[Pixman] [PATCH] sse2: Add a fast path for add_n_8_8888

[Chris Wilson]

This path is being exercised by compositing of trapezoids, for
instance as used in the firefox-asteroids cairo-trace (and appears to be
the last non-SSE2 kernel).

core2 @ 2.66GHz,

reference memcpy speed = 4898.2MB/s (1224.6MP/s for 32bpp fills)

before: add_n_8_8888 = L1:   4.53  L2:   4.39  M:  1.62 (  0.20%)  HT:  1.66  VT:  1.63  R:  1.62  RT:  1.55 (  21Kops/s)

after:  add_n_8_8888 = L1: 531.03  L2: 531.91  M:446.21 ( 54.56%) HT: 298.44  VT:258.56  R:161.87  RT: 63.80 ( 441Kops/s)

firefox-asteroids (xvfb): 6.108s -> 5.943s

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 pixman/pixman-sse2.c |   97 ++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 97 insertions(+)

diff --git a/pixman/pixman-sse2.c b/pixman/pixman-sse2.c
index 8ef6afb..589a100 100644
--- a/pixman/pixman-sse2.c
+++ b/pixman/pixman-sse2.c
@@ -4586,6 +4586,101 @@ sse2_composite_add_n_8888 (pixman_implementation_t *imp,
     }
 }
 
+static void
+sse2_composite_add_n_8_8888 (pixman_implementation_t *imp,
+			     pixman_composite_info_t *info)
+{
+    PIXMAN_COMPOSITE_ARGS (info);
+    uint32_t     *dst_line, *dst;
+    uint8_t     *mask_line, *mask;
+    int dst_stride, mask_stride;
+    int32_t w;
+    uint32_t src;
+
+    __m128i xmm_src;
+
+    src = _pixman_image_get_solid (imp, src_image, dest_image->bits.format);
+    if (src == 0)
+	return;
+    xmm_src = expand_pixel_32_1x128 (src);
+
+    PIXMAN_IMAGE_GET_LINE (
+	dest_image, dest_x, dest_y, uint32_t, dst_stride, dst_line, 1);
+    PIXMAN_IMAGE_GET_LINE (
+	mask_image, mask_x, mask_y, uint8_t, mask_stride, mask_line, 1);
+
+    while (height--)
+    {
+	dst = dst_line;
+	dst_line += dst_stride;
+	mask = mask_line;
+	mask_line += mask_stride;
+	w = width;
+
+	while (w && ((uintptr_t)dst & 15))
+	{
+	    uint8_t m = *mask++;
+	    if (m)
+	    {
+		*dst = pack_1x128_32
+		    (_mm_adds_epu16
+		     (pix_multiply_1x128 (xmm_src, expand_pixel_8_1x128 (m)),
+		      unpack_32_1x128 (*dst)));
+	    }
+	    dst++;
+	    w--;
+	}
+
+	while (w >= 4)
+	{
+	    uint32_t m = *(uint32_t*)mask;
+	    if (m)
+	    {
+		__m128i xmm_mask_lo, xmm_mask_hi;
+		__m128i xmm_dst_lo, xmm_dst_hi;
+
+		__m128i xmm_dst = load_128_aligned ((__m128i*)dst);
+		__m128i xmm_mask =
+		    _mm_unpacklo_epi8 (unpack_32_1x128(m),
+				       _mm_setzero_si128 ());
+
+		unpack_128_2x128 (xmm_mask, &xmm_mask_lo, &xmm_mask_hi);
+		unpack_128_2x128 (xmm_dst, &xmm_dst_lo, &xmm_dst_hi);
+
+		expand_alpha_rev_2x128 (xmm_mask_lo, xmm_mask_hi,
+					&xmm_mask_lo, &xmm_mask_hi);
+
+		pix_multiply_2x128 (&xmm_src, &xmm_src,
+				    &xmm_mask_lo, &xmm_mask_hi,
+				    &xmm_mask_lo, &xmm_mask_hi);
+
+		xmm_dst_lo = _mm_adds_epu16 (xmm_mask_lo, xmm_dst_lo);
+		xmm_dst_hi = _mm_adds_epu16 (xmm_mask_hi, xmm_dst_hi);
+
+		save_128_aligned (
+		    (__m128i*)dst, pack_2x128_128 (xmm_dst_lo, xmm_dst_hi));
+	    }
+
+	    w -= 4;
+	    dst += 4;
+	    mask += 4;
+	}
+
+	while (w)
+	{
+	    uint8_t m = *mask++;
+	    if (m)
+	    {
+		*dst = pack_1x128_32
+		    (_mm_adds_epu16
+		     (pix_multiply_1x128 (xmm_src, expand_pixel_8_1x128 (m)),
+		      unpack_32_1x128 (*dst)));
+	    }
+	    dst++;
+	    w--;
+	}
+    }
+}
 
 static pixman_bool_t
 sse2_blt (pixman_implementation_t *imp,
@@ -5911,6 +6006,8 @@ static const pixman_fast_path_t sse2_fast_paths[] =
     PIXMAN_STD_FAST_PATH (ADD, solid, null, a8, sse2_composite_add_n_8),
     PIXMAN_STD_FAST_PATH (ADD, solid, null, x8r8g8b8, sse2_composite_add_n_8888),
     PIXMAN_STD_FAST_PATH (ADD, solid, null, a8r8g8b8, sse2_composite_add_n_8888),
+    PIXMAN_STD_FAST_PATH (ADD, solid, a8, x8r8g8b8, sse2_composite_add_n_8_8888),
+    PIXMAN_STD_FAST_PATH (ADD, solid, a8, a8r8g8b8, sse2_composite_add_n_8_8888),
 
     /* PIXMAN_OP_SRC */
     PIXMAN_STD_FAST_PATH (SRC, solid, a8, a8r8g8b8, sse2_composite_src_n_8_8888),
-- 
1.7.10.4