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

media/libvpx/vp9/encoder/x86/vp9_sad_sse3.asm@6474c204b198 (annotated)

media/libvpx/vp9/encoder/x86/vp9_sad_sse3.asm

Wed, 31 Dec 2014 06:09:35 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Wed, 31 Dec 2014 06:09:35 +0100

changeset 0

6474c204b198

permissions

-rw-r--r--

Cloned upstream origin tor-browser at tor-browser-31.3.0esr-4.5-1-build1
revision ID fc1c9ff7c1b2defdbc039f12214767608f46423f for hacking purpose.

michael@0	1	;
michael@0	2	; Copyright (c) 2010 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	%include "vpx_ports/x86_abi_support.asm"
michael@0	12
michael@0	13	%macro STACK_FRAME_CREATE_X3 0
michael@0	14	%if ABI_IS_32BIT
michael@0	15	%define src_ptr rsi
michael@0	16	%define src_stride rax
michael@0	17	%define ref_ptr rdi
michael@0	18	%define ref_stride rdx
michael@0	19	%define end_ptr rcx
michael@0	20	%define ret_var rbx
michael@0	21	%define result_ptr arg(4)
michael@0	22	%define max_err arg(4)
michael@0	23	%define height dword ptr arg(4)
michael@0	24	push rbp
michael@0	25	mov rbp, rsp
michael@0	26	push rsi
michael@0	27	push rdi
michael@0	28	push rbx
michael@0	29
michael@0	30	mov rsi, arg(0) ; src_ptr
michael@0	31	mov rdi, arg(2) ; ref_ptr
michael@0	32
michael@0	33	movsxd rax, dword ptr arg(1) ; src_stride
michael@0	34	movsxd rdx, dword ptr arg(3) ; ref_stride
michael@0	35	%else
michael@0	36	%if LIBVPX_YASM_WIN64
michael@0	37	SAVE_XMM 7, u
michael@0	38	%define src_ptr rcx
michael@0	39	%define src_stride rdx
michael@0	40	%define ref_ptr r8
michael@0	41	%define ref_stride r9
michael@0	42	%define end_ptr r10
michael@0	43	%define ret_var r11
michael@0	44	%define result_ptr [rsp+xmm_stack_space+8+4*8]
michael@0	45	%define max_err [rsp+xmm_stack_space+8+4*8]
michael@0	46	%define height dword ptr [rsp+xmm_stack_space+8+4*8]
michael@0	47	%else
michael@0	48	%define src_ptr rdi
michael@0	49	%define src_stride rsi
michael@0	50	%define ref_ptr rdx
michael@0	51	%define ref_stride rcx
michael@0	52	%define end_ptr r9
michael@0	53	%define ret_var r10
michael@0	54	%define result_ptr r8
michael@0	55	%define max_err r8
michael@0	56	%define height r8
michael@0	57	%endif
michael@0	58	%endif
michael@0	59
michael@0	60	%endmacro
michael@0	61
michael@0	62	%macro STACK_FRAME_DESTROY_X3 0
michael@0	63	%define src_ptr
michael@0	64	%define src_stride
michael@0	65	%define ref_ptr
michael@0	66	%define ref_stride
michael@0	67	%define end_ptr
michael@0	68	%define ret_var
michael@0	69	%define result_ptr
michael@0	70	%define max_err
michael@0	71	%define height
michael@0	72
michael@0	73	%if ABI_IS_32BIT
michael@0	74	pop rbx
michael@0	75	pop rdi
michael@0	76	pop rsi
michael@0	77	pop rbp
michael@0	78	%else
michael@0	79	%if LIBVPX_YASM_WIN64
michael@0	80	RESTORE_XMM
michael@0	81	%endif
michael@0	82	%endif
michael@0	83	ret
michael@0	84	%endmacro
michael@0	85
michael@0	86	%macro PROCESS_16X2X3 5
michael@0	87	%if %1==0
michael@0	88	movdqa xmm0, XMMWORD PTR [%2]
michael@0	89	lddqu xmm5, XMMWORD PTR [%3]
michael@0	90	lddqu xmm6, XMMWORD PTR [%3+1]
michael@0	91	lddqu xmm7, XMMWORD PTR [%3+2]
