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"
31 vc4_dump_program(struct qcompile
*c
)
33 fprintf(stderr
, "%s:\n", qir_get_stage_name(c
->stage
));
35 for (int i
= 0; i
< c
->qpu_inst_count
; i
++) {
36 fprintf(stderr
, "0x%016"PRIx64
" ", c
->qpu_insts
[i
]);
37 vc4_qpu_disasm(&c
->qpu_insts
[i
], 1);
38 fprintf(stderr
, "\n");
42 struct queued_qpu_inst
{
43 struct simple_node link
;
48 queue(struct qcompile
*c
, uint64_t inst
)
50 struct queued_qpu_inst
*q
= calloc(1, sizeof(*q
));
52 insert_at_tail(&c
->qpu_inst_list
, &q
->link
);
56 last_inst(struct qcompile
*c
)
58 struct queued_qpu_inst
*q
=
59 (struct queued_qpu_inst
*)last_elem(&c
->qpu_inst_list
);
64 * This is used to resolve the fact that we might register-allocate two
65 * different operands of an instruction to the same physical register file
66 * even though instructions have only one field for the register file source
69 * In that case, we need to move one to a temporary that can be used in the
70 * instruction, instead.
73 fixup_raddr_conflict(struct qcompile
*c
,
74 struct qpu_reg src0
, struct qpu_reg
*src1
)
76 if ((src0
.mux
== QPU_MUX_A
|| src0
.mux
== QPU_MUX_B
) &&
77 (src1
->mux
== QPU_MUX_A
|| src1
->mux
== QPU_MUX_B
) &&
78 src0
.addr
!= src1
->addr
) {
79 queue(c
, qpu_inst(qpu_a_MOV(qpu_r3(), *src1
),
86 serialize_one_inst(struct qcompile
*c
, uint64_t inst
)
88 if (c
->qpu_inst_count
>= c
->qpu_inst_size
) {
89 c
->qpu_inst_size
= MAX2(16, c
->qpu_inst_size
* 2);
90 c
->qpu_insts
= realloc(c
->qpu_insts
,
91 c
->qpu_inst_size
* sizeof(uint64_t));
93 c
->qpu_insts
[c
->qpu_inst_count
++] = inst
;
97 serialize_insts(struct qcompile
*c
)
99 int last_sfu_write
= -10;
100 bool scoreboard_wait_emitted
= false;
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 /* "A scoreboard wait must not occur in the first two
178 * instructions of a fragment shader. This is either the
179 * explicit Wait for Scoreboard signal or an implicit wait
180 * with the first tile-buffer read or write instruction."
182 if (!scoreboard_wait_emitted
&&
183 (waddr_a
== QPU_W_TLB_Z
|| waddr_m
== QPU_W_TLB_Z
||
184 waddr_a
== QPU_W_TLB_COLOR_MS
||
185 waddr_m
== QPU_W_TLB_COLOR_MS
||
186 waddr_a
== QPU_W_TLB_COLOR_ALL
||
187 waddr_m
== QPU_W_TLB_COLOR_ALL
||
188 QPU_GET_FIELD(q
->inst
, QPU_SIG
) == QPU_SIG_COLOR_LOAD
)) {
189 while (c
->qpu_inst_count
< 3 ||
190 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
191 QPU_SIG
) != QPU_SIG_NONE
) {
192 serialize_one_inst(c
, qpu_inst(qpu_a_NOP(),
195 c
->qpu_insts
[c
->qpu_inst_count
- 1] =
196 qpu_set_sig(c
->qpu_insts
[c
->qpu_inst_count
- 1],
197 QPU_SIG_WAIT_FOR_SCOREBOARD
);
198 scoreboard_wait_emitted
= true;
201 serialize_one_inst(c
, q
->inst
);
203 remove_from_list(&q
->link
