projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
i965/fs: set execution size to 8 with simd8 ddy instruction
[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_size != 0);
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_size != 0) {
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_size != 0);
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_size == 0 ||
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_size != 0) {
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_size != 0,
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_size != 0 /* 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_exec_size(p, BRW_EXECUTE_8);
945 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
946 if (negate_value) {
947 brw_ADD(p, firsthalf(dst), firsthalf(src1), negate(firsthalf(src0)));
948 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
949 brw_ADD(p, sechalf(dst), sechalf(src1), negate(sechalf(src0)));
950 } else {
951 brw_ADD(p, firsthalf(dst), firsthalf(src0), negate(firsthalf(src1)));
952 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
953 brw_ADD(p, sechalf(dst), sechalf(src0), negate(sechalf(src1)));
954 }
955 } else {
956 if (negate_value)
957 brw_ADD(p, dst, src1, negate(src0));
958 else
959 brw_ADD(p, dst, src0, negate(src1));
960 }
961 brw_pop_insn_state(p);
962 } else {
963 /* replicate the derivative at the top-left pixel to other pixels */
964 struct brw_reg src0 = brw_reg(src.file, src.nr, 0,
965 src.negate, src.abs,
966 BRW_REGISTER_TYPE_F,
967 BRW_VERTICAL_STRIDE_4,
968 BRW_WIDTH_4,
969 BRW_HORIZONTAL_STRIDE_0,
970 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
971 struct brw_reg src1 = brw_reg(src.file, src.nr, 2,
972 src.negate, src.abs,
973 BRW_REGISTER_TYPE_F,
974 BRW_VERTICAL_STRIDE_4,
975 BRW_WIDTH_4,
976 BRW_HORIZONTAL_STRIDE_0,
977 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
978 if (negate_value)
979 brw_ADD(p, dst, src1, negate(src0));
980 else
981 brw_ADD(p, dst, src0, negate(src1));
982 }
983 }
984
985 void
986 fs_generator::generate_discard_jump(fs_inst *inst)
987 {
988 assert(devinfo->gen >= 6);
989
990 /* This HALT will be patched up at FB write time to point UIP at the end of
991 * the program, and at brw_uip_jip() JIP will be set to the end of the
992 * current block (or the program).
993 */
994 this->discard_halt_patches.push_tail(new(mem_ctx) ip_record(p->nr_insn));
995
996 brw_push_insn_state(p);
997 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
998 gen6_HALT(p);
999 brw_pop_insn_state(p);
1000 }
1001
1002 void
1003 fs_generator::generate_scratch_write(fs_inst *inst, struct brw_reg src)
1004 {
1005 assert(inst->mlen != 0);
1006
1007 brw_MOV(p,
1008 brw_uvec_mrf(inst->exec_size, (inst->base_mrf + 1), 0),
1009 retype(src, BRW_REGISTER_TYPE_UD));
1010 brw_oword_block_write_scratch(p, brw_message_reg(inst->base_mrf),
1011 inst->exec_size / 8, inst->offset);
1012 }
1013
1014 void
1015 fs_generator::generate_scratch_read(fs_inst *inst, struct brw_reg dst)
1016 {
1017 assert(inst->mlen != 0);
1018
1019 brw_oword_block_read_scratch(p, dst, brw_message_reg(inst->base_mrf),
1020 inst->exec_size / 8, inst->offset);
1021 }
1022
1023 void
1024 fs_generator::generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst)
1025 {
1026 gen7_block_read_scratch(p, dst, inst->exec_size / 8, inst->offset);
1027 }
1028
1029 void
1030 fs_generator::generate_uniform_pull_constant_load(fs_inst *inst,
1031 struct brw_reg dst,
1032 struct brw_reg index,
1033 struct brw_reg offset)
1034 {
1035 assert(inst->mlen != 0);
1036
1037 assert(index.file == BRW_IMMEDIATE_VALUE &&
1038 index.type == BRW_REGISTER_TYPE_UD);
1039 uint32_t surf_index = index.dw1.ud;
1040
1041 assert(offset.file == BRW_IMMEDIATE_VALUE &&
1042 offset.type == BRW_REGISTER_TYPE_UD);
1043 uint32_t read_offset = offset.dw1.ud;
1044
1045 brw_oword_block_read(p, dst, brw_message_reg(inst->base_mrf),
1046 read_offset, surf_index);
1047
1048 brw_mark_surface_used(prog_data, surf_index);
1049 }
1050
1051 void
1052 fs_generator::generate_uniform_pull_constant_load_gen7(fs_inst *inst,
1053 struct brw_reg dst,
1054 struct brw_reg index,
1055 struct brw_reg offset)
1056 {
1057 assert(inst->mlen == 0);
1058 assert(index.type == BRW_REGISTER_TYPE_UD);
1059
1060 assert(offset.file == BRW_GENERAL_REGISTER_FILE);
1061 /* Reference just the dword we need, to avoid angering validate_reg(). */
1062 offset = brw_vec1_grf(offset.nr, 0);
1063
1064 /* We use the SIMD4x2 mode because we want to end up with 4 components in
1065 * the destination loaded consecutively from the same offset (which appears
1066 * in the first component, and the rest are ignored).
