swscale/aarch64/rgb2rgb: add neon implementation for rgb24toyv12

                              A55               A76
rgb24toyv12_16_200_c:     36890.6           17275.5
rgb24toyv12_16_200_neon:  12460.1 ( 2.96x)   5360.8 ( 3.22x)
rgb24toyv12_128_60_c:     83205.1           39884.8
rgb24toyv12_128_60_neon:  27468.4 ( 3.03x)  13552.5 ( 2.94x)
rgb24toyv12_512_16_c:     88111.6           42346.8
rgb24toyv12_512_16_neon:  29126.6 ( 3.03x)  14411.2 ( 2.94x)
rgb24toyv12_1920_4_c:     82068.1           39620.0
rgb24toyv12_1920_4_neon:  27011.6 ( 3.04x)  13492.2 ( 2.94x)

2 files changed, 164 insertions, 0 deletions

diff --git a/libswscale/aarch64/rgb2rgb.c b/libswscale/aarch64/rgb2rgb.c
index 31db23bff4..d978a6f173 100644
--- a/libswscale/aarch64/rgb2rgb.c
+++ b/libswscale/aarch64/rgb2rgb.c
@@ -27,6 +27,9 @@
 #include "libswscale/swscale.h"
 #include "libswscale/swscale_internal.h"
 
+void ff_rgb24toyv12_neon(const uint8_t *src, uint8_t *ydst, uint8_t *udst,
+                         uint8_t *vdst, int width, int height, int lumStride,
+                         int chromStride, int srcStride, int32_t *rgb2yuv);
 void ff_interleave_bytes_neon(const uint8_t *src1, const uint8_t *src2,
                               uint8_t *dest, int width, int height,
                               int src1Stride, int src2Stride, int dstStride);
@@ -39,6 +42,7 @@ av_cold void rgb2rgb_init_aarch64(void)
     int cpu_flags = av_get_cpu_flags();
 
     if (have_neon(cpu_flags)) {
+        ff_rgb24toyv12  = ff_rgb24toyv12_neon;
         interleaveBytes = ff_interleave_bytes_neon;
         deinterleaveBytes = ff_deinterleave_bytes_neon;
     }
diff --git a/libswscale/aarch64/rgb2rgb_neon.S b/libswscale/aarch64/rgb2rgb_neon.S
index 2e4f2fb766..1382e00261 100644
--- a/libswscale/aarch64/rgb2rgb_neon.S
+++ b/libswscale/aarch64/rgb2rgb_neon.S
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2020 Martin Storsjo
+ * Copyright (c) 2024 Ramiro Polla
  *
  * This file is part of FFmpeg.
  *
@@ -20,6 +21,165 @@
 
 #include "libavutil/aarch64/asm.S"
 
