Extend DualGemm: support batched mode + decouple B0/B1 layouts (#790)
* Fix MHA kernel
Summary:
ATT
Test Plan:
Reviewers:
Subscribers:
Tasks:
Tags:
* Extend DualGemm to support batched mode (#5)
Following the GemmUniversalMode::kBatched implementation, batched mode is added to the DualGemm (under examples/45_dual_gemm). DualGemmMode::kBatched and SplitKSerial are not compatible: Status::kErrorInvalidProblem is returned if both are set.
* Decouple LayoutB0 and LayoutB1 in DualGemm
The DualGemm template assumed the same layout, LayoutB, for both right operand matrices B0 and B1. This is problematic if the layout of the two matrices is different. In particular, this may be the case when one of the matrices is row-major, while the other is a (column) vector that has to be broadcasted in column-major with zero stride (e.g., as {B1.device_data(), 0}) for the DualGemm implementation to be able to process B0 and B1 simultaneously.
In this commit, LayoutB0 and LayoutB1 are decoupled throughout the DualGemm code (device, kernel, and mma). Additionally, the batch strides of B0 and B1 are also decoupled to accommodate the column vector B1 case described above.
* Remove comment as no longer relevant
* Revert Fix MHA kernel
---------
Co-authored-by: mikeiovine <mikeiovine@fb.com>
2023-02-14 04:27:13 +08:00
|
|
|
/***************************************************************************************************
|
2023-02-14 04:35:33 +08:00
|
|
|
* Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
Extend DualGemm: support batched mode + decouple B0/B1 layouts (#790)
* Fix MHA kernel
Summary:
ATT
Test Plan:
Reviewers:
Subscribers:
Tasks:
Tags:
* Extend DualGemm to support batched mode (#5)
Following the GemmUniversalMode::kBatched implementation, batched mode is added to the DualGemm (under examples/45_dual_gemm). DualGemmMode::kBatched and SplitKSerial are not compatible: Status::kErrorInvalidProblem is returned if both are set.
* Decouple LayoutB0 and LayoutB1 in DualGemm
The DualGemm template assumed the same layout, LayoutB, for both right operand matrices B0 and B1. This is problematic if the layout of the two matrices is different. In particular, this may be the case when one of the matrices is row-major, while the other is a (column) vector that has to be broadcasted in column-major with zero stride (e.g., as {B1.device_data(), 0}) for the DualGemm implementation to be able to process B0 and B1 simultaneously.
In this commit, LayoutB0 and LayoutB1 are decoupled throughout the DualGemm code (device, kernel, and mma). Additionally, the batch strides of B0 and B1 are also decoupled to accommodate the column vector B1 case described above.
* Remove comment as no longer relevant
* Revert Fix MHA kernel
---------
Co-authored-by: mikeiovine <mikeiovine@fb.com>
2023-02-14 04:27:13 +08:00
|
|
|
* SPDX-License-Identifier: BSD-3-Clause
|
|
|
|
*
|
|
|
|
* Redistribution and use in source and binary forms, with or without
|
|
|
|
* modification, are permitted provided that the following conditions are met:
|
|
|
|
*
|
|
|
|
* 1. Redistributions of source code must retain the above copyright notice, this
|
|
|
|
* list of conditions and the following disclaimer.
|
|
|
|
*
|
|
|
|
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
|
|
|
* this list of conditions and the following disclaimer in the documentation
|
|
|
|
* and/or other materials provided with the distribution.
|
|
|
|
*
|
|
|
|
* 3. Neither the name of the copyright holder nor the names of its
|
|
|
|
* contributors may be used to endorse or promote products derived from
|
|
|
|
* this software without specific prior written permission.
|
|
|
|
*
|
|
|
|
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
|
|
|
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
|
|
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
|
|
|
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
|
|
|
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
|
|
|
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
|
|
|
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
|
|
|
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
|
|
|
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
|
|
|
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
|
|
*
|
|
|
|
**************************************************************************************************/
|
|
|
|
/*! \file
|
|
|
|
\brief Defines common types used for all DualGemm operators.
|
|
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
|
|
|
|
namespace cutlass {
|
|
|
|
namespace gemm {
|
|
|
|
|
|
|
|
/////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
|
|
|
|
enum class DualGemmMode {
|
|
|
|
kGemm,
|
|
|
|
kBatched,
|
|
|
|
kInvalid
|
|
|
|
};
|
|
|
|
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
|
|
|
|
|
|
} // namespace gemm
|
|
|
|
} // namespace cutlass
|
|
|
|
|
|
|
|
////////////////////////////////////////////////////////////////////////////////////////////////////
|