ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK TILE] GEMM with packed i4 (#1885)

Browse Source 
* [CK TILE] GEMM with packed i4

* Fixes

* fixes

* fixes

* fixes
This commit is contained in:
Bartłomiej Kocot
2025-02-20 09:59:49 +01:00
committed by GitHub
parent 824e2c1737
commit 4d9973ec8e
32 changed files with 882 additions and 305 deletions
Download Patch File Download Diff File Expand all files Collapse all files

68
include/ck_tile/host/reference/reference_gemm.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -34,11 +34,35 @@ CK_TILE_HOST void reference_gemm(const HostTensor<ADataType>& a_m_k,
for(std::size_t k = 0; k < K; ++k)
{
ADataType v_a = a_element_op(a_m_k(m, k));
BDataType v_b = b_element_op(b_k_n(k, n));
v_acc +=
ck_tile::type_convert<AccDataType>(v_a) * ck_tile::type_convert<AccDataType>(v_b);
AccDataType v_a;
AccDataType v_b;
if constexpr(std::is_same_v<ADataType, pk_int4_t>)
{
const pk_int4_t pk_val = a_element_op(a_m_k(m, k));
const fp32x2_t fp32_val = pk_int4_t_to_fp32x2_t(pk_val);
if(k % 2 == 1)
v_a = fp32_val.hi;
else
v_a = fp32_val.lo;
}
else
{
v_a = ck_tile::type_convert<AccDataType>(a_element_op(a_m_k(m, k)));
}
if constexpr(std::is_same_v<BDataType, pk_int4_t>)
{
const pk_int4_t pk_val = b_element_op(b_k_n(k, n));
const fp32x2_t fp32_val = pk_int4_t_to_fp32x2_t(pk_val);
if(k % 2 == 1)
v_b = fp32_val.hi;
else
v_b = fp32_val.lo;
}
else
{
v_b = ck_tile::type_convert<AccDataType>(b_element_op(b_k_n(k, n)));
}
v_acc += v_a * v_b;
}
c_m_n(m, n) = ck_tile::type_convert<CDataType>(acc_element_op(v_acc));
@@ -73,6 +97,8 @@ __global__ void naive_gemm_kernel(ADataType* A,
AccDataType acc = 0.0;
for(int k = 0; k < K; ++k)
{
constexpr index_t packed_size_a = ck_tile::numeric_traits<ADataType>::PackedSize;
constexpr index_t packed_size_b = ck_tile::numeric_traits<BDataType>::PackedSize;
// Adjust indexing based on matrix layout
int a_index = (std::is_same_v<LayoutA, tensor_layout::gemm::RowMajor>)
? row * strideA + k
@@ -80,8 +106,34 @@ __global__ void naive_gemm_kernel(ADataType* A,
int b_index = (std::is_same_v<LayoutB, tensor_layout::gemm::ColumnMajor>)
? col * strideB + k
: k * strideB + col;
acc += ck_tile::type_convert<AccDataType>(A[a_index]) *
ck_tile::type_convert<AccDataType>(B[b_index]);
AccDataType v_a;
AccDataType v_b;
if constexpr(std::is_same_v<ADataType, pk_int4_t>)
{
const fp32x2_t fp32_val = pk_int4_t_to_fp32x2_t(A[a_index / packed_size_a]);
if(k % 2 == 1)
v_a = fp32_val.hi;
else
v_a = fp32_val.lo;
}
else
{
v_a = ck_tile::type_convert<AccDataType>(A[a_index]);
}
if constexpr(std::is_same_v<BDataType, pk_int4_t>)
{
const fp32x2_t fp32_val = pk_int4_t_to_fp32x2_t(B[b_index / packed_size_b]);
if(k % 2 == 1)
v_b = fp32_val.hi;
else
v_b = fp32_val.lo;
}
else
{
v_b = ck_tile::type_convert<AccDataType>(B[b_index]);
}
acc += v_a * v_b;
}
int c_index = (std::is_same_v<LayoutC, tensor_layout::gemm::RowMajor>)