1 /*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 */
23 #include "r600_sq.h"
24 #include "r600_opcodes.h"
25 #include "r600_formats.h"
26 #include "r600_shader.h"
27 #include "r600d.h"
28
29 #include <errno.h>
30 #include "util/u_bitcast.h"
31 #include "util/u_dump.h"
32 #include "util/u_memory.h"
33 #include "util/u_math.h"
34 #include "pipe/p_shader_tokens.h"
35
36 #include "sb/sb_public.h"
37
38 #define NUM_OF_CYCLES 3
39 #define NUM_OF_COMPONENTS 4
40
alu_writes(struct r600_bytecode_alu * alu)41 static inline bool alu_writes(struct r600_bytecode_alu *alu)
42 {
43 return alu->dst.write || alu->is_op3;
44 }
45
r600_bytecode_get_num_operands(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)46 static inline unsigned int r600_bytecode_get_num_operands(
47 struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
48 {
49 return r600_isa_alu(alu->op)->src_count;
50 }
51
52 int r700_bytecode_alu_build(struct r600_bytecode *bc,
53 struct r600_bytecode_alu *alu, unsigned id);
54
r600_bytecode_cf(void)55 static struct r600_bytecode_cf *r600_bytecode_cf(void)
56 {
57 struct r600_bytecode_cf *cf = CALLOC_STRUCT(r600_bytecode_cf);
58
59 if (!cf)
60 return NULL;
61 LIST_INITHEAD(&cf->list);
62 LIST_INITHEAD(&cf->alu);
63 LIST_INITHEAD(&cf->vtx);
64 LIST_INITHEAD(&cf->tex);
65 LIST_INITHEAD(&cf->gds);
66 return cf;
67 }
68
r600_bytecode_alu(void)69 static struct r600_bytecode_alu *r600_bytecode_alu(void)
70 {
71 struct r600_bytecode_alu *alu = CALLOC_STRUCT(r600_bytecode_alu);
72
73 if (!alu)
74 return NULL;
75 LIST_INITHEAD(&alu->list);
76 return alu;
77 }
78
r600_bytecode_vtx(void)79 static struct r600_bytecode_vtx *r600_bytecode_vtx(void)
80 {
81 struct r600_bytecode_vtx *vtx = CALLOC_STRUCT(r600_bytecode_vtx);
82
83 if (!vtx)
84 return NULL;
85 LIST_INITHEAD(&vtx->list);
86 return vtx;
87 }
88
r600_bytecode_tex(void)89 static struct r600_bytecode_tex *r600_bytecode_tex(void)
90 {
91 struct r600_bytecode_tex *tex = CALLOC_STRUCT(r600_bytecode_tex);
92
93 if (!tex)
94 return NULL;
95 LIST_INITHEAD(&tex->list);
96 return tex;
97 }
98
r600_bytecode_gds(void)99 static struct r600_bytecode_gds *r600_bytecode_gds(void)
100 {
101 struct r600_bytecode_gds *gds = CALLOC_STRUCT(r600_bytecode_gds);
102
103 if (gds == NULL)
104 return NULL;
105 LIST_INITHEAD(&gds->list);
106 return gds;
107 }
108
stack_entry_size(enum radeon_family chip)109 static unsigned stack_entry_size(enum radeon_family chip) {
110 /* Wavefront size:
111 * 64: R600/RV670/RV770/Cypress/R740/Barts/Turks/Caicos/
112 * Aruba/Sumo/Sumo2/redwood/juniper
113 * 32: R630/R730/R710/Palm/Cedar
114 * 16: R610/Rs780
115 *
116 * Stack row size:
117 * Wavefront Size 16 32 48 64
118 * Columns per Row (R6xx/R7xx/R8xx only) 8 8 4 4
119 * Columns per Row (R9xx+) 8 4 4 4 */
120
121 switch (chip) {
122 /* FIXME: are some chips missing here? */
123 /* wavefront size 16 */
124 case CHIP_RV610:
125 case CHIP_RS780:
126 case CHIP_RV620:
127 case CHIP_RS880:
128 /* wavefront size 32 */
129 case CHIP_RV630:
130 case CHIP_RV635:
131 case CHIP_RV730:
132 case CHIP_RV710:
133 case CHIP_PALM:
134 case CHIP_CEDAR:
135 return 8;
136
137 /* wavefront size 64 */
138 default:
139 return 4;
140 }
141 }
142
r600_bytecode_init(struct r600_bytecode * bc,enum chip_class chip_class,enum radeon_family family,bool has_compressed_msaa_texturing)143 void r600_bytecode_init(struct r600_bytecode *bc,
144 enum chip_class chip_class,
145 enum radeon_family family,
146 bool has_compressed_msaa_texturing)
147 {
148 static unsigned next_shader_id = 0;
149
150 bc->debug_id = ++next_shader_id;
151
152 if ((chip_class == R600) &&
153 (family != CHIP_RV670 && family != CHIP_RS780 && family != CHIP_RS880)) {
154 bc->ar_handling = AR_HANDLE_RV6XX;
155 bc->r6xx_nop_after_rel_dst = 1;
156 } else {
157 bc->ar_handling = AR_HANDLE_NORMAL;
158 bc->r6xx_nop_after_rel_dst = 0;
159 }
160
161 LIST_INITHEAD(&bc->cf);
162 bc->chip_class = chip_class;
163 bc->family = family;
164 bc->has_compressed_msaa_texturing = has_compressed_msaa_texturing;
165 bc->stack.entry_size = stack_entry_size(family);
166 }
167
r600_bytecode_add_cf(struct r600_bytecode * bc)168 int r600_bytecode_add_cf(struct r600_bytecode *bc)
169 {
170 struct r600_bytecode_cf *cf = r600_bytecode_cf();
171
172 if (!cf)
173 return -ENOMEM;
174 LIST_ADDTAIL(&cf->list, &bc->cf);
175 if (bc->cf_last) {
176 cf->id = bc->cf_last->id + 2;
177 if (bc->cf_last->eg_alu_extended) {
178 /* take into account extended alu size */
179 cf->id += 2;
180 bc->ndw += 2;
181 }
182 }
183 bc->cf_last = cf;
184 bc->ncf++;
185 bc->ndw += 2;
186 bc->force_add_cf = 0;
187 bc->ar_loaded = 0;
188 return 0;
189 }
190
r600_bytecode_add_output(struct r600_bytecode * bc,const struct r600_bytecode_output * output)191 int r600_bytecode_add_output(struct r600_bytecode *bc,
192 const struct r600_bytecode_output *output)
193 {
194 int r;
195
196 if (output->gpr >= bc->ngpr)
197 bc->ngpr = output->gpr + 1;
198
199 if (bc->cf_last && (bc->cf_last->op == output->op ||
200 (bc->cf_last->op == CF_OP_EXPORT &&
201 output->op == CF_OP_EXPORT_DONE)) &&
202 output->type == bc->cf_last->output.type &&
203 output->elem_size == bc->cf_last->output.elem_size &&
204 output->swizzle_x == bc->cf_last->output.swizzle_x &&
205 output->swizzle_y == bc->cf_last->output.swizzle_y &&
206 output->swizzle_z == bc->cf_last->output.swizzle_z &&
207 output->swizzle_w == bc->cf_last->output.swizzle_w &&
208 output->comp_mask == bc->cf_last->output.comp_mask &&
209 (output->burst_count + bc->cf_last->output.burst_count) <= 16) {
210
211 if ((output->gpr + output->burst_count) == bc->cf_last->output.gpr &&
212 (output->array_base + output->burst_count) == bc->cf_last->output.array_base) {
213
214 bc->cf_last->op = bc->cf_last->output.op = output->op;
215 bc->cf_last->output.gpr = output->gpr;
216 bc->cf_last->output.array_base = output->array_base;
217 bc->cf_last->output.burst_count += output->burst_count;
218 return 0;
219
220 } else if (output->gpr == (bc->cf_last->output.gpr + bc->cf_last->output.burst_count) &&
221 output->array_base == (bc->cf_last->output.array_base + bc->cf_last->output.burst_count)) {
222
223 bc->cf_last->op = bc->cf_last->output.op = output->op;
224 bc->cf_last->output.burst_count += output->burst_count;
225 return 0;
226 }
227 }
228
229 r = r600_bytecode_add_cf(bc);
230 if (r)
231 return r;
232 bc->cf_last->op = output->op;
233 memcpy(&bc->cf_last->output, output, sizeof(struct r600_bytecode_output));
234 bc->cf_last->barrier = 1;
235 return 0;
236 }
237
238 /* alu instructions that can ony exits once per group */
is_alu_once_inst(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)239 static int is_alu_once_inst(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
240 {
241 return r600_isa_alu(alu->op)->flags & (AF_KILL | AF_PRED) || alu->is_lds_idx_op || alu->op == ALU_OP0_GROUP_BARRIER;
242 }
243
is_alu_reduction_inst(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)244 static int is_alu_reduction_inst(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
245 {
246 return (r600_isa_alu(alu->op)->flags & AF_REPL) &&
247 (r600_isa_alu_slots(bc->isa->hw_class, alu->op) == AF_4V);
248 }
249
is_alu_mova_inst(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)250 static int is_alu_mova_inst(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
251 {
252 return r600_isa_alu(alu->op)->flags & AF_MOVA;
253 }
254
alu_uses_rel(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)255 static int alu_uses_rel(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
256 {
257 unsigned num_src = r600_bytecode_get_num_operands(bc, alu);
258 unsigned src;
259
260 if (alu->dst.rel) {
261 return 1;
262 }
263
264 for (src = 0; src < num_src; ++src) {
265 if (alu->src[src].rel) {
266 return 1;
267 }
268 }
269 return 0;
270 }
271
is_lds_read(int sel)272 static int is_lds_read(int sel)
273 {
274 return sel == EG_V_SQ_ALU_SRC_LDS_OQ_A_POP || sel == EG_V_SQ_ALU_SRC_LDS_OQ_B_POP;
275 }
276
alu_uses_lds(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)277 static int alu_uses_lds(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
278 {
279 unsigned num_src = r600_bytecode_get_num_operands(bc, alu);
280 unsigned src;
281
282 for (src = 0; src < num_src; ++src) {
283 if (is_lds_read(alu->src[src].sel)) {
284 return 1;
285 }
286 }
287 return 0;
288 }
289
is_alu_64bit_inst(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)290 static int is_alu_64bit_inst(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
291 {
292 const struct alu_op_info *op = r600_isa_alu(alu->op);
293 return (op->flags & AF_64);
294 }
295
is_alu_vec_unit_inst(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)296 static int is_alu_vec_unit_inst(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
297 {
298 unsigned slots = r600_isa_alu_slots(bc->isa->hw_class, alu->op);
299 return !(slots & AF_S);
300 }
301
is_alu_trans_unit_inst(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)302 static int is_alu_trans_unit_inst(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
303 {
304 unsigned slots = r600_isa_alu_slots(bc->isa->hw_class, alu->op);
305 return !(slots & AF_V);
306 }
307
308 /* alu instructions that can execute on any unit */
is_alu_any_unit_inst(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)309 static int is_alu_any_unit_inst(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
310 {
311 unsigned slots = r600_isa_alu_slots(bc->isa->hw_class, alu->op);
312 return slots == AF_VS;
313 }
314
is_nop_inst(struct r600_bytecode * bc,struct r600_bytecode_alu * alu)315 static int is_nop_inst(struct r600_bytecode *bc, struct r600_bytecode_alu *alu)
316 {
317 return alu->op == ALU_OP0_NOP;
318 }
319
assign_alu_units(struct r600_bytecode * bc,struct r600_bytecode_alu * alu_first,struct r600_bytecode_alu * assignment[5])320 static int assign_alu_units(struct r600_bytecode *bc, struct r600_bytecode_alu *alu_first,
321 struct r600_bytecode_alu *assignment[5])
322 {
323 struct r600_bytecode_alu *alu;
324 unsigned i, chan, trans;
325 int max_slots = bc->chip_class == CAYMAN ? 4 : 5;
326
327 for (i = 0; i < max_slots; i++)
328 assignment[i] = NULL;
329
330 for (alu = alu_first; alu; alu = LIST_ENTRY(struct r600_bytecode_alu, alu->list.next, list)) {
331 chan = alu->dst.chan;
332 if (max_slots == 4)
333 trans = 0;
334 else if (is_alu_trans_unit_inst(bc, alu))
335 trans = 1;
336 else if (is_alu_vec_unit_inst(bc, alu))
337 trans = 0;
338 else if (assignment[chan])
339 trans = 1; /* Assume ALU_INST_PREFER_VECTOR. */
340 else
341 trans = 0;
342
343 if (trans) {
344 if (assignment[4]) {
345 assert(0); /* ALU.Trans has already been allocated. */
346 return -1;
347 }
348 assignment[4] = alu;
349 } else {
350 if (assignment[chan]) {
351 assert(0); /* ALU.chan has already been allocated. */
352 return -1;
353 }
354 assignment[chan] = alu;
355 }
356
357 if (alu->last)
358 break;
359 }
360 return 0;
361 }
362
363 struct alu_bank_swizzle {
364 int hw_gpr[NUM_OF_CYCLES][NUM_OF_COMPONENTS];
365 int hw_cfile_addr[4];
366 int hw_cfile_elem[4];
367 };
368
369 static const unsigned cycle_for_bank_swizzle_vec[][3] = {
370 [SQ_ALU_VEC_012] = { 0, 1, 2 },
371 [SQ_ALU_VEC_021] = { 0, 2, 1 },
372 [SQ_ALU_VEC_120] = { 1, 2, 0 },
373 [SQ_ALU_VEC_102] = { 1, 0, 2 },
374 [SQ_ALU_VEC_201] = { 2, 0, 1 },
375 [SQ_ALU_VEC_210] = { 2, 1, 0 }
376 };
377
378 static const unsigned cycle_for_bank_swizzle_scl[][3] = {
379 [SQ_ALU_SCL_210] = { 2, 1, 0 },
380 [SQ_ALU_SCL_122] = { 1, 2, 2 },
381 [SQ_ALU_SCL_212] = { 2, 1, 2 },
382 [SQ_ALU_SCL_221] = { 2, 2, 1 }
383 };
384
init_bank_swizzle(struct alu_bank_swizzle * bs)385 static void init_bank_swizzle(struct alu_bank_swizzle *bs)
386 {
387 int i, cycle, component;
388 /* set up gpr use */
389 for (cycle = 0; cycle < NUM_OF_CYCLES; cycle++)
390 for (component = 0; component < NUM_OF_COMPONENTS; component++)
391 bs->hw_gpr[cycle][component] = -1;
392 for (i = 0; i < 4; i++)
393 bs->hw_cfile_addr[i] = -1;
394 for (i = 0; i < 4; i++)
395 bs->hw_cfile_elem[i] = -1;
396 }
397
reserve_gpr(struct alu_bank_swizzle * bs,unsigned sel,unsigned chan,unsigned cycle)398 static int reserve_gpr(struct alu_bank_swizzle *bs, unsigned sel, unsigned chan, unsigned cycle)
399 {
400 if (bs->hw_gpr[cycle][chan] == -1)
401 bs->hw_gpr[cycle][chan] = sel;
402 else if (bs->hw_gpr[cycle][chan] != (int)sel) {
403 /* Another scalar operation has already used the GPR read port for the channel. */
404 return -1;
405 }
406 return 0;
407 }
408
reserve_cfile(struct r600_bytecode * bc,struct alu_bank_swizzle * bs,unsigned sel,unsigned chan)409 static int reserve_cfile(struct r600_bytecode *bc, struct alu_bank_swizzle *bs, unsigned sel, unsigned chan)
410 {
411 int res, num_res = 4;
412 if (bc->chip_class >= R700) {
413 num_res = 2;
414 chan /= 2;
415 }
416 for (res = 0; res < num_res; ++res) {
417 if (bs->hw_cfile_addr[res] == -1) {
418 bs->hw_cfile_addr[res] = sel;
419 bs->hw_cfile_elem[res] = chan;
420 return 0;
421 } else if (bs->hw_cfile_addr[res] == sel &&
422 bs->hw_cfile_elem[res] == chan)
423 return 0; /* Read for this scalar element already reserved, nothing to do here. */
424 }
425 /* All cfile read ports are used, cannot reference vector element. */
426 return -1;
427 }
428
is_gpr(unsigned sel)429 static int is_gpr(unsigned sel)
430 {
431 return (sel <= 127);
432 }
433
434 /* CB constants start at 512, and get translated to a kcache index when ALU
435 * clauses are constructed. Note that we handle kcache constants the same way
436 * as (the now gone) cfile constants, is that really required? */
is_cfile(unsigned sel)437 static int is_cfile(unsigned sel)
438 {
439 return (sel > 255 && sel < 512) ||
440 (sel > 511 && sel < 4607) || /* Kcache before translation. */
441 (sel > 127 && sel < 192); /* Kcache after translation. */
442 }
443
is_const(int sel)444 static int is_const(int sel)
445 {
446 return is_cfile(sel) ||
447 (sel >= V_SQ_ALU_SRC_0 &&
448 sel <= V_SQ_ALU_SRC_LITERAL);
449 }
450
check_vector(struct r600_bytecode * bc,struct r600_bytecode_alu * alu,struct alu_bank_swizzle * bs,int bank_swizzle)451 static int check_vector(struct r600_bytecode *bc, struct r600_bytecode_alu *alu,
452 struct alu_bank_swizzle *bs, int bank_swizzle)
453 {
454 int r, src, num_src, sel, elem, cycle;
455
456 num_src = r600_bytecode_get_num_operands(bc, alu);
457 for (src = 0; src < num_src; src++) {
458 sel = alu->src[src].sel;
459 elem = alu->src[src].chan;
460 if (is_gpr(sel)) {
461 cycle = cycle_for_bank_swizzle_vec[bank_swizzle][src];
462 if (src == 1 && sel == alu->src[0].sel && elem == alu->src[0].chan)
463 /* Nothing to do; special-case optimization,
464 * second source uses first source’s reservation. */
465 continue;
466 else {
467 r = reserve_gpr(bs, sel, elem, cycle);
468 if (r)
469 return r;
470 }
471 } else if (is_cfile(sel)) {
472 r = reserve_cfile(bc, bs, (alu->src[src].kc_bank<<16) + sel, elem);
473 if (r)
474 return r;
475 }
476 /* No restrictions on PV, PS, literal or special constants. */
477 }
478 return 0;
479 }
480
check_scalar(struct r600_bytecode * bc,struct r600_bytecode_alu * alu,struct alu_bank_swizzle * bs,int bank_swizzle)481 static int check_scalar(struct r600_bytecode *bc, struct r600_bytecode_alu *alu,
482 struct alu_bank_swizzle *bs, int bank_swizzle)
483 {
484 int r, src, num_src, const_count, sel, elem, cycle;
485
486 num_src = r600_bytecode_get_num_operands(bc, alu);
487 for (const_count = 0, src = 0; src < num_src; ++src) {
488 sel = alu->src[src].sel;
489 elem = alu->src[src].chan;
490 if (is_const(sel)) { /* Any constant, including literal and inline constants. */
491 if (const_count >= 2)
492 /* More than two references to a constant in
493 * transcendental operation. */
494 return -1;
495 else
496 const_count++;
497 }
498 if (is_cfile(sel)) {
499 r = reserve_cfile(bc, bs, (alu->src[src].kc_bank<<16) + sel, elem);
500 if (r)
501 return r;
502 }
503 }
504 for (src = 0; src < num_src; ++src) {
505 sel = alu->src[src].sel;
506 elem = alu->src[src].chan;
507 if (is_gpr(sel)) {
508 cycle = cycle_for_bank_swizzle_scl[bank_swizzle][src];
509 if (cycle < const_count)
510 /* Cycle for GPR load conflicts with
511 * constant load in transcendental operation. */
512 return -1;
513 r = reserve_gpr(bs, sel, elem, cycle);
514 if (r)
515 return r;
516 }
517 /* PV PS restrictions */
518 if (const_count && (sel == 254 || sel == 255)) {
519 cycle = cycle_for_bank_swizzle_scl[bank_swizzle][src];
520 if (cycle < const_count)
521 return -1;
522 }
523 }
524 return 0;
525 }
526
check_and_set_bank_swizzle(struct r600_bytecode * bc,struct r600_bytecode_alu * slots[5])527 static int check_and_set_bank_swizzle(struct r600_bytecode *bc,
528 struct r600_bytecode_alu *slots[5])
529 {
530 struct alu_bank_swizzle bs;
531 int bank_swizzle[5];
532 int i, r = 0, forced = 1;
533 boolean scalar_only = bc->chip_class == CAYMAN ? false : true;
534 int max_slots = bc->chip_class == CAYMAN ? 4 : 5;
535
536 for (i = 0; i < max_slots; i++) {
537 if (slots[i]) {
538 if (slots[i]->bank_swizzle_force) {
539 slots[i]->bank_swizzle = slots[i]->bank_swizzle_force;
540 } else {
541 forced = 0;
542 }
543 }
544
545 if (i < 4 && slots[i])
546 scalar_only = false;
547 }
548 if (forced)
549 return 0;
550
551 /* Just check every possible combination of bank swizzle.
552 * Not very efficent, but works on the first try in most of the cases. */
553 for (i = 0; i < 4; i++)
554 if (!slots[i] || !slots[i]->bank_swizzle_force)
555 bank_swizzle[i] = SQ_ALU_VEC_012;
556 else
557 bank_swizzle[i] = slots[i]->bank_swizzle;
558
559 bank_swizzle[4] = SQ_ALU_SCL_210;
560 while(bank_swizzle[4] <= SQ_ALU_SCL_221) {
561
562 init_bank_swizzle(&bs);
563 if (scalar_only == false) {
564 for (i = 0; i < 4; i++) {
565 if (slots[i]) {
566 r = check_vector(bc, slots[i], &bs, bank_swizzle[i]);
567 if (r)
568 break;
569 }
570 }
571 } else
572 r = 0;
573
574 if (!r && max_slots == 5 && slots[4]) {
575 r = check_scalar(bc, slots[4], &bs, bank_swizzle[4]);
576 }
577 if (!r) {
578 for (i = 0; i < max_slots; i++) {
579 if (slots[i])
580 slots[i]->bank_swizzle = bank_swizzle[i];
581 }
582 return 0;
583 }
584
585 if (scalar_only) {
586 bank_swizzle[4]++;
587 } else {
588 for (i = 0; i < max_slots; i++) {
589 if (!slots[i] || !slots[i]->bank_swizzle_force) {
590 bank_swizzle[i]++;
591 if (bank_swizzle[i] <= SQ_ALU_VEC_210)
592 break;
593 else if (i < max_slots - 1)
594 bank_swizzle[i] = SQ_ALU_VEC_012;
595 else
596 return -1;
597 }
598 }
599 }
600 }
601
602 /* Couldn't find a working swizzle. */
603 return -1;
604 }
605
replace_gpr_with_pv_ps(struct r600_bytecode * bc,struct r600_bytecode_alu * slots[5],struct r600_bytecode_alu * alu_prev)606 static int replace_gpr_with_pv_ps(struct r600_bytecode *bc,
607 struct r600_bytecode_alu *slots[5], struct r600_bytecode_alu *alu_prev)
608 {
609 struct r600_bytecode_alu *prev[5];
610 int gpr[5], chan[5];
611 int i, j, r, src, num_src;
612 int max_slots = bc->chip_class == CAYMAN ? 4 : 5;
613
614 r = assign_alu_units(bc, alu_prev, prev);
615 if (r)
616 return r;
617
618 for (i = 0; i < max_slots; ++i) {
619 if (prev[i] && alu_writes(prev[i]) && !prev[i]->dst.rel) {
620
621 if (is_alu_64bit_inst(bc, prev[i])) {
622 gpr[i] = -1;
623 continue;
624 }
625
626 gpr[i] = prev[i]->dst.sel;
627 /* cube writes more than PV.X */
628 if (is_alu_reduction_inst(bc, prev[i]))
629 chan[i] = 0;
630 else
631 chan[i] = prev[i]->dst.chan;
632 } else
633 gpr[i] = -1;
634 }
635
636 for (i = 0; i < max_slots; ++i) {
637 struct r600_bytecode_alu *alu = slots[i];
638 if (!alu)
639 continue;
640
641 if (is_alu_64bit_inst(bc, alu))
642 continue;
643 num_src = r600_bytecode_get_num_operands(bc, alu);
644 for (src = 0; src < num_src; ++src) {
645 if (!is_gpr(alu->src[src].sel) || alu->src[src].rel)
646 continue;
647
648 if (bc->chip_class < CAYMAN) {
649 if (alu->src[src].sel == gpr[4] &&
650 alu->src[src].chan == chan[4] &&
651 alu_prev->pred_sel == alu->pred_sel) {
652 alu->src[src].sel = V_SQ_ALU_SRC_PS;
653 alu->src[src].chan = 0;
654 continue;
655 }
656 }
657
658 for (j = 0; j < 4; ++j) {
659 if (alu->src[src].sel == gpr[j] &&
660 alu->src[src].chan == j &&
661 alu_prev->pred_sel == alu->pred_sel) {
662 alu->src[src].sel = V_SQ_ALU_SRC_PV;
663 alu->src[src].chan = chan[j];
664 break;
665 }
666 }
667 }
668 }
669
670 return 0;
671 }
672
r600_bytecode_special_constants(uint32_t value,unsigned * sel,unsigned * neg,unsigned abs)673 void r600_bytecode_special_constants(uint32_t value, unsigned *sel, unsigned *neg, unsigned abs)
674 {
675 switch(value) {
676 case 0:
677 *sel = V_SQ_ALU_SRC_0;
678 break;
679 case 1:
680 *sel = V_SQ_ALU_SRC_1_INT;
681 break;
682 case -1:
683 *sel = V_SQ_ALU_SRC_M_1_INT;
684 break;
685 case 0x3F800000: /* 1.0f */
686 *sel = V_SQ_ALU_SRC_1;
687 break;
688 case 0x3F000000: /* 0.5f */
689 *sel = V_SQ_ALU_SRC_0_5;
690 break;
691 case 0xBF800000: /* -1.0f */
692 *sel = V_SQ_ALU_SRC_1;
693 *neg ^= !abs;
694 break;
695 case 0xBF000000: /* -0.5f */
696 *sel = V_SQ_ALU_SRC_0_5;
697 *neg ^= !abs;
698 break;
699 default:
700 *sel = V_SQ_ALU_SRC_LITERAL;
701 break;
702 }
703 }
704
705 /* compute how many literal are needed */
r600_bytecode_alu_nliterals(struct r600_bytecode * bc,struct r600_bytecode_alu * alu,uint32_t literal[4],unsigned * nliteral)706 static int r600_bytecode_alu_nliterals(struct r600_bytecode *bc, struct r600_bytecode_alu *alu,
707 uint32_t literal[4], unsigned *nliteral)
708 {
709 unsigned num_src = r600_bytecode_get_num_operands(bc, alu);
710 unsigned i, j;
711
712 for (i = 0; i < num_src; ++i) {
713 if (alu->src[i].sel == V_SQ_ALU_SRC_LITERAL) {
714 uint32_t value = alu->src[i].value;
715 unsigned found = 0;
716 for (j = 0; j < *nliteral; ++j) {
717 if (literal[j] == value) {
718 found = 1;
719 break;
720 }
721 }
722 if (!found) {
723 if (*nliteral >= 4)
724 return -EINVAL;
725 literal[(*nliteral)++] = value;
726 }
727 }
728 }
729 return 0;
730 }
731
r600_bytecode_alu_adjust_literals(struct r600_bytecode * bc,struct r600_bytecode_alu * alu,uint32_t literal[4],unsigned nliteral)732 static void r600_bytecode_alu_adjust_literals(struct r600_bytecode *bc,
733 struct r600_bytecode_alu *alu,
734 uint32_t literal[4], unsigned nliteral)
735 {
736 unsigned num_src = r600_bytecode_get_num_operands(bc, alu);
737 unsigned i, j;
738
739 for (i = 0; i < num_src; ++i) {
740 if (alu->src[i].sel == V_SQ_ALU_SRC_LITERAL) {
741 uint32_t value = alu->src[i].value;
742 for (j = 0; j < nliteral; ++j) {
743 if (literal[j] == value) {
744 alu->src[i].chan = j;
745 break;
746 }
747 }
748 }
749 }
750 }
751
merge_inst_groups(struct r600_bytecode * bc,struct r600_bytecode_alu * slots[5],struct r600_bytecode_alu * alu_prev)752 static int merge_inst_groups(struct r600_bytecode *bc, struct r600_bytecode_alu *slots[5],
753 struct r600_bytecode_alu *alu_prev)
754 {
755 struct r600_bytecode_alu *prev[5];
756 struct r600_bytecode_alu *result[5] = { NULL };
757
758 uint32_t literal[4], prev_literal[4];
759 unsigned nliteral = 0, prev_nliteral = 0;
760
761 int i, j, r, src, num_src;
762 int num_once_inst = 0;
763 int have_mova = 0, have_rel = 0;
764 int max_slots = bc->chip_class == CAYMAN ? 4 : 5;
765
766 r = assign_alu_units(bc, alu_prev, prev);
767 if (r)
768 return r;
769
770 for (i = 0; i < max_slots; ++i) {
771 if (prev[i]) {
772 if (prev[i]->pred_sel)
773 return 0;
774 if (is_alu_once_inst(bc, prev[i]))
775 return 0;
776 }
777 if (slots[i]) {
778 if (slots[i]->pred_sel)
779 return 0;
780 if (is_alu_once_inst(bc, slots[i]))
781 return 0;
782 }
783 }
784
785 for (i = 0; i < max_slots; ++i) {
786 struct r600_bytecode_alu *alu;
787
788 if (num_once_inst > 0)
789 return 0;
790
791 /* check number of literals */
792 if (prev[i]) {
793 if (r600_bytecode_alu_nliterals(bc, prev[i], literal, &nliteral))
794 return 0;
795 if (r600_bytecode_alu_nliterals(bc, prev[i], prev_literal, &prev_nliteral))
796 return 0;
797 if (is_alu_mova_inst(bc, prev[i])) {
798 if (have_rel)
799 return 0;
800 have_mova = 1;
801 }
802
803 if (alu_uses_rel(bc, prev[i])) {
804 if (have_mova) {
805 return 0;
806 }
807 have_rel = 1;
808 }
809 if (alu_uses_lds(bc, prev[i]))
810 return 0;
811
812 num_once_inst += is_alu_once_inst(bc, prev[i]);
813 }
814 if (slots[i] && r600_bytecode_alu_nliterals(bc, slots[i], literal, &nliteral))
815 return 0;
816
817 /* Let's check used slots. */
818 if (prev[i] && !slots[i]) {
819 result[i] = prev[i];
820 continue;
821 } else if (prev[i] && slots[i]) {
822 if (max_slots == 5 && result[4] == NULL && prev[4] == NULL && slots[4] == NULL) {
823 /* Trans unit is still free try to use it. */
824 if (is_alu_any_unit_inst(bc, slots[i]) && !alu_uses_lds(bc, slots[i])) {
825 result[i] = prev[i];
826 result[4] = slots[i];
827 } else if (is_alu_any_unit_inst(bc, prev[i])) {
828 if (slots[i]->dst.sel == prev[i]->dst.sel &&
829 alu_writes(slots[i]) &&
830 alu_writes(prev[i]))
831 return 0;
832
833 result[i] = slots[i];
834 result[4] = prev[i];
835 } else
836 return 0;
837 } else
838 return 0;
839 } else if(!slots[i]) {
840 continue;
841 } else {
842 if (max_slots == 5 && slots[i] && prev[4] &&
843 slots[i]->dst.sel == prev[4]->dst.sel &&
844 slots[i]->dst.chan == prev[4]->dst.chan &&
845 alu_writes(slots[i]) &&
846 alu_writes(prev[4]))
847 return 0;
848
849 result[i] = slots[i];
850 }
851
852 alu = slots[i];
853 num_once_inst += is_alu_once_inst(bc, alu);
854
855 /* don't reschedule NOPs */
856 if (is_nop_inst(bc, alu))
857 return 0;
858
859 if (is_alu_mova_inst(bc, alu)) {
860 if (have_rel) {
861 return 0;
862 }
863 have_mova = 1;
864 }
865
866 if (alu_uses_rel(bc, alu)) {
867 if (have_mova) {
868 return 0;
869 }
870 have_rel = 1;
871 }
872
873 if (alu->op == ALU_OP0_SET_CF_IDX0 ||
874 alu->op == ALU_OP0_SET_CF_IDX1)
875 return 0; /* data hazard with MOVA */
876
877 /* Let's check source gprs */
878 num_src = r600_bytecode_get_num_operands(bc, alu);
879 for (src = 0; src < num_src; ++src) {
880
881 /* Constants don't matter. */
882 if (!is_gpr(alu->src[src].sel))
883 continue;
884
885 for (j = 0; j < max_slots; ++j) {
886 if (!prev[j] || !alu_writes(prev[j]))
887 continue;
888
889 /* If it's relative then we can't determin which gpr is really used. */
890 if (prev[j]->dst.chan == alu->src[src].chan &&
891 (prev[j]->dst.sel == alu->src[src].sel ||
892 prev[j]->dst.rel || alu->src[src].rel))
893 return 0;
894 }
895 }
896 }
897
898 /* more than one PRED_ or KILL_ ? */
899 if (num_once_inst > 1)
900 return 0;
901
902 /* check if the result can still be swizzlet */
903 r = check_and_set_bank_swizzle(bc, result);
904 if (r)
905 return 0;
906
907 /* looks like everything worked out right, apply the changes */
908
909 /* undo adding previus literals */
910 bc->cf_last->ndw -= align(prev_nliteral, 2);
911
912 /* sort instructions */
913 for (i = 0; i < max_slots; ++i) {
914 slots[i] = result[i];
915 if (result[i]) {
916 LIST_DEL(&result[i]->list);
917 result[i]->last = 0;
918 LIST_ADDTAIL(&result[i]->list, &bc->cf_last->alu);
919 }
920 }
921
922 /* determine new last instruction */
923 LIST_ENTRY(struct r600_bytecode_alu, bc->cf_last->alu.prev, list)->last = 1;
924
925 /* determine new first instruction */
926 for (i = 0; i < max_slots; ++i) {
927 if (result[i]) {
928 bc->cf_last->curr_bs_head = result[i];
929 break;
930 }
931 }
932
933 bc->cf_last->prev_bs_head = bc->cf_last->prev2_bs_head;
934 bc->cf_last->prev2_bs_head = NULL;
935
936 return 0;
937 }
938
939 /* we'll keep kcache sets sorted by bank & addr */
r600_bytecode_alloc_kcache_line(struct r600_bytecode * bc,struct r600_bytecode_kcache * kcache,unsigned bank,unsigned line,unsigned index_mode)940 static int r600_bytecode_alloc_kcache_line(struct r600_bytecode *bc,
941 struct r600_bytecode_kcache *kcache,
942 unsigned bank, unsigned line, unsigned index_mode)
943 {
944 int i, kcache_banks = bc->chip_class >= EVERGREEN ? 4 : 2;
945
946 for (i = 0; i < kcache_banks; i++) {
947 if (kcache[i].mode) {
948 int d;
949
950 if (kcache[i].bank < bank)
951 continue;
952
953 if ((kcache[i].bank == bank && kcache[i].addr > line+1) ||
954 kcache[i].bank > bank) {
955 /* try to insert new line */
956 if (kcache[kcache_banks-1].mode) {
957 /* all sets are in use */
958 return -ENOMEM;
959 }
960
961 memmove(&kcache[i+1],&kcache[i], (kcache_banks-i-1)*sizeof(struct r600_bytecode_kcache));
962 kcache[i].mode = V_SQ_CF_KCACHE_LOCK_1;
963 kcache[i].bank = bank;
964 kcache[i].addr = line;
965 kcache[i].index_mode = index_mode;
966 return 0;
967 }
968
969 d = line - kcache[i].addr;
970
971 if (d == -1) {
972 kcache[i].addr--;
973 if (kcache[i].mode == V_SQ_CF_KCACHE_LOCK_2) {
974 /* we are prepending the line to the current set,
975 * discarding the existing second line,
976 * so we'll have to insert line+2 after it */
977 line += 2;
978 continue;
979 } else if (kcache[i].mode == V_SQ_CF_KCACHE_LOCK_1) {
980 kcache[i].mode = V_SQ_CF_KCACHE_LOCK_2;
981 return 0;
982 } else {
983 /* V_SQ_CF_KCACHE_LOCK_LOOP_INDEX is not supported */
984 return -ENOMEM;
985 }
986 } else if (d == 1) {
987 kcache[i].mode = V_SQ_CF_KCACHE_LOCK_2;
988 return 0;
989 } else if (d == 0)
990 return 0;
991 } else { /* free kcache set - use it */
992 kcache[i].mode = V_SQ_CF_KCACHE_LOCK_1;
993 kcache[i].bank = bank;
994 kcache[i].addr = line;
995 kcache[i].index_mode = index_mode;
996 return 0;
997 }
998 }
999 return -ENOMEM;
1000 }
1001
r600_bytecode_alloc_inst_kcache_lines(struct r600_bytecode * bc,struct r600_bytecode_kcache * kcache,struct r600_bytecode_alu * alu)1002 static int r600_bytecode_alloc_inst_kcache_lines(struct r600_bytecode *bc,
1003 struct r600_bytecode_kcache *kcache,
1004 struct r600_bytecode_alu *alu)
1005 {
1006 int i, r;
1007
1008 for (i = 0; i < 3; i++) {
1009 unsigned bank, line, sel = alu->src[i].sel, index_mode;
1010
1011 if (sel < 512)
1012 continue;
1013
1014 bank = alu->src[i].kc_bank;
1015 line = (sel-512)>>4;
1016 index_mode = alu->src[i].kc_rel ? 1 : 0; // V_SQ_CF_INDEX_0 / V_SQ_CF_INDEX_NONE
1017
1018 if ((r = r600_bytecode_alloc_kcache_line(bc, kcache, bank, line, index_mode)))
1019 return r;
1020 }
1021 return 0;
1022 }
1023
r600_bytecode_assign_kcache_banks(struct r600_bytecode * bc,struct r600_bytecode_alu * alu,struct r600_bytecode_kcache * kcache)1024 static int r600_bytecode_assign_kcache_banks(struct r600_bytecode *bc,
1025 struct r600_bytecode_alu *alu,
1026 struct r600_bytecode_kcache * kcache)
1027 {
1028 int i, j;
1029
1030 /* Alter the src operands to refer to the kcache. */
1031 for (i = 0; i < 3; ++i) {
1032 static const unsigned int base[] = {128, 160, 256, 288};
1033 unsigned int line, sel = alu->src[i].sel, found = 0;
1034
1035 if (sel < 512)
1036 continue;
1037
1038 sel -= 512;
1039 line = sel>>4;
1040
1041 for (j = 0; j < 4 && !found; ++j) {
1042 switch (kcache[j].mode) {
1043 case V_SQ_CF_KCACHE_NOP:
1044 case V_SQ_CF_KCACHE_LOCK_LOOP_INDEX:
1045 R600_ERR("unexpected kcache line mode\n");
1046 return -ENOMEM;
1047 default:
1048 if (kcache[j].bank == alu->src[i].kc_bank &&
1049 kcache[j].addr <= line &&
1050 line < kcache[j].addr + kcache[j].mode) {
1051 alu->src[i].sel = sel - (kcache[j].addr<<4);
1052 alu->src[i].sel += base[j];
1053 found=1;
1054 }
1055 }
1056 }
1057 }
1058 return 0;
1059 }
1060
r600_bytecode_alloc_kcache_lines(struct r600_bytecode * bc,struct r600_bytecode_alu * alu,unsigned type)1061 static int r600_bytecode_alloc_kcache_lines(struct r600_bytecode *bc,
1062 struct r600_bytecode_alu *alu,
1063 unsigned type)
1064 {
1065 struct r600_bytecode_kcache kcache_sets[4];
1066 struct r600_bytecode_kcache *kcache = kcache_sets;
1067 int r;
1068
1069 memcpy(kcache, bc->cf_last->kcache, 4 * sizeof(struct r600_bytecode_kcache));
1070
1071 if ((r = r600_bytecode_alloc_inst_kcache_lines(bc, kcache, alu))) {
1072 /* can't alloc, need to start new clause */
1073 if ((r = r600_bytecode_add_cf(bc))) {
1074 return r;
1075 }
1076 bc->cf_last->op = type;
1077
1078 /* retry with the new clause */
1079 kcache = bc->cf_last->kcache;
1080 if ((r = r600_bytecode_alloc_inst_kcache_lines(bc, kcache, alu))) {
1081 /* can't alloc again- should never happen */
1082 return r;
1083 }
1084 } else {
1085 /* update kcache sets */
1086 memcpy(bc->cf_last->kcache, kcache, 4 * sizeof(struct r600_bytecode_kcache));
1087 }
1088
1089 /* if we actually used more than 2 kcache sets, or have relative indexing - use ALU_EXTENDED on eg+ */
1090 if (kcache[2].mode != V_SQ_CF_KCACHE_NOP ||
1091 kcache[0].index_mode || kcache[1].index_mode || kcache[2].index_mode || kcache[3].index_mode) {
1092 if (bc->chip_class < EVERGREEN)
1093 return -ENOMEM;
1094 bc->cf_last->eg_alu_extended = 1;
1095 }
1096
1097 return 0;
1098 }
1099
insert_nop_r6xx(struct r600_bytecode * bc)1100 static int insert_nop_r6xx(struct r600_bytecode *bc)
1101 {
1102 struct r600_bytecode_alu alu;
1103 int r, i;
1104
1105 for (i = 0; i < 4; i++) {
1106 memset(&alu, 0, sizeof(alu));
1107 alu.op = ALU_OP0_NOP;
1108 alu.src[0].chan = i;
1109 alu.dst.chan = i;
1110 alu.last = (i == 3);
1111 r = r600_bytecode_add_alu(bc, &alu);
1112 if (r)
1113 return r;
1114 }
1115 return 0;
1116 }
1117
1118 /* load AR register from gpr (bc->ar_reg) with MOVA_INT */
load_ar_r6xx(struct r600_bytecode * bc)1119 static int load_ar_r6xx(struct r600_bytecode *bc)
1120 {
1121 struct r600_bytecode_alu alu;
1122 int r;
1123
1124 if (bc->ar_loaded)
1125 return 0;
1126
1127 /* hack to avoid making MOVA the last instruction in the clause */
1128 if ((bc->cf_last->ndw>>1) >= 110)
1129 bc->force_add_cf = 1;
1130
1131 memset(&alu, 0, sizeof(alu));
1132 alu.op = ALU_OP1_MOVA_GPR_INT;
1133 alu.src[0].sel = bc->ar_reg;
1134 alu.src[0].chan = bc->ar_chan;
1135 alu.last = 1;
1136 alu.index_mode = INDEX_MODE_LOOP;
1137 r = r600_bytecode_add_alu(bc, &alu);
1138 if (r)
1139 return r;
1140
1141 /* no requirement to set uses waterfall on MOVA_GPR_INT */
1142 bc->ar_loaded = 1;
1143 return 0;
1144 }
1145
1146 /* load AR register from gpr (bc->ar_reg) with MOVA_INT */
load_ar(struct r600_bytecode * bc)1147 static int load_ar(struct r600_bytecode *bc)
1148 {
1149 struct r600_bytecode_alu alu;
1150 int r;
1151
1152 if (bc->ar_handling)
1153 return load_ar_r6xx(bc);
1154
1155 if (bc->ar_loaded)
1156 return 0;
1157
1158 /* hack to avoid making MOVA the last instruction in the clause */
1159 if ((bc->cf_last->ndw>>1) >= 110)
1160 bc->force_add_cf = 1;
1161
1162 memset(&alu, 0, sizeof(alu));
1163 alu.op = ALU_OP1_MOVA_INT;
1164 alu.src[0].sel = bc->ar_reg;
1165 alu.src[0].chan = bc->ar_chan;
1166 alu.last = 1;
1167 r = r600_bytecode_add_alu(bc, &alu);
1168 if (r)
1169 return r;
1170
1171 bc->cf_last->r6xx_uses_waterfall = 1;
1172 bc->ar_loaded = 1;
1173 return 0;
1174 }
1175
r600_bytecode_add_alu_type(struct r600_bytecode * bc,const struct r600_bytecode_alu * alu,unsigned type)1176 int r600_bytecode_add_alu_type(struct r600_bytecode *bc,
1177 const struct r600_bytecode_alu *alu, unsigned type)
1178 {
1179 struct r600_bytecode_alu *nalu = r600_bytecode_alu();
1180 struct r600_bytecode_alu *lalu;
1181 int i, r;
1182
1183 if (!nalu)
1184 return -ENOMEM;
1185 memcpy(nalu, alu, sizeof(struct r600_bytecode_alu));
1186
1187 if (alu->is_op3) {
1188 /* will fail later since alu does not support it. */
1189 assert(!alu->src[0].abs && !alu->src[1].abs && !alu->src[2].abs);
1190 }
1191
1192 if (bc->cf_last != NULL && bc->cf_last->op != type) {
1193 /* check if we could add it anyway */
1194 if (bc->cf_last->op == CF_OP_ALU &&
1195 type == CF_OP_ALU_PUSH_BEFORE) {
1196 LIST_FOR_EACH_ENTRY(lalu, &bc->cf_last->alu, list) {
1197 if (lalu->execute_mask) {
1198 bc->force_add_cf = 1;
1199 break;
1200 }
1201 }
1202 } else
1203 bc->force_add_cf = 1;
1204 }
1205
1206 /* cf can contains only alu or only vtx or only tex */
1207 if (bc->cf_last == NULL || bc->force_add_cf) {
1208 r = r600_bytecode_add_cf(bc);
1209 if (r) {
1210 free(nalu);
1211 return r;
1212 }
1213 }
1214 bc->cf_last->op = type;
1215
1216 /* Load index register if required */
1217 if (bc->chip_class >= EVERGREEN) {
1218 for (i = 0; i < 3; i++)
1219 if (nalu->src[i].kc_bank && nalu->src[i].kc_rel)
1220 egcm_load_index_reg(bc, 0, true);
1221 }
1222
1223 /* Check AR usage and load it if required */
1224 for (i = 0; i < 3; i++)
1225 if (nalu->src[i].rel && !bc->ar_loaded)
1226 load_ar(bc);
1227
1228 if (nalu->dst.rel && !bc->ar_loaded)
1229 load_ar(bc);
1230
1231 /* Setup the kcache for this ALU instruction. This will start a new
1232 * ALU clause if needed. */
1233 if ((r = r600_bytecode_alloc_kcache_lines(bc, nalu, type))) {
1234 free(nalu);
1235 return r;
1236 }
1237
1238 if (!bc->cf_last->curr_bs_head) {
1239 bc->cf_last->curr_bs_head = nalu;
1240 }
1241 /* number of gpr == the last gpr used in any alu */
1242 for (i = 0; i < 3; i++) {
1243 if (nalu->src[i].sel >= bc->ngpr && nalu->src[i].sel < 128) {
1244 bc->ngpr = nalu->src[i].sel + 1;
1245 }
1246 if (nalu->src[i].sel == V_SQ_ALU_SRC_LITERAL)
1247 r600_bytecode_special_constants(nalu->src[i].value,
1248 &nalu->src[i].sel, &nalu->src[i].neg, nalu->src[i].abs);
1249 }
1250 if (nalu->dst.sel >= bc->ngpr) {
1251 bc->ngpr = nalu->dst.sel + 1;
1252 }
1253 LIST_ADDTAIL(&nalu->list, &bc->cf_last->alu);
1254 /* each alu use 2 dwords */
1255 bc->cf_last->ndw += 2;
1256 bc->ndw += 2;
1257
1258 /* process cur ALU instructions for bank swizzle */
1259 if (nalu->last) {
1260 uint32_t literal[4];
1261 unsigned nliteral;
1262 struct r600_bytecode_alu *slots[5];
1263 int max_slots = bc->chip_class == CAYMAN ? 4 : 5;
1264 r = assign_alu_units(bc, bc->cf_last->curr_bs_head, slots);
1265 if (r)
1266 return r;
1267
1268 if (bc->cf_last->prev_bs_head) {
1269 r = merge_inst_groups(bc, slots, bc->cf_last->prev_bs_head);
1270 if (r)
1271 return r;
1272 }
1273
1274 if (bc->cf_last->prev_bs_head) {
1275 r = replace_gpr_with_pv_ps(bc, slots, bc->cf_last->prev_bs_head);
1276 if (r)
1277 return r;
1278 }
1279
1280 r = check_and_set_bank_swizzle(bc, slots);
1281 if (r)
1282 return r;
1283
1284 for (i = 0, nliteral = 0; i < max_slots; i++) {
1285 if (slots[i]) {
1286 r = r600_bytecode_alu_nliterals(bc, slots[i], literal, &nliteral);
1287 if (r)
1288 return r;
1289 }
1290 }
1291 bc->cf_last->ndw += align(nliteral, 2);
1292
1293 /* at most 128 slots, one add alu can add 5 slots + 4 constants(2 slots)
1294 * worst case */
1295 if ((bc->cf_last->ndw >> 1) >= 120) {
1296 bc->force_add_cf = 1;
1297 }
1298
1299 bc->cf_last->prev2_bs_head = bc->cf_last->prev_bs_head;
1300 bc->cf_last->prev_bs_head = bc->cf_last->curr_bs_head;
1301 bc->cf_last->curr_bs_head = NULL;
1302 }
1303
1304 if (nalu->dst.rel && bc->r6xx_nop_after_rel_dst)
1305 insert_nop_r6xx(bc);
1306
1307 return 0;
1308 }
1309
r600_bytecode_add_alu(struct r600_bytecode * bc,const struct r600_bytecode_alu * alu)1310 int r600_bytecode_add_alu(struct r600_bytecode *bc, const struct r600_bytecode_alu *alu)
1311 {
1312 return r600_bytecode_add_alu_type(bc, alu, CF_OP_ALU);
1313 }
1314
r600_bytecode_num_tex_and_vtx_instructions(const struct r600_bytecode * bc)1315 static unsigned r600_bytecode_num_tex_and_vtx_instructions(const struct r600_bytecode *bc)
1316 {
1317 switch (bc->chip_class) {
1318 case R600:
1319 return 8;
1320
1321 case R700:
1322 case EVERGREEN:
1323 case CAYMAN:
1324 return 16;
1325
1326 default:
1327 R600_ERR("Unknown chip class %d.\n", bc->chip_class);
1328 return 8;
1329 }
1330 }
1331
last_inst_was_not_vtx_fetch(struct r600_bytecode * bc)1332 static inline boolean last_inst_was_not_vtx_fetch(struct r600_bytecode *bc)
1333 {
1334 return !((r600_isa_cf(bc->cf_last->op)->flags & CF_FETCH) &&
1335 (bc->chip_class == CAYMAN ||
1336 bc->cf_last->op != CF_OP_TEX));
1337 }
1338
r600_bytecode_add_vtx(struct r600_bytecode * bc,const struct r600_bytecode_vtx * vtx)1339 int r600_bytecode_add_vtx(struct r600_bytecode *bc, const struct r600_bytecode_vtx *vtx)
1340 {
1341 struct r600_bytecode_vtx *nvtx = r600_bytecode_vtx();
1342 int r;
1343
1344 if (!nvtx)
1345 return -ENOMEM;
1346 memcpy(nvtx, vtx, sizeof(struct r600_bytecode_vtx));
1347
1348 /* Load index register if required */
1349 if (bc->chip_class >= EVERGREEN) {
1350 if (vtx->buffer_index_mode)
1351 egcm_load_index_reg(bc, 0, false);
1352 }
1353
1354 /* cf can contains only alu or only vtx or only tex */
1355 if (bc->cf_last == NULL ||
1356 last_inst_was_not_vtx_fetch(bc) ||
1357 bc->force_add_cf) {
1358 r = r600_bytecode_add_cf(bc);
1359 if (r) {
1360 free(nvtx);
1361 return r;
1362 }
1363 switch (bc->chip_class) {
1364 case R600:
1365 case R700:
1366 case EVERGREEN:
1367 bc->cf_last->op = CF_OP_VTX;
1368 break;
1369 case CAYMAN:
1370 bc->cf_last->op = CF_OP_TEX;
1371 break;
1372 default:
1373 R600_ERR("Unknown chip class %d.\n", bc->chip_class);
1374 free(nvtx);
1375 return -EINVAL;
1376 }
1377 }
1378 LIST_ADDTAIL(&nvtx->list, &bc->cf_last->vtx);
1379 /* each fetch use 4 dwords */
1380 bc->cf_last->ndw += 4;
1381 bc->ndw += 4;
1382 if ((bc->cf_last->ndw / 4) >= r600_bytecode_num_tex_and_vtx_instructions(bc))
1383 bc->force_add_cf = 1;
1384
1385 bc->ngpr = MAX2(bc->ngpr, vtx->src_gpr + 1);
1386 bc->ngpr = MAX2(bc->ngpr, vtx->dst_gpr + 1);
1387
1388 return 0;
1389 }
1390
r600_bytecode_add_tex(struct r600_bytecode * bc,const struct r600_bytecode_tex * tex)1391 int r600_bytecode_add_tex(struct r600_bytecode *bc, const struct r600_bytecode_tex *tex)
1392 {
1393 struct r600_bytecode_tex *ntex = r600_bytecode_tex();
1394 int r;
1395
1396 if (!ntex)
1397 return -ENOMEM;
1398 memcpy(ntex, tex, sizeof(struct r600_bytecode_tex));
1399
1400 /* Load index register if required */
1401 if (bc->chip_class >= EVERGREEN) {
1402 if (tex->sampler_index_mode || tex->resource_index_mode)
1403 egcm_load_index_reg(bc, 1, false);
1404 }
1405
1406 /* we can't fetch data und use it as texture lookup address in the same TEX clause */
1407 if (bc->cf_last != NULL &&
1408 bc->cf_last->op == CF_OP_TEX) {
1409 struct r600_bytecode_tex *ttex;
1410 LIST_FOR_EACH_ENTRY(ttex, &bc->cf_last->tex, list) {
1411 if (ttex->dst_gpr == ntex->src_gpr) {
1412 bc->force_add_cf = 1;
1413 break;
1414 }
1415 }
1416 /* slight hack to make gradients always go into same cf */
1417 if (ntex->op == FETCH_OP_SET_GRADIENTS_H)
1418 bc->force_add_cf = 1;
1419 }
1420
1421 /* cf can contains only alu or only vtx or only tex */
1422 if (bc->cf_last == NULL ||
1423 bc->cf_last->op != CF_OP_TEX ||
1424 bc->force_add_cf) {
1425 r = r600_bytecode_add_cf(bc);
1426 if (r) {
1427 free(ntex);
1428 return r;
1429 }
1430 bc->cf_last->op = CF_OP_TEX;
1431 }
1432 if (ntex->src_gpr >= bc->ngpr) {
1433 bc->ngpr = ntex->src_gpr + 1;
1434 }
1435 if (ntex->dst_gpr >= bc->ngpr) {
1436 bc->ngpr = ntex->dst_gpr + 1;
1437 }
1438 LIST_ADDTAIL(&ntex->list, &bc->cf_last->tex);
1439 /* each texture fetch use 4 dwords */
1440 bc->cf_last->ndw += 4;
1441 bc->ndw += 4;
1442 if ((bc->cf_last->ndw / 4) >= r600_bytecode_num_tex_and_vtx_instructions(bc))
1443 bc->force_add_cf = 1;
1444 return 0;
1445 }
1446
r600_bytecode_add_gds(struct r600_bytecode * bc,const struct r600_bytecode_gds * gds)1447 int r600_bytecode_add_gds(struct r600_bytecode *bc, const struct r600_bytecode_gds *gds)
1448 {
1449 struct r600_bytecode_gds *ngds = r600_bytecode_gds();
1450 int r;
1451
1452 if (ngds == NULL)
1453 return -ENOMEM;
1454 memcpy(ngds, gds, sizeof(struct r600_bytecode_gds));
1455
1456 if (bc->cf_last == NULL ||
1457 bc->cf_last->op != CF_OP_GDS ||
1458 bc->force_add_cf) {
1459 r = r600_bytecode_add_cf(bc);
1460 if (r) {
1461 free(ngds);
1462 return r;
1463 }
1464 bc->cf_last->op = CF_OP_GDS;
1465 }
1466
1467 LIST_ADDTAIL(&ngds->list, &bc->cf_last->gds);
1468 bc->cf_last->ndw += 4; /* each GDS uses 4 dwords */
1469 if ((bc->cf_last->ndw / 4) >= r600_bytecode_num_tex_and_vtx_instructions(bc))
1470 bc->force_add_cf = 1;
1471 return 0;
1472 }
1473
r600_bytecode_add_cfinst(struct r600_bytecode * bc,unsigned op)1474 int r600_bytecode_add_cfinst(struct r600_bytecode *bc, unsigned op)
1475 {
1476 int r;
1477 r = r600_bytecode_add_cf(bc);
1478 if (r)
1479 return r;
1480
1481 bc->cf_last->cond = V_SQ_CF_COND_ACTIVE;
1482 bc->cf_last->op = op;
1483 return 0;
1484 }
1485
cm_bytecode_add_cf_end(struct r600_bytecode * bc)1486 int cm_bytecode_add_cf_end(struct r600_bytecode *bc)
1487 {
1488 return r600_bytecode_add_cfinst(bc, CF_OP_CF_END);
1489 }
1490
1491 /* common to all 3 families */
r600_bytecode_vtx_build(struct r600_bytecode * bc,struct r600_bytecode_vtx * vtx,unsigned id)1492 static int r600_bytecode_vtx_build(struct r600_bytecode *bc, struct r600_bytecode_vtx *vtx, unsigned id)
1493 {
1494 bc->bytecode[id] = S_SQ_VTX_WORD0_BUFFER_ID(vtx->buffer_id) |
1495 S_SQ_VTX_WORD0_FETCH_TYPE(vtx->fetch_type) |
1496 S_SQ_VTX_WORD0_SRC_GPR(vtx->src_gpr) |
1497 S_SQ_VTX_WORD0_SRC_SEL_X(vtx->src_sel_x);
1498 if (bc->chip_class < CAYMAN)
1499 bc->bytecode[id] |= S_SQ_VTX_WORD0_MEGA_FETCH_COUNT(vtx->mega_fetch_count);
1500 id++;
1501 bc->bytecode[id++] = S_SQ_VTX_WORD1_DST_SEL_X(vtx->dst_sel_x) |
1502 S_SQ_VTX_WORD1_DST_SEL_Y(vtx->dst_sel_y) |
1503 S_SQ_VTX_WORD1_DST_SEL_Z(vtx->dst_sel_z) |
1504 S_SQ_VTX_WORD1_DST_SEL_W(vtx->dst_sel_w) |
1505 S_SQ_VTX_WORD1_USE_CONST_FIELDS(vtx->use_const_fields) |
1506 S_SQ_VTX_WORD1_DATA_FORMAT(vtx->data_format) |
1507 S_SQ_VTX_WORD1_NUM_FORMAT_ALL(vtx->num_format_all) |
1508 S_SQ_VTX_WORD1_FORMAT_COMP_ALL(vtx->format_comp_all) |
1509 S_SQ_VTX_WORD1_SRF_MODE_ALL(vtx->srf_mode_all) |
1510 S_SQ_VTX_WORD1_GPR_DST_GPR(vtx->dst_gpr);
1511 bc->bytecode[id] = S_SQ_VTX_WORD2_OFFSET(vtx->offset)|
1512 S_SQ_VTX_WORD2_ENDIAN_SWAP(vtx->endian);
1513 if (bc->chip_class >= EVERGREEN)
1514 bc->bytecode[id] |= ((vtx->buffer_index_mode & 0x3) << 21); // S_SQ_VTX_WORD2_BIM(vtx->buffer_index_mode);
1515 if (bc->chip_class < CAYMAN)
1516 bc->bytecode[id] |= S_SQ_VTX_WORD2_MEGA_FETCH(1);
1517 id++;
1518 bc->bytecode[id++] = 0;
1519 return 0;
1520 }
1521
1522 /* common to all 3 families */
r600_bytecode_tex_build(struct r600_bytecode * bc,struct r600_bytecode_tex * tex,unsigned id)1523 static int r600_bytecode_tex_build(struct r600_bytecode *bc, struct r600_bytecode_tex *tex, unsigned id)
1524 {
1525 bc->bytecode[id] = S_SQ_TEX_WORD0_TEX_INST(
1526 r600_isa_fetch_opcode(bc->isa->hw_class, tex->op)) |
1527 EG_S_SQ_TEX_WORD0_INST_MOD(tex->inst_mod) |
1528 S_SQ_TEX_WORD0_RESOURCE_ID(tex->resource_id) |
1529 S_SQ_TEX_WORD0_SRC_GPR(tex->src_gpr) |
1530 S_SQ_TEX_WORD0_SRC_REL(tex->src_rel);
1531 if (bc->chip_class >= EVERGREEN)
1532 bc->bytecode[id] |= ((tex->sampler_index_mode & 0x3) << 27) | // S_SQ_TEX_WORD0_SIM(tex->sampler_index_mode);
1533 ((tex->resource_index_mode & 0x3) << 25); // S_SQ_TEX_WORD0_RIM(tex->resource_index_mode)
1534 id++;
1535 bc->bytecode[id++] = S_SQ_TEX_WORD1_DST_GPR(tex->dst_gpr) |
1536 S_SQ_TEX_WORD1_DST_REL(tex->dst_rel) |
1537 S_SQ_TEX_WORD1_DST_SEL_X(tex->dst_sel_x) |
1538 S_SQ_TEX_WORD1_DST_SEL_Y(tex->dst_sel_y) |
1539 S_SQ_TEX_WORD1_DST_SEL_Z(tex->dst_sel_z) |
1540 S_SQ_TEX_WORD1_DST_SEL_W(tex->dst_sel_w) |
1541 S_SQ_TEX_WORD1_LOD_BIAS(tex->lod_bias) |
1542 S_SQ_TEX_WORD1_COORD_TYPE_X(tex->coord_type_x) |
1543 S_SQ_TEX_WORD1_COORD_TYPE_Y(tex->coord_type_y) |
1544 S_SQ_TEX_WORD1_COORD_TYPE_Z(tex->coord_type_z) |
1545 S_SQ_TEX_WORD1_COORD_TYPE_W(tex->coord_type_w);
1546 bc->bytecode[id++] = S_SQ_TEX_WORD2_OFFSET_X(tex->offset_x) |
1547 S_SQ_TEX_WORD2_OFFSET_Y(tex->offset_y) |
1548 S_SQ_TEX_WORD2_OFFSET_Z(tex->offset_z) |
1549 S_SQ_TEX_WORD2_SAMPLER_ID(tex->sampler_id) |
1550 S_SQ_TEX_WORD2_SRC_SEL_X(tex->src_sel_x) |
1551 S_SQ_TEX_WORD2_SRC_SEL_Y(tex->src_sel_y) |
1552 S_SQ_TEX_WORD2_SRC_SEL_Z(tex->src_sel_z) |
1553 S_SQ_TEX_WORD2_SRC_SEL_W(tex->src_sel_w);
1554 bc->bytecode[id++] = 0;
1555 return 0;
1556 }
1557
1558 /* r600 only, r700/eg bits in r700_asm.c */
r600_bytecode_alu_build(struct r600_bytecode * bc,struct r600_bytecode_alu * alu,unsigned id)1559 static int r600_bytecode_alu_build(struct r600_bytecode *bc, struct r600_bytecode_alu *alu, unsigned id)
1560 {
1561 unsigned opcode = r600_isa_alu_opcode(bc->isa->hw_class, alu->op);
1562
1563 /* don't replace gpr by pv or ps for destination register */
1564 bc->bytecode[id++] = S_SQ_ALU_WORD0_SRC0_SEL(alu->src[0].sel) |
1565 S_SQ_ALU_WORD0_SRC0_REL(alu->src[0].rel) |
1566 S_SQ_ALU_WORD0_SRC0_CHAN(alu->src[0].chan) |
1567 S_SQ_ALU_WORD0_SRC0_NEG(alu->src[0].neg) |
1568 S_SQ_ALU_WORD0_SRC1_SEL(alu->src[1].sel) |
1569 S_SQ_ALU_WORD0_SRC1_REL(alu->src[1].rel) |
1570 S_SQ_ALU_WORD0_SRC1_CHAN(alu->src[1].chan) |
1571 S_SQ_ALU_WORD0_SRC1_NEG(alu->src[1].neg) |
1572 S_SQ_ALU_WORD0_INDEX_MODE(alu->index_mode) |
1573 S_SQ_ALU_WORD0_PRED_SEL(alu->pred_sel) |
1574 S_SQ_ALU_WORD0_LAST(alu->last);
1575
1576 if (alu->is_op3) {
1577 assert(!alu->src[0].abs && !alu->src[1].abs && !alu->src[2].abs);
1578 bc->bytecode[id++] = S_SQ_ALU_WORD1_DST_GPR(alu->dst.sel) |
1579 S_SQ_ALU_WORD1_DST_CHAN(alu->dst.chan) |
1580 S_SQ_ALU_WORD1_DST_REL(alu->dst.rel) |
1581 S_SQ_ALU_WORD1_CLAMP(alu->dst.clamp) |
1582 S_SQ_ALU_WORD1_OP3_SRC2_SEL(alu->src[2].sel) |
1583 S_SQ_ALU_WORD1_OP3_SRC2_REL(alu->src[2].rel) |
1584 S_SQ_ALU_WORD1_OP3_SRC2_CHAN(alu->src[2].chan) |
1585 S_SQ_ALU_WORD1_OP3_SRC2_NEG(alu->src[2].neg) |
1586 S_SQ_ALU_WORD1_OP3_ALU_INST(opcode) |
1587 S_SQ_ALU_WORD1_BANK_SWIZZLE(alu->bank_swizzle);
1588 } else {
1589 bc->bytecode[id++] = S_SQ_ALU_WORD1_DST_GPR(alu->dst.sel) |
1590 S_SQ_ALU_WORD1_DST_CHAN(alu->dst.chan) |
1591 S_SQ_ALU_WORD1_DST_REL(alu->dst.rel) |
1592 S_SQ_ALU_WORD1_CLAMP(alu->dst.clamp) |
1593 S_SQ_ALU_WORD1_OP2_SRC0_ABS(alu->src[0].abs) |
1594 S_SQ_ALU_WORD1_OP2_SRC1_ABS(alu->src[1].abs) |
1595 S_SQ_ALU_WORD1_OP2_WRITE_MASK(alu->dst.write) |
1596 S_SQ_ALU_WORD1_OP2_OMOD(alu->omod) |
1597 S_SQ_ALU_WORD1_OP2_ALU_INST(opcode) |
1598 S_SQ_ALU_WORD1_BANK_SWIZZLE(alu->bank_swizzle) |
1599 S_SQ_ALU_WORD1_OP2_UPDATE_EXECUTE_MASK(alu->execute_mask) |
1600 S_SQ_ALU_WORD1_OP2_UPDATE_PRED(alu->update_pred);
1601 }
1602 return 0;
1603 }
1604
r600_bytecode_cf_vtx_build(uint32_t * bytecode,const struct r600_bytecode_cf * cf)1605 static void r600_bytecode_cf_vtx_build(uint32_t *bytecode, const struct r600_bytecode_cf *cf)
1606 {
1607 *bytecode++ = S_SQ_CF_WORD0_ADDR(cf->addr >> 1);
1608 *bytecode++ = S_SQ_CF_WORD1_CF_INST(r600_isa_cf_opcode(ISA_CC_R600, cf->op)) |
1609 S_SQ_CF_WORD1_BARRIER(1) |
1610 S_SQ_CF_WORD1_COUNT((cf->ndw / 4) - 1);
1611 }
1612
1613 /* common for r600/r700 - eg in eg_asm.c */
r600_bytecode_cf_build(struct r600_bytecode * bc,struct r600_bytecode_cf * cf)1614 static int r600_bytecode_cf_build(struct r600_bytecode *bc, struct r600_bytecode_cf *cf)
1615 {
1616 unsigned id = cf->id;
1617 const struct cf_op_info *cfop = r600_isa_cf(cf->op);
1618 unsigned opcode = r600_isa_cf_opcode(bc->isa->hw_class, cf->op);
1619
1620
1621 if (cf->op == CF_NATIVE) {
1622 bc->bytecode[id++] = cf->isa[0];
1623 bc->bytecode[id++] = cf->isa[1];
1624 } else if (cfop->flags & CF_ALU) {
1625 bc->bytecode[id++] = S_SQ_CF_ALU_WORD0_ADDR(cf->addr >> 1) |
1626 S_SQ_CF_ALU_WORD0_KCACHE_MODE0(cf->kcache[0].mode) |
1627 S_SQ_CF_ALU_WORD0_KCACHE_BANK0(cf->kcache[0].bank) |
1628 S_SQ_CF_ALU_WORD0_KCACHE_BANK1(cf->kcache[1].bank);
1629
1630 bc->bytecode[id++] = S_SQ_CF_ALU_WORD1_CF_INST(opcode) |
1631 S_SQ_CF_ALU_WORD1_KCACHE_MODE1(cf->kcache[1].mode) |
1632 S_SQ_CF_ALU_WORD1_KCACHE_ADDR0(cf->kcache[0].addr) |
1633 S_SQ_CF_ALU_WORD1_KCACHE_ADDR1(cf->kcache[1].addr) |
1634 S_SQ_CF_ALU_WORD1_BARRIER(1) |
1635 S_SQ_CF_ALU_WORD1_USES_WATERFALL(bc->chip_class == R600 ? cf->r6xx_uses_waterfall : 0) |
1636 S_SQ_CF_ALU_WORD1_COUNT((cf->ndw / 2) - 1);
1637 } else if (cfop->flags & CF_FETCH) {
1638 if (bc->chip_class == R700)
1639 r700_bytecode_cf_vtx_build(&bc->bytecode[id], cf);
1640 else
1641 r600_bytecode_cf_vtx_build(&bc->bytecode[id], cf);
1642 } else if (cfop->flags & CF_EXP) {
1643 bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD0_RW_GPR(cf->output.gpr) |
1644 S_SQ_CF_ALLOC_EXPORT_WORD0_ELEM_SIZE(cf->output.elem_size) |
1645 S_SQ_CF_ALLOC_EXPORT_WORD0_ARRAY_BASE(cf->output.array_base) |
1646 S_SQ_CF_ALLOC_EXPORT_WORD0_TYPE(cf->output.type) |
1647 S_SQ_CF_ALLOC_EXPORT_WORD0_INDEX_GPR(cf->output.index_gpr);
1648 bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(cf->output.burst_count - 1) |
1649 S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_X(cf->output.swizzle_x) |
1650 S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Y(cf->output.swizzle_y) |
1651 S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Z(cf->output.swizzle_z) |
1652 S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_W(cf->output.swizzle_w) |
1653 S_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(cf->barrier) |
1654 S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(opcode) |
1655 S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(cf->end_of_program);
1656 } else if (cfop->flags & CF_MEM) {
1657 bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD0_RW_GPR(cf->output.gpr) |
1658 S_SQ_CF_ALLOC_EXPORT_WORD0_ELEM_SIZE(cf->output.elem_size) |
1659 S_SQ_CF_ALLOC_EXPORT_WORD0_ARRAY_BASE(cf->output.array_base) |
1660 S_SQ_CF_ALLOC_EXPORT_WORD0_TYPE(cf->output.type) |
1661 S_SQ_CF_ALLOC_EXPORT_WORD0_INDEX_GPR(cf->output.index_gpr);
1662 bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(cf->output.burst_count - 1) |
1663 S_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(cf->barrier) |
1664 S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(opcode) |
1665 S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(cf->end_of_program) |
1666 S_SQ_CF_ALLOC_EXPORT_WORD1_BUF_ARRAY_SIZE(cf->output.array_size) |
1667 S_SQ_CF_ALLOC_EXPORT_WORD1_BUF_COMP_MASK(cf->output.comp_mask);
1668 } else {
1669 bc->bytecode[id++] = S_SQ_CF_WORD0_ADDR(cf->cf_addr >> 1);
1670 bc->bytecode[id++] = S_SQ_CF_WORD1_CF_INST(opcode) |
1671 S_SQ_CF_WORD1_BARRIER(1) |
1672 S_SQ_CF_WORD1_COND(cf->cond) |
1673 S_SQ_CF_WORD1_POP_COUNT(cf->pop_count) |
1674 S_SQ_CF_WORD1_END_OF_PROGRAM(cf->end_of_program);
1675 }
1676 return 0;
1677 }
1678
r600_bytecode_build(struct r600_bytecode * bc)1679 int r600_bytecode_build(struct r600_bytecode *bc)
1680 {
1681 struct r600_bytecode_cf *cf;
1682 struct r600_bytecode_alu *alu;
1683 struct r600_bytecode_vtx *vtx;
1684 struct r600_bytecode_tex *tex;
1685 struct r600_bytecode_gds *gds;
1686 uint32_t literal[4];
1687 unsigned nliteral;
1688 unsigned addr;
1689 int i, r;
1690
1691 if (!bc->nstack) // If not 0, Stack_size already provided by llvm
1692 bc->nstack = bc->stack.max_entries;
1693
1694 if ((bc->type == PIPE_SHADER_VERTEX || bc->type == PIPE_SHADER_TESS_EVAL || bc->type == PIPE_SHADER_TESS_CTRL) && !bc->nstack) {
1695 bc->nstack = 1;
1696 }
1697
1698 /* first path compute addr of each CF block */
1699 /* addr start after all the CF instructions */
1700 addr = bc->cf_last->id + 2;
1701 LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) {
1702 if (r600_isa_cf(cf->op)->flags & CF_FETCH) {
1703 addr += 3;
1704 addr &= 0xFFFFFFFCUL;
1705 }
1706 cf->addr = addr;
1707 addr += cf->ndw;
1708 bc->ndw = cf->addr + cf->ndw;
1709 }
1710 free(bc->bytecode);
1711 bc->bytecode = calloc(4, bc->ndw);
1712 if (bc->bytecode == NULL)
1713 return -ENOMEM;
1714 LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) {
1715 const struct cf_op_info *cfop = r600_isa_cf(cf->op);
1716 addr = cf->addr;
1717 if (bc->chip_class >= EVERGREEN)
1718 r = eg_bytecode_cf_build(bc, cf);
1719 else
1720 r = r600_bytecode_cf_build(bc, cf);
1721 if (r)
1722 return r;
1723 if (cfop->flags & CF_ALU) {
1724 nliteral = 0;
1725 memset(literal, 0, sizeof(literal));
1726 LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) {
1727 r = r600_bytecode_alu_nliterals(bc, alu, literal, &nliteral);
1728 if (r)
1729 return r;
1730 r600_bytecode_alu_adjust_literals(bc, alu, literal, nliteral);
1731 r600_bytecode_assign_kcache_banks(bc, alu, cf->kcache);
1732
1733 switch(bc->chip_class) {
1734 case R600:
1735 r = r600_bytecode_alu_build(bc, alu, addr);
1736 break;
1737 case R700:
1738 r = r700_bytecode_alu_build(bc, alu, addr);
1739 break;
1740 case EVERGREEN:
1741 case CAYMAN:
1742 r = eg_bytecode_alu_build(bc, alu, addr);
1743 break;
1744 default:
1745 R600_ERR("unknown chip class %d.\n", bc->chip_class);
1746 return -EINVAL;
1747 }
1748 if (r)
1749 return r;
1750 addr += 2;
1751 if (alu->last) {
1752 for (i = 0; i < align(nliteral, 2); ++i) {
1753 bc->bytecode[addr++] = literal[i];
1754 }
1755 nliteral = 0;
1756 memset(literal, 0, sizeof(literal));
1757 }
1758 }
1759 } else if (cf->op == CF_OP_VTX) {
1760 LIST_FOR_EACH_ENTRY(vtx, &cf->vtx, list) {
1761 r = r600_bytecode_vtx_build(bc, vtx, addr);
1762 if (r)
1763 return r;
1764 addr += 4;
1765 }
1766 } else if (cf->op == CF_OP_GDS) {
1767 assert(bc->chip_class >= EVERGREEN);
1768 LIST_FOR_EACH_ENTRY(gds, &cf->gds, list) {
1769 r = eg_bytecode_gds_build(bc, gds, addr);
1770 if (r)
1771 return r;
1772 addr += 4;
1773 }
1774 } else if (cf->op == CF_OP_TEX) {
1775 LIST_FOR_EACH_ENTRY(vtx, &cf->vtx, list) {
1776 assert(bc->chip_class >= EVERGREEN);
1777 r = r600_bytecode_vtx_build(bc, vtx, addr);
1778 if (r)
1779 return r;
1780 addr += 4;
1781 }
1782 LIST_FOR_EACH_ENTRY(tex, &cf->tex, list) {
1783 r = r600_bytecode_tex_build(bc, tex, addr);
1784 if (r)
1785 return r;
1786 addr += 4;
1787 }
1788 }
1789 }
1790 return 0;
1791 }
1792
r600_bytecode_clear(struct r600_bytecode * bc)1793 void r600_bytecode_clear(struct r600_bytecode *bc)
1794 {
1795 struct r600_bytecode_cf *cf = NULL, *next_cf;
1796
1797 free(bc->bytecode);
1798 bc->bytecode = NULL;
1799
1800 LIST_FOR_EACH_ENTRY_SAFE(cf, next_cf, &bc->cf, list) {
1801 struct r600_bytecode_alu *alu = NULL, *next_alu;
1802 struct r600_bytecode_tex *tex = NULL, *next_tex;
1803 struct r600_bytecode_tex *vtx = NULL, *next_vtx;
1804 struct r600_bytecode_gds *gds = NULL, *next_gds;
1805
1806 LIST_FOR_EACH_ENTRY_SAFE(alu, next_alu, &cf->alu, list) {
1807 free(alu);
1808 }
1809
1810 LIST_INITHEAD(&cf->alu);
1811
1812 LIST_FOR_EACH_ENTRY_SAFE(tex, next_tex, &cf->tex, list) {
1813 free(tex);
1814 }
1815
1816 LIST_INITHEAD(&cf->tex);
1817
1818 LIST_FOR_EACH_ENTRY_SAFE(vtx, next_vtx, &cf->vtx, list) {
1819 free(vtx);
1820 }
1821
1822 LIST_INITHEAD(&cf->vtx);
1823
1824 LIST_FOR_EACH_ENTRY_SAFE(gds, next_gds, &cf->gds, list) {
1825 free(gds);
1826 }
1827
1828 LIST_INITHEAD(&cf->gds);
1829
1830 free(cf);
1831 }
1832
1833 LIST_INITHEAD(&cf->list);
1834 }
1835
print_swizzle(unsigned swz)1836 static int print_swizzle(unsigned swz)
1837 {
1838 const char * swzchars = "xyzw01?_";
1839 assert(swz<8 && swz != 6);
1840 return fprintf(stderr, "%c", swzchars[swz]);
1841 }
1842
print_sel(unsigned sel,unsigned rel,unsigned index_mode,unsigned need_brackets)1843 static int print_sel(unsigned sel, unsigned rel, unsigned index_mode,
1844 unsigned need_brackets)
1845 {
1846 int o = 0;
1847 if (rel && index_mode >= 5 && sel < 128)
1848 o += fprintf(stderr, "G");
1849 if (rel || need_brackets) {
1850 o += fprintf(stderr, "[");
1851 }
1852 o += fprintf(stderr, "%d", sel);
1853 if (rel) {
1854 if (index_mode == 0 || index_mode == 6)
1855 o += fprintf(stderr, "+AR");
1856 else if (index_mode == 4)
1857 o += fprintf(stderr, "+AL");
1858 }
1859 if (rel || need_brackets) {
1860 o += fprintf(stderr, "]");
1861 }
1862 return o;
1863 }
1864
print_dst(struct r600_bytecode_alu * alu)1865 static int print_dst(struct r600_bytecode_alu *alu)
1866 {
1867 int o = 0;
1868 unsigned sel = alu->dst.sel;
1869 char reg_char = 'R';
1870 if (sel > 128 - 4) { /* clause temporary gpr */
1871 sel -= 128 - 4;
1872 reg_char = 'T';
1873 }
1874
1875 if (alu_writes(alu)) {
1876 o += fprintf(stderr, "%c", reg_char);
1877 o += print_sel(alu->dst.sel, alu->dst.rel, alu->index_mode, 0);
1878 } else {
1879 o += fprintf(stderr, "__");
1880 }
1881 o += fprintf(stderr, ".");
1882 o += print_swizzle(alu->dst.chan);
1883 return o;
1884 }
1885
print_src(struct r600_bytecode_alu * alu,unsigned idx)1886 static int print_src(struct r600_bytecode_alu *alu, unsigned idx)
1887 {
1888 int o = 0;
1889 struct r600_bytecode_alu_src *src = &alu->src[idx];
1890 unsigned sel = src->sel, need_sel = 1, need_chan = 1, need_brackets = 0;
1891
1892 if (src->neg)
1893 o += fprintf(stderr,"-");
1894 if (src->abs)
1895 o += fprintf(stderr,"|");
1896
1897 if (sel < 128 - 4) {
1898 o += fprintf(stderr, "R");
1899 } else if (sel < 128) {
1900 o += fprintf(stderr, "T");
1901 sel -= 128 - 4;
1902 } else if (sel < 160) {
1903 o += fprintf(stderr, "KC0");
1904 need_brackets = 1;
1905 sel -= 128;
1906 } else if (sel < 192) {
1907 o += fprintf(stderr, "KC1");
1908 need_brackets = 1;
1909 sel -= 160;
1910 } else if (sel >= 512) {
1911 o += fprintf(stderr, "C%d", src->kc_bank);
1912 need_brackets = 1;
1913 sel -= 512;
1914 } else if (sel >= 448) {
1915 o += fprintf(stderr, "Param");
1916 sel -= 448;
1917 need_chan = 0;
1918 } else if (sel >= 288) {
1919 o += fprintf(stderr, "KC3");
1920 need_brackets = 1;
1921 sel -= 288;
1922 } else if (sel >= 256) {
1923 o += fprintf(stderr, "KC2");
1924 need_brackets = 1;
1925 sel -= 256;
1926 } else {
1927 need_sel = 0;
1928 need_chan = 0;
1929 switch (sel) {
1930 case EG_V_SQ_ALU_SRC_LDS_DIRECT_A:
1931 o += fprintf(stderr, "LDS_A[0x%08X]", src->value);
1932 break;
1933 case EG_V_SQ_ALU_SRC_LDS_DIRECT_B:
1934 o += fprintf(stderr, "LDS_B[0x%08X]", src->value);
1935 break;
1936 case EG_V_SQ_ALU_SRC_LDS_OQ_A:
1937 o += fprintf(stderr, "LDS_OQ_A");
1938 need_chan = 1;
1939 break;
1940 case EG_V_SQ_ALU_SRC_LDS_OQ_B:
1941 o += fprintf(stderr, "LDS_OQ_B");
1942 need_chan = 1;
1943 break;
1944 case EG_V_SQ_ALU_SRC_LDS_OQ_A_POP:
1945 o += fprintf(stderr, "LDS_OQ_A_POP");
1946 need_chan = 1;
1947 break;
1948 case EG_V_SQ_ALU_SRC_LDS_OQ_B_POP:
1949 o += fprintf(stderr, "LDS_OQ_B_POP");
1950 need_chan = 1;
1951 break;
1952 case V_SQ_ALU_SRC_PS:
1953 o += fprintf(stderr, "PS");
1954 break;
1955 case V_SQ_ALU_SRC_PV:
1956 o += fprintf(stderr, "PV");
1957 need_chan = 1;
1958 break;
1959 case V_SQ_ALU_SRC_LITERAL:
1960 o += fprintf(stderr, "[0x%08X %f]", src->value, u_bitcast_u2f(src->value));
1961 break;
1962 case V_SQ_ALU_SRC_0_5:
1963 o += fprintf(stderr, "0.5");
1964 break;
1965 case V_SQ_ALU_SRC_M_1_INT:
1966 o += fprintf(stderr, "-1");
1967 break;
1968 case V_SQ_ALU_SRC_1_INT:
1969 o += fprintf(stderr, "1");
1970 break;
1971 case V_SQ_ALU_SRC_1:
1972 o += fprintf(stderr, "1.0");
1973 break;
1974 case V_SQ_ALU_SRC_0:
1975 o += fprintf(stderr, "0");
1976 break;
1977 default:
1978 o += fprintf(stderr, "??IMM_%d", sel);
1979 break;
1980 }
1981 }
1982
1983 if (need_sel)
1984 o += print_sel(sel, src->rel, alu->index_mode, need_brackets);
1985
1986 if (need_chan) {
1987 o += fprintf(stderr, ".");
1988 o += print_swizzle(src->chan);
1989 }
1990
1991 if (src->abs)
1992 o += fprintf(stderr,"|");
1993
1994 return o;
1995 }
1996
print_indent(int p,int c)1997 static int print_indent(int p, int c)
1998 {
1999 int o = 0;
2000 while (p++ < c)
2001 o += fprintf(stderr, " ");
2002 return o;
2003 }
2004
r600_bytecode_disasm(struct r600_bytecode * bc)2005 void r600_bytecode_disasm(struct r600_bytecode *bc)
2006 {
2007 const char *index_mode[] = {"CF_INDEX_NONE", "CF_INDEX_0", "CF_INDEX_1"};
2008 static int index = 0;
2009 struct r600_bytecode_cf *cf = NULL;
2010 struct r600_bytecode_alu *alu = NULL;
2011 struct r600_bytecode_vtx *vtx = NULL;
2012 struct r600_bytecode_tex *tex = NULL;
2013 struct r600_bytecode_gds *gds = NULL;
2014
2015 unsigned i, id, ngr = 0, last;
2016 uint32_t literal[4];
2017 unsigned nliteral;
2018 char chip = '6';
2019
2020 switch (bc->chip_class) {
2021 case R700:
2022 chip = '7';
2023 break;
2024 case EVERGREEN:
2025 chip = 'E';
2026 break;
2027 case CAYMAN:
2028 chip = 'C';
2029 break;
2030 case R600:
2031 default:
2032 chip = '6';
2033 break;
2034 }
2035 fprintf(stderr, "bytecode %d dw -- %d gprs -- %d nstack -------------\n",
2036 bc->ndw, bc->ngpr, bc->nstack);
2037 fprintf(stderr, "shader %d -- %c\n", index++, chip);
2038
2039 LIST_FOR_EACH_ENTRY(cf, &bc->cf, list) {
2040 id = cf->id;
2041 if (cf->op == CF_NATIVE) {
2042 fprintf(stderr, "%04d %08X %08X CF_NATIVE\n", id, bc->bytecode[id],
2043 bc->bytecode[id + 1]);
2044 } else {
2045 const struct cf_op_info *cfop = r600_isa_cf(cf->op);
2046 if (cfop->flags & CF_ALU) {
2047 if (cf->eg_alu_extended) {
2048 fprintf(stderr, "%04d %08X %08X %s\n", id, bc->bytecode[id],
2049 bc->bytecode[id + 1], "ALU_EXT");
2050 id += 2;
2051 }
2052 fprintf(stderr, "%04d %08X %08X %s ", id, bc->bytecode[id],
2053 bc->bytecode[id + 1], cfop->name);
2054 fprintf(stderr, "%d @%d ", cf->ndw / 2, cf->addr);
2055 for (i = 0; i < 4; ++i) {
2056 if (cf->kcache[i].mode) {
2057 int c_start = (cf->kcache[i].addr << 4);
2058 int c_end = c_start + (cf->kcache[i].mode << 4);
2059 fprintf(stderr, "KC%d[CB%d:%d-%d%s%s] ",
2060 i, cf->kcache[i].bank, c_start, c_end,
2061 cf->kcache[i].index_mode ? " " : "",
2062 cf->kcache[i].index_mode ? index_mode[cf->kcache[i].index_mode] : "");
2063 }
2064 }
2065 fprintf(stderr, "\n");
2066 } else if (cfop->flags & CF_FETCH) {
2067 fprintf(stderr, "%04d %08X %08X %s ", id, bc->bytecode[id],
2068 bc->bytecode[id + 1], cfop->name);
2069 fprintf(stderr, "%d @%d ", cf->ndw / 4, cf->addr);
2070 fprintf(stderr, "\n");
2071 } else if (cfop->flags & CF_EXP) {
2072 int o = 0;
2073 const char *exp_type[] = {"PIXEL", "POS ", "PARAM"};
2074 o += fprintf(stderr, "%04d %08X %08X %s ", id, bc->bytecode[id],
2075 bc->bytecode[id + 1], cfop->name);
2076 o += print_indent(o, 43);
2077 o += fprintf(stderr, "%s ", exp_type[cf->output.type]);
2078 if (cf->output.burst_count > 1) {
2079 o += fprintf(stderr, "%d-%d ", cf->output.array_base,
2080 cf->output.array_base + cf->output.burst_count - 1);
2081
2082 o += print_indent(o, 55);
2083 o += fprintf(stderr, "R%d-%d.", cf->output.gpr,
2084 cf->output.gpr + cf->output.burst_count - 1);
2085 } else {
2086 o += fprintf(stderr, "%d ", cf->output.array_base);
2087 o += print_indent(o, 55);
2088 o += fprintf(stderr, "R%d.", cf->output.gpr);
2089 }
2090
2091 o += print_swizzle(cf->output.swizzle_x);
2092 o += print_swizzle(cf->output.swizzle_y);
2093 o += print_swizzle(cf->output.swizzle_z);
2094 o += print_swizzle(cf->output.swizzle_w);
2095
2096 print_indent(o, 67);
2097
2098 fprintf(stderr, " ES:%X ", cf->output.elem_size);
2099 if (!cf->barrier)
2100 fprintf(stderr, "NO_BARRIER ");
2101 if (cf->end_of_program)
2102 fprintf(stderr, "EOP ");
2103 fprintf(stderr, "\n");
2104 } else if (r600_isa_cf(cf->op)->flags & CF_MEM) {
2105 int o = 0;
2106 const char *exp_type[] = {"WRITE", "WRITE_IND", "WRITE_ACK",
2107 "WRITE_IND_ACK"};
2108 o += fprintf(stderr, "%04d %08X %08X %s ", id,
2109 bc->bytecode[id], bc->bytecode[id + 1], cfop->name);
2110 o += print_indent(o, 43);
2111 o += fprintf(stderr, "%s ", exp_type[cf->output.type]);
2112 if (cf->output.burst_count > 1) {
2113 o += fprintf(stderr, "%d-%d ", cf->output.array_base,
2114 cf->output.array_base + cf->output.burst_count - 1);
2115 o += print_indent(o, 55);
2116 o += fprintf(stderr, "R%d-%d.", cf->output.gpr,
2117 cf->output.gpr + cf->output.burst_count - 1);
2118 } else {
2119 o += fprintf(stderr, "%d ", cf->output.array_base);
2120 o += print_indent(o, 55);
2121 o += fprintf(stderr, "R%d.", cf->output.gpr);
2122 }
2123 for (i = 0; i < 4; ++i) {
2124 if (cf->output.comp_mask & (1 << i))
2125 o += print_swizzle(i);
2126 else
2127 o += print_swizzle(7);
2128 }
2129
2130 if (cf->output.type == V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_WRITE_IND)
2131 o += fprintf(stderr, " R%d", cf->output.index_gpr);
2132
2133 o += print_indent(o, 67);
2134
2135 fprintf(stderr, " ES:%i ", cf->output.elem_size);
2136 if (cf->output.array_size != 0xFFF)
2137 fprintf(stderr, "AS:%i ", cf->output.array_size);
2138 if (!cf->barrier)
2139 fprintf(stderr, "NO_BARRIER ");
2140 if (cf->end_of_program)
2141 fprintf(stderr, "EOP ");
2142 fprintf(stderr, "\n");
2143 } else {
2144 fprintf(stderr, "%04d %08X %08X %s ", id, bc->bytecode[id],
2145 bc->bytecode[id + 1], cfop->name);
2146 fprintf(stderr, "@%d ", cf->cf_addr);
2147 if (cf->cond)
2148 fprintf(stderr, "CND:%X ", cf->cond);
2149 if (cf->pop_count)
2150 fprintf(stderr, "POP:%X ", cf->pop_count);
2151 if (cf->count && (cfop->flags & CF_EMIT))
2152 fprintf(stderr, "STREAM%d ", cf->count);
2153 if (cf->end_of_program)
2154 fprintf(stderr, "EOP ");
2155 fprintf(stderr, "\n");
2156 }
2157 }
2158
2159 id = cf->addr;
2160 nliteral = 0;
2161 last = 1;
2162 LIST_FOR_EACH_ENTRY(alu, &cf->alu, list) {
2163 const char *omod_str[] = {"","*2","*4","/2"};
2164 const struct alu_op_info *aop = r600_isa_alu(alu->op);
2165 int o = 0;
2166
2167 r600_bytecode_alu_nliterals(bc, alu, literal, &nliteral);
2168 o += fprintf(stderr, " %04d %08X %08X ", id, bc->bytecode[id], bc->bytecode[id+1]);
2169 if (last)
2170 o += fprintf(stderr, "%4d ", ++ngr);
2171 else
2172 o += fprintf(stderr, " ");
2173 o += fprintf(stderr, "%c%c %c ", alu->execute_mask ? 'M':' ',
2174 alu->update_pred ? 'P':' ',
2175 alu->pred_sel ? alu->pred_sel==2 ? '0':'1':' ');
2176
2177 o += fprintf(stderr, "%s%s%s ", aop->name,
2178 omod_str[alu->omod], alu->dst.clamp ? "_sat":"");
2179
2180 o += print_indent(o,60);
2181 o += print_dst(alu);
2182 for (i = 0; i < aop->src_count; ++i) {
2183 o += fprintf(stderr, i == 0 ? ", ": ", ");
2184 o += print_src(alu, i);
2185 }
2186
2187 if (alu->bank_swizzle) {
2188 o += print_indent(o,75);
2189 o += fprintf(stderr, " BS:%d", alu->bank_swizzle);
2190 }
2191
2192 fprintf(stderr, "\n");
2193 id += 2;
2194
2195 if (alu->last) {
2196 for (i = 0; i < nliteral; i++, id++) {
2197 float *f = (float*)(bc->bytecode + id);
2198 o = fprintf(stderr, " %04d %08X", id, bc->bytecode[id]);
2199 print_indent(o, 60);
2200 fprintf(stderr, " %f (%d)\n", *f, *(bc->bytecode + id));
2201 }
2202 id += nliteral & 1;
2203 nliteral = 0;
2204 }
2205 last = alu->last;
2206 }
2207
2208 LIST_FOR_EACH_ENTRY(tex, &cf->tex, list) {
2209 int o = 0;
2210 o += fprintf(stderr, " %04d %08X %08X %08X ", id, bc->bytecode[id],
2211 bc->bytecode[id + 1], bc->bytecode[id + 2]);
2212
2213 o += fprintf(stderr, "%s ", r600_isa_fetch(tex->op)->name);
2214
2215 o += print_indent(o, 50);
2216
2217 o += fprintf(stderr, "R%d.", tex->dst_gpr);
2218 o += print_swizzle(tex->dst_sel_x);
2219 o += print_swizzle(tex->dst_sel_y);
2220 o += print_swizzle(tex->dst_sel_z);
2221 o += print_swizzle(tex->dst_sel_w);
2222
2223 o += fprintf(stderr, ", R%d.", tex->src_gpr);
2224 o += print_swizzle(tex->src_sel_x);
2225 o += print_swizzle(tex->src_sel_y);
2226 o += print_swizzle(tex->src_sel_z);
2227 o += print_swizzle(tex->src_sel_w);
2228
2229 o += fprintf(stderr, ", RID:%d", tex->resource_id);
2230 o += fprintf(stderr, ", SID:%d ", tex->sampler_id);
2231
2232 if (tex->sampler_index_mode)
2233 fprintf(stderr, "SQ_%s ", index_mode[tex->sampler_index_mode]);
2234
2235 if (tex->lod_bias)
2236 fprintf(stderr, "LB:%d ", tex->lod_bias);
2237
2238 fprintf(stderr, "CT:%c%c%c%c ",
2239 tex->coord_type_x ? 'N' : 'U',
2240 tex->coord_type_y ? 'N' : 'U',
2241 tex->coord_type_z ? 'N' : 'U',
2242 tex->coord_type_w ? 'N' : 'U');
2243
2244 if (tex->offset_x)
2245 fprintf(stderr, "OX:%d ", tex->offset_x);
2246 if (tex->offset_y)
2247 fprintf(stderr, "OY:%d ", tex->offset_y);
2248 if (tex->offset_z)
2249 fprintf(stderr, "OZ:%d ", tex->offset_z);
2250
2251 id += 4;
2252 fprintf(stderr, "\n");
2253 }
2254
2255 LIST_FOR_EACH_ENTRY(vtx, &cf->vtx, list) {
2256 int o = 0;
2257 const char * fetch_type[] = {"VERTEX", "INSTANCE", ""};
2258 o += fprintf(stderr, " %04d %08X %08X %08X ", id, bc->bytecode[id],
2259 bc->bytecode[id + 1], bc->bytecode[id + 2]);
2260
2261 o += fprintf(stderr, "%s ", r600_isa_fetch(vtx->op)->name);
2262
2263 o += print_indent(o, 50);
2264
2265 o += fprintf(stderr, "R%d.", vtx->dst_gpr);
2266 o += print_swizzle(vtx->dst_sel_x);
2267 o += print_swizzle(vtx->dst_sel_y);
2268 o += print_swizzle(vtx->dst_sel_z);
2269 o += print_swizzle(vtx->dst_sel_w);
2270
2271 o += fprintf(stderr, ", R%d.", vtx->src_gpr);
2272 o += print_swizzle(vtx->src_sel_x);
2273
2274 if (vtx->offset)
2275 fprintf(stderr, " +%db", vtx->offset);
2276
2277 o += print_indent(o, 55);
2278
2279 fprintf(stderr, ", RID:%d ", vtx->buffer_id);
2280
2281 fprintf(stderr, "%s ", fetch_type[vtx->fetch_type]);
2282
2283 if (bc->chip_class < CAYMAN && vtx->mega_fetch_count)
2284 fprintf(stderr, "MFC:%d ", vtx->mega_fetch_count);
2285
2286 if (bc->chip_class >= EVERGREEN && vtx->buffer_index_mode)
2287 fprintf(stderr, "SQ_%s ", index_mode[vtx->buffer_index_mode]);
2288
2289 fprintf(stderr, "UCF:%d ", vtx->use_const_fields);
2290 fprintf(stderr, "FMT(DTA:%d ", vtx->data_format);
2291 fprintf(stderr, "NUM:%d ", vtx->num_format_all);
2292 fprintf(stderr, "COMP:%d ", vtx->format_comp_all);
2293 fprintf(stderr, "MODE:%d)\n", vtx->srf_mode_all);
2294
2295 id += 4;
2296 }
2297
2298 LIST_FOR_EACH_ENTRY(gds, &cf->gds, list) {
2299 int o = 0;
2300 o += fprintf(stderr, " %04d %08X %08X %08X ", id, bc->bytecode[id],
2301 bc->bytecode[id + 1], bc->bytecode[id + 2]);
2302
2303 o += fprintf(stderr, "%s ", r600_isa_fetch(gds->op)->name);
2304
2305 if (gds->op != FETCH_OP_TF_WRITE) {
2306 o += fprintf(stderr, "R%d.", gds->dst_gpr);
2307 o += print_swizzle(gds->dst_sel_x);
2308 o += print_swizzle(gds->dst_sel_y);
2309 o += print_swizzle(gds->dst_sel_z);
2310 o += print_swizzle(gds->dst_sel_w);
2311 }
2312
2313 o += fprintf(stderr, ", R%d.", gds->src_gpr);
2314 o += print_swizzle(gds->src_sel_x);
2315 o += print_swizzle(gds->src_sel_y);
2316 o += print_swizzle(gds->src_sel_z);
2317
2318 if (gds->op != FETCH_OP_TF_WRITE) {
2319 o += fprintf(stderr, ", R%d.", gds->src_gpr2);
2320 }
2321 fprintf(stderr, "\n");
2322 id += 4;
2323 }
2324 }
2325
2326 fprintf(stderr, "--------------------------------------\n");
2327 }
2328
r600_vertex_data_type(enum pipe_format pformat,unsigned * format,unsigned * num_format,unsigned * format_comp,unsigned * endian)2329 void r600_vertex_data_type(enum pipe_format pformat,
2330 unsigned *format,
2331 unsigned *num_format, unsigned *format_comp, unsigned *endian)
2332 {
2333 const struct util_format_description *desc;
2334 unsigned i;
2335
2336 *format = 0;
2337 *num_format = 0;
2338 *format_comp = 0;
2339 *endian = ENDIAN_NONE;
2340
2341 if (pformat == PIPE_FORMAT_R11G11B10_FLOAT) {
2342 *format = FMT_10_11_11_FLOAT;
2343 *endian = r600_endian_swap(32);
2344 return;
2345 }
2346
2347 if (pformat == PIPE_FORMAT_B5G6R5_UNORM) {
2348 *format = FMT_5_6_5;
2349 *endian = r600_endian_swap(16);
2350 return;
2351 }
2352
2353 desc = util_format_description(pformat);
2354 if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN) {
2355 goto out_unknown;
2356 }
2357
2358 /* Find the first non-VOID channel. */
2359 for (i = 0; i < 4; i++) {
2360 if (desc->channel[i].type != UTIL_FORMAT_TYPE_VOID) {
2361 break;
2362 }
2363 }
2364
2365 *endian = r600_endian_swap(desc->channel[i].size);
2366
2367 switch (desc->channel[i].type) {
2368 /* Half-floats, floats, ints */
2369 case UTIL_FORMAT_TYPE_FLOAT:
2370 switch (desc->channel[i].size) {
2371 case 16:
2372 switch (desc->nr_channels) {
2373 case 1:
2374 *format = FMT_16_FLOAT;
2375 break;
2376 case 2:
2377 *format = FMT_16_16_FLOAT;
2378 break;
2379 case 3:
2380 case 4:
2381 *format = FMT_16_16_16_16_FLOAT;
2382 break;
2383 }
2384 break;
2385 case 32:
2386 switch (desc->nr_channels) {
2387 case 1:
2388 *format = FMT_32_FLOAT;
2389 break;
