[vlc-commits] copy: add SSE optimization to U&V planes interleaving
Victorien Le Couviour--Tuffet
git at videolan.org
Fri May 19 09:40:09 CEST 2017
vlc | branch: master | Victorien Le Couviour--Tuffet <victorien.lecouviour.tuffet at gmail.com> | Mon May 15 16:22:34 2017 +0200| [8a5ef6541fcf42734778b99b1bb8bd9ff2f9d3ab] | committer: Thomas Guillem
copy: add SSE optimization to U&V planes interleaving
used in I420 to NV12 conversion
Signed-off-by: Thomas Guillem <thomas at gllm.fr>
> http://git.videolan.org/gitweb.cgi/vlc.git/?a=commit;h=8a5ef6541fcf42734778b99b1bb8bd9ff2f9d3ab
---
modules/video_chroma/copy.c | 166 ++++++++++++++++++++++++++++++++++++++------
1 file changed, 143 insertions(+), 23 deletions(-)
diff --git a/modules/video_chroma/copy.c b/modules/video_chroma/copy.c
index 03918fd44c..defbafb6c0 100644
--- a/modules/video_chroma/copy.c
+++ b/modules/video_chroma/copy.c
@@ -5,6 +5,7 @@
* $Id$
*
* Authors: Laurent Aimar <fenrir _AT_ videolan _DOT_ org>
+ * Victorien Le Couviour--Tuffet <victorien.lecouviour.tuffet at gmail.com>
*
* This program is free software; you can redistribute it and/or modify it
* under the terms of the GNU Lesser General Public License as published by
@@ -174,6 +175,110 @@ static void Copy2d(uint8_t *dst, size_t dst_pitch,
}
VLC_SSE
+static void
+SSE_InterleaveUV(uint8_t *dst, size_t dst_pitch,
+ uint8_t *srcu, size_t srcu_pitch,
+ uint8_t *srcv, size_t srcv_pitch,
+ unsigned int width, unsigned int height,
+ unsigned int cpu)
+{
+ assert(!((intptr_t)srcu & 0xf) && !(srcu_pitch & 0x0f) &&
+ !((intptr_t)srcv & 0xf) && !(srcv_pitch & 0x0f));
+
+#if defined(__SSSE3__) || !defined (CAN_COMPILE_SSSE3)
+ VLC_UNUSED(cpu);
+#endif
+
+ uint8_t const shuffle[] = { 0, 8,
+ 1, 9,
+ 2, 10,
+ 3, 11,
+ 4, 12,
+ 5, 13,
+ 6, 14,
+ 7, 15 };
+
+ for (unsigned int y = 0; y < height; ++y)
+ {
+ unsigned int x;
+
+#define LOAD2X32 \
+ "movhpd 0x00(%[src2]), %%xmm0\n" \
+ "movlpd 0x00(%[src1]), %%xmm0\n" \
+ \
+ "movhpd 0x08(%[src2]), %%xmm1\n" \
+ "movlpd 0x08(%[src1]), %%xmm1\n" \
+ \
+ "movhpd 0x10(%[src2]), %%xmm2\n" \
+ "movlpd 0x10(%[src1]), %%xmm2\n" \
+ \
+ "movhpd 0x18(%[src2]), %%xmm3\n" \
+ "movlpd 0x18(%[src1]), %%xmm3\n"
+
+#define STORE64 \
+ "movdqu %%xmm0, 0x00(%[dst])\n" \
+ "movdqu %%xmm1, 0x10(%[dst])\n" \
+ "movdqu %%xmm2, 0x20(%[dst])\n" \
+ "movdqu %%xmm3, 0x30(%[dst])\n"
+
+#ifdef CAN_COMPILE_SSSE3
+ if (vlc_CPU_SSSE3())
+ for (x = 0; x < (width & ~31); x += 32)
+ asm volatile
+ (
+ "movdqu (%[shuffle]), %%xmm7\n"
+ LOAD2X32
+ "pshufb %%xmm7, %%xmm0\n"
+ "pshufb %%xmm7, %%xmm1\n"
+ "pshufb %%xmm7, %%xmm2\n"
+ "pshufb %%xmm7, %%xmm3\n"
+ STORE64
+ : : [dst]"r"(dst+2*x),
+ [src1]"r"(srcu+x), [src2]"r"(srcv+x),
+ [shuffle]"r"(shuffle)
+ : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm7"
+ );
+ else
+#endif
+
+ {
+ for (x = 0; x < (width & ~31); x += 32)
+ asm volatile
+ (
+ LOAD2X32
+ "movhlps %%xmm0, %%xmm4\n"
+ "punpcklbw %%xmm4, %%xmm0\n"
+
