1 /* Copyright © 2011 Intel Corporation
2  *
3  * Permission is hereby granted, free of charge, to any person obtaining a
4  * copy of this software and associated documentation files (the "Software"),
5  * to deal in the Software without restriction, including without limitation
6  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
7  * and/or sell copies of the Software, and to permit persons to whom the
8  * Software is furnished to do so, subject to the following conditions:
9  *
10  * The above copyright notice and this permission notice (including the next
11  * paragraph) shall be included in all copies or substantial portions of the
12  * Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
20  * IN THE SOFTWARE.
21  */
22 
23 #include "brw_vec4.h"
24 #include "glsl/ir_print_visitor.h"
25 
26 extern "C" {
27 #include "brw_eu.h"
28 #include "main/macros.h"
29 };
30 
31 using namespace brw;
32 
33 namespace brw {
34 
35 int
setup_attributes(int payload_reg)36 vec4_visitor::setup_attributes(int payload_reg)
37 {
38    int nr_attributes;
39    int attribute_map[VERT_ATTRIB_MAX + 1];
40 
41    nr_attributes = 0;
42    for (int i = 0; i < VERT_ATTRIB_MAX; i++) {
43       if (prog_data->inputs_read & BITFIELD64_BIT(i)) {
44 	 attribute_map[i] = payload_reg + nr_attributes;
45 	 nr_attributes++;
46       }
47    }
48 
49    /* VertexID is stored by the VF as the last vertex element, but we
50     * don't represent it with a flag in inputs_read, so we call it
51     * VERT_ATTRIB_MAX.
52     */
53    if (prog_data->uses_vertexid) {
54       attribute_map[VERT_ATTRIB_MAX] = payload_reg + nr_attributes;
55       nr_attributes++;
56    }
57 
58    foreach_list(node, &this->instructions) {
59       vec4_instruction *inst = (vec4_instruction *)node;
60 
61       /* We have to support ATTR as a destination for GL_FIXED fixup. */
62       if (inst->dst.file == ATTR) {
63 	 int grf = attribute_map[inst->dst.reg + inst->dst.reg_offset];
64 
65 	 struct brw_reg reg = brw_vec8_grf(grf, 0);
66 	 reg.dw1.bits.writemask = inst->dst.writemask;
67 
68 	 inst->dst.file = HW_REG;
69 	 inst->dst.fixed_hw_reg = reg;
70       }
71 
72       for (int i = 0; i < 3; i++) {
73 	 if (inst->src[i].file != ATTR)
74 	    continue;
75 
76 	 int grf = attribute_map[inst->src[i].reg + inst->src[i].reg_offset];
77 
78 	 struct brw_reg reg = brw_vec8_grf(grf, 0);
79 	 reg.dw1.bits.swizzle = inst->src[i].swizzle;
80          reg.type = inst->src[i].type;
81 	 if (inst->src[i].abs)
82 	    reg = brw_abs(reg);
83 	 if (inst->src[i].negate)
84 	    reg = negate(reg);
85 
86 	 inst->src[i].file = HW_REG;
87 	 inst->src[i].fixed_hw_reg = reg;
88       }
89    }
90 
91    /* The BSpec says we always have to read at least one thing from
92     * the VF, and it appears that the hardware wedges otherwise.
93     */
94    if (nr_attributes == 0)
95       nr_attributes = 1;
96 
97    prog_data->urb_read_length = (nr_attributes + 1) / 2;
98 
99    unsigned vue_entries = MAX2(nr_attributes, c->prog_data.vue_map.num_slots);
100 
101    if (intel->gen == 6)
102       c->prog_data.urb_entry_size = ALIGN(vue_entries, 8) / 8;
103    else
104       c->prog_data.urb_entry_size = ALIGN(vue_entries, 4) / 4;
105 
106    return payload_reg + nr_attributes;
107 }
108 
109 int
setup_uniforms(int reg)110 vec4_visitor::setup_uniforms(int reg)
111 {
112    /* The pre-gen6 VS requires that some push constants get loaded no
113     * matter what, or the GPU would hang.
114     */
115    if (intel->gen < 6 && this->uniforms == 0) {
116       this->uniform_vector_size[this->uniforms] = 1;
117 
118       for (unsigned int i = 0; i < 4; i++) {
119 	 unsigned int slot = this->uniforms * 4 + i;
120 	 static float zero = 0.0;
121 	 c->prog_data.param[slot] = &zero;
122       }
123 
124       this->uniforms++;
125       reg++;
126    } else {
127       reg += ALIGN(uniforms, 2) / 2;
128    }
129 
130    c->prog_data.nr_params = this->uniforms * 4;
131 
132    c->prog_data.curb_read_length = reg - 1;
133    c->prog_data.uses_new_param_layout = true;
134 
135    return reg;
136 }
137 
138 void
setup_payload(void)139 vec4_visitor::setup_payload(void)
140 {
141    int reg = 0;
142 
143    /* The payload always contains important data in g0, which contains
144     * the URB handles that are passed on to the URB write at the end
145     * of the thread.  So, we always start push constants at g1.
