Go to most recent revision | Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
4358 | Serge | 1 | /* |
2 | * Copyright © 2011 Intel Corporation |
||
3 | * |
||
4 | * Permission is hereby granted, free of charge, to any person obtaining a |
||
5 | * copy of this software and associated documentation files (the "Software"), |
||
6 | * to deal in the Software without restriction, including without limitation |
||
7 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
||
8 | * and/or sell copies of the Software, and to permit persons to whom the |
||
9 | * Software is furnished to do so, subject to the following conditions: |
||
10 | * |
||
11 | * The above copyright notice and this permission notice (including the next |
||
12 | * paragraph) shall be included in all copies or substantial portions of the |
||
13 | * Software. |
||
14 | * |
||
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
||
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
||
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
||
18 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
||
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING |
||
20 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
||
21 | * IN THE SOFTWARE. |
||
22 | */ |
||
23 | |||
24 | #include "brw_vec4.h" |
||
25 | #include "brw_cfg.h" |
||
26 | |||
27 | extern "C" { |
||
28 | #include "main/macros.h" |
||
29 | #include "main/shaderobj.h" |
||
30 | #include "program/prog_print.h" |
||
31 | #include "program/prog_parameter.h" |
||
32 | } |
||
33 | |||
34 | #define MAX_INSTRUCTION (1 << 30) |
||
35 | |||
36 | using namespace brw; |
||
37 | |||
38 | namespace brw { |
||
39 | |||
40 | /** |
||
41 | * Common helper for constructing swizzles. When only a subset of |
||
42 | * channels of a vec4 are used, we don't want to reference the other |
||
43 | * channels, as that will tell optimization passes that those other |
||
44 | * channels are used. |
||
45 | */ |
||
46 | unsigned |
||
47 | swizzle_for_size(int size) |
||
48 | { |
||
49 | static const unsigned size_swizzles[4] = { |
||
50 | BRW_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X), |
||
51 | BRW_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y), |
||
52 | BRW_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_Z), |
||
53 | BRW_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W), |
||
54 | }; |
||
55 | |||
56 | assert((size >= 1) && (size <= 4)); |
||
57 | return size_swizzles[size - 1]; |
||
58 | } |
||
59 | |||
60 | void |
||
61 | src_reg::init() |
||
62 | { |
||
63 | memset(this, 0, sizeof(*this)); |
||
64 | |||
65 | this->file = BAD_FILE; |
||
66 | } |
||
67 | |||
68 | src_reg::src_reg(register_file file, int reg, const glsl_type *type) |
||
69 | { |
||
70 | init(); |
||
71 | |||
72 | this->file = file; |
||
73 | this->reg = reg; |
||
74 | if (type && (type->is_scalar() || type->is_vector() || type->is_matrix())) |
||
75 | this->swizzle = swizzle_for_size(type->vector_elements); |
||
76 | else |
||
77 | this->swizzle = SWIZZLE_XYZW; |
||
78 | } |
||
79 | |||
80 | /** Generic unset register constructor. */ |
||
81 | src_reg::src_reg() |
||
82 | { |
||
83 | init(); |
||
84 | } |
||
85 | |||
86 | src_reg::src_reg(float f) |
||
87 | { |
||
88 | init(); |
||
89 | |||
90 | this->file = IMM; |
||
91 | this->type = BRW_REGISTER_TYPE_F; |
||
92 | this->imm.f = f; |
||
93 | } |
||
94 | |||
95 | src_reg::src_reg(uint32_t u) |
||
96 | { |
||
97 | init(); |
||
98 | |||
99 | this->file = IMM; |
||
100 | this->type = BRW_REGISTER_TYPE_UD; |
||
101 | this->imm.u = u; |
||
102 | } |
||
103 | |||
104 | src_reg::src_reg(int32_t i) |
||
105 | { |
||
106 | init(); |
||
107 | |||
108 | this->file = IMM; |
||
109 | this->type = BRW_REGISTER_TYPE_D; |
||
110 | this->imm.i = i; |
||
111 | } |
||
112 | |||
113 | src_reg::src_reg(dst_reg reg) |
||
114 | { |
||
115 | init(); |
||
116 | |||
117 | this->file = reg.file; |
||
118 | this->reg = reg.reg; |
||
119 | this->reg_offset = reg.reg_offset; |
||
120 | this->type = reg.type; |
||
121 | this->reladdr = reg.reladdr; |
||
122 | this->fixed_hw_reg = reg.fixed_hw_reg; |
||
123 | |||
124 | int swizzles[4]; |
||
125 | int next_chan = 0; |
||
126 | int last = 0; |
||
127 | |||
128 | for (int i = 0; i < 4; i++) { |
||
129 | if (!(reg.writemask & (1 << i))) |
||
130 | continue; |
||
131 | |||
132 | swizzles[next_chan++] = last = i; |
||
133 | } |
||
134 | |||
135 | for (; next_chan < 4; next_chan++) { |
||
136 | swizzles[next_chan] = last; |
||
137 | } |
||
138 | |||
139 | this->swizzle = BRW_SWIZZLE4(swizzles[0], swizzles[1], |
||
140 | swizzles[2], swizzles[3]); |
||
141 | } |
||
142 | |||
143 | void |
||
144 | dst_reg::init() |
||
145 | { |
||
146 | memset(this, 0, sizeof(*this)); |
||
147 | this->file = BAD_FILE; |
||
148 | this->writemask = WRITEMASK_XYZW; |
||
149 | } |
||
150 | |||
151 | dst_reg::dst_reg() |
||
152 | { |
||
153 | init(); |
||
154 | } |
||
155 | |||
156 | dst_reg::dst_reg(register_file file, int reg) |
||
157 | { |
||
158 | init(); |
||
159 | |||
160 | this->file = file; |
||
161 | this->reg = reg; |
||
162 | } |
||
163 | |||
164 | dst_reg::dst_reg(register_file file, int reg, const glsl_type *type, |
||
165 | int writemask) |
||
166 | { |
||
167 | init(); |
||
168 | |||
169 | this->file = file; |
||
170 | this->reg = reg; |
||
171 | this->type = brw_type_for_base_type(type); |
||
172 | this->writemask = writemask; |
||
173 | } |
||
174 | |||
175 | dst_reg::dst_reg(struct brw_reg reg) |
||
176 | { |
||
177 | init(); |
||
178 | |||
179 | this->file = HW_REG; |
||
180 | this->fixed_hw_reg = reg; |
||
181 | } |
||
182 | |||
183 | dst_reg::dst_reg(src_reg reg) |
||
184 | { |
||
185 | init(); |
||
186 | |||
187 | this->file = reg.file; |
||
188 | this->reg = reg.reg; |
||
189 | this->reg_offset = reg.reg_offset; |
||
190 | this->type = reg.type; |
||
191 | /* How should we do writemasking when converting from a src_reg? It seems |
||
192 | * pretty obvious that for src.xxxx the caller wants to write to src.x, but |
||
193 | * what about for src.wx? Just special-case src.xxxx for now. |
||
194 | */ |
||
195 | if (reg.swizzle == BRW_SWIZZLE_XXXX) |
||
196 | this->writemask = WRITEMASK_X; |
||
197 | else |
||
198 | this->writemask = WRITEMASK_XYZW; |
||
199 | this->reladdr = reg.reladdr; |
||
200 | this->fixed_hw_reg = reg.fixed_hw_reg; |
||
201 | } |
||
202 | |||
203 | bool |
||
204 | vec4_instruction::is_send_from_grf() |
||
205 | { |
||
206 | switch (opcode) { |
||
207 | case SHADER_OPCODE_SHADER_TIME_ADD: |
||
208 | case VS_OPCODE_PULL_CONSTANT_LOAD_GEN7: |
||
209 | return true; |
||
210 | default: |
||
211 | return false; |
||
212 | } |
||
213 | } |
||
214 | |||
215 | bool |
||
216 | vec4_visitor::can_do_source_mods(vec4_instruction *inst) |
||
217 | { |
||
218 | if (brw->gen == 6 && inst->is_math()) |
