Improve benchmarks

hoglet67 · Nov 15, 2021 · c55af86 · c55af86
1 parent cb3eaee
commit c55af86
Show file tree

Hide file tree

Showing 2 changed files with 20 additions and 4 deletions.
diff --git a/src/rgb_to_fb.S b/src/rgb_to_fb.S
@@ -2137,8 +2137,10 @@ benchmarkRAM:
         cmp    r0, #2
         beq    gpu_bench
         cmp    r0, #3
-        beq    mbox_bench_1
+        beq    gpio_bench
         cmp    r0, #4
+        beq    mbox_bench_1
+        cmp    r0, #5
         beq    mbox_bench_3
         // RAM address in r0 returns with time in r0
         mov    r1, r0
@@ -2163,6 +2165,19 @@ benchloop2:
         rsbmi  r0, r0, #1
         pop   {r1-r12, pc}
 
+gpio_bench:
+        ldr    r4, =GPLEV0
+        ldr    r1, =100000
+        READ_CYCLE_COUNTER r6
+gpio_bench_loop:
+        ldr    r8, [r4]
+        subs   r1, r1, #1
+        bne    gpio_bench_loop
+        READ_CYCLE_COUNTER r7
+        subs   r0, r7, r6
+        rsbmi  r0, r0, #1
+        pop   {r1-r12, pc}
+
 mbox_bench_1:
         ldr    r4, =GPU_DATA_0
         ldr    r1, =100000

diff --git a/src/rgb_to_hdmi.c b/src/rgb_to_hdmi.c
@@ -3045,9 +3045,10 @@ void rgb_to_hdmi_main() {
    capinfo->sync_type = SYNC_BIT_COMPOSITE_SYNC;
    current_display_buffer = 0;
 
-   log_info("ARM side single MBOX read = %d ns, Arm side triple MBOX read = %d ns", (int)((double) benchmarkRAM(3) * 1000 / cpuspeed / 100000 + 0.5), (int)((double) benchmarkRAM(4) * 1000 / cpuspeed / 300000 + 0.5));
-   log_info("GPU side GPIO read = %d ns, GPU side MBOX write = %d ns", (int)((double) benchmarkRAM(1) * 1000 / cpuspeed / 100000 + 0.5), (int)((double) benchmarkRAM(2) * 1000 / cpuspeed / 100000 + 0.5));
-   log_info("Cached RAM read = %d ns, Uncached screen RAM read = %d ns", (int)((double) benchmarkRAM(0x2000000) * 1000 / cpuspeed / 100000 + 0.5), (int)((double) benchmarkRAM(0x01e000000) * 1000 / cpuspeed / 100000 + 0.5));
+   int triple = (int)((double) benchmarkRAM(5) * 1000 / cpuspeed / 100000 + 0.5);
+   log_info("ARM: GPIO read = %dns, MBOX read = %dns, Triple MBOX read = %dns (%dns/word)", (int)((double) benchmarkRAM(3) * 1000 / cpuspeed / 100000 + 0.5), (int)((double) benchmarkRAM(4) * 1000 / cpuspeed / 100000 + 0.5), triple, triple / 3);
+   log_info("GPU: GPIO read = %dns, MBOX write = %dns", (int)((double) benchmarkRAM(1) * 1000 / cpuspeed / 100000 + 0.5), (int)((double) benchmarkRAM(2) * 1000 / cpuspeed / 100000 + 0.5));
+   log_info("RAM: Cached read = %dns, Uncached screen read = %dns", (int)((double) benchmarkRAM(0x2000000) * 1000 / cpuspeed / 100000 + 0.5), (int)((double) benchmarkRAM(0x01e000000) * 1000 / cpuspeed / 100000 + 0.5));
 
 #ifdef USE_ARM_CAPTURE
    log_info("Running ARM capture build");