146     */
147    reg++;
148 
149    reg = setup_uniforms(reg);
150 
151    reg = setup_attributes(reg);
152 
153    this->first_non_payload_grf = reg;
154 }
155 
156 struct brw_reg
get_dst(void)157 vec4_instruction::get_dst(void)
158 {
159    struct brw_reg brw_reg;
160 
161    switch (dst.file) {
162    case GRF:
163       brw_reg = brw_vec8_grf(dst.reg + dst.reg_offset, 0);
164       brw_reg = retype(brw_reg, dst.type);
165       brw_reg.dw1.bits.writemask = dst.writemask;
166       break;
167 
168    case MRF:
169       brw_reg = brw_message_reg(dst.reg + dst.reg_offset);
170       brw_reg = retype(brw_reg, dst.type);
171       brw_reg.dw1.bits.writemask = dst.writemask;
172       break;
173 
174    case HW_REG:
175       brw_reg = dst.fixed_hw_reg;
176       break;
177 
178    case BAD_FILE:
179       brw_reg = brw_null_reg();
180       break;
181 
182    default:
183       assert(!"not reached");
184       brw_reg = brw_null_reg();
185       break;
186    }
187    return brw_reg;
188 }
189 
190 struct brw_reg
get_src(int i)191 vec4_instruction::get_src(int i)
192 {
193    struct brw_reg brw_reg;
194 
195    switch (src[i].file) {
196    case GRF:
197       brw_reg = brw_vec8_grf(src[i].reg + src[i].reg_offset, 0);
198       brw_reg = retype(brw_reg, src[i].type);
199       brw_reg.dw1.bits.swizzle = src[i].swizzle;
200       if (src[i].abs)
201 	 brw_reg = brw_abs(brw_reg);
202       if (src[i].negate)
203 	 brw_reg = negate(brw_reg);
204       break;
205 
206    case IMM:
207       switch (src[i].type) {
208       case BRW_REGISTER_TYPE_F:
209 	 brw_reg = brw_imm_f(src[i].imm.f);
210 	 break;
211       case BRW_REGISTER_TYPE_D:
212 	 brw_reg = brw_imm_d(src[i].imm.i);
213 	 break;
214       case BRW_REGISTER_TYPE_UD:
215 	 brw_reg = brw_imm_ud(src[i].imm.u);
216 	 break;
217       default:
218 	 assert(!"not reached");
219 	 brw_reg = brw_null_reg();
220 	 break;
221       }
222       break;
223 
224    case UNIFORM:
225       brw_reg = stride(brw_vec4_grf(1 + (src[i].reg + src[i].reg_offset) / 2,
226 				    ((src[i].reg + src[i].reg_offset) % 2) * 4),
227 		       0, 4, 1);
228       brw_reg = retype(brw_reg, src[i].type);
229       brw_reg.dw1.bits.swizzle = src[i].swizzle;
230       if (src[i].abs)
231 	 brw_reg = brw_abs(brw_reg);
232       if (src[i].negate)
233 	 brw_reg = negate(brw_reg);
234 
235       /* This should have been moved to pull constants. */
236       assert(!src[i].reladdr);
237       break;
238 
239    case HW_REG:
240       brw_reg = src[i].fixed_hw_reg;
241       break;
242 
243    case BAD_FILE:
244       /* Probably unused. */
245       brw_reg = brw_null_reg();
246       break;
247    case ATTR:
248    default:
249       assert(!"not reached");
250       brw_reg = brw_null_reg();
251       break;
252    }
253 
254    return brw_reg;
255 }
256 
257 void
generate_math1_gen4(vec4_instruction * inst,struct brw_reg dst,struct brw_reg src)258 vec4_visitor::generate_math1_gen4(vec4_instruction *inst,
259 				  struct brw_reg dst,
260 				  struct brw_reg src)
261 {
262    brw_math(p,
263 	    dst,
264 	    brw_math_function(inst->opcode),
265 	    inst->base_mrf,
266 	    src,
267 	    BRW_MATH_DATA_VECTOR,
268 	    BRW_MATH_PRECISION_FULL);
269 }
270 
271 static void
check_gen6_math_src_arg(struct brw_reg src)272 check_gen6_math_src_arg(struct brw_reg src)
273 {
274    /* Source swizzles are ignored. */
275    assert(!src.abs);
276    assert(!src.negate);
277    assert(src.dw1.bits.swizzle == BRW_SWIZZLE_XYZW);
278 }
279 
280 void
generate_math1_gen6(vec4_instruction * inst,struct brw_reg dst,struct brw_reg src)281 vec4_visitor::generate_math1_gen6(vec4_instruction *inst,
282 				  struct brw_reg dst,
283 				  struct brw_reg src)
284 {
285    /* Can't do writemask because math can't be align16. */
286    assert(dst.dw1.bits.writemask == WRITEMASK_XYZW);
287    check_gen6_math_src_arg(src);
288 
289    brw_set_access_mode(p, BRW_ALIGN_1);
290    brw_math(p,
291 	    dst,
292 	    brw_math_function(inst->opcode),
293 	    inst->base_mrf,
294 	    src,
295 	    BRW_MATH_DATA_SCALAR,
296 	    BRW_MATH_PRECISION_FULL);
297    brw_set_access_mode(p, BRW_ALIGN_16);
298 }
299 
300 void
generate_math2_gen7(vec4_instruction * inst,struct brw_reg dst,struct brw_reg src0,struct brw_reg src1)301 vec4_visitor::generate_math2_gen7(vec4_instruction *inst,
302 				  struct brw_reg dst,
303 				  struct brw_reg src0,
304 				  struct brw_reg src1)
305 {
306    brw_math2(p,
307 	     dst,
308 	     brw_math_function(inst->opcode),
