projects / mesa.git / blob
? search:


8cd1087744b74e8f8856448c944a2ce348f5ce1b
[mesa.git] / src / gallium / drivers / ilo / shader / toy_tgsi.c
1 /*
2 * Mesa 3-D graphics library
3 *
4 * Copyright (C) 2012-2013 LunarG, Inc.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included
14 * in all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22 * DEALINGS IN THE SOFTWARE.
23 *
24 * Authors:
25 * Chia-I Wu <olv@lunarg.com>
26 */
27
28 #include "tgsi/tgsi_parse.h"
29 #include "tgsi/tgsi_info.h"
30 #include "tgsi/tgsi_strings.h"
31 #include "util/u_hash_table.h"
32 #include "toy_helpers.h"
33 #include "toy_tgsi.h"
34
35 /* map TGSI opcode to GEN opcode 1-to-1 */
36 static const struct {
37 int opcode;
38 int num_dst;
39 int num_src;
40 } aos_simple_opcode_map[TGSI_OPCODE_LAST] = {
41 [TGSI_OPCODE_ARL] = { GEN6_OPCODE_RNDD, 1, 1 },
42 [TGSI_OPCODE_MOV] = { GEN6_OPCODE_MOV, 1, 1 },
43 [TGSI_OPCODE_RCP] = { TOY_OPCODE_INV, 1, 1 },
44 [TGSI_OPCODE_RSQ] = { TOY_OPCODE_RSQ, 1, 1 },
45 [TGSI_OPCODE_MUL] = { GEN6_OPCODE_MUL, 1, 2 },
46 [TGSI_OPCODE_ADD] = { GEN6_OPCODE_ADD, 1, 2 },
47 [TGSI_OPCODE_DP3] = { GEN6_OPCODE_DP3, 1, 2 },
48 [TGSI_OPCODE_DP4] = { GEN6_OPCODE_DP4, 1, 2 },
49 [TGSI_OPCODE_MIN] = { GEN6_OPCODE_SEL, 1, 2 },
50 [TGSI_OPCODE_MAX] = { GEN6_OPCODE_SEL, 1, 2 },
51 /* a later pass will move src[2] to accumulator */
52 [TGSI_OPCODE_MAD] = { GEN6_OPCODE_MAC, 1, 3 },
53 [TGSI_OPCODE_SUB] = { GEN6_OPCODE_ADD, 1, 2 },
54 [TGSI_OPCODE_SQRT] = { TOY_OPCODE_SQRT, 1, 1 },
55 [TGSI_OPCODE_FRC] = { GEN6_OPCODE_FRC, 1, 1 },
56 [TGSI_OPCODE_FLR] = { GEN6_OPCODE_RNDD, 1, 1 },
57 [TGSI_OPCODE_ROUND] = { GEN6_OPCODE_RNDE, 1, 1 },
58 [TGSI_OPCODE_EX2] = { TOY_OPCODE_EXP, 1, 1 },
59 [TGSI_OPCODE_LG2] = { TOY_OPCODE_LOG, 1, 1 },
60 [TGSI_OPCODE_POW] = { TOY_OPCODE_POW, 1, 2 },
61 [TGSI_OPCODE_ABS] = { GEN6_OPCODE_MOV, 1, 1 },
62 [TGSI_OPCODE_DPH] = { GEN6_OPCODE_DPH, 1, 2 },
63 [TGSI_OPCODE_COS] = { TOY_OPCODE_COS, 1, 1 },
64 [TGSI_OPCODE_KILL] = { TOY_OPCODE_KIL, 0, 0 },
65 [TGSI_OPCODE_SIN] = { TOY_OPCODE_SIN, 1, 1 },
66 [TGSI_OPCODE_ARR] = { GEN6_OPCODE_RNDZ, 1, 1 },
67 [TGSI_OPCODE_DP2] = { GEN6_OPCODE_DP2, 1, 2 },
68 [TGSI_OPCODE_IF] = { GEN6_OPCODE_IF, 0, 1 },
69 [TGSI_OPCODE_UIF] = { GEN6_OPCODE_IF, 0, 1 },
70 [TGSI_OPCODE_ELSE] = { GEN6_OPCODE_ELSE, 0, 0 },
71 [TGSI_OPCODE_ENDIF] = { GEN6_OPCODE_ENDIF, 0, 0 },
72 [TGSI_OPCODE_I2F] = { GEN6_OPCODE_MOV, 1, 1 },
73 [TGSI_OPCODE_NOT] = { GEN6_OPCODE_NOT, 1, 1 },
74 [TGSI_OPCODE_TRUNC] = { GEN6_OPCODE_RNDZ, 1, 1 },
75 [TGSI_OPCODE_SHL] = { GEN6_OPCODE_SHL, 1, 2 },
76 [TGSI_OPCODE_AND] = { GEN6_OPCODE_AND, 1, 2 },
77 [TGSI_OPCODE_OR] = { GEN6_OPCODE_OR, 1, 2 },
78 [TGSI_OPCODE_MOD] = { TOY_OPCODE_INT_DIV_REMAINDER, 1, 2 },
79 [TGSI_OPCODE_XOR] = { GEN6_OPCODE_XOR, 1, 2 },
80 [TGSI_OPCODE_EMIT] = { TOY_OPCODE_EMIT, 0, 0 },
81 [TGSI_OPCODE_ENDPRIM] = { TOY_OPCODE_ENDPRIM, 0, 0 },
82 [TGSI_OPCODE_NOP] = { GEN6_OPCODE_NOP, 0, 0 },
83 [TGSI_OPCODE_KILL_IF] = { TOY_OPCODE_KIL, 0, 1 },
84 [TGSI_OPCODE_END] = { GEN6_OPCODE_NOP, 0, 0 },
85 [TGSI_OPCODE_F2I] = { GEN6_OPCODE_MOV, 1, 1 },
86 [TGSI_OPCODE_IDIV] = { TOY_OPCODE_INT_DIV_QUOTIENT, 1, 2 },
87 [TGSI_OPCODE_IMAX] = { GEN6_OPCODE_SEL, 1, 2 },
88 [TGSI_OPCODE_IMIN] = { GEN6_OPCODE_SEL, 1, 2 },
89 [TGSI_OPCODE_INEG] = { GEN6_OPCODE_MOV, 1, 1 },
90 [TGSI_OPCODE_ISHR] = { GEN6_OPCODE_ASR, 1, 2 },
91 [TGSI_OPCODE_F2U] = { GEN6_OPCODE_MOV, 1, 1 },
92 [TGSI_OPCODE_U2F] = { GEN6_OPCODE_MOV, 1, 1 },
93 [TGSI_OPCODE_UADD] = { GEN6_OPCODE_ADD, 1, 2 },
94 [TGSI_OPCODE_UDIV] = { TOY_OPCODE_INT_DIV_QUOTIENT, 1, 2 },
95 /* a later pass will move src[2] to accumulator */
96 [TGSI_OPCODE_UMAD] = { GEN6_OPCODE_MAC, 1, 3 },
97 [TGSI_OPCODE_UMAX] = { GEN6_OPCODE_SEL, 1, 2 },
98 [TGSI_OPCODE_UMIN] = { GEN6_OPCODE_SEL, 1, 2 },
99 [TGSI_OPCODE_UMOD] = { TOY_OPCODE_INT_DIV_REMAINDER, 1, 2 },
100 [TGSI_OPCODE_UMUL] = { GEN6_OPCODE_MUL, 1, 2 },
101 [TGSI_OPCODE_USHR] = { GEN6_OPCODE_SHR, 1, 2 },
102 [TGSI_OPCODE_UARL] = { GEN6_OPCODE_MOV, 1, 1 },
103 [TGSI_OPCODE_IABS] = { GEN6_OPCODE_MOV, 1, 1 },
104 };
105
106 static void
107 aos_simple(struct toy_compiler *tc,
108 const struct tgsi_full_instruction *tgsi_inst,
109 struct toy_dst *dst,
110 struct toy_src *src)
111 {
112 struct toy_inst *inst;
113 int opcode;
114 int cond_modifier = GEN6_COND_NONE;
115 int num_dst = tgsi_inst->Instruction.NumDstRegs;
116 int num_src = tgsi_inst->Instruction.NumSrcRegs;
117 int i;
118
119 opcode = aos_simple_opcode_map[tgsi_inst->Instruction.Opcode].opcode;
120 assert(num_dst == aos_simple_opcode_map[tgsi_inst->Instruction.Opcode].num_dst);
121 assert(num_src == aos_simple_opcode_map[tgsi_inst->Instruction.Opcode].num_src);
122 if (!opcode) {
123 assert(!"invalid aos_simple() call");
124 return;
125 }
126
127 /* no need to emit nop */
128 if (opcode == GEN6_OPCODE_NOP)
129 return;
130
131 inst = tc_add(tc);
132 if (!inst)
133 return;
134
135 inst->opcode = opcode;
136
137 switch (tgsi_inst->Instruction.Opcode) {
138 case TGSI_OPCODE_MIN:
139 case TGSI_OPCODE_IMIN:
140 case TGSI_OPCODE_UMIN:
141 cond_modifier = GEN6_COND_L;
142 break;
143 case TGSI_OPCODE_MAX:
144 case TGSI_OPCODE_IMAX:
145 case TGSI_OPCODE_UMAX:
146 cond_modifier = GEN6_COND_GE;
147 break;
148 case TGSI_OPCODE_SUB:
149 src[1] = tsrc_negate(src[1]);
150 break;
151 case TGSI_OPCODE_ABS:
152 case TGSI_OPCODE_IABS:
153 src[0] = tsrc_absolute(src[0]);
154 break;
155 case TGSI_OPCODE_IF:
156 cond_modifier = GEN6_COND_NZ;
157 num_src = 2;
158 assert(src[0].type == TOY_TYPE_F);
159 src[0] = tsrc_swizzle1(src[0], TOY_SWIZZLE_X);
160 src[1] = tsrc_imm_f(0.0f);
161 break;
162 case TGSI_OPCODE_UIF:
163 cond_modifier = GEN6_COND_NZ;
164 num_src = 2;
165 assert(src[0].type == TOY_TYPE_UD);
166 src[0] = tsrc_swizzle1(src[0], TOY_SWIZZLE_X);
167 src[1] = tsrc_imm_d(0);
168 break;
169 case TGSI_OPCODE_INEG:
170 src[0] = tsrc_negate(src[0]);
171 break;
172 case TGSI_OPCODE_RCP:
173 case TGSI_OPCODE_RSQ:
174 case TGSI_OPCODE_EX2:
175 case TGSI_OPCODE_LG2:
176 case TGSI_OPCODE_COS:
177 case TGSI_OPCODE_SIN:
178 src[0] = tsrc_swizzle1(src[0], TOY_SWIZZLE_X);
179 break;
180 case TGSI_OPCODE_POW:
181 src[0] = tsrc_swizzle1(src[0], TOY_SWIZZLE_X);
182 src[1] = tsrc_swizzle1(src[1], TOY_SWIZZLE_X);
183 break;
184 }
185
186 inst->cond_modifier = cond_modifier;
187
188 if (num_dst) {
189 assert(num_dst == 1);
190 inst->dst = dst[0];
191 }
192
193 assert(num_src <= Elements(inst->src));
194 for (i = 0; i < num_src; i++)
195 inst->src[i] = src[i];
196 }
197
198 static void
199 aos_set_on_cond(struct toy_compiler *tc,
200 const struct tgsi_full_instruction *tgsi_inst,
201 struct toy_dst *dst,
202 struct toy_src *src)
203 {
204 struct toy_inst *inst;
205 int cond;
206 struct toy_src zero, one;
207
208 switch (tgsi_inst->Instruction.Opcode) {
209 case TGSI_OPCODE_SLT:
210 case TGSI_OPCODE_ISLT:
211 case TGSI_OPCODE_USLT:
212 case TGSI_OPCODE_FSLT:
213 cond = GEN6_COND_L;
214 break;
215 case TGSI_OPCODE_SGE:
216 case TGSI_OPCODE_ISGE:
217 case TGSI_OPCODE_USGE:
218 case TGSI_OPCODE_FSGE:
219 cond = GEN6_COND_GE;
220 break;
221 case TGSI_OPCODE_SEQ:
222 case TGSI_OPCODE_USEQ:
223 case TGSI_OPCODE_FSEQ:
224 cond = GEN6_COND_Z;
225 break;
226 case TGSI_OPCODE_SGT:
227 cond = GEN6_COND_G;
228 break;
229 case TGSI_OPCODE_SLE:
230 cond = GEN6_COND_LE;
231 break;
232 case TGSI_OPCODE_SNE:
233 case TGSI_OPCODE_USNE:
234 case TGSI_OPCODE_FSNE:
235 cond = GEN6_COND_NZ;
236 break;
237 default:
238 assert(!"invalid aos_set_on_cond() call");
239 return;
240 }
241
242 /* note that for integer versions, all bits are set */
243 switch (dst[0].type) {
244 case TOY_TYPE_F:
245 default:
246 zero = tsrc_imm_f(0.0f);
247 one = tsrc_imm_f(1.0f);
248 break;
249 case TOY_TYPE_D:
250 zero = tsrc_imm_d(0);
251 one = tsrc_imm_d(-1);
252 break;
253 case TOY_TYPE_UD:
254 zero = tsrc_imm_ud(0);
255 one = tsrc_imm_ud(~0);
256 break;
257 }
258
259 tc_MOV(tc, dst[0], zero);
260 tc_CMP(tc, tdst_null(), src[0], src[1], cond);
261 inst = tc_MOV(tc, dst[0], one);
262 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
263 }
264
265 static void
266 aos_compare(struct toy_compiler *tc,
267 const struct tgsi_full_instruction *tgsi_inst,
268 struct toy_dst *dst,
269 struct toy_src *src)
270 {
271 struct toy_inst *inst;
272 struct toy_src zero;
273
274 switch (tgsi_inst->Instruction.Opcode) {
275 case TGSI_OPCODE_CMP:
276 zero = tsrc_imm_f(0.0f);
277 break;
278 case TGSI_OPCODE_UCMP:
279 zero = tsrc_imm_ud(0);
280 break;
281 default:
282 assert(!"invalid aos_compare() call");
283 return;
284 }
285
286 tc_CMP(tc, tdst_null(), src[0], zero, GEN6_COND_L);
287 inst = tc_SEL(tc, dst[0], src[1], src[2], GEN6_COND_NONE);
288 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
289 }
290
291 static void
