2 * Copyright © 2014 Broadcom
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:
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
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
27 #include "vc4_context.h"
32 vc4_dump_program(struct qcompile *c)
34 fprintf(stderr, "%s:\n", qir_get_stage_name(c->stage));
36 for (int i = 0; i < c->qpu_inst_count; i++) {
37 fprintf(stderr, "0x%016"PRIx64" ", c->qpu_insts[i]);
38 vc4_qpu_disasm(&c->qpu_insts[i], 1);
39 fprintf(stderr, "\n");
43 struct queued_qpu_inst {
44 struct simple_node link;
49 queue(struct qcompile *c, uint64_t inst)
51 struct queued_qpu_inst *q = calloc(1, sizeof(*q));
53 insert_at_tail(&c->qpu_inst_list, &q->link);
57 last_inst(struct qcompile *c)
59 struct queued_qpu_inst *q =
60 (struct queued_qpu_inst *)last_elem(&c->qpu_inst_list);
65 * This is used to resolve the fact that we might register-allocate two
66 * different operands of an instruction to the same physical register file
67 * even though instructions have only one field for the register file source
70 * In that case, we need to move one to a temporary that can be used in the
71 * instruction, instead.
74 fixup_raddr_conflict(struct qcompile *c,
75 struct qpu_reg src0, struct qpu_reg *src1)
77 if ((src0.mux == QPU_MUX_A || src0.mux == QPU_MUX_B) &&
78 (src1->mux == QPU_MUX_A || src1->mux == QPU_MUX_B) &&
79 src0.addr != src1->addr) {
80 queue(c, qpu_inst(qpu_a_MOV(qpu_r3(), *src1),
87 serialize_one_inst(struct qcompile *c, uint64_t inst)
89 if (c->qpu_inst_count >= c->qpu_inst_size) {
90 c->qpu_inst_size = MAX2(16, c->qpu_inst_size * 2);
91 c->qpu_insts = realloc(c->qpu_insts,
92 c->qpu_inst_size * sizeof(uint64_t));
94 c->qpu_insts[c->qpu_inst_count++] = inst;
98 serialize_insts(struct qcompile *c)
100 int last_sfu_write = -10;
102 while (!is_empty_list(&c->qpu_inst_list)) {
103 struct queued_qpu_inst *q =
104 (struct queued_qpu_inst *)first_elem(&c->qpu_inst_list);
105 uint32_t last_waddr_a = QPU_W_NOP, last_waddr_b = QPU_W_NOP;
106 uint32_t raddr_a = QPU_GET_FIELD(q->inst, QPU_RADDR_A);
107 uint32_t raddr_b = QPU_GET_FIELD(q->inst, QPU_RADDR_B);
109 if (c->qpu_inst_count > 0) {
110 uint64_t last_inst = c->qpu_insts[c->qpu_inst_count -
112 uint32_t last_waddr_add = QPU_GET_FIELD(last_inst,
114 uint32_t last_waddr_mul = QPU_GET_FIELD(last_inst,
117 if (last_inst & QPU_WS) {
118 last_waddr_a = last_waddr_mul;
119 last_waddr_b = last_waddr_add;
121 last_waddr_a = last_waddr_add;
122 last_waddr_b = last_waddr_mul;
126 uint32_t src_muxes[] = {
127 QPU_GET_FIELD(q->inst, QPU_ADD_A),
128 QPU_GET_FIELD(q->inst, QPU_ADD_B),
129 QPU_GET_FIELD(q->inst, QPU_MUL_A),
130 QPU_GET_FIELD(q->inst, QPU_MUL_B),
133 /* "An instruction must not read from a location in physical
134 * regfile A or B that was written to by the previous
137 bool needs_raddr_vs_waddr_nop = false;
138 bool reads_r4 = false;
139 for (int i = 0; i < ARRAY_SIZE(src_muxes); i++) {
141 src_muxes[i] == QPU_MUX_A &&
142 last_waddr_a == raddr_a) ||
144 src_muxes[i] == QPU_MUX_B &&
145 last_waddr_b == raddr_b)) {
146 needs_raddr_vs_waddr_nop = true;
148 if (src_muxes[i] == QPU_MUX_R4)
152 if (needs_raddr_vs_waddr_nop) {
153 serialize_one_inst(c, qpu_inst(qpu_a_NOP(),
157 /* "After an SFU lookup instruction, accumulator r4 must not
158 * be read in the following two instructions. Any other
159 * instruction that results in r4 being written (that is, TMU
160 * read, TLB read, SFU lookup) cannot occur in the two
161 * instructions following an SFU lookup."
