diff options
author | David Walter Seikel | 2012-01-04 18:41:13 +1000 |
---|---|---|
committer | David Walter Seikel | 2012-01-04 18:41:13 +1000 |
commit | dd7595a3475407a7fa96a97393bae8c5220e8762 (patch) | |
tree | e341e911d7eb911a51684a7412ef7f7c7605d28e /libraries/evas/src/lib/engines/common/evas_op_blend/op_blend_color_sse3.c | |
parent | Add the skeleton. (diff) | |
download | SledjHamr-dd7595a3475407a7fa96a97393bae8c5220e8762.zip SledjHamr-dd7595a3475407a7fa96a97393bae8c5220e8762.tar.gz SledjHamr-dd7595a3475407a7fa96a97393bae8c5220e8762.tar.bz2 SledjHamr-dd7595a3475407a7fa96a97393bae8c5220e8762.tar.xz |
Add the base Enlightenment Foundation Libraries - eina, eet, evas, ecore, embryo, and edje.
Note that embryo wont be used, but I'm not sure yet if you can build edje without it.
Diffstat (limited to '')
-rw-r--r-- | libraries/evas/src/lib/engines/common/evas_op_blend/op_blend_color_sse3.c | 166 |
1 files changed, 166 insertions, 0 deletions
diff --git a/libraries/evas/src/lib/engines/common/evas_op_blend/op_blend_color_sse3.c b/libraries/evas/src/lib/engines/common/evas_op_blend/op_blend_color_sse3.c new file mode 100644 index 0000000..64d5a86 --- /dev/null +++ b/libraries/evas/src/lib/engines/common/evas_op_blend/op_blend_color_sse3.c | |||
@@ -0,0 +1,166 @@ | |||
1 | /* blend color -> dst */ | ||
2 | |||
3 | #ifdef BUILD_SSE3 | ||
4 | |||
5 | static void | ||
6 | _op_blend_c_dp_sse3(DATA32 *s __UNUSED__, DATA8 *m __UNUSED__, DATA32 c, DATA32 *d, int l) { | ||
7 | |||
8 | DATA32 a = 256 - (c >> 24); | ||
9 | |||
10 | const __m128i c_packed = _mm_set_epi32(c, c, c, c); | ||
11 | const __m128i a_packed = _mm_set_epi32(a, a, a, a); | ||
12 | |||
13 | LOOP_ALIGNED_U1_A48_SSE3(d, l, | ||
14 | { /* UOP */ | ||
15 | |||
16 | *d = c + MUL_256(a, *d); | ||
17 | d++; l--; | ||
18 | }, | ||
19 | { /* A4OP */ | ||
20 | |||
21 | __m128i d0 = _mm_load_si128((__m128i *)d); | ||
22 | |||
23 | d0 = mul_256_sse3(a_packed, d0); | ||
24 | d0 = _mm_add_epi32(d0, c_packed); | ||
25 | |||
26 | _mm_store_si128((__m128i *)d, d0); | ||
27 | |||
28 | d += 4; l -= 4; | ||
29 | }, | ||
30 | { /* A8OP */ | ||
31 | |||
32 | __m128i d0 = _mm_load_si128((__m128i *)d); | ||
33 | __m128i d1 = _mm_load_si128((__m128i *)(d+4)); | ||
34 | |||
35 | d0 = mul_256_sse3(a_packed, d0); | ||
36 | d1 = mul_256_sse3(a_packed, d1); | ||
37 | |||
38 | d0 = _mm_add_epi32(d0, c_packed); | ||
39 | d1 = _mm_add_epi32(d1, c_packed); | ||
40 | |||
41 | _mm_store_si128((__m128i *)d, d0); | ||