309 	     src0, src1);
310 }
311 
312 void
generate_math2_gen6(vec4_instruction * inst,struct brw_reg dst,struct brw_reg src0,struct brw_reg src1)313 vec4_visitor::generate_math2_gen6(vec4_instruction *inst,
314 				  struct brw_reg dst,
315 				  struct brw_reg src0,
316 				  struct brw_reg src1)
317 {
318    /* Can't do writemask because math can't be align16. */
319    assert(dst.dw1.bits.writemask == WRITEMASK_XYZW);
320    /* Source swizzles are ignored. */
321    check_gen6_math_src_arg(src0);
322    check_gen6_math_src_arg(src1);
323 
324    brw_set_access_mode(p, BRW_ALIGN_1);
325    brw_math2(p,
326 	     dst,
327 	     brw_math_function(inst->opcode),
328 	     src0, src1);
329    brw_set_access_mode(p, BRW_ALIGN_16);
330 }
331 
332 void
generate_math2_gen4(vec4_instruction * inst,struct brw_reg dst,struct brw_reg src0,struct brw_reg src1)333 vec4_visitor::generate_math2_gen4(vec4_instruction *inst,
334 				  struct brw_reg dst,
335 				  struct brw_reg src0,
336 				  struct brw_reg src1)
337 {
338    /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
339     * "Message Payload":
340     *
341     * "Operand0[7].  For the INT DIV functions, this operand is the
342     *  denominator."
343     *  ...
344     * "Operand1[7].  For the INT DIV functions, this operand is the
345     *  numerator."
346     */
347    bool is_int_div = inst->opcode != SHADER_OPCODE_POW;
348    struct brw_reg &op0 = is_int_div ? src1 : src0;
349    struct brw_reg &op1 = is_int_div ? src0 : src1;
350 
351    brw_push_insn_state(p);
352    brw_set_saturate(p, false);
353    brw_set_predicate_control(p, BRW_PREDICATE_NONE);
354    brw_MOV(p, retype(brw_message_reg(inst->base_mrf + 1), op1.type), op1);
355    brw_pop_insn_state(p);
356 
357    brw_math(p,
358 	    dst,
359 	    brw_math_function(inst->opcode),
360 	    inst->base_mrf,
361 	    op0,
362 	    BRW_MATH_DATA_VECTOR,
363 	    BRW_MATH_PRECISION_FULL);
364 }
365 
366 void
generate_tex(vec4_instruction * inst,struct brw_reg dst,struct brw_reg src)367 vec4_visitor::generate_tex(vec4_instruction *inst,
368 			   struct brw_reg dst,
369 			   struct brw_reg src)
370 {
371    int msg_type = -1;
372 
373    if (intel->gen >= 5) {
374       switch (inst->opcode) {
375       case SHADER_OPCODE_TEX:
376       case SHADER_OPCODE_TXL:
377 	 if (inst->shadow_compare) {
378 	    msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
379 	 } else {
380 	    msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
381 	 }
382 	 break;
383       case SHADER_OPCODE_TXD:
384          if (inst->shadow_compare) {
385             /* Gen7.5+.  Otherwise, lowered by brw_lower_texture_gradients(). */
386             assert(intel->is_haswell);
387             msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE;
388          } else {
389             msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
390          }
391 	 break;
392       case SHADER_OPCODE_TXF:
393 	 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
394 	 break;
395       case SHADER_OPCODE_TXS:
396 	 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
397 	 break;
398       default:
399 	 assert(!"should not get here: invalid VS texture opcode");
400 	 break;
401       }
402    } else {
403       switch (inst->opcode) {
404       case SHADER_OPCODE_TEX:
405       case SHADER_OPCODE_TXL:
406 	 if (inst->shadow_compare) {
407 	    msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD_COMPARE;
408 	    assert(inst->mlen == 3);
409 	 } else {
410 	    msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD;
411 	    assert(inst->mlen == 2);
412 	 }
413 	 break;
414       case SHADER_OPCODE_TXD:
415 	 /* There is no sample_d_c message; comparisons are done manually. */
416 	 msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_GRADIENTS;
417 	 assert(inst->mlen == 4);
418 	 break;
419       case SHADER_OPCODE_TXF:
420 	 msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_LD;
421 	 assert(inst->mlen == 2);
422 	 break;
423       case SHADER_OPCODE_TXS:
424 	 msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_RESINFO;
425 	 assert(inst->mlen == 2);
426 	 break;
427       default:
428 	 assert(!"should not get here: invalid VS texture opcode");
429 	 break;
430       }
431    }
432 
433    assert(msg_type != -1);
434 
435    /* Load the message header if present.  If there's a texture offset, we need
436     * to set it up explicitly and load the offset bitfield.  Otherwise, we can
437     * use an implied move from g0 to the first message register.