164 while (c->qpu_inst_count - last_sfu_write < 3) {
165 serialize_one_inst(c, qpu_inst(qpu_a_NOP(),
170 uint32_t waddr_a = QPU_GET_FIELD(q->inst, QPU_WADDR_ADD);
171 uint32_t waddr_m = QPU_GET_FIELD(q->inst, QPU_WADDR_MUL);
172 if ((waddr_a >= QPU_W_SFU_RECIP && waddr_a <= QPU_W_SFU_LOG) ||
173 (waddr_m >= QPU_W_SFU_RECIP && waddr_m <= QPU_W_SFU_LOG)) {
174 last_sfu_write = c->qpu_inst_count;
177 serialize_one_inst(c, q->inst);
179 remove_from_list(&q->link);
185 vc4_generate_code(struct qcompile *c)
187 struct qpu_reg allocate_to_qpu_reg[3 + 32 + 32];
188 bool reg_in_use[ARRAY_SIZE(allocate_to_qpu_reg)];
189 int *reg_allocated = calloc(c->num_temps, sizeof(*reg_allocated));
190 int *reg_uses_remaining =
191 calloc(c->num_temps, sizeof(*reg_uses_remaining));
193 for (int i = 0; i < ARRAY_SIZE(reg_in_use); i++)
194 reg_in_use[i] = false;
195 for (int i = 0; i < c->num_temps; i++)
196 reg_allocated[i] = -1;
197 for (int i = 0; i < 3; i++)
198 allocate_to_qpu_reg[i] = qpu_rn(i);
199 for (int i = 0; i < 32; i++)
200 allocate_to_qpu_reg[i + 3] = qpu_ra(i);
201 for (int i = 0; i < 32; i++)
202 allocate_to_qpu_reg[i + 3 + 32] = qpu_rb(i);
204 make_empty_list(&c->qpu_inst_list);
206 struct simple_node *node;
207 foreach(node, &c->instructions) {
208 struct qinst *qinst = (struct qinst *)node;
210 if (qinst->dst.file == QFILE_TEMP)
211 reg_uses_remaining[qinst->dst.index]++;
212 for (int i = 0; i < qir_get_op_nsrc(qinst->op); i++) {
213 if (qinst->src[i].file == QFILE_TEMP)
214 reg_uses_remaining[qinst->src[i].index]++;
221 queue(c, qpu_load_imm_ui(qpu_vrsetup(), 0x00401a00));
222 queue(c, qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00));
228 foreach(node, &c->instructions) {
229 struct qinst *qinst = (struct qinst *)node;
232 fprintf(stderr, "translating qinst to qpu: ");
233 qir_dump_inst(qinst);
234 fprintf(stderr, "\n");
237 static const struct {
241 #define A(name) [QOP_##name] = {QPU_A_##name, false}
242 #define M(name) [QOP_##name] = {QPU_M_##name, true}
254 static const uint32_t compareflags[] = {
255 [QOP_SEQ - QOP_SEQ] = QPU_COND_ZS,
256 [QOP_SNE - QOP_SEQ] = QPU_COND_ZC,
257 [QOP_SLT - QOP_SEQ] = QPU_COND_NS,
258 [QOP_SGE - QOP_SEQ] = QPU_COND_NC,
261 struct qpu_reg src[4];
262 for (int i = 0; i < qir_get_op_nsrc(qinst->op); i++) {
263 int index = qinst->src[i].index;
264 switch (qinst->src[i].file) {
269 if (reg_allocated[index] == -1) {
270 fprintf(stderr, "undefined reg use: ");
271 qir_dump_inst(qinst);
272 fprintf(stderr, "\n");
276 src[i] = allocate_to_qpu_reg[reg_allocated[index]];
277 reg_uses_remaining[index]--;
278 if (reg_uses_remaining[index] == 0)
279 reg_in_use[reg_allocated[index]] = false;
292 switch (qinst->dst.file) {
294 dst = qpu_ra(QPU_W_NOP);
298 if (reg_allocated[qinst->dst.index] == -1) {
301 alloc < ARRAY_SIZE(reg_in_use);
303 /* The pack flags require an A-file register. */
304 if (qinst->op == QOP_PACK_SCALED &&
305 allocate_to_qpu_reg[alloc].mux != QPU_MUX_A) {
309 if (!reg_in_use[alloc])
312 assert(alloc != ARRAY_SIZE(reg_in_use) && "need better reg alloc");
313 reg_in_use[alloc] = true;
314 reg_allocated[qinst->dst.index] = alloc;
317 dst = allocate_to_qpu_reg[reg_allocated[qinst->dst.index]];
319 reg_uses_remaining[qinst->dst.index]--;
320 if (reg_uses_remaining[qinst->dst.index] == 0) {
321 reg_in_use[reg_allocated[qinst->dst.index]] =
328 assert(!"not reached");
334 /* Skip emitting the MOV if it's a no-op. */
335 if (dst.mux == QPU_MUX_A || dst.mux == QPU_MUX_B ||
336 dst.mux != src[0].mux || dst.addr != src[0].addr) {
337 queue(c, qpu_inst(qpu_a_MOV(dst, src[0]),
