Avoid branch instructions while in single program flow mode.
[mesa.git] / src / mesa / drivers / dri / i965 / brw_eu_emit.c
1 /*
2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
4 develop this 3D driver.
5
6 Permission is hereby granted, free of charge, to any person obtaining
7 a copy of this software and associated documentation files (the
8 "Software"), to deal in the Software without restriction, including
9 without limitation the rights to use, copy, modify, merge, publish,
10 distribute, sublicense, and/or sell copies of the Software, and to
11 permit persons to whom the Software is furnished to do so, subject to
12 the following conditions:
13
14 The above copyright notice and this permission notice (including the
15 next paragraph) shall be included in all copies or substantial
16 portions of the Software.
17
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25
26 **********************************************************************/
27 /*
28 * Authors:
29 * Keith Whitwell <keith@tungstengraphics.com>
30 */
31
32
33 #include "brw_context.h"
34 #include "brw_defines.h"
35 #include "brw_eu.h"
36
37
38
39
40 /***********************************************************************
41 * Internal helper for constructing instructions
42 */
43
44 static void guess_execution_size( struct brw_instruction *insn,
45 struct brw_reg reg )
46 {
47 if (reg.width == BRW_WIDTH_8 &&
48 insn->header.compression_control == BRW_COMPRESSION_COMPRESSED)
49 insn->header.execution_size = BRW_EXECUTE_16;
50 else
51 insn->header.execution_size = reg.width; /* note - definitions are compatible */
52 }
53
54
55 static void brw_set_dest( struct brw_instruction *insn,
56 struct brw_reg dest )
57 {
58 insn->bits1.da1.dest_reg_file = dest.file;
59 insn->bits1.da1.dest_reg_type = dest.type;
60 insn->bits1.da1.dest_address_mode = dest.address_mode;
61
62 if (dest.address_mode == BRW_ADDRESS_DIRECT) {
63 insn->bits1.da1.dest_reg_nr = dest.nr;
64
65 if (insn->header.access_mode == BRW_ALIGN_1) {
66 insn->bits1.da1.dest_subreg_nr = dest.subnr;
67 insn->bits1.da1.dest_horiz_stride = BRW_HORIZONTAL_STRIDE_1;
68 }
69 else {
70 insn->bits1.da16.dest_subreg_nr = dest.subnr / 16;
71 insn->bits1.da16.dest_writemask = dest.dw1.bits.writemask;
72 }
73 }
74 else {
75 insn->bits1.ia1.dest_subreg_nr = dest.subnr;
76
77 /* These are different sizes in align1 vs align16:
78 */
79 if (insn->header.access_mode == BRW_ALIGN_1) {
80 insn->bits1.ia1.dest_indirect_offset = dest.dw1.bits.indirect_offset;
81 insn->bits1.ia1.dest_horiz_stride = BRW_HORIZONTAL_STRIDE_1;
82 }
83 else {
84 insn->bits1.ia16.dest_indirect_offset = dest.dw1.bits.indirect_offset;
85 }
86 }
87
88 /* NEW: Set the execution size based on dest.width and
89 * insn->compression_control:
90 */
91 guess_execution_size(insn, dest);
92 }
93
94 static void brw_set_src0( struct brw_instruction *insn,
95 struct brw_reg reg )
96 {
97 assert(reg.file != BRW_MESSAGE_REGISTER_FILE);
98
99 insn->bits1.da1.src0_reg_file = reg.file;
100 insn->bits1.da1.src0_reg_type = reg.type;
101 insn->bits2.da1.src0_abs = reg.abs;
102 insn->bits2.da1.src0_negate = reg.negate;
103 insn->bits2.da1.src0_address_mode = reg.address_mode;
104
105 if (reg.file == BRW_IMMEDIATE_VALUE) {
106 insn->bits3.ud = reg.dw1.ud;
107
108 /* Required to set some fields in src1 as well:
109 */
110 insn->bits1.da1.src1_reg_file = 0; /* arf */
111 insn->bits1.da1.src1_reg_type = reg.type;
112 }
113 else
114 {
115 if (reg.address_mode == BRW_ADDRESS_DIRECT) {
116 if (insn->header.access_mode == BRW_ALIGN_1) {
117 insn->bits2.da1.src0_subreg_nr = reg.subnr;
118 insn->bits2.da1.src0_reg_nr = reg.nr;
119 }
120 else {
121 insn->bits2.da16.src0_subreg_nr = reg.subnr / 16;
122 insn->bits2.da16.src0_reg_nr = reg.nr;