292 aos_set_sign(struct toy_compiler *tc,
293 const struct tgsi_full_instruction *tgsi_inst,
294 struct toy_dst *dst,
295 struct toy_src *src)
296 {
297 struct toy_inst *inst;
298 struct toy_src zero, one, neg_one;
299
300 switch (tgsi_inst->Instruction.Opcode) {
301 case TGSI_OPCODE_SSG:
302 zero = tsrc_imm_f(0.0f);
303 one = tsrc_imm_f(1.0f);
304 neg_one = tsrc_imm_f(-1.0f);
305 break;
306 case TGSI_OPCODE_ISSG:
307 zero = tsrc_imm_d(0);
308 one = tsrc_imm_d(1);
309 neg_one = tsrc_imm_d(-1);
310 break;
311 default:
312 assert(!"invalid aos_set_sign() call");
313 return;
314 }
315
316 tc_MOV(tc, dst[0], zero);
317
318 tc_CMP(tc, tdst_null(), src[0], zero, GEN6_COND_G);
319 inst = tc_MOV(tc, dst[0], one);
320 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
321
322 tc_CMP(tc, tdst_null(), src[0], zero, GEN6_COND_L);
323 inst = tc_MOV(tc, dst[0], neg_one);
324 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
325 }
326
327 static void
328 aos_tex(struct toy_compiler *tc,
329 const struct tgsi_full_instruction *tgsi_inst,
330 struct toy_dst *dst,
331 struct toy_src *src)
332 {
333 struct toy_inst *inst;
334 enum toy_opcode opcode;
335 int i;
336
337 switch (tgsi_inst->Instruction.Opcode) {
338 case TGSI_OPCODE_TEX:
339 opcode = TOY_OPCODE_TGSI_TEX;
340 break;
341 case TGSI_OPCODE_TXD:
342 opcode = TOY_OPCODE_TGSI_TXD;
343 break;
344 case TGSI_OPCODE_TXP:
345 opcode = TOY_OPCODE_TGSI_TXP;
346 break;
347 case TGSI_OPCODE_TXB:
348 opcode = TOY_OPCODE_TGSI_TXB;
349 break;
350 case TGSI_OPCODE_TXL:
351 opcode = TOY_OPCODE_TGSI_TXL;
352 break;
353 case TGSI_OPCODE_TXF:
354 opcode = TOY_OPCODE_TGSI_TXF;
355 break;
356 case TGSI_OPCODE_TXQ:
357 opcode = TOY_OPCODE_TGSI_TXQ;
358 break;
359 case TGSI_OPCODE_TXQ_LZ:
360 opcode = TOY_OPCODE_TGSI_TXQ_LZ;
361 break;
362 case TGSI_OPCODE_TEX2:
363 opcode = TOY_OPCODE_TGSI_TEX2;
364 break;
365 case TGSI_OPCODE_TXB2:
366 opcode = TOY_OPCODE_TGSI_TXB2;
367 break;
368 case TGSI_OPCODE_TXL2:
369 opcode = TOY_OPCODE_TGSI_TXL2;
370 break;
371 default:
372 assert(!"unsupported texturing opcode");
373 return;
374 break;
375 }
376
377 assert(tgsi_inst->Instruction.Texture);
378
379 inst = tc_add(tc);
380 inst->opcode = opcode;
381 inst->tex.target = tgsi_inst->Texture.Texture;
382
383 assert(tgsi_inst->Instruction.NumSrcRegs <= Elements(inst->src));
384 assert(tgsi_inst->Instruction.NumDstRegs == 1);
385
386 inst->dst = dst[0];
387 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
388 inst->src[i] = src[i];
389
390 for (i = 0; i < tgsi_inst->Texture.NumOffsets; i++)
391 tc_fail(tc, "texelFetchOffset unsupported");
392 }
393
394 static void
395 aos_sample(struct toy_compiler *tc,
396 const struct tgsi_full_instruction *tgsi_inst,
397 struct toy_dst *dst,
398 struct toy_src *src)
399 {
400 struct toy_inst *inst;
401 enum toy_opcode opcode;
402 int i;
403
404 assert(!"sampling untested");
405
406 switch (tgsi_inst->Instruction.Opcode) {
407 case TGSI_OPCODE_SAMPLE:
408 opcode = TOY_OPCODE_TGSI_SAMPLE;
409 break;
410 case TGSI_OPCODE_SAMPLE_I:
411 opcode = TOY_OPCODE_TGSI_SAMPLE_I;
412 break;
413 case TGSI_OPCODE_SAMPLE_I_MS:
414 opcode = TOY_OPCODE_TGSI_SAMPLE_I_MS;
415 break;
416 case TGSI_OPCODE_SAMPLE_B:
417 opcode = TOY_OPCODE_TGSI_SAMPLE_B;
418 break;
419 case TGSI_OPCODE_SAMPLE_C:
420 opcode = TOY_OPCODE_TGSI_SAMPLE_C;
421 break;
422 case TGSI_OPCODE_SAMPLE_C_LZ:
423 opcode = TOY_OPCODE_TGSI_SAMPLE_C_LZ;
424 break;
425 case TGSI_OPCODE_SAMPLE_D:
426 opcode = TOY_OPCODE_TGSI_SAMPLE_D;
427 break;
428 case TGSI_OPCODE_SAMPLE_L:
429 opcode = TOY_OPCODE_TGSI_SAMPLE_L;
430 break;
431 case TGSI_OPCODE_GATHER4:
432 opcode = TOY_OPCODE_TGSI_GATHER4;
433 break;
434 case TGSI_OPCODE_SVIEWINFO:
435 opcode = TOY_OPCODE_TGSI_SVIEWINFO;
436 break;
437 case TGSI_OPCODE_SAMPLE_POS:
438 opcode = TOY_OPCODE_TGSI_SAMPLE_POS;
439 break;
440 case TGSI_OPCODE_SAMPLE_INFO:
441 opcode = TOY_OPCODE_TGSI_SAMPLE_INFO;
442 break;
443 default:
444 assert(!"unsupported sampling opcode");
445 return;
446 break;
447 }
448
449 inst = tc_add(tc);
450 inst->opcode = opcode;
451
452 assert(tgsi_inst->Instruction.NumSrcRegs <= Elements(inst->src));
453 assert(tgsi_inst->Instruction.NumDstRegs == 1);
454
455 inst->dst = dst[0];
456 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
457 inst->src[i] = src[i];
458 }
459
460 static void
461 aos_LIT(struct toy_compiler *tc,
462 const struct tgsi_full_instruction *tgsi_inst,
463 struct toy_dst *dst,
464 struct toy_src *src)
465 {
466 struct toy_inst *inst;
467
468 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_XW), tsrc_imm_f(1.0f));
469
470 if (!(dst[0].writemask & TOY_WRITEMASK_YZ))
471 return;
472
473 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_YZ), tsrc_imm_f(0.0f));
474
475 tc_CMP(tc, tdst_null(),
476 tsrc_swizzle1(src[0], TOY_SWIZZLE_X),
477 tsrc_imm_f(0.0f),
478 GEN6_COND_G);
479
480 inst = tc_MOV(tc,
481 tdst_writemask(dst[0], TOY_WRITEMASK_Y),
482 tsrc_swizzle1(src[0], TOY_SWIZZLE_X));
483 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
484
485 /* clamp W to (-128, 128)? */
486 inst = tc_POW(tc,
487 tdst_writemask(dst[0], TOY_WRITEMASK_Z),
488 tsrc_swizzle1(src[0], TOY_SWIZZLE_Y),
489 tsrc_swizzle1(src[0], TOY_SWIZZLE_W));
490 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
491 }
492
493 static void
494 aos_EXP(struct toy_compiler *tc,
495 const struct tgsi_full_instruction *tgsi_inst,
496 struct toy_dst *dst,
497 struct toy_src *src)
498 {
499 struct toy_src src0 = tsrc_swizzle1(src[0], TOY_SWIZZLE_X);
500
501 if (dst[0].writemask & TOY_WRITEMASK_X) {
502 struct toy_dst tmp =
503 tdst_d(tdst_writemask(tc_alloc_tmp(tc), TOY_WRITEMASK_X));
504
505 tc_RNDD(tc, tmp, src0);
506
507 /* construct the floating point number manually */
508 tc_ADD(tc, tmp, tsrc_from(tmp), tsrc_imm_d(127));
509 tc_SHL(tc, tdst_d(tdst_writemask(dst[0], TOY_WRITEMASK_X)),
510 tsrc_from(tmp), tsrc_imm_d(23));
511 }
512
513 tc_FRC(tc, tdst_writemask(dst[0], TOY_WRITEMASK_Y), src0);
514 tc_EXP(tc, tdst_writemask(dst[0], TOY_WRITEMASK_Z), src0);
515 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_W), tsrc_imm_f(1.0f));
516 }
517
518 static void
519 aos_LOG(struct toy_compiler *tc,
520 const struct tgsi_full_instruction *tgsi_inst,
521 struct toy_dst *dst,
522 struct toy_src *src)
523 {
524 struct toy_src src0 = tsrc_swizzle1(src[0], TOY_SWIZZLE_X);
525
526 if (dst[0].writemask & TOY_WRITEMASK_XY) {
527 struct toy_dst tmp;
528
529 tmp = tdst_d(tdst_writemask(tc_alloc_tmp(tc), TOY_WRITEMASK_X));
530
531 /* exponent */
532 tc_SHR(tc, tmp, tsrc_absolute(tsrc_d(src0)), tsrc_imm_d(23));
533 tc_ADD(tc, tdst_writemask(dst[0], TOY_WRITEMASK_X),
534 tsrc_from(tmp), tsrc_imm_d(-127));
535
536 /* mantissa */
537 tc_AND(tc, tmp, tsrc_d(src0), tsrc_imm_d((1 << 23) - 1));
538 tc_OR(tc, tdst_writemask(tdst_d(dst[0]), TOY_WRITEMASK_Y),
539 tsrc_from(tmp), tsrc_imm_d(127 << 23));
540 }
541
542 tc_LOG(tc, tdst_writemask(dst[0], TOY_WRITEMASK_Z), src0);
543 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_W), tsrc_imm_f(1.0f));
544 }
545
546 static void
547 aos_DST(struct toy_compiler *tc,
548 const struct tgsi_full_instruction *tgsi_inst,
549 struct toy_dst *dst,
550 struct toy_src *src)
551 {
552 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_X), tsrc_imm_f(1.0f));
553 tc_MUL(tc, tdst_writemask(dst[0], TOY_WRITEMASK_Y), src[0], src[1]);
554 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_Z), src[0]);
555 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_W), src[1]);
556 }
557
558 static void
559 aos_LRP(struct toy_compiler *tc,
560 const struct tgsi_full_instruction *tgsi_inst,
561 struct toy_dst *dst,
562 struct toy_src *src)
563 {
564 struct toy_dst tmp = tc_alloc_tmp(tc);
565
566 tc_ADD(tc, tmp, tsrc_negate(src[0]), tsrc_imm_f(1.0f));
567 tc_MUL(tc, tmp, tsrc_from(tmp), src[2]);
568 tc_MAC(tc, dst[0], src[0], src[1], tsrc_from(tmp));
569 }
570
571 static void
572 aos_CND(struct toy_compiler *tc,
573 const struct tgsi_full_instruction *tgsi_inst,
574 struct toy_dst *dst,
575 struct toy_src *src)
576 {
577 struct toy_inst *inst;
578
579 assert(!"CND untested");
580
581 tc_CMP(tc, tdst_null(), src[2], tsrc_imm_f(0.5f), GEN6_COND_G);
582 inst = tc_SEL(tc, dst[0], src[0], src[1], GEN6_COND_NONE);
583 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
584 }
585
586 static void
587 aos_DP2A(struct toy_compiler *tc,
588 const struct tgsi_full_instruction *tgsi_inst,
589 struct toy_dst *dst,
590 struct toy_src *src)
591 {
592 struct toy_dst tmp = tc_alloc_tmp(tc);
593
594 assert(!"DP2A untested");
595
596 tc_DP2(tc, tmp, src[0], src[1]);
597 tc_ADD(tc, dst[0], tsrc_swizzle1(tsrc_from(tmp), TOY_SWIZZLE_X), src[2]);
598 }
599
600 static void
601 aos_CLAMP(struct toy_compiler *tc,
602 const struct tgsi_full_instruction *tgsi_inst,
