ref: b6012029050ee7c5da7f571f10fd4d9fe9bf09bc
parent: 275c102787e493fbf9b45341c29864b04c27de05
parent: b337c54cc41cc4ca8cb12f027c395089adddd0d5
author: James Zern <[email protected]>
date: Mon Jun 1 16:01:37 EDT 2015
Merge "vp9_reconintra_neon_asm/tm: simplify above_left load"
--- a/vp9/common/arm/neon/vp9_reconintra_neon_asm.asm
+++ b/vp9/common/arm/neon/vp9_reconintra_neon_asm.asm
@@ -345,8 +345,7 @@
|vp9_tm_predictor_8x8_neon| PROC
; Load ytop_left = above[-1];
sub r12, r2, #1
- ldrb r12, [r12]
- vdup.u8 d0, r12
+ vld1.8 {d0[]}, [r12]; preload 8 left
vld1.8 {d30}, [r3]@@ -418,8 +417,7 @@
|vp9_tm_predictor_16x16_neon| PROC
; Load ytop_left = above[-1];
sub r12, r2, #1
- ldrb r12, [r12]
- vdup.u8 q0, r12
+ vld1.8 {d0[]}, [r12]; Load above 8 pixels
vld1.8 {q1}, [r2]@@ -429,7 +427,7 @@
; Compute above - ytop_left
vsubl.u8 q2, d2, d0
- vsubl.u8 q3, d3, d1
+ vsubl.u8 q3, d3, d0
vmovl.u8 q10, d18
@@ -512,8 +510,7 @@
|vp9_tm_predictor_32x32_neon| PROC
; Load ytop_left = above[-1];
sub r12, r2, #1
- ldrb r12, [r12]
- vdup.u8 q0, r12
+ vld1.8 {d0[]}, [r12]; Load above 32 pixels
vld1.8 {q1}, [r2]!@@ -524,9 +521,9 @@
; Compute above - ytop_left
vsubl.u8 q8, d2, d0
- vsubl.u8 q9, d3, d1
+ vsubl.u8 q9, d3, d0
vsubl.u8 q10, d4, d0
- vsubl.u8 q11, d5, d1
+ vsubl.u8 q11, d5, d0
vmovl.u8 q3, d26