projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
dri/i9*5: correctly calculate the amount of system memory
[mesa.git] / src / mesa / drivers / dri / i965 / brw_fs_generator.cpp
1 /*
2 * Copyright © 2010 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 /** @file brw_fs_generator.cpp
25 *
26 * This file supports generating code from the FS LIR to the actual
27 * native instructions.
28 */
29
30 extern "C" {
31 #include "main/macros.h"
32 #include "brw_context.h"
33 #include "brw_eu.h"
34 } /* extern "C" */
35
36 #include "brw_fs.h"
37 #include "brw_cfg.h"
38
39 fs_generator::fs_generator(struct brw_context *brw,
40 struct brw_wm_compile *c,
41 struct gl_shader_program *prog,
42 struct gl_fragment_program *fp,
43 bool dual_source_output)
44
45 : brw(brw), c(c), prog(prog), fp(fp), dual_source_output(dual_source_output)
46 {
47 ctx = &brw->ctx;
48
49 mem_ctx = c;
50
51 p = rzalloc(mem_ctx, struct brw_compile);
52 brw_init_compile(brw, p, mem_ctx);
53 }
54
55 fs_generator::~fs_generator()
56 {
57 }
58
59 void
60 fs_generator::patch_discard_jumps_to_fb_writes()
61 {
62 if (brw->gen < 6 || this->discard_halt_patches.is_empty())
63 return;
64
65 /* There is a somewhat strange undocumented requirement of using
66 * HALT, according to the simulator. If some channel has HALTed to
67 * a particular UIP, then by the end of the program, every channel
68 * must have HALTed to that UIP. Furthermore, the tracking is a
69 * stack, so you can't do the final halt of a UIP after starting
70 * halting to a new UIP.
71 *
72 * Symptoms of not emitting this instruction on actual hardware
73 * included GPU hangs and sparkly rendering on the piglit discard
74 * tests.
75 */
76 struct brw_instruction *last_halt = gen6_HALT(p);
77 last_halt->bits3.break_cont.uip = 2;
78 last_halt->bits3.break_cont.jip = 2;
79
80 int ip = p->nr_insn;
81
82 foreach_list(node, &this->discard_halt_patches) {
83 ip_record *patch_ip = (ip_record *)node;
84 struct brw_instruction *patch = &p->store[patch_ip->ip];
85
86 assert(patch->header.opcode == BRW_OPCODE_HALT);
87 /* HALT takes a half-instruction distance from the pre-incremented IP. */
88 patch->bits3.break_cont.uip = (ip - patch_ip->ip) * 2;
89 }
90
91 this->discard_halt_patches.make_empty();
92 }
93
94 void
95 fs_generator::generate_fb_write(fs_inst *inst)
96 {
97 bool eot = inst->eot;
98 struct brw_reg implied_header;
99 uint32_t msg_control;
100
101 /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
102 * move, here's g1.
103 */
104 brw_push_insn_state(p);
105 brw_set_mask_control(p, BRW_MASK_DISABLE);
106 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
107
108 if (inst->header_present) {
109 /* On HSW, the GPU will use the predicate on SENDC, unless the header is
110 * present.
111 */
112 if (!brw->is_haswell && ((fp && fp->UsesKill) ||
113 c->key.alpha_test_func)) {
114 struct brw_reg pixel_mask;
115
116 if (brw->gen >= 6)
117 pixel_mask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
118 else
119 pixel_mask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
120
121 brw_MOV(p, pixel_mask, brw_flag_reg(0, 1));
122 }
123
124 if (brw->gen >= 6) {
125 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
126 brw_MOV(p,
127 retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD),
128 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
129 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
130
131 if (inst->target > 0 && c->key.replicate_alpha) {
132 /* Set "Source0 Alpha Present to RenderTarget" bit in message
133 * header.
134 */
135 brw_OR(p,
136 vec1(retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD)),
137 vec1(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD)),
138 brw_imm_ud(0x1 << 11));
139 }
140
141 if (inst->target > 0) {
142 /* Set the render target index for choosing BLEND_STATE. */
143 brw_MOV(p, retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE,
144 inst->base_mrf, 2),
145 BRW_REGISTER_TYPE_UD),
146 brw_imm_ud(inst->target));
147 }
148
149 implied_header = brw_null_reg();
150 } else {
151 implied_header = retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW);
152
153 brw_MOV(p,
154 brw_message_reg(inst->base_mrf + 1),
155 brw_vec8_grf(1, 0));
156 }
157 } else {
158 implied_header = brw_null_reg();
159 }
160
161 if (this->dual_source_output)
162 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN01;
163 else if (dispatch_width == 16)
164 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE;
165 else
166 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01;
167
168 brw_pop_insn_state(p);
169
170 uint32_t surf_index =
171 c->prog_data.binding_table.render_target_start + inst->target;
172 brw_fb_WRITE(p,
173 dispatch_width,
174 inst->base_mrf,
175 implied_header,
176 msg_control,
177 surf_index,
178 inst->mlen,
179 0,
180 eot,
181 inst->header_present);
182
183 brw_mark_surface_used(&c->prog_data.base, surf_index);
184 }
185
186 void
187 fs_generator::generate_blorp_fb_write(fs_inst *inst)
188 {
189 brw_fb_WRITE(p,
190 16 /* dispatch_width */,
191 inst->base_mrf,
192 brw_reg_from_fs_reg(&inst->src[0]),
193 BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE,
194 inst->target,
195 inst->mlen,
196 0,
197 true,
198 inst->header_present);
199 }
200
201 /* Computes the integer pixel x,y values from the origin.
202 *
203 * This is the basis of gl_FragCoord computation, but is also used
204 * pre-gen6 for computing the deltas from v0 for computing
205 * interpolation.
206 */
207 void
208 fs_generator::generate_pixel_xy(struct brw_reg dst, bool is_x)
209 {
210 struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
211 struct brw_reg src;
212 struct brw_reg deltas;
213
214 if (is_x) {
215 src = stride(suboffset(g1_uw, 4), 2, 4, 0);
216 deltas = brw_imm_v(0x10101010);
217 } else {
218 src = stride(suboffset(g1_uw, 5), 2, 4, 0);
219 deltas = brw_imm_v(0x11001100);
220 }
221
222 if (dispatch_width == 16) {
223 dst = vec16(dst);
224 }
225
226 /* We do this SIMD8 or SIMD16, but since the destination is UW we
227 * don't do compression in the SIMD16 case.