);
209 vc4_generate_code(struct qcompile
*c
)
211 struct qpu_reg allocate_to_qpu_reg
[3 + 32 + 32];
212 bool reg_in_use
[ARRAY_SIZE(allocate_to_qpu_reg
)];
213 int *reg_allocated
= calloc(c
->num_temps
, sizeof(*reg_allocated
));
214 int *reg_uses_remaining
=
215 calloc(c
->num_temps
, sizeof(*reg_uses_remaining
));
217 for (int i
= 0; i
< ARRAY_SIZE(reg_in_use
); i
++)
218 reg_in_use
[i
] = false;
219 for (int i
= 0; i
< c
->num_temps
; i
++)
220 reg_allocated
[i
] = -1;
221 for (int i
= 0; i
< 3; i
++)
222 allocate_to_qpu_reg
[i
] = qpu_rn(i
);
223 for (int i
= 0; i
< 32; i
++)
224 allocate_to_qpu_reg
[i
+ 3] = qpu_ra(i
);
225 for (int i
= 0; i
< 32; i
++)
226 allocate_to_qpu_reg
[i
+ 3 + 32] = qpu_rb(i
);
228 make_empty_list(&c
->qpu_inst_list
);
230 struct simple_node
*node
;
231 foreach(node
, &c
->instructions
) {
232 struct qinst
*qinst
= (struct qinst
*)node
;
234 if (qinst
->dst
.file
== QFILE_TEMP
)
235 reg_uses_remaining
[qinst
->dst
.index
]++;
236 for (int i
= 0; i
< qir_get_op_nsrc(qinst
->op
); i
++) {
237 if (qinst
->src
[i
].file
== QFILE_TEMP
)
238 reg_uses_remaining
[qinst
->src
[i
].index
]++;
240 if (qinst
->op
== QOP_TLB_PASSTHROUGH_Z_WRITE
||
241 qinst
->op
== QOP_FRAG_Z
)
242 reg_in_use
[3 + 32 + QPU_R_FRAG_PAYLOAD_ZW
] = true;
248 queue(c
, qpu_load_imm_ui(qpu_vrsetup(),
250 0x00100000 * c
->num_inputs
)));
251 queue(c
, qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00));
257 foreach(node
, &c
->instructions
) {
258 struct qinst
*qinst
= (struct qinst
*)node
;
261 fprintf(stderr
, "translating qinst to qpu: ");
262 qir_dump_inst(qinst
);
263 fprintf(stderr
, "\n");
266 static const struct {
270 #define A(name) [QOP_##name] = {QPU_A_##name, false}
271 #define M(name) [QOP_##name] = {QPU_M_##name, true}
284 static const uint32_t compareflags
[] = {
285 [QOP_SEQ
- QOP_SEQ
] = QPU_COND_ZS
,
286 [QOP_SNE
- QOP_SEQ
] = QPU_COND_ZC
,
287 [QOP_SLT
- QOP_SEQ
] = QPU_COND_NS
,
288 [QOP_SGE
- QOP_SEQ
] = QPU_COND_NC
,
291 struct qpu_reg src
[4];
292 for (int i
= 0; i
< qir_get_op_nsrc(qinst
->op
); i
++) {
293 int index
= qinst
->src
[i
].index
;
294 switch (qinst
->src
[i
].file
) {
299 if (reg_allocated
[index
] == -1) {
300 fprintf(stderr
, "undefined reg use: ");
301 qir_dump_inst(qinst
);
302 fprintf(stderr
, "\n");
306 src
[i
] = allocate_to_qpu_reg
[reg_allocated
[index
]];
307 reg_uses_remaining
[index
]--;
308 if (reg_uses_remaining
[index
] == 0)
309 reg_in_use
[reg_allocated
[index
]] = false;
322 switch (qinst
->dst
.file
) {
324 dst
= qpu_ra(QPU_W_NOP
);
328 if (reg_allocated
[qinst
->dst
.index
] == -1) {
331 alloc
< ARRAY_SIZE(reg_in_use
);
333 /* The pack flags require an A-file register. */
334 if (qinst
->op
== QOP_PACK_SCALED
&&
335 allocate_to_qpu_reg
