ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-04 21:51:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

[rocm-libraries] ROCm/rocm-libraries#6305 (commit 19e10a0)

Browse Source 
[CK] Remove obsolete benchmark_fwd_v3.sh script and README
 reference (#6305)

The tile_example_fmha_fwd_v3 target no longer exists in this project,
making this benchmark script non-functional.
This commit is contained in:
Po Yen Chen
2026-04-15 07:38:36 +00:00
committed by assistant-librarian[bot]
parent 7dcc606adc
commit 2b83413b8d
2 changed files with 1 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
example/ck_tile/01_fmha/script/benchmark_fwd_v3.sh
View File

@@ -1,46 +0,0 @@
#!/bin/sh
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
# TODO: run this script from CK root or build directory
EXE="$(find . -name tile_example_fmha_fwd_v3 -type f | head -n 1)"
VALID=0
for causal in 0 1 ; do
for prec in "fp16" "bf16" ; do
for hdim in 128 ; do
for perm in 0 ; do
$EXE -prec=$prec -b=32 -h=16 -s=512 -d=$hdim -causal=$causal -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=16 -h=16 -s=1024 -d=$hdim -causal=$causal -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=8 -h=16 -s=2048 -d=$hdim -causal=$causal -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=4 -h=16 -s=4096 -d=$hdim -causal=$causal -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=2 -h=16 -s=8192 -d=$hdim -causal=$causal -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=1 -h=16 -s=16384 -d=$hdim -causal=$causal -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=1 -h=64 -s=16384 -d=$hdim -causal=$causal -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=1 -h=16 -h_k=1 -s=65536 -d=$hdim -causal=$causal -iperm=$perm -operm=$perm -v=$VALID
$EXE -prec=$prec -b=1 -h=40 -s=37200 -d=$hdim -causal=$causal -iperm=$perm -operm=$perm -v=$VALID
done
done
done
done
# Padding benchmark comparisons for v3 (batch mode only)
# ==== V3 Padding Benchmarks: batch mode (baseline vs low/med/high pad) ====
prec="fp16"
base_v3_args="-prec=$prec -b=4 -h=16 -d=128 -s=1024 -mask=0 -iperm=0 -operm=0 -v=$VALID"
# baseline (no pad)
$EXE $base_v3_args
# low pad (≈90–95% effective)
$EXE $base_v3_args -q_eff_lens=1024,960,992,896 -kv_eff_lens=1024,960,992,896
# medium pad (≈60–75% effective)
$EXE $base_v3_args -q_eff_lens=896,768,512,640 -kv_eff_lens=896,768,512,640
# high pad (≈30–40% effective)
$EXE $base_v3_args -q_eff_lens=512,384,256,320 -kv_eff_lens=512,384,256,320