diff options
Diffstat (limited to 'libvpx/vp9/common/arm/neon')
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_idct16x16_1_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_idct16x16_1_add_neon.asm) | 0 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_idct16x16_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_idct16x16_add_neon.asm) | 0 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_idct32x32_1_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_idct32x32_1_add_neon.asm) | 0 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_idct32x32_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_idct32x32_add_neon.asm) | 10 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_idct4x4_1_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_idct4x4_1_add_neon.asm) | 0 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_idct4x4_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_idct4x4_add_neon.asm) | 0 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_idct8x8_1_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_idct8x8_1_add_neon.asm) | 0 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_idct8x8_add_neon.asm) | 0 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_iht4x4_add_neon.asm) | 0 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.asm (renamed from libvpx/vp9/common/arm/neon/vp9_short_iht8x8_add_neon.asm) | 2 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.asm | 199 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c | 52 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_loopfilter_neon.asm | 70 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_mb_lpf_neon.asm | 33 | ||||
| -rw-r--r-- | libvpx/vp9/common/arm/neon/vp9_reconintra_neon.asm | 634 |
15 files changed, 945 insertions, 55 deletions
diff --git a/libvpx/vp9/common/arm/neon/vp9_short_idct16x16_1_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_idct16x16_1_add_neon.asm index b1fd21b..b1fd21b 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_idct16x16_1_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_idct16x16_1_add_neon.asm diff --git a/libvpx/vp9/common/arm/neon/vp9_short_idct16x16_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_idct16x16_add_neon.asm index a13c0d0..a13c0d0 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_idct16x16_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_idct16x16_add_neon.asm diff --git a/libvpx/vp9/common/arm/neon/vp9_short_idct32x32_1_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_idct32x32_1_add_neon.asm index d290d07..d290d07 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_idct32x32_1_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_idct32x32_1_add_neon.asm diff --git a/libvpx/vp9/common/arm/neon/vp9_short_idct32x32_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_idct32x32_add_neon.asm index 388a7d7..72e933e 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_idct32x32_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_idct32x32_add_neon.asm @@ -72,7 +72,7 @@ cospi_31_64 EQU 804 ; reg1 = output[first_offset] ; reg2 = output[second_offset] ; for proper address calculation, the last offset used when manipulating - ; output, wethere reading or storing) must be passed in. use 0 for first + ; output, whether reading or storing) must be passed in. use 0 for first ; use. MACRO LOAD_FROM_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2 @@ -88,7 +88,7 @@ cospi_31_64 EQU 804 ; output[first_offset] = reg1 ; output[second_offset] = reg2 ; for proper address calculation, the last offset used when manipulating - ; output, wethere reading or storing) must be passed in. use 0 for first + ; output, whether reading or storing) must be passed in. use 0 for first ; use. MACRO STORE_IN_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2 @@ -242,7 +242,7 @@ cospi_31_64 EQU 804 ; TODO(cd): have special case to re-use constants when they are similar for ; consecutive butterflies ; TODO(cd): have special case when both constants are the same, do the - ; additions/substractions before the multiplies. + ; additions/subtractions