1067 */
1068 dst.width = BRW_WIDTH_4;
1069
1070 struct brw_reg src = offset;
1071 bool header_present = false;
1072 int mlen = 1;
1073
1074 if (devinfo->gen >= 9) {
1075 /* Skylake requires a message header in order to use SIMD4x2 mode. */
1076 src = retype(brw_vec4_grf(offset.nr - 1, 0), BRW_REGISTER_TYPE_UD);
1077 mlen = 2;
1078 header_present = true;
1079
1080 brw_push_insn_state(p);
1081 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1082 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1083 brw_MOV(p, vec8(src), retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
1084 brw_set_default_access_mode(p, BRW_ALIGN_1);
1085
1086 brw_MOV(p, get_element_ud(src, 2),
1087 brw_imm_ud(GEN9_SAMPLER_SIMD_MODE_EXTENSION_SIMD4X2));
1088 brw_pop_insn_state(p);
1089 }
1090
1091 if (index.file == BRW_IMMEDIATE_VALUE) {
1092
1093 uint32_t surf_index = index.dw1.ud;
1094
1095 brw_push_insn_state(p);
1096 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1097 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1098 brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
1099 brw_pop_insn_state(p);
1100
1101 brw_set_dest(p, send, dst);
1102 brw_set_src0(p, send, src);
1103 brw_set_sampler_message(p, send,
1104 surf_index,
1105 0, /* LD message ignores sampler unit */
1106 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1107 1, /* rlen */
1108 mlen,
1109 header_present,
1110 BRW_SAMPLER_SIMD_MODE_SIMD4X2,
1111 0);
1112
1113 brw_mark_surface_used(prog_data, surf_index);
1114
1115 } else {
1116
1117 struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
1118
1119 brw_push_insn_state(p);
1120 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1121 brw_set_default_access_mode(p, BRW_ALIGN_1);
1122
1123 /* a0.0 = surf_index & 0xff */
1124 brw_inst *insn_and = brw_next_insn(p, BRW_OPCODE_AND);
1125 brw_inst_set_exec_size(p->devinfo, insn_and, BRW_EXECUTE_1);
1126 brw_set_dest(p, insn_and, addr);
1127 brw_set_src0(p, insn_and, vec1(retype(index, BRW_REGISTER_TYPE_UD)));
1128 brw_set_src1(p, insn_and, brw_imm_ud(0x0ff));
1129
1130 /* dst = send(payload, a0.0 | <descriptor>) */
1131 brw_inst *insn = brw_send_indirect_message(
1132 p, BRW_SFID_SAMPLER, dst, src, addr);
1133 brw_set_sampler_message(p, insn,
1134 0,
1135 0, /* LD message ignores sampler unit */
1136 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1137 1, /* rlen */
1138 mlen,
1139 header_present,
1140 BRW_SAMPLER_SIMD_MODE_SIMD4X2,
1141 0);
1142
1143 brw_pop_insn_state(p);
1144
1145 /* visitor knows more than we do about the surface limit required,
1146 * so has already done marking.
1147 */
1148
1149 }
1150 }
1151
1152 void
1153 fs_generator::generate_varying_pull_constant_load(fs_inst *inst,
1154 struct brw_reg dst,
1155 struct brw_reg index,
1156 struct brw_reg offset)
1157 {
1158 assert(devinfo->gen < 7); /* Should use the gen7 variant. */
1159 assert(inst->header_size != 0);
1160 assert(inst->mlen);
1161
1162 assert(index.file == BRW_IMMEDIATE_VALUE &&
1163 index.type == BRW_REGISTER_TYPE_UD);
1164 uint32_t surf_index = index.dw1.ud;
1165
1166 uint32_t simd_mode, rlen, msg_type;
1167 if (dispatch_width == 16) {
1168 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
1169 rlen = 8;
1170 } else {
1171 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
1172 rlen = 4;
1173 }
1174
1175 if (devinfo->gen >= 5)
1176 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
1177 else {
1178 /* We always use the SIMD16 message so that we only have to load U, and
1179 * not V or R.
1180 */
1181 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_LD;
1182 assert(inst->mlen == 3);
1183 assert(inst->regs_written == 8);
1184 rlen = 8;
1185 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
1186 }
1187
1188 struct brw_reg offset_mrf = retype(brw_message_reg(inst->base_mrf + 1),
1189 BRW_REGISTER_TYPE_D);
1190 brw_MOV(p, offset_mrf, offset);
1191
1192 struct brw_reg header = brw_vec8_grf(0, 0);
1193 gen6_resolve_implied_move(p, &header, inst->base_mrf);
1194
1195 brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
1196 brw_inst_set_qtr_control(p->devinfo, send, BRW_COMPRESSION_NONE);
1197 brw_set_dest(p, send, retype(dst, BRW_REGISTER_TYPE_UW));
1198 brw_set_src0(p, send, header);
1199 if (devinfo->gen < 6)
1200 brw_inst_set_base_mrf(p->devinfo, send, inst->base_mrf);
1201
1202 /* Our surface is set up as floats, regardless of what actual data is
1203 * stored in it.
1204 */
1205 uint32_t return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
1206 brw_set_sampler_message(p, send,
1207 surf_index,
1208 0, /* sampler (unused) */
1209 msg_type,
1210 rlen,
1211 inst->mlen,
1212 inst->header_size != 0,
1213 simd_mode,
1214 return_format);
1215
1216 brw_mark_surface_used(prog_data, surf_index);
1217 }
1218
1219 void
1220 fs_generator::generate_varying_pull_constant_load_gen7(fs_inst *inst,
1221 struct brw_reg dst,
1222 struct brw_reg index,
1223 struct brw_reg offset)
1224 {
1225 assert(devinfo->gen >= 7);
1226 /* Varying-offset pull constant loads are treated as a normal expression on
1227 * gen7, so the fact that it's a send message is hidden at the IR level.
1228 */
1229 assert(inst->header_size == 0);
1230 assert(!inst->mlen);
1231 assert(index.type == BRW_REGISTER_TYPE_UD);
1232
1233 uint32_t simd_mode, rlen, mlen;
1234 if (dispatch_width == 16) {
1235 mlen = 2;
1236 rlen = 8;
1237 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
1238 } else {
1239 mlen = 1;
1240 rlen = 4;
1241 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
1242 }
1243
1244 if (index.file == BRW_IMMEDIATE_VALUE) {
1245
1246 uint32_t surf_index = index.dw1.ud;
1247
1248 brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
1249 brw_set_dest(p, send, retype(dst, BRW_REGISTER_TYPE_UW));
1250 brw_set_src0(p, send, offset);
1251 brw_set_sampler_message(p, send,
1252 surf_index,
1253 0, /* LD message ignores sampler unit */
1254 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1255 rlen,
1256 mlen,
1257 false, /* no header */
1258 simd_mode,
1259 0);
1260
1261 brw_mark_surface_used(prog_data, surf_index);
1262
1263 } else {
1264
1265 struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
1266
1267 brw_push_insn_state(p);
1268 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1269 brw_set_default_access_mode(p, BRW_ALIGN_1);
1270
1271 /* a0.0 = surf_index & 0xff */
1272 brw_inst *insn_and = brw_next_insn(p, BRW_OPCODE_AND);
1273 brw_inst_set_exec_size(p->devinfo, insn_and, BRW_EXECUTE_1);
1274 brw_set_dest(p, insn_and, addr);
1275 brw_set_src0(p, insn_and, vec1(retype(index, BRW_REGISTER_TYPE_UD)));
1276 brw_set_src1(p, insn_and, brw_imm_ud(0x0ff));
1277
1278 brw_pop_insn_state(p);
1279
1280 /* dst = send(offset, a0.0 | <descriptor>) */
1281 brw_inst *insn = brw_send_indirect_message(
1282 p, BRW_SFID_SAMPLER, retype(dst, BRW_REGISTER_TYPE_UW),
1283 offset, addr);
1284 brw_set_sampler_message(p, insn,
1285 0 /* surface */,
1286 0 /* sampler */,
1287 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1288 rlen /* rlen */,
1289 mlen /* mlen */,
1290 false /* header */,
1291 simd_mode,
1292 0);
1293
1294 /* visitor knows more than we do about the surface limit required,
1295 * so has already done marking.
