Skip to content

GitLab

"docs/vscode:/vscode.git/clone" did not exist on "f87b35b227fb1c1309cb9e9fb72bf22939fbfbc9"
Commit 7c56cd01 authored by Astha Rai's avatar Astha Rai
Browse files

fixing merge errors: unary_element_wise_operation.hpp

parents 7d3ee266 cb6c5d39
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2103 additions and 512 deletions
example/ck_tile/01_fmha/fmha_fwd.cpp
View file @ 7c56cd01
......@@ -557,33 +557,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
}
#endif
struct
{
auto operator()(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};
}
auto operator()(bool permute,
ck_tile::index_t ns /*num_splits*/,
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, 5>{ns, b, h, s, d};
else
return std::array<ck_tile::index_t, 5>{ns, b, s, h, d};
}
} get_lengths;
static const 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");
......@@ -635,12 +618,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::HostTensor<LSEDataType> lse_acc_host(
1 < num_splits || use_kvcache
? std::array<ck_tile::index_t, 4>{num_splits, shape_batch, nhead, shape_seqlen_q}
? std::array<ck_tile::index_t, 4>{shape_batch, nhead, num_splits, shape_seqlen_q}
: std::array<ck_tile::index_t, 4>{1, 1, 1, 1});
ck_tile::HostTensor<OaccDataType> o_acc_host(
1 < num_splits || use_kvcache
? get_lengths(o_perm, num_splits, shape_batch, nhead, shape_seqlen_q, hdim_v)
: std::array<ck_tile::index_t, 5>{1, 1, 1, 1, 1});
1 < num_splits || use_kvcache ? std::array<ck_tile::index_t, 5>{shape_batch,
nhead,
num_splits,
shape_seqlen_q,
hdim_v}
: std::array<ck_tile::index_t, 5>{1, 1, 1, 1, 1});
// batch mode of lse data layout is [batch, nhead, seqlen_q]
// group mode of lse data layout is [nhead, total_seqlen_q]
......@@ -880,7 +866,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
}();
const ck_tile::index_t stride_bias = (i_perm ? shape_seqlen_k : 1 * shape_seqlen_k);
const ck_tile::index_t stride_randval = (max_seqlen_k);
const ck_tile::index_t stride_o_acc = (o_perm ? hdim_v : nhead * hdim_v);
const ck_tile::index_t stride_o_acc = (hdim_v);
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);
......@@ -906,8 +892,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
(i_perm ? 0 * shape_seqlen_q * shape_seqlen_k : 0 * shape_seqlen_k);
const ck_tile::index_t nhead_stride_randval = (shape_seqlen_q * max_seqlen_k);
const ck_tile::index_t nhead_stride_lse = shape_seqlen_q;
const ck_tile::index_t nhead_stride_lse_acc = shape_seqlen_q;
const ck_tile::index_t nhead_stride_o_acc = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
const ck_tile::index_t nhead_stride_lse_acc = (num_splits * shape_seqlen_q);
const ck_tile::index_t nhead_stride_o_acc = (num_splits * shape_seqlen_q * hdim_v);
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);
......@@ -922,13 +908,13 @@ bool run(const ck_tile::ArgParser& arg_parser)
const ck_tile::index_t batch_stride_bias = (0 * nhead * shape_seqlen_q * shape_seqlen_k);
const ck_tile::index_t batch_stride_randval = (nhead * shape_seqlen_q * max_seqlen_k);
const ck_tile::index_t batch_stride_lse = (nhead * shape_seqlen_q);
const ck_tile::index_t batch_stride_lse_acc = (nhead * shape_seqlen_q);
const ck_tile::index_t batch_stride_o_acc = (nhead * shape_seqlen_q * hdim_v);
const ck_tile::index_t batch_stride_o = (nhead * shape_seqlen_q * hdim_v);
const ck_tile::index_t batch_stride_lse_acc = (nhead * num_splits * shape_seqlen_q);
const ck_tile::index_t batch_stride_o_acc = (nhead * num_splits * shape_seqlen_q * hdim_v);
const ck_tile::index_t batch_stride_o = (nhead * shape_seqlen_q * hdim_v);
const ck_tile::index_t batch_stride_block_table = (max_num_page_blocks / batch);
// setup split_stride_* arguments (only used in split-kv kernel)
const ck_tile::index_t split_stride_lse_acc = (shape_batch * nhead * shape_seqlen_q);
const ck_tile::index_t split_stride_o_acc = (shape_batch * nhead * shape_seqlen_q * hdim_v);
const ck_tile::index_t split_stride_lse_acc = (shape_seqlen_q);
const ck_tile::index_t split_stride_o_acc = (shape_seqlen_q * hdim_v);
args.q_ptr = q_buf.GetDeviceBuffer();
args.k_ptr = k_buf.GetDeviceBuffer();
......