123 }
124 }
125 else {
126 insn->bits2.ia1.src0_subreg_nr = reg.subnr;
127
128 if (insn->header.access_mode == BRW_ALIGN_1) {
129 insn->bits2.ia1.src0_indirect_offset = reg.dw1.bits.indirect_offset;
130 }
131 else {
132 insn->bits2.ia16.src0_subreg_nr = reg.dw1.bits.indirect_offset;
133 }
134 }
135
136 if (insn->header.access_mode == BRW_ALIGN_1) {
137 if (reg.width == BRW_WIDTH_1 &&
138 insn->header.execution_size == BRW_EXECUTE_1) {
139 insn->bits2.da1.src0_horiz_stride = BRW_HORIZONTAL_STRIDE_0;
140 insn->bits2.da1.src0_width = BRW_WIDTH_1;
141 insn->bits2.da1.src0_vert_stride = BRW_VERTICAL_STRIDE_0;
142 }
143 else {
144 insn->bits2.da1.src0_horiz_stride = reg.hstride;
145 insn->bits2.da1.src0_width = reg.width;
146 insn->bits2.da1.src0_vert_stride = reg.vstride;
147 }
148 }
149 else {
150 insn->bits2.da16.src0_swz_x = BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_X);
151 insn->bits2.da16.src0_swz_y = BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_Y);
152 insn->bits2.da16.src0_swz_z = BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_Z);
153 insn->bits2.da16.src0_swz_w = BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_W);
154
155 /* This is an oddity of the fact we're using the same
156 * descriptions for registers in align_16 as align_1:
157 */
158 if (reg.vstride == BRW_VERTICAL_STRIDE_8)
159 insn->bits2.da16.src0_vert_stride = BRW_VERTICAL_STRIDE_4;
160 else
161 insn->bits2.da16.src0_vert_stride = reg.vstride;
162 }
163 }
164 }
165
166
167 static void brw_set_src1( struct brw_instruction *insn,
168 struct brw_reg reg )
169 {
170 assert(reg.file != BRW_MESSAGE_REGISTER_FILE);
171
172 insn->bits1.da1.src1_reg_file = reg.file;
173 insn->bits1.da1.src1_reg_type = reg.type;
174 insn->bits3.da1.src1_abs = reg.abs;
175 insn->bits3.da1.src1_negate = reg.negate;
176
177 /* Only src1 can be immediate in two-argument instructions.
178 */
179 assert(insn->bits1.da1.src0_reg_file != BRW_IMMEDIATE_VALUE);
180
181 if (reg.file == BRW_IMMEDIATE_VALUE) {
182 insn->bits3.ud = reg.dw1.ud;
183 }
184 else {
185 /* This is a hardware restriction, which may or may not be lifted
186 * in the future:
187 */
188 assert (reg.address_mode == BRW_ADDRESS_DIRECT);
189 assert (reg.file == BRW_GENERAL_REGISTER_FILE);
190
191 if (insn->header.access_mode == BRW_ALIGN_1) {
192 insn->bits3.da1.src1_subreg_nr = reg.subnr;
193 insn->bits3.da1.src1_reg_nr = reg.nr;
194 }
195 else {
196 insn->bits3.da16.src1_subreg_nr = reg.subnr / 16;
197 insn->bits3.da16.src1_reg_nr = reg.nr;
198 }
199
200 if (insn->header.access_mode == BRW_ALIGN_1) {
201 if (reg.width == BRW_WIDTH_1 &&
202 insn->header.execution_size == BRW_EXECUTE_1) {
203 insn->bits3.da1.src1_horiz_stride = BRW_HORIZONTAL_STRIDE_0;
204 insn->bits3.da1.src1_width = BRW_WIDTH_1;
205 insn->bits3.da1.src1_vert_stride = BRW_VERTICAL_STRIDE_0;
206 }
207 else {
208 insn->bits3.da1.src1_horiz_stride = reg.hstride;
209 insn->bits3.da1.src1_width = reg.width;
210 insn->bits3.da1.src1_vert_stride = reg.vstride;
211 }
212 }
213 else {
214 insn->bits3.da16.src1_swz_x = BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_X);
215 insn->bits3.da16.src1_swz_y = BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_Y);
216 insn->bits3.da16.src1_swz_z = BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_Z);
217 insn->bits3.da16.src1_swz_w = BRW_GET_SWZ(reg.dw1.bits.swizzle, BRW_CHANNEL_W);
218
219 /* This is an oddity of the fact we're using the same
220 * descriptions for registers in align_16 as align_1:
221 */
222 if (reg.vstride == BRW_VERTICAL_STRIDE_8)
223 insn->bits3.da16.src1_vert_stride = BRW_VERTICAL_STRIDE_4;
224 else
225 insn->bits3.da16.src1_vert_stride = reg.vstride;
226 }
227 }
228 }
229
