2 * Copyright © 2016 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
25 #include "util/u_format.h"
26 #include "util/u_math.h"
27 #include "util/u_memory.h"
28 #include "util/ralloc.h"
29 #include "util/hash_table.h"
30 #include "compiler/nir/nir.h"
31 #include "compiler/nir/nir_builder.h"
32 #include "common/v3d_device_info.h"
33 #include "v3d_compiler.h"
35 #define GENERAL_TMU_LOOKUP_PER_QUAD (0 << 7)
36 #define GENERAL_TMU_LOOKUP_PER_PIXEL (1 << 7)
37 #define GENERAL_TMU_READ_OP_PREFETCH (0 << 3)
38 #define GENERAL_TMU_READ_OP_CACHE_CLEAR (1 << 3)
39 #define GENERAL_TMU_READ_OP_CACHE_FLUSH (3 << 3)
40 #define GENERAL_TMU_READ_OP_CACHE_CLEAN (3 << 3)
41 #define GENERAL_TMU_READ_OP_CACHE_L1T_CLEAR (4 << 3)
42 #define GENERAL_TMU_READ_OP_CACHE_L1T_FLUSH_AGGREGATION (5 << 3)
43 #define GENERAL_TMU_READ_OP_ATOMIC_INC (8 << 3)
44 #define GENERAL_TMU_READ_OP_ATOMIC_DEC (9 << 3)
45 #define GENERAL_TMU_READ_OP_ATOMIC_NOT (10 << 3)
46 #define GENERAL_TMU_READ_OP_READ (15 << 3)
47 #define GENERAL_TMU_LOOKUP_TYPE_8BIT_I (0 << 0)
48 #define GENERAL_TMU_LOOKUP_TYPE_16BIT_I (1 << 0)
49 #define GENERAL_TMU_LOOKUP_TYPE_VEC2 (2 << 0)
50 #define GENERAL_TMU_LOOKUP_TYPE_VEC3 (3 << 0)
51 #define GENERAL_TMU_LOOKUP_TYPE_VEC4 (4 << 0)
52 #define GENERAL_TMU_LOOKUP_TYPE_8BIT_UI (5 << 0)
53 #define GENERAL_TMU_LOOKUP_TYPE_16BIT_UI (6 << 0)
54 #define GENERAL_TMU_LOOKUP_TYPE_32BIT_UI (7 << 0)
56 #define GENERAL_TMU_WRITE_OP_ATOMIC_ADD_WRAP (0 << 3)
57 #define GENERAL_TMU_WRITE_OP_ATOMIC_SUB_WRAP (1 << 3)
58 #define GENERAL_TMU_WRITE_OP_ATOMIC_XCHG (2 << 3)
59 #define GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG (3 << 3)
60 #define GENERAL_TMU_WRITE_OP_ATOMIC_UMIN (4 << 3)
61 #define GENERAL_TMU_WRITE_OP_ATOMIC_UMAX (5 << 3)
62 #define GENERAL_TMU_WRITE_OP_ATOMIC_SMIN (6 << 3)
63 #define GENERAL_TMU_WRITE_OP_ATOMIC_SMAX (7 << 3)
64 #define GENERAL_TMU_WRITE_OP_ATOMIC_AND (8 << 3)
65 #define GENERAL_TMU_WRITE_OP_ATOMIC_OR (9 << 3)
66 #define GENERAL_TMU_WRITE_OP_ATOMIC_XOR (10 << 3)
67 #define GENERAL_TMU_WRITE_OP_WRITE (15 << 3)
70 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
73 resize_qreg_array(struct v3d_compile
*c
,
78 if (*size
>= decl_size
)
81 uint32_t old_size
= *size
;
82 *size
= MAX2(*size
* 2, decl_size
);
83 *regs
= reralloc(c
, *regs
, struct qreg
, *size
);
85 fprintf(stderr
, "Malloc failure\n");
89 for (uint32_t i
= old_size
; i
< *size
; i
++)
90 (*regs
)[i
] = c
->undef
;
94 vir_emit_thrsw(struct v3d_compile
*c
)
99 /* Always thread switch after each texture operation for now.
101 * We could do better by batching a bunch of texture fetches up and
102 * then doing one thread switch and collecting all their results
105 c
->last_thrsw
= vir_NOP(c
);
106 c
->last_thrsw
->qpu
.sig
.thrsw
= true;
107 c
->last_thrsw_at_top_level
= (c
->execute
.file
== QFILE_NULL
);
111 v3d_general_tmu_op(nir_intrinsic_instr
*instr
)
113 switch (instr
->intrinsic
) {
114 case nir_intrinsic_load_ssbo
:
115 case nir_intrinsic_load_ubo
:
116 case nir_intrinsic_load_uniform
:
117 return GENERAL_TMU_READ_OP_READ
;
118 case nir_intrinsic_store_ssbo
:
119 return GENERAL_TMU_WRITE_OP_WRITE
;
120 case nir_intrinsic_ssbo_atomic_add
:
121 return GENERAL_TMU_WRITE_OP_ATOMIC_ADD_WRAP
;
122 case nir_intrinsic_ssbo_atomic_imin
:
123 return GENERAL_TMU_WRITE_OP_ATOMIC_SMIN
;
124 case nir_intrinsic_ssbo_atomic_umin
:
125 return GENERAL_TMU_WRITE_OP_ATOMIC_UMIN
;
126 case nir_intrinsic_ssbo_atomic_imax
:
127 return GENERAL_TMU_WRITE_OP_ATOMIC_SMAX
;
128 case nir_intrinsic_ssbo_atomic_umax
:
129 return GENERAL_TMU_WRITE_OP_ATOMIC_UMAX
;
130 case nir_intrinsic_ssbo_atomic_and
:
131 return GENERAL_TMU_WRITE_OP_ATOMIC_AND
;
132 case nir_intrinsic_ssbo_atomic_or
:
133 return GENERAL_TMU_WRITE_OP_ATOMIC_OR
;
134 case nir_intrinsic_ssbo_atomic_xor
:
135 return GENERAL_TMU_WRITE_OP_ATOMIC_XOR
;
136 case nir_intrinsic_ssbo_atomic_exchange
:
137 return GENERAL_TMU_WRITE_OP_ATOMIC_XCHG
;
138 case nir_intrinsic_ssbo_atomic_comp_swap
:
139 return GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG
;
141 unreachable("unknown intrinsic op");
146 * Implements indirect uniform loads and SSBO accesses through the TMU general
147 * memory access interface.
150 ntq_emit_tmu_general(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
152 /* XXX perf: We should turn add/sub of 1 to inc/dec. Perhaps NIR
153 * wants to have support for inc/dec?
156 uint32_t tmu_op
= v3d_general_tmu_op(instr
);
157 bool is_store
= instr
->intrinsic
== nir_intrinsic_store_ssbo
;
160 int tmu_writes
= 1; /* address */
161 if (instr
->intrinsic
== nir_intrinsic_load_uniform
) {
163 } else if (instr
->intrinsic
== nir_intrinsic_load_ssbo
||
164 instr
->intrinsic
== nir_intrinsic_load_ubo
) {
166 } else if (is_store
) {
168 for (int i
= 0; i
< instr
->num_components
; i
++) {
170 vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUD
),
171 ntq_get_src(c
, instr
->src
[0], i
));
177 vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUD
),
178 ntq_get_src(c
, instr
->src
[2], 0));
180 if (tmu_op
== GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG
) {
182 vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUD
),
183 ntq_get_src(c
, instr
->src
[3], 0));
188 /* Make sure we won't exceed the 16-entry TMU fifo if each thread is
189 * storing at the same time.
191 while (tmu_writes
> 16 / c
->threads
)
195 if (instr
->intrinsic
== nir_intrinsic_load_uniform
) {
196 offset
= vir_uniform(c
, QUNIFORM_UBO_ADDR
, 0);
198 /* Find what variable in the default uniform block this
199 * uniform load is coming from.
201 uint32_t base
= nir_intrinsic_base(instr
);
203 struct v3d_ubo_range
*range
= NULL
;
204 for (i
= 0; i
< c
->num_ubo_ranges
; i
++) {
205 range
= &c
->ubo_ranges
[i
];
206 if (base
>= range
->src_offset
&&
207 base
< range
->src_offset
+ range
->size
) {
211 /* The driver-location-based offset always has to be within a
212 * declared uniform range.
214 assert(i
!= c
->num_ubo_ranges
);
215 if (!c
->ubo_range_used
[i
]) {
216 c
->ubo_range_used
[i
] = true;
217 range
->dst_offset
= c
->next_ubo_dst_offset
;
218 c
->next_ubo_dst_offset
+= range
->size
;
221 base
= base
- range
->src_offset
+ range
->dst_offset
;
224 offset
= vir_ADD(c
, offset
, vir_uniform_ui(c
, base
));
225 } else if (instr
->intrinsic
== nir_intrinsic_load_ubo
) {
226 /* Note that QUNIFORM_UBO_ADDR takes a UBO index shifted up by
227 * 1 (0 is gallium's constant buffer 0).
