ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-04 05:31:24 +00:00
Code Issues Packages Projects Releases Wiki Activity

gemm/Conv xdlops + dlops quantization (#625)

Browse Source 
* Add conv perlayer quantization

* Add gemm_dlops quantization

* Support int8 for innerproduct

* Refine gemm dlops int8 kernel parameter

* Support gfx908(MI100) and gfx90a(MI200)

* clang-format

* Rename example number

* Support different layout for d tensor

* Add conv dlops perchannel quantization example

* Move to example 40

* Extract the common code for different platform (dlops and xdlops)

* Move ot subfolder. Prepare to add other op of quantization

* Refine the quantization instance library

* Add conv dl instances and client example

* Remove unnecessary type

* Add gemm quantization instance

* Add external api and client example

* Refine num_bytes

* Separete different layout to different cpp

* Add more xdl instances

* Revert "Remove unnecessary type"

This reverts commit 820869182f.

* Remove CShuffleDataType in dlops
Let acc and CShuffleDataType be the same in xdlops

---------

Co-authored-by: zjing14 <zhangjing14@gmail.com>
This commit is contained in:
rocking5566
2023-03-16 04:29:40 +08:00
committed by GitHub
parent a2d5ca8e95
commit 16dc18e0f9
60 changed files with 3186 additions and 717 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View File

@@ -134,7 +134,8 @@ __global__ void
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \
defined(__gfx90a__) || defined(__gfx908__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -314,9 +315,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
const auto in_gemmm_gemmk_desc =
matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
const auto M = in_gemmm_gemmk_desc.GetLength(I0);
const auto K = in_gemmm_gemmk_desc.GetLength(I1);
const auto M = in_gemmm_gemmk_desc.GetLength(I0);
const auto K = in_gemmm_gemmk_desc.GetLength(I1);
const auto AK0 = K / K1;
return transform_tensor_descriptor(
@@ -709,7 +709,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
namespace ctc = tensor_layout::convolution;
// check device
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx908"))
{
return false;
}
@@ -834,6 +835,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
{
return false;
}
// check Gridwise GEMM
return GridwiseGemm::CheckValidity(
arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.e_grid_desc_m_n_);

4
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
View File

@@ -51,7 +51,7 @@ __global__ void
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx1030__))
defined(__gfx90a__) || defined(__gfx1030__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
@@ -552,7 +552,7 @@ struct DeviceGemmMultipleD_Dl : public DeviceGemmMultipleD<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx908" ||
ck::get_device_name() == "gfx1030")
ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx1030")
{
return GridwiseGemm::CheckValidity(
arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_);