1296 */
1297 }
1298 }
1299
1300 /**
1301 * Cause the current pixel/sample mask (from R1.7 bits 15:0) to be transferred
1302 * into the flags register (f0.0).
1303 *
1304 * Used only on Gen6 and above.
1305 */
1306 void
1307 fs_generator::generate_mov_dispatch_to_flags(fs_inst *inst)
1308 {
1309 struct brw_reg flags = brw_flag_reg(0, inst->flag_subreg);
1310 struct brw_reg dispatch_mask;
1311
1312 if (devinfo->gen >= 6)
1313 dispatch_mask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
1314 else
1315 dispatch_mask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
1316
1317 brw_push_insn_state(p);
1318 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1319 brw_MOV(p, flags, dispatch_mask);
1320 brw_pop_insn_state(p);
1321 }
1322
1323 void
1324 fs_generator::generate_pixel_interpolator_query(fs_inst *inst,
1325 struct brw_reg dst,
1326 struct brw_reg src,
1327 struct brw_reg msg_data,
1328 unsigned msg_type)
1329 {
1330 assert(msg_data.file == BRW_IMMEDIATE_VALUE &&
1331 msg_data.type == BRW_REGISTER_TYPE_UD);
1332
1333 brw_pixel_interpolator_query(p,
1334 retype(dst, BRW_REGISTER_TYPE_UW),
1335 src,
1336 inst->pi_noperspective,
1337 msg_type,
1338 msg_data.dw1.ud,
1339 inst->mlen,
1340 inst->regs_written);
1341 }
1342
1343
1344 /**
1345 * Sets the first word of a vgrf for gen7+ simd4x2 uniform pull constant
1346 * sampler LD messages.
1347 *
1348 * We don't want to bake it into the send message's code generation because
1349 * that means we don't get a chance to schedule the instructions.
1350 */
1351 void
1352 fs_generator::generate_set_simd4x2_offset(fs_inst *inst,
1353 struct brw_reg dst,
1354 struct brw_reg value)
1355 {
1356 assert(value.file == BRW_IMMEDIATE_VALUE);
1357
1358 brw_push_insn_state(p);
1359 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1360 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1361 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1362 brw_MOV(p, retype(brw_vec1_reg(dst.file, dst.nr, 0), value.type), value);
1363 brw_pop_insn_state(p);
1364 }
1365
1366 /* Sets vstride=16, width=8, hstride=2 or vstride=0, width=1, hstride=0
1367 * (when mask is passed as a uniform) of register mask before moving it
1368 * to register dst.
1369 */
1370 void
1371 fs_generator::generate_set_omask(fs_inst *inst,
1372 struct brw_reg dst,
1373 struct brw_reg mask)
1374 {
1375 bool stride_8_8_1 =
1376 (mask.vstride == BRW_VERTICAL_STRIDE_8 &&
1377 mask.width == BRW_WIDTH_8 &&
1378 mask.hstride == BRW_HORIZONTAL_STRIDE_1);
1379
1380 bool stride_0_1_0 = has_scalar_region(mask);
1381
1382 assert(stride_8_8_1 || stride_0_1_0);
1383 assert(dst.type == BRW_REGISTER_TYPE_UW);
1384
1385 brw_push_insn_state(p);
1386 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1387 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1388
1389 if (stride_8_8_1) {
1390 brw_MOV(p, dst, retype(stride(mask, 16, 8, 2), dst.type));
1391 } else if (stride_0_1_0) {
1392 brw_MOV(p, dst, retype(mask, dst.type));
1393 }
1394 brw_pop_insn_state(p);
1395 }
1396
1397 /* Sets vstride=1, width=4, hstride=0 of register src1 during
1398 * the ADD instruction.
1399 */
1400 void
1401 fs_generator::generate_set_sample_id(fs_inst *inst,
1402 struct brw_reg dst,
1403 struct brw_reg src0,
1404 struct brw_reg src1)
1405 {
1406 assert(dst.type == BRW_REGISTER_TYPE_D ||
1407 dst.type == BRW_REGISTER_TYPE_UD);
1408 assert(src0.type == BRW_REGISTER_TYPE_D ||
1409 src0.type == BRW_REGISTER_TYPE_UD);
1410
1411 brw_push_insn_state(p);
1412 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1413 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1414 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1415 struct brw_reg reg = retype(stride(src1, 1, 4, 0), BRW_REGISTER_TYPE_UW);
1416 if (dispatch_width == 8) {
1417 brw_ADD(p, dst, src0, reg);
1418 } else if (dispatch_width == 16) {
1419 brw_ADD(p, firsthalf(dst), firsthalf(src0), reg);
1420 brw_ADD(p, sechalf(dst), sechalf(src0), suboffset(reg, 2));
1421 }
1422 brw_pop_insn_state(p);
1423 }
1424
1425 void
1426 fs_generator::generate_pack_half_2x16_split(fs_inst *inst,
1427 struct brw_reg dst,
1428 struct brw_reg x,
1429 struct brw_reg y)
1430 {
1431 assert(devinfo->gen >= 7);
1432 assert(dst.type == BRW_REGISTER_TYPE_UD);
1433 assert(x.type == BRW_REGISTER_TYPE_F);
1434 assert(y.type == BRW_REGISTER_TYPE_F);
1435
1436 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
1437 *
1438 * Because this instruction does not have a 16-bit floating-point type,
1439 * the destination data type must be Word (W).