228 */
229 brw_push_insn_state(p);
230 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
231 brw_ADD(p, dst, src, deltas);
232 brw_pop_insn_state(p);
233 }
234
235 void
236 fs_generator::generate_linterp(fs_inst *inst,
237 struct brw_reg dst, struct brw_reg *src)
238 {
239 struct brw_reg delta_x = src[0];
240 struct brw_reg delta_y = src[1];
241 struct brw_reg interp = src[2];
242
243 if (brw->has_pln &&
244 delta_y.nr == delta_x.nr + 1 &&
245 (brw->gen >= 6 || (delta_x.nr & 1) == 0)) {
246 brw_PLN(p, dst, interp, delta_x);
247 } else {
248 brw_LINE(p, brw_null_reg(), interp, delta_x);
249 brw_MAC(p, dst, suboffset(interp, 1), delta_y);
250 }
251 }
252
253 void
254 fs_generator::generate_math1_gen7(fs_inst *inst,
255 struct brw_reg dst,
256 struct brw_reg src0)
257 {
258 assert(inst->mlen == 0);
259 brw_math(p, dst,
260 brw_math_function(inst->opcode),
261 0, src0,
262 BRW_MATH_DATA_VECTOR,
263 BRW_MATH_PRECISION_FULL);
264 }
265
266 void
267 fs_generator::generate_math2_gen7(fs_inst *inst,
268 struct brw_reg dst,
269 struct brw_reg src0,
270 struct brw_reg src1)
271 {
272 assert(inst->mlen == 0);
273 brw_math2(p, dst, brw_math_function(inst->opcode), src0, src1);
274 }
275
276 void
277 fs_generator::generate_math1_gen6(fs_inst *inst,
278 struct brw_reg dst,
279 struct brw_reg src0)
280 {
281 int op = brw_math_function(inst->opcode);
282
283 assert(inst->mlen == 0);
284
285 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
286 brw_math(p, dst,
287 op,
288 0, src0,
289 BRW_MATH_DATA_VECTOR,
290 BRW_MATH_PRECISION_FULL);
291
292 if (dispatch_width == 16) {
293 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
294 brw_math(p, sechalf(dst),
295 op,
296 0, sechalf(src0),
297 BRW_MATH_DATA_VECTOR,
298 BRW_MATH_PRECISION_FULL);
299 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
300 }
301 }
302
303 void
304 fs_generator::generate_math2_gen6(fs_inst *inst,
305 struct brw_reg dst,
306 struct brw_reg src0,
307 struct brw_reg src1)
308 {
309 int op = brw_math_function(inst->opcode);
310
311 assert(inst->mlen == 0);
312
313 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
314 brw_math2(p, dst, op, src0, src1);
315
316 if (dispatch_width == 16) {
317 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
318 brw_math2(p, sechalf(dst), op, sechalf(src0), sechalf(src1));
319 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
320 }
321 }
322
323 void
324 fs_generator::generate_math_gen4(fs_inst *inst,
325 struct brw_reg dst,
326 struct brw_reg src)
327 {
328 int op = brw_math_function(inst->opcode);
329
330 assert(inst->mlen >= 1);
331
332 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
333 brw_math(p, dst,
334 op,
335 inst->base_mrf, src,
336 BRW_MATH_DATA_VECTOR,
337 BRW_MATH_PRECISION_FULL);
338
339 if (dispatch_width == 16) {
340 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
341 brw_math(p, sechalf(dst),
342 op,
343 inst->base_mrf + 1, sechalf(src),
344 BRW_MATH_DATA_VECTOR,
345 BRW_MATH_PRECISION_FULL);
346
347 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
348 }
349 }
350
351 void
352 fs_generator::generate_math_g45(fs_inst *inst,
353 struct brw_reg dst,
354 struct brw_reg src)
355 {
356 if (inst->opcode == SHADER_OPCODE_POW ||
357 inst->opcode == SHADER_OPCODE_INT_QUOTIENT ||
358 inst->opcode == SHADER_OPCODE_INT_REMAINDER) {
359 generate_math_gen4(inst, dst, src);
360 return;
361 }
362
363 int op = brw_math_function(inst->opcode);
364
365 assert(inst->mlen >= 1);
366
367 brw_math(p, dst,
368 op,
369 inst->base_mrf, src,
370 BRW_MATH_DATA_VECTOR,
371 BRW_MATH_PRECISION_FULL);
372 }
373
374 void
375 fs_generator::generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src)
376 {
377 int msg_type = -1;
378 int rlen = 4;
379 uint32_t simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
380 uint32_t return_format;
381
382 switch (dst.type) {
383 case BRW_REGISTER_TYPE_D:
384 return_format = BRW_SAMPLER_RETURN_FORMAT_SINT32;
385 break;
386 case BRW_REGISTER_TYPE_UD:
387 return_format = BRW_SAMPLER_RETURN_FORMAT_UINT32;
388 break;
389 default:
390 return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
391 break;
392 }
393
394 if (dispatch_width == 16 &&
395 !inst->force_uncompressed && !inst->force_sechalf)
396 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
397
398 if (brw->gen >= 5) {
399 switch (inst->opcode) {
400 case SHADER_OPCODE_TEX:
401 if (inst->shadow_compare) {
402 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE;
403 } else {
404 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE;
405 }
406 break;
407 case FS_OPCODE_TXB:
408 if (inst->shadow_compare) {
409 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE;
410 } else {
411 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS;
412 }
413 break;
414 case SHADER_OPCODE_TXL:
415 if (inst->shadow_compare) {
416 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
417 } else {
418 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
419 }
420 break;
421 case SHADER_OPCODE_TXS:
422 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
423 break;
424 case SHADER_OPCODE_TXD:
425 if (inst->shadow_compare) {
426 /* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
427 assert(brw->is_haswell);
428 msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE;
429 } else {
430 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
431 }
432 break;
433 case SHADER_OPCODE_TXF:
434 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
435 break;
436 case SHADER_OPCODE_TXF_CMS:
437 if (brw->gen >= 7)
438 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
439 else
440 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
441 break;
442 case SHADER_OPCODE_TXF_UMS:
443 assert(brw->gen >= 7);
444 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS;
445 break;
446 case SHADER_OPCODE_TXF_MCS:
447 assert(brw->gen >= 7);
448 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS;
449 break;
450 case SHADER_OPCODE_LOD:
451 msg_type = GEN5_SAMPLER_MESSAGE_LOD;
452 break;
453 case SHADER_OPCODE_TG4:
454 if (inst->shadow_compare) {
455 assert(brw->gen >= 7);
456 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C;
457 } else {
458 assert(brw->gen >= 6);
459 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4;
460 }
461 break;
462 case SHADER_OPCODE_TG4_OFFSET:
463 assert(brw->gen >= 7);
464 if (inst->shadow_compare) {
465 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C;
466 } else {
467 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO;
468 }
469 break;
470 default:
471 assert(!"not reached");
472 break;
473 }
474 } else {
475 switch (inst->opcode) {
476 case SHADER_OPCODE_TEX:
477 /* Note that G45 and older determines shadow compare and dispatch width
478 * from message length for most messages.
479 */
480 assert(dispatch_width == 8);
481 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE;
482 if (inst->shadow_compare) {
483 assert(inst->mlen == 6);
484 } else {
485 assert(inst->mlen <= 4);
486 }
487 break;
488 case FS_OPCODE_TXB:
489 if (inst->shadow_compare) {
490 assert(inst->mlen == 6);
491 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_BIAS_COMPARE;
492 } else {
493 assert(inst->mlen == 9);
494 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS;
495 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
496 }
497 break;
498 case SHADER_OPCODE_TXL:
499 if (inst->shadow_compare) {
500 assert(inst->mlen == 6);
501 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_LOD_COMPARE;
502 } else {
503 assert(inst->mlen == 9);
504 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_LOD;
505 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
506 }
507 break;
508 case SHADER_OPCODE_TXD:
509 /* There is no sample_d_c message; comparisons are done manually */
510 assert(inst->mlen == 7 || inst->mlen == 10);
511 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_GRADIENTS;
512 break;
513 case SHADER_OPCODE_TXF:
514 assert(inst->mlen == 9);
515 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_LD;
516 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
517 break;
518 case SHADER_OPCODE_TXS:
519 assert(inst->mlen == 3);
520 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_RESINFO;
521 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
522 break;
523 default:
524 assert(!"not reached");
525 break;
526 }
527 }
528 assert(msg_type != -1);
529
530 if (simd_mode == BRW_SAMPLER_SIMD_MODE_SIMD16) {
531 rlen = 8;
532 dst = vec16(dst);
533 }
534
535 if (brw->gen >= 7 && inst->header_present && dispatch_width == 16) {
536 /* The send-from-GRF for SIMD16 texturing with a header has an extra
537 * hardware register allocated to it, which we need to skip over (since
538 * our coordinates in the payload are in the even-numbered registers,
539 * and the header comes right before the first one).
540 */
541 assert(src.file == BRW_GENERAL_REGISTER_FILE);
542 src.nr++;
543 }
544
545 /* Load the message header if present. If there's a texture offset,
546 * we need to set it up explicitly and load the offset bitfield.
547 * Otherwise, we can use an implied move from g0 to the first message reg.
548 */
549 if (inst->header_present) {
550 if (brw->gen < 6 && !inst->texture_offset) {
551 /* Set up an implied move from g0 to the MRF. */
552 src = retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW);
553 } else {
554 struct brw_reg header_reg;
555
556 if (brw->gen >= 7) {
557 header_reg = src;
558 } else {
559 assert(inst->base_mrf != -1);
560 header_reg = brw_message_reg(inst->base_mrf);
561 }
562
563 brw_push_insn_state(p);
564 brw_set_mask_control(p, BRW_MASK_DISABLE);
565 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
566 /* Explicitly set up the message header by copying g0 to the MRF. */
567 brw_MOV(p, header_reg, brw_vec8_grf(0, 0));
568
569 if (inst->texture_offset) {
570 /* Set the offset bits in DWord 2. */
571 brw_MOV(p, get_element_ud(header_reg, 2),
572 brw_imm_ud(inst->texture_offset));
573 }
574
575 if (inst->sampler >= 16) {
576 /* The "Sampler Index" field can only store values between 0 and 15.
