tile::half_t;
};
template <>
struct GemmTypeConfig<ck_tile::fp8_t, ck_tile::pk_int4_t, ck_tile::half_t>
{
using ADataType = ck_tile::fp8_t;
using BDataType = ck_tile::pk_int4_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmTypeConfig<ck_tile::bf8_t, ck_tile::pk_int4_t, ck_tile::half_t>
{
using ADataType = ck_tile::bf8_t;
using BDataType = ck_tile::pk_int4_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmTypeConfig<ck_tile::half_t, ck_tile::pk_int4_t, ck_tile::half_t>
{

59
example/ck_tile/03_gemm/run_gemm_example.inc
View File

@@ -1,6 +1,8 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/host/permute_pk_int4.hpp"
template <typename Layout>
static constexpr inline auto is_row_major(Layout layout_)
{
@@ -90,61 +92,6 @@ void permute_tensor_b(Tensor& tensor)
}
}
template <typename Tensor>
void permute_vectors_i4x4_b(Tensor& tensor)
{
const ck_tile::index_t K = tensor.get_length(0);
const ck_tile::index_t N = tensor.get_length(1);
// vector pk_i4x4 permute
for(int i = 0; i < N; i++)
{
for(int j = 0; j < K; j += 8)
{
int8_t input[8];
for(int k = 0; k < 4; k++)
{
int8_t i4x2 = tensor(j + k * 2, i).data;
input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
}
// permute 01234567->20643175
{
int8_t hi = input[2];
int8_t lo = input[0];
int8_t i4x2 = (hi << 4) | lo;
tensor(j + 0, i) = i4x2;
}
{
int8_t hi = input[6];
int8_t lo = input[4];
int8_t i4x2 = (hi << 4) | lo;
tensor(j + 2, i) = i4x2;
}
{
int8_t hi = input[3];
int8_t lo = input[1];
int8_t i4x2 = (hi << 4) | lo;
tensor(j + 4, i) = i4x2;
}
{
int8_t hi = input[7];
int8_t lo = input[5];
int8_t i4x2 = (hi << 4) | lo;
tensor(j + 6, i) = i4x2;
}
}
}
}
template <typename GemmConfig,
typename Invoker,
typename ADataType,
@@ -399,7 +346,7 @@ int run_gemm_example_with_layouts(ck_tile::ArgParser& arg_parser,
BLayout,
CLayout>(b_k_n_dev);
}
permute_vectors_i4x4_b(b_k_n_dev);
ck_tile::permute_vectors_i4x4_b(b_k_n_dev);
b_k_n_dev_buf.ToDevice(b_k_n_dev.data());
}
else

5
example/ck_tile/03_gemm/script/smoke_test_mem_pipeline.sh
View File

@@ -5,7 +5,7 @@ KNAME=1
export CK_WARMUP=0
export CK_REPEAT=1
COMMON_ARGS='-v=2 -warmup=0 -repeat=1'
COMMON_ARGS='-v=1 -warmup=0 -repeat=1'
run_tests() {
for m in 512 1024; do
@@ -32,5 +32,8 @@ run_tests "fp16"
run_tests "bf16"
run_tests "fp8"
run_tests "bf8"
run_tests "fp16i4"
run_tests "fp8i4"
run_tests "bf8i4"
set +x

35
example/ck_tile/03_gemm/universal_gemm.cpp
View File

@@ -5,11 +5,8 @@
#include <cstring>
#include <iostream>
#include <sstream>
#include <string>
#include <tuple>
#include "ck_tile/host.hpp"
#include "gemm_utils.hpp"
#include "run_gemm_example.inc"
#include "run_gemm_example_common.hpp"
@@ -58,7 +55,7 @@ int run_gemm_example(ck_tile::ArgParser& arg_parser)
ck_tile::int8_t,
ck_tile::int32_t>(a_layout, b_layout, arg_parser);
}
else if(data_type == "pk_int4_t")
else if(data_type == "fp16i4")
{
// TODO: Add support for bhalf_t ADataType
if constexpr(GemmConfig<ck_tile::half_t>::Pipeline == CK_TILE_PIPELINE_COMPUTE_V3)
@@ -74,6 +71,36 @@ int run_gemm_example(ck_tile::ArgParser& arg_parser)
throw std::runtime_error("Unsupported pipeline for this operation !!!");
}
}
else if(data_type == "fp8i4")
{
if constexpr(GemmConfig<ck_tile::fp8_t>::Pipeline == CK_TILE_PIPELINE_COMPUTE_V3)
{
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
Invoker,
ck_tile::fp8_t,
ck_tile::pk_int4_t,
ck_tile::half_t>(a_layout, b_layout, arg_parser);
}
else
{
throw std::runtime_error("Unsupported pipeline for this operation !!!");
}
}
else if(data_type == "bf8i4")
{
if constexpr(GemmConfig<ck_tile::bf8_t>::Pipeline == CK_TILE_PIPELINE_COMPUTE_V3)
{
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
Invoker,
ck_tile::bf8_t,
ck_tile::pk_int4_t,
ck_tile::half_t>(a_layout, b_layout, arg_parser);
}
else
{
throw std::runtime_error("Unsupported pipeline for this operation !!!");
}
}
else
{
throw std::runtime_error("Unsupported data type for this operation !!!");