projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallium: Drop the NRM and NRM4 opcodes.
[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_RCC] = aos_unsupported,
833 [TGSI_OPCODE_DPH] = aos_simple,
834 [TGSI_OPCODE_COS] = aos_simple,
835 [TGSI_OPCODE_DDX] = aos_unsupported,
836 [TGSI_OPCODE_DDY] = aos_unsupported,
837 [TGSI_OPCODE_KILL] = aos_simple,
838 [TGSI_OPCODE_PK2H] = aos_PK2H,
839 [TGSI_OPCODE_PK2US] = aos_unsupported,
840 [TGSI_OPCODE_PK4B] = aos_unsupported,
841 [TGSI_OPCODE_PK4UB] = aos_unsupported,
842 [TGSI_OPCODE_RFL] = aos_unsupported,
843 [TGSI_OPCODE_SEQ] = aos_set_on_cond,
844 [TGSI_OPCODE_SFL] = aos_SFL,
845 [TGSI_OPCODE_SGT] = aos_set_on_cond,
846 [TGSI_OPCODE_SIN] = aos_simple,
847 [TGSI_OPCODE_SLE] = aos_set_on_cond,
848 [TGSI_OPCODE_SNE] = aos_set_on_cond,
849 [TGSI_OPCODE_STR] = aos_STR,
850 [TGSI_OPCODE_TEX] = aos_tex,
851 [TGSI_OPCODE_TXD] = aos_tex,
852 [TGSI_OPCODE_TXP] = aos_tex,
853 [TGSI_OPCODE_UP2H] = aos_UP2H,
854 [TGSI_OPCODE_UP2US] = aos_unsupported,
855 [TGSI_OPCODE_UP4B] = aos_unsupported,
856 [TGSI_OPCODE_UP4UB] = aos_unsupported,
857 [TGSI_OPCODE_X2D] = aos_unsupported,
858 [TGSI_OPCODE_ARA] = aos_unsupported,
859 [TGSI_OPCODE_ARR] = aos_simple,
860 [TGSI_OPCODE_BRA] = aos_unsupported,
861 [TGSI_OPCODE_CAL] = aos_unsupported,
862 [TGSI_OPCODE_RET] = aos_unsupported,
863 [TGSI_OPCODE_SSG] = aos_set_sign,
864 [TGSI_OPCODE_CMP] = aos_compare,
865 [TGSI_OPCODE_SCS] = aos_SCS,
866 [TGSI_OPCODE_TXB] = aos_tex,
867 [TGSI_OPCODE_DIV] = aos_DIV,
868 [TGSI_OPCODE_DP2] = aos_simple,
869 [TGSI_OPCODE_TXL] = aos_tex,
870 [TGSI_OPCODE_BRK] = aos_BRK,
871 [TGSI_OPCODE_IF] = aos_simple,
872 [TGSI_OPCODE_UIF] = aos_simple,
873 [TGSI_OPCODE_ELSE] = aos_simple,
874 [TGSI_OPCODE_ENDIF] = aos_simple,
875 [TGSI_OPCODE_PUSHA] = aos_unsupported,
876 [TGSI_OPCODE_POPA] = aos_unsupported,
877 [TGSI_OPCODE_CEIL] = aos_CEIL,
878 [TGSI_OPCODE_I2F] = aos_simple,
879 [TGSI_OPCODE_NOT] = aos_simple,
880 [TGSI_OPCODE_TRUNC] = aos_simple,
881 [TGSI_OPCODE_SHL] = aos_simple,
882 [TGSI_OPCODE_AND] = aos_simple,
883 [TGSI_OPCODE_OR] = aos_simple,
884 [TGSI_OPCODE_MOD] = aos_simple,
885 [TGSI_OPCODE_XOR] = aos_simple,
886 [TGSI_OPCODE_SAD] = aos_SAD,
887 [TGSI_OPCODE_TXF] = aos_tex,
888 [TGSI_OPCODE_TXQ] = aos_tex,
889 [TGSI_OPCODE_CONT] = aos_CONT,
890 [TGSI_OPCODE_EMIT] = aos_simple,
891 [TGSI_OPCODE_ENDPRIM] = aos_simple,
892 [TGSI_OPCODE_BGNLOOP] = aos_BGNLOOP,
893 [TGSI_OPCODE_BGNSUB] = aos_unsupported,
894 [TGSI_OPCODE_ENDLOOP] = aos_ENDLOOP,
895 [TGSI_OPCODE_ENDSUB] = aos_unsupported,
896 [TGSI_OPCODE_TXQ_LZ] = aos_tex,
897 [TGSI_OPCODE_NOP] = aos_simple,
898 [TGSI_OPCODE_FSEQ] = aos_set_on_cond,
899 [TGSI_OPCODE_FSGE] = aos_set_on_cond,
900 [TGSI_OPCODE_FSLT] = aos_set_on_cond,
901 [TGSI_OPCODE_FSNE] = aos_set_on_cond,
902 [TGSI_OPCODE_CALLNZ] = aos_unsupported,
903 [TGSI_OPCODE_BREAKC] = aos_unsupported,
904 [TGSI_OPCODE_KILL_IF] = aos_simple,
905 [TGSI_OPCODE_END] = aos_simple,
906 [TGSI_OPCODE_F2I] = aos_simple,
907 [TGSI_OPCODE_IDIV] = aos_simple,
908 [TGSI_OPCODE_IMAX] = aos_simple,
909 [TGSI_OPCODE_IMIN] = aos_simple,
910 [TGSI_OPCODE_INEG] = aos_simple,
911 [TGSI_OPCODE_ISGE] = aos_set_on_cond,
912 [TGSI_OPCODE_ISHR] = aos_simple,
913 [TGSI_OPCODE_ISLT] = aos_set_on_cond,
914 [TGSI_OPCODE_F2U] = aos_simple,
915 [TGSI_OPCODE_U2F] = aos_simple,
916 [TGSI_OPCODE_UADD] = aos_simple,
917 [TGSI_OPCODE_UDIV] = aos_simple,
918 [TGSI_OPCODE_UMAD] = aos_simple,
919 [TGSI_OPCODE_UMAX] = aos_simple,
920 [TGSI_OPCODE_UMIN] = aos_simple,
921 [TGSI_OPCODE_UMOD] = aos_simple,
922 [TGSI_OPCODE_UMUL] = aos_simple,
923 [TGSI_OPCODE_USEQ] = aos_set_on_cond,
924 [TGSI_OPCODE_USGE] = aos_set_on_cond,
925 [TGSI_OPCODE_USHR] = aos_simple,
926 [TGSI_OPCODE_USLT] = aos_set_on_cond,
927 [TGSI_OPCODE_USNE] = aos_set_on_cond,
928 [TGSI_OPCODE_SWITCH] = aos_unsupported,
929 [TGSI_OPCODE_CASE] = aos_unsupported,
930 [TGSI_OPCODE_DEFAULT] = aos_unsupported,
931 [TGSI_OPCODE_ENDSWITCH] = aos_unsupported,
932 [TGSI_OPCODE_SAMPLE] = aos_sample,
933 [TGSI_OPCODE_SAMPLE_I] = aos_sample,
934 [TGSI_OPCODE_SAMPLE_I_MS] = aos_sample,
935 [TGSI_OPCODE_SAMPLE_B] = aos_sample,
936 [TGSI_OPCODE_SAMPLE_C] = aos_sample,
937 [TGSI_OPCODE_SAMPLE_C_LZ] = aos_sample,
938 [TGSI_OPCODE_SAMPLE_D] = aos_sample,
939 [TGSI_OPCODE_SAMPLE_L] = aos_sample,
940 [TGSI_OPCODE_GATHER4] = aos_sample,
941 [TGSI_OPCODE_SVIEWINFO] = aos_sample,
942 [TGSI_OPCODE_SAMPLE_POS] = aos_sample,
943 [TGSI_OPCODE_SAMPLE_INFO] = aos_sample,
944 [TGSI_OPCODE_UARL] = aos_simple,
945 [TGSI_OPCODE_UCMP] = aos_compare,
946 [TGSI_OPCODE_IABS] = aos_simple,
947 [TGSI_OPCODE_ISSG] = aos_set_sign,
948 [TGSI_OPCODE_LOAD] = aos_unsupported,
949 [TGSI_OPCODE_STORE] = aos_unsupported,
950 [TGSI_OPCODE_MFENCE] = aos_unsupported,
951 [TGSI_OPCODE_LFENCE] = aos_unsupported,
952 [TGSI_OPCODE_SFENCE] = aos_unsupported,
953 [TGSI_OPCODE_BARRIER] = aos_unsupported,
954 [TGSI_OPCODE_ATOMUADD] = aos_unsupported,
955 [TGSI_OPCODE_ATOMXCHG] = aos_unsupported,
956 [TGSI_OPCODE_ATOMCAS] = aos_unsupported,
957 [TGSI_OPCODE_ATOMAND] = aos_unsupported,
958 [TGSI_OPCODE_ATOMOR] = aos_unsupported,
959 [TGSI_OPCODE_ATOMXOR] = aos_unsupported,
960 [TGSI_OPCODE_ATOMUMIN] = aos_unsupported,
961 [TGSI_OPCODE_ATOMUMAX] = aos_unsupported,
962 [TGSI_OPCODE_ATOMIMIN] = aos_unsupported,
963 [TGSI_OPCODE_ATOMIMAX] = aos_unsupported,
964 [TGSI_OPCODE_TEX2] = aos_tex,
965 [TGSI_OPCODE_TXB2] = aos_tex,
966 [TGSI_OPCODE_TXL2] = aos_tex,
967 };
968
969 static void
970 soa_passthrough(struct toy_compiler *tc,
971 const struct tgsi_full_instruction *tgsi_inst,
972 struct toy_dst *dst_,
973 struct toy_src *src_)
974 {