example/ck_tile/02_layernorm2d/CMakeLists.txt
View file @ 7c56cd01
# 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"
add_executable(tile_example_layernorm2d_fwd EXCLUDE_FROM_ALL layernorm2d_fwd.cpp)
target_compile_options(tile_example_layernorm2d_fwd PRIVATE -DSAVE_MEAN_INV_STD)
\ No newline at end of file
set(LAYERNORM2D_FWD_KNOWN_APIS "fwd;bwd")
set(LAYERNORM2D_FWD_ENABLE_APIS "fwd" CACHE STRING
"semicolon-separated list of APIs to generate (${LAYERNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
if(LAYERNORM2D_FWD_ENABLE_APIS STREQUAL "all")
set(LAYERNORM2D_FWD_ENABLE_APIS ${LAYERNORM2D_FWD_KNOWN_APIS})
endif()
# generate a list of kernels, but not actually emit files at config sta
execute_process(
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api ${LAYERNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --list_blobs
RESULT_VARIABLE ret
)
if(ret AND NOT ret EQUAL 0)
message( FATAL_ERROR "Fail to generate kernels via Python. ${ret}")
endif()
file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/layernorm2d_fwd_blobs.txt LAYERNORM2D_FWD_GEN_BLOBS)
add_custom_command(
OUTPUT ${LAYERNORM2D_FWD_GEN_BLOBS}
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api ${LAYERNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --gen_blobs
)
set(EXAMPLE_LAYERNORM2D_FWD "tile_example_layernorm2d_fwd")
message("adding example ${EXAMPLE_LAYERNORM2D_FWD}")
add_executable(${EXAMPLE_LAYERNORM2D_FWD} EXCLUDE_FROM_ALL layernorm2d_fwd.cpp)
target_include_directories(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${LAYERNORM2D_FWD_GEN_BLOBS})
set(EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
target_compile_options(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS})
# TODO: we have to turn off this global prop, otherwise the progress bar generated
# by cmake will print too many files, execvp: /bin/sh: Argument list too long
# however, this property may affect global
# TODO: consider codegen a makefile by us
set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)
example/ck_tile/02_layernorm2d/README.md
View file @ 7c56cd01
# Layernorm2D forward
This folder contains example for Layernorm2D forward using ck_tile tile-programming implementation.
This folder contains example for Layernorm2D forward using `ck_tile` tile-programming implementation.
# Implementation and feature support
## welford online algorithm
We use welfold algorithm to update `mean`/`variance` block by block. For `N <=4096` case we can compute `mean`/`var`/`normalization` within one loop, we call it `one-pass`. For large N case, it is hard to keep `mean`/`var` inside register/LDS and then computation `normalization`, so we need to load input twice, first time to compute `mean`/`var` block-by-block, then load input another time to compute the `normalization`. We call it `two-pass`.
## mean/variance save
In training case the mean/variance need to store out (TBD, not supported yet)
## prenorm/postnorm
![](misc/pnorm.png)
since [prenorm/postnorm](https://arxiv.org/pdf/1906.01787) is quite common in LLM blocks, this example boosts this feature by kernel fusion. Note that `prenorm`/`postnorm` always need to do elementwise-add a `shortcut` before the actual layernorm computation, and optionally store out the result to global. You can use `-fadd=1` to test `pre-add+store`, or `-fadd=2` to test `pre-add` without store out (not codegen by default).
## smooth-quant/dynamic-quant
we support smooth/dynamic quantization for `int8` output, by setting `-fquant=1` and `-prec_o=int8`. In this case the output will doing a rowwise dynamic quantization like below. Note that smooth-quant require input a `(1*N)` size per-channel scale(in fp32 in our example, though this is customizable), then elememt-wise multiply the tensor for each row, then compute the rowwise dynamic quant. if set `-fquant=2` will have the input per-channel scale stage, only the dynamic quant. This case is supported in our kernel but by default not generated (TBD: add some filter in generate.py support on-demand codegen)
![](misc/dquant.png)
```
# assume output int8, hidden_states is [m, n] shape and in fp16/bf16
# [m, 1]
per_token_amax, _ = torch.max(
input=torch.abs(hidden_states),
dim=-1,
keepdim=True
)
per_token_scale = per_token_amax.to(dtype=torch.float32) / 127.0
# quant hidden_states
hidden_states = (hidden_states / per_token_scale).to(dtype=torch.int8)
return hidden_states, per_token_scale
# hidden_states now is int8 will feed to next layer as intput
# per_token_scale will be used as dequant factor later layer
```
## build
```
# in the root of ck_tile
mkdir build && cd build
# you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
sh ../script/cmake-ck-dev.sh ../ <arch>
sh ../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_example_layernorm2d_fwd -j
```
This will result in an executable `build/bin/tile_example_layernorm2d_fwd`
......@@ -16,8 +51,35 @@ This will result in an executable `build/bin/tile_example_layernorm2d_fwd`
```
args:
-m m dimension (default:3328)
-n m dimension (default:4096)
-n n dimension (default:4096)
-stride stride per row, if -1 then equal to n (default:-1)
-e epsilon (default:1e-5)
-save_mv save mean/variance(invstd) or not. set to 1 in training case (default:0)
-v cpu validation or not (default:1)
-prec precision (default:fp16)
```
\ No newline at end of file
-kname print kernel name or not (default:1)
-prec_i input precision (default:fp16)
-prec_o output precision, set auto will be the same as input (default:auto)
-prec_sx output quant scale type, set auto will be the same as input. used when fquant=1 (default:auto)
-prec_sy output quant scale type, set auto will be the same as input. used when fquant=1 or 2 (default:auto)
-fadd fused-add, 0:no fused add, 1:preadd+store, 2:preadd only (default:0)
-fquant fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant (default:0)
-warmup cold iter (default:5)
-repeat hot iter (default:20)
```
## limitations
Note that `fquant=2`, `fadd=2`, `prec_sx/prec_sy` other than `fp32` are not by default generated. Though our kernel template suppor this. (TBD: add some flag in generate.py) to generate those instance on demand. Beside, `N>8192` case will by default using two-pass pipeline, and `-fquant=1/2` are not supported yet. If need suport `N>8192` and `fused+residual+store`, you can use this example together with `12_smoothquant`, to construct layernorm+residual, and smoothquant, 2 kernels for this purpose.