577 * However, we can add an offset to the "Sampler State Pointer"
578 * field, effectively selecting a different set of 16 samplers.
579 *
580 * The "Sampler State Pointer" needs to be aligned to a 32-byte
581 * offset, and each sampler state is only 16-bytes, so we can't
582 * exclusively use the offset - we have to use both.
583 */
584 assert(brw->is_haswell); /* field only exists on Haswell */
585 brw_ADD(p,
586 get_element_ud(header_reg, 3),
587 get_element_ud(brw_vec8_grf(0, 0), 3),
588 brw_imm_ud(16 * (inst->sampler / 16) *
589 sizeof(gen7_sampler_state)));
590 }
591 brw_pop_insn_state(p);
592 }
593 }
594
595 uint32_t surface_index = ((inst->opcode == SHADER_OPCODE_TG4 ||
596 inst->opcode == SHADER_OPCODE_TG4_OFFSET)
597 ? c->prog_data.base.binding_table.gather_texture_start
598 : c->prog_data.base.binding_table.texture_start) + inst->sampler;
599
600 brw_SAMPLE(p,
601 retype(dst, BRW_REGISTER_TYPE_UW),
602 inst->base_mrf,
603 src,
604 surface_index,
605 inst->sampler % 16,
606 msg_type,
607 rlen,
608 inst->mlen,
609 inst->header_present,
610 simd_mode,
611 return_format);
612
613 brw_mark_surface_used(&c->prog_data.base, surface_index);
614 }
615
616
617 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
618 * looking like:
619 *
620 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
621 *
622 * Ideally, we want to produce:
623 *
624 * DDX DDY
625 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
626 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
627 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
628 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
629 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
630 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
631 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
632 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
633 *
634 * and add another set of two more subspans if in 16-pixel dispatch mode.
635 *
636 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
637 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
638 * pair. But the ideal approximation may impose a huge performance cost on
639 * sample_d. On at least Haswell, sample_d instruction does some
640 * optimizations if the same LOD is used for all pixels in the subspan.
641 *
642 * For DDY, we need to use ALIGN16 mode since it's capable of doing the
643 * appropriate swizzling.
644 */
645 void
646 fs_generator::generate_ddx(fs_inst *inst, struct brw_reg dst, struct brw_reg src)
647 {
648 unsigned vstride, width;
649
650 if (c->key.high_quality_derivatives) {
651 /* produce accurate derivatives */
652 vstride = BRW_VERTICAL_STRIDE_2;
653 width = BRW_WIDTH_2;
654 }
655 else {
656 /* replicate the derivative at the top-left pixel to other pixels */
657 vstride = BRW_VERTICAL_STRIDE_4;
658 width = BRW_WIDTH_4;
659 }
660
661 struct brw_reg src0 = brw_reg(src.file, src.nr, 1,
662 BRW_REGISTER_TYPE_F,
663 vstride,
664 width,
665 BRW_HORIZONTAL_STRIDE_0,
666 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
667 struct brw_reg src1 = brw_reg(src.file, src.nr, 0,
668 BRW_REGISTER_TYPE_F,
669 vstride,
670 width,
671 BRW_HORIZONTAL_STRIDE_0,
672 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
673 brw_ADD(p, dst, src0, negate(src1));
674 }
675
676 /* The negate_value boolean is used to negate the derivative computation for
677 * FBOs, since they place the origin at the upper left instead of the lower
678 * left.
679 */
680 void
681 fs_generator::generate_ddy(fs_inst *inst, struct brw_reg dst, struct brw_reg src,
682 bool negate_value)
683 {
684 if (c->key.high_quality_derivatives) {
685 /* From the Ivy Bridge PRM, volume 4 part 3, section 3.3.9 (Register
686 * Region Restrictions):
687 *
688 * In Align16 access mode, SIMD16 is not allowed for DW operations
689 * and SIMD8 is not allowed for DF operations.
690 *
691 * In this context, "DW operations" means "operations acting on 32-bit
692 * values", so it includes operations on floats.
693 *
694 * Gen4 has a similar restriction. From the i965 PRM, section 11.5.3
695 * (Instruction Compression -> Rules and Restrictions):
696 *
697 * A compressed instruction must be in Align1 access mode. Align16
698 * mode instructions cannot be compressed.
699 *
700 * Similar text exists in the g45 PRM.
701 *
702 * On these platforms, if we're building a SIMD16 shader, we need to
703 * manually unroll to a pair of SIMD8 instructions.
704 */
705 bool unroll_to_simd8 =
706 (dispatch_width == 16 &&
707 (brw->gen == 4 || (brw->gen == 7 && !brw->is_haswell)));
708
709 /* produce accurate derivatives */
710 struct brw_reg src0 = brw_reg(src.file, src.nr, 0,
711 BRW_REGISTER_TYPE_F,
712 BRW_VERTICAL_STRIDE_4,
713 BRW_WIDTH_4,
714 BRW_HORIZONTAL_STRIDE_1,
715 BRW_SWIZZLE_XYXY, WRITEMASK_XYZW);
716 struct brw_reg src1 = brw_reg(src.file, src.nr, 0,
717 BRW_REGISTER_TYPE_F,
718 BRW_VERTICAL_STRIDE_4,
719 BRW_WIDTH_4,
720 BRW_HORIZONTAL_STRIDE_1,
721 BRW_SWIZZLE_ZWZW, WRITEMASK_XYZW);
722 brw_push_insn_state(p);
723 brw_set_access_mode(p, BRW_ALIGN_16);
724 if (unroll_to_simd8)
725 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
726 if (negate_value)
727 brw_ADD(p, dst, src1, negate(src0));
728 else
729 brw_ADD(p, dst, src0, negate(src1));
730 if (unroll_to_simd8) {
731 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
732 src0 = sechalf(src0);
733 src1 = sechalf(src1);
734 dst = sechalf(dst);
735 if (negate_value)
736 brw_ADD(p, dst, src1, negate(src0));
737 else
738 brw_ADD(p, dst, src0, negate(src1));
739 }
740 brw_pop_insn_state(p);
741 } else {
742 /* replicate the derivative at the top-left pixel to other pixels */
743 struct brw_reg src0 = brw_reg(src.file, src.nr, 0,
744 BRW_REGISTER_TYPE_F,
745 BRW_VERTICAL_STRIDE_4,
746 BRW_WIDTH_4,
747 BRW_HORIZONTAL_STRIDE_0,
748 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
749 struct brw_reg src1 = brw_reg(src.file, src.nr, 2,
750 BRW_REGISTER_TYPE_F,
751 BRW_VERTICAL_STRIDE_4,
752 BRW_WIDTH_4,
753 BRW_HORIZONTAL_STRIDE_0,
754 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
755 if (negate_value)
756 brw_ADD(p, dst, src1, negate(src0));
757 else
758 brw_ADD(p, dst, src0, negate(src1));
759 }
760 }
761
762 void
763 fs_generator::generate_discard_jump(fs_inst *inst)
764 {
765 assert(brw->gen >= 6);
766
767 /* This HALT will be patched up at FB write time to point UIP at the end of
768 * the program, and at brw_uip_jip() JIP will be set to the end of the
769 * current block (or the program).