603 struct toy_dst *dst,
604 struct toy_src *src)
605 {
606 assert(!"CLAMP untested");
607
608 tc_SEL(tc, dst[0], src[0], src[1], GEN6_COND_GE);
609 tc_SEL(tc, dst[0], src[2], tsrc_from(dst[0]), GEN6_COND_L);
610 }
611
612 static void
613 aos_XPD(struct toy_compiler *tc,
614 const struct tgsi_full_instruction *tgsi_inst,
615 struct toy_dst *dst,
616 struct toy_src *src)
617 {
618 struct toy_dst tmp = tc_alloc_tmp(tc);
619
620 tc_MUL(tc, tdst_writemask(tmp, TOY_WRITEMASK_XYZ),
621 tsrc_swizzle(src[0], TOY_SWIZZLE_Z, TOY_SWIZZLE_X,
622 TOY_SWIZZLE_Y, TOY_SWIZZLE_W),
623 tsrc_swizzle(src[1], TOY_SWIZZLE_Y, TOY_SWIZZLE_Z,
624 TOY_SWIZZLE_X, TOY_SWIZZLE_W));
625
626 tc_MAC(tc, tdst_writemask(dst[0], TOY_WRITEMASK_XYZ),
627 tsrc_swizzle(src[0], TOY_SWIZZLE_Y, TOY_SWIZZLE_Z,
628 TOY_SWIZZLE_X, TOY_SWIZZLE_W),
629 tsrc_swizzle(src[1], TOY_SWIZZLE_Z, TOY_SWIZZLE_X,
630 TOY_SWIZZLE_Y, TOY_SWIZZLE_W),
631 tsrc_negate(tsrc_from(tmp)));
632
633 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_W),
634 tsrc_imm_f(1.0f));
635 }
636
637 static void
638 aos_PK2H(struct toy_compiler *tc,
639 const struct tgsi_full_instruction *tgsi_inst,
640 struct toy_dst *dst,
641 struct toy_src *src)
642 {
643 const struct toy_src h1 = tsrc_ud(tsrc_swizzle1(src[0], TOY_SWIZZLE_X));
644 const struct toy_src h2 = tsrc_ud(tsrc_swizzle1(src[0], TOY_SWIZZLE_Y));
645 struct toy_dst tmp = tdst_ud(tc_alloc_tmp(tc));
646
647 assert(!"PK2H untested");
648
649 tc_SHL(tc, tmp, h2, tsrc_imm_ud(16));
650 tc_OR(tc, tdst_ud(dst[0]), h1, tsrc_from(tmp));
651 }
652
653 static void
654 aos_SFL(struct toy_compiler *tc,
655 const struct tgsi_full_instruction *tgsi_inst,
656 struct toy_dst *dst,
657 struct toy_src *src)
658 {
659 assert(!"SFL untested");
660
661 tc_MOV(tc, dst[0], tsrc_imm_f(0.0f));
662 }
663
664 static void
665 aos_STR(struct toy_compiler *tc,
666 const struct tgsi_full_instruction *tgsi_inst,
667 struct toy_dst *dst,
668 struct toy_src *src)
669 {
670 assert(!"STR untested");
671
672 tc_MOV(tc, dst[0], tsrc_imm_f(1.0f));
673 }
674
675 static void
676 aos_UP2H(struct toy_compiler *tc,
677 const struct tgsi_full_instruction *tgsi_inst,
678 struct toy_dst *dst,
679 struct toy_src *src)
680 {
681 assert(!"UP2H untested");
682
683 tc_AND(tc, tdst_writemask(tdst_ud(dst[0]), TOY_WRITEMASK_XZ),
684 tsrc_ud(src[0]), tsrc_imm_ud(0xffff));
685 tc_SHR(tc, tdst_writemask(tdst_ud(dst[0]), TOY_WRITEMASK_YW),
686 tsrc_ud(src[0]), tsrc_imm_ud(16));
687 }
688
689 static void
690 aos_SCS(struct toy_compiler *tc,
691 const struct tgsi_full_instruction *tgsi_inst,
692 struct toy_dst *dst,
693 struct toy_src *src)
694 {
695 assert(!"SCS untested");
696
697 tc_add1(tc, TOY_OPCODE_COS,
698 tdst_writemask(dst[0], TOY_WRITEMASK_X), src[0]);
699
700 tc_add1(tc, TOY_OPCODE_SIN,
701 tdst_writemask(dst[0], TOY_WRITEMASK_Y), src[0]);
702
703 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_Z), tsrc_imm_f(0.0f));
704 tc_MOV(tc, tdst_writemask(dst[0], TOY_WRITEMASK_W), tsrc_imm_f(1.0f));
705 }
706
707 static void
708 aos_DIV(struct toy_compiler *tc,
709 const struct tgsi_full_instruction *tgsi_inst,
710 struct toy_dst *dst,
711 struct toy_src *src)
712 {
713 struct toy_dst tmp = tc_alloc_tmp(tc);
714
715 assert(!"DIV untested");
716
717 tc_INV(tc, tmp, src[1]);
718 tc_MUL(tc, dst[0], src[0], tsrc_from(tmp));
719 }
720
721 static void
722 aos_BRK(struct toy_compiler *tc,
723 const struct tgsi_full_instruction *tgsi_inst,
724 struct toy_dst *dst,
725 struct toy_src *src)
726 {
727 tc_add0(tc, GEN6_OPCODE_BREAK);
728 }
729
730 static void
731 aos_CEIL(struct toy_compiler *tc,
732 const struct tgsi_full_instruction *tgsi_inst,
733 struct toy_dst *dst,
734 struct toy_src *src)
735 {
736 struct toy_dst tmp = tc_alloc_tmp(tc);
737
738 tc_RNDD(tc, tmp, tsrc_negate(src[0]));
739 tc_MOV(tc, dst[0], tsrc_negate(tsrc_from(tmp)));
740 }
741
742 static void
743 aos_SAD(struct toy_compiler *tc,
744 const struct tgsi_full_instruction *tgsi_inst,
745 struct toy_dst *dst,
746 struct toy_src *src)
747 {
748 struct toy_dst tmp = tc_alloc_tmp(tc);
749
750 assert(!"SAD untested");
751
752 tc_ADD(tc, tmp, src[0], tsrc_negate(src[1]));
753 tc_ADD(tc, dst[0], tsrc_absolute(tsrc_from(tmp)), src[2]);
754 }
755
756 static void
757 aos_CONT(struct toy_compiler *tc,
758 const struct tgsi_full_instruction *tgsi_inst,
759 struct toy_dst *dst,
760 struct toy_src *src)
761 {
762 tc_add0(tc, GEN6_OPCODE_CONT);
763 }
764
765 static void
766 aos_BGNLOOP(struct toy_compiler *tc,
767 const struct tgsi_full_instruction *tgsi_inst,
768 struct toy_dst *dst,
769 struct toy_src *src)
770 {
771 struct toy_inst *inst;
772
773 inst = tc_add0(tc, TOY_OPCODE_DO);
774 /* this is just a marker */
775 inst->marker = true;
776 }
777
778 static void
779 aos_ENDLOOP(struct toy_compiler *tc,
780 const struct tgsi_full_instruction *tgsi_inst,
781 struct toy_dst *dst,
782 struct toy_src *src)
783 {
784 tc_add0(tc, GEN6_OPCODE_WHILE);
785 }
786
787 static void
788 aos_unsupported(struct toy_compiler *tc,
789 const struct tgsi_full_instruction *tgsi_inst,
790 struct toy_dst *dst,
791 struct toy_src *src)
792 {
793 const char *name = tgsi_get_opcode_name(tgsi_inst->Instruction.Opcode);
794
795 ilo_warn("unsupported TGSI opcode: TGSI_OPCODE_%s\n", name);
796
797 tc_fail(tc, "unsupported TGSI instruction");
798 }
799
800 static const toy_tgsi_translate aos_translate_table[TGSI_OPCODE_LAST] = {
801 [TGSI_OPCODE_ARL] = aos_simple,
802 [TGSI_OPCODE_MOV] = aos_simple,
803 [TGSI_OPCODE_LIT] = aos_LIT,
804 [TGSI_OPCODE_RCP] = aos_simple,
805 [TGSI_OPCODE_RSQ] = aos_simple,
806 [TGSI_OPCODE_EXP] = aos_EXP,
807 [TGSI_OPCODE_LOG] = aos_LOG,
808 [TGSI_OPCODE_MUL] = aos_simple,
809 [TGSI_OPCODE_ADD] = aos_simple,
810 [TGSI_OPCODE_DP3] = aos_simple,
811 [TGSI_OPCODE_DP4] = aos_simple,
812 [TGSI_OPCODE_DST] = aos_DST,
813 [TGSI_OPCODE_MIN] = aos_simple,
814 [TGSI_OPCODE_MAX] = aos_simple,
815 [TGSI_OPCODE_SLT] = aos_set_on_cond,
816 [TGSI_OPCODE_SGE] = aos_set_on_cond,
817 [TGSI_OPCODE_MAD] = aos_simple,
818 [TGSI_OPCODE_SUB] = aos_simple,
819 [TGSI_OPCODE_LRP] = aos_LRP,
820 [TGSI_OPCODE_CND] = aos_CND,
821 [TGSI_OPCODE_SQRT] = aos_simple,
822 [TGSI_OPCODE_DP2A] = aos_DP2A,
823 [TGSI_OPCODE_FRC] = aos_simple,
824 [TGSI_OPCODE_CLAMP] = aos_CLAMP,
825 [TGSI_OPCODE_FLR] = aos_simple,
826 [TGSI_OPCODE_ROUND] = aos_simple,
827 [TGSI_OPCODE_EX2] = aos_simple,
828 [TGSI_OPCODE_LG2] = aos_simple,
829 [TGSI_OPCODE_POW] = aos_simple,
830 [TGSI_OPCODE_XPD] = aos_XPD,
831 [TGSI_OPCODE_ABS] = aos_simple,
832 [TGSI_OPCODE_DPH] = aos_simple,
833 [TGSI_OPCODE_COS] = aos_simple,
834 [TGSI_OPCODE_DDX] = aos_unsupported,
835 [TGSI_OPCODE_DDY] = aos_unsupported,
836 [TGSI_OPCODE_KILL] = aos_simple,
837 [TGSI_OPCODE_PK2H] = aos_PK2H,
838 [TGSI_OPCODE_PK2US] = aos_unsupported,
839 [TGSI_OPCODE_PK4B] = aos_unsupported,
840 [TGSI_OPCODE_PK4UB] = aos_unsupported,
841 [TGSI_OPCODE_SEQ] = aos_set_on_cond,
842 [TGSI_OPCODE_SFL] = aos_SFL,
843 [TGSI_OPCODE_SGT] = aos_set_on_cond,
844 [TGSI_OPCODE_SIN] = aos_simple,
845 [TGSI_OPCODE_SLE] = aos_set_on_cond,
846 [TGSI_OPCODE_SNE] = aos_set_on_cond,
847 [TGSI_OPCODE_STR] = aos_STR,
848 [TGSI_OPCODE_TEX] = aos_tex,
849 [TGSI_OPCODE_TXD] = aos_tex,
850 [TGSI_OPCODE_TXP] = aos_tex,
851 [TGSI_OPCODE_UP2H] = aos_UP2H,
852 [TGSI_OPCODE_UP2US] = aos_unsupported,
853 [TGSI_OPCODE_UP4B] = aos_unsupported,
854 [TGSI_OPCODE_UP4UB] = aos_unsupported,
855 [TGSI_OPCODE_ARR] = aos_simple,
856 [TGSI_OPCODE_BRA] = aos_unsupported,
857 [TGSI_OPCODE_CAL] = aos_unsupported,
858 [TGSI_OPCODE_RET] = aos_unsupported,
859 [TGSI_OPCODE_SSG] = aos_set_sign,
860 [TGSI_OPCODE_CMP] = aos_compare,
861 [TGSI_OPCODE_SCS] = aos_SCS,
862 [TGSI_OPCODE_TXB] = aos_tex,
863 [TGSI_OPCODE_DIV] = aos_DIV,
864 [TGSI_OPCODE_DP2] = aos_simple,
865 [TGSI_OPCODE_TXL] = aos_tex,
866 [TGSI_OPCODE_BRK] = aos_BRK,
867 [TGSI_OPCODE_IF] = aos_simple,
868 [TGSI_OPCODE_UIF] = aos_simple,
869 [TGSI_OPCODE_ELSE] = aos_simple,
870 [TGSI_OPCODE_ENDIF] = aos_simple,
871 [TGSI_OPCODE_PUSHA] = aos_unsupported,
872 [TGSI_OPCODE_POPA] = aos_unsupported,
873 [TGSI_OPCODE_CEIL] = aos_CEIL,
874 [TGSI_OPCODE_I2F] = aos_simple,
875 [TGSI_OPCODE_NOT] = aos_simple,
876 [TGSI_OPCODE_TRUNC] = aos_simple,
877 [TGSI_OPCODE_SHL] = aos_simple,
878 [TGSI_OPCODE_AND] = aos_simple,
879 [TGSI_OPCODE_OR] = aos_simple,
880 [TGSI_OPCODE_MOD] = aos_simple,
881 [TGSI_OPCODE_XOR] = aos_simple,
882 [TGSI_OPCODE_SAD] = aos_SAD,
883 [TGSI_OPCODE_TXF] = aos_tex,
884 [TGSI_OPCODE_TXQ] = aos_tex,
885 [TGSI_OPCODE_CONT] = aos_CONT,
886 [TGSI_OPCODE_EMIT] = aos_simple,
887 [TGSI_OPCODE_ENDPRIM] = aos_simple,
888 [TGSI_OPCODE_BGNLOOP] = aos_BGNLOOP,
889 [TGSI_OPCODE_BGNSUB] = aos_unsupported,
890 [TGSI_OPCODE_ENDLOOP] = aos_ENDLOOP,
