Skip to content

GitLab

Commit 175a17f8 authored by Rostyslav Geyyer's avatar Rostyslav Geyyer
Browse files

Merge branch 'gfx950' of https://github.com/ROCm/composable_kernel-internal into lwpck-2390

parents 3e520bbd 1504c3e8
Hide whitespace changes
Inline Side-by-side
Showing with 1622 additions and 112 deletions
example/62_convnd_activ/multi_AB/CMakeLists.txt
View file @ 175a17f8
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/62_convnd_activ/unary/CMakeLists.txt
View file @ 175a17f8
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
......
example/CMakeLists.txt
View file @ 175a17f8
......@@ -54,6 +54,13 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any DPP examples if DL_KERNELS not set
foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dpp")
message("removing dpp example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any XDL examples if gfx9 targets are not on the list
foreach(source IN LISTS FILE_NAME)
if(NOT EX_TARGETS MATCHES "gfx9" AND source MATCHES "_xdl")
......@@ -85,7 +92,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
#only continue if there are some source files left on the list
if(FILE_NAME)
if(FILE_NAME MATCHES "_xdl")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(FILE_NAME MATCHES "_wmma")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx950)
endif()
......@@ -169,7 +176,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
#only continue if there are some source files left on the list
if(FILE_NAME)
if(FILE_NAME MATCHES "_xdl")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(FILE_NAME MATCHES "_wmma")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx950)
endif()
......
example/ck_tile/01_fmha/script/smoke_test_fwd.sh
View file @ 175a17f8
......@@ -29,14 +29,14 @@ while getopts ":sa" opt; do
done
run_fp16_bf16_tests() {
local NUM_SPLITS=(1)
local PAGE_BLOCK_SIZE=(0)
local CACHE_BATCH_IDX=(0)
local NUM_SPLITS="1"
local PAGE_BLOCK_SIZE="0"
local CACHE_BATCH_IDX="0"
if [ $TEST_SPLITKV -eq 1 ] ; then
NUM_SPLITS+=(2 3)
PAGE_BLOCK_SIZE+=(128)
CACHE_BATCH_IDX+=(1)
NUM_SPLITS="$NUM_SPLITS 2 3"
PAGE_BLOCK_SIZE="$PAGE_BLOCK_SIZE 128"
CACHE_BATCH_IDX="$CACHE_BATCH_IDX 1"
fi
for prec in "fp16" "bf16" ; do
......@@ -47,9 +47,9 @@ run_fp16_bf16_tests() {
for lse in 0 1 ; do
for bias in "n" "e" "a" ; do
for p_drop in 0.0 0.2 ; do
for num_splits in "${NUM_SPLITS[@]}" ; do
for page_block_size in "${PAGE_BLOCK_SIZE[@]}" ; do
for cache_batch_idx in "${CACHE_BATCH_IDX[@]}" ; do
for num_splits in $NUM_SPLITS ; do
for page_block_size in $PAGE_BLOCK_SIZE ; do
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=$perm -vlayout=$vlayout -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=$perm -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS
......@@ -103,4 +103,4 @@ if [ $TEST_APPENDKV -eq 1 ] ; then
run_fp16_appendkv_tests
fi
set +x
\ No newline at end of file
set +x
example/ck_tile/02_layernorm2d/generate.py
View file @ 175a17f8
......@@ -57,6 +57,7 @@ template <typename XDataType_,
ck_tile::index_t Vector_N_, // vector size along N
bool kPadN_,
bool kSaveMeanInvStd_,
bool kFastFDiv_,
bool kTwoPass_,
ck_tile::index_t kFusedAdd_ = 0,
ck_tile::index_t kFusedQuant_ = 0>
......@@ -118,6 +119,7 @@ struct layernorm2d_fwd_traits_
static constexpr bool kPadN = kPadN_;
static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_;
static constexpr bool kFastFDiv = kFastFDiv_;
static constexpr bool kTwoPass = kTwoPass_;
static constexpr ck_tile::index_t kFusedAdd = kFusedAdd_;
static constexpr ck_tile::index_t kFusedQuant = kFusedQuant_;
......@@ -134,6 +136,7 @@ template <typename XDataType_,
ck_tile::index_t Vector_N_, // vector size along N
bool kPadN_,
bool kSaveMeanInvStd_,
bool kFastFDiv_,
bool kTwoPass_,
int kFusedAdd_,
int kFusedQuant_>
......@@ -148,6 +151,7 @@ using traits_ = layernorm2d_fwd_traits_<XDataType_,
Vector_N_,
kPadN_,
kSaveMeanInvStd_,
kFastFDiv_,
kTwoPass_,
kFusedAdd_,
kFusedQuant_>;
......@@ -179,6 +183,7 @@ float layernorm2d_fwd_(const S& s, A a)
using PipelineTraits = ck_tile::Layernorm2dFwdTraits<Traits_::kPadN,
Traits_::kSaveMeanInvStd,
Traits_::kFastFDiv,
Traits_::kTwoPass,
static_cast<ck_tile::Layernorm2dFusedAddEnum>(Traits_::kFusedAdd),
static_cast<ck_tile::Layernorm2dFusedQuantEnum>(Traits_::kFusedQuant)>;
......@@ -269,7 +274,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
#include "layernorm2d_fwd_api_common.hpp"
// clang-format off
// prec_i prec_o prec_sy rm rn tm tn vn pd mv 2p add sweep
// prec_i prec_o prec_sy rm rn tm tn vn pd mv rpcf 2p add sweep
{F_instance_def}
// clang-format on
......@@ -356,6 +361,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
F_Vector_N : int
F_kPadN : bool
F_kSaveMeanInvStd_ : bool
F_kFastFDiv_ : bool
F_kTwoPass_ : bool
F_kFusedAdd : int
F_kFusedQuant : int
......@@ -363,7 +369,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
@property
def trait_name(self) ->str:
t_ = f'{DATA_TYPE_MAP[self.F_XDataType]}, {DATA_TYPE_MAP[self.F_YDataType]}, {DATA_TYPE_MAP[self.F_XScaleDataType]}, {DATA_TYPE_MAP[self.F_YScaleDataType]}, {self.F_Repeat_M:2}, {self.F_Repeat_N:2}, {self.F_ThreadPerBlock_M:2}, {self.F_ThreadPerBlock_N:4}'
t_ += f', {self.F_Vector_N:2}, {BOOL_MAP(self.F_kPadN):5}, {BOOL_MAP(self.F_kSaveMeanInvStd_):5}'
t_ += f', {self.F_Vector_N:2}, {BOOL_MAP(self.F_kPadN):5}, {BOOL_MAP(self.F_kSaveMeanInvStd_):5}, {BOOL_MAP(self.F_kFastFDiv_):5}'
t_ += f', {BOOL_MAP(self.F_kTwoPass_):5}, {self.F_kFusedAdd:4}, {self.F_kFusedQuant:4}'
return t_
......@@ -483,52 +489,55 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
fused_add_list = [0, 1]
fused_sweep_list = [0, 1] # NOTE: only single pass can use fused dynamic quant
# rm rn tm tn vn pd mv 2p add sweep
h_trait_dict = {'64' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 1, True, False, False, 0, 0)],
'128' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 1, True, False, False, 0, 0)],
'256' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 4, 64, 1, True, False, False, 0, 0)],
'512' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 8, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 4, 64, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 4, 64, 1, True, False, False, 0, 0)],
'768' : [ h_traits('x', 'y', 'xs', 'ys', 1, 3, 4, 64, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 4, 64, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 12, 4, 64, 1, True, False, False, 0, 0)],
'1024' :[ h_traits('x', 'y', 'xs', 'ys', 1, 1, 2, 128, 8, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 2, 128, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 2, 128, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 1, True, False, False, 0, 0)],
'1536' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 4, 64, 8, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 2, 128, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1, 256, 1, True, False, False, 0, 0)],
'2048' :[ h_traits('x', 'y', 'xs', 'ys', 1, 1, 1, 256, 8, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 1, 256, 1, True, False, False, 0, 0)],
'3072' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 128, 8, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1, 256, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 1, True, False, False, 0, 0)],
'4096' :[ h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 8, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1,1024, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 1, True, False, False, 0, 0)],
'6144' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 8, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 512, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1,1024, 1, True, False, False, 0, 0)],
'8192' :[ h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 8, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 512, 4, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 2, True, False, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 1,1024, 1, True, False, False, 0, 0)],
'big' :[ h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 8, True, False, True, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 4, True, False, True, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1,1024, 2, True, False, True, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 1, True, False, True, 0, 0)]}
# rm rn tm tn vn pd mv fdiv 2p add sweep
h_trait_dict = {'64' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 8, 8, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 16, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 1, True, False, True, False, 0, 0)],
'128' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 16, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 1, True, False, True, False, 0, 0)],
'256' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 4, 64, 1, True, False, True, False, 0, 0)],
'512' : [ h_traits('x', 'y', 'xs', 'ys', 1, 1, 4, 64, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 4, 64, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 4, 64, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 4, 64, 1, True, False, True, False, 0, 0)],
'768' : [ h_traits('x', 'y', 'xs', 'ys', 1, 3, 4, 64, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 4, 64, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 12, 4, 64, 1, True, False, True, False, 0, 0)],
'1024' :[ h_traits('x', 'y', 'xs', 'ys', 1, 1, 2, 128, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 2, 128, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 2, 128, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 1, True, False, True, False, 0, 0)],
'1536' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 4, 64, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 2, 128, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1, 256, 1, True, False, True, False, 0, 0)],
'2048' :[ h_traits('x', 'y', 'xs', 'ys', 1, 1, 1, 256, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 1, 256, 1, True, False, True, False, 0, 0)],
'3072' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 128, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1, 256, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 1, True, False, True, False, 0, 0)],
'4096' :[ h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1,1024, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 1, True, False, True, False, 0, 0)],
'6144' :[ h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 512, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 6, 1,1024, 1, True, False, True, False, 0, 0)],
'8192' :[ h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 8, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 512, 4, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 2, True, False, True, False, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 8, 1,1024, 1, True, False, True, False, 0, 0)],
'big' :[ h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 8, True, False, True, True, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 4, True, False, True, True, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 2, 1,1024, 2, True, False, True, True, 0, 0),
h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 1, True, False, True, True, 0, 0)]}
total_blob = list()
for hs_key in h_trait_dict:
hs = h_trait_dict[hs_key]
......