```
# some case
# standard fp16 layernorm 2d, m=10. n=1024
./build/bin/tile_example_layernorm2d_fwd -m=10 -n=1024
# standard fp16 layernorm 2d, m=10. n=1024, fused-smooth-quant, output in int8
./build/bin/tile_example_layernorm2d_fwd -m=10 -n=1024 -prec_o=int8 -fquant=1
# standard fp16 layernorm 2d, m=10. n=1024, fused-smooth-quant+fused-add-store, output in int8
./build/bin/tile_example_layernorm2d_fwd -m=10 -n=1024 -prec_o=int8 -fquant=1 -fadd=1
```
example/ck_tile/02_layernorm2d/generate.py 0 → 100644
View file @ 7c56cd01
This diff is collapsed.
example/ck_tile/02_layernorm2d/layernorm2d_fwd.hpp
View file @ 7c56cd01
......@@ -8,23 +8,57 @@
#include "ck_tile/ops/layernorm2d.hpp"
#include <string>
struct layernorm2d_fwd_traits
template <typename InType, typename OutType, typename XScaleDataType_, typename YScaleDataType_>
struct LayerNormTypeConfig;
template <typename OutType, typename XScaleDataType_, typename YScaleDataType_>
struct LayerNormTypeConfig<ck_tile::half_t, OutType, XScaleDataType_, YScaleDataType_>
{
using XDataType = ck_tile::half_t;
using YDataType = OutType;
using GammaDataType = ck_tile::half_t;
using BetaDataType = ck_tile::half_t;
using MeanDataType = ck_tile::half_t;
using InvStdDataType = ck_tile::half_t;
using ComputeDataType = float;
using XScaleDataType = XScaleDataType_;
using YScaleDataType = YScaleDataType_;
};
template <typename OutType, typename XScaleDataType_, typename YScaleDataType_>
struct LayerNormTypeConfig<ck_tile::bf16_t, OutType, XScaleDataType_, YScaleDataType_>
{
std::string data_type;
using XDataType = ck_tile::bf16_t;
using YDataType = OutType;
using GammaDataType = ck_tile::bf16_t;
using BetaDataType = ck_tile::bf16_t;
using MeanDataType = ck_tile::bf16_t;
using InvStdDataType = ck_tile::bf16_t;
using ComputeDataType = float;
using XScaleDataType = XScaleDataType_;
using YScaleDataType = YScaleDataType_;
};
struct layernorm2d_fwd_args
// runtime args
struct layernorm2d_fwd_args : public ck_tile::Layernorm2dFwdHostArgs
{
const void* p_x;
const void* p_gamma;
const void* p_beta;
void* p_y;
void* p_mean;
void* p_invStd;
float epsilon;
ck_tile::index_t M;
ck_tile::index_t N;
};
// host API
// This is the public API, will be generated by script
struct layernorm2d_fwd_traits
{
std::string prec_i; // input precision
std::string prec_o; // output precision
// if fused_quant == 1, need set prec_sx/prec_sy to proper string, otherwise can set
// arbitrary(will skip check) if fused_quant == 2, need set prec_sy to proper string, otherwise
// can set arbitrary(will skip check)
std::string prec_sx; // x-scale, used for [1*N] input smooth quant
std::string prec_sy; // y-scale, used for [M*1] output for next layer
bool save_mean_var; //
int fused_add; // 0:no-add, 1:pre-add-store, 2:pre-add
int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
};
float layernorm2d_fwd(layernorm2d_fwd_traits, layernorm2d_fwd_args, const ck_tile::stream_config&);
example/ck_tile/02_layernorm2d/script/perf_test.sh 0 → 100755
View file @ 7c56cd01
#!/bin/sh
EXE="$(find . -name tile_example_layernorm2d_fwd -type f | head -n 1)"
$EXE -m=1 -n=1 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
\ No newline at end of file
example/ck_tile/02_layernorm2d/script/smoke_test.sh 0 → 100755
View file @ 7c56cd01
#!/bin/sh
EXE="$(find . -name tile_example_layernorm2d_fwd -type f | head -n 1)"
for fquant in "" "-fquant=1 -prec_o=int8"; do
for pr_i in "fp16" "bf16" ; do
for fadd in "0" "1"; do
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=99 -n=13
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=17 -n=16
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1 -n=100
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=4 -n=128
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=80 -n=127
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=22 -n=255 -stride=256
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7 -n=599
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=19 -n=512
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=33 -n=313 -stride=1000
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=11 -n=510
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=171 -n=676 -stride=818
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=91 -n=636
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=12 -n=768 -stride=800
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=100 -n=766 -stride=812
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=31 -n=1024
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=64 -n=1000 -stride=1004
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=8 -n=1501
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3 -n=1826
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=5 -n=2040
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7 -n=2734
