Merge "ppc: Add vpx_iwht4x4_16_add_vsx"

James Zern committed 8 years ago

Commit c9bb5922 parents 3d8ea90 233aef03 files · +88 −10Browse files

Unified · Split

3 files

test/dct_test.cc +10 −0
vpx_dsp/ppc/inv_txfm_vsx.c +77 −9
vpx_dsp/vpx_dsp_rtcd_defs.pl +1 −1

@@ -725,4 +725,14 @@ INSTANTIATE_TEST_CASE_P(SSE2, TransWHT, · 724 unmodified lines
725	725	::testing::Values(make_tuple(0, &wht_sse2_func_info, 0,
726	726	VPX_BITS_8)));
727	727	#endif // HAVE_SSE2 && !CONFIG_EMULATE_HARDWARE
	728	+
	729	+#if HAVE_VSX && !CONFIG_EMULATE_HARDWARE && !CONFIG_VP9_HIGHBITDEPTH
	730	+static const FuncInfo wht_vsx_func_info = {
	731	+ &fdct_wrapper<vp9_fwht4x4_c>, &idct_wrapper<vpx_iwht4x4_16_add_vsx>, 4, 1
	732	+};
	733	+
	734	+INSTANTIATE_TEST_CASE_P(VSX, TransWHT,
	735	+ ::testing::Values(make_tuple(0, &wht_vsx_func_info, 0,
	736	+ VPX_BITS_8)));
	737	+#endif // HAVE_VSX && !CONFIG_EMULATE_HARDWARE
728	738	} // namespace