230
231
232 static void brw_set_math_message( struct brw_instruction *insn,
233 GLuint msg_length,
234 GLuint response_length,
235 GLuint function,
236 GLuint integer_type,
237 GLboolean low_precision,
238 GLboolean saturate,
239 GLuint dataType )
240 {
241 brw_set_src1(insn, brw_imm_d(0));
242
243 insn->bits3.math.function = function;
244 insn->bits3.math.int_type = integer_type;
245 insn->bits3.math.precision = low_precision;
246 insn->bits3.math.saturate = saturate;
247 insn->bits3.math.data_type = dataType;
248 insn->bits3.math.response_length = response_length;
249 insn->bits3.math.msg_length = msg_length;
250 insn->bits3.math.msg_target = BRW_MESSAGE_TARGET_MATH;
251 insn->bits3.math.end_of_thread = 0;
252 }
253
254 static void brw_set_urb_message( struct brw_instruction *insn,
255 GLboolean allocate,
256 GLboolean used,
257 GLuint msg_length,
258 GLuint response_length,
259 GLboolean end_of_thread,
260 GLboolean complete,
261 GLuint offset,
262 GLuint swizzle_control )
263 {
264 brw_set_src1(insn, brw_imm_d(0));
265
266 insn->bits3.urb.opcode = 0; /* ? */
267 insn->bits3.urb.offset = offset;
268 insn->bits3.urb.swizzle_control = swizzle_control;
269 insn->bits3.urb.allocate = allocate;
270 insn->bits3.urb.used = used; /* ? */
271 insn->bits3.urb.complete = complete;
272 insn->bits3.urb.response_length = response_length;
273 insn->bits3.urb.msg_length = msg_length;
274 insn->bits3.urb.msg_target = BRW_MESSAGE_TARGET_URB;
275 insn->bits3.urb.end_of_thread = end_of_thread;
276 }
277
278 static void brw_set_dp_write_message( struct brw_instruction *insn,
279 GLuint binding_table_index,
280 GLuint msg_control,
281 GLuint msg_type,
282 GLuint msg_length,
283 GLuint pixel_scoreboard_clear,
284 GLuint response_length,
285 GLuint end_of_thread )
286 {
287 brw_set_src1(insn, brw_imm_d(0));
288
289 insn->bits3.dp_write.binding_table_index = binding_table_index;
290 insn->bits3.dp_write.msg_control = msg_control;
291 insn->bits3.dp_write.pixel_scoreboard_clear = pixel_scoreboard_clear;
292 insn->bits3.dp_write.msg_type = msg_type;
293 insn->bits3.dp_write.send_commit_msg = 0;
294 insn->bits3.dp_write.response_length = response_length;
295 insn->bits3.dp_write.msg_length = msg_length;
296 insn->bits3.dp_write.msg_target = BRW_MESSAGE_TARGET_DATAPORT_WRITE;
297 insn->bits3.urb.end_of_thread = end_of_thread;
298 }
299
300 static void brw_set_dp_read_message( struct brw_instruction *insn,
301 GLuint binding_table_index,
302 GLuint msg_control,
303 GLuint msg_type,
304 GLuint target_cache,
305 GLuint msg_length,
306 GLuint response_length,
307 GLuint end_of_thread )
308 {
309 brw_set_src1(insn, brw_imm_d(0));
310
311 insn->bits3.dp_read.binding_table_index = binding_table_index;
312 insn->bits3.dp_read.msg_control = msg_control;
313 insn->bits3.dp_read.msg_type = msg_type;
314 insn->bits3.dp_read.target_cache = target_cache;
315 insn->bits3.dp_read.response_length = response_length;
316 insn->bits3.dp_read.msg_length = msg_length;
317 insn->bits3.dp_read.msg_target = BRW_MESSAGE_TARGET_DATAPORT_READ;
318 insn->bits3.dp_read.end_of_thread = end_of_thread;
319 }
320
321 static void brw_set_sampler_message( struct brw_instruction *insn,
322 GLuint binding_table_index,
323 GLuint sampler,
324 GLuint msg_type,
325 GLuint response_length,
326 GLuint msg_length,
327 GLboolean eot)
328 {
329 brw_set_src1(insn, brw_imm_d(0));
330
331 insn->bits3.sampler.binding_table_index = binding_table_index;
332 insn->bits3.sampler.sampler = sampler;
333 insn->bits3.sampler.msg_type = msg_type;
334 insn->bits3.sampler.return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
335 insn->bits3.sampler.response_length = response_length;
336 insn->bits3.sampler.msg_length = msg_length;
