projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
i965/fs: Fix stride for immediate registers.
[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_inst *inst, 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 (inst->exec_size < 8) {
61 brw_reg = brw_vec8_reg(brw_file_from_reg(reg), reg->reg, 0);
62 brw_reg = stride(brw_reg, inst->exec_size * reg->stride,
63 inst->exec_size, 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 assert(reg->stride == ((reg->type == BRW_REGISTER_TYPE_V ||
83 reg->type == BRW_REGISTER_TYPE_UV ||
84 reg->type == BRW_REGISTER_TYPE_VF) ? 1 : 0));
85
86 switch (reg->type) {
87 case BRW_REGISTER_TYPE_F:
88 brw_reg = brw_imm_f(reg->fixed_hw_reg.dw1.f);
89 break;
90 case BRW_REGISTER_TYPE_D:
91 brw_reg = brw_imm_d(reg->fixed_hw_reg.dw1.d);
92 break;
93 case BRW_REGISTER_TYPE_UD:
94 brw_reg = brw_imm_ud(reg->fixed_hw_reg.dw1.ud);
95 break;
96 case BRW_REGISTER_TYPE_W:
97 brw_reg = brw_imm_w(reg->fixed_hw_reg.dw1.d);
98 break;
99 case BRW_REGISTER_TYPE_UW:
100 brw_reg = brw_imm_uw(reg->fixed_hw_reg.dw1.ud);
101 break;
102 case BRW_REGISTER_TYPE_VF:
103 brw_reg = brw_imm_vf(reg->fixed_hw_reg.dw1.ud);
104 break;
105 default:
106 unreachable("not reached");
107 }
108 break;
109 case HW_REG:
110 assert(reg->type == reg->fixed_hw_reg.type);
111 brw_reg = reg->fixed_hw_reg;
112 break;
113 case BAD_FILE:
114 /* Probably unused. */
115 brw_reg = brw_null_reg();
116 break;
117 default:
118 unreachable("not reached");
119 }
120 if (reg->abs)
121 brw_reg = brw_abs(brw_reg);
122 if (reg->negate)
123 brw_reg = negate(brw_reg);
124
125 return brw_reg;
126 }
127
128 fs_generator::fs_generator(const struct brw_compiler *compiler, void *log_data,
129 void *mem_ctx,
130 const void *key,
131 struct brw_stage_prog_data *prog_data,
132 struct gl_program *prog,
133 unsigned promoted_constants,
134 bool runtime_check_aads_emit,
135 const char *stage_abbrev)
136
137 : compiler(compiler), log_data(log_data),
138 devinfo(compiler->devinfo), key(key),
139 prog_data(prog_data),
140 prog(prog), promoted_constants(promoted_constants),
141 runtime_check_aads_emit(runtime_check_aads_emit), debug_flag(false),
142 stage_abbrev(stage_abbrev), mem_ctx(mem_ctx)
143 {
144 p = rzalloc(mem_ctx, struct brw_codegen);
145 brw_init_codegen(devinfo, p, mem_ctx);
146 }
147
148 fs_generator::~fs_generator()
149 {
150 }
151
152 class ip_record : public exec_node {
153 public:
154 DECLARE_RALLOC_CXX_OPERATORS(ip_record)
155
156 ip_record(int ip)
157 {
158 this->ip = ip;
159 }
160
161 int ip;
162 };
163
164 bool
165 fs_generator::patch_discard_jumps_to_fb_writes()
166 {
167 if (devinfo->gen < 6 || this->discard_halt_patches.is_empty())
168 return false;
169
170 int scale = brw_jump_scale(p->devinfo);
171
172 /* There is a somewhat strange undocumented requirement of using
173 * HALT, according to the simulator. If some channel has HALTed to
174 * a particular UIP, then by the end of the program, every channel
175 * must have HALTed to that UIP. Furthermore, the tracking is a
176 * stack, so you can't do the final halt of a UIP after starting
177 * halting to a new UIP.
178 *
179 * Symptoms of not emitting this instruction on actual hardware
180 * included GPU hangs and sparkly rendering on the piglit discard
181 * tests.
182 */
183 brw_inst *last_halt = gen6_HALT(p);
184 brw_inst_set_uip(p->devinfo, last_halt, 1 * scale);
185 brw_inst_set_jip(p->devinfo, last_halt, 1 * scale);
186
187 int ip = p->nr_insn;
188
189 foreach_in_list(ip_record, patch_ip, &discard_halt_patches) {
190 brw_inst *patch = &p->store[patch_ip->ip];
191
192 assert(brw_inst_opcode(p->devinfo, patch) == BRW_OPCODE_HALT);
193 /* HALT takes a half-instruction distance from the pre-incremented IP. */
194 brw_inst_set_uip(p->devinfo, patch, (ip - patch_ip->ip) * scale);
195 }
196
197 this->discard_halt_patches.make_empty();
198 return true;
199 }
200
201 void
202 fs_generator::fire_fb_write(fs_inst *inst,
203 struct brw_reg payload,
204 struct brw_reg implied_header,
205 GLuint nr)
206 {
207 uint32_t msg_control;
208
209 brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data;
210
211 if (devinfo->gen < 6) {
212 brw_push_insn_state(p);
213 brw_set_default_exec_size(p, BRW_EXECUTE_8);
214 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
215 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
216 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
217 brw_MOV(p, offset(payload, 1), brw_vec8_grf(1, 0));
218 brw_pop_insn_state(p);
219 }
220
221 if (inst->opcode == FS_OPCODE_REP_FB_WRITE)
222 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE_REPLICATED;
223 else if (prog_data->dual_src_blend) {
224 if (dispatch_width == 8 || !inst->eot)
225 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN01;
226 else
227 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN23;
228 } else if (dispatch_width == 16)
229 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE;
230 else
231 msg_control = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01;
232
233 uint32_t surf_index =
234 prog_data->binding_table.render_target_start + inst->target;
235
236 bool last_render_target = inst->eot ||
237 (prog_data->dual_src_blend && dispatch_width == 16);
238
239
240 brw_fb_WRITE(p,
241 dispatch_width,
242 payload,
243 implied_header,
244 msg_control,
245 surf_index,
246 nr,
247 0,
248 inst->eot,
249 last_render_target,
250 inst->header_size != 0);
251
252 brw_mark_surface_used(&prog_data->base, surf_index);
253 }
254
255 void
256 fs_generator::generate_fb_write(fs_inst *inst, struct brw_reg payload)
257 {
258 brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data;
259 const brw_wm_prog_key * const key = (brw_wm_prog_key * const) this->key;
260 struct brw_reg implied_header;
261
262 if (devinfo->gen < 8 && !devinfo->is_haswell) {
263 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
264 }
265
266 if (inst->base_mrf >= 0)
267 payload = brw_message_reg(inst->base_mrf);
268
269 /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
270 * move, here's g1.
271 */
272 if (inst->header_size != 0) {
273 brw_push_insn_state(p);
274 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
275 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
276 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
277 brw_set_default_flag_reg(p, 0, 0);
278
279 /* On HSW, the GPU will use the predicate on SENDC, unless the header is
280 * present.
281 */
282 if (prog_data->uses_kill) {
283 struct brw_reg pixel_mask;
284
285 if (devinfo->gen >= 6)
286 pixel_mask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
287 else
288 pixel_mask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
289
290 brw_MOV(p, pixel_mask, brw_flag_reg(0, 1));
291 }
292
293 if (devinfo->gen >= 6) {
294 brw_push_insn_state(p);
295 brw_set_default_exec_size(p, BRW_EXECUTE_16);
296 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
297 brw_MOV(p,
298 retype(payload, BRW_REGISTER_TYPE_UD),
299 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
300 brw_pop_insn_state(p);
301
302 if (inst->target > 0 && key->replicate_alpha) {
303 /* Set "Source0 Alpha Present to RenderTarget" bit in message
304 * header.
305 */
306 brw_OR(p,
307 vec1(retype(payload, BRW_REGISTER_TYPE_UD)),
308 vec1(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD)),
309 brw_imm_ud(0x1 << 11));
310 }
311
312 if (inst->target > 0) {
313 /* Set the render target index for choosing BLEND_STATE. */
314 brw_MOV(p, retype(vec1(suboffset(payload, 2)),
315 BRW_REGISTER_TYPE_UD),
316 brw_imm_ud(inst->target));
317 }
318
319 implied_header = brw_null_reg();
320 } else {
321 implied_header = retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW);
322 }
323
324 brw_pop_insn_state(p);
325 } else {
326 implied_header = brw_null_reg();
327 }
328
329 if (!runtime_check_aads_emit) {
330 fire_fb_write(inst, payload, implied_header, inst->mlen);
331 } else {
332 /* This can only happen in gen < 6 */
333 assert(devinfo->gen < 6);
334
335 struct brw_reg v1_null_ud = vec1(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD));
336
337 /* Check runtime bit to detect if we have to send AA data or not */
338 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
339 brw_AND(p,
340 v1_null_ud,
341 retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD),
342 brw_imm_ud(1<<26));
343 brw_inst_set_cond_modifier(p->devinfo, brw_last_inst, BRW_CONDITIONAL_NZ);
344
345 int jmp = brw_JMPI(p, brw_imm_ud(0), BRW_PREDICATE_NORMAL) - p->store;
346 brw_inst_set_exec_size(p->devinfo, brw_last_inst, BRW_EXECUTE_1);
347 {
348 /* Don't send AA data */
349 fire_fb_write(inst, offset(payload, 1), implied_header, inst->mlen-1);
350 }
351 brw_land_fwd_jump(p, jmp);
352 fire_fb_write(inst, payload, implied_header, inst->mlen);
353 }
354 }
355
356 void
357 fs_generator::generate_urb_write(fs_inst *inst, struct brw_reg payload)
358 {
359 brw_inst *insn;
360
361 insn = brw_next_insn(p, BRW_OPCODE_SEND);
362
363 brw_set_dest(p, insn, brw_null_reg());
364 brw_set_src0(p, insn, payload);
365 brw_set_src1(p, insn, brw_imm_d(0));
366
367 brw_inst_set_sfid(p->devinfo, insn, BRW_SFID_URB);
368 brw_inst_set_urb_opcode(p->devinfo, insn, GEN8_URB_OPCODE_SIMD8_WRITE);
369
370 brw_inst_set_mlen(p->devinfo, insn, inst->mlen);
371 brw_inst_set_rlen(p->devinfo, insn, 0);
372 brw_inst_set_eot(p->devinfo, insn, inst->eot);
373 brw_inst_set_header_present(p->devinfo, insn, true);
374 brw_inst_set_urb_global_offset(p->devinfo, insn, inst->offset);
375 }
376
377 void
378 fs_generator::generate_cs_terminate(fs_inst *inst, struct brw_reg payload)
379 {
380 struct brw_inst *insn;
381
382 insn = brw_next_insn(p, BRW_OPCODE_SEND);
383
384 brw_set_dest(p, insn, brw_null_reg());
385 brw_set_src0(p, insn, payload);
386 brw_set_src1(p, insn, brw_imm_d(0));
387
388 /* Terminate a compute shader by sending a message to the thread spawner.