229 offset
= vir_uniform(c
, QUNIFORM_UBO_ADDR
,
230 nir_src_as_uint(instr
->src
[0]) + 1);
232 offset
= vir_uniform(c
, QUNIFORM_SSBO_OFFSET
,
233 nir_src_as_uint(instr
->src
[is_store
?
237 uint32_t config
= (0xffffff00 |
239 GENERAL_TMU_LOOKUP_PER_PIXEL
);
240 if (instr
->num_components
== 1) {
241 config
|= GENERAL_TMU_LOOKUP_TYPE_32BIT_UI
;
243 config
|= (GENERAL_TMU_LOOKUP_TYPE_VEC2
+
244 instr
->num_components
- 2);
247 if (c
->execute
.file
!= QFILE_NULL
)
248 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
252 dest
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUA
);
254 dest
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUAU
);
257 if (nir_src_is_const(instr
->src
[offset_src
]) &&
258 nir_src_as_uint(instr
->src
[offset_src
]) == 0) {
259 tmu
= vir_MOV_dest(c
, dest
, offset
);
261 tmu
= vir_ADD_dest(c
, dest
,
263 ntq_get_src(c
, instr
->src
[offset_src
], 0));
267 tmu
->src
[vir_get_implicit_uniform_src(tmu
)] =
268 vir_uniform_ui(c
, config
);
271 if (c
->execute
.file
!= QFILE_NULL
)
272 vir_set_cond(tmu
, V3D_QPU_COND_IFA
);
276 /* Read the result, or wait for the TMU op to complete. */
277 for (int i
= 0; i
< nir_intrinsic_dest_components(instr
); i
++)
278 ntq_store_dest(c
, &instr
->dest
, i
, vir_MOV(c
, vir_LDTMU(c
)));
280 if (nir_intrinsic_dest_components(instr
) == 0)
285 ntq_init_ssa_def(struct v3d_compile
*c
, nir_ssa_def
*def
)
287 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
288 def
->num_components
);
289 _mesa_hash_table_insert(c
->def_ht
, def
, qregs
);
294 * This function is responsible for getting VIR results into the associated
295 * storage for a NIR instruction.
297 * If it's a NIR SSA def, then we just set the associated hash table entry to
300 * If it's a NIR reg, then we need to update the existing qreg assigned to the
301 * NIR destination with the incoming value. To do that without introducing
302 * new MOVs, we require that the incoming qreg either be a uniform, or be
303 * SSA-defined by the previous VIR instruction in the block and rewritable by
304 * this function. That lets us sneak ahead and insert the SF flag beforehand
305 * (knowing that the previous instruction doesn't depend on flags) and rewrite
306 * its destination to be the NIR reg's destination
309 ntq_store_dest(struct v3d_compile
*c
, nir_dest
*dest
, int chan
,
312 struct qinst
*last_inst
= NULL
;
313 if (!list_empty(&c
->cur_block
->instructions
))
314 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
316 assert(result
.file
== QFILE_UNIF
||
317 (result
.file
== QFILE_TEMP
&&
318 last_inst
&& last_inst
== c
->defs
[result
.index
]));
321 assert(chan
< dest
->ssa
.num_components
);
324 struct hash_entry
*entry
=
325 _mesa_hash_table_search(c
->def_ht
, &dest
->ssa
);
330 qregs
= ntq_init_ssa_def(c
, &dest
->ssa
);
332 qregs
[chan
] = result
;
334 nir_register
*reg
= dest
->reg
.reg
;
335 assert(dest
->reg
.base_offset
== 0);
336 assert(reg
->num_array_elems
== 0);
337 struct hash_entry
*entry
=
338 _mesa_hash_table_search(c
->def_ht
, reg
);
339 struct qreg
*qregs
= entry
->data
;
341 /* Insert a MOV if the source wasn't an SSA def in the
342 * previous instruction.
344 if (result
.file
== QFILE_UNIF
) {
345 result
= vir_MOV(c
, result
);
346 last_inst
= c
->defs
[result
.index
];
349 /* We know they're both temps, so just rewrite index. */
350 c
->defs
[last_inst
->dst
.index
] = NULL
;
351 last_inst
->dst
.index
= qregs
[chan
].index
;
353 /* If we're in control flow, then make this update of the reg
354 * conditional on the execution mask.
356 if (c
->execute
.file
!= QFILE_NULL
) {
357 last_inst
->dst
.index
= qregs
[chan
].index
;
359 /* Set the flags to the current exec mask.
361 c
->cursor
= vir_before_inst(last_inst
);
362 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
363 c
->cursor
= vir_after_inst(last_inst
);
365 vir_set_cond(last_inst
, V3D_QPU_COND_IFA
);
366 last_inst
->cond_is_exec_mask
= true;
372 ntq_get_src(struct v3d_compile
*c
, nir_src src
, int i
)
374 struct hash_entry
*entry
;
376 entry
= _mesa_hash_table_search(c
->def_ht
, src
.ssa
);
377 assert(i
< src
.ssa
->num_components
);
379 nir_register
*reg
= src
.reg
.reg
;
380 entry
= _mesa_hash_table_search(c
->def_ht
, reg
);
381 assert(reg
->num_array_elems
== 0);
382 assert(src
.reg
.base_offset
== 0);
383 assert(i
< reg
->num_components
);
386 struct qreg
*qregs
= entry
->data
;
391 ntq_get_alu_src(struct v3d_compile
*c
, nir_alu_instr
*instr
,
394 assert(util_is_power_of_two_or_zero(instr
->dest
.write_mask
));
395 unsigned chan
= ffs(instr
->dest
.write_mask
) - 1;
396 struct qreg r
= ntq_get_src(c
, instr
->src
[src
].src
,
397 instr
->src
[src
].swizzle
[chan
]);
399 assert(!instr
->src
[src
].abs
);
400 assert(!instr
->src
[src
].negate
);
406 ntq_minify(struct v3d_compile
*c
, struct qreg size
, struct qreg level
)
408 return vir_MAX(c
, vir_SHR(c
, size
, level
), vir_uniform_ui(c
, 1));
412 ntq_emit_txs(struct v3d_compile
*c
, nir_tex_instr
*instr
)
414 unsigned unit
= instr
->texture_index
;
415 int lod_index
= nir_tex_instr_src_index(instr
, nir_tex_src_lod
);
416 int dest_size
= nir_tex_instr_dest_size(instr
);
418 struct qreg lod
= c
->undef
;
420 lod
= ntq_get_src(c
, instr
->src
[lod_index
].src
, 0);
422 for (int i
= 0; i
< dest_size
; i
++) {
424 enum quniform_contents contents
;
426 if (instr
->is_array
&& i
== dest_size
- 1)
427 contents
= QUNIFORM_TEXTURE_ARRAY_SIZE
;
429 contents
= QUNIFORM_TEXTURE_WIDTH
+ i
;
431 struct qreg size
= vir_uniform(c
, contents
, unit
);
433 switch (instr
->sampler_dim
) {
434 case GLSL_SAMPLER_DIM_1D
:
435 case GLSL_SAMPLER_DIM_2D
:
436 case GLSL_SAMPLER_DIM_MS
:
437 case GLSL_SAMPLER_DIM_3D
:
438 case GLSL_SAMPLER_DIM_CUBE
:
439 /* Don't minify the array size. */
440 if (!(instr
->is_array
&& i
== dest_size
- 1)) {
441 size
= ntq_minify(c
, size
, lod
);
445 case GLSL_SAMPLER_DIM_RECT
:
446 /* There's no LOD field for rects */
450 unreachable("Bad sampler type");
453 ntq_store_dest(c
, &instr
->dest
, i
, size
);
458 ntq_emit_tex(struct v3d_compile
*c
, nir_tex_instr
*instr
)
460 unsigned unit
= instr
->texture_index
;
462 /* Since each texture sampling op requires uploading uniforms to
463 * reference the texture, there's no HW support for texture size and
464 * you just upload uniforms containing the size.