1440 *
1441 * The destination must be DWord-aligned and specify a horizontal stride
1442 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
1443 * each destination channel and the upper word is not modified.
1444 */
1445 struct brw_reg dst_w = spread(retype(dst, BRW_REGISTER_TYPE_W), 2);
1446
1447 /* Give each 32-bit channel of dst the form below, where "." means
1448 * unchanged.
1449 * 0x....hhhh
1450 */
1451 brw_F32TO16(p, dst_w, y);
1452
1453 /* Now the form:
1454 * 0xhhhh0000
1455 */
1456 brw_SHL(p, dst, dst, brw_imm_ud(16u));
1457
1458 /* And, finally the form of packHalf2x16's output:
1459 * 0xhhhhllll
1460 */
1461 brw_F32TO16(p, dst_w, x);
1462 }
1463
1464 void
1465 fs_generator::generate_unpack_half_2x16_split(fs_inst *inst,
1466 struct brw_reg dst,
1467 struct brw_reg src)
1468 {
1469 assert(devinfo->gen >= 7);
1470 assert(dst.type == BRW_REGISTER_TYPE_F);
1471 assert(src.type == BRW_REGISTER_TYPE_UD);
1472
1473 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
1474 *
1475 * Because this instruction does not have a 16-bit floating-point type,
1476 * the source data type must be Word (W). The destination type must be
1477 * F (Float).
1478 */
1479 struct brw_reg src_w = spread(retype(src, BRW_REGISTER_TYPE_W), 2);
1480
1481 /* Each channel of src has the form of unpackHalf2x16's input: 0xhhhhllll.
1482 * For the Y case, we wish to access only the upper word; therefore
1483 * a 16-bit subregister offset is needed.
1484 */
1485 assert(inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X ||
1486 inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y);
1487 if (inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y)
1488 src_w.subnr += 2;
1489
1490 brw_F16TO32(p, dst, src_w);
1491 }
1492
1493 void
1494 fs_generator::generate_shader_time_add(fs_inst *inst,
1495 struct brw_reg payload,
1496 struct brw_reg offset,
1497 struct brw_reg value)
1498 {
1499 assert(devinfo->gen >= 7);
1500 brw_push_insn_state(p);
1501 brw_set_default_mask_control(p, true);
1502
1503 assert(payload.file == BRW_GENERAL_REGISTER_FILE);
1504 struct brw_reg payload_offset = retype(brw_vec1_grf(payload.nr, 0),
1505 offset.type);
1506 struct brw_reg payload_value = retype(brw_vec1_grf(payload.nr + 1, 0),
1507 value.type);
1508
1509 assert(offset.file == BRW_IMMEDIATE_VALUE);
1510 if (value.file == BRW_GENERAL_REGISTER_FILE) {
1511 value.width = BRW_WIDTH_1;
1512 value.hstride = BRW_HORIZONTAL_STRIDE_0;
1513 value.vstride = BRW_VERTICAL_STRIDE_0;
1514 } else {
1515 assert(value.file == BRW_IMMEDIATE_VALUE);
1516 }
1517
1518 /* Trying to deal with setup of the params from the IR is crazy in the FS8
1519 * case, and we don't really care about squeezing every bit of performance
1520 * out of this path, so we just emit the MOVs from here.
1521 */
1522 brw_MOV(p, payload_offset, offset);
1523 brw_MOV(p, payload_value, value);
1524 brw_shader_time_add(p, payload,
1525 prog_data->binding_table.shader_time_start);
1526 brw_pop_insn_state(p);
1527
1528 brw_mark_surface_used(prog_data,
1529 prog_data->binding_table.shader_time_start);
1530 }
1531
1532 void
1533 fs_generator::enable_debug(const char *shader_name)
1534 {
1535 debug_flag = true;
1536 this->shader_name = shader_name;
1537 }
1538
1539 int
1540 fs_generator::generate_code(const cfg_t *cfg, int dispatch_width)
1541 {
1542 /* align to 64 byte boundary. */
1543 while (p->next_insn_offset % 64)
1544 brw_NOP(p);
1545
1546 this->dispatch_width = dispatch_width;
1547 if (dispatch_width == 16)
1548 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1549
1550 int start_offset = p->next_insn_offset;
1551 int spill_count = 0, fill_count = 0;
1552 int loop_count = 0;
1553
1554 struct annotation_info annotation;
1555 memset(&annotation, 0, sizeof(annotation));
1556
1557 foreach_block_and_inst (block, fs_inst, inst, cfg) {
1558 struct brw_reg src[3], dst;
1559 unsigned int last_insn_offset = p->next_insn_offset;
1560 bool multiple_instructions_emitted = false;
1561
1562 if (unlikely(debug_flag))
1563 annotate(p->devinfo, &annotation, cfg, inst, p->next_insn_offset);
1564
1565 for (unsigned int i = 0; i < inst->sources; i++) {
1566 src[i] = brw_reg_from_fs_reg(&inst->src[i]);
1567
1568 /* The accumulator result appears to get used for the
1569 * conditional modifier generation. When negating a UD
1570 * value, there is a 33rd bit generated for the sign in the
1571 * accumulator value, so now you can't check, for example,
1572 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
1573 */
1574 assert(!inst->conditional_mod ||
1575 inst->src[i].type != BRW_REGISTER_TYPE_UD ||
1576 !inst->src[i].negate);
1577 }
1578 dst = brw_reg_from_fs_reg(&inst->dst);
1579
1580 brw_set_default_predicate_control(p, inst->predicate);
1581 brw_set_default_predicate_inverse(p, inst->predicate_inverse);
1582 brw_set_default_flag_reg(p, 0, inst->flag_subreg);
1583 brw_set_default_saturate(p, inst->saturate);
1584 brw_set_default_mask_control(p, inst->force_writemask_all);
1585 brw_set_default_acc_write_control(p, inst->writes_accumulator);
1586 brw_set_default_exec_size(p, cvt(inst->exec_size) - 1);
1587
1588 switch (inst->exec_size) {
1589 case 1:
1590 case 2:
1591 case 4:
1592 assert(inst->force_writemask_all);
1593 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1594 break;