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1 -n=3182
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=9 -n=4096
$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3 -n=8192
#$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1 -n=10547
#$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3 -n=17134
done
done
done
example/ck_tile/03_gemm/CMakeLists.txt
View file @ 7c56cd01
set(CMAKE_BUILD_TYPE Debug)
add_executable(tile_example_gemm_basic EXCLUDE_FROM_ALL gemm_basic.cpp)
\ No newline at end of file
add_executable(tile_example_gemm_basic EXCLUDE_FROM_ALL gemm_basic.cpp)
add_executable(tile_example_gemm_mem_pipeline EXCLUDE_FROM_ALL gemm_mem_pipeline.cpp)
example/ck_tile/03_gemm/gemm_basic.cpp
View file @ 7c56cd01
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gemm_basic.hpp"
#include <hip/hip_runtime.h>
#include <cstring>
......@@ -10,51 +9,48 @@
#include <string>
#include <tuple>
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("b", "1", "batch size")
.insert("m", "1024", "m dimension")
.insert("n", "2048", "n dimension")
.insert("k", "64", "k dimension")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")
.insert("stride_c", "0", "Tensor C stride")
.insert("v", "2", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
.insert("e", "1e-5", "Absolute error tolerance")
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
.insert("warmup", "10", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/host.hpp"
#include "gemm_basic.hpp"
template <typename LayoutA,
typename LayoutB,
typename LayoutC,
typename PipelineProblem,
typename GemmPipeline,
typename GemmShape>
template <typename ALayout, typename BLayout, typename CLayout>
float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
{
// The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadA = true;
constexpr bool kPadB = true;
constexpr bool kPadC = true;
constexpr bool kTilePermute = false;
// The rank and permutation will also be generate out by the CodeGen part.
constexpr ck_tile::index_t kOutputRank = 2;
constexpr int kBlockPerCu = 1;
using TilePartitioner = ck_tile::GemmTilePartitioner<GemmShape>;
// This part comes from the Codegen
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 128;
constexpr ck_tile::index_t K_Tile = 32;
// The rank and permutation will also be generate out by the CodeGen part.
constexpr ck_tile::index_t kOutputRank = 2;
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
// Whether doing the CShuffle (transpose before the global memory), depending on the output
// layout.
constexpr bool CShuffleEpilogue =
std::is_same_v<LayoutC, ck_tile::tensor_layout::gemm::ColumnMajor>;
std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTilePartitioner<CodegenGemmShape>;
using GemmEpilogue = std::conditional_t<
CShuffleEpilogue,
......@@ -70,14 +66,21 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
TilePartitioner::kN>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadA, kPadB>>>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadA, kPadB, kPadC, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy<ALayout, BLayout, CLayout>;
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenGemmPolicy>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args.p_a,
args.p_b,
args.p_c,
args.epsilon,
args.M,
args.N,
args.K,
......@@ -88,299 +91,20 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
const dim3 grids = Kernel::GridSize(args.M, args.N, args.kbatch);
constexpr dim3 blocks = Kernel::BlockSize();
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}
template <typename DataType,
typename LayoutA,
typename LayoutB,
typename LayoutC,
typename PipelineProblem,
typename GemmPipeline,
typename GemmShape>
float invoke_gemm(ck_tile::DeviceMem& a_buf,
ck_tile::DeviceMem& b_buf,
ck_tile::DeviceMem& c_buf,
const ck_tile::ArgParser& arg_parser)
{
std::string data_type = arg_parser.get_str("prec");
if(data_type != DataTypeTraits<DataType>::name)
{
std::cerr << "Data type mismatch: expected " << DataTypeTraits<DataType>::name << ", got "
<< data_type << std::endl;
return -1; // Or handle the error appropriately
}
float epsilon = arg_parser.get_float("e");
ck_tile::index_t batch_size = arg_parser.get_int("b");
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
ck_tile::index_t K = arg_parser.get_int("k");
ck_tile::index_t stride_a = arg_parser.get_int("stride_a");
ck_tile::index_t stride_b = arg_parser.get_int("stride_b");
ck_tile::index_t stride_c = arg_parser.get_int("stride_c");
gemm_basic_args args;
args.p_a = a_buf.GetDeviceBuffer();
args.p_b = b_buf.GetDeviceBuffer();
args.p_c = c_buf.GetDeviceBuffer();
args.epsilon = epsilon;
args.kbatch = batch_size;
args.M = M;
args.N = N;
args.K = K;
// Only set stride_M and stride_N if they are non-zero and not equal to K.