467 case nir_texop_query_levels
:
468 ntq_store_dest(c
, &instr
->dest
, 0,
469 vir_uniform(c
, QUNIFORM_TEXTURE_LEVELS
, unit
));
472 ntq_emit_txs(c
, instr
);
478 if (c
->devinfo
->ver
>= 40)
479 v3d40_vir_emit_tex(c
, instr
);
481 v3d33_vir_emit_tex(c
, instr
);
485 ntq_fsincos(struct v3d_compile
*c
, struct qreg src
, bool is_cos
)
487 struct qreg input
= vir_FMUL(c
, src
, vir_uniform_f(c
, 1.0f
/ M_PI
));
489 input
= vir_FADD(c
, input
, vir_uniform_f(c
, 0.5));
491 struct qreg periods
= vir_FROUND(c
, input
);
492 struct qreg sin_output
= vir_SIN(c
, vir_FSUB(c
, input
, periods
));
493 return vir_XOR(c
, sin_output
, vir_SHL(c
,
494 vir_FTOIN(c
, periods
),
495 vir_uniform_ui(c
, -1)));
499 ntq_fsign(struct v3d_compile
*c
, struct qreg src
)
501 struct qreg t
= vir_get_temp(c
);
503 vir_MOV_dest(c
, t
, vir_uniform_f(c
, 0.0));
504 vir_PF(c
, vir_FMOV(c
, src
), V3D_QPU_PF_PUSHZ
);
505 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_f(c
, 1.0));
506 vir_PF(c
, vir_FMOV(c
, src
), V3D_QPU_PF_PUSHN
);
507 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_f(c
, -1.0));
508 return vir_MOV(c
, t
);
512 ntq_isign(struct v3d_compile
*c
, struct qreg src
)
514 struct qreg t
= vir_get_temp(c
);
516 vir_MOV_dest(c
, t
, vir_uniform_ui(c
, 0));
517 vir_PF(c
, vir_MOV(c
, src
), V3D_QPU_PF_PUSHZ
);
518 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_ui(c
, 1));
519 vir_PF(c
, vir_MOV(c
, src
), V3D_QPU_PF_PUSHN
);
520 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_ui(c
, -1));
521 return vir_MOV(c
, t
);
525 emit_fragcoord_input(struct v3d_compile
*c
, int attr
)
527 c
->inputs
[attr
* 4 + 0] = vir_FXCD(c
);
528 c
->inputs
[attr
* 4 + 1] = vir_FYCD(c
);
529 c
->inputs
[attr
* 4 + 2] = c
->payload_z
;
530 c
->inputs
[attr
* 4 + 3] = vir_RECIP(c
, c
->payload_w
);
534 emit_fragment_varying(struct v3d_compile
*c
, nir_variable
*var
,
535 uint8_t swizzle
, int array_index
)
537 struct qreg r3
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R3
);
538 struct qreg r5
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R5
);
541 if (c
->devinfo
->ver
>= 41) {
542 struct qinst
*ldvary
= vir_add_inst(V3D_QPU_A_NOP
, c
->undef
,
544 ldvary
->qpu
.sig
.ldvary
= true;
545 vary
= vir_emit_def(c
, ldvary
);
547 vir_NOP(c
)->qpu
.sig
.ldvary
= true;
551 /* For gl_PointCoord input or distance along a line, we'll be called
552 * with no nir_variable, and we don't count toward VPM size so we
553 * don't track an input slot.
556 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
559 int i
= c
->num_inputs
++;
561 v3d_slot_from_slot_and_component(var
->data
.location
+
562 array_index
, swizzle
);
564 switch (var
->data
.interpolation
) {
565 case INTERP_MODE_NONE
:
566 /* If a gl_FrontColor or gl_BackColor input has no interp
567 * qualifier, then if we're using glShadeModel(GL_FLAT) it
568 * needs to be flat shaded.
570 switch (var
->data
.location
+ array_index
) {
571 case VARYING_SLOT_COL0
:
572 case VARYING_SLOT_COL1
:
573 case VARYING_SLOT_BFC0
:
574 case VARYING_SLOT_BFC1
:
575 if (c
->fs_key
->shade_model_flat
) {
576 BITSET_SET(c
->flat_shade_flags
, i
);
577 vir_MOV_dest(c
, c
->undef
, vary
);
578 return vir_MOV(c
, r5
);
580 return vir_FADD(c
, vir_FMUL(c
, vary
,
587 case INTERP_MODE_SMOOTH
:
588 if (var
->data
.centroid
) {
589 BITSET_SET(c
->centroid_flags
, i
);
590 return vir_FADD(c
, vir_FMUL(c
, vary
,
591 c
->payload_w_centroid
), r5
);
593 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
595 case INTERP_MODE_NOPERSPECTIVE
:
596 BITSET_SET(c
->noperspective_flags
, i
);
597 return vir_FADD(c
, vir_MOV(c
, vary
), r5
);
598 case INTERP_MODE_FLAT
:
599 BITSET_SET(c
->flat_shade_flags
, i
);
600 vir_MOV_dest(c
, c
->undef
, vary
);
601 return vir_MOV(c
, r5
);
603 unreachable("Bad interp mode");
608 emit_fragment_input(struct v3d_compile
*c
, int attr
, nir_variable
*var
,
611 for (int i
= 0; i
< glsl_get_vector_elements(var
->type
); i
++) {
612 int chan
= var
->data
.location_frac
+ i
;
613 c
->inputs
[attr
* 4 + chan
] =
614 emit_fragment_varying(c
, var
, chan
, array_index
);
619 add_output(struct v3d_compile
*c
,
620 uint32_t decl_offset
,
624 uint32_t old_array_size
= c
->outputs_array_size
;
625 resize_qreg_array(c
, &c
->outputs
, &c
->outputs_array_size
,
628 if (old_array_size
!= c
->outputs_array_size
) {
629 c
->output_slots
= reralloc(c
,
631 struct v3d_varying_slot
,
632 c
->outputs_array_size
);
635 c
->output_slots
[decl_offset
] =
636 v3d_slot_from_slot_and_component(slot
, swizzle
);
640 declare_uniform_range(struct v3d_compile
*c
, uint32_t start
, uint32_t size
)
642 unsigned array_id
= c
->num_ubo_ranges
++;
643 if (array_id
>= c
->ubo_ranges_array_size
) {
644 c
->ubo_ranges_array_size
= MAX2(c
->ubo_ranges_array_size
* 2,
646 c
->ubo_ranges
= reralloc(c
, c
->ubo_ranges
,
647 struct v3d_ubo_range
,
648 c
->ubo_ranges_array_size
);
649 c
->ubo_range_used
= reralloc(c
, c
->ubo_range_used
,
651 c
->ubo_ranges_array_size
);
654 c
->ubo_ranges
[array_id
].dst_offset
= 0;
655 c
->ubo_ranges
[array_id
].src_offset
= start
;
656 c
->ubo_ranges
[array_id
].size
= size
;
657 c
->ubo_range_used
[array_id
] = false;
661 * If compare_instr is a valid comparison instruction, emits the
662 * compare_instr's comparison and returns the sel_instr's return value based
663 * on the compare_instr's result.
666 ntq_emit_comparison(struct v3d_compile
*c
,
667 nir_alu_instr
*compare_instr
,
668 enum v3d_qpu_cond
*out_cond
)
670 struct qreg src0
= ntq_get_alu_src(c
, compare_instr
, 0);
672 if (nir_op_infos
[compare_instr
->op
].num_inputs
> 1)
673 src1
= ntq_get_alu_src(c
, compare_instr
, 1);
674 bool cond_invert
= false;
675 struct qreg nop
= vir_reg(QFILE_NULL
, 0);
677 switch (compare_instr
->op
) {
680 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
683 vir_set_pf(vir_XOR_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
688 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
692 vir_set_pf(vir_XOR_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
698 vir_set_pf(vir_FCMP_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
701 vir_set_pf(vir_MIN_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
705 vir_set_pf(vir_SUB_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHC
);
711 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHN
);
714 vir_set_pf(vir_MIN_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
717 vir_set_pf(vir_SUB_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHC
);
724 *out_cond
= cond_invert
? V3D_QPU_COND_IFNA
: V3D_QPU_COND_IFA
;
729 /* Finds an ALU instruction that generates our src value that could
730 * (potentially) be greedily emitted in the consuming instruction.
732 static struct nir_alu_instr
*
733 ntq_get_alu_parent(nir_src src
)
735 if (!src
.is_ssa
|| src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
737 nir_alu_instr
*instr
= nir_instr_as_alu(src
.ssa
->parent_instr
);
741 /* If the ALU instr's srcs are non-SSA, then we would have to avoid
742 * moving emission of the ALU instr down past another write of the
745 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
746 if (!instr
->src
[i
].src
.is_ssa
)
754 * Attempts to fold a comparison generating a boolean result into the
755 * condition code for selecting between two values, instead of comparing the
756 * boolean result against 0 to generate the condition code.
758 static struct qreg
ntq_emit_bcsel(struct v3d_compile
*c
, nir_alu_instr
*instr
,
761 nir_alu_instr
*compare
= ntq_get_alu_parent(instr
->src
[0].src
);
765 enum v3d_qpu_cond cond
;
766 if (ntq_emit_comparison(c
, compare
, &cond
))
767 return vir_MOV(c
, vir_SEL(c
, cond
, src
[1], src
[2]));
770 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
771 return vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
, src
[1], src
[2]));
776 ntq_emit_alu(struct v3d_compile
*c
, nir_alu_instr
*instr
)
778 /* This should always be lowered to ALU operations for V3D. */
779 assert(!instr
->dest
.saturate
);
781 /* Vectors are special in that they have non-scalarized writemasks,
782 * and just take the first swizzle channel for each argument in order
783 * into each writemask channel.