1595 case 8:
1596 if (inst->force_sechalf) {
1597 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1598 } else {
1599 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1600 }
1601 break;
1602 case 16:
1603 case 32:
1604 if (type_sz(inst->dst.type) < sizeof(float))
1605 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1606 else
1607 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1608 break;
1609 default:
1610 unreachable("Invalid instruction width");
1611 }
1612
1613 switch (inst->opcode) {
1614 case BRW_OPCODE_MOV:
1615 brw_MOV(p, dst, src[0]);
1616 break;
1617 case BRW_OPCODE_ADD:
1618 brw_ADD(p, dst, src[0], src[1]);
1619 break;
1620 case BRW_OPCODE_MUL:
1621 brw_MUL(p, dst, src[0], src[1]);
1622 break;
1623 case BRW_OPCODE_AVG:
1624 brw_AVG(p, dst, src[0], src[1]);
1625 break;
1626 case BRW_OPCODE_MACH:
1627 brw_MACH(p, dst, src[0], src[1]);
1628 break;
1629
1630 case BRW_OPCODE_LINE:
1631 brw_LINE(p, dst, src[0], src[1]);
1632 break;
1633
1634 case BRW_OPCODE_MAD:
1635 assert(devinfo->gen >= 6);
1636 brw_set_default_access_mode(p, BRW_ALIGN_16);
1637 if (dispatch_width == 16 && !devinfo->supports_simd16_3src) {
1638 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1639 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1640 brw_inst *f = brw_MAD(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1641 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1642 brw_inst *s = brw_MAD(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1643 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1644
1645 if (inst->conditional_mod) {
1646 brw_inst_set_cond_modifier(p->devinfo, f, inst->conditional_mod);
1647 brw_inst_set_cond_modifier(p->devinfo, s, inst->conditional_mod);
1648 multiple_instructions_emitted = true;
1649 }
1650 } else {
1651 brw_MAD(p, dst, src[0], src[1], src[2]);
1652 }
1653 brw_set_default_access_mode(p, BRW_ALIGN_1);
1654 break;
1655
1656 case BRW_OPCODE_LRP:
1657 assert(devinfo->gen >= 6);
1658 brw_set_default_access_mode(p, BRW_ALIGN_16);
1659 if (dispatch_width == 16 && !devinfo->supports_simd16_3src) {
1660 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1661 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1662 brw_inst *f = brw_LRP(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1663 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1664 brw_inst *s = brw_LRP(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1665 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1666
1667 if (inst->conditional_mod) {
1668 brw_inst_set_cond_modifier(p->devinfo, f, inst->conditional_mod);
1669 brw_inst_set_cond_modifier(p->devinfo, s, inst->conditional_mod);
1670 multiple_instructions_emitted = true;
1671 }
1672 } else {
1673 brw_LRP(p, dst, src[0], src[1], src[2]);
1674 }
1675 brw_set_default_access_mode(p, BRW_ALIGN_1);
1676 break;
1677
1678 case BRW_OPCODE_FRC:
1679 brw_FRC(p, dst, src[0]);
1680 break;
1681 case BRW_OPCODE_RNDD:
1682 brw_RNDD(p, dst, src[0]);
1683 break;
1684 case BRW_OPCODE_RNDE:
1685 brw_RNDE(p, dst, src[0]);
1686 break;
1687 case BRW_OPCODE_RNDZ:
1688 brw_RNDZ(p, dst, src[0]);
1689 break;
1690
1691 case BRW_OPCODE_AND:
1692 brw_AND(p, dst, src[0], src[1]);
1693 break;
1694 case BRW_OPCODE_OR:
1695 brw_OR(p, dst, src[0], src[1]);
1696 break;
1697 case BRW_OPCODE_XOR:
1698 brw_XOR(p, dst, src[0], src[1]);
1699 break;
1700 case BRW_OPCODE_NOT:
1701 brw_NOT(p, dst, src[0]);
1702 break;
1703 case BRW_OPCODE_ASR:
1704 brw_ASR(p, dst, src[0], src[1]);
1705 break;
1706 case BRW_OPCODE_SHR:
1707 brw_SHR(p, dst, src[0], src[1]);
1708 break;
1709 case BRW_OPCODE_SHL:
1710 brw_SHL(p, dst, src[0], src[1]);
1711 break;
1712 case BRW_OPCODE_F32TO16:
1713 assert(devinfo->gen >= 7);
1714 brw_F32TO16(p, dst, src[0]);
1715 break;
1716 case BRW_OPCODE_F16TO32:
1717 assert(devinfo->gen >= 7);
1718 brw_F16TO32(p, dst, src[0]);
1719 break;
1720 case BRW_OPCODE_CMP:
1721 /* The Ivybridge/BayTrail WaCMPInstFlagDepClearedEarly workaround says
1722 * that when the destination is a GRF that the dependency-clear bit on
1723 * the flag register is cleared early.
1724 *
1725 * Suggested workarounds are to disable coissuing CMP instructions
1726 * or to split CMP(16) instructions into two CMP(8) instructions.
1727 *
1728 * We choose to split into CMP(8) instructions since disabling
1729 * coissuing would affect CMP instructions not otherwise affected by
1730 * the errata.
1731 */
1732 if (dispatch_width == 16 && devinfo->gen == 7 && !devinfo->is_haswell) {
1733 if (dst.file == BRW_GENERAL_REGISTER_FILE) {
1734 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1735 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1736 brw_CMP(p, firsthalf(dst), inst->conditional_mod,
1737 firsthalf(src[0]), firsthalf(src[1]));
1738 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1739 brw_CMP(p, sechalf(dst), inst->conditional_mod,
1740 sechalf(src[0]), sechalf(src[1]));
1741 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1742
1743 multiple_instructions_emitted = true;
1744 } else if (dst.file == BRW_ARCHITECTURE_REGISTER_FILE) {
1745 /* For unknown reasons, the aforementioned workaround is not
1746 * sufficient. Overriding the type when the destination is the
1747 * null register is necessary but not sufficient by itself.