891 [TGSI_OPCODE_ENDSUB] = aos_unsupported,
892 [TGSI_OPCODE_TXQ_LZ] = aos_tex,
893 [TGSI_OPCODE_NOP] = aos_simple,
894 [TGSI_OPCODE_FSEQ] = aos_set_on_cond,
895 [TGSI_OPCODE_FSGE] = aos_set_on_cond,
896 [TGSI_OPCODE_FSLT] = aos_set_on_cond,
897 [TGSI_OPCODE_FSNE] = aos_set_on_cond,
898 [TGSI_OPCODE_CALLNZ] = aos_unsupported,
899 [TGSI_OPCODE_BREAKC] = aos_unsupported,
900 [TGSI_OPCODE_KILL_IF] = aos_simple,
901 [TGSI_OPCODE_END] = aos_simple,
902 [TGSI_OPCODE_F2I] = aos_simple,
903 [TGSI_OPCODE_IDIV] = aos_simple,
904 [TGSI_OPCODE_IMAX] = aos_simple,
905 [TGSI_OPCODE_IMIN] = aos_simple,
906 [TGSI_OPCODE_INEG] = aos_simple,
907 [TGSI_OPCODE_ISGE] = aos_set_on_cond,
908 [TGSI_OPCODE_ISHR] = aos_simple,
909 [TGSI_OPCODE_ISLT] = aos_set_on_cond,
910 [TGSI_OPCODE_F2U] = aos_simple,
911 [TGSI_OPCODE_U2F] = aos_simple,
912 [TGSI_OPCODE_UADD] = aos_simple,
913 [TGSI_OPCODE_UDIV] = aos_simple,
914 [TGSI_OPCODE_UMAD] = aos_simple,
915 [TGSI_OPCODE_UMAX] = aos_simple,
916 [TGSI_OPCODE_UMIN] = aos_simple,
917 [TGSI_OPCODE_UMOD] = aos_simple,
918 [TGSI_OPCODE_UMUL] = aos_simple,
919 [TGSI_OPCODE_USEQ] = aos_set_on_cond,
920 [TGSI_OPCODE_USGE] = aos_set_on_cond,
921 [TGSI_OPCODE_USHR] = aos_simple,
922 [TGSI_OPCODE_USLT] = aos_set_on_cond,
923 [TGSI_OPCODE_USNE] = aos_set_on_cond,
924 [TGSI_OPCODE_SWITCH] = aos_unsupported,
925 [TGSI_OPCODE_CASE] = aos_unsupported,
926 [TGSI_OPCODE_DEFAULT] = aos_unsupported,
927 [TGSI_OPCODE_ENDSWITCH] = aos_unsupported,
928 [TGSI_OPCODE_SAMPLE] = aos_sample,
929 [TGSI_OPCODE_SAMPLE_I] = aos_sample,
930 [TGSI_OPCODE_SAMPLE_I_MS] = aos_sample,
931 [TGSI_OPCODE_SAMPLE_B] = aos_sample,
932 [TGSI_OPCODE_SAMPLE_C] = aos_sample,
933 [TGSI_OPCODE_SAMPLE_C_LZ] = aos_sample,
934 [TGSI_OPCODE_SAMPLE_D] = aos_sample,
935 [TGSI_OPCODE_SAMPLE_L] = aos_sample,
936 [TGSI_OPCODE_GATHER4] = aos_sample,
937 [TGSI_OPCODE_SVIEWINFO] = aos_sample,
938 [TGSI_OPCODE_SAMPLE_POS] = aos_sample,
939 [TGSI_OPCODE_SAMPLE_INFO] = aos_sample,
940 [TGSI_OPCODE_UARL] = aos_simple,
941 [TGSI_OPCODE_UCMP] = aos_compare,
942 [TGSI_OPCODE_IABS] = aos_simple,
943 [TGSI_OPCODE_ISSG] = aos_set_sign,
944 [TGSI_OPCODE_LOAD] = aos_unsupported,
945 [TGSI_OPCODE_STORE] = aos_unsupported,
946 [TGSI_OPCODE_MFENCE] = aos_unsupported,
947 [TGSI_OPCODE_LFENCE] = aos_unsupported,
948 [TGSI_OPCODE_SFENCE] = aos_unsupported,
949 [TGSI_OPCODE_BARRIER] = aos_unsupported,
950 [TGSI_OPCODE_ATOMUADD] = aos_unsupported,
951 [TGSI_OPCODE_ATOMXCHG] = aos_unsupported,
952 [TGSI_OPCODE_ATOMCAS] = aos_unsupported,
953 [TGSI_OPCODE_ATOMAND] = aos_unsupported,
954 [TGSI_OPCODE_ATOMOR] = aos_unsupported,
955 [TGSI_OPCODE_ATOMXOR] = aos_unsupported,
956 [TGSI_OPCODE_ATOMUMIN] = aos_unsupported,
957 [TGSI_OPCODE_ATOMUMAX] = aos_unsupported,
958 [TGSI_OPCODE_ATOMIMIN] = aos_unsupported,
959 [TGSI_OPCODE_ATOMIMAX] = aos_unsupported,
960 [TGSI_OPCODE_TEX2] = aos_tex,
961 [TGSI_OPCODE_TXB2] = aos_tex,
962 [TGSI_OPCODE_TXL2] = aos_tex,
963 };
964
965 static void
966 soa_passthrough(struct toy_compiler *tc,
967 const struct tgsi_full_instruction *tgsi_inst,
968 struct toy_dst *dst_,
969 struct toy_src *src_)
970 {
971 const toy_tgsi_translate translate =
972 aos_translate_table[tgsi_inst->Instruction.Opcode];
973
974 translate(tc, tgsi_inst, dst_, src_);
975 }
976
977 static void
978 soa_per_channel(struct toy_compiler *tc,
979 const struct tgsi_full_instruction *tgsi_inst,
980 struct toy_dst *dst_,
981 struct toy_src *src_)
982 {
983 struct toy_dst dst[TGSI_FULL_MAX_DST_REGISTERS][4];
984 struct toy_src src[TGSI_FULL_MAX_SRC_REGISTERS][4];
985 int i, ch;
986
987 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++)
988 tdst_transpose(dst_[i], dst[i]);
989 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
990 tsrc_transpose(src_[i], src[i]);
991
992 /* emit the same instruction four times for the four channels */
993 for (ch = 0; ch < 4; ch++) {
994 struct toy_dst aos_dst[TGSI_FULL_MAX_DST_REGISTERS];
995 struct toy_src aos_src[TGSI_FULL_MAX_SRC_REGISTERS];
996
997 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++)
998 aos_dst[i] = dst[i][ch];
999 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
1000 aos_src[i] = src[i][ch];
1001
1002 aos_translate_table[tgsi_inst->Instruction.Opcode](tc,
1003 tgsi_inst, aos_dst, aos_src);
1004 }
1005 }
1006
1007 static void
1008 soa_scalar_replicate(struct toy_compiler *tc,
1009 const struct tgsi_full_instruction *tgsi_inst,
1010 struct toy_dst *dst_,
1011 struct toy_src *src_)
1012 {
1013 struct toy_dst dst0[4], tmp;
1014 struct toy_src srcx[TGSI_FULL_MAX_SRC_REGISTERS];
1015 int opcode, i;
1016
1017 assert(tgsi_inst->Instruction.NumDstRegs == 1);
1018
1019 tdst_transpose(dst_[0], dst0);
1020 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++) {
1021 struct toy_src tmp[4];
1022
1023 tsrc_transpose(src_[i], tmp);
1024 /* only the X channels */
1025 srcx[i] = tmp[0];
1026 }
1027
1028 tmp = tc_alloc_tmp(tc);
1029
1030 opcode = aos_simple_opcode_map[tgsi_inst->Instruction.Opcode].opcode;
1031 assert(opcode);
1032
1033 switch (tgsi_inst->Instruction.Opcode) {
1034 case TGSI_OPCODE_RCP:
1035 case TGSI_OPCODE_RSQ:
1036 case TGSI_OPCODE_SQRT:
1037 case TGSI_OPCODE_EX2:
1038 case TGSI_OPCODE_LG2:
1039 case TGSI_OPCODE_COS:
1040 case TGSI_OPCODE_SIN:
1041 tc_add1(tc, opcode, tmp, srcx[0]);
1042 break;
1043 case TGSI_OPCODE_POW:
1044 tc_add2(tc, opcode, tmp, srcx[0], srcx[1]);
1045 break;
1046 default:
1047 assert(!"invalid soa_scalar_replicate() call");
1048 return;
1049 }
1050
1051 /* replicate the result */
1052 for (i = 0; i < 4; i++)
1053 tc_MOV(tc, dst0[i], tsrc_from(tmp));
1054 }
1055
1056 static void
1057 soa_dot_product(struct toy_compiler *tc,
1058 const struct tgsi_full_instruction *tgsi_inst,
1059 struct toy_dst *dst_,
1060 struct toy_src *src_)
1061 {
1062 struct toy_dst dst0[4], tmp;
1063 struct toy_src src[TGSI_FULL_MAX_SRC_REGISTERS][4];
1064 int i;
1065
1066 tdst_transpose(dst_[0], dst0);
1067 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
1068 tsrc_transpose(src_[i], src[i]);
1069
1070 tmp = tc_alloc_tmp(tc);
1071
1072 switch (tgsi_inst->Instruction.Opcode) {
1073 case TGSI_OPCODE_DP2:
1074 tc_MUL(tc, tmp, src[0][1], src[1][1]);
1075 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1076 break;
1077 case TGSI_OPCODE_DP2A:
1078 tc_MAC(tc, tmp, src[0][1], src[1][1], src[2][0]);
1079 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1080 break;
1081 case TGSI_OPCODE_DP3:
1082 tc_MUL(tc, tmp, src[0][2], src[1][2]);
1083 tc_MAC(tc, tmp, src[0][1], src[1][1], tsrc_from(tmp));
1084 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1085 break;
1086 case TGSI_OPCODE_DPH:
1087 tc_MAC(tc, tmp, src[0][2], src[1][2], src[1][3]);
1088 tc_MAC(tc, tmp, src[0][1], src[1][1], tsrc_from(tmp));
1089 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1090 break;
1091 case TGSI_OPCODE_DP4:
1092 tc_MUL(tc, tmp, src[0][3], src[1][3]);
1093 tc_MAC(tc, tmp, src[0][2], src[1][2], tsrc_from(tmp));
1094 tc_MAC(tc, tmp, src[0][1], src[1][1], tsrc_from(tmp));
1095 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1096 break;
1097 default:
1098 assert(!"invalid soa_dot_product() call");
1099 return;
1100 }
1101
1102 for (i = 0; i < 4; i++)
1103 tc_MOV(tc, dst0[i], tsrc_from(tmp));
1104 }
1105
1106 static void
1107 soa_partial_derivative(struct toy_compiler *tc,
1108 const struct tgsi_full_instruction *tgsi_inst,
1109 struct toy_dst *dst_,
1110 struct toy_src *src_)
1111 {
1112 if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_DDX)
1113 tc_add1(tc, TOY_OPCODE_DDX, dst_[0], src_[0]);
1114 else
1115 tc_add1(tc, TOY_OPCODE_DDY, dst_[0], src_[0]);
1116 }
1117
1118 static void
1119 soa_if(struct toy_compiler *tc,
1120 const struct tgsi_full_instruction *tgsi_inst,
1121 struct toy_dst *dst_,
1122 struct toy_src *src_)
1123 {
1124 struct toy_src src0[4];
1125
1126 assert(tsrc_is_swizzle1(src_[0]));
1127 tsrc_transpose(src_[0], src0);
1128
1129 if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_IF)
1130 tc_IF(tc, tdst_null(), src0[0], tsrc_imm_f(0.0f), GEN6_COND_NZ);
1131 else
1132 tc_IF(tc, tdst_null(), src0[0], tsrc_imm_d(0), GEN6_COND_NZ);
1133 }
1134
1135 static void
1136 soa_LIT(struct toy_compiler *tc,
1137 const struct tgsi_full_instruction *tgsi_inst,
1138 struct toy_dst *dst_,
1139 struct toy_src *src_)
1140 {
1141 struct toy_inst *inst;
1142 struct toy_dst dst0[4];
1143 struct toy_src src0[4];
1144
1145 tdst_transpose(dst_[0], dst0);
1146 tsrc_transpose(src_[0], src0);
1147
1148 tc_MOV(tc, dst0[0], tsrc_imm_f(1.0f));
1149 tc_MOV(tc, dst0[1], src0[0]);
1150 tc_POW(tc, dst0[2], src0[1], src0[3]);
1151 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1152
1153 /*
1154 * POW is calculated first because math with pred_ctrl is broken here.
1155 * But, why?
