Merge commit 'origin/perrtblend'
[mesa.git] / src / gallium / drivers / llvmpipe / lp_bld_blend_soa.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28
29 /**
30 * @file
31 * Blend LLVM IR generation -- SoA layout.
32 *
33 * Blending in SoA is much faster than AoS, especially when separate rgb/alpha
34 * factors/functions are used, since no channel masking/shuffling is necessary
35 * and we can achieve the full throughput of the SIMD operations. Furthermore
36 * the fragment shader output is also in SoA, so it fits nicely with the rest of
37 * the fragment pipeline.
38 *
39 * The drawback is that to be displayed the color buffer needs to be in AoS
40 * layout, so we need to tile/untile the color buffer before/after rendering.
41 * A color buffer like
42 *
43 * R11 G11 B11 A11 R12 G12 B12 A12 R13 G13 B13 A13 R14 G14 B14 A14 ...
44 * R21 G21 B21 A21 R22 G22 B22 A22 R23 G23 B23 A23 R24 G24 B24 A24 ...
45 *
46 * R31 G31 B31 A31 R32 G32 B32 A32 R33 G33 B33 A33 R34 G34 B34 A34 ...
47 * R41 G41 B41 A41 R42 G42 B42 A42 R43 G43 B43 A43 R44 G44 B44 A44 ...
48 *
49 * ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
50 *
51 * will actually be stored in memory as
52 *
53 * R11 R12 R21 R22 R13 R14 R23 R24 ... G11 G12 G21 G22 G13 G14 G23 G24 ... B11 B12 B21 B22 B13 B14 B23 B24 ... A11 A12 A21 A22 A13 A14 A23 A24 ...
54 * R31 R32 R41 R42 R33 R34 R43 R44 ... G31 G32 G41 G42 G33 G34 G43 G44 ... B31 B32 B41 B42 B33 B34 B43 B44 ... A31 A32 A41 A42 A33 A34 A43 A44 ...
55 * ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
56 *
57 * NOTE: Run lp_blend_test after any change to this file.
58 *
59 * You can also run lp_blend_test to obtain AoS vs SoA benchmarks. Invoking it
60 * as:
61 *
62 * lp_blend_test -o blend.tsv
63 *
64 * will generate a tab-seperated-file with the test results and performance
65 * measurements.
66 *
67 * @author Jose Fonseca <jfonseca@vmware.com>
68 */
69
70
71 #include "pipe/p_state.h"
72
73 #include "lp_bld_type.h"
74 #include "lp_bld_arit.h"
75 #include "lp_bld_blend.h"
76
77
78 /**
79 * We may the same values several times, so we keep them here to avoid
80 * recomputing them. Also reusing the values allows us to do simplifications
81 * that LLVM optimization passes wouldn't normally be able to do.
82 */
83 struct lp_build_blend_soa_context
84 {
85 struct lp_build_context base;
86
87 LLVMValueRef src[4];
88 LLVMValueRef dst[4];
89 LLVMValueRef con[4];
90
91 LLVMValueRef inv_src[4];
92 LLVMValueRef inv_dst[4];
93 LLVMValueRef inv_con[4];
94
95 LLVMValueRef src_alpha_saturate;
96
97 /**
98 * We store all factors in a table in order to eliminate redundant
99 * multiplications later.
100 */
101 LLVMValueRef factor[2][2][4];
102
103 /**
104 * Table with all terms.
105 */
106 LLVMValueRef term[2][4];
107 };
108
109
110 static LLVMValueRef
111 lp_build_blend_soa_factor(struct lp_build_blend_soa_context *bld,
112 unsigned factor, unsigned i)
113 {
114 /*
115 * Compute src/first term RGB
116 */
117 switch (factor) {
118 case PIPE_BLENDFACTOR_ONE:
119 return bld->base.one;
120 case PIPE_BLENDFACTOR_SRC_COLOR:
121 return bld->src[i];
122 case PIPE_BLENDFACTOR_SRC_ALPHA:
123 return bld->src[3];
124 case PIPE_BLENDFACTOR_DST_COLOR:
125 return bld->dst[i];
126 case PIPE_BLENDFACTOR_DST_ALPHA:
127 return bld->dst[3];
128 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
129 if(i == 3)
130 return bld->base.one;
131 else {
132 if(!bld->inv_dst[3])
133 bld->inv_dst[3] = lp_build_comp(&bld->base, bld->dst[3]);
134 if(!bld->src_alpha_saturate)
135 bld->src_alpha_saturate = lp_build_min(&bld->base, bld->src[3], bld->inv_dst[3]);
136 return bld->src_alpha_saturate;
137 }
138 case PIPE_BLENDFACTOR_CONST_COLOR:
139 return bld->con[i];
140 case PIPE_BLENDFACTOR_CONST_ALPHA:
141 return bld->con[3];
142 case PIPE_BLENDFACTOR_SRC1_COLOR:
143 /* TODO */
144 assert(0);
145 return bld->base.zero;
146 case PIPE_BLENDFACTOR_SRC1_ALPHA:
147 /* TODO */
148 assert(0);
149 return bld->base.zero;
150 case PIPE_BLENDFACTOR_ZERO:
151 return bld->base.zero;
152 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
153 if(!bld->inv_src[i])
154 bld->inv_src[i] = lp_build_comp(&bld->base, bld->src[i]);
155 return bld->inv_src[i];
156 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
157 if(!bld->inv_src[3])
158 bld->inv_src[3] = lp_build_comp(&bld->base, bld->src[3]);
159 return bld->inv_src[3];
160 case PIPE_BLENDFACTOR_INV_DST_COLOR:
161 if(!bld->inv_dst[i])
162 bld->inv_dst[i] = lp_build_comp(&bld->base, bld->dst[i]);
163 return bld->inv_dst[i];
164 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
165 if(!bld->inv_dst[3])
166 bld->inv_dst[3] = lp_build_comp(&bld->base, bld->dst[3]);
167 return bld->inv_dst[3];
168 case PIPE_BLENDFACTOR_INV_CONST_COLOR:
169 if(!bld->inv_con[i])
170 bld->inv_con[i] = lp_build_comp(&bld->base, bld->con[i]);
171 return bld->inv_con[i];
172 case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
173 if(!bld->inv_con[3])
174 bld->inv_con[3] = lp_build_comp(&bld->base, bld->con[3]);
175 return bld->inv_con[3];
176 case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
177 /* TODO */
178 assert(0);
179 return bld->base.zero;
180 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
181 /* TODO */
182 assert(0);
183 return bld->base.zero;
184 default:
185 assert(0);
186 return bld->base.zero;
187 }
188 }
189
190
191 /**
192 * Generate blend code in SOA mode.
