projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
i965: Fix the untyped surface opcodes to deal with indirect surface access.
[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 #include "main/macros.h"
31 #include "brw_context.h"
32 #include "brw_eu.h"
33 #include "brw_fs.h"
34 #include "brw_cfg.h"
35
36 static uint32_t brw_file_from_reg(fs_reg *reg)
37 {
38 switch (reg->file) {
39 case GRF:
40 return BRW_GENERAL_REGISTER_FILE;
41 case MRF:
42 return BRW_MESSAGE_REGISTER_FILE;
43 case IMM:
44 return BRW_IMMEDIATE_VALUE;
45 default:
46 unreachable("not reached");
47 }
48 }
49
50 static struct brw_reg
51 brw_reg_from_fs_reg(fs_reg *reg)
52 {
53 struct brw_reg brw_reg;
54
55 switch (reg->file) {
56 case GRF:
57 case MRF:
58 if (reg->stride == 0) {
59 brw_reg = brw_vec1_reg(brw_file_from_reg(reg), reg->reg, 0);
60 } else if (reg->width < 8) {
61 brw_reg = brw_vec8_reg(brw_file_from_reg(reg), reg->reg, 0);
62 brw_reg = stride(brw_reg, reg->width * reg->stride,
63 reg->width, reg->stride);
64 } else {
65 /* From the Haswell PRM:
66 *
67 * VertStride must be used to cross GRF register boundaries. This
68 * rule implies that elements within a 'Width' cannot cross GRF
69 * boundaries.
70 *
71 * So, for registers with width > 8, we have to use a width of 8
72 * and trust the compression state to sort out the exec size.
73 */
74 brw_reg = brw_vec8_reg(brw_file_from_reg(reg), reg->reg, 0);
75 brw_reg = stride(brw_reg, 8 * reg->stride, 8, reg->stride);
76 }
77
78 brw_reg = retype(brw_reg, reg->type);
79 brw_reg = byte_offset(brw_reg, reg->subreg_offset);
80 break;
81 case IMM:
82 switch (reg->type) {
83 case BRW_REGISTER_TYPE_F:
84 brw_reg = brw_imm_f(reg->fixed_hw_reg.dw1.f);
85 break;
86 case BRW_REGISTER_TYPE_D:
87 brw_reg = brw_imm_d(reg->fixed_hw_reg.dw1.d);
88 break;
89 case BRW_REGISTER_TYPE_UD:
90 brw_reg = brw_imm_ud(reg->fixed_hw_reg.dw1.ud);
91 break;
92 case BRW_REGISTER_TYPE_W:
93 brw_reg = brw_imm_w(reg->fixed_hw_reg.dw1.d);
94 break;
95 case BRW_REGISTER_TYPE_UW:
96 brw_reg = brw_imm_uw(reg->fixed_hw_reg.dw1.ud);
97 break;
98 case BRW_REGISTER_TYPE_VF:
99 brw_reg = brw_imm_vf(reg->fixed_hw_reg.dw1.ud);
100 break;
101 default:
102 unreachable("not reached");
103 }
104 break;
105 case HW_REG:
106 assert(reg->type == reg->fixed_hw_reg.type);
107 brw_reg = reg->fixed_hw_reg;
108 break;
109 case BAD_FILE:
110 /* Probably unused. */
111 brw_reg = brw_null_reg();
112 break;
113 default:
114 unreachable("not reached");
115 }
116 if (reg->abs)
117 brw_reg = brw_abs(brw_reg);
118 if (reg->negate)
119 brw_reg = negate(brw_reg);
120
121 return brw_reg;
122 }
123
124 fs_generator::fs_generator(struct brw_context *brw,
125 void *mem_ctx,
126 const void *key,
127 struct brw_stage_prog_data *prog_data,
128 struct gl_program *prog,
129 unsigned promoted_constants,
130 bool runtime_check_aads_emit,
131 const char *stage_abbrev)
132
133 : brw(brw), devinfo(brw->intelScreen->devinfo), key(key),
134 prog_data(prog_data),
135 prog(prog), promoted_constants(promoted_constants),
136 runtime_check_aads_emit(runtime_check_aads_emit), debug_flag(false),
137 stage_abbrev(stage_abbrev), mem_ctx(mem_ctx)
138 {
139 p = rzalloc(mem_ctx, struct brw_codegen);
140 brw_init_codegen(devinfo, p, mem_ctx);
141 }
142
143 fs_generator::~fs_generator()
144 {
145 }
146
147 class ip_record : public exec_node {
148 public:
149 DECLARE_RALLOC_CXX_OPERATORS(ip_record)
150
151 ip_record(int ip)
152 {
153 this->ip = ip;
154 }
155
156 int ip;
157 };
158
159 bool
160 fs_generator::patch_discard_jumps_to_fb_writes()
161 {
162 if (devinfo->gen < 6 || this->discard_halt_patches.is_empty())
163 return false;
164
165 int scale = brw_jump_scale(p->devinfo);
166
167 /* There is a somewhat strange undocumented requirement of using
168 * HALT, according to the simulator. If some channel has HALTed to
169 * a particular UIP, then by the end of the program, every channel
170 * must have HALTed to that UIP. Furthermore, the tracking is a
171 * stack, so you can't do the final halt of a UIP after starting
172 * halting to a new UIP.
173 *
174 * Symptoms of not emitting this instruction on actual hardware
175 * included GPU hangs and sparkly rendering on the piglit discard
176 * tests.
177 */
178 brw_inst *last_halt = gen6_HALT(p);
179 brw_inst_set_uip(p->devinfo, last_halt, 1 * scale);
180 brw_inst_set_jip(p->devinfo, last_halt, 1 * scale);
181
182 int ip = p->nr_insn;
183
184 foreach_in_list(ip_record, patch_ip, &discard_halt_patches) {
185 brw_inst *patch = &p->store[patch_ip->ip];
186
187 assert(brw_inst_opcode(p->devinfo, patch) == BRW_OPCODE_HALT);
188 /* HALT takes a half-instruction distance from the pre-incremented IP. */
189 brw_inst_set_uip(p->devinfo, patch, (ip - patch_ip->ip) * scale);
190 }
191
192 this->discard_halt_patches.make_empty();
193 return true;
194 }
195
196 void
197 fs_generator::fire_fb_write(fs_inst *inst,
198 struct brw_reg payload,
199 struct brw_reg implied_header,
200 GLuint nr)
201 {
202 uint32_t msg_control;
203
204 brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data;
205
206 if (devinfo->gen < 6) {
207 brw_push_insn_state(p);
208 brw_set_default_exec_size(p, BRW_EXECUTE_8);
209 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
210 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
211 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
212 brw_MOV(p, offset(payload, 1), brw_vec8_grf(1, 0));
213 brw_pop_insn_state(p);
214 }
215
216 if (inst->opcode == FS_OPCODE_REP_FB_WRITE)
217 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED;
218 else if (prog_data->dual_src_blend) {
219 if (dispatch_width == 8 || !inst->eot)
220 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN01;
221 else
222 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN23;
223 } else if (dispatch_width == 16)
224 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE;
225 else
226 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01;
227
228 uint32_t surf_index =
229 prog_data->binding_table.render_target_start + inst->target;
230
231 bool last_render_target = inst->eot ||
232 (prog_data->dual_src_blend && dispatch_width == 16);
233
234
235 brw_fb_WRITE(p,
236 dispatch_width,
237 payload,
238 implied_header,
239 msg_control,
240 surf_index,
241 nr,
242 0,
243 inst->eot,
244 last_render_target,
245 inst->header_present);
246
247 brw_mark_surface_used(&prog_data->base, surf_index);
248 }
249
250 void
251 fs_generator::generate_fb_write(fs_inst *inst, struct brw_reg payload)
252 {
253 brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data;
254 const brw_wm_prog_key * const key = (brw_wm_prog_key * const) this->key;
255 struct brw_reg implied_header;
256
257 if (devinfo->gen < 8 && !devinfo->is_haswell) {
258 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
259 }
260
261 if (inst->base_mrf >= 0)
262 payload = brw_message_reg(inst->base_mrf);
263
264 /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
265 * move, here's g1.
266 */
267 if (inst->header_present) {
268 brw_push_insn_state(p);
269 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
270 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
271 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
272 brw_set_default_flag_reg(p, 0, 0);
273
274 /* On HSW, the GPU will use the predicate on SENDC, unless the header is
275 * present.
276 */
277 if (prog_data->uses_kill) {
278 struct brw_reg pixel_mask;
279
280 if (devinfo->gen >= 6)
281 pixel_mask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
282 else
283 pixel_mask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
284
285 brw_MOV(p, pixel_mask, brw_flag_reg(0, 1));
286 }
287
288 if (devinfo->gen >= 6) {
289 brw_push_insn_state(p);
290 brw_set_default_exec_size(p, BRW_EXECUTE_16);
291 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
292 brw_MOV(p,
293 retype(payload, BRW_REGISTER_TYPE_UD),
294 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
295 brw_pop_insn_state(p);
296
297 if (inst->target > 0 && key->replicate_alpha) {
298 /* Set "Source0 Alpha Present to RenderTarget" bit in message
299 * header.
