cedric pushed a commit to branch master.
http://git.enlightenment.org/core/efl.git/commit/?id=a0d0c9883995e0e04979f5382fc8954941b19edc
commit a0d0c9883995e0e04979f5382fc8954941b19edc
Author: Yury Usishchev <y.usishc...@samsung.com>
Date: Thu Apr 16 19:23:29 2015 +0200
evas: improve _op_blend_mas_c_dp_neon intrinsics implementation.
Summary:
Use vceqq and vbsl instead of twice as much vmovl and vadd instructions.
Replace vaddq_u8 with vaddq_u32.
This allows NEON code to behave exactly like C version.
Reviewers: cedric, raster
Projects: #efl
Differential Revision: https://phab.enlightenment.org/D2362
Signed-off-by: Cedric BAIL <ced...@osg.samsung.com>
---
.../evas_op_blend/op_blend_mask_color_neon.c | 37 +++++++++++-----------
1 file changed, 18 insertions(+), 19 deletions(-)
diff --git a/src/lib/evas/common/evas_op_blend/op_blend_mask_color_neon.c
b/src/lib/evas/common/evas_op_blend/op_blend_mask_color_neon.c
index 0bc8c5c..a09277e 100644
--- a/src/lib/evas/common/evas_op_blend/op_blend_mask_color_neon.c
+++ b/src/lib/evas/common/evas_op_blend/op_blend_mask_color_neon.c
@@ -25,8 +25,6 @@
static void
_op_blend_mas_c_dp_neon(DATA32 *s EINA_UNUSED, DATA8 *m, DATA32 c, DATA32 *d,
int l) {
#ifdef BUILD_NEON_INTRINSICS
- uint16x8_t d0_16x8;
- uint16x8_t d1_16x8;
uint16x8_t m_16x8;
uint16x8_t mc0_16x8;
uint16x8_t mc1_16x8;
@@ -36,14 +34,20 @@ _op_blend_mas_c_dp_neon(DATA32 *s EINA_UNUSED, DATA8 *m,
DATA32 c, DATA32 *d, in
uint32x2_t c_32x2;
uint32x2_t m_32x2;
uint32x4_t a_32x4;
+ uint32x4_t ad_32x4;
+ uint32x4_t cond_32x4;
uint32x4_t d_32x4;
uint32x4_t m_32x4;
+ uint32x4_t temp_32x4;
+ uint32x4_t mc_32x4;
+ uint32x4_t x0_32x4;
uint32x4_t x1_32x4;
uint8x16_t a_8x16;
uint8x16_t d_8x16;
uint8x16_t m_8x16;
uint8x16_t mc_8x16;
uint8x16_t temp_8x16;
+ uint8x16_t x0_8x16;
uint8x16_t x1_8x16;
uint8x8_t a0_8x8;
uint8x8_t a1_8x8;
@@ -59,6 +63,8 @@ _op_blend_mas_c_dp_neon(DATA32 *s EINA_UNUSED, DATA8 *m,
DATA32 c, DATA32 *d, in
uint8x8_t temp1_8x8;
x1_8x16 = vdupq_n_u8(0x1);
+ x0_8x16 = vdupq_n_u8(0x0);
+ x0_32x4 = vreinterpretq_u32_u8(x0_8x16);
x255_16x8 = vdupq_n_u16(0xff);
x1_32x4 = vreinterpretq_u32_u8(x1_8x16);
c_32x2 = vdup_n_u32(c);
@@ -66,7 +72,7 @@ _op_blend_mas_c_dp_neon(DATA32 *s EINA_UNUSED, DATA8 *m,
DATA32 c, DATA32 *d, in
DATA32 *start = d;
int size = l;
- DATA32 *end = start + (size & ~7);
+ DATA32 *end = start + (size & ~3);
while (start < end) {
int k = *((int *)m);
if (k == 0)
@@ -77,7 +83,6 @@ _op_blend_mas_c_dp_neon(DATA32 *s EINA_UNUSED, DATA8 *m,
DATA32 c, DATA32 *d, in
}
m_32x2 = vld1_lane_u32((DATA32*)m, m_32x2, 0);
-
d_32x4 = vld1q_u32(start);
m_8x8 = vreinterpret_u8_u32(m_32x2);
@@ -94,15 +99,15 @@ _op_blend_mas_c_dp_neon(DATA32 *s EINA_UNUSED, DATA8 *m,
DATA32 c, DATA32 *d, in
mc0_16x8 = vmull_u8(m0_8x8, c_8x8);
mc1_16x8 = vmull_u8(m1_8x8, c_8x8);
-
mc0_16x8 = vaddq_u16(mc0_16x8, x255_16x8);
mc1_16x8 = vaddq_u16(mc1_16x8, x255_16x8);
mc0_8x8 = vshrn_n_u16(mc0_16x8, 8);
mc1_8x8 = vshrn_n_u16(mc1_16x8, 8);
-
mc_8x16 = vcombine_u8(mc0_8x8, mc1_8x8);
- a_8x16 = vmvnq_u8(mc_8x16);
+
+ a_8x16 = vsubq_u8(x0_8x16, mc_8x16);
+
a_32x4 = vreinterpretq_u32_u8(a_8x16);
a_32x4 = vshrq_n_u32(a_32x4, 24);
a_32x4 = vmulq_u32(a_32x4, x1_32x4);
@@ -112,35 +117,29 @@ _op_blend_mas_c_dp_neon(DATA32 *s EINA_UNUSED, DATA8 *m,
DATA32 c, DATA32 *d, in
a1_8x8 = vget_high_u8(a_8x16);
d_8x16 = vreinterpretq_u8_u32(d_32x4);
-
d0_8x8 = vget_low_u8(d_8x16);
d1_8x8 = vget_high_u8(d_8x16);
- d0_16x8 = vmovl_u8(d0_8x8);
- d1_16x8 = vmovl_u8(d1_8x8);
-
temp0_16x8 = vmull_u8(a0_8x8, d0_8x8);
temp1_16x8 = vmull_u8(a1_8x8, d1_8x8);
-
- temp0_16x8 = vaddq_u16(temp0_16x8, d0_16x8);
- temp1_16x8 = vaddq_u16(temp1_16x8, d1_16x8);
-
temp0_8x8 = vshrn_n_u16(temp0_16x8,8);
temp1_8x8 = vshrn_n_u16(temp1_16x8,8);
temp_8x16 = vcombine_u8(temp0_8x8, temp1_8x8);
+ temp_32x4 = vreinterpretq_u32_u8(temp_8x16);
- d_8x16 = vaddq_u8(mc_8x16, temp_8x16);
+ cond_32x4 = vceqq_u32(a_32x4, x0_32x4);
+ ad_32x4 = vbslq_u32(cond_32x4, d_32x4, temp_32x4);
- d_32x4 = vreinterpretq_u32_u8(d_8x16);
+ mc_32x4 = vreinterpretq_u32_u8(mc_8x16);
+ d_32x4 = vaddq_u32(mc_32x4, ad_32x4);
vst1q_u32(start, d_32x4);
start+=4;
m+=4;
-
}
- end += (size & 7);
+ end += (size & 3);
while (start < end) {
DATA32 a = *m;
DATA32 mc = MUL_SYM(a, c);
--