389 */
390 brw_inst_set_sfid(devinfo, insn, BRW_SFID_THREAD_SPAWNER);
391 brw_inst_set_mlen(devinfo, insn, 1);
392 brw_inst_set_rlen(devinfo, insn, 0);
393 brw_inst_set_eot(devinfo, insn, inst->eot);
394 brw_inst_set_header_present(devinfo, insn, false);
395
396 brw_inst_set_ts_opcode(devinfo, insn, 0); /* Dereference resource */
397 brw_inst_set_ts_request_type(devinfo, insn, 0); /* Root thread */
398
399 /* Note that even though the thread has a URB resource associated with it,
400 * we set the "do not dereference URB" bit, because the URB resource is
401 * managed by the fixed-function unit, so it will free it automatically.
402 */
403 brw_inst_set_ts_resource_select(devinfo, insn, 1); /* Do not dereference URB */
404
405 brw_inst_set_mask_control(devinfo, insn, BRW_MASK_DISABLE);
406 }
407
408 void
409 fs_generator::generate_barrier(fs_inst *inst, struct brw_reg src)
410 {
411 brw_barrier(p, src);
412 brw_WAIT(p);
413 }
414
415 void
416 fs_generator::generate_blorp_fb_write(fs_inst *inst)
417 {
418 brw_fb_WRITE(p,
419 16 /* dispatch_width */,
420 brw_message_reg(inst->base_mrf),
421 brw_reg_from_fs_reg(inst, &inst->src[0]),
422 BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE,
423 inst->target,
424 inst->mlen,
425 0,
426 true,
427 true,
428 inst->header_size != 0);
429 }
430
431 void
432 fs_generator::generate_linterp(fs_inst *inst,
433 struct brw_reg dst, struct brw_reg *src)
434 {
435 /* PLN reads:
436 * / in SIMD16 \
437 * -----------------------------------
438 * | src1+0 | src1+1 | src1+2 | src1+3 |
439 * |-----------------------------------|
440 * |(x0, x1)|(y0, y1)|(x2, x3)|(y2, y3)|
441 * -----------------------------------
442 *
443 * but for the LINE/MAC pair, the LINE reads Xs and the MAC reads Ys:
444 *
445 * -----------------------------------
446 * | src1+0 | src1+1 | src1+2 | src1+3 |
447 * |-----------------------------------|
448 * |(x0, x1)|(y0, y1)| | | in SIMD8
449 * |-----------------------------------|
450 * |(x0, x1)|(x2, x3)|(y0, y1)|(y2, y3)| in SIMD16
451 * -----------------------------------
452 *
453 * See also: emit_interpolation_setup_gen4().
454 */
455 struct brw_reg delta_x = src[0];
456 struct brw_reg delta_y = offset(src[0], dispatch_width / 8);
457 struct brw_reg interp = src[1];
458
459 if (devinfo->has_pln &&
460 (devinfo->gen >= 7 || (delta_x.nr & 1) == 0)) {
461 brw_PLN(p, dst, interp, delta_x);
462 } else {
463 brw_LINE(p, brw_null_reg(), interp, delta_x);
464 brw_MAC(p, dst, suboffset(interp, 1), delta_y);
465 }
466 }
467
468 void
469 fs_generator::generate_math_gen6(fs_inst *inst,
470 struct brw_reg dst,
471 struct brw_reg src0,
472 struct brw_reg src1)
473 {
474 int op = brw_math_function(inst->opcode);
475 bool binop = src1.file != BRW_ARCHITECTURE_REGISTER_FILE;
476
477 if (dispatch_width == 8) {
478 gen6_math(p, dst, op, src0, src1);
479 } else if (dispatch_width == 16) {
480 brw_push_insn_state(p);
481 brw_set_default_exec_size(p, BRW_EXECUTE_8);
482 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
483 gen6_math(p, firsthalf(dst), op, firsthalf(src0), firsthalf(src1));
484 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
485 gen6_math(p, sechalf(dst), op, sechalf(src0),
486 binop ? sechalf(src1) : brw_null_reg());
487 brw_pop_insn_state(p);
488 }
489 }
490
491 void
492 fs_generator::generate_math_gen4(fs_inst *inst,
493 struct brw_reg dst,
494 struct brw_reg src)
495 {
496 int op = brw_math_function(inst->opcode);
497
498 assert(inst->mlen >= 1);
499
500 if (dispatch_width == 8) {
501 gen4_math(p, dst,
502 op,
503 inst->base_mrf, src,
504 BRW_MATH_PRECISION_FULL);
505 } else if (dispatch_width == 16) {
506 brw_set_default_exec_size(p, BRW_EXECUTE_8);
507 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
508 gen4_math(p, firsthalf(dst),
509 op,
510 inst->base_mrf, firsthalf(src),
511 BRW_MATH_PRECISION_FULL);
512 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
513 gen4_math(p, sechalf(dst),
514 op,
515 inst->base_mrf + 1, sechalf(src),
516 BRW_MATH_PRECISION_FULL);
517
518 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
519 }
520 }
521
522 void
523 fs_generator::generate_math_g45(fs_inst *inst,
524 struct brw_reg dst,
525 struct brw_reg src)
526 {
527 if (inst->opcode == SHADER_OPCODE_POW ||
528 inst->opcode == SHADER_OPCODE_INT_QUOTIENT ||
529 inst->opcode == SHADER_OPCODE_INT_REMAINDER) {
530 generate_math_gen4(inst, dst, src);
531 return;
532 }
533
534 int op = brw_math_function(inst->opcode);
535
536 assert(inst->mlen >= 1);
537
538 gen4_math(p, dst,
539 op,
540 inst->base_mrf, src,
541 BRW_MATH_PRECISION_FULL);
542 }
543
544 void
545 fs_generator::generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src,
546 struct brw_reg sampler_index)
547 {
548 int msg_type = -1;
549 int rlen = 4;
550 uint32_t simd_mode;
551 uint32_t return_format;
552 bool is_combined_send = inst->eot;
553
554 switch (dst.type) {
555 case BRW_REGISTER_TYPE_D:
556 return_format = BRW_SAMPLER_RETURN_FORMAT_SINT32;
557 break;
558 case BRW_REGISTER_TYPE_UD:
559 return_format = BRW_SAMPLER_RETURN_FORMAT_UINT32;
560 break;
561 default:
562 return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
563 break;
564 }
565
566 switch (inst->exec_size) {
567 case 8:
568 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
569 break;
570 case 16:
571 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
572 break;
573 default:
574 unreachable("Invalid width for texture instruction");
575 }
576
577 if (devinfo->gen >= 5) {
578 switch (inst->opcode) {
579 case SHADER_OPCODE_TEX:
580 if (inst->shadow_compare) {
581 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE;
582 } else {
583 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE;
584 }
585 break;
586 case FS_OPCODE_TXB:
587 if (inst->shadow_compare) {
588 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE;
589 } else {
590 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS;
591 }
592 break;
593 case SHADER_OPCODE_TXL:
594 if (inst->shadow_compare) {
595 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
596 } else {
597 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
598 }
599 break;
600 case SHADER_OPCODE_TXS:
601 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
602 break;
603 case SHADER_OPCODE_TXD:
604 if (inst->shadow_compare) {
605 /* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
606 assert(devinfo->gen >= 8 || devinfo->is_haswell);
607 msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE;
608 } else {
609 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
610 }
611 break;
612 case SHADER_OPCODE_TXF:
613 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
614 break;
615 case SHADER_OPCODE_TXF_CMS:
616 if (devinfo->gen >= 7)
617 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
618 else
619 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
620 break;
621 case SHADER_OPCODE_TXF_UMS:
622 assert(devinfo->gen >= 7);
623 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS;
624 break;
625 case SHADER_OPCODE_TXF_MCS:
626 assert(devinfo->gen >= 7);
627 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS;
628 break;
629 case SHADER_OPCODE_LOD:
630 msg_type = GEN5_SAMPLER_MESSAGE_LOD;
631 break;
632 case SHADER_OPCODE_TG4:
633 if (inst->shadow_compare) {
634 assert(devinfo->gen >= 7);
635 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C;
636 } else {
637 assert(devinfo->gen >= 6);
638 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4;
639 }
640 break;
641 case SHADER_OPCODE_TG4_OFFSET:
642 assert(devinfo->gen >= 7);
643 if (inst->shadow_compare) {
644 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C;
645 } else {
646 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO;
647 }
648 break;
649 default:
650 unreachable("not reached");
651 }
652 } else {
653 switch (inst->opcode) {
654 case SHADER_OPCODE_TEX:
655 /* Note that G45 and older determines shadow compare and dispatch width
656 * from message length for most messages.