||
219 | return false; |
||
220 | |||
221 | if (inst->is_send_from_grf()) |
||
222 | return false; |
||
223 | |||
224 | return true; |
||
225 | } |
||
226 | |||
227 | /** |
||
228 | * Returns how many MRFs an opcode will write over. |
||
229 | * |
||
230 | * Note that this is not the 0 or 1 implied writes in an actual gen |
||
231 | * instruction -- the generate_* functions generate additional MOVs |
||
232 | * for setup. |
||
233 | */ |
||
234 | int |
||
235 | vec4_visitor::implied_mrf_writes(vec4_instruction *inst) |
||
236 | { |
||
237 | if (inst->mlen == 0) |
||
238 | return 0; |
||
239 | |||
240 | switch (inst->opcode) { |
||
241 | case SHADER_OPCODE_RCP: |
||
242 | case SHADER_OPCODE_RSQ: |
||
243 | case SHADER_OPCODE_SQRT: |
||
244 | case SHADER_OPCODE_EXP2: |
||
245 | case SHADER_OPCODE_LOG2: |
||
246 | case SHADER_OPCODE_SIN: |
||
247 | case SHADER_OPCODE_COS: |
||
248 | return 1; |
||
249 | case SHADER_OPCODE_INT_QUOTIENT: |
||
250 | case SHADER_OPCODE_INT_REMAINDER: |
||
251 | case SHADER_OPCODE_POW: |
||
252 | return 2; |
||
253 | case VS_OPCODE_URB_WRITE: |
||
254 | return 1; |
||
255 | case VS_OPCODE_PULL_CONSTANT_LOAD: |
||
256 | return 2; |
||
257 | case VS_OPCODE_SCRATCH_READ: |
||
258 | return 2; |
||
259 | case VS_OPCODE_SCRATCH_WRITE: |
||
260 | return 3; |
||
261 | case SHADER_OPCODE_SHADER_TIME_ADD: |
||
262 | return 0; |
||
263 | case SHADER_OPCODE_TEX: |
||
264 | case SHADER_OPCODE_TXL: |
||
265 | case SHADER_OPCODE_TXD: |
||
266 | case SHADER_OPCODE_TXF: |
||
267 | case SHADER_OPCODE_TXF_MS: |
||
268 | case SHADER_OPCODE_TXS: |
||
269 | return inst->header_present ? 1 : 0; |
||
270 | default: |
||
271 | assert(!"not reached"); |
||
272 | return inst->mlen; |
||
273 | } |
||
274 | } |
||
275 | |||
276 | bool |
||
277 | src_reg::equals(src_reg *r) |
||
278 | { |
||
279 | return (file == r->file && |
||
280 | reg == r->reg && |
||
281 | reg_offset == r->reg_offset && |
||
282 | type == r->type && |
||
283 | negate == r->negate && |
||
284 | abs == r->abs && |
||
285 | swizzle == r->swizzle && |
||
286 | !reladdr && !r->reladdr && |
||
287 | memcmp(&fixed_hw_reg, &r->fixed_hw_reg, |
||
288 | sizeof(fixed_hw_reg)) == 0 && |
||
289 | imm.u == r->imm.u); |
||
290 | } |
||
291 | |||
292 | /** |
||
293 | * Must be called after calculate_live_intervales() to remove unused |
||
294 | * writes to registers -- register allocation will fail otherwise |
||
295 | * because something deffed but not used won't be considered to |
||
296 | * interfere with other regs. |
||
297 | */ |
||
298 | bool |
||
299 | vec4_visitor::dead_code_eliminate() |
||
300 | { |
||
301 | bool progress = false; |
||
302 | int pc = 0; |
||
303 | |||
304 | calculate_live_intervals(); |
||
305 | |||
306 | foreach_list_safe(node, &this->instructions) { |
||
307 | vec4_instruction *inst = (vec4_instruction *)node; |
||
308 | |||
309 | if (inst->dst.file == GRF) { |
||
310 | assert(this->virtual_grf_end[inst->dst.reg] >= pc); |
||
311 | if (this->virtual_grf_end[inst->dst.reg] == pc) { |
||
312 | inst->remove(); |
||
313 | progress = true; |
||
314 | } |
||
315 | } |
||
316 | |||
317 | pc++; |
||
318 | } |
||
319 | |||
320 | if (progress) |
||
321 | live_intervals_valid = false; |
||
322 | |||
323 | return progress; |
||
324 | } |
||
325 | |||
326 | void |
||
327 | vec4_visitor::split_uniform_registers() |
||
328 | { |
||
329 | /* Prior to this, uniforms have been in an array sized according to |
||
330 | * the number of vector uniforms present, sparsely filled (so an |
||
331 | * aggregate results in reg indices being skipped over). Now we're |
||
332 | * going to cut those aggregates up so each .reg index is one |
||
333 | * vector. The goal is to make elimination of unused uniform |
||
334 | * components easier later. |
||
335 | */ |
||
336 | foreach_list(node, &this->instructions) { |
||
337 | vec4_instruction *inst = (vec4_instruction *)node; |
||
338 | |||
339 | for (int i = 0 ; i < 3; i++) { |
||
340 | if (inst->src[i].file != UNIFORM) |
||
341 | continue; |
||
342 | |||
343 | assert(!inst->src[i].reladdr); |
||
344 | |||
345 | inst->src[i].reg += inst->src[i].reg_offset; |
||
346 | inst->src[i].reg_offset = 0; |
||
347 | } |
||
348 | } |
||
349 | |||
350 | /* Update that everything is now vector-sized. */ |
||
351 | for (int i = 0; i < this->uniforms; i++) { |
||
352 | this->uniform_size[i] = 1; |
||
353 | } |
||
354 | } |
||
355 | |||
356 | void |
||
357 | vec4_visitor::pack_uniform_registers() |
||
358 | { |
||
359 | bool uniform_used[this->uniforms]; |
||
360 | int new_loc[this->uniforms]; |
||
361 | int new_chan[this->uniforms]; |
||
362 | |||
363 | memset(uniform_used, 0, sizeof(uniform_used)); |
||
364 | memset(new_loc, 0, sizeof(new_loc)); |
||
365 | memset(new_chan, 0, sizeof(new_chan)); |
||
366 | |||
367 | /* Find which uniform vectors are actually used by the program. We |
||
368 | * expect unused vector elements when we've moved array access out |
||
369 | * to pull constants, and from some GLSL code generators like wine. |
||
370 | */ |
||
371 | foreach_list(node, &this->instructions) { |
||
372 | vec4_instruction *inst = (vec4_instruction *)node; |
||
373 | |||
374 | for (int i = 0 ; i < 3; i++) { |
||
375 | if (inst->src[i].file != UNIFORM) |
||
376 | continue; |
||
377 | |||
378 | uniform_used[inst->src[i].reg] = true; |
||
379 | } |
||
380 | } |
||
381 | |||
382 | int new_uniform_count = 0; |
||
383 | |||
384 | /* Now, figure out a packing of the live uniform vectors into our |
||
385 | * push constants. |
||
386 | */ |
||
387 | for (int src = 0; src < uniforms; src++) { |
||
388 | int size = this->uniform_vector_size[src]; |
||
389 | |||
390 | if (!uniform_used[src]) { |
||
391 | this->uniform_vector_size[src] = 0; |
||
392 | continue; |
||
393 | } |
||
394 | |||
395 | int dst; |
||
396 | /* Find the lowest place we can slot this uniform in. */ |
||
397 | for (dst = 0; dst < src; dst++) { |
||
398 | if (this->uniform_vector_size[dst] + size <= 4) |
||
399 | break; |
||
400 | } |
||
401 | |||
402 | if (src == dst) { |
||
403 | new_loc[src] = dst; |
||
404 | new_chan[src] = 0; |
||
405 | } else { |
||
406 | new_loc[src] = dst; |
||
407 | new_chan[src] = this->uniform_vector_size[dst]; |
||
408 | |||
409 | /* Move the references to the data */ |
||
410 | for (int j = 0; j < size; j++) { |
||
411 | prog_data->param[dst * 4 + new_chan[src] + j] = |
||
412 | prog_data->param[src * 4 + j]; |
||
413 | } |
||
414 | |||
415 | this->uniform_vector_size[dst] += size; |
||
416 | this->uniform_vector_size[src] = 0; |
||
417 | } |
||
418 | |||
419 | new_uniform_count = MAX2(new_uniform_count, dst + 1); |
||
420 | } |
||
421 | |||
422 | this->uniforms = new_uniform_count; |
||
423 | |||
424 | /* Now, update the instructions for our repacked uniforms. */ |