337 insn->bits3.sampler.end_of_thread = eot;
338 insn->bits3.sampler.msg_target = BRW_MESSAGE_TARGET_SAMPLER;
339 }
340
341
342
343 static struct brw_instruction *next_insn( struct brw_compile *p,
344 GLuint opcode )
345 {
346 struct brw_instruction *insn;
347
348 assert(p->nr_insn + 1 < BRW_EU_MAX_INSN);
349
350 insn = &p->store[p->nr_insn++];
351 memcpy(insn, p->current, sizeof(*insn));
352
353 /* Reset this one-shot flag:
354 */
355
356 if (p->current->header.destreg__conditonalmod) {
357 p->current->header.destreg__conditonalmod = 0;
358 p->current->header.predicate_control = BRW_PREDICATE_NORMAL;
359 }
360
361 insn->header.opcode = opcode;
362 return insn;
363 }
364
365
366 static struct brw_instruction *brw_alu1( struct brw_compile *p,
367 GLuint opcode,
368 struct brw_reg dest,
369 struct brw_reg src )
370 {
371 struct brw_instruction *insn = next_insn(p, opcode);
372 brw_set_dest(insn, dest);
373 brw_set_src0(insn, src);
374 return insn;
375 }
376
377 static struct brw_instruction *brw_alu2(struct brw_compile *p,
378 GLuint opcode,
379 struct brw_reg dest,
380 struct brw_reg src0,
381 struct brw_reg src1 )
382 {
383 struct brw_instruction *insn = next_insn(p, opcode);
384 brw_set_dest(insn, dest);
385 brw_set_src0(insn, src0);
386 brw_set_src1(insn, src1);
387 return insn;
388 }
389
390
391 /***********************************************************************
392 * Convenience routines.
393 */
394 #define ALU1(OP) \
395 struct brw_instruction *brw_##OP(struct brw_compile *p, \
396 struct brw_reg dest, \
397 struct brw_reg src0) \
398 { \
399 return brw_alu1(p, BRW_OPCODE_##OP, dest, src0); \
400 }
401
402 #define ALU2(OP) \
403 struct brw_instruction *brw_##OP(struct brw_compile *p, \
404 struct brw_reg dest, \
405 struct brw_reg src0, \
406 struct brw_reg src1) \
407 { \
408 return brw_alu2(p, BRW_OPCODE_##OP, dest, src0, src1); \
409 }
410
411
412 ALU1(MOV)
413 ALU2(SEL)
414 ALU1(NOT)
415 ALU2(AND)
416 ALU2(OR)
417 ALU2(XOR)
418 ALU2(SHR)
419 ALU2(SHL)
420 ALU2(RSR)
421 ALU2(RSL)
422 ALU2(ASR)
423 ALU2(ADD)
424 ALU2(MUL)
425 ALU1(FRC)
426 ALU1(RNDD)
427 ALU2(MAC)
428 ALU2(MACH)
429 ALU1(LZD)
430 ALU2(DP4)
431 ALU2(DPH)
432 ALU2(DP3)
433 ALU2(DP2)
434 ALU2(LINE)
435
436
437
438
439 void brw_NOP(struct brw_compile *p)
440 {
441 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_NOP);
442 brw_set_dest(insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
443 brw_set_src0(insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
444 brw_set_src1(insn, brw_imm_ud(0x0));
445 }
446
447
448
449
450
451 /***********************************************************************
452 * Comparisons, if/else/endif
453 */
454
455 struct brw_instruction *brw_JMPI(struct brw_compile *p,
456 struct brw_reg dest,
457 struct brw_reg src0,
458 struct brw_reg src1)
459 {
460 struct brw_instruction *insn = brw_alu2(p, BRW_OPCODE_JMPI, dest, src0, src1);
461
462 p->current->header.predicate_control = BRW_PREDICATE_NONE;
463
464 return insn;
465 }
466
467 /* EU takes the value from the flag register and pushes it onto some
468 * sort of a stack (presumably merging with any flag value already on
469 * the stack). Within an if block, the flags at the top of the stack
470 * control execution on each channel of the unit, eg. on each of the
471 * 16 pixel values in our wm programs.
472 *
473 * When the matching 'else' instruction is reached (presumably by
474 * countdown of the instruction count patched in by our ELSE/ENDIF
475 * functions), the relevent flags are inverted.
476 *
477 * When the matching 'endif' instruction is reached, the flags are
478 * popped off. If the stack is now empty, normal execution resumes.
479 *
480 * No attempt is made to deal with stack overflow (14 elements?).