657 */
658 if (dispatch_width == 8) {
659 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE;
660 if (inst->shadow_compare) {
661 assert(inst->mlen == 6);
662 } else {
663 assert(inst->mlen <= 4);
664 }
665 } else {
666 if (inst->shadow_compare) {
667 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE;
668 assert(inst->mlen == 9);
669 } else {
670 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE;
671 assert(inst->mlen <= 7 && inst->mlen % 2 == 1);
672 }
673 }
674 break;
675 case FS_OPCODE_TXB:
676 if (inst->shadow_compare) {
677 assert(dispatch_width == 8);
678 assert(inst->mlen == 6);
679 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_BIAS_COMPARE;
680 } else {
681 assert(inst->mlen == 9);
682 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS;
683 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
684 }
685 break;
686 case SHADER_OPCODE_TXL:
687 if (inst->shadow_compare) {
688 assert(dispatch_width == 8);
689 assert(inst->mlen == 6);
690 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_LOD_COMPARE;
691 } else {
692 assert(inst->mlen == 9);
693 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_LOD;
694 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
695 }
696 break;
697 case SHADER_OPCODE_TXD:
698 /* There is no sample_d_c message; comparisons are done manually */
699 assert(dispatch_width == 8);
700 assert(inst->mlen == 7 || inst->mlen == 10);
701 msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_GRADIENTS;
702 break;
703 case SHADER_OPCODE_TXF:
704 assert(inst->mlen <= 9 && inst->mlen % 2 == 1);
705 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_LD;
706 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
707 break;
708 case SHADER_OPCODE_TXS:
709 assert(inst->mlen == 3);
710 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_RESINFO;
711 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
712 break;
713 default:
714 unreachable("not reached");
715 }
716 }
717 assert(msg_type != -1);
718
719 if (simd_mode == BRW_SAMPLER_SIMD_MODE_SIMD16) {
720 rlen = 8;
721 dst = vec16(dst);
722 }
723
724 if (is_combined_send) {
725 assert(devinfo->gen >= 9 || devinfo->is_cherryview);
726 rlen = 0;
727 }
728
729 assert(devinfo->gen < 7 || inst->header_size == 0 ||
730 src.file == BRW_GENERAL_REGISTER_FILE);
731
732 assert(sampler_index.type == BRW_REGISTER_TYPE_UD);
733
734 /* Load the message header if present. If there's a texture offset,
735 * we need to set it up explicitly and load the offset bitfield.
736 * Otherwise, we can use an implied move from g0 to the first message reg.
737 */
738 if (inst->header_size != 0) {
739 if (devinfo->gen < 6 && !inst->offset) {
740 /* Set up an implied move from g0 to the MRF. */
741 src = retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW);
742 } else {
743 struct brw_reg header_reg;
744
745 if (devinfo->gen >= 7) {
746 header_reg = src;
747 } else {
748 assert(inst->base_mrf != -1);
749 header_reg = brw_message_reg(inst->base_mrf);
750 }
751
752 brw_push_insn_state(p);
753 brw_set_default_exec_size(p, BRW_EXECUTE_8);
754 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
755 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
756 /* Explicitly set up the message header by copying g0 to the MRF. */
757 brw_MOV(p, header_reg, brw_vec8_grf(0, 0));
758
759 if (inst->offset) {
760 /* Set the offset bits in DWord 2. */
761 brw_MOV(p, get_element_ud(header_reg, 2),
762 brw_imm_ud(inst->offset));
763 }
764
765 brw_adjust_sampler_state_pointer(p, header_reg, sampler_index);
766 brw_pop_insn_state(p);
767 }
768 }
769
770 uint32_t base_binding_table_index = (inst->opcode == SHADER_OPCODE_TG4 ||
771 inst->opcode == SHADER_OPCODE_TG4_OFFSET)
772 ? prog_data->binding_table.gather_texture_start
773 : prog_data->binding_table.texture_start;
774
775 if (sampler_index.file == BRW_IMMEDIATE_VALUE) {
776 uint32_t sampler = sampler_index.dw1.ud;
777
778 brw_SAMPLE(p,
779 retype(dst, BRW_REGISTER_TYPE_UW),
780 inst->base_mrf,
781 src,
782 sampler + base_binding_table_index,
783 sampler % 16,
784 msg_type,
785 rlen,
786 inst->mlen,
787 inst->header_size != 0,
788 simd_mode,
789 return_format);
790
791 brw_mark_surface_used(prog_data, sampler + base_binding_table_index);
792 } else {
793 /* Non-const sampler index */
794
795 struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
796 struct brw_reg sampler_reg = vec1(retype(sampler_index, BRW_REGISTER_TYPE_UD));
797
798 brw_push_insn_state(p);
799 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
800 brw_set_default_access_mode(p, BRW_ALIGN_1);
801
802 /* addr = ((sampler * 0x101) + base_binding_table_index) & 0xfff */
803 brw_MUL(p, addr, sampler_reg, brw_imm_uw(0x101));
804 if (base_binding_table_index)
805 brw_ADD(p, addr, addr, brw_imm_ud(base_binding_table_index));
806 brw_AND(p, addr, addr, brw_imm_ud(0xfff));
807
808 brw_pop_insn_state(p);
809
810 /* dst = send(offset, a0.0 | <descriptor>) */
811 brw_inst *insn = brw_send_indirect_message(
812 p, BRW_SFID_SAMPLER, dst, src, addr);
813 brw_set_sampler_message(p, insn,
814 0 /* surface */,
815 0 /* sampler */,
816 msg_type,
817 rlen,
818 inst->mlen /* mlen */,
819 inst->header_size != 0 /* header */,
820 simd_mode,
821 return_format);
822
823 /* visitor knows more than we do about the surface limit required,
824 * so has already done marking.
825 */
826 }
827
828 if (is_combined_send) {
829 brw_inst_set_eot(p->devinfo, brw_last_inst, true);
830 brw_inst_set_opcode(p->devinfo, brw_last_inst, BRW_OPCODE_SENDC);
831 }
832 }
833
834
835 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
836 * looking like:
837 *
838 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
839 *
840 * Ideally, we want to produce:
841 *
842 * DDX DDY
843 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
844 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
845 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
846 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
847 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
848 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
849 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
850 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
851 *
852 * and add another set of two more subspans if in 16-pixel dispatch mode.
853 *
854 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
855 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
856 * pair. But the ideal approximation may impose a huge performance cost on
857 * sample_d. On at least Haswell, sample_d instruction does some
858 * optimizations if the same LOD is used for all pixels in the subspan.
859 *
860 * For DDY, we need to use ALIGN16 mode since it's capable of doing the
861 * appropriate swizzling.
862 */
863 void
864 fs_generator::generate_ddx(enum opcode opcode,
865 struct brw_reg dst, struct brw_reg src)
866 {
867 unsigned vstride, width;
868
869 if (opcode == FS_OPCODE_DDX_FINE) {
870 /* produce accurate derivatives */
871 vstride = BRW_VERTICAL_STRIDE_2;
872 width = BRW_WIDTH_2;
873 } else {
874 /* replicate the derivative at the top-left pixel to other pixels */
875 vstride = BRW_VERTICAL_STRIDE_4;
876 width = BRW_WIDTH_4;
877 }
878
879 struct brw_reg src0 = brw_reg(src.file, src.nr, 1,
880 src.negate, src.abs,
881 BRW_REGISTER_TYPE_F,
882 vstride,
883 width,
884 BRW_HORIZONTAL_STRIDE_0,
885 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
886 struct brw_reg src1 = brw_reg(src.file, src.nr, 0,
887 src.negate, src.abs,
888 BRW_REGISTER_TYPE_F,
889 vstride,
890 width,
891 BRW_HORIZONTAL_STRIDE_0,
892 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
893 brw_ADD(p, dst, src0, negate(src1));
894 }
895
896 /* The negate_value boolean is used to negate the derivative computation for
897 * FBOs, since they place the origin at the upper left instead of the lower
898 * left.
899 */
900 void
901 fs_generator::generate_ddy(enum opcode opcode,
902 struct brw_reg dst, struct brw_reg src,
903 bool negate_value)
904 {
905 if (opcode == FS_OPCODE_DDY_FINE) {
906 /* From the Ivy Bridge PRM, volume 4 part 3, section 3.3.9 (Register
907 * Region Restrictions):
908 *
909 * In Align16 access mode, SIMD16 is not allowed for DW operations
910 * and SIMD8 is not allowed for DF operations.
911 *
912 * In this context, "DW operations" means "operations acting on 32-bit
913 * values", so it includes operations on floats.
914 *
915 * Gen4 has a similar restriction. From the i965 PRM, section 11.5.3
916 * (Instruction Compression -> Rules and Restrictions):
917 *
918 * A compressed instruction must be in Align1 access mode. Align16
919 * mode instructions cannot be compressed.
920 *
921 * Similar text exists in the g45 PRM.
922 *
923 * On these platforms, if we're building a SIMD16 shader, we need to
924 * manually unroll to a pair of SIMD8 instructions.