vpx_dsp/ppc/inv_txfm_vsx.c+77−9

@@ -76,6 +76,8 @@ static int16x8_t cospi29_v = { 2404, 2404, 2404, 2404, 2404, 2404, 2404, 2404 }; · 75 unmodified lines
76	76	static int16x8_t cospi30_v = { 1606, 1606, 1606, 1606, 1606, 1606, 1606, 1606 };
77	77	static int16x8_t cospi31_v = { 804, 804, 804, 804, 804, 804, 804, 804 };
78	78
	79	+static uint8x16_t mask1 = { 0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7,
	80	+ 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17 };
79	81	#define ROUND_SHIFT_INIT \
80	82	const int32x4_t shift = vec_sl(vec_splat_s32(1), vec_splat_u32(13)); \
81	83	const uint32x4_t shift14 = vec_splat_u32(14);
@@ -107,6 +109,15 @@ static int16x8_t cospi31_v = { 804, 804, 804, 804, 804, 804, 804, 804 }; · 25 unmodified lines
107	109	out1 = vec_sub(step0, step1); \
108	110	out1 = vec_perm(out1, out1, mask0);
109	111
	112	+#define PACK_STORE(v0, v1) \
	113	+ tmp16_0 = vec_add(vec_perm(d_u0, d_u1, mask1), v0); \
	114	+ tmp16_1 = vec_add(vec_perm(d_u2, d_u3, mask1), v1); \
	115	+ output_v = vec_packsu(tmp16_0, tmp16_1); \
	116	+ \
	117	+ vec_vsx_st(output_v, 0, tmp_dest); \
	118	+ for (i = 0; i < 4; i++) \
	119	+ for (j = 0; j < 4; j++) dest[j * stride + i] = tmp_dest[j * 4 + i];
	120	+
110	121	void vpx_idct4x4_16_add_vsx(const tran_low_t input, uint8_t dest,
111	122	int stride) {
112	123	int i, j;
@@ -114,13 +125,10 @@ void vpx_idct4x4_16_add_vsx(const tran_low_t input, uint8_t dest, · 1 unmodified line
114	125	int16x8_t step0, step1, tmp16_0, tmp16_1, t_out0, t_out1;
115	126	uint8x16_t mask0 = { 0x8, 0x9, 0xA, 0xB, 0xC, 0xD, 0xE, 0xF,
116	127	0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7 };
117		− uint8x16_t mask1 = { 0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7,
118		− 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17 };
119	128	int16x8_t v0 = load_tran_low(0, input);
120	129	int16x8_t v1 = load_tran_low(8 * sizeof(*input), input);
121	130	int16x8_t t0 = vec_mergeh(v0, v1);
122	131	int16x8_t t1 = vec_mergel(v0, v1);
123		−
124	132	uint8x16_t dest0 = vec_vsx_ld(0, dest);
125	133	uint8x16_t dest1 = vec_vsx_ld(stride, dest);
126	134	uint8x16_t dest2 = vec_vsx_ld(2 * stride, dest);
@@ -130,6 +138,7 @@ void vpx_idct4x4_16_add_vsx(const tran_low_t input, uint8_t dest, · 3 unmodified lines
130	138	int16x8_t d_u1 = (int16x8_t)vec_mergeh(dest1, zerov);
131	139	int16x8_t d_u2 = (int16x8_t)vec_mergeh(dest2, zerov);
132	140	int16x8_t d_u3 = (int16x8_t)vec_mergeh(dest3, zerov);
	141	+
133	142	uint8x16_t output_v;
134	143	uint8_t tmp_dest[16];
135	144	ROUND_SHIFT_INIT
@@ -148,13 +157,8 @@ void vpx_idct4x4_16_add_vsx(const tran_low_t input, uint8_t dest, · 12 unmodified lines
148	157
149	158	PIXEL_ADD4(v0, t_out0);
150	159	PIXEL_ADD4(v1, t_out1);
151		− tmp16_0 = vec_add(vec_perm(d_u0, d_u1, mask1), v0);
152		− tmp16_1 = vec_add(vec_perm(d_u2, d_u3, mask1), v1);
153		− output_v = vec_packsu(tmp16_0, tmp16_1);
154	160
155		− vec_vsx_st(output_v, 0, tmp_dest);
156		− for (i = 0; i < 4; i++)
157		− for (j = 0; j < 4; j++) dest[j * stride + i] = tmp_dest[j * 4 + i];
	161	+ PACK_STORE(v0, v1);
158	162	}
159	163
160	164	#define TRANSPOSE8x8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, \
@@ -1062,3 +1066,67 @@ void vpx_idct32x32_1024_add_vsx(const tran_low_t input, uint8_t dest, · 901 unmodified lines
1062	1066	ADD_STORE_BLOCK(src2, 16);
1063	1067	ADD_STORE_BLOCK(src3, 24);
1064	1068	}
	1069	+
	1070	+#define TRANSFORM_COLS \
	1071	+ v32_a = vec_add(v32_a, v32_c); \
	1072	+ v32_d = vec_sub(v32_d, v32_b); \
	1073	+ v32_e = vec_sub(v32_a, v32_d); \
	1074	+ v32_e = vec_sra(v32_e, one); \
	1075	+ v32_b = vec_sub(v32_e, v32_b); \
	1076	+ v32_c = vec_sub(v32_e, v32_c); \
	1077	+ v32_a = vec_sub(v32_a, v32_b); \
	1078	+ v32_d = vec_add(v32_d, v32_c); \
	1079	+ v_a = vec_packs(v32_a, v32_b); \
	1080	+ v_c = vec_packs(v32_c, v32_d);
	1081	+
	1082	+#define TRANSPOSE_WHT \
	1083	+ tmp_a = vec_mergeh(v_a, v_c); \
	1084	+ tmp_c = vec_mergel(v_a, v_c); \
	1085	+ v_a = vec_mergeh(tmp_a, tmp_c); \
	1086	+ v_c = vec_mergel(tmp_a, tmp_c);
	1087	+
	1088	+void vpx_iwht4x4_16_add_vsx(const tran_low_t input, uint8_t dest,
	1089	+ int stride) {
	1090	+ int16x8_t v_a = load_tran_low(0, input);
	1091	+ int16x8_t v_c = load_tran_low(8 * sizeof(*input), input);
	1092	+ int16x8_t tmp_a, tmp_c;
	1093	+ uint16x8_t two = vec_splat_u16(2);
	1094	+ uint32x4_t one = vec_splat_u32(1);
	1095	+ int16x8_t tmp16_0, tmp16_1;
	1096	+ int32x4_t v32_a, v32_c, v32_d, v32_b, v32_e;
	1097	+ uint8x16_t dest0 = vec_vsx_ld(0, dest);
	1098	+ uint8x16_t dest1 = vec_vsx_ld(stride, dest);
	1099	+ uint8x16_t dest2 = vec_vsx_ld(2 * stride, dest);
	1100	+ uint8x16_t dest3 = vec_vsx_ld(3 * stride, dest);
	1101	+ int16x8_t d_u0 = (int16x8_t)unpack_to_u16_h(dest0);
	1102	+ int16x8_t d_u1 = (int16x8_t)unpack_to_u16_h(dest1);
	1103	+ int16x8_t d_u2 = (int16x8_t)unpack_to_u16_h(dest2);
	1104	+ int16x8_t d_u3 = (int16x8_t)unpack_to_u16_h(dest3);
	1105	+ uint8x16_t output_v;
	1106	+ uint8_t tmp_dest[16];
	1107	+ int i, j;
	1108	+
	1109	+ v_a = vec_sra(v_a, two);
	1110	+ v_c = vec_sra(v_c, two);
	1111	+
	1112	+ TRANSPOSE_WHT;
	1113	+
	1114	+ v32_a = vec_unpackh(v_a);
	1115	+ v32_c = vec_unpackl(v_a);
	1116	+
	1117	+ v32_d = vec_unpackh(v_c);
	1118	+ v32_b = vec_unpackl(v_c);
	1119	+
	1120	+ TRANSFORM_COLS;
	1121	+
	1122	+ TRANSPOSE_WHT;
	1123	+
	1124	+ v32_a = vec_unpackh(v_a);
	1125	+ v32_c = vec_unpackl(v_a);
	1126	+ v32_d = vec_unpackh(v_c);
	1127	+ v32_b = vec_unpackl(v_c);
	1128	+
	1129	+ TRANSFORM_COLS;
	1130	+
	1131	+ PACK_STORE(v_a, v_c);
	1132	+}

vpx_dsp/vpx_dsp_rtcd_defs.pl+1−1

@@ -626,7 +626,7 @@ if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") { · 625 unmodified lines
626	626	specialize qw/vpx_idct32x32_135_add neon sse2 ssse3/;
627	627	specialize qw/vpx_idct32x32_34_add neon sse2 ssse3/;
628	628	specialize qw/vpx_idct32x32_1_add neon sse2/;
629		− specialize qw/vpx_iwht4x4_16_add sse2/;
	629	+ specialize qw/vpx_iwht4x4_16_add sse2 vsx/;
630	630
631	631	if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") ne "yes") {
632	632	# Note that these specializations are appended to the above ones.

Merge "ppc: Add vpx_iwht4x4_16_add_vsx"

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code