michael@0	92
michael@0	93	psadbw xmm5, xmm0
michael@0	94	psadbw xmm6, xmm0
michael@0	95	psadbw xmm7, xmm0
michael@0	96	%else
michael@0	97	movdqa xmm0, XMMWORD PTR [%2]
michael@0	98	lddqu xmm1, XMMWORD PTR [%3]
michael@0	99	lddqu xmm2, XMMWORD PTR [%3+1]
michael@0	100	lddqu xmm3, XMMWORD PTR [%3+2]
michael@0	101
michael@0	102	psadbw xmm1, xmm0
michael@0	103	psadbw xmm2, xmm0
michael@0	104	psadbw xmm3, xmm0
michael@0	105
michael@0	106	paddw xmm5, xmm1
michael@0	107	paddw xmm6, xmm2
michael@0	108	paddw xmm7, xmm3
michael@0	109	%endif
michael@0	110	movdqa xmm0, XMMWORD PTR [%2+%4]
michael@0	111	lddqu xmm1, XMMWORD PTR [%3+%5]
michael@0	112	lddqu xmm2, XMMWORD PTR [%3+%5+1]
michael@0	113	lddqu xmm3, XMMWORD PTR [%3+%5+2]
michael@0	114
michael@0	115	%if %1==0 || %1==1
michael@0	116	lea %2, [%2+%4*2]
michael@0	117	lea %3, [%3+%5*2]
michael@0	118	%endif
michael@0	119
michael@0	120	psadbw xmm1, xmm0
michael@0	121	psadbw xmm2, xmm0
michael@0	122	psadbw xmm3, xmm0
michael@0	123
michael@0	124	paddw xmm5, xmm1
michael@0	125	paddw xmm6, xmm2
michael@0	126	paddw xmm7, xmm3
michael@0	127	%endmacro
michael@0	128
michael@0	129	%macro PROCESS_8X2X3 5
michael@0	130	%if %1==0
michael@0	131	movq mm0, QWORD PTR [%2]
michael@0	132	movq mm5, QWORD PTR [%3]
michael@0	133	movq mm6, QWORD PTR [%3+1]
michael@0	134	movq mm7, QWORD PTR [%3+2]
michael@0	135
michael@0	136	psadbw mm5, mm0
michael@0	137	psadbw mm6, mm0
michael@0	138	psadbw mm7, mm0
michael@0	139	%else
michael@0	140	movq mm0, QWORD PTR [%2]
michael@0	141	movq mm1, QWORD PTR [%3]
michael@0	142	movq mm2, QWORD PTR [%3+1]
michael@0	143	movq mm3, QWORD PTR [%3+2]
michael@0	144
michael@0	145	psadbw mm1, mm0
michael@0	146	psadbw mm2, mm0
michael@0	147	psadbw mm3, mm0
michael@0	148
michael@0	149	paddw mm5, mm1
michael@0	150	paddw mm6, mm2
michael@0	151	paddw mm7, mm3
michael@0	152	%endif
michael@0	153	movq mm0, QWORD PTR [%2+%4]
michael@0	154	movq mm1, QWORD PTR [%3+%5]
michael@0	155	movq mm2, QWORD PTR [%3+%5+1]
michael@0	156	movq mm3, QWORD PTR [%3+%5+2]
michael@0	157
michael@0	158	%if %1==0 || %1==1
michael@0	159	lea %2, [%2+%4*2]
michael@0	160	lea %3, [%3+%5*2]
michael@0	161	%endif
michael@0	162
michael@0	163	psadbw mm1, mm0
michael@0	164	psadbw mm2, mm0
michael@0	165	psadbw mm3, mm0
michael@0	166
michael@0	167	paddw mm5, mm1
michael@0	168	paddw mm6, mm2
michael@0	169	paddw mm7, mm3
michael@0	170	%endmacro
michael@0	171
michael@0	172	;void int vp9_sad16x16x3_sse3(
michael@0	173	; unsigned char *src_ptr,
michael@0	174	; int src_stride,
michael@0	175	; unsigned char *ref_ptr,
michael@0	176	; int ref_stride,
michael@0	177	; int *results)
michael@0	178	global sym(vp9_sad16x16x3_sse3) PRIVATE
michael@0	179	sym(vp9_sad16x16x3_sse3):
michael@0	180
michael@0	181	STACK_FRAME_CREATE_X3
michael@0	182
michael@0	183	PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	184	PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	185	PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	186	PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	187	PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	188	PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	189	PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	190	PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	191
