1 /**************************************************************************
2 *
3 * Copyright 2009-2010 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
37 * of 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 #include "util/u_debug.h"
73
74 #include "gallivm/lp_bld_type.h"
75 #include "gallivm/lp_bld_arit.h"
76 #include "gallivm/lp_bld_init.h"
77 #include "lp_bld_blend.h"
78
79
80 /**
81 * We may use the same values several times, so we keep them here to avoid
82 * recomputing them. Also reusing the values allows us to do simplifications
83 * that LLVM optimization passes wouldn't normally be able to do.
84 */
85 struct lp_build_blend_soa_context
86 {
87 struct lp_build_context base;
88
89 LLVMValueRef src[4];
90 LLVMValueRef dst[4];
91 LLVMValueRef con[4];
92
93 LLVMValueRef inv_src[4];
94 LLVMValueRef inv_dst[4];
95 LLVMValueRef inv_con[4];
96
97 LLVMValueRef src_alpha_saturate;
98
99 /**
100 * We store all factors in a table in order to eliminate redundant
101 * multiplications later.
102 * Indexes are: factor[src,dst][color,term][r,g,b,a]
103 */
104 LLVMValueRef factor[2][2][4];
105
106 /**
107 * Table with all terms.
108 * Indexes are: term[src,dst][r,g,b,a]
109 */
110 LLVMValueRef term[2][4];
111 };
112
113
114 /**
115 * Build a single SOA blend factor for a color channel.
116 * \param i the color channel in [0,3]
117 */
118 static LLVMValueRef
lp_build_blend_soa_factor(struct lp_build_blend_soa_context * bld,unsigned factor,unsigned i)119 lp_build_blend_soa_factor(struct lp_build_blend_soa_context *bld,
120 unsigned factor, unsigned i)
121 {
122 /*
123 * Compute src/first term RGB
124 */
125 switch (factor) {
126 case PIPE_BLENDFACTOR_ONE:
127 return bld->base.one;
128 case PIPE_BLENDFACTOR_SRC_COLOR:
129 return bld->src[i];
130 case PIPE_BLENDFACTOR_SRC_ALPHA:
131 return bld->src[3];
132 case PIPE_BLENDFACTOR_DST_COLOR:
133 return bld->dst[i];
134 case PIPE_BLENDFACTOR_DST_ALPHA:
135 return bld->dst[3];
136 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
137 if(i == 3)
138 return bld->base.one;
139 else {
140 if(!bld->inv_dst[3])
141 bld->inv_dst[3] = lp_build_comp(&bld->base, bld->dst[3]);
142 if(!bld->src_alpha_saturate)
143 bld->src_alpha_saturate = lp_build_min(&bld->base, bld->src[3], bld->inv_dst[3]);
144 return bld->src_alpha_saturate;
145 }
146 case PIPE_BLENDFACTOR_CONST_COLOR:
147 return bld->con[i];
148 case PIPE_BLENDFACTOR_CONST_ALPHA:
149 return bld->con[3];
150 case PIPE_BLENDFACTOR_SRC1_COLOR:
151 /* TODO */
152 assert(0);
153 return bld->base.zero;
154 case PIPE_BLENDFACTOR_SRC1_ALPHA:
155 /* TODO */
156 assert(0);
157 return bld->base.zero;
158 case PIPE_BLENDFACTOR_ZERO:
159 return bld->base.zero;
160 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
161 if(!bld->inv_src[i])
162 bld->inv_src[i] = lp_build_comp(&bld->base, bld->src[i]);
163 return bld->inv_src[i];
164 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
165 if(!bld->inv_src[3])
166 bld->inv_src[3] = lp_build_comp(&bld->base, bld->src[3]);
167 return bld->inv_src[3];
168 case PIPE_BLENDFACTOR_INV_DST_COLOR:
169 if(!bld->inv_dst[i])
170 bld->inv_dst[i] = lp_build_comp(&bld->base, bld->dst[i]);
171 return bld->inv_dst[i];
172 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
173 if(!bld->inv_dst[3])
174 bld->inv_dst[3] = lp_build_comp(&bld->base, bld->dst[3]);
175 return bld->inv_dst[3];
176 case PIPE_BLENDFACTOR_INV_CONST_COLOR:
177 if(!bld->inv_con[i])
178 bld->inv_con[i] = lp_build_comp(&bld->base, bld->con[i]);
179 return bld->inv_con[i];
180 case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
181 if(!bld->inv_con[3])
182 bld->inv_con[3] = lp_build_comp(&bld->base, bld->con[3]);
183 return bld->inv_con[3];
184 case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
185 /* TODO */
186 assert(0);
187 return bld->base.zero;
188 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
189 /* TODO */
190 assert(0);
191 return bld->base.zero;
192 default:
193 assert(0);
194 return bld->base.zero;
195 }
196 }
197
198
199 /**
200 * Generate blend code in SOA mode.
