swscale/ppc: VSX-optimize non-full-chroma yuv2rgb_2

./ffmpeg -f lavfi -i yuvtestsrc=duration=1:size=1200x1440 -sws_flags fast_bilinear \
        -s 1200x720 -f null -vframes 100 -pix_fmt $i -nostats \
        -cpuflags 0 -v error -

32-bit mul, power8 only.

~2x speedup:

rgb24
  24431 UNITS in yuv2packed2,   16384 runs,      0 skips
  13783 UNITS in yuv2packed2,   16383 runs,      1 skips
bgr24
  24396 UNITS in yuv2packed2,   16384 runs,      0 skips
  14059 UNITS in yuv2packed2,   16384 runs,      0 skips
rgba
  26815 UNITS in yuv2packed2,   16383 runs,      1 skips
  12797 UNITS in yuv2packed2,   16383 runs,      1 skips
bgra
  27060 UNITS in yuv2packed2,   16384 runs,      0 skips
  13138 UNITS in yuv2packed2,   16384 runs,      0 skips
argb
  26998 UNITS in yuv2packed2,   16384 runs,      0 skips
  12728 UNITS in yuv2packed2,   16381 runs,      3 skips
bgra
  26651 UNITS in yuv2packed2,   16384 runs,      0 skips
  13124 UNITS in yuv2packed2,   16384 runs,      0 skips

This is a low speedup, but the x86 mmx version also gets only ~2x. The mmx version
is also heavily inaccurate, while the vsx version has high accuracy.

1 files changed, 188 insertions, 0 deletions

diff --git a/libswscale/ppc/swscale_vsx.c b/libswscale/ppc/swscale_vsx.c
index e05f9ecdd3..ba00791dcb 100644
--- a/libswscale/ppc/swscale_vsx.c
+++ b/libswscale/ppc/swscale_vsx.c
@@ -793,6 +793,180 @@ yuv2rgb_full_2_vsx_template(SwsContext *c, const int16_t *buf[2],
     }
 }
 