42 | _mm_store_si128((__m128i *)(d+4), d1); | ||
43 | |||
44 | d += 8; l -= 8; | ||
45 | }) | ||
46 | } | ||
47 | |||
48 | #define _op_blend_caa_dp_sse3 _op_blend_c_dp_sse3 | ||
49 | |||
50 | #define _op_blend_c_dpan_sse3 _op_blend_c_dp_sse3 | ||
51 | #define _op_blend_caa_dpan_sse3 _op_blend_c_dpan_sse3 | ||
52 | |||
53 | static void | ||
54 | init_blend_color_span_funcs_sse3(void) | ||
55 | { | ||
56 | op_blend_span_funcs[SP_N][SM_N][SC][DP][CPU_SSE3] = _op_blend_c_dp_sse3; | ||
57 | op_blend_span_funcs[SP_N][SM_N][SC_AA][DP][CPU_SSE3] = _op_blend_caa_dp_sse3; | ||
58 | |||
59 | // FIXME: BUGGY BUGGY Core i5 750 (32bit), 4.5.2 (Ubuntu/Linaro 4.5.2-8ubuntu4), ello (text and rectangle) | ||
60 | // op_blend_span_funcs[SP_N][SM_N][SC][DP_AN][CPU_SSE3] = _op_blend_c_dpan_sse3; | ||
61 | op_blend_span_funcs[SP_N][SM_N][SC_AA][DP_AN][CPU_SSE3] = _op_blend_caa_dpan_sse3; | ||
62 | } | ||
63 | |||
64 | #define _op_blend_pt_c_dp_sse3 NULL | ||
65 | #define _op_blend_pt_caa_dp_sse3 _op_blend_pt_c_dp_sse3 | ||
66 | |||
67 | #define _op_blend_pt_c_dpan_sse3 _op_blend_pt_c_dp_sse3 | ||
68 | #define _op_blend_pt_caa_dpan_sse3 _op_blend_pt_c_dpan_sse3 | ||
69 | |||
70 | #define _op_blend_pt_c_dpas_sse3 _op_blend_pt_c_dp_sse3 | ||
71 | #define _op_blend_pt_caa_dpas_sse3 _op_blend_pt_c_dp_sse3 | ||
72 | |||
73 | static void | ||
74 | init_blend_color_pt_funcs_sse3(void) | ||
75 | { | ||
76 | op_blend_pt_funcs[SP_N][SM_N][SC][DP][CPU_SSE3] = _op_blend_pt_c_dp_sse3; | ||
77 | op_blend_pt_funcs[SP_N][SM_N][SC_AA][DP][CPU_SSE3] = _op_blend_pt_caa_dp_sse3; | ||
78 | |||
79 | op_blend_pt_funcs[SP_N][SM_N][SC][DP_AN][CPU_SSE3] = _op_blend_pt_c_dpan_sse3; | ||
80 | op_blend_pt_funcs[SP_N][SM_N][SC_AA][DP_AN][CPU_SSE3] = _op_blend_pt_caa_dpan_sse3; | ||
81 | } | ||
82 | |||
83 | |||
84 | /*-----*/ | ||
85 | |||
86 | /* blend_rel color -> dst */ | ||
87 | |||
88 | static void | ||
89 | _op_blend_rel_c_dp_sse3(DATA32 *s __UNUSED__, DATA8 *m __UNUSED__, DATA32 c, DATA32 *d, int l) { | ||
90 | |||
91 | int alpha = 256 - (c >> 24); | ||
92 | |||
93 | const __m128i c_packed = _mm_set_epi32(c, c, c, c); | ||
94 | const __m128i alpha_packed = _mm_set_epi32(alpha, alpha, alpha, alpha); | ||
95 | |||
96 | LOOP_ALIGNED_U1_A48_SSE3(d, l, | ||
97 | { /* UOP */ | ||
98 | |||
99 | *d = MUL_SYM(*d >> 24, c) + MUL_256(alpha, *d); | ||
100 | d++; l--; | ||
101 | }, | ||
102 | { /* A4OP */ | ||
103 | |||
104 | __m128i d0 = _mm_load_si128((__m128i *)d); | ||
105 | |||
106 | __m128i mul0 = mul_256_sse3(alpha_packed, d0); | ||