438     */
439    if (inst->texture_offset) {
440       /* Explicitly set up the message header by copying g0 to the MRF. */
441       brw_MOV(p, retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD),
442 	         retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
443 
444       /* Then set the offset bits in DWord 2. */
445       brw_set_access_mode(p, BRW_ALIGN_1);
446       brw_MOV(p,
447 	      retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE, inst->base_mrf, 2),
448 		     BRW_REGISTER_TYPE_UD),
449 	      brw_imm_uw(inst->texture_offset));
450       brw_set_access_mode(p, BRW_ALIGN_16);
451    } else if (inst->header_present) {
452       /* Set up an implied move from g0 to the MRF. */
453       src = brw_vec8_grf(0, 0);
454    }
455 
456    uint32_t return_format;
457 
458    switch (dst.type) {
459    case BRW_REGISTER_TYPE_D:
460       return_format = BRW_SAMPLER_RETURN_FORMAT_SINT32;
461       break;
462    case BRW_REGISTER_TYPE_UD:
463       return_format = BRW_SAMPLER_RETURN_FORMAT_UINT32;
464       break;
465    default:
466       return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
467       break;
468    }
469 
470    brw_SAMPLE(p,
471 	      dst,
472 	      inst->base_mrf,
473 	      src,
474 	      SURF_INDEX_VS_TEXTURE(inst->sampler),
475 	      inst->sampler,
476 	      WRITEMASK_XYZW,
477 	      msg_type,
478 	      1, /* response length */
479 	      inst->mlen,
480 	      inst->header_present,
481 	      BRW_SAMPLER_SIMD_MODE_SIMD4X2,
482 	      return_format);
483 }
484 
485 void
generate_urb_write(vec4_instruction * inst)486 vec4_visitor::generate_urb_write(vec4_instruction *inst)
487 {
488    brw_urb_WRITE(p,
489 		 brw_null_reg(), /* dest */
490 		 inst->base_mrf, /* starting mrf reg nr */
491 		 brw_vec8_grf(0, 0), /* src */
492 		 false,		/* allocate */
493 		 true,		/* used */
494 		 inst->mlen,
495 		 0,		/* response len */
496 		 inst->eot,	/* eot */
497 		 inst->eot,	/* writes complete */
498 		 inst->offset,	/* urb destination offset */
499 		 BRW_URB_SWIZZLE_INTERLEAVE);
500 }
501 
502 void
generate_oword_dual_block_offsets(struct brw_reg m1,struct brw_reg index)503 vec4_visitor::generate_oword_dual_block_offsets(struct brw_reg m1,
504 						struct brw_reg index)
505 {
506    int second_vertex_offset;
507 
508    if (intel->gen >= 6)
509       second_vertex_offset = 1;
510    else
511       second_vertex_offset = 16;
512 
513    m1 = retype(m1, BRW_REGISTER_TYPE_D);
514 
515    /* Set up M1 (message payload).  Only the block offsets in M1.0 and
516     * M1.4 are used, and the rest are ignored.
517     */
518    struct brw_reg m1_0 = suboffset(vec1(m1), 0);
519    struct brw_reg m1_4 = suboffset(vec1(m1), 4);
520    struct brw_reg index_0 = suboffset(vec1(index), 0);
521    struct brw_reg index_4 = suboffset(vec1(index), 4);
522 
523    brw_push_insn_state(p);
524    brw_set_mask_control(p, BRW_MASK_DISABLE);
525    brw_set_access_mode(p, BRW_ALIGN_1);
526 
527    brw_MOV(p, m1_0, index_0);
528 
529    brw_set_predicate_inverse(p, true);
530    if (index.file == BRW_IMMEDIATE_VALUE) {
531       index_4.dw1.ud += second_vertex_offset;
532       brw_MOV(p, m1_4, index_4);
533    } else {
534       brw_ADD(p, m1_4, index_4, brw_imm_d(second_vertex_offset));
535    }
536 
537    brw_pop_insn_state(p);
538 }
539 
540 void
generate_scratch_read(vec4_instruction * inst,struct brw_reg dst,struct brw_reg index)541 vec4_visitor::generate_scratch_read(vec4_instruction *inst,
542 				    struct brw_reg dst,
543 				    struct brw_reg index)
544 {
545    struct brw_reg header = brw_vec8_grf(0, 0);
546 
547    gen6_resolve_implied_move(p, &header, inst->base_mrf);
548 
549    generate_oword_dual_block_offsets(brw_message_reg(inst->base_mrf + 1),
550 				     index);
551 
552    uint32_t msg_type;
553 
554    if (intel->gen >= 6)
555       msg_type = GEN6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ;
556    else if (intel->gen == 5 || intel->is_g4x)
557       msg_type = G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ;
558    else
559       msg_type = BRW_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ;
560 
561    /* Each of the 8 channel enables is considered for whether each
562     * dword is written.