300 */
301 brw_OR(p,
302 vec1(retype(payload, BRW_REGISTER_TYPE_UD)),
303 vec1(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD)),
304 brw_imm_ud(0x1 << 11));
305 }
306
307 if (inst->target > 0) {
308 /* Set the render target index for choosing BLEND_STATE. */
309 brw_MOV(p, retype(vec1(suboffset(payload, 2)),
310 BRW_REGISTER_TYPE_UD),
311 brw_imm_ud(inst->target));
312 }
313
314 implied_header = brw_null_reg();
315 } else {
316 implied_header = retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW);
317 }
318
319 brw_pop_insn_state(p);
320 } else {
321 implied_header = brw_null_reg();
322 }
323
324 if (!runtime_check_aads_emit) {
325 fire_fb_write(inst, payload, implied_header, inst->mlen);
326 } else {
327 /* This can only happen in gen < 6 */
328 assert(devinfo->gen < 6);
329
330 struct brw_reg v1_null_ud = vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD));
331
332 /* Check runtime bit to detect if we have to send AA data or not */
333 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
334 brw_AND(p,
335 v1_null_ud,
336 retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD),
337 brw_imm_ud(1<<26));
338 brw_inst_set_cond_modifier(p->devinfo, brw_last_inst, BRW_CONDITIONAL_NZ);
339
340 int jmp = brw_JMPI(p, brw_imm_ud(0), BRW_PREDICATE_NORMAL) - p->store;
341 brw_inst_set_exec_size(p->devinfo, brw_last_inst, BRW_EXECUTE_1);
342 {
343 /* Don't send AA data */
344 fire_fb_write(inst, offset(payload, 1), implied_header, inst->mlen-1);
345 }
346 brw_land_fwd_jump(p, jmp);
347 fire_fb_write(inst, payload, implied_header, inst->mlen);
348 }
349 }
350
351 void
352 fs_generator::generate_urb_write(fs_inst *inst, struct brw_reg payload)
353 {
354 brw_inst *insn;
355
356 insn = brw_next_insn(p, BRW_OPCODE_SEND);
357
358 brw_set_dest(p, insn, brw_null_reg());
359 brw_set_src0(p, insn, payload);
360 brw_set_src1(p, insn, brw_imm_d(0));
361
362 brw_inst_set_sfid(p->devinfo, insn, BRW_SFID_URB);
363 brw_inst_set_urb_opcode(p->devinfo, insn, GEN8_URB_OPCODE_SIMD8_WRITE);
364
365 brw_inst_set_mlen(p->devinfo, insn, inst->mlen);
366 brw_inst_set_rlen(p->devinfo, insn, 0);
367 brw_inst_set_eot(p->devinfo, insn, inst->eot);
368 brw_inst_set_header_present(p->devinfo, insn, true);
369 brw_inst_set_urb_global_offset(p->devinfo, insn, inst->offset);
370 }
371
372 void
373 fs_generator::generate_cs_terminate(fs_inst *inst, struct brw_reg payload)
374 {
375 struct brw_inst *insn;
376
377 insn = brw_next_insn(p, BRW_OPCODE_SEND);
378
379 brw_set_dest(p, insn, brw_null_reg());
380 brw_set_src0(p, insn, payload);
381 brw_set_src1(p, insn, brw_imm_d(0));
382
383 /* Terminate a compute shader by sending a message to the thread spawner.
384 */
385 brw_inst_set_sfid(devinfo, insn, BRW_SFID_THREAD_SPAWNER);
386 brw_inst_set_mlen(devinfo, insn, 1);
387 brw_inst_set_rlen(devinfo, insn, 0);
388 brw_inst_set_eot(devinfo, insn, inst->eot);
389 brw_inst_set_header_present(devinfo, insn, false);
390
391 brw_inst_set_ts_opcode(devinfo, insn, 0); /* Dereference resource */
392 brw_inst_set_ts_request_type(devinfo, insn, 0); /* Root thread */
393
394 /* Note that even though the thread has a URB resource associated with it,
395 * we set the "do not dereference URB" bit, because the URB resource is
396 * managed by the fixed-function unit, so it will free it automatically.
397 */
398 brw_inst_set_ts_resource_select(devinfo, insn, 1); /* Do not dereference URB */
399
400 brw_inst_set_mask_control(devinfo, insn, BRW_MASK_DISABLE);
401 }
402
403 void
404 fs_generator::generate_blorp_fb_write(fs_inst *inst)
405 {
406 brw_fb_WRITE(p,
407 16 /* dispatch_width */,
408 brw_message_reg(inst->base_mrf),
409 brw_reg_from_fs_reg(&inst->src[0]),
410 BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE,
411 inst->target,
412 inst->mlen,
413 0,
414 true,
415 true,
416 inst->header_present);
417 }
418
419 void
420 fs_generator::generate_linterp(fs_inst *inst,
421 struct brw_reg dst, struct brw_reg *src)
422 {
423 /* PLN reads:
424 * / in SIMD16 \
425 * -----------------------------------
426 * | src1+0 | src1+1 | src1+2 | src1+3 |
427 * |-----------------------------------|
428 * |(x0, x1)|(y0, y1)|(x2, x3)|(y2, y3)|
429 * -----------------------------------
430 *
431 * but for the LINE/MAC pair, the LINE reads Xs and the MAC reads Ys:
432 *
433 * -----------------------------------
434 * | src1+0 | src1+1 | src1+2 | src1+3 |
435 * |-----------------------------------|
436 * |(x0, x1)|(y0, y1)| | | in SIMD8
437 * |-----------------------------------|
438 * |(x0, x1)|(x2, x3)|(y0, y1)|(y2, y3)| in SIMD16
439 * -----------------------------------
440 *
441 * See also: emit_interpolation_setup_gen4().
442 */
443 struct brw_reg delta_x = src[0];
444 struct brw_reg delta_y = offset(src[0], dispatch_width / 8);
445 struct brw_reg interp = src[1];
446
447 if (devinfo->has_pln &&
448 (devinfo->gen >= 7 || (delta_x.nr & 1) == 0)) {
449 brw_PLN(p, dst, interp, delta_x);
450 } else {
451 brw_LINE(p, brw_null_reg(), interp, delta_x);
452 brw_MAC(p, dst, suboffset(interp, 1), delta_y);
453 }
454 }
455
456 void
457 fs_generator::generate_math_gen6(fs_inst *inst,
458 struct brw_reg dst,
459 struct brw_reg src0,
460 struct brw_reg src1)
461 {
462 int op = brw_math_function(inst->opcode);
463 bool binop = src1.file != BRW_ARCHITECTURE_REGISTER_FILE;
464
465 if (dispatch_width == 8) {
466 gen6_math(p, dst, op, src0, src1);
467 } else if (dispatch_width == 16) {
468 brw_push_insn_state(p);
469 brw_set_default_exec_size(p, BRW_EXECUTE_8);
470 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
471 gen6_math(p, firsthalf(dst), op, firsthalf(src0), firsthalf(src1));
472 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
473 gen6_math(p, sechalf(dst), op, sechalf(src0),
474 binop ? sechalf(src1) : brw_null_reg());
475 brw_pop_insn_state(p);
476 }
477 }
478
479 void
480 fs_generator::generate_math_gen4(fs_inst *inst,
481 struct brw_reg dst,
482 struct brw_reg src)
483 {
484 int op = brw_math_function(inst->opcode);
485
486 assert(inst->mlen >= 1);
487
488 if (dispatch_width == 8) {
489 gen4_math(p, dst,
490 op,
491 inst->base_mrf, src,
492 BRW_MATH_PRECISION_FULL);
493 } else if (dispatch_width == 16) {
494 brw_set_default_exec_size(p, BRW_EXECUTE_8);
495 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
496 gen4_math(p, firsthalf(dst),
497 op,
498 inst->base_mrf, firsthalf(src),
499 BRW_MATH_PRECISION_FULL);
500 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
501 gen4_math(p, sechalf(dst),
502 op,
503 inst->base_mrf + 1, sechalf(src),
504 BRW_MATH_PRECISION_FULL);
505
506 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
507 }
508 }
509
510 void
511 fs_generator::generate_math_g45(fs_inst *inst,
512 struct brw_reg dst,
513 struct brw_reg src)
514 {
515 if (inst->opcode == SHADER_OPCODE_POW ||
516 inst->opcode == SHADER_OPCODE_INT_QUOTIENT ||
517 inst->opcode == SHADER_OPCODE_INT_REMAINDER) {
518 generate_math_gen4(inst, dst, src);
519 return;
520 }
521
522 int op = brw_math_function(inst->opcode);
523
524 assert(inst->mlen >= 1);
525
526 gen4_math(p, dst,
527 op,
528 inst->base_mrf, src,
529 BRW_MATH_PRECISION_FULL);
530 }
531
532 void
533 fs_generator::generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src,
534 struct brw_reg sampler_index)
535 {
536 int msg_type = -1;
537 int rlen = 4;
538 uint32_t simd_mode;
539 uint32_t return_format;
540 bool is_combined_send = inst->eot;
541
542 switch (dst.type) {
543 case BRW_REGISTER_TYPE_D:
544 return_format = BRW_SAMPLER_RETURN_FORMAT_SINT32;
545 break;
546 case BRW_REGISTER_TYPE_UD:
547 return_format = BRW_SAMPLER_RETURN_FORMAT_UINT32;
548 break;
549 default:
550 return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
551 break;
552 }
553
554 switch (inst->exec_size) {
555 case 8:
556 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
557 break;
558 case 16:
559 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
560 break;
561 default:
562 unreachable("Invalid width for texture instruction");
563 }
564
565 if (devinfo->gen >= 5) {
566 switch (inst->opcode) {
567 case SHADER_OPCODE_TEX:
568 if (inst->shadow_compare) {
569 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE;
570 } else {
571 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE;
572 }
573 break;
574 case FS_OPCODE_TXB:
575 if (inst->shadow_compare) {
576 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE;
577 } else {
578 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS;
579 }
580 break;
581 case SHADER_OPCODE_TXL:
582 if (inst->shadow_compare) {
583 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
584 } else {
585 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
586 }
587 break;
588 case SHADER_OPCODE_TXS:
589 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
590 break;
591 case SHADER_OPCODE_TXD:
592 if (inst->shadow_compare) {
593 /* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
594 assert(devinfo->gen >= 8 || devinfo->is_haswell);
595 msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE;
596 } else {
597 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
598 }
599 break;
600 case SHADER_OPCODE_TXF:
601 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
602 break;
603 case SHADER_OPCODE_TXF_CMS:
604 if (devinfo->gen >= 7)
605 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
606 else
607 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
608 break;
609 case SHADER_OPCODE_TXF_UMS:
610 assert(devinfo->gen >= 7);
611 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS;
612 break;
613 case SHADER_OPCODE_TXF_MCS:
614 assert(devinfo->gen >= 7);
615 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS;
616 break;
617 case SHADER_OPCODE_LOD:
618 msg_type = GEN5_SAMPLER_MESSAGE_LOD;
619 break;
620 case SHADER_OPCODE_TG4:
621 if (inst->shadow_compare) {
622 assert(devinfo->gen >= 7);
623 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C;
624 } else {
625 assert(devinfo->gen >= 6);
626 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4;
627 }
628 break;
629 case SHADER_OPCODE_TG4_OFFSET:
630 assert(devinfo->gen >= 7);
631 if (inst->shadow_compare) {
632 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C;
633 } else {
634 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO;
635 }
636 break;
637 default:
638 unreachable("not reached");
639 }
640 } else {
641 switch (inst->opcode) {
642 case SHADER_OPCODE_TEX:
643 /* Note that G45 and older determines shadow compare and dispatch width
644 * from message length for most messages.
