• file
• revisions
• annotate
• diff
• comparison
• raw

media/libvpx/vp9/common/arm/neon/vp9_convolve8_neon.asm@b8a032363ba2 (annotated)

media/libvpx/vp9/common/arm/neon/vp9_convolve8_neon.asm

Thu, 22 Jan 2015 13:21:57 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Thu, 22 Jan 2015 13:21:57 +0100

branch

TOR_BUG_9701

changeset 15

b8a032363ba2

permissions

-rw-r--r--

Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6

michael@0	1	;
michael@0	2	; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
michael@0	3	;
michael@0	4	; Use of this source code is governed by a BSD-style license
michael@0	5	; that can be found in the LICENSE file in the root of the source
michael@0	6	; tree. An additional intellectual property rights grant can be found
michael@0	7	; in the file PATENTS. All contributing project authors may
michael@0	8	; be found in the AUTHORS file in the root of the source tree.
michael@0	9	;
michael@0	10
michael@0	11
michael@0	12	; These functions are only valid when:
michael@0	13	; x_step_q4 == 16
michael@0	14	; w%4 == 0
michael@0	15	; h%4 == 0
michael@0	16	; taps == 8
michael@0	17	; VP9_FILTER_WEIGHT == 128
michael@0	18	; VP9_FILTER_SHIFT == 7
michael@0	19
michael@0	20	EXPORT |vp9_convolve8_horiz_neon|
michael@0	21	EXPORT |vp9_convolve8_vert_neon|
michael@0	22	IMPORT |vp9_convolve8_horiz_c|
michael@0	23	IMPORT |vp9_convolve8_vert_c|
michael@0	24	ARM
michael@0	25	REQUIRE8
michael@0	26	PRESERVE8
michael@0	27
michael@0	28	AREA ||.text||, CODE, READONLY, ALIGN=2
michael@0	29
michael@0	30	; Multiply and accumulate by q0
michael@0	31	MACRO
michael@0	32	MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
michael@0	33	vmull.s16 $dst, $src0, d0[0]
michael@0	34	vmlal.s16 $dst, $src1, d0[1]
michael@0	35	vmlal.s16 $dst, $src2, d0[2]
michael@0	36	vmlal.s16 $dst, $src3, d0[3]
michael@0	37	vmlal.s16 $dst, $src4, d1[0]
michael@0	38	vmlal.s16 $dst, $src5, d1[1]
michael@0	39	vmlal.s16 $dst, $src6, d1[2]
michael@0	40	vmlal.s16 $dst, $src7, d1[3]
michael@0	41	MEND
michael@0	42
michael@0	43	; r0 const uint8_t *src
michael@0	44	; r1 int src_stride
michael@0	45	; r2 uint8_t *dst
michael@0	46	; r3 int dst_stride
michael@0	47	; sp[]const int16_t *filter_x
michael@0	48	; sp[]int x_step_q4
michael@0	49	; sp[]const int16_t *filter_y ; unused
michael@0	50	; sp[]int y_step_q4 ; unused
michael@0	51	; sp[]int w
michael@0	52	; sp[]int h
michael@0	53
michael@0	54	|vp9_convolve8_horiz_neon| PROC
michael@0	55	ldr r12, [sp, #4] ; x_step_q4
michael@0	56	cmp r12, #16
michael@0	57	bne vp9_convolve8_horiz_c
michael@0	58
michael@0	59	push {r4-r10, lr}
michael@0	60
michael@0	61	sub r0, r0, #3 ; adjust for taps
michael@0	62
michael@0	63	ldr r5, [sp, #32] ; filter_x
michael@0	64	ldr r6, [sp, #48] ; w
michael@0	65	ldr r7, [sp, #52] ; h
michael@0	66
michael@0	67	vld1.s16 {q0}, [r5] ; filter_x
michael@0	68
michael@0	69	sub r8, r1, r1, lsl #2 ; -src_stride * 3
michael@0	70	add r8, r8, #4 ; -src_stride * 3 + 4
michael@0	71
michael@0	72	sub r4, r3, r3, lsl #2 ; -dst_stride * 3
michael@0	73	add r4, r4, #4 ; -dst_stride * 3 + 4
michael@0	74
michael@0	75	rsb r9, r6, r1, lsl #2 ; reset src for outer loop
