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_predict4x4_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_predict4x4_neon| PROC
2690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    push            {r4, lr}
2790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
28d35fe0269d77984b383b6bdc051f26b72da15277Ard Biesheuvel    adr             r12, bifilter4_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 (5x4)
3690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.u8         {d2}, [r0], r1          ;load src data
3790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    add             r2, r12, r2, lsl #3     ;calculate Hfilter location (2coeffsx4bytes=8bytes)
3890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
3990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.u8         {d3}, [r0], r1
4090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.u32        {d31}, [r2]             ;first_pass filter
4190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
4290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.u8         {d4}, [r0], r1
4390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vdup.8          d0, d31[0]              ;first_pass filter (d0-d1)
4490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.u8         {d5}, [r0], r1
4590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vdup.8          d1, d31[4]
4690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.u8         {d6}, [r0], r1
4790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
4890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vshr.u64        q4, q1, #8              ;construct src_ptr[1]
4990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vshr.u64        q5, q2, #8
5090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vshr.u64        d12, d6, #8
5190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
5290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vzip.32         d2, d3                  ;put 2-line data in 1 register (src_ptr[0])
5390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vzip.32         d4, d5
5490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vzip.32         d8, d9                  ;put 2-line data in 1 register (src_ptr[1])
5590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vzip.32         d10, d11
5690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
5790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmull.u8        q7, d2, d0              ;(src_ptr[0] * vp8_filter[0])
5890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmull.u8        q8, d4, d0
5990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmull.u8        q9, d6, d0
6090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
6190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmlal.u8        q7, d8, d1              ;(src_ptr[1] * vp8_filter[1])
6290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmlal.u8        q8, d10, d1
6390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmlal.u8        q9, d12, d1
6490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
6590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vqrshrn.u16    d28, q7, #7              ;shift/round/saturate to u8
6690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vqrshrn.u16    d29, q8, #7
6790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vqrshrn.u16    d30, q9, #7
6890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
6990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;Second pass: 4x4
7090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Hubersecondpass_filter
7190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    cmp             r3, #0                  ;skip second_pass filter if yoffset=0
7290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    beq             skip_secondpass_filter
7390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
7490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    add             r3, r12, r3, lsl #3 ;calculate Vfilter location
7590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.u32        {d31}, [r3]         ;load second_pass filter
7690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
7790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vdup.8          d0, d31[0]              ;second_pass filter parameters (d0-d5)
7890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vdup.8          d1, d31[4]
7990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
8090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmull.u8        q1, d28, d0
8190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmull.u8        q2, d29, d0
8290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
8390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vext.8          d26, d28, d29, #4       ;construct src_ptr[pixel_step]
8490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vext.8          d27, d29, d30, #4
8590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
8690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmlal.u8        q1, d26, d1
8790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vmlal.u8        q2, d27, d1
8890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
8990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    add             r0, r4, lr
9090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    add             r1, r0, lr
9190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    add             r2, r1, lr
9290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
9390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vqrshrn.u16    d2, q1, #7               ;shift/round/saturate to u8
9490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vqrshrn.u16    d3, q2, #7
9590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
9690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vst1.32         {d2[0]}, [r4]           ;store result
9790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vst1.32         {d2[1]}, [r0]
9890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vst1.32         {d3[0]}, [r1]
9990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vst1.32         {d3[1]}, [r2]
10090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
10190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop             {r4, pc}
10290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
10390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;--------------------
10490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberskip_firstpass_filter
10590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
10690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.32         {d28[0]}, [r0], r1      ;load src data
10790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.32         {d28[1]}, [r0], r1
10890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.32         {d29[0]}, [r0], r1
10990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.32         {d29[1]}, [r0], r1
11090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vld1.32         {d30[0]}, [r0], r1
11190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
11290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    b               secondpass_filter
11390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
11490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;---------------------
11590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberskip_secondpass_filter
11690d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vst1.32         {d28[0]}, [r4], lr      ;store result
11790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vst1.32         {d28[1]}, [r4], lr
11890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vst1.32         {d29[0]}, [r4], lr
11990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    vst1.32         {d29[1]}, [r4], lr
12090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
12190d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    pop             {r4, pc}
12290d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
12390d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    ENDP
12490d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
12590d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber;-----------------
12679f15823c34ae1e423108295e416213200bb280fAndreas Huber
12790d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huberbifilter4_coeff
12890d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    DCD     128, 0, 112, 16, 96, 32, 80, 48, 64, 64, 48, 80, 32, 96, 16, 112
12990d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber
13090d3ed91ae9228e1c8bab561b6138d4cb8c1e4fdAndreas Huber    END
131