Hi,
attached set of patches implements optimization for luma scale in matte
transition,
first patch change scaling equation that avoid division and only use
shift and multiplication
second patch implement using SSE code for scaling and copying luma.
--
________________________________________
Maksym Veremeyenko
>From fd58ca781aa2d8f68aaedabe7b0f428c5baf6853 Mon Sep 17 00:00:00 2001
From: Maksym Veremeyenko <ve...@m1.tv>
Date: Fri, 27 Jun 2014 12:14:45 +0300
Subject: [PATCH 1/2] update scaling equantion to avoid division
---
src/modules/core/transition_matte.c | 4 +++-
1 files changed, 3 insertions(+), 1 deletions(-)
diff --git a/src/modules/core/transition_matte.c b/src/modules/core/transition_matte.c
index c127761..25e93c4 100644
--- a/src/modules/core/transition_matte.c
+++ b/src/modules/core/transition_matte.c
@@ -57,7 +57,9 @@ static void copy_Y_to_A_scaled_luma(uint8_t* alpha_a, int stride_a, uint8_t* ima
p = 16;
if(p > 235)
p = 235;
- p = (p - 16) * 255 / 219;
+ /* p = (p - 16) * 255 / 219; */
+ p -= 16;
+ p = ((p << 8) + (p * 43)) >> 8;
alpha_a[i] = p;
};
--
1.7.7.6
>From 466e71bd8f7f9fd7ec7fb800bc312c5d0305b16b Mon Sep 17 00:00:00 2001
From: Maksym Veremeyenko <ve...@m1.tv>
Date: Fri, 27 Jun 2014 17:05:50 +0300
Subject: [PATCH 2/2] implement SSE optimized luma copy/scale functions
---
src/modules/core/transition_matte.c | 151 ++++++++++++++++++++++++++++++++++-
1 files changed, 149 insertions(+), 2 deletions(-)
diff --git a/src/modules/core/transition_matte.c b/src/modules/core/transition_matte.c
index 25e93c4..2ea0acd 100644
--- a/src/modules/core/transition_matte.c
+++ b/src/modules/core/transition_matte.c
@@ -30,26 +30,173 @@
typedef void ( *copy_luma_fn )(uint8_t* alpha_a, int stride_a, uint8_t* image_b, int stride_b, int width, int height);
+#if defined(USE_SSE)
+static void __attribute__((noinline)) copy_Y_to_A_full_luma_sse(uint8_t* alpha_a, uint8_t* image_b, int cnt)
+{
+ const static unsigned char const4[] =
+ {
+ 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0
+ };
+
+ __asm__ volatile
+ (
+ "movdqu (%[equ255]), %%xmm4 \n\t" /* load bottom value 0xff */
+
+ "loop_start1: \n\t"
+
+ /* load pixels block 1 */
+ "movdqu 0(%[image_b]), %%xmm0 \n\t"
+ "add $0x10, %[image_b] \n\t"
+
+ /* load pixels block 2 */
+ "movdqu 0(%[image_b]), %%xmm1 \n\t"
+ "add $0x10, %[image_b] \n\t"
+
+ /* leave only Y */
+ "pand %%xmm4, %%xmm0 \n\t"
+ "pand %%xmm4, %%xmm1 \n\t"
+
+ /* pack to 8 bit value */
+ "packuswb %%xmm1, %%xmm0 \n\t"
+
+ /* store */
+ "movdqu %%xmm0, (%[alpha_a]) \n\t"
+ "add $0x10, %[alpha_a] \n\t"
+
+ /* loop if we done */
+ "dec %[cnt] \n\t"
+ "jnz loop_start1 \n\t"
+ :
+ : [cnt]"r" (cnt), [alpha_a]"r"(alpha_a), [image_b]"r"(image_b), [equ255]"r"(const4)
+ );
+};
+#endif
+
static void copy_Y_to_A_full_luma(uint8_t* alpha_a, int stride_a, uint8_t* image_b, int stride_b, int width, int height)
{
int i, j;
for(j = 0; j < height; j++)
{
- for(i = 0; i < width; i++)
+ i = 0;
+#if defined(USE_SSE)
+ if(width >= 16)
+ {
+ copy_Y_to_A_full_luma_sse(alpha_a, image_b, width >> 4);
+ i = (width >> 4) << 4;
+ }
+#endif
+ for(; i < width; i++)
alpha_a[i] = image_b[2*i];
alpha_a += stride_a;
