Merge branch 'ae' into ae-volta

kernels/flash_attention/compile_flash.sh

@@ -26,7 +26,8 @@ ln -sf input.b.rand.fp32.seqlen1024headdim64.row.bin input.b.bin
 ln -sf input.c.rand.fp32.seqlen1024headdim64.row.bin input.c.bin
 
 for arch in "${archs[@]}"; do
-    git checkout ae-$arch
+    git checkout ae-flash-$arch
+    git pull
 
     # re-compile libvortexrt.a
     # FIXME after restructure
@@ -34,7 +35,6 @@ for arch in "${archs[@]}"; do
     make
     popd
 
-    for dim in "${dims[@]}"; do
     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
@@ -42,5 +42,4 @@ for arch in "${archs[@]}"; do
     touch kernel.cpp
 
     make CONFIG=flash.$arch.seqlen1024.headdim64
-    done
 done

kernels/sgemm_gemmini_dma/compile_virgo.sh

@@ -1,5 +1,9 @@
 #!/bin/sh
 
+# hopper and virgo has the same SIMT configurations
+git checkout ae-hopper
+git pull
+
 if [ ! -f input.a.rand01.fp16.m256n256k256.row.bin ]; then
     echo "input binaries not found, generating operands"
     python3 generate_operands.py

kernels/sgemm_gemmini_dma/kernel.cpp

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

kernels/sgemm_tcore/compile_tcore.sh

@@ -53,6 +53,7 @@ done
 
 for arch in "${archs[@]}"; do
     git checkout ae-$arch
+    git pull
 
     # re-compile libvortexrt.a
     # FIXME after restructure