1156 */
1157 tc_CMP(tc, tdst_null(), src0[0], tsrc_imm_f(0.0f), GEN6_COND_L);
1158 inst = tc_MOV(tc, dst0[1], tsrc_imm_f(0.0f));
1159 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
1160 inst = tc_MOV(tc, dst0[2], tsrc_imm_f(0.0f));
1161 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
1162 }
1163
1164 static void
1165 soa_EXP(struct toy_compiler *tc,
1166 const struct tgsi_full_instruction *tgsi_inst,
1167 struct toy_dst *dst_,
1168 struct toy_src *src_)
1169 {
1170 struct toy_dst dst0[4];
1171 struct toy_src src0[4];
1172
1173 assert(!"SoA EXP untested");
1174
1175 tdst_transpose(dst_[0], dst0);
1176 tsrc_transpose(src_[0], src0);
1177
1178 if (!tdst_is_null(dst0[0])) {
1179 struct toy_dst tmp = tdst_d(tc_alloc_tmp(tc));
1180
1181 tc_RNDD(tc, tmp, src0[0]);
1182
1183 /* construct the floating point number manually */
1184 tc_ADD(tc, tmp, tsrc_from(tmp), tsrc_imm_d(127));
1185 tc_SHL(tc, tdst_d(dst0[0]), tsrc_from(tmp), tsrc_imm_d(23));
1186 }
1187
1188 tc_FRC(tc, dst0[1], src0[0]);
1189 tc_EXP(tc, dst0[2], src0[0]);
1190 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1191 }
1192
1193 static void
1194 soa_LOG(struct toy_compiler *tc,
1195 const struct tgsi_full_instruction *tgsi_inst,
1196 struct toy_dst *dst_,
1197 struct toy_src *src_)
1198 {
1199 struct toy_dst dst0[4];
1200 struct toy_src src0[4];
1201
1202 assert(!"SoA LOG untested");
1203
1204 tdst_transpose(dst_[0], dst0);
1205 tsrc_transpose(src_[0], src0);
1206
1207 if (dst_[0].writemask & TOY_WRITEMASK_XY) {
1208 struct toy_dst tmp = tdst_d(tc_alloc_tmp(tc));
1209
1210 /* exponent */
1211 tc_SHR(tc, tmp, tsrc_absolute(tsrc_d(src0[0])), tsrc_imm_d(23));
1212 tc_ADD(tc, dst0[0], tsrc_from(tmp), tsrc_imm_d(-127));
1213
1214 /* mantissa */
1215 tc_AND(tc, tmp, tsrc_d(src0[0]), tsrc_imm_d((1 << 23) - 1));
1216 tc_OR(tc, dst0[1], tsrc_from(tmp), tsrc_imm_d(127 << 23));
1217 }
1218
1219 tc_LOG(tc, dst0[2], src0[0]);
1220 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1221 }
1222
1223 static void
1224 soa_DST(struct toy_compiler *tc,
1225 const struct tgsi_full_instruction *tgsi_inst,
1226 struct toy_dst *dst_,
1227 struct toy_src *src_)
1228 {
1229 struct toy_dst dst0[4];
1230 struct toy_src src[2][4];
1231
1232 tdst_transpose(dst_[0], dst0);
1233 tsrc_transpose(src_[0], src[0]);
1234 tsrc_transpose(src_[1], src[1]);
1235
1236 tc_MOV(tc, dst0[0], tsrc_imm_f(1.0f));
1237 tc_MUL(tc, dst0[1], src[0][1], src[1][1]);
1238 tc_MOV(tc, dst0[2], src[0][2]);
1239 tc_MOV(tc, dst0[3], src[1][3]);
1240 }
1241
1242 static void
1243 soa_XPD(struct toy_compiler *tc,
1244 const struct tgsi_full_instruction *tgsi_inst,
1245 struct toy_dst *dst_,
1246 struct toy_src *src_)
1247 {
1248 struct toy_dst dst0[4];
1249 struct toy_src src[2][4];
1250
1251 tdst_transpose(dst_[0], dst0);
1252 tsrc_transpose(src_[0], src[0]);
1253 tsrc_transpose(src_[1], src[1]);
1254
1255 /* dst.x = src0.y * src1.z - src1.y * src0.z */
1256 tc_MUL(tc, dst0[0], src[0][2], src[1][1]);
1257 tc_MAC(tc, dst0[0], src[0][1], src[1][2], tsrc_negate(tsrc_from(dst0[0])));
1258
1259 /* dst.y = src0.z * src1.x - src1.z * src0.x */
1260 tc_MUL(tc, dst0[1], src[0][0], src[1][2]);
1261 tc_MAC(tc, dst0[1], src[0][2], src[1][0], tsrc_negate(tsrc_from(dst0[1])));
1262
1263 /* dst.z = src0.x * src1.y - src1.x * src0.y */
1264 tc_MUL(tc, dst0[2], src[0][1], src[1][0]);
1265 tc_MAC(tc, dst0[2], src[0][0], src[1][1], tsrc_negate(tsrc_from(dst0[2])));
1266
1267 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1268 }
1269
1270 static void
1271 soa_PK2H(struct toy_compiler *tc,
1272 const struct tgsi_full_instruction *tgsi_inst,
1273 struct toy_dst *dst_,
1274 struct toy_src *src_)
1275 {
1276 struct toy_dst tmp = tdst_ud(tc_alloc_tmp(tc));
1277 struct toy_dst dst0[4];
1278 struct toy_src src0[4];
1279 int i;
1280
1281 assert(!"SoA PK2H untested");
1282
1283 tdst_transpose(dst_[0], dst0);
1284 tsrc_transpose(src_[0], src0);
1285
1286 tc_SHL(tc, tmp, src0[1], tsrc_imm_ud(16));
1287 tc_OR(tc, tmp, src0[0], tsrc_from(tmp));
1288
1289 for (i = 0; i < 4; i++)
1290 tc_MOV(tc, dst0[i], tsrc_from(tmp));
1291 }
1292
1293 static void
1294 soa_UP2H(struct toy_compiler *tc,
1295 const struct tgsi_full_instruction *tgsi_inst,
1296 struct toy_dst *dst_,
1297 struct toy_src *src_)
1298 {
1299 struct toy_dst dst0[4];
1300 struct toy_src src0[4];
1301
1302 assert(!"SoA UP2H untested");
1303
1304 tdst_transpose(dst_[0], dst0);
1305 tsrc_transpose(src_[0], src0);
1306
1307 tc_AND(tc, tdst_ud(dst0[0]), tsrc_ud(src0[0]), tsrc_imm_ud(0xffff));
1308 tc_SHR(tc, tdst_ud(dst0[1]), tsrc_ud(src0[1]), tsrc_imm_ud(16));
1309 tc_AND(tc, tdst_ud(dst0[2]), tsrc_ud(src0[2]), tsrc_imm_ud(0xffff));
1310 tc_SHR(tc, tdst_ud(dst0[3]), tsrc_ud(src0[3]), tsrc_imm_ud(16));
1311
1312 }
1313
1314 static void
1315 soa_SCS(struct toy_compiler *tc,
1316 const struct tgsi_full_instruction *tgsi_inst,
1317 struct toy_dst *dst_,
1318 struct toy_src *src_)
1319 {
1320 struct toy_dst dst0[4];
1321 struct toy_src src0[4];
1322
1323 tdst_transpose(dst_[0], dst0);
1324 tsrc_transpose(src_[0], src0);
1325
1326 tc_add1(tc, TOY_OPCODE_COS, dst0[0], src0[0]);
1327 tc_add1(tc, TOY_OPCODE_SIN, dst0[1], src0[0]);
1328 tc_MOV(tc, dst0[2], tsrc_imm_f(0.0f));
1329 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1330 }
1331
1332 static void
1333 soa_unsupported(struct toy_compiler *tc,
1334 const struct tgsi_full_instruction *tgsi_inst,
1335 struct toy_dst *dst_,
1336 struct toy_src *src_)
1337 {
1338 const struct tgsi_opcode_info *info =
1339 tgsi_get_opcode_info(tgsi_inst->Instruction.Opcode);
1340
1341 ilo_warn("unsupported TGSI opcode in SoA form: TGSI_OPCODE_%s\n",
1342 info->mnemonic);
1343
1344 tc_fail(tc, "unsupported TGSI instruction in SoA form");
1345 }
1346
1347 static const toy_tgsi_translate soa_translate_table[TGSI_OPCODE_LAST] = {
1348 [TGSI_OPCODE_ARL] = soa_per_channel,
1349 [TGSI_OPCODE_MOV] = soa_per_channel,
1350 [TGSI_OPCODE_LIT] = soa_LIT,
1351 [TGSI_OPCODE_RCP] = soa_scalar_replicate,
1352 [TGSI_OPCODE_RSQ] = soa_scalar_replicate,
1353 [TGSI_OPCODE_EXP] = soa_EXP,
1354 [TGSI_OPCODE_LOG] = soa_LOG,
1355 [TGSI_OPCODE_MUL] = soa_per_channel,
1356 [TGSI_OPCODE_ADD] = soa_per_channel,
1357 [TGSI_OPCODE_DP3] = soa_dot_product,
1358 [TGSI_OPCODE_DP4] = soa_dot_product,
1359 [TGSI_OPCODE_DST] = soa_DST,
1360 [TGSI_OPCODE_MIN] = soa_per_channel,
1361 [TGSI_OPCODE_MAX] = soa_per_channel,
1362 [TGSI_OPCODE_SLT] = soa_per_channel,
1363 [TGSI_OPCODE_SGE] = soa_per_channel,
1364 [TGSI_OPCODE_MAD] = soa_per_channel,
1365 [TGSI_OPCODE_SUB] = soa_per_channel,
1366 [TGSI_OPCODE_LRP] = soa_per_channel,
1367 [TGSI_OPCODE_CND] = soa_per_channel,
1368 [TGSI_OPCODE_SQRT] = soa_scalar_replicate,
1369 [TGSI_OPCODE_DP2A] = soa_dot_product,
1370 [TGSI_OPCODE_FRC] = soa_per_channel,
1371 [TGSI_OPCODE_CLAMP] = soa_per_channel,
1372 [TGSI_OPCODE_FLR] = soa_per_channel,
1373 [TGSI_OPCODE_ROUND] = soa_per_channel,
1374 [TGSI_OPCODE_EX2] = soa_scalar_replicate,
1375 [TGSI_OPCODE_LG2] = soa_scalar_replicate,
1376 [TGSI_OPCODE_POW] = soa_scalar_replicate,
1377 [TGSI_OPCODE_XPD] = soa_XPD,
1378 [TGSI_OPCODE_ABS] = soa_per_channel,
1379 [TGSI_OPCODE_DPH] = soa_dot_product,
1380 [TGSI_OPCODE_COS] = soa_scalar_replicate,
1381 [TGSI_OPCODE_DDX] = soa_partial_derivative,
1382 [TGSI_OPCODE_DDY] = soa_partial_derivative,
1383 [TGSI_OPCODE_KILL] = soa_passthrough,
1384 [TGSI_OPCODE_PK2H] = soa_PK2H,
1385 [TGSI_OPCODE_PK2US] = soa_unsupported,
1386 [TGSI_OPCODE_PK4B] = soa_unsupported,
1387 [TGSI_OPCODE_PK4UB] = soa_unsupported,
1388 [TGSI_OPCODE_SEQ] = soa_per_channel,
1389 [TGSI_OPCODE_SFL] = soa_per_channel,
1390 [TGSI_OPCODE_SGT] = soa_per_channel,
1391 [TGSI_OPCODE_SIN] = soa_scalar_replicate,
1392 [TGSI_OPCODE_SLE] = soa_per_channel,
1393 [TGSI_OPCODE_SNE] = soa_per_channel,
1394 [TGSI_OPCODE_STR] = soa_per_channel,
1395 [TGSI_OPCODE_TEX] = soa_passthrough,
1396 [TGSI_OPCODE_TXD] = soa_passthrough,
1397 [TGSI_OPCODE_TXP] = soa_passthrough,
1398 [TGSI_OPCODE_UP2H] = soa_UP2H,
1399 [TGSI_OPCODE_UP2US] = soa_unsupported,
1400 [TGSI_OPCODE_UP4B] = soa_unsupported,
1401 [TGSI_OPCODE_UP4UB] = soa_unsupported,
1402 [TGSI_OPCODE_ARR] = soa_per_channel,
1403 [TGSI_OPCODE_BRA] = soa_unsupported,
1404 [TGSI_OPCODE_CAL] = soa_unsupported,
1405 [TGSI_OPCODE_RET] = soa_unsupported,
1406 [TGSI_OPCODE_SSG] = soa_per_channel,
1407 [TGSI_OPCODE_CMP] = soa_per_channel,
1408 [TGSI_OPCODE_SCS] = soa_SCS,
1409 [TGSI_OPCODE_TXB] = soa_passthrough,
1410 [TGSI_OPCODE_DIV] = soa_per_channel,
1411 [TGSI_OPCODE_DP2] = soa_dot_product,
1412 [TGSI_OPCODE_TXL] = soa_passthrough,
1413 [TGSI_OPCODE_BRK] = soa_passthrough,
1414 [TGSI_OPCODE_IF] = soa_if,
1415 [TGSI_OPCODE_UIF] = soa_if,
1416 [TGSI_OPCODE_ELSE] = soa_passthrough,
1417 [TGSI_OPCODE_ENDIF] = soa_passthrough,
1418 [TGSI_OPCODE_PUSHA] = soa_unsupported,
1419 [TGSI_OPCODE_POPA] = soa_unsupported,
1420 [TGSI_OPCODE_CEIL] = soa_per_channel,
1421 [TGSI_OPCODE_I2F] = soa_per_channel,
1422 [TGSI_OPCODE_NOT] = soa_per_channel,
1423 [TGSI_OPCODE_TRUNC] = soa_per_channel,
1424 [TGSI_OPCODE_SHL] = soa_per_channel,
1425 [TGSI_OPCODE_AND] = soa_per_channel,
1426 [TGSI_OPCODE_OR] = soa_per_channel,
1427 [TGSI_OPCODE_MOD] = soa_per_channel,
1428 [TGSI_OPCODE_XOR] = soa_per_channel,
1429 [TGSI_OPCODE_SAD] = soa_per_channel,
1430 [TGSI_OPCODE_TXF] = soa_passthrough,
1431 [TGSI_OPCODE_TXQ] = soa_passthrough,
1432 [TGSI_OPCODE_CONT] = soa_passthrough,
1433 [TGSI_OPCODE_EMIT] = soa_unsupported,
1434 [TGSI_OPCODE_ENDPRIM] = soa_unsupported,
1435 [TGSI_OPCODE_BGNLOOP] = soa_passthrough,
1436 [TGSI_OPCODE_BGNSUB] = soa_unsupported,
1437 [TGSI_OPCODE_ENDLOOP] = soa_passthrough,
1438 [TGSI_OPCODE_ENDSUB] = soa_unsupported,
1439 [TGSI_OPCODE_TXQ_LZ] = soa_passthrough,
1440 [TGSI_OPCODE_NOP] = soa_passthrough,
1441 [TGSI_OPCODE_FSEQ] = soa_per_channel,
1442 [TGSI_OPCODE_FSGE] = soa_per_channel,
1443 [TGSI_OPCODE_FSLT] = soa_per_channel,
1444 [TGSI_OPCODE_FSNE] = soa_per_channel,
1445 [TGSI_OPCODE_CALLNZ] = soa_unsupported,
1446 [TGSI_OPCODE_BREAKC] = soa_unsupported,
1447 [TGSI_OPCODE_KILL_IF] = soa_passthrough,
1448 [TGSI_OPCODE_END] = soa_passthrough,
1449 [TGSI_OPCODE_F2I] = soa_per_channel,
1450 [TGSI_OPCODE_IDIV] = soa_per_channel,
1451 [TGSI_OPCODE_IMAX] = soa_per_channel,
1452 [TGSI_OPCODE_IMIN] = soa_per_channel,
1453 [TGSI_OPCODE_INEG] = soa_per_channel,
1454 [TGSI_OPCODE_ISGE] = soa_per_channel,
1455 [TGSI_OPCODE_ISHR] = soa_per_channel,
1456 [TGSI_OPCODE_ISLT] = soa_per_channel,
1457 [TGSI_OPCODE_F2U] = soa_per_channel,
1458 [TGSI_OPCODE_U2F] = soa_per_channel,
1459 [TGSI_OPCODE_UADD] = soa_per_channel,
1460 [TGSI_OPCODE_UDIV] = soa_per_channel,
1461 [TGSI_OPCODE_UMAD] = soa_per_channel,
1462 [TGSI_OPCODE_UMAX] = soa_per_channel,
1463 [TGSI_OPCODE_UMIN] = soa_per_channel,
1464 [TGSI_OPCODE_UMOD] = soa_per_channel,
1465 [TGSI_OPCODE_UMUL] = soa_per_channel,
1466 [TGSI_OPCODE_USEQ] = soa_per_channel,
1467 [TGSI_OPCODE_USGE] = soa_per_channel,
1468 [TGSI_OPCODE_USHR] = soa_per_channel,
1469 [TGSI_OPCODE_USLT] = soa_per_channel,
1470 [TGSI_OPCODE_USNE] = soa_per_channel,
1471 [TGSI_OPCODE_SWITCH] = soa_unsupported,
1472 [TGSI_OPCODE_CASE] = soa_unsupported,
1473 [TGSI_OPCODE_DEFAULT] = soa_unsupported,
1474 [TGSI_OPCODE_ENDSWITCH] = soa_unsupported,
1475 [TGSI_OPCODE_SAMPLE] = soa_passthrough,
1476 [TGSI_OPCODE_SAMPLE_I] = soa_passthrough,
1477 [TGSI_OPCODE_SAMPLE_I_MS] = soa_passthrough,
1478 [TGSI_OPCODE_SAMPLE_B] = soa_passthrough,
1479 [TGSI_OPCODE_SAMPLE_C] = soa_passthrough,
1480 [TGSI_OPCODE_SAMPLE_C_LZ] = soa_passthrough,
1481 [TGSI_OPCODE_SAMPLE_D] = soa_passthrough,
1482 [TGSI_OPCODE_SAMPLE_L] = soa_passthrough,
1483 [TGSI_OPCODE_GATHER4] = soa_passthrough,
1484 [TGSI_OPCODE_SVIEWINFO] = soa_passthrough,
1485 [TGSI_OPCODE_SAMPLE_POS] = soa_passthrough,
1486 [TGSI_OPCODE_SAMPLE_INFO] = soa_passthrough,
1487 [TGSI_OPCODE_UARL] = soa_per_channel,
1488 [TGSI_OPCODE_UCMP] = soa_per_channel,
1489 [TGSI_OPCODE_IABS] = soa_per_channel,
1490 [TGSI_OPCODE_ISSG] = soa_per_channel,
1491 [TGSI_OPCODE_LOAD] = soa_unsupported,