975 const toy_tgsi_translate translate =
976 aos_translate_table[tgsi_inst->Instruction.Opcode];
977
978 translate(tc, tgsi_inst, dst_, src_);
979 }
980
981 static void
982 soa_per_channel(struct toy_compiler *tc,
983 const struct tgsi_full_instruction *tgsi_inst,
984 struct toy_dst *dst_,
985 struct toy_src *src_)
986 {
987 struct toy_dst dst[TGSI_FULL_MAX_DST_REGISTERS][4];
988 struct toy_src src[TGSI_FULL_MAX_SRC_REGISTERS][4];
989 int i, ch;
990
991 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++)
992 tdst_transpose(dst_[i], dst[i]);
993 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
994 tsrc_transpose(src_[i], src[i]);
995
996 /* emit the same instruction four times for the four channels */
997 for (ch = 0; ch < 4; ch++) {
998 struct toy_dst aos_dst[TGSI_FULL_MAX_DST_REGISTERS];
999 struct toy_src aos_src[TGSI_FULL_MAX_SRC_REGISTERS];
1000
1001 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++)
1002 aos_dst[i] = dst[i][ch];
1003 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
1004 aos_src[i] = src[i][ch];
1005
1006 aos_translate_table[tgsi_inst->Instruction.Opcode](tc,
1007 tgsi_inst, aos_dst, aos_src);
1008 }
1009 }
1010
1011 static void
1012 soa_scalar_replicate(struct toy_compiler *tc,
1013 const struct tgsi_full_instruction *tgsi_inst,
1014 struct toy_dst *dst_,
1015 struct toy_src *src_)
1016 {
1017 struct toy_dst dst0[4], tmp;
1018 struct toy_src srcx[TGSI_FULL_MAX_SRC_REGISTERS];
1019 int opcode, i;
1020
1021 assert(tgsi_inst->Instruction.NumDstRegs == 1);
1022
1023 tdst_transpose(dst_[0], dst0);
1024 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++) {
1025 struct toy_src tmp[4];
1026
1027 tsrc_transpose(src_[i], tmp);
1028 /* only the X channels */
1029 srcx[i] = tmp[0];
1030 }
1031
1032 tmp = tc_alloc_tmp(tc);
1033
1034 opcode = aos_simple_opcode_map[tgsi_inst->Instruction.Opcode].opcode;
1035 assert(opcode);
1036
1037 switch (tgsi_inst->Instruction.Opcode) {
1038 case TGSI_OPCODE_RCP:
1039 case TGSI_OPCODE_RSQ:
1040 case TGSI_OPCODE_SQRT:
1041 case TGSI_OPCODE_EX2:
1042 case TGSI_OPCODE_LG2:
1043 case TGSI_OPCODE_COS:
1044 case TGSI_OPCODE_SIN:
1045 tc_add1(tc, opcode, tmp, srcx[0]);
1046 break;
1047 case TGSI_OPCODE_POW:
1048 tc_add2(tc, opcode, tmp, srcx[0], srcx[1]);
1049 break;
1050 default:
1051 assert(!"invalid soa_scalar_replicate() call");
1052 return;
1053 }
1054
1055 /* replicate the result */
1056 for (i = 0; i < 4; i++)
1057 tc_MOV(tc, dst0[i], tsrc_from(tmp));
1058 }
1059
1060 static void
1061 soa_dot_product(struct toy_compiler *tc,
1062 const struct tgsi_full_instruction *tgsi_inst,
1063 struct toy_dst *dst_,
1064 struct toy_src *src_)
1065 {
1066 struct toy_dst dst0[4], tmp;
1067 struct toy_src src[TGSI_FULL_MAX_SRC_REGISTERS][4];
1068 int i;
1069
1070 tdst_transpose(dst_[0], dst0);
1071 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
1072 tsrc_transpose(src_[i], src[i]);
1073
1074 tmp = tc_alloc_tmp(tc);
1075
1076 switch (tgsi_inst->Instruction.Opcode) {
1077 case TGSI_OPCODE_DP2:
1078 tc_MUL(tc, tmp, src[0][1], src[1][1]);
1079 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1080 break;
1081 case TGSI_OPCODE_DP2A:
1082 tc_MAC(tc, tmp, src[0][1], src[1][1], src[2][0]);
1083 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1084 break;
1085 case TGSI_OPCODE_DP3:
1086 tc_MUL(tc, tmp, src[0][2], src[1][2]);
1087 tc_MAC(tc, tmp, src[0][1], src[1][1], tsrc_from(tmp));
1088 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1089 break;
1090 case TGSI_OPCODE_DPH:
1091 tc_MAC(tc, tmp, src[0][2], src[1][2], src[1][3]);
1092 tc_MAC(tc, tmp, src[0][1], src[1][1], tsrc_from(tmp));
1093 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1094 break;
1095 case TGSI_OPCODE_DP4:
1096 tc_MUL(tc, tmp, src[0][3], src[1][3]);
1097 tc_MAC(tc, tmp, src[0][2], src[1][2], tsrc_from(tmp));
1098 tc_MAC(tc, tmp, src[0][1], src[1][1], tsrc_from(tmp));
1099 tc_MAC(tc, tmp, src[0][0], src[1][0], tsrc_from(tmp));
1100 break;
1101 default:
1102 assert(!"invalid soa_dot_product() call");
1103 return;
1104 }
1105
1106 for (i = 0; i < 4; i++)
1107 tc_MOV(tc, dst0[i], tsrc_from(tmp));
1108 }
1109
1110 static void
1111 soa_partial_derivative(struct toy_compiler *tc,
1112 const struct tgsi_full_instruction *tgsi_inst,
1113 struct toy_dst *dst_,
1114 struct toy_src *src_)
1115 {
1116 if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_DDX)
1117 tc_add1(tc, TOY_OPCODE_DDX, dst_[0], src_[0]);
1118 else
1119 tc_add1(tc, TOY_OPCODE_DDY, dst_[0], src_[0]);
1120 }
1121
1122 static void
1123 soa_if(struct toy_compiler *tc,
1124 const struct tgsi_full_instruction *tgsi_inst,
1125 struct toy_dst *dst_,
1126 struct toy_src *src_)
1127 {
1128 struct toy_src src0[4];
1129
1130 assert(tsrc_is_swizzle1(src_[0]));
1131 tsrc_transpose(src_[0], src0);
1132
1133 if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_IF)
1134 tc_IF(tc, tdst_null(), src0[0], tsrc_imm_f(0.0f), GEN6_COND_NZ);
1135 else
1136 tc_IF(tc, tdst_null(), src0[0], tsrc_imm_d(0), GEN6_COND_NZ);
1137 }
1138
1139 static void
1140 soa_LIT(struct toy_compiler *tc,
1141 const struct tgsi_full_instruction *tgsi_inst,
1142 struct toy_dst *dst_,
1143 struct toy_src *src_)
1144 {
1145 struct toy_inst *inst;
1146 struct toy_dst dst0[4];
1147 struct toy_src src0[4];
1148
1149 tdst_transpose(dst_[0], dst0);
1150 tsrc_transpose(src_[0], src0);
1151
1152 tc_MOV(tc, dst0[0], tsrc_imm_f(1.0f));
1153 tc_MOV(tc, dst0[1], src0[0]);
1154 tc_POW(tc, dst0[2], src0[1], src0[3]);
1155 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1156
1157 /*
1158 * POW is calculated first because math with pred_ctrl is broken here.
1159 * But, why?