||
425 | foreach_list(node, &this->instructions) { |
||
426 | vec4_instruction *inst = (vec4_instruction *)node; |
||
427 | |||
428 | for (int i = 0 ; i < 3; i++) { |
||
429 | int src = inst->src[i].reg; |
||
430 | |||
431 | if (inst->src[i].file != UNIFORM) |
||
432 | continue; |
||
433 | |||
434 | inst->src[i].reg = new_loc[src]; |
||
435 | |||
436 | int sx = BRW_GET_SWZ(inst->src[i].swizzle, 0) + new_chan[src]; |
||
437 | int sy = BRW_GET_SWZ(inst->src[i].swizzle, 1) + new_chan[src]; |
||
438 | int sz = BRW_GET_SWZ(inst->src[i].swizzle, 2) + new_chan[src]; |
||
439 | int sw = BRW_GET_SWZ(inst->src[i].swizzle, 3) + new_chan[src]; |
||
440 | inst->src[i].swizzle = BRW_SWIZZLE4(sx, sy, sz, sw); |
||
441 | } |
||
442 | } |
||
443 | } |
||
444 | |||
445 | bool |
||
446 | src_reg::is_zero() const |
||
447 | { |
||
448 | if (file != IMM) |
||
449 | return false; |
||
450 | |||
451 | if (type == BRW_REGISTER_TYPE_F) { |
||
452 | return imm.f == 0.0; |
||
453 | } else { |
||
454 | return imm.i == 0; |
||
455 | } |
||
456 | } |
||
457 | |||
458 | bool |
||
459 | src_reg::is_one() const |
||
460 | { |
||
461 | if (file != IMM) |
||
462 | return false; |
||
463 | |||
464 | if (type == BRW_REGISTER_TYPE_F) { |
||
465 | return imm.f == 1.0; |
||
466 | } else { |
||
467 | return imm.i == 1; |
||
468 | } |
||
469 | } |
||
470 | |||
471 | /** |
||
472 | * Does algebraic optimizations (0 * a = 0, 1 * a = a, a + 0 = a). |
||
473 | * |
||
474 | * While GLSL IR also performs this optimization, we end up with it in |
||
475 | * our instruction stream for a couple of reasons. One is that we |
||
476 | * sometimes generate silly instructions, for example in array access |
||
477 | * where we'll generate "ADD offset, index, base" even if base is 0. |
||
478 | * The other is that GLSL IR's constant propagation doesn't track the |
||
479 | * components of aggregates, so some VS patterns (initialize matrix to |
||
480 | * 0, accumulate in vertex blending factors) end up breaking down to |
||
481 | * instructions involving 0. |
||
482 | */ |
||
483 | bool |
||
484 | vec4_visitor::opt_algebraic() |
||
485 | { |
||
486 | bool progress = false; |
||
487 | |||
488 | foreach_list(node, &this->instructions) { |
||
489 | vec4_instruction *inst = (vec4_instruction *)node; |
||
490 | |||
491 | switch (inst->opcode) { |
||
492 | case BRW_OPCODE_ADD: |
||
493 | if (inst->src[1].is_zero()) { |
||
494 | inst->opcode = BRW_OPCODE_MOV; |
||
495 | inst->src[1] = src_reg(); |
||
496 | progress = true; |
||
497 | } |
||
498 | break; |
||
499 | |||
500 | case BRW_OPCODE_MUL: |
||
501 | if (inst->src[1].is_zero()) { |
||
502 | inst->opcode = BRW_OPCODE_MOV; |
||
503 | switch (inst->src[0].type) { |
||
504 | case BRW_REGISTER_TYPE_F: |
||
505 | inst->src[0] = src_reg(0.0f); |
||
506 | break; |
||
507 | case BRW_REGISTER_TYPE_D: |
||
508 | inst->src[0] = src_reg(0); |
||
509 | break; |
||
510 | case BRW_REGISTER_TYPE_UD: |
||
511 | inst->src[0] = src_reg(0u); |
||
512 | break; |
||
513 | default: |
||
514 | assert(!"not reached"); |
||
515 | inst->src[0] = src_reg(0.0f); |
||
516 | break; |
||
517 | } |
||
518 | inst->src[1] = src_reg(); |
||
519 | progress = true; |
||
520 | } else if (inst->src[1].is_one()) { |
||
521 | inst->opcode = BRW_OPCODE_MOV; |
||
522 | inst->src[1] = src_reg(); |
||
523 | progress = true; |
||
524 | } |
||
525 | break; |
||
526 | default: |
||
527 | break; |
||
528 | } |
||
529 | } |
||
530 | |||
531 | if (progress) |
||
532 | this->live_intervals_valid = false; |
||
533 | |||
534 | return progress; |
||
535 | } |
||
536 | |||
537 | /** |
||
538 | * Only a limited number of hardware registers may be used for push |
||
539 | * constants, so this turns access to the overflowed constants into |
||
540 | * pull constants. |
||
541 | */ |
||
542 | void |
||
543 | vec4_visitor::move_push_constants_to_pull_constants() |
||
544 | { |
||
545 | int pull_constant_loc[this->uniforms]; |
||
546 | |||
547 | /* Only allow 32 registers (256 uniform components) as push constants, |
||
548 | * which is the limit on gen6. |
||
549 | */ |
||
550 | int max_uniform_components = 32 * 8; |
||
551 | if (this->uniforms * 4 <= max_uniform_components) |
||
552 | return; |
||
553 | |||
554 | /* Make some sort of choice as to which uniforms get sent to pull |
||
555 | * constants. We could potentially do something clever here like |
||
556 | * look for the most infrequently used uniform vec4s, but leave |
||
557 | * that for later. |
||
558 | */ |
||
559 | for (int i = 0; i < this->uniforms * 4; i += 4) { |
||
560 | pull_constant_loc[i / 4] = -1; |
||
561 | |||
562 | if (i >= max_uniform_components) { |
||
563 | const float **values = &prog_data->param[i]; |
||
564 | |||
565 | /* Try to find an existing copy of this uniform in the pull |
||
566 | * constants if it was part of an array access already. |
||
567 | */ |
||
568 | for (unsigned int j = 0; j < prog_data->nr_pull_params; j += 4) { |
||
569 | int matches; |
||
570 | |||
571 | for (matches = 0; matches < 4; matches++) { |
||
572 | if (prog_data->pull_param[j + matches] != values[matches]) |
||
573 | break; |
||
574 | } |
||
575 | |||
576 | if (matches == 4) { |
||
577 | pull_constant_loc[i / 4] = j / 4; |
||
578 | break; |
||
579 | } |
||
580 | } |
||
581 | |||
582 | if (pull_constant_loc[i / 4] == -1) { |
||
583 | assert(prog_data->nr_pull_params % 4 == 0); |
||
584 | pull_constant_loc[i / 4] = prog_data->nr_pull_params / 4; |
||
585 | |||
586 | for (int j = 0; j < 4; j++) { |
||
587 | prog_data->pull_param[prog_data->nr_pull_params++] = values[j]; |
||
588 | } |
||
589 | } |
||
590 | } |
||
591 | } |
||
592 | |||
593 | /* Now actually rewrite usage of the things we've moved to pull |
||
594 | * constants. |
||
595 | */ |
||
596 | foreach_list_safe(node, &this->instructions) { |
||
597 | vec4_instruction *inst = (vec4_instruction *)node; |
||
598 | |||
599 | for (int i = 0 ; i < 3; i++) { |
||
600 | if (inst->src[i].file != UNIFORM || |
||
601 | pull_constant_loc[inst->src[i].reg] == -1) |
||
602 | continue; |
||
603 | |||
604 | int uniform = inst->src[i].reg; |
||
605 | |||
606 | dst_reg temp = dst_reg(this, glsl_type::vec4_type); |
||
607 | |||
608 | emit_pull_constant_load(inst, temp, inst->src[i], |
||
609 | pull_constant_loc[uniform]); |
||
610 | |||
611 | inst->src[i].file = temp.file; |
||
612 | inst->src[i].reg = temp.reg; |
||
613 | inst->src[i].reg_offset = temp.reg_offset; |
||
614 | inst->src[i].reladdr = NULL; |
||
615 | } |
||
616 | } |
||
617 | |||
618 | /* Repack push constants to remove the now-unused ones. */ |
||
619 | pack_uniform_registers(); |
||
620 | } |
||
621 | |||
622 | /** |
||
623 | * Sets the dependency control fields on instructions after register |
||
624 | * allocation and before the generator is run. |
||
625 | * |