925 */
926 bool unroll_to_simd8 =
927 (dispatch_width == 16 &&
928 (devinfo->gen == 4 || (devinfo->gen == 7 && !devinfo->is_haswell)));
929
930 /* produce accurate derivatives */
931 struct brw_reg src0 = brw_reg(src.file, src.nr, 0,
932 src.negate, src.abs,
933 BRW_REGISTER_TYPE_F,
934 BRW_VERTICAL_STRIDE_4,
935 BRW_WIDTH_4,
936 BRW_HORIZONTAL_STRIDE_1,
937 BRW_SWIZZLE_XYXY, WRITEMASK_XYZW);
938 struct brw_reg src1 = brw_reg(src.file, src.nr, 0,
939 src.negate, src.abs,
940 BRW_REGISTER_TYPE_F,
941 BRW_VERTICAL_STRIDE_4,
942 BRW_WIDTH_4,
943 BRW_HORIZONTAL_STRIDE_1,
944 BRW_SWIZZLE_ZWZW, WRITEMASK_XYZW);
945 brw_push_insn_state(p);
946 brw_set_default_access_mode(p, BRW_ALIGN_16);
947 if (unroll_to_simd8) {
948 brw_set_default_exec_size(p, BRW_EXECUTE_8);
949 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
950 if (negate_value) {
951 brw_ADD(p, firsthalf(dst), firsthalf(src1), negate(firsthalf(src0)));
952 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
953 brw_ADD(p, sechalf(dst), sechalf(src1), negate(sechalf(src0)));
954 } else {
955 brw_ADD(p, firsthalf(dst), firsthalf(src0), negate(firsthalf(src1)));
956 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
957 brw_ADD(p, sechalf(dst), sechalf(src0), negate(sechalf(src1)));
958 }
959 } else {
960 if (negate_value)
961 brw_ADD(p, dst, src1, negate(src0));
962 else
963 brw_ADD(p, dst, src0, negate(src1));
964 }
965 brw_pop_insn_state(p);
966 } else {
967 /* replicate the derivative at the top-left pixel to other pixels */
968 struct brw_reg src0 = brw_reg(src.file, src.nr, 0,
969 src.negate, src.abs,
970 BRW_REGISTER_TYPE_F,
971 BRW_VERTICAL_STRIDE_4,
972 BRW_WIDTH_4,
973 BRW_HORIZONTAL_STRIDE_0,
974 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
975 struct brw_reg src1 = brw_reg(src.file, src.nr, 2,
976 src.negate, src.abs,
977 BRW_REGISTER_TYPE_F,
978 BRW_VERTICAL_STRIDE_4,
979 BRW_WIDTH_4,
980 BRW_HORIZONTAL_STRIDE_0,
981 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
982 if (negate_value)
983 brw_ADD(p, dst, src1, negate(src0));
984 else
985 brw_ADD(p, dst, src0, negate(src1));
986 }
987 }
988
989 void
990 fs_generator::generate_discard_jump(fs_inst *inst)
991 {
992 assert(devinfo->gen >= 6);
993
994 /* This HALT will be patched up at FB write time to point UIP at the end of
995 * the program, and at brw_uip_jip() JIP will be set to the end of the
996 * current block (or the program).
997 */
998 this->discard_halt_patches.push_tail(new(mem_ctx) ip_record(p->nr_insn));
999
1000 brw_push_insn_state(p);
1001 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1002 gen6_HALT(p);
1003 brw_pop_insn_state(p);
1004 }
1005
1006 void
1007 fs_generator::generate_scratch_write(fs_inst *inst, struct brw_reg src)
1008 {
1009 assert(inst->mlen != 0);
1010
1011 brw_MOV(p,
1012 brw_uvec_mrf(inst->exec_size, (inst->base_mrf + 1), 0),
1013 retype(src, BRW_REGISTER_TYPE_UD));
1014 brw_oword_block_write_scratch(p, brw_message_reg(inst->base_mrf),
1015 inst->exec_size / 8, inst->offset);
1016 }
1017
1018 void
1019 fs_generator::generate_scratch_read(fs_inst *inst, struct brw_reg dst)
1020 {
1021 assert(inst->mlen != 0);
1022
1023 brw_oword_block_read_scratch(p, dst, brw_message_reg(inst->base_mrf),
1024 inst->exec_size / 8, inst->offset);
1025 }
1026
1027 void
1028 fs_generator::generate_scratch_read_gen7(fs_inst *inst, struct brw_reg dst)
1029 {
1030 gen7_block_read_scratch(p, dst, inst->exec_size / 8, inst->offset);
1031 }
1032
1033 void
1034 fs_generator::generate_uniform_pull_constant_load(fs_inst *inst,
1035 struct brw_reg dst,
1036 struct brw_reg index,
1037 struct brw_reg offset)
1038 {
1039 assert(inst->mlen != 0);
1040
1041 assert(index.file == BRW_IMMEDIATE_VALUE &&
1042 index.type == BRW_REGISTER_TYPE_UD);
1043 uint32_t surf_index = index.dw1.ud;
1044
1045 assert(offset.file == BRW_IMMEDIATE_VALUE &&
1046 offset.type == BRW_REGISTER_TYPE_UD);
1047 uint32_t read_offset = offset.dw1.ud;
1048
1049 brw_oword_block_read(p, dst, brw_message_reg(inst->base_mrf),
1050 read_offset, surf_index);
1051
1052 brw_mark_surface_used(prog_data, surf_index);
1053 }
1054
1055 void
1056 fs_generator::generate_uniform_pull_constant_load_gen7(fs_inst *inst,
1057 struct brw_reg dst,
1058 struct brw_reg index,
1059 struct brw_reg offset)
1060 {
1061 assert(index.type == BRW_REGISTER_TYPE_UD);
1062
1063 assert(offset.file == BRW_GENERAL_REGISTER_FILE);
1064 /* Reference just the dword we need, to avoid angering validate_reg(). */
1065 offset = brw_vec1_grf(offset.nr, 0);
1066
1067 /* We use the SIMD4x2 mode because we want to end up with 4 components in
1068 * the destination loaded consecutively from the same offset (which appears
1069 * in the first component, and the rest are ignored).
1070 */
1071 dst.width = BRW_WIDTH_4;
1072
1073 struct brw_reg src = offset;
1074 bool header_present = false;
1075
1076 if (devinfo->gen >= 9) {
1077 /* Skylake requires a message header in order to use SIMD4x2 mode. */
1078 src = retype(brw_vec4_grf(offset.nr, 0), BRW_REGISTER_TYPE_UD);
1079 header_present = true;
1080
1081 brw_push_insn_state(p);
1082 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1083 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1084 brw_MOV(p, vec8(src), retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
1085 brw_set_default_access_mode(p, BRW_ALIGN_1);
1086
1087 brw_MOV(p, get_element_ud(src, 2),
1088 brw_imm_ud(GEN9_SAMPLER_SIMD_MODE_EXTENSION_SIMD4X2));
1089 brw_pop_insn_state(p);
1090 }
1091
1092 if (index.file == BRW_IMMEDIATE_VALUE) {
1093
1094 uint32_t surf_index = index.dw1.ud;
1095
1096 brw_push_insn_state(p);
1097 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1098 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1099 brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
1100 brw_pop_insn_state(p);
1101
1102 brw_set_dest(p, send, dst);
1103 brw_set_src0(p, send, src);
1104 brw_set_sampler_message(p, send,
1105 surf_index,
1106 0, /* LD message ignores sampler unit */
1107 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1108 1, /* rlen */
1109 inst->mlen,
1110 header_present,
1111 BRW_SAMPLER_SIMD_MODE_SIMD4X2,
1112 0);
1113
1114 brw_mark_surface_used(prog_data, surf_index);
1115
1116 } else {
1117
1118 struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
1119
1120 brw_push_insn_state(p);
1121 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1122 brw_set_default_access_mode(p, BRW_ALIGN_1);
1123
1124 /* a0.0 = surf_index & 0xff */
1125 brw_inst *insn_and = brw_next_insn(p, BRW_OPCODE_AND);
1126 brw_inst_set_exec_size(p->devinfo, insn_and, BRW_EXECUTE_1);
1127 brw_set_dest(p, insn_and, addr);
1128 brw_set_src0(p, insn_and, vec1(retype(index, BRW_REGISTER_TYPE_UD)));
1129 brw_set_src1(p, insn_and, brw_imm_ud(0x0ff));
1130
1131 /* dst = send(payload, a0.0 | <descriptor>) */
1132 brw_inst *insn = brw_send_indirect_message(
1133 p, BRW_SFID_SAMPLER, dst, src, addr);
1134 brw_set_sampler_message(p, insn,
1135 0,
1136 0, /* LD message ignores sampler unit */
1137 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1138 1, /* rlen */
1139 inst->mlen,
1140 header_present,
1141 BRW_SAMPLER_SIMD_MODE_SIMD4X2,
1142 0);
1143
1144 brw_pop_insn_state(p);
1145
1146 /* visitor knows more than we do about the surface limit required,
1147 * so has already done marking.
1148 */
1149
1150 }
1151 }
1152
1153 void
1154 fs_generator::generate_varying_pull_constant_load(fs_inst *inst,
1155 struct brw_reg dst,
1156 struct brw_reg index,
1157 struct brw_reg offset)
1158 {
1159 assert(devinfo->gen < 7); /* Should use the gen7 variant. */
1160 assert(inst->header_size != 0);
1161 assert(inst->mlen);
1162
1163 assert(index.file == BRW_IMMEDIATE_VALUE &&
1164 index.type == BRW_REGISTER_TYPE_UD);
1165 uint32_t surf_index = index.dw1.ud;
1166
1167 uint32_t simd_mode, rlen, msg_type;
1168 if (dispatch_width == 16) {
1169 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
1170 rlen = 8;
1171 } else {
1172 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
1173 rlen = 4;
1174 }
1175
1176 if (devinfo->gen >= 5)
1177 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
1178 else {
1179 /* We always use the SIMD16 message so that we only have to load U, and
1180 * not V or R.
1181 */
1182 msg_type = BRW_SAMPLER_MESSAGE_SIMD16_LD;
1183 assert(inst->mlen == 3);
1184 assert(inst->regs_written == 8);
1185 rlen = 8;
1186 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
1187 }
1188
1189 struct brw_reg offset_mrf = retype(brw_message_reg(inst->base_mrf + 1),
1190 BRW_REGISTER_TYPE_D);
1191 brw_MOV(p, offset_mrf, offset);
1192
1193 struct brw_reg header = brw_vec8_grf(0, 0);
1194 gen6_resolve_implied_move(p, &header, inst->base_mrf);
1195
1196 brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
1197 brw_inst_set_qtr_control(p->devinfo, send, BRW_COMPRESSION_NONE);
1198 brw_set_dest(p, send, retype(dst, BRW_REGISTER_TYPE_UW));
1199 brw_set_src0(p, send, header);
1200 if (devinfo->gen < 6)
1201 brw_inst_set_base_mrf(p->devinfo, send, inst->base_mrf);
1202
1203 /* Our surface is set up as floats, regardless of what actual data is
1204 * stored in it.
