wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'ae' into ae-hopper

Browse Source
This commit is contained in:
Richard Yan
2025-01-30 01:35:10 -08:00
parent e4f8f3481c 9cdee597b6
commit 3cd6aacc17
2 changed files with 7 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
kernels/flash_attention/compile_flash.sh
View File

@@ -26,7 +26,7 @@ ln -sf input.b.rand.fp32.seqlen1024headdim64.row.bin input.b.bin
ln -sf input.c.rand.fp32.seqlen1024headdim64.row.bin input.c.bin ln -sf input.c.rand.fp32.seqlen1024headdim64.row.bin input.c.bin
for arch in "${archs[@]}"; do for arch in "${archs[@]}"; do
git checkout ae-$arch git checkout ae-flash-$arch
# re-compile libvortexrt.a # re-compile libvortexrt.a
# FIXME after restructure # FIXME after restructure
@@ -34,13 +34,11 @@ for arch in "${archs[@]}"; do
make make
popd popd
for dim in "${dims[@]}"; do echo "compiling flash_attn kernel for $arch with seqlen 1024, headdim 64"
echo "compiling flash_attn kernel for $arch with seqlen 1024, headdim 64"
# touch source file to force re-building, as the Makefile does not track # touch source file to force re-building, as the Makefile does not track
# binary changes # binary changes
touch kernel.cpp touch kernel.cpp
make CONFIG=flash.$arch.seqlen1024.headdim64 make CONFIG=flash.$arch.seqlen1024.headdim64
done
done done

2
kernels/sgemm_gemmini_dma/kernel.cpp
View File

@@ -53,7 +53,7 @@
#define PRINTF(...) sprintf(PRINT_BUF, __VA_ARGS__) #define PRINTF(...) sprintf(PRINT_BUF, __VA_ARGS__)
// #define PRINTF(...) vx_printf(__VA_ARGS__) // #define PRINTF(...) vx_printf(__VA_ARGS__)
#define SWISH(beta, x) ((x) / (1 + exp(-(beta) * (x)))) #define SWISH(beta, x) ((x) / (1 + exp(-(beta) * (x))))
// #define POWER #define POWER
typedef uint16_t smem_elem_t; typedef uint16_t smem_elem_t;
// typedef float smem_elem_t; // typedef float smem_elem_t;