563     */
564    struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
565    brw_set_dest(p, send, dst);
566    brw_set_src0(p, send, header);
567    if (intel->gen < 6)
568       send->header.destreg__conditionalmod = inst->base_mrf;
569    brw_set_dp_read_message(p, send,
570 			   255, /* binding table index: stateless access */
571 			   BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD,
572 			   msg_type,
573 			   BRW_DATAPORT_READ_TARGET_RENDER_CACHE,
574 			   2, /* mlen */
575 			   1 /* rlen */);
576 }
577 
578 void
generate_scratch_write(vec4_instruction * inst,struct brw_reg dst,struct brw_reg src,struct brw_reg index)579 vec4_visitor::generate_scratch_write(vec4_instruction *inst,
580 				     struct brw_reg dst,
581 				     struct brw_reg src,
582 				     struct brw_reg index)
583 {
584    struct brw_reg header = brw_vec8_grf(0, 0);
585    bool write_commit;
586 
587    /* If the instruction is predicated, we'll predicate the send, not
588     * the header setup.
589     */
590    brw_set_predicate_control(p, false);
591 
592    gen6_resolve_implied_move(p, &header, inst->base_mrf);
593 
594    generate_oword_dual_block_offsets(brw_message_reg(inst->base_mrf + 1),
595 				     index);
596 
597    brw_MOV(p,
598 	   retype(brw_message_reg(inst->base_mrf + 2), BRW_REGISTER_TYPE_D),
599 	   retype(src, BRW_REGISTER_TYPE_D));
600 
601    uint32_t msg_type;
602 
603    if (intel->gen >= 7)
604       msg_type = GEN7_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE;
605    else if (intel->gen == 6)
606       msg_type = GEN6_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE;
607    else
608       msg_type = BRW_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE;
609 
610    brw_set_predicate_control(p, inst->predicate);
611 
612    /* Pre-gen6, we have to specify write commits to ensure ordering
613     * between reads and writes within a thread.  Afterwards, that's
614     * guaranteed and write commits only matter for inter-thread
615     * synchronization.
616     */
617    if (intel->gen >= 6) {
618       write_commit = false;
619    } else {
620       /* The visitor set up our destination register to be g0.  This
621        * means that when the next read comes along, we will end up
622        * reading from g0 and causing a block on the write commit.  For
623        * write-after-read, we are relying on the value of the previous
624        * read being used (and thus blocking on completion) before our
625        * write is executed.  This means we have to be careful in
626        * instruction scheduling to not violate this assumption.
627        */
628       write_commit = true;
629    }
630 
631    /* Each of the 8 channel enables is considered for whether each
632     * dword is written.
633     */
634    struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
635    brw_set_dest(p, send, dst);
636    brw_set_src0(p, send, header);
637    if (intel->gen < 6)
638       send->header.destreg__conditionalmod = inst->base_mrf;
639    brw_set_dp_write_message(p, send,
640 			    255, /* binding table index: stateless access */
641 			    BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD,
642 			    msg_type,
643 			    3, /* mlen */
644 			    true, /* header present */
645 			    false, /* not a render target write */
646 			    write_commit, /* rlen */
647 			    false, /* eot */
648 			    write_commit);
649 }
650 
651 void
generate_pull_constant_load(vec4_instruction * inst,struct brw_reg dst,struct brw_reg index,struct brw_reg offset)652 vec4_visitor::generate_pull_constant_load(vec4_instruction *inst,
653 					  struct brw_reg dst,
654 					  struct brw_reg index,
655 					  struct brw_reg offset)
656 {
657    assert(index.file == BRW_IMMEDIATE_VALUE &&
658 	  index.type == BRW_REGISTER_TYPE_UD);
659    uint32_t surf_index = index.dw1.ud;
660 
661    if (intel->gen == 7) {
662       gen6_resolve_implied_move(p, &offset, inst->base_mrf);
663       brw_instruction *insn = brw_next_insn(p, BRW_OPCODE_SEND);
664       brw_set_dest(p, insn, dst);
665       brw_set_src0(p, insn, offset);
666       brw_set_sampler_message(p, insn,
667                               surf_index,
668                               0, /* LD message ignores sampler unit */
669                               GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
670                               1, /* rlen */
671                               1, /* mlen */
672                               false, /* no header */
673                               BRW_SAMPLER_SIMD_MODE_SIMD4X2,
674                               0);
675       return;
676    }
677 
678    struct brw_reg header = brw_vec8_grf(0, 0);
679 
680    gen6_resolve_implied_move(p, &header, inst->base_mrf);
681 
682    brw_MOV(p, retype(brw_message_reg(inst->base_mrf + 1), BRW_REGISTER_TYPE_D),
683 	   offset);
684 
685    uint32_t msg_type;
686 
687    if (intel->gen >= 6)
688       msg_type = GEN6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ;
689    else if (intel->gen == 5 || intel->is_g4x)
690       msg_type = G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ;
691    else
692       msg_type = BRW_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ;
693 
694    /* Each of the 8 channel enables is considered for whether each
695     * dword is written.