1205 */
1206 uint32_t return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
1207 brw_set_sampler_message(p, send,
1208 surf_index,
1209 0, /* sampler (unused) */
1210 msg_type,
1211 rlen,
1212 inst->mlen,
1213 inst->header_size != 0,
1214 simd_mode,
1215 return_format);
1216
1217 brw_mark_surface_used(prog_data, surf_index);
1218 }
1219
1220 void
1221 fs_generator::generate_varying_pull_constant_load_gen7(fs_inst *inst,
1222 struct brw_reg dst,
1223 struct brw_reg index,
1224 struct brw_reg offset)
1225 {
1226 assert(devinfo->gen >= 7);
1227 /* Varying-offset pull constant loads are treated as a normal expression on
1228 * gen7, so the fact that it's a send message is hidden at the IR level.
1229 */
1230 assert(inst->header_size == 0);
1231 assert(!inst->mlen);
1232 assert(index.type == BRW_REGISTER_TYPE_UD);
1233
1234 uint32_t simd_mode, rlen, mlen;
1235 if (dispatch_width == 16) {
1236 mlen = 2;
1237 rlen = 8;
1238 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
1239 } else {
1240 mlen = 1;
1241 rlen = 4;
1242 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
1243 }
1244
1245 if (index.file == BRW_IMMEDIATE_VALUE) {
1246
1247 uint32_t surf_index = index.dw1.ud;
1248
1249 brw_inst *send = brw_next_insn(p, BRW_OPCODE_SEND);
1250 brw_set_dest(p, send, retype(dst, BRW_REGISTER_TYPE_UW));
1251 brw_set_src0(p, send, offset);
1252 brw_set_sampler_message(p, send,
1253 surf_index,
1254 0, /* LD message ignores sampler unit */
1255 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1256 rlen,
1257 mlen,
1258 false, /* no header */
1259 simd_mode,
1260 0);
1261
1262 brw_mark_surface_used(prog_data, surf_index);
1263
1264 } else {
1265
1266 struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
1267
1268 brw_push_insn_state(p);
1269 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1270 brw_set_default_access_mode(p, BRW_ALIGN_1);
1271
1272 /* a0.0 = surf_index & 0xff */
1273 brw_inst *insn_and = brw_next_insn(p, BRW_OPCODE_AND);
1274 brw_inst_set_exec_size(p->devinfo, insn_and, BRW_EXECUTE_1);
1275 brw_set_dest(p, insn_and, addr);
1276 brw_set_src0(p, insn_and, vec1(retype(index, BRW_REGISTER_TYPE_UD)));
1277 brw_set_src1(p, insn_and, brw_imm_ud(0x0ff));
1278
1279 brw_pop_insn_state(p);
1280
1281 /* dst = send(offset, a0.0 | <descriptor>) */
1282 brw_inst *insn = brw_send_indirect_message(
1283 p, BRW_SFID_SAMPLER, retype(dst, BRW_REGISTER_TYPE_UW),
1284 offset, addr);
1285 brw_set_sampler_message(p, insn,
1286 0 /* surface */,
1287 0 /* sampler */,
1288 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1289 rlen /* rlen */,
1290 mlen /* mlen */,
1291 false /* header */,
1292 simd_mode,
1293 0);
1294
1295 /* visitor knows more than we do about the surface limit required,
1296 * so has already done marking.
1297 */
1298 }
1299 }
1300
1301 /**
1302 * Cause the current pixel/sample mask (from R1.7 bits 15:0) to be transferred
1303 * into the flags register (f0.0).
1304 *
1305 * Used only on Gen6 and above.
1306 */
1307 void
1308 fs_generator::generate_mov_dispatch_to_flags(fs_inst *inst)
1309 {
1310 struct brw_reg flags = brw_flag_reg(0, inst->flag_subreg);
1311 struct brw_reg dispatch_mask;
1312
1313 if (devinfo->gen >= 6)
1314 dispatch_mask = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
1315 else
1316 dispatch_mask = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
1317
1318 brw_push_insn_state(p);
1319 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1320 brw_MOV(p, flags, dispatch_mask);
1321 brw_pop_insn_state(p);
1322 }
1323
1324 void
1325 fs_generator::generate_pixel_interpolator_query(fs_inst *inst,
1326 struct brw_reg dst,
1327 struct brw_reg src,
1328 struct brw_reg msg_data,
1329 unsigned msg_type)
1330 {
1331 assert(msg_data.file == BRW_IMMEDIATE_VALUE &&
1332 msg_data.type == BRW_REGISTER_TYPE_UD);
1333
1334 brw_pixel_interpolator_query(p,
1335 retype(dst, BRW_REGISTER_TYPE_UW),
1336 src,
1337 inst->pi_noperspective,
1338 msg_type,
1339 msg_data.dw1.ud,
1340 inst->mlen,
1341 inst->regs_written);
1342 }
1343
1344
1345 /**
1346 * Sets the first word of a vgrf for gen7+ simd4x2 uniform pull constant
1347 * sampler LD messages.
1348 *
1349 * We don't want to bake it into the send message's code generation because
1350 * that means we don't get a chance to schedule the instructions.
1351 */
1352 void
1353 fs_generator::generate_set_simd4x2_offset(fs_inst *inst,
1354 struct brw_reg dst,
1355 struct brw_reg value)
1356 {
1357 assert(value.file == BRW_IMMEDIATE_VALUE);
1358
1359 brw_push_insn_state(p);
1360 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1361 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1362 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1363 brw_MOV(p, retype(brw_vec1_reg(dst.file, dst.nr, 0), value.type), value);
1364 brw_pop_insn_state(p);
1365 }
1366
1367 /* Sets vstride=16, width=8, hstride=2 or vstride=0, width=1, hstride=0
1368 * (when mask is passed as a uniform) of register mask before moving it
1369 * to register dst.
1370 */
1371 void
1372 fs_generator::generate_set_omask(fs_inst *inst,
1373 struct brw_reg dst,
1374 struct brw_reg mask)
1375 {
1376 bool stride_8_8_1 =
1377 (mask.vstride == BRW_VERTICAL_STRIDE_8 &&
1378 mask.width == BRW_WIDTH_8 &&
1379 mask.hstride == BRW_HORIZONTAL_STRIDE_1);
1380
1381 bool stride_0_1_0 = has_scalar_region(mask);
1382
1383 assert(stride_8_8_1 || stride_0_1_0);
1384 assert(dst.type == BRW_REGISTER_TYPE_UW);
1385
1386 brw_push_insn_state(p);
1387 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1388 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1389
1390 if (stride_8_8_1) {
1391 brw_MOV(p, dst, retype(stride(mask, 16, 8, 2), dst.type));
1392 } else if (stride_0_1_0) {
1393 brw_MOV(p, dst, retype(mask, dst.type));
1394 }
1395 brw_pop_insn_state(p);
1396 }
1397
1398 /* Sets vstride=1, width=4, hstride=0 of register src1 during
1399 * the ADD instruction.
1400 */
1401 void
1402 fs_generator::generate_set_sample_id(fs_inst *inst,
1403 struct brw_reg dst,
1404 struct brw_reg src0,
1405 struct brw_reg src1)
1406 {
1407 assert(dst.type == BRW_REGISTER_TYPE_D ||
1408 dst.type == BRW_REGISTER_TYPE_UD);
1409 assert(src0.type == BRW_REGISTER_TYPE_D ||
1410 src0.type == BRW_REGISTER_TYPE_UD);
1411
1412 brw_push_insn_state(p);
1413 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1414 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1415 brw_set_default_mask_control(p, BRW_MASK_DISABLE);
1416 struct brw_reg reg = retype(stride(src1, 1, 4, 0), BRW_REGISTER_TYPE_UW);
1417 if (dispatch_width == 8) {
1418 brw_ADD(p, dst, src0, reg);
1419 } else if (dispatch_width == 16) {
1420 brw_ADD(p, firsthalf(dst), firsthalf(src0), reg);
1421 brw_ADD(p, sechalf(dst), sechalf(src0), suboffset(reg, 2));
1422 }
1423 brw_pop_insn_state(p);
1424 }
1425
1426 void
1427 fs_generator::generate_pack_half_2x16_split(fs_inst *inst,
1428 struct brw_reg dst,
1429 struct brw_reg x,
1430 struct brw_reg y)
1431 {
1432 assert(devinfo->gen >= 7);
1433 assert(dst.type == BRW_REGISTER_TYPE_UD);
1434 assert(x.type == BRW_REGISTER_TYPE_F);
1435 assert(y.type == BRW_REGISTER_TYPE_F);
1436
1437 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
1438 *
1439 * Because this instruction does not have a 16-bit floating-point type,
1440 * the destination data type must be Word (W).
1441 *
1442 * The destination must be DWord-aligned and specify a horizontal stride
1443 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
1444 * each destination channel and the upper word is not modified.
1445 */
1446 struct brw_reg dst_w = spread(retype(dst, BRW_REGISTER_TYPE_W), 2);
1447
1448 /* Give each 32-bit channel of dst the form below, where "." means
1449 * unchanged.
1450 * 0x....hhhh
1451 */
1452 brw_F32TO16(p, dst_w, y);
1453
1454 /* Now the form:
1455 * 0xhhhh0000
1456 */
1457 brw_SHL(p, dst, dst, brw_imm_ud(16u));
1458
1459 /* And, finally the form of packHalf2x16's output:
1460 * 0xhhhhllll
1461 */
1462 brw_F32TO16(p, dst_w, x);
1463 }
1464
1465 void
1466 fs_generator::generate_unpack_half_2x16_split(fs_inst *inst,
1467 struct brw_reg dst,
1468 struct brw_reg src)
1469 {
1470 assert(devinfo->gen >= 7);
1471 assert(dst.type == BRW_REGISTER_TYPE_F);
1472 assert(src.type == BRW_REGISTER_TYPE_UD);
1473
1474 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
1475 *
1476 * Because this instruction does not have a 16-bit floating-point type,
1477 * the source data type must be Word (W). The destination type must be
1478 * F (Float).