645 */
646 if (dispatch_width == 8) {
647 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE;
648 if (inst->shadow_compare) {
649 assert(inst->mlen == 6);
650 } else {
651 assert(inst->mlen <= 4);
652 }
653 } else {
654 if (inst->shadow_compare) {
655 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE;
656 assert(inst->mlen == 9);
657 } else {
658 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE;
659 assert(inst->mlen <= 7 && inst->mlen % 2 == 1);
660 }
661 }
662 break;
663 case FS_OPCODE_TXB:
664 if (inst->shadow_compare) {
665 assert(dispatch_width == 8);
666 assert(inst->mlen == 6);
667 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_BIAS_COMPARE;
668 } else {
669 assert(inst->mlen == 9);
670 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS;
671 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
672 }
673 break;
674 case SHADER_OPCODE_TXL:
675 if (inst->shadow_compare) {
676 assert(dispatch_width == 8);
677 assert(inst->mlen == 6);
678 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_LOD_COMPARE;
679 } else {
680 assert(inst->mlen == 9);
681 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_LOD;
682 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
683 }
684 break;
685 case SHADER_OPCODE_TXD:
686 /* There is no sample_d_c message; comparisons are done manually */
687 assert(dispatch_width == 8);
688 assert(inst->mlen == 7 || inst->mlen == 10);
689 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_GRADIENTS;
690 break;
691 case SHADER_OPCODE_TXF:
692 assert(inst->mlen <= 9 && inst->mlen % 2 == 1);
693 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_LD;
694 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
695 break;
696 case SHADER_OPCODE_TXS:
697 assert(inst->mlen == 3);
698 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_RESINFO;
699 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
700 break;
701 default:
702 unreachable("not reached");
703 }
704 }
705 assert(msg_type != -1);
706
707 if (simd_mode == BRW_SAMPLER_SIMD_MODE_SIMD16) {
708 rlen = 8;
709 dst = vec16(dst);
710 }
711
712 if (is_combined_send) {
713 assert(devinfo->gen >= 9 || devinfo->is_cherryview);
714 rlen = 0;
715 }
716
717 assert(devinfo->gen < 7 || !inst->header_present ||
718 src.file == BRW_GENERAL_REGISTER_FILE);
719
720 assert(sampler_index.type == BRW_REGISTER_TYPE_UD);
721
722 /* Load the message header if present. If there's a texture offset,
723 * we need to set it up explicitly and load the offset bitfield.
724 * Otherwise, we can use an implied move from g0 to the first message reg.
725 */
726 if (inst->header_present) {
727 if (devinfo->gen < 6 && !inst->offset) {
728 /* Set up an implied move from g0 to the MRF. */
729 src = retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW);
730 } else {
731 struct brw_reg header_reg;
732
733 if (devinfo->gen >= 7) {
734 header_reg = src;
735 } else {
736 assert(inst->base_mrf != -1);
737 header_reg = brw_message_reg(inst->base_mrf);
738 }
739
740 brw_push_insn_state(p);
741 brw_set_default_exec_size(p, BRW_EXECUTE_8);
742 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
743 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
744 /* Explicitly set up the message header by copying g0 to the MRF. */
745 brw_MOV(p, header_reg, brw_vec8_grf(0, 0));
746
747 if (inst->offset) {
748 /* Set the offset bits in DWord 2. */
749 brw_MOV(p, get_element_ud(header_reg, 2),
750 brw_imm_ud(inst->offset));
751 }
752
753 brw_adjust_sampler_state_pointer(p, header_reg, sampler_index);
754 brw_pop_insn_state(p);
755 }
756 }
757
758 uint32_t base_binding_table_index = (inst->opcode == SHADER_OPCODE_TG4 ||
759 inst->opcode == SHADER_OPCODE_TG4_OFFSET)
760 ? prog_data->binding_table.gather_texture_start
761 : prog_data->binding_table.texture_start;
762
763 if (sampler_index.file == BRW_IMMEDIATE_VALUE) {
764 uint32_t sampler = sampler_index.dw1.ud;
765
766 brw_SAMPLE(p,
767 retype(dst, BRW_REGISTER_TYPE_UW),
768 inst->base_mrf,
769 src,
770 sampler + base_binding_table_index,
771 sampler % 16,
772 msg_type,
773 rlen,
774 inst->mlen,
775 inst->header_present,
776 simd_mode,
777 return_format);
778
779 brw_mark_surface_used(prog_data, sampler + base_binding_table_index);
780 } else {
781 /* Non-const sampler index */
782 /* Note: this clobbers `dst` as a temporary before emitting the send */
783
784 struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
785 struct brw_reg temp = vec1(retype(dst, BRW_REGISTER_TYPE_UD));
786
787 struct brw_reg sampler_reg = vec1(retype(sampler_index, BRW_REGISTER_TYPE_UD));
788
789 brw_push_insn_state(p);
790 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
791 brw_set_default_access_mode(p, BRW_ALIGN_1);
792
793 /* Some care required: `sampler` and `temp` may alias:
794 * addr = sampler & 0xff
795 * temp = (sampler << 8) & 0xf00
796 * addr = addr | temp
797 */
798 brw_ADD(p, addr, sampler_reg, brw_imm_ud(base_binding_table_index));
799 brw_SHL(p, temp, sampler_reg, brw_imm_ud(8u));
800 brw_AND(p, temp, temp, brw_imm_ud(0x0f00));
801 brw_AND(p, addr, addr, brw_imm_ud(0x0ff));
802 brw_OR(p, addr, addr, temp);
803
804 brw_pop_insn_state(p);
805
806 /* dst = send(offset, a0.0 | <descriptor>) */
807 brw_inst *insn = brw_send_indirect_message(
808 p, BRW_SFID_SAMPLER, dst, src, addr);
809 brw_set_sampler_message(p, insn,
810 0 /* surface */,
811 0 /* sampler */,
812 msg_type,
813 rlen,
814 inst->mlen /* mlen */,
815 inst->header_present /* header */,
816 simd_mode,
817 return_format);
818
819 /* visitor knows more than we do about the surface limit required,
820 * so has already done marking.
821 */
822 }
823
824 if (is_combined_send) {
825 brw_inst_set_eot(p->devinfo, brw_last_inst, true);
826 brw_inst_set_opcode(p->devinfo, brw_last_inst, BRW_OPCODE_SENDC);
827 }
828 }
829
830
831 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
832 * looking like:
833 *
834 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
835 *
836 * Ideally, we want to produce:
837 *
838 * DDX DDY
839 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
840 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
841 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
842 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
843 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
844 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
845 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
846 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
847 *
848 * and add another set of two more subspans if in 16-pixel dispatch mode.
849 *
850 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
851 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
852 * pair. But the ideal approximation may impose a huge performance cost on
853 * sample_d. On at least Haswell, sample_d instruction does some
854 * optimizations if the same LOD is used for all pixels in the subspan.
855 *
856 * For DDY, we need to use ALIGN16 mode since it's capable of doing the
857 * appropriate swizzling.
858 */
859 void
860 fs_generator::generate_ddx(enum opcode opcode,
861 struct brw_reg dst, struct brw_reg src)
862 {
863 unsigned vstride, width;
864
865 if (opcode == FS_OPCODE_DDX_FINE) {
866 /* produce accurate derivatives */
867 vstride = BRW_VERTICAL_STRIDE_2;
868 width = BRW_WIDTH_2;
869 } else {
870 /* replicate the derivative at the top-left pixel to other pixels */
871 vstride = BRW_VERTICAL_STRIDE_4;
872 width = BRW_WIDTH_4;
873 }
874
875 struct brw_reg src0 = brw_reg(src.file, src.nr, 1,
876 src.negate, src.abs,
877 BRW_REGISTER_TYPE_F,
878 vstride,
879 width,
880 BRW_HORIZONTAL_STRIDE_0,
881 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
882 struct brw_reg src1 = brw_reg(src.file, src.nr, 0,
883 src.negate, src.abs,
884 BRW_REGISTER_TYPE_F,
885 vstride,
886 width,
887 BRW_HORIZONTAL_STRIDE_0,
888 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
889 brw_ADD(p, dst, src0, negate(src1));
890 }
891
892 /* The negate_value boolean is used to negate the derivative computation for
893 * FBOs, since they place the origin at the upper left instead of the lower
894 * left.
895 */
896 void
897 fs_generator::generate_ddy(enum opcode opcode,
898 struct brw_reg dst, struct brw_reg src,
899 bool negate_value)
900 {
901 if (opcode == FS_OPCODE_DDY_FINE) {
902 /* From the Ivy Bridge PRM, volume 4 part 3, section 3.3.9 (Register
903 * Region Restrictions):
904 *
905 * In Align16 access mode, SIMD16 is not allowed for DW operations
906 * and SIMD8 is not allowed for DF operations.