770 */
771 this->discard_halt_patches.push_tail(new(mem_ctx) ip_record(p->nr_insn));
772
773 brw_push_insn_state(p);
774 brw_set_mask_control(p, BRW_MASK_DISABLE);
775 gen6_HALT(p);
776 brw_pop_insn_state(p);
777 }
778
779 void
780 fs_generator::generate_scratch_write(fs_inst *inst, struct brw_reg src)
781 {
782 assert(inst->mlen != 0);
783
784 brw_MOV(p,
785 retype(brw_message_reg(inst->base_mrf + 1), BRW_REGISTER_TYPE_UD),
786 retype(src, BRW_REGISTER_TYPE_UD));
787 brw_oword_block_write_scratch(p, brw_message_reg(inst->base_mrf),
788 dispatch_width / 8, inst->offset);
789 }
790
791 void
792 fs_generator::generate_scratch_read(fs_inst *inst, struct brw_reg dst)
793 {
794 assert(inst->mlen != 0);
795
796 brw_oword_block_read_scratch(p, dst, brw_message_reg(inst->base_mrf),
797 dispatch_width / 8, inst->offset);
798 }
799
800 void
801 fs_generator::generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst)
802 {
803 gen7_block_read_scratch(p, dst, dispatch_width / 8, inst->offset);
804 }
805
806 void
807 fs_generator::generate_uniform_pull_constant_load(fs_inst *inst,
808 struct brw_reg dst,
809 struct brw_reg index,
810 struct brw_reg offset)
811 {
812 assert(inst->mlen != 0);
813
814 assert(index.file == BRW_IMMEDIATE_VALUE &&
815 index.type == BRW_REGISTER_TYPE_UD);
816 uint32_t surf_index = index.dw1.ud;
817
818 assert(offset.file == BRW_IMMEDIATE_VALUE &&
819 offset.type == BRW_REGISTER_TYPE_UD);
820 uint32_t read_offset = offset.dw1.ud;
821
822 brw_oword_block_read(p, dst, brw_message_reg(inst->base_mrf),
823 read_offset, surf_index);
824
825 brw_mark_surface_used(&c->prog_data.base, surf_index);
826 }
827
828 void
829 fs_generator::generate_uniform_pull_constant_load_gen7(fs_inst *inst,
830 struct brw_reg dst,
831 struct brw_reg index,
832 struct brw_reg offset)
833 {
834 assert(inst->mlen == 0);
835
836 assert(index.file == BRW_IMMEDIATE_VALUE &&
837 index.type == BRW_REGISTER_TYPE_UD);
838 uint32_t surf_index = index.dw1.ud;
839
840 assert(offset.file == BRW_GENERAL_REGISTER_FILE);
841 /* Reference just the dword we need, to avoid angering validate_reg(). */
842 offset = brw_vec1_grf(offset.nr, 0);
843
844 brw_push_insn_state(p);
845 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
846 brw_set_mask_control(p, BRW_MASK_DISABLE);
847 struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
848 brw_pop_insn_state(p);
849
850 /* We use the SIMD4x2 mode because we want to end up with 4 components in
851 * the destination loaded consecutively from the same offset (which appears
852 * in the first component, and the rest are ignored).
853 */
854 dst.width = BRW_WIDTH_4;
855 brw_set_dest(p, send, dst);
856 brw_set_src0(p, send, offset);
857 brw_set_sampler_message(p, send,
858 surf_index,
859 0, /* LD message ignores sampler unit */
860 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
861 1, /* rlen */
862 1, /* mlen */
863 false, /* no header */
864 BRW_SAMPLER_SIMD_MODE_SIMD4X2,
865 0);
866
867 brw_mark_surface_used(&c->prog_data.base, surf_index);
868 }
869
870 void
871 fs_generator::generate_varying_pull_constant_load(fs_inst *inst,
872 struct brw_reg dst,
873 struct brw_reg index,
874 struct brw_reg offset)
875 {
876 assert(brw->gen < 7); /* Should use the gen7 variant. */
877 assert(inst->header_present);
878 assert(inst->mlen);
879
880 assert(index.file == BRW_IMMEDIATE_VALUE &&
881 index.type == BRW_REGISTER_TYPE_UD);
882 uint32_t surf_index = index.dw1.ud;
883
884 uint32_t simd_mode, rlen, msg_type;
885 if (dispatch_width == 16) {
886 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
887 rlen = 8;
888 } else {
889 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
890 rlen = 4;
891 }
892
893 if (brw->gen >= 5)
894 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
895 else {
896 /* We always use the SIMD16 message so that we only have to load U, and
897 * not V or R.
898 */
899 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_LD;
900 assert(inst->mlen == 3);
901 assert(inst->regs_written == 8);
902 rlen = 8;
903 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
904 }
905
906 struct brw_reg offset_mrf = retype(brw_message_reg(inst->base_mrf + 1),
907 BRW_REGISTER_TYPE_D);
908 brw_MOV(p, offset_mrf, offset);
909
910 struct brw_reg header = brw_vec8_grf(0, 0);
911 gen6_resolve_implied_move(p, &header, inst->base_mrf);
912
913 struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
914 send->header.compression_control = BRW_COMPRESSION_NONE;
915 brw_set_dest(p, send, dst);
916 brw_set_src0(p, send, header);
917 if (brw->gen < 6)
918 send->header.destreg__conditionalmod = inst->base_mrf;
919
920 /* Our surface is set up as floats, regardless of what actual data is
921 * stored in it.
922 */
923 uint32_t return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
924 brw_set_sampler_message(p, send,
925 surf_index,
926 0, /* sampler (unused) */
927 msg_type,
928 rlen,
929 inst->mlen,
930 inst->header_present,
931 simd_mode,
932 return_format);
933
934 brw_mark_surface_used(&c->prog_data.base, surf_index);
935 }
936
937 void
938 fs_generator::generate_varying_pull_constant_load_gen7(fs_inst *inst,
939 struct brw_reg dst,
940 struct brw_reg index,
941 struct brw_reg offset)
942 {
943 assert(brw->gen >= 7);
944 /* Varying-offset pull constant loads are treated as a normal expression on
945 * gen7, so the fact that it's a send message is hidden at the IR level.
946 */
947 assert(!inst->header_present);
948 assert(!inst->mlen);
949
950 assert(index.file == BRW_IMMEDIATE_VALUE &&
951 index.type == BRW_REGISTER_TYPE_UD);
952 uint32_t surf_index = index.dw1.ud;
953
954 uint32_t simd_mode, rlen, mlen;
955 if (dispatch_width == 16) {
956 mlen = 2;
957 rlen = 8;
958 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
959 } else {
960 mlen = 1;
961 rlen = 4;
962 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
963 }
964
965 struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
966 brw_set_dest(p, send, dst);
967 brw_set_src0(p, send, offset);
968 brw_set_sampler_message(p, send,
969 surf_index,
970 0, /* LD message ignores sampler unit */
971 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
972 rlen,
973 mlen,
974 false, /* no header */
975 simd_mode,
976 0);
977
978 brw_mark_surface_used(&c->prog_data.base, surf_index);
979 }
980
981 /**
982 * Cause the current pixel/sample mask (from R1.7 bits 15:0) to be transferred
983 * into the flags register (f0.0).
984 *
985 * Used only on Gen6 and above.