+static av_always_inline void
+yuv2rgb_2_vsx_template(SwsContext *c, const int16_t *buf[2],
+                     const int16_t *ubuf[2], const int16_t *vbuf[2],
+                     const int16_t *abuf[2], uint8_t *dest, int dstW,
+                     int yalpha, int uvalpha, int y,
+                     enum AVPixelFormat target, int hasAlpha)
+{
+    const int16_t *buf0  = buf[0],  *buf1  = buf[1],
+                  *ubuf0 = ubuf[0], *ubuf1 = ubuf[1],
+                  *vbuf0 = vbuf[0], *vbuf1 = vbuf[1],
+                  *abuf0 = hasAlpha ? abuf[0] : NULL,
+                  *abuf1 = hasAlpha ? abuf[1] : NULL;
+    const int16_t  yalpha1 = 4096 - yalpha;
+    const int16_t uvalpha1 = 4096 - uvalpha;
+    vector int16_t vy, vu, vv, A = vec_splat_s16(0);
+    vector int32_t vy32_l, vy32_r, vu32_l, vu32_r, vv32_l, vv32_r, tmp32;
+    vector int32_t R_l, R_r, G_l, G_r, B_l, B_r, vud32_l, vud32_r, vvd32_l, vvd32_r;
+    vector int32_t tmp, tmp2, tmp3, tmp4, tmp5, tmp6;
+    vector uint16_t rd16, gd16, bd16;
+    vector uint8_t rd, bd, gd, ad, out0, out1, tmp8;
+    const vector int16_t vyalpha1 = vec_splats(yalpha1);
+    const vector int16_t vuvalpha1 = vec_splats(uvalpha1);
+    const vector int16_t vyalpha = vec_splats((int16_t) yalpha);
+    const vector int16_t vuvalpha = vec_splats((int16_t) uvalpha);
+    const vector uint16_t zero16 = vec_splat_u16(0);
+    const vector int32_t y_offset = vec_splats(c->yuv2rgb_y_offset);
+    const vector int32_t y_coeff = vec_splats(c->yuv2rgb_y_coeff);
+    const vector int32_t y_add = vec_splats(1 << 21);
+    const vector int32_t v2r_coeff = vec_splats(c->yuv2rgb_v2r_coeff);
+    const vector int32_t v2g_coeff = vec_splats(c->yuv2rgb_v2g_coeff);
+    const vector int32_t u2g_coeff = vec_splats(c->yuv2rgb_u2g_coeff);
+    const vector int32_t u2b_coeff = vec_splats(c->yuv2rgb_u2b_coeff);
+    const vector int32_t rgbclip = vec_splats(1 << 30);
+    const vector int32_t zero32 = vec_splat_s32(0);
+    const vector uint32_t shift19 = vec_splats(19U);
+    const vector uint32_t shift22 = vec_splats(22U);
+    const vector uint32_t shift10 = vec_splat_u32(10);
+    const vector int32_t dec128 = vec_splats(128 << 19);
+    const vector int32_t add18 = vec_splats(1 << 18);
+    int i;
+
+    // Various permutations
+    const vector uint8_t doubleleft = (vector uint8_t) {0, 1, 2, 3,
+                                                        0, 1, 2, 3,
+                                                        4, 5, 6, 7,
+                                                        4, 5, 6, 7 };
+    const vector uint8_t doubleright = (vector uint8_t) {8, 9, 10, 11,
+                                                        8, 9, 10, 11,
+                                                        12, 13, 14, 15,
+                                                        12, 13, 14, 15 };
+    const vector uint8_t perm3rg0 = (vector uint8_t) {0x0, 0x10, 0,
+                                                      0x1, 0x11, 0,
+                                                      0x2, 0x12, 0,
+                                                      0x3, 0x13, 0,
+                                                      0x4, 0x14, 0,
+                                                      0x5 };
+    const vector uint8_t perm3rg1 = (vector uint8_t) {     0x15, 0,
+                                                      0x6, 0x16, 0,
+                                                      0x7, 0x17, 0 };
+    const vector uint8_t perm3tb0 = (vector uint8_t) {0x0, 0x1, 0x10,
+                                                      0x3, 0x4, 0x11,
+                                                      0x6, 0x7, 0x12,
+                                                      0x9, 0xa, 0x13,
+                                                      0xc, 0xd, 0x14,
+                                                      0xf };
+    const vector uint8_t perm3tb1 = (vector uint8_t) {     0x0, 0x15,