785 if (instr
->op
== nir_op_vec2
||
786 instr
->op
== nir_op_vec3
||
787 instr
->op
== nir_op_vec4
) {
789 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
790 srcs
[i
] = ntq_get_src(c
, instr
->src
[i
].src
,
791 instr
->src
[i
].swizzle
[0]);
792 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
793 ntq_store_dest(c
, &instr
->dest
.dest
, i
,
794 vir_MOV(c
, srcs
[i
]));
798 /* General case: We can just grab the one used channel per src. */
799 struct qreg src
[nir_op_infos
[instr
->op
].num_inputs
];
800 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
801 src
[i
] = ntq_get_alu_src(c
, instr
, i
);
809 result
= vir_MOV(c
, src
[0]);
813 result
= vir_XOR(c
, src
[0], vir_uniform_ui(c
, 1 << 31));
816 result
= vir_NEG(c
, src
[0]);
820 result
= vir_FMUL(c
, src
[0], src
[1]);
823 result
= vir_FADD(c
, src
[0], src
[1]);
826 result
= vir_FSUB(c
, src
[0], src
[1]);
829 result
= vir_FMIN(c
, src
[0], src
[1]);
832 result
= vir_FMAX(c
, src
[0], src
[1]);
836 result
= vir_FTOIZ(c
, src
[0]);
839 result
= vir_FTOUZ(c
, src
[0]);
842 result
= vir_ITOF(c
, src
[0]);
845 result
= vir_UTOF(c
, src
[0]);
848 result
= vir_AND(c
, src
[0], vir_uniform_f(c
, 1.0));
851 result
= vir_AND(c
, src
[0], vir_uniform_ui(c
, 1));
855 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
856 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
857 vir_uniform_ui(c
, ~0),
858 vir_uniform_ui(c
, 0)));
862 result
= vir_ADD(c
, src
[0], src
[1]);
865 result
= vir_SHR(c
, src
[0], src
[1]);
868 result
= vir_SUB(c
, src
[0], src
[1]);
871 result
= vir_ASR(c
, src
[0], src
[1]);
874 result
= vir_SHL(c
, src
[0], src
[1]);
877 result
= vir_MIN(c
, src
[0], src
[1]);
880 result
= vir_UMIN(c
, src
[0], src
[1]);
883 result
= vir_MAX(c
, src
[0], src
[1]);
886 result
= vir_UMAX(c
, src
[0], src
[1]);
889 result
= vir_AND(c
, src
[0], src
[1]);
892 result
= vir_OR(c
, src
[0], src
[1]);
895 result
= vir_XOR(c
, src
[0], src
[1]);
898 result
= vir_NOT(c
, src
[0]);
901 case nir_op_ufind_msb
:
902 result
= vir_SUB(c
, vir_uniform_ui(c
, 31), vir_CLZ(c
, src
[0]));
906 result
= vir_UMUL(c
, src
[0], src
[1]);
913 enum v3d_qpu_cond cond
;
914 MAYBE_UNUSED
bool ok
= ntq_emit_comparison(c
, instr
, &cond
);
916 result
= vir_MOV(c
, vir_SEL(c
, cond
,
917 vir_uniform_f(c
, 1.0),
918 vir_uniform_f(c
, 0.0)));
932 enum v3d_qpu_cond cond
;
933 MAYBE_UNUSED
bool ok
= ntq_emit_comparison(c
, instr
, &cond
);
935 result
= vir_MOV(c
, vir_SEL(c
, cond
,
936 vir_uniform_ui(c
, ~0),
937 vir_uniform_ui(c
, 0)));
942 result
= ntq_emit_bcsel(c
, instr
, src
);
945 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
946 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
951 result
= vir_RECIP(c
, src
[0]);
954 result
= vir_RSQRT(c
, src
[0]);
957 result
= vir_EXP(c
, src
[0]);
960 result
= vir_LOG(c
, src
[0]);
964 result
= vir_FCEIL(c
, src
[0]);
967 result
= vir_FFLOOR(c
, src
[0]);
969 case nir_op_fround_even
:
970 result
= vir_FROUND(c
, src
[0]);
973 result
= vir_FTRUNC(c
, src
[0]);
976 result
= vir_FSUB(c
, src
[0], vir_FFLOOR(c
, src
[0]));
980 result
= ntq_fsincos(c
, src
[0], false);
983 result
= ntq_fsincos(c
, src
[0], true);
987 result
= ntq_fsign(c
, src
[0]);
990 result
= ntq_isign(c
, src
[0]);
994 result
= vir_FMOV(c
, src
[0]);
995 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_ABS
);
1000 result
= vir_MAX(c
, src
[0],
1001 vir_SUB(c
, vir_uniform_ui(c
, 0), src
[0]));
1005 case nir_op_fddx_coarse
:
1006 case nir_op_fddx_fine
:
1007 result
= vir_FDX(c
, src
[0]);
1011 case nir_op_fddy_coarse
:
1012 case nir_op_fddy_fine
:
1013 result
= vir_FDY(c
, src
[0]);
1016 case nir_op_uadd_carry
:
1017 vir_PF(c
, vir_ADD(c
, src
[0], src
[1]), V3D_QPU_PF_PUSHC
);
1018 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFA
,
1019 vir_uniform_ui(c
, ~0),
1020 vir_uniform_ui(c
, 0)));
1023 case nir_op_pack_half_2x16_split
:
1024 result
= vir_VFPACK(c
, src
[0], src
[1]);
1027 case nir_op_unpack_half_2x16_split_x
:
1028 /* XXX perf: It would be good to be able to merge this unpack
1029 * with whatever uses our result.
1031 result
= vir_FMOV(c
, src
[0]);
1032 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_L
);
1035 case nir_op_unpack_half_2x16_split_y
:
1036 result
= vir_FMOV(c
, src
[0]);
1037 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_H
);
1041 fprintf(stderr
, "unknown NIR ALU inst: ");
1042 nir_print_instr(&instr
->instr
, stderr
);
1043 fprintf(stderr
, "\n");
1047 /* We have a scalar result, so the instruction should only have a
1048 * single channel written to.
1050 assert(util_is_power_of_two_or_zero(instr
->dest
.write_mask
));
1051 ntq_store_dest(c
, &instr
->dest
.dest
,
1052 ffs(instr
->dest
.write_mask
) - 1, result
);
1055 /* Each TLB read/write setup (a render target or depth buffer) takes an 8-bit
1056 * specifier. They come from a register that's preloaded with 0xffffffff
1057 * (0xff gets you normal vec4 f16 RT0 writes), and when one is neaded the low
1058 * 8 bits are shifted off the bottom and 0xff shifted in from the top.
1060 #define TLB_TYPE_F16_COLOR (3 << 6)
1061 #define TLB_TYPE_I32_COLOR (1 << 6)
1062 #define TLB_TYPE_F32_COLOR (0 << 6)
1063 #define TLB_RENDER_TARGET_SHIFT 3 /* Reversed! 7 = RT 0, 0 = RT 7. */
1064 #define TLB_SAMPLE_MODE_PER_SAMPLE (0 << 2)
1065 #define TLB_SAMPLE_MODE_PER_PIXEL (1 << 2)
1066 #define TLB_F16_SWAP_HI_LO (1 << 1)
1067 #define TLB_VEC_SIZE_4_F16 (1 << 0)
1068 #define TLB_VEC_SIZE_2_F16 (0 << 0)
1069 #define TLB_VEC_SIZE_MINUS_1_SHIFT 0
1071 /* Triggers Z/Stencil testing, used when the shader state's "FS modifies Z"
1074 #define TLB_TYPE_DEPTH ((2 << 6) | (0 << 4))
1075 #define TLB_DEPTH_TYPE_INVARIANT (0 << 2) /* Unmodified sideband input used */
1076 #define TLB_DEPTH_TYPE_PER_PIXEL (1 << 2) /* QPU result used */
1077 #define TLB_V42_DEPTH_TYPE_INVARIANT (0 << 3) /* Unmodified sideband input used */
1078 #define TLB_V42_DEPTH_TYPE_PER_PIXEL (1 << 3) /* QPU result used */
1080 /* Stencil is a single 32-bit write. */
1081 #define TLB_TYPE_STENCIL_ALPHA ((2 << 6) | (1 << 4))
1084 emit_frag_end(struct v3d_compile
*c
)
1087 if (c->output_sample_mask_index != -1) {
1088 vir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
1092 bool has_any_tlb_color_write
= false;
1093 for (int rt
= 0; rt
< c
->fs_key
->nr_cbufs
; rt
++) {
1094 if (c
->output_color_var
[rt
])
1095 has_any_tlb_color_write
= true;
1098 if (c
->fs_key
->sample_alpha_to_coverage
&& c
->output_color_var
[0]) {
1099 struct nir_variable
*var
= c
->output_color_var
[0];
1100 struct qreg
*color
= &c
->outputs
[var
->data
.driver_location
* 4];
1102 vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1105 vir_FTOC(c
, color
[3])));
1108 if (c
->output_position_index
!= -1) {
1109 struct qinst
*inst
= vir_MOV_dest(c
,
1110 vir_reg(QFILE_TLBU
, 0),
1111 c
->outputs
[c
->output_position_index
]);
1112 uint8_t tlb_specifier
= TLB_TYPE_DEPTH
;
1114 if (c
->devinfo
->ver
>= 42) {
1115 tlb_specifier
|= (TLB_V42_DEPTH_TYPE_PER_PIXEL
|
1116 TLB_SAMPLE_MODE_PER_PIXEL
);
1118 tlb_specifier
|= TLB_DEPTH_TYPE_PER_PIXEL
;
1120 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1121 vir_uniform_ui(c
, tlb_specifier
| 0xffffff00);
1122 } else if (c
->s
->info
.fs
.uses_discard
||
1123 c
->fs_key
->sample_alpha_to_coverage
||
1124 !has_any_tlb_color_write
) {
1125 /* Emit passthrough Z if it needed to be delayed until shader
1126 * end due to potential discards.
