cedric pushed a commit to branch master.
http://git.enlightenment.org/core/efl.git/commit/?id=10ece61dbf6d77d0a42df05c88742114c0ad6ef2
commit 10ece61dbf6d77d0a42df05c88742114c0ad6ef2
Author: Cedric BAIL <ced...@osg.samsung.com>
Date: Tue Apr 28 23:38:34 2015 +0200
evas: implement _op_blend_rel_c_dp_neon using NEON intrinsics
Summary: NEON intrinsics can be built both for armv7 and armv8.
Reviewers: raster, cedric
Reviewed By: cedric
Subscribers: cedric
Projects: #efl
Differential Revision: https://phab.enlightenment.org/D2440
Signed-off-by: Cedric BAIL <ced...@osg.samsung.com>
---
.../common/evas_op_blend/op_blend_color_neon.c | 98 ++++++++++++++++++++--
1 file changed, 89 insertions(+), 9 deletions(-)
diff --git a/src/lib/evas/common/evas_op_blend/op_blend_color_neon.c
b/src/lib/evas/common/evas_op_blend/op_blend_color_neon.c
index 7ba2ffd..076bad9 100644
--- a/src/lib/evas/common/evas_op_blend/op_blend_color_neon.c
+++ b/src/lib/evas/common/evas_op_blend/op_blend_color_neon.c
@@ -1,8 +1,6 @@
#ifdef BUILD_NEON
-#ifdef BUILD_NEON_INTRINSICS
#include <arm_neon.h>
#endif
-#endif
/* blend color --> dst */
#ifdef BUILD_NEON
@@ -278,13 +276,95 @@ init_blend_color_pt_funcs_neon(void)
#ifdef BUILD_NEON
static void
_op_blend_rel_c_dp_neon(DATA32 *s EINA_UNUSED, DATA8 *m EINA_UNUSED, DATA32 c,
DATA32 *d, int l) {
- DATA32 *e;
- int alpha = 256 - (c >> 24);
- UNROLL8_PLD_WHILE(d, l, e,
- {
- *d = MUL_SYM(*d >> 24, c) + MUL_256(alpha, *d);
- d++;
- });
+ uint16x8_t ad0_16x8;
+ uint16x8_t ad1_16x8;
+ uint16x8_t dc0_16x8;
+ uint16x8_t dc1_16x8;
+ uint16x8_t x255_16x8;
+ uint32x2_t c_32x2;
+ uint32x4_t ad_32x4;
+ uint32x4_t d_32x4;
+ uint32x4_t dc_32x4;
+ uint32x4_t x1_32x4;
+ uint8x16_t ad_8x16;
+ uint8x16_t d_8x16;
+ uint8x16_t dc_8x16;
+ uint8x16_t x1_8x16;
+ uint8x8_t ad0_8x8;
+ uint8x8_t ad1_8x8;
+ uint8x8_t alpha_8x8;
+ uint8x8_t c_8x8;
+ uint8x8_t d0_8x8;
+ uint8x8_t d1_8x8;
+ uint8x8_t dc0_8x8;
+ uint8x8_t dc1_8x8;
+
+ // alpha can only be 0 if color is 0x0. In that case we can just return.
+ // Otherwise we can assume alpha != 0. This allows more optimization in
+ // NEON code.
+
+ if(!c)
+ return;
+
+ unsigned char alpha;
+ alpha = ~(c >> 24) + 1; // 256 - (c >> 24)
+
+ alpha_8x8 = vdup_n_u8(alpha);
+ c_32x2 = vdup_n_u32(c);
+ c_8x8 = vreinterpret_u8_u32(c_32x2);
+ x1_8x16 = vdupq_n_u8(0x1);
+ x1_32x4 = vreinterpretq_u32_u8(x1_8x16);
+ x255_16x8 = vdupq_n_u16(0xff);
+
+ DATA32 *end = d + (l & ~3);
+ while (d < end)
+ {
+ // load 4 elements from d
+ d_32x4 = vld1q_u32(d);
+ d_8x16 = vreinterpretq_u8_u32(d_32x4);
+ d0_8x8 = vget_low_u8(d_8x16);
+ d1_8x8 = vget_high_u8(d_8x16);
+
+ // multiply MUL_256(alpha, *d);
+ ad0_16x8 = vmull_u8(alpha_8x8, d0_8x8);
+ ad1_16x8 = vmull_u8(alpha_8x8, d1_8x8);
+ ad0_8x8 = vshrn_n_u16(ad0_16x8,8);
+ ad1_8x8 = vshrn_n_u16(ad1_16x8,8);
+ ad_8x16 = vcombine_u8(ad0_8x8, ad1_8x8);
+ ad_32x4 = vreinterpretq_u32_u8(ad_8x16);
+
+ // shift (*d >> 24)
+ dc_32x4 = vshrq_n_u32(d_32x4, 24);
+ dc_32x4 = vmulq_u32(x1_32x4, dc_32x4);
+ dc_8x16 = vreinterpretq_u8_u32(dc_32x4);
+ dc0_8x8 = vget_low_u8(dc_8x16);
+ dc1_8x8 = vget_high_u8(dc_8x16);
+
+ // multiply MUL_256(*d >> 24, sc);
+ dc0_16x8 = vmull_u8(dc0_8x8, c_8x8);
+ dc1_16x8 = vmull_u8(dc1_8x8, c_8x8);
+ dc0_16x8 = vaddq_u16(dc0_16x8, x255_16x8);
+ dc1_16x8 = vaddq_u16(dc1_16x8, x255_16x8);
+ dc0_8x8 = vshrn_n_u16(dc0_16x8, 8);
+ dc1_8x8 = vshrn_n_u16(dc1_16x8, 8);
+ dc_8x16 = vcombine_u8(dc0_8x8, dc1_8x8);
+
+ // add up everything
+ dc_32x4 = vreinterpretq_u32_u8(dc_8x16);
+ d_32x4 = vaddq_u32(dc_32x4, ad_32x4);
+
+ // save result
+ vst1q_u32(d, d_32x4);
+
+ d+=4;
+ }
+
+ end += (l & 3);
+ while (d < end)
+ {
+ *d = MUL_SYM(*d >> 24, c) + MUL_256(alpha, *d);
+ d++;
+ }
}
#define _op_blend_rel_caa_dp_neon _op_blend_rel_c_dp_neon
--