986 */
987 void
988 fs_generator::generate_mov_dispatch_to_flags(fs_inst *inst)
989 {
990 struct brw_reg flags = brw_flag_reg(0, inst->flag_subreg);
991 struct brw_reg dispatch_mask;
992
993 if (brw->gen >= 6)
994 dispatch_mask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
995 else
996 dispatch_mask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
997
998 brw_push_insn_state(p);
999 brw_set_mask_control(p, BRW_MASK_DISABLE);
1000 brw_MOV(p, flags, dispatch_mask);
1001 brw_pop_insn_state(p);
1002 }
1003
1004
1005 static uint32_t brw_file_from_reg(fs_reg *reg)
1006 {
1007 switch (reg->file) {
1008 case GRF:
1009 return BRW_GENERAL_REGISTER_FILE;
1010 case MRF:
1011 return BRW_MESSAGE_REGISTER_FILE;
1012 case IMM:
1013 return BRW_IMMEDIATE_VALUE;
1014 default:
1015 assert(!"not reached");
1016 return BRW_GENERAL_REGISTER_FILE;
1017 }
1018 }
1019
1020 struct brw_reg
1021 brw_reg_from_fs_reg(fs_reg *reg)
1022 {
1023 struct brw_reg brw_reg;
1024
1025 switch (reg->file) {
1026 case GRF:
1027 case MRF:
1028 if (reg->stride == 0) {
1029 brw_reg = brw_vec1_reg(brw_file_from_reg(reg), reg->reg, 0);
1030 } else {
1031 brw_reg = brw_vec8_reg(brw_file_from_reg(reg), reg->reg, 0);
1032 brw_reg = stride(brw_reg, 8 * reg->stride, 8, reg->stride);
1033 }
1034
1035 brw_reg = retype(brw_reg, reg->type);
1036 brw_reg = byte_offset(brw_reg, reg->subreg_offset);
1037 break;
1038 case IMM:
1039 switch (reg->type) {
1040 case BRW_REGISTER_TYPE_F:
1041 brw_reg = brw_imm_f(reg->imm.f);
1042 break;
1043 case BRW_REGISTER_TYPE_D:
1044 brw_reg = brw_imm_d(reg->imm.i);
1045 break;
1046 case BRW_REGISTER_TYPE_UD:
1047 brw_reg = brw_imm_ud(reg->imm.u);
1048 break;
1049 default:
1050 assert(!"not reached");
1051 brw_reg = brw_null_reg();
1052 break;
1053 }
1054 break;
1055 case HW_REG:
1056 assert(reg->type == reg->fixed_hw_reg.type);
1057 brw_reg = reg->fixed_hw_reg;
1058 break;
1059 case BAD_FILE:
1060 /* Probably unused. */
1061 brw_reg = brw_null_reg();
1062 break;
1063 case UNIFORM:
1064 assert(!"not reached");
1065 brw_reg = brw_null_reg();
1066 break;
1067 default:
1068 assert(!"not reached");
1069 brw_reg = brw_null_reg();
1070 break;
1071 }
1072 if (reg->abs)
1073 brw_reg = brw_abs(brw_reg);
1074 if (reg->negate)
1075 brw_reg = negate(brw_reg);
1076
1077 return brw_reg;
1078 }
1079
1080 /**
1081 * Sets the first word of a vgrf for gen7+ simd4x2 uniform pull constant
1082 * sampler LD messages.
1083 *
1084 * We don't want to bake it into the send message's code generation because
1085 * that means we don't get a chance to schedule the instructions.
1086 */
1087 void
1088 fs_generator::generate_set_simd4x2_offset(fs_inst *inst,
1089 struct brw_reg dst,
1090 struct brw_reg value)
1091 {
1092 assert(value.file == BRW_IMMEDIATE_VALUE);
1093
1094 brw_push_insn_state(p);
1095 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
1096 brw_set_mask_control(p, BRW_MASK_DISABLE);
1097 brw_MOV(p, retype(brw_vec1_reg(dst.file, dst.nr, 0), value.type), value);
1098 brw_pop_insn_state(p);
1099 }
1100
1101 /* Sets vstride=16, width=8, hstride=2 or vstride=0, width=1, hstride=0
1102 * (when mask is passed as a uniform) of register mask before moving it
1103 * to register dst.
1104 */
1105 void
1106 fs_generator::generate_set_omask(fs_inst *inst,
1107 struct brw_reg dst,
1108 struct brw_reg mask)
1109 {
1110 bool stride_8_8_1 =
1111 (mask.vstride == BRW_VERTICAL_STRIDE_8 &&
1112 mask.width == BRW_WIDTH_8 &&
1113 mask.hstride == BRW_HORIZONTAL_STRIDE_1);
1114
1115 bool stride_0_1_0 =
1116 (mask.vstride == BRW_VERTICAL_STRIDE_0 &&
1117 mask.width == BRW_WIDTH_1 &&
1118 mask.hstride == BRW_HORIZONTAL_STRIDE_0);
1119
1120 assert(stride_8_8_1 || stride_0_1_0);
1121 assert(dst.type == BRW_REGISTER_TYPE_UW);
1122
1123 if (dispatch_width == 16)
1124 dst = vec16(dst);
1125 brw_push_insn_state(p);
1126 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
1127 brw_set_mask_control(p, BRW_MASK_DISABLE);
1128
1129 if (stride_8_8_1) {
1130 brw_MOV(p, dst, retype(stride(mask, 16, 8, 2), dst.type));
1131 } else if (stride_0_1_0) {
1132 brw_MOV(p, dst, retype(mask, dst.type));
1133 }
1134 brw_pop_insn_state(p);
1135 }
1136
1137 /* Sets vstride=1, width=4, hstride=0 of register src1 during
1138 * the ADD instruction.
1139 */
1140 void
1141 fs_generator::generate_set_sample_id(fs_inst *inst,
1142 struct brw_reg dst,
1143 struct brw_reg src0,
1144 struct brw_reg src1)
1145 {
1146 assert(dst.type == BRW_REGISTER_TYPE_D ||
1147 dst.type == BRW_REGISTER_TYPE_UD);
1148 assert(src0.type == BRW_REGISTER_TYPE_D ||
1149 src0.type == BRW_REGISTER_TYPE_UD);
1150
1151 brw_push_insn_state(p);
1152 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
1153 brw_set_mask_control(p, BRW_MASK_DISABLE);
1154 struct brw_reg reg = retype(stride(src1, 1, 4, 0), BRW_REGISTER_TYPE_UW);
1155 brw_ADD(p, dst, src0, reg);
1156 if (dispatch_width == 16)
1157 brw_ADD(p, offset(dst, 1), offset(src0, 1), suboffset(reg, 2));
1158 brw_pop_insn_state(p);
1159 }
1160
1161 /**
1162 * Change the register's data type from UD to W, doubling the strides in order
1163 * to compensate for halving the data type width.
1164 */
1165 static struct brw_reg
1166 ud_reg_to_w(struct brw_reg r)
1167 {
1168 assert(r.type == BRW_REGISTER_TYPE_UD);
1169 r.type = BRW_REGISTER_TYPE_W;
1170
1171 /* The BRW_*_STRIDE enums are defined so that incrementing the field
1172 * doubles the real stride.
1173 */
1174 if (r.hstride != 0)
1175 ++r.hstride;
1176 if (r.vstride != 0)
1177 ++r.vstride;
1178
1179 return r;
1180 }
1181
1182 void
1183 fs_generator::generate_pack_half_2x16_split(fs_inst *inst,
1184 struct brw_reg dst,
1185 struct brw_reg x,
1186 struct brw_reg y)
1187 {
1188 assert(brw->gen >= 7);
1189 assert(dst.type == BRW_REGISTER_TYPE_UD);
1190 assert(x.type == BRW_REGISTER_TYPE_F);
1191 assert(y.type == BRW_REGISTER_TYPE_F);
1192
1193 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
1194 *
1195 * Because this instruction does not have a 16-bit floating-point type,
1196 * the destination data type must be Word (W).
1197 *
1198 * The destination must be DWord-aligned and specify a horizontal stride
1199 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
1200 * each destination channel and the upper word is not modified.
1201 */
1202 struct brw_reg dst_w = ud_reg_to_w(dst);
1203
1204 /* Give each 32-bit channel of dst the form below , where "." means
1205 * unchanged.
1206 * 0x....hhhh
1207 */
1208 brw_F32TO16(p, dst_w, y);
1209
1210 /* Now the form:
1211 * 0xhhhh0000
1212 */
1213 brw_SHL(p, dst, dst, brw_imm_ud(16u));
1214
1215 /* And, finally the form of packHalf2x16's output:
1216 * 0xhhhhllll
1217 */
1218 brw_F32TO16(p, dst_w, x);
1219 }
1220
1221 void
1222 fs_generator::generate_unpack_half_2x16_split(fs_inst *inst,
1223 struct brw_reg dst,
1224 struct brw_reg src)
1225 {
1226 assert(brw->gen >= 7);
1227 assert(dst.type == BRW_REGISTER_TYPE_F);
1228 assert(src.type == BRW_REGISTER_TYPE_UD);
1229
1230 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
1231 *
1232 * Because this instruction does not have a 16-bit floating-point type,
1233 * the source data type must be Word (W). The destination type must be
1234 * F (Float).
1235 */
1236 struct brw_reg src_w = ud_reg_to_w(src);
1237
1238 /* Each channel of src has the form of unpackHalf2x16's input: 0xhhhhllll.
1239 * For the Y case, we wish to access only the upper word; therefore
1240 * a 16-bit subregister offset is needed.