201 * \param rt render target index (to index the blend / colormask state)
202 * \param src src/fragment color
203 * \param dst dst/framebuffer color
204 * \param con constant blend color
205 * \param res the result/output
206 */
207 void
lp_build_blend_soa(struct gallivm_state * gallivm,const struct pipe_blend_state * blend,struct lp_type type,unsigned rt,LLVMValueRef src[4],LLVMValueRef dst[4],LLVMValueRef con[4],LLVMValueRef res[4])208 lp_build_blend_soa(struct gallivm_state *gallivm,
209 const struct pipe_blend_state *blend,
210 struct lp_type type,
211 unsigned rt,
212 LLVMValueRef src[4],
213 LLVMValueRef dst[4],
214 LLVMValueRef con[4],
215 LLVMValueRef res[4])
216 {
217 LLVMBuilderRef builder = gallivm->builder;
218 struct lp_build_blend_soa_context bld;
219 unsigned i, j, k;
220
221 assert(rt < PIPE_MAX_COLOR_BUFS);
222
223 /* Setup build context */
224 memset(&bld, 0, sizeof bld);
225 lp_build_context_init(&bld.base, gallivm, type);
226 for (i = 0; i < 4; ++i) {
227 bld.src[i] = src[i];
228 bld.dst[i] = dst[i];
229 bld.con[i] = con[i];
230 }
231
232 for (i = 0; i < 4; ++i) {
233 /* only compute blending for the color channels enabled for writing */
234 if (blend->rt[rt].colormask & (1 << i)) {
235 if (blend->logicop_enable) {
236 if(!type.floating) {
237 res[i] = lp_build_logicop(builder, blend->logicop_func, src[i], dst[i]);
238 }
239 else
240 res[i] = dst[i];
241 }
242 else if (blend->rt[rt].blend_enable) {
243 unsigned src_factor = i < 3 ? blend->rt[rt].rgb_src_factor : blend->rt[rt].alpha_src_factor;
244 unsigned dst_factor = i < 3 ? blend->rt[rt].rgb_dst_factor : blend->rt[rt].alpha_dst_factor;
245 unsigned func = i < 3 ? blend->rt[rt].rgb_func : blend->rt[rt].alpha_func;
246 boolean func_commutative = lp_build_blend_func_commutative(func);
247
248 /*
249 * Compute src/dst factors.
250 */
251
252 bld.factor[0][0][i] = src[i];
253 bld.factor[0][1][i] = lp_build_blend_soa_factor(&bld, src_factor, i);
254 bld.factor[1][0][i] = dst[i];
255 bld.factor[1][1][i] = lp_build_blend_soa_factor(&bld, dst_factor, i);
256
257 /*
258 * Check if lp_build_blend can perform any optimisations
259 */
260 res[i] = lp_build_blend(&bld.base,
261 func,
262 src_factor,
263 dst_factor,
264 bld.factor[0][0][i],
265 bld.factor[1][0][i],
266 bld.factor[0][1][i],
267 bld.factor[1][1][i],
268 true,
269 true);
270
271 if (res[i]) {
272 continue;
273 }
274
275 /*
276 * Compute src/dst terms
277 */
278
279 for(k = 0; k < 2; ++k) {
280 /* See if this multiplication has been previously computed */
281 for(j = 0; j < i; ++j) {
282 if((bld.factor[k][0][j] == bld.factor[k][0][i] &&
283 bld.factor[k][1][j] == bld.factor[k][1][i]) ||
284 (bld.factor[k][0][j] == bld.factor[k][1][i] &&
285 bld.factor[k][1][j] == bld.factor[k][0][i]))
286 break;
287 }
288
289 if(j < i && bld.term[k][j])
290 bld.term[k][i] = bld.term[k][j];
291 else
292 bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i], bld.factor[k][1][i]);
293
294 if (src_factor == PIPE_BLENDFACTOR_ZERO &&
295 (dst_factor == PIPE_BLENDFACTOR_DST_ALPHA ||
296 dst_factor == PIPE_BLENDFACTOR_INV_DST_ALPHA)) {
297 /* XXX special case these combos to work around an apparent
298 * bug in LLVM.
299 * This hack disables the check for multiplication by zero
300 * in lp_bld_mul(). When we optimize away the
301 * multiplication, something goes wrong during code
302 * generation and we segfault at runtime.
303 */
304 LLVMValueRef zeroSave = bld.base.zero;
305 bld.base.zero = NULL;
306 bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i],
307 bld.factor[k][1][i]);
308 bld.base.zero = zeroSave;
309 }
310 }
311
312 /*
313 * Combine terms
314 */
315
316 /* See if this function has been previously applied */
317 for(j = 0; j < i; ++j) {
318 unsigned prev_func = j < 3 ? blend->rt[rt].rgb_func : blend->rt[rt].alpha_func;
319 unsigned func_reverse = lp_build_blend_func_reverse(func, prev_func);
320
321 if((!func_reverse &&
322 bld.term[0][j] == bld.term[0][i] &&
323 bld.term[1][j] == bld.term[1][i]) ||
324 ((func_commutative || func_reverse) &&
325 bld.term[0][j] == bld.term[1][i] &&
326 bld.term[1][j] == bld.term[0][i]))
327 break;
328 }
329
330 if(j < i)
331 res[i] = res[j];
332 else
333 res[i] = lp_build_blend_func(&bld.base, func, bld.term[0][i], bld.term[1][i]);
334 }
335 else {
336 res[i] = src[i];
337 }
338 }
339 else {
340 res[i] = dst[i];
341 }
342 }
343 }
344