ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-18 20:09:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'develop' into hstu_attention_mi350_fwd_bwd and change in using ck_tile::make_kernel

Browse Source
This commit is contained in:
Qianfeng Zhang
2025-09-01 07:13:16 +00:00
parent d4f43f0653 fcff0043ae
commit 4bf65d9fe5
266 changed files with 9617 additions and 3260 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
example/ck_tile/01_fmha/README.md
View File

@@ -7,7 +7,7 @@ This folder contains example for fmha(fused multi-head attention) using ck_tile
# 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>
../script/cmake-ck-dev.sh ../ <arch>
make tile_example_fmha_fwd -j
```
This will result in an executable `build/bin/tile_example_fmha_fwd`

4
example/ck_tile/01_fmha/codegen/ops/fmha_batch_prefill.py
View File

@@ -110,9 +110,9 @@ float fmha_batch_prefill_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_b
if(s.log_level_ > 0)
std::cout << ", " << k_::GetName() << std::flush;
auto [kargs, grids] = fmha_batch_prefill_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
return ck_tile::launch_kernel(s, ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
}}
"""

24
example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py
View File

@@ -136,10 +136,10 @@ float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_{F_idx}>(const ck_tile::stream_config& s
if(s.log_level_ > 0)
std::cout << ", " << k_::GetName() << std::flush;
auto [kargs, grids] = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
}}
template <>
@@ -148,9 +148,9 @@ void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_{F_idx}>(const ck_tile::stream_co
{{
using k_ = fmha_bwd_dq_dk_dv_kernel_{F_idx};
auto [kargs, grids] = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(
ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(
ck_tile::stream_config{{s.stream_id_}});
}}
@@ -425,10 +425,10 @@ float fmha_bwd_dot_do_o_<dot_do_o_trait_{F_idx}>(const ck_tile::stream_config& s
if(s.log_level_ > 0)
std::cout << ", " << k_::GetName() << std::flush;
auto [kargs, grids] = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
}}
template <>
@@ -436,9 +436,9 @@ void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_{F_idx}>(const ck_tile::stream_co
{{
using k_ = fmha_bwd_dot_do_o_kernel_{F_idx};
auto [kargs, grids] = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(
ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(
ck_tile::stream_config{{s.stream_id_}});
}}
@@ -530,10 +530,10 @@ float fmha_bwd_convert_dq_<convert_dq_trait_{F_idx}>(const ck_tile::stream_confi
if(s.log_level_ > 0)
std::cout << ", " << k_::GetName() << std::flush;
auto [kargs, grids] = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
}}
template <>
@@ -542,9 +542,9 @@ void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_{F_idx}>(const ck_tile::strea
{{
using k_ = fmha_bwd_convert_dq_kernel_{F_idx};
auto [kargs, grids] = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(
ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(
ck_tile::stream_config{{s.stream_id_}});
}}

6
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View File

@@ -110,9 +110,9 @@ float fmha_fwd_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_args a)
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();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
return ck_tile::launch_kernel(s, ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
}}
"""
@@ -385,7 +385,7 @@ class FmhaFwdApiPool:
for i, dtype in enumerate(self.pool.keys()):
per_hdim_case=str()
for j, (hdim, hdim_v) in enumerate(self.pool[dtype].keys()):
traits=self.pool[dtype][(hdim, hdim_v)]
traits=[t for t in self.pool[dtype][(hdim, hdim_v)] if tr_load == t.tr_load]
inners=str()
for k, trait in enumerate(traits):
if_k = 'if' if k == 0 else 'else if'

4
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_appendkv.py
View File

@@ -60,9 +60,9 @@ float fmha_fwd_appendkv_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fw
if(s.log_level_ > 0)
std::cout << ", " << k_::GetName() << std::flush;
auto [kargs, grids] = fmha_fwd_appendkv_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
return ck_tile::launch_kernel(s, ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
}}
"""

8
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py
View File

@@ -108,9 +108,9 @@ static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
{{
using k_ = fmha_kernel;
auto [kargs, grids] = fmha_fwd_splitkv_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
}}
}};
}}
@@ -208,9 +208,9 @@ static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
{{
using k_ = fmha_kernel;
auto [kargs, grids] = fmha_fwd_splitkv_combine_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
}}
}};
}}

4
example/ck_tile/01_fmha/codegen/ops/fmha_pagedkv_prefill.py
View File

@@ -109,9 +109,9 @@ float fmha_fwd_pagedkv_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd
if(s.log_level_ > 0)
std::cout << ", " << k_::GetName() << std::flush;
auto [kargs, grids] = fmha_fwd_pagedkv_create_kargs_and_grids<k_>(a);
constexpr dim3 blocks = k_::BlockSize();
const dim3 blocks = k_::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
return ck_tile::launch_kernel(s, ck_tile::make_kernel<kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
}}
"""

14
example/ck_tile/01_fmha/fmha_bwd.cpp
View File

@@ -809,20 +809,6 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::stream_config stream_config_v{
nullptr, true, 0, 0, 1, arg_parser.get_str("timer") == std::string("gpu")};
printf("\nfmha_bwd_traits: hdim_q=%d, hdim_v=%d, data_type=%s, is_group_mode=%d, mask_type=%d, "
"bias_type=%d, has_dbias=%d, has_dropout=%d, is_store_randval=%d, is_deterministic=%d\n",
fmha_traits.hdim_q,
fmha_traits.hdim_v,
fmha_traits.data_type.c_str(),
fmha_traits.is_group_mode,
static_cast<int>(fmha_traits.mask_type),
static_cast<int>(fmha_traits.bias_type),
fmha_traits.has_dbias,
fmha_traits.has_dropout,
fmha_traits.is_store_randval,
fmha_traits.is_deterministic);
fflush(stdout);
fmha_bwd(fmha_traits, fmha_args, stream_config_v);
dq_buf.FromDevice(dq_host.data());

2
example/ck_tile/01_fmha/script/run_full_test.sh
View File

@@ -9,6 +9,8 @@
# host name : $hostname
# gpu architecture: e.g., gfx90a, or gfx942, etc.
set -euo pipefail
#get the command line arguments:
export env_type=$1
echo 'Environment type: ' $env_type

16
example/ck_tile/01_fmha/script/smoke_test_bwd.sh
View File

@@ -1,5 +1,7 @@
#!/bin/sh
#!/bin/bash
# TODO: run this script from CK root or build directory
set -euo pipefail
EXE="$(find . -name tile_example_fmha_bwd -type f | head -n 1)"
KNAME=1
@@ -17,12 +19,12 @@ for dbias in 0 ; do
for p_drop in 0.0 0.2 ; do
for deterministic in 0 ; do
$EXE -prec=$prec -b=1 -h=4 -h_k=2 -d=$hdim -s=259 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=2 -d=$hdim -s=516 -s_k=253 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=4 -h_k=1 -d=$hdim -s=500 -s_k=251 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=1 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=2 -d=$hdim -s=900 -s_k=258 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=2 -v=1 -deterministic=$deterministic -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=1 -d=$hdim -s=987 -s_k=219 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=t:128,30 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=3 -h_k=1 -d=$hdim -s=244 -s_k=499 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=b:4,35 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=4 -h_k=2 -d=$hdim -s=259 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=2 -d=$hdim -s=516 -s_k=253 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=4 -h_k=1 -d=$hdim -s=500 -s_k=251 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=1 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=1 -h=2 -d=$hdim -s=900 -s_k=258 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=2 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=1 -d=$hdim -s=987 -s_k=219 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=t:128,30 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -b=2 -h=3 -h_k=1 -d=$hdim -s=244 -s_k=499 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=b:4,35 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
done
done

21
example/ck_tile/01_fmha/script/smoke_test_fwd.sh
View File

@@ -1,5 +1,7 @@
#!/bin/bash
# TODO: run this script from CK root or build directory
set -euo pipefail
EXE="$(find . -name tile_example_fmha_fwd -type f | head -n 1)"
KNAME=1
@@ -51,19 +53,18 @@ run_fp16_bf16_tests() {
for cache_batch_idx in $CACHE_BATCH_IDX ; do
# $EXE -prec=$prec -mode=$mode -b=1 -h=1 -d=$hdim -s=1024 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -num_splits=$num_splits -page_block_size=$page_block_size -kname=$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=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=3 -d=$hdim -s=100 -s_k=51 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=16 -d_v=$hdim -s=99 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=1 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1024 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -d_v=24 -s=3 -s_k=99 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=3 -h=2 -h_k=1 -d=$hdim -s=200 -s_k=520 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=t:128,30 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -s=99 -s_k=32 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=b:4,35 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=33 -s_k=0 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1 -s_k=10 -s_kpad=32 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -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=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=3 -d=$hdim -s=100 -s_k=51 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=16 -d_v=$hdim -s=99 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=1 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1024 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -d_v=24 -s=3 -s_k=99 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=3 -h=2 -h_k=1 -d=$hdim -s=200 -s_k=520 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=t:128,30 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -s=99 -s_k=32 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=b:4,35 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=33 -s_k=0 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1 -s_k=10 -s_kpad=32 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
done ; done ; done ; done ; done
done ; done ; done ; done ; done
done ;
}
run_fp8_tests() {

2
example/ck_tile/02_layernorm2d/README.md
View File

@@ -42,7 +42,7 @@ return hidden_states, per_token_scale
```
# 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...
../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`

4
example/ck_tile/02_layernorm2d/generate.py
View File

@@ -235,7 +235,7 @@ float layernorm2d_fwd_(const S& s, A a)
using Kernel = ck_tile::Layernorm2dFwd<Pipeline, Epilogue>;
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto kargs = Kernel::MakeKargs(a);
@@ -243,7 +243,7 @@ float layernorm2d_fwd_(const S& s, A a)
std::cout << ", " << Kernel::GetName() << std::flush;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{{}}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{{}}, grids, blocks, 0, kargs));
}}
"""

2
example/ck_tile/03_gemm/README.md
View File

@@ -7,7 +7,7 @@ This folder contains example for GEMM using ck_tile tile-programming implementat
# 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>
../script/cmake-ck-dev.sh ../ <arch>
# The basic pipeline method on the gemm calculation
make tile_example_gemm_basic -j
# The memory bound pipeline on the gemm calculation

21
example/ck_tile/03_gemm/gemm_basic.cpp
View File

@@ -26,6 +26,15 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
constexpr ck_tile::index_t N_Tile = 256;
constexpr ck_tile::index_t K_Tile = 64;
#if CK_TILE_USE_WMMA
constexpr ck_tile::index_t M_Warp = 4;
constexpr ck_tile::index_t N_Warp = 2;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp_Tile = 16;
constexpr ck_tile::index_t N_Warp_Tile = 16;
constexpr ck_tile::index_t K_Warp_Tile = 16;
#else
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
constexpr ck_tile::index_t K_Warp = 1;
@@ -33,6 +42,7 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
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 = 16;
#endif
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
@@ -65,7 +75,6 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
ck_tile::tuple<>,
CLayout,
ck_tile::element_wise::PassThrough,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
@@ -81,8 +90,8 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
using Kernel = ck_tile::GemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -100,10 +109,8 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
<< std::endl;
}
float ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};

17
example/ck_tile/03_gemm/gemm_splitk_two_stage_reduce.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2025 Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
@@ -208,7 +208,6 @@ float gemm_stage1(const GemmSplitKHostArgs& args, const ck_tile::stream_config&
DsLayout,
ELayout,
CDEElementWise,
UniversalGemmProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
@@ -232,7 +231,7 @@ float gemm_stage1(const GemmSplitKHostArgs& args, const ck_tile::stream_config&
{
grids = Kernel::GridSize(args.M, args.N, args.k_batch);
}
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -279,15 +278,13 @@ float gemm_stage1(const GemmSplitKHostArgs& args, const ck_tile::stream_config&
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
@@ -373,7 +370,7 @@ float reduce_stage2(const GemmSplitKHostArgs& args, const ck_tile::stream_config
float ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<kBlockSize, kBlockPerCu>(
ck_tile::make_kernel<kBlockPerCu>(
Kernel{},
kGridSize,
kBlockSize,