michael@0	192	mov rcx, result_ptr
michael@0	193
michael@0	194	movq xmm0, xmm5
michael@0	195	psrldq xmm5, 8
michael@0	196
michael@0	197	paddw xmm0, xmm5
michael@0	198	movd [rcx], xmm0
michael@0	199	;-
michael@0	200	movq xmm0, xmm6
michael@0	201	psrldq xmm6, 8
michael@0	202
michael@0	203	paddw xmm0, xmm6
michael@0	204	movd [rcx+4], xmm0
michael@0	205	;-
michael@0	206	movq xmm0, xmm7
michael@0	207	psrldq xmm7, 8
michael@0	208
michael@0	209	paddw xmm0, xmm7
michael@0	210	movd [rcx+8], xmm0
michael@0	211
michael@0	212	STACK_FRAME_DESTROY_X3
michael@0	213
michael@0	214	;void int vp9_sad16x8x3_sse3(
michael@0	215	; unsigned char *src_ptr,
michael@0	216	; int src_stride,
michael@0	217	; unsigned char *ref_ptr,
michael@0	218	; int ref_stride,
michael@0	219	; int *results)
michael@0	220	global sym(vp9_sad16x8x3_sse3) PRIVATE
michael@0	221	sym(vp9_sad16x8x3_sse3):
michael@0	222
michael@0	223	STACK_FRAME_CREATE_X3
michael@0	224
michael@0	225	PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	226	PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	227	PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	228	PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	229
michael@0	230	mov rcx, result_ptr
michael@0	231
michael@0	232	movq xmm0, xmm5
michael@0	233	psrldq xmm5, 8
michael@0	234
michael@0	235	paddw xmm0, xmm5
michael@0	236	movd [rcx], xmm0
michael@0	237	;-
michael@0	238	movq xmm0, xmm6
michael@0	239	psrldq xmm6, 8
michael@0	240
michael@0	241	paddw xmm0, xmm6
michael@0	242	movd [rcx+4], xmm0
michael@0	243	;-
michael@0	244	movq xmm0, xmm7
michael@0	245	psrldq xmm7, 8
michael@0	246
michael@0	247	paddw xmm0, xmm7
michael@0	248	movd [rcx+8], xmm0
michael@0	249
michael@0	250	STACK_FRAME_DESTROY_X3
michael@0	251
michael@0	252	;void int vp9_sad8x16x3_sse3(
michael@0	253	; unsigned char *src_ptr,
michael@0	254	; int src_stride,
michael@0	255	; unsigned char *ref_ptr,
michael@0	256	; int ref_stride,
michael@0	257	; int *results)
michael@0	258	global sym(vp9_sad8x16x3_sse3) PRIVATE
michael@0	259	sym(vp9_sad8x16x3_sse3):
michael@0	260
michael@0	261	STACK_FRAME_CREATE_X3
michael@0	262
michael@0	263	PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	264	PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	265	PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	266	PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	267	PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	268	PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	269	PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	270	PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	271
michael@0	272	mov rcx, result_ptr
michael@0	273
michael@0	274	punpckldq mm5, mm6
michael@0	275
michael@0	276	movq [rcx], mm5
michael@0	277	movd [rcx+8], mm7
michael@0	278
michael@0	279	STACK_FRAME_DESTROY_X3
michael@0	280
michael@0	281	;void int vp9_sad8x8x3_sse3(