1748 */
1749 assert(dst.nr == BRW_ARF_NULL);
1750 dst.type = BRW_REGISTER_TYPE_D;
1751 brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
1752 } else {
1753 unreachable("not reached");
1754 }
1755 } else {
1756 brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
1757 }
1758 break;
1759 case BRW_OPCODE_SEL:
1760 brw_SEL(p, dst, src[0], src[1]);
1761 break;
1762 case BRW_OPCODE_BFREV:
1763 assert(devinfo->gen >= 7);
1764 /* BFREV only supports UD type for src and dst. */
1765 brw_BFREV(p, retype(dst, BRW_REGISTER_TYPE_UD),
1766 retype(src[0], BRW_REGISTER_TYPE_UD));
1767 break;
1768 case BRW_OPCODE_FBH:
1769 assert(devinfo->gen >= 7);
1770 /* FBH only supports UD type for dst. */
1771 brw_FBH(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1772 break;
1773 case BRW_OPCODE_FBL:
1774 assert(devinfo->gen >= 7);
1775 /* FBL only supports UD type for dst. */
1776 brw_FBL(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1777 break;
1778 case BRW_OPCODE_CBIT:
1779 assert(devinfo->gen >= 7);
1780 /* CBIT only supports UD type for dst. */
1781 brw_CBIT(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1782 break;
1783 case BRW_OPCODE_ADDC:
1784 assert(devinfo->gen >= 7);
1785 brw_ADDC(p, dst, src[0], src[1]);
1786 break;
1787 case BRW_OPCODE_SUBB:
1788 assert(devinfo->gen >= 7);
1789 brw_SUBB(p, dst, src[0], src[1]);
1790 break;
1791 case BRW_OPCODE_MAC:
1792 brw_MAC(p, dst, src[0], src[1]);
1793 break;
1794
1795 case BRW_OPCODE_BFE:
1796 assert(devinfo->gen >= 7);
1797 brw_set_default_access_mode(p, BRW_ALIGN_16);
1798 if (dispatch_width == 16 && !devinfo->supports_simd16_3src) {
1799 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1800 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1801 brw_BFE(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1802 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1803 brw_BFE(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1804 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1805 } else {
1806 brw_BFE(p, dst, src[0], src[1], src[2]);
1807 }
1808 brw_set_default_access_mode(p, BRW_ALIGN_1);
1809 break;
1810
1811 case BRW_OPCODE_BFI1:
1812 assert(devinfo->gen >= 7);
1813 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
1814 * should
1815 *
1816 * "Force BFI instructions to be executed always in SIMD8."
1817 */
1818 if (dispatch_width == 16 && devinfo->is_haswell) {
1819 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1820 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1821 brw_BFI1(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]));
1822 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1823 brw_BFI1(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]));
1824 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1825 } else {
1826 brw_BFI1(p, dst, src[0], src[1]);
1827 }
1828 break;
1829 case BRW_OPCODE_BFI2:
1830 assert(devinfo->gen >= 7);
1831 brw_set_default_access_mode(p, BRW_ALIGN_16);
1832 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
1833 * should
1834 *
1835 * "Force BFI instructions to be executed always in SIMD8."
1836 *
1837 * Otherwise we would be able to emit compressed instructions like we
1838 * do for the other three-source instructions.
1839 */
1840 if (dispatch_width == 16 &&
1841 (devinfo->is_haswell || !devinfo->supports_simd16_3src)) {
1842 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1843 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1844 brw_BFI2(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1845 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1846 brw_BFI2(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1847 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1848 } else {
1849 brw_BFI2(p, dst, src[0], src[1], src[2]);
1850 }
1851 brw_set_default_access_mode(p, BRW_ALIGN_1);
1852 break;
1853
1854 case BRW_OPCODE_IF:
1855 if (inst->src[0].file != BAD_FILE) {
1856 /* The instruction has an embedded compare (only allowed on gen6) */
1857 assert(devinfo->gen == 6);
1858 gen6_IF(p, inst->conditional_mod, src[0], src[1]);
1859 } else {
1860 brw_IF(p, dispatch_width == 16 ? BRW_EXECUTE_16 : BRW_EXECUTE_8);
1861 }
1862 break;
1863
1864 case BRW_OPCODE_ELSE:
1865 brw_ELSE(p);
1866 break;
1867 case BRW_OPCODE_ENDIF:
1868 brw_ENDIF(p);
1869 break;
1870
1871 case BRW_OPCODE_DO:
1872 brw_DO(p, BRW_EXECUTE_8);
1873 break;
1874
1875 case BRW_OPCODE_BREAK:
1876 brw_BREAK(p);
1877 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
1878 break;
1879 case BRW_OPCODE_CONTINUE:
1880 brw_CONT(p);
1881 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
1882 break;
1883
1884 case BRW_OPCODE_WHILE:
1885 brw_WHILE(p);
1886 loop_count++;
1887 break;
1888
1889 case SHADER_OPCODE_RCP:
1890 case SHADER_OPCODE_RSQ:
1891 case SHADER_OPCODE_SQRT:
1892 case SHADER_OPCODE_EXP2:
1893 case SHADER_OPCODE_LOG2:
1894 case SHADER_OPCODE_SIN:
1895 case SHADER_OPCODE_COS:
1896 assert(devinfo->gen < 6 || inst->mlen == 0);
1897 assert(inst->conditional_mod == BRW_CONDITIONAL_NONE);
1898 if (devinfo->gen >= 7) {
1899 gen6_math(p, dst, brw_math_function(inst->opcode), src[0],
1900 brw_null_reg());
1901 } else if (devinfo->gen == 6) {
1902 generate_math_gen6(inst, dst, src[0], brw_null_reg());
1903 } else if (devinfo->gen == 5 || devinfo->is_g4x) {
1904 generate_math_g45(inst, dst, src[0]);
1905 } else {
1906 generate_math_gen4(inst, dst, src[0]);
1907 }
1908 break;
1909 case SHADER_OPCODE_INT_QUOTIENT:
1910 case SHADER_OPCODE_INT_REMAINDER:
1911 case SHADER_OPCODE_POW:
1912 assert(devinfo->gen < 6 || inst->mlen == 0);
1913 assert(inst->conditional_mod == BRW_CONDITIONAL_NONE);
1914 if (devinfo->gen >= 7 && inst->opcode == SHADER_OPCODE_POW) {
1915 gen6_math(p, dst, brw_math_function(inst->opcode), src[0], src[1]);
1916 } else if (devinfo->gen >= 6) {
1917 generate_math_gen6(inst, dst, src[0], src[1]);
1918 } else {
1919 generate_math_gen4(inst, dst, src[0]);
1920 }
1921 break;
1922 case FS_OPCODE_CINTERP:
1923 brw_MOV(p, dst, src[0]);
1924 break;
1925 case FS_OPCODE_LINTERP:
1926 generate_linterp(inst, dst, src);
1927 break;
1928 case FS_OPCODE_PIXEL_X:
1929 assert(src[0].type == BRW_REGISTER_TYPE_UW);
1930 src[0].subnr = 0 * type_sz(src[0].type);
1931 brw_MOV(p, dst, stride(src[0], 8, 4, 1));
1932 break;
1933 case FS_OPCODE_PIXEL_Y:
1934 assert(src[0].type == BRW_REGISTER_TYPE_UW);
1935 src[0].subnr = 4 * type_sz(src[0].type);
1936 brw_MOV(p, dst, stride(src[0], 8, 4, 1));
1937 break;
1938 case SHADER_OPCODE_TEX:
1939 case FS_OPCODE_TXB:
1940 case SHADER_OPCODE_TXD:
1941 case SHADER_OPCODE_TXF:
1942 case SHADER_OPCODE_TXF_CMS:
1943 case SHADER_OPCODE_TXF_UMS:
1944 case SHADER_OPCODE_TXF_MCS:
1945 case SHADER_OPCODE_TXL:
1946 case SHADER_OPCODE_TXS:
1947 case SHADER_OPCODE_LOD:
1948 case SHADER_OPCODE_TG4:
1949 case SHADER_OPCODE_TG4_OFFSET:
1950 generate_tex(inst, dst, src[0], src[1]);
1951 break;
1952 case FS_OPCODE_DDX_COARSE:
1953 case FS_OPCODE_DDX_FINE:
1954 generate_ddx(inst->opcode, dst, src[0]);
1955 break;
1956 case FS_OPCODE_DDY_COARSE:
1957 case FS_OPCODE_DDY_FINE:
1958 assert(src[1].file == BRW_IMMEDIATE_VALUE);
1959 generate_ddy(inst->opcode, dst, src[0], src[1].dw1.ud);
1960 break;
1961
1962 case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
1963 generate_scratch_write(inst, src[0]);
1964 spill_count++;
1965 break;
1966
1967 case SHADER_OPCODE_GEN4_SCRATCH_READ:
1968 generate_scratch_read(inst, dst);
1969 fill_count++;
1970 break;
1971
1972 case SHADER_OPCODE_GEN7_SCRATCH_READ:
1973 generate_scratch_read_gen7(inst, dst);
1974 fill_count++;
1975 break;
1976
1977 case SHADER_OPCODE_URB_WRITE_SIMD8:
1978 generate_urb_write(inst, src[0]);
1979 break;
1980
1981 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
1982 generate_uniform_pull_constant_load(inst, dst, src[0], src[1]);
1983 break;
1984
1985 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7:
1986 generate_uniform_pull_constant_load_gen7(inst, dst, src[0], src[1]);
1987 break;
1988
1989 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
1990 generate_varying_pull_constant_load(inst, dst, src[0], src[1]);
1991 break;
1992
1993 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
1994 generate_varying_pull_constant_load_gen7(inst, dst, src[0], src[1]);
1995 break;
1996
1997 case FS_OPCODE_REP_FB_WRITE:
1998 case FS_OPCODE_FB_WRITE:
1999 generate_fb_write(inst, src[0]);
2000 break;
2001
2002 case FS_OPCODE_BLORP_FB_WRITE:
2003 generate_blorp_fb_write(inst);
2004 break;
2005
2006 case FS_OPCODE_MOV_DISPATCH_TO_FLAGS:
2007 generate_mov_dispatch_to_flags(inst);
2008 break;
2009
2010 case FS_OPCODE_DISCARD_JUMP:
2011 generate_discard_jump(inst);
2012 break;
2013
2014 case SHADER_OPCODE_SHADER_TIME_ADD:
2015 generate_shader_time_add(inst, src[0], src[1], src[2]);
2016 break;
2017
2018 case SHADER_OPCODE_UNTYPED_ATOMIC:
2019 assert(src[1].file == BRW_IMMEDIATE_VALUE &&
2020 src[2].file == BRW_IMMEDIATE_VALUE);
2021 brw_untyped_atomic(p, dst, src[0], src[1], src[2].dw1.ud,
2022 inst->mlen, !inst->dst.is_null());
2023 brw_mark_surface_used(prog_data, src[1].dw1.ud);
2024 break;
2025
2026 case SHADER_OPCODE_UNTYPED_SURFACE_READ:
2027 assert(src[1].file == BRW_IMMEDIATE_VALUE &&
2028 src[2].file == BRW_IMMEDIATE_VALUE);
2029 brw_untyped_surface_read(p, dst, src[0], src[1],
2030 inst->mlen, src[2].dw1.ud);
2031 brw_mark_surface_used(prog_data, src[1].dw1.ud);
2032 break;
2033
2034 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE:
2035 assert(src[2].file == BRW_IMMEDIATE_VALUE);
2036 brw_untyped_surface_write(p, src[0], src[1],
2037 inst->mlen, src[2].dw1.ud);
2038 break;
2039
2040 case SHADER_OPCODE_TYPED_ATOMIC:
2041 assert(src[2].file == BRW_IMMEDIATE_VALUE);