481 */
482 struct brw_instruction *brw_IF(struct brw_compile *p, GLuint execute_size)
483 {
484 struct brw_instruction *insn;
485
486 if (p->single_program_flow) {
487 assert(execute_size == BRW_EXECUTE_1);
488
489 insn = next_insn(p, BRW_OPCODE_ADD);
490 insn->header.predicate_inverse = 1;
491 } else {
492 insn = next_insn(p, BRW_OPCODE_IF);
493 }
494
495 /* Override the defaults for this instruction:
496 */
497 brw_set_dest(insn, brw_ip_reg());
498 brw_set_src0(insn, brw_ip_reg());
499 brw_set_src1(insn, brw_imm_d(0x0));
500
501 insn->header.execution_size = execute_size;
502 insn->header.compression_control = BRW_COMPRESSION_NONE;
503 insn->header.predicate_control = BRW_PREDICATE_NORMAL;
504 insn->header.mask_control = BRW_MASK_ENABLE;
505
506 p->current->header.predicate_control = BRW_PREDICATE_NONE;
507
508 return insn;
509 }
510
511
512 struct brw_instruction *brw_ELSE(struct brw_compile *p,
513 struct brw_instruction *if_insn)
514 {
515 struct brw_instruction *insn;
516
517 if (p->single_program_flow) {
518 insn = next_insn(p, BRW_OPCODE_ADD);
519 } else {
520 insn = next_insn(p, BRW_OPCODE_ELSE);
521 }
522
523 brw_set_dest(insn, brw_ip_reg());
524 brw_set_src0(insn, brw_ip_reg());
525 brw_set_src1(insn, brw_imm_d(0x0));
526
527 insn->header.compression_control = BRW_COMPRESSION_NONE;
528 insn->header.execution_size = if_insn->header.execution_size;
529 insn->header.mask_control = BRW_MASK_ENABLE;
530
531 /* Patch the if instruction to point at this instruction.
532 */
533 if (p->single_program_flow) {
534 assert(if_insn->header.opcode == BRW_OPCODE_ADD);
535
536 if_insn->bits3.ud = (insn - if_insn + 1) * 16;
537 } else {
538 assert(if_insn->header.opcode == BRW_OPCODE_IF);
539
540 if_insn->bits3.if_else.jump_count = insn - if_insn;
541 if_insn->bits3.if_else.pop_count = 1;
542 if_insn->bits3.if_else.pad0 = 0;
543 }
544
545 return insn;
546 }
547
548 void brw_ENDIF(struct brw_compile *p,
549 struct brw_instruction *patch_insn)
550 {
551 if (p->single_program_flow) {
552 /* In single program flow mode, there's no need to execute an ENDIF,
553 * since we don't need to do any stack operations, and if we're executing
554 * currently, we want to just continue executing.
555 */
556 struct brw_instruction *next = &p->store[p->nr_insn];
557
558 assert(patch_insn->header.opcode == BRW_OPCODE_ADD);
559
560 patch_insn->bits3.ud = (next - patch_insn) * 16;
561 } else {
562 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_ENDIF);
563
564 brw_set_dest(insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
565 brw_set_src0(insn, retype(brw_vec4_grf(0,0), BRW_REGISTER_TYPE_UD));
566 brw_set_src1(insn, brw_imm_d(0x0));
567
568 insn->header.compression_control = BRW_COMPRESSION_NONE;
569 insn->header.execution_size = patch_insn->header.execution_size;
570 insn->header.mask_control = BRW_MASK_ENABLE;
571
572 assert(patch_insn->bits3.if_else.jump_count == 0);
573
574 /* Patch the if or else instructions to point at this or the next
575 * instruction respectively.
576 */
577 if (patch_insn->header.opcode == BRW_OPCODE_IF) {
578 /* Automagically turn it into an IFF:
579 */
580 patch_insn->header.opcode = BRW_OPCODE_IFF;
581 patch_insn->bits3.if_else.jump_count = insn - patch_insn + 1;
582 patch_insn->bits3.if_else.pop_count = 0;
583 patch_insn->bits3.if_else.pad0 = 0;
584 } else if (patch_insn->header.opcode == BRW_OPCODE_ELSE) {
585 patch_insn->bits3.if_else.jump_count = insn - patch_insn + 1;
586 patch_insn->bits3.if_else.pop_count = 1;
587 patch_insn->bits3.if_else.pad0 = 0;
588 } else {
589 assert(0);
590 }
591
592 /* Also pop item off the stack in the endif instruction:
593 */
594 insn->bits3.if_else.jump_count = 0;
595 insn->bits3.if_else.pop_count = 1;
596 insn->bits3.if_else.pad0 = 0;