1160 */
1161 tc_CMP(tc, tdst_null(), src0[0], tsrc_imm_f(0.0f), GEN6_COND_L);
1162 inst = tc_MOV(tc, dst0[1], tsrc_imm_f(0.0f));
1163 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
1164 inst = tc_MOV(tc, dst0[2], tsrc_imm_f(0.0f));
1165 inst->pred_ctrl = GEN6_PREDCTRL_NORMAL;
1166 }
1167
1168 static void
1169 soa_EXP(struct toy_compiler *tc,
1170 const struct tgsi_full_instruction *tgsi_inst,
1171 struct toy_dst *dst_,
1172 struct toy_src *src_)
1173 {
1174 struct toy_dst dst0[4];
1175 struct toy_src src0[4];
1176
1177 assert(!"SoA EXP untested");
1178
1179 tdst_transpose(dst_[0], dst0);
1180 tsrc_transpose(src_[0], src0);
1181
1182 if (!tdst_is_null(dst0[0])) {
1183 struct toy_dst tmp = tdst_d(tc_alloc_tmp(tc));
1184
1185 tc_RNDD(tc, tmp, src0[0]);
1186
1187 /* construct the floating point number manually */
1188 tc_ADD(tc, tmp, tsrc_from(tmp), tsrc_imm_d(127));
1189 tc_SHL(tc, tdst_d(dst0[0]), tsrc_from(tmp), tsrc_imm_d(23));
1190 }
1191
1192 tc_FRC(tc, dst0[1], src0[0]);
1193 tc_EXP(tc, dst0[2], src0[0]);
1194 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1195 }
1196
1197 static void
1198 soa_LOG(struct toy_compiler *tc,
1199 const struct tgsi_full_instruction *tgsi_inst,
1200 struct toy_dst *dst_,
1201 struct toy_src *src_)
1202 {
1203 struct toy_dst dst0[4];
1204 struct toy_src src0[4];
1205
1206 assert(!"SoA LOG untested");
1207
1208 tdst_transpose(dst_[0], dst0);
1209 tsrc_transpose(src_[0], src0);
1210
1211 if (dst_[0].writemask & TOY_WRITEMASK_XY) {
1212 struct toy_dst tmp = tdst_d(tc_alloc_tmp(tc));
1213
1214 /* exponent */
1215 tc_SHR(tc, tmp, tsrc_absolute(tsrc_d(src0[0])), tsrc_imm_d(23));
1216 tc_ADD(tc, dst0[0], tsrc_from(tmp), tsrc_imm_d(-127));
1217
1218 /* mantissa */
1219 tc_AND(tc, tmp, tsrc_d(src0[0]), tsrc_imm_d((1 << 23) - 1));
1220 tc_OR(tc, dst0[1], tsrc_from(tmp), tsrc_imm_d(127 << 23));
1221 }
1222
1223 tc_LOG(tc, dst0[2], src0[0]);
1224 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1225 }
1226
1227 static void
1228 soa_DST(struct toy_compiler *tc,
1229 const struct tgsi_full_instruction *tgsi_inst,
1230 struct toy_dst *dst_,
1231 struct toy_src *src_)
1232 {
1233 struct toy_dst dst0[4];
1234 struct toy_src src[2][4];
1235
1236 tdst_transpose(dst_[0], dst0);
1237 tsrc_transpose(src_[0], src[0]);
1238 tsrc_transpose(src_[1], src[1]);
1239
1240 tc_MOV(tc, dst0[0], tsrc_imm_f(1.0f));
1241 tc_MUL(tc, dst0[1], src[0][1], src[1][1]);
1242 tc_MOV(tc, dst0[2], src[0][2]);
1243 tc_MOV(tc, dst0[3], src[1][3]);
1244 }
1245
1246 static void
1247 soa_XPD(struct toy_compiler *tc,
1248 const struct tgsi_full_instruction *tgsi_inst,
1249 struct toy_dst *dst_,
1250 struct toy_src *src_)
1251 {
1252 struct toy_dst dst0[4];
1253 struct toy_src src[2][4];
1254
1255 tdst_transpose(dst_[0], dst0);
1256 tsrc_transpose(src_[0], src[0]);
1257 tsrc_transpose(src_[1], src[1]);
1258
1259 /* dst.x = src0.y * src1.z - src1.y * src0.z */
1260 tc_MUL(tc, dst0[0], src[0][2], src[1][1]);
1261 tc_MAC(tc, dst0[0], src[0][1], src[1][2], tsrc_negate(tsrc_from(dst0[0])));
1262
1263 /* dst.y = src0.z * src1.x - src1.z * src0.x */
1264 tc_MUL(tc, dst0[1], src[0][0], src[1][2]);
1265 tc_MAC(tc, dst0[1], src[0][2], src[1][0], tsrc_negate(tsrc_from(dst0[1])));
1266
1267 /* dst.z = src0.x * src1.y - src1.x * src0.y */
1268 tc_MUL(tc, dst0[2], src[0][1], src[1][0]);
1269 tc_MAC(tc, dst0[2], src[0][0], src[1][1], tsrc_negate(tsrc_from(dst0[2])));
1270
1271 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1272 }
1273
1274 static void
1275 soa_PK2H(struct toy_compiler *tc,
1276 const struct tgsi_full_instruction *tgsi_inst,
1277 struct toy_dst *dst_,
1278 struct toy_src *src_)
1279 {
1280 struct toy_dst tmp = tdst_ud(tc_alloc_tmp(tc));
1281 struct toy_dst dst0[4];
1282 struct toy_src src0[4];
1283 int i;
1284
1285 assert(!"SoA PK2H untested");
1286
1287 tdst_transpose(dst_[0], dst0);
1288 tsrc_transpose(src_[0], src0);
1289
1290 tc_SHL(tc, tmp, src0[1], tsrc_imm_ud(16));
1291 tc_OR(tc, tmp, src0[0], tsrc_from(tmp));
1292
1293 for (i = 0; i < 4; i++)
1294 tc_MOV(tc, dst0[i], tsrc_from(tmp));
1295 }
1296
1297 static void
1298 soa_UP2H(struct toy_compiler *tc,
1299 const struct tgsi_full_instruction *tgsi_inst,
1300 struct toy_dst *dst_,
1301 struct toy_src *src_)
1302 {
1303 struct toy_dst dst0[4];
1304 struct toy_src src0[4];
1305
1306 assert(!"SoA UP2H untested");
1307
1308 tdst_transpose(dst_[0], dst0);
1309 tsrc_transpose(src_[0], src0);
1310
1311 tc_AND(tc, tdst_ud(dst0[0]), tsrc_ud(src0[0]), tsrc_imm_ud(0xffff));
1312 tc_SHR(tc, tdst_ud(dst0[1]), tsrc_ud(src0[1]), tsrc_imm_ud(16));
1313 tc_AND(tc, tdst_ud(dst0[2]), tsrc_ud(src0[2]), tsrc_imm_ud(0xffff));
1314 tc_SHR(tc, tdst_ud(dst0[3]), tsrc_ud(src0[3]), tsrc_imm_ud(16));
1315
1316 }
1317
1318 static void
1319 soa_SCS(struct toy_compiler *tc,
1320 const struct tgsi_full_instruction *tgsi_inst,
1321 struct toy_dst *dst_,
1322 struct toy_src *src_)
1323 {
1324 struct toy_dst dst0[4];
1325 struct toy_src src0[4];
1326
1327 tdst_transpose(dst_[0], dst0);
1328 tsrc_transpose(src_[0], src0);
1329
1330 tc_add1(tc, TOY_OPCODE_COS, dst0[0], src0[0]);
1331 tc_add1(tc, TOY_OPCODE_SIN, dst0[1], src0[0]);
1332 tc_MOV(tc, dst0[2], tsrc_imm_f(0.0f));
1333 tc_MOV(tc, dst0[3], tsrc_imm_f(1.0f));
1334 }
1335
1336 static void
1337 soa_unsupported(struct toy_compiler *tc,
1338 const struct tgsi_full_instruction *tgsi_inst,
1339 struct toy_dst *dst_,
1340 struct toy_src *src_)
1341 {
1342 const struct tgsi_opcode_info *info =
1343 tgsi_get_opcode_info(tgsi_inst->Instruction.Opcode);
1344
1345 ilo_warn("unsupported TGSI opcode in SoA form: TGSI_OPCODE_%s\n",
1346 info->mnemonic);
1347
1348 tc_fail(tc, "unsupported TGSI instruction in SoA form");
1349 }
1350
1351 static const toy_tgsi_translate soa_translate_table[TGSI_OPCODE_LAST] = {
1352 [TGSI_OPCODE_ARL] = soa_per_channel,
1353 [TGSI_OPCODE_MOV] = soa_per_channel,
1354 [TGSI_OPCODE_LIT] = soa_LIT,
1355 [TGSI_OPCODE_RCP] = soa_scalar_replicate,
1356 [TGSI_OPCODE_RSQ] = soa_scalar_replicate,
1357 [TGSI_OPCODE_EXP] = soa_EXP,
1358 [TGSI_OPCODE_LOG] = soa_LOG,
1359 [TGSI_OPCODE_MUL] = soa_per_channel,
1360 [TGSI_OPCODE_ADD] = soa_per_channel,
1361 [TGSI_OPCODE_DP3] = soa_dot_product,
1362 [TGSI_OPCODE_DP4] = soa_dot_product,
1363 [TGSI_OPCODE_DST] = soa_DST,
1364 [TGSI_OPCODE_MIN] = soa_per_channel,
1365 [TGSI_OPCODE_MAX] = soa_per_channel,
1366 [TGSI_OPCODE_SLT] = soa_per_channel,
1367 [TGSI_OPCODE_SGE] = soa_per_channel,
1368 [TGSI_OPCODE_MAD] = soa_per_channel,
1369 [TGSI_OPCODE_SUB] = soa_per_channel,
1370 [TGSI_OPCODE_LRP] = soa_per_channel,
1371 [TGSI_OPCODE_CND] = soa_per_channel,
1372 [TGSI_OPCODE_SQRT] = soa_scalar_replicate,
1373 [TGSI_OPCODE_DP2A] = soa_dot_product,
1374 [TGSI_OPCODE_FRC] = soa_per_channel,
1375 [TGSI_OPCODE_CLAMP] = soa_per_channel,
1376 [TGSI_OPCODE_FLR] = soa_per_channel,
1377 [TGSI_OPCODE_ROUND] = soa_per_channel,
1378 [TGSI_OPCODE_EX2] = soa_scalar_replicate,
1379 [TGSI_OPCODE_LG2] = soa_scalar_replicate,
1380 [TGSI_OPCODE_POW] = soa_scalar_replicate,
1381 [TGSI_OPCODE_XPD] = soa_XPD,
1382 [TGSI_OPCODE_ABS] = soa_per_channel,
1383 [TGSI_OPCODE_RCC] = soa_unsupported,