||
626 | * When you have a sequence of instructions like: |
||
627 | * |
||
628 | * DP4 temp.x vertex uniform[0] |
||
629 | * DP4 temp.y vertex uniform[0] |
||
630 | * DP4 temp.z vertex uniform[0] |
||
631 | * DP4 temp.w vertex uniform[0] |
||
632 | * |
||
633 | * The hardware doesn't know that it can actually run the later instructions |
||
634 | * while the previous ones are in flight, producing stalls. However, we have |
||
635 | * manual fields we can set in the instructions that let it do so. |
||
636 | */ |
||
637 | void |
||
638 | vec4_visitor::opt_set_dependency_control() |
||
639 | { |
||
640 | vec4_instruction *last_grf_write[BRW_MAX_GRF]; |
||
641 | uint8_t grf_channels_written[BRW_MAX_GRF]; |
||
642 | vec4_instruction *last_mrf_write[BRW_MAX_GRF]; |
||
643 | uint8_t mrf_channels_written[BRW_MAX_GRF]; |
||
644 | |||
645 | cfg_t cfg(this); |
||
646 | |||
647 | assert(prog_data->total_grf || |
||
648 | !"Must be called after register allocation"); |
||
649 | |||
650 | for (int i = 0; i < cfg.num_blocks; i++) { |
||
651 | bblock_t *bblock = cfg.blocks[i]; |
||
652 | vec4_instruction *inst; |
||
653 | |||
654 | memset(last_grf_write, 0, sizeof(last_grf_write)); |
||
655 | memset(last_mrf_write, 0, sizeof(last_mrf_write)); |
||
656 | |||
657 | for (inst = (vec4_instruction *)bblock->start; |
||
658 | inst != (vec4_instruction *)bblock->end->next; |
||
659 | inst = (vec4_instruction *)inst->next) { |
||
660 | /* If we read from a register that we were doing dependency control |
||
661 | * on, don't do dependency control across the read. |
||
662 | */ |
||
663 | for (int i = 0; i < 3; i++) { |
||
664 | int reg = inst->src[i].reg + inst->src[i].reg_offset; |
||
665 | if (inst->src[i].file == GRF) { |
||
666 | last_grf_write[reg] = NULL; |
||
667 | } else if (inst->src[i].file == HW_REG) { |
||
668 | memset(last_grf_write, 0, sizeof(last_grf_write)); |
||
669 | break; |
||
670 | } |
||
671 | assert(inst->src[i].file != MRF); |
||
672 | } |
||
673 | |||
674 | /* In the presence of send messages, totally interrupt dependency |
||
675 | * control. They're long enough that the chance of dependency |
||
676 | * control around them just doesn't matter. |
||
677 | */ |
||
678 | if (inst->mlen) { |
||
679 | memset(last_grf_write, 0, sizeof(last_grf_write)); |
||
680 | memset(last_mrf_write, 0, sizeof(last_mrf_write)); |
||
681 | continue; |
||
682 | } |
||
683 | |||
684 | /* It looks like setting dependency control on a predicated |
||
685 | * instruction hangs the GPU. |
||
686 | */ |
||
687 | if (inst->predicate) { |
||
688 | memset(last_grf_write, 0, sizeof(last_grf_write)); |
||
689 | memset(last_mrf_write, 0, sizeof(last_mrf_write)); |
||
690 | continue; |
||
691 | } |
||
692 | |||
693 | /* Now, see if we can do dependency control for this instruction |
||
694 | * against a previous one writing to its destination. |
||
695 | */ |
||
696 | int reg = inst->dst.reg + inst->dst.reg_offset; |
||
697 | if (inst->dst.file == GRF) { |
||
698 | if (last_grf_write[reg] && |
||
699 | !(inst->dst.writemask & grf_channels_written[reg])) { |
||
700 | last_grf_write[reg]->no_dd_clear = true; |
||
701 | inst->no_dd_check = true; |
||
702 | } else { |
||
703 | grf_channels_written[reg] = 0; |
||
704 | } |
||
705 | |||
706 | last_grf_write[reg] = inst; |
||
707 | grf_channels_written[reg] |= inst->dst.writemask; |
||
708 | } else if (inst->dst.file == MRF) { |
||
709 | if (last_mrf_write[reg] && |
||
710 | !(inst->dst.writemask & mrf_channels_written[reg])) { |
||
711 | last_mrf_write[reg]->no_dd_clear = true; |
||
712 | inst->no_dd_check = true; |
||
713 | } else { |
||
714 | mrf_channels_written[reg] = 0; |
||
715 | } |
||
716 | |||
717 | last_mrf_write[reg] = inst; |
||
718 | mrf_channels_written[reg] |= inst->dst.writemask; |
||
719 | } else if (inst->dst.reg == HW_REG) { |
||
720 | if (inst->dst.fixed_hw_reg.file == BRW_GENERAL_REGISTER_FILE) |
||
721 | memset(last_grf_write, 0, sizeof(last_grf_write)); |
||
722 | if (inst->dst.fixed_hw_reg.file == BRW_MESSAGE_REGISTER_FILE) |
||
723 | memset(last_mrf_write, 0, sizeof(last_mrf_write)); |
||
724 | } |
||
725 | } |
||
726 | } |
||
727 | } |
||
728 | |||
729 | bool |
||
730 | vec4_instruction::can_reswizzle_dst(int dst_writemask, |
||
731 | int swizzle, |
||
732 | int swizzle_mask) |
||
733 | { |
||
734 | /* If this instruction sets anything not referenced by swizzle, then we'd |
||
735 | * totally break it when we reswizzle. |
||
736 | */ |
||
737 | if (dst.writemask & ~swizzle_mask) |
||
738 | return false; |
||
739 | |||
740 | switch (opcode) { |
||
741 | case BRW_OPCODE_DP4: |
||
742 | case BRW_OPCODE_DP3: |
||
743 | case BRW_OPCODE_DP2: |
||
744 | return true; |
||
745 | default: |
||
746 | /* Check if there happens to be no reswizzling required. */ |
||
747 | for (int c = 0; c < 4; c++) { |
||
748 | int bit = 1 << BRW_GET_SWZ(swizzle, c); |
||
749 | /* Skip components of the swizzle not used by the dst. */ |
||
750 | if (!(dst_writemask & (1 << c))) |
||
751 | continue; |
||
752 | |||
753 | /* We don't do the reswizzling yet, so just sanity check that we |
||
754 | * don't have to. |
||
755 | */ |
||
756 | if (bit != (1 << c)) |
||
757 | return false; |
||
758 | } |
||
759 | return true; |
||
760 | } |
||
761 | } |
||
762 | |||
763 | /** |
||
764 | * For any channels in the swizzle's source that were populated by this |
||
765 | * instruction, rewrite the instruction to put the appropriate result directly |
||
766 | * in those channels. |
||
767 | * |
||
768 | * e.g. for swizzle=yywx, MUL a.xy b c -> MUL a.yy_x b.yy z.yy_x |
||
769 | */ |
||
770 | void |
||
771 | vec4_instruction::reswizzle_dst(int dst_writemask, int swizzle) |
||
772 | { |
||
773 | int new_writemask = 0; |
||
774 | |||
775 | switch (opcode) { |
||
776 | case BRW_OPCODE_DP4: |
||
777 | case BRW_OPCODE_DP3: |
||
778 | case BRW_OPCODE_DP2: |
||
779 | for (int c = 0; c < 4; c++) { |
||
780 | int bit = 1 << BRW_GET_SWZ(swizzle, c); |
||
781 | /* Skip components of the swizzle not used by the dst. */ |
||
782 | if (!(dst_writemask & (1 << c))) |
||
783 | continue; |
||
784 | /* If we were populating this component, then populate the |
||
785 | * corresponding channel of the new dst. |
||
786 | */ |
||
787 | if (dst.writemask & bit) |
||
788 | new_writemask |= (1 << c); |
||
789 | } |
||
790 | dst.writemask = new_writemask; |
||
791 | break; |
||
792 | default: |
||
793 | for (int c = 0; c < 4; c++) { |
||
794 | /* Skip components of the swizzle not used by the dst. */ |
||
795 | if (!(dst_writemask & (1 << c))) |
||
796 | continue; |
||
797 | |||
798 | /* We don't do the reswizzling yet, so just sanity check that we |
||
799 | * don't have to. |
||
800 | */ |
||
801 | assert((1 << BRW_GET_SWZ(swizzle, c)) == (1 << c)); |
||
802 | } |
||
803 | break; |