michael@0	282	; unsigned char *src_ptr,
michael@0	283	; int src_stride,
michael@0	284	; unsigned char *ref_ptr,
michael@0	285	; int ref_stride,
michael@0	286	; int *results)
michael@0	287	global sym(vp9_sad8x8x3_sse3) PRIVATE
michael@0	288	sym(vp9_sad8x8x3_sse3):
michael@0	289
michael@0	290	STACK_FRAME_CREATE_X3
michael@0	291
michael@0	292	PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	293	PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	294	PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	295	PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
michael@0	296
michael@0	297	mov rcx, result_ptr
michael@0	298
michael@0	299	punpckldq mm5, mm6
michael@0	300
michael@0	301	movq [rcx], mm5
michael@0	302	movd [rcx+8], mm7
michael@0	303
michael@0	304	STACK_FRAME_DESTROY_X3
michael@0	305
michael@0	306	;void int vp9_sad4x4x3_sse3(
michael@0	307	; unsigned char *src_ptr,
michael@0	308	; int src_stride,
michael@0	309	; unsigned char *ref_ptr,
michael@0	310	; int ref_stride,
michael@0	311	; int *results)
michael@0	312	global sym(vp9_sad4x4x3_sse3) PRIVATE
michael@0	313	sym(vp9_sad4x4x3_sse3):
michael@0	314
michael@0	315	STACK_FRAME_CREATE_X3
michael@0	316
michael@0	317	movd mm0, DWORD PTR [src_ptr]
michael@0	318	movd mm1, DWORD PTR [ref_ptr]
michael@0	319
michael@0	320	movd mm2, DWORD PTR [src_ptr+src_stride]
michael@0	321	movd mm3, DWORD PTR [ref_ptr+ref_stride]
michael@0	322
michael@0	323	punpcklbw mm0, mm2
michael@0	324	punpcklbw mm1, mm3
michael@0	325
michael@0	326	movd mm4, DWORD PTR [ref_ptr+1]
michael@0	327	movd mm5, DWORD PTR [ref_ptr+2]
michael@0	328
michael@0	329	movd mm2, DWORD PTR [ref_ptr+ref_stride+1]
michael@0	330	movd mm3, DWORD PTR [ref_ptr+ref_stride+2]
michael@0	331
michael@0	332	psadbw mm1, mm0
michael@0	333
michael@0	334	punpcklbw mm4, mm2
michael@0	335	punpcklbw mm5, mm3
michael@0	336
michael@0	337	psadbw mm4, mm0
michael@0	338	psadbw mm5, mm0
michael@0	339
michael@0	340	lea src_ptr, [src_ptr+src_stride*2]
michael@0	341	lea ref_ptr, [ref_ptr+ref_stride*2]
michael@0	342
michael@0	343	movd mm0, DWORD PTR [src_ptr]
michael@0	344	movd mm2, DWORD PTR [ref_ptr]
michael@0	345
michael@0	346	movd mm3, DWORD PTR [src_ptr+src_stride]
michael@0	347	movd mm6, DWORD PTR [ref_ptr+ref_stride]
michael@0	348
michael@0	349	punpcklbw mm0, mm3
michael@0	350	punpcklbw mm2, mm6
michael@0	351
michael@0	352	movd mm3, DWORD PTR [ref_ptr+1]
michael@0	353	movd mm7, DWORD PTR [ref_ptr+2]
michael@0	354
michael@0	355	psadbw mm2, mm0
michael@0	356
michael@0	357	paddw mm1, mm2
michael@0	358
michael@0	359	movd mm2, DWORD PTR [ref_ptr+ref_stride+1]
michael@0	360	movd mm6, DWORD PTR [ref_ptr+ref_stride+2]
michael@0	361
michael@0	362	punpcklbw mm3, mm2
michael@0	363	punpcklbw mm7, mm6
michael@0	364
michael@0	365	psadbw mm3, mm0
michael@0	366	psadbw mm7, mm0
michael@0	367
michael@0	368	paddw mm3, mm4
michael@0	369	paddw mm7, mm5
michael@0	370
michael@0	371	mov rcx, result_ptr
michael@0	372
michael@0	373	punpckldq mm1, mm3
michael@0	374
michael@0	375	movq [rcx], mm1
michael@0	376	movd [rcx+8], mm7
michael@0	377
michael@0	378	STACK_FRAME_DESTROY_X3