+                                                      0x2, 0x3, 0x16,
+                                                      0x5, 0x6, 0x17 };
+
+    av_assert2(yalpha  <= 4096U);
+    av_assert2(uvalpha <= 4096U);
+
+    for (i = 0; i < (dstW + 1) >> 1; i += 8) {
+        SETUP(vy, &buf0[i * 2], vyalpha1, &buf1[i * 2], vyalpha);
+        vy32_l = vec_sra(tmp3, shift10);
+        vy32_r = vec_sra(tmp4, shift10);
+
+        SETUP(vu, &ubuf0[i], vuvalpha1, &ubuf1[i], vuvalpha);
+        tmp3 = vec_sub(tmp3, dec128);
+        tmp4 = vec_sub(tmp4, dec128);
+        vu32_l = vec_sra(tmp3, shift10);
+        vu32_r = vec_sra(tmp4, shift10);
+
+        SETUP(vv, &vbuf0[i], vuvalpha1, &vbuf1[i], vuvalpha);
+        tmp3 = vec_sub(tmp3, dec128);
+        tmp4 = vec_sub(tmp4, dec128);
+        vv32_l = vec_sra(tmp3, shift10);
+        vv32_r = vec_sra(tmp4, shift10);
+
+        if (hasAlpha) {
+            SETUP(A, &abuf0[i], vyalpha1, &abuf1[i], vyalpha);
+            tmp3 = vec_add(tmp3, add18);
+            tmp4 = vec_add(tmp4, add18);
+            tmp3 = vec_sra(tmp3, shift19);
+            tmp4 = vec_sra(tmp4, shift19);
+            A = vec_packs(tmp3, tmp4);
+            ad = vec_packsu(A, (vector int16_t) zero16);
+        } else {
+            ad = vec_splats((uint8_t) 255);
+        }
+
+        vy32_l = vec_sub(vy32_l, y_offset);
+        vy32_r = vec_sub(vy32_r, y_offset);
+        vy32_l = vec_mul(vy32_l, y_coeff);
+        vy32_r = vec_mul(vy32_r, y_coeff);
+        vy32_l = vec_add(vy32_l, y_add);
+        vy32_r = vec_add(vy32_r, y_add);
+
+        // Use the first UV half
+        vud32_l = vec_perm(vu32_l, vu32_l, doubleleft);
+        vud32_r = vec_perm(vu32_l, vu32_l, doubleright);
+        vvd32_l = vec_perm(vv32_l, vv32_l, doubleleft);
+        vvd32_r = vec_perm(vv32_l, vv32_l, doubleright);
+
+        R_l = vec_mul(vvd32_l, v2r_coeff);
+        R_l = vec_add(R_l, vy32_l);
+        R_r = vec_mul(vvd32_r, v2r_coeff);
+        R_r = vec_add(R_r, vy32_r);
+        G_l = vec_mul(vvd32_l, v2g_coeff);
+        tmp32 = vec_mul(vud32_l, u2g_coeff);
+        G_l = vec_add(G_l, vy32_l);
+        G_l = vec_add(G_l, tmp32);
+        G_r = vec_mul(vvd32_r, v2g_coeff);
+        tmp32 = vec_mul(vud32_r, u2g_coeff);
+        G_r = vec_add(G_r, vy32_r);
+        G_r = vec_add(G_r, tmp32);
+
+        B_l = vec_mul(vud32_l, u2b_coeff);
+        B_l = vec_add(B_l, vy32_l);
+        B_r = vec_mul(vud32_r, u2b_coeff);
+        B_r = vec_add(B_r, vy32_r);
+
+        WRITERGB
+
+        // New Y for the second half
+        SETUP(vy, &buf0[i * 2 + 8], vyalpha1, &buf1[i * 2 + 8], vyalpha);
+        vy32_l = vec_sra(tmp3, shift10);
+        vy32_r = vec_sra(tmp4, shift10);
+
+        vy32_l = vec_sub(vy32_l, y_offset);
+        vy32_r = vec_sub(vy32_r, y_offset);
+        vy32_l = vec_mul(vy32_l, y_coeff);
+        vy32_r = vec_mul(vy32_r, y_coeff);
+        vy32_l = vec_add(vy32_l, y_add);
+        vy32_r = vec_add(vy32_r, y_add);
+
+        // Second UV half
+        vud32_l = vec_perm(vu32_r, vu32_r, doubleleft);
+        vud32_r = vec_perm(vu32_r, vu32_r, doubleright);
+        vvd32_l = vec_perm(vv32_r, vv32_r, doubleleft);
+        vvd32_r = vec_perm(vv32_r, vv32_r, doubleright);
+
+        R_l = vec_mul(vvd32_l, v2r_coeff);
+        R_l = vec_add(R_l, vy32_l);
+        R_r = vec_mul(vvd32_r, v2r_coeff);
+        R_r = vec_add(R_r, vy32_r);
+        G_l = vec_mul(vvd32_l, v2g_coeff);
+        tmp32 = vec_mul(vud32_l, u2g_coeff);
+        G_l = vec_add(G_l, vy32_l);
+        G_l = vec_add(G_l, tmp32);
+        G_r = vec_mul(vvd32_r, v2g_coeff);
+        tmp32 = vec_mul(vud32_r, u2g_coeff);
+        G_r = vec_add(G_r, vy32_r);
+        G_r = vec_add(G_r, tmp32);
+
+        B_l = vec_mul(vud32_l, u2b_coeff);
+        B_l = vec_add(B_l, vy32_l);
+        B_r = vec_mul(vud32_r, u2b_coeff);
+        B_r = vec_add(B_r, vy32_r);
+
+        WRITERGB
+    }
+}
+
 #undef SETUP
 