||
804 | } |
||
805 | } |
||
806 | |||
807 | /* |
||
808 | * Tries to reduce extra MOV instructions by taking temporary GRFs that get |
||
809 | * just written and then MOVed into another reg and making the original write |
||
810 | * of the GRF write directly to the final destination instead. |
||
811 | */ |
||
812 | bool |
||
813 | vec4_visitor::opt_register_coalesce() |
||
814 | { |
||
815 | bool progress = false; |
||
816 | int next_ip = 0; |
||
817 | |||
818 | calculate_live_intervals(); |
||
819 | |||
820 | foreach_list_safe(node, &this->instructions) { |
||
821 | vec4_instruction *inst = (vec4_instruction *)node; |
||
822 | |||
823 | int ip = next_ip; |
||
824 | next_ip++; |
||
825 | |||
826 | if (inst->opcode != BRW_OPCODE_MOV || |
||
827 | (inst->dst.file != GRF && inst->dst.file != MRF) || |
||
828 | inst->predicate || |
||
829 | inst->src[0].file != GRF || |
||
830 | inst->dst.type != inst->src[0].type || |
||
831 | inst->src[0].abs || inst->src[0].negate || inst->src[0].reladdr) |
||
832 | continue; |
||
833 | |||
834 | bool to_mrf = (inst->dst.file == MRF); |
||
835 | |||
836 | /* Can't coalesce this GRF if someone else was going to |
||
837 | * read it later. |
||
838 | */ |
||
839 | if (this->virtual_grf_end[inst->src[0].reg] > ip) |
||
840 | continue; |
||
841 | |||
842 | /* We need to check interference with the final destination between this |
||
843 | * instruction and the earliest instruction involved in writing the GRF |
||
844 | * we're eliminating. To do that, keep track of which of our source |
||
845 | * channels we've seen initialized. |
||
846 | */ |
||
847 | bool chans_needed[4] = {false, false, false, false}; |
||
848 | int chans_remaining = 0; |
||
849 | int swizzle_mask = 0; |
||
850 | for (int i = 0; i < 4; i++) { |
||
851 | int chan = BRW_GET_SWZ(inst->src[0].swizzle, i); |
||
852 | |||
853 | if (!(inst->dst.writemask & (1 << i))) |
||
854 | continue; |
||
855 | |||
856 | swizzle_mask |= (1 << chan); |
||
857 | |||
858 | if (!chans_needed[chan]) { |
||
859 | chans_needed[chan] = true; |
||
860 | chans_remaining++; |
||
861 | } |
||
862 | } |
||
863 | |||
864 | /* Now walk up the instruction stream trying to see if we can rewrite |
||
865 | * everything writing to the temporary to write into the destination |
||
866 | * instead. |
||
867 | */ |
||
868 | vec4_instruction *scan_inst; |
||
869 | for (scan_inst = (vec4_instruction *)inst->prev; |
||
870 | scan_inst->prev != NULL; |
||
871 | scan_inst = (vec4_instruction *)scan_inst->prev) { |
||
872 | if (scan_inst->dst.file == GRF && |
||
873 | scan_inst->dst.reg == inst->src[0].reg && |
||
874 | scan_inst->dst.reg_offset == inst->src[0].reg_offset) { |
||
875 | /* Found something writing to the reg we want to coalesce away. */ |
||
876 | if (to_mrf) { |
||
877 | /* SEND instructions can't have MRF as a destination. */ |
||
878 | if (scan_inst->mlen) |
||
879 | break; |
||
880 | |||
881 | if (brw->gen == 6) { |
||
882 | /* gen6 math instructions must have the destination be |
||
883 | * GRF, so no compute-to-MRF for them. |
||
884 | */ |
||
885 | if (scan_inst->is_math()) { |
||
886 | break; |
||
887 | } |
||
888 | } |
||
889 | } |
||
890 | |||
891 | /* If we can't handle the swizzle, bail. */ |
||
892 | if (!scan_inst->can_reswizzle_dst(inst->dst.writemask, |
||
893 | inst->src[0].swizzle, |
||
894 | swizzle_mask)) { |
||
895 | break; |
||
896 | } |
||
897 | |||
898 | /* Mark which channels we found unconditional writes for. */ |
||
899 | if (!scan_inst->predicate) { |
||
900 | for (int i = 0; i < 4; i++) { |
||
901 | if (scan_inst->dst.writemask & (1 << i) && |
||
902 | chans_needed[i]) { |
||
903 | chans_needed[i] = false; |
||
904 | chans_remaining--; |
||
905 | } |
||
906 | } |
||
907 | } |
||
908 | |||
909 | if (chans_remaining == 0) |
||
910 | break; |
||
911 | } |
||
912 | |||
913 | /* We don't handle flow control here. Most computation of values |
||
914 | * that could be coalesced happens just before their use. |
||
915 | */ |
||
916 | if (scan_inst->opcode == BRW_OPCODE_DO || |
||
917 | scan_inst->opcode == BRW_OPCODE_WHILE || |
||
918 | scan_inst->opcode == BRW_OPCODE_ELSE || |
||
919 | scan_inst->opcode == BRW_OPCODE_ENDIF) { |
||
920 | break; |
||
921 | } |
||
922 | |||
923 | /* You can't read from an MRF, so if someone else reads our MRF's |
||
924 | * source GRF that we wanted to rewrite, that stops us. If it's a |
||
925 | * GRF we're trying to coalesce to, we don't actually handle |
||
926 | * rewriting sources so bail in that case as well. |
||
927 | */ |
||
928 | bool interfered = false; |
||
929 | for (int i = 0; i < 3; i++) { |
||
930 | if (scan_inst->src[i].file == GRF && |
||
931 | scan_inst->src[i].reg == inst->src[0].reg && |
||
932 | scan_inst->src[i].reg_offset == inst->src[0].reg_offset) { |
||
933 | interfered = true; |
||
934 | } |
||
935 | } |
||
936 | if (interfered) |
||
937 | break; |
||
938 | |||
939 | /* If somebody else writes our destination here, we can't coalesce |
||
940 | * before that. |
||
941 | */ |
||
942 | if (scan_inst->dst.file == inst->dst.file && |
||
943 | scan_inst->dst.reg == inst->dst.reg) { |
||
944 | break; |
||
945 | } |
||
946 | |||
947 | /* Check for reads of the register we're trying to coalesce into. We |
||
948 | * can't go rewriting instructions above that to put some other value |
||
949 | * in the register instead. |
||
950 | */ |
||
951 | if (to_mrf && scan_inst->mlen > 0) { |
||
952 | if (inst->dst.reg >= scan_inst->base_mrf && |
||
953 | inst->dst.reg < scan_inst->base_mrf + scan_inst->mlen) { |
||
954 | break; |
||
955 | } |
||
956 | } else { |
||
957 | for (int i = 0; i < 3; i++) { |
||
958 | if (scan_inst->src[i].file == inst->dst.file && |
||
959 | scan_inst->src[i].reg == inst->dst.reg && |
||
960 | scan_inst->src[i].reg_offset == inst->src[0].reg_offset) { |
||
961 | interfered = true; |
||
962 | } |
||
963 | } |
||
964 | if (interfered) |
||
965 | break; |
||
966 | } |
||
967 | } |
||
968 | |||
969 | if (chans_remaining == 0) { |
||
970 | /* If we've made it here, we have an MOV we want to coalesce out, and |
||
971 | * a scan_inst pointing to the earliest instruction involved in |
||
972 | * computing the value. Now go rewrite the instruction stream |
||
973 | * between the two. |
||
974 | */ |
||
975 | |||
976 | while (scan_inst != inst) { |
||
977 | if (scan_inst->dst.file == GRF && |
||
978 | scan_inst->dst.reg == inst->src[0].reg && |
||
979 | scan_inst->dst.reg_offset == inst->src[0].reg_offset) { |
||
980 | scan_inst->reswizzle_dst(inst->dst.writemask, |
||
981 | inst->src[0].swizzle); |
||
982 | scan_inst->dst.file = inst->dst.file; |
||
983 | scan_inst->dst.reg = inst->dst.reg; |
||
984 | scan_inst->dst.reg_offset = inst->dst.reg_offset; |
||
985 | scan_inst->saturate |= inst->saturate; |
||