1241 */
1242 assert(inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X ||
1243 inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y);
1244 if (inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y)
1245 src_w.subnr += 2;
1246
1247 brw_F16TO32(p, dst, src_w);
1248 }
1249
1250 void
1251 fs_generator::generate_shader_time_add(fs_inst *inst,
1252 struct brw_reg payload,
1253 struct brw_reg offset,
1254 struct brw_reg value)
1255 {
1256 assert(brw->gen >= 7);
1257 brw_push_insn_state(p);
1258 brw_set_mask_control(p, true);
1259
1260 assert(payload.file == BRW_GENERAL_REGISTER_FILE);
1261 struct brw_reg payload_offset = retype(brw_vec1_grf(payload.nr, 0),
1262 offset.type);
1263 struct brw_reg payload_value = retype(brw_vec1_grf(payload.nr + 1, 0),
1264 value.type);
1265
1266 assert(offset.file == BRW_IMMEDIATE_VALUE);
1267 if (value.file == BRW_GENERAL_REGISTER_FILE) {
1268 value.width = BRW_WIDTH_1;
1269 value.hstride = BRW_HORIZONTAL_STRIDE_0;
1270 value.vstride = BRW_VERTICAL_STRIDE_0;
1271 } else {
1272 assert(value.file == BRW_IMMEDIATE_VALUE);
1273 }
1274
1275 /* Trying to deal with setup of the params from the IR is crazy in the FS8
1276 * case, and we don't really care about squeezing every bit of performance
1277 * out of this path, so we just emit the MOVs from here.
1278 */
1279 brw_MOV(p, payload_offset, offset);
1280 brw_MOV(p, payload_value, value);
1281 brw_shader_time_add(p, payload,
1282 c->prog_data.base.binding_table.shader_time_start);
1283 brw_pop_insn_state(p);
1284
1285 brw_mark_surface_used(&c->prog_data.base,
1286 c->prog_data.base.binding_table.shader_time_start);
1287 }
1288
1289 void
1290 fs_generator::generate_untyped_atomic(fs_inst *inst, struct brw_reg dst,
1291 struct brw_reg atomic_op,
1292 struct brw_reg surf_index)
1293 {
1294 assert(atomic_op.file == BRW_IMMEDIATE_VALUE &&
1295 atomic_op.type == BRW_REGISTER_TYPE_UD &&
1296 surf_index.file == BRW_IMMEDIATE_VALUE &&
1297 surf_index.type == BRW_REGISTER_TYPE_UD);
1298
1299 brw_untyped_atomic(p, dst, brw_message_reg(inst->base_mrf),
1300 atomic_op.dw1.ud, surf_index.dw1.ud,
1301 inst->mlen, dispatch_width / 8);
1302
1303 brw_mark_surface_used(&c->prog_data.base, surf_index.dw1.ud);
1304 }
1305
1306 void
1307 fs_generator::generate_untyped_surface_read(fs_inst *inst, struct brw_reg dst,
1308 struct brw_reg surf_index)
1309 {
1310 assert(surf_index.file == BRW_IMMEDIATE_VALUE &&
1311 surf_index.type == BRW_REGISTER_TYPE_UD);
1312
1313 brw_untyped_surface_read(p, dst, brw_message_reg(inst->base_mrf),
1314 surf_index.dw1.ud,
1315 inst->mlen, dispatch_width / 8);
1316
1317 brw_mark_surface_used(&c->prog_data.base, surf_index.dw1.ud);
1318 }
1319
1320 void
1321 fs_generator::generate_code(exec_list *instructions, FILE *dump_file)
1322 {
1323 int last_native_insn_offset = p->next_insn_offset;
1324 const char *last_annotation_string = NULL;
1325 const void *last_annotation_ir = NULL;
1326
1327 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
1328 if (prog) {
1329 fprintf(stderr,
1330 "Native code for %s fragment shader %d (SIMD%d dispatch):\n",
1331 prog->Label ? prog->Label : "unnamed",
1332 prog->Name, dispatch_width);
1333 } else if (fp) {
1334 fprintf(stderr,
1335 "Native code for fragment program %d (SIMD%d dispatch):\n",
1336 fp->Base.Id, dispatch_width);
1337 } else {
1338 fprintf(stderr, "Native code for blorp program (SIMD%d dispatch):\n",
1339 dispatch_width);
1340 }
1341 }
1342
1343 cfg_t *cfg = NULL;
1344 if (unlikely(INTEL_DEBUG & DEBUG_WM))
1345 cfg = new(mem_ctx) cfg_t(instructions);
1346
1347 foreach_list(node, instructions) {
1348 fs_inst *inst = (fs_inst *)node;
1349 struct brw_reg src[3], dst;
1350
1351 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
1352 foreach_list(node, &cfg->block_list) {
1353 bblock_link *link = (bblock_link *)node;
1354 bblock_t *block = link->block;
1355
1356 if (block->start == inst) {
1357 fprintf(stderr, " START B%d", block->block_num);
1358 foreach_list(predecessor_node, &block->parents) {
1359 bblock_link *predecessor_link =
1360 (bblock_link *)predecessor_node;
1361 bblock_t *predecessor_block = predecessor_link->block;
1362 fprintf(stderr, " <-B%d", predecessor_block->block_num);
1363 }
1364 fprintf(stderr, "\n");
1365 }
1366 }
1367
1368 if (last_annotation_ir != inst->ir) {
1369 last_annotation_ir = inst->ir;
1370 if (last_annotation_ir) {
1371 fprintf(stderr, " ");
1372 if (prog)
1373 ((ir_instruction *)inst->ir)->fprint(stderr);
1374 else {
1375 const prog_instruction *fpi;
1376 fpi = (const prog_instruction *)inst->ir;
1377 fprintf(stderr, "%d: ",
1378 (int)(fpi - (fp ? fp->Base.Instructions : 0)));
1379 _mesa_fprint_instruction_opt(stderr,
1380 fpi,
1381 0, PROG_PRINT_DEBUG, NULL);
1382 }
1383 fprintf(stderr, "\n");
1384 }
1385 }
1386 if (last_annotation_string != inst->annotation) {
1387 last_annotation_string = inst->annotation;
1388 if (last_annotation_string)
1389 fprintf(stderr, " %s\n", last_annotation_string);
1390 }
1391 }
1392
1393 for (unsigned int i = 0; i < 3; i++) {
1394 src[i] = brw_reg_from_fs_reg(&inst->src[i]);
1395
1396 /* The accumulator result appears to get used for the
1397 * conditional modifier generation. When negating a UD
1398 * value, there is a 33rd bit generated for the sign in the
1399 * accumulator value, so now you can't check, for example,
1400 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
1401 */
1402 assert(!inst->conditional_mod ||
1403 inst->src[i].type != BRW_REGISTER_TYPE_UD ||
1404 !inst->src[i].negate);
1405 }
1406 dst = brw_reg_from_fs_reg(&inst->dst);
1407
1408 brw_set_conditionalmod(p, inst->conditional_mod);
1409 brw_set_predicate_control(p, inst->predicate);
1410 brw_set_predicate_inverse(p, inst->predicate_inverse);
1411 brw_set_flag_reg(p, 0, inst->flag_subreg);
1412 brw_set_saturate(p, inst->saturate);
1413 brw_set_mask_control(p, inst->force_writemask_all);
1414
1415 if (inst->force_uncompressed || dispatch_width == 8) {
1416 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
1417 } else if (inst->force_sechalf) {
1418 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
1419 } else {
1420 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1421 }
1422
1423 switch (inst->opcode) {
1424 case BRW_OPCODE_MOV:
1425 brw_MOV(p, dst, src[0]);
1426 break;
1427 case BRW_OPCODE_ADD:
1428 brw_ADD(p, dst, src[0], src[1]);
1429 break;
1430 case BRW_OPCODE_MUL:
1431 brw_MUL(p, dst, src[0], src[1]);
1432 break;
1433 case BRW_OPCODE_AVG:
1434 brw_AVG(p, dst, src[0], src[1]);
1435 break;
1436 case BRW_OPCODE_MACH:
1437 brw_set_acc_write_control(p, 1);
1438 brw_MACH(p, dst, src[0], src[1]);
1439 brw_set_acc_write_control(p, 0);
1440 break;
1441
1442 case BRW_OPCODE_MAD:
1443 assert(brw->gen >= 6);
1444 brw_set_access_mode(p, BRW_ALIGN_16);
1445 if (dispatch_width == 16 && !brw->is_haswell) {
1446 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
1447 brw_MAD(p, dst, src[0], src[1], src[2]);
1448 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
1449 brw_MAD(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1450 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1451 } else {
1452 brw_MAD(p, dst, src[0], src[1], src[2]);
1453 }
1454 brw_set_access_mode(p, BRW_ALIGN_1);
1455 break;
1456
1457 case BRW_OPCODE_LRP:
1458 assert(brw->gen >= 6);
1459 brw_set_access_mode(p, BRW_ALIGN_16);
1460 if (dispatch_width == 16 && !brw->is_haswell) {
1461 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