343 queue(c, qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_NOP),
346 *last_inst(c) |= QPU_SF;
348 if (dst.mux <= QPU_MUX_R3) {
349 fixup_raddr_conflict(c, src[1], &src[2]);
350 queue(c, qpu_inst(qpu_a_MOV(dst, src[1]),
351 qpu_m_MOV(dst, src[2])));
352 *last_inst(c) = ((*last_inst(c) & ~(QPU_COND_ADD_MASK |
354 | QPU_SET_FIELD(QPU_COND_NS,
356 | QPU_SET_FIELD(QPU_COND_NC,
359 if (dst.mux == src[1].mux &&
360 dst.addr == src[1].addr) {
361 queue(c, qpu_inst(qpu_a_MOV(dst, src[1]),
364 queue(c, qpu_inst(qpu_a_MOV(dst, src[2]),
366 *last_inst(c) = ((*last_inst(c) & ~(QPU_COND_ADD_MASK))
367 | QPU_SET_FIELD(QPU_COND_NC,
370 queue(c, qpu_inst(qpu_a_MOV(dst, src[2]),
373 queue(c, qpu_inst(qpu_a_MOV(dst, src[1]),
375 *last_inst(c) = ((*last_inst(c) & ~(QPU_COND_ADD_MASK))
376 | QPU_SET_FIELD(QPU_COND_NS,
386 fixup_raddr_conflict(c, src[0], &src[1]);
387 queue(c, qpu_inst(qpu_a_SUB(qpu_ra(QPU_W_NOP),
390 *last_inst(c) |= QPU_SF;
392 queue(c, qpu_load_imm_f(dst, 0.0));
393 queue(c, qpu_load_imm_f(dst, 1.0));
394 *last_inst(c) = ((*last_inst(c) & ~QPU_COND_ADD_MASK)
395 | QPU_SET_FIELD(compareflags[qinst->op - QOP_SEQ],
401 queue(c, qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_VPM), src[0]),
406 queue(c, qpu_inst(qpu_a_MOV(dst, qpu_ra(QPU_R_VPM)),
416 queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP),
421 queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT),
426 queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP),
431 queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG),
439 queue(c, qpu_inst(qpu_a_MOV(dst, qpu_r4()),
444 case QOP_PACK_COLORS:
445 for (int i = 0; i < 4; i++) {
446 queue(c, qpu_inst(qpu_a_NOP(),
447 qpu_m_MOV(qpu_r3(), src[i])));
448 *last_inst(c) |= QPU_PM;
449 *last_inst(c) |= QPU_SET_FIELD(QPU_PACK_MUL_8A + i,
453 queue(c, qpu_inst(qpu_a_MOV(dst, qpu_r3()),
458 case QOP_TLB_COLOR_WRITE:
459 queue(c, qpu_inst(qpu_a_MOV(qpu_tlbc(),
465 queue(c, qpu_inst(qpu_a_FADD(dst,
470 case QOP_PACK_SCALED:
471 queue(c, qpu_inst(qpu_a_MOV(dst, src[0]),
473 *last_inst(c) |= QPU_SET_FIELD(QPU_PACK_A_16A,
476 queue(c, qpu_inst(qpu_a_MOV(dst, src[1]),
478 *last_inst(c) |= QPU_SET_FIELD(QPU_PACK_A_16B,
484 assert(qinst->op < ARRAY_SIZE(translate));
485 assert(translate[qinst->op].op != 0); /* NOPs */
487 /* If we have only one source, put it in the second
488 * argument slot as well so that we don't take up
489 * another raddr just to get unused data.
491 if (qir_get_op_nsrc(qinst->op) == 1)
494 fixup_raddr_conflict(c, src[0], &src[1]);
496 if (translate[qinst->op].is_mul) {
497 queue(c, qpu_inst(qpu_a_NOP(),
498 qpu_m_alu2(translate[qinst->op].op,
502 queue(c, qpu_inst(qpu_a_alu2(translate[qinst->op].op,
513 /* thread end can't have VPM write */
514 if (QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
515 QPU_WADDR_ADD) == QPU_W_VPM ||
516 QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
517 QPU_WADDR_MUL) == QPU_W_VPM) {
518 serialize_one_inst(c, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
521 c->qpu_insts[c->qpu_inst_count - 1] =
522 qpu_set_sig(c->qpu_insts[c->qpu_inst_count - 1],
524 serialize_one_inst(c, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
525 serialize_one_inst(c, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
532 c->qpu_insts[2] = qpu_set_sig(c->qpu_insts[2],
533 QPU_SIG_WAIT_FOR_SCOREBOARD);
534 c->qpu_insts[c->qpu_inst_count - 1] =
535 qpu_set_sig(c->qpu_insts[c->qpu_inst_count - 1],
536 QPU_SIG_SCOREBOARD_UNLOCK);
540 if (vc4_debug & VC4_DEBUG_QPU)
543 vc4_qpu_validate(c->qpu_insts, c->qpu_inst_count);