diff mbox series

[FFmpeg-devel,v1,4/6] swscale: RGB24->YUV allow odd widths & improve C rounding

Message ID 20230820151022.2204421-5-jc@kynesim.co.uk
State New
Headers show
Series swscale: Add dedicated RGB->YUV unscaled functions & aarch64 asm | expand

Checks

Context Check Description
yinshiyou/make_loongarch64 success Make finished
yinshiyou/make_fate_loongarch64 success Make fate finished
andriy/make_x86 success Make finished
andriy/make_fate_x86 success Make fate finished

Commit Message

John Cox Aug. 20, 2023, 3:10 p.m. UTC 
Allow odd widths for conversion it costs very little and simplifies
setup slightly. x86 asm will fall back to the C code if width is odd.
Round to nearest rather than just down. This reduces the Y error
reported by tests/swscale from 3 to 1. x86 asm doesn't mirror the C so
exact correspondence isn't an issue there.

Signed-off-by: John Cox <jc@kynesim.co.uk>
---
 libswscale/rgb2rgb_template.c     | 42 ++++++++++++++++++-------------
 libswscale/swscale_unscaled.c     |  5 ++--
 libswscale/x86/rgb2rgb_template.c |  5 ++++
 3 files changed, 32 insertions(+), 20 deletions(-)
diff mbox series

Patch

diff --git a/libswscale/rgb2rgb_template.c b/libswscale/rgb2rgb_template.c
index e57bfa6545..5503e58a29 100644
--- a/libswscale/rgb2rgb_template.c
+++ b/libswscale/rgb2rgb_template.c
@@ -656,6 +656,8 @@  static void rgb24toyv12_x(const uint8_t *src, uint8_t *ydst, uint8_t *udst,
     int32_t rv = rgb2yuv[x[6]], gv = rgb2yuv[x[7]], bv = rgb2yuv[x[8]];
     int y;
     const int chromWidth = width >> 1;
+    const int32_t ky = ((16 << 1) + 1) << (RGB2YUV_SHIFT - 1);
+    const int32_t kc = ((128 << 1) + 1) << (RGB2YUV_SHIFT - 1);
 
     for (y = 0; y < height; y += 2) {
         int i;
@@ -664,9 +666,9 @@  static void rgb24toyv12_x(const uint8_t *src, uint8_t *ydst, uint8_t *udst,
             unsigned int g = src[6 * i + 1];
             unsigned int r = src[6 * i + 2];
 
-            unsigned int Y = ((ry * r + gy * g + by * b) >> RGB2YUV_SHIFT) +  16;
-            unsigned int V = ((rv * r + gv * g + bv * b) >> RGB2YUV_SHIFT) + 128;
-            unsigned int U = ((ru * r + gu * g + bu * b) >> RGB2YUV_SHIFT) + 128;
+            unsigned int Y = (ry * r + gy * g + by * b + ky) >> RGB2YUV_SHIFT;
+            unsigned int V = (rv * r + gv * g + bv * b + kc) >> RGB2YUV_SHIFT;
+            unsigned int U = (ru * r + gu * g + bu * b + kc) >> RGB2YUV_SHIFT;
 
             udst[i]     = U;
             vdst[i]     = V;
@@ -676,30 +678,36 @@  static void rgb24toyv12_x(const uint8_t *src, uint8_t *ydst, uint8_t *udst,
             g = src[6 * i + 4];
             r = src[6 * i + 5];
 
-            Y = ((ry * r + gy * g + by * b) >> RGB2YUV_SHIFT) + 16;
+            Y = ((ry * r + gy * g + by * b + ky) >> RGB2YUV_SHIFT);
             ydst[2 * i + 1] = Y;
         }
-        ydst += lumStride;
-        src  += srcStride;
-
-        if (y+1 == height)
-            break;
-
-        for (i = 0; i < chromWidth; i++) {
+        if ((width & 1) != 0) {
             unsigned int b = src[6 * i + 0];
             unsigned int g = src[6 * i + 1];
             unsigned int r = src[6 * i + 2];
 