4
example/ck_tile/03_gemm/gemm_utils.hpp Executable file → Normal file
View File

@@ -172,6 +172,7 @@ struct GemmConfigComputeV3_2 : public GemmConfigBase
static constexpr int kBlockPerCu = 2;
};
#if CK_TILE_USE_WMMA
template <typename PrecType>
struct GemmConfigComputeV3_WMMA : public GemmConfigBase
{
@@ -192,6 +193,7 @@ struct GemmConfigComputeV3_WMMA : public GemmConfigBase
static constexpr int kBlockPerCu = 2;
};
#endif
template <typename PrecType>
struct GemmConfigComputeV4 : public GemmConfigBase
@@ -484,7 +486,7 @@ auto create_args(int argc, char* argv[])
.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("prec", "fp16", "data type. fp16/bf16/fp8/bf8/pk_int4_t")
.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")

13
example/ck_tile/03_gemm/gemm_weight_preshuffle.cpp
View File

@@ -103,7 +103,6 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
DsLayout,
ELayout,
CDEElementWise,
UniversalGemmProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
@@ -126,7 +125,7 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{
grids = Kernel::GridSize(args.M, args.N, args.k_batch);
}
constexpr dim3 blocks = Kernel::BlockSize();
dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -172,15 +171,13 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};

17
example/ck_tile/03_gemm/universal_gemm.cpp
View File

@@ -103,7 +103,6 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
DsLayout,
ELayout,
CDEElementWise,
UniversalGemmProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
@@ -127,7 +126,7 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{
grids = Kernel::GridSize(args.M, args.N, args.k_batch);
}
constexpr dim3 blocks = Kernel::BlockSize();
dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -173,15 +172,13 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
@@ -338,7 +335,11 @@ int main(int argc, char* argv[])
try
{
#if CK_TILE_USE_WMMA
return !run_gemm_example<GemmConfigComputeV3_WMMA>(arg_parser);
#else
return !run_gemm_example<GemmConfigComputeV3>(arg_parser);
#endif
}
catch(const std::runtime_error& e)
{

2
example/ck_tile/04_img2col/README.md
View File

@@ -7,7 +7,7 @@ This folder contains example for Image to Column using ck_tile tile-programming
# 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>
../script/cmake-ck-dev.sh ../ <arch>
make tile_example_img2col -j
```
This will result in an executable `build/bin/tile_example_img2col`

5
example/ck_tile/04_img2col/image_to_column.cpp
View File

@@ -55,13 +55,12 @@ float image_to_column(const image_to_column_traits& traits,
args.N * args.output_spatial_lengths[0] * args.output_spatial_lengths[1],
args.filter_spatial_lengths[0] * args.filter_spatial_lengths[1] * args.C,
args.G);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 2;
float ave_time = ck_tile::launch_kernel(
stream_conf,
ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
stream_conf, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}

24
example/ck_tile/05_reduce/reduce.cpp
View File

@@ -94,18 +94,18 @@ bool run(const ck_tile::ArgParser& arg_parser)
throw std::runtime_error("Wrong! Arguments not supported!\n");
}
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()),
input_shape,
input_strides,
kept_dim,
reduce_dims));
float ave_time = launch_kernel(
ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockPerCu>(Kernel{},
kGridSize,
kBlockSize,
0,
static_cast<XDataType*>(x_buf.GetDeviceBuffer()),
static_cast<YDataType*>(y_buf.GetDeviceBuffer()),
input_shape,
input_strides,
kept_dim,
reduce_dims));
std::size_t num_btype = sizeof(XDataType) * N * C * H * W + sizeof(YDataType) * N * C;

2
example/ck_tile/06_permute/README.md
View File

@@ -15,7 +15,7 @@ args:
```
# 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...
../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`

5
example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle_kernel.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -115,11 +115,12 @@ struct matrix_core_swizzle_kernel
__host__ void operator()(const ck_tile::stream_config& s) const
{
ck_tile::kentry<BLOCK_SIZE, 1, kernel><<<grids, BLOCK_SIZE, 0, s.stream_id_>>>(a);
ck_tile::kentry<1, kernel><<<grids, BLOCK_SIZE, 0, s.stream_id_>>>(a);
}
struct kernel
{
static constexpr int kBlockSize = BLOCK_SIZE;
__device__ static constexpr auto get_src_dist()
{
using namespace ck_tile;

24
example/ck_tile/06_permute/permute.cpp
View File

@@ -53,11 +53,11 @@ float permute(permute_traits t, permute_args a, const ck_tile::stream_config& s)
auto kargs = Kernel::MakeKargs(a);
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(a);
const dim3 blocks = Kernel::BlockSize();
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, 1>(Kernel{}, grids, blocks, 0, kargs));
float ave_time =
ck_tile::launch_kernel(s, ck_tile::make_kernel<1>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}
@@ -69,11 +69,11 @@ float permute(permute_traits t, permute_args a, const ck_tile::stream_config& s)
auto kargs = Kernel::MakeKargs(a);
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(a);
const dim3 blocks = Kernel::BlockSize();
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, 1>(Kernel{}, grids, blocks, 0, kargs));
float ave_time =
ck_tile::launch_kernel(s, ck_tile::make_kernel<1>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}
@@ -85,11 +85,11 @@ float permute(permute_traits t, permute_args a, const ck_tile::stream_config& s)
auto kargs = Kernel::MakeKargs(a);
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(a);
const dim3 blocks = Kernel::BlockSize();
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, 1>(Kernel{}, grids, blocks, 0, kargs));
float ave_time =
ck_tile::launch_kernel(s, ck_tile::make_kernel<1>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
}

2
example/ck_tile/09_topk_softmax/README.md
View File

@@ -6,7 +6,7 @@ This folder contains example for topk-softmax kernel using ck_tile tile-programm
```
# 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...
../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_example_topk_softmax -j
```
This will result in an executable `build/bin/tile_example_topk_softmax`

8
example/ck_tile/09_topk_softmax/topk_softmax_api.cpp
View File

@@ -13,11 +13,11 @@
\
auto kargs = kernel::MakeKargs(a); \
\
const dim3 grids = kernel::GridSize(a); \
constexpr dim3 blocks = kernel::BlockSize(); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(); \
\
float ave_time = ck_tile::launch_kernel( \
s, ck_tile::make_kernel<blocks.x, 1>(kernel{}, grids, blocks, 0, kargs)); \
float ave_time = \
ck_tile::launch_kernel(s, ck_tile::make_kernel<1>(kernel{}, grids, blocks, 0, kargs)); \
\
return ave_time;

2
example/ck_tile/10_rmsnorm2d/README.md
View File

@@ -6,7 +6,7 @@ This folder contains example for Rmsnorm2D forward using ck_tile tile-programmin
```
# 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...
../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_rmsnorm2d_fwd -j
```
This will result in an executable `build/bin/tile_rmsnorm2d_fwd`

5
example/ck_tile/10_rmsnorm2d/example_rmsnorm2d_fwd.cpp
View File

@@ -138,12 +138,11 @@ bool run(const ck_tile::ArgParser& arg_parser)
auto kargs = Kernel::MakeKargs(args);
const dim3 grids = Kernel::GridSize(args);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto s = ck_tile::stream_config{nullptr, true, 0, warmup, repeat};
ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
ck_tile::launch_kernel(s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
bool pass = true;

4
example/ck_tile/10_rmsnorm2d/generate.py
View File

@@ -249,7 +249,7 @@ float rmsnorm2d_fwd_(const S& s, A a)
using Kernel = ck_tile::Rmsnorm2dFwd<Pipeline, Epilogue>;
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto kargs = Kernel::MakeKargs(a);
@@ -257,7 +257,7 @@ float rmsnorm2d_fwd_(const S& s, A a)
std::cout << ", " << Kernel::GetName() << std::flush;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{{}}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{{}}, grids, blocks, 0, kargs));
}}
"""

2
example/ck_tile/11_add_rmsnorm2d_rdquant/README.md
View File

@@ -6,7 +6,7 @@ This folder contains example for add + Rmsnorm2D + rowwise dynamic quantization
```
# 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...
../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_add_rmsnorm2d_rdquant_fwd -j
```
This will result in an executable `build/bin/tile_add_rmsnorm2d_rdquant_fwd`

5
example/ck_tile/11_add_rmsnorm2d_rdquant/example_add_rmsnorm2d_rdquant_fwd.cpp
View File

@@ -136,12 +136,11 @@ bool run(const ck_tile::ArgParser& arg_parser)
auto kargs = Kernel::MakeKargs(args);
const dim3 grids = Kernel::GridSize(args);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto s = ck_tile::stream_config{nullptr, true, 0, warmup, repeat};
ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
ck_tile::launch_kernel(s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
bool pass = true;

4
example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_instance_common.hpp
View File

@@ -58,7 +58,7 @@ float add_rmsnorm2d_rdquant_fwd_(const S& s, A a)
using Kernel = ck_tile::AddRmsnorm2dRdquantFwd<Pipeline>;
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto kargs = Kernel::MakeKargs(a);
@@ -66,5 +66,5 @@ float add_rmsnorm2d_rdquant_fwd_(const S& s, A a)
std::cout << ", " << Kernel::GetName() << std::flush;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}

2
example/ck_tile/12_smoothquant/README.md
View File

@@ -6,7 +6,7 @@ This folder contains example for smoothquant using ck_tile tile-programming impl
```
# 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...
../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_smoothquant -j
```
This will result in an executable `build/bin/tile_smoothquant`

5
example/ck_tile/12_smoothquant/example_smoothquant.cpp
View File

@@ -126,12 +126,11 @@ bool run(const ck_tile::ArgParser& arg_parser)
auto kargs = Kernel::MakeKargs(args);
const dim3 grids = Kernel::GridSize(args);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto s = ck_tile::stream_config{nullptr, true, 1, warmup, repeat};
ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
ck_tile::launch_kernel(s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
bool pass = true;

4
example/ck_tile/12_smoothquant/instances/smoothquant_instance_common.hpp
View File

@@ -50,7 +50,7 @@ float smoothquant_(const S& s, A a)
using Kernel = ck_tile::Smoothquant<Pipeline>;
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto kargs = Kernel::MakeKargs(a);
@@ -58,5 +58,5 @@ float smoothquant_(const S& s, A a)
std::cout << ", " << Kernel::GetName() << std::flush;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}

2
example/ck_tile/13_moe_sorting/README.md
View File

@@ -6,7 +6,7 @@ This folder contains example for moe-sorting kernel using ck_tile tile-programmi
```
# 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...
../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_example_moe_sorting -j
```
This will result in an executable `build/bin/tile_example_moe_sorting`