example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp
View file @ 175a17f8
......@@ -25,7 +25,10 @@ auto create_args(int argc, char* argv[])
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension")
.insert("stride", "-1", "stride per row, if -1 then equal to n")
.insert("x_stride", "-1", "x row_stride, if -1 then equal to n")
.insert("xr_stride", "-1", "x residule row_stride, if -1 then equal to n")
.insert("y_stride", "-1", "y row_stride, if -1 then equal to n")
.insert("yr_stride", "-1", "y residule row_stride, if -1 then equal to n")
.insert("e", "1e-5", "epsilon")
.insert("save_mv", "0", "save mean/variance(invstd) or not. set to 1 in training case")
.insert("v", "1", "cpu validation or not")
......@@ -54,11 +57,20 @@ template <typename InDataType,
bool SaveMeanVar>
bool run(const ck_tile::ArgParser& arg_parser)
{
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t stride = arg_parser.get_int("stride");
if(stride < 0)
stride = n;
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
if(x_stride < 0)
x_stride = n;
ck_tile::index_t xr_stride = arg_parser.get_int("xr_stride");
if(xr_stride < 0)
xr_stride = n;
ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
if(y_stride < 0)
y_stride = n;
ck_tile::index_t yr_stride = arg_parser.get_int("yr_stride");
if(yr_stride < 0)
yr_stride = n;
float epsilon = arg_parser.get_float("e");
std::string prec_i = arg_parser.get_str("prec_i");
std::string prec_o = arg_parser.get_str("prec_o");
......@@ -89,7 +101,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
return false;
}
assert(stride >= n);
assert(x_stride >= n);
using TypeConfig = LayerNormTypeConfig<InDataType, OutDataType, XScaleDataType, YScaleDataType>;
......@@ -108,15 +120,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
using ComputeDataType = typename TypeConfig::ComputeDataType;
// host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1});
ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<GammaDataType> gamma_host({n});
ck_tile::HostTensor<BetaDataType> beta_host({n});
ck_tile::HostTensor<XResidualDataType> x_residual_host({m, n}, {stride, 1});
ck_tile::HostTensor<YResidualDataType> y_residual_host({m, n}, {stride, 1});
ck_tile::HostTensor<XResidualDataType> x_residual_host({m, n}, {xr_stride, 1});
ck_tile::HostTensor<YResidualDataType> y_residual_host({m, n}, {yr_stride, 1});
ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {stride, 1});
ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {stride, 1});
ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {y_stride, 1});
ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {y_stride, 1});
ck_tile::HostTensor<MeanDataType> mean_host_ref({m});
ck_tile::HostTensor<InvStdDataType> invStd_host_ref({m});
......@@ -162,7 +174,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
}();
std::cout << "[" << prec_str << "]"
<< " m:" << m << ", n:" << n << ", stride:" << stride << std::flush;
<< " m:" << m << ", n:" << n << ", x_stride:" << x_stride
<< ", xr_stride:" << xr_stride << ", y_stride:" << y_stride
<< ", yr_stride:" << yr_stride << std::flush;
layernorm2d_fwd_traits traits{
prec_i, prec_o, prec_sx, prec_sy, SaveMeanVar, fused_add, fused_quant};
......@@ -182,7 +196,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
epsilon,
m,
n,
stride};
x_stride, // x row_stride
xr_stride, // x residule row stride
y_stride, // y row stride
yr_stride}; // y residule row stride
float ave_time = layernorm2d_fwd(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
......@@ -285,7 +302,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
y_buf.FromDevice(y_host_dev.data());
ck_tile::HostTensor<YResidualDataType> y_residual_host_dev({m, n}, {stride, 1});
ck_tile::HostTensor<YResidualDataType> y_residual_host_dev({m, n}, {yr_stride, 1});
if(fused_add == 1)
{
y_residual_buf.FromDevice(y_residual_host_dev.data());
......@@ -293,7 +310,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
auto [rtol, atol] = get_elimit<InDataType>();
if(stride == n)
if(x_stride == n)
{
pass = ck_tile::check_err(
y_host_dev, y_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol);
......@@ -310,10 +327,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
{
for(int i_r = 0; i_r < m; i_r++)
{
std::vector<YDataType> y_host_dev_row(y_host_dev.begin() + i_r * stride,
y_host_dev.begin() + i_r * stride + n);
std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * stride,
y_host_ref.begin() + i_r * stride + n);
std::vector<YDataType> y_host_dev_row(y_host_dev.begin() + i_r * y_stride,
y_host_dev.begin() + i_r * y_stride + n);
std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * y_stride,
y_host_ref.begin() + i_r * y_stride + n);
pass &= ck_tile::check_err(y_host_dev_row,
y_host_ref_row,
std::string("OUT[") + std::to_string(i_r) +
......@@ -323,10 +340,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(fused_add == 1)
{
std::vector<YResidualDataType> y_residual_host_dev_row(
y_residual_host_dev.begin() + i_r * stride,
y_residual_host_dev.begin() + i_r * stride + n);
y_residual_host_dev.begin() + i_r * yr_stride,
y_residual_host_dev.begin() + i_r * yr_stride + n);
std::vector<YResidualDataType> y_residual_host_ref_row(
x_host.begin() + i_r * stride, x_host.begin() + i_r * stride + n);
x_host.begin() + i_r * yr_stride, x_host.begin() + i_r * yr_stride + n);
pass &= ck_tile::check_err(y_residual_host_dev_row,
y_residual_host_ref_row,
std::string("ADD[") + std::to_string(i_r) +
......
example/ck_tile/03_gemm/README.md
View file @ 175a17f8
......@@ -8,7 +8,10 @@ This folder contains example for GEMM using ck_tile tile-programming implementat
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>
# The basic pipeline method on the gemm calculation
make tile_example_gemm_basic -j
# The memory bound pipeline on the gemm calculation
make tile_example_gemm_mem_pipeline -j
```
This will result in an executable `build/bin/tile_example_gemm_basic`
......
example/ck_tile/03_gemm/gemm_basic.cpp
View file @ 175a17f8
......@@ -17,10 +17,11 @@
template <typename ALayout, typename BLayout, typename CLayout>
float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
{
// The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadA = true;
constexpr bool kPadB = true;
constexpr bool kPadC = true;
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr bool kTilePermute = false;
// The rank and permutation will also be generate out by the CodeGen part.
constexpr ck_tile::index_t kOutputRank = 2;
......@@ -56,8 +57,8 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
CShuffleEpilogue,
ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
kPadA,
kPadB,
kPadM,
kPadN,
kTilePermute,
kOutputRank,
1,
......@@ -65,13 +66,13 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
TilePartitioner::kM,
TilePartitioner::kN>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadA, kPadB>>>;
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadA, kPadB, kPadC, ALayout, BLayout, CLayout>;
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy<ALayout, BLayout, CLayout>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy;
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenGemmPolicy>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
......