if(stride_a != 0)
{
args.stride_A = stride_a;
}
else
{
args.stride_A = [&]() {
if constexpr(std::is_same_v<LayoutA, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
return M;
}
else
{
return K;
}
}();
}
if(stride_b != 0)
{
args.stride_B = stride_b;
}
else
if(s.log_level_ > 0)
{
args.stride_B = [&]() {
if constexpr(std::is_same_v<LayoutB, ck_tile::tensor_layout::gemm::RowMajor>)
{
return N;
}
else
{
return K;
}
}();
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
if(stride_c != 0)
{
args.stride_C = stride_c;
}
else
{
args.stride_C = [&]() {
if constexpr(std::is_same_v<LayoutC, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
return M;
}
else
{
return N;
}
}();
}
float ave_time = gemm_calc<LayoutA, LayoutB, LayoutC, PipelineProblem, GemmPipeline, GemmShape>(
args, ck_tile::stream_config{nullptr, true});
std::size_t num_byte =
sizeof(ADataType) * M * K + sizeof(BDataType) * N * K + sizeof(CDataType) * M * N;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "The overall perfomance of the GEMM with "
<< "[" << data_type << "]"
<< "batch size: " << batch_size << ". m:" << M << ", n:" << N << ", k:" << K
<< " is: \n";
std::cout << "Running time: " << ave_time << "ms, Throughput " << gb_per_sec << "GB/s \n"
<< std::flush;
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
ck_tile::index_t K = arg_parser.get_int("k");
// The Matrix Multiplication goes with Matrix A (M, K), Matrix B (N, K) = Matrix C (M, N).
using matrix_a_layout = ck_tile::tensor_layout::gemm::RowMajor;
using matrix_b_layout = ck_tile::tensor_layout::gemm::ColumnMajor;
using matrix_c_layout = ck_tile::tensor_layout::gemm::RowMajor;
// host verify
std::vector<int> a_dimensions =
(std::is_same_v<matrix_a_layout, ck_tile::tensor_layout::gemm::RowMajor>)
? std::vector<int>{M, K}
: std::vector<int>{K, M};
std::vector<int> b_dimensions =
(std::is_same_v<matrix_b_layout, ck_tile::tensor_layout::gemm::ColumnMajor>)
? std::vector<int>{N, K}
: std::vector<int>{K, N};
std::vector<int> c_dimensions =
(std::is_same_v<matrix_c_layout, ck_tile::tensor_layout::gemm::RowMajor>)
? std::vector<int>{M, N}
: std::vector<int>{N, M};
ck_tile::HostTensor<ADataType> a_host(a_dimensions);
ck_tile::HostTensor<BDataType> b_host(b_dimensions);
ck_tile::HostTensor<CDataType> c_host_ref(c_dimensions);
ck_tile::HostTensor<CDataType> c_host_dev(c_dimensions);
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_host);
ck_tile::DeviceMem a_buf(a_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem b_buf(b_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem c_buf(c_host_dev.get_element_space_size_in_bytes());
a_buf.ToDevice(a_host.data());
b_buf.ToDevice(b_host.data());
// The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadA = true;
constexpr bool kPadB = true;
constexpr bool kPadC = true;
// This part comes from the Codegen
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 128;
constexpr ck_tile::index_t K_Tile = 32;
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using CodegenGemmTraits = ck_tile::
TileGemmTraits<kPadA, kPadB, kPadC, matrix_a_layout, matrix_b_layout, matrix_c_layout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPolicy = ck_tile::
UniversalGemmPipelineAgBgCrPolicy<matrix_a_layout, matrix_b_layout, matrix_c_layout>;
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenGemmPolicy>;
invoke_gemm<ck_tile::half_t,
matrix_a_layout,
matrix_b_layout,
matrix_c_layout,
CodegenPipelineProblem,
CodegenGemmPipeline,
CodegenGemmShape>(a_buf, b_buf, c_buf, arg_parser);
c_buf.FromDevice(c_host_dev.data());
bool pass_cpu = true;
if(arg_parser.get_int("v") == 1)
{
// ToDo: Will Add the Element Op (bias) verification in the future.
ck_tile::reference_gemm<ADataType,
BDataType,
AccDataType,
CDataType,
matrix_a_layout,
matrix_b_layout,
matrix_c_layout>(a_host, b_host, c_host_ref);
pass_cpu = ck_tile::check_err(c_host_dev, c_host_ref);
std::cout << "The CPU veification result is:" << (pass_cpu ? "correct" : "fail")
<< std::flush;
}
bool pass_gpu = true;
if(arg_parser.get_int("v") == 2)
{
ck_tile::index_t stride_a = arg_parser.get_int("stride_a");
ck_tile::index_t stride_b = arg_parser.get_int("stride_b");
ck_tile::index_t stride_c = arg_parser.get_int("stride_c");
if(stride_a == 0)
{
if constexpr(std::is_same_v<matrix_a_layout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
stride_a = M;
}
else
{
stride_a = K;
}
}
if(stride_b == 0)
{
if constexpr(std::is_same_v<matrix_b_layout, ck_tile::tensor_layout::gemm::RowMajor>)
{
stride_b = N;
}
else
{
stride_b = K;
}
}
if(stride_c == 0)
{
if constexpr(std::is_same_v<matrix_c_layout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
stride_c = M;
}
else
{
stride_c = N;
}
}
ck_tile::HostTensor<CDataType> c_host_gpu_ref(c_dimensions);
ck_tile::DeviceMem c_gpu_buf(c_host_gpu_ref.get_element_space_size_in_bytes());
#include "run_gemm_example.inc"
ck_tile::reference_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
matrix_a_layout,
matrix_b_layout,
matrix_c_layout>(
a_buf, b_buf, c_gpu_buf, M, N, K, stride_a, stride_b, stride_c);
c_buf.FromDevice(c_host_gpu_ref.data());
pass_gpu = ck_tile::check_err(c_host_dev, c_host_gpu_ref);
std::cout << "The GPU veification result is: " << (pass_gpu ? "correct" : "fail")
<< std::flush;
}
std::cout << std::endl << std::flush;
return !pass_gpu;
}
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/ck_tile/03_gemm/gemm_basic.hpp
View file @ 7c56cd01
......@@ -4,12 +4,10 @@
#pragma once
#include <string>
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/host.hpp"
#include <string>
template <typename DataType>
struct GemmBasicTypeConfig;
......@@ -20,7 +18,7 @@ struct GemmBasicTypeConfig<ck_tile::half_t>
using ADataType = ck_tile::half_t;
using BDataType = ck_tile::half_t;
using AccDataType = float;
using CDataType = ck_tile::half_t; // type convert
using CDataType = ck_tile::half_t;
// ToDo: Add more bias config to support different categories of GEMM.
