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
26 #include "vc4_context.h"
29 #include "util/ralloc.h"
32 vc4_dump_program(struct vc4_compile
*c
)
34 fprintf(stderr
, "%s prog %d/%d QPU:\n",
35 qir_get_stage_name(c
->stage
),
36 c
->program_id
, c
->variant_id
);
38 for (int i
= 0; i
< c
->qpu_inst_count
; i
++) {
39 fprintf(stderr
, "0x%016"PRIx64
" ", c
->qpu_insts
[i
]);
40 vc4_qpu_disasm(&c
->qpu_insts
[i
], 1);
41 fprintf(stderr
, "\n");
46 queue(struct vc4_compile
*c
, uint64_t inst
)
48 struct queued_qpu_inst
*q
= rzalloc(c
, struct queued_qpu_inst
);
50 list_addtail(&q
->link
, &c
->qpu_inst_list
);
54 last_inst(struct vc4_compile
*c
)
56 struct queued_qpu_inst
*q
=
57 (struct queued_qpu_inst
*)c
->qpu_inst_list
.prev
;
62 set_last_cond_add(struct vc4_compile
*c
, uint32_t cond
)
64 *last_inst(c
) = qpu_set_cond_add(*last_inst(c
), cond
);
68 * Some special registers can be read from either file, which lets us resolve
69 * raddr conflicts without extra MOVs.
72 swap_file(struct qpu_reg
*src
)
77 if (src
->mux
== QPU_MUX_SMALL_IMM
) {
80 if (src
->mux
== QPU_MUX_A
)
93 * This is used to resolve the fact that we might register-allocate two
94 * different operands of an instruction to the same physical register file
95 * even though instructions have only one field for the register file source
98 * In that case, we need to move one to a temporary that can be used in the
99 * instruction, instead. We reserve ra31/rb31 for this purpose.
102 fixup_raddr_conflict(struct vc4_compile
*c
,
104 struct qpu_reg
*src0
, struct qpu_reg
*src1
,
105 struct qinst
*inst
, uint64_t *unpack
)
107 uint32_t mux0
= src0
->mux
== QPU_MUX_SMALL_IMM
? QPU_MUX_B
: src0
->mux
;
108 uint32_t mux1
= src1
->mux
== QPU_MUX_SMALL_IMM
? QPU_MUX_B
: src1
->mux
;
110 if (mux0
<= QPU_MUX_R5
||
112 (src0
->addr
== src1
->addr
&&
113 src0
->mux
== src1
->mux
)) {
117 if (swap_file(src0
) || swap_file(src1
))
120 if (mux0
== QPU_MUX_A
) {
121 /* Make sure we use the same type of MOV as the instruction,
122 * in case of unpacks.
124 if (qir_is_float_input(inst
))
125 queue(c
, qpu_a_FMAX(qpu_rb(31), *src0
, *src0
));
127 queue(c
, qpu_a_MOV(qpu_rb(31), *src0
));
129 /* If we had an unpack on this A-file source, we need to put
130 * it into this MOV, not into the later move from regfile B.
132 if (inst
->src
[0].pack
) {
133 *last_inst(c
) |= *unpack
;
138 queue(c
, qpu_a_MOV(qpu_ra(31), *src0
));
144 set_last_dst_pack(struct vc4_compile
*c
, struct qinst
*inst
)
146 bool had_pm
= *last_inst(c
) & QPU_PM
;
147 bool had_ws
= *last_inst(c
) & QPU_WS
;
148 uint32_t unpack
= QPU_GET_FIELD(*last_inst(c
), QPU_UNPACK
);
153 *last_inst(c
) |= QPU_SET_FIELD(inst
->dst
.pack
, QPU_PACK
);
155 if (qir_is_mul(inst
)) {
156 assert(!unpack
|| had_pm
);
157 *last_inst(c
) |= QPU_PM
;
159 assert(!unpack
|| !had_pm
);
160 assert(!had_ws
); /* dst must be a-file to pack. */
165 vc4_generate_code(struct vc4_context
*vc4
, struct vc4_compile
*c
)
167 struct qpu_reg
*temp_registers
= vc4_register_allocate(vc4
, c
);
168 bool discard
= false;
169 uint32_t inputs_remaining
= c
->num_inputs
;
170 uint32_t vpm_read_fifo_count
= 0;
171 uint32_t vpm_read_offset
= 0;
172 int last_vpm_read_index
= -1;
174 list_inithead(&c
->qpu_inst_list
);
179 /* There's a 4-entry FIFO for VPMVCD reads, each of which can
180 * load up to 16 dwords (4 vec4s) per vertex.