example/ck_tile/03_gemm/gemm_mem_pipeline.cpp
View file @ 175a17f8
......@@ -31,9 +31,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
constexpr ck_tile::index_t K_Warp_Tile = 8;
// The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadA = true;
constexpr bool kPadB = true;
constexpr bool kPadC = true;
constexpr bool kPadM = true;
constexpr bool kPadN = true;
constexpr bool kPadK = true;
constexpr int kBlockPerCu = 1;
......@@ -46,9 +46,9 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
using TilePartitioner = ck_tile::GemmTilePartitioner<GemmShape>;
using GemmEpilogue = ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, false, kPadC>>;
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>;
using Traits = ck_tile::TileGemmTraits<kPadA, kPadB, kPadC, ALayout, BLayout, CLayout>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>;
......
example/ck_tile/13_moe_sorting/CMakeLists.txt 0 → 100644
View file @ 175a17f8
add_executable(tile_example_moe_sorting EXCLUDE_FROM_ALL moe_sorting.cpp moe_sorting_api.cpp)
target_include_directories(tile_example_moe_sorting PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/)
set(EXAMPLE_MOE_SORTING_COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND EXAMPLE_MOE_SORTING_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
# list(APPEND EXAMPLE_MOE_SORTING_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker)
target_compile_options(tile_example_moe_sorting PRIVATE ${EXAMPLE_MOE_SORTING_COMPILE_OPTIONS})
example/ck_tile/13_moe_sorting/README.md 0 → 100644
View file @ 175a17f8
# moe-sorting
This folder contains example for moe-sorting kernel using ck_tile tile-programming implementation. This kernel is often used in Moe model, before launching the fused-moe-gemm block. The input&weight is a `token*topk` 2d matrix. The op rearange the input weight ids into different experts and feed into fuse moe gemm kernel.
## build
```
# in the root of ck_tile
mkdir build && cd build
sh ../script/cmake-ck-dev.sh ../ <arch> # you can replace this <arch> to gfx90a, gfx942...
make tile_example_moe_sorting -j
```
This will result in an executable `build/bin/tile_example_moe_sorting`
## example
```
args:
-v weather do CPU validation or not (default:1)
-pr_i index data type. (currently only fp32 supported now) (default:int32)
-pr_w output weight data type(currently only fp32 supported now) (default:fp32)
-t number of input tokens (default:32)
-e number of experts (default:8)
-k topk (default:2)
-st_i row stride of input, -1 means same as experts (default:-1)
-seed seed to be used, -1 means random every time (default:-1)
-kname when set to 1 it will print kernel name (default:0)
```
example/ck_tile/13_moe_sorting/moe_sorting.cpp 0 → 100644
View file @ 175a17f8
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <set>
#include <vector>
#include <iostream>
#include <numeric>
#include <cassert>
#include <cstdlib>
#include <iostream>
#include <time.h>
#include <unordered_set>
#include "ck_tile/core.hpp"
#include "ck_tile/ops/reduce.hpp"
#include "moe_sorting_api.hpp"
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("v", "1", "weather do CPU validation or not")
.insert("pr_i", "int32", "index data type. (currently only int32 supported now)")
.insert("pr_w", "fp32", "output weight data type(currently only fp32 supported now)")
.insert("t", "128", "number of input tokens")
.insert("e", "8", "number of num_experts")
.insert("k", "4", "topk")
.insert("unit", "32", "unit_size")
.insert("moe_buf_size", "0", "moe_buf_size")
.insert("seed", "-1", "seed to be used, -1 means random every time")
.insert("kname", "0", "when set to 1 it will print kernel name")
.insert("warmup", "5", "number of iterations before benchmark the kernel")
.insert("repeat", "20", "number of iterations to benchmark the kernel");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
template <typename IndexType>
void topid_unique_gen(
std::vector<IndexType>& host_tensor, int tokens, int topk, int num_expert, int seed)
{
size_t total_size = topk * tokens;
std::srand(seed);
std::set<IndexType> unique_set;
IndexType current_v;
for(size_t i = 0; i < total_size; i++)
{
if(i % topk == 0)
{
unique_set.clear();
}
current_v = std::rand() % num_expert;
while(unique_set.find(current_v) != unique_set.end())
{
current_v = std::rand() % num_expert;
}
unique_set.insert(current_v);
host_tensor[i] = current_v;
}
}
template <typename WeightType, typename IndexType = ck_tile::index_t>
bool test_moe_sorting(ck_tile::ArgParser args)
{
int validate = args.get_int("v");
std::string index_prec = args.get_str("pr_i");
std::string weight_prec = args.get_str("pr_w");
int tokens = args.get_int("t");
int num_experts = args.get_int("e");
int topk = args.get_int("k");
int seed = args.get_int("seed");
int unit_size = args.get_int("unit");
int moe_buf_size = args.get_int("moe_buf_size");
int kname = args.get_int("kname");
int warmup = args.get_int("warmup");
int repeat = args.get_int("repeat");
int max_output_ids =
ck_tile::integer_least_multiple(topk * tokens + num_experts * unit_size - topk, unit_size);
if(seed < 0)
{
seed = std::time(nullptr);
}
if(topk > num_experts)
{
printf("topk:%d value should be smaller than, or equal to number of num_experts:%d\n",
topk,
num_experts);
return false;
}
// tokens already considered batch size
ck_tile::HostTensor<IndexType> topk_ids_host({tokens, topk}, {topk, 1});
ck_tile::HostTensor<WeightType> weights_host({tokens, topk}, {topk, 1});
ck_tile::HostTensor<IndexType> sorted_ids_host({max_output_ids}, {1});
ck_tile::HostTensor<WeightType> sorted_weights_host({max_output_ids}, {1});
ck_tile::HostTensor<IndexType> sorted_expert_ids_host({max_output_ids / unit_size}, {1});
ck_tile::HostTensor<IndexType> sorted_id_cnt_host({1}, {1});
ck_tile::HostTensor<float> moe_buf_host({moe_buf_size});
ck_tile::FillUniformDistribution<WeightType>{-.5f, .5f}(weights_host);
ck_tile::FillUniformDistribution<WeightType>{-.5f, .5f}(moe_buf_host);
topid_unique_gen<IndexType>(topk_ids_host.mData, tokens, topk, num_experts, seed);
ck_tile::DeviceMem topk_ids_dev(topk_ids_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem weights_dev(weights_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem sorted_ids_dev(sorted_ids_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem sorted_weights_dev(sorted_weights_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem sorted_expert_ids_dev(
sorted_expert_ids_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem sorted_id_cnt_dev(sorted_id_cnt_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem moe_buf_dev(moe_buf_host.get_element_space_size_in_bytes());
topk_ids_dev.ToDevice(topk_ids_host.data());
weights_dev.ToDevice(weights_host.data());
if(moe_buf_size > 0)
{
moe_buf_dev.ToDevice(moe_buf_host.data());
}
moe_sorting_trait trait{index_prec, weight_prec};
moe_sorting_args karg{topk_ids_dev.GetDeviceBuffer(),
weights_dev.GetDeviceBuffer(),
sorted_ids_dev.GetDeviceBuffer(),
sorted_weights_dev.GetDeviceBuffer(),
sorted_expert_ids_dev.GetDeviceBuffer(),
sorted_id_cnt_dev.GetDeviceBuffer(),
moe_buf_size > 0 ? moe_buf_dev.GetDeviceBuffer() : nullptr,
tokens,
unit_size,
num_experts,
topk,
static_cast<ck_tile::index_t>(moe_buf_size * sizeof(float))};
ck_tile::stream_config sc{nullptr,
true,
/* log_level = */ (kname ? 1 : 0),
warmup,
repeat};
auto ms = moe_sorting(trait, karg, sc);
printf("[%s|%s]tokens:%d, num_experts:%d, topk:%d, ms:%f , ",
index_prec.c_str(),
weight_prec.c_str(),
tokens,
num_experts,
topk,
ms);
if(ms < 0)
printf("not supported\n");
fflush(stdout);
if(ms < 0)
{
return false;
}
sorted_ids_dev.FromDevice(sorted_ids_host.data());
sorted_weights_dev.FromDevice(sorted_weights_host.data());
sorted_expert_ids_dev.FromDevice(sorted_expert_ids_host.data());
sorted_id_cnt_dev.FromDevice(sorted_id_cnt_host.data());
if(moe_buf_size > 0)
{
moe_buf_dev.FromDevice(moe_buf_host.data());
}
bool rtn = true;
if(validate)
{
ck_tile::HostTensor<IndexType> sorted_ids_ref({max_output_ids}, {1});
ck_tile::HostTensor<WeightType> sorted_weights_ref({max_output_ids}, {1});
ck_tile::HostTensor<IndexType> sorted_expert_ids_ref({max_output_ids / unit_size}, {1});
int32_t ref_total_tokens_post_pad = 0;
ck_tile::reference_moe_sorting<WeightType, IndexType>(topk_ids_host,
weights_host,
sorted_ids_ref,
sorted_weights_ref,
sorted_expert_ids_ref,
ref_total_tokens_post_pad,
num_experts,
unit_size);
rtn &= ck_tile::check_err(
sorted_ids_host, sorted_ids_ref, std::string("OUT Error: Incorrect ids!"), 1e-6, 1e-6);
rtn &= ck_tile::check_err(sorted_weights_host,
sorted_weights_ref,
std::string("OUT Error: Incorrect w!"),
1e-6,
1e-6);
rtn &= ck_tile::check_err(sorted_expert_ids_host,
sorted_expert_ids_ref,
std::string("OUT Error: Incorrect eid!"),
1e-6,
1e-6);
if(moe_buf_size)
{
ck_tile::HostTensor<WeightType> moe_buf_ref({moe_buf_size});
rtn &= ck_tile::check_err(
moe_buf_host, moe_buf_ref, std::string("OUT Error: Incorrect zero buf!"), 0, 0);
}
rtn &= ref_total_tokens_post_pad == sorted_id_cnt_host.mData[0];
}
printf("valid:%s\n", rtn ? "y" : "n");
fflush(stdout);
return rtn;
}
int main(int argc, char** argv)
{
auto [result, args] = create_args(argc, argv);
if(!result)
return -1;
std::string index_prec = args.get_str("pr_i");
std::string weight_prec = args.get_str("pr_w");
bool r = true;
if(weight_prec.compare("fp32") == 0 && index_prec.compare("int32") == 0)
{
r &= test_moe_sorting<float, ck_tile::index_t>(args);
}
return r ? 0 : -1;
}
example/ck_tile/13_moe_sorting/moe_sorting_api.cpp 0 → 100644
View file @ 175a17f8
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "moe_sorting_api.hpp"
#define MOE_SORTING_DISPATCH(unroll_num_) \
constexpr ck_tile::index_t unroll_num = unroll_num_; \
using ms_problem = ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num>; \
using kernel = ck_tile::MoeSortingKernel<ms_problem>; \
auto kargs = kernel::MakeKargs(a); \
const dim3 grids = kernel::GridSize(a); \
const dim3 blocks = kernel::BlockSize(a); \
const auto lds_bytes = kernel::GetSmemSize(a); \
float ave_time = ck_tile::launch_kernel( \
s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \
return ave_time;
float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s)
{
if(t.weight_type == "fp32" && t.index_type == "int32")
{
if(a.num_experts > 127)
{
printf("lds size exceed, only support experts <127 \n");
return -1;
}
if(a.moe_buf_bytes % 16)
{
printf("buf set size %d unaligned, must be multiple of 16\n", a.moe_buf_bytes);
return -1;
}
using index_t = ck_tile::index_t;
using ms_weight_type = float;
index_t smem_io_unroll_num = ck_tile::integer_divide_ceil(a.tokens * a.topk, 64);
switch(smem_io_unroll_num)
{
case(1): {
MOE_SORTING_DISPATCH(1);
}
case(2): {
MOE_SORTING_DISPATCH(2);
}
case(3): {
MOE_SORTING_DISPATCH(3);
}
case(5): {
MOE_SORTING_DISPATCH(5);
}
case(6): {
MOE_SORTING_DISPATCH(6);
}
case(7): {
MOE_SORTING_DISPATCH(7);
}
case(8): {
MOE_SORTING_DISPATCH(8);
}
case(9): {
MOE_SORTING_DISPATCH(9);
}
case(10): {
MOE_SORTING_DISPATCH(10);
}
case(11): {
MOE_SORTING_DISPATCH(11);
}
default: {
MOE_SORTING_DISPATCH(4);
}
}
}
return -1;
}
example/ck_tile/13_moe_sorting/moe_sorting_api.hpp 0 → 100644
View file @ 175a17f8
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <string>
#include "ck_tile/core.hpp"
#include "ck_tile/host.hpp"
#include "ck_tile/ops/moe_sorting.hpp"
struct moe_sorting_trait
{
std::string index_type;
std::string weight_type; // currently always float
};
struct moe_sorting_args : public ck_tile::MoeSortingHostArgs
{
};
float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s);
example/ck_tile/13_moe_sorting/script/smoke_test.sh 0 → 100644
View file @ 175a17f8
# #!/bin/sh
EXE=./build/bin/tile_example_moe_sorting
$EXE -t=80 -e=17 -moe_buf_size=16
$EXE -t=111 -e=117 -moe_buf_size=4
$EXE -t=1000 -e=55 -moe_buf_size=1024
$EXE -t=99 -e=120 -moe_buf_size=10244
$EXE -t=175 -e=64 -k=8
$EXE -t=65 -e=8 -k=2
$EXE -t=1 -e=25
$EXE -t=31 -e=19 -k=15
$EXE -t=81 -e=37 -k=7
$EXE -t=23 -e=1 -k=1
$EXE -t=127 -e=99 -k=19
$EXE -t=71 -e=11 -k=11
$EXE -t=1 -e=1 -k=1
$EXE -t=99 -e=2 -k=1
$EXE -t=333 -e=99 -k=13
\ No newline at end of file
example/ck_tile/CMakeLists.txt
View file @ 175a17f8
......@@ -12,3 +12,4 @@ add_subdirectory(09_topk_softmax)
add_subdirectory(10_rmsnorm2d)
add_subdirectory(11_add_rmsnorm2d_rdquant)
add_subdirectory(12_smoothquant)
add_subdirectory(13_moe_sorting)
include/ck/ck.hpp
View file @ 175a17f8
......@@ -63,13 +63,15 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
#define __gfx101__
#endif
#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__)
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || \
defined(__gfx10_3_generic__)
#define __gfx103__
#endif
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__)
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || \
defined(__gfx1103__) || defined(__gfx11_generic__)
#define __gfx11__
#endif
#if defined(__gfx1200__) || defined(__gfx1201__)
#if defined(__gfx1200__) || defined(__gfx1201__) || defined(__gfx12_generic__)
#define __gfx12__
#endif
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
View file @ 175a17f8
......@@ -381,10 +381,6 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
{
tildes = {i_ztilde, i_ytilde, i_xtilde};
}
else
{
throw std::runtime_error("wrong! only implemented for 2D and 3D now");
}
const auto a_grid_desc_ak0_m_ak1 =
transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
......@@ -750,6 +746,12 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
}
}
// check number of dimension, only implemented for 2D and 3D now
if(NDimSpatial != 2 && NDimSpatial != 3)
{
return false;
}
return true;
}
......
include/ck/utility/amd_buffer_addressing.hpp
View file @ 175a17f8
......@@ -549,8 +549,10 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
(is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, bhalf_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, f8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, bf8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, f8_fnuz_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, bf8_fnuz_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, fp8_storage_t>::value &&
(N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
"wrong! not implemented");
......@@ -843,8 +845,8 @@ amd_buffer_load_invalid_element_return_zero(const T* p_src_wave,
#else
vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size, coherence>(
src_wave_buffer_resource, src_thread_addr_offset, 0);
vector_t tmp{amd_buffer_load_impl<scalar_t, vector_size, coherence>(
src_wave_buffer_resource, src_thread_addr_offset, 0)};
return src_thread_element_valid ? tmp : vector_t(0);
#endif
}
......@@ -873,8 +875,8 @@ amd_buffer_load_invalid_element_return_customized_value(const T* p_src_wave,
constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size, coherence>(
src_wave_buffer_resource, src_thread_addr_offset, 0);
vector_t tmp{amd_buffer_load_impl<scalar_t, vector_size, coherence>(
src_wave_buffer_resource, src_thread_addr_offset, 0)};
return src_thread_element_valid ? tmp : vector_t(customized_value);
}
......
include/ck/utility/amd_ck_fp8.hpp 0 → 100644
View file @ 175a17f8
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/random_gen.hpp"
#include "ck/utility/type.hpp"
#ifdef CK_USE_FNUZ_FP8
#define CK_USE_FNUZ_FP8 1
#else
#define CK_USE_FNUZ_FP8 0
#endif
#ifdef CK_USE_OCP_FP8
#define CK_USE_OCP_FP8 1
#else
#define CK_USE_OCP_FP8 0
#endif
namespace ck {
using f8_fnuz_t = _BitInt(8);
using bf8_fnuz_t = unsigned _BitInt(8);
#if(defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx1200__) || \
defined(__gfx1201__) || defined(__gfx950__)) && \
__HIP_DEVICE_COMPILE__
#define CK_FP8_CVT_FAST_PATH 1
#else
#define CK_FP8_CVT_FAST_PATH 0
#endif
#if(defined(__gfx1200__) || defined(__gfx1201__) || defined(__gfx950__)) && __HIP_DEVICE_COMPILE__
#define CK_OCP_FP8_CVT_FAST_PATH 1
#else
#define CK_OCP_FP8_CVT_FAST_PATH 0
#endif
typedef unsigned char fp8_storage_t;
/**
* \brief Describes FP8 interpretation
*/
enum class ck_fp8_interpretation_t
{
CK_E4M3_OCP = 0, // OCP E4M3
CK_E5M2_OCP = 1, // OCP E5M2
CK_E4M3_FNUZ = 2, // FP8
CK_E5M2_FNUZ = 3, // BF8
};
/**
* \brief Describes saturation behavior
*/
enum class ck_saturation_t
{
CK_NOSAT = 0, // No saturation - replace with NaN or Inf
CK_SATFINITE = 1, // Saturate to finite
};
namespace fp8_impl {
typedef fp8_storage_t fp8x2_storage_t __attribute__((ext_vector_type(2)));
typedef float float2_t __attribute__((ext_vector_type(2)));
__host__ __device__ static inline constexpr bool fnuz_f8_is_nan(f8_fnuz_t a)
{
return static_cast<unsigned char>(a) == 0x80;
}
__host__ __device__ static inline constexpr bool fnuz_bf8_is_nan(bf8_fnuz_t a)
{
return static_cast<unsigned char>(a) == 0x80;
}
__host__ __device__ static inline constexpr bool ocp_f8_is_nan(fp8_storage_t a)
{
return (a & 0x7f) == 0x7f;
}
__host__ __device__ static inline constexpr bool ocp_bf8_is_nan(fp8_storage_t a)
{
return (a & 0x7f) > 0x7c;
}
// The conversion function is from rocblas
// https://github.com/ROCm/rocBLAS/blob/9b7f692abe3c54b88d1e77e045a7db7f1f188b69/library/include/internal/rocblas_hip_f8_impl.h#L220
// This has been modified to handle double types as well
template <typename T, int wm, int we, bool is_fnuz, bool clip = false>
__host__ __device__ static inline T cast_from_f8(fp8_storage_t x)
{
constexpr bool is_half = __hip_internal::is_same<T, _Float16>::value;
constexpr bool is_float = __hip_internal::is_same<T, float>::value;
constexpr bool is_double = __hip_internal::is_same<T, double>::value;
static_assert(is_half || is_float || is_double, "only half, float and double are supported");
constexpr int weo = is_half ? 5 : (is_float ? 8 : 11);
constexpr int wmo = is_half ? 10 : (is_float ? 23 : 52);
T fInf, fNegInf, fNaN, fNeg0, fmax, fmin;
if constexpr(is_half)
{
const unsigned short int ihInf = 0x7C00;
const unsigned short int ihNegInf = 0xFC00;
const unsigned short int ihNaN = 0x7C01;
const unsigned short int ihNeg0 = 0x8000;
/* Max number in e5m2 57344*/
const unsigned short int ifmax = 0x7B00;
const unsigned short int ifmin = 0xFB00;
fInf = bit_cast<_Float16>(ihInf);
fNegInf = bit_cast<_Float16>(ihNegInf);
fNaN = bit_cast<_Float16>(ihNaN);
fNeg0 = bit_cast<_Float16>(ihNeg0);
fmax = bit_cast<_Float16>(ifmax);
fmin = bit_cast<_Float16>(ifmin);
}
else if constexpr(is_float)
{
const unsigned int ifInf = 0x7F800000;
const unsigned int ifNegInf = 0xFF800000;
const unsigned int ifNaN = 0x7F800001;
const unsigned int ifNeg0 = 0x80000000;
/* Max number in e5m2 57344*/
const unsigned int ifmax = 0x47600000;
const unsigned int ifmin = 0xC7600000;
fInf = bit_cast<float>(ifInf);
fNegInf = bit_cast<float>(ifNegInf);
fNaN = bit_cast<float>(ifNaN);
fNeg0 = bit_cast<float>(ifNeg0);
fmax = bit_cast<float>(ifmax);
fmin = bit_cast<float>(ifmin);
}
else if constexpr(is_double)
{
const unsigned long long ifInf = 0x7FF0000000000000ull;
const unsigned long long ifNegInf = 0xFFF0000000000000ull;
const unsigned long long ifNaN = 0x7FF0000000000001ull;
const unsigned long long ifNeg0 = 0x8000000000000000ull;
/* Max number in e5m2 57344*/
const unsigned long long ifmax = 0x40EC000000000000ull;
const unsigned long long ifmin = 0xC0EC000000000000ull;
fInf = bit_cast<double>(ifInf);
fNegInf = bit_cast<double>(ifNegInf);
fNaN = bit_cast<double>(ifNaN);
fNeg0 = bit_cast<double>(ifNeg0);
fmax = bit_cast<double>(ifmax);
fmin = bit_cast<double>(ifmin);
}
if(x == 0)
{
return 0;
}
unsigned long long sign = x >> 7;
unsigned long long mantissa = x & ((1 << wm) - 1);
int exponent = (x & 0x7F) >> wm;
if constexpr(is_fnuz)
{
if(x == 0x80)
{
return fNaN;
}
}
else
{
if(x == 0x80)
{
return fNeg0;
}
if constexpr(we == 4)
{ // e4m3
if((x & 0x7F) == 0x7F)
{
return fNaN;
}
}
else if((x & 0x7C) == 0x7C)
{ // e5m2
if((x & 0x3) == 0)
{
if constexpr(clip)
{
return sign ? fmin : fmax;
}
return sign ? fNegInf : fInf;
}
return fNaN;
}
}
typename __hip_internal::conditional<
sizeof(T) == 2,
unsigned short int,
typename __hip_internal::conditional<sizeof(T) == 4, unsigned int, unsigned long long>::
type>::type retval;
if constexpr(we == 5 && is_half && !is_fnuz)
{
retval = x << 8;
return bit_cast<T>(retval);
}
const int exp_low_cutoff = (1 << (weo - 1)) - (1 << (we - 1)) + 1 - (is_fnuz ? 1 : 0);
// subnormal input
if(exponent == 0)
{
#if defined(__HIP_DEVICE_COMPILE__) && __HIP_DEVICE_COMPILE__
// guaranteed mantissa!=0 since cases 0x0 and 0x80 are handled above
int sh = 1 + __clz(mantissa) - (32 - wm);
#else
int sh = 1 + __builtin_clz(mantissa) - (32 - wm);
#endif
mantissa <<= sh;
exponent += 1 - sh;
mantissa &= ((1ull << wm) - 1);
}
exponent += exp_low_cutoff - 1;
mantissa <<= wmo - wm;
// subnormal output (occurs when T=half, we=5, negative_zero_nan=true)
if(exponent <= 0)
{
mantissa |= 1 << wmo;
mantissa >>= 1 - exponent;
exponent = 0;
}
if constexpr(sizeof(T) == 2)
retval = (sign << 15) | (exponent << 10) | mantissa;
else if constexpr(sizeof(T) == 4)
retval = (sign << 31) | (exponent << 23) | mantissa;
else
retval = (sign << 63) | (static_cast<unsigned long long>(exponent) << 52) | mantissa;
return bit_cast<T>(retval);
}
#if CK_FP8_CVT_FAST_PATH
template <ck_fp8_interpretation_t interpret>
static __device__ float cast_to_f32_from_f8(fp8_storage_t v)
{
union
{
unsigned int i32val;
unsigned char i8val[4];
} val;
val.i8val[0] = v;
static_assert(interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ ||
interpret == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interpret == ck_fp8_interpretation_t::CK_E5M2_FNUZ ||
interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only FNUZ and OCP interpretations are supported");
if constexpr((interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ) ||
(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP))
{
return __builtin_amdgcn_cvt_f32_fp8(val.i32val, 0);
}
else
{
return __builtin_amdgcn_cvt_f32_bf8(val.i32val, 0);
}
}
template <ck_fp8_interpretation_t interpret>
static __device__ float2_t cast_to_f32x2_from_f8x2(fp8x2_storage_t v)
{
const auto i16val = bit_cast<uint16_t>(v);
static_assert(interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ ||
interpret == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interpret == ck_fp8_interpretation_t::CK_E5M2_FNUZ ||
interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only FNUZ and OCP interpretations are supported");
if constexpr((interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ) ||
(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP))
{
return __builtin_amdgcn_cvt_pk_f32_fp8(i16val, false);
}
else
{
return __builtin_amdgcn_cvt_pk_f32_bf8(i16val, false);
}
}
#endif
} // namespace fp8_impl
template <typename T, index_t N>
struct non_native_vector_base;
struct f8_ocp_t
{
using data_type = fp8_storage_t;
data_type data;
static constexpr ck_saturation_t default_saturation = ck_saturation_t::CK_SATFINITE;
static constexpr ck_fp8_interpretation_t default_interpret =
ck_fp8_interpretation_t::CK_E4M3_OCP;
static constexpr unsigned int we = 4; // exponent width
static constexpr unsigned int wm = 3; // mantissa width
__host__ __device__ constexpr bool operator==(const f8_ocp_t& other) const
{
return (data == other.data) && (fp8_impl::ocp_f8_is_nan(data) == false); // NaN != NaN
}
#if CK_USE_OCP_FP8
__host__ __device__ explicit operator float() const
#else
__host__ explicit operator float() const
#endif
{
#if CK_OCP_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data);
#else
return fp8_impl::cast_from_f8<float, wm, we, false>(
this->data); // XXX: clip==false must be consistent with operator _Float16
#endif
}
#if CK_USE_OCP_FP8
__host__ __device__ explicit operator _Float16() const
#else
__host__ explicit operator _Float16() const
#endif
{
#if CK_OCP_FP8_CVT_FAST_PATH
return static_cast<_Float16>(fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data));
#else
return fp8_impl::cast_from_f8<_Float16, wm, we, false>(
this->data); // XXX: clip==false must be consistent with operator float
#endif
}
};
struct bf8_ocp_t
{
using data_type = fp8_storage_t;
data_type data;
static constexpr ck_saturation_t default_saturation = ck_saturation_t::CK_SATFINITE;
static constexpr ck_fp8_interpretation_t default_interpret =
ck_fp8_interpretation_t::CK_E5M2_OCP;
static constexpr unsigned int we = 5; // exponent width
static constexpr unsigned int wm = 2; // mantissa width
__host__ __device__ constexpr bool operator==(const bf8_ocp_t& other) const
{
return (data == other.data) && (fp8_impl::ocp_bf8_is_nan(data) == false); // NaN != NaN
}
#if CK_USE_OCP_FP8
__host__ __device__ explicit operator float() const
#else
__host__ explicit operator float() const
#endif
{
#if defined(__gfx950__) || defined(__gfx1200__) || defined(__gfx1201__)
return fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data);
#else
return fp8_impl::cast_from_f8<float, wm, we, false>(
this->data); // XXX: clip==false must be consistent with operator _Float16
#endif
}
#if CK_USE_OCP_FP8
__host__ __device__ explicit operator _Float16() const
#else
__host__ explicit operator _Float16() const
#endif
{
#if defined(__gfx950__) || defined(__gfx1200__) || defined(__gfx1201__)
return static_cast<_Float16>(fp8_impl::cast_to_f32_from_f8<default_interpret>(this->data));
#else
return fp8_impl::cast_from_f8<_Float16, wm, we, false>(
this->data); // XXX: clip==false must be consistent with operator float
#endif
}
};
template <index_t N>
struct non_native_vector_base<f8_ocp_t, N>
{
using data_t = f8_ocp_t::data_type;
static_assert(sizeof(f8_ocp_t) == sizeof(data_t),
"non_native_vector_base storage size mismatch");
using data_v = data_t __attribute__((ext_vector_type(sizeof(data_t) * N)));
using type = non_native_vector_base<f8_ocp_t, N>;
data_v d; // storage vector
__host__ __device__ non_native_vector_base() = default;
__host__ __device__ non_native_vector_base(data_t a) : d{a} {}
__host__ __device__ non_native_vector_base(f8_ocp_t f) : non_native_vector_base(f.data) {}
__host__ __device__ non_native_vector_base(data_v v) : d{v} {}
__host__ __device__ operator data_v() const { return d; }
};
template <>
struct non_native_vector_base<f8_ocp_t, 1>
{
using data_t = f8_ocp_t::data_type;
using data_v = data_t __attribute__((ext_vector_type(sizeof(data_t))));
using type = non_native_vector_base<f8_ocp_t, 1>;
data_v d; // storage vector
__host__ __device__ non_native_vector_base() = default;
__host__ __device__ non_native_vector_base(data_t a) : d{a} {}
__host__ __device__ non_native_vector_base(f8_ocp_t f) : non_native_vector_base(f.data) {}
__host__ __device__ non_native_vector_base(data_v v) : d{v} {}
__host__ __device__ operator data_v() const { return d; }
__host__ __device__ operator data_t() const { return d[0]; }
__host__ __device__ operator f8_ocp_t() const { return f8_ocp_t{d[0]}; }
};
template <>
struct non_native_vector_base<f8_ocp_t, 2>
{
using data_t = f8_ocp_t::data_type;
using type = non_native_vector_base<f8_ocp_t, 2>;
using data_v = fp8_impl::fp8x2_storage_t; // type of storage vector
data_v d; // storage vector
__host__ __device__ non_native_vector_base() = default;
__host__ __device__ non_native_vector_base(data_t a) : d{a} {}
__host__ __device__ non_native_vector_base(f8_ocp_t f) : non_native_vector_base(f.data) {}
__host__ __device__ non_native_vector_base(data_v v) : d{v} {}
__host__ __device__ operator data_v() const { return d; }
using float2_t = fp8_impl::float2_t;
#if CK_USE_OCP_FP8
__host__ __device__ explicit operator float2_t() const
#else
__host__ explicit operator float2_t() const
#endif
{
#if CK_OCP_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32x2_from_f8x2<f8_ocp_t::default_interpret>(d);
#else
return float2_t{fp8_impl::cast_from_f8<float, f8_ocp_t::wm, f8_ocp_t::we, false>(d[0]),
fp8_impl::cast_from_f8<float, f8_ocp_t::wm, f8_ocp_t::we, false>(d[1])};
#endif
}
};
template <index_t N>
struct non_native_vector_base<bf8_ocp_t, N>
{
using data_t = bf8_ocp_t::data_type;
using data_v = data_t __attribute__((ext_vector_type(sizeof(data_t) * N)));
using type = non_native_vector_base<bf8_ocp_t, N>;
data_v d; // storage vector
__host__ __device__ non_native_vector_base() = default;
__host__ __device__ non_native_vector_base(data_t a) : d{a} {}
__host__ __device__ non_native_vector_base(data_v v) : d{v} {}
__host__ __device__ operator data_v() const { return d; }
};
template <>
struct non_native_vector_base<bf8_ocp_t, 1>
{
using data_t = bf8_ocp_t::data_type;
using data_v = data_t __attribute__((ext_vector_type(sizeof(data_t))));
using type = non_native_vector_base<bf8_ocp_t, 1>;
data_v d; // storage vector
__host__ __device__ non_native_vector_base() = default;
__host__ __device__ non_native_vector_base(data_t a) : d{a} {}
__host__ __device__ non_native_vector_base(bf8_ocp_t f) : non_native_vector_base(f.data) {}
__host__ __device__ non_native_vector_base(data_v v) : d{v} {}
__host__ __device__ operator data_v() const { return d; }
__host__ __device__ operator data_t() const { return d[0]; }
__host__ __device__ operator bf8_ocp_t() const { return bf8_ocp_t{d[0]}; }
};
template <typename T>
__host__ __device__ static inline constexpr bool fp8_is_nan(T);
template <>
__host__ __device__ inline constexpr bool fp8_is_nan(f8_ocp_t a)
{
return fp8_impl::ocp_f8_is_nan(a.data);
}
template <>
__host__ __device__ inline constexpr bool fp8_is_nan(bf8_ocp_t a)
{
return fp8_impl::ocp_bf8_is_nan(a.data);
}
template <>
__host__ __device__ inline constexpr bool fp8_is_nan(f8_fnuz_t a)
{
return fp8_impl::fnuz_f8_is_nan(a);
}
template <>
__host__ __device__ inline constexpr bool fp8_is_nan(bf8_fnuz_t a)
{
return fp8_impl::fnuz_bf8_is_nan(a);
}
template <typename T,
std::enable_if_t<std::is_same_v<T, bf8_ocp_t> || std::is_same_v<T, f8_ocp_t> ||
std::is_same_v<T, bf8_fnuz_t> || std::is_same_v<T, f8_fnuz_t>,
bool> = true>
__host__ __device__ static inline constexpr bool fp8_is_inf(T)
{
return false;
}
template <>
__host__ __device__ inline constexpr bool fp8_is_inf(bf8_ocp_t a)
{
return (a.data & 0x7f) == 0x7c;
}
namespace fp8_impl {
// Assertions to check for supported conversion types
#define __assert_ocp_support(interp) \
{ \
if(interp != ck_fp8_interpretation_t::CK_E4M3_OCP && \
interp != ck_fp8_interpretation_t::CK_E5M2_OCP) \
{ \
__hip_assert(false && "type is unsupported by current target device"); \
} \
}
#define __assert_fnuz_support(interp) \
{ \
if(interp != ck_fp8_interpretation_t::CK_E4M3_FNUZ && \
interp != ck_fp8_interpretation_t::CK_E5M2_FNUZ) \
{ \
__hip_assert(false && "type is unsupported by current target device"); \
} \
}
__host__ __device__ static inline void
__is_interpret_supported([[maybe_unused]] ck_fp8_interpretation_t interp)
{
#if defined(__HIP_DEVICE_COMPILE__) && __HIP_DEVICE_COMPILE__
#if CK_USE_OCP_FP8
__assert_ocp_support(interp);
#endif
#if CK_USE_FNUZ_FP8
__assert_fnuz_support(interp);
#endif
#endif
}
#if CK_FP8_CVT_FAST_PATH
// The conversion function is from rocblas
// https://github.com/ROCm/rocBLAS/blob/9b7f692abe3c54b88d1e77e045a7db7f1f188b69/library/include/internal/rocblas_float8.h#L79
template <ck_fp8_interpretation_t interpret, bool saturate, bool stochastic_rounding = false>
static __device__ fp8_storage_t cast_to_f8_from_f32(float v, unsigned int rng = 0)
{
fp8_storage_t i8data;
union
{
float fval;
unsigned int i32val;
unsigned char i8val[4]; // NOTE: not endian independent
} val;
unsigned int ival = 0;
val.fval = v;
if constexpr(saturate)
{
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ)
{
if((val.i32val & 0x7F800000) != 0x7F800000)
{ /// propagate NAN/INF, no clipping
val.fval = __builtin_amdgcn_fmed3f(val.fval, 240.0, -240.0);
}
}
else if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{ // OCP type
if((val.i32val & 0x7F800000) != 0x7F800000)
{ /// propagate NAN/INF, no clipping
val.fval = __builtin_amdgcn_fmed3f(val.fval, 448.0, -448.0);
}
}
else
{
if((val.i32val & 0x7F800000) != 0x7F800000)
{ /// propagate NAN/INF, no clipping
val.fval = __builtin_amdgcn_fmed3f(val.fval, 57344.0, -57344.0);
}
}
}
if constexpr(stochastic_rounding)
{
ival = (interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ) ||
(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
? __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0)
: __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
i8data = val.i8val[0]; // little endian
}
else
{ // RNE CVT
ival = (interpret == ck_fp8_interpretation_t::CK_E4M3_FNUZ) ||
(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
? __builtin_amdgcn_cvt_pk_fp8_f32(val.fval, val.fval, ival, false)
: __builtin_amdgcn_cvt_pk_bf8_f32(val.fval,
val.fval,
ival,
false); // false -> WORD0
val.i32val = ival;
i8data = val.i8val[0];
}
return i8data;
}
#endif // CK_FP8_CVT_FAST_PATH
// The conversion function is from rocblas
// https://github.com/ROCm/rocBLAS/blob/9b7f692abe3c54b88d1e77e045a7db7f1f188b69/library/include/internal/rocblas_hip_f8_impl.h#L39
// This has been modified to add double types conversion as well
template <typename T, int wm, int we, bool is_fnuz, bool clip = false, bool stoch = false>
__host__ __device__ static inline fp8_storage_t cast_to_f8(T _x, unsigned int rng = 0)
{
constexpr bool is_half = __hip_internal::is_same<T, _Float16>::value;
constexpr bool is_float = __hip_internal::is_same<T, float>::value;
constexpr bool is_double = __hip_internal::is_same<T, double>::value;
static_assert(is_half || is_float || is_double,
"Only half, float and double can be cast to f8");
constexpr int mfmt = (sizeof(T) == 8) ? 52 : ((sizeof(T) == 4) ? 23 : 10);
using T_bitwise = typename __hip_internal::conditional<
sizeof(T) == 2,
unsigned short int,
typename __hip_internal::conditional<sizeof(T) == 4, unsigned int, unsigned long long>::
type>::type;
T_bitwise x_bitwise = bit_cast<T_bitwise>(_x);
unsigned long long x{x_bitwise};
unsigned long long head, mantissa;
int exponent, bias;
unsigned int sign;
unsigned long long fInf, mask;
if constexpr(sizeof(T) == 8)
{
head = x & 0xFFF0000000000000ull;
mantissa = x & 0xFFFFFFFFFFFFFull;
exponent = (head >> 52) & 0x7FF;
sign = head >> 63;
bias = 1023;
fInf = 0x7FF0000000000000ull;
mask = 0x7FFFFFFFFFFFFFFFull;
}
else if constexpr(sizeof(T) == 4)
{
head = x & 0xFF800000;
mantissa = x & 0x7FFFFF;
exponent = (head >> 23) & 0xFF;
sign = head >> 31;
bias = 127;
fInf = 0x7F800000;
mask = 0x7FFFFFFF;
}
else
{
head = x & 0xFC00;
mantissa = x & 0x3FF;
exponent = (head >> 10) & 0x1F;
sign = head >> 15;
bias = 15;
fInf = 0x7C00;
mask = 0x7FFF;
}
unsigned int signed_inf = 0;
unsigned int nan = 0;
if constexpr(is_fnuz)
{
signed_inf = clip ? ((sign << 7) + 0x7f) : 0x80;
nan = 0x80;
}
else
{
if constexpr(we == 4)
{ // e4m3
signed_inf = (sign << 7) + (clip ? 0x7e : 0x7f);
}
else
{ // e5m2
signed_inf = (sign << 7) + (clip ? 0x7b : 0x7c);
}
nan = (sign << 7) + 0x7f;
}
// Max values
unsigned long long ifmax = 0;
if constexpr(sizeof(T) == 8)
{
if constexpr(we == 5)
{ // 57344
ifmax = 0x40EC000000000000ull;
}
else
{
if constexpr(is_fnuz)
{ // 240
ifmax = 0x406E000000000000ull;
}
else
{ // 448
ifmax = 0x407C000000000000ull;
}
}
}
else if(sizeof(T) == 4)
{
if constexpr(we == 5)
{
ifmax = 0x47600000;
}
else
{
if constexpr(is_fnuz)
{
ifmax = 0x43700000;
}
else
{
ifmax = 0x43E00000;
}
}
}
else
{
if constexpr(we == 5)
{
ifmax = 0x7B00;
}
else
{
if constexpr(is_fnuz)
{
ifmax = 0x5B80;
}
else
{
ifmax = 0x5F00;
}
}
}
// Deal with inf and NaNs
if((x & fInf) == fInf)
{
if constexpr(is_fnuz)
return signed_inf;
return mantissa != 0 ? nan : signed_inf;
}
if((x & mask) > ifmax)
{
return signed_inf;
}
if(x == 0)
{
return 0;
}
// First need to check if it is normal or denorm as there is a difference of
// implicit 1 Then need to adjust the exponent to align with the F8 exponent,
// in the meanwhile, shift The mantissa. Then for stochastic rounding, add rng
// to mantissa and truncate. And for RNE, no need to add rng. Then probably
// need to check whether there is carry and adjust exponent and mantissa again
// For IEEE bias mode, the bias is 2^(k-1) -1 where k is the width of exponent
// bits
const int f8_bias = (1 << (we - 1)) - 1 + (is_fnuz ? 1 : 0);
const int f8_denormal_act_exponent = 1 - f8_bias; // actual exponent of f8 denormal
// act_exponent is the actual exponent of fp32/fp16 (after subtracting bias)
// f8_exponent is the converted f8 exponent with bias encoding
// exponent_diff is the diff between fp32/fp16 exponent and f8 exponent,
// the difference needs to be adjusted and mantissa shifted
int act_exponent, f8_exponent, exponent_diff;
if(exponent == 0)
{ // fp32/fp16 is in denormal.
/* fp32 denormal is below 2^-127 so it is usually not a concern here, we
mostly concern fp16 here. In this case, f8 is usually in denormal. But there
could be exceptions. fp16 denormal has exponent bias 15 while bf8 with NANOO has
exponent bias 16. It means that there are some numbers in fp16 denormal but they
are bf8 (NANOO) normals - smallest bf8 (NANOO) normal is 2^-15. fp16 numbers
where exponent==0 (actual exponent -14) and highest bit of mantissa is 1 are bf8
(NANOO) normal. In this case, the fp16 mantissa should be shift left by 1 */
act_exponent = exponent - bias + 1;
exponent_diff = f8_denormal_act_exponent -
act_exponent; // actual exponent is exponent-bias+1 as it is denormal
}
else
{ // fp32/fp16 is normal with implicit 1
act_exponent = exponent - bias;
if(act_exponent <= f8_denormal_act_exponent)
{
/* This is the case where fp32/fp16 is normal but it is in f8 denormal
range. For example fp8 nanoo mode, denormal exponent is -7, but if the fp32/fp16
actual exponent is -7, it is actually larger due to the implicit 1,
Therefore it needs to be adjust to -6 and mantissa shift right by 1.
So for fp32/fp16, exponent -8 is the cut point to convert to fp8 nanoo */
exponent_diff = f8_denormal_act_exponent - act_exponent;
}
else
{ // both fp32/fp16 and f8 are in normal range
exponent_diff = 0; // exponent_diff=0 does not mean there is no difference
// for this case, act_exponent could be larger. Just
// that it does not need shift mantissa
}
mantissa += (1ull << mfmt); // Add the implicit 1 into mantissa
}
bool midpoint = (mantissa & ((1ull << (mfmt - wm + exponent_diff)) - 1)) ==
(1ull << (mfmt - wm + exponent_diff - 1));
/* This part is a bit tricky. The judgment of whether it is a tie needs to be
done before we shift right as shift right could rip off some residual part and
make something not midpoint look like midpoint. For example, the fp16 number
0x1002 (0 00100 0000000010), it is larger than midpoint, but after shift right
by 4 bits, it would look like midpoint.
*/
if(exponent_diff > 0)
mantissa >>= exponent_diff;
else if(exponent_diff == -1)
mantissa <<= -exponent_diff;
bool implicit_one = mantissa & (1ull << mfmt);
// if there is no implicit 1, it means the f8 is denormal and need to adjust
// to denorm exponent
f8_exponent =
(act_exponent + exponent_diff) /*actual f8 exponent*/ + f8_bias - (implicit_one ? 0 : 1);
// Now we have the exponent and mantissa adjusted
unsigned long long drop_mask = (1ull << (mfmt - wm)) - 1;
bool odd =
mantissa & (1ull << (mfmt - wm)); // if the least significant bit that is not truncated is 1
mantissa +=
(stoch ? rng : (midpoint ? (odd ? mantissa : mantissa - 1ull) : mantissa)) & drop_mask;
// Now we deal with overflow
if(f8_exponent == 0)
{
if((1ull << mfmt) & mantissa)
{
f8_exponent = 1; // denormal overflow to become normal, promote exponent
}
}
else
{
if((1ull << (mfmt + 1)) & mantissa)
{
mantissa >>= 1;
f8_exponent++;
}
}
mantissa >>= (mfmt - wm);
// above range: quantize to maximum possible float of the same sign
const int max_exp = (1 << we) - 1;
if(f8_exponent > max_exp)
{
if constexpr(clip)
{
mantissa = (1 << wm) - 1;
f8_exponent = max_exp;
}
else
{
return signed_inf;
}
}
if(f8_exponent == 0 && mantissa == 0)
return is_fnuz ? 0 : (sign << 7);
mantissa &= (1 << wm) - 1;
return (sign << 7) | (f8_exponent << wm) | mantissa;
}
/**
* \brief convert float to @p fp8_storage_t
*
* \tparam interp interpretation of fp8
* \tparam sat saturation of fp8
* \param f float number
* \return fp8_storage_t
*/
template <ck_fp8_interpretation_t interp,
ck_saturation_t sat = ck_saturation_t::CK_SATFINITE,
bool stochastic_rounding = false>
#if CK_FP8_CVT_FAST_PATH
__host__ __device__ static inline fp8_storage_t cvt_float_to_fp8(const float f)
{
__is_interpret_supported(interp);
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f);
}
return cast_to_f8_from_f32<interp, sat == ck_saturation_t::CK_SATFINITE, stochastic_rounding>(
f, rng);
#else
#if CK_USE_OCP_FP8
__host__ __device__ static inline fp8_storage_t cvt_float_to_fp8(const float f)
{
#else
__host__ static inline fp8_storage_t cvt_float_to_fp8(const float f)
{
#endif
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f);
}
if constexpr(interp == ck_fp8_interpretation_t::CK_E4M3_FNUZ)
{
return cast_to_f8<float,
3,
4,
true,
sat == ck_saturation_t::CK_SATFINITE,
stochastic_rounding>(f, rng);
}
else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_FNUZ)
{
return cast_to_f8<float,
2,
5,
true,
sat == ck_saturation_t::CK_SATFINITE,
stochastic_rounding>(f, rng);
}
else if constexpr(interp == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
return cast_to_f8<float,
3,
4,
false,
sat == ck_saturation_t::CK_SATFINITE,
stochastic_rounding>(f, rng);
}
else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_OCP)
{
return cast_to_f8<float,
2,
5,
false,
sat == ck_saturation_t::CK_SATFINITE,
stochastic_rounding>(f, rng);
}
else
{
__hip_assert(false && "FP8 type is not supported by current target device");
return 0;
}
#endif // CK_FP8_CVT_FAST_PATH
}
/**
* \brief convert _Float16 to @p fp8_storage_t
*
* \tparam sat saturation of fp8
* \tparam interp interpretation of fp8
* \tparam stochastic_rounding switch between RNE and SR
* \param x _Float16 value
* \return fp8_storage_t
*/
template <ck_fp8_interpretation_t interp,
ck_saturation_t sat = ck_saturation_t::CK_SATFINITE,
bool stochastic_rounding = false>
#if CK_FP8_CVT_FAST_PATH || CK_USE_OCP_FP8
__host__ __device__ static inline fp8_storage_t cvt_half_t_to_fp8(const _Float16 x)
#else
__host__ static inline fp8_storage_t cvt_half_t_to_fp8(const _Float16 x)
#endif
{
return cvt_float_to_fp8<interp, sat, stochastic_rounding>(static_cast<float>(x));
}
} // namespace fp8_impl
// Declare a template function for fp8 conversion using RNE
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_rne(X x);
// convert fp32 to fp8 with rounding to nearest even
template <>
inline __host__ __device__ f8_ocp_t f8_convert_rne<f8_ocp_t, float>(float x)
{
return f8_ocp_t{
fp8_impl::cvt_float_to_fp8<f8_ocp_t::default_interpret, f8_ocp_t::default_saturation>(x)};
}
// convert fp32 to bf8 with rounding to nearest even
template <>
inline __host__ __device__ bf8_ocp_t f8_convert_rne<bf8_ocp_t, float>(float x)
{
return bf8_ocp_t{
fp8_impl::cvt_float_to_fp8<bf8_ocp_t::default_interpret, bf8_ocp_t::default_saturation>(x)};
}
// convert _Float16 to fp8 with rounding to nearest even
template <>
inline __host__ __device__ f8_ocp_t f8_convert_rne<f8_ocp_t, _Float16>(_Float16 x)
{
return f8_ocp_t{
fp8_impl::cvt_half_t_to_fp8<f8_ocp_t::default_interpret, f8_ocp_t::default_saturation>(x)};
}
template <>
inline __host__ __device__ bf8_ocp_t f8_convert_rne<bf8_ocp_t, _Float16>(_Float16 x)
{
return bf8_ocp_t{
fp8_impl::cvt_half_t_to_fp8<bf8_ocp_t::default_interpret, bf8_ocp_t::default_saturation>(
x)};
}
// Declare a template function for fp8 conversion using RNE
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_sr(X x);
// convert fp32 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_ocp_t f8_convert_sr<f8_ocp_t, float>(float x)
{
return f8_ocp_t{
fp8_impl::cvt_float_to_fp8<f8_ocp_t::default_interpret, f8_ocp_t::default_saturation, true>(
x)};
}
// convert fp32 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_ocp_t f8_convert_sr<bf8_ocp_t, float>(float x)
{
return bf8_ocp_t{fp8_impl::cvt_float_to_fp8<bf8_ocp_t::default_interpret,
bf8_ocp_t::default_saturation,
true>(x)};
}
// convert _Float16 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_ocp_t f8_convert_sr<f8_ocp_t, _Float16>(_Float16 x)
{
return f8_ocp_t{fp8_impl::cvt_half_t_to_fp8<f8_ocp_t::default_interpret,
f8_ocp_t::default_saturation,
true>(x)};
}
// convert _Float16 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_ocp_t f8_convert_sr<bf8_ocp_t, _Float16>(_Float16 x)
{
return bf8_ocp_t{fp8_impl::cvt_half_t_to_fp8<bf8_ocp_t::default_interpret,
bf8_ocp_t::default_saturation,
true>(x)};
}
#if CK_USE_OCP_FP8
using f8_t = f8_ocp_t;
using bf8_t = bf8_ocp_t;
#define CK_FP8_TYPE_FNUZ 0
#define CK_FP8_TYPE_OCP 1
#else
using f8_t = f8_fnuz_t;
using bf8_t = bf8_fnuz_t;
#define CK_FP8_TYPE_FNUZ 1
#define CK_FP8_TYPE_OCP 0
#endif
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment