diff --git a/benchmark.c b/benchmark.c
index f6a21da97d..082348ec18 100644
--- a/benchmark.c
+++ b/benchmark.c
@@ -26,6 +26,26 @@
 #include <sys/time.h>
 #endif
 
+#define MAX_COUNTERS 64
+static struct rarch_perf_counter *perf_counters[MAX_COUNTERS];
+static unsigned perf_ptr;
+
+void rarch_perf_register(struct rarch_perf_counter *perf)
+{
+   if (perf_ptr >= MAX_COUNTERS)
+      return;
+
+   perf_counters[perf_ptr++] = perf;
+   perf->registered = true;
+}
+
+void rarch_perf_log(void)
+{
+   RARCH_LOG("[PERF]: Performance counters:\n");
+   for (unsigned i = 0; i < perf_ptr; i++)
+      RARCH_PERFORMANCE_LOG(perf_counters[i]->ident, *perf_counters[i]);
+}
+
 rarch_perf_tick_t rarch_get_perf_counter(void)
 {
    rarch_perf_tick_t time = 0;
diff --git a/benchmark.h b/benchmark.h
index c46c60a628..31375181ce 100644
--- a/benchmark.h
+++ b/benchmark.h
@@ -28,23 +28,45 @@ typedef unsigned long long rarch_perf_tick_t;
 
 typedef struct rarch_perf_counter
 {
+   const char *ident;
    rarch_perf_tick_t start;
    rarch_perf_tick_t total;
-   unsigned call_cnt;
+   uint64_t call_cnt;
+
+   bool registered;
 } rarch_perf_counter_t;
 
 rarch_perf_tick_t rarch_get_perf_counter(void);
+void rarch_perf_register(struct rarch_perf_counter *perf);
+void rarch_perf_log(void);
 
 #ifdef PERF_TEST
 
-#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) /* TODO: we should increment at PERFORMANCE_START instead of PERFORMANCE_STOP since sometimes we cannot assume that we don't prematurely exit out of the function */
+#define RARCH_PERFORMANCE_INIT(X)  static rarch_perf_counter_t X = {#X}; \
+   do { \
+      if (!(X).registered) \
+         rarch_perf_register(&(X)); \
+   } while(0)
+
+#define RARCH_PERFORMANCE_START(X) do { \
+   (X).call_cnt++; \
+   (X).start  = rarch_get_perf_counter(); \
+} while(0)
+
+#define RARCH_PERFORMANCE_STOP(X) do { \
+   (X).total += rarch_get_perf_counter() - (X).start; \
+} while(0)
 
 #ifdef _WIN32
-#define RARCH_PERFORMANCE_LOG(functionname, X) RARCH_LOG("[PERF]: Avg (%s): %I64u ticks.\n", functionname, (X).total / (X).call_cnt)
+#define RARCH_PERFORMANCE_LOG(functionname, X) RARCH_LOG("[PERF]: Avg (%s): %I64u ticks, %I64u runs.\n", \
+      functionname, \
+      (X).total / (X).call_cnt, \
+      (X).call_cnt)
 #else
-#define RARCH_PERFORMANCE_LOG(functionname, X) RARCH_LOG("[PERF]: Avg (%s): %llu ticks.\n", functionname, (X).total / (X).call_cnt)
+#define RARCH_PERFORMANCE_LOG(functionname, X) RARCH_LOG("[PERF]: Avg (%s): %llu ticks, %llu runs.\n", \
+      functionname, \
+      (long long unsigned)((X).total / (X).call_cnt), \
+      (long long unsigned)((X).call_cnt))
 #endif
 
 #else
diff --git a/gfx/gl.c b/gfx/gl.c
index 797ec5fd40..c758166612 100644
--- a/gfx/gl.c
+++ b/gfx/gl.c
@@ -1107,7 +1107,6 @@ static bool gl_frame(void *data, const void *frame, unsigned width, unsigned hei
       RARCH_PERFORMANCE_START(copy_frame);
       gl_copy_frame(gl, frame, width, height, pitch);
       RARCH_PERFORMANCE_STOP(copy_frame);
-      RARCH_PERFORMANCE_LOG("gl_copy_frame", copy_frame);
    }
 
    struct gl_tex_info tex_info = {0};
@@ -1144,7 +1143,6 @@ static bool gl_frame(void *data, const void *frame, unsigned width, unsigned hei
 #endif
 
    RARCH_PERFORMANCE_STOP(frame_run);
-   RARCH_PERFORMANCE_LOG("gl_frame", frame_run);
 
 #ifdef RARCH_CONSOLE
    if (!gl->block_swap)
diff --git a/gfx/scaler/scaler.c b/gfx/scaler/scaler.c
index bd92a7c95b..5f2bc6d978 100644
--- a/gfx/scaler/scaler.c
+++ b/gfx/scaler/scaler.c
@@ -192,9 +192,6 @@ void scaler_ctx_gen_reset(struct scaler_ctx *ctx)
 void scaler_ctx_scale(struct scaler_ctx *ctx,
       void *output, const void *input)
 {
-   RARCH_PERFORMANCE_INIT(scaler_perf);
-   RARCH_PERFORMANCE_START(scaler_perf);
-
    if (ctx->unscaled) // Just perform straight pixel conversion.
    {
       ctx->direct_pixconv(output, input,
@@ -262,8 +259,5 @@ void scaler_ctx_scale(struct scaler_ctx *ctx,
       else
          ctx->scaler_vert(ctx, output, ctx->out_stride);
    }
-
-   RARCH_PERFORMANCE_STOP(scaler_perf);
-   RARCH_PERFORMANCE_LOG("Scaler", scaler_perf);
 }
 
diff --git a/gfx/vg.c b/gfx/vg.c
index 582e314df6..792f557270 100644
--- a/gfx/vg.c
+++ b/gfx/vg.c
@@ -397,7 +397,6 @@ static bool vg_frame(void *data, const void *frame, unsigned width, unsigned hei
    RARCH_PERFORMANCE_START(vg_image);
    vg_copy_frame(vg, frame, width, height, pitch);
    RARCH_PERFORMANCE_STOP(vg_image);
-   RARCH_PERFORMANCE_LOG("vg_copy_frame", vg_image);
 
    vgDrawImage(vg->mImage);
 
@@ -409,8 +408,6 @@ static bool vg_frame(void *data, const void *frame, unsigned width, unsigned hei
 #endif
 
    RARCH_PERFORMANCE_STOP(vg_fr);
-   RARCH_PERFORMANCE_LOG("vg_frame", vg_fr);
-
    vg->driver->swap_buffers();
 
    return true;
diff --git a/retroarch.c b/retroarch.c
index 428758cdad..e9a5f441b9 100644
--- a/retroarch.c
+++ b/retroarch.c
@@ -272,7 +272,6 @@ static void video_frame(const void *data, unsigned width, unsigned height, size_
       data = driver.scaler_out;
       pitch = driver.scaler.out_stride;
       RARCH_PERFORMANCE_STOP(video_frame_conv);
-      RARCH_PERFORMANCE_LOG("video_frame_conv()", video_frame_conv);
    }
 
    // Slightly messy code,
@@ -2772,6 +2771,11 @@ int rarch_main(int argc, char *argv[])
    while (rarch_main_iterate());
    rarch_main_deinit();
    rarch_deinit_msg_queue();
+
+#ifdef PERF_TEST
+   rarch_perf_log();
+#endif
+
    rarch_main_clear_state();
    return 0;
 }