-            unsigned int Y = ((ry * r + gy * g + by * b) >> RGB2YUV_SHIFT) + 16;
+            unsigned int Y = (ry * r + gy * g + by * b + ky) >> RGB2YUV_SHIFT;
+            unsigned int V = (rv * r + gv * g + bv * b + kc) >> RGB2YUV_SHIFT;
+            unsigned int U = (ru * r + gu * g + bu * b + kc) >> RGB2YUV_SHIFT;
 
+            udst[i]     = U;
+            vdst[i]     = V;
             ydst[2 * i] = Y;
+        }
+        ydst += lumStride;
+        src  += srcStride;
 
-            b = src[6 * i + 3];
-            g = src[6 * i + 4];
-            r = src[6 * i + 5];
+        if (y+1 == height)
+            break;
 
-            Y = ((ry * r + gy * g + by * b) >> RGB2YUV_SHIFT) + 16;
-            ydst[2 * i + 1] = Y;
+        for (i = 0; i < width; i++) {
+            unsigned int b = src[3 * i + 0];
+            unsigned int g = src[3 * i + 1];
+            unsigned int r = src[3 * i + 2];
+
+            unsigned int Y = (ry * r + gy * g + by * b + ky) >> RGB2YUV_SHIFT;
+
+            ydst[i] = Y;
         }
         udst += chromStride;
         vdst += chromStride;
diff --git a/libswscale/swscale_unscaled.c b/libswscale/swscale_unscaled.c
index 751bdcb2e4..e10f967755 100644
--- a/libswscale/swscale_unscaled.c
+++ b/libswscale/swscale_unscaled.c
@@ -1994,7 +1994,6 @@  void ff_get_unscaled_swscale(SwsContext *c)
     const enum AVPixelFormat dstFormat = c->dstFormat;
     const int flags = c->flags;
     const int dstH = c->dstH;
-    const int dstW = c->dstW;
     int needsDither;
 
     needsDither = isAnyRGB(dstFormat) &&
@@ -2052,12 +2051,12 @@  void ff_get_unscaled_swscale(SwsContext *c)
     /* bgr24toYV12 */
     if (srcFormat == AV_PIX_FMT_BGR24 &&
         (dstFormat == AV_PIX_FMT_YUV420P || dstFormat == AV_PIX_FMT_YUVA420P) &&
-        !(flags & (SWS_ACCURATE_RND | SWS_BITEXACT)) && !(dstW&1))
+        !(flags & (SWS_ACCURATE_RND | SWS_BITEXACT)))
         c->convert_unscaled = bgr24ToYv12Wrapper;
     /* rgb24toYV12 */
     if (srcFormat == AV_PIX_FMT_RGB24 &&
         (dstFormat == AV_PIX_FMT_YUV420P || dstFormat == AV_PIX_FMT_YUVA420P) &&
-        !(flags & (SWS_ACCURATE_RND | SWS_BITEXACT)) && !(dstW&1))
+        !(flags & (SWS_ACCURATE_RND | SWS_BITEXACT)))
         c->convert_unscaled = rgb24ToYv12Wrapper;
 
     /* RGB/BGR -> RGB/BGR (no dither needed forms) */
diff --git a/libswscale/x86/rgb2rgb_template.c b/libswscale/x86/rgb2rgb_template.c
index dc2b4e205a..f90527aa08 100644
--- a/libswscale/x86/rgb2rgb_template.c
+++ b/libswscale/x86/rgb2rgb_template.c
@@ -1555,6 +1555,11 @@  static inline void RENAME(bgr24toyv12)(const uint8_t *src, uint8_t *ydst, uint8_
     int y;
     const x86_reg chromWidth= width>>1;
 
+    if ((width & 1) != 0) {
+        ff_bgr24toyv12_c(src, ydst, udst, vdst, width, height, lumStride, chromStride, srcStride, rgb2yuv);
+        return;
+    }
+
     if (height > 2) {
         ff_bgr24toyv12_c(src, ydst, udst, vdst, width, 2, lumStride, chromStride, srcStride, rgb2yuv);
         src  += 2*srcStride;