before the multiplies. ; generate the constants ; generate scalar constants mov r8, #$first_constant & 0xFF00 @@ -260,7 +260,7 @@ cospi_31_64 EQU 804 vmull.s16 q11, $regB, d31 vmull.s16 q12, $regC, d31 ; (used) five for intermediate (q8-q12), one for constants (q15) - ; do some addition/substractions (to get back two register) + ; do some addition/subtractions (to get back two register) vsub.s32 q8, q8, q10 vsub.s32 q9, q9, q11 ; do more multiplications (ordered for maximum latency hiding) @@ -268,7 +268,7 @@ cospi_31_64 EQU 804 vmull.s16 q11, $regA, d30 vmull.s16 q15, $regB, d30 ; (used) six for intermediate (q8-q12, q15) - ; do more addition/substractions + ; do more addition/subtractions vadd.s32 q11, q12, q11 vadd.s32 q10, q10, q15 ; (used) four for intermediate (q8-q11) diff --git a/libvpx/vp9/common/arm/neon/vp9_short_idct4x4_1_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_idct4x4_1_add_neon.asm index 0d4a721..0d4a721 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_idct4x4_1_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_idct4x4_1_add_neon.asm diff --git a/libvpx/vp9/common/arm/neon/vp9_short_idct4x4_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_idct4x4_add_neon.asm index 00283fc..00283fc 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_idct4x4_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_idct4x4_add_neon.asm diff --git a/libvpx/vp9/common/arm/neon/vp9_short_idct8x8_1_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_idct8x8_1_add_neon.asm index 421d202..421d202 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_idct8x8_1_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_idct8x8_1_add_neon.asm diff --git a/libvpx/vp9/common/arm/neon/vp9_short_idct8x8_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.asm index 5476400..5476400 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_idct8x8_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.asm diff --git a/libvpx/vp9/common/arm/neon/vp9_short_iht4x4_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.asm index 2f326e2..2f326e2 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_iht4x4_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.asm diff --git a/libvpx/vp9/common/arm/neon/vp9_short_iht8x8_add_neon.asm b/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.asm index 93d3af3..b41f566 100644 --- a/libvpx/vp9/common/arm/neon/vp9_short_iht8x8_add_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.asm @@ -576,6 +576,7 @@ vld1.s16 {q14,q15}, [r0]! push {r0-r10} + vpush {d8-d15} ; transpose the input data TRANSPOSE8X8 @@ -636,6 +637,7 @@ iadst_iadst IADST8X8_1D end_vp9_iht8x8_64_add_neon + vpop {d8-d15} pop {r0-r10} ; ROUND_POWER_OF_TWO(temp_out[j], 5) diff --git a/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.asm b/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.asm new file mode 100644 index 0000000..5b8ec20 --- /dev/null +++ b/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.asm @@ -0,0 +1,199 @@ +; +; Copyright (c) 2013 The WebM project authors. All Rights Reserved. +; +; Use of this source code is governed by a BSD-style license +; that can be found in the LICENSE file in the root of the source +; tree. An additional intellectual property rights grant can be found +; in the file PATENTS. All contributing project authors may +; be found in the AUTHORS file in the root of the source tree. +; + + EXPORT |vp9_lpf_horizontal_4_dual_neon| + ARM + + AREA ||.text||, CODE, READONLY, ALIGN=2 + +;void vp9_lpf_horizontal_4_dual_neon(uint8_t *s, int p, +; const uint8_t *blimit0, +; const uint8_t *limit0, +; const uint8_t *thresh0, +; const uint8_t *blimit1, +; const uint8_t *limit1, +; const uint8_t *thresh1) +; r0 uint8_t *s, +; r1 int p, +; r2 const uint8_t *blimit0, +; r3 const uint8_t *limit0, +; sp const uint8_t *thresh0, +; sp+4 const uint8_t *blimit1, +; sp+8 const uint8_t *limit1, +; sp+12 const uint8_t *thresh1, + +|vp9_lpf_horizontal_4_dual_neon| PROC + push {lr} + + ldr r12, [sp, #4] ; load thresh0 + vld1.8 {d0}, [r2] ; load blimit0 to first half q + vld1.8 {d2}, [r3] ; load limit0 to first half q + + add r1, r1, r1 ; double pitch + ldr r2, [sp, #8] ; load blimit1 + + vld1.8 {d4}, [r12] ; load thresh0 to first half q + + ldr r3, [sp, #12] ; load limit1 + ldr r12, [sp, #16] ; load thresh1 + vld1.8 {d1}, [r2] ; load blimit1 to 2nd half q + + sub r2, r0, r1, lsl #1 ; s[-4 * p] + + vld1.8 {d3}, [r3] ; load limit1 to 2nd half q + vld1.8 {d5}, [r12] ; load thresh1 to 2nd half q + + vpush {d8-d15} ; save neon registers + + add r3, r2, r1, lsr #1 ; s[-3 * p] + + vld1.u8 {q3}, [r2@64], r1 ; p3 + vld1.u8 {q4}, [r3@64], r1 ; p2 + vld1.u8 {q5}, [r2@64], r1 ; p1 + vld1.u8 {q6}, [r3@64], r1 ; p0 + vld1.u8 {q7}, [r2@64], r1 ; q0 + vld1.u8 {q8}, [r3@64], r1 ; q1 + vld1.u8 {q9}, [r2@64] ; q2 + vld1.u8 {q10}, [r3@64] ; q3 + + sub r2, r2, r1, lsl #1 + sub r3, r3, r1, lsl #1 + + bl vp9_loop_filter_neon_16 + + vst1.u8 {q5}, [r2@64], r1 ; store op1 + vst1.u8 {q6}, [r3@64], r1 ; store op0 + vst1.u8 {q7}, [r2@64], r1 ; store oq0 + vst1.u8 {q8}, [r3@64], r1 ; store oq1 + + vpop {d8-d15} ; restore neon registers + + pop {pc} + ENDP ; |vp9_lpf_horizontal_4_dual_neon| + +; void vp9_loop_filter_neon_16(); +; This is a helper function for the loopfilters. The invidual functions do the +; necessary load, transpose (if necessary) and store. This function uses +; registers d8-d15, so the calling function must save those registers. +; +; r0-r3, r12 PRESERVE +; q0 blimit +; q1 limit +; q2 thresh +; q3 p3 +; q4 p2 +; q5 p1 +; q6 p0 +; q7 q0 +; q8 q1 +; q9 q2 +; q10 q3 +; +; Outputs: +; q5 op1 +; q6 op0 +; q7 oq0 +; q8 oq1 +|vp9_loop_filter_neon_16| PROC + + ; filter_mask + vabd.u8 q11, q3, q4 ; m1 = abs(p3 - p2) + vabd.u8 q12, q4, q5 ; m2 = abs(p2 - p1) + vabd.u8 q13, q5, q6 ; m3 = abs(p1 - p0) + vabd.u8 q14, q8, q7 ; m4 = abs(q1 - q0) + vabd.u8 q3, q9, q8 ; m5 = abs(q2 - q1) + vabd.u8 q4, q10, q9 ; m6 = abs(q3 - q2) + + ; only compare the largest value to limit + vmax.u8 q11, q11, q12 ; m7 = max(m1, m2) + vmax.u8 q12, q13, q14 ; m8 = max(m3, m4) + + vabd.u8 q9, q6, q7 ; abs(p0 - q0) + + vmax.u8 q3, q3, q4 ; m9 = max(m5, m6) + + vmov.u8 q10, #0x80 + + vmax.u8 q15, q11, q12 ; m10 = max(m7, m8) + + vcgt.u8 q13, q13, q2 ; (abs(p1 - p0) > thresh)*-1 + vcgt.u8 q14, q14, q2 ; (abs(q1 - q0) > thresh)*-1 + vmax.u8 q15, q15, q3 ; m11 = max(m10, m9) + + vabd.u8 q2, q5, q8 ; a = abs(p1 - q1) + vqadd.u8 q9, q9, q9 ; b = abs(p0 - q0) * 2 + + veor q7, q7, q10 ; qs0 + + vcge.u8 q15, q1, q15 ; abs(m11) > limit + + vshr.u8 q2, q2, #1 ; a = a / 2 + veor q6, q6, q10 ; ps0 + + veor q5, q5, q10 ; ps1 + vqadd.u8 q9, q9, q2 ; a = b + a + + veor q8, q8, q10 ; qs1 + + vmov.u16 q4, #3 + + vsubl.s8 q2, d14, d12 ; ( qs0 - ps0) + vsubl.s8 q11, d15, d13 + + vcge.u8 q9, q0, q9 ; a > blimit + + vqsub.s8 q1, q5, q8 ; filter = clamp(ps1-qs1) + vorr q14, q13, q14 ; hev + + vmul.i16 q2, q2, q4 ; 3 * ( qs0 - ps0) + vmul.i16 q11, q11, q4 + + vand q1, q1, q14 ; filter &= hev + vand q15, q15, q9 ; mask + + vmov.u8 q4, #3 + + vaddw.s8 q2, q2, d2 ; filter + 3 * (qs0 - ps0) + vaddw.s8 q11, q11, d3 + + vmov.u8 q9, #4 + + ; filter = clamp(filter + 3 * ( qs0 - ps0)) + vqmovn.s16 d2, q2 + vqmovn.s16 d3, q11 + vand q1, q1, q15 ; filter &= mask + + vqadd.s8 q2, q1, q4 ; filter2 = clamp(filter+3) + vqadd.s8 q1, q1, q9 ; filter1 = clamp(filter+4) + vshr.s8 q2, q2, #3 ; filter2 >>= 3 + vshr.s8 q1, q1, #3 ; filter1 >>= 3 + + + vqadd.s8 q11, q6, q2 ; u = clamp(ps0 + filter2) + vqsub.s8 q0, q7, q1 ; u = clamp(qs0 - filter1) + + ; outer tap adjustments + vrshr.s8 q1, q1, #1 ; filter = ++filter1 >> 1 + + veor q7, q0, q10 ; *oq0 = u^0x80 + + vbic q1, q1, q14 ; filter &= ~hev + + vqadd.s8 q13, q5, q1 ; u = clamp(ps1 + filter) + vqsub.s8 q12, q8, q1 ; u = clamp(qs1 - filter) + + veor q6, q11, q10 ; *op0 = u^0x80 + veor q5, q13, q10 ; *op1 = u^0x80 + veor q8, q12, q10 ; *oq1 = u^0x80 + + bx lr + ENDP ; |vp9_loop_filter_neon_16| + + END diff --git a/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c b/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c new file mode 100644 index 0000000..0820db2 --- /dev/null +++ b/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c @@ -0,0 +1,52 @@ +/* + * Copyright (c) 2013 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "./vp9_rtcd.h" + +void vp9_lpf_horizontal_8_dual_neon(uint8_t *s, int p /* pitch */, + const uint8_t *blimit0, + const uint8_t *limit0, + const uint8_t *thresh0, + const uint8_t *blimit1, + const uint8_t *limit1, + const uint8_t *thresh1) { + vp9_lpf_horizontal_8(s, p, blimit0, limit0, thresh0, 1); + vp9_lpf_horizontal_8(s + 8, p, blimit1, limit1, thresh1, 1); +} + +void vp9_lpf_vertical_4_dual_neon(uint8_t *s, int p, + const uint8_t *blimit0, + const uint8_t *limit0, + const uint8_t *thresh0, + const uint8_t *blimit1, + const uint8_t *limit1, + const uint8_t *thresh1) { + vp9_lpf_vertical_4_neon(s, p, blimit0, limit0, thresh0, 1); + vp9_lpf_vertical_4_neon(s + 8 * p, p, blimit1, limit1, thresh1, 1); +} + +void vp9_lpf_vertical_8_dual_neon(uint8_t *s, int p, + const uint8_t *blimit0, + const uint8_t *limit0, + const uint8_t *thresh0, + const uint8_t *blimit1, + const uint8_t *limit1, + const uint8_t *thresh1) { + vp9_lpf_vertical_8_neon(s, p, blimit0, limit0, thresh0, 1); + vp9_lpf_vertical_8_neon(s + 8 * p, p, blimit1, limit1, thresh1, 1); +} + +void vp9_lpf_vertical_16_dual_neon(uint8_t *s, int p, + const uint8_t *blimit, + const uint8_t *limit, + const uint8_t *thresh) { + vp9_lpf_vertical_16_neon(s, p, blimit, limit, thresh); + vp9_lpf_vertical_16_neon(s + 8 * p, p, blimit, limit, thresh); +} diff --git a/libvpx/vp9/common/arm/neon/vp9_loopfilter_neon.asm b/libvpx/vp9/common/arm/neon/vp9_loopfilter_neon.asm index 8b4fe5d..4430322 100644 --- a/libvpx/vp9/common/arm/neon/vp9_loopfilter_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_loopfilter_neon.asm @@ -8,10 +8,10 @@ ; be found in the AUTHORS file in the root of the source tree. ; - EXPORT |vp9_loop_filter_horizontal_edge_neon| - EXPORT |vp9_loop_filter_vertical_edge_neon| - EXPORT |vp9_mbloop_filter_horizontal_edge_neon| - EXPORT |vp9_mbloop_filter_vertical_edge_neon| + EXPORT |vp9_lpf_horizontal_4_neon| + EXPORT |vp9_lpf_vertical_4_neon| + EXPORT |vp9_lpf_horizontal_8_neon| + EXPORT |vp9_lpf_vertical_8_neon| ARM AREA ||.text||, CODE, READONLY, ALIGN=2 @@ -21,12 +21,12 @@ ; TODO(fgalligan): See about removing the count code as this function is only ; called with a count of 1. ; -; void vp9_loop_filter_horizontal_edge_neon(uint8_t *s, -; int p /* pitch */, -; const uint8_t *blimit, -; const uint8_t *limit, -; const uint8_t *thresh, -; int count) +; void vp9_lpf_horizontal_4_neon(uint8_t *s, +; int p /* pitch */, +; const uint8_t *blimit, +; const uint8_t *limit, +; const uint8_t *thresh, +; int count) ; ; r0 uint8_t *s, ; r1 int p, /* pitch */ @@ -34,7 +34,7 @@ ; r3 const uint8_t *limit, ; sp const uint8_t *thresh, ; sp+4 int count -|vp9_loop_filter_horizontal_edge_neon| PROC +|vp9_lpf_horizontal_4_neon| PROC push {lr} vld1.8 {d0[]}, [r2] ; duplicate *blimit @@ -77,19 +77,19 @@ count_lf_h_loop end_vp9_lf_h_edge pop {pc} - ENDP ; |vp9_loop_filter_horizontal_edge_neon| + ENDP ; |vp9_lpf_horizontal_4_neon| ; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter ; works on 16 iterations at a time. ; TODO(fgalligan): See about removing the count code as this function is only ; called with a count of 1. ; -; void vp9_loop_filter_vertical_edge_neon(uint8_t *s, -; int p /* pitch */, -; const uint8_t *blimit, -; const uint8_t *limit, -; const uint8_t *thresh, -; int count) +; void vp9_lpf_vertical_4_neon(uint8_t *s, +; int p /* pitch */, +; const uint8_t *blimit, +; const uint8_t *limit, +; const uint8_t *thresh, +; int count) ; ; r0 uint8_t *s, ; r1 int p, /* pitch */ @@ -97,7 +97,7 @@ end_vp9_lf_h_edge ; r3 const uint8_t *limit, ; sp const uint8_t *thresh, ; sp+4 int count -|vp9_loop_filter_vertical_edge_neon| PROC +|vp9_lpf_vertical_4_neon| PROC push {lr} vld1.8 {d0[]}, [r2] ; duplicate *blimit @@ -158,7 +158,7 @@ count_lf_v_loop end_vp9_lf_v_edge pop {pc} - ENDP ; |vp9_loop_filter_vertical_edge_neon| + ENDP ; |vp9_lpf_vertical_4_neon| ; void vp9_loop_filter_neon(); ; This is a helper function for the loopfilters. The invidual functions do the @@ -276,18 +276,18 @@ end_vp9_lf_v_edge bx lr ENDP ; |vp9_loop_filter_neon| -; void vp9_mbloop_filter_horizontal_edge_neon(uint8_t *s, int p, -; const uint8_t *blimit, -; const uint8_t *limit, -; const uint8_t *thresh, -; int count) +; void vp9_lpf_horizontal_8_neon(uint8_t *s, int p, +; const uint8_t *blimit, +; const uint8_t *limit, +; const uint8_t *thresh, +; int count) ; r0 uint8_t *s, ; r1 int p, /* pitch */ ; r2 const uint8_t *blimit, ; r3 const uint8_t *limit, ; sp const uint8_t *thresh, ; sp+4 int count -|vp9_mbloop_filter_horizontal_edge_neon| PROC +|vp9_lpf_horizontal_8_neon| PROC push {r4-r5, lr} vld1.8 {d0[]}, [r2] ; duplicate *blimit @@ -333,14 +333,14 @@ count_mblf_h_loop end_vp9_mblf_h_edge pop {r4-r5, pc} - ENDP ; |vp9_mbloop_filter_horizontal_edge_neon| + ENDP ; |vp9_lpf_horizontal_8_neon| -; void vp9_mbloop_filter_vertical_edge_neon(uint8_t *s, -; int pitch, -; const uint8_t *blimit, -; const uint8_t *limit, -; const uint8_t *thresh, -; int count) +; void vp9_lpf_vertical_8_neon(uint8_t *s, +; int pitch, +; const uint8_t *blimit, +; const uint8_t *limit, +; const uint8_t *thresh, +; int count) ; ; r0 uint8_t *s, ; r1 int pitch, @@ -348,7 +348,7 @@ end_vp9_mblf_h_edge ; r3 const uint8_t *limit, ; sp const uint8_t *thresh, ; sp+4 int count -|vp9_mbloop_filter_vertical_edge_neon| PROC +|vp9_lpf_vertical_8_neon| PROC push {r4-r5, lr} vld1.8 {d0[]}, [r2] ; duplicate *blimit @@ -420,7 +420,7 @@ count_mblf_v_loop end_vp9_mblf_v_edge pop {r4-r5, pc} - ENDP ; |vp9_mbloop_filter_vertical_edge_neon| + ENDP ; |vp9_lpf_vertical_8_neon| ; void vp9_mbloop_filter_neon(); ; This is a helper function for the loopfilters. The invidual functions do the diff --git a/libvpx/vp9/common/arm/neon/vp9_mb_lpf_neon.asm b/libvpx/vp9/common/arm/neon/vp9_mb_lpf_neon.asm index 2e8001b..5fe2bba 100644 --- a/libvpx/vp9/common/arm/neon/vp9_mb_lpf_neon.asm +++ b/libvpx/vp9/common/arm/neon/vp9_mb_lpf_neon.asm @@ -8,23 +8,23 @@ ; be found in the AUTHORS file in the root of the source tree. ; - EXPORT |vp9_mb_lpf_horizontal_edge_w_neon| - EXPORT |vp9_mb_lpf_vertical_edge_w_neon| + EXPORT |vp9_lpf_horizontal_16_neon| + EXPORT |vp9_lpf_vertical_16_neon| ARM AREA ||.text||, CODE, READONLY, ALIGN=2 -; void vp9_mb_lpf_horizontal_edge_w_neon(uint8_t *s, int p, -; const uint8_t *blimit, -; const uint8_t *limit, -; const uint8_t *thresh -; int count) +; void vp9_lpf_horizontal_16_neon(uint8_t *s, int p, +; const uint8_t *blimit, +; const uint8_t *limit, +; const uint8_t *thresh +; int count) ; r0 uint8_t *s, ; r1 int p, /* pitch */ ; r2 const uint8_t *blimit, ; r3 const uint8_t *limit, ; sp const uint8_t *thresh, -|vp9_mb_lpf_horizontal_edge_w_neon| PROC +|vp9_lpf_horizontal_16_neon| PROC push {r4-r8, lr} vpush {d8-d15} ldr r4, [sp, #88] ; load thresh @@ -115,18 +115,18 @@ h_next vpop {d8-d15} pop {r4-r8, pc} - ENDP ; |vp9_mb_lpf_horizontal_edge_w_neon| + ENDP ; |vp9_lpf_horizontal_16_neon| -; void vp9_mb_lpf_vertical_edge_w_neon(uint8_t *s, int p, -; const uint8_t *blimit, -; const uint8_t *limit, -; const uint8_t *thresh) +; void vp9_lpf_vertical_16_neon(uint8_t *s, int p, +; const uint8_t *blimit, +; const uint8_t *limit, +; const uint8_t *thresh) ; r0 uint8_t *s, ; r1 int p, /* pitch */ ; r2 const uint8_t *blimit, ; r3 const uint8_t *limit, ; sp const uint8_t *thresh, -|vp9_mb_lpf_vertical_edge_w_neon| PROC +|vp9_lpf_vertical_16_neon| PROC push {r4-r8, lr} vpush {d8-d15} ldr r4, [sp, #88] ; load thresh @@ -279,7 +279,7 @@ v_end vpop {d8-d15} pop {r4-r8, pc} - ENDP ; |vp9_mb_lpf_vertical_edge_w_neon| + ENDP ; |vp9_lpf_vertical_16_neon| ; void vp9_wide_mbfilter_neon(); ; This is a helper function for the loopfilters. The invidual functions do the @@ -439,6 +439,9 @@ v_end tst r7, #1 bxne lr + orrs r5, r5, r6 ; Check for 0 + orreq r7, r7, #2 ; Only do mbfilter branch + ; mbfilter flat && mask branch ; TODO(fgalligan): Can I decrease the cycles shifting to consective d's ; and using vibt on the q's? diff --git a/libvpx/vp9/common/arm/neon/vp9_reconintra_neon.asm b/libvpx/vp9/common/arm/neon/vp9_reconintra_neon.asm new file mode 100644 index 0000000..dc9856f --- /dev/null +++ b/libvpx/vp9/common/arm/neon/vp9_reconintra_neon.asm @@ -0,0 +1,634 @@ +; +; Copyright (c) 2014 The WebM project authors. All Rights Reserved. +; +; Use of this source code is governed by a BSD-style license +; that can be found in the LICENSE file in the root of the source +; tree. An additional intellectual property rights grant can be found +; in the file PATENTS. All contributing project authors may +; be found in the AUTHORS file in the root of the source tree. +; + + EXPORT |vp9_v_predictor_4x4_neon| + EXPORT |vp9_v_predictor_8x8_neon| + EXPORT |vp9_v_predictor_16x16_neon| + EXPORT |vp9_v_predictor_32x32_neon| + EXPORT |vp9_h_predictor_4x4_neon| + EXPORT |vp9_h_predictor_8x8_neon| + EXPORT |vp9_h_predictor_16x16_neon| + EXPORT |vp9_h_predictor_32x32_neon| + EXPORT |vp9_tm_predictor_4x4_neon| + EXPORT |vp9_tm_predictor_8x8_neon| + EXPORT |vp9_tm_predictor_16x16_neon| + EXPORT |vp9_tm_predictor_32x32_neon| + ARM + REQUIRE8 + PRESERVE8 + + AREA ||.text||, CODE, READONLY, ALIGN=2 + +;void vp9_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_v_predictor_4x4_neon| PROC + vld1.32 {d0[0]}, [r2] + vst1.32 {d0[0]}, [r0], r1 + vst1.32 {d0[0]}, [r0], r1 + vst1.32 {d0[0]}, [r0], r1 + vst1.32 {d0[0]}, [r0], r1 + bx lr + ENDP ; |vp9_v_predictor_4x4_neon| + +;void vp9_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_v_predictor_8x8_neon| PROC + vld1.8 {d0}, [r2] + vst1.8 {d0}, [r0], r1 + vst1.8 {d0}, [r0], r1 + vst1.8 {d0}, [r0], r1 + vst1.8 {d0}, [r0], r1 + vst1.8 {d0}, [r0], r1 + vst1.8 {d0}, [r0], r1 + vst1.8 {d0}, [r0], r1 + vst1.8 {d0}, [r0], r1 + bx lr + ENDP ; |vp9_v_predictor_8x8_neon| + +;void vp9_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_v_predictor_16x16_neon| PROC + vld1.8 {q0}, [r2] + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + vst1.8 {q0}, [r0], r1 + bx lr + ENDP ; |vp9_v_predictor_16x16_neon| + +;void vp9_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_v_predictor_32x32_neon| PROC + vld1.8 {q0, q1}, [r2] + mov r2, #2 +loop_v + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + vst1.8 {q0, q1}, [r0], r1 + subs r2, r2, #1 + bgt loop_v + bx lr + ENDP ; |vp9_v_predictor_32x32_neon| + +;void vp9_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_h_predictor_4x4_neon| PROC + vld1.32 {d1[0]}, [r3] + vdup.8 d0, d1[0] + vst1.32 {d0[0]}, [r0], r1 + vdup.8 d0, d1[1] + vst1.32 {d0[0]}, [r0], r1 + vdup.8 d0, d1[2] + vst1.32 {d0[0]}, [r0], r1 + vdup.8 d0, d1[3] + vst1.32 {d0[0]}, [r0], r1 + bx lr + ENDP ; |vp9_h_predictor_4x4_neon| + +;void vp9_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_h_predictor_8x8_neon| PROC + vld1.64 {d1}, [r3] + vdup.8 d0, d1[0] + vst1.64 {d0}, [r0], r1 + vdup.8 d0, d1[1] + vst1.64 {d0}, [r0], r1 + vdup.8 d0, d1[2] + vst1.64 {d0}, [r0], r1 + vdup.8 d0, d1[3] + vst1.64 {d0}, [r0], r1 + vdup.8 d0, d1[4] + vst1.64 {d0}, [r0], r1 + vdup.8 d0, d1[5] + vst1.64 {d0}, [r0], r1 + vdup.8 d0, d1[6] + vst1.64 {d0}, [r0], r1 + vdup.8 d0, d1[7] + vst1.64 {d0}, [r0], r1 + bx lr + ENDP ; |vp9_h_predictor_8x8_neon| + +;void vp9_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_h_predictor_16x16_neon| PROC + vld1.8 {q1}, [r3] + vdup.8 q0, d2[0] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[1] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[2] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[3] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[4] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[5] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[6] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[7] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[0] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[1] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[2] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[3] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[4] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[5] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[6] + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[7] + vst1.8 {q0}, [r0], r1 + bx lr + ENDP ; |vp9_h_predictor_16x16_neon| + +;void vp9_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_h_predictor_32x32_neon| PROC + sub r1, r1, #16 + mov r2, #2 +loop_h + vld1.8 {q1}, [r3]! + vdup.8 q0, d2[0] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[1] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[2] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[3] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[4] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[5] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[6] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d2[7] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[0] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[1] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[2] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[3] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[4] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[5] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[6] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + vdup.8 q0, d3[7] + vst1.8 {q0}, [r0]! + vst1.8 {q0}, [r0], r1 + subs r2, r2, #1 + bgt loop_h + bx lr + ENDP ; |vp9_h_predictor_32x32_neon| + +;void vp9_tm_predictor_4x4_neon (uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_tm_predictor_4x4_neon| PROC + ; Load ytop_left = above[-1]; + sub r12, r2, #1 + ldrb r12, [r12] + vdup.u8 d0, r12 + + ; Load above 4 pixels + vld1.32 {d2[0]}, [r2] + + ; Compute above - ytop_left + vsubl.u8 q3, d2, d0 + + ; Load left row by row and compute left + (above - ytop_left) + ; 1st row and 2nd row + ldrb r12, [r3], #1 + ldrb r2, [r3], #1 + vdup.u16 q1, r12 + vdup.u16 q2, r2 + vadd.s16 q1, q1, q3 + vadd.s16 q2, q2, q3 + vqmovun.s16 d0, q1 + vqmovun.s16 d1, q2 + vst1.32 {d0[0]}, [r0], r1 + vst1.32 {d1[0]}, [r0], r1 + + ; 3rd row and 4th row + ldrb r12, [r3], #1 + ldrb r2, [r3], #1 + vdup.u16 q1, r12 + vdup.u16 q2, r2 + vadd.s16 q1, q1, q3 + vadd.s16 q2, q2, q3 + vqmovun.s16 d0, q1 + vqmovun.s16 d1, q2 + vst1.32 {d0[0]}, [r0], r1 + vst1.32 {d1[0]}, [r0], r1 + bx lr + ENDP ; |vp9_tm_predictor_4x4_neon| + +;void vp9_tm_predictor_8x8_neon (uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_tm_predictor_8x8_neon| PROC + ; Load ytop_left = above[-1]; + sub r12, r2, #1 + ldrb r12, [r12] + vdup.u8 d0, r12 + + ; preload 8 left + vld1.8 {d30}, [r3] + + ; Load above 8 pixels + vld1.64 {d2}, [r2] + + vmovl.u8 q10, d30 + + ; Compute above - ytop_left + vsubl.u8 q3, d2, d0 + + ; Load left row by row and compute left + (above - ytop_left) + ; 1st row and 2nd row + vdup.16 q0, d20[0] + vdup.16 q1, d20[1] + vadd.s16 q0, q3, q0 + vadd.s16 q1, q3, q1 + + ; 3rd row and 4th row + vdup.16 q8, d20[2] + vdup.16 q9, d20[3] + vadd.s16 q8, q3, q8 + vadd.s16 q9, q3, q9 + + vqmovun.s16 d0, q0 + vqmovun.s16 d1, q1 + vqmovun.s16 d2, q8 + vqmovun.s16 d3, q9 + + vst1.64 {d0}, [r0], r1 + vst1.64 {d1}, [r0], r1 + vst1.64 {d2}, [r0], r1 + vst1.64 {d3}, [r0], r1 + + ; 5th row and 6th row + vdup.16 q0, d21[0] + vdup.16 q1, d21[1] + vadd.s16 q0, q3, q0 + vadd.s16 q1, q3, q1 + + ; 7th row and 8th row + vdup.16 q8, d21[2] + vdup.16 q9, d21[3] + vadd.s16 q8, q3, q8 + vadd.s16 q9, q3, q9 + + vqmovun.s16 d0, q0 + vqmovun.s16 d1, q1 + vqmovun.s16 d2, q8 + vqmovun.s16 d3, q9 + + vst1.64 {d0}, [r0], r1 + vst1.64 {d1}, [r0], r1 + vst1.64 {d2}, [r0], r1 + vst1.64 {d3}, [r0], r1 + + bx lr + ENDP ; |vp9_tm_predictor_8x8_neon| + +;void vp9_tm_predictor_16x16_neon (uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_tm_predictor_16x16_neon| PROC + ; Load ytop_left = above[-1]; + sub r12, r2, #1 + ldrb r12, [r12] + vdup.u8 q0, r12 + + ; Load above 8 pixels + vld1.8 {q1}, [r2] + + ; preload 8 left into r12 + vld1.8 {d18}, [r3]! + + ; Compute above - ytop_left + vsubl.u8 q2, d2, d0 + vsubl.u8 q3, d3, d1 + + vmovl.u8 q10, d18 + + ; Load left row by row and compute left + (above - ytop_left) + ; Process 8 rows in each single loop and loop 2 times to process 16 rows. + mov r2, #2 + +loop_16x16_neon + ; Process two rows. + vdup.16 q0, d20[0] + vdup.16 q8, d20[1] + vadd.s16 q1, q0, q2 + vadd.s16 q0, q0, q3 + vadd.s16 q11, q8, q2 + vadd.s16 q8, q8, q3 + vqmovun.s16 d2, q1 + vqmovun.s16 d3, q0 + vqmovun.s16 d22, q11 + vqmovun.s16 d23, q8 + vdup.16 q0, d20[2] ; proload next 2 rows data + vdup.16 q8, d20[3] + vst1.64 {d2,d3}, [r0], r1 + vst1.64 {d22,d23}, [r0], r1 + + ; Process two rows. + vadd.s16 q1, q0, q2 + vadd.s16 q0, q0, q3 + vadd.s16 q11, q8, q2 + vadd.s16 q8, q8, q3 + vqmovun.s16 d2, q1 + vqmovun.s16 d3, q0 + vqmovun.s16 d22, q11 + vqmovun.s16 d23, q8 + vdup.16 q0, d21[0] ; proload next 2 rows data + vdup.16 q8, d21[1] + vst1.64 {d2,d3}, [r0], r1 + vst1.64 {d22,d23}, [r0], r1 + + vadd.s16 q1, q0, q2 + vadd.s16 q0, q0, q3 + vadd.s16 q11, q8, q2 + vadd.s16 q8, q8, q3 + vqmovun.s16 d2, q1 + vqmovun.s16 d3, q0 + vqmovun.s16 d22, q11 + vqmovun.s16 d23, q8 + vdup.16 q0, d21[2] ; proload next 2 rows data + vdup.16 q8, d21[3] + vst1.64 {d2,d3}, [r0], r1 + vst1.64 {d22,d23}, [r0], r1 + + + vadd.s16 q1, q0, q2 + vadd.s16 q0, q0, q3 + vadd.s16 q11, q8, q2 + vadd.s16 q8, q8, q3 + vqmovun.s16 d2, q1 + vqmovun.s16 d3, q0 + vqmovun.s16 d22, q11 + vqmovun.s16 d23, q8 + vld1.8 {d18}, [r3]! ; preload 8 left into r12 + vmovl.u8 q10, d18 + vst1.64 {d2,d3}, [r0], r1 + vst1.64 {d22,d23}, [r0], r1 + + subs r2, r2, #1 + bgt loop_16x16_neon + + bx lr + ENDP ; |vp9_tm_predictor_16x16_neon| + +;void vp9_tm_predictor_32x32_neon (uint8_t *dst, ptrdiff_t y_stride, +; const uint8_t *above, +; const uint8_t *left) +; r0 uint8_t *dst +; r1 ptrdiff_t y_stride +; r2 const uint8_t *above +; r3 const uint8_t *left + +|vp9_tm_predictor_32x32_neon| PROC + ; Load ytop_left = above[-1]; + sub r12, r2, #1 + ldrb r12, [r12] + vdup.u8 q0, r12 + + ; Load above 32 pixels + vld1.8 {q1}, [r2]! + vld1.8 {q2}, [r2] + + ; preload 8 left pixels + vld1.8 {d26}, [r3]! + + ; Compute above - ytop_left + vsubl.u8 q8, d2, d0 + vsubl.u8 q9, d3, d1 + vsubl.u8 q10, d4, d0 + vsubl.u8 q11, d5, d1 + + vmovl.u8 q3, d26 + + ; Load left row by row and compute left + (above - ytop_left) + ; Process 8 rows in each single loop and loop 4 times to process 32 rows. + mov r2, #4 + +loop_32x32_neon + ; Process two rows. + vdup.16 q0, d6[0] + vdup.16 q2, d6[1] + vadd.s16 q12, q0, q8 + vadd.s16 q13, q0, q9 + vadd.s16 q14, q0, q10 + vadd.s16 q15, q0, q11 + vqmovun.s16 d0, q12 + vqmovun.s16 d1, q13 + vadd.s16 q12, q2, q8 + vadd.s16 q13, q2, q9 + vqmovun.s16 d2, q14 + vqmovun.s16 d3, q15 + vadd.s16 q14, q2, q10 + vadd.s16 q15, q2, q11 + vst1.64 {d0-d3}, [r0], r1 + vqmovun.s16 d24, q12 + vqmovun.s16 d25, q13 + vqmovun.s16 d26, q14 + vqmovun.s16 d27, q15 + vdup.16 q1, d6[2] + vdup.16 q2, d6[3] + vst1.64 {d24-d27}, [r0], r1 + + ; Process two rows. + vadd.s16 q12, q1, q8 + vadd.s16 q13, q1, q9 + vadd.s16 q14, q1, q10 + vadd.s16 q15, q1, q11 + vqmovun.s16 d0, q12 + vqmovun.s16 d1, q13 + vadd.s16 q12, q2, q8 + vadd.s16 q13, q2, q9 + vqmovun.s16 d2, q14 + vqmovun.s16 d3, q15 + vadd.s16 q14, q2, q10 + vadd.s16 q15, q2, q11 + vst1.64 {d0-d3}, [r0], r1 + vqmovun.s16 d24, q12 + vqmovun.s16 d25, q13 + vqmovun.s16 d26, q14 + vqmovun.s16 d27, q15 + vdup.16 q0, d7[0] + vdup.16 q2, d7[1] + vst1.64 {d24-d27}, [r0], r1 + + ; Process two rows. + vadd.s16 q12, q0, q8 + vadd.s16 q13, q0, q9 + vadd.s16 q14, q0, q10 + vadd.s16 q15, q0, q11 + vqmovun.s16 d0, q12 + vqmovun.s16 d1, q13 + vadd.s16 q12, q2, q8 + vadd.s16 q13, q2, q9 + vqmovun.s16 d2, q14 + vqmovun.s16 d3, q15 + vadd.s16 q14, q2, q10 + vadd.s16 q15, q2, q11 + vst1.64 {d0-d3}, [r0], r1 + vqmovun.s16 d24, q12 + vqmovun.s16 d25, q13 + vqmovun.s16 d26, q14 + vqmovun.s16 d27, q15 + vdup.16 q0, d7[2] + vdup.16 q2, d7[3] + vst1.64 {d24-d27}, [r0], r1 + + ; Process two rows. + vadd.s16 q12, q0, q8 + vadd.s16 q13, q0, q9 + vadd.s16 q14, q0, q10 + vadd.s16 q15, q0, q11 + vqmovun.s16 d0, q12 + vqmovun.s16 d1, q13 + vadd.s16 q12, q2, q8 + vadd.s16 q13, q2, q9 + vqmovun.s16 d2, q14 + vqmovun.s16 d3, q15 + vadd.s16 q14, q2, q10 + vadd.s16 q15, q2, q11 + vst1.64 {d0-d3}, [r0], r1 + vqmovun.s16 d24, q12 + vqmovun.s16 d25, q13 + vld1.8 {d0}, [r3]! ; preload 8 left pixels + vqmovun.s16 d26, q14 + vqmovun.s16 d27, q15 + vmovl.u8 q3, d0 + vst1.64 {d24-d27}, [r0], r1 + + subs r2, r2, #1 + bgt loop_32x32_neon + + bx lr + ENDP ; |vp9_tm_predictor_32x32_neon| + + END |