• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2014 Broadcom
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  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * 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 NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  */
23 
24 #include "util/ralloc.h"
25 #include "util/register_allocate.h"
26 #include "vc4_context.h"
27 #include "vc4_qir.h"
28 #include "vc4_qpu.h"
29 
30 #define QPU_R(file, index) { QPU_MUX_##file, index }
31 
32 static const struct qpu_reg vc4_regs[] = {
33         { QPU_MUX_R0, 0},
34         { QPU_MUX_R1, 0},
35         { QPU_MUX_R2, 0},
36         { QPU_MUX_R3, 0},
37         { QPU_MUX_R4, 0},
38         QPU_R(A, 0),
39         QPU_R(B, 0),
40         QPU_R(A, 1),
41         QPU_R(B, 1),
42         QPU_R(A, 2),
43         QPU_R(B, 2),
44         QPU_R(A, 3),
45         QPU_R(B, 3),
46         QPU_R(A, 4),
47         QPU_R(B, 4),
48         QPU_R(A, 5),
49         QPU_R(B, 5),
50         QPU_R(A, 6),
51         QPU_R(B, 6),
52         QPU_R(A, 7),
53         QPU_R(B, 7),
54         QPU_R(A, 8),
55         QPU_R(B, 8),
56         QPU_R(A, 9),
57         QPU_R(B, 9),
58         QPU_R(A, 10),
59         QPU_R(B, 10),
60         QPU_R(A, 11),
61         QPU_R(B, 11),
62         QPU_R(A, 12),
63         QPU_R(B, 12),
64         QPU_R(A, 13),
65         QPU_R(B, 13),
66         QPU_R(A, 14),
67         QPU_R(B, 14),
68         QPU_R(A, 15),
69         QPU_R(B, 15),
70         QPU_R(A, 16),
71         QPU_R(B, 16),
72         QPU_R(A, 17),
73         QPU_R(B, 17),
74         QPU_R(A, 18),
75         QPU_R(B, 18),
76         QPU_R(A, 19),
77         QPU_R(B, 19),
78         QPU_R(A, 20),
79         QPU_R(B, 20),
80         QPU_R(A, 21),
81         QPU_R(B, 21),
82         QPU_R(A, 22),
83         QPU_R(B, 22),
84         QPU_R(A, 23),
85         QPU_R(B, 23),
86         QPU_R(A, 24),
87         QPU_R(B, 24),
88         QPU_R(A, 25),
89         QPU_R(B, 25),
90         QPU_R(A, 26),
91         QPU_R(B, 26),
92         QPU_R(A, 27),
93         QPU_R(B, 27),
94         QPU_R(A, 28),
95         QPU_R(B, 28),
96         QPU_R(A, 29),
97         QPU_R(B, 29),
98         QPU_R(A, 30),
99         QPU_R(B, 30),
100         QPU_R(A, 31),
101         QPU_R(B, 31),
102 };
103 #define ACC_INDEX     0
104 #define ACC_COUNT     5
105 #define AB_INDEX      (ACC_INDEX + ACC_COUNT)
106 #define AB_COUNT      64
107 
108 static void
vc4_alloc_reg_set(struct vc4_context * vc4)109 vc4_alloc_reg_set(struct vc4_context *vc4)
110 {
111         assert(vc4_regs[AB_INDEX].addr == 0);
112         assert(vc4_regs[AB_INDEX + 1].addr == 0);
113         STATIC_ASSERT(ARRAY_SIZE(vc4_regs) == AB_INDEX + 64);
114 
115         if (vc4->regs)
116                 return;
117 
118         vc4->regs = ra_alloc_reg_set(vc4, ARRAY_SIZE(vc4_regs), true);
119 
120         /* The physical regfiles split us into two classes, with [0] being the
121          * whole space and [1] being the bottom half (for threaded fragment
122          * shaders).
123          */
124         for (int i = 0; i < 2; i++) {
125                 vc4->reg_class_any[i] = ra_alloc_reg_class(vc4->regs);
126                 vc4->reg_class_a_or_b[i] = ra_alloc_reg_class(vc4->regs);
127                 vc4->reg_class_a_or_b_or_acc[i] = ra_alloc_reg_class(vc4->regs);
128                 vc4->reg_class_r4_or_a[i] = ra_alloc_reg_class(vc4->regs);
129                 vc4->reg_class_a[i] = ra_alloc_reg_class(vc4->regs);
130         }
131         vc4->reg_class_r0_r3 = ra_alloc_reg_class(vc4->regs);
132 
133         /* r0-r3 */
134         for (uint32_t i = ACC_INDEX; i < ACC_INDEX + 4; i++) {
135                 ra_class_add_reg(vc4->regs, vc4->reg_class_r0_r3, i);
136                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc[0], i);
137                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc[1], i);
138         }
139 
140         /* R4 gets a special class because it can't be written as a general
141          * purpose register. (it's TMU_NOSWAP as a write address).
142          */
143         for (int i = 0; i < 2; i++) {
144                 ra_class_add_reg(vc4->regs, vc4->reg_class_r4_or_a[i],
145                                  ACC_INDEX + 4);
146                 ra_class_add_reg(vc4->regs, vc4->reg_class_any[i],
147                                  ACC_INDEX + 4);
148         }
149 
150         /* A/B */
151         for (uint32_t i = AB_INDEX; i < AB_INDEX + 64; i ++) {
152                 /* Reserve ra14/rb14 for spilling fixup_raddr_conflict() in
153                  * vc4_qpu_emit.c
154                  */
155                 if (vc4_regs[i].addr == 14)
156                         continue;
157 
158                 ra_class_add_reg(vc4->regs, vc4->reg_class_any[0], i);
159                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b[0], i);
160                 ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc[0], i);
161 
162                 if (vc4_regs[i].addr < 16) {
163                         ra_class_add_reg(vc4->regs, vc4->reg_class_any[1], i);
164                         ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b[1], i);
165                         ra_class_add_reg(vc4->regs, vc4->reg_class_a_or_b_or_acc[1], i);
166                 }
167 
168 
169                 /* A only */
170                 if (((i - AB_INDEX) & 1) == 0) {
171                         ra_class_add_reg(vc4->regs, vc4->reg_class_a[0], i);
172                         ra_class_add_reg(vc4->regs, vc4->reg_class_r4_or_a[0], i);
173 
174                         if (vc4_regs[i].addr < 16) {
175                                 ra_class_add_reg(vc4->regs,
176                                                  vc4->reg_class_a[1], i);
177                                 ra_class_add_reg(vc4->regs,
178                                                  vc4->reg_class_r4_or_a[1], i);
179                         }
180                 }
181         }
182 
183         ra_set_finalize(vc4->regs, NULL);
184 }
185 
186 struct node_to_temp_map {
187         uint32_t temp;
188         uint32_t priority;
189 };
190 
191 static int
node_to_temp_priority(const void * in_a,const void * in_b)192 node_to_temp_priority(const void *in_a, const void *in_b)
193 {
194         const struct node_to_temp_map *a = in_a;
195         const struct node_to_temp_map *b = in_b;
196 
197         return a->priority - b->priority;
198 }
199 
200 #define CLASS_BIT_A			(1 << 0)
201 #define CLASS_BIT_B			(1 << 1)
202 #define CLASS_BIT_R4			(1 << 2)
203 #define CLASS_BIT_R0_R3			(1 << 4)
204 
205 struct vc4_ra_select_callback_data {
206         uint32_t next_acc;
207         uint32_t next_ab;
208 };
209 
210 static unsigned int
vc4_ra_select_callback(struct ra_graph * g,BITSET_WORD * regs,void * data)211 vc4_ra_select_callback(struct ra_graph *g, BITSET_WORD *regs, void *data)
212 {
213         struct vc4_ra_select_callback_data *vc4_ra = data;
214 
215         /* If r4 is available, always choose it -- few other things can go
216          * there, and choosing anything else means inserting a mov.
217          */
218         if (BITSET_TEST(regs, ACC_INDEX + 4))
219                 return ACC_INDEX + 4;
220 
221         /* Choose an accumulator if possible (no delay between write and
222          * read), but round-robin through them to give post-RA instruction
223          * selection more options.
224          */
225         for (int i = 0; i < ACC_COUNT; i++) {
226                 int acc_off = (vc4_ra->next_acc + i) % ACC_COUNT;
227                 int acc = ACC_INDEX + acc_off;
228 
229                 if (BITSET_TEST(regs, acc)) {
230                         vc4_ra->next_acc = acc_off + 1;
231                         return acc;
232                 }
233         }
234 
235         for (int i = 0; i < AB_COUNT; i++) {
236                 int ab_off = (vc4_ra->next_ab + i) % AB_COUNT;
237                 int ab = AB_INDEX + ab_off;
238 
239                 if (BITSET_TEST(regs, ab)) {
240                         vc4_ra->next_ab = ab_off + 1;
241                         return ab;
242                 }
243         }
244 
245         unreachable("RA must pass us at least one possible reg.");
246 }
247 
248 /**
249  * Returns a mapping from QFILE_TEMP indices to struct qpu_regs.
250  *
251  * The return value should be freed by the caller.
252  */
253 struct qpu_reg *
vc4_register_allocate(struct vc4_context * vc4,struct vc4_compile * c)254 vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
255 {
256         struct node_to_temp_map map[c->num_temps];
257         uint32_t temp_to_node[c->num_temps];
258         uint8_t class_bits[c->num_temps];
259         struct qpu_reg *temp_registers = calloc(c->num_temps,
260                                                 sizeof(*temp_registers));
261         struct vc4_ra_select_callback_data callback_data = {
262                 .next_acc = 0,
263                 .next_ab = 0,
264         };
265 
266         /* If things aren't ever written (undefined values), just read from
267          * r0.
268          */
269         for (uint32_t i = 0; i < c->num_temps; i++)
270                 temp_registers[i] = qpu_rn(0);
271 
272         vc4_alloc_reg_set(vc4);
273 
274         struct ra_graph *g = ra_alloc_interference_graph(vc4->regs,
275                                                          c->num_temps);
276 
277         /* Compute the live ranges so we can figure out interference. */
278         qir_calculate_live_intervals(c);
279 
280         ra_set_select_reg_callback(g, vc4_ra_select_callback, &callback_data);
281 
282         for (uint32_t i = 0; i < c->num_temps; i++) {
283                 map[i].temp = i;
284                 map[i].priority = c->temp_end[i] - c->temp_start[i];
285         }
286         qsort(map, c->num_temps, sizeof(map[0]), node_to_temp_priority);
287         for (uint32_t i = 0; i < c->num_temps; i++) {
288                 temp_to_node[map[i].temp] = i;
289         }
290 
291         /* Figure out our register classes and preallocated registers.  We
292          * start with any temp being able to be in any file, then instructions
293          * incrementally remove bits that the temp definitely can't be in.
294          */
295         memset(class_bits,
296                CLASS_BIT_A | CLASS_BIT_B | CLASS_BIT_R4 | CLASS_BIT_R0_R3,
297                sizeof(class_bits));
298 
299         int ip = 0;
300         qir_for_each_inst_inorder(inst, c) {
301                 if (qir_writes_r4(inst)) {
302                         /* This instruction writes r4 (and optionally moves
303                          * its result to a temp), so nothing else can be
304                          * stored in r4 across it.
305                          */
306                         for (int i = 0; i < c->num_temps; i++) {
307                                 if (c->temp_start[i] < ip && c->temp_end[i] > ip)
308                                         class_bits[i] &= ~CLASS_BIT_R4;
309                         }
310 
311                         /* If we're doing a conditional write of something
312                          * writing R4 (math, tex results), then make sure that
313                          * we store in a temp so that we actually
314                          * conditionally move the result.
315                          */
316                         if (inst->cond != QPU_COND_ALWAYS)
317                                 class_bits[inst->dst.index] &= ~CLASS_BIT_R4;
318                 } else {
319                         /* R4 can't be written as a general purpose
320                          * register. (it's TMU_NOSWAP as a write address).
321                          */
322                         if (inst->dst.file == QFILE_TEMP)
323                                 class_bits[inst->dst.index] &= ~CLASS_BIT_R4;
324                 }
325 
326                 switch (inst->op) {
327                 case QOP_FRAG_Z:
328                         ra_set_node_reg(g, temp_to_node[inst->dst.index],
329                                         AB_INDEX + QPU_R_FRAG_PAYLOAD_ZW * 2 + 1);
330                         break;
331 
332                 case QOP_FRAG_W:
333                         ra_set_node_reg(g, temp_to_node[inst->dst.index],
334                                         AB_INDEX + QPU_R_FRAG_PAYLOAD_ZW * 2);
335                         break;
336 
337                 case QOP_ROT_MUL:
338                         assert(inst->src[0].file == QFILE_TEMP);
339                         class_bits[inst->src[0].index] &= CLASS_BIT_R0_R3;
340                         break;
341 
342                 case QOP_THRSW:
343                         /* All accumulators are invalidated across a thread
344                          * switch.
345                          */
346                         for (int i = 0; i < c->num_temps; i++) {
347                                 if (c->temp_start[i] < ip && c->temp_end[i] > ip)
348                                         class_bits[i] &= ~(CLASS_BIT_R0_R3 |
349                                                            CLASS_BIT_R4);
350                         }
351                         break;
352 
353                 default:
354                         break;
355                 }
356 
357                 if (inst->dst.pack && !qir_is_mul(inst)) {
358                         /* The non-MUL pack flags require an A-file dst
359                          * register.
360                          */
361                         class_bits[inst->dst.index] &= CLASS_BIT_A;
362                 }
363 
364                 /* Apply restrictions for src unpacks.  The integer unpacks
365                  * can only be done from regfile A, while float unpacks can be
366                  * either A or R4.
367                  */
368                 for (int i = 0; i < qir_get_nsrc(inst); i++) {
369                         if (inst->src[i].file == QFILE_TEMP &&
370                             inst->src[i].pack) {
371                                 if (qir_is_float_input(inst)) {
372                                         class_bits[inst->src[i].index] &=
373                                                 CLASS_BIT_A | CLASS_BIT_R4;
374                                 } else {
375                                         class_bits[inst->src[i].index] &=
376                                                 CLASS_BIT_A;
377                                 }
378                         }
379                 }
380 
381                 ip++;
382         }
383 
384         for (uint32_t i = 0; i < c->num_temps; i++) {
385                 int node = temp_to_node[i];
386 
387                 switch (class_bits[i]) {
388                 case CLASS_BIT_A | CLASS_BIT_B | CLASS_BIT_R4 | CLASS_BIT_R0_R3:
389                         ra_set_node_class(g, node,
390                                           vc4->reg_class_any[c->fs_threaded]);
391                         break;
392                 case CLASS_BIT_A | CLASS_BIT_B:
393                         ra_set_node_class(g, node,
394                                           vc4->reg_class_a_or_b[c->fs_threaded]);
395                         break;
396                 case CLASS_BIT_A | CLASS_BIT_B | CLASS_BIT_R0_R3:
397                         ra_set_node_class(g, node,
398                                           vc4->reg_class_a_or_b_or_acc[c->fs_threaded]);
399                         break;
400                 case CLASS_BIT_A | CLASS_BIT_R4:
401                         ra_set_node_class(g, node,
402                                           vc4->reg_class_r4_or_a[c->fs_threaded]);
403                         break;
404                 case CLASS_BIT_A:
405                         ra_set_node_class(g, node,
406                                           vc4->reg_class_a[c->fs_threaded]);
407                         break;
408                 case CLASS_BIT_R0_R3:
409                         ra_set_node_class(g, node, vc4->reg_class_r0_r3);
410                         break;
411 
412                 default:
413                         /* DDX/DDY used across thread switched might get us
414                          * here.
415                          */
416                         if (c->fs_threaded) {
417                                 c->failed = true;
418                                 free(temp_registers);
419                                 return NULL;
420                         }
421 
422                         fprintf(stderr, "temp %d: bad class bits: 0x%x\n",
423                                 i, class_bits[i]);
424                         abort();
425                         break;
426                 }
427         }
428 
429         for (uint32_t i = 0; i < c->num_temps; i++) {
430                 for (uint32_t j = i + 1; j < c->num_temps; j++) {
431                         if (!(c->temp_start[i] >= c->temp_end[j] ||
432                               c->temp_start[j] >= c->temp_end[i])) {
433                                 ra_add_node_interference(g,
434                                                          temp_to_node[i],
435                                                          temp_to_node[j]);
436                         }
437                 }
438         }
439 
440         bool ok = ra_allocate(g);
441         if (!ok) {
442                 if (!c->fs_threaded) {
443                         fprintf(stderr, "Failed to register allocate:\n");
444                         qir_dump(c);
445                 }
446 
447                 c->failed = true;
448                 free(temp_registers);
449                 return NULL;
450         }
451 
452         for (uint32_t i = 0; i < c->num_temps; i++) {
453                 temp_registers[i] = vc4_regs[ra_get_node_reg(g, temp_to_node[i])];
454 
455                 /* If the value's never used, just write to the NOP register
456                  * for clarity in debug output.
457                  */
458                 if (c->temp_start[i] == c->temp_end[i])
459                         temp_registers[i] = qpu_ra(QPU_W_NOP);
460         }
461 
462         ralloc_free(g);
463 
464         return temp_registers;
465 }
466