ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-02 04:31:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

introducing ck_tile! (#1216)

Browse Source 
* enable gfx940

* switch between intrinsic mfma routines on mi100/200 and mi300

* fix mfma_int8 on MI300

* disable 2 int8 examples on MI300

* Update cmake-ck-dev.sh

* restore gitignore file

* modify Jenkinsfile to the internal repo

* Bump rocm-docs-core from 0.24.0 to 0.29.0 in /docs/sphinx

Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.24.0 to 0.29.0.
- [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases)
- [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md)
- [Commits](https://github.com/RadeonOpenCompute/rocm-docs-core/compare/v0.24.0...v0.29.0)

---
updated-dependencies:
- dependency-name: rocm-docs-core
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>

* initial enablement of gfx950

* fix clang format

* disable examples 31 and 41 int8 on gfx950

* add code

* fix build wip

* fix xx

* now can build

* naming

* minor fix

* wip fix

* fix macro for exp2; fix warpgemm a/b in transposedC

* unify as tuple_array

* Update the required Python version to 3.9

* Update executable name in test scripts

* re-structure tuple/array to avoid spill

* Merge function templates

* Fix format

* Add constraint to array<> ctor

* Re-use function

* Some minor changes

* remove wrong code in store_raw()

* fix compile issue in transpose

* Rename enum
Rename 'cood_transform_enum' to 'coord_transform_enum'

* let more integral_constant->constant, and formating

* make sure thread_buffer can be tuple/array

* temp fix buffer_store spill

* not using custom data type by default, now we can have ISA-level same code as opt_padding

* fix compile error, fp8 not ready now

* fix fp8 duplicated move/shift/and/or problem

* Default use CK_TILE_FLOAT_TO_FP8_STOCHASTIC rounding mode

* fix scratch in fp8 kernel

* update some readme

* fix merge from upstream

* sync with upstream

* sync upstream again

* sync 22

* remove unused

* fix clang-format

* update README of ck_tile example

* fix several issue

* let python version to be 3.8 as minimal

* remove ck_tile example from default cmake target like all/install/check

* remove mistake

* 1).support receipe in generate.py 2).use simplified mask type 3).change left/right to pass into karg

* fix some bug in group-mode masking and codegen. update README

* F8 quantization for FMHA forward (#1224)

* Add SAccElementFunction, PComputeElementFunction, OAccElementFunction in pipeline

* Add element function to fmha api

* Adjust P elementwise function

* Fix bug of elementwise op, our elementwise op is not inout

* Add some elementwise op, prepare to quantization

* Let generate.py can generate different elementwise function

* To prevent compiler issue, remove the elementwise function we have not used.

* Remove f8 pipeline, we should share the same pipeline even in f8

* Remove remove_cvref_t

* Avoid warning

* Fix wrong fp8 QK/KV block gemm setting

* Check fp8 rounding error in check_err()

* Set fp8 rounding error for check_err()

* Use CK_TILE_FLOAT_TO_FP8_STANDARD as default fp8 rounding mode

* 1. codgen the f8 api and kernel
2. f8 host code

* prevent warning in filter mode

* Remove not-in-use elementwise function kargs

* Remove more not-in-use elementwise function kargs

* Small refinements in C++ source files

* Use conditional_t<> to simplify code

* Support heterogeneous argument for binary function types

* Re-use already-existing scales<> functor template

* Fix wrong value produced by saturating

* Generalize the composes<> template

* Unify saturates<> implementation

* Fix type errors in composes<>

* Extend less_equal<>

* Reuse the existing template less_equal<> in check_err()

* Add equal<float> & equal<double>

* Rename check_err() parameter

* Rename check_err() parameter

* Add FIXME comment for adding new macro in future

* Remove unnecessary cast to void

* Eliminate duplicated code

* Avoid dividing api pool into more than 2 groups

* Use more clear variable names

* Use affirmative condition in if stmt

* Remove blank lines

* Donot perfect forwarding in composes<>

* To fix compile error, revert generate.py back to 4439cc107d

* Fix bug of p element function

* Add compute element op to host softmax

* Remove element function in api interface

* Extract user parameter

* Rename pscale and oscale variable

* rename f8 to fp8

* rename more f8 to fp8

* Add pipeline::operator() without element_functor

* 1. Remove deprecated pipeline enum
2. Refine host code parameter

* Use quantization range as input

* 1. Rename max_dtype to dtype_max.
2. Rename scale to scale_s
3.Add init description

* Refine description

* prevent early return

* unify _squant kernel name in cpp, update README

* Adjust the default range.

* Refine error message and bias range

* Add fp8 benchmark and smoke test

* fix fp8 swizzle_factor=4 case

---------

Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Co-authored-by: carlushuang <carlus.huang@amd.com>

---------

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: illsilin <Illia.Silin@amd.com>
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: Jing Zhang <jizha@amd.com>
Co-authored-by: zjing14 <zhangjing14@gmail.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Po-Yen, Chen <PoYen.Chen@amd.com>
Co-authored-by: rocking <ChunYu.Lai@amd.com>
This commit is contained in:
carlushuang
2024-04-16 08:27:12 +08:00
committed by GitHub
parent dd34ab6e64
commit db376dd8a4
141 changed files with 30623 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
example/ck_tile/01_fmha/CMakeLists.txt Normal file
View File

@@ -0,0 +1,44 @@
# generate a list of kernels, but not actually emit files at config stage
execute_process(
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--list_blobs ${CMAKE_CURRENT_BINARY_DIR}/blob_list.txt
)
# NOTE: for cmake, the FMHA_FWD_GEN_BLOBS files must be in the same directory
# as current cmake list, otherwise will not figure out the dependency properly
file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/blob_list.txt FMHA_FWD_GEN_BLOBS)
add_custom_command(
OUTPUT ${FMHA_FWD_GEN_BLOBS}
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--output_dir ${CMAKE_CURRENT_BINARY_DIR}
)
set(EXAMPLE_FMHA_FWD "tile_example_fmha_fwd")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message("adding tile_example ${EXAMPLE_NAME}")
add_executable(${EXAMPLE_FMHA_FWD} EXCLUDE_FROM_ALL fmha_fwd.cpp)
target_include_directories(${EXAMPLE_FMHA_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${EXAMPLE_FMHA_FWD} PRIVATE ${FMHA_FWD_GEN_BLOBS})
# NOTE: this is dangerous since will change the whole kernel to flush denormals
# WIP with compiler team for an exp2 intrinsic..., then remove this
if(NOT DEFINED FMHA_FWD_FAST_EXP2)
set(FMHA_FWD_FAST_EXP2 true)
endif()
set(EXAMPLE_FMHA_FWD_COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
# ... because they are auto-generated
if(FMHA_FWD_FAST_EXP2)
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero)
else()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=0)
endif()
# Allow comparing floating points directly in order to check sentinel values
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-float-equal)
target_compile_options(${EXAMPLE_FMHA_FWD} PRIVATE ${EXAMPLE_FMHA_FWD_COMPILE_OPTIONS})