1492 [TGSI_OPCODE_STORE] = soa_unsupported,
1493 [TGSI_OPCODE_MFENCE] = soa_unsupported,
1494 [TGSI_OPCODE_LFENCE] = soa_unsupported,
1495 [TGSI_OPCODE_SFENCE] = soa_unsupported,
1496 [TGSI_OPCODE_BARRIER] = soa_unsupported,
1497 [TGSI_OPCODE_ATOMUADD] = soa_unsupported,
1498 [TGSI_OPCODE_ATOMXCHG] = soa_unsupported,
1499 [TGSI_OPCODE_ATOMCAS] = soa_unsupported,
1500 [TGSI_OPCODE_ATOMAND] = soa_unsupported,
1501 [TGSI_OPCODE_ATOMOR] = soa_unsupported,
1502 [TGSI_OPCODE_ATOMXOR] = soa_unsupported,
1503 [TGSI_OPCODE_ATOMUMIN] = soa_unsupported,
1504 [TGSI_OPCODE_ATOMUMAX] = soa_unsupported,
1505 [TGSI_OPCODE_ATOMIMIN] = soa_unsupported,
1506 [TGSI_OPCODE_ATOMIMAX] = soa_unsupported,
1507 [TGSI_OPCODE_TEX2] = soa_passthrough,
1508 [TGSI_OPCODE_TXB2] = soa_passthrough,
1509 [TGSI_OPCODE_TXL2] = soa_passthrough,
1510 };
1511
1512 static bool
1513 ra_dst_is_indirect(const struct tgsi_full_dst_register *d)
1514 {
1515 return (d->Register.Indirect ||
1516 (d->Register.Dimension && d->Dimension.Indirect));
1517 }
1518
1519 static int
1520 ra_dst_index(const struct tgsi_full_dst_register *d)
1521 {
1522 assert(!d->Register.Indirect);
1523 return d->Register.Index;
1524 }
1525
1526 static int
1527 ra_dst_dimension(const struct tgsi_full_dst_register *d)
1528 {
1529 if (d->Register.Dimension) {
1530 assert(!d->Dimension.Indirect);
1531 return d->Dimension.Index;
1532 }
1533 else {
1534 return 0;
1535 }
1536 }
1537
1538 static bool
1539 ra_is_src_indirect(const struct tgsi_full_src_register *s)
1540 {
1541 return (s->Register.Indirect ||
1542 (s->Register.Dimension && s->Dimension.Indirect));
1543 }
1544
1545 static int
1546 ra_src_index(const struct tgsi_full_src_register *s)
1547 {
1548 assert(!s->Register.Indirect);
1549 return s->Register.Index;
1550 }
1551
1552 static int
1553 ra_src_dimension(const struct tgsi_full_src_register *s)
1554 {
1555 if (s->Register.Dimension) {
1556 assert(!s->Dimension.Indirect);
1557 return s->Dimension.Index;
1558 }
1559 else {
1560 return 0;
1561 }
1562 }
1563
1564 /**
1565 * Infer the type of either the sources or the destination.
1566 */
1567 static enum toy_type
1568 ra_infer_opcode_type(int tgsi_opcode, bool is_dst)
1569 {
1570 enum tgsi_opcode_type type;
1571
1572 if (is_dst)
1573 type = tgsi_opcode_infer_dst_type(tgsi_opcode);
1574 else
1575 type = tgsi_opcode_infer_src_type(tgsi_opcode);
1576
1577 switch (type) {
1578 case TGSI_TYPE_UNSIGNED:
1579 return TOY_TYPE_UD;
1580 case TGSI_TYPE_SIGNED:
1581 return TOY_TYPE_D;
1582 case TGSI_TYPE_FLOAT:
1583 return TOY_TYPE_F;
1584 case TGSI_TYPE_UNTYPED:
1585 case TGSI_TYPE_VOID:
1586 case TGSI_TYPE_DOUBLE:
1587 default:
1588 assert(!"unsupported TGSI type");
1589 return TOY_TYPE_UD;
1590 }
1591 }
1592
1593 /**
1594 * Return the type of an operand of the specified instruction.
1595 */
1596 static enum toy_type
1597 ra_get_type(struct toy_tgsi *tgsi, const struct tgsi_full_instruction *tgsi_inst,
1598 int operand, bool is_dst)
1599 {
1600 enum toy_type type;
1601 enum tgsi_file_type file;
1602
1603 /* we need to look at both src and dst for MOV */
1604 /* XXX it should not be this complex */
1605 if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_MOV) {
1606 const enum tgsi_file_type dst_file = tgsi_inst->Dst[0].Register.File;
1607 const enum tgsi_file_type src_file = tgsi_inst->Src[0].Register.File;
1608
1609 if (dst_file == TGSI_FILE_ADDRESS || src_file == TGSI_FILE_ADDRESS) {
1610 type = TOY_TYPE_D;
1611 }
1612 else if (src_file == TGSI_FILE_IMMEDIATE &&
1613 !tgsi_inst->Src[0].Register.Indirect) {
1614 const int src_idx = tgsi_inst->Src[0].Register.Index;
1615 type = tgsi->imm_data.types[src_idx];
1616 }
1617 else {
1618 /* this is the best we can do */
1619 type = TOY_TYPE_F;
1620 }
1621
1622 return type;
1623 }
1624 else if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_UCMP) {
1625 if (!is_dst && operand == 0)
1626 type = TOY_TYPE_UD;
1627 else
1628 type = TOY_TYPE_F;
1629
1630 return type;
1631 }
1632
1633 type = ra_infer_opcode_type(tgsi_inst->Instruction.Opcode, is_dst);
1634
1635 /* fix the type */
1636 file = (is_dst) ?
1637 tgsi_inst->Dst[operand].Register.File :
1638 tgsi_inst->Src[operand].Register.File;
1639 switch (file) {
1640 case TGSI_FILE_SAMPLER:
1641 case TGSI_FILE_RESOURCE:
1642 case TGSI_FILE_SAMPLER_VIEW:
1643 type = TOY_TYPE_D;
1644 break;
1645 case TGSI_FILE_ADDRESS:
1646 assert(type == TOY_TYPE_D);
1647 break;
1648 default:
1649 break;
1650 }
1651
1652 return type;
1653 }
1654
1655 /**
1656 * Allocate a VRF register.
1657 */
1658 static int
1659 ra_alloc_reg(struct toy_tgsi *tgsi, enum tgsi_file_type file)
1660 {
1661 const int count = (tgsi->aos) ? 1 : 4;
1662 return tc_alloc_vrf(tgsi->tc, count);
1663 }
1664
1665 /**
1666 * Construct the key for VRF mapping look-up.
1667 */
1668 static void *
1669 ra_get_map_key(enum tgsi_file_type file, unsigned dim, unsigned index)
1670 {
1671 intptr_t key;
1672
1673 /* this is ugly... */
1674 assert(file < 1 << 4);
1675 assert(dim < 1 << 12);
1676 assert(index < 1 << 16);
1677 key = (file << 28) | (dim << 16) | index;
1678
1679 return intptr_to_pointer(key);
1680 }
1681
1682 /**
1683 * Map a TGSI register to a VRF register.
1684 */
1685 static int
1686 ra_map_reg(struct toy_tgsi *tgsi, enum tgsi_file_type file,
1687 int dim, int index, bool *is_new)
1688 {
1689 void *key, *val;
1690 intptr_t vrf;
1691
1692 key = ra_get_map_key(file, dim, index);
1693
1694 /*
1695 * because we allocate vrf from 1 and on, val is never NULL as long as the
1696 * key exists
1697 */
1698 val = util_hash_table_get(tgsi->reg_mapping, key);
1699 if (val) {
1700 vrf = pointer_to_intptr(val);
1701
1702 if (is_new)
1703 *is_new = false;
1704 }
1705 else {
1706 vrf = (intptr_t) ra_alloc_reg(tgsi, file);
1707
1708 /* add to the mapping */
1709 val = intptr_to_pointer(vrf);
1710 util_hash_table_set(tgsi->reg_mapping, key, val);
1711
1712 if (is_new)
1713 *is_new = true;
1714 }
1715
1716 return (int) vrf;
1717 }
1718
1719 /**
1720 * Return true if the destination aliases any of the sources.
1721 */
1722 static bool
1723 ra_dst_is_aliasing(const struct tgsi_full_instruction *tgsi_inst, int dst_index)
1724 {
1725 const struct tgsi_full_dst_register *d = &tgsi_inst->Dst[dst_index];
1726 int i;
1727
1728 /* we need a scratch register for indirect dst anyway */
1729 if (ra_dst_is_indirect(d))
1730 return true;
1731
1732 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++) {
1733 const struct tgsi_full_src_register *s = &tgsi_inst->Src[i];
1734
1735 if (s->Register.File != d->Register.File)
1736 continue;
1737
1738 /*
1739 * we can go on to check dimension and index respectively, but
1740 * keep it simple for now
1741 */
1742 if (ra_is_src_indirect(s))
1743 return true;
1744 if (ra_src_dimension(s) == ra_dst_dimension(d) &&
1745 ra_src_index(s) == ra_dst_index(d))
1746 return true;
1747 }
1748
1749 return false;
1750 }
1751
1752 /**
1753 * Return the toy register for a TGSI destination operand.
1754 */
1755 static struct toy_dst
1756 ra_get_dst(struct toy_tgsi *tgsi,
1757 const struct tgsi_full_instruction *tgsi_inst, int dst_index,
1758 bool *is_scratch)
1759 {
1760 const struct tgsi_full_dst_register *d = &tgsi_inst->Dst[dst_index];
1761 bool need_vrf = false;
1762 struct toy_dst dst;
1763
1764 switch (d->Register.File) {
1765 case TGSI_FILE_NULL:
1766 dst = tdst_null();
1767 break;
1768 case TGSI_FILE_OUTPUT:
1769 case TGSI_FILE_TEMPORARY:
1770 case TGSI_FILE_ADDRESS:
1771 case TGSI_FILE_PREDICATE:
1772 need_vrf = true;
1773 break;
1774 default:
1775 assert(!"unhandled dst file");
1776 dst = tdst_null();
1777 break;
1778 }
1779
1780 if (need_vrf) {
1781 /* XXX we do not always need a scratch given the conditions... */
1782 const bool need_scratch =
1783 (ra_dst_is_indirect(d) || ra_dst_is_aliasing(tgsi_inst, dst_index) ||
1784 tgsi_inst->Instruction.Saturate);
1785 const enum toy_type type = ra_get_type(tgsi, tgsi_inst, dst_index, true);
1786 int vrf;
1787
1788 if (need_scratch) {
1789 vrf = ra_alloc_reg(tgsi, d->Register.File);
1790 }
1791 else {
1792 vrf = ra_map_reg(tgsi, d->Register.File,
1793 ra_dst_dimension(d), ra_dst_index(d), NULL);
1794 }
1795
1796 if (is_scratch)
1797 *is_scratch = need_scratch;
1798
1799 dst = tdst_full(TOY_FILE_VRF, type, TOY_RECT_LINEAR,
1800 false, 0, d->Register.WriteMask, vrf * TOY_REG_WIDTH);
1801 }
1802
1803 return dst;
1804 }
1805
1806 static struct toy_src
1807 ra_get_src_for_vrf(const struct tgsi_full_src_register *s,
1808 enum toy_type type, int vrf)
1809 {
1810 return tsrc_full(TOY_FILE_VRF, type, TOY_RECT_LINEAR,
1811 false, 0,
1812 s->Register.SwizzleX, s->Register.SwizzleY,
1813 s->Register.SwizzleZ, s->Register.SwizzleW,
1814 s->Register.Absolute, s->Register.Negate,
1815 vrf * TOY_REG_WIDTH);
1816 }
1817
1818 static int
1819 init_tgsi_reg(struct toy_tgsi *tgsi, struct toy_inst *inst,
1820 enum tgsi_file_type file, int index,
1821 const struct tgsi_ind_register *indirect,
1822 const struct tgsi_dimension *dimension,
1823 const struct tgsi_ind_register *dim_indirect)
1824 {
1825 struct toy_src src;
1826 int num_src = 0;
1827
1828 /* src[0]: TGSI file */
1829 inst->src[num_src++] = tsrc_imm_d(file);
1830
1831 /* src[1]: TGSI dimension */
1832 inst->src[num_src++] = tsrc_imm_d((dimension) ? dimension->Index : 0);
1833
1834 /* src[2]: TGSI dimension indirection */
1835 if (dim_indirect) {
1836 const int vrf = ra_map_reg(tgsi, dim_indirect->File, 0,
1837 dim_indirect->Index, NULL);
1838
1839 src = tsrc(TOY_FILE_VRF, vrf, 0);
1840 src = tsrc_swizzle1(tsrc_d(src), indirect->Swizzle);
1841 }
1842 else {
1843 src = tsrc_imm_d(0);
1844 }
1845
1846 inst->src[num_src++] = src;
1847
1848 /* src[3]: TGSI index */
1849 inst->src[num_src++] = tsrc_imm_d(index);
1850
1851 /* src[4]: TGSI index indirection */
1852 if (indirect) {
1853 const int vrf = ra_map_reg(tgsi, indirect->File, 0,
1854 indirect->Index, NULL);
1855
1856 src = tsrc(TOY_FILE_VRF, vrf, 0);
1857 src = tsrc_swizzle1(tsrc_d(src), indirect->Swizzle);
1858 }
1859 else {
1860 src = tsrc_imm_d(0);
1861 }
1862
1863 inst->src[num_src++] = src;
1864
1865 return num_src;
1866 }
1867
1868 static struct toy_src
1869 ra_get_src_indirect(struct toy_tgsi *tgsi,
1870 const struct tgsi_full_instruction *tgsi_inst,
1871 int src_index)
1872 {
1873 const struct tgsi_full_src_register *s = &tgsi_inst->Src[src_index];
1874 bool need_vrf = false, is_resource = false;
1875 struct toy_src src;
1876
1877 switch (s->Register.File) {
1878 case TGSI_FILE_NULL:
1879 src = tsrc_null();
1880 break;
1881 case TGSI_FILE_SAMPLER:
1882 case TGSI_FILE_RESOURCE:
1883 case TGSI_FILE_SAMPLER_VIEW:
1884 is_resource = true;
1885 /* fall through */
1886 case TGSI_FILE_CONSTANT:
1887 case TGSI_FILE_INPUT:
1888 case TGSI_FILE_SYSTEM_VALUE:
1889 case TGSI_FILE_TEMPORARY:
1890 case TGSI_FILE_ADDRESS:
1891 case TGSI_FILE_IMMEDIATE:
1892 case TGSI_FILE_PREDICATE:
1893 need_vrf = true;
1894 break;
1895 default:
1896 assert(!"unhandled src file");
1897 src = tsrc_null();
1898 break;
1899 }
1900
1901 if (need_vrf) {
1902 const enum toy_type type = ra_get_type(tgsi, tgsi_inst, src_index, false);
1903 int vrf;
1904
1905 if (is_resource) {
1906 assert(!s->Register.Dimension);
1907 assert(s->Register.Indirect);
1908
1909 vrf = ra_map_reg(tgsi, s->Indirect.File, 0, s->Indirect.Index, NULL);
1910 }
1911 else {
1912 vrf = ra_alloc_reg(tgsi, s->Register.File);
1913 }
1914
1915 src = ra_get_src_for_vrf(s, type, vrf);
1916
1917 /* emit indirect fetch */
1918 if (!is_resource) {
1919 struct toy_inst *inst;
1920
1921 inst = tc_add(tgsi->tc);
1922 inst->opcode = TOY_OPCODE_TGSI_INDIRECT_FETCH;
1923 inst->dst = tdst_from(src);
1924 inst->dst.writemask = TOY_WRITEMASK_XYZW;
1925
1926 init_tgsi_reg(tgsi, inst, s->Register.File, s->Register.Index,
1927 (s->Register.Indirect) ? &s->Indirect : NULL,
1928 (s->Register.Dimension) ? &s->Dimension : NULL,
1929 (s->Dimension.Indirect) ? &s->DimIndirect : NULL);
1930 }
1931 }
1932
1933 return src;
1934 }
1935
1936 /**
1937 * Return the toy register for a TGSI source operand.
