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] smoothquant (#1617)

Browse Source 
* fix compile error

* fix typo of padding

* Add smoothquant op

* Add smoothquant instance library

* refine type

* add test script

* Re-generate smoothquant.hpp

* Always use 'current year' in copyright

* use Generic2dBlockShape instead

* Add vector = 8 instance back

* Find exe path automatically

* Simplify the api condition

* Remove debugging code

* update year

* Add blank line between function declaration

* explicitly cast return value to dim3

* refine return value

* Fix default warmup and repeat value

* Add comment

* refactor sommthquant cmake

* Add README

* Fix typo

---------

Co-authored-by: Po Yen, Chen <PoYen.Chen@amd.com>
This commit is contained in:
rocking
2024-11-01 13:51:56 +08:00
committed by GitHub
parent 550248deec
commit fbd654545a
62 changed files with 1758 additions and 219 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
include/ck_tile/ops/rmsnorm2d/kernel/rmsnorm2d_fwd_kernel.hpp
View File

@@ -11,11 +11,11 @@ namespace ck_tile {
// host side args
struct Rmsnorm2dFwdHostArgs
{
const void* p_x;
const void* p_gamma;
const void* p_x; // [m ,n], input, fp16/bf16
const void* p_gamma; // [1, n], gamma, prec same as input
void* p_y;
void* p_invRms;
void* p_y; // [m, n], output, fp16/bf16
void* p_invRms; // [m, 1], output inv-rms, prec same as input, nullptr if not used
float epsilon;
@@ -83,7 +83,7 @@ struct Rmsnorm2dFwd
CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
{
return (hargs.m + Block_M - 1) / Block_M;
return dim3(integer_divide_ceil(hargs.m, Block_M));
}
CK_TILE_HOST static constexpr auto BlockSize() { return Problem::BlockShape::BlockSize; }
@@ -149,7 +149,7 @@ struct Rmsnorm2dFwd
number<1>{});
const auto tmp2_ =
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<kPadM>{});
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<kPadN>{});
return make_tile_window(tmp2_, make_tuple(number<Block_N>{}), {0});
}();

78
include/ck_tile/ops/rmsnorm2d/kernel/rmsnorm2d_fwd_shape.hpp
View File

@@ -1,78 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
namespace ck_tile {
/*
// clang-format off
4-level descriptor: BlockTile-> WarpPerBlock-> WarpTile-> Vector
Block_N (Warp_N * WarpPerBlock_N * Repeat_N )
+<----------------------< Repeat_N(2)>--------------------->+
| |
+<-- <WarpPerBlock_N(2)> -->+
Warp_N
+--------------+--------------+--------------+--------------+----+----------------+
Warp_M | wrap_0 | wrap_1 | | ^ ^
+--------------+--------------+ | <WarpPerBlock_M(2)> |
| wrap_2 | wrap_3 | | v
+--------------+--------------+--------------+--------------+----+ Block_M
| | |
+ + |
| | | v
+--------------+--------------+--------------+--------------+ +
each Warp-tile (e.g 16 thrd per row)
Vector_N (contiguous pixels each thrd holds along N, or vector size)
+-----------+-----------+-----------+-----------+-----------+
| thrd_0 | thrd_1 | thrd_2 | thrd_3 | ... Vector_M
+-----------+-----------+-----------+-----------+-----------+
| thrd_16 | thrd_17 | thrd_18 | thrd_19 | ...
+-----------+-----------+-----------+-----------+-----------+
// clang-format on
*/
template <typename BlockTile_, // block size, seq<M, N>
typename WarpPerBlock_, // num warps along seq<M, N>
typename WarpTile_, // warp size, seq<M, N>
typename Vector_, // contiguous pixels(vector size) along seq<M, N>
index_t BlockSize_ =
warpSize* reduce_on_sequence(WarpPerBlock_{}, multiplies{}, number<1>{})>
struct Rmsnorm2dShape
{
// block size
static constexpr index_t Block_M = BlockTile_::at(number<0>{});
static constexpr index_t Block_N = BlockTile_::at(number<1>{});
// num warps along seq<M, N>, within each block
static constexpr index_t WarpPerBlock_M = WarpPerBlock_::at(number<0>{});
static constexpr index_t WarpPerBlock_N = WarpPerBlock_::at(number<1>{});
// warp size
static constexpr index_t Warp_M = WarpTile_::at(number<0>{});
static constexpr index_t Warp_N = WarpTile_::at(number<1>{});
static_assert(Block_M % (WarpPerBlock_M * Warp_M) == 0);
static_assert(Block_N % (WarpPerBlock_N * Warp_N) == 0);
// repeat of each thread along seq<M, N>
static constexpr index_t Repeat_M = Block_M / (WarpPerBlock_M * Warp_M);
static constexpr index_t Repeat_N = Block_N / (WarpPerBlock_N * Warp_N);
// vector size along seq<M, N>
static constexpr index_t Vector_M = Vector_::at(number<0>{});
static constexpr index_t Vector_N = Vector_::at(number<1>{});
static_assert(Warp_M % Vector_M == 0);
static_assert(Warp_N % Vector_N == 0);
// num of threads along seq<M, N>, within each warp
static constexpr index_t ThreadPerWarp_M = Warp_M / Vector_M;
static constexpr index_t ThreadPerWarp_N = Warp_N / Vector_N;
static constexpr index_t BlockSize = BlockSize_;
};
} // namespace ck_tile