907 *
908 * In this context, "DW operations" means "operations acting on 32-bit
909 * values", so it includes operations on floats.
910 *
911 * Gen4 has a similar restriction. From the i965 PRM, section 11.5.3
912 * (Instruction Compression -> Rules and Restrictions):
913 *
914 * A compressed instruction must be in Align1 access mode. Align16
915 * mode instructions cannot be compressed.
916 *
917 * Similar text exists in the g45 PRM.
918 *
919 * On these platforms, if we're building a SIMD16 shader, we need to
920 * manually unroll to a pair of SIMD8 instructions.
921 */
922 bool unroll_to_simd8 =
923 (dispatch_width == 16 &&
924 (devinfo->gen == 4 || (devinfo->gen == 7 && !devinfo->is_haswell)));
925
926 /* produce accurate derivatives */
927 struct brw_reg src0 = brw_reg(src.file, src.nr, 0,
928 src.negate, src.abs,
929 BRW_REGISTER_TYPE_F,
930 BRW_VERTICAL_STRIDE_4,
931 BRW_WIDTH_4,
932 BRW_HORIZONTAL_STRIDE_1,
933 BRW_SWIZZLE_XYXY, WRITEMASK_XYZW);
934 struct brw_reg src1 = brw_reg(src.file, src.nr, 0,
935 src.negate, src.abs,
936 BRW_REGISTER_TYPE_F,
937 BRW_VERTICAL_STRIDE_4,
938 BRW_WIDTH_4,
939 BRW_HORIZONTAL_STRIDE_1,
940 BRW_SWIZZLE_ZWZW, WRITEMASK_XYZW);
941 brw_push_insn_state(p);
942 brw_set_default_access_mode(p, BRW_ALIGN_16);
943 if (unroll_to_simd8) {
944 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
945 if (negate_value) {
946 brw_ADD(p, firsthalf(dst), firsthalf(src1), negate(firsthalf(src0)));
947 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
948 brw_ADD(p, sechalf(dst), sechalf(src1), negate(sechalf(src0)));
949 } else {
950 brw_ADD(p, firsthalf(dst), firsthalf(src0), negate(firsthalf(src1)));
951 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
952 brw_ADD(p, sechalf(dst), sechalf(src0), negate(sechalf(src1)));
953 }
954 } else {
955 if (negate_value)
956 brw_ADD(p, dst, src1, negate(src0));
957 else
958 brw_ADD(p, dst, src0, negate(src1));
959 }
960 brw_pop_insn_state(p);
961 } else {
962 /* replicate the derivative at the top-left pixel to other pixels */
963 struct brw_reg src0 = brw_reg(src.file, src.nr, 0,
964 src.negate, src.abs,
965 BRW_REGISTER_TYPE_F,
966 BRW_VERTICAL_STRIDE_4,
967 BRW_WIDTH_4,
968 BRW_HORIZONTAL_STRIDE_0,
969 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
970 struct brw_reg src1 = brw_reg(src.file, src.nr, 2,
971 src.negate, src.abs,
972 BRW_REGISTER_TYPE_F,
973 BRW_VERTICAL_STRIDE_4,
974 BRW_WIDTH_4,
975 BRW_HORIZONTAL_STRIDE_0,
976 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
977 if (negate_value)
978 brw_ADD(p, dst, src1, negate(src0));
979 else
980 brw_ADD(p, dst, src0, negate(src1));
981 }
982 }
983
984 void
985 fs_generator::generate_discard_jump(fs_inst *inst)
986 {
987 assert(devinfo->gen >= 6);
988
989 /* This HALT will be patched up at FB write time to point UIP at the end of
990 * the program, and at brw_uip_jip() JIP will be set to the end of the
991 * current block (or the program).
992 */
993 this->discard_halt_patches.push_tail(new(mem_ctx) ip_record(p->nr_insn));
994
995 brw_push_insn_state(p);
996 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
997 gen6_HALT(p);
998 brw_pop_insn_state(p);
999 }
1000
1001 void
1002 fs_generator::generate_scratch_write(fs_inst *inst, struct brw_reg src)
1003 {
1004 assert(inst->mlen != 0);
1005
1006 brw_MOV(p,
1007 brw_uvec_mrf(inst->exec_size, (inst->base_mrf + 1), 0),
1008 retype(src, BRW_REGISTER_TYPE_UD));
1009 brw_oword_block_write_scratch(p, brw_message_reg(inst->base_mrf),
1010 inst->exec_size / 8, inst->offset);
1011 }
1012
1013 void
1014 fs_generator::generate_scratch_read(fs_inst *inst, struct brw_reg dst)
1015 {
1016 assert(inst->mlen != 0);
1017
1018 brw_oword_block_read_scratch(p, dst, brw_message_reg(inst->base_mrf),
1019 inst->exec_size / 8, inst->offset);
1020 }
1021
1022 void
1023 fs_generator::generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst)
1024 {
1025 gen7_block_read_scratch(p, dst, inst->exec_size / 8, inst->offset);
1026 }
1027
1028 void
1029 fs_generator::generate_uniform_pull_constant_load(fs_inst *inst,
1030 struct brw_reg dst,
1031 struct brw_reg index,
1032 struct brw_reg offset)
1033 {
1034 assert(inst->mlen != 0);
1035
1036 assert(index.file == BRW_IMMEDIATE_VALUE &&
1037 index.type == BRW_REGISTER_TYPE_UD);
1038 uint32_t surf_index = index.dw1.ud;
1039
1040 assert(offset.file == BRW_IMMEDIATE_VALUE &&
1041 offset.type == BRW_REGISTER_TYPE_UD);
1042 uint32_t read_offset = offset.dw1.ud;
1043
1044 brw_oword_block_read(p, dst, brw_message_reg(inst->base_mrf),
1045 read_offset, surf_index);
1046
1047 brw_mark_surface_used(prog_data, surf_index);
1048 }
1049
1050 void
1051 fs_generator::generate_uniform_pull_constant_load_gen7(fs_inst *inst,
1052 struct brw_reg dst,
1053 struct brw_reg index,
1054 struct brw_reg offset)
1055 {
1056 assert(inst->mlen == 0);
1057 assert(index.type == BRW_REGISTER_TYPE_UD);
1058
1059 assert(offset.file == BRW_GENERAL_REGISTER_FILE);
1060 /* Reference just the dword we need, to avoid angering validate_reg(). */
1061 offset = brw_vec1_grf(offset.nr, 0);
1062
1063 /* We use the SIMD4x2 mode because we want to end up with 4 components in
1064 * the destination loaded consecutively from the same offset (which appears
1065 * in the first component, and the rest are ignored).
1066 */
1067 dst.width = BRW_WIDTH_4;
1068
1069 struct brw_reg src = offset;
1070 bool header_present = false;
1071 int mlen = 1;
1072
1073 if (devinfo->gen >= 9) {
1074 /* Skylake requires a message header in order to use SIMD4x2 mode. */
1075 src = retype(brw_vec4_grf(offset.nr - 1, 0), BRW_REGISTER_TYPE_UD);
1076 mlen = 2;
1077 header_present = true;
1078
1079 brw_push_insn_state(p);
1080 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1081 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1082 brw_MOV(p, vec8(src), retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
1083 brw_set_default_access_mode(p, BRW_ALIGN_1);
1084
1085 brw_MOV(p, get_element_ud(src, 2),
1086 brw_imm_ud(GEN9_SAMPLER_SIMD_MODE_EXTENSION_SIMD4X2));
1087 brw_pop_insn_state(p);
1088 }
1089
1090 if (index.file == BRW_IMMEDIATE_VALUE) {
1091
1092 uint32_t surf_index = index.dw1.ud;
1093
1094 brw_push_insn_state(p);
1095 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1096 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1097 brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
1098 brw_pop_insn_state(p);
1099
1100 brw_set_dest(p, send, dst);
1101 brw_set_src0(p, send, src);
1102 brw_set_sampler_message(p, send,
1103 surf_index,
1104 0, /* LD message ignores sampler unit */
1105 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1106 1, /* rlen */
1107 mlen,
1108 header_present,
1109 BRW_SAMPLER_SIMD_MODE_SIMD4X2,
1110 0);
1111
1112 brw_mark_surface_used(prog_data, surf_index);
1113
1114 } else {
1115
1116 struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
1117
1118 brw_push_insn_state(p);
1119 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1120 brw_set_default_access_mode(p, BRW_ALIGN_1);
1121
1122 /* a0.0 = surf_index & 0xff */
1123 brw_inst *insn_and = brw_next_insn(p, BRW_OPCODE_AND);
1124 brw_inst_set_exec_size(p->devinfo, insn_and, BRW_EXECUTE_1);
1125 brw_set_dest(p, insn_and, addr);
1126 brw_set_src0(p, insn_and, vec1(retype(index, BRW_REGISTER_TYPE_UD)));
1127 brw_set_src1(p, insn_and, brw_imm_ud(0x0ff));
1128
1129 /* dst = send(payload, a0.0 | <descriptor>) */
1130 brw_inst *insn = brw_send_indirect_message(
1131 p, BRW_SFID_SAMPLER, dst, src, addr);
1132 brw_set_sampler_message(p, insn,
1133 0,
1134 0, /* LD message ignores sampler unit */
1135 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1136 1, /* rlen */
1137 mlen,
1138 header_present,
1139 BRW_SAMPLER_SIMD_MODE_SIMD4X2,
1140 0);
1141
1142 brw_pop_insn_state(p);
1143
1144 /* visitor knows more than we do about the surface limit required,
1145 * so has already done marking.