1462 brw_LRP(p, dst, src[0], src[1], src[2]);
1463 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
1464 brw_LRP(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1465 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1466 } else {
1467 brw_LRP(p, dst, src[0], src[1], src[2]);
1468 }
1469 brw_set_access_mode(p, BRW_ALIGN_1);
1470 break;
1471
1472 case BRW_OPCODE_FRC:
1473 brw_FRC(p, dst, src[0]);
1474 break;
1475 case BRW_OPCODE_RNDD:
1476 brw_RNDD(p, dst, src[0]);
1477 break;
1478 case BRW_OPCODE_RNDE:
1479 brw_RNDE(p, dst, src[0]);
1480 break;
1481 case BRW_OPCODE_RNDZ:
1482 brw_RNDZ(p, dst, src[0]);
1483 break;
1484
1485 case BRW_OPCODE_AND:
1486 brw_AND(p, dst, src[0], src[1]);
1487 break;
1488 case BRW_OPCODE_OR:
1489 brw_OR(p, dst, src[0], src[1]);
1490 break;
1491 case BRW_OPCODE_XOR:
1492 brw_XOR(p, dst, src[0], src[1]);
1493 break;
1494 case BRW_OPCODE_NOT:
1495 brw_NOT(p, dst, src[0]);
1496 break;
1497 case BRW_OPCODE_ASR:
1498 brw_ASR(p, dst, src[0], src[1]);
1499 break;
1500 case BRW_OPCODE_SHR:
1501 brw_SHR(p, dst, src[0], src[1]);
1502 break;
1503 case BRW_OPCODE_SHL:
1504 brw_SHL(p, dst, src[0], src[1]);
1505 break;
1506 case BRW_OPCODE_F32TO16:
1507 assert(brw->gen >= 7);
1508 brw_F32TO16(p, dst, src[0]);
1509 break;
1510 case BRW_OPCODE_F16TO32:
1511 assert(brw->gen >= 7);
1512 brw_F16TO32(p, dst, src[0]);
1513 break;
1514 case BRW_OPCODE_CMP:
1515 brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
1516 break;
1517 case BRW_OPCODE_SEL:
1518 brw_SEL(p, dst, src[0], src[1]);
1519 break;
1520 case BRW_OPCODE_BFREV:
1521 assert(brw->gen >= 7);
1522 /* BFREV only supports UD type for src and dst. */
1523 brw_BFREV(p, retype(dst, BRW_REGISTER_TYPE_UD),
1524 retype(src[0], BRW_REGISTER_TYPE_UD));
1525 break;
1526 case BRW_OPCODE_FBH:
1527 assert(brw->gen >= 7);
1528 /* FBH only supports UD type for dst. */
1529 brw_FBH(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1530 break;
1531 case BRW_OPCODE_FBL:
1532 assert(brw->gen >= 7);
1533 /* FBL only supports UD type for dst. */
1534 brw_FBL(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1535 break;
1536 case BRW_OPCODE_CBIT:
1537 assert(brw->gen >= 7);
1538 /* CBIT only supports UD type for dst. */
1539 brw_CBIT(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1540 break;
1541 case BRW_OPCODE_ADDC:
1542 assert(brw->gen >= 7);
1543 brw_set_acc_write_control(p, 1);
1544 brw_ADDC(p, dst, src[0], src[1]);
1545 brw_set_acc_write_control(p, 0);
1546 break;
1547 case BRW_OPCODE_SUBB:
1548 assert(brw->gen >= 7);
1549 brw_set_acc_write_control(p, 1);
1550 brw_SUBB(p, dst, src[0], src[1]);
1551 brw_set_acc_write_control(p, 0);
1552 break;
1553
1554 case BRW_OPCODE_BFE:
1555 assert(brw->gen >= 7);
1556 brw_set_access_mode(p, BRW_ALIGN_16);
1557 if (dispatch_width == 16 && !brw->is_haswell) {
1558 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
1559 brw_BFE(p, dst, src[0], src[1], src[2]);
1560 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
1561 brw_BFE(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1562 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1563 } else {
1564 brw_BFE(p, dst, src[0], src[1], src[2]);
1565 }
1566 brw_set_access_mode(p, BRW_ALIGN_1);
1567 break;
1568
1569 case BRW_OPCODE_BFI1:
1570 assert(brw->gen >= 7);
1571 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
1572 * should
1573 *
1574 * "Force BFI instructions to be executed always in SIMD8."
1575 */
1576 if (dispatch_width == 16 && brw->is_haswell) {
1577 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
1578 brw_BFI1(p, dst, src[0], src[1]);
1579 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
1580 brw_BFI1(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]));
1581 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1582 } else {
1583 brw_BFI1(p, dst, src[0], src[1]);
1584 }
1585 break;
1586 case BRW_OPCODE_BFI2:
1587 assert(brw->gen >= 7);
1588 brw_set_access_mode(p, BRW_ALIGN_16);
1589 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
1590 * should
1591 *
1592 * "Force BFI instructions to be executed always in SIMD8."
1593 *
1594 * Otherwise we would be able to emit compressed instructions like we
1595 * do for the other three-source instructions.
1596 */
1597 if (dispatch_width == 16) {
1598 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
1599 brw_BFI2(p, dst, src[0], src[1], src[2]);
1600 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
1601 brw_BFI2(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1602 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1603 } else {
1604 brw_BFI2(p, dst, src[0], src[1], src[2]);
1605 }
1606 brw_set_access_mode(p, BRW_ALIGN_1);
1607 break;
1608
1609 case BRW_OPCODE_IF:
1610 if (inst->src[0].file != BAD_FILE) {
1611 /* The instruction has an embedded compare (only allowed on gen6) */
1612 assert(brw->gen == 6);
1613 gen6_IF(p, inst->conditional_mod, src[0], src[1]);
1614 } else {
1615 brw_IF(p, dispatch_width == 16 ? BRW_EXECUTE_16 : BRW_EXECUTE_8);
1616 }
1617 break;
1618
1619 case BRW_OPCODE_ELSE:
1620 brw_ELSE(p);
1621 break;
1622 case BRW_OPCODE_ENDIF:
1623 brw_ENDIF(p);
1624 break;
1625
1626 case BRW_OPCODE_DO:
1627 brw_DO(p, BRW_EXECUTE_8);
1628 break;
1629
1630 case BRW_OPCODE_BREAK:
1631 brw_BREAK(p);
1632 brw_set_predicate_control(p, BRW_PREDICATE_NONE);
1633 break;
1634 case BRW_OPCODE_CONTINUE:
1635 /* FINISHME: We need to write the loop instruction support still. */
1636 if (brw->gen >= 6)
1637 gen6_CONT(p);
1638 else
1639 brw_CONT(p);
1640 brw_set_predicate_control(p, BRW_PREDICATE_NONE);
1641 break;
1642
1643 case BRW_OPCODE_WHILE:
1644 brw_WHILE(p);
1645 break;
1646
1647 case SHADER_OPCODE_RCP:
1648 case SHADER_OPCODE_RSQ:
1649 case SHADER_OPCODE_SQRT:
1650 case SHADER_OPCODE_EXP2:
1651 case SHADER_OPCODE_LOG2:
1652 case SHADER_OPCODE_SIN:
1653 case SHADER_OPCODE_COS:
1654 if (brw->gen >= 7) {
1655 generate_math1_gen7(inst, dst, src[0]);
1656 } else if (brw->gen == 6) {
1657 generate_math1_gen6(inst, dst, src[0]);
1658 } else if (brw->gen == 5 || brw->is_g4x) {
1659 generate_math_g45(inst, dst, src[0]);
1660 } else {
1661 generate_math_gen4(inst, dst, src[0]);
1662 }
1663 break;
1664 case SHADER_OPCODE_INT_QUOTIENT:
1665 case SHADER_OPCODE_INT_REMAINDER:
1666 case SHADER_OPCODE_POW:
1667 if (brw->gen >= 7) {
1668 generate_math2_gen7(inst, dst, src[0], src[1]);
1669 } else if (brw->gen == 6) {
1670 generate_math2_gen6(inst, dst, src[0], src[1]);
1671 } else {
1672 generate_math_gen4(inst, dst, src[0]);
1673 }
1674 break;
1675 case FS_OPCODE_PIXEL_X:
1676 generate_pixel_xy(dst, true);
1677 break;
1678 case FS_OPCODE_PIXEL_Y:
1679 generate_pixel_xy(dst, false);
1680 break;
1681 case FS_OPCODE_CINTERP:
1682 brw_MOV(p, dst, src[0]);
1683 break;
1684 case FS_OPCODE_LINTERP:
1685 generate_linterp(inst, dst, src);
1686 break;
1687 case SHADER_OPCODE_TEX:
1688 case FS_OPCODE_TXB:
1689 case SHADER_OPCODE_TXD:
1690 case SHADER_OPCODE_TXF:
1691 case SHADER_OPCODE_TXF_CMS:
1692 case SHADER_OPCODE_TXF_UMS:
1693 case SHADER_OPCODE_TXF_MCS:
1694 case SHADER_OPCODE_TXL:
1695 case SHADER_OPCODE_TXS:
1696 case SHADER_OPCODE_LOD:
1697 case SHADER_OPCODE_TG4:
1698 case SHADER_OPCODE_TG4_OFFSET:
1699 generate_tex(inst, dst, src[0]);
1700 break;
1701 case FS_OPCODE_DDX:
1702 generate_ddx(inst, dst, src[0]);
1703 break;
1704 case FS_OPCODE_DDY:
1705 /* Make sure fp->UsesDFdy flag got set (otherwise there's no
1706 * guarantee that c->key.render_to_fbo is set).