986 | } |
||
987 | scan_inst = (vec4_instruction *)scan_inst->next; |
||
988 | } |
||
989 | inst->remove(); |
||
990 | progress = true; |
||
991 | } |
||
992 | } |
||
993 | |||
994 | if (progress) |
||
995 | live_intervals_valid = false; |
||
996 | |||
997 | return progress; |
||
998 | } |
||
999 | |||
1000 | /** |
||
1001 | * Splits virtual GRFs requesting more than one contiguous physical register. |
||
1002 | * |
||
1003 | * We initially create large virtual GRFs for temporary structures, arrays, |
||
1004 | * and matrices, so that the dereference visitor functions can add reg_offsets |
||
1005 | * to work their way down to the actual member being accessed. But when it |
||
1006 | * comes to optimization, we'd like to treat each register as individual |
||
1007 | * storage if possible. |
||
1008 | * |
||
1009 | * So far, the only thing that might prevent splitting is a send message from |
||
1010 | * a GRF on IVB. |
||
1011 | */ |
||
1012 | void |
||
1013 | vec4_visitor::split_virtual_grfs() |
||
1014 | { |
||
1015 | int num_vars = this->virtual_grf_count; |
||
1016 | int new_virtual_grf[num_vars]; |
||
1017 | bool split_grf[num_vars]; |
||
1018 | |||
1019 | memset(new_virtual_grf, 0, sizeof(new_virtual_grf)); |
||
1020 | |||
1021 | /* Try to split anything > 0 sized. */ |
||
1022 | for (int i = 0; i < num_vars; i++) { |
||
1023 | split_grf[i] = this->virtual_grf_sizes[i] != 1; |
||
1024 | } |
||
1025 | |||
1026 | /* Check that the instructions are compatible with the registers we're trying |
||
1027 | * to split. |
||
1028 | */ |
||
1029 | foreach_list(node, &this->instructions) { |
||
1030 | vec4_instruction *inst = (vec4_instruction *)node; |
||
1031 | |||
1032 | /* If there's a SEND message loading from a GRF on gen7+, it needs to be |
||
1033 | * contiguous. |
||
1034 | */ |
||
1035 | if (inst->is_send_from_grf()) { |
||
1036 | for (int i = 0; i < 3; i++) { |
||
1037 | if (inst->src[i].file == GRF) { |
||
1038 | split_grf[inst->src[i].reg] = false; |
||
1039 | } |
||
1040 | } |
||
1041 | } |
||
1042 | } |
||
1043 | |||
1044 | /* Allocate new space for split regs. Note that the virtual |
||
1045 | * numbers will be contiguous. |
||
1046 | */ |
||
1047 | for (int i = 0; i < num_vars; i++) { |
||
1048 | if (!split_grf[i]) |
||
1049 | continue; |
||
1050 | |||
1051 | new_virtual_grf[i] = virtual_grf_alloc(1); |
||
1052 | for (int j = 2; j < this->virtual_grf_sizes[i]; j++) { |
||
1053 | int reg = virtual_grf_alloc(1); |
||
1054 | assert(reg == new_virtual_grf[i] + j - 1); |
||
1055 | (void) reg; |
||
1056 | } |
||
1057 | this->virtual_grf_sizes[i] = 1; |
||
1058 | } |
||
1059 | |||
1060 | foreach_list(node, &this->instructions) { |
||
1061 | vec4_instruction *inst = (vec4_instruction *)node; |
||
1062 | |||
1063 | if (inst->dst.file == GRF && split_grf[inst->dst.reg] && |
||
1064 | inst->dst.reg_offset != 0) { |
||
1065 | inst->dst.reg = (new_virtual_grf[inst->dst.reg] + |
||
1066 | inst->dst.reg_offset - 1); |
||
1067 | inst->dst.reg_offset = 0; |
||
1068 | } |
||
1069 | for (int i = 0; i < 3; i++) { |
||
1070 | if (inst->src[i].file == GRF && split_grf[inst->src[i].reg] && |
||
1071 | inst->src[i].reg_offset != 0) { |
||
1072 | inst->src[i].reg = (new_virtual_grf[inst->src[i].reg] + |
||
1073 | inst->src[i].reg_offset - 1); |
||
1074 | inst->src[i].reg_offset = 0; |
||
1075 | } |
||
1076 | } |
||
1077 | } |
||
1078 | this->live_intervals_valid = false; |
||
1079 | } |
||
1080 | |||
1081 | void |
||
1082 | vec4_visitor::dump_instruction(backend_instruction *be_inst) |
||
1083 | { |
||
1084 | vec4_instruction *inst = (vec4_instruction *)be_inst; |
||
1085 | |||
1086 | printf("%s ", brw_instruction_name(inst->opcode)); |
||
1087 | |||
1088 | switch (inst->dst.file) { |
||
1089 | case GRF: |
||
1090 | printf("vgrf%d.%d", inst->dst.reg, inst->dst.reg_offset); |
||
1091 | break; |
||
1092 | case MRF: |
||
1093 | printf("m%d", inst->dst.reg); |
||
1094 | break; |
||
1095 | case BAD_FILE: |
||
1096 | printf("(null)"); |
||
1097 | break; |
||
1098 | default: |
||
1099 | printf("???"); |
||
1100 | break; |
||
1101 | } |
||
1102 | if (inst->dst.writemask != WRITEMASK_XYZW) { |
||
1103 | printf("."); |
||
1104 | if (inst->dst.writemask & 1) |
||
1105 | printf("x"); |
||
1106 | if (inst->dst.writemask & 2) |
||
1107 | printf("y"); |
||
1108 | if (inst->dst.writemask & 4) |
||
1109 | printf("z"); |
||
1110 | if (inst->dst.writemask & 8) |
||
1111 | printf("w"); |
||
1112 | } |
||
1113 | printf(", "); |
||
1114 | |||
1115 | for (int i = 0; i < 3; i++) { |
||
1116 | switch (inst->src[i].file) { |
||
1117 | case GRF: |
||
1118 | printf("vgrf%d", inst->src[i].reg); |
||
1119 | break; |
||
1120 | case ATTR: |
||
1121 | printf("attr%d", inst->src[i].reg); |
||
1122 | break; |
||
1123 | case UNIFORM: |
||
1124 | printf("u%d", inst->src[i].reg); |
||
1125 | break; |
||
1126 | case IMM: |
||
1127 | switch (inst->src[i].type) { |
||
1128 | case BRW_REGISTER_TYPE_F: |
||
1129 | printf("%fF", inst->src[i].imm.f); |
||
1130 | break; |
||
1131 | case BRW_REGISTER_TYPE_D: |
||
1132 | printf("%dD", inst->src[i].imm.i); |
||
1133 | break; |
||
1134 | case BRW_REGISTER_TYPE_UD: |
||
1135 | printf("%uU", inst->src[i].imm.u); |
||
1136 | break; |
||
1137 | default: |
||
1138 | printf("???"); |
||
1139 | break; |
||
1140 | } |
||
1141 | break; |
||
1142 | case BAD_FILE: |
||
1143 | printf("(null)"); |
||
1144 | break; |
||
1145 | default: |
||
1146 | printf("???"); |
||
1147 | break; |
||
1148 | } |
||
1149 | |||
1150 | if (inst->src[i].reg_offset) |
||
1151 | printf(".%d", inst->src[i].reg_offset); |
||
1152 | |||
1153 | static const char *chans[4] = {"x", "y", "z", "w"}; |
||
1154 | printf("."); |
||
1155 | for (int c = 0; c < 4; c++) { |
||
1156 | printf("%s", chans[BRW_GET_SWZ(inst->src[i].swizzle, c)]); |
||
1157 | } |
||
1158 | |||
1159 | if (i < 3) |
||
1160 | printf(", "); |
||
1161 | } |
||
1162 | |||
1163 | printf("\n"); |
||
1164 | } |
||
1165 | |||
1166 | /** |
||
1167 | * Replace each register of type ATTR in this->instructions with a reference |
||
1168 | * to a fixed HW register. |
||
1169 | */ |
||
1170 | void |
||
1171 | vec4_visitor::lower_attributes_to_hw_regs(const int *attribute_map) |
||
1172 | { |
||
1173 | foreach_list(node, &this->instructions) { |
||
1174 | vec4_instruction *inst = (vec4_instruction *)node; |
||
1175 | |||
1176 | /* We have to support ATTR as a destination for GL_FIXED fixup. */ |
||
1177 | if (inst->dst.file == ATTR) { |
||
1178 | int grf = attribute_map[inst->dst.reg + inst->dst.reg_offset]; |
||
1179 | |||
1180 | /* All attributes used in the shader need to have been assigned a |
||
1181 | * hardware register by the caller |
||
1182 | */ |
||
1183 | assert(grf != 0); |
||
1184 | |||
1185 | struct brw_reg reg = brw_vec8_grf(grf, 0); |
||
