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
)
106 uint32_t mux0
= src0
->mux
== QPU_MUX_SMALL_IMM
? QPU_MUX_B
: src0
->mux
;
107 uint32_t mux1
= src1
->mux
== QPU_MUX_SMALL_IMM
? QPU_MUX_B
: src1
->mux
;
109 if (mux0
<= QPU_MUX_R5
||
111 (src0
->addr
== src1
->addr
&&
112 src0
->mux
== src1
->mux
)) {
116 if (swap_file(src0
) || swap_file(src1
))
119 if (mux0
== QPU_MUX_A
) {
120 queue(c
, qpu_a_MOV(qpu_rb(31), *src0
));
123 queue(c
, qpu_a_MOV(qpu_ra(31), *src0
));
129 vc4_generate_code(struct vc4_context
*vc4
, struct vc4_compile
*c
)
131 struct qpu_reg
*temp_registers
= vc4_register_allocate(vc4
, c
);
132 bool discard
= false;
133 uint32_t inputs_remaining
= c
->num_inputs
;
134 uint32_t vpm_read_fifo_count
= 0;
135 uint32_t vpm_read_offset
= 0;
136 int last_vpm_read_index
= -1;
138 list_inithead(&c
->qpu_inst_list
);
143 /* There's a 4-entry FIFO for VPMVCD reads, each of which can
144 * load up to 16 dwords (4 vec4s) per vertex.
146 while (inputs_remaining
) {
147 uint32_t num_entries
= MIN2(inputs_remaining
, 16);
148 queue(c
, qpu_load_imm_ui(qpu_vrsetup(),
151 ((num_entries
& 0xf) << 20)));
152 inputs_remaining
-= num_entries
;
153 vpm_read_offset
+= num_entries
;
154 vpm_read_fifo_count
++;
156 assert(vpm_read_fifo_count
<= 4);
158 queue(c
, qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00));
164 list_for_each_entry(struct qinst
, qinst
, &c
->instructions
, link
) {
166 fprintf(stderr
, "translating qinst to qpu: ");
167 qir_dump_inst(qinst
);
168 fprintf(stderr
, "\n");
171 static const struct {
174 #define A(name) [QOP_##name] = {QPU_A_##name}
175 #define M(name) [QOP_##name] = {QPU_M_##name}
204 /* If we replicate src[0] out to src[1], this works
205 * out the same as a MOV.
207 [QOP_MOV
] = { QPU_A_OR
},
208 [QOP_FMOV
] = { QPU_A_FMAX
},
212 struct qpu_reg src
[4];
213 for (int i
= 0; i
< qir_get_op_nsrc(qinst
->op
); i
++) {
214 int index
= qinst
->src
[i
].index
;
215 switch (qinst
->src
[i
].file
) {
220 src
[i
] = temp_registers
[index
];
221 if (qinst
->src
[i
].pack
) {
223 unpack
== qinst
->src
[i
].pack
);
224 unpack
= QPU_SET_FIELD(qinst
->src
[i
].pack
,
226 if (src
[i
].mux
== QPU_MUX_R4
)
236 case QFILE_SMALL_IMM
:
237 src
[i
].mux
= QPU_MUX_SMALL_IMM
;
238 src
[i
].addr
= qpu_encode_small_immediate(qinst
->src
[i
].index
);
239 /* This should only have returned a valid
240 * small immediate field, not ~0 for failure.
242 assert(src
[i
].addr
<= 47);
245 assert((int)qinst
->src
[i
].index
>=
246 last_vpm_read_index
);
247 (void)last_vpm_read_index
;
248 last_vpm_read_index
= qinst
->src
[i
].index
;
249 src
[i
] = qpu_ra(QPU_R_VPM
);
255 switch (qinst
->dst
.file
) {
257 dst
= qpu_ra(QPU_W_NOP
);
260 dst
= temp_registers
[qinst
->dst
.index
];
263 dst
= qpu_ra(QPU_W_VPM
);
267 case QFILE_SMALL_IMM
:
268 assert(!"not reached");
277 queue(c
, qpu_a_MOV(dst
, src
[0]));
278 set_last_cond_add(c
, qinst
->op
- QOP_SEL_X_0_ZS
+
281 queue(c
, qpu_a_XOR(dst
, qpu_r0(), qpu_r0()));
282 set_last_cond_add(c
, ((qinst
->op
- QOP_SEL_X_0_ZS
) ^
290 queue(c
, qpu_a_MOV(dst
, src
[0]));
291 set_last_cond_add(c
, qinst
->op
- QOP_SEL_X_Y_ZS
+
294 queue(c
, qpu_a_MOV(dst
, src
[1]));
295 set_last_cond_add(c
, ((qinst
->op
- QOP_SEL_X_Y_ZS
) ^
306 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP
),
310 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT
),
314 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP
),
318 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG
),
325 if (dst
.mux
!= QPU_MUX_R4
)
326 queue(c
, qpu_a_MOV(dst
, qpu_r4()));
330 case QOP_PACK_8888_F
:
331 queue(c
, qpu_m_MOV(dst
, src
[0]));
332 *last_inst(c
) |= QPU_PM
;
333 *last_inst(c
) |= QPU_SET_FIELD(QPU_PACK_MUL_8888
,
342 qpu_m_MOV(dst
, src
[0]) |
344 QPU_SET_FIELD(QPU_PACK_MUL_8A
+
345 qinst
->op
- QOP_PACK_8A_F
,
350 queue(c
, qpu_a_ITOF(dst
,
351 qpu_ra(QPU_R_XY_PIXEL_COORD
)));
355 queue(c
, qpu_a_ITOF(dst
,
356 qpu_rb(QPU_R_XY_PIXEL_COORD
)));
359 case QOP_FRAG_REV_FLAG
:
360 queue(c
, qpu_a_ITOF(dst
,
361 qpu_rb(QPU_R_MS_REV_FLAGS
)));
366 /* QOP_FRAG_Z/W don't emit instructions, just allocate
367 * the register to the Z/W payload.