1128 * Since (single-threaded) fragment shaders always need a TLB
1129 * write, emit passthrouh Z if we didn't have any color
1130 * buffers and flag us as potentially discarding, so that we
1131 * can use Z as the TLB write.
1133 c
->s
->info
.fs
.uses_discard
= true;
1135 struct qinst
*inst
= vir_MOV_dest(c
,
1136 vir_reg(QFILE_TLBU
, 0),
1137 vir_reg(QFILE_NULL
, 0));
1138 uint8_t tlb_specifier
= TLB_TYPE_DEPTH
;
1140 if (c
->devinfo
->ver
>= 42) {
1141 /* The spec says the PER_PIXEL flag is ignored for
1142 * invariant writes, but the simulator demands it.
1144 tlb_specifier
|= (TLB_V42_DEPTH_TYPE_INVARIANT
|
1145 TLB_SAMPLE_MODE_PER_PIXEL
);
1147 tlb_specifier
|= TLB_DEPTH_TYPE_INVARIANT
;
1150 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1151 vir_uniform_ui(c
, tlb_specifier
| 0xffffff00);
1154 /* XXX: Performance improvement: Merge Z write and color writes TLB
1158 for (int rt
= 0; rt
< c
->fs_key
->nr_cbufs
; rt
++) {
1159 if (!c
->output_color_var
[rt
])
1162 nir_variable
*var
= c
->output_color_var
[rt
];
1163 struct qreg
*color
= &c
->outputs
[var
->data
.driver_location
* 4];
1164 int num_components
= glsl_get_vector_elements(var
->type
);
1165 uint32_t conf
= 0xffffff00;
1168 conf
|= TLB_SAMPLE_MODE_PER_PIXEL
;
1169 conf
|= (7 - rt
) << TLB_RENDER_TARGET_SHIFT
;
1171 if (c
->fs_key
->swap_color_rb
& (1 << rt
))
1172 num_components
= MAX2(num_components
, 3);
1174 assert(num_components
!= 0);
1175 switch (glsl_get_base_type(var
->type
)) {
1176 case GLSL_TYPE_UINT
:
1178 /* The F32 vs I32 distinction was dropped in 4.2. */
1179 if (c
->devinfo
->ver
< 42)
1180 conf
|= TLB_TYPE_I32_COLOR
;
1182 conf
|= TLB_TYPE_F32_COLOR
;
1183 conf
|= ((num_components
- 1) <<
1184 TLB_VEC_SIZE_MINUS_1_SHIFT
);
1186 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLBU
, 0), color
[0]);
1187 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1188 vir_uniform_ui(c
, conf
);
1190 for (int i
= 1; i
< num_components
; i
++) {
1191 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0),
1197 struct qreg r
= color
[0];
1198 struct qreg g
= color
[1];
1199 struct qreg b
= color
[2];
1200 struct qreg a
= color
[3];
1202 if (c
->fs_key
->f32_color_rb
& (1 << rt
)) {
1203 conf
|= TLB_TYPE_F32_COLOR
;
1204 conf
|= ((num_components
- 1) <<
1205 TLB_VEC_SIZE_MINUS_1_SHIFT
);
1207 conf
|= TLB_TYPE_F16_COLOR
;
1208 conf
|= TLB_F16_SWAP_HI_LO
;
1209 if (num_components
>= 3)
1210 conf
|= TLB_VEC_SIZE_4_F16
;
1212 conf
|= TLB_VEC_SIZE_2_F16
;
1215 if (c
->fs_key
->swap_color_rb
& (1 << rt
)) {
1220 if (c
->fs_key
->sample_alpha_to_one
)
1221 a
= vir_uniform_f(c
, 1.0);
1223 if (c
->fs_key
->f32_color_rb
& (1 << rt
)) {
1224 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLBU
, 0), r
);
1225 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1226 vir_uniform_ui(c
, conf
);
1228 if (num_components
>= 2)
1229 vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0), g
);
1230 if (num_components
>= 3)
1231 vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0), b
);
1232 if (num_components
>= 4)
1233 vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0), a
);
1235 inst
= vir_VFPACK_dest(c
, vir_reg(QFILE_TLB
, 0), r
, g
);
1237 inst
->dst
.file
= QFILE_TLBU
;
1238 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1239 vir_uniform_ui(c
, conf
);
1242 if (num_components
>= 3)
1243 inst
= vir_VFPACK_dest(c
, vir_reg(QFILE_TLB
, 0), b
, a
);
1252 vir_VPM_WRITE(struct v3d_compile
*c
, struct qreg val
, uint32_t *vpm_index
)
1254 if (c
->devinfo
->ver
>= 40) {
1255 vir_STVPMV(c
, vir_uniform_ui(c
, *vpm_index
), val
);
1256 *vpm_index
= *vpm_index
+ 1;
1258 vir_MOV_dest(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_VPM
), val
);
1261 c
->num_vpm_writes
++;
1265 emit_scaled_viewport_write(struct v3d_compile
*c
, struct qreg rcp_w
,
1266 uint32_t *vpm_index
)
1268 for (int i
= 0; i
< 2; i
++) {
1269 struct qreg coord
= c
->outputs
[c
->output_position_index
+ i
];
1270 coord
= vir_FMUL(c
, coord
,
1271 vir_uniform(c
, QUNIFORM_VIEWPORT_X_SCALE
+ i
,
1273 coord
= vir_FMUL(c
, coord
, rcp_w
);
1274 vir_VPM_WRITE(c
, vir_FTOIN(c
, coord
), vpm_index
);
1280 emit_zs_write(struct v3d_compile
*c
, struct qreg rcp_w
, uint32_t *vpm_index
)
1282 struct qreg zscale
= vir_uniform(c
, QUNIFORM_VIEWPORT_Z_SCALE
, 0);
1283 struct qreg zoffset
= vir_uniform(c
, QUNIFORM_VIEWPORT_Z_OFFSET
, 0);
1285 struct qreg z
= c
->outputs
[c
->output_position_index
+ 2];
1286 z
= vir_FMUL(c
, z
, zscale
);
1287 z
= vir_FMUL(c
, z
, rcp_w
);
1288 z
= vir_FADD(c
, z
, zoffset
);
1289 vir_VPM_WRITE(c
, z
, vpm_index
);
1293 emit_rcp_wc_write(struct v3d_compile
*c
, struct qreg rcp_w
, uint32_t *vpm_index
)
1295 vir_VPM_WRITE(c
, rcp_w
, vpm_index
);
1299 emit_point_size_write(struct v3d_compile
*c
, uint32_t *vpm_index
)
1301 struct qreg point_size
;
1303 if (c
->output_point_size_index
!= -1)
1304 point_size
= c
->outputs
[c
->output_point_size_index
];
1306 point_size
= vir_uniform_f(c
, 1.0);
1308 /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
1311 point_size
= vir_FMAX(c
, point_size
, vir_uniform_f(c
, .125));
1313 vir_VPM_WRITE(c
, point_size
, vpm_index
);
1317 emit_vpm_write_setup(struct v3d_compile
*c
)
1319 if (c
->devinfo
->ver
>= 40)
1322 v3d33_vir_vpm_write_setup(c
);
1326 * Sets up c->outputs[c->output_position_index] for the vertex shader
1327 * epilogue, if an output vertex position wasn't specified in the user's
1328 * shader. This may be the case for transform feedback with rasterizer
1332 setup_default_position(struct v3d_compile
*c
)
1334 if (c
->output_position_index
!= -1)
1337 c
->output_position_index
= c
->outputs_array_size
;
1338 for (int i
= 0; i
< 4; i
++) {
1340 c
->output_position_index
+ i
,
1341 VARYING_SLOT_POS
, i
);
1346 emit_vert_end(struct v3d_compile
*c
)
1348 setup_default_position(c
);
1350 uint32_t vpm_index
= 0;
1351 struct qreg rcp_w
= vir_RECIP(c
,
1352 c
->outputs
[c
->output_position_index
+ 3]);
1354 emit_vpm_write_setup(c
);
1356 if (c
->vs_key
->is_coord
) {
1357 for (int i
= 0; i
< 4; i
++)
1358 vir_VPM_WRITE(c
, c
->outputs
[c
->output_position_index
+ i
],
1360 emit_scaled_viewport_write(c
, rcp_w
, &vpm_index
);
1361 if (c
->vs_key
->per_vertex_point_size
) {
1362 emit_point_size_write(c
, &vpm_index
);
1363 /* emit_rcp_wc_write(c, rcp_w); */
1365 /* XXX: Z-only rendering */
1367 emit_zs_write(c
, rcp_w
, &vpm_index
);
1369 emit_scaled_viewport_write(c
, rcp_w
, &vpm_index
);
1370 emit_zs_write(c
, rcp_w
, &vpm_index
);
1371 emit_rcp_wc_write(c
, rcp_w
, &vpm_index
);
1372 if (c
->vs_key
->per_vertex_point_size
)
1373 emit_point_size_write(c
, &vpm_index
);
1376 for (int i
= 0; i
< c
->vs_key
->num_fs_inputs
; i
++) {
1377 struct v3d_varying_slot input
= c
->vs_key
->fs_inputs
[i
];
1380 for (j
= 0; j
< c
->num_outputs
; j
++) {
1381 struct v3d_varying_slot output
= c
->output_slots
[j
];
1383 if (!memcmp(&input
, &output
, sizeof(input
))) {
1384 vir_VPM_WRITE(c
, c
->outputs
[j
],
1389 /* Emit padding if we didn't find a declared VS output for
1392 if (j
== c
->num_outputs
)
1393 vir_VPM_WRITE(c
, vir_uniform_f(c
, 0.0),
1397 /* GFXH-1684: VPM writes need to be complete by the end of the shader.