696     */
697    struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
698    brw_set_dest(p, send, dst);
699    brw_set_src0(p, send, header);
700    if (intel->gen < 6)
701       send->header.destreg__conditionalmod = inst->base_mrf;
702    brw_set_dp_read_message(p, send,
703 			   surf_index,
704 			   BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD,
705 			   msg_type,
706 			   BRW_DATAPORT_READ_TARGET_DATA_CACHE,
707 			   2, /* mlen */
708 			   1 /* rlen */);
709 }
710 
711 void
generate_vs_instruction(vec4_instruction * instruction,struct brw_reg dst,struct brw_reg * src)712 vec4_visitor::generate_vs_instruction(vec4_instruction *instruction,
713 				      struct brw_reg dst,
714 				      struct brw_reg *src)
715 {
716    vec4_instruction *inst = (vec4_instruction *)instruction;
717 
718    switch (inst->opcode) {
719    case SHADER_OPCODE_RCP:
720    case SHADER_OPCODE_RSQ:
721    case SHADER_OPCODE_SQRT:
722    case SHADER_OPCODE_EXP2:
723    case SHADER_OPCODE_LOG2:
724    case SHADER_OPCODE_SIN:
725    case SHADER_OPCODE_COS:
726       if (intel->gen == 6) {
727 	 generate_math1_gen6(inst, dst, src[0]);
728       } else {
729 	 /* Also works for Gen7. */
730 	 generate_math1_gen4(inst, dst, src[0]);
731       }
732       break;
733 
734    case SHADER_OPCODE_POW:
735    case SHADER_OPCODE_INT_QUOTIENT:
736    case SHADER_OPCODE_INT_REMAINDER:
737       if (intel->gen >= 7) {
738 	 generate_math2_gen7(inst, dst, src[0], src[1]);
739       } else if (intel->gen == 6) {
740 	 generate_math2_gen6(inst, dst, src[0], src[1]);
741       } else {
742 	 generate_math2_gen4(inst, dst, src[0], src[1]);
743       }
744       break;
745 
746    case SHADER_OPCODE_TEX:
747    case SHADER_OPCODE_TXD:
748    case SHADER_OPCODE_TXF:
749    case SHADER_OPCODE_TXL:
750    case SHADER_OPCODE_TXS:
751       generate_tex(inst, dst, src[0]);
752       break;
753 
754    case VS_OPCODE_URB_WRITE:
755       generate_urb_write(inst);
756       break;
757 
758    case VS_OPCODE_SCRATCH_READ:
759       generate_scratch_read(inst, dst, src[0]);
760       break;
761 
762    case VS_OPCODE_SCRATCH_WRITE:
763       generate_scratch_write(inst, dst, src[0], src[1]);
764       break;
765 
766    case VS_OPCODE_PULL_CONSTANT_LOAD:
767       generate_pull_constant_load(inst, dst, src[0], src[1]);
768       break;
769 
770    default:
771       if (inst->opcode < (int)ARRAY_SIZE(brw_opcodes)) {
772 	 fail("unsupported opcode in `%s' in VS\n",
773 	      brw_opcodes[inst->opcode].name);
774       } else {
775 	 fail("Unsupported opcode %d in VS", inst->opcode);
776       }
777    }
778 }
779 
780 bool
run()781 vec4_visitor::run()
782 {
783    if (c->key.userclip_active && !c->key.uses_clip_distance)
784       setup_uniform_clipplane_values();
785 
786    /* Generate VS IR for main().  (the visitor only descends into
787     * functions called "main").
788     */
789    visit_instructions(shader->ir);
790 
791    emit_urb_writes();
792 
793    /* Before any optimization, push array accesses out to scratch
794     * space where we need them to be.  This pass may allocate new
795     * virtual GRFs, so we want to do it early.  It also makes sure
796     * that we have reladdr computations available for CSE, since we'll
797     * often do repeated subexpressions for those.