+ "movhlps %%xmm1, %%xmm4\n"
+ "punpcklbw %%xmm4, %%xmm1\n"
+
+ "movhlps %%xmm2, %%xmm4\n"
+ "punpcklbw %%xmm4, %%xmm2\n"
+
+ "movhlps %%xmm3, %%xmm4\n"
+ "punpcklbw %%xmm4, %%xmm3\n"
+ STORE64
+ : : [dst]"r"(dst+2*x),
+ [src1]"r"(srcu+x), [src2]"r"(srcv+x)
+ : "memory",
+ "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm7"
+ );
+ }
+#undef LOAD2X32
+#undef STORE64
+
+ for (; x < width; ++x)
+ {
+ dst[2*x+0] = srcu[x];
+ dst[2*x+1] = srcv[x];
+ }
+ srcu += srcu_pitch;
+ srcv += srcv_pitch;
+ dst += dst_pitch;
+ }
+}
+
+VLC_SSE
static void SSE_SplitUV(uint8_t *dstu, size_t dstu_pitch,
uint8_t *dstv, size_t dstv_pitch,
const uint8_t *src, size_t src_pitch,
@@ -292,6 +397,40 @@ static void SSE_CopyPlane(uint8_t *dst, size_t dst_pitch,
}
}
+static void
+SSE_InterleavePlanes(uint8_t *dst, size_t dst_pitch,
+ uint8_t *srcu, size_t srcu_pitch,
+ uint8_t *srcv, size_t srcv_pitch,
+ uint8_t *cache, size_t cache_size,
+ unsigned int height,
+ unsigned int cpu)
+{
+ assert(srcu_pitch == srcv_pitch);
+ unsigned int const w16 = (srcu_pitch+15) & ~15;
+ unsigned int const hstep = (cache_size) / (2*w16);
+ assert(hstep > 0);
+
+ for (unsigned int y = 0; y < height; y += hstep)
+ {
+ unsigned int const hblock = __MIN(hstep, height - y);
+
+ /* Copy a bunch of line into our cache */
+ CopyFromUswc(cache, w16, srcu, srcu_pitch,
+ srcu_pitch, hblock, cpu);
+ CopyFromUswc(cache+w16*hblock, w16, srcv, srcv_pitch,
+ srcv_pitch, hblock, cpu);
+
+ /* Copy from our cache to the destination */
+ SSE_InterleaveUV(dst, dst_pitch, cache, w16,
+ cache+w16*hblock, w16, srcu_pitch, hblock, cpu);
+
+ /* */
+ srcu += hblock * srcu_pitch;
+ srcv += hblock * srcv_pitch;
+ dst += hblock * dst_pitch;
+ }
+}
+
static void SSE_SplitPlanes(uint8_t *dstu, size_t dstu_pitch,
uint8_t *dstv, size_t dstv_pitch,
const uint8_t *src, size_t src_pitch,
@@ -393,29 +532,10 @@ static void SSE_CopyFromI420ToNv12(picture_t *dst,
src[0], src_pitch[0],
cache->buffer, cache->size,
height, cpu);
-
- /* TODO optimise the plane merging */
- const unsigned copy_lines = height / 2;
- const unsigned copy_pitch = src_pitch[1];
-
- const int i_extra_pitch_uv = dst->p[1].i_pitch - 2 * copy_pitch;
- const int i_extra_pitch_u = src_pitch[U_PLANE] - copy_pitch;
- const int i_extra_pitch_v = src_pitch[V_PLANE] - copy_pitch;
-
- uint8_t *dstUV = dst->p[1].p_pixels;
- uint8_t *srcU = src[U_PLANE];
- uint8_t *srcV = src[V_PLANE];
- for ( unsigned int line = 0; line < copy_lines; line++ )
- {
- for ( unsigned int col = 0; col < copy_pitch; col++ )
- {
- *dstUV++ = *srcU++;
- *dstUV++ = *srcV++;
- }
- dstUV += i_extra_pitch_uv;
- srcU += i_extra_pitch_u;
- srcV += i_extra_pitch_v;
- }
+ SSE_InterleavePlanes(dst->p[1].p_pixels, dst->p[1].i_pitch,
+ src[U_PLANE], src_pitch[U_PLANE],
+ src[V_PLANE], src_pitch[V_PLANE],
+ cache->buffer, cache->size, height / 2, cpu);
asm volatile ("emms");
}
#undef COPY64
More information about the vlc-commits
mailing list