1938 */
1939 static struct toy_src
1940 ra_get_src(struct toy_tgsi *tgsi,
1941 const struct tgsi_full_instruction *tgsi_inst,
1942 int src_index)
1943 {
1944 const struct tgsi_full_src_register *s = &tgsi_inst->Src[src_index];
1945 bool need_vrf = false;
1946 struct toy_src src;
1947
1948 if (ra_is_src_indirect(s))
1949 return ra_get_src_indirect(tgsi, tgsi_inst, src_index);
1950
1951 switch (s->Register.File) {
1952 case TGSI_FILE_NULL:
1953 src = tsrc_null();
1954 break;
1955 case TGSI_FILE_CONSTANT:
1956 case TGSI_FILE_INPUT:
1957 case TGSI_FILE_SYSTEM_VALUE:
1958 need_vrf = true;
1959 break;
1960 case TGSI_FILE_TEMPORARY:
1961 case TGSI_FILE_ADDRESS:
1962 case TGSI_FILE_PREDICATE:
1963 need_vrf = true;
1964 break;
1965 case TGSI_FILE_SAMPLER:
1966 case TGSI_FILE_RESOURCE:
1967 case TGSI_FILE_SAMPLER_VIEW:
1968 assert(!s->Register.Dimension);
1969 src = tsrc_imm_d(s->Register.Index);
1970 break;
1971 case TGSI_FILE_IMMEDIATE:
1972 {
1973 const uint32_t *imm;
1974 enum toy_type imm_type;
1975 bool is_scalar;
1976
1977 imm = toy_tgsi_get_imm(tgsi, s->Register.Index, &imm_type);
1978
1979 is_scalar =
1980 (imm[s->Register.SwizzleX] == imm[s->Register.SwizzleY] &&
1981 imm[s->Register.SwizzleX] == imm[s->Register.SwizzleZ] &&
1982 imm[s->Register.SwizzleX] == imm[s->Register.SwizzleW]);
1983
1984 if (is_scalar) {
1985 const enum toy_type type =
1986 ra_get_type(tgsi, tgsi_inst, src_index, false);
1987
1988 /* ignore imm_type */
1989 src = tsrc_imm_ud(imm[s->Register.SwizzleX]);
1990 src.type = type;
1991 src.absolute = s->Register.Absolute;
1992 src.negate = s->Register.Negate;
1993 }
1994 else {
1995 need_vrf = true;
1996 }
1997 }
1998 break;
1999 default:
2000 assert(!"unhandled src file");
2001 src = tsrc_null();
2002 break;
2003 }
2004
2005 if (need_vrf) {
2006 const enum toy_type type = ra_get_type(tgsi, tgsi_inst, src_index, false);
2007 bool is_new;
2008 int vrf;
2009
2010 vrf = ra_map_reg(tgsi, s->Register.File,
2011 ra_src_dimension(s), ra_src_index(s), &is_new);
2012
2013 src = ra_get_src_for_vrf(s, type, vrf);
2014
2015 if (is_new) {
2016 switch (s->Register.File) {
2017 case TGSI_FILE_TEMPORARY:
2018 case TGSI_FILE_ADDRESS:
2019 case TGSI_FILE_PREDICATE:
2020 {
2021 struct toy_dst dst = tdst_from(src);
2022 dst.writemask = TOY_WRITEMASK_XYZW;
2023
2024 /* always initialize registers before use */
2025 if (tgsi->aos) {
2026 tc_MOV(tgsi->tc, dst, tsrc_type(tsrc_imm_d(0), type));
2027 }
2028 else {
2029 struct toy_dst tdst[4];
2030 int i;
2031
2032 tdst_transpose(dst, tdst);
2033
2034 for (i = 0; i < 4; i++) {
2035 tc_MOV(tgsi->tc, tdst[i],
2036 tsrc_type(tsrc_imm_d(0), type));
2037 }
2038 }
2039 }
2040 break;
2041 default:
2042 break;
2043 }
2044 }
2045
2046 }
2047
2048 return src;
2049 }
2050
2051 static void
2052 parse_instruction(struct toy_tgsi *tgsi,
2053 const struct tgsi_full_instruction *tgsi_inst)
2054 {
2055 struct toy_dst dst[TGSI_FULL_MAX_DST_REGISTERS];
2056 struct toy_src src[TGSI_FULL_MAX_SRC_REGISTERS];
2057 bool dst_is_scratch[TGSI_FULL_MAX_DST_REGISTERS];
2058 toy_tgsi_translate translate;
2059 int i;
2060
2061 /* convert TGSI registers to toy registers */
2062 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
2063 src[i] = ra_get_src(tgsi, tgsi_inst, i);
2064 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++)
2065 dst[i] = ra_get_dst(tgsi, tgsi_inst, i, &dst_is_scratch[i]);
2066
2067 /* translate the instruction */
2068 translate = tgsi->translate_table[tgsi_inst->Instruction.Opcode];
2069 if (!translate) {
2070 if (tgsi->translate_table == soa_translate_table)
2071 soa_unsupported(tgsi->tc, tgsi_inst, dst, src);
2072 else
2073 aos_unsupported(tgsi->tc, tgsi_inst, dst, src);
2074 }
2075 translate(tgsi->tc, tgsi_inst, dst, src);
2076
2077 /* write the result to the real destinations if needed */
2078 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++) {
2079 const struct tgsi_full_dst_register *d = &tgsi_inst->Dst[i];
2080
2081 if (!dst_is_scratch[i])
2082 continue;
2083
2084 if (tgsi_inst->Instruction.Saturate == TGSI_SAT_MINUS_PLUS_ONE)
2085 tc_fail(tgsi->tc, "TGSI_SAT_MINUS_PLUS_ONE unhandled");
2086
2087 tgsi->tc->templ.saturate = tgsi_inst->Instruction.Saturate;
2088
2089 /* emit indirect store */
2090 if (ra_dst_is_indirect(d)) {
2091 struct toy_inst *inst;
2092
2093 inst = tc_add(tgsi->tc);
2094 inst->opcode = TOY_OPCODE_TGSI_INDIRECT_STORE;
2095 inst->dst = dst[i];
2096
2097 init_tgsi_reg(tgsi, inst, d->Register.File, d->Register.Index,
2098 (d->Register.Indirect) ? &d->Indirect : NULL,
2099 (d->Register.Dimension) ? &d->Dimension : NULL,
2100 (d->Dimension.Indirect) ? &d->DimIndirect : NULL);
2101 }
2102 else {
2103 const enum toy_type type = ra_get_type(tgsi, tgsi_inst, i, true);
2104 struct toy_dst real_dst;
2105 int vrf;
2106
2107 vrf = ra_map_reg(tgsi, d->Register.File,
2108 ra_dst_dimension(d), ra_dst_index(d), NULL);
2109 real_dst = tdst_full(TOY_FILE_VRF, type, TOY_RECT_LINEAR,
2110 false, 0, d->Register.WriteMask, vrf * TOY_REG_WIDTH);
2111
2112 if (tgsi->aos) {
2113 tc_MOV(tgsi->tc, real_dst, tsrc_from(dst[i]));
2114 }
2115 else {
2116 struct toy_dst tdst[4];
2117 struct toy_src tsrc[4];
2118 int j;
2119
2120 tdst_transpose(real_dst, tdst);
2121 tsrc_transpose(tsrc_from(dst[i]), tsrc);
2122
2123 for (j = 0; j < 4; j++)
2124 tc_MOV(tgsi->tc, tdst[j], tsrc[j]);
2125 }
2126 }
2127
2128 tgsi->tc->templ.saturate = false;
2129 }
2130
2131 switch (tgsi_inst->Instruction.Opcode) {
2132 case TGSI_OPCODE_KILL_IF:
2133 case TGSI_OPCODE_KILL:
2134 tgsi->uses_kill = true;
2135 break;
2136 }
2137
2138 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++) {
2139 const struct tgsi_full_src_register *s = &tgsi_inst->Src[i];
2140 if (s->Register.File == TGSI_FILE_CONSTANT && s->Register.Indirect)
2141 tgsi->const_indirect = true;
2142 }
2143
2144 /* remember channels written */
2145 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++) {
2146 const struct tgsi_full_dst_register *d = &tgsi_inst->Dst[i];
2147
2148 if (d->Register.File != TGSI_FILE_OUTPUT)
2149 continue;
2150 for (i = 0; i < tgsi->num_outputs; i++) {
2151 if (tgsi->outputs[i].index == d->Register.Index) {
2152 tgsi->outputs[i].undefined_mask &= ~d->Register.WriteMask;
2153 break;
2154 }
2155 }
2156 }
2157 }
2158
2159 static void
2160 decl_add_in(struct toy_tgsi *tgsi, const struct tgsi_full_declaration *decl)
2161 {
2162 static const struct tgsi_declaration_interp default_interp = {
2163 TGSI_INTERPOLATE_PERSPECTIVE, false, 0,
2164 };
2165 const struct tgsi_declaration_interp *interp =
2166 (decl->Declaration.Interpolate) ? &decl->Interp: &default_interp;
2167 int index;
2168
2169 if (decl->Range.Last >= Elements(tgsi->inputs)) {
2170 assert(!"invalid IN");
2171 return;
2172 }
2173
2174 for (index = decl->Range.First; index <= decl->Range.Last; index++) {
2175 const int slot = tgsi->num_inputs++;
2176
2177 tgsi->inputs[slot].index = index;
2178 tgsi->inputs[slot].usage_mask = decl->Declaration.UsageMask;
2179 if (decl->Declaration.Semantic) {
2180 tgsi->inputs[slot].semantic_name = decl->Semantic.Name;
2181 tgsi->inputs[slot].semantic_index = decl->Semantic.Index;
2182 }
2183 else {
2184 tgsi->inputs[slot].semantic_name = TGSI_SEMANTIC_GENERIC;
2185 tgsi->inputs[slot].semantic_index = index;
2186 }
2187 tgsi->inputs[slot].interp = interp->Interpolate;
2188 tgsi->inputs[slot].centroid = interp->Location == TGSI_INTERPOLATE_LOC_CENTROID;
2189 }
2190 }
2191
2192 static void
2193 decl_add_out(struct toy_tgsi *tgsi, const struct tgsi_full_declaration *decl)
2194 {
2195 int index;
2196
2197 if (decl->Range.Last >= Elements(tgsi->outputs)) {
2198 assert(!"invalid OUT");
2199 return;
2200 }
2201
2202 assert(decl->Declaration.Semantic);
2203
2204 for (index = decl->Range.First; index <= decl->Range.Last; index++) {
2205 const int slot = tgsi->num_outputs++;
2206
2207 tgsi->outputs[slot].index = index;
2208 tgsi->outputs[slot].undefined_mask = TOY_WRITEMASK_XYZW;
2209 tgsi->outputs[slot].usage_mask = decl->Declaration.UsageMask;
2210 tgsi->outputs[slot].semantic_name = decl->Semantic.Name;
2211 tgsi->outputs[slot].semantic_index = decl->Semantic.Index;
2212 }
2213 }
2214
2215 static void
2216 decl_add_sv(struct toy_tgsi *tgsi, const struct tgsi_full_declaration *decl)
2217 {
2218 int index;
2219
2220 if (decl->Range.Last >= Elements(tgsi->system_values)) {
2221 assert(!"invalid SV");
2222 return;
2223 }
2224
2225 for (index = decl->Range.First; index <= decl->Range.Last; index++) {
2226 const int slot = tgsi->num_system_values++;
2227
2228 tgsi->system_values[slot].index = index;
2229 if (decl->Declaration.Semantic) {
2230 tgsi->system_values[slot].semantic_name = decl->Semantic.Name;
2231 tgsi->system_values[slot].semantic_index = decl->Semantic.Index;
2232 }
2233 else {
2234 tgsi->system_values[slot].semantic_name = TGSI_SEMANTIC_GENERIC;
2235 tgsi->system_values[slot].semantic_index = index;
2236 }
2237 }
2238 }
2239
2240 /**
2241 * Emit an instruction to fetch the value of a TGSI register.
