mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-11 17:00:18 +00:00
ggemm tile_loop multD bf16 int8 (#1258)
* Overload output stream operator for LoopScheduler and PiplineVersion * Add Run overload accepting grid descriptors MK. * Add __device__ keyword for CalculateGridSize * Create device op GroupedGemmMultipleD * Add GroupedGemm MultipleD Tile Loop implementation. * Add an example for GroupedGemm MultipleD tile loop. * Device Op GroupedGEMMTileLoop. * Bunch of small changes in exmaple. * CkProfiler * Remove unused tparam. * changed the copy function to v7r2 * adding multi_abd * in-progress * add post-load oob check * Fix include statement. * Fix output stream overloads. * Do not make descriptors and check validity untill we find group. * Fix gemm desc initialization. * debugging * adjust instances * add run_lds * add elemntwise_op * replace multi_abd_device with v3 * clean up * clean * clean * Revert device op * Fix compilation for DTYPES=FP16 * Validate tensor transfers paramters. * Added LDSType * profiling * adjust oobcheck * add missing file * Validate on host only NK dims if M is not known. * add * clean * refactor * clean * add examples * add fuse * add fusion and client example * Fix bug. * A convenient debug func for selecting threads. * Fix has main k block loop bug. * Make sure that b2c has up to date tile offset. * Output stream operator for Sequence type. * Cmake file formatting. * clean --------- Co-authored-by: Adam Osewski <Adam.Osewski@amd.com>
This commit is contained in:
zjing14
@@ -27,14 +27,16 @@ using Empty_Tuple = ck::Tuple<>;
|
||||
|
||||
using BF16_Tuple = ck::Tuple<BF16>;
|
||||
|
||||
using F16_Tuple = ck::Tuple<F16>;
|
||||
using F16_F16_Tuple = ck::Tuple<F16, F16>;
|
||||
using F16_Tuple = ck::Tuple<F16>;
|
||||
using F16_F16_Tuple = ck::Tuple<F16, F16>;
|
||||
using BF16_BF16_Tuple = ck::Tuple<BF16, BF16>;
|
||||
|
||||
using F64_Tuple = ck::Tuple<F64>;
|
||||
using F32_Tuple = ck::Tuple<F32>;
|
||||
using I32_Tuple = ck::Tuple<I32>;
|
||||
using I32_F32_Tuple = ck::Tuple<I32, F32>;
|
||||
using I8_Tuple = ck::Tuple<I8>;
|
||||
using BF16_Tuple = ck::Tuple<BF16>;
|
||||
|
||||
using F32_F32_Tuple = ck::Tuple<F32, F32>;
|
||||
|
||||
|
||||
@@ -0,0 +1,287 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <vector>
|
||||
#include <memory>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_tile_loop.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
|
||||
Row,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
BF16_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
Multiply>>>& instances);
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename D0Layout,
|
||||
typename ELayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename D0DataType,
|
||||
typename EDataType>
|
||||
struct DeviceOperationInstanceFactory<
|
||||
ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<D0Layout>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<D0DataType>,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
Multiply>>
|
||||
{
|
||||
using DeviceOp = DeviceGroupedGemmTileLoop<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<D0Layout>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<D0DataType>,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
Multiply>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
// fp16_output
|
||||
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
|
||||
is_same_v<EDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
void add_device_grouped_gemm_xdl_tile_loop_multiply_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
|
||||
Row,
|
||||
Row_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
BF16_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyFastGelu>>>& instances);
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename D0Layout,
|
||||
typename ELayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename D0DataType,
|
||||
typename EDataType>
|
||||
struct DeviceOperationInstanceFactory<
|
||||
ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<D0Layout>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<D0DataType>,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyFastGelu>>
|
||||
{
|
||||
using DeviceOp = DeviceGroupedGemmTileLoop<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<D0Layout>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<D0DataType>,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyFastGelu>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
// fp16_output
|
||||
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
|
||||
is_same_v<EDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_xdl_tile_loop_multiply_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
void add_device_grouped_gemm_xdl_tile_loop_multiply_bias_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
|
||||
Row,
|
||||
Row_Row_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
BF16_BF16_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyAdd>>>& instances);
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename D0Layout,
|
||||
typename D1Layout,
|
||||
typename ELayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename D0DataType,
|
||||
typename D1DataType,
|
||||
typename EDataType>
|
||||
struct DeviceOperationInstanceFactory<
|
||||
ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<D0Layout, D1Layout>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<D0DataType, D1DataType>,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyAdd>>
|
||||
{
|
||||
using DeviceOp = DeviceGroupedGemmTileLoop<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<D0Layout, D1Layout>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<D0DataType, D1DataType>,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyAdd>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
// fp16_output
|
||||
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
|
||||
is_same_v<EDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_xdl_tile_loop_multiply_bias_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
void add_device_grouped_gemm_xdl_tile_loop_multiply_bias_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
|
||||
Row,
|
||||
Row_Row_Tuple,
|
||||
Row,
|
||||
BF16,
|
||||
I8,
|
||||
BF16_BF16_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyAddFastGelu>>>& instances);
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename D0Layout,
|
||||
typename D1Layout,
|
||||
typename ELayout,
|
||||
typename ADataType,
|
||||
typename BDataType,
|
||||
typename D0DataType,
|
||||
typename D1DataType,
|
||||
typename EDataType>
|
||||
struct DeviceOperationInstanceFactory<
|
||||
ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<D0Layout, D1Layout>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<D0DataType, D1DataType>,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyAddFastGelu>>
|
||||
{
|
||||
using DeviceOp = DeviceGroupedGemmTileLoop<ALayout,
|
||||
BLayout,
|
||||
ck::Tuple<D0Layout, D1Layout>,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
ck::Tuple<D0DataType, D1DataType>,
|
||||
EDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyAddFastGelu>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
// fp16_output
|
||||
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
|
||||
is_same_v<EDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
|
||||
is_same_v<ELayout, Row>)
|
||||
{
|
||||
add_device_grouped_gemm_xdl_tile_loop_multiply_bias_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user