1399 if (c
->devinfo
->ver
>= 40 && c
->devinfo
->ver
<= 42)
1404 v3d_optimize_nir(struct nir_shader
*s
)
1411 NIR_PASS_V(s
, nir_lower_vars_to_ssa
);
1412 NIR_PASS(progress
, s
, nir_lower_alu_to_scalar
);
1413 NIR_PASS(progress
, s
, nir_lower_phis_to_scalar
);
1414 NIR_PASS(progress
, s
, nir_copy_prop
);
1415 NIR_PASS(progress
, s
, nir_opt_remove_phis
);
1416 NIR_PASS(progress
, s
, nir_opt_dce
);
1417 NIR_PASS(progress
, s
, nir_opt_dead_cf
);
1418 NIR_PASS(progress
, s
, nir_opt_cse
);
1419 NIR_PASS(progress
, s
, nir_opt_peephole_select
, 8, true, true);
1420 NIR_PASS(progress
, s
, nir_opt_algebraic
);
1421 NIR_PASS(progress
, s
, nir_opt_constant_folding
);
1422 NIR_PASS(progress
, s
, nir_opt_undef
);
1425 NIR_PASS(progress
, s
, nir_opt_move_load_ubo
);
1429 driver_location_compare(const void *in_a
, const void *in_b
)
1431 const nir_variable
*const *a
= in_a
;
1432 const nir_variable
*const *b
= in_b
;
1434 return (*a
)->data
.driver_location
- (*b
)->data
.driver_location
;
1438 ntq_emit_vpm_read(struct v3d_compile
*c
,
1439 uint32_t *num_components_queued
,
1440 uint32_t *remaining
,
1443 struct qreg vpm
= vir_reg(QFILE_VPM
, vpm_index
);
1445 if (c
->devinfo
->ver
>= 40 ) {
1446 return vir_LDVPMV_IN(c
,
1448 (*num_components_queued
)++));
1451 if (*num_components_queued
!= 0) {
1452 (*num_components_queued
)--;
1454 return vir_MOV(c
, vpm
);
1457 uint32_t num_components
= MIN2(*remaining
, 32);
1459 v3d33_vir_vpm_read_setup(c
, num_components
);
1461 *num_components_queued
= num_components
- 1;
1462 *remaining
-= num_components
;
1465 return vir_MOV(c
, vpm
);
1469 ntq_setup_vpm_inputs(struct v3d_compile
*c
)
1471 /* Figure out how many components of each vertex attribute the shader
1472 * uses. Each variable should have been split to individual
1473 * components and unused ones DCEed. The vertex fetcher will load
1474 * from the start of the attribute to the number of components we
1475 * declare we need in c->vattr_sizes[].
1477 nir_foreach_variable(var
, &c
->s
->inputs
) {
1478 /* No VS attribute array support. */
1479 assert(MAX2(glsl_get_length(var
->type
), 1) == 1);
1481 unsigned loc
= var
->data
.driver_location
;
1482 int start_component
= var
->data
.location_frac
;
1483 int num_components
= glsl_get_components(var
->type
);
1485 c
->vattr_sizes
[loc
] = MAX2(c
->vattr_sizes
[loc
],
1486 start_component
+ num_components
);
1489 unsigned num_components
= 0;
1490 uint32_t vpm_components_queued
= 0;
1491 bool uses_iid
= c
->s
->info
.system_values_read
&
1492 (1ull << SYSTEM_VALUE_INSTANCE_ID
);
1493 bool uses_vid
= c
->s
->info
.system_values_read
&
1494 (1ull << SYSTEM_VALUE_VERTEX_ID
);
1495 num_components
+= uses_iid
;
1496 num_components
+= uses_vid
;
1498 for (int i
= 0; i
< ARRAY_SIZE(c
->vattr_sizes
); i
++)
1499 num_components
+= c
->vattr_sizes
[i
];
1502 c
->iid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1503 &num_components
, ~0);
1507 c
->vid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1508 &num_components
, ~0);
1511 for (int loc
= 0; loc
< ARRAY_SIZE(c
->vattr_sizes
); loc
++) {
1512 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1515 for (int i
= 0; i
< c
->vattr_sizes
[loc
]; i
++) {
1516 c
->inputs
[loc
* 4 + i
] =
1517 ntq_emit_vpm_read(c
,
1518 &vpm_components_queued
,
1525 if (c
->devinfo
->ver
>= 40) {
1526 assert(vpm_components_queued
== num_components
);
1528 assert(vpm_components_queued
== 0);
1529 assert(num_components
== 0);
1534 ntq_setup_fs_inputs(struct v3d_compile
*c
)
1536 unsigned num_entries
= 0;
1537 unsigned num_components
= 0;
1538 nir_foreach_variable(var
, &c
->s
->inputs
) {
1540 num_components
+= glsl_get_components(var
->type
);
1543 nir_variable
*vars
[num_entries
];
1546 nir_foreach_variable(var
, &c
->s
->inputs
)
1549 /* Sort the variables so that we emit the input setup in
1550 * driver_location order. This is required for VPM reads, whose data
1551 * is fetched into the VPM in driver_location (TGSI register index)
1554 qsort(&vars
, num_entries
, sizeof(*vars
), driver_location_compare
);
1556 for (unsigned i
= 0; i
< num_entries
; i
++) {
1557 nir_variable
*var
= vars
[i
];
1558 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1559 unsigned loc
= var
->data
.driver_location
;
1561 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1562 (loc
+ array_len
) * 4);
1564 if (var
->data
.location
== VARYING_SLOT_POS
) {
1565 emit_fragcoord_input(c
, loc
);
1566 } else if (var
->data
.location
== VARYING_SLOT_PNTC
||
1567 (var
->data
.location
>= VARYING_SLOT_VAR0
&&
1568 (c
->fs_key
->point_sprite_mask
&
1569 (1 << (var
->data
.location
-
1570 VARYING_SLOT_VAR0
))))) {
1571 c
->inputs
[loc
* 4 + 0] = c
->point_x
;
1572 c
->inputs
[loc
* 4 + 1] = c
->point_y
;
1574 for (int j
= 0; j
< array_len
; j
++)
1575 emit_fragment_input(c
, loc
+ j
, var
, j
);
1581 ntq_setup_outputs(struct v3d_compile
*c
)
1583 nir_foreach_variable(var
, &c
->s
->outputs
) {
1584 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1585 unsigned loc
= var
->data
.driver_location
* 4;
1587 assert(array_len
== 1);
1590 for (int i
= 0; i
< 4 - var
->data
.location_frac
; i
++) {
1591 add_output(c
, loc
+ var
->data
.location_frac
+ i
,
1593 var
->data
.location_frac
+ i
);
1596 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1597 switch (var
->data
.location
) {
1598 case FRAG_RESULT_COLOR
:
1599 c
->output_color_var
[0] = var
;
1600 c
->output_color_var
[1] = var
;
1601 c
->output_color_var
[2] = var
;
1602 c
->output_color_var
[3] = var
;
1604 case FRAG_RESULT_DATA0
:
1605 case FRAG_RESULT_DATA1
:
1606 case FRAG_RESULT_DATA2
:
1607 case FRAG_RESULT_DATA3
:
1608 c
->output_color_var
[var
->data
.location
-
1609 FRAG_RESULT_DATA0
] = var
;
1611 case FRAG_RESULT_DEPTH
:
1612 c
->output_position_index
= loc
;
1614 case FRAG_RESULT_SAMPLE_MASK
:
1615 c
->output_sample_mask_index
= loc
;
1619 switch (var
->data
.location
) {
1620 case VARYING_SLOT_POS
:
1621 c
->output_position_index
= loc
;
1623 case VARYING_SLOT_PSIZ
:
1624 c
->output_point_size_index
= loc
;
1632 ntq_setup_uniforms(struct v3d_compile
*c
)
1634 nir_foreach_variable(var
, &c
->s
->uniforms
) {
1635 uint32_t vec4_count
= glsl_count_attribute_slots(var
->type
,
1637 unsigned vec4_size
= 4 * sizeof(float);
1639 if (var
->data
.mode
!= nir_var_uniform
)
1642 declare_uniform_range(c
, var
->data
.driver_location
* vec4_size
,
1643 vec4_count
* vec4_size
);
1649 * Sets up the mapping from nir_register to struct qreg *.