 static av_always_inline void
@@ -1151,6 +1325,14 @@ YUV2RGBWRAPPER(yuv2, rgb, xbgr32, AV_PIX_FMT_ABGR,  0)
 YUV2RGBWRAPPER(yuv2, rgb, rgb24, AV_PIX_FMT_RGB24,   0)
 YUV2RGBWRAPPER(yuv2, rgb, bgr24, AV_PIX_FMT_BGR24,   0)
 
+YUV2RGBWRAPPERX2(yuv2, rgb, bgrx32, AV_PIX_FMT_BGRA,  0)
+YUV2RGBWRAPPERX2(yuv2, rgb, rgbx32, AV_PIX_FMT_RGBA,  0)
+YUV2RGBWRAPPERX2(yuv2, rgb, xrgb32, AV_PIX_FMT_ARGB,  0)
+YUV2RGBWRAPPERX2(yuv2, rgb, xbgr32, AV_PIX_FMT_ABGR,  0)
+
+YUV2RGBWRAPPERX2(yuv2, rgb, rgb24, AV_PIX_FMT_RGB24,   0)
+YUV2RGBWRAPPERX2(yuv2, rgb, bgr24, AV_PIX_FMT_BGR24,   0)
+
 YUV2RGBWRAPPER(yuv2, rgb_full, bgrx32_full, AV_PIX_FMT_BGRA,  0)
 YUV2RGBWRAPPER(yuv2, rgb_full, rgbx32_full, AV_PIX_FMT_RGBA,  0)
 YUV2RGBWRAPPER(yuv2, rgb_full, xrgb32_full, AV_PIX_FMT_ARGB,  0)
@@ -1616,6 +1798,7 @@ av_cold void ff_sws_init_swscale_vsx(SwsContext *c)
                 if (HAVE_POWER8 && cpu_flags & AV_CPU_FLAG_POWER8) {
                     if (!c->needAlpha) {
                         c->yuv2packed1 = yuv2bgrx32_1_vsx;
+                        c->yuv2packed2 = yuv2bgrx32_2_vsx;
                     }
                 }
             break;
@@ -1623,6 +1806,7 @@ av_cold void ff_sws_init_swscale_vsx(SwsContext *c)
                 if (HAVE_POWER8 && cpu_flags & AV_CPU_FLAG_POWER8) {
                     if (!c->needAlpha) {
                         c->yuv2packed1 = yuv2rgbx32_1_vsx;
+                        c->yuv2packed2 = yuv2rgbx32_2_vsx;
                     }
                 }
             break;
@@ -1630,6 +1814,7 @@ av_cold void ff_sws_init_swscale_vsx(SwsContext *c)
                 if (HAVE_POWER8 && cpu_flags & AV_CPU_FLAG_POWER8) {
                     if (!c->needAlpha) {
                         c->yuv2packed1 = yuv2xrgb32_1_vsx;
+                        c->yuv2packed2 = yuv2xrgb32_2_vsx;
                     }
                 }
             break;
@@ -1637,17 +1822,20 @@ av_cold void ff_sws_init_swscale_vsx(SwsContext *c)
                 if (HAVE_POWER8 && cpu_flags & AV_CPU_FLAG_POWER8) {
                     if (!c->needAlpha) {
                         c->yuv2packed1 = yuv2xbgr32_1_vsx;
+                        c->yuv2packed2 = yuv2xbgr32_2_vsx;
                     }
                 }
             break;
             case AV_PIX_FMT_RGB24:
                 if (HAVE_POWER8 && cpu_flags & AV_CPU_FLAG_POWER8) {
                     c->yuv2packed1 = yuv2rgb24_1_vsx;
+                    c->yuv2packed2 = yuv2rgb24_2_vsx;
                 }
             break;
             case AV_PIX_FMT_BGR24:
                 if (HAVE_POWER8 && cpu_flags & AV_CPU_FLAG_POWER8) {
                     c->yuv2packed1 = yuv2bgr24_1_vsx;
+                    c->yuv2packed2 = yuv2bgr24_2_vsx;
                 }
             break;
         }