1146 */
1147
1148 }
1149 }
1150
1151 void
1152 fs_generator::generate_varying_pull_constant_load(fs_inst *inst,
1153 struct brw_reg dst,
1154 struct brw_reg index,
1155 struct brw_reg offset)
1156 {
1157 assert(devinfo->gen < 7); /* Should use the gen7 variant. */
1158 assert(inst->header_present);
1159 assert(inst->mlen);
1160
1161 assert(index.file == BRW_IMMEDIATE_VALUE &&
1162 index.type == BRW_REGISTER_TYPE_UD);
1163 uint32_t surf_index = index.dw1.ud;
1164
1165 uint32_t simd_mode, rlen, msg_type;
1166 if (dispatch_width == 16) {
1167 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
1168 rlen = 8;
1169 } else {
1170 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
1171 rlen = 4;
1172 }
1173
1174 if (devinfo->gen >= 5)
1175 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
1176 else {
1177 /* We always use the SIMD16 message so that we only have to load U, and
1178 * not V or R.
1179 */
1180 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_LD;
1181 assert(inst->mlen == 3);
1182 assert(inst->regs_written == 8);
1183 rlen = 8;
1184 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
1185 }
1186
1187 struct brw_reg offset_mrf = retype(brw_message_reg(inst->base_mrf + 1),
1188 BRW_REGISTER_TYPE_D);
1189 brw_MOV(p, offset_mrf, offset);
1190
1191 struct brw_reg header = brw_vec8_grf(0, 0);
1192 gen6_resolve_implied_move(p, &header, inst->base_mrf);
1193
1194 brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
1195 brw_inst_set_qtr_control(p->devinfo, send, BRW_COMPRESSION_NONE);
1196 brw_set_dest(p, send, retype(dst, BRW_REGISTER_TYPE_UW));
1197 brw_set_src0(p, send, header);
1198 if (devinfo->gen < 6)
1199 brw_inst_set_base_mrf(p->devinfo, send, inst->base_mrf);
1200
1201 /* Our surface is set up as floats, regardless of what actual data is
1202 * stored in it.
1203 */
1204 uint32_t return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
1205 brw_set_sampler_message(p, send,
1206 surf_index,
1207 0, /* sampler (unused) */
1208 msg_type,
1209 rlen,
1210 inst->mlen,
1211 inst->header_present,
1212 simd_mode,
1213 return_format);
1214
1215 brw_mark_surface_used(prog_data, surf_index);
1216 }
1217
1218 void
1219 fs_generator::generate_varying_pull_constant_load_gen7(fs_inst *inst,
1220 struct brw_reg dst,
1221 struct brw_reg index,
1222 struct brw_reg offset)
1223 {
1224 assert(devinfo->gen >= 7);
1225 /* Varying-offset pull constant loads are treated as a normal expression on
1226 * gen7, so the fact that it's a send message is hidden at the IR level.
1227 */
1228 assert(!inst->header_present);
1229 assert(!inst->mlen);
1230 assert(index.type == BRW_REGISTER_TYPE_UD);
1231
1232 uint32_t simd_mode, rlen, mlen;
1233 if (dispatch_width == 16) {
1234 mlen = 2;
1235 rlen = 8;
1236 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
1237 } else {
1238 mlen = 1;
1239 rlen = 4;
1240 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
1241 }
1242
1243 if (index.file == BRW_IMMEDIATE_VALUE) {
1244
1245 uint32_t surf_index = index.dw1.ud;
1246
1247 brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
1248 brw_set_dest(p, send, retype(dst, BRW_REGISTER_TYPE_UW));
1249 brw_set_src0(p, send, offset);
1250 brw_set_sampler_message(p, send,
1251 surf_index,
1252 0, /* LD message ignores sampler unit */
1253 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1254 rlen,
1255 mlen,
1256 false, /* no header */
1257 simd_mode,
1258 0);
1259
1260 brw_mark_surface_used(prog_data, surf_index);
1261
1262 } else {
1263
1264 struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
1265
1266 brw_push_insn_state(p);
1267 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1268 brw_set_default_access_mode(p, BRW_ALIGN_1);
1269
1270 /* a0.0 = surf_index & 0xff */
1271 brw_inst *insn_and = brw_next_insn(p, BRW_OPCODE_AND);
1272 brw_inst_set_exec_size(p->devinfo, insn_and, BRW_EXECUTE_1);
1273 brw_set_dest(p, insn_and, addr);
1274 brw_set_src0(p, insn_and, vec1(retype(index, BRW_REGISTER_TYPE_UD)));
1275 brw_set_src1(p, insn_and, brw_imm_ud(0x0ff));
1276
1277 brw_pop_insn_state(p);
1278
1279 /* dst = send(offset, a0.0 | <descriptor>) */
1280 brw_inst *insn = brw_send_indirect_message(
1281 p, BRW_SFID_SAMPLER, retype(dst, BRW_REGISTER_TYPE_UW),
1282 offset, addr);
1283 brw_set_sampler_message(p, insn,
1284 0 /* surface */,
1285 0 /* sampler */,
1286 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1287 rlen /* rlen */,
1288 mlen /* mlen */,
1289 false /* header */,
1290 simd_mode,
1291 0);
1292
1293 /* visitor knows more than we do about the surface limit required,
1294 * so has already done marking.
1295 */
1296 }
1297 }
1298
1299 /**
1300 * Cause the current pixel/sample mask (from R1.7 bits 15:0) to be transferred
1301 * into the flags register (f0.0).
1302 *
1303 * Used only on Gen6 and above.
1304 */
1305 void
1306 fs_generator::generate_mov_dispatch_to_flags(fs_inst *inst)
1307 {
1308 struct brw_reg flags = brw_flag_reg(0, inst->flag_subreg);
1309 struct brw_reg dispatch_mask;
1310
1311 if (devinfo->gen >= 6)
1312 dispatch_mask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
1313 else
1314 dispatch_mask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
1315
1316 brw_push_insn_state(p);
1317 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1318 brw_MOV(p, flags, dispatch_mask);
1319 brw_pop_insn_state(p);
1320 }
1321
1322 void
1323 fs_generator::generate_pixel_interpolator_query(fs_inst *inst,
1324 struct brw_reg dst,
1325 struct brw_reg src,
1326 struct brw_reg msg_data,
1327 unsigned msg_type)
1328 {
1329 assert(msg_data.file == BRW_IMMEDIATE_VALUE &&
1330 msg_data.type == BRW_REGISTER_TYPE_UD);
1331
1332 brw_pixel_interpolator_query(p,
1333 retype(dst, BRW_REGISTER_TYPE_UW),
1334 src,
1335 inst->pi_noperspective,
1336 msg_type,
1337 msg_data.dw1.ud,
1338 inst->mlen,
1339 inst->regs_written);
1340 }
1341
1342
1343 /**
1344 * Sets the first word of a vgrf for gen7+ simd4x2 uniform pull constant
1345 * sampler LD messages.
1346 *
1347 * We don't want to bake it into the send message's code generation because
1348 * that means we don't get a chance to schedule the instructions.
1349 */
1350 void
1351 fs_generator::generate_set_simd4x2_offset(fs_inst *inst,
1352 struct brw_reg dst,
1353 struct brw_reg value)
1354 {
1355 assert(value.file == BRW_IMMEDIATE_VALUE);
1356
1357 brw_push_insn_state(p);
1358 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1359 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1360 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1361 brw_MOV(p, retype(brw_vec1_reg(dst.file, dst.nr, 0), value.type), value);
1362 brw_pop_insn_state(p);
1363 }
1364
1365 /* Sets vstride=16, width=8, hstride=2 or vstride=0, width=1, hstride=0
1366 * (when mask is passed as a uniform) of register mask before moving it
1367 * to register dst.
1368 */
1369 void
1370 fs_generator::generate_set_omask(fs_inst *inst,
1371 struct brw_reg dst,
1372 struct brw_reg mask)
1373 {
1374 bool stride_8_8_1 =
1375 (mask.vstride == BRW_VERTICAL_STRIDE_8 &&
1376 mask.width == BRW_WIDTH_8 &&
1377 mask.hstride == BRW_HORIZONTAL_STRIDE_1);
1378
1379 bool stride_0_1_0 = has_scalar_region(mask);
1380
1381 assert(stride_8_8_1 || stride_0_1_0);
1382 assert(dst.type == BRW_REGISTER_TYPE_UW);
1383
1384 brw_push_insn_state(p);
1385 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1386 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1387
1388 if (stride_8_8_1) {
1389 brw_MOV(p, dst, retype(stride(mask, 16, 8, 2), dst.type));
1390 } else if (stride_0_1_0) {
1391 brw_MOV(p, dst, retype(mask, dst.type));
1392 }
1393 brw_pop_insn_state(p);
1394 }
1395
1396 /* Sets vstride=1, width=4, hstride=0 of register src1 during
1397 * the ADD instruction.
1398 */
1399 void
1400 fs_generator::generate_set_sample_id(fs_inst *inst,
1401 struct brw_reg dst,
1402 struct brw_reg src0,
1403 struct brw_reg src1)
1404 {
1405 assert(dst.type == BRW_REGISTER_TYPE_D ||
1406 dst.type == BRW_REGISTER_TYPE_UD);
1407 assert(src0.type == BRW_REGISTER_TYPE_D ||
1408 src0.type == BRW_REGISTER_TYPE_UD);
1409
1410 brw_push_insn_state(p);
1411 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1412 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1413 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1414 struct brw_reg reg = retype(stride(src1, 1, 4, 0), BRW_REGISTER_TYPE_UW);
1415 if (dispatch_width == 8) {
1416 brw_ADD(p, dst, src0, reg);
1417 } else if (dispatch_width == 16) {
1418 brw_ADD(p, firsthalf(dst), firsthalf(src0), reg);
1419 brw_ADD(p, sechalf(dst), sechalf(src0), suboffset(reg, 2));
1420 }
1421 brw_pop_insn_state(p);
1422 }
1423
1424 void
1425 fs_generator::generate_pack_half_2x16_split(fs_inst *inst,
1426 struct brw_reg dst,
1427 struct brw_reg x,
1428 struct brw_reg y)
1429 {
1430 assert(devinfo->gen >= 7);
1431 assert(dst.type == BRW_REGISTER_TYPE_UD);
1432 assert(x.type == BRW_REGISTER_TYPE_F);
1433 assert(y.type == BRW_REGISTER_TYPE_F);
1434
1435 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
1436 *
1437 * Because this instruction does not have a 16-bit floating-point type,
1438 * the destination data type must be Word (W).
