C++/FastMathFun: Avoid SSE-AVX transition penalties

Author: GHScan

C++/FastMathFun/fma_dot.s

@@ -45,7 +45,7 @@ fma_dot:
     vaddps %ymm0,%ymm8,%ymm0
 
     /* TODO: horizontally sum ymm0 */
-
+    vzeroupper
     retq
 
 .data

C++/FastMathFun/fma_gemm48.s

@@ -68,6 +68,7 @@ fma_gemm48:
         dec %rcx
         jnz loop_row
 
+    vzeroupper
     pop %rbx
     retq
 

C++/FastMathFun/fma_gemm96.s

@@ -186,6 +186,7 @@ fma_gemm96:
         dec %rcx
         jnz loop_row
 
+    vzeroupper
     pop %rbx
     retq
 

C++/FastMathFun/gen.py

@@ -28,6 +28,7 @@ def gen_gemm48():
     print('        dec %rcx')
     print('        jnz loop_row')
 
+    print('    vzeroupper')
     print('    retq')
 
     print('.data')
@@ -69,6 +70,7 @@ def gen_gemm96():
     print('        dec %rcx')
     print('        jnz loop_row')
 
+    print('    vzeroupper')
     print('    retq')
 
     print('.data')
@@ -129,6 +131,7 @@ def gen_transform_cols(working_set):
 
     if working_set != 'L1':
         print('    sfence')
+    print('    vzeroupper')
     print('    pop %rbx')
     print('    retq')