182 while (inputs_remaining
) {
183 uint32_t num_entries
= MIN2(inputs_remaining
, 16);
184 queue(c
, qpu_load_imm_ui(qpu_vrsetup(),
187 ((num_entries
& 0xf) << 20)));
188 inputs_remaining
-= num_entries
;
189 vpm_read_offset
+= num_entries
;
190 vpm_read_fifo_count
++;
192 assert(vpm_read_fifo_count
<= 4);
194 queue(c
, qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00));
200 list_for_each_entry(struct qinst
, qinst
, &c
->instructions
, link
) {
202 fprintf(stderr
, "translating qinst to qpu: ");
203 qir_dump_inst(qinst
);
204 fprintf(stderr
, "\n");
207 static const struct {
210 #define A(name) [QOP_##name] = {QPU_A_##name}
211 #define M(name) [QOP_##name] = {QPU_M_##name}
240 /* If we replicate src[0] out to src[1], this works
241 * out the same as a MOV.
243 [QOP_MOV
] = { QPU_A_OR
},
244 [QOP_FMOV
] = { QPU_A_FMAX
},
245 [QOP_MMOV
] = { QPU_M_V8MIN
},
249 struct qpu_reg src
[4];
250 for (int i
= 0; i
< qir_get_op_nsrc(qinst
->op
); i
++) {
251 int index
= qinst
->src
[i
].index
;
252 switch (qinst
->src
[i
].file
) {
257 src
[i
] = temp_registers
[index
];
258 if (qinst
->src
[i
].pack
) {
260 unpack
== qinst
->src
[i
].pack
);
261 unpack
= QPU_SET_FIELD(qinst
->src
[i
].pack
,
263 if (src
[i
].mux
== QPU_MUX_R4
)
273 case QFILE_SMALL_IMM
:
274 src
[i
].mux
= QPU_MUX_SMALL_IMM
;
275 src
[i
].addr
= qpu_encode_small_immediate(qinst
->src
[i
].index
);
276 /* This should only have returned a valid
277 * small immediate field, not ~0 for failure.
279 assert(src
[i
].addr
<= 47);
282 assert((int)qinst
->src
[i
].index
>=
283 last_vpm_read_index
);
284 (void)last_vpm_read_index
;
285 last_vpm_read_index
= qinst
->src
[i
].index
;
286 src
[i
] = qpu_ra(QPU_R_VPM
);
292 switch (qinst
->dst
.file
) {
294 dst
= qpu_ra(QPU_W_NOP
);
297 dst
= temp_registers
[qinst
->dst
.index
];
300 dst
= qpu_ra(QPU_W_VPM
);
304 case QFILE_SMALL_IMM
:
305 assert(!"not reached");
316 queue(c
, qpu_a_MOV(dst
, src
[0]) | unpack
);
317 set_last_cond_add(c
, qinst
->op
- QOP_SEL_X_0_ZS
+
320 queue(c
, qpu_a_XOR(dst
, qpu_r0(), qpu_r0()));
321 set_last_cond_add(c
, ((qinst
->op
- QOP_SEL_X_0_ZS
) ^
331 queue(c
, qpu_a_MOV(dst
, src
[0]));
332 if (qinst
->src
[0].pack
)
333 *(last_inst(c
)) |= unpack
;
334 set_last_cond_add(c
, qinst
->op
- QOP_SEL_X_Y_ZS
+
337 queue(c
, qpu_a_MOV(dst
, src
[1]));
338 if (qinst
->src
[1].pack
)
339 *(last_inst(c
)) |= unpack
;
340 set_last_cond_add(c
, ((qinst
->op
- QOP_SEL_X_Y_ZS
) ^
351 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP
),
355 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT
),
359 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP
),
363 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG
),
370 if (dst
.mux
!= QPU_MUX_R4
)
371 queue(c
, qpu_a_MOV(dst
, qpu_r4()));
376 queue(c
, qpu_a_ITOF(dst
,
377 qpu_ra(QPU_R_XY_PIXEL_COORD
)));
381 queue(c
, qpu_a_ITOF(dst
,
382 qpu_rb(QPU_R_XY_PIXEL_COORD
)));
385 case QOP_FRAG_REV_FLAG
:
386 queue(c
, qpu_a_ITOF(dst
,
387 qpu_rb(QPU_R_MS_REV_FLAGS
)));
391 src
[1] = qpu_ra(QPU_R_MS_REV_FLAGS
);
392 fixup_raddr_conflict(c
, dst
, &src
[0], &src
[1],
394 queue(c
, qpu_a_AND(qpu_ra(QPU_W_MS_FLAGS
),
395 src
[0], src
[1]) | unpack
);
400 /* QOP_FRAG_Z/W don't emit instructions, just allocate
401 * the register to the Z/W payload.