1384 [TGSI_OPCODE_DPH] = soa_dot_product,
1385 [TGSI_OPCODE_COS] = soa_scalar_replicate,
1386 [TGSI_OPCODE_DDX] = soa_partial_derivative,
1387 [TGSI_OPCODE_DDY] = soa_partial_derivative,
1388 [TGSI_OPCODE_KILL] = soa_passthrough,
1389 [TGSI_OPCODE_PK2H] = soa_PK2H,
1390 [TGSI_OPCODE_PK2US] = soa_unsupported,
1391 [TGSI_OPCODE_PK4B] = soa_unsupported,
1392 [TGSI_OPCODE_PK4UB] = soa_unsupported,
1393 [TGSI_OPCODE_RFL] = soa_unsupported,
1394 [TGSI_OPCODE_SEQ] = soa_per_channel,
1395 [TGSI_OPCODE_SFL] = soa_per_channel,
1396 [TGSI_OPCODE_SGT] = soa_per_channel,
1397 [TGSI_OPCODE_SIN] = soa_scalar_replicate,
1398 [TGSI_OPCODE_SLE] = soa_per_channel,
1399 [TGSI_OPCODE_SNE] = soa_per_channel,
1400 [TGSI_OPCODE_STR] = soa_per_channel,
1401 [TGSI_OPCODE_TEX] = soa_passthrough,
1402 [TGSI_OPCODE_TXD] = soa_passthrough,
1403 [TGSI_OPCODE_TXP] = soa_passthrough,
1404 [TGSI_OPCODE_UP2H] = soa_UP2H,
1405 [TGSI_OPCODE_UP2US] = soa_unsupported,
1406 [TGSI_OPCODE_UP4B] = soa_unsupported,
1407 [TGSI_OPCODE_UP4UB] = soa_unsupported,
1408 [TGSI_OPCODE_X2D] = soa_unsupported,
1409 [TGSI_OPCODE_ARA] = soa_unsupported,
1410 [TGSI_OPCODE_ARR] = soa_per_channel,
1411 [TGSI_OPCODE_BRA] = soa_unsupported,
1412 [TGSI_OPCODE_CAL] = soa_unsupported,
1413 [TGSI_OPCODE_RET] = soa_unsupported,
1414 [TGSI_OPCODE_SSG] = soa_per_channel,
1415 [TGSI_OPCODE_CMP] = soa_per_channel,
1416 [TGSI_OPCODE_SCS] = soa_SCS,
1417 [TGSI_OPCODE_TXB] = soa_passthrough,
1418 [TGSI_OPCODE_DIV] = soa_per_channel,
1419 [TGSI_OPCODE_DP2] = soa_dot_product,
1420 [TGSI_OPCODE_TXL] = soa_passthrough,
1421 [TGSI_OPCODE_BRK] = soa_passthrough,
1422 [TGSI_OPCODE_IF] = soa_if,
1423 [TGSI_OPCODE_UIF] = soa_if,
1424 [TGSI_OPCODE_ELSE] = soa_passthrough,
1425 [TGSI_OPCODE_ENDIF] = soa_passthrough,
1426 [TGSI_OPCODE_PUSHA] = soa_unsupported,
1427 [TGSI_OPCODE_POPA] = soa_unsupported,
1428 [TGSI_OPCODE_CEIL] = soa_per_channel,
1429 [TGSI_OPCODE_I2F] = soa_per_channel,
1430 [TGSI_OPCODE_NOT] = soa_per_channel,
1431 [TGSI_OPCODE_TRUNC] = soa_per_channel,
1432 [TGSI_OPCODE_SHL] = soa_per_channel,
1433 [TGSI_OPCODE_AND] = soa_per_channel,
1434 [TGSI_OPCODE_OR] = soa_per_channel,
1435 [TGSI_OPCODE_MOD] = soa_per_channel,
1436 [TGSI_OPCODE_XOR] = soa_per_channel,
1437 [TGSI_OPCODE_SAD] = soa_per_channel,
1438 [TGSI_OPCODE_TXF] = soa_passthrough,
1439 [TGSI_OPCODE_TXQ] = soa_passthrough,
1440 [TGSI_OPCODE_CONT] = soa_passthrough,
1441 [TGSI_OPCODE_EMIT] = soa_unsupported,
1442 [TGSI_OPCODE_ENDPRIM] = soa_unsupported,
1443 [TGSI_OPCODE_BGNLOOP] = soa_passthrough,
1444 [TGSI_OPCODE_BGNSUB] = soa_unsupported,
1445 [TGSI_OPCODE_ENDLOOP] = soa_passthrough,
1446 [TGSI_OPCODE_ENDSUB] = soa_unsupported,
1447 [TGSI_OPCODE_TXQ_LZ] = soa_passthrough,
1448 [TGSI_OPCODE_NOP] = soa_passthrough,
1449 [TGSI_OPCODE_FSEQ] = soa_per_channel,
1450 [TGSI_OPCODE_FSGE] = soa_per_channel,
1451 [TGSI_OPCODE_FSLT] = soa_per_channel,
1452 [TGSI_OPCODE_FSNE] = soa_per_channel,
1453 [TGSI_OPCODE_CALLNZ] = soa_unsupported,
1454 [TGSI_OPCODE_BREAKC] = soa_unsupported,
1455 [TGSI_OPCODE_KILL_IF] = soa_passthrough,
1456 [TGSI_OPCODE_END] = soa_passthrough,
1457 [TGSI_OPCODE_F2I] = soa_per_channel,
1458 [TGSI_OPCODE_IDIV] = soa_per_channel,
1459 [TGSI_OPCODE_IMAX] = soa_per_channel,
1460 [TGSI_OPCODE_IMIN] = soa_per_channel,
1461 [TGSI_OPCODE_INEG] = soa_per_channel,
1462 [TGSI_OPCODE_ISGE] = soa_per_channel,
1463 [TGSI_OPCODE_ISHR] = soa_per_channel,
1464 [TGSI_OPCODE_ISLT] = soa_per_channel,
1465 [TGSI_OPCODE_F2U] = soa_per_channel,
1466 [TGSI_OPCODE_U2F] = soa_per_channel,
1467 [TGSI_OPCODE_UADD] = soa_per_channel,
1468 [TGSI_OPCODE_UDIV] = soa_per_channel,
1469 [TGSI_OPCODE_UMAD] = soa_per_channel,
1470 [TGSI_OPCODE_UMAX] = soa_per_channel,
1471 [TGSI_OPCODE_UMIN] = soa_per_channel,
1472 [TGSI_OPCODE_UMOD] = soa_per_channel,
1473 [TGSI_OPCODE_UMUL] = soa_per_channel,
1474 [TGSI_OPCODE_USEQ] = soa_per_channel,
1475 [TGSI_OPCODE_USGE] = soa_per_channel,
1476 [TGSI_OPCODE_USHR] = soa_per_channel,
1477 [TGSI_OPCODE_USLT] = soa_per_channel,
1478 [TGSI_OPCODE_USNE] = soa_per_channel,
1479 [TGSI_OPCODE_SWITCH] = soa_unsupported,
1480 [TGSI_OPCODE_CASE] = soa_unsupported,
1481 [TGSI_OPCODE_DEFAULT] = soa_unsupported,
1482 [TGSI_OPCODE_ENDSWITCH] = soa_unsupported,
1483 [TGSI_OPCODE_SAMPLE] = soa_passthrough,
1484 [TGSI_OPCODE_SAMPLE_I] = soa_passthrough,
1485 [TGSI_OPCODE_SAMPLE_I_MS] = soa_passthrough,
1486 [TGSI_OPCODE_SAMPLE_B] = soa_passthrough,
1487 [TGSI_OPCODE_SAMPLE_C] = soa_passthrough,
1488 [TGSI_OPCODE_SAMPLE_C_LZ] = soa_passthrough,
1489 [TGSI_OPCODE_SAMPLE_D] = soa_passthrough,
1490 [TGSI_OPCODE_SAMPLE_L] = soa_passthrough,
1491 [TGSI_OPCODE_GATHER4] = soa_passthrough,
1492 [TGSI_OPCODE_SVIEWINFO] = soa_passthrough,
1493 [TGSI_OPCODE_SAMPLE_POS] = soa_passthrough,
1494 [TGSI_OPCODE_SAMPLE_INFO] = soa_passthrough,
1495 [TGSI_OPCODE_UARL] = soa_per_channel,
1496 [TGSI_OPCODE_UCMP] = soa_per_channel,
1497 [TGSI_OPCODE_IABS] = soa_per_channel,
1498 [TGSI_OPCODE_ISSG] = soa_per_channel,
1499 [TGSI_OPCODE_LOAD] = soa_unsupported,
1500 [TGSI_OPCODE_STORE] = soa_unsupported,
1501 [TGSI_OPCODE_MFENCE] = soa_unsupported,
1502 [TGSI_OPCODE_LFENCE] = soa_unsupported,
1503 [TGSI_OPCODE_SFENCE] = soa_unsupported,
1504 [TGSI_OPCODE_BARRIER] = soa_unsupported,
1505 [TGSI_OPCODE_ATOMUADD] = soa_unsupported,
1506 [TGSI_OPCODE_ATOMXCHG] = soa_unsupported,
1507 [TGSI_OPCODE_ATOMCAS] = soa_unsupported,
1508 [TGSI_OPCODE_ATOMAND] = soa_unsupported,
1509 [TGSI_OPCODE_ATOMOR] = soa_unsupported,
1510 [TGSI_OPCODE_ATOMXOR] = soa_unsupported,
1511 [TGSI_OPCODE_ATOMUMIN] = soa_unsupported,
1512 [TGSI_OPCODE_ATOMUMAX] = soa_unsupported,
1513 [TGSI_OPCODE_ATOMIMIN] = soa_unsupported,
1514 [TGSI_OPCODE_ATOMIMAX] = soa_unsupported,
1515 [TGSI_OPCODE_TEX2] = soa_passthrough,
1516 [TGSI_OPCODE_TXB2] = soa_passthrough,
1517 [TGSI_OPCODE_TXL2] = soa_passthrough,
1518 };
1519
1520 static bool
1521 ra_dst_is_indirect(const struct tgsi_full_dst_register *d)
1522 {
1523 return (d->Register.Indirect ||
1524 (d->Register.Dimension && d->Dimension.Indirect));
1525 }
1526
1527 static int
1528 ra_dst_index(const struct tgsi_full_dst_register *d)
1529 {
1530 assert(!d->Register.Indirect);
1531 return d->Register.Index;
1532 }
1533
1534 static int
1535 ra_dst_dimension(const struct tgsi_full_dst_register *d)
1536 {
1537 if (d->Register.Dimension) {
1538 assert(!d->Dimension.Indirect);
1539 return d->Dimension.Index;
1540 }
1541 else {
1542 return 0;
1543 }
1544 }
1545
1546 static bool
1547 ra_is_src_indirect(const struct tgsi_full_src_register *s)
1548 {
1549 return (s->Register.Indirect ||
1550 (s->Register.Dimension && s->Dimension.Indirect));
1551 }
1552
1553 static int
1554 ra_src_index(const struct tgsi_full_src_register *s)
1555 {
1556 assert(!s->Register.Indirect);
1557 return s->Register.Index;
1558 }
1559
1560 static int
1561 ra_src_dimension(const struct tgsi_full_src_register *s)
1562 {
1563 if (s->Register.Dimension) {
1564 assert(!s->Dimension.Indirect);
1565 return s->Dimension.Index;
1566 }
1567 else {
1568 return 0;
1569 }
1570 }
1571
1572 /**
1573 * Infer the type of either the sources or the destination.