24
example/ck_tile/13_moe_sorting/moe_sorting_api.cpp
View File

@@ -209,7 +209,7 @@ float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_confi
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
return ck_tile::make_kernel<kernel::BLOCK_SIZE>(kernel{}, grids, blocks, 0, kargs); \
return ck_tile::make_kernel(kernel{}, grids, blocks, 0, kargs); \
}()
#define MOE_SORTING_MP_1(mesh_type_, unroll_num_, expert_masking_, local_token_) \
@@ -227,7 +227,7 @@ float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_confi
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
return ck_tile::make_kernel<kernel::BLOCK_SIZE>(kernel{}, grids, blocks, 0, kargs); \
return ck_tile::make_kernel(kernel{}, grids, blocks, 0, kargs); \
}()
#if MOE_SORTING_SUPPORT_LARGE_EXPERT
#define MOE_SORTING_MP_2(mesh_type_, unroll_num_, expert_masking_, local_token_) \
@@ -283,7 +283,7 @@ float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_confi
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
const auto lds_size = kernel::GetSmemSize(a); \
return ck_tile::make_kernel<kernel::BLOCK_SIZE>(kernel{}, grids, blocks, lds_size, kargs); \
return ck_tile::make_kernel(kernel{}, grids, blocks, lds_size, kargs); \
}()
#define MOR_SORTING_MP_DISPATCH_(mesh_type_, token_vec_0_, token_vec_1_, token_vec_23_) \
@@ -334,15 +334,15 @@ float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_confi
} \
}
#define MOR_SORTING_CLEAR_WS_DISPATCH_(is_local_token_, block_size_, occu_) \
[&]() { \
using problem_ = \
ck_tile::MoeSortingClearWorkspaceProblem<is_local_token_, block_size_, occu_>; \
using kernel = ck_tile::MoeSortingClearWorkspaceKernel<problem_>; \
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
return ck_tile::make_kernel<kernel::BLOCK_SIZE>(kernel{}, grids, blocks, 0, kargs); \
#define MOR_SORTING_CLEAR_WS_DISPATCH_(is_local_token_, block_size_, occu_) \
[&]() { \
using problem_ = \
ck_tile::MoeSortingClearWorkspaceProblem<is_local_token_, block_size_, occu_>; \
using kernel = ck_tile::MoeSortingClearWorkspaceKernel<problem_>; \
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
return ck_tile::make_kernel(kernel{}, grids, blocks, 0, kargs); \
}()
float moe_sorting_mp(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s)

2
example/ck_tile/14_moe_smoothquant/README.md
View File

@@ -9,7 +9,7 @@ Unlike standard smoothquant op, the input scale is from different expert `[exper
```
# 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...
../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_example_moe_smoothquant -j
```
This will result in an executable `build/bin/tile_example_moe_smoothquant`

4
example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_instance_common.hpp
View File

@@ -53,7 +53,7 @@ float moe_smoothquant_(const S& s, A a)
using Kernel = ck_tile::MoeSmoothquant<Pipeline>;
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto kargs = Kernel::MakeKargs(a);
@@ -61,5 +61,5 @@ float moe_smoothquant_(const S& s, A a)
std::cout << ", " << Kernel::GetName() << std::flush;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}

4
example/ck_tile/15_fused_moe/instances/fused_moegemm_api_internal.hpp
View File

@@ -53,7 +53,7 @@ float fused_moegemm_(const ck_tile::stream_config& s, fused_moegemm_args a)
using f_kernel = ck_tile::FusedMoeGemmKernel<f_partitioner, f_pipeline, void>;
const dim3 grids = f_kernel::GridSize(a);
constexpr dim3 blocks = f_kernel::BlockSize();
const dim3 blocks = f_kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
static int printed = 0;
@@ -66,5 +66,5 @@ float fused_moegemm_(const ck_tile::stream_config& s, fused_moegemm_args a)
}
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(f_kernel{}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(f_kernel{}, grids, blocks, 0, kargs));
}

6
example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp
View File

@@ -213,7 +213,7 @@ float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_til
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
return ck_tile::make_kernel<kernel::BLOCK_SIZE>(kernel{}, grids, blocks, 0, kargs); \
return ck_tile::make_kernel(kernel{}, grids, blocks, 0, kargs); \
}()
#define MOE_SORTING_MP_1(mesh_type_, unroll_num_, expert_masking_, local_token_) \
@@ -231,7 +231,7 @@ float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_til
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
return ck_tile::make_kernel<kernel::BLOCK_SIZE>(kernel{}, grids, blocks, 0, kargs); \
return ck_tile::make_kernel(kernel{}, grids, blocks, 0, kargs); \
}()
#if MOE_SORTING_SUPPORT_LARGE_EXPERT
#define MOE_SORTING_MP_2(mesh_type_, unroll_num_, expert_masking_, local_token_) \
@@ -287,7 +287,7 @@ float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_til
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
const auto lds_size = kernel::GetSmemSize(a); \
return ck_tile::make_kernel<kernel::BLOCK_SIZE>(kernel{}, grids, blocks, lds_size, kargs); \
return ck_tile::make_kernel(kernel{}, grids, blocks, lds_size, kargs); \
}()
#define MOR_SORTING_MP_DISPATCH_(mesh_type_, token_vec_0_, token_vec_1_, token_vec_23_) \

2
example/ck_tile/16_batched_gemm/README.md
View File

@@ -7,7 +7,7 @@ This folder contains example for batched GEMM using ck_tile tile-programming imp
# 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>
../script/cmake-ck-dev.sh ../ <arch>
make tile_example_batched_gemm -j
```
This will result in an executable `build/bin/tile_example_batched_gemm`

7
example/ck_tile/16_batched_gemm/batched_gemm.cpp
View File

@@ -142,7 +142,6 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
DsLayout,
CLayout,
CDEElementWise,
GemmPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
@@ -156,8 +155,8 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch, args.batch_count);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch, args.batch_count);
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -176,7 +175,7 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
}
ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};

2
example/ck_tile/17_grouped_gemm/README.md
View File

@@ -148,7 +148,7 @@ All the necessary parameters are set, the tiling is computed, the GEMM pipeline
# 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>
../script/cmake-ck-dev.sh ../ <arch>
# The basic pipeline method on the gemm calculation
make tile_example_grouped_gemm -j
```

16
example/ck_tile/17_grouped_gemm/grouped_gemm.cpp
View File

@@ -29,10 +29,6 @@ float grouped_gemm_tileloop(const ck_tile::stream_config& s,
void* kargs_ptr,
bool splitk)
{
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr ck_tile::index_t TileParitionerGroupNum = 8;
constexpr ck_tile::index_t TileParitionerM01 = 4;
@@ -44,7 +40,6 @@ float grouped_gemm_tileloop(const ck_tile::stream_config& s,
using TilePartitioner = ck_tile::
GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using GemmUniversalTraits =
ck_tile::PersistentTileGemmUniversalTraits<GemmConfig::kPadM,
GemmConfig::kPadN,
@@ -53,8 +48,6 @@ float grouped_gemm_tileloop(const ck_tile::stream_config& s,
ALayout,
BLayout,
CLayout>;
using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
float ave_time{0};
@@ -82,7 +75,6 @@ float grouped_gemm_tileloop(const ck_tile::stream_config& s,
ck_tile::tuple<>,
CLayout,
ck_tile::element_wise::PassThrough,
GemmPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
@@ -92,9 +84,9 @@ float grouped_gemm_tileloop(const ck_tile::stream_config& s,
GemmConfig::K_Warp_Tile,
UniversalGemmProblem::TransposeC,
memory_operation>>;
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::MaxOccupancyGridSize(s);
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
const dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::MaxOccupancyGridSize(s);
if(s.log_level_ > 0)
{
@@ -105,7 +97,7 @@ float grouped_gemm_tileloop(const ck_tile::stream_config& s,
ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(
Kernel{},
grids,
blocks,

2
example/ck_tile/18_flatmm/README.md
View File

@@ -7,7 +7,7 @@ This folder contains example for FLATMM using ck_tile tile-programming implement
# 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>
../script/cmake-ck-dev.sh ../ <arch>
# The basic pipeline method on the flatmm calculation
make tile_example_flatmm_basic -j
```

15
example/ck_tile/18_flatmm/flatmm_basic.cpp
View File

@@ -101,7 +101,6 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_c
DsLayout,
ELayout,
CDEElementWise,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
FlatmmConfig::M_Warp,
@@ -119,8 +118,8 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_c
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -171,15 +170,13 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs<>& args, const ck_tile::stream_c
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<blocks.x, FlatmmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, FlatmmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<FlatmmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};

18
example/ck_tile/18_flatmm/run_flatmm_example.inc
View File

@@ -40,9 +40,11 @@ template <typename FlatmmConfig, typename T>
auto shuffle_b(const ck_tile::HostTensor<T>& t)
{
assert(t.get_lengths().size() == 2);
int n_ = t.get_lengths()[1];
int k_ = t.get_lengths()[0];
constexpr int divisor = FlatmmConfig::N_Warp_Tile == 32 ? 2 : 4;
int n_ = t.get_lengths()[1];
int k_ = t.get_lengths()[0];
int divisor = ck_tile::is_wave32() ? (FlatmmConfig::N_Warp_Tile == 32 ? 1 : 2)
: (FlatmmConfig::N_Warp_Tile == 32 ? 2 : 4);
ck_tile::HostTensor<T> t_view({n_ / FlatmmConfig::N_Warp_Tile,
FlatmmConfig::N_Warp_Tile,
k_ / FlatmmConfig::K_Warp_Tile,
@@ -213,6 +215,16 @@ int run_flatmm_example_with_layouts(int argc,
ck_tile::FillUniformDistribution<ADataType>{1.f, 1.f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_origin_host);
}
else if(init_method == 3)
{
ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_origin_host);
}
else if(init_method == 4)
{
ck_tile::FillUniformDistribution<ADataType>{1.f, 1.f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host);
}
else
{
a_host.SetZero();

6
example/ck_tile/18_hstu_attention/hstu_attention_batched_forward_dispatch.hpp
View File

@@ -135,9 +135,9 @@ struct batched_forward_causal_local_bias_dropout_dispatch
constexpr dim3 kBlockSize = HstuKernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = HstuKernel::kBlockPerCu;
(void)ck_tile::launch_kernel(ck_tile::stream_config{stream, false},
ck_tile::make_kernel<kBlockSize.x, kBlockPerCu>(
HstuKernel{}, kGridSize, kBlockSize, 0, kargs));
(void)ck_tile::launch_kernel(
ck_tile::stream_config{stream, false},
ck_tile::make_kernel<kBlockPerCu>(HstuKernel{}, kGridSize, kBlockSize, 0, kargs));
};
};

6
example/ck_tile/18_hstu_attention/hstu_attention_jagged_forward_dispatch.hpp
View File

@@ -126,9 +126,9 @@ struct jagged_forward_causal_local_bias_dropout_dispatch
constexpr dim3 kBlockSize = HstuKernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = HstuKernel::kBlockPerCu;
(void)ck_tile::launch_kernel(ck_tile::stream_config{stream, false},
ck_tile::make_kernel<kBlockSize.x, kBlockPerCu>(
HstuKernel{}, kGridSize, kBlockSize, 0, kargs));
(void)ck_tile::launch_kernel(
ck_tile::stream_config{stream, false},
ck_tile::make_kernel<kBlockPerCu>(HstuKernel{}, kGridSize, kBlockSize, 0, kargs));
};
};

2
example/ck_tile/19_gemm_multi_d/README.md
View File

@@ -8,7 +8,7 @@ This folder contains example for Multiple D GEMM using ck_tile tile-programming
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>
../script/cmake-ck-dev.sh ../ <arch>
#The basic pipeline method on the gemm calculation
make tile_example_gemm_multi_d_fp16 -j
```

7
example/ck_tile/19_gemm_multi_d/gemm_multi_d_fp16.cpp
View File

@@ -146,7 +146,6 @@ auto gemm_multi_d(const gemm_multi_d_kargs& args, const ck_tile::stream_config&
DsLayout,
CLayout,
CDEElementWise,
GemmPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
@@ -160,8 +159,8 @@ auto gemm_multi_d(const gemm_multi_d_kargs& args, const ck_tile::stream_config&
using Kernel = ck_tile::GemmKernelMultiD<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -176,7 +175,7 @@ auto gemm_multi_d(const gemm_multi_d_kargs& args, const ck_tile::stream_config&
}
ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};