405 case QOP_TLB_DISCARD_SETUP
:
407 queue(c
, qpu_a_MOV(src
[0], src
[0]) | unpack
);
408 *last_inst(c
) |= QPU_SF
;
411 case QOP_TLB_STENCIL_SETUP
:
413 queue(c
, qpu_a_MOV(qpu_ra(QPU_W_TLB_STENCIL_SETUP
),
417 case QOP_TLB_Z_WRITE
:
418 queue(c
, qpu_a_MOV(qpu_ra(QPU_W_TLB_Z
),
421 set_last_cond_add(c
, QPU_COND_ZS
);
425 case QOP_TLB_COLOR_READ
:
427 *last_inst(c
) = qpu_set_sig(*last_inst(c
),
430 if (dst
.mux
!= QPU_MUX_R4
)
431 queue(c
, qpu_a_MOV(dst
, qpu_r4()));
434 case QOP_TLB_COLOR_WRITE
:
435 queue(c
, qpu_a_MOV(qpu_tlbc(), src
[0]) | unpack
);
437 set_last_cond_add(c
, QPU_COND_ZS
);
441 case QOP_TLB_COLOR_WRITE_MS
:
442 queue(c
, qpu_a_MOV(qpu_tlbc_ms(), src
[0]));
444 set_last_cond_add(c
, QPU_COND_ZS
);
449 queue(c
, qpu_a_FADD(dst
, src
[0], qpu_r5()) | unpack
);
456 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_TMU0_S
+
457 (qinst
->op
- QOP_TEX_S
)),
462 fixup_raddr_conflict(c
, dst
, &src
[0], &src
[1],
464 queue(c
, qpu_a_ADD(qpu_rb(QPU_W_TMU0_S
),
465 src
[0], src
[1]) | unpack
);
470 *last_inst(c
) = qpu_set_sig(*last_inst(c
),
472 if (dst
.mux
!= QPU_MUX_R4
)
473 queue(c
, qpu_a_MOV(dst
, qpu_r4()));
477 assert(qinst
->op
< ARRAY_SIZE(translate
));
478 assert(translate
[qinst
->op
].op
!= 0); /* NOPs */
480 /* Skip emitting the MOV if it's a no-op. */
481 if (qir_is_raw_mov(qinst
) &&
482 dst
.mux
== src
[0].mux
&& dst
.addr
== src
[0].addr
) {
486 /* If we have only one source, put it in the second
487 * argument slot as well so that we don't take up
488 * another raddr just to get unused data.
490 if (qir_get_op_nsrc(qinst
->op
) == 1)
493 fixup_raddr_conflict(c
, dst
, &src
[0], &src
[1],
496 if (qir_is_mul(qinst
)) {
497 queue(c
, qpu_m_alu2(translate
[qinst
->op
].op
,
499 src
[0], src
[1]) | unpack
);
501 queue(c
, qpu_a_alu2(translate
[qinst
->op
].op
,
503 src
[0], src
[1]) | unpack
);
505 set_last_dst_pack(c
, qinst
);
511 assert(!qir_is_multi_instruction(qinst
));
512 *last_inst(c
) |= QPU_SF
;
516 uint32_t cycles
= qpu_schedule_instructions(c
);
517 uint32_t inst_count_at_schedule_time
= c
->qpu_inst_count
;
519 /* thread end can't have VPM write or read */
520 if (QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
521 QPU_WADDR_ADD
) == QPU_W_VPM
||
522 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
523 QPU_WADDR_MUL
) == QPU_W_VPM
||
524 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
525 QPU_RADDR_A
) == QPU_R_VPM
||
526 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
527 QPU_RADDR_B
) == QPU_R_VPM
) {
528 qpu_serialize_one_inst(c
, qpu_NOP());
531 /* thread end can't have uniform read */
532 if (QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
533 QPU_RADDR_A
) == QPU_R_UNIF
||
534 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
535 QPU_RADDR_B
) == QPU_R_UNIF
) {
536 qpu_serialize_one_inst(c
, qpu_NOP());
539 /* thread end can't have TLB operations */
540 if (qpu_inst_is_tlb(c
->qpu_insts
[c
->qpu_inst_count
- 1]))
541 qpu_serialize_one_inst(c
, qpu_NOP());
543 c
->qpu_insts
[c
->qpu_inst_count
- 1] =
544 qpu_set_sig(c
->qpu_insts
[c
->qpu_inst_count
- 1],
546 qpu_serialize_one_inst(c
, qpu_NOP());
547 qpu_serialize_one_inst(c
, qpu_NOP());
554 c
->qpu_insts
[c
->qpu_inst_count
- 1] =
555 qpu_set_sig(c
->qpu_insts
[c
->qpu_inst_count
- 1],
556 QPU_SIG_SCOREBOARD_UNLOCK
);
560 cycles
+= c
->qpu_inst_count
- inst_count_at_schedule_time
;
562 if (vc4_debug
& VC4_DEBUG_SHADERDB
) {
563 fprintf(stderr
, "SHADER-DB: %s prog %d/%d: %d estimated cycles\n",
564 qir_get_stage_name(c
->stage
),
565 c
->program_id
, c
->variant_id
,
569 if (vc4_debug
& VC4_DEBUG_QPU
)
572 vc4_qpu_validate(c
->qpu_insts
, c
->qpu_inst_count
);
574 free(temp_registers
);