1479 */
1480 struct brw_reg src_w = spread(retype(src, BRW_REGISTER_TYPE_W), 2);
1481
1482 /* Each channel of src has the form of unpackHalf2x16's input: 0xhhhhllll.
1483 * For the Y case, we wish to access only the upper word; therefore
1484 * a 16-bit subregister offset is needed.
1485 */
1486 assert(inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X ||
1487 inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y);
1488 if (inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y)
1489 src_w.subnr += 2;
1490
1491 brw_F16TO32(p, dst, src_w);
1492 }
1493
1494 void
1495 fs_generator::generate_shader_time_add(fs_inst *inst,
1496 struct brw_reg payload,
1497 struct brw_reg offset,
1498 struct brw_reg value)
1499 {
1500 assert(devinfo->gen >= 7);
1501 brw_push_insn_state(p);
1502 brw_set_default_mask_control(p, true);
1503
1504 assert(payload.file == BRW_GENERAL_REGISTER_FILE);
1505 struct brw_reg payload_offset = retype(brw_vec1_grf(payload.nr, 0),
1506 offset.type);
1507 struct brw_reg payload_value = retype(brw_vec1_grf(payload.nr + 1, 0),
1508 value.type);
1509
1510 assert(offset.file == BRW_IMMEDIATE_VALUE);
1511 if (value.file == BRW_GENERAL_REGISTER_FILE) {
1512 value.width = BRW_WIDTH_1;
1513 value.hstride = BRW_HORIZONTAL_STRIDE_0;
1514 value.vstride = BRW_VERTICAL_STRIDE_0;
1515 } else {
1516 assert(value.file == BRW_IMMEDIATE_VALUE);
1517 }
1518
1519 /* Trying to deal with setup of the params from the IR is crazy in the FS8
1520 * case, and we don't really care about squeezing every bit of performance
1521 * out of this path, so we just emit the MOVs from here.
1522 */
1523 brw_MOV(p, payload_offset, offset);
1524 brw_MOV(p, payload_value, value);
1525 brw_shader_time_add(p, payload,
1526 prog_data->binding_table.shader_time_start);
1527 brw_pop_insn_state(p);
1528
1529 brw_mark_surface_used(prog_data,
1530 prog_data->binding_table.shader_time_start);
1531 }
1532
1533 void
1534 fs_generator::enable_debug(const char *shader_name)
1535 {
1536 debug_flag = true;
1537 this->shader_name = shader_name;
1538 }
1539
1540 int
1541 fs_generator::generate_code(const cfg_t *cfg, int dispatch_width)
1542 {
1543 /* align to 64 byte boundary. */
1544 while (p->next_insn_offset % 64)
1545 brw_NOP(p);
1546
1547 this->dispatch_width = dispatch_width;
1548 if (dispatch_width == 16)
1549 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1550
1551 int start_offset = p->next_insn_offset;
1552 int spill_count = 0, fill_count = 0;
1553 int loop_count = 0;
1554
1555 struct annotation_info annotation;
1556 memset(&annotation, 0, sizeof(annotation));
1557
1558 foreach_block_and_inst (block, fs_inst, inst, cfg) {
1559 struct brw_reg src[3], dst;
1560 unsigned int last_insn_offset = p->next_insn_offset;
1561 bool multiple_instructions_emitted = false;
1562
1563 if (unlikely(debug_flag))
1564 annotate(p->devinfo, &annotation, cfg, inst, p->next_insn_offset);
1565
1566 for (unsigned int i = 0; i < inst->sources; i++) {
1567 src[i] = brw_reg_from_fs_reg(inst, &inst->src[i]);
1568
1569 /* The accumulator result appears to get used for the
1570 * conditional modifier generation. When negating a UD
1571 * value, there is a 33rd bit generated for the sign in the
1572 * accumulator value, so now you can't check, for example,
1573 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
1574 */
1575 assert(!inst->conditional_mod ||
1576 inst->src[i].type != BRW_REGISTER_TYPE_UD ||
1577 !inst->src[i].negate);
1578 }
1579 dst = brw_reg_from_fs_reg(inst, &inst->dst);
1580
1581 brw_set_default_predicate_control(p, inst->predicate);
1582 brw_set_default_predicate_inverse(p, inst->predicate_inverse);
1583 brw_set_default_flag_reg(p, 0, inst->flag_subreg);
1584 brw_set_default_saturate(p, inst->saturate);
1585 brw_set_default_mask_control(p, inst->force_writemask_all);
1586 brw_set_default_acc_write_control(p, inst->writes_accumulator);
1587 brw_set_default_exec_size(p, cvt(inst->exec_size) - 1);
1588
1589 switch (inst->exec_size) {
1590 case 1:
1591 case 2:
1592 case 4:
1593 assert(inst->force_writemask_all);
1594 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1595 break;
1596 case 8:
1597 if (inst->force_sechalf) {
1598 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1599 } else {
1600 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1601 }
1602 break;
1603 case 16:
1604 case 32:
1605 /* If the instruction writes to more than one register, it needs to
1606 * be a "compressed" instruction on Gen <= 5.
1607 */
1608 if (inst->dst.component_size(inst->exec_size) > REG_SIZE)
1609 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1610 else
1611 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1612 break;
1613 default:
1614 unreachable("Invalid instruction width");
1615 }
1616
1617 switch (inst->opcode) {
1618 case BRW_OPCODE_MOV:
1619 brw_MOV(p, dst, src[0]);
1620 break;
1621 case BRW_OPCODE_ADD:
1622 brw_ADD(p, dst, src[0], src[1]);
1623 break;
1624 case BRW_OPCODE_MUL:
1625 brw_MUL(p, dst, src[0], src[1]);
1626 break;
1627 case BRW_OPCODE_AVG:
1628 brw_AVG(p, dst, src[0], src[1]);
1629 break;
1630 case BRW_OPCODE_MACH:
1631 brw_MACH(p, dst, src[0], src[1]);
1632 break;
1633
1634 case BRW_OPCODE_LINE:
1635 brw_LINE(p, dst, src[0], src[1]);
1636 break;
1637
1638 case BRW_OPCODE_MAD:
1639 assert(devinfo->gen >= 6);
1640 brw_set_default_access_mode(p, BRW_ALIGN_16);
1641 if (dispatch_width == 16 && !devinfo->supports_simd16_3src) {
1642 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1643 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1644 brw_inst *f = brw_MAD(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1645 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1646 brw_inst *s = brw_MAD(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1647 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1648
1649 if (inst->conditional_mod) {
1650 brw_inst_set_cond_modifier(p->devinfo, f, inst->conditional_mod);
1651 brw_inst_set_cond_modifier(p->devinfo, s, inst->conditional_mod);
1652 multiple_instructions_emitted = true;
1653 }
1654 } else {
1655 brw_MAD(p, dst, src[0], src[1], src[2]);
1656 }
1657 brw_set_default_access_mode(p, BRW_ALIGN_1);
1658 break;
1659
1660 case BRW_OPCODE_LRP:
1661 assert(devinfo->gen >= 6);
1662 brw_set_default_access_mode(p, BRW_ALIGN_16);
1663 if (dispatch_width == 16 && !devinfo->supports_simd16_3src) {
1664 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1665 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1666 brw_inst *f = brw_LRP(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1667 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1668 brw_inst *s = brw_LRP(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1669 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1670
1671 if (inst->conditional_mod) {
1672 brw_inst_set_cond_modifier(p->devinfo, f, inst->conditional_mod);
1673 brw_inst_set_cond_modifier(p->devinfo, s, inst->conditional_mod);
1674 multiple_instructions_emitted = true;
1675 }
1676 } else {
1677 brw_LRP(p, dst, src[0], src[1], src[2]);
1678 }
1679 brw_set_default_access_mode(p, BRW_ALIGN_1);
1680 break;
1681
1682 case BRW_OPCODE_FRC:
1683 brw_FRC(p, dst, src[0]);
1684 break;
1685 case BRW_OPCODE_RNDD:
1686 brw_RNDD(p, dst, src[0]);
1687 break;
1688 case BRW_OPCODE_RNDE:
1689 brw_RNDE(p, dst, src[0]);
1690 break;
1691 case BRW_OPCODE_RNDZ:
1692 brw_RNDZ(p, dst, src[0]);
1693 break;
1694
1695 case BRW_OPCODE_AND:
1696 brw_AND(p, dst, src[0], src[1]);
1697 break;
1698 case BRW_OPCODE_OR:
1699 brw_OR(p, dst, src[0], src[1]);
1700 break;
1701 case BRW_OPCODE_XOR:
1702 brw_XOR(p, dst, src[0], src[1]);
1703 break;
1704 case BRW_OPCODE_NOT:
1705 brw_NOT(p, dst, src[0]);
1706 break;
1707 case BRW_OPCODE_ASR:
1708 brw_ASR(p, dst, src[0], src[1]);
1709 break;
1710 case BRW_OPCODE_SHR:
1711 brw_SHR(p, dst, src[0], src[1]);
1712 break;
1713 case BRW_OPCODE_SHL:
1714 brw_SHL(p, dst, src[0], src[1]);
1715 break;
1716 case BRW_OPCODE_F32TO16:
1717 assert(devinfo->gen >= 7);
1718 brw_F32TO16(p, dst, src[0]);
1719 break;
1720 case BRW_OPCODE_F16TO32:
1721 assert(devinfo->gen >= 7);
1722 brw_F16TO32(p, dst, src[0]);
1723 break;
1724 case BRW_OPCODE_CMP:
1725 /* The Ivybridge/BayTrail WaCMPInstFlagDepClearedEarly workaround says
1726 * that when the destination is a GRF that the dependency-clear bit on
1727 * the flag register is cleared early.