193 * \param src src/fragment color
194 * \param dst dst/framebuffer color
195 * \param con constant blend color
196 * \param res the result/output
197 */
198 void
199 lp_build_blend_soa(LLVMBuilderRef builder,
200 const struct pipe_blend_state *blend,
201 struct lp_type type,
202 LLVMValueRef src[4],
203 LLVMValueRef dst[4],
204 LLVMValueRef con[4],
205 LLVMValueRef res[4])
206 {
207 struct lp_build_blend_soa_context bld;
208 unsigned i, j, k;
209
210 /* Setup build context */
211 memset(&bld, 0, sizeof bld);
212 lp_build_context_init(&bld.base, builder, type);
213 for (i = 0; i < 4; ++i) {
214 bld.src[i] = src[i];
215 bld.dst[i] = dst[i];
216 bld.con[i] = con[i];
217 }
218
219 for (i = 0; i < 4; ++i) {
220 if (blend->rt[0].colormask & (1 << i)) {
221 if (blend->logicop_enable) {
222 if(!type.floating) {
223 res[i] = lp_build_logicop(builder, blend->logicop_func, src[i], dst[i]);
224 }
225 else
226 res[i] = dst[i];
227 }
228 else if (blend->rt[0].blend_enable) {
229 unsigned src_factor = i < 3 ? blend->rt[0].rgb_src_factor : blend->rt[0].alpha_src_factor;
230 unsigned dst_factor = i < 3 ? blend->rt[0].rgb_dst_factor : blend->rt[0].alpha_dst_factor;
231 unsigned func = i < 3 ? blend->rt[0].rgb_func : blend->rt[0].alpha_func;
232 boolean func_commutative = lp_build_blend_func_commutative(func);
233
234 /* It makes no sense to blend unless values are normalized */
235 assert(type.norm);
236
237 /*
238 * Compute src/dst factors.
239 */
240
241 bld.factor[0][0][i] = src[i];
242 bld.factor[0][1][i] = lp_build_blend_soa_factor(&bld, src_factor, i);
243 bld.factor[1][0][i] = dst[i];
244 bld.factor[1][1][i] = lp_build_blend_soa_factor(&bld, dst_factor, i);
245
246 /*
247 * Compute src/dst terms
248 */
249
250 for(k = 0; k < 2; ++k) {
251 /* See if this multiplication has been previously computed */
252 for(j = 0; j < i; ++j) {
253 if((bld.factor[k][0][j] == bld.factor[k][0][i] &&
254 bld.factor[k][1][j] == bld.factor[k][1][i]) ||
255 (bld.factor[k][0][j] == bld.factor[k][1][i] &&
256 bld.factor[k][1][j] == bld.factor[k][0][i]))
257 break;
258 }
259
260 if(j < i)
261 bld.term[k][i] = bld.term[k][j];
262 else
263 bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i], bld.factor[k][1][i]);
264 }
265
266 /*
267 * Combine terms
268 */
269
270 /* See if this function has been previously applied */
271 for(j = 0; j < i; ++j) {
272 unsigned prev_func = j < 3 ? blend->rt[0].rgb_func : blend->rt[0].alpha_func;
273 unsigned func_reverse = lp_build_blend_func_reverse(func, prev_func);
274
275 if((!func_reverse &&
276 bld.term[0][j] == bld.term[0][i] &&
277 bld.term[1][j] == bld.term[1][i]) ||
278 ((func_commutative || func_reverse) &&
279 bld.term[0][j] == bld.term[1][i] &&
280 bld.term[1][j] == bld.term[0][i]))
281 break;
282 }
283
284 if(j < i)
285 res[i] = res[j];
286 else
287 res[i] = lp_build_blend_func(&bld.base, func, bld.term[0][i], bld.term[1][i]);
288 }
289 else {
290 res[i] = src[i];
291 }
292 }
293 else {
294 res[i] = dst[i];
295 }
296 }
297 }