3
example/ck_tile/20_grouped_convolution/CMakeLists.txt
View File

@@ -6,3 +6,6 @@ target_compile_options(tile_example_grouped_conv_fwd PRIVATE ${EXAMPLE_GEMM_COMP
add_executable(tile_example_grouped_conv_bwd_weight EXCLUDE_FROM_ALL grouped_convolution_backward_weight.cpp)
target_compile_options(tile_example_grouped_conv_bwd_weight PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
add_executable(tile_example_grouped_conv_bwd_data EXCLUDE_FROM_ALL grouped_convolution_backward_data.cpp)
target_compile_options(tile_example_grouped_conv_bwd_data PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})

215
example/ck_tile/20_grouped_convolution/grouped_convolution_backward_data.cpp Normal file
View File

@@ -0,0 +1,215 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
#include <cstring>
#include <iostream>
#include <ostream>
#include <string>
#include <tuple>
#include "ck_tile/host.hpp"
#include "grouped_convolution_utils.hpp"
template <ck_tile::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename AccDataType,
typename OutDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout,
typename DsDataType = ck_tile::tuple<>,
typename DsLayout = ck_tile::tuple<>,
typename CDEElementWise = ck_tile::element_wise::PassThrough>
float grouped_conv_bwd_data(const ck_tile::GroupedConvBwdDataHostArgs& args,
const ck_tile::stream_config& s)
{
constexpr int kBlockPerCu = 1;
constexpr ck_tile::index_t M_Tile = 64;
constexpr ck_tile::index_t N_Tile = 64;
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 = 16;
constexpr ck_tile::index_t VectorSizeA = 8;
constexpr ck_tile::index_t VectorSizeB = 8;
constexpr ck_tile::index_t VectorSizeC = 8;
// Implicit GEMM Traits
using CodegenShape =
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>>;
constexpr auto ConvSpec = ck_tile::ConvolutionSpecialization::Default;
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenShape>;
using GroupedConvTraitsType =
ck_tile::GroupedConvTraits<NDimSpatial, ConvSpec, InLayout, WeiLayout, DsLayout, OutLayout>;
using CodegenPipelineProblem =
ck_tile::GemmPipelineProblem<InDataType,
WeiDataType,
AccDataType,
CodegenShape,
typename GroupedConvTraitsType::GroupedConvImplicitGemmTraits,
InDataType,
true,
VectorSizeA,
VectorSizeB>;
using CodegenPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
const auto Run = [&](const auto memory_operation_) {
constexpr auto memory_operation = memory_operation_.value;
using ConvEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<InDataType,
WeiDataType,
DsDataType,
AccDataType,
OutDataType,
typename GroupedConvTraitsType::ImplicitGemmDsLayout,
ck_tile::tensor_layout::gemm::RowMajor,
CDEElementWise,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
CodegenPipelineProblem::TransposeC,
memory_operation,
1,
true,
VectorSizeC>>;
using Kernel = ck_tile::GroupedConvolutionBackwardDataKernel<GroupedConvTraitsType,
TilePartitioner,
CodegenPipeline,
ConvEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args);
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping conv!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args: " << Kernel::GetName() << '\n'
<< "shape: " << CodegenShape::GetName() << '\n'
<< "problem: " << CodegenPipelineProblem::GetName() << '\n'
<< "pipeline: " << CodegenPipeline::GetName() << '\n'
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< '\n'
<< "Vector size A: " << CodegenPipeline::GetVectorSizeA()
<< ", Vector size B: " << CodegenPipeline::GetVectorSizeB()
<< ", Vector size C: " << ConvEpilogue::GetVectorSizeC() << std::endl;
}
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};
if(args.k_batch == 1)
{
return Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::set>{});
}
else
{
return Run(ck_tile::integral_constant<ck_tile::memory_operation_enum,
ck_tile::memory_operation_enum::atomic_add>{});
}
}
#include "run_grouped_convolution_bwd_data_example.inc"
template <typename InPrecType, typename WeiPrecType = InPrecType, typename OutPrecType = InPrecType>
int run_grouped_conv_bwd_data_example_prec_type(
std::string in_layout, std::string wei_layout, std::string out_layout, int argc, char* argv[])
{
using NWGC = ck_tile::tensor_layout::convolution::NWGC;
using NHWGC = ck_tile::tensor_layout::convolution::NHWGC;
using NDHWGC = ck_tile::tensor_layout::convolution::NDHWGC;
using GKXC = ck_tile::tensor_layout::convolution::GKXC;
using GKYXC = ck_tile::tensor_layout::convolution::GKYXC;
using GKZYXC = ck_tile::tensor_layout::convolution::GKZYXC;
using NWGK = ck_tile::tensor_layout::convolution::NWGK;
using NHWGK = ck_tile::tensor_layout::convolution::NHWGK;
using NDHWGK = ck_tile::tensor_layout::convolution::NDHWGK;
if(in_layout == "NWGC" && wei_layout == "GKXC" && out_layout == "NWGK")
{
return run_grouped_conv_bwd_data_example_with_layouts<ck_tile::number<1>{},
InPrecType,
WeiPrecType,
OutPrecType>(
argc, argv, NWGC{}, GKXC{}, NWGK{});
}
else if(in_layout == "NHWGC" && wei_layout == "GKYXC" && out_layout == "NHWGK")
{
return run_grouped_conv_bwd_data_example_with_layouts<ck_tile::number<2>{},
InPrecType,
WeiPrecType,
OutPrecType>(
argc, argv, NHWGC{}, GKYXC{}, NHWGK{});
}
else if(in_layout == "NDHWGC" && wei_layout == "GKZYXC" && out_layout == "NDHWGK")
{
return run_grouped_conv_bwd_data_example_with_layouts<ck_tile::number<3>{},
InPrecType,
WeiPrecType,
OutPrecType>(
argc, argv, NDHWGC{}, GKZYXC{}, NDHWGK{});
}
else
{
throw std::runtime_error("Unsupported memory layout!");
}
}
int run_grouped_conv_bwd_data_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
std::string data_type = arg_parser.get_str("prec");
std::string in_layout = arg_parser.get_str("in_layout");
std::string wei_layout = arg_parser.get_str("wei_layout");
std::string out_layout = arg_parser.get_str("out_layout");
if(data_type == "fp16")
{
return run_grouped_conv_bwd_data_example_prec_type<ck_tile::half_t>(
in_layout, wei_layout, out_layout, argc, argv);
}
else if(data_type == "bf16")
{
return run_grouped_conv_bwd_data_example_prec_type<ck_tile::bf16_t>(
in_layout, wei_layout, out_layout, argc, argv);
}
else
{
throw std::runtime_error("Unsupported data type for this operation!");
}
}
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_data_example(argc, argv); }

7
example/ck_tile/20_grouped_convolution/grouped_convolution_backward_weight.cpp
View File

@@ -78,7 +78,6 @@ float grouped_conv_bwd_weight(const ck_tile::GroupedConvBwdWeightHostArgs& args,
typename GroupedConvTraitsType::ImplicitGemmDsLayout,
ck_tile::tensor_layout::gemm::RowMajor,
CDEElementWise,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
@@ -98,8 +97,8 @@ float grouped_conv_bwd_weight(const ck_tile::GroupedConvBwdWeightHostArgs& args,
ConvEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(kargs);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(kargs);
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -123,7 +122,7 @@ float grouped_conv_bwd_weight(const ck_tile::GroupedConvBwdWeightHostArgs& args,
float ave_time = ck_tile::launch_kernel_time_mask(
s,
Kernel::Preprocess(kargs, s),
ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};

7
example/ck_tile/20_grouped_convolution/grouped_convolution_forward.cpp
View File

@@ -77,7 +77,6 @@ float grouped_conv_fwd(const ck_tile::GroupedConvFwdHostArgs& args, const ck_til
typename GroupedConvTraitsType::ImplicitGemmDsLayout,
ck_tile::tensor_layout::gemm::RowMajor,
CDEElementWise,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
@@ -97,8 +96,8 @@ float grouped_conv_fwd(const ck_tile::GroupedConvFwdHostArgs& args, const ck_til
ConvEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(kargs);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(kargs);
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -120,7 +119,7 @@ float grouped_conv_fwd(const ck_tile::GroupedConvFwdHostArgs& args, const ck_til
}
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};

186
example/ck_tile/20_grouped_convolution/run_grouped_convolution_bwd_data_example.inc Normal file
View File

@@ -0,0 +1,186 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <ck_tile::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename AccDataType,
typename OutDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
float invoke_grouped_conv_bwd_data(ck_tile::GroupedConvBwdDataHostArgs& args,
int n_warmup,
int n_repeat)
{
float ave_time = grouped_conv_bwd_data<NDimSpatial,
InDataType,
WeiDataType,
AccDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::size_t flop = args.GetFlops();
std::size_t num_byte = args.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< std::endl;
return ave_time;
}
template <ck_tile::index_t NDimSpatial,
typename InDataType,
typename WeiDataType = InDataType,
typename OutDataType = InDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
int run_grouped_conv_bwd_data_example_with_layouts(
int argc, char* argv[], const InLayout, const WeiLayout, const OutLayout)
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using AccDataType = float;
std::vector<ck_tile::index_t> filter_spatial_lengths;
std::vector<ck_tile::index_t> image_spatial_lengths;
std::vector<ck_tile::index_t> strides;
std::vector<ck_tile::index_t> dilations;
std::vector<ck_tile::index_t> lpads;
std::vector<ck_tile::index_t> rpads;
const ck_tile::index_t num_dim_sp = fill_spatial_dimensions(filter_spatial_lengths,
image_spatial_lengths,
strides,
dilations,
lpads,
rpads,
arg_parser);
ck_tile::conv::ConvParam conv_param{num_dim_sp,
arg_parser.get_int("g"),
arg_parser.get_int("n"),
arg_parser.get_int("k"),
arg_parser.get_int("c"),
filter_spatial_lengths,
image_spatial_lengths,
strides,
dilations,
lpads,
rpads};
ck_tile::index_t kbatch = arg_parser.get_int("split_k");
int n_warmup = arg_parser.get_int("warmup");
int n_repeat = arg_parser.get_int("repeat");
ck_tile::index_t init_method = arg_parser.get_int("init");
const auto in_g_n_c_wis_desc =
ck_tile::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
const auto wei_g_k_c_xs_desc =
ck_tile::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
const auto out_g_n_k_wos_desc =
ck_tile::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
ck_tile::HostTensor<InDataType> input(in_g_n_c_wis_desc);
ck_tile::HostTensor<WeiDataType> weight(wei_g_k_c_xs_desc);
ck_tile::HostTensor<OutDataType> output(out_g_n_k_wos_desc);
if(init_method == 0)
{
ck_tile::FillUniformDistribution<WeiDataType>{-1.f, 1.f}(weight);
ck_tile::FillUniformDistribution<OutDataType>{-1.f, 1.f}(output);
}
else if(init_method == 1)
{
ck_tile::FillMonotonicSeq<WeiDataType>{}(weight);
ck_tile::FillMonotonicSeq<OutDataType>{}(output);
}
else if(init_method == 2)
{
ck_tile::FillUniformDistribution<WeiDataType>{1.f, 1.f}(weight);
ck_tile::FillUniformDistribution<OutDataType>{1.f, 1.f}(output);
}
else
{
weight.SetZero();
output.SetZero();
}
ck_tile::DeviceMem input_dev_buf(input.get_element_space_size_in_bytes());
ck_tile::DeviceMem weight_dev_buf(weight.get_element_space_size_in_bytes());
ck_tile::DeviceMem output_dev_buf(output.get_element_space_size_in_bytes());
input_dev_buf.SetZero();
weight_dev_buf.ToDevice(weight.data());
output_dev_buf.ToDevice(output.data());
ck_tile::GroupedConvBwdDataHostArgs args(conv_param,
input_dev_buf.GetDeviceBuffer(),
weight_dev_buf.GetDeviceBuffer(),
{},
output_dev_buf.GetDeviceBuffer(),
kbatch);
std::cout << "Run Grouped Conv Bwd Data kernel" << std::endl;
std::cout << "input: " << input.mDesc << std::endl;
std::cout << "weight: " << weight.mDesc << std::endl;
std::cout << "output: " << output.mDesc << std::endl;
invoke_grouped_conv_bwd_data<NDimSpatial,
InDataType,
WeiDataType,
AccDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout>(args, n_warmup, n_repeat);
input_dev_buf.FromDevice(input.data());
bool pass = true;
if(arg_parser.get_int("v") == 1)
{
ck_tile::HostTensor<InDataType> input_host_ref(in_g_n_c_wis_desc);
input_host_ref.SetZero();
ck_tile::reference_grouped_conv_bwd_data<NDimSpatial, InDataType, WeiDataType, OutDataType>(
input_host_ref,
weight,
output,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_);
const ck_tile::index_t GemmK = weight.get_element_size() / (conv_param.G_ * conv_param.K_);
const float max_accumulated_value =
*std::max_element(input_host_ref.mData.begin(), input_host_ref.mData.end());
const auto rtol_atol =
calculate_rtol_atol<InDataType, WeiDataType, AccDataType, OutDataType>(
GemmK, kbatch, max_accumulated_value);
pass = ck_tile::check_err(input,
input_host_ref,
"Error: Incorrect results!",
rtol_atol.at(ck_tile::number<0>{}),
rtol_atol.at(ck_tile::number<1>{}));
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
<< std::endl;
std::cout << "The CPU verification result is:" << (pass ? "correct" : "fail") << std::endl;
}
else if(arg_parser.get_int("v") == 2)
{
throw std::runtime_error("Unsupported gpu verification !!!");
}
return pass;
}