};
......@@ -58,7 +56,6 @@ struct gemm_basic_args
const void* p_a;
const void* p_b;
void* p_c;
float epsilon;
ck_tile::index_t kbatch;
ck_tile::index_t M;
ck_tile::index_t N;
......@@ -68,5 +65,28 @@ struct gemm_basic_args
ck_tile::index_t stride_C;
};
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("b", "1", "batch size")
.insert("m", "3840", "m dimension")
.insert("n", "4096", "n dimension")
.insert("k", "2048", "k dimension")
.insert("a_layout", "R", "A tensor data layout - Row by default")
.insert("b_layout", "R", "B tensor data layout - Row by default")
.insert("c_layout", "R", "C tensor data layout - Row by default")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")
.insert("stride_c", "0", "Tensor C stride")
.insert("v", "2", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
.insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
.insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
// host API
float gemm_calc(gemm_basic_args args, const ck_tile::stream_config& s);
example/ck_tile/03_gemm/gemm_mem_pipeline.cpp 0 → 100644
View file @ 7c56cd01
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
#include <cstring>
#include <iostream>
#include <sstream>
#include <string>
#include <tuple>
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/host.hpp"
#include "gemm_basic.hpp"
template <typename ALayout, typename BLayout, typename CLayout>
float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
{
// ToDo: This will be modified by the codegen code later.
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 128;
constexpr ck_tile::index_t K_Tile = 32;
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
// The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadA = true;
constexpr bool kPadB = true;
constexpr bool kPadC = true;
constexpr int kBlockPerCu = 1;
// ===============================================
using GemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTilePartitioner<GemmShape>;
using GemmEpilogue = ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, false, kPadC>>;
using Traits = ck_tile::TileGemmTraits<kPadA, kPadB, kPadC, ALayout, BLayout, CLayout>;
using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(args.K);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
float ave_time{0};
const auto Run = [&](const auto has_hot_loop_, const auto tail_number_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<
ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
Traits,
ck_tile::GemmPipelineScheduler::Intrawave,
has_hot_loop_v,
tail_number_v>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKargs(args.p_a,
args.p_b,
args.p_c,
args.M,
args.N,
args.K,
args.stride_A,
args.stride_B,
args.stride_C);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.kbatch);
constexpr dim3 blocks = Kernel::BlockSize();
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};
if(has_hot_loop)
{
// Tail pipeline One to Seven
if(tail_num == ck_tile::TailNumber::One)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::One>{});
}
else if(tail_num == ck_tile::TailNumber::Full)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
}
if constexpr(BaseGemmPipeline::PrefetchStages > 2)
{
if(tail_num == ck_tile::TailNumber::Two)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Two>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 3)
{
if(tail_num == ck_tile::TailNumber::Three)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Three>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 4)
{
if(tail_num == ck_tile::TailNumber::Four)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Four>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 5)
{
if(tail_num == ck_tile::TailNumber::Five)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Five>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 6)
{
if(tail_num == ck_tile::TailNumber::Six)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Six>{});
}
}
if constexpr(BaseGemmPipeline::PrefetchStages > 7)
{
if(tail_num == ck_tile::TailNumber::Seven)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Seven>{});
}
}
}
else
{
// Tail number always Full - #PrefetchStages
if(tail_num == ck_tile::TailNumber::Full)
{
Run(ck_tile::bool_constant<false>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
}
else
{
std::ostringstream err;
err << "When there's no hot loop, this tail number \"" << tail_num
<< "\" is not supported! " << __FILE__ << ":" << __LINE__
<< ", in function: " << __func__;
throw std::runtime_error(err.str());
}
}
return ave_time;
}
#include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/ck_tile/03_gemm/run_gemm_example.inc 0 → 100644
View file @ 7c56cd01
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <typename ALayout, typename BLayout, typename CLayout>
float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
ck_tile::DeviceMem& b_k_n_dev_buf,
ck_tile::DeviceMem& c_m_n_dev_buf,
ck_tile::index_t M,
ck_tile::index_t N,
ck_tile::index_t K,
ck_tile::index_t stride_A,
ck_tile::index_t stride_B,
ck_tile::index_t stride_C,
ck_tile::index_t kbatch,
int n_warmup,
int n_repeat)
{
gemm_basic_args args;
args.p_a = a_m_k_dev_buf.GetDeviceBuffer();
args.p_b = b_k_n_dev_buf.GetDeviceBuffer();
args.p_c = c_m_n_dev_buf.GetDeviceBuffer();
