1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
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:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
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.
26 **************************************************************************/
31 * Blend LLVM IR generation -- SoA layout.
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.
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.
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 ...
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 ...
49 * ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
51 * will actually be stored in memory as
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 * ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
57 * NOTE: Run lp_blend_test after any change to this file.
59 * You can also run lp_blend_test to obtain AoS vs SoA benchmarks. Invoking it
62 * lp_blend_test -o blend.tsv
64 * will generate a tab-seperated-file with the test results and performance
67 * @author Jose Fonseca <jfonseca@vmware.com>
71 #include "pipe/p_state.h"
72 #include "util/u_debug.h"
74 #include "gallivm/lp_bld_type.h"
75 #include "gallivm/lp_bld_arit.h"
76 #include "lp_bld_blend.h"
80 * We may use the same values several times, so we keep them here to avoid
81 * recomputing them. Also reusing the values allows us to do simplifications
82 * that LLVM optimization passes wouldn't normally be able to do.
84 struct lp_build_blend_soa_context
86 struct lp_build_context base
;
92 LLVMValueRef inv_src
[4];
93 LLVMValueRef inv_dst
[4];
94 LLVMValueRef inv_con
[4];
96 LLVMValueRef src_alpha_saturate
;
99 * We store all factors in a table in order to eliminate redundant
100 * multiplications later.
101 * Indexes are: factor[src,dst][color,term][r,g,b,a]
103 LLVMValueRef factor
[2][2][4];
106 * Table with all terms.
107 * Indexes are: term[src,dst][r,g,b,a]
109 LLVMValueRef term
[2][4];
114 * Build a single SOA blend factor for a color channel.
115 * \param i the color channel in [0,3]
118 lp_build_blend_soa_factor(struct lp_build_blend_soa_context
*bld
,
119 unsigned factor
, unsigned i
)
122 * Compute src/first term RGB
125 case PIPE_BLENDFACTOR_ONE
:
126 return bld
->base
.one
;
127 case PIPE_BLENDFACTOR_SRC_COLOR
:
129 case PIPE_BLENDFACTOR_SRC_ALPHA
:
131 case PIPE_BLENDFACTOR_DST_COLOR
:
133 case PIPE_BLENDFACTOR_DST_ALPHA
:
135 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE
:
137 return bld
->base
.one
;
140 bld
->inv_dst
[3] = lp_build_comp(&bld
->base
, bld
->dst
[3]);
141 if(!bld
->src_alpha_saturate
)
142 bld
->src_alpha_saturate
= lp_build_min(&bld
->base
, bld
->src
[3], bld
->inv_dst
[3]);
143 return bld
->src_alpha_saturate
;
145 case PIPE_BLENDFACTOR_CONST_COLOR
:
147 case PIPE_BLENDFACTOR_CONST_ALPHA
:
149 case PIPE_BLENDFACTOR_SRC1_COLOR
:
152 return bld
->base
.zero
;
153 case PIPE_BLENDFACTOR_SRC1_ALPHA
:
156 return bld
->base
.zero
;
157 case PIPE_BLENDFACTOR_ZERO
:
158 return bld
->base
.zero
;
159 case PIPE_BLENDFACTOR_INV_SRC_COLOR
:
161 bld
->inv_src
[i
] = lp_build_comp(&bld
->base
, bld
->src
[i
]);
162 return bld
->inv_src
[i
];
163 case PIPE_BLENDFACTOR_INV_SRC_ALPHA
:
165 bld
->inv_src
[3] = lp_build_comp(&bld
->base
, bld
->src
[3]);
166 return bld
->inv_src
[3];
167 case PIPE_BLENDFACTOR_INV_DST_COLOR
:
169 bld
->inv_dst
[i
] = lp_build_comp(&bld
->base
, bld
->dst
[i
]);
170 return bld
->inv_dst
[i
];
171 case PIPE_BLENDFACTOR_INV_DST_ALPHA
:
173 bld
->inv_dst
[3] = lp_build_comp(&bld
->base
, bld
->dst
[3]);
174 return bld
->inv_dst
[3];
175 case PIPE_BLENDFACTOR_INV_CONST_COLOR
:
177 bld
->inv_con
[i
] = lp_build_comp(&bld
->base
, bld
->con
[i
]);
178 return bld
->inv_con
[i
];
179 case PIPE_BLENDFACTOR_INV_CONST_ALPHA
:
181 bld
->inv_con
[3] = lp_build_comp(&bld
->base
, bld
->con
[3]);
182 return bld
->inv_con
[3];
183 case PIPE_BLENDFACTOR_INV_SRC1_COLOR
:
186 return bld
->base
.zero
;
187 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA
:
190 return bld
->base
.zero
;
193 return bld
->base
.zero
;
199 * Generate blend code in SOA mode.