22
example/ck_tile/21_elementwise/elementwise_example.cpp
View File

@@ -167,17 +167,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
}
// 4. Run the kernel
float ave_time = launch_kernel(ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockSize, kBlockPerCu>(
Kernel{},
kGridSize,
kBlockSize,
0,
input_size,
ck_tile::make_tuple(N, 1), // Input Stride
ck_tile::make_tuple(N, 1), // Output Stride
input_tensors,
static_cast<YDataType*>(y_buf.GetDeviceBuffer())));
float ave_time = launch_kernel(
ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockPerCu>(Kernel{},
kGridSize,
kBlockSize,
0,
input_size,
ck_tile::make_tuple(N, 1), // Input Stride
ck_tile::make_tuple(N, 1), // Output Stride
input_tensors,
static_cast<YDataType*>(y_buf.GetDeviceBuffer())));
std::cout << "Average time: " << ave_time << " ms" << std::endl;

2
example/ck_tile/21_elementwise/elementwise_example_add_4d.cpp
View File

@@ -113,7 +113,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
// Run the kernel
float ave_time = launch_kernel(
ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockSize, kBlockPerCu>(
ck_tile::make_kernel<kBlockPerCu>(
Kernel{},
kGridSize,
kBlockSize,

22
example/ck_tile/21_elementwise/elementwise_example_transpose.cpp
View File

@@ -112,17 +112,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
}
// 4. Run the kernel
float ave_time = launch_kernel(ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockSize, kBlockPerCu>(
Kernel{},
kGridSize,
kBlockSize,
0, // Shared memory
op_lengths, // Logical dimensions for the operation (M, N)
input_strides, // Strides for input tensor(s)
output_strides, // Strides for output tensor (N, M)
input_tensors,
static_cast<YDataType*>(y_buf.GetDeviceBuffer())));
float ave_time = launch_kernel(
ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockPerCu>(Kernel{},
kGridSize,
kBlockSize,
0, // Shared memory
op_lengths, // Logical dimensions for the operation (M, N)
input_strides, // Strides for input tensor(s)
output_strides, // Strides for output tensor (N, M)
input_tensors,
static_cast<YDataType*>(y_buf.GetDeviceBuffer())));
std::cout << "Average time: " << ave_time << " ms" << std::endl;

22
example/ck_tile/21_elementwise/elementwise_example_unary.cpp
View File

@@ -99,17 +99,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
}
// 4. Run the kernel
float ave_time = launch_kernel(ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockSize, kBlockPerCu>(
Kernel{},
kGridSize,
kBlockSize,
0,
input_size,
ck_tile::make_tuple(N, 1), // Input Stride
ck_tile::make_tuple(N, 1), // Output Stride
input_tensors,
static_cast<YDataType*>(y_buf.GetDeviceBuffer())));
float ave_time = launch_kernel(
ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
ck_tile::make_kernel<kBlockPerCu>(Kernel{},
kGridSize,
kBlockSize,
0,
input_size,
ck_tile::make_tuple(N, 1), // Input Stride
ck_tile::make_tuple(N, 1), // Output Stride
input_tensors,
static_cast<YDataType*>(y_buf.GetDeviceBuffer())));
std::cout << "Average time: " << ave_time << " ms" << std::endl;

2
example/ck_tile/35_batched_transpose/README.md
View File

@@ -6,7 +6,7 @@ This folder contains example for batched Transpose using ck_tile tile-programmin
# 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>
../script/cmake-ck-dev.sh ../ <arch>
# Make the transpose executable
make tile_example_batched_transpose -j
```

8
example/ck_tile/35_batched_transpose/batched_transpose_api.cpp
View File

@@ -74,8 +74,8 @@ float batched_transpose_dispatch(batched_transpose_kargs& a, ck_tile::stream_con
auto kargs = kernel::MakeKargs(a);
const dim3 grids = kernel::GridSize(a);
constexpr dim3 blocks = kernel::BlockSize();
const dim3 grids = kernel::GridSize(a);
const dim3 blocks = kernel::BlockSize();
printf("Pipeline: %d\n", Config::kPipelineId);
printf("Grid: x=%u y=%u z=%u\n", grids.x, grids.y, grids.z);
@@ -96,8 +96,8 @@ float batched_transpose_dispatch(batched_transpose_kargs& a, ck_tile::stream_con
printf("Launching Kernel...\n");
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, 1>(kernel{}, grids, blocks, 0, kargs));
float ave_time =
ck_tile::launch_kernel(s, ck_tile::make_kernel<1>(kernel{}, grids, blocks, 0, kargs));
printf("Kernel finished...\n");

5
example/ck_tile/38_block_scale_gemm/CMakeLists.txt
View File

@@ -8,9 +8,8 @@ list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -mllvm -enable-noalias-to-md-conversion
if(GPU_TARGETS MATCHES "gfx94" OR GPU_TARGETS MATCHES "gfx95")
add_executable(tile_example_gemm_aquant_basic EXCLUDE_FROM_ALL gemm_aquant_basic.cpp)
target_compile_options(tile_example_gemm_aquant_basic PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
add_executable(tile_example_gemm_aquant_preshuffle EXCLUDE_FROM_ALL gemm_aquant_preshuffle.cpp)
target_compile_options(tile_example_gemm_aquant_preshuffle PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
add_executable(tile_example_gemm_bquant_basic EXCLUDE_FROM_ALL gemm_bquant_basic.cpp)
target_compile_options(tile_example_gemm_bquant_basic PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
else()
message(DEBUG "Skipping ck_tile quant gemm tests for current target")
endif()

3
example/ck_tile/38_block_scale_gemm/README.md
View File

@@ -7,9 +7,10 @@ This folder contains example for Block Scale GEMM using ck_tile tile-programming
# 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>
../script/cmake-ck-dev.sh ../ <arch>
# The aquant pipeline method on the gemm calculation
make tile_example_gemm_aquant_basic -j
make tile_example_gemm_bquant_basic -j
```
This will result in an executable `build/bin/tile_example_gemm_aquant_basic`

76
example/ck_tile/38_block_scale_gemm/gemm_aquant_basic.cpp
View File

@@ -8,11 +8,10 @@
#include <string>
#include <tuple>
#include "ck_tile/core/config.hpp"
#include "ck_tile/host.hpp"
#include "gemm_utils.hpp"
template <typename ADataType,
template <typename GemmConfig,
typename ADataType,
typename AQDataType,
typename BDataType,
typename AccDataType,
@@ -21,29 +20,26 @@ template <typename ADataType,
typename ALayout,
typename BLayout,
typename CLayout,
uint32_t QuantGroupSize,
bool Preshuffle = false>
uint32_t QuantGroupSize>
float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::stream_config& s)
{
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr int kBlockPerCu = 1;
static_assert(std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr ck_tile::index_t M_Tile = 16;
constexpr ck_tile::index_t N_Tile = 64;
constexpr ck_tile::index_t K_Tile = 256;
constexpr ck_tile::index_t M_Tile = GemmConfig::M_Tile;
constexpr ck_tile::index_t N_Tile = GemmConfig::N_Tile;
constexpr ck_tile::index_t K_Tile = GemmConfig::K_Tile;
constexpr ck_tile::index_t M_Warp = 1;
constexpr ck_tile::index_t N_Warp = 4;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp = GemmConfig::M_Warp;
constexpr ck_tile::index_t N_Warp = GemmConfig::N_Warp;
constexpr ck_tile::index_t K_Warp = GemmConfig::K_Warp;
constexpr ck_tile::index_t M_Warp_Tile = 16;
constexpr ck_tile::index_t N_Warp_Tile = 16;
constexpr ck_tile::index_t K_Warp_Tile = 32;
constexpr ck_tile::index_t M_Warp_Tile = GemmConfig::M_Warp_Tile;
constexpr ck_tile::index_t N_Warp_Tile = GemmConfig::N_Warp_Tile;
constexpr ck_tile::index_t K_Warp_Tile = GemmConfig::K_Warp_Tile;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
@@ -52,8 +48,13 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
using CodegenGemmTraits =
ck_tile::TileGemmAQuantTraits<kPadM, kPadN, kPadK, Preshuffle, ALayout, BLayout, CLayout>;
using CodegenGemmTraits = ck_tile::TileGemmAQuantTraits<kPadM,
kPadN,
kPadK,
GemmConfig::PreshuffleQuant,
ALayout,
BLayout,
CLayout>;
using GemmPipelineProblem = ck_tile::GemmPipelineProblemBase<ADataType,
BDataType,
@@ -68,7 +69,7 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
constexpr bool transposed_warp_gemm = false;
constexpr bool transposed_warp_gemm = true;
const auto Run = [&](const auto has_hot_loop_, const auto tail_number_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
@@ -82,6 +83,7 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
CodegenGemmShape,
CodegenGemmTraits,
QuantGroupSize,
transposed_warp_gemm,
ComputeDataType,
ck_tile::GemmPipelineScheduler::Intrawave,
has_hot_loop_v,
@@ -96,7 +98,6 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
ck_tile::tuple<>,
CLayout,
ck_tile::element_wise::PassThrough,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
@@ -111,8 +112,8 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
const dim3 blocks = Kernel::BlockSize();
if(args.k_batch != 1)
{
@@ -136,7 +137,7 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
}
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};
@@ -187,13 +188,14 @@ int run_gemm_example(int argc, char* argv[])
if(data_type == "fp8")
{
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t>{});
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "bf8")
{
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, float>{});
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
@@ -201,32 +203,18 @@ int run_gemm_example(int argc, char* argv[])
{
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t,
ck_tile::fp8_t,
float,
ck_tile::half_t,
ck_tile::fp8_t>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "i4bf8")
{
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t,
ck_tile::bf8_t,
float,
ck_tile::half_t,
ck_tile::bf8_t>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "i4f32fp8")
{
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::fp8_t, float, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "i4f32bf8")
{
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::bf8_t, float, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else
@@ -235,4 +223,4 @@ int run_gemm_example(int argc, char* argv[])
}
}
int main(int argc, char* argv[]) { return !run_gemm_example<GemmConfigComputeV3>(argc, argv); }
int main(int argc, char* argv[]) { return !run_gemm_example<GemmConfigDecode>(argc, argv); }