args.kbatch = kbatch;
args.M = M;
args.N = N;
args.K = K;
args.stride_A = stride_A;
args.stride_B = stride_B;
args.stride_C = stride_C;
float ave_time = gemm_calc<ALayout, BLayout, CLayout>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::string op_name{"Gemm{MemBoundPipeline}"};
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_byte =
sizeof(ADataType) * M * K + sizeof(BDataType) * N * K + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Run " << op_name << "kernel with M =" << M << " N =" << N << " K =" << K
<< " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideC =" << stride_C
<< " : " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< std::endl;
return ave_time;
}
template <typename ALayout, typename BLayout, typename CLayout>
int run_gemm_example_with_layouts(int argc,
char* argv[],
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
ck_tile::index_t K = arg_parser.get_int("k");
ck_tile::index_t stride_A = arg_parser.get_int("stride_a");
ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
ck_tile::index_t batch_size = arg_parser.get_int("b");
int n_warmup = arg_parser.get_int("warmup");
int n_repeat = arg_parser.get_int("repeat");
using namespace ck_tile::literals;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>)
{
return ck_tile::HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return ck_tile::HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
auto f_get_default_stride = [](std::size_t row,
std::size_t col,
std::size_t stride,
auto layout) {
if(stride == 0)
{
// give a chance if stride is zero, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>)
{
return col;
}
else
{
return row;
}
}
else
return stride;
};
stride_A = f_get_default_stride(M, K, stride_A, a_layout);
stride_B = f_get_default_stride(K, N, stride_B, b_layout);
stride_C = f_get_default_stride(M, N, stride_C, CLayout{});
ck_tile::HostTensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, stride_A, a_layout));
ck_tile::HostTensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, stride_B, b_layout));
ck_tile::HostTensor<CDataType> c_m_n_dev_result(
f_host_tensor_descriptor(M, N, stride_C, CLayout{}));
// TODO: add different init types
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k);
ck_tile::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_k_n);
ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());
a_m_k_dev_buf.ToDevice(a_m_k.data());
b_k_n_dev_buf.ToDevice(b_k_n.data());
c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero();
invoke_gemm<ALayout, BLayout, CLayout>(a_m_k_dev_buf,
b_k_n_dev_buf,
c_m_n_dev_buf,
M,
N,
K,
stride_A,
stride_B,
stride_C,
batch_size,
n_warmup,
n_repeat);
c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
bool pass = true;
if(arg_parser.get_int("v") == 1)
{
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
f_host_tensor_descriptor(M, N, stride_C, CLayout{}));
c_m_n_host_ref.SetZero();
ck_tile::reference_gemm<ADataType, BDataType, AccDataType, CDataType>(
a_m_k, b_k_n, c_m_n_host_ref);
pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_host_ref);
std::cout << "The CPU veification result is:" << (pass ? "correct" : "fail") << std::endl;
}
else if(arg_parser.get_int("v") == 2)
{
ck_tile::HostTensor<CDataType> c_m_n_gpu_ref(
f_host_tensor_descriptor(M, N, stride_C, CLayout{}));
ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_gpu_ref.get_element_space_size_in_bytes());
c_m_n_gpu_ref.SetZero();
c_m_n_gpu_buf_ref.SetZero();
ck_tile::reference_gemm_gpu<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(
a_m_k_dev_buf, b_k_n_dev_buf, c_m_n_gpu_buf_ref, M, N, K, stride_A, stride_B, stride_C);
c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data());
pass = ck_tile::check_err(c_m_n_dev_result, c_m_n_gpu_ref);
std::cout << "The GPU veification result is: " << (pass ? "correct" : "fail") << std::endl;
}
return pass;
}
int run_gemm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "R")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
}
else if(a_layout == "R" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}
else if(a_layout == "C" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Col{}, Col{}, Row{});
}
else if(a_layout == "C" && b_layout == "R")
{
return run_gemm_example_with_layouts(argc, argv, Col{}, Row{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
}
example/ck_tile/05_reduce/CMakeLists.txt 0 → 100644
View file @ 7c56cd01
set(EXAMPLE_REDUCE "tile_example_reduce")
# 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 example ${EXAMPLE_REDUCE}")
add_executable(${EXAMPLE_REDUCE} EXCLUDE_FROM_ALL reduce.cpp)
target_include_directories(${EXAMPLE_REDUCE} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
set(EXAMPLE_REDUCE_COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND EXAMPLE_REDUCE_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
target_compile_options(${EXAMPLE_REDUCE} PRIVATE ${EXAMPLE_REDUCE_COMPILE_OPTIONS})
# TODO: we have to turn off this global prop, otherwise the progress bar generated
# by cmake will print too many files, execvp: /bin/sh: Argument list too long
# however, this property may affect global
# TODO: consider codegen a makefile by us
set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)
