From a81768561ca1de2a947a45f224bf91d3913c1e2a Mon Sep 17 00:00:00 2001
From: elliottk <elliottk@google.com>
Date: Wed, 19 Dec 2018 13:35:30 -0800
Subject: [PATCH] Improve accuracy of benchmarking

For small code regions, readtsc can give inaccurate results because it does
not account for out-of-order execution. Add x86_tsc_start and x86_tsc_end
that account for this, according to the white paper at

https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/ia-32-ia-64-benchmark-code-execution-paper.pdf

Using x86_tsc_start/end will also add in several more instructions; I imagine
this is negligible.

Change-Id: I54a1c8fa7977c34bf91b422369c96f036c93a08a
---
 vpx_ports/x86.h | 60 ++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 55 insertions(+), 5 deletions(-)

diff --git a/vpx_ports/x86.h b/vpx_ports/x86.h
index f6aac1832..58eeb7b63 100644
--- a/vpx_ports/x86.h
+++ b/vpx_ports/x86.h
@@ -223,11 +223,26 @@ static INLINE int x86_simd_caps(void) {
   return flags & mask;
 }
 
-// Note:
-//  32-bit CPU cycle counter is light-weighted for most function performance
-//  measurement. For large function (CPU time > a couple of seconds), 64-bit
-//  counter should be used.
-// 32-bit CPU cycle counter
+// Fine-Grain Measurement Functions
+//
+// If you are timing a small region of code, access the timestamp counter
+// (TSC) via:
+//
+// unsigned int start = x86_tsc_start();
+//   ...
+// unsigned int end = x86_tsc_end();
+// unsigned int diff = end - start;
+//
+// The start/end functions introduce a few more instructions than using
+// x86_readtsc directly, but prevent the CPU's out-of-order execution from
+// affecting the measurement (by having earlier/later instructions be evaluated
+// in the time interval). See the white paper, "How to Benchmark Code
+// Execution Times on Intel® IA-32 and IA-64 Instruction Set Architectures" by
+// Gabriele Paoloni for more information.
+//
+// If you are timing a large function (CPU time > a couple of seconds), use
+// x86_readtsc64 to read the timestamp counter in a 64-bit integer. The
+// out-of-order leakage that can occur is minimal compared to total runtime.
 static INLINE unsigned int x86_readtsc(void) {
 #if defined(__GNUC__) && __GNUC__
   unsigned int tsc;
@@ -264,6 +279,41 @@ static INLINE uint64_t x86_readtsc64(void) {
 #endif
 }
 
+// 32-bit CPU cycle counter with a partial fence against out-of-order execution.
+static INLINE unsigned int x86_readtscp(void) {
+#if defined(__GNUC__) && __GNUC__
+  unsigned int tscp;
+  __asm__ __volatile__("rdtscp\n\t" : "=a"(tscp) :);
+  return tscp;
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+  unsigned int tscp;
+  asm volatile("rdtscp\n\t" : "=a"(tscp) :);
+  return tscp;
+#elif defined(_MSC_VER)
+  unsigned int ui;
+  return (unsigned int)__rdtscp(&ui);
+#else
+#if ARCH_X86_64
+  return (unsigned int)__rdtscp();
+#else
+  __asm rdtscp;
+#endif
+#endif
+}
+
+static INLINE unsigned int x86_tsc_start(void) {
+  unsigned int reg_eax, reg_ebx, reg_ecx, reg_edx;
+  cpuid(0, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
+  return x86_readtsc();
+}
+
+static INLINE unsigned int x86_tsc_end(void) {
+  uint32_t v = x86_readtscp();
+  unsigned int reg_eax, reg_ebx, reg_ecx, reg_edx;
+  cpuid(0, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
+  return v;
+}
+
 #if defined(__GNUC__) && __GNUC__
 #define x86_pause_hint() __asm__ __volatile__("pause \n\t")
 #elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
-- 
2.40.0