x86/aacpsdsp: add ff_ps_hybrid_synthesis_deint_{sse,sse4}

About 2x faster than the c version.

James Almer committed 9 years ago

Commit e229df91 parent 3c5a53c3 files · +140 −6Browse files

Unified · Split

3 files

libavcodec/x86/aacpsdsp.asm +123 −0
libavcodec/x86/aacpsdsp_init.c +8 −0
libavutil/x86/x86util.asm +9 −6

@@ -166,6 +166,129 @@ align 16 · 165 unmodified lines
166	166	jl .loop
167	167	REP_RET
168	168
	169	+;***********************************************************
	170	+;void ps_hybrid_synthesis_deint_sse4(float out[2][38][64],
	171	+; float (*in)[32][2],
	172	+; int i, int len)
	173	+;***********************************************************
	174	+%macro HYBRID_SYNTHESIS_DEINT 0
	175	+cglobal ps_hybrid_synthesis_deint, 3, 7, 5, out, in, i, len, out0, out1, tmp
	176	+%if cpuflag(sse4)
	177	+%define MOVH movsd
	178	+%else
	179	+%define MOVH movlps
	180	+%endif
	181	+ movsxdifnidn iq, id
	182	+ mov lend, 32 << 3
	183	+ lea outq, [outq+iq*4]
	184	+ mov tmpd, id
	185	+ shl tmpd, 8
	186	+ add inq, tmpq
	187	+ mov tmpd, 64
	188	+ sub tmpd, id
	189	+ mov id, tmpd
	190	+
	191	+ test id, 1
	192	+ jne .loop4
	193	+ test id, 2
	194	+ jne .loop8
	195	+
	196	+align 16
	197	+.loop16:
	198	+ mov out0q, outq
	199	+ mov out1q, 38644
	200	+ add out1q, out0q
	201	+ mov tmpd, lend
	202	+
	203	+.inner_loop16:
	204	+ movaps m0, [inq]
	205	+ movaps m1, [inq+lenq]
	206	+ movaps m2, [inq+lenq*2]
	207	+ movaps m3, [inq+3322*4]
	208	+ TRANSPOSE4x4PS 0, 1, 2, 3, 4
	209	+ movaps [out0q], m0
	210	+ movaps [out1q], m1
	211	+ movaps [out0q+lenq], m2
	212	+ movaps [out1q+lenq], m3
	213	+ lea out0q, [out0q+lenq*2]
	214	+ lea out1q, [out1q+lenq*2]
	215	+ add inq, mmsize
	216	+ sub tmpd, mmsize
	217	+ jg .inner_loop16
	218	+ add outq, 16
	219	+ add inq, 3322*4
	220	+ sub id, 4
	221	+ jg .loop16
	222	+ RET
	223	+
	224	+align 16
	225	+.loop8:
	226	+ mov out0q, outq
	227	+ mov out1q, 38644
	228	+ add out1q, out0q
	229	+ mov tmpd, lend
	230	+
	231	+.inner_loop8:
	232	+ movaps m0, [inq]
	233	+ movaps m1, [inq+lenq]
	234	+ SBUTTERFLYPS 0, 1, 2
	235	+ SBUTTERFLYPD 0, 1, 2
	236	+ MOVH [out0q], m0
	237	+ MOVH [out1q], m1
	238	+ movhps [out0q+lenq], m0
	239	+ movhps [out1q+lenq], m1
	240	+ lea out0q, [out0q+lenq*2]
	241	+ lea out1q, [out1q+lenq*2]
	242	+ add inq, mmsize
	243	+ sub tmpd, mmsize
	244	+ jg .inner_loop8
	245	+ add outq, 8
	246	+ add inq, lenq
	247	+ sub id, 2
	248	+ jg .loop16
	249	+ RET
	250	+
	251	+align 16
	252	+.loop4:
	253	+ mov out0q, outq
	254	+ mov out1q, 38644
	255	+ add out1q, out0q
	256	+ mov tmpd, lend
	257	+
	258	+.inner_loop4:
	259	+ movaps m0, [inq]
	260	+ movss [out0q], m0
	261	+%if cpuflag(sse4)
	262	+ extractps [out1q], m0, 1
	263	+ extractps [out0q+lenq], m0, 2
	264	+ extractps [out1q+lenq], m0, 3
	265	+%else
	266	+ movhlps m1, m0
	267	+ movss [out0q+lenq], m1
	268	+ shufps m0, m0, 0xb1
	269	+ movss [out1q], m0
	270	+ movhlps m1, m0
	271	+ movss [out1q+lenq], m1
	272	+%endif
	273	+ lea out0q, [out0q+lenq*2]
	274	+ lea out1q, [out1q+lenq*2]
	275	+ add inq, mmsize
	276	+ sub tmpd, mmsize
	277	+ jg .inner_loop4
	278	+ add outq, 4
	279	+ sub id, 1
	280	+ test id, 2
	281	+ jne .loop8
	282	+ cmp id, 4
	283	+ jge .loop16
	284	+ RET
	285	+%endmacro
	286	+
	287	+INIT_XMM sse
	288	+HYBRID_SYNTHESIS_DEINT
	289	+INIT_XMM sse4
	290	+HYBRID_SYNTHESIS_DEINT
	291	+
169	292	;*******************************************************************
170	293	;void ff_ps_hybrid_analysis_<opt>(float (out)[2], float (in)[2],
171	294	; const float (*filter)[8][2],