1439 *
1440 * The destination must be DWord-aligned and specify a horizontal stride
1441 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
1442 * each destination channel and the upper word is not modified.
1443 */
1444 struct brw_reg dst_w = spread(retype(dst, BRW_REGISTER_TYPE_W), 2);
1445
1446 /* Give each 32-bit channel of dst the form below, where "." means
1447 * unchanged.
1448 * 0x....hhhh
1449 */
1450 brw_F32TO16(p, dst_w, y);
1451
1452 /* Now the form:
1453 * 0xhhhh0000
1454 */
1455 brw_SHL(p, dst, dst, brw_imm_ud(16u));
1456
1457 /* And, finally the form of packHalf2x16's output:
1458 * 0xhhhhllll
1459 */
1460 brw_F32TO16(p, dst_w, x);
1461 }
1462
1463 void
1464 fs_generator::generate_unpack_half_2x16_split(fs_inst *inst,
1465 struct brw_reg dst,
1466 struct brw_reg src)
1467 {
1468 assert(devinfo->gen >= 7);
1469 assert(dst.type == BRW_REGISTER_TYPE_F);
1470 assert(src.type == BRW_REGISTER_TYPE_UD);
1471
1472 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
1473 *
1474 * Because this instruction does not have a 16-bit floating-point type,
1475 * the source data type must be Word (W). The destination type must be
1476 * F (Float).
1477 */
1478 struct brw_reg src_w = spread(retype(src, BRW_REGISTER_TYPE_W), 2);
1479
1480 /* Each channel of src has the form of unpackHalf2x16's input: 0xhhhhllll.
1481 * For the Y case, we wish to access only the upper word; therefore
1482 * a 16-bit subregister offset is needed.
1483 */
1484 assert(inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X ||
1485 inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y);
1486 if (inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y)
1487 src_w.subnr += 2;
1488
1489 brw_F16TO32(p, dst, src_w);
1490 }
1491
1492 void
1493 fs_generator::generate_shader_time_add(fs_inst *inst,
1494 struct brw_reg payload,
1495 struct brw_reg offset,
1496 struct brw_reg value)
1497 {
1498 assert(devinfo->gen >= 7);
1499 brw_push_insn_state(p);
1500 brw_set_default_mask_control(p, true);
1501
1502 assert(payload.file == BRW_GENERAL_REGISTER_FILE);
1503 struct brw_reg payload_offset = retype(brw_vec1_grf(payload.nr, 0),
1504 offset.type);
1505 struct brw_reg payload_value = retype(brw_vec1_grf(payload.nr + 1, 0),
1506 value.type);
1507
1508 assert(offset.file == BRW_IMMEDIATE_VALUE);
1509 if (value.file == BRW_GENERAL_REGISTER_FILE) {
1510 value.width = BRW_WIDTH_1;
1511 value.hstride = BRW_HORIZONTAL_STRIDE_0;
1512 value.vstride = BRW_VERTICAL_STRIDE_0;
1513 } else {
1514 assert(value.file == BRW_IMMEDIATE_VALUE);
1515 }
1516
1517 /* Trying to deal with setup of the params from the IR is crazy in the FS8
1518 * case, and we don't really care about squeezing every bit of performance
1519 * out of this path, so we just emit the MOVs from here.
1520 */
1521 brw_MOV(p, payload_offset, offset);
1522 brw_MOV(p, payload_value, value);
1523 brw_shader_time_add(p, payload,
1524 prog_data->binding_table.shader_time_start);
1525 brw_pop_insn_state(p);
1526
1527 brw_mark_surface_used(prog_data,
1528 prog_data->binding_table.shader_time_start);
1529 }
1530
1531 void
1532 fs_generator::generate_untyped_atomic(fs_inst *inst, struct brw_reg dst,
1533 struct brw_reg payload,
1534 struct brw_reg atomic_op,
1535 struct brw_reg surf_index)
1536 {
1537 assert(atomic_op.file == BRW_IMMEDIATE_VALUE &&
1538 atomic_op.type == BRW_REGISTER_TYPE_UD &&
1539 surf_index.file == BRW_IMMEDIATE_VALUE &&
1540 surf_index.type == BRW_REGISTER_TYPE_UD);
1541
1542 brw_untyped_atomic(p, dst, payload,
1543 surf_index, atomic_op.dw1.ud,
1544 inst->mlen, true);
1545
1546 brw_mark_surface_used(prog_data, surf_index.dw1.ud);
1547 }
1548
1549 void
1550 fs_generator::generate_untyped_surface_read(fs_inst *inst, struct brw_reg dst,
1551 struct brw_reg payload,
1552 struct brw_reg surf_index)
1553 {
1554 assert(surf_index.file == BRW_IMMEDIATE_VALUE &&
1555 surf_index.type == BRW_REGISTER_TYPE_UD);
1556
1557 brw_untyped_surface_read(p, dst, payload, surf_index, inst->mlen, 1);
1558
1559 brw_mark_surface_used(prog_data, surf_index.dw1.ud);
1560 }
1561
1562 void
1563 fs_generator::enable_debug(const char *shader_name)
1564 {
1565 debug_flag = true;
1566 this->shader_name = shader_name;
1567 }
1568
1569 int
1570 fs_generator::generate_code(const cfg_t *cfg, int dispatch_width)
1571 {
1572 /* align to 64 byte boundary. */
1573 while (p->next_insn_offset % 64)
1574 brw_NOP(p);
1575
1576 this->dispatch_width = dispatch_width;
1577 if (dispatch_width == 16)
1578 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1579
1580 int start_offset = p->next_insn_offset;
1581 int spill_count = 0, fill_count = 0;
1582 int loop_count = 0;
1583
1584 struct annotation_info annotation;
1585 memset(&annotation, 0, sizeof(annotation));
1586
1587 foreach_block_and_inst (block, fs_inst, inst, cfg) {
1588 struct brw_reg src[3], dst;
1589 unsigned int last_insn_offset = p->next_insn_offset;
1590 bool multiple_instructions_emitted = false;
1591
1592 if (unlikely(debug_flag))
1593 annotate(p->devinfo, &annotation, cfg, inst, p->next_insn_offset);
1594
1595 for (unsigned int i = 0; i < inst->sources; i++) {
1596 src[i] = brw_reg_from_fs_reg(&inst->src[i]);
1597
1598 /* The accumulator result appears to get used for the
1599 * conditional modifier generation. When negating a UD
1600 * value, there is a 33rd bit generated for the sign in the
1601 * accumulator value, so now you can't check, for example,
1602 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
1603 */
1604 assert(!inst->conditional_mod ||
1605 inst->src[i].type != BRW_REGISTER_TYPE_UD ||
1606 !inst->src[i].negate);
1607 }
1608 dst = brw_reg_from_fs_reg(&inst->dst);
1609
1610 brw_set_default_predicate_control(p, inst->predicate);
1611 brw_set_default_predicate_inverse(p, inst->predicate_inverse);
1612 brw_set_default_flag_reg(p, 0, inst->flag_subreg);
1613 brw_set_default_saturate(p, inst->saturate);
1614 brw_set_default_mask_control(p, inst->force_writemask_all);
1615 brw_set_default_acc_write_control(p, inst->writes_accumulator);
1616 brw_set_default_exec_size(p, cvt(inst->exec_size) - 1);
1617
1618 switch (inst->exec_size) {
1619 case 1:
1620 case 2:
1621 case 4:
1622 assert(inst->force_writemask_all);
1623 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1624 break;
1625 case 8:
1626 if (inst->force_sechalf) {
1627 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1628 } else {
1629 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1630 }
1631 break;
1632 case 16:
1633 case 32:
1634 if (type_sz(inst->dst.type) < sizeof(float))
1635 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1636 else
1637 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1638 break;
1639 default:
1640 unreachable("Invalid instruction width");
1641 }
1642
1643 switch (inst->opcode) {
1644 case BRW_OPCODE_MOV:
1645 brw_MOV(p, dst, src[0]);
1646 break;
1647 case BRW_OPCODE_ADD:
1648 brw_ADD(p, dst, src[0], src[1]);
1649 break;
1650 case BRW_OPCODE_MUL:
1651 brw_MUL(p, dst, src[0], src[1]);
1652 break;
1653 case BRW_OPCODE_AVG:
1654 brw_AVG(p, dst, src[0], src[1]);
1655 break;
1656 case BRW_OPCODE_MACH:
1657 brw_MACH(p, dst, src[0], src[1]);
1658 break;
1659
1660 case BRW_OPCODE_LINE:
1661 brw_LINE(p, dst, src[0], src[1]);
1662 break;
1663
1664 case BRW_OPCODE_MAD:
1665 assert(devinfo->gen >= 6);
1666 brw_set_default_access_mode(p, BRW_ALIGN_16);
1667 if (dispatch_width == 16 && !devinfo->supports_simd16_3src) {
1668 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1669 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1670 brw_inst *f = brw_MAD(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1671 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1672 brw_inst *s = brw_MAD(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1673 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1674
1675 if (inst->conditional_mod) {
1676 brw_inst_set_cond_modifier(p->devinfo, f, inst->conditional_mod);
1677 brw_inst_set_cond_modifier(p->devinfo, s, inst->conditional_mod);
1678 multiple_instructions_emitted = true;
1679 }
1680 } else {
1681 brw_MAD(p, dst, src[0], src[1], src[2]);
1682 }
1683 brw_set_default_access_mode(p, BRW_ALIGN_1);
1684 break;
1685
1686 case BRW_OPCODE_LRP:
1687 assert(devinfo->gen >= 6);
1688 brw_set_default_access_mode(p, BRW_ALIGN_16);
1689 if (dispatch_width == 16 && !devinfo->supports_simd16_3src) {
1690 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1691 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1692 brw_inst *f = brw_LRP(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1693 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1694 brw_inst *s = brw_LRP(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1695 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1696
1697 if (inst->conditional_mod) {
1698 brw_inst_set_cond_modifier(p->devinfo, f, inst->conditional_mod);
1699 brw_inst_set_cond_modifier(p->devinfo, s, inst->conditional_mod);
1700 multiple_instructions_emitted = true;
1701 }
1702 } else {
1703 brw_LRP(p, dst, src[0], src[1], src[2]);
1704 }
1705 brw_set_default_access_mode(p, BRW_ALIGN_1);
1706 break;
1707
1708 case BRW_OPCODE_FRC:
1709 brw_FRC(p, dst, src[0]);
1710 break;
1711 case BRW_OPCODE_RNDD:
1712 brw_RNDD(p, dst, src[0]);
1713 break;
1714 case BRW_OPCODE_RNDE:
1715 brw_RNDE(p, dst, src[0]);
1716 break;
1717 case BRW_OPCODE_RNDZ:
1718 brw_RNDZ(p, dst, src[0]);
1719 break;
1720
1721 case BRW_OPCODE_AND:
1722 brw_AND(p, dst, src[0], src[1]);
1723 break;
1724 case BRW_OPCODE_OR:
1725 brw_OR(p, dst, src[0], src[1]);
1726 break;
1727 case BRW_OPCODE_XOR:
1728 brw_XOR(p, dst, src[0], src[1]);
1729 break;
1730 case BRW_OPCODE_NOT:
1731 brw_NOT(p, dst, src[0]);
1732 break;
1733 case BRW_OPCODE_ASR:
1734 brw_ASR(p, dst, src[0], src[1]);
1735 break;
1736 case BRW_OPCODE_SHR:
1737 brw_SHR(p, dst, src[0], src[1]);
1738 break;
1739 case BRW_OPCODE_SHL:
1740 brw_SHL(p, dst, src[0], src[1]);
1741 break;
1742 case BRW_OPCODE_F32TO16:
1743 assert(devinfo->gen >= 7);
1744 brw_F32TO16(p, dst, src[0]);
1745 break;
1746 case BRW_OPCODE_F16TO32:
1747 assert(devinfo->gen >= 7);
1748 brw_F16TO32(p, dst, src[0]);
1749 break;
1750 case BRW_OPCODE_CMP:
1751 /* The Ivybridge/BayTrail WaCMPInstFlagDepClearedEarly workaround says
1752 * that when the destination is a GRF that the dependency-clear bit on
1753 * the flag register is cleared early.