1707 */
1708 assert(fp->UsesDFdy);
1709 generate_ddy(inst, dst, src[0], c->key.render_to_fbo);
1710 break;
1711
1712 case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
1713 generate_scratch_write(inst, src[0]);
1714 break;
1715
1716 case SHADER_OPCODE_GEN4_SCRATCH_READ:
1717 generate_scratch_read(inst, dst);
1718 break;
1719
1720 case SHADER_OPCODE_GEN7_SCRATCH_READ:
1721 generate_scratch_read_gen7(inst, dst);
1722 break;
1723
1724 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
1725 generate_uniform_pull_constant_load(inst, dst, src[0], src[1]);
1726 break;
1727
1728 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7:
1729 generate_uniform_pull_constant_load_gen7(inst, dst, src[0], src[1]);
1730 break;
1731
1732 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
1733 generate_varying_pull_constant_load(inst, dst, src[0], src[1]);
1734 break;
1735
1736 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
1737 generate_varying_pull_constant_load_gen7(inst, dst, src[0], src[1]);
1738 break;
1739
1740 case FS_OPCODE_FB_WRITE:
1741 generate_fb_write(inst);
1742 break;
1743
1744 case FS_OPCODE_BLORP_FB_WRITE:
1745 generate_blorp_fb_write(inst);
1746 break;
1747
1748 case FS_OPCODE_MOV_DISPATCH_TO_FLAGS:
1749 generate_mov_dispatch_to_flags(inst);
1750 break;
1751
1752 case FS_OPCODE_DISCARD_JUMP:
1753 generate_discard_jump(inst);
1754 break;
1755
1756 case SHADER_OPCODE_SHADER_TIME_ADD:
1757 generate_shader_time_add(inst, src[0], src[1], src[2]);
1758 break;
1759
1760 case SHADER_OPCODE_UNTYPED_ATOMIC:
1761 generate_untyped_atomic(inst, dst, src[0], src[1]);
1762 break;
1763
1764 case SHADER_OPCODE_UNTYPED_SURFACE_READ:
1765 generate_untyped_surface_read(inst, dst, src[0]);
1766 break;
1767
1768 case FS_OPCODE_SET_SIMD4X2_OFFSET:
1769 generate_set_simd4x2_offset(inst, dst, src[0]);
1770 break;
1771
1772 case FS_OPCODE_SET_OMASK:
1773 generate_set_omask(inst, dst, src[0]);
1774 break;
1775
1776 case FS_OPCODE_SET_SAMPLE_ID:
1777 generate_set_sample_id(inst, dst, src[0], src[1]);
1778 break;
1779
1780 case FS_OPCODE_PACK_HALF_2x16_SPLIT:
1781 generate_pack_half_2x16_split(inst, dst, src[0], src[1]);
1782 break;
1783
1784 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X:
1785 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y:
1786 generate_unpack_half_2x16_split(inst, dst, src[0]);
1787 break;
1788
1789 case FS_OPCODE_PLACEHOLDER_HALT:
1790 /* This is the place where the final HALT needs to be inserted if
1791 * we've emitted any discards. If not, this will emit no code.
1792 */
1793 patch_discard_jumps_to_fb_writes();
1794 break;
1795
1796 default:
1797 if (inst->opcode < (int) ARRAY_SIZE(opcode_descs)) {
1798 _mesa_problem(ctx, "Unsupported opcode `%s' in FS",
1799 opcode_descs[inst->opcode].name);
1800 } else {
1801 _mesa_problem(ctx, "Unsupported opcode %d in FS", inst->opcode);
1802 }
1803 abort();
1804 }
1805
1806 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
1807 brw_dump_compile(p, stderr,
1808 last_native_insn_offset, p->next_insn_offset);
1809
1810 foreach_list(node, &cfg->block_list) {
1811 bblock_link *link = (bblock_link *)node;
1812 bblock_t *block = link->block;
1813
1814 if (block->end == inst) {
1815 fprintf(stderr, " END B%d", block->block_num);
1816 foreach_list(successor_node, &block->children) {
1817 bblock_link *successor_link =
1818 (bblock_link *)successor_node;
1819 bblock_t *successor_block = successor_link->block;
1820 fprintf(stderr, " ->B%d", successor_block->block_num);
1821 }
1822 fprintf(stderr, "\n");
1823 }
1824 }
1825 }
1826
1827 last_native_insn_offset = p->next_insn_offset;
1828 }
1829
1830 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
1831 fprintf(stderr, "\n");
1832 }
1833
1834 brw_set_uip_jip(p);
1835
1836 /* OK, while the INTEL_DEBUG=wm above is very nice for debugging FS
1837 * emit issues, it doesn't get the jump distances into the output,
1838 * which is often something we want to debug. So this is here in
1839 * case you're doing that.
1840 */
1841 if (dump_file) {
1842 brw_dump_compile(p, dump_file, 0, p->next_insn_offset);
1843 }
1844 }
1845
1846 const unsigned *
1847 fs_generator::generate_assembly(exec_list *simd8_instructions,
1848 exec_list *simd16_instructions,
1849 unsigned *assembly_size,
1850 FILE *dump_file)
1851 {
1852 assert(simd8_instructions || simd16_instructions);
1853
1854 if (simd8_instructions) {
1855 dispatch_width = 8;
1856 generate_code(simd8_instructions, dump_file);
1857 }
1858
1859 if (simd16_instructions) {
1860 /* We have to do a compaction pass now, or the one at the end of
1861 * execution will squash down where our prog_offset start needs
1862 * to be.
1863 */
1864 brw_compact_instructions(p);
1865
1866 /* align to 64 byte boundary. */
1867 while ((p->nr_insn * sizeof(struct brw_instruction)) % 64) {
1868 brw_NOP(p);
1869 }
1870
1871 /* Save off the start of this SIMD16 program */
1872 c->prog_data.prog_offset_16 = p->nr_insn * sizeof(struct brw_instruction);
1873
1874 brw_set_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1875
1876 dispatch_width = 16;
1877 generate_code(simd16_instructions, dump_file);
1878 }
1879
1880 return brw_get_program(p, assembly_size);
1881 }