yuvbench: Apple Accelerate

3 files changed, 120 insertions, 7 deletions

diff --git a/yuvbench/build-macos-aarch64-clang.sh b/yuvbench/build-macos-aarch64-clang.sh
index 0a794fa..3aa3943 100755
--- a/yuvbench/build-macos-aarch64-clang.sh
+++ b/yuvbench/build-macos-aarch64-clang.sh
@@ -1,7 +1,11 @@
 #!/bin/sh
-CFLAGS="-Wall -Wextra -Wpedantic -O3 -g -DYUVBENCH_BAD"
-LFLAGS=""
+CFLAGS="-Wall -Wextra -Wpedantic -O3 -g -DYUVBENCH_ACCELERATE -DYUVBENCH_BAD"
+LFLAGS="-framework Accelerate"
 mkdir -p build
+set -x
 clang -o build/yuvbench.o $CFLAGS -c ./yuvbench.c
+clang -o build/yuvbench_accelerate.o $CFLAGS -c ./yuvbench_accelerate.c
+clang -o build/yuvbench_accelerate.S $CFLAGS -S ./yuvbench_accelerate.c
 clang -o build/yuvbench_bad.o $CFLAGS -c ./yuvbench_bad.c
-clang -o build/yuvbench $LFLAGS build/yuvbench.o build/yuvbench_bad.o
+clang -o build/yuvbench_bad.S $CFLAGS -S ./yuvbench_bad.c
+clang -o build/yuvbench $LFLAGS build/yuvbench.o build/yuvbench_accelerate.o build/yuvbench_bad.o
diff --git a/yuvbench/yuvbench.c b/yuvbench/yuvbench.c
index 1d34f0b..5ef344c 100644
--- a/yuvbench/yuvbench.c
+++ b/yuvbench/yuvbench.c
@@ -3,6 +3,9 @@
 #define KBENCH_IMPLEMENTATION
 #include "kbench.h"
 
+#ifdef YUVBENCH_ACCELERATE
+Backend yuvbench_accelerate(void);
+#endif
 #ifdef YUVBENCH_BAD
 Backend yuvbench_bad(void);
 #endif
@@ -44,7 +47,7 @@ static void run_backend(Backend b)
     }
 
     printf("testing...\n");
-    int tests = 200;
+    int tests = 1000;
     double* tests_table = calloc(tests, sizeof(double)); assert(tests_table);
     for (int i = 0; i < tests; ++i) {
         uintptr_t t0 = KBenchTS();
@@ -72,9 +75,9 @@ static void run_backend(Backend b)
         }
         ts_avg += (tests_table[i] / (double)tests);
     }
-    printf("min result: %f\n", ts_min);
-    printf("max result: %f\n", ts_max);
-    printf("avg result: %f\n", ts_avg);
+    printf("min result: %fms\n", ts_min * 1000.0f);
+    printf("max result: %fms\n", ts_max * 1000.0f);
+    printf("avg result: %fms\n", ts_avg * 1000.0f);
 
 #if 0 && (defined(__APPLE__) || defined(__linux__))
     // Display last result
@@ -167,6 +170,9 @@ int main(int argc, char** argv)
     G.out_len = G.inp_w * G.inp_h * 3; // RGB888
     G.out_buf = calloc(1, G.out_len);
 
