Update benchmarks for GCC + x86.

Compile with make PERF_TEST=1 to enable performance logging.

Makefile

@@ -137,8 +137,9 @@ ifeq ($(SCALER_NO_SIMD), 1)
    DEFINES += -DSCALER_NO_SIMD
 endif
 
-ifeq ($(SCALER_PERF), 1)
+ifeq ($(PERF_TEST), 1)
-   DEFINES += -DSCALER_PERF
+   DEFINES += -DPERF_TEST
+   OBJ += benchmark.o
 endif
 
 ifeq ($(HAVE_SDL), 1)

Makefile.win

@@ -68,6 +68,11 @@ ifeq ($(SCALER_NO_SIMD), 1)
    DEFINES += -DSCALER_NO_SIMD
 endif
 
+ifeq ($(PERF_TEST), 1)
+   DEFINES += -DPERF_TEST
+   OBJ += benchmark.o
+endif
+
 ifeq ($(HAVE_SDL), 1)
    OBJ += gfx/scaler/scaler.o gfx/scaler/pixconv.o gfx/scaler/scaler_int.o gfx/scaler/filter.o
 else ifeq ($(HAVE_FFMPEG), 1)

benchmark.c

@@ -14,7 +14,7 @@
  *  If not, see <http://www.gnu.org/licenses/>.
  */
 
-#include <stdio.h>
+#include "benchmark.h"
 
 #if defined(__CELLOS_LV2__) || defined(GEKKO)
 #ifndef _PPU_INTRINSICS_H
@@ -24,23 +24,31 @@
 #include <PPCIntrinsics.h>
 #endif
 
-unsigned long long rarch_get_performance_counter(void)
+rarch_perf_tick_t rarch_get_perf_counter(void)
 {
-   unsigned long long time = 0;
+   rarch_perf_tick_t time = 0;
 #ifdef _XBOX1
+
 #define rdtsc	__asm __emit 0fh __asm __emit 031h
    LARGE_INTEGER time_tmp;
    rdtsc;
    __asm	mov	time_tmp.LowPart, eax;
    __asm	mov	time_tmp.HighPart, edx;
    time = time_tmp.QuadPart;
-#elif defined(__i386__) || defined(__i486__) || defined(__x86_64__)
+
-   uint64_t lo, hi;
+#elif defined(__GNUC__)
-   __asm__ __volatile__("rdtsc" : "=a" (lo), "=d" (hi));
+
-   time =  ((((uint64 t)hi) << 32) | ((uint64 t)lo) );
+#if defined(__i386__) || defined(__i486__)
+   asm volatile ("rdtsc" : "=A" (time));
+#elif defined(__x86_64__)
+   unsigned a, d;
+   asm volatile ("rdtsc" : "=a" (a), "=d" (d));
+   time = (rarch_perf_tick_t)a | ((rarch_perf_tick_t)d << 32);
+#endif
+
 #elif defined(__CELLOS_LV2__) || defined(GEKKO) || defined(_XBOX360)
    time = __mftb();
 #endif
-   (void)time;
+
    return time;
 }

benchmark.h

@@ -17,22 +17,44 @@
 #ifndef _RARCH_BENCHMARK_H
 #define _RARCH_BENCHMARK_H
 
-typedef struct performance_counter_t
+#include "general.h"
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <stdint.h>
+typedef unsigned long long rarch_perf_tick_t;
+
+typedef struct rarch_perf_counter
 {
-   unsigned long long start;
+   rarch_perf_tick_t start;
-   unsigned long long stop;
+   rarch_perf_tick_t total;
-} performance_counter_t;
+   unsigned call_cnt;
+} rarch_perf_counter_t;
 
-unsigned long long rarch_get_performance_counter(void);
+rarch_perf_tick_t rarch_get_perf_counter(void);
 
-#define RARCH_PERFORMANCE_INIT(X)  performance_counter_t (X)
+#ifdef PERF_TEST
-#define RARCH_PERFORMANCE_START(X) ((X).start = rarch_get_performance_counter())
+
-#define RARCH_PERFORMANCE_STOP(X)  ((X).stop  = rarch_get_performance_counter() - (X).start)
+#define RARCH_PERFORMANCE_INIT(X)  static rarch_perf_counter_t X
+#define RARCH_PERFORMANCE_START(X) ((X).start  = rarch_get_perf_counter())
+#define RARCH_PERFORMANCE_STOP(X)  do { (X).total += rarch_get_perf_counter() - (X).start; (X).call_cnt++; } while(0)
 
 #ifdef _WIN32
-#define RARCH_PERFORMANCE_LOG(functionname, X)   RARCH_LOG("Time taken (%s): %I64u.\n", functionname, (X).stop)
+#define RARCH_PERFORMANCE_LOG(functionname, X) RARCH_LOG("[PERF]: Avg (%s): %I64u ticks.\n", functionname, (X).total / (X).call_cnt)
 #else
-#define RARCH_PERFORMANCE_LOG(functionname, X)   RARCH_LOG("Time taken (%s): %llu.\n", functionname, (X).stop)
+#define RARCH_PERFORMANCE_LOG(functionname, X) RARCH_LOG("[PERF]: Avg (%s): %llu ticks.\n", functionname, (X).total / (X).call_cnt)
+#endif
+
+#else
+
+#define RARCH_PERFORMANCE_INIT(X)
+#define RARCH_PERFORMANCE_START(X)
+#define RARCH_PERFORMANCE_STOP(X)
+#define RARCH_PERFORMANCE_LOG(functionname, X)
+
 #endif
 
 #endif
+

gfx/scaler/scaler.c

@@ -21,6 +21,7 @@
 #include <string.h>
 #include <stdio.h>
 #include <math.h>
+#include "../../benchmark.h"
 
 #ifdef SCALER_PERF
 #include <time.h>
@@ -189,10 +190,8 @@ void scaler_ctx_gen_reset(struct scaler_ctx *ctx)
 void scaler_ctx_scale(struct scaler_ctx *ctx,
       void *output, const void *input)
 {
-#ifdef SCALER_PERF
+   RARCH_PERFORMANCE_INIT(scaler_perf);
-   struct timespec start_tv, end_tv;
+   RARCH_PERFORMANCE_START(scaler_perf);
-   clock_gettime(CLOCK_MONOTONIC, &start_tv);
-#endif
 
    if (ctx->unscaled) // Just perform straight pixel conversion.
    {
@@ -262,10 +261,7 @@ void scaler_ctx_scale(struct scaler_ctx *ctx,
          ctx->scaler_vert(ctx, output, ctx->out_stride);
    }
 
-#ifdef SCALER_PERF
+   RARCH_PERFORMANCE_STOP(scaler_perf);
-   clock_gettime(CLOCK_MONOTONIC, &end_tv);
+   RARCH_PERFORMANCE_LOG("Scaler", scaler_perf);
-   ctx->elapsed_time_ms += (end_tv.tv_sec - start_tv.tv_sec) * 1000.0 + (end_tv.tv_nsec - start_tv.tv_nsec) / 1000000.0;
-   ctx->elapsed_frames++;
-#endif
 }