+#define RGB2YUV_COEFFS 16*4+16*32
+#define BY v0.h[0]
+#define GY v0.h[1]
+#define RY v0.h[2]
+#define BU v1.h[0]
+#define GU v1.h[1]
+#define RU v1.h[2]
+#define BV v2.h[0]
+#define GV v2.h[1]
+#define RV v2.h[2]
+#define Y_OFFSET  v22
+#define UV_OFFSET v23
+
+// convert rgb to 16-bit y, u, or v
+// uses v3 and v4
+.macro rgbconv16 dst, b, g, r, bc, gc, rc
+        smull           v3.4s, \b\().4h, \bc
+        smlal           v3.4s, \g\().4h, \gc
+        smlal           v3.4s, \r\().4h, \rc
+        smull2          v4.4s, \b\().8h, \bc
+        smlal2          v4.4s, \g\().8h, \gc
+        smlal2          v4.4s, \r\().8h, \rc        // v3:v4 = b * bc + g * gc + r * rc (32-bit)
+        shrn            \dst\().4h, v3.4s, #7
+        shrn2           \dst\().8h, v4.4s, #7       // dst = b * bc + g * gc + r * rc (16-bit)
+.endm
+
+// void ff_rgb24toyv12_neon(const uint8_t *src, uint8_t *ydst, uint8_t *udst,
+//                          uint8_t *vdst, int width, int height, int lumStride,
+//                          int chromStride, int srcStride, int32_t *rgb2yuv);
+function ff_rgb24toyv12_neon, export=1
+// x0  const uint8_t *src
+// x1  uint8_t *ydst
+// x2  uint8_t *udst
+// x3  uint8_t *vdst
+// w4  int width
+// w5  int height
+// w6  int lumStride
+// w7  int chromStride
+        ldrsw           x14, [sp]
+        ldr             x15, [sp, #8]
+// x14 int srcStride
+// x15 int32_t *rgb2yuv
+
+        // extend width and stride parameters
+        uxtw            x4, w4
+        sxtw            x6, w6
+        sxtw            x7, w7
+
+        // src1 = x0
+        // src2 = x10
+        add             x10, x0,  x14               // x10 = src + srcStride
+        lsl             x14, x14, #1                // srcStride *= 2
+        add             x11, x4,  x4, lsl #1        // x11 = 3 * width
+        sub             x14, x14, x11               // srcPadding = (2 * srcStride) - (3 * width)
+
+        // ydst1 = x1
+        // ydst2 = x11
+        add             x11, x1,  x6                // x11 = ydst + lumStride
+        lsl             x6,  x6,  #1                // lumStride *= 2
+        sub             x6,  x6,  x4                // lumPadding = (2 * lumStride) - width
+
+        sub             x7,  x7,  x4, lsr #1        // chromPadding = chromStride - (width / 2)
+
+        // load rgb2yuv coefficients into v0, v1, and v2
+        add             x15, x15, #RGB2YUV_COEFFS
+        ld1             {v0.8h-v2.8h}, [x15]        // load 24 values
+
+        // load offset constants
+        movi            Y_OFFSET.8h,  #0x10, lsl #8
+        movi            UV_OFFSET.8h, #0x80, lsl #8
+
+1:
+        mov             w15, w4                     // w15 = width
+
+2:
+        // load first line
+        ld3             {v26.16b, v27.16b, v28.16b}, [x0], #48
+
+        // widen first line to 16-bit
+        uxtl            v16.8h, v26.8b              // v16 = B11
+        uxtl            v17.8h, v27.8b              // v17 = G11
+        uxtl            v18.8h, v28.8b              // v18 = R11
+        uxtl2           v19.8h, v26.16b             // v19 = B12
+        uxtl2           v20.8h, v27.16b             // v20 = G12
+        uxtl2           v21.8h, v28.16b             // v21 = R12
+
+        // calculate Y values for first line
+        rgbconv16       v24, v16, v17, v18, BY, GY, RY // v24 = Y11
+        rgbconv16       v25, v19, v20, v21, BY, GY, RY // v25 = Y12
+
+        // load second line
+        ld3             {v26.16b, v27.16b, v28.16b}, [x10], #48
+
+        // pairwise add and save rgb values to calculate average
+        addp            v5.8h, v16.8h, v19.8h
+        addp            v6.8h, v17.8h, v20.8h
+        addp            v7.8h, v18.8h, v21.8h
+
+        // widen second line to 16-bit
+        uxtl            v16.8h, v26.8b              // v16 = B21
+        uxtl            v17.8h, v27.8b              // v17 = G21
+        uxtl            v18.8h, v28.8b              // v18 = R21
+        uxtl2           v19.8h, v26.16b             // v19 = B22
+        uxtl2           v20.8h, v27.16b             // v20 = G22
+        uxtl2           v21.8h, v28.16b             // v21 = R22
+
+        // calculate Y values for second line
+        rgbconv16       v26, v16, v17, v18, BY, GY, RY // v26 = Y21
+        rgbconv16       v27, v19, v20, v21, BY, GY, RY // v27 = Y22
+
+        // pairwise add rgb values to calculate average
+        addp            v16.8h, v16.8h, v19.8h
+        addp            v17.8h, v17.8h, v20.8h
+        addp            v18.8h, v18.8h, v21.8h
+
+        // calculate average
+        add             v16.8h, v16.8h, v5.8h
+        add             v17.8h, v17.8h, v6.8h
+        add             v18.8h, v18.8h, v7.8h
+        ushr            v16.8h, v16.8h, #2
+        ushr            v17.8h, v17.8h, #2
+        ushr            v18.8h, v18.8h, #2
+
+        // calculate U and V values
+        rgbconv16       v28, v16, v17, v18, BU, GU, RU // v28 = U
+        rgbconv16       v29, v16, v17, v18, BV, GV, RV // v29 = V
+
+        // add offsets and narrow all values
+        addhn           v24.8b, v24.8h, Y_OFFSET.8h
+        addhn           v25.8b, v25.8h, Y_OFFSET.8h
+        addhn           v26.8b, v26.8h, Y_OFFSET.8h
+        addhn           v27.8b, v27.8h, Y_OFFSET.8h
+        addhn           v28.8b, v28.8h, UV_OFFSET.8h
+        addhn           v29.8b, v29.8h, UV_OFFSET.8h
+
+        subs            w15, w15, #16
+
+        // store output
+        st1             {v24.8b, v25.8b}, [x1], #16 // store ydst1
+        st1             {v26.8b, v27.8b}, [x11], #16 // store ydst2
+        st1             {v28.8b}, [x2], #8          // store udst
+        st1             {v29.8b}, [x3], #8          // store vdst
+
+        b.gt            2b
+
+        subs            w5,  w5,  #2
+
+        // row += 2
+        add             x0,  x0,  x14               // src1  += srcPadding
+        add             x10, x10, x14               // src2  += srcPadding
+        add             x1,  x1,  x6                // ydst1 += lumPadding
+        add             x11, x11, x6                // ydst2 += lumPadding
+        add             x2,  x2,  x7                // udst  += chromPadding
+        add             x3,  x3,  x7                // vdst  += chromPadding
+        b.gt            1b
+
+        ret
+endfunc
+
 // void ff_interleave_bytes_neon(const uint8_t *src1, const uint8_t *src2,
 //                               uint8_t *dest, int width, int height,
 //                               int src1Stride, int src2Stride, int dstStride);