71
example/ck_tile/38_block_scale_gemm/gemm_aquant_preshuffle.cpp
View File

@@ -8,11 +8,10 @@
#include <string>
#include <tuple>
#include "ck_tile/core/config.hpp"
#include "ck_tile/host.hpp"
#include "gemm_utils.hpp"
template <typename ADataType,
template <typename GemmConfig,
typename ADataType,
typename AQDataType,
typename BDataType,
typename AccDataType,
@@ -21,8 +20,7 @@ template <typename ADataType,
typename ALayout,
typename BLayout,
typename CLayout,
uint32_t QuantGroupSize,
bool Preshuffle = false>
uint32_t QuantGroupSize>
float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::stream_config& s)
{
constexpr bool kPadM = false;
@@ -33,17 +31,17 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
static_assert(std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr ck_tile::index_t M_Tile = 16;
constexpr ck_tile::index_t N_Tile = 64;
constexpr ck_tile::index_t K_Tile = 256;
constexpr ck_tile::index_t M_Tile = GemmConfig::M_Tile;
constexpr ck_tile::index_t N_Tile = GemmConfig::N_Tile;
constexpr ck_tile::index_t K_Tile = GemmConfig::K_Tile;
constexpr ck_tile::index_t M_Warp = 1;
constexpr ck_tile::index_t N_Warp = 4;
constexpr ck_tile::index_t K_Warp = 1;
constexpr ck_tile::index_t M_Warp = GemmConfig::M_Warp;
constexpr ck_tile::index_t N_Warp = GemmConfig::N_Warp;
constexpr ck_tile::index_t K_Warp = GemmConfig::K_Warp;
constexpr ck_tile::index_t M_Warp_Tile = 16;
constexpr ck_tile::index_t N_Warp_Tile = 16;
constexpr ck_tile::index_t K_Warp_Tile = 32;
constexpr ck_tile::index_t M_Warp_Tile = GemmConfig::M_Warp_Tile;
constexpr ck_tile::index_t N_Warp_Tile = GemmConfig::N_Warp_Tile;
constexpr ck_tile::index_t K_Warp_Tile = GemmConfig::K_Warp_Tile;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
@@ -52,8 +50,13 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
using CodegenGemmTraits =
ck_tile::TileGemmAQuantTraits<kPadM, kPadN, kPadK, Preshuffle, ALayout, BLayout, CLayout>;
using CodegenGemmTraits = ck_tile::TileGemmAQuantTraits<kPadM,
kPadN,
kPadK,
GemmConfig::PreshuffleQuant,
ALayout,
BLayout,
CLayout>;
using GemmPipelineProblem = ck_tile::GemmPipelineProblemBase<ADataType,
BDataType,
@@ -82,6 +85,7 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
CodegenGemmShape,
CodegenGemmTraits,
QuantGroupSize,
transposed_warp_gemm,
ComputeDataType,
ck_tile::GemmPipelineScheduler::Intrawave,
has_hot_loop_v,
@@ -96,7 +100,6 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
ck_tile::tuple<>,
CLayout,
ck_tile::element_wise::PassThrough,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
@@ -111,8 +114,8 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
const dim3 blocks = Kernel::BlockSize();
if(args.k_batch != 1)
{
@@ -136,7 +139,7 @@ float gemm_calc_aquant(const ck_tile::AQuantGemmHostArgs& args, const ck_tile::s
}
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};
@@ -187,13 +190,14 @@ int run_gemm_example(int argc, char* argv[])
if(data_type == "fp8")
{
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t>{});
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "bf8")
{
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, float>{});
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
@@ -201,7 +205,7 @@ int run_gemm_example(int argc, char* argv[])
{
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t,
ck_tile::fp8_t,
float,
ck_tile::half_t,
ck_tile::fp8_t>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
@@ -210,29 +214,18 @@ int run_gemm_example(int argc, char* argv[])
{
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t,
ck_tile::bf8_t,
float,
ck_tile::half_t,
ck_tile::bf8_t>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "i4f32fp8")
{
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::fp8_t, float, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "i4f32bf8")
{
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::bf8_t, float, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else
{
throw std::runtime_error("Unsupported data type for this operation !!!");
}
}
int main(int argc, char* argv[]) { return !run_gemm_example<GemmConfigPreshufle_AQ>(argc, argv); }
int main(int argc, char* argv[])
{
return !run_gemm_example<GemmConfigPreshuffleQuant>(argc, argv);
}

228
example/ck_tile/38_block_scale_gemm/gemm_bquant_basic.cpp Normal file
View File

@@ -0,0 +1,228 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <cstring>
#include <iostream>
#include <ostream>
#include <stdexcept>
#include <string>
#include <tuple>
#include "ck_tile/core/config.hpp"
#include "ck_tile/host.hpp"
#include "gemm_utils.hpp"
template <typename GemmConfig,
typename ADataType,
typename BDataType,
typename BQDataType,
typename AccDataType,
typename CDataType,
typename ComputeDataType,
typename ALayout,
typename BLayout,
typename CLayout,
uint32_t QuantGroupSize>
float gemm_calc_bquant(const ck_tile::BQuantGemmHostArgs& args, const ck_tile::stream_config& s)
{
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
static_assert(std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr ck_tile::index_t M_Tile = GemmConfig::M_Tile;
constexpr ck_tile::index_t N_Tile = GemmConfig::N_Tile;
constexpr ck_tile::index_t K_Tile = GemmConfig::K_Tile;
constexpr ck_tile::index_t M_Warp = GemmConfig::M_Warp;
constexpr ck_tile::index_t N_Warp = GemmConfig::N_Warp;
constexpr ck_tile::index_t K_Warp = GemmConfig::K_Warp;
constexpr ck_tile::index_t M_Warp_Tile = GemmConfig::M_Warp_Tile;
constexpr ck_tile::index_t N_Warp_Tile = GemmConfig::N_Warp_Tile;
constexpr ck_tile::index_t K_Warp_Tile = GemmConfig::K_Warp_Tile;
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::GemmTile1DPartitioner<CodegenGemmShape>;
using CodegenGemmTraits = ck_tile::TileGemmBQuantTraits<kPadM,
kPadN,
kPadK,
GemmConfig::PreshuffleQuant,
ALayout,
BLayout,
CLayout>;
using GemmPipelineProblem = ck_tile::GemmPipelineProblemBase<ADataType,
BDataType,
AccDataType,
CodegenGemmShape,
CodegenGemmTraits,
ComputeDataType>;
using BaseGemmPipeline = ck_tile::BaseBQuantGemmPipelineAgBgCrCompV3<GemmPipelineProblem>;
const ck_tile::index_t K_split = (args.K + K_Tile - 1) / K_Tile * K_Tile;
const ck_tile::index_t num_loop = TilePartitioner::GetLoopNum(K_split);
const bool has_hot_loop = BaseGemmPipeline::BlockHasHotloop(num_loop);
const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
constexpr bool transposed_warp_gemm = false;
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 CodegenPipelineProblem =
ck_tile::GemmBQuantPipelineProblem<ADataType,
BDataType,
BQDataType,
AccDataType,
CodegenGemmShape,
CodegenGemmTraits,
QuantGroupSize,
ComputeDataType,
ck_tile::GemmPipelineScheduler::Intrawave,
has_hot_loop_v,
tail_number_v>;
using CodegenGemmPipeline = ck_tile::BQuantGemmPipelineAgBgCrCompV3<CodegenPipelineProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ck_tile::tuple<>,
CLayout,
ck_tile::element_wise::PassThrough,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
transposed_warp_gemm,
ck_tile::memory_operation_enum::set>>;
using Kernel =
ck_tile::BQuantGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
const dim3 blocks = Kernel::BlockSize();
if(args.k_batch != 1)
{
throw std::runtime_error("split-k is not supported yet!");
}
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args: " << Kernel::GetName() << '\n'
<< "shape: " << CodegenGemmShape::GetName() << '\n'
<< "problem: " << CodegenPipelineProblem::GetName() << '\n'
<< "pipeline: " << CodegenGemmPipeline::GetName() << '\n'
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};
return BaseGemmPipeline::TailHandler(Run, has_hot_loop, tail_num);
;
}
#include "run_gemm_bquant_example.inc"
template <typename GemmConfig, typename TypeConfig, uint32_t QuantGroupSize>
int run_gemm_example_prec_type(std::string a_layout, std::string b_layout, int argc, char* argv[])
{
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
if constexpr(std::is_same_v<typename TypeConfig::BDataType, ck_tile::pk_int4_t> ||
std::is_same_v<typename TypeConfig::BDataType, ck_tile::fp8_t> ||
std::is_same_v<typename TypeConfig::BDataType, ck_tile::bf8_t>)
{
if(a_layout == "R" && b_layout == "C")
{
return run_gemm_example_with_layouts<GemmConfig, TypeConfig, QuantGroupSize>(
argc, argv, Row{}, Col{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported memory layout for the input matrices!");
}
}
else
{
throw std::runtime_error("Unsupported data type for B.");
}
return 0;
}
template <template <typename PreType> typename GemmConfig>
int run_gemm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
std::string data_type = arg_parser.get_str("prec");
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(data_type == "fp8")
{
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "bf8")
{
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "fp8i4")
{
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::fp8_t,
ck_tile::pk_int4_t,
ck_tile::half_t,
ck_tile::fp8_t>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else if(data_type == "bf8i4")
{
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::bf8_t,
ck_tile::pk_int4_t,
ck_tile::half_t,
ck_tile::bf8_t>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>, TypeConfig, 128>(
a_layout, b_layout, argc, argv);
}
else
{
throw std::runtime_error("Unsupported data type for this operation !!!");
}
}
int main(int argc, char* argv[]) { return !run_gemm_example<GemmConfigDecode>(argc, argv); }