597 }
598 }
599
600 /* DO/WHILE loop:
601 */
602 struct brw_instruction *brw_DO(struct brw_compile *p, GLuint execute_size)
603 {
604 if (p->single_program_flow) {
605 return &p->store[p->nr_insn];
606 } else {
607 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_DO);
608
609 /* Override the defaults for this instruction:
610 */
611 brw_set_dest(insn, retype(brw_vec1_grf(0,0), BRW_REGISTER_TYPE_UD));
612 brw_set_src0(insn, retype(brw_vec1_grf(0,0), BRW_REGISTER_TYPE_UD));
613 brw_set_src1(insn, retype(brw_vec1_grf(0,0), BRW_REGISTER_TYPE_UD));
614
615 insn->header.compression_control = BRW_COMPRESSION_NONE;
616 insn->header.execution_size = execute_size;
617 /* insn->header.mask_control = BRW_MASK_ENABLE; */
618
619 return insn;
620 }
621 }
622
623
624
625 void brw_WHILE(struct brw_compile *p,
626 struct brw_instruction *do_insn)
627 {
628 struct brw_instruction *insn;
629
630 if (p->single_program_flow)
631 insn = next_insn(p, BRW_OPCODE_ADD);
632 else
633 insn = next_insn(p, BRW_OPCODE_WHILE);
634
635 brw_set_dest(insn, brw_ip_reg());
636 brw_set_src0(insn, brw_ip_reg());
637 brw_set_src1(insn, brw_imm_d(0x0));
638
639 insn->header.compression_control = BRW_COMPRESSION_NONE;
640
641 if (p->single_program_flow) {
642 insn->header.execution_size = BRW_EXECUTE_1;
643
644 insn->bits3.d = (do_insn - insn) * 16;
645 } else {
646 insn->header.execution_size = do_insn->header.execution_size;
647
648 assert(do_insn->header.opcode == BRW_OPCODE_DO);
649 insn->bits3.if_else.jump_count = do_insn - insn;
650 insn->bits3.if_else.pop_count = 0;
651 insn->bits3.if_else.pad0 = 0;
652 }
653
654 /* insn->header.mask_control = BRW_MASK_ENABLE; */
655
656 p->current->header.predicate_control = BRW_PREDICATE_NONE;
657 }
658
659
660 /* FORWARD JUMPS:
661 */
662 void brw_land_fwd_jump(struct brw_compile *p,
663 struct brw_instruction *jmp_insn)
664 {
665 struct brw_instruction *landing = &p->store[p->nr_insn];
666
667 assert(jmp_insn->header.opcode == BRW_OPCODE_JMPI);
668 assert(jmp_insn->bits1.da1.src1_reg_file = BRW_IMMEDIATE_VALUE);
669
670 jmp_insn->bits3.ud = (landing - jmp_insn) - 1;
671 }
672
673
674
675 /* To integrate with the above, it makes sense that the comparison
676 * instruction should populate the flag register. It might be simpler
677 * just to use the flag reg for most WM tasks?
678 */
679 void brw_CMP(struct brw_compile *p,
680 struct brw_reg dest,
681 GLuint conditional,
682 struct brw_reg src0,
683 struct brw_reg src1)
684 {
685 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_CMP);
686
687 insn->header.destreg__conditonalmod = conditional;
688 brw_set_dest(insn, dest);
689 brw_set_src0(insn, src0);
690 brw_set_src1(insn, src1);
691
692 /* guess_execution_size(insn, src0); */
693
694
695 /* Make it so that future instructions will use the computed flag
696 * value until brw_set_predicate_control_flag_value() is called
697 * again.
698 */
699 if (dest.file == BRW_ARCHITECTURE_REGISTER_FILE &&
700 dest.nr == 0) {
701 p->current->header.predicate_control = BRW_PREDICATE_NORMAL;
702 p->flag_value = 0xff;
703 }
704 }
705
706
707
708 /***********************************************************************
709 * Helpers for the various SEND message types:
710 */
711
712 /* Invert 8 values
713 */
714 void brw_math( struct brw_compile *p,
715 struct brw_reg dest,
716 GLuint function,
717 GLuint saturate,
718 GLuint msg_reg_nr,
719 struct brw_reg src,
720 GLuint data_type,
721 GLuint precision )
722 {
723 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
724 GLuint msg_length = (function == BRW_MATH_FUNCTION_POW) ? 2 : 1;
725 GLuint response_length = (function == BRW_MATH_FUNCTION_SINCOS) ? 2 : 1;
726
727 /* Example code doesn't set predicate_control for send
728 * instructions.