117
example/ck_tile/01_fmha/README.md Normal file
View File

@@ -0,0 +1,117 @@
# fused multi-head attention
This folder contains example for fmha(fused multi-head attention) using ck_tile tile-programming implementation. It is a good example to demonstrate the usage of tile-programming API, as well as illustrate the new approach to construct a kernel template and instantiate it(them) while keeping compile time fast.
## build
```
# in the root of ck_tile
mkdir build && cd build
sh ../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_example_fmha_fwd -j
```
This will result in an executable `build/bin/tile_example_fmha_fwd`
## kernel
The kernel template is `fmha_fwd_kernel.hpp`, this is the grid-wise op in old ck_tile's terminology. We put it here purposely, to demonstrate one can construct a kernel by using various internal component from ck_tile. We may still have an implementation under ck_tile's include path (in the future) for the kernel template.
There are 3 template parameters for this kernel template.
* `TilePartitioner` is used to map the workgroup to corresponding tile, `fmha_fwd_tile_partitioner.hpp` in this folder served as this purpose.
* `FmhaPipeline` is one of the block_tile_pipeline(under `include/ck_tile/tile_program/block_tile_pipeline`) which is a performance critical component. Indeed, we did a lot of optimization and trials to optimize the pipeline and may still workout more performance pipeline and update into that folder. People only need to replace this pipeline type and would be able to enjoy the benefit of different performant implementations (stay tuned for updated pipeline(s)).
* `EpiloguePipeline` will modify and store out the result in the last phase. People usually will do lot of post-fusion at this stage, so we also abstract this concept. Currently we didn't do much thing at the epilogue stage but leave the room for future possible support.
## codegen
To speed up compile time, we instantiate the kernels into separate file. In this way we can benefit from parallel building from CMake/Make system. This is achieved by `generate.py` script. Besides, you can look into this script to learn how to instantiate a kernel instance step by step, which is described in `FMHA_FWD_KERNEL_BODY` variable.
## executable
`tile_example_fmha_fwd` is the example executable, implemented in `fmha_fwd.cpp`. You can type `./bin/tile_example_fmha_fwd -?` to list all supported args. Below is an example of the output (may subject to change)
```
args:
-v weather do CPU validation or not (default:1)
-mode kernel mode. 0:batch, 1:group (default:0)
-b batch size (default:2)
-h num of head, for q (default:8)
-h_k num of head, for k/v, 0 means equal to h (default:0)
if not equal to h, then this is GQA/MQA case
-s seqlen_q. if group-mode, means the average value of seqlen_q (default:3328)
total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary
-s_k seqlen_k, 0 means equal to s (default:0)
-d head dim for q, k (default:128)
-d_v head dim for v, 0 means equal to d (default:0)
-scale_s scale factor of S. 0 means equal to 1/sqrt(hdim). (default:0)
note when squant=1, this value will be modified by range_q/k
-range_q per-tensor quantization range of q. used if squant=1. (default:2)
-range_k per-tensor quantization range of k. used if squant=1. (default:2)
-range_v per-tensor quantization range of v. used if squant=1. (default:2)
-range_p per-tensor quantization range of p [e^(s-m)]. used if squant=1. (default:1)
-range_o per-tensor quantization range of o (p*v). used if squant=1. (default:2)
-squant if using static quantization fusion or not. 0: original flow(not prefered) (default:0)
1: apply scale_p and scale_o with respect to P and O. calculate scale_s, scale_p,
scale_o according to range_q, range_k, range_v, range_p, range_o
-iperm permute input (default:1)
if true, will be b*h*s*d, else b*s*h*d
-operm permute output (default:1)
-bias add bias or not (default:0)
-prec data type. fp16/bf16/fp8/bf8 (default:fp16)
-mask 0: no mask, 1: top-left(same as 't'), 2:bottom-right(same as 'b') (default:0)
't', top-left causal mask, 'b', bottom-r causal mask
't:l,r', top-left sliding window attn(swa) with FA style left right size
'b:l,r', bottom-r sliding window attn(swa) with FA style left right size
'xt:window_size', xformer style masking from top-left, window_size negative is causal, positive is swa
'xb:window_size', xformer style masking from bottom-r, window_size negative is causal, positive is swa
'g:y,x', generic attention mask coordinate with y/x size (only debug purpose for now)
-vlayout r for row-major(seqlen*hdim), c for col-major(hdim*seqlen) (default:r)
-lse 0 not store lse, 1 store lse (default:0)
-kname if set to 1 will print kernel name (default:0)
-init init method. 0:random int, 1:random float, 2:trig float, 3:quantization (default:1)
```
Example: `./bin/tile_example_fmha_fwd -b=1 -h=16 -s=16384 -d=128` will run a fmha case with batch=1, nhead=16, sequence length=16384, hdim=128, fp16 case.
## support features
Currently we are still in rapid development stage, so more features/optimizations will be coming soon.
### hdim
Currently we support `32/64/128/256` hdim for `fp16`/`bf16`, within which `64`/`128` is better optimized. hdim should be multiple of 8, while seqlen_s can be arbitrary. For hdim be arbitrary number, it can be support through padding kernel of `qr` pipeline (we didn't generate this in generate.py by default)
### group/batch mode
Currently we support both `batch mode` and `group mode` (or `varlen`, in FA's term), by setting `-mode` = `0` or `1`. In `group mode` different kind of attention mask is also supported(see below)
### MQA/GQA
By setting `-h`(nhead for q) and `-h_k`(nhead for k/v) with different number, you can achieve MQA/GQA. Please pay attention that `h % h_K == 0` when you set different numbers.
### input/output permute, and `b*s*3*h*d`
If you look at the kernel argument inside `fmha_fwd_kernel.hpp`, we support providing arbitrary stride for seqlen(stride_q/k/v), nhead, batch of q/k/v matrix, hence it is very flexible to support `b*h*s*d` or `b*s*h*d` input/output permute. The `-iperm=0/1`, `-operm=0/1` is a convenient way to achieve this through the executable. We didn't provide a command-line arg to test `b*s*3*h*d` layout which is by default used by torch/FA, but it's trivial to achieve this if one set the proper `stride_q/k/v` value as `3*h*d`.
### attention bias
Attention bias is supported with the layout of `1*1*s*s`(similiar to input/output, different layout can be supported by changing the stride value for bias, or even extend to `b*h*s*s`) and bias value in float number.
### lse
For training kernels, "log sum exp" need to store out in forward and used in backward. We support this by setting `-lse=1`
### vlayout
We support v matrix in both row-major(`seqlen*hdim`) and col-major(`hdim*seqlen`). Since the accumulate(reduce) dimension for V is along `seqlen`, for current AMD's mfma layout which expect each thread to have contiguous register holding pixels along reduce dimension, it's easier to support col-major V layout. However, the performance of col-major is not necessarily faster than row-major, there are many factors that may affect the overall performance. We still provide the `-vlayout=r/c` here to switch/test between different layouts.
### attention mask
we support `causal mask` and `sliding window attention(swa)` mask in both batch and group mode, either from top-left or bottom-right.
Underneath, we unify the mask expression into `generic attention mask coordinate`, providing an uniformed approach for each batch to locate the corresponding pixel need to be masked out.
![](misc/gamc.png)
Since FA/xformer style with window_size_left/right is more popular, we accept window_size as parameter and convert that internally to our generic coordinate(this coordinate can express more cases). Below shows some example of how to achieve different kind of mask through cmdline.
| mask case| cmdline | FA style | xformer style |
|----------|:-------------:|:-------------:|:-------------:|
| no mask | `-mask=0`(default) | | |
| causal mask from top-left | `-mask=1` or `-mask=t` | `-mask=t:-1,0` | `-mask=xt:-1` |
| causal mask from bottom-right | `-mask=2` or `-mask=b` | `-mask=b:-1,0` | `-mask=xb:-1` |
| swa from top-left | | `-mask=t:3,5` | `-mask=xt:4` |
| swa from bottom-right | | `-mask=b:10,11` | `-mask=xb:16` |
Note FA use bottom-right by default to express swa case, here we require you explicitly specify top-left/bottom-right.
### dropout
TBD
## FP8 experimental support
As described in [this blog](https://blog.hippoml.com/8bit-hippoattention-up-to-3x-faster-compared-to-flashattentionv2-8f9def90b482), we have an experimental support for fp8 fmha kernels, you can evaluate the performance by setting the arg `-prec=fp8` to the `tile_example_fmha_fwd`, on a gfx940/941/942 machine and ROCm 6.0+.
Currently we only support `-vlayout=c`( `hdim*seqlen` for V matrix) and `-squant=1`(static quantization) with `hdim=128` for fp8 now. Full feature support will come later.