2390 case 2:
2391 *format = FMT_32_32_FLOAT;
2392 break;
2393 case 3:
2394 *format = FMT_32_32_32_FLOAT;
2395 break;
2396 case 4:
2397 *format = FMT_32_32_32_32_FLOAT;
2398 break;
2399 }
2400 break;
2401 default:
2402 goto out_unknown;
2403 }
2404 break;
2405 /* Unsigned ints */
2406 case UTIL_FORMAT_TYPE_UNSIGNED:
2407 /* Signed ints */
2408 case UTIL_FORMAT_TYPE_SIGNED:
2409 switch (desc->channel[i].size) {
2410 case 8:
2411 switch (desc->nr_channels) {
2412 case 1:
2413 *format = FMT_8;
2414 break;
2415 case 2:
2416 *format = FMT_8_8;
2417 break;
2418 case 3:
2419 case 4:
2420 *format = FMT_8_8_8_8;
2421 break;
2422 }
2423 break;
2424 case 10:
2425 if (desc->nr_channels != 4)
2426 goto out_unknown;
2427
2428 *format = FMT_2_10_10_10;
2429 break;
2430 case 16:
2431 switch (desc->nr_channels) {
2432 case 1:
2433 *format = FMT_16;
2434 break;
2435 case 2:
2436 *format = FMT_16_16;
2437 break;
2438 case 3:
2439 case 4:
2440 *format = FMT_16_16_16_16;
2441 break;
2442 }
2443 break;
2444 case 32:
2445 switch (desc->nr_channels) {
2446 case 1:
2447 *format = FMT_32;
2448 break;
2449 case 2:
2450 *format = FMT_32_32;
2451 break;
2452 case 3:
2453 *format = FMT_32_32_32;
2454 break;
2455 case 4:
2456 *format = FMT_32_32_32_32;
2457 break;
2458 }
2459 break;
2460 default:
2461 goto out_unknown;
2462 }
2463 break;
2464 default:
2465 goto out_unknown;
2466 }
2467
2468 if (desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) {
2469 *format_comp = 1;
2470 }
2471
2472 *num_format = 0;
2473 if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED ||
2474 desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) {
2475 if (!desc->channel[i].normalized) {
2476 if (desc->channel[i].pure_integer)
2477 *num_format = 1;
2478 else
2479 *num_format = 2;
2480 }
2481 }
2482 return;
2483 out_unknown:
2484 R600_ERR("unsupported vertex format %s\n", util_format_name(pformat));
2485 }
2486
r600_create_vertex_fetch_shader(struct pipe_context * ctx,unsigned count,const struct pipe_vertex_element * elements)2487 void *r600_create_vertex_fetch_shader(struct pipe_context *ctx,
2488 unsigned count,
2489 const struct pipe_vertex_element *elements)
2490 {
2491 struct r600_context *rctx = (struct r600_context *)ctx;
2492 struct r600_bytecode bc;
2493 struct r600_bytecode_vtx vtx;
2494 const struct util_format_description *desc;
2495 unsigned fetch_resource_start = rctx->b.chip_class >= EVERGREEN ? 0 : 160;
2496 unsigned format, num_format, format_comp, endian;
2497 uint32_t *bytecode;
2498 int i, j, r, fs_size;
2499 struct r600_fetch_shader *shader;
2500 unsigned no_sb = rctx->screen->b.debug_flags & DBG_NO_SB;
2501 unsigned sb_disasm = !no_sb || (rctx->screen->b.debug_flags & DBG_SB_DISASM);
2502
2503 assert(count < 32);
2504
2505 memset(&bc, 0, sizeof(bc));
2506 r600_bytecode_init(&bc, rctx->b.chip_class, rctx->b.family,
2507 rctx->screen->has_compressed_msaa_texturing);
2508
2509 bc.isa = rctx->isa;
2510
2511 for (i = 0; i < count; i++) {
2512 if (elements[i].instance_divisor > 1) {
2513 if (rctx->b.chip_class == CAYMAN) {
2514 for (j = 0; j < 4; j++) {
2515 struct r600_bytecode_alu alu;
2516 memset(&alu, 0, sizeof(alu));
2517 alu.op = ALU_OP2_MULHI_UINT;
2518 alu.src[0].sel = 0;
2519 alu.src[0].chan = 3;
2520 alu.src[1].sel = V_SQ_ALU_SRC_LITERAL;
2521 alu.src[1].value = (1ll << 32) / elements[i].instance_divisor + 1;
2522 alu.dst.sel = i + 1;
2523 alu.dst.chan = j;
2524 alu.dst.write = j == 3;
2525 alu.last = j == 3;
2526 if ((r = r600_bytecode_add_alu(&bc, &alu))) {
2527 r600_bytecode_clear(&bc);
2528 return NULL;
2529 }
2530 }
2531 } else {
2532 struct r600_bytecode_alu alu;
2533 memset(&alu, 0, sizeof(alu));
2534 alu.op = ALU_OP2_MULHI_UINT;
2535 alu.src[0].sel = 0;
2536 alu.src[0].chan = 3;
2537 alu.src[1].sel = V_SQ_ALU_SRC_LITERAL;
2538 alu.src[1].value = (1ll << 32) / elements[i].instance_divisor + 1;
2539 alu.dst.sel = i + 1;
2540 alu.dst.chan = 3;
2541 alu.dst.write = 1;
2542 alu.last = 1;
2543 if ((r = r600_bytecode_add_alu(&bc, &alu))) {
2544 r600_bytecode_clear(&bc);
2545 return NULL;
2546 }
2547 }
2548 }
2549 }
2550
2551 for (i = 0; i < count; i++) {
2552 r600_vertex_data_type(elements[i].src_format,
2553 &format, &num_format, &format_comp, &endian);
2554
2555 desc = util_format_description(elements[i].src_format);
2556 if (!desc) {
2557 r600_bytecode_clear(&bc);
2558 R600_ERR("unknown format %d\n", elements[i].src_format);
2559 return NULL;
2560 }
2561
2562 if (elements[i].src_offset > 65535) {
2563 r600_bytecode_clear(&bc);
2564 R600_ERR("too big src_offset: %u\n", elements[i].src_offset);
2565 return NULL;
2566 }
2567
2568 memset(&vtx, 0, sizeof(vtx));
2569 vtx.buffer_id = elements[i].vertex_buffer_index + fetch_resource_start;
2570 vtx.fetch_type = elements[i].instance_divisor ? SQ_VTX_FETCH_INSTANCE_DATA : SQ_VTX_FETCH_VERTEX_DATA;
2571 vtx.src_gpr = elements[i].instance_divisor > 1 ? i + 1 : 0;
2572 vtx.src_sel_x = elements[i].instance_divisor ? 3 : 0;
2573 vtx.mega_fetch_count = 0x1F;
2574 vtx.dst_gpr = i + 1;
2575 vtx.dst_sel_x = desc->swizzle[0];
2576 vtx.dst_sel_y = desc->swizzle[1];
2577 vtx.dst_sel_z = desc->swizzle[2];
2578 vtx.dst_sel_w = desc->swizzle[3];
2579 vtx.data_format = format;
2580 vtx.num_format_all = num_format;
2581 vtx.format_comp_all = format_comp;
2582 vtx.offset = elements[i].src_offset;
2583 vtx.endian = endian;
2584
2585 if ((r = r600_bytecode_add_vtx(&bc, &vtx))) {
2586 r600_bytecode_clear(&bc);
2587 return NULL;
2588 }
2589 }
2590
2591 r600_bytecode_add_cfinst(&bc, CF_OP_RET);
2592
2593 if ((r = r600_bytecode_build(&bc))) {
2594 r600_bytecode_clear(&bc);
2595 return NULL;
2596 }
2597
2598 if (rctx->screen->b.debug_flags & DBG_FS) {
2599 fprintf(stderr, "--------------------------------------------------------------\n");
2600 fprintf(stderr, "Vertex elements state:\n");
2601 for (i = 0; i < count; i++) {
2602 fprintf(stderr, " ");
2603 util_dump_vertex_element(stderr, elements+i);
2604 fprintf(stderr, "\n");
2605 }
2606
2607 if (!sb_disasm) {
2608 r600_bytecode_disasm(&bc);
2609
2610 fprintf(stderr, "______________________________________________________________\n");
2611 } else {
2612 r600_sb_bytecode_process(rctx, &bc, NULL, 1 /*dump*/, 0 /*optimize*/);
2613 }
2614 }
2615
2616 fs_size = bc.ndw*4;
2617
2618 /* Allocate the CSO. */
2619 shader = CALLOC_STRUCT(r600_fetch_shader);
2620 if (!shader) {
2621 r600_bytecode_clear(&bc);
2622 return NULL;
2623 }
2624
2625 u_suballocator_alloc(rctx->allocator_fetch_shader, fs_size, 256,
2626 &shader->offset,
2627 (struct pipe_resource**)&shader->buffer);
2628 if (!shader->buffer) {
2629 r600_bytecode_clear(&bc);
2630 FREE(shader);
2631 return NULL;
2632 }
2633
2634 bytecode = r600_buffer_map_sync_with_rings(&rctx->b, shader->buffer, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_UNSYNCHRONIZED);
2635 bytecode += shader->offset / 4;
2636
2637 if (R600_BIG_ENDIAN) {
2638 for (i = 0; i < fs_size / 4; ++i) {
2639 bytecode[i] = util_cpu_to_le32(bc.bytecode[i]);
2640 }
2641 } else {
2642 memcpy(bytecode, bc.bytecode, fs_size);
2643 }
2644 rctx->b.ws->buffer_unmap(shader->buffer->buf);
2645
2646 r600_bytecode_clear(&bc);
2647 return shader;
2648 }
2649
r600_bytecode_alu_read(struct r600_bytecode * bc,struct r600_bytecode_alu * alu,uint32_t word0,uint32_t word1)2650 void r600_bytecode_alu_read(struct r600_bytecode *bc,
2651 struct r600_bytecode_alu *alu, uint32_t word0, uint32_t word1)
2652 {
2653 /* WORD0 */
2654 alu->src[0].sel = G_SQ_ALU_WORD0_SRC0_SEL(word0);
2655 alu->src[0].rel = G_SQ_ALU_WORD0_SRC0_REL(word0);
2656 alu->src[0].chan = G_SQ_ALU_WORD0_SRC0_CHAN(word0);
2657 alu->src[0].neg = G_SQ_ALU_WORD0_SRC0_NEG(word0);
2658 alu->src[1].sel = G_SQ_ALU_WORD0_SRC1_SEL(word0);
2659 alu->src[1].rel = G_SQ_ALU_WORD0_SRC1_REL(word0);
2660 alu->src[1].chan = G_SQ_ALU_WORD0_SRC1_CHAN(word0);
2661 alu->src[1].neg = G_SQ_ALU_WORD0_SRC1_NEG(word0);
2662 alu->index_mode = G_SQ_ALU_WORD0_INDEX_MODE(word0);
2663 alu->pred_sel = G_SQ_ALU_WORD0_PRED_SEL(word0);
2664 alu->last = G_SQ_ALU_WORD0_LAST(word0);
2665
2666 /* WORD1 */
2667 alu->bank_swizzle = G_SQ_ALU_WORD1_BANK_SWIZZLE(word1);
2668 if (alu->bank_swizzle)
2669 alu->bank_swizzle_force = alu->bank_swizzle;
2670 alu->dst.sel = G_SQ_ALU_WORD1_DST_GPR(word1);
2671 alu->dst.rel = G_SQ_ALU_WORD1_DST_REL(word1);
2672 alu->dst.chan = G_SQ_ALU_WORD1_DST_CHAN(word1);
2673 alu->dst.clamp = G_SQ_ALU_WORD1_CLAMP(word1);
2674 if (G_SQ_ALU_WORD1_ENCODING(word1)) /*ALU_DWORD1_OP3*/
2675 {
2676 alu->is_op3 = 1;
2677 alu->src[2].sel = G_SQ_ALU_WORD1_OP3_SRC2_SEL(word1);
2678 alu->src[2].rel = G_SQ_ALU_WORD1_OP3_SRC2_REL(word1);
2679 alu->src[2].chan = G_SQ_ALU_WORD1_OP3_SRC2_CHAN(word1);
2680 alu->src[2].neg = G_SQ_ALU_WORD1_OP3_SRC2_NEG(word1);
2681 alu->op = r600_isa_alu_by_opcode(bc->isa,
2682 G_SQ_ALU_WORD1_OP3_ALU_INST(word1), /* is_op3 = */ 1);
2683
2684 }
2685 else /*ALU_DWORD1_OP2*/
2686 {
2687 alu->src[0].abs = G_SQ_ALU_WORD1_OP2_SRC0_ABS(word1);
2688 alu->src[1].abs = G_SQ_ALU_WORD1_OP2_SRC1_ABS(word1);
2689 alu->op = r600_isa_alu_by_opcode(bc->isa,
2690 G_SQ_ALU_WORD1_OP2_ALU_INST(word1), /* is_op3 = */ 0);
2691 alu->omod = G_SQ_ALU_WORD1_OP2_OMOD(word1);
2692 alu->dst.write = G_SQ_ALU_WORD1_OP2_WRITE_MASK(word1);
2693 alu->update_pred = G_SQ_ALU_WORD1_OP2_UPDATE_PRED(word1);
2694 alu->execute_mask =
2695 G_SQ_ALU_WORD1_OP2_UPDATE_EXECUTE_MASK(word1);
2696 }
2697 }
2698
2699 #if 0
2700 void r600_bytecode_export_read(struct r600_bytecode *bc,
2701 struct r600_bytecode_output *output, uint32_t word0, uint32_t word1)
2702 {
2703 output->array_base = G_SQ_CF_ALLOC_EXPORT_WORD0_ARRAY_BASE(word0);
2704 output->type = G_SQ_CF_ALLOC_EXPORT_WORD0_TYPE(word0);
2705 output->gpr = G_SQ_CF_ALLOC_EXPORT_WORD0_RW_GPR(word0);
2706 output->elem_size = G_SQ_CF_ALLOC_EXPORT_WORD0_ELEM_SIZE(word0);
2707
2708 output->swizzle_x = G_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_X(word1);
2709 output->swizzle_y = G_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Y(word1);
2710 output->swizzle_z = G_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Z(word1);
2711 output->swizzle_w = G_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_W(word1);
2712 output->burst_count = G_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(word1);
2713 output->end_of_program = G_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(word1);
2714 output->op = r600_isa_cf_by_opcode(bc->isa,
2715 G_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(word1), 0);
2716 output->barrier = G_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(word1);
2717 output->array_size = G_SQ_CF_ALLOC_EXPORT_WORD1_BUF_ARRAY_SIZE(word1);
2718 output->comp_mask = G_SQ_CF_ALLOC_EXPORT_WORD1_BUF_COMP_MASK(word1);
2719 }
2720 #endif
2721