1728 *
1729 * Suggested workarounds are to disable coissuing CMP instructions
1730 * or to split CMP(16) instructions into two CMP(8) instructions.
1731 *
1732 * We choose to split into CMP(8) instructions since disabling
1733 * coissuing would affect CMP instructions not otherwise affected by
1734 * the errata.
1735 */
1736 if (dispatch_width == 16 && devinfo->gen == 7 && !devinfo->is_haswell) {
1737 if (dst.file == BRW_GENERAL_REGISTER_FILE) {
1738 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1739 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1740 brw_CMP(p, firsthalf(dst), inst->conditional_mod,
1741 firsthalf(src[0]), firsthalf(src[1]));
1742 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1743 brw_CMP(p, sechalf(dst), inst->conditional_mod,
1744 sechalf(src[0]), sechalf(src[1]));
1745 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1746
1747 multiple_instructions_emitted = true;
1748 } else if (dst.file == BRW_ARCHITECTURE_REGISTER_FILE) {
1749 /* For unknown reasons, the aforementioned workaround is not
1750 * sufficient. Overriding the type when the destination is the
1751 * null register is necessary but not sufficient by itself.
1752 */
1753 assert(dst.nr == BRW_ARF_NULL);
1754 dst.type = BRW_REGISTER_TYPE_D;
1755 brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
1756 } else {
1757 unreachable("not reached");
1758 }
1759 } else {
1760 brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
1761 }
1762 break;
1763 case BRW_OPCODE_SEL:
1764 brw_SEL(p, dst, src[0], src[1]);
1765 break;
1766 case BRW_OPCODE_BFREV:
1767 assert(devinfo->gen >= 7);
1768 /* BFREV only supports UD type for src and dst. */
1769 brw_BFREV(p, retype(dst, BRW_REGISTER_TYPE_UD),
1770 retype(src[0], BRW_REGISTER_TYPE_UD));
1771 break;
1772 case BRW_OPCODE_FBH:
1773 assert(devinfo->gen >= 7);
1774 /* FBH only supports UD type for dst. */
1775 brw_FBH(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1776 break;
1777 case BRW_OPCODE_FBL:
1778 assert(devinfo->gen >= 7);
1779 /* FBL only supports UD type for dst. */
1780 brw_FBL(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1781 break;
1782 case BRW_OPCODE_CBIT:
1783 assert(devinfo->gen >= 7);
1784 /* CBIT only supports UD type for dst. */
1785 brw_CBIT(p, retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1786 break;
1787 case BRW_OPCODE_ADDC:
1788 assert(devinfo->gen >= 7);
1789 brw_ADDC(p, dst, src[0], src[1]);
1790 break;
1791 case BRW_OPCODE_SUBB:
1792 assert(devinfo->gen >= 7);
1793 brw_SUBB(p, dst, src[0], src[1]);
1794 break;
1795 case BRW_OPCODE_MAC:
1796 brw_MAC(p, dst, src[0], src[1]);
1797 break;
1798
1799 case BRW_OPCODE_BFE:
1800 assert(devinfo->gen >= 7);
1801 brw_set_default_access_mode(p, BRW_ALIGN_16);
1802 if (dispatch_width == 16 && !devinfo->supports_simd16_3src) {
1803 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1804 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1805 brw_BFE(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1806 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1807 brw_BFE(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1808 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1809 } else {
1810 brw_BFE(p, dst, src[0], src[1], src[2]);
1811 }
1812 brw_set_default_access_mode(p, BRW_ALIGN_1);
1813 break;
1814
1815 case BRW_OPCODE_BFI1:
1816 assert(devinfo->gen >= 7);
1817 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
1818 * should
1819 *
1820 * "Force BFI instructions to be executed always in SIMD8."
1821 */
1822 if (dispatch_width == 16 && devinfo->is_haswell) {
1823 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1824 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1825 brw_BFI1(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]));
1826 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1827 brw_BFI1(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]));
1828 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1829 } else {
1830 brw_BFI1(p, dst, src[0], src[1]);
1831 }
1832 break;
1833 case BRW_OPCODE_BFI2:
1834 assert(devinfo->gen >= 7);
1835 brw_set_default_access_mode(p, BRW_ALIGN_16);
1836 /* The Haswell WaForceSIMD8ForBFIInstruction workaround says that we
1837 * should
1838 *
1839 * "Force BFI instructions to be executed always in SIMD8."
1840 *
1841 * Otherwise we would be able to emit compressed instructions like we
1842 * do for the other three-source instructions.
1843 */
1844 if (dispatch_width == 16 &&
1845 (devinfo->is_haswell || !devinfo->supports_simd16_3src)) {
1846 brw_set_default_exec_size(p, BRW_EXECUTE_8);
1847 brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
1848 brw_BFI2(p, firsthalf(dst), firsthalf(src[0]), firsthalf(src[1]), firsthalf(src[2]));
1849 brw_set_default_compression_control(p, BRW_COMPRESSION_2NDHALF);
1850 brw_BFI2(p, sechalf(dst), sechalf(src[0]), sechalf(src[1]), sechalf(src[2]));
1851 brw_set_default_compression_control(p, BRW_COMPRESSION_COMPRESSED);
1852 } else {
1853 brw_BFI2(p, dst, src[0], src[1], src[2]);
1854 }
1855 brw_set_default_access_mode(p, BRW_ALIGN_1);
1856 break;
1857
1858 case BRW_OPCODE_IF:
1859 if (inst->src[0].file != BAD_FILE) {
1860 /* The instruction has an embedded compare (only allowed on gen6) */
1861 assert(devinfo->gen == 6);
1862 gen6_IF(p, inst->conditional_mod, src[0], src[1]);
1863 } else {
1864 brw_IF(p, dispatch_width == 16 ? BRW_EXECUTE_16 : BRW_EXECUTE_8);
1865 }
1866 break;
1867
1868 case BRW_OPCODE_ELSE:
1869 brw_ELSE(p);
1870 break;
1871 case BRW_OPCODE_ENDIF:
1872 brw_ENDIF(p);
1873 break;
1874
1875 case BRW_OPCODE_DO:
1876 brw_DO(p, dispatch_width == 16 ? BRW_EXECUTE_16 : BRW_EXECUTE_8);
1877 break;
1878
1879 case BRW_OPCODE_BREAK:
1880 brw_BREAK(p);
1881 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
1882 break;
1883 case BRW_OPCODE_CONTINUE:
1884 brw_CONT(p);
1885 brw_set_default_predicate_control(p, BRW_PREDICATE_NONE);
1886 break;
1887
1888 case BRW_OPCODE_WHILE:
1889 brw_WHILE(p);
1890 loop_count++;
1891 break;
1892
1893 case SHADER_OPCODE_RCP:
1894 case SHADER_OPCODE_RSQ:
1895 case SHADER_OPCODE_SQRT:
1896 case SHADER_OPCODE_EXP2:
1897 case SHADER_OPCODE_LOG2:
1898 case SHADER_OPCODE_SIN:
1899 case SHADER_OPCODE_COS:
1900 assert(devinfo->gen < 6 || inst->mlen == 0);
1901 assert(inst->conditional_mod == BRW_CONDITIONAL_NONE);
1902 if (devinfo->gen >= 7) {
1903 gen6_math(p, dst, brw_math_function(inst->opcode), src[0],
1904 brw_null_reg());
1905 } else if (devinfo->gen == 6) {
1906 generate_math_gen6(inst, dst, src[0], brw_null_reg());
1907 } else if (devinfo->gen == 5 || devinfo->is_g4x) {
1908 generate_math_g45(inst, dst, src[0]);
1909 } else {
1910 generate_math_gen4(inst, dst, src[0]);
1911 }
1912 break;
1913 case SHADER_OPCODE_INT_QUOTIENT:
1914 case SHADER_OPCODE_INT_REMAINDER:
1915 case SHADER_OPCODE_POW:
1916 assert(devinfo->gen < 6 || inst->mlen == 0);
1917 assert(inst->conditional_mod == BRW_CONDITIONAL_NONE);
1918 if (devinfo->gen >= 7 && inst->opcode == SHADER_OPCODE_POW) {
1919 gen6_math(p, dst, brw_math_function(inst->opcode), src[0], src[1]);
1920 } else if (devinfo->gen >= 6) {
1921 generate_math_gen6(inst, dst, src[0], src[1]);
1922 } else {
1923 generate_math_gen4(inst, dst, src[0]);
1924 }
1925 break;
1926 case FS_OPCODE_CINTERP:
1927 brw_MOV(p, dst, src[0]);
1928 break;
1929 case FS_OPCODE_LINTERP:
1930 generate_linterp(inst, dst, src);
1931 break;
1932 case FS_OPCODE_PIXEL_X:
1933 assert(src[0].type == BRW_REGISTER_TYPE_UW);
1934 src[0].subnr = 0 * type_sz(src[0].type);
1935 brw_MOV(p, dst, stride(src[0], 8, 4, 1));
1936 break;
1937 case FS_OPCODE_PIXEL_Y:
1938 assert(src[0].type == BRW_REGISTER_TYPE_UW);
1939 src[0].subnr = 4 * type_sz(src[0].type);
1940 brw_MOV(p, dst, stride(src[0], 8, 4, 1));
1941 break;
1942 case SHADER_OPCODE_TEX:
1943 case FS_OPCODE_TXB:
1944 case SHADER_OPCODE_TXD:
1945 case SHADER_OPCODE_TXF:
1946 case SHADER_OPCODE_TXF_CMS:
1947 case SHADER_OPCODE_TXF_UMS:
1948 case SHADER_OPCODE_TXF_MCS:
1949 case SHADER_OPCODE_TXL:
1950 case SHADER_OPCODE_TXS:
1951 case SHADER_OPCODE_LOD:
1952 case SHADER_OPCODE_TG4:
1953 case SHADER_OPCODE_TG4_OFFSET:
1954 generate_tex(inst, dst, src[0], src[1]);
1955 break;
1956 case FS_OPCODE_DDX_COARSE:
1957 case FS_OPCODE_DDX_FINE:
1958 generate_ddx(inst->opcode, dst, src[0]);
1959 break;
1960 case FS_OPCODE_DDY_COARSE:
1961 case FS_OPCODE_DDY_FINE:
1962 assert(src[1].file == BRW_IMMEDIATE_VALUE);
1963 generate_ddy(inst->opcode, dst, src[0], src[1].dw1.ud);
1964 break;
1965
1966 case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
1967 generate_scratch_write(inst, src[0]);
1968 spill_count++;
1969 break;
1970
1971 case SHADER_OPCODE_GEN4_SCRATCH_READ:
1972 generate_scratch_read(inst, dst);
1973 fill_count++;
1974 break;
1975
1976 case SHADER_OPCODE_GEN7_SCRATCH_READ:
1977 generate_scratch_read_gen7(inst, dst);
1978 fill_count++;
1979 break;
1980
1981 case SHADER_OPCODE_URB_WRITE_SIMD8:
1982 generate_urb_write(inst, src[0]);
1983 break;
1984
1985 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
1986 generate_uniform_pull_constant_load(inst, dst, src[0], src[1]);
1987 break;
1988
1989 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7:
1990 generate_uniform_pull_constant_load_gen7(inst, dst, src[0], src[1]);
1991 break;
1992
1993 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD:
1994 generate_varying_pull_constant_load(inst, dst, src[0], src[1]);
1995 break;
1996
1997 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
1998 generate_varying_pull_constant_load_gen7(inst, dst, src[0], src[1]);
1999 break;
2000
2001 case FS_OPCODE_REP_FB_WRITE:
2002 case FS_OPCODE_FB_WRITE:
2003 generate_fb_write(inst, src[0]);
2004 break;
2005
2006 case FS_OPCODE_BLORP_FB_WRITE:
2007 generate_blorp_fb_write(inst);
2008 break;
2009
2010 case FS_OPCODE_MOV_DISPATCH_TO_FLAGS:
2011 generate_mov_dispatch_to_flags(inst);
2012 break;
2013
2014 case FS_OPCODE_DISCARD_JUMP:
2015 generate_discard_jump(inst);
2016 break;
2017
2018 case SHADER_OPCODE_SHADER_TIME_ADD:
2019 generate_shader_time_add(inst, src[0], src[1], src[2]);
2020 break;
2021
2022 case SHADER_OPCODE_UNTYPED_ATOMIC:
2023 assert(src[1].file == BRW_IMMEDIATE_VALUE &&
2024 src[2].file == BRW_IMMEDIATE_VALUE);
2025 brw_untyped_atomic(p, dst, src[0], src[1], src[2].dw1.ud,
2026 inst->mlen, !inst->dst.is_null());
2027 brw_mark_surface_used(prog_data, src[1].dw1.ud);
2028 break;
2029
2030 case SHADER_OPCODE_UNTYPED_SURFACE_READ:
2031 assert(src[1].file == BRW_IMMEDIATE_VALUE &&
2032 src[2].file == BRW_IMMEDIATE_VALUE);
2033 brw_untyped_surface_read(p, dst, src[0], src[1],
2034 inst->mlen, src[2].dw1.ud);
2035 brw_mark_surface_used(prog_data, src[1].dw1.ud);
2036 break;
2037
2038 case SHADER_OPCODE_UNTYPED_SURFACE_WRITE:
2039 assert(src[2].file == BRW_IMMEDIATE_VALUE);
2040 brw_untyped_surface_write(p, src[0], src[1],
2041 inst->mlen, src[2].dw1.ud);
2042 break;
2043
2044 case SHADER_OPCODE_TYPED_ATOMIC:
2045 assert(src[2].file == BRW_IMMEDIATE_VALUE);
2046 brw_typed_atomic(p, dst, src[0], src[1],
2047 src[2].dw1.ud, inst->mlen, !inst->dst.is_null());
2048 break;
2049
2050 case SHADER_OPCODE_TYPED_SURFACE_READ:
2051 assert(src[2].file == BRW_IMMEDIATE_VALUE);
2052 brw_typed_surface_read(p, dst, src[0], src[1],
2053 inst->mlen, src[2].dw1.ud);
2054 break;
2055
2056 case SHADER_OPCODE_TYPED_SURFACE_WRITE:
2057 assert(src[2].file == BRW_IMMEDIATE_VALUE);
2058 brw_typed_surface_write(p, src[0], src[1], inst->mlen, src[2].dw1.ud);
2059 break;
2060
2061 case SHADER_OPCODE_MEMORY_FENCE:
2062 brw_memory_fence(p, dst);
2063 break;
2064
2065 case FS_OPCODE_SET_SIMD4X2_OFFSET:
2066 generate_set_simd4x2_offset(inst, dst, src[0]);
2067 break;
2068
2069 case SHADER_OPCODE_FIND_LIVE_CHANNEL:
2070 brw_find_live_channel(p, dst);
2071 break;
2072
2073 case SHADER_OPCODE_BROADCAST:
2074 brw_broadcast(p, dst, src[0], src[1]);
2075 break;
2076
2077 case FS_OPCODE_SET_OMASK:
2078 generate_set_omask(inst, dst, src[0]);
2079 break;
2080
2081 case FS_OPCODE_SET_SAMPLE_ID:
2082 generate_set_sample_id(inst, dst, src[0], src[1]);
2083 break;
2084
2085 case FS_OPCODE_PACK_HALF_2x16_SPLIT:
2086 generate_pack_half_2x16_split(inst, dst, src[0], src[1]);
2087 break;
2088
2089 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X:
2090 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y:
2091 generate_unpack_half_2x16_split(inst, dst, src[0]);
2092 break;
2093
2094 case FS_OPCODE_PLACEHOLDER_HALT:
2095 /* This is the place where the final HALT needs to be inserted if
2096 * we've emitted any discards. If not, this will emit no code.
2097 */
2098 if (!patch_discard_jumps_to_fb_writes()) {
2099 if (unlikely(debug_flag)) {
2100 annotation.ann_count--;
2101 }
2102 }
2103 break;
2104
2105 case FS_OPCODE_INTERPOLATE_AT_CENTROID:
2106 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2107 GEN7_PIXEL_INTERPOLATOR_LOC_CENTROID);
2108 break;
2109
2110 case FS_OPCODE_INTERPOLATE_AT_SAMPLE:
2111 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2112 GEN7_PIXEL_INTERPOLATOR_LOC_SAMPLE);
2113 break;
2114
2115 case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET:
2116 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2117 GEN7_PIXEL_INTERPOLATOR_LOC_SHARED_OFFSET);
2118 break;
2119
2120 case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET:
2121 generate_pixel_interpolator_query(inst, dst, src[0], src[1],
2122 GEN7_PIXEL_INTERPOLATOR_LOC_PER_SLOT_OFFSET);
2123 break;
2124
2125 case CS_OPCODE_CS_TERMINATE:
2126 generate_cs_terminate(inst, src[0]);
2127 break;
2128
2129 case SHADER_OPCODE_BARRIER:
2130 generate_barrier(inst, src[0]);
2131 break;
2132
2133 default:
2134 unreachable("Unsupported opcode");
2135
2136 case SHADER_OPCODE_LOAD_PAYLOAD:
2137 unreachable("Should be lowered by lower_load_payload()");
2138 }
2139
2140 if (multiple_instructions_emitted)
2141 continue;
2142
2143 if (inst->no_dd_clear || inst->no_dd_check || inst->conditional_mod) {
2144 assert(p->next_insn_offset == last_insn_offset + 16 ||
2145 !"conditional_mod, no_dd_check, or no_dd_clear set for IR "
2146 "emitting more than 1 instruction");
2147
2148 brw_inst *last = &p->store[last_insn_offset / 16];
2149
2150 if (inst->conditional_mod)
2151 brw_inst_set_cond_modifier(p->devinfo, last, inst->conditional_mod);
2152 brw_inst_set_no_dd_clear(p->devinfo, last, inst->no_dd_clear);
2153 brw_inst_set_no_dd_check(p->devinfo, last, inst->no_dd_check);
2154 }
2155 }
2156
2157 brw_set_uip_jip(p);
2158 annotation_finalize(&annotation, p->next_insn_offset);
2159
2160 int before_size = p->next_insn_offset - start_offset;
2161 brw_compact_instructions(p, start_offset, annotation.ann_count,
2162 annotation.ann);
2163 int after_size = p->next_insn_offset - start_offset;
2164
2165 if (unlikely(debug_flag)) {
2166 fprintf(stderr, "Native code for %s\n"
2167 "SIMD%d shader: %d instructions. %d loops. %d:%d spills:fills. Promoted %u constants. Compacted %d to %d"
2168 " bytes (%.0f%%)\n",
2169 shader_name, dispatch_width, before_size / 16, loop_count,
2170 spill_count, fill_count, promoted_constants, before_size, after_size,
2171 100.0f * (before_size - after_size) / before_size);
2172
2173 dump_assembly(p->store, annotation.ann_count, annotation.ann,
2174 p->devinfo, prog);
2175 ralloc_free(annotation.ann);
2176 }
2177
2178 compiler->shader_debug_log(log_data,
2179 "%s SIMD%d shader: %d inst, %d loops, "
2180 "%d:%d spills:fills, Promoted %u constants, "
2181 "compacted %d to %d bytes.\n",
2182 stage_abbrev, dispatch_width, before_size / 16,
2183 loop_count, spill_count, fill_count,
2184 promoted_constants, before_size, after_size);
2185
2186 return start_offset;
2187 }
2188
2189 const unsigned *
2190 fs_generator::get_assembly(unsigned int *assembly_size)
2191 {
2192 return brw_get_program(p, assembly_size);
2193 }