107 | __m128i sym0 = mul_sym_sse3(_mm_srli_epi32(d0, 24), c_packed); | ||
108 | |||
109 | d0 = _mm_add_epi32(mul0, sym0); | ||
110 | |||
111 | _mm_store_si128((__m128i *)d, d0); | ||
112 | |||
113 | d += 4; l -= 4; | ||
114 | }, | ||
115 | { /* A8OP */ | ||
116 | |||
117 | __m128i d0 = _mm_load_si128((__m128i *)d); | ||
118 | __m128i d1 = _mm_load_si128((__m128i *)(d+4)); | ||
119 | |||
120 | __m128i mul0 = mul_256_sse3(alpha_packed, d0); | ||
121 | __m128i mul1 = mul_256_sse3(alpha_packed, d1); | ||
122 | |||
123 | __m128i sym0 = mul_sym_sse3(_mm_srli_epi32(d0, 24), c_packed); | ||
124 | __m128i sym1 = mul_sym_sse3(_mm_srli_epi32(d1, 24), c_packed); | ||
125 | |||
126 | d0 = _mm_add_epi32(mul0, sym0); | ||
127 | d1 = _mm_add_epi32(mul1, sym1); | ||
128 | |||
129 | _mm_store_si128((__m128i *)d, d0); | ||
130 | _mm_store_si128((__m128i *)(d+4), d1); | ||
131 | |||
132 | d += 8; l -= 8; | ||
133 | }) | ||
134 | } | ||
135 | |||
136 | #define _op_blend_rel_caa_dp_sse3 _op_blend_rel_c_dp_sse3 | ||
137 | #define _op_blend_rel_c_dpan_sse3 _op_blend_c_dpan_sse3 | ||
138 | #define _op_blend_rel_caa_dpan_sse3 _op_blend_caa_dpan_sse3 | ||
139 | |||
140 | static void | ||
141 | init_blend_rel_color_span_funcs_sse3(void) | ||
142 | { | ||
143 | op_blend_rel_span_funcs[SP_N][SM_N][SC][DP][CPU_SSE3] = _op_blend_rel_c_dp_sse3; | ||
144 | op_blend_rel_span_funcs[SP_N][SM_N][SC_AA][DP][CPU_SSE3] = _op_blend_rel_caa_dp_sse3; | ||
145 | |||
146 | op_blend_rel_span_funcs[SP_N][SM_N][SC][DP_AN][CPU_SSE3] = _op_blend_rel_c_dpan_sse3; | ||
147 | op_blend_rel_span_funcs[SP_N][SM_N][SC_AA][DP_AN][CPU_SSE3] = _op_blend_rel_caa_dpan_sse3; | ||
148 | } | ||
149 | |||
150 | #define _op_blend_rel_pt_c_dp_sse3 NULL | ||
151 | #define _op_blend_rel_pt_caa_dp_sse3 _op_blend_rel_pt_c_dp_sse3 | ||
152 | |||
153 | #define _op_blend_rel_pt_c_dpan_sse3 _op_blend_pt_c_dpan_sse3 | ||
154 | #define _op_blend_rel_pt_caa_dpan_sse3 _op_blend_pt_caa_dpan_sse3 | ||
155 | |||
156 | static void | ||
157 | init_blend_rel_color_pt_funcs_sse3(void) | ||
158 | { | ||
159 | op_blend_rel_pt_funcs[SP_N][SM_N][SC][DP][CPU_SSE3] = _op_blend_rel_pt_c_dp_sse3; | ||
160 | op_blend_rel_pt_funcs[SP_N][SM_N][SC_AA][DP][CPU_SSE3] = _op_blend_rel_pt_caa_dp_sse3; | ||
161 | |||
162 | op_blend_rel_pt_funcs[SP_N][SM_N][SC][DP_AN][CPU_SSE3] = _op_blend_rel_pt_c_dpan_sse3; | ||
163 | op_blend_rel_pt_funcs[SP_N][SM_N][SC_AA][DP_AN][CPU_SSE3] = _op_blend_rel_pt_caa_dpan_sse3; | ||
164 | } | ||
165 | |||
166 | #endif | ||