2042 brw_typed_atomic(p, dst, src[0], src[1],
2043 src[2].dw1.ud, inst->mlen, !inst->dst.is_null());
2044 break;
2045
2046 case SHADER_OPCODE_TYPED_SURFACE_READ:
2047 assert(src[2].file == BRW_IMMEDIATE_VALUE);
2048 brw_typed_surface_read(p, dst, src[0], src[1],
2049 inst->mlen, src[2].dw1.ud);
2050 break;
2051
2052 case SHADER_OPCODE_TYPED_SURFACE_WRITE:
2053 assert(src[2].file == BRW_IMMEDIATE_VALUE);
2054 brw_typed_surface_write(p, src[0], src[1], inst->mlen, src[2].dw1.ud);
2055 break;
2056
2057 case SHADER_OPCODE_MEMORY_FENCE:
2058 brw_memory_fence(p, dst);
2059 break;
2060
2061 case FS_OPCODE_SET_SIMD4X2_OFFSET:
2062 generate_set_simd4x2_offset(inst, dst, src[0]);
2063 break;
2064
2065 case SHADER_OPCODE_FIND_LIVE_CHANNEL:
2066 brw_find_live_channel(p, dst);
2067 break;
2068
2069 case SHADER_OPCODE_BROADCAST:
2070 brw_broadcast(p, dst, src[0], src[1]);
2071 break;
2072
2073 case FS_OPCODE_SET_OMASK:
2074 generate_set_omask(inst, dst, src[0]);
2075 break;
2076
2077 case FS_OPCODE_SET_SAMPLE_ID:
2078 generate_set_sample_id(inst, dst, src[0], src[1]);
2079 break;
2080
2081 case FS_OPCODE_PACK_HALF_2x16_SPLIT:
2082 generate_pack_half_2x16_split(inst, dst, src[0], src[1]);
2083 break;
2084
2085 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X:
2086 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y:
2087 generate_unpack_half_2x16_split(inst, dst, src[0]);
2088 break;
2089
2090 case FS_OPCODE_PLACEHOLDER_HALT:
2091 /* This is the place where the final HALT needs to be inserted if
2092 * we've emitted any discards. If not, this will emit no code.
2093 */
2094 if (!patch_discard_jumps_to_fb_writes()) {
2095 if (unlikely(debug_flag)) {
2096 annotation.ann_count--;
2097 }
2098 }
2099 break;
2100
2101 case FS_OPCODE_INTERPOLATE_AT_CENTROID:
2102 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2103 GEN7_PIXEL_INTERPOLATOR_LOC_CENTROID);
2104 break;
2105
2106 case FS_OPCODE_INTERPOLATE_AT_SAMPLE:
2107 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2108 GEN7_PIXEL_INTERPOLATOR_LOC_SAMPLE);
2109 break;
2110
2111 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET:
2112 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2113 GEN7_PIXEL_INTERPOLATOR_LOC_SHARED_OFFSET);
2114 break;
2115
2116 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET:
2117 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2118 GEN7_PIXEL_INTERPOLATOR_LOC_PER_SLOT_OFFSET);
2119 break;
2120
2121 case CS_OPCODE_CS_TERMINATE:
2122 generate_cs_terminate(inst, src[0]);
2123 break;
2124
2125 default:
2126 unreachable("Unsupported opcode");
2127
2128 case SHADER_OPCODE_LOAD_PAYLOAD:
2129 unreachable("Should be lowered by lower_load_payload()");
2130 }
2131
2132 if (multiple_instructions_emitted)
2133 continue;
2134
2135 if (inst->no_dd_clear || inst->no_dd_check || inst->conditional_mod) {
2136 assert(p->next_insn_offset == last_insn_offset + 16 ||
2137 !"conditional_mod, no_dd_check, or no_dd_clear set for IR "
2138 "emitting more than 1 instruction");
2139
2140 brw_inst *last = &p->store[last_insn_offset / 16];
2141
2142 if (inst->conditional_mod)
2143 brw_inst_set_cond_modifier(p->devinfo, last, inst->conditional_mod);
2144 brw_inst_set_no_dd_clear(p->devinfo, last, inst->no_dd_clear);
2145 brw_inst_set_no_dd_check(p->devinfo, last, inst->no_dd_check);
2146 }
2147 }
2148
2149 brw_set_uip_jip(p);
2150 annotation_finalize(&annotation, p->next_insn_offset);
2151
2152 int before_size = p->next_insn_offset - start_offset;
2153 brw_compact_instructions(p, start_offset, annotation.ann_count,
2154 annotation.ann);
2155 int after_size = p->next_insn_offset - start_offset;
2156
2157 if (unlikely(debug_flag)) {
2158 fprintf(stderr, "Native code for %s\n"
2159 "SIMD%d shader: %d instructions. %d loops. %d:%d spills:fills. Promoted %u constants. Compacted %d to %d"
2160 " bytes (%.0f%%)\n",
2161 shader_name, dispatch_width, before_size / 16, loop_count,
2162 spill_count, fill_count, promoted_constants, before_size, after_size,
2163 100.0f * (before_size - after_size) / before_size);
2164
2165 dump_assembly(p->store, annotation.ann_count, annotation.ann,
2166 p->devinfo, prog);
2167 ralloc_free(annotation.ann);
2168 }
2169
2170 static GLuint msg_id = 0;
2171 _mesa_gl_debug(&brw->ctx, &msg_id,
2172 MESA_DEBUG_SOURCE_SHADER_COMPILER,
2173 MESA_DEBUG_TYPE_OTHER,
2174 MESA_DEBUG_SEVERITY_NOTIFICATION,
2175 "%s SIMD%d shader: %d inst, %d loops, %d:%d spills:fills, "
2176 "Promoted %u constants, compacted %d to %d bytes.\n",
2177 stage_abbrev, dispatch_width, before_size / 16, loop_count,
2178 spill_count, fill_count, promoted_constants, before_size, after_size);
2179
2180 return start_offset;
2181 }
2182
2183 const unsigned *
2184 fs_generator::get_assembly(unsigned int *assembly_size)
2185 {
2186 return brw_get_program(p, assembly_size);
2187 }