1574 */
1575 static enum toy_type
1576 ra_infer_opcode_type(int tgsi_opcode, bool is_dst)
1577 {
1578 enum tgsi_opcode_type type;
1579
1580 if (is_dst)
1581 type = tgsi_opcode_infer_dst_type(tgsi_opcode);
1582 else
1583 type = tgsi_opcode_infer_src_type(tgsi_opcode);
1584
1585 switch (type) {
1586 case TGSI_TYPE_UNSIGNED:
1587 return TOY_TYPE_UD;
1588 case TGSI_TYPE_SIGNED:
1589 return TOY_TYPE_D;
1590 case TGSI_TYPE_FLOAT:
1591 return TOY_TYPE_F;
1592 case TGSI_TYPE_UNTYPED:
1593 case TGSI_TYPE_VOID:
1594 case TGSI_TYPE_DOUBLE:
1595 default:
1596 assert(!"unsupported TGSI type");
1597 return TOY_TYPE_UD;
1598 }
1599 }
1600
1601 /**
1602 * Return the type of an operand of the specified instruction.
1603 */
1604 static enum toy_type
1605 ra_get_type(struct toy_tgsi *tgsi, const struct tgsi_full_instruction *tgsi_inst,
1606 int operand, bool is_dst)
1607 {
1608 enum toy_type type;
1609 enum tgsi_file_type file;
1610
1611 /* we need to look at both src and dst for MOV */
1612 /* XXX it should not be this complex */
1613 if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_MOV) {
1614 const enum tgsi_file_type dst_file = tgsi_inst->Dst[0].Register.File;
1615 const enum tgsi_file_type src_file = tgsi_inst->Src[0].Register.File;
1616
1617 if (dst_file == TGSI_FILE_ADDRESS || src_file == TGSI_FILE_ADDRESS) {
1618 type = TOY_TYPE_D;
1619 }
1620 else if (src_file == TGSI_FILE_IMMEDIATE &&
1621 !tgsi_inst->Src[0].Register.Indirect) {
1622 const int src_idx = tgsi_inst->Src[0].Register.Index;
1623 type = tgsi->imm_data.types[src_idx];
1624 }
1625 else {
1626 /* this is the best we can do */
1627 type = TOY_TYPE_F;
1628 }
1629
1630 return type;
1631 }
1632 else if (tgsi_inst->Instruction.Opcode == TGSI_OPCODE_UCMP) {
1633 if (!is_dst && operand == 0)
1634 type = TOY_TYPE_UD;
1635 else
1636 type = TOY_TYPE_F;
1637
1638 return type;
1639 }
1640
1641 type = ra_infer_opcode_type(tgsi_inst->Instruction.Opcode, is_dst);
1642
1643 /* fix the type */
1644 file = (is_dst) ?
1645 tgsi_inst->Dst[operand].Register.File :
1646 tgsi_inst->Src[operand].Register.File;
1647 switch (file) {
1648 case TGSI_FILE_SAMPLER:
1649 case TGSI_FILE_RESOURCE:
1650 case TGSI_FILE_SAMPLER_VIEW:
1651 type = TOY_TYPE_D;
1652 break;
1653 case TGSI_FILE_ADDRESS:
1654 assert(type == TOY_TYPE_D);
1655 break;
1656 default:
1657 break;
1658 }
1659
1660 return type;
1661 }
1662
1663 /**
1664 * Allocate a VRF register.
1665 */
1666 static int
1667 ra_alloc_reg(struct toy_tgsi *tgsi, enum tgsi_file_type file)
1668 {
1669 const int count = (tgsi->aos) ? 1 : 4;
1670 return tc_alloc_vrf(tgsi->tc, count);
1671 }
1672
1673 /**
1674 * Construct the key for VRF mapping look-up.
1675 */
1676 static void *
1677 ra_get_map_key(enum tgsi_file_type file, unsigned dim, unsigned index)
1678 {
1679 intptr_t key;
1680
1681 /* this is ugly... */
1682 assert(file < 1 << 4);
1683 assert(dim < 1 << 12);
1684 assert(index < 1 << 16);
1685 key = (file << 28) | (dim << 16) | index;
1686
1687 return intptr_to_pointer(key);
1688 }
1689
1690 /**
1691 * Map a TGSI register to a VRF register.
1692 */
1693 static int
1694 ra_map_reg(struct toy_tgsi *tgsi, enum tgsi_file_type file,
1695 int dim, int index, bool *is_new)
1696 {
1697 void *key, *val;
1698 intptr_t vrf;
1699
1700 key = ra_get_map_key(file, dim, index);
1701
1702 /*
1703 * because we allocate vrf from 1 and on, val is never NULL as long as the
1704 * key exists
1705 */
1706 val = util_hash_table_get(tgsi->reg_mapping, key);
1707 if (val) {
1708 vrf = pointer_to_intptr(val);
1709
1710 if (is_new)
1711 *is_new = false;
1712 }
1713 else {
1714 vrf = (intptr_t) ra_alloc_reg(tgsi, file);
1715
1716 /* add to the mapping */
1717 val = intptr_to_pointer(vrf);
1718 util_hash_table_set(tgsi->reg_mapping, key, val);
1719
1720 if (is_new)
1721 *is_new = true;
1722 }
1723
1724 return (int) vrf;
1725 }
1726
1727 /**
1728 * Return true if the destination aliases any of the sources.
1729 */
1730 static bool
1731 ra_dst_is_aliasing(const struct tgsi_full_instruction *tgsi_inst, int dst_index)
1732 {
1733 const struct tgsi_full_dst_register *d = &tgsi_inst->Dst[dst_index];
1734 int i;
1735
1736 /* we need a scratch register for indirect dst anyway */
1737 if (ra_dst_is_indirect(d))
1738 return true;
1739
1740 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++) {
1741 const struct tgsi_full_src_register *s = &tgsi_inst->Src[i];
1742
1743 if (s->Register.File != d->Register.File)
1744 continue;
1745
1746 /*
1747 * we can go on to check dimension and index respectively, but
1748 * keep it simple for now
1749 */
1750 if (ra_is_src_indirect(s))
1751 return true;
1752 if (ra_src_dimension(s) == ra_dst_dimension(d) &&
1753 ra_src_index(s) == ra_dst_index(d))
1754 return true;
1755 }
1756
1757 return false;
1758 }
1759
1760 /**
1761 * Return the toy register for a TGSI destination operand.