716
example/ck_tile/01_fmha/fmha_fwd.cpp Normal file
View File

@@ -0,0 +1,716 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "fmha_fwd.hpp"
#include "ck_tile/host.hpp"
#include "mask.hpp"
#include "utils.hpp"
#include <array>
#include <cstring>
#include <functional>
#include <numeric>
#include <ostream>
#include <string>
#include <tuple>
#include <utility>
#include <vector>
template <typename T>
std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
{
using size_type = typename std::vector<T>::size_type;
os << "[";
for(size_type idx = 0; idx < v.size(); ++idx)
{
if(0 < idx)
{
os << ", ";
}
os << v[idx];
}
return os << "]";
}
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("v", "1", "weather do CPU validation or not")
.insert("mode", "0", "kernel mode. 0:batch, 1:group")
.insert("b", "2", "batch size")
.insert("h", "8", "num of head, for q")
.insert("h_k",
"0",
"num of head, for k/v, 0 means equal to h\n"
"if not equal to h, then this is GQA/MQA case")
.insert("s",
"3328",
"seqlen_q. if group-mode, means the average value of seqlen_q\n"
"total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary")
.insert("s_k", "0", "seqlen_k, 0 means equal to s")
.insert("d", "128", "head dim for q, k")
.insert("d_v", "0", "head dim for v, 0 means equal to d")
.insert("scale_s",
"0",
"scale factor of S. 0 means equal to 1/sqrt(hdim).\n"
"note when squant=1, this value will be modified by range_q/k")
.insert("range_q", "16", "per-tensor quantization range of q. used if squant=1.")
.insert("range_k", "16", "per-tensor quantization range of k. used if squant=1.")
.insert("range_v", "16", "per-tensor quantization range of v. used if squant=1.")
.insert("range_p", "1", "per-tensor quantization range of p [e^(s-m)]. used if squant=1.")
.insert("range_o", "16", "per-tensor quantization range of o (p*v). used if squant=1.")
.insert(
"squant",
"0",
"if using static quantization fusion or not. 0: original flow(not prefered)\n"
"1: apply scale_p and scale_o with respect to P and O. calculate scale_s, scale_p,\n"
"scale_o according to range_q, range_k, range_v, range_p, range_o")
.insert("iperm",
"1",
"permute input\n"
"if true, will be b*h*s*d, else b*s*h*d")
.insert("operm", "1", "permute output")
.insert("bias", "0", "add bias or not")
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
.insert("mask",
"0",
"0: no mask, 1: top-left(same as 't'), 2:bottom-right(same as 'b')\n"
"'t', top-left causal mask, 'b', bottom-r causal mask\n"
"'t:l,r', top-left sliding window attn(swa) with FA style left right size\n"
"'b:l,r', bottom-r sliding window attn(swa) with FA style left right size\n"
"'xt:window_size', xformer style masking from top-left, window_size negative is "
"causal, positive is swa\n"
"'xb:window_size', xformer style masking from bottom-r, window_size negative is "
"causal, positive is swa\n"
"'g:y,x', generic attention mask coordinate with y/x size (only debug purpose for "
"now)")
.insert("vlayout", "r", "r for row-major(seqlen*hdim), c for col-major(hdim*seqlen)")
.insert("lse", "0", "0 not store lse, 1 store lse")
.insert("kname", "0", "if set to 1 will print kernel name")
.insert(
"init", "1", "init method. 0:random int, 1:random float, 2:trig float, 3:quantization")
.insert("seed",
"11939",
"random seed used for initializing input tensors. 0 for "
"non-deterministic seed")
.insert("warmup", "5", "number of iterations before benchmark the kernel")
.insert("repeat", "20", "number of iterations to benchmark the kernel");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
// different threshold for different dtype
template <typename DataType>
auto get_elimit(int /*init_method*/)
{
double rtol = 1e-3;
double atol = 1e-3;
return ck_tile::make_tuple(rtol, atol);
}
template <>
auto get_elimit<ck_tile::bf16_t>(int init_method)
{
if(init_method == 0)
{
double rtol = 1e-2;
double atol = 1e-2;
return ck_tile::make_tuple(rtol, atol);
}
else
{
double rtol = 3e-3;
double atol = 3e-3;
return ck_tile::make_tuple(rtol, atol);
}
}
template <>
auto get_elimit<ck_tile::fp8_t>(int init_method)
{
if(init_method == 0)
{
unsigned max_rounding_point_distance = 0;
double atol = 2e-3;
return ck_tile::make_tuple(max_rounding_point_distance, atol);
}
else
{
unsigned max_rounding_point_distance = 1;
double atol = 0.0625;
return ck_tile::make_tuple(max_rounding_point_distance, atol);
}
}
template <typename DataType>
bool run(const ck_tile::ArgParser& arg_parser)
{
std::string data_type = arg_parser.get_str("prec");
int do_validation = arg_parser.get_int("v");
auto mode = static_cast<mode_enum>(arg_parser.get_uint32("mode"));
ck_tile::index_t batch = arg_parser.get_int("b");
ck_tile::index_t nhead = arg_parser.get_int("h");
ck_tile::index_t nhead_k = arg_parser.get_int("h_k");
if(nhead_k == 0)
nhead_k = nhead;
if(nhead % nhead_k != 0)
{
std::cerr << "nhead:" << nhead << " must be multiple of nhead_k:" << nhead_k << std::endl;
return false;
}
ck_tile::index_t seqlen_q = arg_parser.get_int("s");
ck_tile::index_t seqlen_k = arg_parser.get_int("s_k");
if(seqlen_k == 0)
seqlen_k = seqlen_q;
ck_tile::index_t hdim_q = arg_parser.get_int("d");
ck_tile::index_t hdim_v = arg_parser.get_int("d_v");
if(hdim_v == 0)
hdim_v = hdim_q;
bool i_perm = arg_parser.get_bool("iperm"); // if true, will be batch * nhead * seqlen * hdim
bool o_perm = arg_parser.get_bool("operm"); // if false, will be batch * seqlen * nhead * hdim
float scale_s = arg_parser.get_float("scale_s");
if(scale_s == .0f)
scale_s = 1.0 / ck_tile::sqrt(static_cast<float>(hdim_q)); // TODO: q ? v ?
bool squant = arg_parser.get_bool("squant");
if constexpr(!std::is_same_v<DataType, ck_tile::fp8_t>)
{
if(squant)
{
std::cerr << "static quantization only support fp8 for now" << std::endl;
return false;
}
}
float range_q = arg_parser.get_float("range_q");
float range_k = arg_parser.get_float("range_k");
float range_v = arg_parser.get_float("range_v");
float range_p = arg_parser.get_float("range_p");
float range_o = arg_parser.get_float("range_o");
float dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<DataType>::max());
float scale_p = 1.f;
float scale_o = 1.f;
if(squant)
{
scale_s = scale_s * (range_q / dtype_max) * (range_k / dtype_max);
scale_p = dtype_max / range_p;
// scale_p = [max(fp8_t)/range_o] * [range_p/max(fp8_t)] * [range_v/max(fp8_t)]
scale_o = range_p * range_v / range_o / dtype_max;
}
std::string vlayout = arg_parser.get_str("vlayout");
bool use_bias = arg_parser.get_bool("bias");
bool lse = arg_parser.get_bool("lse");
mask_info mask = mask_info::decode(arg_parser.get_str("mask"), seqlen_q, seqlen_k);
int init_method = arg_parser.get_int("init");
std::optional<uint32_t> seed = arg_parser.get_uint32("seed");
if(*seed == 0)
{
seed.reset();
}
int stream_warmup = arg_parser.get_int("warmup");
int stream_repeat = arg_parser.get_int("repeat");
bool kname = arg_parser.get_bool("kname");
ck_tile::stream_config stream_config{
nullptr, true, /* log_level = */ (kname ? 1 : 0), stream_warmup, stream_repeat};
const auto seqstart_q_host = generate_seqstarts(mode, batch, seqlen_q);
const auto seqstart_k_host = generate_seqstarts(mode, batch, seqlen_k);
using TypeConfig = FmhaFwdTypeConfig<DataType>;
using QDataType = typename TypeConfig::QDataType;
using KDataType = typename TypeConfig::KDataType;
using VDataType = typename TypeConfig::VDataType;
using BiasDataType = typename TypeConfig::BiasDataType;
using LSEDataType = typename TypeConfig::LSEDataType;
using SaccDataType = typename TypeConfig::SaccDataType;
using SMPLComputeDataType = typename TypeConfig::SMPLComputeDataType;
using PDataType = typename TypeConfig::PDataType;
using OaccDataType = typename TypeConfig::OaccDataType;
using ODataType = typename TypeConfig::ODataType;
// accumulation numbers for performance evaluation
std::size_t flop = 0, num_byte = 0;
auto max_seqlen_q =
std::numeric_limits<int32_t>::min(); // we will use max seqlen to decide grid size
{
for(ck_tile::index_t wb = 0; wb < batch; ++wb)
{
const int32_t real_seqlen_q = seqstart_q_host[wb + 1] - seqstart_q_host[wb];
const int32_t real_seqlen_k = seqstart_k_host[wb + 1] - seqstart_k_host[wb];
if(max_seqlen_q < real_seqlen_q)
{
max_seqlen_q = real_seqlen_q;
}
flop += nhead * (static_cast<std::size_t>(2) * real_seqlen_q * real_seqlen_k * hdim_q +
static_cast<std::size_t>(2) * real_seqlen_q * hdim_v * real_seqlen_k);
num_byte += nhead * (sizeof(QDataType) * real_seqlen_q * hdim_q +
sizeof(KDataType) * real_seqlen_k * hdim_q +
sizeof(VDataType) * hdim_v * real_seqlen_k +
sizeof(ODataType) * real_seqlen_q * hdim_v);
}
}
auto get_lengths = [&](bool permute,
ck_tile::index_t b /*batch*/,
ck_tile::index_t h /*nhead*/,
ck_tile::index_t s /*seqlen*/,
ck_tile::index_t d /*hdim*/) {
if(permute)
return std::array<ck_tile::index_t, 4>{b, h, s, d};
else
return std::array<ck_tile::index_t, 4>{b, s, h, d};
};
bool is_v_rowmajor = vlayout == std::string("r");
// host memory for storing all the tensor elements
const ck_tile::index_t shape_batch = (mode == mode_enum::batch ? batch : 1);
const ck_tile::index_t shape_seqlen_q =
(mode == mode_enum::batch ? seqlen_q : seqstart_q_host.back());
const ck_tile::index_t shape_seqlen_k =
(mode == mode_enum::batch ? seqlen_k : seqstart_k_host.back());
ck_tile::HostTensor<QDataType> q_host(
get_lengths(i_perm, shape_batch, nhead, shape_seqlen_q, hdim_q));
ck_tile::HostTensor<KDataType> k_host(
get_lengths(i_perm, shape_batch, nhead_k, shape_seqlen_k, hdim_q));
ck_tile::HostTensor<VDataType> v_host(
is_v_rowmajor ? get_lengths(i_perm, shape_batch, nhead_k, shape_seqlen_k, hdim_v)
: get_lengths(i_perm, shape_batch, nhead_k, hdim_v, shape_seqlen_k));
// use bias shape = [1, 1, shape_seqlen_q, shape_seqlen_k]. if use_bias=false, the bias_host
// will not be used for verification at all (but will be copied to device anyway).
ck_tile::HostTensor<BiasDataType> bias_host(
use_bias
? get_lengths(i_perm, 1, 1, shape_seqlen_q, shape_seqlen_k)
: std::array<ck_tile::index_t, 4>{1, 1, 1, 1} /* dummy shape for simplifying code */);
// self define lse data layout as [shape_batch, nhead, shape_seqlen_q]
ck_tile::HostTensor<LSEDataType> lse_host(
lse ? std::array<ck_tile::index_t, 3>{shape_batch, nhead, shape_seqlen_q}
: std::array<ck_tile::index_t, 3>{1, 1, 1} /* dummy shape for simplifying code */);
ck_tile::HostTensor<ODataType> o_host(
get_lengths(o_perm, shape_batch, nhead, shape_seqlen_q, hdim_v));
if(init_method == 0)
{
ck_tile::FillUniformDistributionIntegerValue<QDataType>{-2.f, 2.f, seed}(q_host);
ck_tile::FillUniformDistributionIntegerValue<KDataType>{-2.f, 2.f, seed}(k_host);
ck_tile::FillUniformDistributionIntegerValue<VDataType>{-2.f, 2.f, seed}(v_host);
ck_tile::FillUniformDistributionIntegerValue<BiasDataType>{-2.f, 2.f, seed}(bias_host);
}
else if(init_method == 1)
{
ck_tile::FillUniformDistribution<QDataType>{0.f, 1.f, seed}(q_host);
ck_tile::FillUniformDistribution<KDataType>{0.f, 1.f, seed}(k_host);
ck_tile::FillUniformDistribution<VDataType>{0.f, 1.f, seed}(v_host);
ck_tile::FillUniformDistribution<BiasDataType>{0.f, 1.f, seed}(bias_host);
}
else if(init_method == 2)
{
ck_tile::FillTrigValue<QDataType>{}(q_host);
ck_tile::FillTrigValue<KDataType>{}(k_host);
ck_tile::FillTrigValue<VDataType>{}(v_host);
ck_tile::FillTrigValue<BiasDataType>{}(bias_host);
}
else if(init_method == 3) // suitable for fp8 quantization
{
ck_tile::FillUniformDistribution<QDataType>{-dtype_max, dtype_max, seed}(q_host);
ck_tile::FillUniformDistribution<KDataType>{-dtype_max, dtype_max, seed}(k_host);
ck_tile::FillUniformDistribution<VDataType>{-dtype_max, dtype_max, seed}(v_host);
// bias_fp8 = qscale_bias * bias_fp32
float qscale_bias = (dtype_max / range_q) * (dtype_max / range_k);
// Assume bias is in [-1.f, 1.f] in original fp32
ck_tile::FillUniformDistribution<BiasDataType>{-qscale_bias, qscale_bias, seed}(bias_host);
}