729 */
730 insn->header.predicate_control = 0;
731 insn->header.destreg__conditonalmod = msg_reg_nr;
732
733 brw_set_dest(insn, dest);
734 brw_set_src0(insn, src);
735 brw_set_math_message(insn,
736 msg_length, response_length,
737 function,
738 BRW_MATH_INTEGER_UNSIGNED,
739 precision,
740 saturate,
741 data_type);
742 }
743
744 /* Use 2 send instructions to invert 16 elements
745 */
746 void brw_math_16( struct brw_compile *p,
747 struct brw_reg dest,
748 GLuint function,
749 GLuint saturate,
750 GLuint msg_reg_nr,
751 struct brw_reg src,
752 GLuint precision )
753 {
754 struct brw_instruction *insn;
755 GLuint msg_length = (function == BRW_MATH_FUNCTION_POW) ? 2 : 1;
756 GLuint response_length = (function == BRW_MATH_FUNCTION_SINCOS) ? 2 : 1;
757
758 /* First instruction:
759 */
760 brw_push_insn_state(p);
761 brw_set_predicate_control_flag_value(p, 0xff);
762 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
763
764 insn = next_insn(p, BRW_OPCODE_SEND);
765 insn->header.destreg__conditonalmod = msg_reg_nr;
766
767 brw_set_dest(insn, dest);
768 brw_set_src0(insn, src);
769 brw_set_math_message(insn,
770 msg_length, response_length,
771 function,
772 BRW_MATH_INTEGER_UNSIGNED,
773 precision,
774 saturate,
775 BRW_MATH_DATA_VECTOR);
776
777 /* Second instruction:
778 */
779 insn = next_insn(p, BRW_OPCODE_SEND);
780 insn->header.compression_control = BRW_COMPRESSION_2NDHALF;
781 insn->header.destreg__conditonalmod = msg_reg_nr+1;
782
783 brw_set_dest(insn, offset(dest,1));
784 brw_set_src0(insn, src);
785 brw_set_math_message(insn,
786 msg_length, response_length,
787 function,
788 BRW_MATH_INTEGER_UNSIGNED,
789 precision,
790 saturate,
791 BRW_MATH_DATA_VECTOR);
792
793 brw_pop_insn_state(p);
794 }
795
796
797
798
799 void brw_dp_WRITE_16( struct brw_compile *p,
800 struct brw_reg src,
801 GLuint msg_reg_nr,
802 GLuint scratch_offset )
803 {
804 {
805 brw_push_insn_state(p);
806 brw_set_mask_control(p, BRW_MASK_DISABLE);
807 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
808
809 brw_MOV(p,
810 retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_D),
811 brw_imm_d(scratch_offset));
812
813 brw_pop_insn_state(p);
814 }
815
816 {
817 GLuint msg_length = 3;
818 struct brw_reg dest = retype(brw_null_reg(), BRW_REGISTER_TYPE_UW);
819 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
820
821 insn->header.predicate_control = 0; /* XXX */
822 insn->header.compression_control = BRW_COMPRESSION_NONE;
823 insn->header.destreg__conditonalmod = msg_reg_nr;
824
825 brw_set_dest(insn, dest);
826 brw_set_src0(insn, src);
827
828 brw_set_dp_write_message(insn,
829 255, /* bti */
830 BRW_DATAPORT_OWORD_BLOCK_4_OWORDS, /* msg_control */
831 BRW_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE, /* msg_type */
832 msg_length,
833 0, /* pixel scoreboard */
834 0, /* response_length */
835 0); /* eot */
836 }
837
838 }
839
840
841 void brw_dp_READ_16( struct brw_compile *p,
842 struct brw_reg dest,
843 GLuint msg_reg_nr,
844 GLuint scratch_offset )
845 {
846 {
847 brw_push_insn_state(p);
848 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
849 brw_set_mask_control(p, BRW_MASK_DISABLE);
850
851 brw_MOV(p,
852 retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_D),
853 brw_imm_d(scratch_offset));
854
855 brw_pop_insn_state(p);
856 }
857
858 {
859 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
860
861 insn->header.predicate_control = 0; /* XXX */
862 insn->header.compression_control = BRW_COMPRESSION_NONE;
863 insn->header.destreg__conditonalmod = msg_reg_nr;
864
865 brw_set_dest(insn, dest); /* UW? */
866 brw_set_src0(insn, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW));
867
868 brw_set_dp_read_message(insn,
869 255, /* bti */
870 3, /* msg_control */
871 BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ, /* msg_type */
872 1, /* target cache */
873 1, /* msg_length */
874 2, /* response_length */
875 0); /* eot */
876 }
877 }
878
879
880 void brw_fb_WRITE(struct brw_compile *p,
881 struct brw_reg dest,
882 GLuint msg_reg_nr,
883 struct brw_reg src0,
884 GLuint binding_table_index,
885 GLuint msg_length,
886 GLuint response_length,
887 GLboolean eot)
888 {
889 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
890
891 insn->header.predicate_control = 0; /* XXX */
892 insn->header.compression_control = BRW_COMPRESSION_NONE;
893 insn->header.destreg__conditonalmod = msg_reg_nr;
894
895 brw_set_dest(insn, dest);
896 brw_set_src0(insn, src0);
897 brw_set_dp_write_message(insn,
898 binding_table_index,
899 BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE, /* msg_control */
900 BRW_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE, /* msg_type */
901 msg_length,
902 1, /* pixel scoreboard */
903 response_length,
904 eot);
905 }
906
907
908
909 void brw_SAMPLE(struct brw_compile *p,
910 struct brw_reg dest,
911 GLuint msg_reg_nr,
912 struct brw_reg src0,
913 GLuint binding_table_index,
914 GLuint sampler,
915 GLuint writemask,
916 GLuint msg_type,
917 GLuint response_length,
918 GLuint msg_length,
919 GLboolean eot)
920 {
921 GLboolean need_stall = 0;
922
923 if(writemask == 0) {
924 /* _mesa_printf("%s: zero writemask??\n", __FUNCTION__); */
925 return;
926 }
927
928 /* Hardware doesn't do destination dependency checking on send
929 * instructions properly. Add a workaround which generates the
930 * dependency by other means. In practice it seems like this bug
931 * only crops up for texture samples, and only where registers are
932 * written by the send and then written again later without being
933 * read in between. Luckily for us, we already track that
934 * information and use it to modify the writemask for the
935 * instruction, so that is a guide for whether a workaround is
936 * needed.