1762 */
1763 static struct toy_dst
1764 ra_get_dst(struct toy_tgsi *tgsi,
1765 const struct tgsi_full_instruction *tgsi_inst, int dst_index,
1766 bool *is_scratch)
1767 {
1768 const struct tgsi_full_dst_register *d = &tgsi_inst->Dst[dst_index];
1769 bool need_vrf = false;
1770 struct toy_dst dst;
1771
1772 switch (d->Register.File) {
1773 case TGSI_FILE_NULL:
1774 dst = tdst_null();
1775 break;
1776 case TGSI_FILE_OUTPUT:
1777 case TGSI_FILE_TEMPORARY:
1778 case TGSI_FILE_ADDRESS:
1779 case TGSI_FILE_PREDICATE:
1780 need_vrf = true;
1781 break;
1782 default:
1783 assert(!"unhandled dst file");
1784 dst = tdst_null();
1785 break;
1786 }
1787
1788 if (need_vrf) {
1789 /* XXX we do not always need a scratch given the conditions... */
1790 const bool need_scratch =
1791 (ra_dst_is_indirect(d) || ra_dst_is_aliasing(tgsi_inst, dst_index) ||
1792 tgsi_inst->Instruction.Saturate);
1793 const enum toy_type type = ra_get_type(tgsi, tgsi_inst, dst_index, true);
1794 int vrf;
1795
1796 if (need_scratch) {
1797 vrf = ra_alloc_reg(tgsi, d->Register.File);
1798 }
1799 else {
1800 vrf = ra_map_reg(tgsi, d->Register.File,
1801 ra_dst_dimension(d), ra_dst_index(d), NULL);
1802 }
1803
1804 if (is_scratch)
1805 *is_scratch = need_scratch;
1806
1807 dst = tdst_full(TOY_FILE_VRF, type, TOY_RECT_LINEAR,
1808 false, 0, d->Register.WriteMask, vrf * TOY_REG_WIDTH);
1809 }
1810
1811 return dst;
1812 }
1813
1814 static struct toy_src
1815 ra_get_src_for_vrf(const struct tgsi_full_src_register *s,
1816 enum toy_type type, int vrf)
1817 {
1818 return tsrc_full(TOY_FILE_VRF, type, TOY_RECT_LINEAR,
1819 false, 0,
1820 s->Register.SwizzleX, s->Register.SwizzleY,
1821 s->Register.SwizzleZ, s->Register.SwizzleW,
1822 s->Register.Absolute, s->Register.Negate,
1823 vrf * TOY_REG_WIDTH);
1824 }
1825
1826 static int
1827 init_tgsi_reg(struct toy_tgsi *tgsi, struct toy_inst *inst,
1828 enum tgsi_file_type file, int index,
1829 const struct tgsi_ind_register *indirect,
1830 const struct tgsi_dimension *dimension,
1831 const struct tgsi_ind_register *dim_indirect)
1832 {
1833 struct toy_src src;
1834 int num_src = 0;
1835
1836 /* src[0]: TGSI file */
1837 inst->src[num_src++] = tsrc_imm_d(file);
1838
1839 /* src[1]: TGSI dimension */
1840 inst->src[num_src++] = tsrc_imm_d((dimension) ? dimension->Index : 0);
1841
1842 /* src[2]: TGSI dimension indirection */
1843 if (dim_indirect) {
1844 const int vrf = ra_map_reg(tgsi, dim_indirect->File, 0,
1845 dim_indirect->Index, NULL);
1846
1847 src = tsrc(TOY_FILE_VRF, vrf, 0);
1848 src = tsrc_swizzle1(tsrc_d(src), indirect->Swizzle);
1849 }
1850 else {
1851 src = tsrc_imm_d(0);
1852 }
1853
1854 inst->src[num_src++] = src;
1855
1856 /* src[3]: TGSI index */
1857 inst->src[num_src++] = tsrc_imm_d(index);
1858
1859 /* src[4]: TGSI index indirection */
1860 if (indirect) {
1861 const int vrf = ra_map_reg(tgsi, indirect->File, 0,
1862 indirect->Index, NULL);
1863
1864 src = tsrc(TOY_FILE_VRF, vrf, 0);
1865 src = tsrc_swizzle1(tsrc_d(src), indirect->Swizzle);
1866 }
1867 else {
1868 src = tsrc_imm_d(0);
1869 }
1870
1871 inst->src[num_src++] = src;
1872
1873 return num_src;
1874 }
1875
1876 static struct toy_src
1877 ra_get_src_indirect(struct toy_tgsi *tgsi,
1878 const struct tgsi_full_instruction *tgsi_inst,
1879 int src_index)
1880 {
1881 const struct tgsi_full_src_register *s = &tgsi_inst->Src[src_index];
1882 bool need_vrf = false, is_resource = false;
1883 struct toy_src src;
1884
1885 switch (s->Register.File) {
1886 case TGSI_FILE_NULL:
1887 src = tsrc_null();
1888 break;
1889 case TGSI_FILE_SAMPLER:
1890 case TGSI_FILE_RESOURCE:
1891 case TGSI_FILE_SAMPLER_VIEW:
1892 is_resource = true;
1893 /* fall through */
1894 case TGSI_FILE_CONSTANT:
1895 case TGSI_FILE_INPUT:
1896 case TGSI_FILE_SYSTEM_VALUE:
1897 case TGSI_FILE_TEMPORARY:
1898 case TGSI_FILE_ADDRESS:
1899 case TGSI_FILE_IMMEDIATE:
1900 case TGSI_FILE_PREDICATE:
1901 need_vrf = true;
1902 break;
1903 default:
1904 assert(!"unhandled src file");
1905 src = tsrc_null();
1906 break;
1907 }
1908
1909 if (need_vrf) {
1910 const enum toy_type type = ra_get_type(tgsi, tgsi_inst, src_index, false);
1911 int vrf;
1912
1913 if (is_resource) {
1914 assert(!s->Register.Dimension);
1915 assert(s->Register.Indirect);
1916
1917 vrf = ra_map_reg(tgsi, s->Indirect.File, 0, s->Indirect.Index, NULL);
1918 }
1919 else {
1920 vrf = ra_alloc_reg(tgsi, s->Register.File);
1921 }
1922
1923 src = ra_get_src_for_vrf(s, type, vrf);
1924
1925 /* emit indirect fetch */
1926 if (!is_resource) {
1927 struct toy_inst *inst;
1928
1929 inst = tc_add(tgsi->tc);
1930 inst->opcode = TOY_OPCODE_TGSI_INDIRECT_FETCH;
1931 inst->dst = tdst_from(src);
1932 inst->dst.writemask = TOY_WRITEMASK_XYZW;
1933
1934 init_tgsi_reg(tgsi, inst, s->Register.File, s->Register.Index,
1935 (s->Register.Indirect) ? &s->Indirect : NULL,
1936 (s->Register.Dimension) ? &s->Dimension : NULL,
1937 (s->Dimension.Indirect) ? &s->DimIndirect : NULL);
1938 }
1939 }
1940
1941 return src;
1942 }
1943
1944 /**
1945 * Return the toy register for a TGSI source operand.
1946 */
1947 static struct toy_src
1948 ra_get_src(struct toy_tgsi *tgsi,
1949 const struct tgsi_full_instruction *tgsi_inst,
1950 int src_index)
1951 {
1952 const struct tgsi_full_src_register *s = &tgsi_inst->Src[src_index];
1953 bool need_vrf = false;
1954 struct toy_src src;
1955
1956 if (ra_is_src_indirect(s))
1957 return ra_get_src_indirect(tgsi, tgsi_inst, src_index);
1958
1959 switch (s->Register.File) {
1960 case TGSI_FILE_NULL:
1961 src = tsrc_null();
1962 break;
1963 case TGSI_FILE_CONSTANT:
1964 case TGSI_FILE_INPUT:
1965 case TGSI_FILE_SYSTEM_VALUE:
1966 need_vrf = true;
1967 break;
1968 case TGSI_FILE_TEMPORARY:
1969 case TGSI_FILE_ADDRESS:
1970 case TGSI_FILE_PREDICATE:
1971 need_vrf = true;
1972 break;
1973 case TGSI_FILE_SAMPLER:
1974 case TGSI_FILE_RESOURCE:
1975 case TGSI_FILE_SAMPLER_VIEW:
1976 assert(!s->Register.Dimension);
1977 src = tsrc_imm_d(s->Register.Index);
1978 break;
1979 case TGSI_FILE_IMMEDIATE:
1980 {
1981 const uint32_t *imm;
1982 enum toy_type imm_type;
1983 bool is_scalar;
1984
1985 imm = toy_tgsi_get_imm(tgsi, s->Register.Index, &imm_type);
1986
1987 is_scalar =
1988 (imm[s->Register.SwizzleX] == imm[s->Register.SwizzleY] &&
1989 imm[s->Register.SwizzleX] == imm[s->Register.SwizzleZ] &&
1990 imm[s->Register.SwizzleX] == imm[s->Register.SwizzleW]);
1991
1992 if (is_scalar) {
1993 const enum toy_type type =
1994 ra_get_type(tgsi, tgsi_inst, src_index, false);
1995
1996 /* ignore imm_type */
1997 src = tsrc_imm_ud(imm[s->Register.SwizzleX]);
1998 src.type = type;
1999 src.absolute = s->Register.Absolute;
2000 src.negate = s->Register.Negate;
2001 }
2002 else {
2003 need_vrf = true;
2004 }
2005 }
2006 break;
2007 default:
2008 assert(!"unhandled src file");
2009 src = tsrc_null();
2010 break;
2011 }
2012
2013 if (need_vrf) {
2014 const enum toy_type type = ra_get_type(tgsi, tgsi_inst, src_index, false);
2015 bool is_new;
2016 int vrf;
2017
2018 vrf = ra_map_reg(tgsi, s->Register.File,
2019 ra_src_dimension(s), ra_src_index(s), &is_new);
2020
2021 src = ra_get_src_for_vrf(s, type, vrf);
2022
2023 if (is_new) {
2024 switch (s->Register.File) {
2025 case TGSI_FILE_TEMPORARY:
2026 case TGSI_FILE_ADDRESS:
2027 case TGSI_FILE_PREDICATE:
2028 {
2029 struct toy_dst dst = tdst_from(src);
2030 dst.writemask = TOY_WRITEMASK_XYZW;
2031
2032 /* always initialize registers before use */
2033 if (tgsi->aos) {
2034 tc_MOV(tgsi->tc, dst, tsrc_type(tsrc_imm_d(0), type));
2035 }
2036 else {
2037 struct toy_dst tdst[4];
2038 int i;
2039
2040 tdst_transpose(dst, tdst);
2041
2042 for (i = 0; i < 4; i++) {
2043 tc_MOV(tgsi->tc, tdst[i],
2044 tsrc_type(tsrc_imm_d(0), type));
2045 }
2046 }
2047 }
2048 break;
2049 default:
2050 break;
2051 }