427
example/ck_tile/38_block_scale_gemm/gemm_utils.hpp
View File

@@ -11,11 +11,9 @@
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/ops/gemm_group_quant.hpp"
#define CK_TILE_PIPELINE_COMPUTE_V3 1
#define CK_TILE_PIPELINE_MEMORY 2
#define CK_TILE_PIPELINE_COMPUTE_V4 3
#define CK_TILE_PIPELINE_COMPUTE_V5 4
#define CK_TILE_PIPELINE_PRESHUFFLE 5
#define CK_TILE_PIPELINE_PREFILL 1
#define CK_TILE_PIPELINE_DECODE 2
#define CK_TILE_PIPELINE_PRESHUFFLEQUANT 3
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_warp_tile()
@@ -83,200 +81,37 @@ struct GemmConfigBase
static constexpr bool TransposeC = false;
static constexpr bool UseStructuredSparsity = false;
static constexpr int kBlockPerCu = 1;
static constexpr ck_tile::index_t TileParitionerGroupNum = 8;
static constexpr ck_tile::index_t TileParitionerM01 = 4;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Intrawave;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V3;
static constexpr ck_tile::index_t NumWaveGroups = 1;
static constexpr bool Preshuffle = false;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Intrawave;
static constexpr bool PreshuffleQuant = false;
static constexpr bool DoubleSmemBuffer = false;
};
template <typename PrecType>
struct GemmConfigMemoryInterwave : public GemmConfigBase
struct GemmConfigDecode : public GemmConfigBase
{
// Memory friendly for Interwave scheduler
static constexpr ck_tile::index_t M_Tile = 128;
static constexpr ck_tile::index_t N_Tile = 32;
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 4;
static constexpr ck_tile::index_t N_Warp = 1;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(PrecType) == 2 ? 8 : 16;
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_MEMORY;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Interwave;
};
template <typename PrecType>
struct GemmConfigMemoryIntrawave : public GemmConfigBase
{
static constexpr ck_tile::index_t M_Tile = 128;
static constexpr ck_tile::index_t N_Tile = 32;
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 4;
static constexpr ck_tile::index_t N_Warp = 1;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = sizeof(PrecType) == 2 ? 8 : 16;
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_MEMORY;
};
template <typename PrecType>
struct GemmConfigComputeV3 : public GemmConfigBase
{
// Compute V3 only support Intrawave scheduler
static constexpr ck_tile::index_t M_Tile = 32;
static constexpr ck_tile::index_t N_Tile = 128;
static constexpr ck_tile::index_t M_Tile = 16;
static constexpr ck_tile::index_t N_Tile = 64;
static constexpr ck_tile::index_t K_Tile = 256 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 1;
static constexpr ck_tile::index_t N_Warp = 4;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V3;
};
template <typename PrecType>
struct GemmConfigComputeV3_1 : public GemmConfigBase
{
static constexpr ck_tile::index_t M_Tile = 256;
static constexpr ck_tile::index_t N_Tile = 256;
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 2;
static constexpr ck_tile::index_t N_Warp = 2;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V3;
};
template <typename PrecType>
struct GemmConfigComputeV3_2 : public GemmConfigBase
{
static constexpr ck_tile::index_t M_Tile = 128;
static constexpr ck_tile::index_t N_Tile = 128;
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 2;
static constexpr ck_tile::index_t N_Warp = 2;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V3;
static constexpr int kBlockPerCu = 1;
static constexpr int kBlockPerCu = 2;
};
template <typename PrecType>
struct GemmConfigComputeV4 : public GemmConfigBase
{
// Compute V4 only support Intrawave scheduler
// Using the ping pong reader in the lds level
static constexpr ck_tile::index_t M_Tile = 256;
static constexpr ck_tile::index_t N_Tile = 256;
static constexpr ck_tile::index_t K_Tile = 64 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 2;
static constexpr ck_tile::index_t N_Warp = 2;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = true;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V4;
};
template <typename PrecType>
struct GemmConfigComputeV4_1 : public GemmConfigBase
{
static constexpr ck_tile::index_t M_Tile = 256;
static constexpr ck_tile::index_t N_Tile = 256;
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 2;
static constexpr ck_tile::index_t N_Warp = 2;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = true;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V4;
};
template <typename PrecType>
struct GemmConfigComputeV5 : public GemmConfigBase
{
static constexpr ck_tile::index_t M_Tile = 128;
static constexpr ck_tile::index_t N_Tile = 128;
static constexpr ck_tile::index_t K_Tile = 64 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 1;
static constexpr ck_tile::index_t N_Warp = 1;
static constexpr ck_tile::index_t K_Warp = 2;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_COMPUTE_V5;
static constexpr ck_tile::index_t NumWaNumWaveGroups = 2;
};
template <typename PrecType>
struct GemmConfigPreshufle_1 : public GemmConfigBase
{
static constexpr ck_tile::index_t M_Tile = 128;
static constexpr ck_tile::index_t N_Tile = 128;
static constexpr ck_tile::index_t K_Tile = 128 / sizeof(PrecType);
static constexpr ck_tile::index_t M_Warp = 1;
static constexpr ck_tile::index_t N_Warp = 4;
static constexpr ck_tile::index_t K_Warp = 1;
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile =
get_k_from_preshuffled_warp_tile<PrecType, M_Warp_Tile>();
static constexpr int kBlockPerCu = 2;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_PRESHUFFLE;
static constexpr bool Preshuffle = true;
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_DECODE;
};
template <typename PrecType>
struct GemmConfigPreshufle_2 : public GemmConfigBase
struct GemmConfigPrefill : public GemmConfigBase
{
static constexpr ck_tile::index_t M_Tile = 128;
static constexpr ck_tile::index_t N_Tile = 128;
@@ -288,18 +123,15 @@ struct GemmConfigPreshufle_2 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile =
get_k_from_preshuffled_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr int kBlockPerCu = 2;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_PRESHUFFLE;
static constexpr bool Preshuffle = true;
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_PREFILL;
};
template <typename PrecType>
struct GemmConfigPreshufle_AQ : public GemmConfigBase
struct GemmConfigPreshuffleQuant : public GemmConfigBase
{
static constexpr ck_tile::index_t M_Tile = 16;
static constexpr ck_tile::index_t N_Tile = 64;
@@ -314,9 +146,11 @@ struct GemmConfigPreshufle_AQ : public GemmConfigBase
static constexpr ck_tile::index_t K_Warp_Tile =
get_k_from_preshuffled_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_PRESHUFFLE;
static constexpr bool Preshuffle = true;
static constexpr bool DoubleSmemBuffer = false;
static constexpr int kBlockPerCu = 1;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
static constexpr ck_tile::index_t Pipeline = CK_TILE_PIPELINE_PRESHUFFLEQUANT;
static constexpr bool PreshuffleQuant = true;
};
template <typename ADataType_,
@@ -332,176 +166,6 @@ struct GemmQuantTypeConfig
using CDataType = CDataType_;
};
template <>
struct GemmQuantTypeConfig<ck_tile::half_t>
{
using ADataType = ck_tile::half_t;
using QDataType = float;
using BDataType = ck_tile::half_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t>
{
using ADataType = ck_tile::bf16_t;
using QDataType = float;
using BDataType = ck_tile::bf16_t;
using AccDataType = float;
using CDataType = ck_tile::bf16_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t>
{
using ADataType = ck_tile::fp8_t;
using QDataType = float;
using BDataType = ck_tile::fp8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t>
{
using ADataType = ck_tile::bf8_t;
using QDataType = float;
using BDataType = ck_tile::bf8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::half_t, ck_tile::pk_int4_t, ck_tile::half_t>
{
using ADataType = ck_tile::half_t;
using QDataType = float;
using BDataType = ck_tile::pk_int4_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, float>
{
using ADataType = ck_tile::fp8_t;
using QDataType = float;
using BDataType = ck_tile::fp8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, float>
{
using ADataType = ck_tile::bf8_t;
using QDataType = float;
using BDataType = ck_tile::bf8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::fp8_t, float, ck_tile::fp8_t>
{
using ADataType = ck_tile::pk_int4_t;
using QDataType = ck_tile::fp8_t;
using BDataType = ck_tile::fp8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, float, ck_tile::fp8_t>
{
using ADataType = ck_tile::fp8_t;
using QDataType = ck_tile::fp8_t;
using BDataType = ck_tile::fp8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, float, ck_tile::bf8_t>
{
using ADataType = ck_tile::bf8_t;
using QDataType = ck_tile::bf8_t;
using BDataType = ck_tile::bf8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::bf8_t, float, ck_tile::bf8_t>
{
using ADataType = ck_tile::pk_int4_t;
using QDataType = ck_tile::bf8_t;
using BDataType = ck_tile::bf8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::fp8_t, float, float>
{
using ADataType = ck_tile::pk_int4_t;
using QDataType = float;
using BDataType = ck_tile::fp8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::bf8_t, float, float>
{
using ADataType = ck_tile::pk_int4_t;
using QDataType = float;
using BDataType = ck_tile::bf8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::pk_int4_t, float, ck_tile::fp8_t>
{
using ADataType = ck_tile::fp8_t;
using QDataType = ck_tile::fp8_t;
using BDataType = ck_tile::pk_int4_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::pk_int4_t, float, ck_tile::bf8_t>
{
using ADataType = ck_tile::bf8_t;
using QDataType = ck_tile::bf8_t;
using BDataType = ck_tile::pk_int4_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::pk_int4_t, float, float>
{
using ADataType = ck_tile::fp8_t;
using QDataType = float;
using BDataType = ck_tile::pk_int4_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::pk_int4_t, float, float>
{
using ADataType = ck_tile::bf8_t;
using QDataType = float;
using BDataType = ck_tile::pk_int4_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <typename T>
struct DataTypeTraits;
@@ -559,55 +223,6 @@ struct DataTypeTraits<ck_tile::int8_t>
static constexpr const char* name = "int8";
};
template <ck_tile::index_t PipelineId>
struct PipelineTypeTraits;
template <>
struct PipelineTypeTraits<CK_TILE_PIPELINE_MEMORY>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<PipelineProblem>;
template <typename PipelineProblem>
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<PipelineProblem>;
};
template <>
struct PipelineTypeTraits<CK_TILE_PIPELINE_COMPUTE_V3>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV3<PipelineProblem>;
template <typename PipelineProblem>
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV3<PipelineProblem>;
};
template <>
struct PipelineTypeTraits<CK_TILE_PIPELINE_COMPUTE_V4>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV4<PipelineProblem>;
template <typename PipelineProblem>
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV4<PipelineProblem>;
};
template <>
struct PipelineTypeTraits<CK_TILE_PIPELINE_COMPUTE_V5>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV5<PipelineProblem>;
template <typename PipelineProblem>
using UniversalGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV5<PipelineProblem>;
};
template <>
struct PipelineTypeTraits<CK_TILE_PIPELINE_PRESHUFFLE>
{
template <typename PipelineProblem>
using GemmPipeline = ck_tile::WeightPreshufflePipelineAGmemBGmemCRegV1<PipelineProblem>;
template <typename PipelineProblem>
using UniversalGemmPipeline =
ck_tile::BaseWeightPreshufflePipelineAGmemBGmemCRegV1<PipelineProblem>;
};
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;

49
example/ck_tile/38_block_scale_gemm/run_gemm_aquant_example.inc
View File

@@ -1,3 +1,4 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
@@ -31,7 +32,8 @@ auto shuffle_aq(const ck_tile::HostTensor<T>& t, int block_aq_k)
return ck_tile::reference_permute(t_view, {1, 0, 2});
}
template <typename ADataType,
template <typename GemmConfig,
typename ADataType,
typename AQDataType,
typename BDataType,
typename AccDataType,
@@ -40,8 +42,7 @@ template <typename ADataType,
typename AQLayout,
typename BLayout,
typename CLayout,
uint32_t QuantGroupSize,
bool Preshuffle = false>
uint32_t QuantGroupSize>
float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
ck_tile::DeviceMem& aq_m_aqk_dev_buf,
ck_tile::DeviceMem& b_k_n_dev_buf,
@@ -73,7 +74,8 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
args.stride_C = stride_C;
args.stride_AQ = stride_AQ;
float ave_time = gemm_calc_aquant<ADataType,
float ave_time = gemm_calc_aquant<GemmConfig,
ADataType,
AQDataType,
BDataType,
AccDataType,
@@ -82,8 +84,7 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
ALayout,
BLayout,
CLayout,
QuantGroupSize,
Preshuffle>(
QuantGroupSize>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::size_t flop = std::size_t(2) * M * N * K;
@@ -206,7 +207,7 @@ int run_gemm_example_with_layouts(int argc,
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());
if constexpr(GemmConfig::Preshuffle)
if constexpr(GemmConfig::PreshuffleQuant)
{
ck_tile::HostTensor<AQDataType> aq_shuffle_host =
shuffle_aq(aq_m_aqk, GemmConfig::K_Tile / QuantGroupSize);
@@ -222,7 +223,8 @@ int run_gemm_example_with_layouts(int argc,
c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero();
invoke_gemm<ADataType,
invoke_gemm<GemmConfig,
ADataType,
AQDataType,
BDataType,
AccDataType,
@@ -231,22 +233,21 @@ int run_gemm_example_with_layouts(int argc,
AQLayout,
BLayout,
CLayout,
QuantGroupSize,
GemmConfig::Preshuffle>(a_m_k_dev_buf,
aq_m_aqk_dev_buf,
b_k_n_dev_buf,
c_m_n_dev_buf,
M,
N,
K,
AQK,
stride_A,
stride_AQ,
stride_B,
stride_C,
kbatch,
n_warmup,
n_repeat);
QuantGroupSize>(a_m_k_dev_buf,
aq_m_aqk_dev_buf,
b_k_n_dev_buf,
c_m_n_dev_buf,
M,
N,
K,
AQK,
stride_A,
stride_AQ,
stride_B,
stride_C,
kbatch,
n_warmup,
n_repeat);
c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
bool pass = true;