937 */
938 if (writemask != WRITEMASK_XYZW) {
939 GLuint dst_offset = 0;
940 GLuint i, newmask = 0, len = 0;
941
942 for (i = 0; i < 4; i++) {
943 if (writemask & (1<<i))
944 break;
945 dst_offset += 2;
946 }
947 for (; i < 4; i++) {
948 if (!(writemask & (1<<i)))
949 break;
950 newmask |= 1<<i;
951 len++;
952 }
953
954 if (newmask != writemask) {
955 need_stall = 1;
956 /* _mesa_printf("need stall %x %x\n", newmask , writemask); */
957 }
958 else {
959 struct brw_reg m1 = brw_message_reg(msg_reg_nr);
960
961 newmask = ~newmask & WRITEMASK_XYZW;
962
963 brw_push_insn_state(p);
964
965 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
966 brw_set_mask_control(p, BRW_MASK_DISABLE);
967
968 brw_MOV(p, m1, brw_vec8_grf(0,0));
969 brw_MOV(p, get_element_ud(m1, 2), brw_imm_ud(newmask << 12));
970
971 brw_pop_insn_state(p);
972
973 src0 = retype(brw_null_reg(), BRW_REGISTER_TYPE_UW);
974 dest = offset(dest, dst_offset);
975 response_length = len * 2;
976 }
977 }
978
979 {
980 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
981
982 insn->header.predicate_control = 0; /* XXX */
983 insn->header.compression_control = BRW_COMPRESSION_NONE;
984 insn->header.destreg__conditonalmod = msg_reg_nr;
985
986 brw_set_dest(insn, dest);
987 brw_set_src0(insn, src0);
988 brw_set_sampler_message(insn,
989 binding_table_index,
990 sampler,
991 msg_type,
992 response_length,
993 msg_length,
994 eot);
995 }
996
997 if (need_stall)
998 {
999 struct brw_reg reg = vec8(offset(dest, response_length-1));
1000
1001 /* mov (8) r9.0<1>:f r9.0<8;8,1>:f { Align1 }
1002 */
1003 brw_push_insn_state(p);
1004 brw_set_compression_control(p, GL_FALSE);
1005 brw_MOV(p, reg, reg);
1006 brw_pop_insn_state(p);
1007 }
1008
1009 }
1010
1011 /* All these variables are pretty confusing - we might be better off
1012 * using bitmasks and macros for this, in the old style. Or perhaps
1013 * just having the caller instantiate the fields in dword3 itself.
1014 */
1015 void brw_urb_WRITE(struct brw_compile *p,
1016 struct brw_reg dest,
1017 GLuint msg_reg_nr,
1018 struct brw_reg src0,
1019 GLboolean allocate,
1020 GLboolean used,
1021 GLuint msg_length,
1022 GLuint response_length,
1023 GLboolean eot,
1024 GLboolean writes_complete,
1025 GLuint offset,
1026 GLuint swizzle)
1027 {
1028 struct brw_instruction *insn = next_insn(p, BRW_OPCODE_SEND);
1029
1030 assert(msg_length < 16);
1031
1032 brw_set_dest(insn, dest);
1033 brw_set_src0(insn, src0);
1034 brw_set_src1(insn, brw_imm_d(0));
1035
1036 insn->header.destreg__conditonalmod = msg_reg_nr;
1037
1038 brw_set_urb_message(insn,
1039 allocate,
1040 used,
1041 msg_length,
1042 response_length,
1043 eot,
1044 writes_complete,
1045 offset,
1046 swizzle);
1047 }
1048