ck_tile::DeviceMem q_buf(q_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem k_buf(k_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem v_buf(v_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem bias_buf(bias_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem lse_buf(lse_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem o_buf(o_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t));
ck_tile::DeviceMem seqstart_k(seqstart_k_host.size() * sizeof(int32_t));
q_buf.ToDevice(q_host.data());
k_buf.ToDevice(k_host.data());
v_buf.ToDevice(v_host.data());
bias_buf.ToDevice(bias_host.data());
seqstart_q.ToDevice(seqstart_q_host.data());
seqstart_k.ToDevice(seqstart_k_host.data());
// clang-format off
auto layout_str = [&](bool permute){
if (permute) return std::string("bhsd");
else return std::string("bshd");
};
auto io_layout = [&](bool iperm_, bool operm_) {
if (iperm_ == operm_) return layout_str(iperm_);
else return layout_str(iperm_) + std::string("-") + layout_str(operm_);
};
// clang-format on
const std::string prec = arg_parser.get_str("prec");
std::cout << "[" << prec << "|" << mode << "|" << io_layout(i_perm, o_perm) << "] b:" << batch
<< ", h:" << nhead << "/" << nhead_k << ", s:" << seqlen_q << "/" << seqlen_k
<< ", d:" << hdim_q << "/" << hdim_v << ", scale_s:" << scale_s
<< ", bias:" << use_bias << ", lse:" << lse << ", squant:" << squant
<< ", mask:" << mask << ", v:" << vlayout << std::flush;
auto fmha_traits = fmha_fwd_traits{hdim_q,
hdim_v,
data_type,
mode == mode_enum::group,
is_v_rowmajor,
mask.type,
use_bias,
lse,
squant};
auto p_compute_element_func = [&]() {
if constexpr(std::is_same_v<DataType, ck_tile::fp8_t>)
return ck_tile::scales{scale_p};
else
return ck_tile::identity{};
}();
auto oacc_element_func = [&]() {
if constexpr(std::is_same_v<DataType, ck_tile::fp8_t>)
return ck_tile::composes(ck_tile::saturates<ck_tile::fp8_t>{},
ck_tile::scales{scale_o});
else
return ck_tile::identity{};
}();
auto fmha_args = [&]() {
assert(nhead % nhead_k == 0);
/// NOTE: we broadcast bias from [1, 1, seqlen_q, seqlen_k] to [batch, nhead, seqlen_q,
/// seqlen_k] in this example, hence both the 'batch_stride_bias' &
/// 'nhead_stride_bias' are 0.
// setup stride_* arguments
const ck_tile::index_t stride_q = (i_perm ? hdim_q : nhead * hdim_q);
const ck_tile::index_t stride_k = (i_perm ? hdim_q : nhead_k * hdim_q);
const ck_tile::index_t stride_v = [&]() {
if(is_v_rowmajor)
return i_perm ? hdim_v : nhead_k * hdim_v;
else
return i_perm ? shape_seqlen_k : nhead_k * shape_seqlen_k;
}();
const ck_tile::index_t stride_bias = (i_perm ? shape_seqlen_k : 1 * shape_seqlen_k);
const ck_tile::index_t stride_o = (o_perm ? hdim_v : nhead * hdim_v);
// setup nhead_stride_* arguments
const ck_tile::index_t nhead_stride_q = (i_perm ? shape_seqlen_q * hdim_q : hdim_q);
const ck_tile::index_t nhead_stride_k = (i_perm ? shape_seqlen_k * hdim_q : hdim_q);
const ck_tile::index_t nhead_stride_v = [&]() {
if(is_v_rowmajor)
return i_perm ? shape_seqlen_k * hdim_v : hdim_v;
else
return i_perm ? hdim_v * shape_seqlen_k : shape_seqlen_k;
}();
const ck_tile::index_t nhead_stride_bias =
(i_perm ? 0 * shape_seqlen_q * shape_seqlen_k : 0 * shape_seqlen_k);
const ck_tile::index_t nhead_stride_lse = (shape_seqlen_q * 1);
const ck_tile::index_t nhead_stride_o = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
// setup batch_stride_* arguments
const ck_tile::index_t batch_stride_q = (nhead * shape_seqlen_q * hdim_q);
const ck_tile::index_t batch_stride_k = (nhead_k * shape_seqlen_k * hdim_q);
const ck_tile::index_t batch_stride_v = (nhead_k * hdim_v * shape_seqlen_k);
const ck_tile::index_t batch_stride_bias = (0 * nhead * shape_seqlen_q * shape_seqlen_k);
const ck_tile::index_t batch_stride_lse = (nhead * shape_seqlen_q * 1);
const ck_tile::index_t batch_stride_o = (nhead * shape_seqlen_q * hdim_v);
return fmha_fwd_args{q_buf.GetDeviceBuffer(),
k_buf.GetDeviceBuffer(),
v_buf.GetDeviceBuffer(),
bias_buf.GetDeviceBuffer(),
lse_buf.GetDeviceBuffer(),
o_buf.GetDeviceBuffer(),
seqstart_q.GetDeviceBuffer(),
seqstart_k.GetDeviceBuffer(),
nullptr,
shape_seqlen_q,
shape_seqlen_k,
batch,
max_seqlen_q,
hdim_q,
hdim_v,
nhead,
nhead_k,
scale_s,
scale_p,
scale_o,
stride_q,
stride_k,
stride_v,
stride_bias,
stride_o,
nhead_stride_q,
nhead_stride_k,
nhead_stride_v,
nhead_stride_bias,
nhead_stride_lse,
nhead_stride_o,
batch_stride_q,
batch_stride_k,
batch_stride_v,
batch_stride_bias,
batch_stride_lse,
batch_stride_o,
mask.left,
mask.right,
static_cast<ck_tile::index_t>(mask.type)};
}();
float ave_time = fmha_fwd(fmha_traits, fmha_args, stream_config);
if(ave_time < 0)
{
std::cout << ", not supported yet" << std::flush << std::endl;
return false;
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << std::fixed << ", " << std::setprecision(3) << ave_time << " ms, "
<< std::setprecision(2) << tflops << " TFlops, " << std::setprecision(2) << gb_per_sec
<< " GB/s" << std::flush;
if(!do_validation)
{
std::cout << std::flush << std::endl;
return true;
}
o_buf.FromDevice(o_host.data());
lse_buf.FromDevice(lse_host.data());
bool pass = true;
for(ck_tile::index_t wb = 0; wb < batch; ++wb)
{
const ck_tile::index_t real_seqlen_q = seqstart_q_host[wb + 1] - seqstart_q_host[wb];
const ck_tile::index_t real_seqlen_k = seqstart_k_host[wb + 1] - seqstart_k_host[wb];
// adjust matrix index according to the mode
const ck_tile::index_t b = (mode == mode_enum::batch ? wb : 0);
const ck_tile::index_t query_offset = (mode == mode_enum::batch ? 0 : seqstart_q_host[wb]);
const ck_tile::index_t key_offset = (mode == mode_enum::batch ? 0 : seqstart_k_host[wb]);
const auto v_host_ref_lengths =
std::array<ck_tile::index_t, 3>{nhead, hdim_v, real_seqlen_k};
const auto v_host_ref_strides =
is_v_rowmajor
? std::array<ck_tile::index_t, 3>{hdim_v * real_seqlen_k, 1, hdim_v}
: std::array<ck_tile::index_t, 3>{hdim_v * real_seqlen_k, real_seqlen_k, 1};
ck_tile::HostTensor<QDataType> q_host_ref({nhead, real_seqlen_q, hdim_q});
ck_tile::HostTensor<KDataType> k_host_ref({nhead, real_seqlen_k, hdim_q});
ck_tile::HostTensor<VDataType> v_host_ref(v_host_ref_lengths, v_host_ref_strides);
ck_tile::HostTensor<ODataType> o_host_ref({nhead, real_seqlen_q, hdim_v});
ck_tile::HostTensor<SMPLComputeDataType> s_host_ref({nhead, real_seqlen_q, real_seqlen_k});
ck_tile::HostTensor<PDataType> p_host_ref({nhead, real_seqlen_q, real_seqlen_k});
ck_tile::HostTensor<SMPLComputeDataType> lse_host_ref({nhead, real_seqlen_q});
ck_tile::index_t nr = nhead / nhead_k;
// clang-format off
// permute
if(i_perm) q_host_ref.ForEach([&](auto& self, auto i) { self(i) = q_host(b, i[0], i[1] + query_offset, i[2]); });
else q_host_ref.ForEach([&](auto& self, auto i) { self(i) = q_host(b, i[1] + query_offset, i[0], i[2]); });
if(i_perm) k_host_ref.ForEach([&](auto& self, auto i) { self(i) = k_host(b, i[0] / nr, i[1] + key_offset, i[2]); });
else k_host_ref.ForEach([&](auto& self, auto i) { self(i) = k_host(b, i[1] + key_offset, i[0] / nr, i[2]); });
if (is_v_rowmajor) {
// v_host_ref: [nhead, hdim, seq], v_host: [b, h_k, s, d]
if(i_perm) v_host_ref.ForEach([&](auto& self, auto i) { self(i) = v_host(b, i[0] / nr, i[2] + key_offset, i[1]); });
// v_host_ref: [nhead, hdim, seq], v_host: [b, s, h_k, d]
else v_host_ref.ForEach([&](auto& self, auto i) { self(i) = v_host(b, i[2] + key_offset, i[0] / nr, i[1]); });
}
else {
if(i_perm) v_host_ref.ForEach([&](auto& self, auto i) { self(i) = v_host(b, i[0] / nr, i[1], i[2] + key_offset); });
else v_host_ref.ForEach([&](auto& self, auto i) { self(i) = v_host(b, i[1], i[0] / nr, i[2] + key_offset); });
}
// clang-format on
// reference
ck_tile::reference_batched_gemm<QDataType, KDataType, SaccDataType, SMPLComputeDataType>(
q_host_ref,
k_host_ref,
s_host_ref,
ck_tile::identity{},
ck_tile::identity{},
ck_tile::scales(scale_s));
if(use_bias)
{
ck_tile::HostTensor<BiasDataType> bias_host_ref({1, real_seqlen_q, real_seqlen_k});
// clang-format off
if(i_perm)
bias_host_ref.ForEach([&](auto& self, auto i) { self(i) = bias_host(0, 0, i[1] + query_offset, i[2] + key_offset); });
else
bias_host_ref.ForEach([&](auto& self, auto i) { self(i) = bias_host(0, i[1] + query_offset, 0, i[2] + key_offset); });
// clang-format on
// broadcast from [1, real_seqlen_q, real_seqlen_k] to [nhead, real_seqlen_q,
// real_seqlen_k]
ck_tile::reference_batched_elementwise<SMPLComputeDataType,
BiasDataType,
SMPLComputeDataType,
SMPLComputeDataType>(
s_host_ref, bias_host_ref, s_host_ref);
}
if(mask.type == mask_enum::no_mask)
{
ck_tile::reference_batched_masking<SaccDataType>(
s_host_ref, FmhaMasks::NoMask{real_seqlen_q, real_seqlen_k});
}
else if(mask.type == mask_enum::window_generic)
{
ck_tile::reference_batched_masking<SaccDataType>(
s_host_ref,
ck_tile::make_generic_attention_mask_from_lr_window<FmhaMasks::GenericMask>(
mask.left, mask.right, real_seqlen_q, real_seqlen_k));
}
else
{
// if left window size is negative, means causal
// else means generic (for current batch)
if(mask.left < 0)
ck_tile::reference_batched_masking<SaccDataType>(
s_host_ref,
ck_tile::make_generic_attention_mask_from_lr_window<FmhaMasks::CausalMask>(
mask.left,
mask.right,
real_seqlen_q,
real_seqlen_k,
mask.type == mask_enum::mask_top_left));
else
ck_tile::reference_batched_masking<SaccDataType>(
s_host_ref,
ck_tile::make_generic_attention_mask_from_lr_window<FmhaMasks::GenericMask>(
mask.left,
mask.right,
real_seqlen_q,
real_seqlen_k,
mask.type == mask_enum::mask_top_left));
}
if(lse)
{
ck_tile::reference_batched_softmax<SMPLComputeDataType, SMPLComputeDataType, PDataType>(
s_host_ref, p_host_ref, p_compute_element_func, lse_host_ref);
}
else
{
ck_tile::reference_batched_softmax<SMPLComputeDataType, SMPLComputeDataType, PDataType>(
s_host_ref, p_host_ref, p_compute_element_func);
}
ck_tile::reference_batched_gemm<PDataType, VDataType, OaccDataType, ODataType>(
p_host_ref,
v_host_ref,
o_host_ref,
ck_tile::identity{},
ck_tile::identity{},
oacc_element_func);
ck_tile::HostTensor<ODataType> o_host_result({nhead, real_seqlen_q, hdim_v});
// clang-format off
// permute
if(o_perm) o_host_result.ForEach([&](auto& self, auto idx) { self(idx) = o_host(b, idx[0], idx[1] + query_offset, idx[2]); });
else o_host_result.ForEach([&](auto& self, auto idx) { self(idx) = o_host(b, idx[1] + query_offset, idx[0], idx[2]); });
// clang-format on
auto [rtol, atol] = get_elimit<DataType>(init_method);
bool cur_pass = ck_tile::check_err(
o_host_result, o_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol);
pass &= cur_pass;
if(!cur_pass)
{
std::cerr << "OUT mismatch found at batch: " << wb << std::endl
<< "\tseqlen_q: " << real_seqlen_q << std::endl
<< "\tseqlen_k: " << real_seqlen_k << std::endl
<< "\tseqstart_q: " << seqstart_q_host << std::endl
<< "\tseqstart_k: " << seqstart_k_host << std::endl;
break;
}
if(lse)
{
ck_tile::HostTensor<SMPLComputeDataType> lse_host_result({nhead, real_seqlen_q});
lse_host_result.ForEach([&](auto& self, auto idx) {
self(idx) = lse_host(b, idx[0], idx[1] + query_offset);
});
bool lse_pass = ck_tile::check_err(lse_host_result,
lse_host_ref,
"LSE Error: Incorrect results!",
rtol,
atol,
/* allow_infinity_ref = */ true);
pass &= lse_pass;
if(!cur_pass)
{
std::cerr << "LSE mismatch found at batch: " << wb << std::endl
<< "\tseqlen_q: " << real_seqlen_q << std::endl
<< "\tseqlen_k: " << real_seqlen_k << std::endl
<< "\tseqstart_q: " << seqstart_q_host << std::endl
<< "\tseqstart_k: " << seqstart_k_host << std::endl;
break;
}
}
}
std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
return pass;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
const std::string data_type = arg_parser.get_str("prec");
if(data_type == "fp16")
{
return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
}
else if(data_type == "bf16")
{
return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2;
}
else if(data_type == "fp8")
{
return run<ck_tile::fp8_t>(arg_parser) ? 0 : -2;
}
return -3;
}