image_b += stride_b;
};
};
+#if defined(USE_SSE)
+static void __attribute__((noinline)) copy_Y_to_A_scaled_luma_sse(uint8_t* alpha_a, uint8_t* image_b, int cnt)
+{
+ const static unsigned char const1[] =
+ {
+ 43, 0, 43, 0, 43, 0, 43, 0, 43, 0, 43, 0, 43, 0, 43, 0
+ };
+ const static unsigned char const2[] =
+ {
+ 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0
+ };
+ const static unsigned char const3[] =
+ {
+ 235, 0, 235, 0, 235, 0, 235, 0, 235, 0, 235, 0, 235, 0, 235, 0
+ };
+ const static unsigned char const4[] =
+ {
+ 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0
+ };
+
+ __asm__ volatile
+ (
+ "movdqu (%[equ43]), %%xmm7 \n\t" /* load multiplier 43 */
+ "movdqu (%[equ16]), %%xmm6 \n\t" /* load bottom value 16 */
+ "movdqu (%[equ235]), %%xmm5 \n\t" /* load bottom value 235 */
+ "movdqu (%[equ255]), %%xmm4 \n\t" /* load bottom value 0xff */
+
+ "loop_start: \n\t"
+
+ /* load pixels block 1 */
+ "movdqu 0(%[image_b]), %%xmm0 \n\t"
+ "add $0x10, %[image_b] \n\t"
+
+ /* load pixels block 2 */
+ "movdqu 0(%[image_b]), %%xmm1 \n\t"
+ "add $0x10, %[image_b] \n\t"
+
+ /* leave only Y */
+ "pand %%xmm4, %%xmm0 \n\t"
+ "pand %%xmm4, %%xmm1 \n\t"
+
+ /* upper range clip */
+ "pminsw %%xmm5, %%xmm0 \n\t"
+ "pminsw %%xmm5, %%xmm1 \n\t"
+
+ /* upper range clip */
+ "pmaxsw %%xmm6, %%xmm0 \n\t"
+ "pmaxsw %%xmm6, %%xmm1 \n\t"
+
+ /* upper range clip */
+ "psubw %%xmm6, %%xmm0 \n\t"
+ "psubw %%xmm6, %%xmm1 \n\t"
+
+ /* duplicate values */
+ "movdqa %%xmm0,%%xmm2 \n\t"
+ "movdqa %%xmm1,%%xmm3 \n\t"
+
+ /* regA = regA << 8 */
+ "psllw $8, %%xmm0 \n\t"
+ "psllw $8, %%xmm1 \n\t"
+
+ /* regB = regB * 47 */
+ "pmullw %%xmm7, %%xmm2 \n\t"
+ "pmullw %%xmm7, %%xmm3 \n\t"
+
+ /* regA = regA + regB */
+ "paddw %%xmm2, %%xmm0 \n\t"
+ "paddw %%xmm3, %%xmm1 \n\t"
+
+ /* regA = regA >> 8 */
+ "psrlw $8, %%xmm0 \n\t"
+ "psrlw $8, %%xmm1 \n\t"
+
+ /* pack to 8 bit value */
+ "packuswb %%xmm1, %%xmm0 \n\t"
+
+ /* store */
+ "movdqu %%xmm0, (%[alpha_a]) \n\t"
+ "add $0x10, %[alpha_a] \n\t"
+
+ /* loop if we done */
+ "dec %[cnt] \n\t"
+ "jnz loop_start \n\t"
+ :
+ : [cnt]"r" (cnt), [alpha_a]"r"(alpha_a), [image_b]"r"(image_b), [equ43]"r"(const1), [equ16]"r"(const2), [equ235]"r"(const3), [equ255]"r"(const4)
+ );
+};
+#endif
+
static void copy_Y_to_A_scaled_luma(uint8_t* alpha_a, int stride_a, uint8_t* image_b, int stride_b, int width, int height)
{
int i, j;
for(j = 0; j < height; j++)
{
- for(i = 0; i < width; i++)
+ i = 0;
+#if defined(USE_SSE)
+ if(width >= 16)
+ {
+ copy_Y_to_A_scaled_luma_sse(alpha_a, image_b, width >> 4);
+ i = (width >> 4) << 4;
+ }
+#endif
+ for(; i < width; i++)
{
unsigned int p = image_b[2*i];
--
1.7.7.6
------------------------------------------------------------------------------
Open source business process management suite built on Java and Eclipse
Turn processes into business applications with Bonita BPM Community Edition
Quickly connect people, data, and systems into organized workflows
Winner of BOSSIE, CODIE, OW2 and Gartner awards
http://p.sf.net/sfu/Bonitasoft
_______________________________________________
Mlt-devel mailing list
Mlt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/mlt-devel