michael@0	76	sub r9, r9, #7
michael@0	77	rsb r12, r6, r3, lsl #2 ; reset dst for outer loop
michael@0	78
michael@0	79	mov r10, r6 ; w loop counter
michael@0	80
michael@0	81	loop_horiz_v
michael@0	82	vld1.8 {d24}, [r0], r1
michael@0	83	vld1.8 {d25}, [r0], r1
michael@0	84	vld1.8 {d26}, [r0], r1
michael@0	85	vld1.8 {d27}, [r0], r8
michael@0	86
michael@0	87	vtrn.16 q12, q13
michael@0	88	vtrn.8 d24, d25
michael@0	89	vtrn.8 d26, d27
michael@0	90
michael@0	91	pld [r0, r1, lsl #2]
michael@0	92
michael@0	93	vmovl.u8 q8, d24
michael@0	94	vmovl.u8 q9, d25
michael@0	95	vmovl.u8 q10, d26
michael@0	96	vmovl.u8 q11, d27
michael@0	97
michael@0	98	; save a few instructions in the inner loop
michael@0	99	vswp d17, d18
michael@0	100	vmov d23, d21
michael@0	101
michael@0	102	add r0, r0, #3
michael@0	103
michael@0	104	loop_horiz
michael@0	105	add r5, r0, #64
michael@0	106
michael@0	107	vld1.32 {d28[]}, [r0], r1
michael@0	108	vld1.32 {d29[]}, [r0], r1
michael@0	109	vld1.32 {d31[]}, [r0], r1
michael@0	110	vld1.32 {d30[]}, [r0], r8
michael@0	111
michael@0	112	pld [r5]
michael@0	113
michael@0	114	vtrn.16 d28, d31
michael@0	115	vtrn.16 d29, d30
michael@0	116	vtrn.8 d28, d29
michael@0	117	vtrn.8 d31, d30
michael@0	118
michael@0	119	pld [r5, r1]
michael@0	120
michael@0	121	; extract to s16
michael@0	122	vtrn.32 q14, q15
michael@0	123	vmovl.u8 q12, d28
michael@0	124	vmovl.u8 q13, d29
michael@0	125
michael@0	126	pld [r5, r1, lsl #1]
michael@0	127
michael@0	128	; src[] * filter_x
michael@0	129	MULTIPLY_BY_Q0 q1, d16, d17, d20, d22, d18, d19, d23, d24
michael@0	130	MULTIPLY_BY_Q0 q2, d17, d20, d22, d18, d19, d23, d24, d26
michael@0	131	MULTIPLY_BY_Q0 q14, d20, d22, d18, d19, d23, d24, d26, d27
michael@0	132	MULTIPLY_BY_Q0 q15, d22, d18, d19, d23, d24, d26, d27, d25
michael@0	133
michael@0	134	pld [r5, -r8]
michael@0	135
michael@0	136	; += 64 >> 7
michael@0	137	vqrshrun.s32 d2, q1, #7
michael@0	138	vqrshrun.s32 d3, q2, #7
michael@0	139	vqrshrun.s32 d4, q14, #7
michael@0	140	vqrshrun.s32 d5, q15, #7
michael@0	141
michael@0	142	; saturate
michael@0	143	vqmovn.u16 d2, q1
michael@0	144	vqmovn.u16 d3, q2
michael@0	145
michael@0	146	; transpose
michael@0	147	vtrn.16 d2, d3
michael@0	148	vtrn.32 d2, d3
michael@0	149	vtrn.8 d2, d3
michael@0	150
michael@0	151	vst1.u32 {d2[0]}, [r2@32], r3
michael@0	152	vst1.u32 {d3[0]}, [r2@32], r3
michael@0	153	vst1.u32 {d2[1]}, [r2@32], r3
michael@0	154	vst1.u32 {d3[1]}, [r2@32], r4
michael@0	155
michael@0	156	vmov q8, q9
michael@0	157	vmov d20, d23
michael@0	158	vmov q11, q12
michael@0	159	vmov q9, q13
michael@0	160
michael@0	161	subs r6, r6, #4 ; w -= 4
michael@0	162	bgt loop_horiz
michael@0	163
michael@0	164	; outer loop
michael@0	165	mov r6, r10 ; restore w counter
michael@0	166	add r0, r0, r9 ; src += src_stride * 4 - w
michael@0	167	add r2, r2, r12 ; dst += dst_stride * 4 - w
michael@0	168	subs r7, r7, #4 ; h -= 4
michael@0	169	bgt loop_horiz_v
michael@0	170
michael@0	171	pop {r4-r10, pc}
michael@0	172
michael@0	173	ENDP
michael@0	174
michael@0	175	|vp9_convolve8_vert_neon| PROC
michael@0	176	ldr r12, [sp, #12]
michael@0	177	cmp r12, #16