2242 */
2243 static void
2244 fetch_source(struct toy_tgsi *tgsi, enum tgsi_file_type file, int dim, int idx)
2245 {
2246 struct toy_dst dst;
2247 int vrf;
2248 enum toy_opcode opcode;
2249 enum toy_type type = TOY_TYPE_F;
2250
2251 switch (file) {
2252 case TGSI_FILE_INPUT:
2253 opcode = TOY_OPCODE_TGSI_IN;
2254 break;
2255 case TGSI_FILE_CONSTANT:
2256 opcode = TOY_OPCODE_TGSI_CONST;
2257 break;
2258 case TGSI_FILE_SYSTEM_VALUE:
2259 opcode = TOY_OPCODE_TGSI_SV;
2260 break;
2261 case TGSI_FILE_IMMEDIATE:
2262 opcode = TOY_OPCODE_TGSI_IMM;
2263 toy_tgsi_get_imm(tgsi, idx, &type);
2264 break;
2265 default:
2266 /* no need to fetch */
2267 return;
2268 break;
2269 }
2270
2271 vrf = ra_map_reg(tgsi, file, dim, idx, NULL);
2272 dst = tdst(TOY_FILE_VRF, vrf, 0);
2273 dst = tdst_type(dst, type);
2274
2275 tc_add2(tgsi->tc, opcode, dst, tsrc_imm_d(dim), tsrc_imm_d(idx));
2276 }
2277
2278 static void
2279 parse_declaration(struct toy_tgsi *tgsi,
2280 const struct tgsi_full_declaration *decl)
2281 {
2282 int i;
2283
2284 switch (decl->Declaration.File) {
2285 case TGSI_FILE_INPUT:
2286 decl_add_in(tgsi, decl);
2287 break;
2288 case TGSI_FILE_OUTPUT:
2289 decl_add_out(tgsi, decl);
2290 break;
2291 case TGSI_FILE_SYSTEM_VALUE:
2292 decl_add_sv(tgsi, decl);
2293 break;
2294 case TGSI_FILE_IMMEDIATE:
2295 /* immediates should be declared with TGSI_TOKEN_TYPE_IMMEDIATE */
2296 assert(!"unexpected immediate declaration");
2297 break;
2298 case TGSI_FILE_CONSTANT:
2299 if (tgsi->const_count <= decl->Range.Last)
2300 tgsi->const_count = decl->Range.Last + 1;
2301 break;
2302 case TGSI_FILE_NULL:
2303 case TGSI_FILE_TEMPORARY:
2304 case TGSI_FILE_SAMPLER:
2305 case TGSI_FILE_PREDICATE:
2306 case TGSI_FILE_ADDRESS:
2307 case TGSI_FILE_RESOURCE:
2308 case TGSI_FILE_SAMPLER_VIEW:
2309 /* nothing to do */
2310 break;
2311 default:
2312 assert(!"unhandled TGSI file");
2313 break;
2314 }
2315
2316 /* fetch the registers now */
2317 for (i = decl->Range.First; i <= decl->Range.Last; i++) {
2318 const int dim = (decl->Declaration.Dimension) ? decl->Dim.Index2D : 0;
2319 fetch_source(tgsi, decl->Declaration.File, dim, i);
2320 }
2321 }
2322
2323 static int
2324 add_imm(struct toy_tgsi *tgsi, enum toy_type type, const uint32_t *buf)
2325 {
2326 /* reallocate the buffer if necessary */
2327 if (tgsi->imm_data.cur >= tgsi->imm_data.size) {
2328 const int cur_size = tgsi->imm_data.size;
2329 int new_size;
2330 enum toy_type *new_types;
2331 uint32_t (*new_buf)[4];
2332
2333 new_size = (cur_size) ? cur_size << 1 : 16;
2334 while (new_size <= tgsi->imm_data.cur)
2335 new_size <<= 1;
2336
2337 new_buf = REALLOC(tgsi->imm_data.buf,
2338 cur_size * sizeof(new_buf[0]),
2339 new_size * sizeof(new_buf[0]));
2340 new_types = REALLOC(tgsi->imm_data.types,
2341 cur_size * sizeof(new_types[0]),
2342 new_size * sizeof(new_types[0]));
2343 if (!new_buf || !new_types) {
2344 if (new_buf)
2345 FREE(new_buf);
2346 if (new_types)
2347 FREE(new_types);
2348 return -1;
2349 }
2350
2351 tgsi->imm_data.buf = new_buf;
2352 tgsi->imm_data.types = new_types;
2353 tgsi->imm_data.size = new_size;
2354 }
2355
2356 tgsi->imm_data.types[tgsi->imm_data.cur] = type;
2357 memcpy(&tgsi->imm_data.buf[tgsi->imm_data.cur],
2358 buf, sizeof(tgsi->imm_data.buf[0]));
2359
2360 return tgsi->imm_data.cur++;
2361 }
2362
2363 static void
2364 parse_immediate(struct toy_tgsi *tgsi, const struct tgsi_full_immediate *imm)
2365 {
2366 enum toy_type type;
2367 uint32_t imm_buf[4];
2368 int idx;
2369
2370 switch (imm->Immediate.DataType) {
2371 case TGSI_IMM_FLOAT32:
2372 type = TOY_TYPE_F;
2373 imm_buf[0] = fui(imm->u[0].Float);
2374 imm_buf[1] = fui(imm->u[1].Float);
2375 imm_buf[2] = fui(imm->u[2].Float);
2376 imm_buf[3] = fui(imm->u[3].Float);
2377 break;
2378 case TGSI_IMM_INT32:
2379 type = TOY_TYPE_D;
2380 imm_buf[0] = (uint32_t) imm->u[0].Int;
2381 imm_buf[1] = (uint32_t) imm->u[1].Int;
2382 imm_buf[2] = (uint32_t) imm->u[2].Int;
2383 imm_buf[3] = (uint32_t) imm->u[3].Int;
2384 break;
2385 case TGSI_IMM_UINT32:
2386 type = TOY_TYPE_UD;
2387 imm_buf[0] = imm->u[0].Uint;
2388 imm_buf[1] = imm->u[1].Uint;
2389 imm_buf[2] = imm->u[2].Uint;
2390 imm_buf[3] = imm->u[3].Uint;
2391 break;
2392 default:
2393 assert(!"unhandled TGSI imm type");
2394 type = TOY_TYPE_F;
2395 memset(imm_buf, 0, sizeof(imm_buf));
2396 break;
2397 }
2398
2399 idx = add_imm(tgsi, type, imm_buf);
2400 if (idx >= 0)
2401 fetch_source(tgsi, TGSI_FILE_IMMEDIATE, 0, idx);
2402 else
2403 tc_fail(tgsi->tc, "failed to add TGSI imm");
2404 }
2405
2406 static void
2407 parse_property(struct toy_tgsi *tgsi, const struct tgsi_full_property *prop)
2408 {
2409 switch (prop->Property.PropertyName) {
2410 case TGSI_PROPERTY_VS_PROHIBIT_UCPS:
2411 tgsi->props.vs_prohibit_ucps = prop->u[0].Data;
2412 break;
2413 case TGSI_PROPERTY_FS_COORD_ORIGIN:
2414 tgsi->props.fs_coord_origin = prop->u[0].Data;
2415 break;
2416 case TGSI_PROPERTY_FS_COORD_PIXEL_CENTER:
2417 tgsi->props.fs_coord_pixel_center = prop->u[0].Data;
2418 break;
2419 case TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS:
2420 tgsi->props.fs_color0_writes_all_cbufs = prop->u[0].Data;
2421 break;
2422 case TGSI_PROPERTY_FS_DEPTH_LAYOUT:
2423 tgsi->props.fs_depth_layout = prop->u[0].Data;
2424 break;
2425 case TGSI_PROPERTY_GS_INPUT_PRIM:
2426 tgsi->props.gs_input_prim = prop->u[0].Data;
2427 break;
2428 case TGSI_PROPERTY_GS_OUTPUT_PRIM:
2429 tgsi->props.gs_output_prim = prop->u[0].Data;
2430 break;
2431 case TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES:
2432 tgsi->props.gs_max_output_vertices = prop->u[0].Data;
2433 break;
2434 default:
2435 assert(!"unhandled TGSI property");
2436 break;
2437 }
2438 }
2439
2440 static void
2441 parse_token(struct toy_tgsi *tgsi, const union tgsi_full_token *token)
2442 {
2443 switch (token->Token.Type) {
2444 case TGSI_TOKEN_TYPE_DECLARATION:
2445 parse_declaration(tgsi, &token->FullDeclaration);
2446 break;
2447 case TGSI_TOKEN_TYPE_IMMEDIATE:
2448 parse_immediate(tgsi, &token->FullImmediate);
2449 break;
2450 case TGSI_TOKEN_TYPE_INSTRUCTION:
2451 parse_instruction(tgsi, &token->FullInstruction);
2452 break;
2453 case TGSI_TOKEN_TYPE_PROPERTY:
2454 parse_property(tgsi, &token->FullProperty);
2455 break;
2456 default:
2457 assert(!"unhandled TGSI token type");
2458 break;
2459 }
2460 }
2461
2462 static enum pipe_error
2463 dump_reg_mapping(void *key, void *val, void *data)
2464 {
2465 int tgsi_file, tgsi_dim, tgsi_index;
2466 uint32_t sig, vrf;
2467
2468 sig = (uint32_t) pointer_to_intptr(key);
2469 vrf = (uint32_t) pointer_to_intptr(val);
2470
2471 /* see ra_get_map_key() */
2472 tgsi_file = (sig >> 28) & 0xf;
2473 tgsi_dim = (sig >> 16) & 0xfff;
2474 tgsi_index = (sig >> 0) & 0xffff;
2475
2476 if (tgsi_dim) {
2477 ilo_printf(" v%d:\t%s[%d][%d]\n", vrf,
2478 tgsi_file_name(tgsi_file), tgsi_dim, tgsi_index);
2479 }
2480 else {
2481 ilo_printf(" v%d:\t%s[%d]\n", vrf,
2482 tgsi_file_name(tgsi_file), tgsi_index);
2483 }
2484
2485 return PIPE_OK;
2486 }
2487
2488 /**
2489 * Dump the TGSI translator, currently only the register mapping.
2490 */
2491 void
2492 toy_tgsi_dump(const struct toy_tgsi *tgsi)
2493 {
2494 util_hash_table_foreach(tgsi->reg_mapping, dump_reg_mapping, NULL);
2495 }
2496
2497 /**
2498 * Clean up the TGSI translator.
2499 */
2500 void
2501 toy_tgsi_cleanup(struct toy_tgsi *tgsi)
2502 {
2503 FREE(tgsi->imm_data.buf);
2504 FREE(tgsi->imm_data.types);
2505
2506 util_hash_table_destroy(tgsi->reg_mapping);
2507 }
2508
2509 static unsigned
2510 reg_mapping_hash(void *key)
2511 {
2512 return (unsigned) pointer_to_intptr(key);
2513 }
2514
2515 static int
2516 reg_mapping_compare(void *key1, void *key2)
2517 {
2518 return (key1 != key2);
2519 }
2520
2521 /**
2522 * Initialize the TGSI translator.
2523 */
2524 static bool
2525 init_tgsi(struct toy_tgsi *tgsi, struct toy_compiler *tc, bool aos)
2526 {
2527 memset(tgsi, 0, sizeof(*tgsi));
2528
2529 tgsi->tc = tc;
2530 tgsi->aos = aos;
2531 tgsi->translate_table = (aos) ? aos_translate_table : soa_translate_table;
2532
2533 /* create a mapping of TGSI registers to VRF reigsters */
2534 tgsi->reg_mapping =
2535 util_hash_table_create(reg_mapping_hash, reg_mapping_compare);
2536
2537 return (tgsi->reg_mapping != NULL);
2538 }
2539
2540 /**
2541 * Translate TGSI tokens into toy instructions.
2542 */
2543 void
2544 toy_compiler_translate_tgsi(struct toy_compiler *tc,
2545 const struct tgsi_token *tokens, bool aos,
2546 struct toy_tgsi *tgsi)
2547 {
2548 struct tgsi_parse_context parse;
2549
2550 if (!init_tgsi(tgsi, tc, aos)) {
2551 tc_fail(tc, "failed to initialize TGSI translator");
2552 return;
2553 }
2554
2555 tgsi_parse_init(&parse, tokens);
2556 while (!tgsi_parse_end_of_tokens(&parse)) {
2557 tgsi_parse_token(&parse);
2558 parse_token(tgsi, &parse.FullToken);
2559 }
2560 tgsi_parse_free(&parse);
2561 }
2562
2563 /**
2564 * Map the TGSI register to VRF register.
2565 */
2566 int
2567 toy_tgsi_get_vrf(const struct toy_tgsi *tgsi,
2568 enum tgsi_file_type file, int dimension, int index)
2569 {
2570 void *key, *val;
2571
2572 key = ra_get_map_key(file, dimension, index);
2573
2574 val = util_hash_table_get(tgsi->reg_mapping, key);
2575
2576 return (val) ? pointer_to_intptr(val) : -1;
2577 }