200 * \param rt render target index (to index the blend / colormask state)
201 * \param src src/fragment color
202 * \param dst dst/framebuffer color
203 * \param con constant blend color
204 * \param res the result/output
207 lp_build_blend_soa(LLVMBuilderRef builder
,
208 const struct pipe_blend_state
*blend
,
216 struct lp_build_blend_soa_context bld
;
219 assert(rt
< PIPE_MAX_COLOR_BUFS
);
221 /* Setup build context */
222 memset(&bld
, 0, sizeof bld
);
223 lp_build_context_init(&bld
.base
, builder
, type
);
224 for (i
= 0; i
< 4; ++i
) {
230 for (i
= 0; i
< 4; ++i
) {
231 /* only compute blending for the color channels enabled for writing */
232 if (blend
->rt
[rt
].colormask
& (1 << i
)) {
233 if (blend
->logicop_enable
) {
235 res
[i
] = lp_build_logicop(builder
, blend
->logicop_func
, src
[i
], dst
[i
]);
240 else if (blend
->rt
[rt
].blend_enable
) {
241 unsigned src_factor
= i
< 3 ? blend
->rt
[rt
].rgb_src_factor
: blend
->rt
[rt
].alpha_src_factor
;
242 unsigned dst_factor
= i
< 3 ? blend
->rt
[rt
].rgb_dst_factor
: blend
->rt
[rt
].alpha_dst_factor
;
243 unsigned func
= i
< 3 ? blend
->rt
[rt
].rgb_func
: blend
->rt
[rt
].alpha_func
;
244 boolean func_commutative
= lp_build_blend_func_commutative(func
);
246 /* It makes no sense to blend unless values are normalized */
250 * Compute src/dst factors.
253 bld
.factor
[0][0][i
] = src
[i
];
254 bld
.factor
[0][1][i
] = lp_build_blend_soa_factor(&bld
, src_factor
, i
);
255 bld
.factor
[1][0][i
] = dst
[i
];
256 bld
.factor
[1][1][i
] = lp_build_blend_soa_factor(&bld
, dst_factor
, i
);
259 * Compute src/dst terms
262 for(k
= 0; k
< 2; ++k
) {
263 /* See if this multiplication has been previously computed */
264 for(j
= 0; j
< i
; ++j
) {
265 if((bld
.factor
[k
][0][j
] == bld
.factor
[k
][0][i
] &&
266 bld
.factor
[k
][1][j
] == bld
.factor
[k
][1][i
]) ||
267 (bld
.factor
[k
][0][j
] == bld
.factor
[k
][1][i
] &&
268 bld
.factor
[k
][1][j
] == bld
.factor
[k
][0][i
]))
273 bld
.term
[k
][i
] = bld
.term
[k
][j
];
275 bld
.term
[k
][i
] = lp_build_mul(&bld
.base
, bld
.factor
[k
][0][i
], bld
.factor
[k
][1][i
]);
277 if (src_factor
== PIPE_BLENDFACTOR_ZERO
&&
278 (dst_factor
== PIPE_BLENDFACTOR_DST_ALPHA
||
279 dst_factor
== PIPE_BLENDFACTOR_INV_DST_ALPHA
)) {
280 /* XXX special case these combos to work around an apparent
282 * This hack disables the check for multiplication by zero
283 * in lp_bld_mul(). When we optimize away the
284 * multiplication, something goes wrong during code
285 * generation and we segfault at runtime.
287 LLVMValueRef zeroSave
= bld
.base
.zero
;
288 bld
.base
.zero
= NULL
;
289 bld
.term
[k
][i
] = lp_build_mul(&bld
.base
, bld
.factor
[k
][0][i
],
290 bld
.factor
[k
][1][i
]);
291 bld
.base
.zero
= zeroSave
;
299 /* See if this function has been previously applied */
300 for(j
= 0; j
< i
; ++j
) {
301 unsigned prev_func
= j
< 3 ? blend
->rt
[rt
].rgb_func
: blend
->rt
[rt
].alpha_func
;
302 unsigned func_reverse
= lp_build_blend_func_reverse(func
, prev_func
);
305 bld
.term
[0][j
] == bld
.term
[0][i
] &&
306 bld
.term
[1][j
] == bld
.term
[1][i
]) ||
307 ((func_commutative
|| func_reverse
) &&
308 bld
.term
[0][j
] == bld
.term
[1][i
] &&
309 bld
.term
[1][j
] == bld
.term
[0][i
]))
316 res
[i
] = lp_build_blend_func(&bld
.base
, func
, bld
.term
[0][i
], bld
.term
[1][i
]);