michael@0	178	bne vp9_convolve8_vert_c
michael@0	179
michael@0	180	push {r4-r8, lr}
michael@0	181
michael@0	182	; adjust for taps
michael@0	183	sub r0, r0, r1
michael@0	184	sub r0, r0, r1, lsl #1
michael@0	185
michael@0	186	ldr r4, [sp, #32] ; filter_y
michael@0	187	ldr r6, [sp, #40] ; w
michael@0	188	ldr lr, [sp, #44] ; h
michael@0	189
michael@0	190	vld1.s16 {q0}, [r4] ; filter_y
michael@0	191
michael@0	192	lsl r1, r1, #1
michael@0	193	lsl r3, r3, #1
michael@0	194
michael@0	195	loop_vert_h
michael@0	196	mov r4, r0
michael@0	197	add r7, r0, r1, asr #1
michael@0	198	mov r5, r2
michael@0	199	add r8, r2, r3, asr #1
michael@0	200	mov r12, lr ; h loop counter
michael@0	201
michael@0	202	vld1.u32 {d16[0]}, [r4], r1
michael@0	203	vld1.u32 {d16[1]}, [r7], r1
michael@0	204	vld1.u32 {d18[0]}, [r4], r1
michael@0	205	vld1.u32 {d18[1]}, [r7], r1
michael@0	206	vld1.u32 {d20[0]}, [r4], r1
michael@0	207	vld1.u32 {d20[1]}, [r7], r1
michael@0	208	vld1.u32 {d22[0]}, [r4], r1
michael@0	209
michael@0	210	vmovl.u8 q8, d16
michael@0	211	vmovl.u8 q9, d18
michael@0	212	vmovl.u8 q10, d20
michael@0	213	vmovl.u8 q11, d22
michael@0	214
michael@0	215	loop_vert
michael@0	216	; always process a 4x4 block at a time
michael@0	217	vld1.u32 {d24[0]}, [r7], r1
michael@0	218	vld1.u32 {d26[0]}, [r4], r1
michael@0	219	vld1.u32 {d26[1]}, [r7], r1
michael@0	220	vld1.u32 {d24[1]}, [r4], r1
michael@0	221
michael@0	222	; extract to s16
michael@0	223	vmovl.u8 q12, d24
michael@0	224	vmovl.u8 q13, d26
michael@0	225
michael@0	226	pld [r5]
michael@0	227	pld [r8]
michael@0	228
michael@0	229	; src[] * filter_y
michael@0	230	MULTIPLY_BY_Q0 q1, d16, d17, d18, d19, d20, d21, d22, d24
michael@0	231
michael@0	232	pld [r5, r3]
michael@0	233	pld [r8, r3]
michael@0	234
michael@0	235	MULTIPLY_BY_Q0 q2, d17, d18, d19, d20, d21, d22, d24, d26
michael@0	236
michael@0	237	pld [r7]
michael@0	238	pld [r4]
michael@0	239
michael@0	240	MULTIPLY_BY_Q0 q14, d18, d19, d20, d21, d22, d24, d26, d27
michael@0	241
michael@0	242	pld [r7, r1]
michael@0	243	pld [r4, r1]
michael@0	244
michael@0	245	MULTIPLY_BY_Q0 q15, d19, d20, d21, d22, d24, d26, d27, d25
michael@0	246
michael@0	247	; += 64 >> 7
michael@0	248	vqrshrun.s32 d2, q1, #7
michael@0	249	vqrshrun.s32 d3, q2, #7
michael@0	250	vqrshrun.s32 d4, q14, #7
michael@0	251	vqrshrun.s32 d5, q15, #7
michael@0	252
michael@0	253	; saturate
michael@0	254	vqmovn.u16 d2, q1
michael@0	255	vqmovn.u16 d3, q2
michael@0	256
michael@0	257	vst1.u32 {d2[0]}, [r5@32], r3
michael@0	258	vst1.u32 {d2[1]}, [r8@32], r3
michael@0	259	vst1.u32 {d3[0]}, [r5@32], r3
michael@0	260	vst1.u32 {d3[1]}, [r8@32], r3
michael@0	261
michael@0	262	vmov q8, q10
michael@0	263	vmov d18, d22
michael@0	264	vmov d19, d24
michael@0	265	vmov q10, q13
michael@0	266	vmov d22, d25
michael@0	267
michael@0	268	subs r12, r12, #4 ; h -= 4
michael@0	269	bgt loop_vert
michael@0	270
michael@0	271	; outer loop
michael@0	272	add r0, r0, #4
michael@0	273	add r2, r2, #4
michael@0	274	subs r6, r6, #4 ; w -= 4
michael@0	275	bgt loop_vert_h
michael@0	276
michael@0	277	pop {r4-r8, pc}
michael@0	278
michael@0	279	ENDP
michael@0	280	END