1651 * Each nir_register gets a struct qreg per 32-bit component being stored.
1654 ntq_setup_registers(struct v3d_compile
*c
, struct exec_list
*list
)
1656 foreach_list_typed(nir_register
, nir_reg
, node
, list
) {
1657 unsigned array_len
= MAX2(nir_reg
->num_array_elems
, 1);
1658 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
1660 nir_reg
->num_components
);
1662 _mesa_hash_table_insert(c
->def_ht
, nir_reg
, qregs
);
1664 for (int i
= 0; i
< array_len
* nir_reg
->num_components
; i
++)
1665 qregs
[i
] = vir_get_temp(c
);
1670 ntq_emit_load_const(struct v3d_compile
*c
, nir_load_const_instr
*instr
)
1672 /* XXX perf: Experiment with using immediate loads to avoid having
1673 * these end up in the uniform stream. Watch out for breaking the
1674 * small immediates optimization in the process!
1676 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1677 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1678 qregs
[i
] = vir_uniform_ui(c
, instr
->value
.u32
[i
]);
1680 _mesa_hash_table_insert(c
->def_ht
, &instr
->def
, qregs
);
1684 ntq_emit_ssa_undef(struct v3d_compile
*c
, nir_ssa_undef_instr
*instr
)
1686 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1688 /* VIR needs there to be *some* value, so pick 0 (same as for
1689 * ntq_setup_registers().
1691 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1692 qregs
[i
] = vir_uniform_ui(c
, 0);
1696 ntq_emit_image_size(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1698 assert(instr
->intrinsic
== nir_intrinsic_image_deref_size
);
1699 nir_variable
*var
= nir_intrinsic_get_var(instr
, 0);
1700 unsigned image_index
= var
->data
.driver_location
;
1701 const struct glsl_type
*sampler_type
= glsl_without_array(var
->type
);
1702 bool is_array
= glsl_sampler_type_is_array(sampler_type
);
1704 ntq_store_dest(c
, &instr
->dest
, 0,
1705 vir_uniform(c
, QUNIFORM_IMAGE_WIDTH
, image_index
));
1706 if (instr
->num_components
> 1) {
1707 ntq_store_dest(c
, &instr
->dest
, 1,
1708 vir_uniform(c
, QUNIFORM_IMAGE_HEIGHT
,
1711 if (instr
->num_components
> 2) {
1712 ntq_store_dest(c
, &instr
->dest
, 2,
1715 QUNIFORM_IMAGE_ARRAY_SIZE
:
1716 QUNIFORM_IMAGE_DEPTH
,
1722 ntq_emit_intrinsic(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1726 switch (instr
->intrinsic
) {
1727 case nir_intrinsic_load_uniform
:
1728 if (nir_src_is_const(instr
->src
[0])) {
1729 int offset
= (nir_intrinsic_base(instr
) +
1730 nir_src_as_uint(instr
->src
[0]));
1731 assert(offset
% 4 == 0);
1732 /* We need dwords */
1733 offset
= offset
/ 4;
1734 for (int i
= 0; i
< instr
->num_components
; i
++) {
1735 ntq_store_dest(c
, &instr
->dest
, i
,
1736 vir_uniform(c
, QUNIFORM_UNIFORM
,
1740 ntq_emit_tmu_general(c
, instr
);
1744 case nir_intrinsic_load_ubo
:
1745 ntq_emit_tmu_general(c
, instr
);
1748 case nir_intrinsic_ssbo_atomic_add
:
1749 case nir_intrinsic_ssbo_atomic_imin
:
1750 case nir_intrinsic_ssbo_atomic_umin
:
1751 case nir_intrinsic_ssbo_atomic_imax
:
1752 case nir_intrinsic_ssbo_atomic_umax
:
1753 case nir_intrinsic_ssbo_atomic_and
:
1754 case nir_intrinsic_ssbo_atomic_or
:
1755 case nir_intrinsic_ssbo_atomic_xor
:
1756 case nir_intrinsic_ssbo_atomic_exchange
:
1757 case nir_intrinsic_ssbo_atomic_comp_swap
:
1758 case nir_intrinsic_load_ssbo
:
1759 case nir_intrinsic_store_ssbo
:
1760 ntq_emit_tmu_general(c
, instr
);
1763 case nir_intrinsic_image_deref_load
:
1764 case nir_intrinsic_image_deref_store
:
1765 case nir_intrinsic_image_deref_atomic_add
:
1766 case nir_intrinsic_image_deref_atomic_min
:
1767 case nir_intrinsic_image_deref_atomic_max
:
1768 case nir_intrinsic_image_deref_atomic_and
:
1769 case nir_intrinsic_image_deref_atomic_or
:
1770 case nir_intrinsic_image_deref_atomic_xor
:
1771 case nir_intrinsic_image_deref_atomic_exchange
:
1772 case nir_intrinsic_image_deref_atomic_comp_swap
:
1773 v3d40_vir_emit_image_load_store(c
, instr
);
1776 case nir_intrinsic_get_buffer_size
:
1777 ntq_store_dest(c
, &instr
->dest
, 0,
1778 vir_uniform(c
, QUNIFORM_GET_BUFFER_SIZE
,
1779 nir_src_as_uint(instr
->src
[0])));
1782 case nir_intrinsic_load_user_clip_plane
:
1783 for (int i
= 0; i
< instr
->num_components
; i
++) {
1784 ntq_store_dest(c
, &instr
->dest
, i
,
1785 vir_uniform(c
, QUNIFORM_USER_CLIP_PLANE
,
1786 nir_intrinsic_ucp_id(instr
) *
1791 case nir_intrinsic_load_alpha_ref_float
:
1792 ntq_store_dest(c
, &instr
->dest
, 0,
1793 vir_uniform(c
, QUNIFORM_ALPHA_REF
, 0));
1796 case nir_intrinsic_load_sample_mask_in
:
1797 ntq_store_dest(c
, &instr
->dest
, 0, vir_MSF(c
));
1800 case nir_intrinsic_load_helper_invocation
:
1801 vir_PF(c
, vir_MSF(c
), V3D_QPU_PF_PUSHZ
);
1802 ntq_store_dest(c
, &instr
->dest
, 0,
1803 vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFA
,
1804 vir_uniform_ui(c
, ~0),
1805 vir_uniform_ui(c
, 0))));
1808 case nir_intrinsic_load_front_face
:
1809 /* The register contains 0 (front) or 1 (back), and we need to
1810 * turn it into a NIR bool where true means front.