269
example/ck_tile/01_fmha/fmha_fwd.hpp Normal file
View File

@@ -0,0 +1,269 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/fmha.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "mask.hpp"
#include <type_traits>
template <typename DataType>
struct FmhaFwdTypeConfig;
template <>
struct FmhaFwdTypeConfig<ck_tile::half_t>
{
using QDataType = ck_tile::half_t;
using KDataType = ck_tile::half_t;
using VDataType = ck_tile::half_t;
using BiasDataType = ck_tile::half_t;
using LSEDataType = float; // data type for lse(logsumexp L_j = max_j + log(l_j))
using SaccDataType = float; // data type for first gemm accumulation
using SMPLComputeDataType = float; // data type for reduction, softmax
using PDataType = ck_tile::half_t; // data type for A matrix of second gemm
using OaccDataType = float; // data type for second gemm accumulation
using ODataType = ck_tile::half_t;
};
template <>
struct FmhaFwdTypeConfig<ck_tile::bf16_t>
{
using QDataType = ck_tile::bf16_t;
using KDataType = ck_tile::bf16_t;
using VDataType = ck_tile::bf16_t;
using BiasDataType = ck_tile::bf16_t;
using LSEDataType = float; // data type for lse(logsumexp L_j = max_j + log(l_j))
using SaccDataType = float; // data type for first gemm accumulation
using SMPLComputeDataType = float; // data type for reduction, softmax
using PDataType = ck_tile::bf16_t; // data type for A matrix of second gemm
using OaccDataType = float; // data type for second gemm accumulation
using ODataType = ck_tile::bf16_t;
};
template <>
struct FmhaFwdTypeConfig<ck_tile::fp8_t>
{
using QDataType = ck_tile::fp8_t;
using KDataType = ck_tile::fp8_t;
using VDataType = ck_tile::fp8_t;
using BiasDataType = float;
using LSEDataType = float; // data type for lse(logsumexp L_j = max_j + log(l_j))
using SaccDataType = float; // data type for first gemm accumulation
using SMPLComputeDataType = float; // data type for reduction, softmax
using PDataType = ck_tile::fp8_t; // data type for A matrix of second gemm
using OaccDataType = float; // data type for second gemm accumulation
using ODataType = ck_tile::fp8_t;
};
template <>
struct FmhaFwdTypeConfig<ck_tile::bf8_t>
{
using QDataType = ck_tile::bf8_t;
using KDataType = ck_tile::bf8_t;
using VDataType = ck_tile::bf8_t;
using BiasDataType = ck_tile::bf8_t;
using LSEDataType = float; // data type for lse(logsumexp L_j = max_j + log(l_j))
using SaccDataType = float; // data type for first gemm accumulation
using SMPLComputeDataType = float; // data type for reduction, softmax
using PDataType = ck_tile::bf8_t; // data type for A matrix of second gemm
using OaccDataType = float; // data type for second gemm accumulation
using ODataType = ck_tile::bf8_t;
};
struct FmhaMasks
{
using NoMask = ck_tile::GenericAttentionMask<false>;
using GenericMask = ck_tile::GenericAttentionMask<true, true>;
using CausalMask = ck_tile::GenericAttentionMask<true, false>;
};
// runtime args, some will passed to karg, some will used to compute grids/blocks
struct fmha_fwd_args
{
const void* q_ptr;
const void* k_ptr;
const void* v_ptr;
const void* bias_ptr;
void* lse_ptr;
void* o_ptr;
const void* seqstart_q_ptr;
const void* seqstart_k_ptr;
const void* seqlen_k_ptr;
ck_tile::index_t seqlen_q;
ck_tile::index_t seqlen_k;
ck_tile::index_t batch;
ck_tile::index_t max_seqlen_q;
ck_tile::index_t hdim_q;
ck_tile::index_t hdim_v;
ck_tile::index_t nhead_q;
ck_tile::index_t nhead_k;
float scale_s;
float scale_p;
float scale_o;
ck_tile::index_t stride_q;
ck_tile::index_t stride_k;
ck_tile::index_t stride_v;
ck_tile::index_t stride_bias;
ck_tile::index_t stride_o;
ck_tile::index_t nhead_stride_q;
ck_tile::index_t nhead_stride_k;
ck_tile::index_t nhead_stride_v;
ck_tile::index_t nhead_stride_bias;
ck_tile::index_t nhead_stride_lse;
ck_tile::index_t nhead_stride_o;
ck_tile::index_t batch_stride_q;
ck_tile::index_t batch_stride_k;
ck_tile::index_t batch_stride_v;
ck_tile::index_t batch_stride_bias;
ck_tile::index_t batch_stride_lse;
ck_tile::index_t batch_stride_o;
ck_tile::index_t window_size_left;
ck_tile::index_t window_size_right;
ck_tile::index_t mask_type;
};
template <typename FmhaKernel>
auto fmha_fwd_create_kargs_and_grids(fmha_fwd_args args)
{
assert(args.nhead_q % args.nhead_k == 0);
auto kargs = [&] {
// create group mode kernel arguments
if constexpr(FmhaKernel::kIsGroupMode)
{
return FmhaKernel::MakeKargs(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.lse_ptr,
args.o_ptr,
args.seqstart_q_ptr,
args.seqstart_k_ptr,
args.seqlen_k_ptr,
args.hdim_q,
args.hdim_v,
args.nhead_q / args.nhead_k,
args.scale_s,
args.scale_p,
args.scale_o,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_o,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_lse,
args.nhead_stride_o,
args.window_size_left,
args.window_size_right,
args.mask_type);
}
else
{ // create batch mode kernel arguments
return FmhaKernel::MakeKargs(args.q_ptr,
args.k_ptr,
args.v_ptr,
args.bias_ptr,
args.lse_ptr,
args.o_ptr,
args.seqlen_q,
args.seqlen_k,
args.hdim_q,
args.hdim_v,
args.nhead_q / args.nhead_k,
args.scale_s,
args.scale_p,
args.scale_o,
args.stride_q,
args.stride_k,
args.stride_v,
args.stride_bias,
args.stride_o,
args.nhead_stride_q,
args.nhead_stride_k,
args.nhead_stride_v,
args.nhead_stride_bias,
args.nhead_stride_lse,
args.nhead_stride_o,
args.batch_stride_q,
args.batch_stride_k,
args.batch_stride_v,
args.batch_stride_bias,
args.batch_stride_lse,
args.batch_stride_o,
args.window_size_left,
args.window_size_right,
args.mask_type);
}
}();
dim3 grids = FmhaKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v);
return ck_tile::make_tuple(kargs, grids);
}
// this is used to pattern-match internl kernel implementation, not to instantiate kernel
template <ck_tile::index_t HDim_,
typename DataType_,
bool kIsGroupMode_,
ck_tile::index_t kM0_,
ck_tile::index_t kN0_,
ck_tile::index_t kK0_,
ck_tile::index_t kN1_,
ck_tile::index_t kK1_,
ck_tile::index_t kK0BlockLength_,
bool kIsVLayoutRowMajor_,
ck_tile::BlockFmhaPipelineEnum FmhaPipelineEnum_,
typename FmhaMask_,
bool kHasBias_,
bool kStoreLse_,
bool kDoFp8StaticQuant_,
bool kPadS_,
bool kPadSK_,
bool kPadD_,
bool kPadDv_>
struct fmha_fwd_traits_
{
static constexpr ck_tile::index_t HDim = HDim_;
using DataType = ck_tile::remove_cvref_t<DataType_>;
static constexpr bool kIsGroupMode = kIsGroupMode_;
static constexpr ck_tile::index_t kM0 = kM0_;
static constexpr ck_tile::index_t kN0 = kN0_;
static constexpr ck_tile::index_t kK0 = kK0_;
static constexpr ck_tile::index_t kN1 = kN1_;
static constexpr ck_tile::index_t kK1 = kK1_;
static constexpr ck_tile::index_t kK0BlockLength = kK0BlockLength_;
static constexpr bool kIsVLayoutRowMajor = kIsVLayoutRowMajor_;
static constexpr auto FmhaPipelineEnum = FmhaPipelineEnum_;
using FmhaMask = ck_tile::remove_cvref_t<FmhaMask_>;
static constexpr bool kHasBias = kHasBias_;
static constexpr bool kStoreLse = kStoreLse_;
static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_;
static constexpr bool kPadS = kPadS_;
static constexpr bool kPadSK = kPadSK_;
static constexpr bool kPadD = kPadD_;
static constexpr bool kPadDv = kPadDv_;
};
template <typename Traits_>
float fmha_fwd_(const ck_tile::stream_config&, fmha_fwd_args);
// This is the public API, will be generated by script
struct fmha_fwd_traits
{
int hdim_q;
int hdim_v;
std::string data_type;
bool is_group_mode;
bool is_v_rowmajor;
mask_enum mask_type;
bool has_bias;
bool has_lse;
bool do_fp8_static_quant;
// TODO: padding check is inside this api
};
float fmha_fwd(fmha_fwd_traits, fmha_fwd_args, const ck_tile::stream_config&);

