Fix seg fault on avx2: rbp was over-written by new code

r-devulap · Sep 11, 2024 · 6743659 · 6743659
1 parent 9a0b45b
commit 6743659
Showing 1 changed file with 4 additions and 4 deletions.
diff --git a/onnxruntime/core/mlas/lib/amd64/QgemmU8X8KernelAvx2.asm b/onnxruntime/core/mlas/lib/amd64/QgemmU8X8KernelAvx2.asm
@@ -1074,7 +1074,7 @@ LSkipAccumulateOutputMasked8xNBlock1:
         push_reg r12
         push_reg r13
         push_reg r14
-        alloc_stack (GemmU8X8KernelFrame.SavedR13)
+        alloc_stack (GemmU8X8KernelFrame.SavedR14)
         save_xmm128 xmm6,GemmU8X8KernelFrame.SavedXmm6
         save_xmm128 xmm7,GemmU8X8KernelFrame.SavedXmm7
         save_xmm128 xmm8,GemmU8X8KernelFrame.SavedXmm8
@@ -1101,8 +1101,8 @@ LSkipAccumulateOutputMasked8xNBlock1:
         mov     r13,GemmU8X8KernelFrame.ZeroPointB[rsp]
         vpcmpeqw ymm12,ymm12,ymm12          ; generate 256-bit word vector [0xFFFF]
         vpsrlw  ymm12,ymm12,15              ; generate 256-bit word vector [0x0001]
-        lea     rbp,[r9*8]
-        lea     r14,[rbp*2]
+        lea     r14,[r9*8]
+        lea     r14,[r14*2]
         cmp     DWORD PTR GemmU8X8KernelFrame.PreviousP1Home[rsp],0
         je      CheckCountM4OrMore          ; U8S8 AVX2 kernel requires extra registers
 
@@ -1147,7 +1147,7 @@ ExitKernel:
         movaps  xmm13,GemmU8X8KernelFrame.SavedXmm13[rsp]
         movaps  xmm14,GemmU8X8KernelFrame.SavedXmm14[rsp]
         movaps  xmm15,GemmU8X8KernelFrame.SavedXmm15[rsp]
-        add     rsp,(GemmU8X8KernelFrame.SavedR13)
+        add     rsp,(GemmU8X8KernelFrame.SavedR14)
 
         BEGIN_EPILOGUE