286
example/ck_tile/38_block_scale_gemm/run_gemm_bquant_example.inc Normal file
View File

@@ -0,0 +1,286 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <bit>
#include <random>
template <typename Layout>
static constexpr inline auto is_row_major(Layout layout_)
{
return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
ck_tile::tensor_layout::gemm::RowMajor>>{};
}
template <typename T>
auto shuffle_bq(const ck_tile::HostTensor<T>& t, int block_bq_k)
{
if(t.get_lengths().size() != 2)
{
throw std::runtime_error("Host tensor is not rank 2 tensor.");
}
int n_ = t.get_lengths()[0];
int bqk_ = t.get_lengths()[1];
if(bqk_ % block_bq_k != 0)
{
throw std::runtime_error("shuffle_aq needs a bqk of multiple times of block_bq_k.");
}
ck_tile::HostTensor<T> t_view({n_, bqk_ / block_bq_k, block_bq_k});
std::copy(t.begin(), t.end(), t_view.begin());
return ck_tile::reference_permute(t_view, {1, 0, 2});
}
template <typename GemmConfig,
typename ADataType,
typename BDataType,
typename BQDataType,
typename DsDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename BQLayout,
typename DsLayout,
typename CLayout,
uint32_t QuantGroupSize,
typename CDEElementWise = ck_tile::element_wise::PassThrough>
float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
ck_tile::DeviceMem& b_k_n_dev_buf,
ck_tile::DeviceMem& bq_bqk_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 BQK,
ck_tile::index_t stride_A,
ck_tile::index_t stride_B,
ck_tile::index_t stride_BQ,
ck_tile::index_t stride_C,
ck_tile::index_t kbatch,
int n_warmup,
int n_repeat)
{
ck_tile::BQuantGemmHostArgs args;
args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer();
args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer();
args.bq_ptr = bq_bqk_n_dev_buf.GetDeviceBuffer();
args.c_ptr = c_m_n_dev_buf.GetDeviceBuffer();
args.k_batch = kbatch;
args.M = M;
args.N = N;
args.K = K;
args.QK = BQK;
args.stride_A = stride_A;
args.stride_B = stride_B;
args.stride_C = stride_C;
args.stride_BQ = stride_BQ;
float ave_time = gemm_calc_bquant<GemmConfig,
ADataType,
BDataType,
BQDataType,
AccDataType,
CDataType,
ADataType, // computeDatatype
ALayout,
BLayout,
CLayout,
QuantGroupSize>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
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(BQDataType) * BQK * N + 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 Gemm kernel with M =" << M << " N =" << N << " K =" << K
<< " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideBQ =" << stride_BQ
<< " StrideC =" << stride_C << " A_Layout =" << ALayout::name
<< " B_Layout =" << BLayout::name << " C_Layout =" << CLayout::name
<< " A_Type = " << DataTypeTraits<ADataType>::name
<< " B_Type = " << DataTypeTraits<BDataType>::name
<< " BQ_Type = " << DataTypeTraits<BQDataType>::name
<< " Acc_Type = " << DataTypeTraits<AccDataType>::name
<< " C_Type = " << DataTypeTraits<CDataType>::name << " : " << ave_time << " ms, "
<< tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
return ave_time;
}
template <typename GemmConfig,
typename TypeConfig,
uint32_t QuantGroupSize,
typename ALayout,
typename BLayout,
typename BQLayout,
typename CLayout>
int run_gemm_example_with_layouts(int argc,
char* argv[],
const ALayout a_layout = ALayout{},
const BLayout b_layout = BLayout{},
const BQLayout bq_layout = BQLayout{},
[[maybe_unused]] const CLayout c_layout = CLayout{})
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using ADataType = typename TypeConfig::ADataType;
using BDataType = typename TypeConfig::BDataType;
using BQDataType = typename TypeConfig::QDataType;
using AccDataType = typename TypeConfig::AccDataType;
using CDataType = typename TypeConfig::CDataType;
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");
if(K % QuantGroupSize != 0)
{
throw std::runtime_error("K must be aligned with QuantGroupSize");
}
ck_tile::index_t BQK = K / QuantGroupSize;
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_BQ = arg_parser.get_int("stride_q");
ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
ck_tile::index_t kbatch = arg_parser.get_int("split_k");
int n_warmup = arg_parser.get_int("warmup");
int n_repeat = arg_parser.get_int("repeat");
ck_tile::index_t init_method = arg_parser.get_int("init");
stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
stride_BQ = ck_tile::get_default_stride(BQK, N, stride_BQ, is_row_major(bq_layout));
stride_C = ck_tile::get_default_stride(M, N, stride_C, is_row_major(CLayout{}));
ck_tile::HostTensor<ADataType> a_m_k(
ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
ck_tile::HostTensor<BDataType> b_k_n(
ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
ck_tile::HostTensor<BQDataType> bq_bqk_n(
ck_tile::host_tensor_descriptor(BQK, N, stride_BQ, is_row_major(bq_layout)));
ck_tile::HostTensor<CDataType> c_m_n_dev_result(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<std::uint32_t> fill_seed(0, 500);
if(init_method == 0)
{
if constexpr(std::is_same_v<BDataType, ck_tile::pk_int4_t>)
{
ck_tile::FillUniformDistribution<ck_tile::pk_int4_t>{-5.0f, 5.0f, fill_seed(gen)}(
b_k_n);
}
else
{
ck_tile::FillUniformDistribution<BDataType>{-2.0f, 3.0f, fill_seed(gen)}(b_k_n);
}
ck_tile::FillUniformDistribution<BQDataType>{-2.0f, 2.0f, fill_seed(gen)}(bq_bqk_n);
ck_tile::FillUniformDistribution<ADataType>{-5.0f, 5.0f, fill_seed(gen)}(a_m_k);
}
else if(init_method == 1)
{
std::cout << "Monotonic initialization is not supported." << std::endl;
return 0;
}
else if(init_method == 2)
{
ck_tile::FillConstant<ADataType>{static_cast<ADataType>(0x38)}(a_m_k);
ck_tile::FillConstant<BDataType>{static_cast<BDataType>(0x22)}(b_k_n);
ck_tile::FillConstant<BQDataType>{static_cast<BQDataType>(0.5f)}(bq_bqk_n);
}
else
{
a_m_k.SetZero();
b_k_n.SetZero();
bq_bqk_n.SetZero();
}
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 bq_bqk_n_dev_buf(bq_bqk_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());
bq_bqk_n_dev_buf.ToDevice(bq_bqk_n.data());
c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero();
invoke_gemm<GemmConfig,
ADataType,
BDataType,
BQDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
BQLayout,
ck_tile::tuple<>,
CLayout,
QuantGroupSize>(a_m_k_dev_buf,
b_k_n_dev_buf,
bq_bqk_n_dev_buf,
c_m_n_dev_buf,
M,
N,
K,
BQK,
stride_A,
stride_B,
stride_BQ,
stride_C,
kbatch,
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(
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
c_m_n_host_ref.SetZero();
ck_tile::reference_gemm_quant<ADataType,
BQDataType,
BDataType,
AccDataType,
CDataType,
QuantGroupSize,
false>(a_m_k, bq_bqk_n, b_k_n, c_m_n_host_ref);
const float max_accumulated_value =
*std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_host_ref,
"Error: Incorrect results!",
rtol_atol.at(ck_tile::number<0>{}),
rtol_atol.at(ck_tile::number<1>{}));
if(!pass)
{
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
<< std::endl;
}
std::cout << "CPU verification " << (pass ? "Passed!" : "Failed ...") << std::endl;
}
else if(arg_parser.get_int("v") == 2)
{
std::cout << "GPU verification is not implemented yet. Re-run with -v=1" << std::endl;
return false;
}
return pass;
}

2
example/ck_tile/39_copy/README.md
View File

@@ -12,7 +12,7 @@ This experimental kernel is intended for novice CK developers. It introduces the
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>
../script/cmake-ck-dev.sh ../ <arch>
# Make the copy kernel executable
make tile_example_copy -j
```

24
example/ck_tile/39_copy/copy_basic.cpp
View File

@@ -77,10 +77,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
// we intentionally do not use pipeline for this example and let the kernel be composite of
// Problem and Policy
constexpr ck_tile::index_t kBlockSize = Shape::BlockSize;
auto blockSize = Kernel::BlockSize();
// Print configuration information
std::cout << "block size (number of threads per block) " << kBlockSize << std::endl;
std::cout << "block size (number of threads per block) " << blockSize << std::endl;
std::cout << "wave size (number of threads per wave) " << ck_tile::get_warp_size() << std::endl;
std::cout << "block waves (number of waves per block) " << BlockWaves::at(ck_tile::number<0>{})
<< " " << BlockWaves::at(ck_tile::number<1>{}) << std::endl;
@@ -99,16 +99,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
<< ")" << std::endl;
// Launch kernel
float ave_time = launch_kernel(
ck_tile::stream_config{nullptr, true, warmup, repeat, 1},
ck_tile::make_kernel<kBlockSize, 1>(Kernel{},
kGridSize,
kBlockSize,
0,
static_cast<XDataType*>(x_buf.GetDeviceBuffer()),
static_cast<YDataType*>(y_buf.GetDeviceBuffer()),
m,
n));
float ave_time =
launch_kernel(ck_tile::stream_config{nullptr, true, warmup, repeat, 1},
ck_tile::make_kernel<1>(Kernel{},
kGridSize,
blockSize,
0,
static_cast<XDataType*>(x_buf.GetDeviceBuffer()),
static_cast<YDataType*>(y_buf.GetDeviceBuffer()),
m,
n));
// Calculate and print performance metrics
std::size_t num_btype = sizeof(XDataType) * m * n + sizeof(YDataType) * m * n;

19
example/ck_tile/39_copy/copy_basic.hpp
View File

@@ -27,8 +27,9 @@ struct TileCopyShape
static constexpr index_t ThreadTile_N = ThreadTile::at(number<1>{});
// Wave tile dimensions
static constexpr index_t Wave_Tile_M = WaveTile::at(number<0>{});
static constexpr index_t WaveSize = get_warp_size();
static constexpr index_t Wave_Tile_N = WaveTile::at(number<1>{});
static constexpr index_t Wave_Tile_M = ThreadTile_M * ThreadTile_N * WaveSize / Wave_Tile_N;
// Block tile dimensions
static constexpr index_t Block_Tile_M = BlockTile::at(number<0>{});
@@ -45,7 +46,6 @@ struct TileCopyShape
Block_Tile_N / (Waves_Per_Block_N * Wave_Tile_N);
// Hardware configuration
static constexpr index_t WaveSize = get_warp_size();
static constexpr index_t BlockSize = Waves_Per_Block_M * Waves_Per_Block_N * WaveSize;
// Configuration validation
@@ -60,8 +60,10 @@ struct TileCopyShape
"Invalid wave configuration for N dimension");
// Ensure wave tile dimensions align with wave size
#if defined(__HIP_DEVICE_COMPILE__)
static_assert(Wave_Tile_M / ThreadTile_M * Wave_Tile_N / ThreadTile_N == WaveSize,
"(Wave_Tile_M/ThreadTile_M) * (Wave_Tile_N/ThreadTile_N) != WaveSize");
#endif
};
/**
@@ -200,6 +202,19 @@ struct ElementWiseTileCopyKernel
using XDataType = typename Problem::XDataType;
using Policy = ck_tile::remove_cvref_t<Policy_>;
static constexpr index_t kBlockSize = Problem::BlockShape::BlockSize;
CK_TILE_HOST static auto BlockSize()
{
if(ck_tile::is_wave32())
{
return kBlockSize / 2;
}
else
{
return kBlockSize;
}
}
CK_TILE_DEVICE void operator()(const XDataType* p_x, XDataType* p_y, index_t M, index_t N) const
{
using S = typename Problem::BlockShape;