1186 | reg.type = inst->dst.type; |
||
1187 | reg.dw1.bits.writemask = inst->dst.writemask; |
||
1188 | |||
1189 | inst->dst.file = HW_REG; |
||
1190 | inst->dst.fixed_hw_reg = reg; |
||
1191 | } |
||
1192 | |||
1193 | for (int i = 0; i < 3; i++) { |
||
1194 | if (inst->src[i].file != ATTR) |
||
1195 | continue; |
||
1196 | |||
1197 | int grf = attribute_map[inst->src[i].reg + inst->src[i].reg_offset]; |
||
1198 | |||
1199 | /* All attributes used in the shader need to have been assigned a |
||
1200 | * hardware register by the caller |
||
1201 | */ |
||
1202 | assert(grf != 0); |
||
1203 | |||
1204 | struct brw_reg reg = brw_vec8_grf(grf, 0); |
||
1205 | reg.dw1.bits.swizzle = inst->src[i].swizzle; |
||
1206 | reg.type = inst->src[i].type; |
||
1207 | if (inst->src[i].abs) |
||
1208 | reg = brw_abs(reg); |
||
1209 | if (inst->src[i].negate) |
||
1210 | reg = negate(reg); |
||
1211 | |||
1212 | inst->src[i].file = HW_REG; |
||
1213 | inst->src[i].fixed_hw_reg = reg; |
||
1214 | } |
||
1215 | } |
||
1216 | } |
||
1217 | |||
1218 | int |
||
1219 | vec4_vs_visitor::setup_attributes(int payload_reg) |
||
1220 | { |
||
1221 | int nr_attributes; |
||
1222 | int attribute_map[VERT_ATTRIB_MAX + 1]; |
||
1223 | memset(attribute_map, 0, sizeof(attribute_map)); |
||
1224 | |||
1225 | nr_attributes = 0; |
||
1226 | for (int i = 0; i < VERT_ATTRIB_MAX; i++) { |
||
1227 | if (vs_prog_data->inputs_read & BITFIELD64_BIT(i)) { |
||
1228 | attribute_map[i] = payload_reg + nr_attributes; |
||
1229 | nr_attributes++; |
||
1230 | } |
||
1231 | } |
||
1232 | |||
1233 | /* VertexID is stored by the VF as the last vertex element, but we |
||
1234 | * don't represent it with a flag in inputs_read, so we call it |
||
1235 | * VERT_ATTRIB_MAX. |
||
1236 | */ |
||
1237 | if (vs_prog_data->uses_vertexid) { |
||
1238 | attribute_map[VERT_ATTRIB_MAX] = payload_reg + nr_attributes; |
||
1239 | nr_attributes++; |
||
1240 | } |
||
1241 | |||
1242 | lower_attributes_to_hw_regs(attribute_map); |
||
1243 | |||
1244 | /* The BSpec says we always have to read at least one thing from |
||
1245 | * the VF, and it appears that the hardware wedges otherwise. |
||
1246 | */ |
||
1247 | if (nr_attributes == 0) |
||
1248 | nr_attributes = 1; |
||
1249 | |||
1250 | prog_data->urb_read_length = (nr_attributes + 1) / 2; |
||
1251 | |||
1252 | unsigned vue_entries = |
||
1253 | MAX2(nr_attributes, prog_data->vue_map.num_slots); |
||
1254 | |||
1255 | if (brw->gen == 6) |
||
1256 | prog_data->urb_entry_size = ALIGN(vue_entries, 8) / 8; |
||
1257 | else |
||
1258 | prog_data->urb_entry_size = ALIGN(vue_entries, 4) / 4; |
||
1259 | |||
1260 | return payload_reg + nr_attributes; |
||
1261 | } |
||
1262 | |||
1263 | int |
||
1264 | vec4_visitor::setup_uniforms(int reg) |
||
1265 | { |
||
1266 | /* The pre-gen6 VS requires that some push constants get loaded no |
||
1267 | * matter what, or the GPU would hang. |
||
1268 | */ |
||
1269 | if (brw->gen < 6 && this->uniforms == 0) { |
||
1270 | this->uniform_vector_size[this->uniforms] = 1; |
||
1271 | |||
1272 | for (unsigned int i = 0; i < 4; i++) { |
||
1273 | unsigned int slot = this->uniforms * 4 + i; |
||
1274 | static float zero = 0.0; |
||
1275 | prog_data->param[slot] = &zero; |
||
1276 | } |
||
1277 | |||
1278 | this->uniforms++; |
||
1279 | reg++; |
||
1280 | } else { |
||
1281 | reg += ALIGN(uniforms, 2) / 2; |
||
1282 | } |
||
1283 | |||
1284 | prog_data->nr_params = this->uniforms * 4; |
||
1285 | |||
1286 | prog_data->curb_read_length = reg - 1; |
||
1287 | |||
1288 | return reg; |
||
1289 | } |
||
1290 | |||
1291 | void |
||
1292 | vec4_visitor::setup_payload(void) |
||
1293 | { |
||
1294 | int reg = 0; |
||
1295 | |||
1296 | /* The payload always contains important data in g0, which contains |
||
1297 | * the URB handles that are passed on to the URB write at the end |
||
1298 | * of the thread. So, we always start push constants at g1. |
||
1299 | */ |
||
1300 | reg++; |
||
1301 | |||
1302 | reg = setup_uniforms(reg); |
||
1303 | |||
1304 | reg = setup_attributes(reg); |
||
1305 | |||
1306 | this->first_non_payload_grf = reg; |
||
1307 | } |
||
1308 | |||
1309 | src_reg |
||
1310 | vec4_visitor::get_timestamp() |
||
1311 | { |
||
1312 | assert(brw->gen >= 7); |
||
1313 | |||
1314 | src_reg ts = src_reg(brw_reg(BRW_ARCHITECTURE_REGISTER_FILE, |
||
1315 | BRW_ARF_TIMESTAMP, |
||
1316 | 0, |
||
1317 | BRW_REGISTER_TYPE_UD, |
||
1318 | BRW_VERTICAL_STRIDE_0, |
||
1319 | BRW_WIDTH_4, |
||
1320 | BRW_HORIZONTAL_STRIDE_4, |
||
1321 | BRW_SWIZZLE_XYZW, |
||
1322 | WRITEMASK_XYZW)); |
||
1323 | |||
1324 | dst_reg dst = dst_reg(this, glsl_type::uvec4_type); |
||
1325 | |||
1326 | vec4_instruction *mov = emit(MOV(dst, ts)); |
||
1327 | /* We want to read the 3 fields we care about (mostly field 0, but also 2) |
||
1328 | * even if it's not enabled in the dispatch. |
||
1329 | */ |
||
1330 | mov->force_writemask_all = true; |
||
1331 | |||
1332 | return src_reg(dst); |
||
1333 | } |
||
1334 | |||
1335 | void |
||
1336 | vec4_visitor::emit_shader_time_begin() |
||
1337 | { |
||
1338 | current_annotation = "shader time start"; |
||
1339 | shader_start_time = get_timestamp(); |
||
1340 | } |
||
1341 | |||
1342 | void |
||
1343 | vec4_visitor::emit_shader_time_end() |
||
1344 | { |
||
1345 | current_annotation = "shader time end"; |
||
1346 | src_reg shader_end_time = get_timestamp(); |
||
1347 | |||
1348 | |||
1349 | /* Check that there weren't any timestamp reset events (assuming these |
||
1350 | * were the only two timestamp reads that happened). |
||
1351 | */ |
||
1352 | src_reg reset_end = shader_end_time; |
||
1353 | reset_end.swizzle = BRW_SWIZZLE_ZZZZ; |
||
1354 | vec4_instruction *test = emit(AND(dst_null_d(), reset_end, src_reg(1u))); |
||
1355 | test->conditional_mod = BRW_CONDITIONAL_Z; |
||
1356 | |||
1357 | emit(IF(BRW_PREDICATE_NORMAL)); |
||
1358 | |||
1359 | /* Take the current timestamp and get the delta. */ |
||
1360 | shader_start_time.negate = true; |
||
1361 | dst_reg diff = dst_reg(this, glsl_type::uint_type); |
||
1362 | emit(ADD(diff, shader_start_time, shader_end_time)); |
||
1363 | |||
1364 | /* If there were no instructions between the two timestamp gets, the diff |
||
1365 | * is 2 cycles. Remove that overhead, so I can forget about that when |
||
1366 | * trying to determine the time taken for single instructions. |
||
1367 | */ |
||
1368 | emit(ADD(diff, src_reg(diff), src_reg(-2u))); |
||
1369 | |||
1370 | emit_shader_time_write(ST_VS, src_reg(diff)); |
||
1371 | emit_shader_time_write(ST_VS_WRITTEN, src_reg(1u)); |
||
1372 | emit(BRW_OPCODE_ELSE); |
||
1373 | emit_shader_time_write(ST_VS_RESET, src_reg(1u)); |
||
1374 | emit(BRW_OPCODE_ENDIF); |
||
1375 | } |
||
1376 | |||
1377 | void |
||
1378 | vec4_visitor::emit_shader_time_write(enum shader_time_shader_type type, |