\ No newline at end of file
example/ck_tile/05_reduce/reduce.cpp 0 → 100644
View file @ 7c56cd01
#include "ck_tile/host.hpp"
#include "reduce.hpp"
#include <cstring>
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension")
.insert("v", "1", "cpu validation or not")
.insert("prec", "fp16", "precision")
.insert("warmup", "5", "cold iter")
.insert("repeat", "20", "hot iter");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
template <typename DataType>
bool run(const ck_tile::ArgParser& arg_parser)
{
using XDataType = DataType;
using ComputeDataType = float;
using YDataType = DataType;
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
ck_tile::HostTensor<XDataType> x_host({m, n});
ck_tile::HostTensor<YDataType> y_host_ref({m});
ck_tile::HostTensor<YDataType> y_host_dev({m});
ck_tile::FillUniformDistribution<XDataType>{-5.f, 5.f}(x_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem y_buf(y_host_dev.get_element_space_size_in_bytes());
x_buf.ToDevice(x_host.data());
using ReduceOp = ck_tile::ReduceOp::Add;
using BlockWarps = ck_tile::sequence<4, 1>;
using BlockTile = ck_tile::sequence<128, 128>;
using WarpTile = ck_tile::sequence<32, 128>;
using Vector = ck_tile::sequence<8, 8>;
// cross warp-reduce
// using BlockWarps = ck_tile::sequence<2, 2>;
// using BlockTile = ck_tile::sequence<2, 1024>;
// using WarpTile = ck_tile::sequence<1, 512>;
// using Vector = ck_tile::sequence<1, 8>;
constexpr ck_tile::index_t kBlockSize = 512;
constexpr ck_tile::index_t kBlockPerCu = 1;
ck_tile::index_t kGridSize = (m / BlockTile::at(ck_tile::number<0>{}));
std::cout << "grid size " << kGridSize << std::endl;
using Shape = ck_tile::Reduce2dShape<BlockWarps, BlockTile, WarpTile, Vector>;
using Porblem =
ck_tile::Reduce2dProblem<XDataType, ComputeDataType, YDataType, Shape, ReduceOp>;
using Kernel = ck_tile::Reduce<Porblem>;
float ave_time = launch_kernel(ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockSize, kBlockPerCu>(
Kernel{},
kGridSize,
kBlockSize,
0,
static_cast<XDataType*>(x_buf.GetDeviceBuffer()),
static_cast<YDataType*>(y_buf.GetDeviceBuffer()),
m,
n));
std::size_t num_btype = sizeof(XDataType) * m * n + sizeof(YDataType) * m;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
bool pass = true;
if(do_validation)
{
// reference
ck_tile::reference_reduce<XDataType, ComputeDataType, YDataType>(
x_host, y_host_ref, ReduceOp{});
y_buf.FromDevice(y_host_dev.mData.data());
pass = ck_tile::check_err(y_host_dev, y_host_ref);
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;
// }
}
example/ck_tile/05_reduce/reduce.hpp 0 → 100644
View file @ 7c56cd01
This diff is collapsed.
example/ck_tile/06_permute/CMakeLists.txt 0 → 100644
View file @ 7c56cd01
# 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"
add_executable(tile_example_permute EXCLUDE_FROM_ALL permute.cpp)
if(NOT DEFINED PERMUTE_USE_ALTERNATIVE_IMPL)
# set(PERMUTE_USE_ALTERNATIVE_IMPL false)
set(PERMUTE_USE_ALTERNATIVE_IMPL true)
endif()
if(PERMUTE_USE_ALTERNATIVE_IMPL)
target_compile_options(tile_example_permute PRIVATE -DPERMUTE_USE_ALTERNATIVE_IMPL)
target_sources(tile_example_permute PRIVATE alternative_impl/matrix_core_swizzle.cpp)
endif()
# target_compile_options(tile_example_permute PRIVATE -v --save-temps -Wno-gnu-line-marker)
example/ck_tile/06_permute/README.md 0 → 100644
View file @ 7c56cd01
# permute
This folder contains example for permute kernel, which is similiar to [torch.permute](https://pytorch.org/docs/stable/generated/torch.permute.html) (combined with [torch.contiguous](https://pytorch.org/docs/stable/generated/torch.Tensor.contiguous.html)). Currently we implement a generic permute kernel that support up to rank 8 arbitrary permutation with a single kernel instance. Performance is not the first consideration, we prefer a simple and general kernel implementation using `ck_tile` in this example.
```
args:
-v weather do CPU validation or not (default:1)
-prec data type. fp16/bf16/fp32 (default:fp16)
-shape the shape of the input tensor (default:2,3,4)
-perm permute perm (default:2,1,0)
```
## 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_permute -j
```
This will result in an executable `build/bin/tile_example_permute`
## some examples
```
# torch
x=torch.randn(2,3,4,6)
y=x.permute(0,3,2,1).contiguous()
# ck_tile
./build/bin/tile_example_permute -shape=2,3,4,6 -perm=0,3,2,1
```
or you can try the smoke_test
```
# in the root of ck_tile, after you build this example
sh example/ck_tile/06_permute/script/smoke_test.sh
```
### alternative implementation
we have an alternative implementation under `alternative_impl/` folder, that can swizzle the tensor to be more friendly for data loading for matrix core layout. This can be enabled when dealing with a `rank-7` tensor, with a fixed pattern of either `0,1,4,2,5,3,6` or `0,1,2,4,5,3,6`. There are other shape limitation of this implementation, check the source code of `permute.cpp` for detail.
```
# example
./build/bin/tile_example_permute -shape=3,6,4,32,16,2,8 -perm=0,1,4,2,5,3,6 # b_n0_k0_n1_k1_n2_k2
./build/bin/tile_example_permute -shape=3,8,4,16,16,4,8 -perm=0,1,2,4,5,3,6 # b_n0_n1_k0_k1_n2_k2
```
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment