Skip to content

GitLab

Commit aea62819 authored by Chaitanya Inumella's avatar Chaitanya Inumella
Browse files

Rebase branch 'develop' of... 

Rebase branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into contraction_hipTENSOR
parents 75af5450 75ab874e
Hide whitespace changes
Inline Side-by-side
Showing with 874 additions and 1394 deletions
script/profile_gemm_bilinear.sh 0 → 100755
View file @ aea62819
#!/bin/bash
## GPU visibility
export HIP_VISIBLE_DEVICES=0
DRIVER="../build/bin/ckProfiler"
OP=$1
DATATYPE=$2
LAYOUT=$3
VERIFY=$4
INIT=$5
LOG=$6
TIME=$7
######## op datatype layout verify init log time M___ N___ K___ StrideA StrideB StrideD StrideE Alpha Beta
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 960 1024 1024 -1 -1 -1 -1 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920 2048 2048 -1 -1 -1 -1 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840 4096 4096 -1 -1 -1 -1 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680 8192 8192 -1 -1 -1 -1 1 1
######## op datatype layout verify init log time M___ N___ K___ StrideA StrideB StrideD StrideE Alpha Beta
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 960 1024 1024 -1 -1 0 -1 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920 2048 2048 -1 -1 0 -1 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840 4096 4096 -1 -1 0 -1 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680 8192 8192 -1 -1 0 -1 1 1
######## op datatype layout verify init log time M___ N___ K___ StrideA StrideB StrideD StrideE Alpha Beta
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1000 1000 1000 -1 -1 0 -1 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2000 2000 2000 -1 -1 0 -1 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4000 4000 4000 -1 -1 0 -1 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8000 8000 8000 -1 -1 0 -1 1 1
######## op datatype layout verify init log time M___ N___ K___ StrideA StrideB StrideD StrideE Alpha Beta
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024 1024 1024 1056 1056 1056 1056 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 2048 2048 2080 2080 2080 2080 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096 4096 4096 4128 4128 4128 4128 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 8192 8192 8224 8224 8224 8224 1 1
######## op datatype layout verify init log time M___ N___ K___ StrideA StrideB StrideD StrideE Alpha Beta
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024 1024 1024 1088 1088 1088 1088 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 2048 2048 2112 2112 2112 2112 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096 4096 4096 4160 4160 4160 4160 1 1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 8192 8192 8256 8256 8256 8256 1 1
\ No newline at end of file
script/profile_grouped_gemm.sh 0 → 100755
View file @ aea62819
#!/bin/bash
## GPU visibility
export HIP_VISIBLE_DEVICES=0
DRIVER="../build/bin/ckProfiler"
OP=$1
DATATYPE=$2
LAYOUT=$3
VERIFY=$4
INIT=$5
LOG=$6
TIME=$7
######## op datatype layout verify init log time Ms______________ Ns______________ Ks_____________ StrideAs___________ StrideBs__________ StrideCs___________
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 256,512,1024,768 128,256,384,1024 128,192,256,512 1024,1025,1044,1026 1024,1024,1024,1024 1025,1024,1028,1024
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 512,768,2048,128 128,256,384,1024 128,192,256,512 1024,1025,2053,1026 1024,1024,1024,1024 1025,1024,2054,1024
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 256,512,1024,768 512,256,768,1024 128,192,256,512 1024,1045,1034,1026 1024,1024,1024,1024 1025,1063,1028,1024
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 512,768,4096,768 128,768,512,2048 128,192,256,512 1024,1027,4096,2050 1024,1024,1024,2048 1025,1024,4099,2049
script/profile_reduce_no_index.sh
View file @ aea62819
#!/bin/bash
PRECISION=
DRIVER="../build/bin/ckProfiler"
VERIFY="-v $1"
INIT=$2
NREPEAT=$3
PRECISION=$4
##PRECISION=--half
##PRECISION=--double
##PRECISION=--int8
......@@ -12,19 +15,11 @@ elif [ -n $PRECISION ] && [ "$PRECISION" = "--int8" ]; then
ACCTYPE="-C 2"
fi
driver="./bin/ckProfiler"
VERIFY="-v $1"
INIT=$2
NREPEAT=$3
#### 0 - ADD, 5 - AVG, 7 - NORM2
Operations="0 5 7"
Operations="0 5"
#### 0 - ADD, 5 - AVG, for int8, no NORM2 supported
if [ -n $PRECISION ] && [ "$PRECISION" = "--int8" ]; then
if [ -n $PRECISION ] && [ "$PRECISION" = "--int8" -o "$PRECISION" = "--half" ]; then
Operations=5
fi
......@@ -32,19 +27,19 @@ fi
for op in $Operations; do
set -x
####### datatype layout reduce dims op acctype verify init repeats
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,2,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 1 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 1,2,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,2,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,22960 -R 0 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,22960 -R 1 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 4,1469440 -R 0 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 4,1469440 -R 1 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0,1,2,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 1 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 1,2,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0,2,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0,1,3 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,22960 -R 0 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,22960 -R 1 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 4,1469440 -R 0 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 4,1469440 -R 1 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
set +x
done
......@@ -55,29 +50,29 @@ Operations=5
for op in $Operations; do
set -x
####### datatype layout reduce dims op acctype verify init repeats
$driver reduce $PRECISION -D 256,14,14,1024 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,28,28,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,58,58,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,7,7,2048 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,14,14,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,30,30,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,56,56,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,16,16,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,28,28,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,7,7,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,56,56,64 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,230,230,3 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,14,14,1024 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,28,28,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,58,58,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,7,7,2048 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,14,14,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,30,30,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,56,56,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,16,16,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,28,28,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,7,7,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,56,56,64 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,14,14,1024 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,28,28,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,58,58,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,7,7,2048 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,14,14,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,30,30,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,56,56,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,16,16,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,28,28,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,7,7,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,56,56,64 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,230,230,3 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,14,14,1024 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,28,28,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,58,58,128 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,7,7,2048 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,14,14,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,30,30,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,56,56,256 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,16,16,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,28,28,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,7,7,512 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,56,56,64 -R 0,1,2 -O $op $ACCTYPE $VERIFY $INIT $NREPEAT
set +x
done
script/profile_reduce_with_index.sh
View file @ aea62819
#!/bin/bash
PRECISION=
DRIVER="../build/bin/ckProfiler"
VERIFY="-v $1"
INIT=$2
NREPEAT=$3
PRECISION=$4
##PRECISION=--half
##PRECISION=--double
##PRECISION=--int8
##PRECISION=--bf16
driver="./bin/ckProfiler"
VERIFY="-v $1"
INIT=$2
NREPEAT=$3
#### 2 - MIN, 3 - MAX, 4 - AMAX
Operations="2 4"
......@@ -20,19 +17,19 @@ for op in $Operations; do
for use_idx in 0 1; do
set -x
####### datatype layout reduce dims op use index verify init repeats
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,2,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 1 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 1,2,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,2,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 64,4,280,82 -R 0,1,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,22960 -R 0 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,22960 -R 1 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 4,1469440 -R 0 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 4,1469440 -R 1 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0,1,2,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 1 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 1,2,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0,2,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 64,4,280,82 -R 0,1,3 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,22960 -R 0 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,22960 -R 1 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 4,1469440 -R 0 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 4,1469440 -R 1 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
set +x
done
done
......@@ -44,29 +41,29 @@ for op in $Operations; do
for use_idx in 0 1; do
set -x
####### datatype layout reduce dims op use index verify init repeats
$driver reduce $PRECISION -D 256,14,14,1024 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,28,28,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,58,58,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,7,7,2048 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,14,14,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,30,30,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,56,56,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,16,16,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,28,28,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,7,7,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,56,56,64 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 256,230,230,3 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,14,14,1024 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,28,28,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,58,58,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,7,7,2048 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,14,14,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,30,30,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,56,56,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,16,16,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,28,28,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,7,7,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$driver reduce $PRECISION -D 128,56,56,64 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,14,14,1024 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,28,28,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,58,58,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,7,7,2048 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,14,14,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,30,30,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,56,56,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,16,16,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,28,28,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,7,7,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,56,56,64 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 256,230,230,3 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,14,14,1024 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,28,28,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,58,58,128 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,7,7,2048 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,14,14,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,30,30,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,56,56,256 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,16,16,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,28,28,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,7,7,512 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
$DRIVER reduce $PRECISION -D 128,56,56,64 -R 0,1,2 -O $op -I $use_idx $VERIFY $INIT $NREPEAT
set +x
done
done
......
script/profile_resnet50.sh 0 → 100755
View file @ aea62819
#!/bin/bash
## GPU visibility
export HIP_VISIBLE_DEVICES=0
DRIVER="../build/bin/ckProfiler"
OP=$1
DATATYPE=$2
IN_LAYOUT=$3
WEI_LAYOUT=$4
OUT_LAYOUT=$5
VERIFY=$6
INIT=$7
LOG=$8
TIME=$9
N=${10}
# Resnet50
######## op____________________ datatype in_layout wei_layout out_layout verify init log time N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 64 3 7 7 224 224 2 2 1 1 3 3 3 3
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 64 64 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 128 256 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 128 128 3 3 56 56 2 2 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 512 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 256 3 3 28 28 2 2 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 512 3 3 14 14 2 2 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
script/run_full_performance_tests.sh 0 → 100755
View file @ aea62819
#!/bin/bash
#
# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
# you would also need to set up some environment variables in order to
# post your new test results to the database and compare them to the baseline
# please contact Illia.Silin@amd.com for more details
#
# run the script as "./run_full_performance_tests.sh <verification> <tag for your test environment> <gpu_arch> <branch name> < node name>
# input arguments:
# verification = 0 : do not verify result correctness on CPU
# = 1 : verifuy correctness on CPU (may take a long time)
# environment tag : a string describing the specifics of your test environment
# gpu_arch : a string for GPU architecture, e.g. "gfx908" or "gfx90a".
# branch name : name of the branch in git repo (git status | grep -e 'On branch')
# node name : $hostname
#get the command line arguments:
export verify=$1
echo 'Verification: ' $verify
export env_type=$2
echo 'Environment type: ' $env_type
export gpu_arch=$3
echo 'GPU architecture: ' $gpu_arch
export branch=$4
echo 'Branch name: ' $branch
export host_name=$5
echo 'Host name: ' $host_name
function print_log_header(){
rm -f $1;
echo 'On branch ' $3 &> $1;
echo 'Node name: ' $4 >> $1;
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
rocminfo | grep "Compute Unit:" >> $1;
hipcc --version | grep -e 'HIP version' >> $1;
echo 'Environment type: ' $2 >> $1;
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
}
#run gemm tests
export gemm_log="perf_gemm_${gpu_arch}.log"
print_log_header $gemm_log $env_type $branch $host_name
./profile_gemm.sh gemm 0 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 $verify 1 0 1 | tee -a $gemm_log
#run batched_gemm tests
export batched_gemm_log="perf_batched_gemm_${gpu_arch}.log"
print_log_header $batched_gemm_log $env_type $branch $host_name
./profile_batched_gemm.sh batched_gemm 0 0 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 1 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 2 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 3 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 0 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 1 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 2 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 3 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 0 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 1 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 2 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 3 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 0 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 1 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 2 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 3 $verify 1 0 1 | tee -a $batched_gemm_log
#run grouped_gemm tests
export grouped_gemm_log="perf_grouped_gemm_${gpu_arch}.log"
print_log_header $grouped_gemm_log $env_type $branch $host_name
./profile_grouped_gemm.sh grouped_gemm 1 0 $verify 1 0 1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 1 $verify 1 0 1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 2 $verify 1 0 1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 3 $verify 1 0 1 | tee -a $grouped_gemm_log
#run GEMM+Bilinear tests
export gemm_bilinear_log="perf_gemm_bilinear_${gpu_arch}.log"
print_log_header $gemm_bilinear_log $env_type $branch $host_name
./profile_gemm_bilinear.sh gemm_bilinear 1 0 $verify 1 0 1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 1 $verify 1 0 1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 2 $verify 1 0 1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 3 $verify 1 0 1 | tee -a $gemm_bilinear_log
#run conv_fwd tests
export conv_fwd_log="perf_conv_fwd_${gpu_arch}.log"
print_log_header $conv_fwd_log $env_type $branch $host_name
./profile_conv_fwd.sh conv_fwd 0 1 $verify 1 0 1 256 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 1 1 $verify 1 0 1 256 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 2 1 $verify 1 0 1 256 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 3 1 $verify 1 0 1 256 | tee -a $conv_fwd_log
#run conv_bwd_data tests
export conv_bwd_data_log="perf_conv_bwd_data_${gpu_arch}.log"
print_log_header $conv_bwd_data_log $env_type $branch $host_name
./profile_conv_bwd_data.sh conv_bwd_data 0 1 $verify 1 0 1 256 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 1 1 $verify 1 0 1 256 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 2 1 $verify 1 0 1 256 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 3 1 $verify 1 0 1 256 | tee -a $conv_bwd_data_log
#run resnet50 tests
export resnet256_log="perf_resnet50_N256_${gpu_arch}.log"
print_log_header $resnet256_log $env_type $branch $host_name
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 256 | tee -a $resnet256_log
export resnet4_log="perf_resnet50_N4_${gpu_arch}.log"
print_log_header $resnet4_log $env_type $branch $host_name
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 4 | tee -a $resnet4_log
#run reduction tests
export reduction_log="perf_reduction_${gpu_arch}.log"
print_log_header $reduction_log $env_type $branch $host_name
./profile_reduce_with_index.sh $verify 2 10 --half | tee -a $reduction_log
./profile_reduce_no_index.sh $verify 2 10 --half | tee -a $reduction_log
script/run_performance_tests.sh
View file @ aea62819
#!/bin/bash
#
# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
# and make sure the following python packages are installed in your environment:
# run the script as "./run_performance_tests.sh <verification> <tag for your test environment> <gpu_arch> <branch name> < node name>
# input arguments:
# verification = 0 : do not verify result correctness on CPU
# = 1 : verify correctness on CPU (may take a long time)
# environment tag : a string describing the specifics of your test environment
# gpu_arch : a string for GPU architecture, e.g. "gfx908" or "gfx90a".
# branch name : name of the branch in git repo (git status | grep -e 'On branch')
# node name : $hostname
pip3 install --upgrade pip
pip3 install sqlalchemy pymysql pandas sshtunnel
#get the command line arguments:
export verify=$1
echo 'Verification: ' $verify
export env_type=$2
echo 'Environment type: ' $env_type
export gpu_arch=$3
echo 'GPU architecture: ' $gpu_arch
export branch=$4
echo 'Branch name: ' $branch
export host_name=$5
echo 'Host name: ' $host_name
# you would also need to set up some environment variables in order to
# post your new test results to the database and compare them to the baseline
# please contact Illia.Silin@amd.com for more details
#
export gemm_log="perf_gemm.log"
rm -f $gemm_log
git status | grep -e 'On branch' > ${gemm_log}
echo -n 'Node name: ' >>${gemm_log}; hostname >> ${gemm_log}
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> ${gemm_log}; rocminfo | grep "Name:" | grep "gfx" >> ${gemm_log}
rocminfo | grep "Compute Unit:" >> ${gemm_log}
hipcc --version | grep -e 'HIP version' >> ${gemm_log}
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${gemm_log}
./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a ${gemm_log}
./profile_gemm.sh gemm 1 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 0 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 0 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 0 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 0 1 0 5 | tee -a $gemm_log
function print_log_header(){
rm -f $1;
echo 'On branch ' $3 &> $1;
echo 'Node name: ' $4 >> $1;
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
rocminfo | grep "Compute Unit:" >> $1;
hipcc --version | grep -e 'HIP version' >> $1;
echo 'Environment type: ' $2 >> $1;
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
}
python3 parse_perf_data.py ${gemm_log}
#run gemm tests
export gemm_log="perf_gemm_${gpu_arch}.log"
print_log_header $gemm_log $env_type $branch $host_name
./profile_gemm.sh gemm 0 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 $verify 1 0 1 | tee -a $gemm_log
#run resnet50 test
export resnet_log="perf_resnet50.log"
rm -f $resnet_log
git status | grep -e 'On branch' > ${resnet_log}
echo -n 'Node name: '>>${resnet_log}; hostname >>${resnet_log}
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> ${resnet_log}; rocminfo | grep "Name:" | grep "gfx" >> ${resnet_log}
rocminfo | grep "Compute Unit:" >> ${resnet_log}
hipcc --version | grep -e 'HIP version' >> ${resnet_log}
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${resnet_log}
#first run tests with N=256
./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 256 | tee -a ${resnet_log}
#then run with N=4
./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 4 | tee -a ${resnet_log}
#the script will put the results from N=256 and N=4 runs into separate tables
python3 parse_perf_data.py ${resnet_log}
#run resnet50 tests
export resnet256_log="perf_resnet50_N256_${gpu_arch}.log"
print_log_header $resnet256_log $env_type $branch $host_name
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 256 | tee -a $resnet256_log
export resnet4_log="perf_resnet50_N4_${gpu_arch}.log"
print_log_header $resnet4_log $env_type $branch $host_name
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 4 | tee -a $resnet4_log
test/CMakeLists.txt
View file @ aea62819
......@@ -41,10 +41,11 @@ add_subdirectory(gemm_reduce)
add_subdirectory(batched_gemm)
add_subdirectory(batched_gemm_reduce)
add_subdirectory(grouped_gemm)
add_subdirectory(convnd_fwd)
add_subdirectory(reduce)
add_subdirectory(conv2d_bwd_weight)
add_subdirectory(convnd_fwd)
add_subdirectory(convnd_bwd_weight)
add_subdirectory(convnd_bwd_data)
add_subdirectory(grouped_convnd_fwd)
add_subdirectory(block_to_ctile_map)
add_subdirectory(softmax)
add_subdirectory(layernorm)
test/batched_gemm/CMakeLists.txt
View file @ aea62819
add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp)
target_link_libraries(test_batched_gemm_fp16 PRIVATE host_tensor)
target_link_libraries(test_batched_gemm_fp16 PRIVATE utility)
target_link_libraries(test_batched_gemm_fp16 PRIVATE device_batched_gemm_instance)
test/batched_gemm_reduce/CMakeLists.txt
View file @ aea62819
add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE host_tensor)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE utility)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance)
test/conv2d_bwd_data/CMakeLists.txt deleted 100644 → 0
View file @ 75af5450
add_test_executable(test_conv2d_bwd_data conv2d_bwd_data.cpp)
target_link_libraries(test_conv2d_bwd_data PRIVATE host_tensor)
target_link_libraries(test_conv2d_bwd_data PRIVATE device_conv2d_bwd_data_instance)
test/conv2d_bwd_data/conv2d_bwd_data.cpp deleted 100644 → 0
View file @ 75af5450
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_bwd_data.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_bwd_data.hpp"
using F16 = ck::half_t;
using F32 = float;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using DeviceConvBwdDataNoOpPtr =
DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
template <typename T>
static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
{
float max_diff = 1e-6;
for(int i = 0; i < ref.mData.size(); ++i)
{
float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
if(max_diff < diff)
{
return false;
}
}
return true;
}
int main(int argc, char* argv[])
{
int data_type = 0;
int init_method = 0;
// Conv shape
ck::index_t N = 128;
ck::index_t K = 256;
ck::index_t C = 192;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 71;
ck::index_t Wi = 71;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
if(argc == 1)
{
data_type = 1;
init_method = 1;
}
else if(argc == 3)
{
data_type = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
}
else if(argc == 18)
{
data_type = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
N = std::stoi(argv[3]);
K = std::stoi(argv[4]);
C = std::stoi(argv[5]);
Y = std::stoi(argv[6]);
X = std::stoi(argv[7]);
Hi = std::stoi(argv[8]);
Wi = std::stoi(argv[9]);
conv_stride_h = std::stoi(argv[10]);
conv_stride_w = std::stoi(argv[11]);
conv_dilation_h = std::stoi(argv[12]);
conv_dilation_w = std::stoi(argv[13]);
in_left_pad_h = std::stoi(argv[14]);
in_left_pad_w = std::stoi(argv[15]);
in_right_pad_h = std::stoi(argv[16]);
in_right_pad_w = std::stoi(argv[17]);
}
else
{
printf("arg1: data type (0=fp32, 1=fp16, 2= bfp16, 3= int8_t )\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3 to 17: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
auto Run = [&](auto input_type, auto wei_type, auto out_type, auto acc_type) {
using InDataType = decltype(input_type);
using WeiDataType = decltype(wei_type);
using OutDataType = decltype(out_type);
using AccDataType = decltype(acc_type);
using ReferenceConvBwdInstance =
ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const std::vector<ck::index_t> input_spatial_lengths{{Hi, Wi}};
const std::vector<ck::index_t> filter_spatial_lengths{{Y, X}};
const std::vector<ck::index_t> output_spatial_lengths{{Ho, Wo}};
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
};
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X));
Tensor<InDataType> in_n_c_hi_wi_host_result(f_host_tensor_descriptor(N, C, Hi, Wi));
Tensor<InDataType> in_n_c_hi_wi_device_result(f_host_tensor_descriptor(N, C, Hi, Wi));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi_host_result.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) *
in_n_c_hi_wi_device_result.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
// reset input to zero
in_n_c_hi_wi_device_result.GenerateTensorValue(GeneratorTensor_1<InDataType>{0});
in_device_buf.ToDevice(in_n_c_hi_wi_device_result.mData.data());
// get host result
{
auto ref_conv = ReferenceConvBwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result,
wei_k_c_y_x,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvBwdDataNoOpPtr = ck::tensor_operation::device::
DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>;
// add device Conv instances
std::vector<DeviceConvBwdDataNoOpPtr> conv_ptrs;
if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
{
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::bhalf_t>)
{
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, int8_t>)
{
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
// profile device Conv instances
bool success = true;
for(auto& conv_ptr : conv_ptrs)
{
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
invoker_ptr->Run(argument_ptr.get(), 1);
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
if(!check_out(in_n_c_hi_wi_host_result, in_n_c_hi_wi_device_result))
{
std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
success = false;
}
else
{
std::cout << "Pass Info: " << conv_ptr->GetTypeString() << std::endl;
}
}
else
{
std::cout << "Not support Info: " << conv_ptr->GetTypeString() << std::endl;
}
}
if(success)
{
std::cout << "test conv2d bwd : Pass" << std::endl;
return 0;
}
else
{
std::cout << "test conv2d bwd: Fail " << std::endl;
return -1;
}
};
if(data_type == 0)
{
return Run(F32(), F32(), F32(), F32());
}
else if(data_type == 1)
{
return Run(F16(), F16(), F16(), F32());
}
else if(data_type == 2)
{
return Run(BF16(), BF16(), BF16(), F32());
}
else if(data_type == 3)
{
return Run(INT8(), INT8(), INT8(), int());
}
else
{
return 1;
}
}
test/conv2d_bwd_weight/CMakeLists.txt deleted 100644 → 0
View file @ 75af5450
add_test_executable(test_conv2d_bwd_weight conv2d_bwd_weight.cpp)
target_link_libraries(test_conv2d_bwd_weight PRIVATE host_tensor device_conv2d_bwd_weight_instance conv_util)
test/conv2d_bwd_weight/conv2d_bwd_weight.cpp deleted 100644 → 0
View file @ 75af5450
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <vector>
#include "test/convnd_fwd/conv_util.hpp"
#include "profiler/include/profile_conv_bwd_weight_impl.hpp"
int test_self()
{
bool pass = true;
std::vector<ck::utils::conv::ConvParams> params;
params.push_back({2, 128, 256, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
params.push_back({2, 128, 256, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
params.push_back({2, 128, 256, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
for(auto& param : params)
{
// f32
pass &= ck::profiler::profile_conv_bwd_weight_impl<2,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
// fp16
pass &= ck::profiler::profile_conv_bwd_weight_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
}
return pass;
}
int main(int argc, char* argv[])
{
int data_type = 1;
int init_method = 1;
// Conv shape
ck::index_t N = 128;
ck::index_t K = 256;
ck::index_t C = 192;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 71;
ck::index_t Wi = 71;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
ck::index_t split_k = 1;
bool pass = true;
if(argc == 1)
{
pass = test_self();
}
else
{
if(argc == 3)
{
data_type = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
}
else if(argc == 19)
{
data_type = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
N = std::stoi(argv[3]);
K = std::stoi(argv[4]);
C = std::stoi(argv[5]);
Y = std::stoi(argv[6]);
X = std::stoi(argv[7]);
Hi = std::stoi(argv[8]);
Wi = std::stoi(argv[9]);
conv_stride_h = std::stoi(argv[10]);
conv_stride_w = std::stoi(argv[11]);
conv_dilation_h = std::stoi(argv[12]);
conv_dilation_w = std::stoi(argv[13]);
in_left_pad_h = std::stoi(argv[14]);
in_left_pad_w = std::stoi(argv[15]);
in_right_pad_h = std::stoi(argv[16]);
in_right_pad_w = std::stoi(argv[17]);
split_k = std::stoi(argv[18]);
}
else
{
printf("arg1: data type (0=fp32, 1=fp16, 2= bfp16, 3= int8_t )\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3 to 17: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
ck::utils::conv::ConvParams param{2,
N,
K,
C,
{Y, X},
{Hi, Wi},
{conv_stride_h, conv_stride_w},
{conv_dilation_h, conv_dilation_w},
{in_left_pad_h, in_left_pad_w},
{in_right_pad_h, in_right_pad_w}};
if(data_type == 0)
{
pass = ck::profiler::profile_conv_bwd_weight_impl<2,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
init_method,
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
split_k);
}
else if(data_type == 1)
{
pass = ck::profiler::profile_conv_bwd_weight_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
init_method,
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
split_k);
}
else
{
std::cout << "Not support data type" << std::endl;
return 1;
}
}
if(pass)
{
std::cout << "test conv2d bwd weight : Pass" << std::endl;
return 0;
}
else
{
std::cout << "test conv2d bwd weight: Fail " << std::endl;
return -1;
}
}
test/conv_util/CMakeLists.txt
View file @ aea62819
add_gtest_executable(test_conv_util conv_util.cpp)
target_link_libraries(test_conv_util PRIVATE host_tensor conv_util)
target_link_libraries(test_conv_util PRIVATE utility)
test/conv_util/conv_util.cpp
View file @ aea62819
......@@ -10,198 +10,147 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
namespace {
class TestConvUtil : public ::testing::Test
{
public:
void SetNDParams(std::size_t ndims)
void SetNDParams(std::size_t ndims, std::size_t s, std::size_t d, std::size_t p)
{
conv_params.num_dim_spatial_ = ndims;
conv_params.filter_spatial_lengths_ = std::vector<ck::index_t>(ndims, 3);
conv_params.input_spatial_lengths_ = std::vector<ck::index_t>(ndims, 71);
conv_params.conv_filter_strides_ = std::vector<ck::index_t>(ndims, 2);
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>(ndims, 1);
conv_params.input_left_pads_ = std::vector<ck::index_t>(ndims, 1);
conv_params.input_right_pads_ = std::vector<ck::index_t>(ndims, 1);
conv_params = ck::utils::conv::ConvParam(ndims,
2,
128,
192,
256,
std::vector<ck::index_t>(ndims, 3),
std::vector<ck::index_t>(ndims, 71),
std::vector<ck::index_t>(ndims, s),
std::vector<ck::index_t>(ndims, d),
std::vector<ck::index_t>(ndims, p),
std::vector<ck::index_t>(ndims, p));
}
protected:
// ------- default 2D -------
// input NCHW {128,192,71,71},
// weights KCYX {256,192,3,3},
// stride {2,2},
// dilations {1,1},
// padding {{1,1}, {1,1}}
ck::utils::conv::ConvParams conv_params;
// input GNCHW {2, 128, 192, 71, 71},
// weights GKCYX {2, 256, 192, 3, 3},
// stride {s, s},
// dilations {d, d},
// padding {{p, p}, {p, p}
ck::utils::conv::ConvParam conv_params;
};
} // namespace
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths2D)
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths1D)
{
ck::utils::conv::ConvParams conv_params;
// stride 2, dilation 1, pad 1
SetNDParams(1, 2, 1, 1);
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36},
"Error: ConvParams 2D default constructor."));
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 1, dilation 1, pad 1
SetNDParams(1, 1, 1, 1);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71, 71}, "Error: ConvParams 2D stride {1,1}."));
out_spatial_len, std::vector<ck::index_t>{71}, "Error: ConvParams 1D stride {1}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{2, 2};
conv_params.input_left_pads_ = std::vector<ck::index_t>{2, 2};
conv_params.input_right_pads_ = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 2, dilation 1, pad 2
SetNDParams(1, 2, 1, 2);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37},
"Error: ConvParams 2D padding left/right {2,2}."));
std::vector<ck::index_t>{37},
"Error: ConvParams 1D padding left/right {2}."));
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 2, dilation 2, pad 2
SetNDParams(1, 2, 2, 2);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36}, "Error: ConvParams 2D dilation {2,2}."));
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D dilation {2}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{3, 3};
conv_params.input_left_pads_ = std::vector<ck::index_t>{1, 1};
conv_params.input_right_pads_ = std::vector<ck::index_t>{1, 1};
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 3, dilation 2, pad 1
SetNDParams(1, 3, 2, 1);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(
ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{23, 23},
"Error: ConvParams 2D strides{3,3}, padding {1,1}, dilations {2,2}."));
std::vector<ck::index_t>{23},
"Error: ConvParams 1D strides{3}, padding {1}, dilations {2}."));
}
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths1D)
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths2D)
{
SetNDParams(1);
// stride 2, dilation 1, pad 1
SetNDParams(2, 2, 1, 1);
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D."));
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36},
"Error: ConvParams 2D default constructor."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 1, dilation 1, pad 1
SetNDParams(2, 1, 1, 1);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71}, "Error: ConvParams 1D stride {1}."));
out_spatial_len, std::vector<ck::index_t>{71, 71}, "Error: ConvParams 2D stride {1,1}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{2};
conv_params.input_left_pads_ = std::vector<ck::index_t>{2};
conv_params.input_right_pads_ = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 2, dilation 1, pad 2
SetNDParams(2, 2, 1, 2);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37},
"Error: ConvParams 1D padding left/right {2}."));
std::vector<ck::index_t>{37, 37},
"Error: ConvParams 2D padding left/right {2,2}."));
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 2, dilation 2, pad 2
SetNDParams(2, 2, 2, 2);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D dilation {2}."));
out_spatial_len, std::vector<ck::index_t>{36, 36}, "Error: ConvParams 2D dilation {2,2}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{3};
conv_params.input_left_pads_ = std::vector<ck::index_t>{1};
conv_params.input_right_pads_ = std::vector<ck::index_t>{1};
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 3, dilation 2, pad 1
SetNDParams(2, 3, 2, 1);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(
ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{23},
"Error: ConvParams 1D strides{3}, padding {1}, dilations {2}."));
std::vector<ck::index_t>{23, 23},
"Error: ConvParams 2D strides{3,3}, padding {1,1}, dilations {2,2}."));
}
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths3D)
{
SetNDParams(3);
// stride 2, dilation 1, pad 1
SetNDParams(3, 2, 1, 1);
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36, 36}, "Error: ConvParams 3D."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{1, 1, 1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 1, dilation 1, pad 1
SetNDParams(3, 1, 1, 1);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{71, 71, 71},
"Error: ConvParams 3D stride {1, 1, 1}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{2, 2, 2};
conv_params.input_left_pads_ = std::vector<ck::index_t>{2, 2, 2};
conv_params.input_right_pads_ = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 2, dilation 1, pad 2
SetNDParams(3, 2, 1, 2);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37, 37},
"Error: ConvParams 3D padding left/right {2, 2, 2}."));
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 2, dilation 2, pad 2
SetNDParams(3, 2, 2, 2);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36, 36},
"Error: ConvParams 3D dilation {2, 2, 2}."));
conv_params.conv_filter_strides_ = std::vector<ck::index_t>{3, 3, 3};
conv_params.input_left_pads_ = std::vector<ck::index_t>{1, 1, 1};
conv_params.input_right_pads_ = std::vector<ck::index_t>{1, 1, 1};
conv_params.conv_filter_dilations_ = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
// stride 3, dilation 2, pad 1
SetNDParams(3, 3, 2, 1);
out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len,
std::vector<ck::index_t>{23, 23, 23},
"Error: ConvParams 3D strides{3, 3, 3}, padding {1, 1, 1}, dilations {2, 2, 2}."));
}
TEST(ConvUtil, GetHostTensorDescriptor)
{
namespace tl = ck::tensor_layout::convolution;
std::vector<std::size_t> dims{2, 3, 4, 5};
HostTensorDescriptor h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NHWC{});
EXPECT_TRUE(ck::utils::check_err(
h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NHWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 1, 3 * 5, 3}, "Error: wrong NHWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCHW{});
EXPECT_TRUE(ck::utils::check_err(
h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NCHW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 4 * 5, 5, 1}, "Error: wrong NCHW dimensions strides!"));
dims = std::vector<std::size_t>{2, 3, 4};
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NWC{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4, 1, 3}, "Error: wrong NWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCW{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NCW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4, 4, 1}, "Error: wrong NCW dimensions strides!"));
dims = std::vector<std::size_t>{2, 3, 4, 5, 6};
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NDHWC{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NDHWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
1, // C
3 * 5 * 6, // D
3 * 6, // H
3}, // W
"Error: wrong NDHWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCDHW{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NCDHW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
4 * 5 * 6, // C
5 * 6, // D
6, // H
1}, // W
"Error: wrong NCDHW dimensions strides!"));
}
test/convnd_bwd_data/CMakeLists.txt
View file @ aea62819
add_test_executable(test_convnd_bwd_data convnd_bwd_data.cpp)
target_link_libraries(test_convnd_bwd_data PRIVATE host_tensor device_convnd_bwd_data_instance conv_util)
add_gtest_executable(test_convnd_bwd_data convnd_bwd_data.cpp)
target_link_libraries(test_convnd_bwd_data PRIVATE utility device_conv1d_bwd_data_instance device_conv2d_bwd_data_instance device_conv3d_bwd_data_instance)
test/convnd_bwd_data/convnd_bwd_data.cpp
View file @ aea62819
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/include/profile_convnd_bwd_data_impl.hpp"
#include "profiler/include/profile_conv_bwd_data_impl.hpp"
int main()
class TestConvndBwdData : public ::testing::Test
{
bool pass = true;
// check 1d
std::vector<ck::utils::conv::ConvParams> params;
params.push_back({1, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
params.push_back({1, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
params.push_back({1, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
for(auto& param : params)
{
pass &= ck::profiler::profile_convnd_bwd_data_impl<1,
float,
float,
float,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
pass &= ck::profiler::profile_convnd_bwd_data_impl<1,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
pass &= ck::profiler::profile_convnd_bwd_data_impl<1,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
pass &= ck::profiler::profile_convnd_bwd_data_impl<1,
int8_t,
int8_t,
int8_t,
int,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
}
protected:
std::vector<ck::utils::conv::ConvParam> conv_params;
};
// check 2d
params.clear();
params.push_back({2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
params.push_back({2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
params.push_back({2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
// 1d
TEST_F(TestConvndBwdData, Conv1dBwdData)
{
conv_params.clear();
conv_params.push_back({1, 1, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
conv_params.push_back({1, 1, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
conv_params.push_back({1, 1, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
for(auto& param : params)
for(auto& param : conv_params)
{
pass &= ck::profiler::profile_convnd_bwd_data_impl<2,
float,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
pass &= ck::profiler::profile_convnd_bwd_data_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
pass &= ck::profiler::profile_convnd_bwd_data_impl<2,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
pass &= ck::profiler::profile_convnd_bwd_data_impl<2,
int8_t,
int8_t,
int8_t,
int,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
bool pass;
// fp32
pass = ck::profiler::profile_conv_bwd_data_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// fp16
pass = ck::profiler::profile_conv_bwd_data_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// bf16
pass = ck::profiler::profile_conv_bwd_data_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// int8
pass = ck::profiler::profile_conv_bwd_data_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
int8_t,
int8_t,
int8_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
// check 3d
params.clear();
params.push_back(
{3, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
params.push_back(
{3, 128, 128, 256, {3, 3, 3}, {14, 14, 14}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
params.push_back(
{3, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
// 2d
TEST_F(TestConvndBwdData, Conv2dBwdData)
{
conv_params.clear();
conv_params.push_back({2, 1, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 1, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
for(auto& param : params)
for(auto& param : conv_params)
{
pass &= ck::profiler::profile_convnd_bwd_data_impl<3,
float,
float,
float,
float,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
pass &= ck::profiler::profile_convnd_bwd_data_impl<3,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
pass &= ck::profiler::profile_convnd_bwd_data_impl<3,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
pass &= ck::profiler::profile_convnd_bwd_data_impl<3,
int8_t,
int8_t,
int8_t,
int,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_);
bool pass;
// fp32
pass = ck::profiler::profile_conv_bwd_data_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// fp16
pass = ck::profiler::profile_conv_bwd_data_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// bf16
pass = ck::profiler::profile_conv_bwd_data_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// int8
pass = ck::profiler::profile_conv_bwd_data_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
int8_t,
int8_t,
int8_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
if(pass)
{
std::cout << "test convnd bwd : Pass" << std::endl;
return 0;
}
else
// 3d
TEST_F(TestConvndBwdData, Conv3dBwdData)
{
conv_params.clear();
conv_params.push_back(
{3, 1, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
conv_params.push_back(
{3, 1, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
conv_params.push_back(
{3, 1, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
for(auto& param : conv_params)
{
std::cout << "test convnd bwd: Fail " << std::endl;
return -1;
bool pass;
// fp32
pass = ck::profiler::profile_conv_bwd_data_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// fp16
pass = ck::profiler::profile_conv_bwd_data_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// bf16
pass = ck::profiler::profile_conv_bwd_data_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
// int8
pass = ck::profiler::profile_conv_bwd_data_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
int8_t,
int8_t,
int8_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param);
EXPECT_TRUE(pass);
}
}
test/convnd_bwd_weight/CMakeLists.txt
View file @ aea62819
add_test_executable(test_convnd_bwd_weight convnd_bwd_weight.cpp)
target_link_libraries(test_convnd_bwd_weight PRIVATE host_tensor device_convnd_bwd_weight_instance conv_util)
add_gtest_executable(test_convnd_bwd_weight convnd_bwd_weight.cpp)
target_link_libraries(test_convnd_bwd_weight PRIVATE utility device_conv1d_bwd_weight_instance device_conv2d_bwd_weight_instance device_conv3d_bwd_weight_instance)
test/convnd_bwd_weight/convnd_bwd_weight.cpp
View file @ aea62819
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <vector>
#include <gtest/gtest.h>
#include "test/convnd_fwd/conv_util.hpp"
#include "profiler/include/profile_convnd_bwd_weight_impl.hpp"
#include "profiler/include/profile_conv_bwd_weight_impl.hpp"
int test_self()
class TestConvndBwdWeight : public ::testing::Test
{
bool pass = true;
std::vector<ck::utils::conv::ConvParams> params;
protected:
std::vector<ck::utils::conv::ConvParam> conv_params;
};
params.push_back({1, 128, 256, 256, {1}, {7}, {2}, {1}, {0}, {0}});
params.push_back({1, 128, 256, 256, {3}, {14}, {1}, {1}, {1}, {1}});
params.push_back({1, 128, 256, 256, {1}, {3}, {1}, {1}, {0}, {0}});
// 1d
TEST_F(TestConvndBwdWeight, Conv1dBwdWeight)
{
conv_params.clear();
conv_params.push_back({1, 1, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
conv_params.push_back({1, 1, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
conv_params.push_back({1, 1, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
for(auto& param : params)
for(auto& param : conv_params)
{
// f32
pass &= ck::profiler::profile_convnd_bwd_weight_impl<1,
float,
float,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
bool pass;
// fp32
pass = ck::profiler::profile_conv_bwd_weight_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param,
2);
EXPECT_TRUE(pass);
// fp16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<1,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
pass = ck::profiler::profile_conv_bwd_weight_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param,
2);
EXPECT_TRUE(pass);
// bf16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<1,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
pass = ck::profiler::profile_conv_bwd_weight_impl<1,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param,
2);
EXPECT_TRUE(pass);
}
}
// check 2d
params.clear();
params.push_back({2, 128, 256, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
params.push_back({2, 128, 256, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
params.push_back({2, 128, 256, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
// 2d
TEST_F(TestConvndBwdWeight, Conv2dBwdWeight)
{
conv_params.clear();
conv_params.push_back({2, 1, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
conv_params.push_back({2, 1, 32, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
conv_params.push_back({2, 1, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
for(auto& param : params)
for(auto& param : conv_params)
{
// f32
pass &= ck::profiler::profile_convnd_bwd_weight_impl<2,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
bool pass;
// fp32
pass = ck::profiler::profile_conv_bwd_weight_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param,
2);
EXPECT_TRUE(pass);
// fp16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
pass = ck::profiler::profile_conv_bwd_weight_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param,
2);
EXPECT_TRUE(pass);
// bf16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<2,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
pass = ck::profiler::profile_conv_bwd_weight_impl<2,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param,
2);
EXPECT_TRUE(pass);
}
}
// check 2d
params.clear();
params.push_back(
{3, 128, 256, 256, {1, 1, 1}, {4, 4, 4}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
params.push_back(
{3, 128, 256, 256, {3, 3, 3}, {4, 4, 8}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
params.push_back(
{3, 128, 256, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
for(auto& param : params)
// 3d
TEST_F(TestConvndBwdWeight, Conv3dBwdWeight)
{
conv_params.clear();
conv_params.push_back(
{3, 1, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
conv_params.push_back(
{3, 1, 32, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
conv_params.push_back(
{3, 1, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
for(auto& param : conv_params)
{
// f32
pass &= ck::profiler::profile_convnd_bwd_weight_impl<3,
float,
float,
float,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
bool pass;
// fp32
pass = ck::profiler::profile_conv_bwd_weight_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
float,
float,
float>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param,
2);
EXPECT_TRUE(pass);
// fp16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<3,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
pass = ck::profiler::profile_conv_bwd_weight_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
ck::half_t,
ck::half_t,
ck::half_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param,
2);
EXPECT_TRUE(pass);
// bf16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<3,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
}
return pass;
}
int main()
{
// int data_type = 1;
// int init_method = 1;
bool pass = true;
pass = test_self();
if(pass)
{
std::cout << "test conv2d bwd weight : Pass" << std::endl;
return 0;
}
else
{
std::cout << "test conv2d bwd weight: Fail " << std::endl;
return -1;
pass = ck::profiler::profile_conv_bwd_weight_impl<3,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t>(true, // do_verification
1, // init_method
false, // do_log
false, // time_kernel
param,
2);
EXPECT_TRUE(pass);
}
}
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