libavcodec/x86/aacpsdsp_init.c+8−0

@@ -40,6 +40,10 @@ void ff_ps_stereo_interpolate_sse3(float (l)[2], float (r)[2], · 39 unmodified lines
40	40	void ff_ps_stereo_interpolate_ipdopd_sse3(float (l)[2], float (r)[2],
41	41	float h[2][4], float h_step[2][4],
42	42	int len);
	43	+void ff_ps_hybrid_synthesis_deint_sse(float out[2][38][64], float (*in)[32][2],
	44	+ int i, int len);
	45	+void ff_ps_hybrid_synthesis_deint_sse4(float out[2][38][64], float (*in)[32][2],
	46	+ int i, int len);
43	47
44	48	av_cold void ff_psdsp_init_x86(PSDSPContext *s)
45	49	{
@@ -48,6 +52,7 @@ av_cold void ff_psdsp_init_x86(PSDSPContext *s) · 2 unmodified lines
48	52	if (EXTERNAL_SSE(cpu_flags)) {
49	53	s->add_squares = ff_ps_add_squares_sse;
50	54	s->mul_pair_single = ff_ps_mul_pair_single_sse;
	55	+ s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse;
51	56	s->hybrid_analysis = ff_ps_hybrid_analysis_sse;
52	57	}
53	58	if (EXTERNAL_SSE3(cpu_flags)) {
@@ -56,4 +61,7 @@ av_cold void ff_psdsp_init_x86(PSDSPContext *s) · 2 unmodified lines
56	61	s->stereo_interpolate[1] = ff_ps_stereo_interpolate_ipdopd_sse3;
57	62	s->hybrid_analysis = ff_ps_hybrid_analysis_sse3;
58	63	}
	64	+ if (EXTERNAL_SSE4(cpu_flags)) {
	65	+ s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse4;
	66	+ }
59	67	}

libavutil/x86/x86util.asm+9−6

@@ -71,6 +71,12 @@ · 70 unmodified lines
71	71	SWAP %1, %3, %2
72	72	%endmacro
73	73
	74	+%macro SBUTTERFLYPD 3
	75	+ movlhps m%3, m%1, m%2
	76	+ movhlps m%2, m%2, m%1
	77	+ SWAP %1, %3
	78	+%endmacro
	79	+
74	80	%macro TRANSPOSE4x4B 5
75	81	SBUTTERFLY bw, %1, %2, %5
76	82	SBUTTERFLY bw, %3, %4, %5
@@ -117,12 +123,9 @@ · 40 unmodified lines
117	123	%macro TRANSPOSE4x4PS 5
118	124	SBUTTERFLYPS %1, %2, %5
119	125	SBUTTERFLYPS %3, %4, %5
120		− movlhps m%5, m%1, m%3
121		− movhlps m%3, m%1
122		− SWAP %5, %1
123		− movlhps m%5, m%2, m%4
124		− movhlps m%4, m%2
125		− SWAP %5, %2, %3
	126	+ SBUTTERFLYPD %1, %3, %5
	127	+ SBUTTERFLYPD %2, %4, %5
	128	+ SWAP %2, %3
126	129	%endmacro
127	130
128	131	%macro TRANSPOSE8x4D 9-11

x86/aacpsdsp: add ff_ps_hybrid_synthesis_deint_{sse,sse4}

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code