||
1379 | src_reg value) |
||
1380 | { |
||
1381 | int shader_time_index = |
||
1382 | brw_get_shader_time_index(brw, shader_prog, prog, type); |
||
1383 | |||
1384 | dst_reg dst = |
||
1385 | dst_reg(this, glsl_type::get_array_instance(glsl_type::vec4_type, 2)); |
||
1386 | |||
1387 | dst_reg offset = dst; |
||
1388 | dst_reg time = dst; |
||
1389 | time.reg_offset++; |
||
1390 | |||
1391 | offset.type = BRW_REGISTER_TYPE_UD; |
||
1392 | emit(MOV(offset, src_reg(shader_time_index * SHADER_TIME_STRIDE))); |
||
1393 | |||
1394 | time.type = BRW_REGISTER_TYPE_UD; |
||
1395 | emit(MOV(time, src_reg(value))); |
||
1396 | |||
1397 | emit(SHADER_OPCODE_SHADER_TIME_ADD, dst_reg(), src_reg(dst)); |
||
1398 | } |
||
1399 | |||
1400 | bool |
||
1401 | vec4_visitor::run() |
||
1402 | { |
||
1403 | sanity_param_count = prog->Parameters->NumParameters; |
||
1404 | |||
1405 | if (INTEL_DEBUG & DEBUG_SHADER_TIME) |
||
1406 | emit_shader_time_begin(); |
||
1407 | |||
1408 | emit_prolog(); |
||
1409 | |||
1410 | /* Generate VS IR for main(). (the visitor only descends into |
||
1411 | * functions called "main"). |
||
1412 | */ |
||
1413 | if (shader) { |
||
1414 | visit_instructions(shader->ir); |
||
1415 | } else { |
||
1416 | emit_program_code(); |
||
1417 | } |
||
1418 | base_ir = NULL; |
||
1419 | |||
1420 | if (key->userclip_active && !key->uses_clip_distance) |
||
1421 | setup_uniform_clipplane_values(); |
||
1422 | |||
1423 | emit_thread_end(); |
||
1424 | |||
1425 | /* Before any optimization, push array accesses out to scratch |
||
1426 | * space where we need them to be. This pass may allocate new |
||
1427 | * virtual GRFs, so we want to do it early. It also makes sure |
||
1428 | * that we have reladdr computations available for CSE, since we'll |
||
1429 | * often do repeated subexpressions for those. |
||
1430 | */ |
||
1431 | if (shader) { |
||
1432 | move_grf_array_access_to_scratch(); |
||
1433 | move_uniform_array_access_to_pull_constants(); |
||
1434 | } else { |
||
1435 | /* The ARB_vertex_program frontend emits pull constant loads directly |
||
1436 | * rather than using reladdr, so we don't need to walk through all the |
||
1437 | * instructions looking for things to move. There isn't anything. |
||
1438 | * |
||
1439 | * We do still need to split things to vec4 size. |
||
1440 | */ |
||
1441 | split_uniform_registers(); |
||
1442 | } |
||
1443 | pack_uniform_registers(); |
||
1444 | move_push_constants_to_pull_constants(); |
||
1445 | split_virtual_grfs(); |
||
1446 | |||
1447 | bool progress; |
||
1448 | do { |
||
1449 | progress = false; |
||
1450 | progress = dead_code_eliminate() || progress; |
||
1451 | progress = opt_copy_propagation() || progress; |
||
1452 | progress = opt_algebraic() || progress; |
||
1453 | progress = opt_register_coalesce() || progress; |
||
1454 | } while (progress); |
||
1455 | |||
1456 | |||
1457 | if (failed) |
||
1458 | return false; |
||
1459 | |||
1460 | setup_payload(); |
||
1461 | |||
1462 | if (false) { |
||
1463 | /* Debug of register spilling: Go spill everything. */ |
||
1464 | const int grf_count = virtual_grf_count; |
||
1465 | float spill_costs[virtual_grf_count]; |
||
1466 | bool no_spill[virtual_grf_count]; |
||
1467 | evaluate_spill_costs(spill_costs, no_spill); |
||
1468 | for (int i = 0; i < grf_count; i++) { |
||
1469 | if (no_spill[i]) |
||
1470 | continue; |
||
1471 | spill_reg(i); |
||
1472 | } |
||
1473 | } |
||
1474 | |||
1475 | while (!reg_allocate()) { |
||
1476 | if (failed) |
||
1477 | break; |
||
1478 | } |
||
1479 | |||
1480 | opt_schedule_instructions(); |
||
1481 | |||
1482 | opt_set_dependency_control(); |
||
1483 | |||
1484 | /* If any state parameters were appended, then ParameterValues could have |
||
1485 | * been realloced, in which case the driver uniform storage set up by |
||
1486 | * _mesa_associate_uniform_storage() would point to freed memory. Make |
||
1487 | * sure that didn't happen. |
||
1488 | */ |
||
1489 | assert(sanity_param_count == prog->Parameters->NumParameters); |
||
1490 | |||
1491 | return !failed; |
||
1492 | } |
||
1493 | |||
1494 | } /* namespace brw */ |
||
1495 | |||
1496 | extern "C" { |
||
1497 | |||
1498 | /** |
||
1499 | * Compile a vertex shader. |
||
1500 | * |
||
1501 | * Returns the final assembly and the program's size. |
||
1502 | */ |
||
1503 | const unsigned * |
||
1504 | brw_vs_emit(struct brw_context *brw, |
||
1505 | struct gl_shader_program *prog, |
||
1506 | struct brw_vs_compile *c, |
||
1507 | struct brw_vs_prog_data *prog_data, |
||
1508 | void *mem_ctx, |
||
1509 | unsigned *final_assembly_size) |
||
1510 | { |
||
1511 | bool start_busy = false; |
||
1512 | float start_time = 0; |
||
1513 | |||
1514 | if (unlikely(brw->perf_debug)) { |
||
1515 | start_busy = (brw->batch.last_bo && |
||
1516 | drm_intel_bo_busy(brw->batch.last_bo)); |
||
1517 | start_time = get_time(); |
||
1518 | } |
||
1519 | |||
1520 | struct brw_shader *shader = NULL; |
||
1521 | if (prog) |
||
1522 | shader = (brw_shader *) prog->_LinkedShaders[MESA_SHADER_VERTEX]; |
||
1523 | |||
1524 | if (unlikely(INTEL_DEBUG & DEBUG_VS)) { |
||
1525 | if (prog) { |
||
1526 | printf("GLSL IR for native vertex shader %d:\n", prog->Name); |
||
1527 | _mesa_print_ir(shader->ir, NULL); |
||
1528 | printf("\n\n"); |
||
1529 | } else { |
||
1530 | printf("ARB_vertex_program %d for native vertex shader\n", |
||
1531 | c->vp->program.Base.Id); |
||
1532 | _mesa_print_program(&c->vp->program.Base); |
||
1533 | } |
||
1534 | } |
||
1535 | |||
1536 | vec4_vs_visitor v(brw, c, prog_data, prog, shader, mem_ctx); |
||
1537 | if (!v.run()) { |
||
1538 | if (prog) { |
||
1539 | prog->LinkStatus = false; |
||
1540 | ralloc_strcat(&prog->InfoLog, v.fail_msg); |
||
1541 | } |
||
1542 | |||
1543 | _mesa_problem(NULL, "Failed to compile vertex shader: %s\n", |
||
1544 | v.fail_msg); |
||
1545 | |||
1546 | return NULL; |
||
1547 | } |
||
1548 | |||
1549 | vec4_generator g(brw, prog, &c->vp->program.Base, mem_ctx, |
||
1550 | INTEL_DEBUG & DEBUG_VS); |
||
1551 | const unsigned *generated =g.generate_assembly(&v.instructions, |
||
1552 | final_assembly_size); |
||
1553 | |||
1554 | if (unlikely(brw->perf_debug) && shader) { |
||
1555 | if (shader->compiled_once) { |
||
1556 | brw_vs_debug_recompile(brw, prog, &c->key); |
||
1557 | } |
||
1558 | if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) { |
||
1559 | perf_debug("VS compile took %.03f ms and stalled the GPU\n", |
||
1560 | (get_time() - start_time) * 1000); |
||
1561 | } |
||
1562 | shader->compiled_once = true; |
||
1563 | } |
||
1564 | |||
1565 | return generated; |
||
1566 | } |
||
1567 | |||
1568 | } /* extern "C" */>>>>>>>>>>>>>>>>><>>><>><>>><>><>><>>><>><>><>>><>><>><>>>>>>>>>=>>>=>>>>>>>><>>=>><> |