1754 *
1755 * Suggested workarounds are to disable coissuing CMP instructions
1756 * or to split CMP(16) instructions into two CMP(8) instructions.
1757 *
1758 * We choose to split into CMP(8) instructions since disabling
1759 * coissuing would affect CMP instructions not otherwise affected by
1760 * the errata.
1761 */
1762 if (dispatch_width == 16 && devinfo->gen == 7 && !devinfo->is_haswell) {
1763 if (dst.file == BRW_GENERAL_REGISTER_FILE) {
1764 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1765 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1766 brw_CMP(p, firsthalf(dst), inst->conditional_mod,
1767 firsthalf(src[0]), firsthalf(src[1]));
1768 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1769 brw_CMP(p, sechalf(dst), inst->conditional_mod,
1770 sechalf(src[0]), sechalf(src[1]));
1771 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1772
1773 multiple_instructions_emitted = true;
1774 } else if (dst.file == BRW_ARCHITECTURE_REGISTER_FILE) {
1775 /* For unknown reasons, the aforementioned workaround is not
1776 * sufficient. Overriding the type when the destination is the
1777 * null register is necessary but not sufficient by itself.
1778 */
1779 assert(dst.nr == BRW_ARF_NULL);
1780 dst.type = BRW_REGISTER_TYPE_D;
1781 brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
1782 } else {
1783 unreachable("not reached");
1784 }
1785 } else {
1786 brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
1787 }
1788 break;
1789 case BRW_OPCODE_SEL:
1790 brw_SEL(p, dst, src[0], src[1]);
1791 break;
1792 case BRW_OPCODE_BFREV:
1793 assert(devinfo->gen >= 7);
1794 /* BFREV only supports UD type for src and dst. */
1795 brw_BFREV(p, retype(dst, BRW_REGISTER_TYPE_UD),
1796 retype(src[0], BRW_REGISTER_TYPE_UD));
1797 break;
1798 case BRW_OPCODE_FBH:
1799 assert(devinfo->gen >= 7);
1800 /* FBH only supports UD type for dst. */
1801 brw_FBH(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1802 break;
1803 case BRW_OPCODE_FBL:
1804 assert(devinfo->gen >= 7);
1805 /* FBL only supports UD type for dst. */
1806 brw_FBL(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1807 break;
1808 case BRW_OPCODE_CBIT:
1809 assert(devinfo->gen >= 7);
1810 /* CBIT only supports UD type for dst. */
1811 brw_CBIT(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1812 break;
1813 case BRW_OPCODE_ADDC:
1814 assert(devinfo->gen >= 7);
1815 brw_ADDC(p, dst, src[0], src[1]);
1816 break;
1817 case BRW_OPCODE_SUBB:
1818 assert(devinfo->gen >= 7);
1819 brw_SUBB(p, dst, src[0], src[1]);
1820 break;
1821 case BRW_OPCODE_MAC:
1822 brw_MAC(p, dst, src[0], src[1]);
1823 break;
1824
1825 case BRW_OPCODE_BFE:
1826 assert(devinfo->gen >= 7);
1827 brw_set_default_access_mode(p, BRW_ALIGN_16);
1828 if (dispatch_width == 16 && !devinfo->supports_simd16_3src) {
1829 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1830 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1831 brw_BFE(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1832 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1833 brw_BFE(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1834 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1835 } else {
1836 brw_BFE(p, dst, src[0], src[1], src[2]);
1837 }
1838 brw_set_default_access_mode(p, BRW_ALIGN_1);
1839 break;
1840
1841 case BRW_OPCODE_BFI1:
1842 assert(devinfo->gen >= 7);
1843 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
1844 * should
1845 *
1846 * "Force BFI instructions to be executed always in SIMD8."
1847 */
1848 if (dispatch_width == 16 && devinfo->is_haswell) {
1849 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1850 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1851 brw_BFI1(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]));
1852 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1853 brw_BFI1(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]));
1854 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1855 } else {
1856 brw_BFI1(p, dst, src[0], src[1]);
1857 }
1858 break;
1859 case BRW_OPCODE_BFI2:
1860 assert(devinfo->gen >= 7);
1861 brw_set_default_access_mode(p, BRW_ALIGN_16);
1862 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
1863 * should
1864 *
1865 * "Force BFI instructions to be executed always in SIMD8."
1866 *
1867 * Otherwise we would be able to emit compressed instructions like we
1868 * do for the other three-source instructions.