798     */
799    move_grf_array_access_to_scratch();
800    move_uniform_array_access_to_pull_constants();
801    pack_uniform_registers();
802    move_push_constants_to_pull_constants();
803 
804    bool progress;
805    do {
806       progress = false;
807       progress = dead_code_eliminate() || progress;
808       progress = opt_copy_propagation() || progress;
809       progress = opt_algebraic() || progress;
810       progress = opt_compute_to_mrf() || progress;
811    } while (progress);
812 
813 
814    if (failed)
815       return false;
816 
817    setup_payload();
818 
819    if (false) {
820       /* Debug of register spilling: Go spill everything. */
821       const int grf_count = virtual_grf_count;
822       float spill_costs[virtual_grf_count];
823       bool no_spill[virtual_grf_count];
824       evaluate_spill_costs(spill_costs, no_spill);
825       for (int i = 0; i < grf_count; i++) {
826          if (no_spill[i])
827             continue;
828          spill_reg(i);
829       }
830    }
831 
832    while (!reg_allocate()) {
833       if (failed)
834          break;
835    }
836 
837    if (failed)
838       return false;
839 
840    brw_set_access_mode(p, BRW_ALIGN_16);
841 
842    generate_code();
843 
844    return !failed;
845 }
846 
847 void
generate_code()848 vec4_visitor::generate_code()
849 {
850    int last_native_inst = 0;
851    const char *last_annotation_string = NULL;
852    ir_instruction *last_annotation_ir = NULL;
853 
854    if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
855       printf("Native code for vertex shader %d:\n", prog->Name);
856    }
857 
858    foreach_list(node, &this->instructions) {
859       vec4_instruction *inst = (vec4_instruction *)node;
860       struct brw_reg src[3], dst;
861 
862       if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
863 	 if (last_annotation_ir != inst->ir) {
864 	    last_annotation_ir = inst->ir;
865 	    if (last_annotation_ir) {
866 	       printf("   ");
867 	       last_annotation_ir->print();
868 	       printf("\n");
869 	    }
870 	 }
871 	 if (last_annotation_string != inst->annotation) {
872 	    last_annotation_string = inst->annotation;
873 	    if (last_annotation_string)
874 	       printf("   %s\n", last_annotation_string);
875 	 }
876       }
877 
878       for (unsigned int i = 0; i < 3; i++) {
879 	 src[i] = inst->get_src(i);
880       }
881       dst = inst->get_dst();
882 
883       brw_set_conditionalmod(p, inst->conditional_mod);
884       brw_set_predicate_control(p, inst->predicate);
885       brw_set_predicate_inverse(p, inst->predicate_inverse);
886       brw_set_saturate(p, inst->saturate);
887 
888       switch (inst->opcode) {
889       case BRW_OPCODE_MOV:
890 	 brw_MOV(p, dst, src[0]);
891 	 break;
892       case BRW_OPCODE_ADD:
893 	 brw_ADD(p, dst, src[0], src[1]);
894 	 break;
895       case BRW_OPCODE_MUL:
896 	 brw_MUL(p, dst, src[0], src[1]);
897 	 break;
898       case BRW_OPCODE_MACH:
899 	 brw_set_acc_write_control(p, 1);
900 	 brw_MACH(p, dst, src[0], src[1]);
901 	 brw_set_acc_write_control(p, 0);
902 	 break;
903 
904       case BRW_OPCODE_FRC:
905 	 brw_FRC(p, dst, src[0]);
906 	 break;
907       case BRW_OPCODE_RNDD:
908 	 brw_RNDD(p, dst, src[0]);
909 	 break;
910       case BRW_OPCODE_RNDE:
911 	 brw_RNDE(p, dst, src[0]);
912 	 break;
913       case BRW_OPCODE_RNDZ:
914 	 brw_RNDZ(p, dst, src[0]);
915 	 break;
916 
917       case BRW_OPCODE_AND:
918 	 brw_AND(p, dst, src[0], src[1]);
919 	 break;
920       case BRW_OPCODE_OR:
921 	 brw_OR(p, dst, src[0], src[1]);
922 	 break;
923       case BRW_OPCODE_XOR:
924 	 brw_XOR(p, dst, src[0], src[1]);
925 	 break;
926       case BRW_OPCODE_NOT:
927 	 brw_NOT(p, dst, src[0]);
928 	 break;
929       case BRW_OPCODE_ASR:
930 	 brw_ASR(p, dst, src[0], src[1]);
931 	 break;
932       case BRW_OPCODE_SHR:
933 	 brw_SHR(p, dst, src[0], src[1]);
934 	 break;
935       case BRW_OPCODE_SHL:
936 	 brw_SHL(p, dst, src[0], src[1]);
937 	 break;
938 
939       case BRW_OPCODE_CMP:
940 	 brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
941 	 break;
942       case BRW_OPCODE_SEL:
943 	 brw_SEL(p, dst, src[0], src[1]);
944 	 break;
945 
946       case BRW_OPCODE_DP4:
947 	 brw_DP4(p, dst, src[0], src[1]);
948 	 break;
949 
950       case BRW_OPCODE_DP3:
951 	 brw_DP3(p, dst, src[0], src[1]);
952 	 break;
953 
954       case BRW_OPCODE_DP2:
955 	 brw_DP2(p, dst, src[0], src[1]);
956 	 break;
957 
958       case BRW_OPCODE_IF:
959 	 if (inst->src[0].file != BAD_FILE) {
960 	    /* The instruction has an embedded compare (only allowed on gen6) */
961 	    assert(intel->gen == 6);
962 	    gen6_IF(p, inst->conditional_mod, src[0], src[1]);
963 	 } else {
964 	    struct brw_instruction *brw_inst = brw_IF(p, BRW_EXECUTE_8);
965 	    brw_inst->header.predicate_control = inst->predicate;
966 	 }
967 	 break;
968 
969       case BRW_OPCODE_ELSE:
970 	 brw_ELSE(p);
971 	 break;
972       case BRW_OPCODE_ENDIF:
973 	 brw_ENDIF(p);
974 	 break;
975 
976       case BRW_OPCODE_DO:
977 	 brw_DO(p, BRW_EXECUTE_8);
978 	 break;
979 
980       case BRW_OPCODE_BREAK:
981 	 brw_BREAK(p);
982 	 brw_set_predicate_control(p, BRW_PREDICATE_NONE);
983 	 break;
984       case BRW_OPCODE_CONTINUE:
985 	 /* FINISHME: We need to write the loop instruction support still. */
986 	 if (intel->gen >= 6)
987 	    gen6_CONT(p);
988 	 else
989 	    brw_CONT(p);
990 	 brw_set_predicate_control(p, BRW_PREDICATE_NONE);
991 	 break;
992 
993       case BRW_OPCODE_WHILE:
994 	 brw_WHILE(p);
995 	 break;
996 
997       default:
998 	 generate_vs_instruction(inst, dst, src);
999 	 break;
1000       }
1001 
1002       if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
1003 	 for (unsigned int i = last_native_inst; i < p->nr_insn; i++) {
1004 	    if (0) {
1005 	       printf("0x%08x 0x%08x 0x%08x 0x%08x ",
1006 		      ((uint32_t *)&p->store[i])[3],
1007 		      ((uint32_t *)&p->store[i])[2],
1008 		      ((uint32_t *)&p->store[i])[1],
1009 		      ((uint32_t *)&p->store[i])[0]);
1010 	    }
1011 	    brw_disasm(stdout, &p->store[i], intel->gen);
1012 	 }
1013       }
1014 
1015       last_native_inst = p->nr_insn;
1016    }
1017 
1018    if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
1019       printf("\n");
1020    }
1021 
1022    brw_set_uip_jip(p);
1023 
1024    /* OK, while the INTEL_DEBUG=vs above is very nice for debugging VS
1025     * emit issues, it doesn't get the jump distances into the output,
1026     * which is often something we want to debug.  So this is here in
1027     * case you're doing that.
1028     */
1029    if (0) {
1030       if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
1031 	 for (unsigned int i = 0; i < p->nr_insn; i++) {
1032 	    printf("0x%08x 0x%08x 0x%08x 0x%08x ",
1033 		   ((uint32_t *)&p->store[i])[3],
1034 		   ((uint32_t *)&p->store[i])[2],
1035 		   ((uint32_t *)&p->store[i])[1],
1036 		   ((uint32_t *)&p->store[i])[0]);
1037 	    brw_disasm(stdout, &p->store[i], intel->gen);
1038 	 }
1039       }
1040    }
1041 }
1042 
1043 extern "C" {
1044 
1045 bool
brw_vs_emit(struct gl_shader_program * prog,struct brw_vs_compile * c)1046 brw_vs_emit(struct gl_shader_program *prog, struct brw_vs_compile *c)
1047 {
1048    struct brw_context *brw = c->func.brw;
1049    struct intel_context *intel = &c->func.brw->intel;
1050    bool start_busy = false;
1051    float start_time = 0;
1052 
1053    if (!prog)
1054       return false;
1055 
1056    if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
1057       start_busy = (intel->batch.last_bo &&
1058                     drm_intel_bo_busy(intel->batch.last_bo));
1059       start_time = get_time();
1060    }
1061 
1062    struct brw_shader *shader =
1063      (brw_shader *) prog->_LinkedShaders[MESA_SHADER_VERTEX];
1064    if (!shader)
1065       return false;
1066 
1067    if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
1068       printf("GLSL IR for native vertex shader %d:\n", prog->Name);
1069       _mesa_print_ir(shader->ir, NULL);
1070       printf("\n\n");
1071    }
1072 
1073    if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
1074       if (shader->compiled_once) {
1075          brw_vs_debug_recompile(brw, prog, &c->key);
1076       }
1077       if (start_busy && !drm_intel_bo_busy(intel->batch.last_bo)) {
1078          perf_debug("VS compile took %.03f ms and stalled the GPU\n",
1079                     (get_time() - start_time) * 1000);
1080       }
1081    }
1082 
1083    vec4_visitor v(c, prog, shader);
1084    if (!v.run()) {
1085       prog->LinkStatus = false;
1086       ralloc_strcat(&prog->InfoLog, v.fail_msg);
1087       return false;
1088    }
1089 
1090    shader->compiled_once = true;
1091 
1092    return true;
1093 }
1094 
1095 } /* extern "C" */
1096 
1097 } /* namespace brw */
1098