605
example/ck_tile/01_fmha/generate.py Normal file
View File

@@ -0,0 +1,605 @@
# SPDX-License-Identifier: MIT
# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
# generate kernel instances to speed up compilation
import argparse
import itertools
from pathlib import Path
from typing import List, Optional, Tuple
from dataclasses import dataclass
import copy
import fnmatch
DTYPE_MAP = {
"fp16": "ck_tile::fp16_t",
"bf16": "ck_tile::bf16_t",
"fp8" : "ck_tile::fp8_t"
}
DTYPE_BITS = {
"fp32": 32,
"fp16": 16,
"bf16": 16,
"fp8" : 8,
"bf8" : 8
}
MASK_IMPL = {
"generic" : "ck_tile::GenericAttentionMask",
"simplified" : "ck_tile::SimplifiedGenericAttentionMask"
}
MASK_SIMPLIFIED_MAP = {
"s_no" : "ck_tile::SimplifiedGenericAttentionMask<false>",
"s_mask" : "ck_tile::SimplifiedGenericAttentionMask<true>",
}
MASK_MAP = {
"no" : "FmhaMasks::NoMask",
"causal" : "FmhaMasks::CausalMask",
"generic" : "FmhaMasks::GenericMask"
}
MODE_MAP = {
"batch" : "false",
"group" : "true"
}
LAYOUT_MAP = {
"row" : "true",
"col" : "false"
}
PIPELINE_MAP = {
"qr" : "ck_tile::BlockFmhaPipelineQRKSVS",
"qr_async" : "ck_tile::BlockFmhaPipelineQRKSVSAsync",
}
PIPELINE_ENUM_MAP = {
"qr" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS",
"qr_async" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC",
}
BOOL_MAP = {
"t" : "true",
"f" : "false"
}
DIRECTIONS = ["fwd"]
GEN_DIR = "" # in Cmake, have to generate files in same folder
FMHA_FWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.\n
// auto generated by generate.py
#include "fmha_fwd.hpp"
"""
FMHA_FWD_KERNEL_BODY="""
using fmha_dtype_{F_idx} = {F_dtype};
using fmha_block_tile_{F_idx} = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}>;
using fmha_block_warps_{F_idx} = ck_tile::sequence<{F_rm}, {F_rn}, {F_rk}>;
using fmha_warp_tile_{F_idx} = ck_tile::sequence<{F_wm}, {F_wn}, {F_wk}>;
using fmha_shape_{F_idx} = ck_tile::TileFmhaShape<fmha_block_tile_{F_idx},
fmha_block_warps_{F_idx},
fmha_warp_tile_{F_idx},
fmha_block_warps_{F_idx},
fmha_warp_tile_{F_idx},
{F_vlayout}>;
using fmha_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad},
{F_skpad},
{F_dpad},
{F_dvpad},
{F_bias},
{F_lse},
{F_squant},
{F_occupancy}>;
using fmha_mask_{F_idx} = {F_mask};
using fmha_pipeline_problem_{F_idx} = ck_tile::BlockFmhaPipelineProblem<
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::QDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::KDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::VDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SaccDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SMPLComputeDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::BiasDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::PDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
fmha_shape_{F_idx},
{F_mode},
fmha_mask_{F_idx},
fmha_trait_{F_idx}>;
using fmha_pipeline_{F_idx} = {F_pipeline}<
fmha_pipeline_problem_{F_idx}>;
using fmha_epilogue_{F_idx} =
ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
typename FmhaFwdTypeConfig<{F_dtype}>::ODataType,
{F_spad}, {F_dvpad}>>;
using fmha_kernel_{F_idx} =
ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner<fmha_shape_{F_idx}>,
fmha_pipeline_{F_idx},
fmha_epilogue_{F_idx}>;
using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout},
{F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
#include <iostream>
template<>
float fmha_fwd_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_args a)
{{
using k_ = fmha_kernel_{F_idx};
if(s.log_level_ > 0)
std::cout << ", " << k_::GetName() << std::flush;
auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
return ck_tile::launch_kernel<blocks.x, kBlockPerCu>(s, k_{{}}, grids, blocks, 0, kargs);
}}
"""
FMHA_FWD_API_FILENAME="fmha_fwd_api.cpp"
FMHA_FWD_API="""
float fmha_fwd(fmha_fwd_traits t, fmha_fwd_args a, const ck_tile::stream_config& s){{
float r = -1;
{F_dispatch}
return r;
}}
"""
FMHA_FWD_API_PER_DTYPE=""" {F_if}(t.data_type.compare(\"{F_dtype}\") == 0){{
{F_hdim_case}
}}
"""
FMHA_FWD_API_PER_HDIM_CASE=""" {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v <= {F_hdim}) {{
{F_inner_dispatch}
}}
"""
MASK_CHECK_MAP = {
"no" : "t.mask_type == mask_enum::no_mask",
"causal" : "t.mask_type == mask_enum::mask_top_left || t.mask_type == mask_enum::mask_bottom_right",
"generic" : "t.mask_type == mask_enum::window_generic",
}
MASK_SIMPLIFIED_CHECK_MAP = {
"s_no" : "t.mask_type == mask_enum::no_mask",
"s_mask" : "t.mask_type != mask_enum::no_mask",
}
FMHA_FWD_API_INNER_DISPATCH=""" {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.has_bias == {F_bias}) && (t.has_lse == {F_lse}) && (t.do_fp8_static_quant == {F_squant}) &&
({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
return fmha_fwd_<trait_>(s, a);
}}
"""
def get_mask_map(mask : str):
if mask == "generic":
return MASK_MAP
elif mask == "simplified":
return MASK_SIMPLIFIED_MAP
else:
assert False
return None
def get_mask_check_map(mask : str):
if mask == "generic":
return MASK_CHECK_MAP
elif mask == "simplified":
return MASK_SIMPLIFIED_CHECK_MAP
else:
assert False
return None
@dataclass
class FmhaFwdApiTrait:
pipeline_tag : str
# sync with fmha_fwd_traits<>, to generate fallback calls
hdim : str
dtype : str # data type
mode : str # value from MODE_MAP
bm0 : int # tile size along q seqlen (block size)
bn0 : int # tile size along qk seqlen
bk0 : int # tile size along qk gemm unroll
bn1 : int # tile size along v head_dim
bk1 : int # tile size along kv gemm unroll
bk0blen : int
vlayout : str
mask : str
bias : str # true/false
lse : str #
squant : str #
spad : str
skpad : str
dpad : str
dvpad : str
@property
def name(self) -> str:
return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0blen}-'+\
f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}'
@property
def scheck(self) -> str:
if self.mode == 'group': return 'true/*group mode spad always true*/' # group mode only generate spad/skpad == true
if self.pipeline_tag == 'qr_async':
if self.spad == 't' : return 'true' # always support
else : return 'true'
elif self.pipeline_tag in ['qr']:
if self.spad == 't' : return f'true /*a.seqlen_q % {self.bm0} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
else : return f'a.seqlen_q % {self.bm0} == 0'
else: assert False
@property
def skcheck(self) -> str:
if self.mode == 'group': return 'true/*group mode skpad always true*/' # group mode only generate spad/skpad == true
if self.pipeline_tag == 'qr_async':
if self.skpad == 't' : return f'a.seqlen_k % {self.bn0} != 0'
else : return f'a.seqlen_k % {self.bn0} == 0'
elif self.pipeline_tag in ['qr', 'qr_fp8']:
if self.skpad == 't' : return f'true /*a.seqlen_k % {self.bn0} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
else : return f'a.seqlen_k % {self.bn0} == 0'
else: assert False
@property
def dcheck(self) -> str:
if self.pipeline_tag == 'qr_async':
vec = int((32 * 4) / DTYPE_BITS[self.dtype])
if self.dpad == 't': return f'a.hdim_q % {vec} == 0'
else : assert False
elif self.pipeline_tag in ['qr']:
if self.dpad == 't': return f'true /*a.hdim_q % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
else : return f'a.hdim_q % {self.bk0blen} == 0'
else: assert False
@property
def dvcheck(self) -> str:
if self.pipeline_tag == 'qr_async':
vec = int((32 * 4) / DTYPE_BITS[self.dtype])
if self.dvpad == 't': return f'a.hdim_v % {vec} == 0'
else : assert False
elif self.pipeline_tag in ['qr']:
if self.dvpad == 't': return f'true /*a.hdim_v % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
else : return f'a.hdim_v % {self.bk0blen} == 0'
else: assert False
@dataclass
class FmhaFwdPipeline:
tag : str
F_vlayout : str # row/col
F_spad : str # true/false
F_skpad : str #
F_dpad : str #
F_dvpad : str #
F_bias : str # true/false
F_lse : str #
F_squant : str #
F_mask : str # value from MASK_MAP
@property
def name(self) -> str:
def pad_name() -> str:
n = ''
if self.F_spad == 't': n += 's'
if self.F_skpad == 't' : n += 'sk'
if self.F_dpad == 't' : n += 'd'
if self.F_dvpad == 't' : n += 'dv'
if n != '' : n = 'p' + n
return n
pn = pad_name()
n = f'{self.tag}_v{self.F_vlayout[0]}'
if pn != '' : n += f'_{pn}'
if self.F_bias == 't' : n += '_bias'
if self.F_mask[0:2] == 's_':
if self.F_mask == 's_mask': n += f'_mask'
else:
if self.F_mask != 'no' : n += f'_m{self.F_mask[0]}'
if self.F_lse == 't' : n += '_lse'
if self.F_squant == 't' : n += '_squant'
return n
class FmhaFwdApiPool:
def __init__(self, mask_impl):
self.pool = dict()
self.mask_impl = mask_impl
def register_traits(self, trait : FmhaFwdApiTrait) -> None:
# TODO: do we need to check duplication?
if trait.dtype not in self.pool.keys():
self.pool[trait.dtype] = dict()
if trait.hdim not in self.pool[trait.dtype].keys():
self.pool[trait.dtype][trait.hdim] = list()
self.pool[trait.dtype][trait.hdim].append(copy.copy(trait))
@property
def api(self) -> str:
per_dtypes=str()
for i, dtype in enumerate(self.pool.keys()):
per_hdim_case=str()
for j, hdim in enumerate(self.pool[dtype].keys()):
traits=self.pool[dtype][hdim]
inners=str()
for k, trait in enumerate(traits):
if_k = 'if' if k == 0 else 'else if'
inners = inners + FMHA_FWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_vlayout=LAYOUT_MAP[trait.vlayout],
F_pipeline_enum=PIPELINE_ENUM_MAP[trait.pipeline_tag], F_mask=get_mask_map(self.mask_impl)[trait.mask],