[alloc
].mux
!= QPU_MUX_A
) {
339 if (!reg_in_use
[alloc
])
342 assert(alloc
!= ARRAY_SIZE(reg_in_use
) && "need better reg alloc");
343 reg_in_use
[alloc
] = true;
344 reg_allocated
[qinst
->dst
.index
] = alloc
;
347 dst
= allocate_to_qpu_reg
[reg_allocated
[qinst
->dst
.index
]];
349 reg_uses_remaining
[qinst
->dst
.index
]--;
350 if (reg_uses_remaining
[qinst
->dst
.index
] == 0) {
351 reg_in_use
[reg_allocated
[qinst
->dst
.index
]] =
358 assert(!"not reached");
364 /* Skip emitting the MOV if it's a no-op. */
365 if (dst
.mux
== QPU_MUX_A
|| dst
.mux
== QPU_MUX_B
||
366 dst
.mux
!= src
[0].mux
|| dst
.addr
!= src
[0].addr
) {
367 queue(c
, qpu_inst(qpu_a_MOV(dst
, src
[0]),
373 queue(c
, qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_NOP
),
376 *last_inst(c
) |= QPU_SF
;
378 if (dst
.mux
<= QPU_MUX_R3
) {
379 fixup_raddr_conflict(c
, src
[1], &src
[2]);
380 queue(c
, qpu_inst(qpu_a_MOV(dst
, src
[1]),
381 qpu_m_MOV(dst
, src
[2])));
382 *last_inst(c
) = ((*last_inst(c
) & ~(QPU_COND_ADD_MASK
|
384 | QPU_SET_FIELD(QPU_COND_NS
,
386 | QPU_SET_FIELD(QPU_COND_NC
,
389 if (dst
.mux
== src
[1].mux
&&
390 dst
.addr
== src
[1].addr
) {
391 queue(c
, qpu_inst(qpu_a_MOV(dst
, src
[1]),
394 queue(c
, qpu_inst(qpu_a_MOV(dst
, src
[2]),
396 *last_inst(c
) = ((*last_inst(c
) & ~(QPU_COND_ADD_MASK
))
397 | QPU_SET_FIELD(QPU_COND_NC
,
400 queue(c
, qpu_inst(qpu_a_MOV(dst
, src
[2]),
403 queue(c
, qpu_inst(qpu_a_MOV(dst
, src
[1]),
405 *last_inst(c
) = ((*last_inst(c
) & ~(QPU_COND_ADD_MASK
))
406 | QPU_SET_FIELD(QPU_COND_NS
,
416 fixup_raddr_conflict(c
, src
[0], &src
[1]);
417 queue(c
, qpu_inst(qpu_a_SUB(qpu_ra(QPU_W_NOP
),
420 *last_inst(c
) |= QPU_SF
;
422 queue(c
, qpu_load_imm_f(dst
, 0.0));
423 queue(c
, qpu_load_imm_f(dst
, 1.0));
424 *last_inst(c
) = ((*last_inst(c
) & ~QPU_COND_ADD_MASK
)
425 | QPU_SET_FIELD(compareflags
[qinst
->op
- QOP_SEQ
],
431 queue(c
, qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_VPM
), src
[0]),
436 queue(c
, qpu_inst(qpu_a_MOV(dst
, qpu_ra(QPU_R_VPM
)),
446 queue(c
, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP
),
451 queue(c
, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT
),
456 queue(c
, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP
),
461 queue(c
, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG
),
469 queue(c
, qpu_inst(qpu_a_MOV(dst
, qpu_r4()),
474 case QOP_PACK_COLORS
:
475 for (int i
= 0; i
< 4; i
++) {
476 queue(c
, qpu_inst(qpu_a_NOP(),
477 qpu_m_MOV(qpu_r3(), src
[i
])));
478 *last_inst(c
) |= QPU_PM
;
479 *last_inst(c
) |= QPU_SET_FIELD(QPU_PACK_MUL_8A
+ i
,
483 queue(c
, qpu_inst(qpu_a_MOV(dst
, qpu_r3()),
489 queue(c
, qpu_inst(qpu_a_ITOF(dst
,
490 qpu_ra(QPU_R_XY_PIXEL_COORD
)),
495 queue(c
, qpu_inst(qpu_a_ITOF(dst
,
496 qpu_rb(QPU_R_XY_PIXEL_COORD
)),
501 queue(c
, qpu_inst(qpu_a_ITOF(dst
,
502 qpu_rb(QPU_R_FRAG_PAYLOAD_ZW
)),
507 queue(c
, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP
),
508 qpu_ra(QPU_R_FRAG_PAYLOAD_ZW
)),
511 queue(c
, qpu_inst(qpu_a_MOV(dst
, qpu_r4()),
515 case QOP_TLB_PASSTHROUGH_Z_WRITE
:
516 queue(c
, qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_TLB_Z
),
517 qpu_rb(QPU_R_FRAG_PAYLOAD_ZW
)),
521 case QOP_TLB_COLOR_READ
:
522 queue(c
, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
523 *last_inst(c
) = qpu_set_sig(*last_inst(c
),
528 case QOP_TLB_COLOR_WRITE
:
529 queue(c
, qpu_inst(qpu_a_MOV(qpu_tlbc(),
535 queue(c
, qpu_inst(qpu_a_FADD(dst
,
540 case QOP_PACK_SCALED
: {
541 uint64_t a
= (qpu_inst(qpu_a_MOV(dst
, src
[0]),
543 QPU_SET_FIELD(QPU_PACK_A_16A
,
545 uint64_t b
= (qpu_inst(qpu_a_MOV(dst
, src
[1]),
547 QPU_SET_FIELD(QPU_PACK_A_16B
,
550 if (dst
.mux
== src
[1].mux
&& dst
.addr
== src
[1].addr
) {
564 queue(c
, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_TMU0_S
+
572 queue(c
, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
573 *last_inst(c
) = qpu_set_sig(*last_inst(c
),
578 case QOP_R4_UNPACK_A
:
579 case QOP_R4_UNPACK_B
:
580 case QOP_R4_UNPACK_C
:
581 case QOP_R4_UNPACK_D
:
582 queue(c
, qpu_inst(qpu_a_MOV(dst
, qpu_r4()),
584 *last_inst(c
) |= QPU_PM
;
585 *last_inst(c
) |= QPU_SET_FIELD(QPU_UNPACK_R4_8A
+
593 assert(qinst
->op
< ARRAY_SIZE(translate
));
594 assert(translate
[qinst
->op
].op
!= 0); /* NOPs */
596 /* If we have only one source, put it in the second
597 * argument slot as well so that we don't take up
598 * another raddr just to get unused data.
600 if (qir_get_op_nsrc(qinst
->op
) == 1)
603 fixup_raddr_conflict(c
, src
[0], &src
[1]);
605 if (translate
[qinst
->op
].is_mul
) {
606 queue(c
, qpu_inst(qpu_a_NOP(),
607 qpu_m_alu2(translate
[qinst
->op
].op
,
611 queue(c
, qpu_inst(qpu_a_alu2(translate
[qinst
->op
].op
,
622 /* thread end can't have VPM write */
623 if (QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
624 QPU_WADDR_ADD
) == QPU_W_VPM
||
625 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
626 QPU_WADDR_MUL
) == QPU_W_VPM
) {
627 serialize_one_inst(c
, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
630 c
->qpu_insts
[c
->qpu_inst_count
- 1] =
631 qpu_set_sig(c
->qpu_insts
[c
->qpu_inst_count
- 1],
633 serialize_one_inst(c
, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
634 serialize_one_inst(c
, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
641 c
->qpu_insts
[c
->qpu_inst_count
- 1] =
642 qpu_set_sig(c
->qpu_insts
[c
->qpu_inst_count
- 1],
643 QPU_SIG_SCOREBOARD_UNLOCK
);
647 if (vc4_debug
& VC4_DEBUG_QPU
)
650 vc4_qpu_validate(c
->qpu_insts
, c
->qpu_inst_count
);