1869 */
1870 if (dispatch_width == 16 &&
1871 (devinfo->is_haswell || !devinfo->supports_simd16_3src)) {
1872 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1873 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1874 brw_BFI2(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1875 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1876 brw_BFI2(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1877 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1878 } else {
1879 brw_BFI2(p, dst, src[0], src[1], src[2]);
1880 }
1881 brw_set_default_access_mode(p, BRW_ALIGN_1);
1882 break;
1883
1884 case BRW_OPCODE_IF:
1885 if (inst->src[0].file != BAD_FILE) {
1886 /* The instruction has an embedded compare (only allowed on gen6) */
1887 assert(devinfo->gen == 6);
1888 gen6_IF(p, inst->conditional_mod, src[0], src[1]);
1889 } else {
1890 brw_IF(p, dispatch_width == 16 ? BRW_EXECUTE_16 : BRW_EXECUTE_8);
1891 }
1892 break;
1893
1894 case BRW_OPCODE_ELSE:
1895 brw_ELSE(p);
1896 break;
1897 case BRW_OPCODE_ENDIF:
1898 brw_ENDIF(p);
1899 break;
1900
1901 case BRW_OPCODE_DO:
1902 brw_DO(p, BRW_EXECUTE_8);
1903 break;
1904
1905 case BRW_OPCODE_BREAK:
1906 brw_BREAK(p);
1907 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
1908 break;
1909 case BRW_OPCODE_CONTINUE:
1910 brw_CONT(p);
1911 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
1912 break;
1913
1914 case BRW_OPCODE_WHILE:
1915 brw_WHILE(p);
1916 loop_count++;
1917 break;
1918
1919 case SHADER_OPCODE_RCP:
1920 case SHADER_OPCODE_RSQ:
1921 case SHADER_OPCODE_SQRT:
1922 case SHADER_OPCODE_EXP2:
1923 case SHADER_OPCODE_LOG2:
1924 case SHADER_OPCODE_SIN:
1925 case SHADER_OPCODE_COS:
1926 assert(devinfo->gen < 6 || inst->mlen == 0);
1927 assert(inst->conditional_mod == BRW_CONDITIONAL_NONE);
1928 if (devinfo->gen >= 7) {
1929 gen6_math(p, dst, brw_math_function(inst->opcode), src[0],
1930 brw_null_reg());
1931 } else if (devinfo->gen == 6) {
1932 generate_math_gen6(inst, dst, src[0], brw_null_reg());
1933 } else if (devinfo->gen == 5 || devinfo->is_g4x) {
1934 generate_math_g45(inst, dst, src[0]);
1935 } else {
1936 generate_math_gen4(inst, dst, src[0]);
1937 }
1938 break;
1939 case SHADER_OPCODE_INT_QUOTIENT:
1940 case SHADER_OPCODE_INT_REMAINDER:
1941 case SHADER_OPCODE_POW:
1942 assert(devinfo->gen < 6 || inst->mlen == 0);
1943 assert(inst->conditional_mod == BRW_CONDITIONAL_NONE);
1944 if (devinfo->gen >= 7 && inst->opcode == SHADER_OPCODE_POW) {
1945 gen6_math(p, dst, brw_math_function(inst->opcode), src[0], src[1]);
1946 } else if (devinfo->gen >= 6) {
1947 generate_math_gen6(inst, dst, src[0], src[1]);
1948 } else {
1949 generate_math_gen4(inst, dst, src[0]);
1950 }
1951 break;
1952 case FS_OPCODE_CINTERP:
1953 brw_MOV(p, dst, src[0]);
1954 break;
1955 case FS_OPCODE_LINTERP:
1956 generate_linterp(inst, dst, src);
1957 break;
1958 case FS_OPCODE_PIXEL_X:
1959 assert(src[0].type == BRW_REGISTER_TYPE_UW);
1960 src[0].subnr = 0 * type_sz(src[0].type);
1961 brw_MOV(p, dst, stride(src[0], 8, 4, 1));
1962 break;
1963 case FS_OPCODE_PIXEL_Y:
1964 assert(src[0].type == BRW_REGISTER_TYPE_UW);
1965 src[0].subnr = 4 * type_sz(src[0].type);
1966 brw_MOV(p, dst, stride(src[0], 8, 4, 1));
1967 break;
1968 case SHADER_OPCODE_TEX:
1969 case FS_OPCODE_TXB:
1970 case SHADER_OPCODE_TXD:
1971 case SHADER_OPCODE_TXF:
1972 case SHADER_OPCODE_TXF_CMS:
1973 case SHADER_OPCODE_TXF_UMS:
1974 case SHADER_OPCODE_TXF_MCS:
1975 case SHADER_OPCODE_TXL:
1976 case SHADER_OPCODE_TXS:
1977 case SHADER_OPCODE_LOD:
1978 case SHADER_OPCODE_TG4:
1979 case SHADER_OPCODE_TG4_OFFSET:
1980 generate_tex(inst, dst, src[0], src[1]);
1981 break;
1982 case FS_OPCODE_DDX_COARSE:
1983 case FS_OPCODE_DDX_FINE:
1984 generate_ddx(inst->opcode, dst, src[0]);
1985 break;
1986 case FS_OPCODE_DDY_COARSE:
1987 case FS_OPCODE_DDY_FINE:
1988 assert(src[1].file == BRW_IMMEDIATE_VALUE);
1989 generate_ddy(inst->opcode, dst, src[0], src[1].dw1.ud);
1990 break;
1991
1992 case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
1993 generate_scratch_write(inst, src[0]);
1994 spill_count++;
1995 break;
1996
1997 case SHADER_OPCODE_GEN4_SCRATCH_READ:
1998 generate_scratch_read(inst, dst);
1999 fill_count++;
2000 break;
2001
2002 case SHADER_OPCODE_GEN7_SCRATCH_READ:
2003 generate_scratch_read_gen7(inst, dst);
2004 fill_count++;
2005 break;
2006
2007 case SHADER_OPCODE_URB_WRITE_SIMD8:
2008 generate_urb_write(inst, src[0]);
2009 break;
2010
2011 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
2012 generate_uniform_pull_constant_load(inst, dst, src[0], src[1]);
2013 break;
2014
2015 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7:
2016 generate_uniform_pull_constant_load_gen7(inst, dst, src[0], src[1]);
2017 break;
2018
2019 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
2020 generate_varying_pull_constant_load(inst, dst, src[0], src[1]);
2021 break;
2022
2023 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
2024 generate_varying_pull_constant_load_gen7(inst, dst, src[0], src[1]);
2025 break;
2026
2027 case FS_OPCODE_REP_FB_WRITE:
2028 case FS_OPCODE_FB_WRITE:
2029 generate_fb_write(inst, src[0]);
2030 break;
2031
2032 case FS_OPCODE_BLORP_FB_WRITE:
2033 generate_blorp_fb_write(inst);
2034 break;
2035
2036 case FS_OPCODE_MOV_DISPATCH_TO_FLAGS:
2037 generate_mov_dispatch_to_flags(inst);
2038 break;
2039
2040 case FS_OPCODE_DISCARD_JUMP:
2041 generate_discard_jump(inst);
2042 break;
2043
2044 case SHADER_OPCODE_SHADER_TIME_ADD:
2045 generate_shader_time_add(inst, src[0], src[1], src[2]);
2046 break;
2047
2048 case SHADER_OPCODE_UNTYPED_ATOMIC:
2049 generate_untyped_atomic(inst, dst, src[0], src[1], src[2]);
2050 break;
2051
2052 case SHADER_OPCODE_UNTYPED_SURFACE_READ:
2053 generate_untyped_surface_read(inst, dst, src[0], src[1]);
2054 break;
2055
2056 case FS_OPCODE_SET_SIMD4X2_OFFSET:
2057 generate_set_simd4x2_offset(inst, dst, src[0]);
2058 break;
2059
2060 case FS_OPCODE_SET_OMASK:
2061 generate_set_omask(inst, dst, src[0]);
2062 break;
2063
2064 case FS_OPCODE_SET_SAMPLE_ID:
2065 generate_set_sample_id(inst, dst, src[0], src[1]);
2066 break;
2067
2068 case FS_OPCODE_PACK_HALF_2x16_SPLIT:
2069 generate_pack_half_2x16_split(inst, dst, src[0], src[1]);
2070 break;
2071
2072 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X:
2073 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y:
2074 generate_unpack_half_2x16_split(inst, dst, src[0]);
2075 break;
2076
2077 case FS_OPCODE_PLACEHOLDER_HALT:
2078 /* This is the place where the final HALT needs to be inserted if
2079 * we've emitted any discards. If not, this will emit no code.
2080 */
2081 if (!patch_discard_jumps_to_fb_writes()) {
2082 if (unlikely(debug_flag)) {
2083 annotation.ann_count--;
2084 }
2085 }
2086 break;
2087
2088 case FS_OPCODE_INTERPOLATE_AT_CENTROID:
2089 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2090 GEN7_PIXEL_INTERPOLATOR_LOC_CENTROID);
2091 break;
2092
2093 case FS_OPCODE_INTERPOLATE_AT_SAMPLE:
2094 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2095 GEN7_PIXEL_INTERPOLATOR_LOC_SAMPLE);
2096 break;
2097
2098 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET:
2099 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2100 GEN7_PIXEL_INTERPOLATOR_LOC_SHARED_OFFSET);
2101 break;
2102
2103 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET:
2104 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2105 GEN7_PIXEL_INTERPOLATOR_LOC_PER_SLOT_OFFSET);
2106 break;
2107
2108 case CS_OPCODE_CS_TERMINATE:
2109 generate_cs_terminate(inst, src[0]);
2110 break;
2111
2112 default:
2113 unreachable("Unsupported opcode");
2114
2115 case SHADER_OPCODE_LOAD_PAYLOAD:
2116 unreachable("Should be lowered by lower_load_payload()");
2117 }
2118
2119 if (multiple_instructions_emitted)
2120 continue;
2121
2122 if (inst->no_dd_clear || inst->no_dd_check || inst->conditional_mod) {
2123 assert(p->next_insn_offset == last_insn_offset + 16 ||
2124 !"conditional_mod, no_dd_check, or no_dd_clear set for IR "
2125 "emitting more than 1 instruction");
2126
2127 brw_inst *last = &p->store[last_insn_offset / 16];
2128
2129 if (inst->conditional_mod)
2130 brw_inst_set_cond_modifier(p->devinfo, last, inst->conditional_mod);
2131 brw_inst_set_no_dd_clear(p->devinfo, last, inst->no_dd_clear);
2132 brw_inst_set_no_dd_check(p->devinfo, last, inst->no_dd_check);
2133 }
2134 }
2135
2136 brw_set_uip_jip(p);
2137 annotation_finalize(&annotation, p->next_insn_offset);
2138
2139 int before_size = p->next_insn_offset - start_offset;
2140 brw_compact_instructions(p, start_offset, annotation.ann_count,
2141 annotation.ann);
2142 int after_size = p->next_insn_offset - start_offset;
2143
2144 if (unlikely(debug_flag)) {
2145 fprintf(stderr, "Native code for %s\n"
2146 "SIMD%d shader: %d instructions. %d loops. %d:%d spills:fills. Promoted %u constants. Compacted %d to %d"
2147 " bytes (%.0f%%)\n",
2148 shader_name, dispatch_width, before_size / 16, loop_count,
2149 spill_count, fill_count, promoted_constants, before_size, after_size,
2150 100.0f * (before_size - after_size) / before_size);
2151
2152 dump_assembly(p->store, annotation.ann_count, annotation.ann,
2153 p->devinfo, prog);
2154 ralloc_free(annotation.ann);
2155 }
2156
2157 static GLuint msg_id = 0;
2158 _mesa_gl_debug(&brw->ctx, &msg_id,
2159 MESA_DEBUG_SOURCE_SHADER_COMPILER,
2160 MESA_DEBUG_TYPE_OTHER,
2161 MESA_DEBUG_SEVERITY_NOTIFICATION,
2162 "%s SIMD%d shader: %d inst, %d loops, %d:%d spills:fills, "
2163 "Promoted %u constants, compacted %d to %d bytes.\n",
2164 stage_abbrev, dispatch_width, before_size / 16, loop_count,
2165 spill_count, fill_count, promoted_constants, before_size, after_size);
2166
2167 return start_offset;
2168 }
2169
2170 const unsigned *
2171 fs_generator::get_assembly(unsigned int *assembly_size)
2172 {
2173 return brw_get_program(p, assembly_size);
2174 }