371 case QOP_TLB_DISCARD_SETUP
:
373 queue(c
, qpu_a_MOV(src
[0], src
[0]));
374 *last_inst(c
) |= QPU_SF
;
377 case QOP_TLB_STENCIL_SETUP
:
378 queue(c
, qpu_a_MOV(qpu_ra(QPU_W_TLB_STENCIL_SETUP
), src
[0]));
381 case QOP_TLB_Z_WRITE
:
382 queue(c
, qpu_a_MOV(qpu_ra(QPU_W_TLB_Z
), src
[0]));
384 set_last_cond_add(c
, QPU_COND_ZS
);
388 case QOP_TLB_COLOR_READ
:
390 *last_inst(c
) = qpu_set_sig(*last_inst(c
),
393 if (dst
.mux
!= QPU_MUX_R4
)
394 queue(c
, qpu_a_MOV(dst
, qpu_r4()));
397 case QOP_TLB_COLOR_WRITE
:
398 queue(c
, qpu_a_MOV(qpu_tlbc(), src
[0]));
400 set_last_cond_add(c
, QPU_COND_ZS
);
405 queue(c
, qpu_a_FADD(dst
, src
[0], qpu_r5()));
412 queue(c
, qpu_a_MOV(qpu_rb(QPU_W_TMU0_S
+
413 (qinst
->op
- QOP_TEX_S
)),
418 fixup_raddr_conflict(c
, dst
, &src
[0], &src
[1]);
419 queue(c
, qpu_a_ADD(qpu_rb(QPU_W_TMU0_S
), src
[0], src
[1]));
424 *last_inst(c
) = qpu_set_sig(*last_inst(c
),
426 if (dst
.mux
!= QPU_MUX_R4
)
427 queue(c
, qpu_a_MOV(dst
, qpu_r4()));
431 assert(qinst
->op
< ARRAY_SIZE(translate
));
432 assert(translate
[qinst
->op
].op
!= 0); /* NOPs */
434 /* Skip emitting the MOV if it's a no-op. */
435 if (qir_is_raw_mov(qinst
) &&
436 dst
.mux
== src
[0].mux
&& dst
.addr
== src
[0].addr
) {
440 /* If we have only one source, put it in the second
441 * argument slot as well so that we don't take up
442 * another raddr just to get unused data.
444 if (qir_get_op_nsrc(qinst
->op
) == 1)
447 fixup_raddr_conflict(c
, dst
, &src
[0], &src
[1]);
449 if (qir_is_mul(qinst
)) {
450 queue(c
, qpu_m_alu2(translate
[qinst
->op
].op
,
453 if (qinst
->dst
.pack
) {
454 *last_inst(c
) |= QPU_PM
;
455 *last_inst(c
) |= QPU_SET_FIELD(qinst
->dst
.pack
,
459 queue(c
, qpu_a_alu2(translate
[qinst
->op
].op
,
462 if (qinst
->dst
.pack
) {
463 assert(dst
.mux
== QPU_MUX_A
);
464 *last_inst(c
) |= QPU_SET_FIELD(qinst
->dst
.pack
,
473 assert(!qir_is_multi_instruction(qinst
));
474 *last_inst(c
) |= QPU_SF
;
478 qpu_schedule_instructions(c
);
480 /* thread end can't have VPM write or read */
481 if (QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
482 QPU_WADDR_ADD
) == QPU_W_VPM
||
483 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
484 QPU_WADDR_MUL
) == QPU_W_VPM
||
485 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
486 QPU_RADDR_A
) == QPU_R_VPM
||
487 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
488 QPU_RADDR_B
) == QPU_R_VPM
) {
489 qpu_serialize_one_inst(c
, qpu_NOP());
492 /* thread end can't have uniform read */
493 if (QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
494 QPU_RADDR_A
) == QPU_R_UNIF
||
495 QPU_GET_FIELD(c
->qpu_insts
[c
->qpu_inst_count
- 1],
496 QPU_RADDR_B
) == QPU_R_UNIF
) {
497 qpu_serialize_one_inst(c
, qpu_NOP());
500 /* thread end can't have TLB operations */
501 if (qpu_inst_is_tlb(c
->qpu_insts
[c
->qpu_inst_count
- 1]))
502 qpu_serialize_one_inst(c
, qpu_NOP());
504 c
->qpu_insts
[c
->qpu_inst_count
- 1] =
505 qpu_set_sig(c
->qpu_insts
[c
->qpu_inst_count
- 1],
507 qpu_serialize_one_inst(c
, qpu_NOP());
508 qpu_serialize_one_inst(c
, qpu_NOP());
515 c
->qpu_insts
[c
->qpu_inst_count
- 1] =
516 qpu_set_sig(c
->qpu_insts
[c
->qpu_inst_count
- 1],
517 QPU_SIG_SCOREBOARD_UNLOCK
);
521 if (vc4_debug
& VC4_DEBUG_QPU
)
524 vc4_qpu_validate(c
->qpu_insts
, c
->qpu_inst_count
);
526 free(temp_registers
);