F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_bias=BOOL_MAP[trait.bias], F_lse=BOOL_MAP[trait.lse],
F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0blen=trait.bk0blen,
F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if'
per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_API.format(F_dispatch = per_dtypes)
@dataclass
class FmhaFwdTileSize:
F_bm0 : int # tile size along q seqlen (block size)
F_bn0 : int # tile size along qk seqlen
F_bk0 : int # tile size along qk gemm unroll
F_bn1 : int # tile size along v head_dim
F_bk1 : int # tile size along kv gemm unroll
F_bk0blen : int # total length of K0, used for pipeline that need load Q at once (or repeately load Q as a whole tile)
F_rm : int # number of warps along q seqlen (block warps)
F_rn : int # number of warps along k seqlen(not used)
F_rk : int # number of warps along gemm-k(not used)
F_wm : int # warp size along m (warp size)
F_wn : int # warp size along n
F_wk : int # warp size along k
F_occupancy : int # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy
@property
def name(self) -> str:
return f"b{self.F_bm0}x{self.F_bn0}x{self.F_bk0}x{self.F_bn1}x{self.F_bk1}x{self.F_bk0blen}" +\
f"_r{self.F_rm}x{self.F_rn}x{self.F_rk}_w{self.F_wm}x{self.F_wn}x{self.F_wk}" +\
("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
@dataclass
class FmhaFwdKernel:
direction : str
F_idx : int # this is not a tunable, but a counter to differentiate symbol
F_hdim : int # hdim
F_dtype : str # data type
F_mode : str # value from MODE_MAP
F_tile : FmhaFwdTileSize
F_pipeline : FmhaFwdPipeline
mask_impl : str
@property
def template(self) -> str:
kernel_body = str()
return FMHA_FWD_KERNEL_HEADER + \
FMHA_FWD_KERNEL_BODY.format(
F_idx = self.F_idx,
F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0,
F_bn0 = self.F_tile.F_bn0,
F_bk0 = self.F_tile.F_bk0,
F_bn1 = self.F_tile.F_bn1,
F_bk1 = self.F_tile.F_bk1,
F_bk0blen = self.F_tile.F_bk0blen,
F_rm = self.F_tile.F_rm,
F_rn = self.F_tile.F_rn,
F_rk = self.F_tile.F_rk,
F_wm = self.F_tile.F_wm,
F_wn = self.F_tile.F_wn,
F_wk = self.F_tile.F_wk,
F_vlayout = LAYOUT_MAP[self.F_pipeline.F_vlayout],
F_spad = BOOL_MAP[self.F_pipeline.F_spad],
F_skpad = BOOL_MAP[self.F_pipeline.F_skpad],
F_dpad = BOOL_MAP[self.F_pipeline.F_dpad],
F_dvpad = BOOL_MAP[self.F_pipeline.F_dvpad],
F_bias = BOOL_MAP[self.F_pipeline.F_bias],
F_lse = BOOL_MAP[self.F_pipeline.F_lse],
F_squant = BOOL_MAP[self.F_pipeline.F_squant],
F_occupancy = self.F_tile.F_occupancy,
F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag],
F_mask = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask],
F_mode = MODE_MAP[self.F_mode],
F_pipeline = PIPELINE_MAP[self.F_pipeline.tag])
@property
def name(self) -> str:
# TODO: we don't encode idx here
return f"fmha_{self.direction}_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" +\
self.F_tile.name + '_' + self.F_pipeline.name
@property
def filename(self) -> str:
return self.name + ".cpp"
def api_trait(self) -> FmhaFwdApiTrait:
return FmhaFwdApiTrait(
pipeline_tag=self.F_pipeline.tag,
hdim=str(self.F_hdim),
dtype=self.F_dtype,
mode=self.F_mode,
bm0=self.F_tile.F_bm0,
bn0=self.F_tile.F_bn0,
bk0=self.F_tile.F_bk0,
bn1=self.F_tile.F_bn1,
bk1=self.F_tile.F_bk1,
bk0blen=self.F_tile.F_bk0blen,
vlayout=self.F_pipeline.F_vlayout,
mask=self.F_pipeline.F_mask,
bias=self.F_pipeline.F_bias,
lse=self.F_pipeline.F_lse,
squant=self.F_pipeline.F_squant,
spad=self.F_pipeline.F_spad,
skpad=self.F_pipeline.F_skpad,
dpad=self.F_pipeline.F_dpad,
dvpad=self.F_pipeline.F_dvpad)
# TODO: design a more practical way to do it
# this is current supported tile size per hdim
def get_fmha_fwd_tile_dict_from_dtype(direction : str, dtype : str) -> Optional[dict]:
if direction == 'fwd':
if dtype == 'fp16' or dtype == 'bf16':
return {
'32' : FmhaFwdTileSize(128, 64, 16, 32, 32, 32, 2, 1, 1, 32, 32, 16, -1),
'64' : FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 4, 1, 1, 32, 32, 16, -1),
'128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 4, 1, 1, 32, 32, 16, -1),
'256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256, 4, 1, 1, 32, 32, 16, -1),
}
elif dtype == 'fp8' or dtype == 'bf8':
return {
'128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128, 4, 1, 1, 32, 32, 32, -1)
}
else:
return None
else:
return None
def get_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[FmhaFwdApiPool, List[FmhaFwdKernel]]:
# TODO: we don't support tuning yet, so pick up one value for vlayout/pipeline/pad
# support this in future
def get_pipelines(dtype, hdim) -> List[FmhaFwdPipeline]:
# this function will populate a list possible pipelines
# TODO: the order of List matters! the later in this list will be also be checked later
# TODO: currently for qr pipeline, let 't' padding to appear later!!
# TODO: how to design this more generic?
squant = 't' if dtype == 'fp8' else 'f'
pipelines = []
if dtype in ['fp16', 'bf16']:
for mask, bias, lse in itertools.product(get_mask_map(mask_impl).keys(), ["t", "f"], ["t", "f"]):
if hdim == 256:
# if True:
pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', bias, lse, squant, mask))
pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, lse, squant, mask))
pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, squant, mask))
pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', bias, lse, squant, mask))
else:
pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', bias, lse, squant, mask))
pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', bias, lse, squant, mask))
pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', bias, lse, squant, mask))
pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', bias, lse, squant, mask))
if receipt == 1:
pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, squant, mask)) # TODO: cover arbitraty hdim
pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 'f', 't', 't', bias, lse, squant, mask)) # TODO: cover arbitraty hdim
elif dtype in ['fp8', 'bf8']:
# no need lse kernels
for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), ["t", "f"]):
pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 'f', squant, mask))
else:
assert False
return pipelines
gen = list()
api_pool = FmhaFwdApiPool(mask_impl)
for direction, dtype in itertools.product(DIRECTIONS, DTYPE_MAP.keys()):
d = get_fmha_fwd_tile_dict_from_dtype(direction, dtype)
if d == None:
continue
#for hdim_str, mode, mask, bias, lse in itertools.product(d.keys(), MODE_MAP.keys(), MASK_MAP.keys(), ["t", "f"], ["t", "f"]):
for hdim_str, mode in itertools.product(d.keys(), MODE_MAP.keys()):
tile = d[hdim_str]
hdim = int(hdim_str)
for pipeline in get_pipelines(dtype, hdim):
if mode == "group":
if pipeline.F_spad != 't' or pipeline.F_skpad != 't':
# in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
continue
k = FmhaFwdKernel(direction=direction,
F_idx=0,
F_hdim=hdim,
F_dtype=dtype,
F_mode=mode,
F_tile=tile,
F_pipeline=pipeline,
mask_impl=mask_impl)
if kernel_filter != None:
if not fnmatch.fnmatch(k.name, kernel_filter):
continue
api_pool.register_traits(k.api_trait())
gen.append(k)
return (api_pool, gen)
def write_single_kernel(kernel: FmhaFwdKernel, autogen_dir: Path) -> None:
(autogen_dir / kernel.filename).write_text(kernel.template)
def write_api(api_pool : FmhaFwdApiPool, autogen_dir: Path) -> None:
(autogen_dir / FMHA_FWD_API_FILENAME).write_text(api_pool.api)
def write_blobs(output_dir : Optional[str], kernel_filter : Optional[str], receipt, mask_impl) -> None:
if output_dir is None:
output_dir = Path(__file__).parent
else:
output_dir = Path(output_dir) / GEN_DIR
output_dir.mkdir(parents=True, exist_ok=True)
api_pool, kernels = get_blobs(kernel_filter, receipt, mask_impl)
for kernel in kernels:
write_single_kernel(kernel, output_dir)
write_api(api_pool, output_dir)
# list all the files that will be generated
def list_blobs(output_file : Optional[str], kernel_filter : Optional[str], receipt, mask_impl) -> None:
assert output_file is not None
file_path = Path(output_file)
with file_path.open('a') as f:
_, kernels = get_blobs(kernel_filter, receipt, mask_impl)
for kernel in kernels:
f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_API_FILENAME) + "\n")
if __name__ == "__main__":
parser = argparse.ArgumentParser(
prog="generate",
description="gen api for CK fmha kernel",
)
parser.add_argument(
"-o",
"--output_dir",
required=False,
help="write all the blobs into a directory"
)
parser.add_argument(
"-l",
"--list_blobs",
required=False,
help="list all the kernels to a file"
)
# TODO: if using filter, must apply same value to output_dir and list_blobs
parser.add_argument(
"-f",
"--filter",
required=False,
help="filter out kernels that need to generate, using fnmatch module"
)
parser.add_argument(
"-m",
"--mask",
default="simplified",
required=False,
help="mask implementation, simplified/generic"
)
parser.add_argument(
"-r",
"--receipt",
default=0,
required=False,
help="codegen receipt. 0: generate only 8xhdim coverage\n" + \
" 1: generate more instance to cover all hdim"
)
args = parser.parse_args()
if args.list_blobs is not None:
list_blobs(args.list_blobs, args.filter, args.receipt, mask_impl=args.mask)
else:
write_blobs(args.output_dir, args.filter, args.receipt, mask_impl=args.mask)