+#ifdef YUVBENCH_ACCELERATE
+    run_backend(yuvbench_accelerate());
+#endif
 #ifdef YUVBENCH_BAD
     run_backend(yuvbench_bad());
 #endif
diff --git a/yuvbench/yuvbench_accelerate.c b/yuvbench/yuvbench_accelerate.c
new file mode 100644
index 0000000..5f2e794
--- /dev/null
+++ b/yuvbench/yuvbench_accelerate.c
@@ -0,0 +1,103 @@
+#include "yuvbench.h"
+
+#include <Accelerate/Accelerate.h>
+
+typedef struct AccelerateCtx AccelerateCtx;
+struct AccelerateCtx
+{
+    void* rgba_buf;
+};
+
+static bool yuvbench_accelerate_init(Ctx* ctx)
+{
+    AccelerateCtx* accel = calloc(1, sizeof(AccelerateCtx));
+    accel->rgba_buf = calloc(4, ctx->inp_w * ctx->inp_h);
+    ctx->user = accel;
+    return true;
+}
+
+static void yuvbench_accelerate_deinit(Ctx* ctx)
+{
+    AccelerateCtx* accel = (AccelerateCtx*)ctx->user;
+    free(accel->rgba_buf);
+    free(accel);
+}
+
+static bool yuvbench_accelerate_convert(Ctx* ctx)
+{
+    AccelerateCtx* accel = (AccelerateCtx*)ctx->user;
+    const uint32_t w = ctx->inp_w;
+    const uint32_t h = ctx->inp_h;
+    const uint8_t* Y  = (const uint8_t*)ctx->inp_buf;
+    const uint8_t* Cb = Y + (w * h);
+    const uint8_t* Cr = Cb + (w / 2 * h / 2);
+    // BT.709
+    vImage_YpCbCrToARGBMatrix matrix = { 0 };
+    matrix.Yp = 1.0f;
+    matrix.Cb_G = -0.1873f;
+    matrix.Cb_B = 1.8556f;
+    matrix.Cr_R = 1.5748f;
+    matrix.Cr_G = -0.4681f;
+    //
+    vImage_YpCbCrPixelRange pr = { 0 };
+    pr.Yp_bias = 16;
+    pr.CbCr_bias = 128;
+    pr.YpRangeMax = 235;
+    pr.CbCrRangeMax = 240;
+    pr.YpMax = 235;
+    pr.YpMin = 16;
+    pr.CbCrMax = 240;
+    pr.CbCrMin = 16;
+    //
+    vImage_YpCbCrToARGB out = { 0 };
+    //
+    if (vImageConvert_YpCbCrToARGB_GenerateConversion(&matrix, &pr, &out, kvImage420Yp8_Cb8_Cr8, kvImageARGB8888, kvImageNoFlags) != kvImageNoError) {
+        return false;
+    }
+    //
+    vImage_Buffer ypbuf = { 0 };
+    ypbuf.data = (void*)Y;
+    ypbuf.width = w;
+    ypbuf.height = h;
+    ypbuf.rowBytes = w;
+    vImage_Buffer cbbuf = { 0 };
+    cbbuf.data = (void*)Cb;
+    cbbuf.width = w / 2;
+    cbbuf.height = h / 2;
+    cbbuf.rowBytes = w / 2;
+    vImage_Buffer crbuf = { 0 };
+    crbuf.data = (void*)Cr;
+    crbuf.width = w / 2;
+    crbuf.height = h / 2;
+    crbuf.rowBytes = w / 2;
+    vImage_Buffer rgbabuf = { 0 };
+    rgbabuf.data = accel->rgba_buf;
+    rgbabuf.width = w;
+    rgbabuf.height = h;
+    rgbabuf.rowBytes = w * 4;
+    //
+    uint8_t permute[4] = { 0, 1, 2, 3 };
+    //
+    if (vImageConvert_420Yp8_Cb8_Cr8ToARGB8888(&ypbuf, &cbbuf, &crbuf, &rgbabuf, &out, permute, 0xFF, kvImageNoFlags) != kvImageNoError) {
+        return false;
+    }
+    //
+    vImage_Buffer rgbbuf = { 0 };
+    rgbbuf.data = ctx->out_buf;
+    rgbbuf.width = w;
+    rgbbuf.height = h;
+    rgbbuf.rowBytes = w * 3;
+    if (vImageConvert_ARGB8888toRGB888(&rgbabuf, &rgbbuf, kvImageNoFlags) != kvImageNoError) {
+        return false;
+    }
+    return true;
+}
+
+Backend yuvbench_accelerate(void)
+{
+  Backend b = { 0 };
+  b.init_fn = yuvbench_accelerate_init;
+  b.deinit_fn = yuvbench_accelerate_deinit;
+  b.convert_fn = yuvbench_accelerate_convert;
+  return b;
+}