190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber; 2f71323e297a928af368937089d3ed71239786f86Andreas Huber; Copyright (c) 2010 The WebM project authors. All Rights Reserved. 390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber; 4f71323e297a928af368937089d3ed71239786f86Andreas Huber; Use of this source code is governed by a BSD-style license 5f71323e297a928af368937089d3ed71239786f86Andreas Huber; that can be found in the LICENSE file in the root of the source 6f71323e297a928af368937089d3ed71239786f86Andreas Huber; tree. An additional intellectual property rights grant can be found 7f71323e297a928af368937089d3ed71239786f86Andreas Huber; in the file PATENTS. All contributing project authors may 8f71323e297a928af368937089d3ed71239786f86Andreas Huber; be found in the AUTHORS file in the root of the source tree. 990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber; 1090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 1190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 1290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber EXPORT |vp8_bilinear_predict8x8_neon| 1390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber ARM 1490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber REQUIRE8 1590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber PRESERVE8 1690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 1790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber AREA ||.text||, CODE, READONLY, ALIGN=2 1890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber; r0 unsigned char *src_ptr, 1990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber; r1 int src_pixels_per_line, 2090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber; r2 int xoffset, 2190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber; r3 int yoffset, 2290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber; r4 unsigned char *dst_ptr, 2390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber; stack(lr) int dst_pitch 2490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 2590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber|vp8_bilinear_predict8x8_neon| PROC 2690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber push {r4, lr} 2790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 28d35fe0269d77984b383b6bdc051f26b72da15277Ard Biesheuvel adr r12, bifilter8_coeff 2990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber ldr r4, [sp, #8] ;load parameters from stack 3090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber ldr lr, [sp, #12] ;load parameters from stack 3190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 3290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber cmp r2, #0 ;skip first_pass filter if xoffset=0 3390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber beq skip_firstpass_filter 3490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 3590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;First pass: output_height lines x output_width columns (9x8) 3690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber add r2, r12, r2, lsl #3 ;calculate filter location 3790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 3890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {q1}, [r0], r1 ;load src data 3990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u32 {d31}, [r2] ;load first_pass filter 4090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {q2}, [r0], r1 4190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vdup.8 d0, d31[0] ;first_pass filter (d0 d1) 4290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {q3}, [r0], r1 4390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vdup.8 d1, d31[4] 4490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {q4}, [r0], r1 4590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 4690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q6, d2, d0 ;(src_ptr[0] * vp8_filter[0]) 4790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q7, d4, d0 4890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q8, d6, d0 4990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q9, d8, d0 5090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 5190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vext.8 d3, d2, d3, #1 ;construct src_ptr[-1] 5290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vext.8 d5, d4, d5, #1 5390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vext.8 d7, d6, d7, #1 5490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vext.8 d9, d8, d9, #1 5590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 5690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q6, d3, d1 ;(src_ptr[1] * vp8_filter[1]) 5790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q7, d5, d1 5890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q8, d7, d1 5990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q9, d9, d1 6090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 6190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {q1}, [r0], r1 ;load src data 6290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d22, q6, #7 ;shift/round/saturate to u8 6390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {q2}, [r0], r1 6490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d23, q7, #7 6590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {q3}, [r0], r1 6690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d24, q8, #7 6790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {q4}, [r0], r1 6890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d25, q9, #7 6990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 7090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber ;first_pass filtering on the rest 5-line data 7190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {q5}, [r0], r1 7290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 7390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q6, d2, d0 ;(src_ptr[0] * vp8_filter[0]) 7490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q7, d4, d0 7590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q8, d6, d0 7690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q9, d8, d0 7790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q10, d10, d0 7890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 7990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vext.8 d3, d2, d3, #1 ;construct src_ptr[-1] 8090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vext.8 d5, d4, d5, #1 8190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vext.8 d7, d6, d7, #1 8290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vext.8 d9, d8, d9, #1 8390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vext.8 d11, d10, d11, #1 8490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 8590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q6, d3, d1 ;(src_ptr[1] * vp8_filter[1]) 8690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q7, d5, d1 8790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q8, d7, d1 8890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q9, d9, d1 8990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q10, d11, d1 9090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 9190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d26, q6, #7 ;shift/round/saturate to u8 9290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d27, q7, #7 9390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d28, q8, #7 9490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d29, q9, #7 9590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d30, q10, #7 9690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 9790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;Second pass: 8x8 9890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Hubersecondpass_filter 9990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber cmp r3, #0 ;skip second_pass filter if yoffset=0 10090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber beq skip_secondpass_filter 10190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 10290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber add r3, r12, r3, lsl #3 10390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber add r0, r4, lr 10490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 10590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u32 {d31}, [r3] ;load second_pass filter 10690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber add r1, r0, lr 10790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 10890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vdup.8 d0, d31[0] ;second_pass filter parameters (d0 d1) 10990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vdup.8 d1, d31[4] 11090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 11190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q1, d22, d0 ;(src_ptr[0] * vp8_filter[0]) 11290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q2, d23, d0 11390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q3, d24, d0 11490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q4, d25, d0 11590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q5, d26, d0 11690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q6, d27, d0 11790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q7, d28, d0 11890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmull.u8 q8, d29, d0 11990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 12090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q1, d23, d1 ;(src_ptr[pixel_step] * vp8_filter[1]) 12190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q2, d24, d1 12290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q3, d25, d1 12390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q4, d26, d1 12490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q5, d27, d1 12590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q6, d28, d1 12690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q7, d29, d1 12790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vmlal.u8 q8, d30, d1 12890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 12990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d2, q1, #7 ;shift/round/saturate to u8 13090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d3, q2, #7 13190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d4, q3, #7 13290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d5, q4, #7 13390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d6, q5, #7 13490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d7, q6, #7 13590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d8, q7, #7 13690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vqrshrn.u16 d9, q8, #7 13790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 13890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d2}, [r4] ;store result 13990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d3}, [r0] 14090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d4}, [r1], lr 14190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d5}, [r1], lr 14290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d6}, [r1], lr 14390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d7}, [r1], lr 14490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d8}, [r1], lr 14590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d9}, [r1], lr 14690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 14790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber pop {r4, pc} 14890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 14990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;-------------------- 15090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberskip_firstpass_filter 15190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {d22}, [r0], r1 ;load src data 15290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {d23}, [r0], r1 15390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {d24}, [r0], r1 15490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {d25}, [r0], r1 15590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {d26}, [r0], r1 15690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {d27}, [r0], r1 15790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {d28}, [r0], r1 15890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {d29}, [r0], r1 15990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vld1.u8 {d30}, [r0], r1 16090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 16190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber b secondpass_filter 16290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 16390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;--------------------- 16490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberskip_secondpass_filter 16590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d22}, [r4], lr ;store result 16690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d23}, [r4], lr 16790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d24}, [r4], lr 16890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d25}, [r4], lr 16990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d26}, [r4], lr 17090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d27}, [r4], lr 17190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d28}, [r4], lr 17290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber vst1.u8 {d29}, [r4], lr 17390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 17490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber pop {r4, pc} 17590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 17690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber ENDP 17790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 17890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;----------------- 17979f15823c34ae1e423108295e416213200bb280fAndreas Huber 18090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberbifilter8_coeff 18190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber DCD 128, 0, 112, 16, 96, 32, 80, 48, 64, 64, 48, 80, 32, 96, 16, 112 18290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber 18390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber END 184