159
example/ck_tile/01_fmha/mask.hpp Normal file
View File

@@ -0,0 +1,159 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <ostream>
#include <string>
#include "ck_tile/core.hpp"
#include "ck_tile/ops/fmha.hpp"
// keep this in sync with ck_tile::GenericAttentionMaskEnum
enum class mask_enum
{
no_mask = 0,
mask_top_left,
mask_bottom_right,
window_generic,
};
struct mask_info
{
mask_enum type;
ck_tile::index_t y, x;
ck_tile::index_t left, right; // FA style SWA left/right
void serialize(std::ostream& os) const
{
if(type == mask_enum::no_mask)
os << "n";
else if(type == mask_enum::mask_top_left)
os << "t(" << left << ":" << right << ")";
else if(type == mask_enum::mask_bottom_right)
os << "b(" << left << ":" << right << ")";
else
{
os << "g(" << y << ":" << x << ")";
}
}
static mask_info decode(std::string str, ck_tile::index_t seqlen_q, ck_tile::index_t seqlen_k)
{
ck_tile::index_t x_total = seqlen_k;
ck_tile::index_t y_total = seqlen_q;
mask_info tmp;
auto found_0 = str.find(':');
if(found_0 != std::string::npos)
{
std::string t = str.substr(0, found_0);
std::string v = str.substr(found_0 + 1);
if(t == "xt" || t == "xb")
{
// xformer style sliding window attn from top-left
ck_tile::index_t window_size = atoi(v.c_str());
ck_tile::index_t left_size = -1;
ck_tile::index_t right_size = 0;
if(window_size > 0)
{
left_size = window_size / 2;
right_size = window_size - 1 - left_size;
}
auto r = ck_tile::make_generic_attention_mask_coordinates_from_lr_window(
left_size, right_size, y_total, x_total, t == "xt");
tmp.type = t == "xt" ? mask_enum::mask_top_left : mask_enum::mask_bottom_right;
tmp.y = r.at(ck_tile::number<0>{});
tmp.x = r.at(ck_tile::number<1>{});
tmp.left = left_size;
tmp.right = right_size;
}
else
{
auto found_1 = v.find(",");
if(found_1 == std::string::npos)
{
printf("not supported value %s, %s\n", v.c_str(), str.c_str());
assert(0);
}
tmp.type = mask_enum::window_generic;
ck_tile::index_t v0 = atoi(v.substr(0, found_1).c_str());
ck_tile::index_t v1 = atoi(v.substr(found_1 + 1).c_str());
// TODO: some validation
if(t == "t")
{
tmp.type = mask_enum::mask_top_left;
auto r = ck_tile::make_generic_attention_mask_coordinates_from_lr_window(
v0, v1, y_total, x_total, true);
tmp.y = r.at(ck_tile::number<0>{});
tmp.x = r.at(ck_tile::number<1>{});
tmp.left = v0;
tmp.right = v1;
}
else if(t == "b")
{
tmp.type = mask_enum::mask_bottom_right;
auto r = ck_tile::make_generic_attention_mask_coordinates_from_lr_window(
v0, v1, y_total, x_total, false);
tmp.y = r.at(ck_tile::number<0>{});
tmp.x = r.at(ck_tile::number<1>{});
tmp.left = v0;
tmp.right = v1;
}
else if(t == "g")
{
tmp.y = v0;
tmp.x = v1;
tmp.left = v0; // TODO: don't use this?
tmp.right = v1;
}
else
{
printf("not supported type %s, %s\n", t.c_str(), str.c_str());
assert(0);
}
}
}
else
{
auto set_causal_top_left = [&]() {
tmp.type = mask_enum::mask_top_left;
tmp.y = seqlen_q;
tmp.x = 1;
tmp.left = -1;
tmp.right = 0;
};
auto set_causal_bottom_right = [&]() {
tmp.type = mask_enum::mask_bottom_right;
tmp.y = seqlen_q;
tmp.x = seqlen_k - seqlen_q + 1;
tmp.left = -1;
tmp.right = 0;
};
if(str == "t")
set_causal_top_left();
else if(str == "b")
set_causal_bottom_right();
else
{
tmp.type = static_cast<mask_enum>(atoi(str.c_str()));
if(tmp.type == mask_enum::mask_top_left)
{
set_causal_top_left();
}
else if(tmp.type == mask_enum::mask_bottom_right)
{
set_causal_bottom_right();
}
}
}
return tmp;
}
friend std::ostream& operator<<(std::ostream& os, const mask_info& mi);
};
inline std::ostream& operator<<(std::ostream& os, const mask_info& mi)
{
mi.serialize(os);
return os;
}

BIN
example/ck_tile/01_fmha/misc/gamc.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 29 KiB

32
example/ck_tile/01_fmha/script/benchmark.sh Executable file
View File

@@ -0,0 +1,32 @@
#!/bin/sh
# TODO: run this script from CK root
BUILD=build
EXE=$BUILD/bin/tile_example_fmha_fwd
VALID=0
for prec in "fp16" "bf16" ; do
for perm in 0 1 ; do
for hdim in 64 128 256 ; do
nhead=$((2048 / $hdim)) # follow fav2 setup
$EXE -prec=$prec -b=32 -h=$nhead -d=$hdim -s=512 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=16 -h=$nhead -d=$hdim -s=1024 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=8 -h=$nhead -d=$hdim -s=2048 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=4 -h=$nhead -d=$hdim -s=4096 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=2 -h=$nhead -d=$hdim -s=8192 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
$EXE -prec=$prec -b=1 -h=$nhead -d=$hdim -s=16384 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
done
done
done
for perm in 0 1 ; do
$EXE -prec=fp8 -squant=1 -b=32 -h=16 -d=128 -s=512 -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
$EXE -prec=fp8 -squant=1 -b=16 -h=16 -d=128 -s=1024 -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
$EXE -prec=fp8 -squant=1 -b=8 -h=16 -d=128 -s=2048 -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
$EXE -prec=fp8 -squant=1 -b=4 -h=16 -d=128 -s=4096 -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
$EXE -prec=fp8 -squant=1 -b=2 -h=16 -d=128 -s=8192 -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
$EXE -prec=fp8 -squant=1 -b=1 -h=16 -d=128 -s=16384 -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
done

45
example/ck_tile/01_fmha/script/smoke_test.sh Executable file
View File

@@ -0,0 +1,45 @@
#!/bin/sh
# TODO: run this script from CK root
BUILD=build
EXE=$BUILD/bin/tile_example_fmha_fwd
KNAME=1
export CK_WARMUP=0
export CK_REPEAT=1
COMMON_ARGS='-v=1 -warmup=0 -repeat=1'
# mode=0
# export HIP_VISIBLE_DEVICES=4
for prec in "fp16" "bf16" ; do
for mode in 1 0 ; do
for perm in 0 1 ; do
for vlayout in "r" "c" ; do
for hdim in 32 64 128 256 ; do
for lse in 0 1 ; do
for bias in 0 1 ; do
# $EXE -prec=$prec -mode=$mode -b=1 -h=1 -d=$hdim -s=1024 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=2 -h_k=1 -d=16, -d_v=$hdim -s=55 -s_k=256 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=3 -d=$hdim -s=100 -s_k=51 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=16 -d_v=$hdim -s=99 -s_k=256 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=1 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1024 -s_k=256 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -d_v=24 -s=3 -s_k=99 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=3 -h=2 -h_k=1 -d=$hdim -s=200 -s_k=520 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=t:128,30 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -s=99 -s_k=32 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=b:4,35 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
done
done
done
done
done
done
done
for perm in 0 1 ; do
for bias in 0 1 ; do
for b in 1 2 ; do
$EXE -prec=fp8 -init=3 -b=$b -h=1 -d=128 -s=128 -bias=$bias -iperm=$perm -operm=$perm -vlayout=c -squant=1 -kname=$KNAME $COMMON_ARGS
done
done
done

92
example/ck_tile/01_fmha/utils.hpp Normal file
View File

@@ -0,0 +1,92 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdint>
#include <optional>
#include <ostream>
#include <tuple>
#include <utility>
#include <vector>
#include <functional>
#include "ck_tile/core/container/span.hpp"
enum class mode_enum
{
batch = 0,
group
};
std::ostream& operator<<(std::ostream& stream, mode_enum mode)
{
return stream << (mode == mode_enum::batch ? "batch" : "group");
}
std::vector<int32_t> to_seqstarts(ck_tile::span<const int32_t> seqlens)
{
std::vector<int32_t> seqstarts = {0};
for(int32_t seqlen : seqlens)
{
seqstarts.push_back(seqstarts.back() + seqlen);
}
assert(seqstarts.size() == seqlens.size() + 1);
return seqstarts;
}
std::vector<int32_t> generate_seqlens(mode_enum mode,
unsigned count,
int32_t seqlens_sum,
std::optional<unsigned> seed = std::nullopt)
{
assert(0 < count);
std::vector<int32_t> seqlens(count, seqlens_sum);
if(mode == mode_enum::group && 1 < count)
{
using size_type = std::vector<int32_t>::size_type;
std::mt19937 random_engine(seed.has_value() ? *seed : std::random_device{}());
std::uniform_int_distribution<size_type> idx_dist(0, count - 1);
auto next_idx = std::bind(idx_dist, std::ref(random_engine));
std::uniform_int_distribution<size_type> step_dist(1, count - 1);
auto next_step = std::bind(step_dist, std::ref(random_engine));
for(unsigned repeat = seqlens_sum * (count / 2); 0 < repeat; --repeat)
{
const size_type to_decrease = next_idx();
// make sure each elements of seqlens is always greater than 0
if(seqlens[to_decrease] == 1)
{
continue;
}
const size_type to_increase = (to_decrease + next_step()) % count;
--seqlens[to_decrease];
++seqlens[to_increase];
}
}
return seqlens;
}
std::vector<int32_t> generate_seqstarts(mode_enum mode,
unsigned count,
int32_t seqlens_sum,
std::optional<unsigned> seed = std::nullopt)
{
return to_seqstarts(generate_seqlens(mode, count, seqlens_sum, seed));
}
int env_get_int(const char* var_name, int default_int)
{
char* v = getenv(var_name);
int r = default_int;
if(v)
r = atoi(v);
return r;
}