2052 }
2053
2054 }
2055
2056 return src;
2057 }
2058
2059 static void
2060 parse_instruction(struct toy_tgsi *tgsi,
2061 const struct tgsi_full_instruction *tgsi_inst)
2062 {
2063 struct toy_dst dst[TGSI_FULL_MAX_DST_REGISTERS];
2064 struct toy_src src[TGSI_FULL_MAX_SRC_REGISTERS];
2065 bool dst_is_scratch[TGSI_FULL_MAX_DST_REGISTERS];
2066 toy_tgsi_translate translate;
2067 int i;
2068
2069 /* convert TGSI registers to toy registers */
2070 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++)
2071 src[i] = ra_get_src(tgsi, tgsi_inst, i);
2072 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++)
2073 dst[i] = ra_get_dst(tgsi, tgsi_inst, i, &dst_is_scratch[i]);
2074
2075 /* translate the instruction */
2076 translate = tgsi->translate_table[tgsi_inst->Instruction.Opcode];
2077 if (!translate) {
2078 if (tgsi->translate_table == soa_translate_table)
2079 soa_unsupported(tgsi->tc, tgsi_inst, dst, src);
2080 else
2081 aos_unsupported(tgsi->tc, tgsi_inst, dst, src);
2082 }
2083 translate(tgsi->tc, tgsi_inst, dst, src);
2084
2085 /* write the result to the real destinations if needed */
2086 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++) {
2087 const struct tgsi_full_dst_register *d = &tgsi_inst->Dst[i];
2088
2089 if (!dst_is_scratch[i])
2090 continue;
2091
2092 if (tgsi_inst->Instruction.Saturate == TGSI_SAT_MINUS_PLUS_ONE)
2093 tc_fail(tgsi->tc, "TGSI_SAT_MINUS_PLUS_ONE unhandled");
2094
2095 tgsi->tc->templ.saturate = tgsi_inst->Instruction.Saturate;
2096
2097 /* emit indirect store */
2098 if (ra_dst_is_indirect(d)) {
2099 struct toy_inst *inst;
2100
2101 inst = tc_add(tgsi->tc);
2102 inst->opcode = TOY_OPCODE_TGSI_INDIRECT_STORE;
2103 inst->dst = dst[i];
2104
2105 init_tgsi_reg(tgsi, inst, d->Register.File, d->Register.Index,
2106 (d->Register.Indirect) ? &d->Indirect : NULL,
2107 (d->Register.Dimension) ? &d->Dimension : NULL,
2108 (d->Dimension.Indirect) ? &d->DimIndirect : NULL);
2109 }
2110 else {
2111 const enum toy_type type = ra_get_type(tgsi, tgsi_inst, i, true);
2112 struct toy_dst real_dst;
2113 int vrf;
2114
2115 vrf = ra_map_reg(tgsi, d->Register.File,
2116 ra_dst_dimension(d), ra_dst_index(d), NULL);
2117 real_dst = tdst_full(TOY_FILE_VRF, type, TOY_RECT_LINEAR,
2118 false, 0, d->Register.WriteMask, vrf * TOY_REG_WIDTH);
2119
2120 if (tgsi->aos) {
2121 tc_MOV(tgsi->tc, real_dst, tsrc_from(dst[i]));
2122 }
2123 else {
2124 struct toy_dst tdst[4];
2125 struct toy_src tsrc[4];
2126 int j;
2127
2128 tdst_transpose(real_dst, tdst);
2129 tsrc_transpose(tsrc_from(dst[i]), tsrc);
2130
2131 for (j = 0; j < 4; j++)
2132 tc_MOV(tgsi->tc, tdst[j], tsrc[j]);
2133 }
2134 }
2135
2136 tgsi->tc->templ.saturate = false;
2137 }
2138
2139 switch (tgsi_inst->Instruction.Opcode) {
2140 case TGSI_OPCODE_KILL_IF:
2141 case TGSI_OPCODE_KILL:
2142 tgsi->uses_kill = true;
2143 break;
2144 }
2145
2146 for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++) {
2147 const struct tgsi_full_src_register *s = &tgsi_inst->Src[i];
2148 if (s->Register.File == TGSI_FILE_CONSTANT && s->Register.Indirect)
2149 tgsi->const_indirect = true;
2150 }
2151
2152 /* remember channels written */
2153 for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++) {
2154 const struct tgsi_full_dst_register *d = &tgsi_inst->Dst[i];
2155
2156 if (d->Register.File != TGSI_FILE_OUTPUT)
2157 continue;
2158 for (i = 0; i < tgsi->num_outputs; i++) {
2159 if (tgsi->outputs[i].index == d->Register.Index) {
2160 tgsi->outputs[i].undefined_mask &= ~d->Register.WriteMask;
2161 break;
2162 }
2163 }
2164 }
2165 }
2166
2167 static void
2168 decl_add_in(struct toy_tgsi *tgsi, const struct tgsi_full_declaration *decl)
2169 {
2170 static const struct tgsi_declaration_interp default_interp = {
2171 TGSI_INTERPOLATE_PERSPECTIVE, false, 0,
2172 };
2173 const struct tgsi_declaration_interp *interp =
2174 (decl->Declaration.Interpolate) ? &decl->Interp: &default_interp;
2175 int index;
2176
2177 if (decl->Range.Last >= Elements(tgsi->inputs)) {
2178 assert(!"invalid IN");
2179 return;
2180 }
2181
2182 for (index = decl->Range.First; index <= decl->Range.Last; index++) {
2183 const int slot = tgsi->num_inputs++;
2184
2185 tgsi->inputs[slot].index = index;
2186 tgsi->inputs[slot].usage_mask = decl->Declaration.UsageMask;
2187 if (decl->Declaration.Semantic) {
2188 tgsi->inputs[slot].semantic_name = decl->Semantic.Name;
2189 tgsi->inputs[slot].semantic_index = decl->Semantic.Index;
2190 }
2191 else {
2192 tgsi->inputs[slot].semantic_name = TGSI_SEMANTIC_GENERIC;
2193 tgsi->inputs[slot].semantic_index = index;
2194 }
2195 tgsi->inputs[slot].interp = interp->Interpolate;
2196 tgsi->inputs[slot].centroid = interp->Location == TGSI_INTERPOLATE_LOC_CENTROID;
2197 }
2198 }
2199
2200 static void
2201 decl_add_out(struct toy_tgsi *tgsi, const struct tgsi_full_declaration *decl)
2202 {
2203 int index;
2204
2205 if (decl->Range.Last >= Elements(tgsi->outputs)) {
2206 assert(!"invalid OUT");
2207 return;
2208 }
2209
2210 assert(decl->Declaration.Semantic);
2211
2212 for (index = decl->Range.First; index <= decl->Range.Last; index++) {
2213 const int slot = tgsi->num_outputs++;
2214
2215 tgsi->outputs[slot].index = index;
2216 tgsi->outputs[slot].undefined_mask = TOY_WRITEMASK_XYZW;
2217 tgsi->outputs[slot].usage_mask = decl->Declaration.UsageMask;
2218 tgsi->outputs[slot].semantic_name = decl->Semantic.Name;
2219 tgsi->outputs[slot].semantic_index = decl->Semantic.Index;
2220 }
2221 }
2222
2223 static void
2224 decl_add_sv(struct toy_tgsi *tgsi, const struct tgsi_full_declaration *decl)
2225 {
2226 int index;
2227
2228 if (decl->Range.Last >= Elements(tgsi->system_values)) {
2229 assert(!"invalid SV");
2230 return;
2231 }
2232
2233 for (index = decl->Range.First; index <= decl->Range.Last; index++) {
2234 const int slot = tgsi->num_system_values++;
2235
2236 tgsi->system_values[slot].index = index;
2237 if (decl->Declaration.Semantic) {
2238 tgsi->system_values[slot].semantic_name = decl->Semantic.Name;
2239 tgsi->system_values[slot].semantic_index = decl->Semantic.Index;
2240 }
2241 else {
2242 tgsi->system_values[slot].semantic_name = TGSI_SEMANTIC_GENERIC;
2243 tgsi->system_values[slot].semantic_index = index;
2244 }
2245 }
2246 }
2247
2248 /**
2249 * Emit an instruction to fetch the value of a TGSI register.
2250 */
2251 static void
2252 fetch_source(struct toy_tgsi *tgsi, enum tgsi_file_type file, int dim, int idx)
2253 {
2254 struct toy_dst dst;
2255 int vrf;
2256 enum toy_opcode opcode;
2257 enum toy_type type = TOY_TYPE_F;
2258
2259 switch (file) {
2260 case TGSI_FILE_INPUT:
2261 opcode = TOY_OPCODE_TGSI_IN;
2262 break;
2263 case TGSI_FILE_CONSTANT:
2264 opcode = TOY_OPCODE_TGSI_CONST;
2265 break;
2266 case TGSI_FILE_SYSTEM_VALUE:
2267 opcode = TOY_OPCODE_TGSI_SV;
2268 break;
2269 case TGSI_FILE_IMMEDIATE:
2270 opcode = TOY_OPCODE_TGSI_IMM;
2271 toy_tgsi_get_imm(tgsi, idx, &type);
2272 break;
2273 default:
2274 /* no need to fetch */
2275 return;
2276 break;
2277 }
2278
2279 vrf = ra_map_reg(tgsi, file, dim, idx, NULL);
2280 dst = tdst(TOY_FILE_VRF, vrf, 0);
2281 dst = tdst_type(dst, type);
2282
2283 tc_add2(tgsi->tc, opcode, dst, tsrc_imm_d(dim), tsrc_imm_d(idx));
2284 }
2285
2286 static void
2287 parse_declaration(struct toy_tgsi *tgsi,
2288 const struct tgsi_full_declaration *decl)
2289 {
2290 int i;
2291
2292 switch (decl->Declaration.File) {
2293 case TGSI_FILE_INPUT:
2294 decl_add_in(tgsi, decl);
2295 break;
2296 case TGSI_FILE_OUTPUT:
2297 decl_add_out(tgsi, decl);
2298 break;
2299 case TGSI_FILE_SYSTEM_VALUE:
2300 decl_add_sv(tgsi, decl);
2301 break;
2302 case TGSI_FILE_IMMEDIATE:
2303 /* immediates should be declared with TGSI_TOKEN_TYPE_IMMEDIATE */
2304 assert(!"unexpected immediate declaration");
2305 break;
2306 case TGSI_FILE_CONSTANT:
2307 if (tgsi->const_count <= decl->Range.Last)
2308 tgsi->const_count = decl->Range.Last + 1;
2309 break;
2310 case TGSI_FILE_NULL:
2311 case TGSI_FILE_TEMPORARY:
2312 case TGSI_FILE_SAMPLER:
2313 case TGSI_FILE_PREDICATE:
2314 case TGSI_FILE_ADDRESS:
2315 case TGSI_FILE_RESOURCE:
2316 case TGSI_FILE_SAMPLER_VIEW:
2317 /* nothing to do */
2318 break;
2319 default:
2320 assert(!"unhandled TGSI file");
2321 break;
2322 }
2323
2324 /* fetch the registers now */
2325 for (i = decl->Range.First; i <= decl->Range.Last; i++) {
2326 const int dim = (decl->Declaration.Dimension) ? decl->Dim.Index2D : 0;
2327 fetch_source(tgsi, decl->Declaration.File, dim, i);
2328 }
2329 }
2330
2331 static int
2332 add_imm(struct toy_tgsi *tgsi, enum toy_type type, const uint32_t *buf)
2333 {
2334 /* reallocate the buffer if necessary */
2335 if (tgsi->imm_data.cur >= tgsi->imm_data.size) {
2336 const int cur_size = tgsi->imm_data.size;
2337 int new_size;
2338 enum toy_type *new_types;
2339 uint32_t (*new_buf)[4];
2340
2341 new_size = (cur_size) ? cur_size << 1 : 16;
2342 while (new_size <= tgsi->imm_data.cur)
2343 new_size <<= 1;
2344
2345 new_buf = REALLOC(tgsi->imm_data.buf,
2346 cur_size * sizeof(new_buf[0]),
2347 new_size * sizeof(new_buf[0]));
2348 new_types = REALLOC(tgsi->imm_data.types,
2349 cur_size * sizeof(new_types[0]),
2350 new_size * sizeof(new_types[0]));
2351 if (!new_buf || !new_types) {
2352 if (new_buf)
2353 FREE(new_buf);
2354 if (new_types)
2355 FREE(new_types);
2356 return -1;
2357 }
2358
2359 tgsi->imm_data.buf = new_buf;
2360 tgsi->imm_data.types = new_types;
2361 tgsi->imm_data.size = new_size;
2362 }
2363
2364 tgsi->imm_data.types[tgsi->imm_data.cur] = type;
2365 memcpy(&tgsi->imm_data.buf[tgsi->imm_data.cur],
2366 buf, sizeof(tgsi->imm_data.buf[0]));
2367
2368 return tgsi->imm_data.cur++;
2369 }
2370
2371 static void
2372 parse_immediate(struct toy_tgsi *tgsi, const struct tgsi_full_immediate *imm)
2373 {
2374 enum toy_type type;
2375 uint32_t imm_buf[4];
2376 int idx;
2377
2378 switch (imm->Immediate.DataType) {
2379 case TGSI_IMM_FLOAT32:
2380 type = TOY_TYPE_F;
2381 imm_buf[0] = fui(imm->u[0].Float);
2382 imm_buf[1] = fui(imm->u[1].Float);
2383 imm_buf[2] = fui(imm->u[2].Float);
2384 imm_buf[3] = fui(imm->u[3].Float);
2385 break;
2386 case TGSI_IMM_INT32:
2387 type = TOY_TYPE_D;
2388 imm_buf[0] = (uint32_t) imm->u[0].Int;
2389 imm_buf[1] = (uint32_t) imm->u[1].Int;
2390 imm_buf[2] = (uint32_t) imm->u[2].Int;
2391 imm_buf[3] = (uint32_t) imm->u[3].Int;
2392 break;
2393 case TGSI_IMM_UINT32:
2394 type = TOY_TYPE_UD;
2395 imm_buf[0] = imm->u[0].Uint;
2396 imm_buf[1] = imm->u[1].Uint;
2397 imm_buf[2] = imm->u[2].Uint;
2398 imm_buf[3] = imm->u[3].Uint;
2399 break;
2400 default:
2401 assert(!"unhandled TGSI imm type");
2402 type = TOY_TYPE_F;
2403 memset(imm_buf, 0, sizeof(imm_buf));
2404 break;
2405 }
2406
2407 idx = add_imm(tgsi, type, imm_buf);
2408 if (idx >= 0)
2409 fetch_source(tgsi, TGSI_FILE_IMMEDIATE, 0, idx);
2410 else
2411 tc_fail(tgsi->tc, "failed to add TGSI imm");
2412 }
2413
2414 static void
2415 parse_property(struct toy_tgsi *tgsi, const struct tgsi_full_property *prop)
2416 {
2417 switch (prop->Property.PropertyName) {
2418 case TGSI_PROPERTY_VS_PROHIBIT_UCPS:
2419 tgsi->props.vs_prohibit_ucps = prop->u[0].Data;
2420 break;
2421 case TGSI_PROPERTY_FS_COORD_ORIGIN:
2422 tgsi->props.fs_coord_origin = prop->u[0].Data;
2423 break;
2424 case TGSI_PROPERTY_FS_COORD_PIXEL_CENTER:
2425 tgsi->props.fs_coord_pixel_center = prop->u[0].Data;
2426 break;
2427 case TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS:
2428 tgsi->props.fs_color0_writes_all_cbufs = prop->u[0].Data;
2429 break;
2430 case TGSI_PROPERTY_FS_DEPTH_LAYOUT:
2431 tgsi->props.fs_depth_layout = prop->u[0].Data;
2432 break;
2433 case TGSI_PROPERTY_GS_INPUT_PRIM:
2434 tgsi->props.gs_input_prim = prop->u[0].Data;
2435 break;
2436 case TGSI_PROPERTY_GS_OUTPUT_PRIM:
2437 tgsi->props.gs_output_prim = prop->u[0].Data;
2438 break;
2439 case TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES:
2440 tgsi->props.gs_max_output_vertices = prop->u[0].Data;
2441 break;
2442 default:
2443 assert(!"unhandled TGSI property");
2444 break;
2445 }
2446 }
2447
2448 static void
2449 parse_token(struct toy_tgsi *tgsi, const union tgsi_full_token *token)
2450 {
2451 switch (token->Token.Type) {
2452 case TGSI_TOKEN_TYPE_DECLARATION:
2453 parse_declaration(tgsi, &token->FullDeclaration);
2454 break;
2455 case TGSI_TOKEN_TYPE_IMMEDIATE:
2456 parse_immediate(tgsi, &token->FullImmediate);
2457 break;
2458 case TGSI_TOKEN_TYPE_INSTRUCTION:
2459 parse_instruction(tgsi, &token->FullInstruction);
2460 break;
2461 case TGSI_TOKEN_TYPE_PROPERTY:
2462 parse_property(tgsi, &token->FullProperty);
2463 break;
2464 default:
2465 assert(!"unhandled TGSI token type");
2466 break;
2467 }
2468 }
2469
2470 static enum pipe_error
2471 dump_reg_mapping(void *key, void *val, void *data)
2472 {
2473 int tgsi_file, tgsi_dim, tgsi_index;
2474 uint32_t sig, vrf;
2475
2476 sig = (uint32_t) pointer_to_intptr(key);
2477 vrf = (uint32_t) pointer_to_intptr(val);
2478
2479 /* see ra_get_map_key() */
2480 tgsi_file = (sig >> 28) & 0xf;
2481 tgsi_dim = (sig >> 16) & 0xfff;
2482 tgsi_index = (sig >> 0) & 0xffff;
2483
2484 if (tgsi_dim) {
2485 ilo_printf(" v%d:\t%s[%d][%d]\n", vrf,
2486 tgsi_file_name(tgsi_file), tgsi_dim, tgsi_index);
2487 }
2488 else {
2489 ilo_printf(" v%d:\t%s[%d]\n", vrf,
2490 tgsi_file_name(tgsi_file), tgsi_index);
2491 }
2492
2493 return PIPE_OK;
2494 }
2495
2496 /**
2497 * Dump the TGSI translator, currently only the register mapping.
2498 */
2499 void
2500 toy_tgsi_dump(const struct toy_tgsi *tgsi)
2501 {
2502 util_hash_table_foreach(tgsi->reg_mapping, dump_reg_mapping, NULL);
2503 }
2504
2505 /**
2506 * Clean up the TGSI translator.
2507 */
2508 void
2509 toy_tgsi_cleanup(struct toy_tgsi *tgsi)
2510 {
2511 FREE(tgsi->imm_data.buf);
2512 FREE(tgsi->imm_data.types);
2513
2514 util_hash_table_destroy(tgsi->reg_mapping);
2515 }
2516
2517 static unsigned
2518 reg_mapping_hash(void *key)
2519 {
2520 return (unsigned) pointer_to_intptr(key);
2521 }
2522
2523 static int
2524 reg_mapping_compare(void *key1, void *key2)
2525 {
2526 return (key1 != key2);
2527 }
2528
2529 /**
2530 * Initialize the TGSI translator.
2531 */
2532 static bool
2533 init_tgsi(struct toy_tgsi *tgsi, struct toy_compiler *tc, bool aos)
2534 {
2535 memset(tgsi, 0, sizeof(*tgsi));
2536
2537 tgsi->tc = tc;
2538 tgsi->aos = aos;
2539 tgsi->translate_table = (aos) ? aos_translate_table : soa_translate_table;
2540
2541 /* create a mapping of TGSI registers to VRF reigsters */
2542 tgsi->reg_mapping =
2543 util_hash_table_create(reg_mapping_hash, reg_mapping_compare);
2544
2545 return (tgsi->reg_mapping != NULL);
2546 }
2547
2548 /**
2549 * Translate TGSI tokens into toy instructions.
2550 */
2551 void
2552 toy_compiler_translate_tgsi(struct toy_compiler *tc,
2553 const struct tgsi_token *tokens, bool aos,
2554 struct toy_tgsi *tgsi)
2555 {
2556 struct tgsi_parse_context parse;
2557
2558 if (!init_tgsi(tgsi, tc, aos)) {
2559 tc_fail(tc, "failed to initialize TGSI translator");
2560 return;
2561 }
2562
2563 tgsi_parse_init(&parse, tokens);
2564 while (!tgsi_parse_end_of_tokens(&parse)) {
2565 tgsi_parse_token(&parse);
2566 parse_token(tgsi, &parse.FullToken);
2567 }
2568 tgsi_parse_free(&parse);
2569 }
2570
2571 /**
2572 * Map the TGSI register to VRF register.
2573 */
2574 int
2575 toy_tgsi_get_vrf(const struct toy_tgsi *tgsi,
2576 enum tgsi_file_type file, int dimension, int index)
2577 {
2578 void *key, *val;
2579
2580 key = ra_get_map_key(file, dimension, index);
2581
2582 val = util_hash_table_get(tgsi->reg_mapping, key);
2583
2584 return (val) ? pointer_to_intptr(val) : -1;
2585 }