checkasm: x86: More accurate ymm/zmm measurements

YMM and ZMM registers on x86 are turned off to save power when they haven't
been used for some period of time. When they are used there will be a
"warmup" period during which performance will be reduced and inconsistent
which is problematic when trying to benchmark individual functions.

Periodically issue "dummy" instructions that uses those registers to
prevent them from being powered down. The end result is more consitent
benchmark results.

diff --git a/tools/checkasm-a.asm b/tools/checkasm-a.asm
index d1da9b7900bdcc5b26403ead138307efbf390237..a9f749315afea88586299e22cc10599b6c665103 100644 (file)
--- a/tools/checkasm-a.asm
+++ b/tools/checkasm-a.asm
@@ -225,3 +225,14 @@ cglobal stack_pagealign, 2,2
     leave
     RET
 
+; Trigger a warmup of vector units
+%macro WARMUP 0
+cglobal checkasm_warmup, 0,0
+    xorps m0, m0
+    RET
+%endmacro
+
+INIT_YMM avx
+WARMUP
+INIT_ZMM avx512
+WARMUP

diff --git a/tools/checkasm.c b/tools/checkasm.c
index a2c2e492a6de69a7bea1ee8d18467491cdcf40b3..75899dfef8395c967debea7bce6cf72f98fc5255 100644 (file)
--- a/tools/checkasm.c
+++ b/tools/checkasm.c
@@ -222,8 +222,18 @@ static void print_bench(void)
         }
 }
 
+/* YMM and ZMM registers on x86 are turned off to save power when they haven't been
+ * used for some period of time. When they are used there will be a "warmup" period
+ * during which performance will be reduced and inconsistent which is problematic when
+ * trying to benchmark individual functions. We can work around this by periodically
+ * issuing "dummy" instructions that uses those registers to keep them powered on. */
+static void (*simd_warmup_func)( void ) = NULL;
+#define simd_warmup() do { if( simd_warmup_func ) simd_warmup_func(); } while( 0 )
+
 #if ARCH_X86 || ARCH_X86_64
 int x264_stack_pagealign( int (*func)(), int align );
+void x264_checkasm_warmup_avx( void );
+void x264_checkasm_warmup_avx512( void );
 
 /* detect when callee-saved regs aren't saved
  * needs an explicit asm check because it only sometimes crashes in normal use. */
@@ -258,6 +268,7 @@ void x264_checkasm_stack_clobber( uint64_t clobber, ... );
 #define call_a1(func,...) ({ \
     uint64_t r = (rand() & 0xffff) * 0x0001000100010001ULL; \
     x264_checkasm_stack_clobber( r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r ); /* max_args+6 */ \
+    simd_warmup(); \
     x264_checkasm_call(( intptr_t(*)())func, &ok, 0, 0, 0, 0, __VA_ARGS__ ); })
 #elif ARCH_AARCH64 && !defined(__APPLE__)
 void x264_checkasm_stack_clobber( uint64_t clobber, ... );
@@ -285,6 +296,7 @@ void x264_checkasm_stack_clobber( uint64_t clobber, ... );
         call_a1(func, __VA_ARGS__);\
         for( int ti = 0; ti < (cpu?BENCH_RUNS:BENCH_RUNS/4); ti++ )\
         {\
+            simd_warmup();\
             uint32_t t = read_time();\
             func(__VA_ARGS__);\
             func(__VA_ARGS__);\
@@ -2785,6 +2797,14 @@ static int check_all_flags( void )
     int ret = 0;
     int cpu0 = 0, cpu1 = 0;
     uint32_t cpu_detect = x264_cpu_detect();
+#if ARCH_X86 || ARCH_X86_64
+    if( cpu_detect & X264_CPU_AVX512 )
+        simd_warmup_func = x264_checkasm_warmup_avx512;
+    else if( cpu_detect & X264_CPU_AVX )
+        simd_warmup_func = x264_checkasm_warmup_avx;
+#endif
+    simd_warmup();
+
 #if HAVE_MMX
     if( cpu_detect & X264_CPU_MMX2 )
     {