1812 ntq_store_dest(c
, &instr
->dest
, 0,
1814 vir_uniform_ui(c
, -1),
1818 case nir_intrinsic_load_instance_id
:
1819 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->iid
));
1822 case nir_intrinsic_load_vertex_id
:
1823 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->vid
));
1826 case nir_intrinsic_load_input
:
1827 for (int i
= 0; i
< instr
->num_components
; i
++) {
1828 offset
= (nir_intrinsic_base(instr
) +
1829 nir_src_as_uint(instr
->src
[0]));
1830 int comp
= nir_intrinsic_component(instr
) + i
;
1831 ntq_store_dest(c
, &instr
->dest
, i
,
1832 vir_MOV(c
, c
->inputs
[offset
* 4 + comp
]));
1836 case nir_intrinsic_store_output
:
1837 offset
= ((nir_intrinsic_base(instr
) +
1838 nir_src_as_uint(instr
->src
[1])) * 4 +
1839 nir_intrinsic_component(instr
));
1841 for (int i
= 0; i
< instr
->num_components
; i
++) {
1842 c
->outputs
[offset
+ i
] =
1843 vir_MOV(c
, ntq_get_src(c
, instr
->src
[0], i
));
1845 c
->num_outputs
= MAX2(c
->num_outputs
,
1846 offset
+ instr
->num_components
);
1849 case nir_intrinsic_image_deref_size
:
1850 ntq_emit_image_size(c
, instr
);
1853 case nir_intrinsic_discard
:
1854 if (c
->execute
.file
!= QFILE_NULL
) {
1855 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1856 vir_set_cond(vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1857 vir_uniform_ui(c
, 0)),
1860 vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1861 vir_uniform_ui(c
, 0));
1865 case nir_intrinsic_discard_if
: {
1866 /* true (~0) if we're discarding */
1867 struct qreg cond
= ntq_get_src(c
, instr
->src
[0], 0);
1869 if (c
->execute
.file
!= QFILE_NULL
) {
1870 /* execute == 0 means the channel is active. Invert
1871 * the condition so that we can use zero as "executing
1874 vir_PF(c
, vir_OR(c
, c
->execute
, vir_NOT(c
, cond
)),
1876 vir_set_cond(vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1877 vir_uniform_ui(c
, 0)),
1880 vir_PF(c
, cond
, V3D_QPU_PF_PUSHZ
);
1881 vir_set_cond(vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1882 vir_uniform_ui(c
, 0)),
1889 case nir_intrinsic_memory_barrier
:
1890 case nir_intrinsic_memory_barrier_atomic_counter
:
1891 case nir_intrinsic_memory_barrier_buffer
:
1892 case nir_intrinsic_memory_barrier_image
:
1893 /* We don't do any instruction scheduling of these NIR
1894 * instructions between each other, so we just need to make
1895 * sure that the TMU operations before the barrier are flushed
1896 * before the ones after the barrier. That is currently
1897 * handled by having a THRSW in each of them and a LDTMU
1898 * series or a TMUWT after.
1903 fprintf(stderr
, "Unknown intrinsic: ");
1904 nir_print_instr(&instr
->instr
, stderr
);
1905 fprintf(stderr
, "\n");
1910 /* Clears (activates) the execute flags for any channels whose jump target
1911 * matches this block.
1913 * XXX perf: Could we be using flpush/flpop somehow for our execution channel
1916 * XXX perf: For uniform control flow, we should be able to skip c->execute
1917 * handling entirely.
1920 ntq_activate_execute_for_block(struct v3d_compile
*c
)
1922 vir_set_pf(vir_XOR_dest(c
, vir_reg(QFILE_NULL
, 0),
1923 c
->execute
, vir_uniform_ui(c
, c
->cur_block
->index
)),
1926 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
1930 ntq_emit_uniform_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
1932 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
1933 bool empty_else_block
=
1934 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
1935 exec_list_is_empty(&nir_else_block
->instr_list
));
1937 struct qblock
*then_block
= vir_new_block(c
);
1938 struct qblock
*after_block
= vir_new_block(c
);
1939 struct qblock
*else_block
;
1940 if (empty_else_block
)
1941 else_block
= after_block
;
1943 else_block
= vir_new_block(c
);
1945 /* Set up the flags for the IF condition (taking the THEN branch). */
1946 nir_alu_instr
*if_condition_alu
= ntq_get_alu_parent(if_stmt
->condition
);
1947 enum v3d_qpu_cond cond
;
1948 if (!if_condition_alu
||
1949 !ntq_emit_comparison(c
, if_condition_alu
, &cond
)) {
1950 vir_PF(c
, ntq_get_src(c
, if_stmt
->condition
, 0),
1952 cond
= V3D_QPU_COND_IFNA
;
1956 vir_BRANCH(c
, cond
== V3D_QPU_COND_IFA
?
1957 V3D_QPU_BRANCH_COND_ALLNA
:
1958 V3D_QPU_BRANCH_COND_ALLA
);
1959 vir_link_blocks(c
->cur_block
, else_block
);
1960 vir_link_blocks(c
->cur_block
, then_block
);
1962 /* Process the THEN block. */
1963 vir_set_emit_block(c
, then_block
);
1964 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
1966 if (!empty_else_block
) {
1967 /* At the end of the THEN block, jump to ENDIF */
1968 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALWAYS
);
1969 vir_link_blocks(c
->cur_block
, after_block
);
1971 /* Emit the else block. */
1972 vir_set_emit_block(c
, else_block
);
1973 ntq_activate_execute_for_block(c
);
1974 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
1977 vir_link_blocks(c
->cur_block
, after_block
);
1979 vir_set_emit_block(c
, after_block
);
1983 ntq_emit_nonuniform_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
1985 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
1986 bool empty_else_block
=
1987 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
1988 exec_list_is_empty(&nir_else_block
->instr_list
));
1990 struct qblock
*then_block
= vir_new_block(c
);
1991 struct qblock
*after_block
= vir_new_block(c
);
1992 struct qblock
*else_block
;
1993 if (empty_else_block
)
1994 else_block
= after_block
;
1996 else_block
= vir_new_block(c
);
1998 bool was_top_level
= false;
1999 if (c
->execute
.file
== QFILE_NULL
) {
2000 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
2001 was_top_level
= true;
2004 /* Set up the flags for the IF condition (taking the THEN branch). */
2005 nir_alu_instr
*if_condition_alu
= ntq_get_alu_parent(if_stmt
->condition
);
2006 enum v3d_qpu_cond cond
;
2007 if (!if_condition_alu
||
2008 !ntq_emit_comparison(c
, if_condition_alu
, &cond
)) {
2009 vir_PF(c
, ntq_get_src(c
, if_stmt
->condition
, 0),
2011 cond
= V3D_QPU_COND_IFNA
;
2014 /* Update the flags+cond to mean "Taking the ELSE branch (!cond) and
2015 * was previously active (execute Z) for updating the exec flags.
2017 if (was_top_level
) {
2018 cond
= v3d_qpu_cond_invert(cond
);
2020 struct qinst
*inst
= vir_MOV_dest(c
, vir_reg(QFILE_NULL
, 0),
2022 if (cond
== V3D_QPU_COND_IFA
) {
2023 vir_set_uf(inst
, V3D_QPU_UF_NORNZ
);
2025 vir_set_uf(inst
, V3D_QPU_UF_ANDZ
);
2026 cond
= V3D_QPU_COND_IFA
;
2030 vir_MOV_cond(c
, cond
,
2032 vir_uniform_ui(c
, else_block
->index
));
2034 /* Jump to ELSE if nothing is active for THEN, otherwise fall
2037 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
2038 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLNA
);
2039 vir_link_blocks(c
->cur_block
, else_block
);
2040 vir_link_blocks(c
->cur_block
, then_block
);
2042 /* Process the THEN block. */
2043 vir_set_emit_block(c
, then_block
);
2044 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
2046 if (!empty_else_block
) {
2047 /* Handle the end of the THEN block. First, all currently
2048 * active channels update their execute flags to point to
2051 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
2052 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
2053 vir_uniform_ui(c
, after_block
->index
));
2055 /* If everything points at ENDIF, then jump there immediately. */
2056 vir_PF(c
, vir_XOR(c
, c
->execute
,
2057 vir_uniform_ui(c
, after_block
->index
)),
2059 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLA
);
2060 vir_link_blocks(c
->cur_block
, after_block
);
2061 vir_link_blocks(c
->cur_block
, else_block
);
2063 vir_set_emit_block(c
, else_block
);
2064 ntq_activate_execute_for_block(c
);
2065 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
2068 vir_link_blocks(c
->cur_block
, after_block
);
2070 vir_set_emit_block(c
, after_block
);
2072 c
->execute
= c
->undef
;
2074 ntq_activate_execute_for_block(c
);
2078 ntq_emit_if(struct v3d_compile
*c
, nir_if
*nif
)
2080 if (c
->execute
.file
== QFILE_NULL
&&
2081 nir_src_is_dynamically_uniform(nif
->condition
)) {
2082 ntq_emit_uniform_if(c
, nif
);
2084 ntq_emit_nonuniform_if(c
, nif
);
2089 ntq_emit_jump(struct v3d_compile
*c
, nir_jump_instr
*jump
)
2091 switch (jump
->type
) {
2092 case nir_jump_break
:
2093 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
2094 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
2095 vir_uniform_ui(c
, c
->loop_break_block
->index
));
2098 case nir_jump_continue
:
2099 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
2100 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
2101 vir_uniform_ui(c
, c
->loop_cont_block
->index
));
2104 case nir_jump_return
:
2105 unreachable("All returns shouold be lowered\n");
2110 ntq_emit_instr(struct v3d_compile
*c
, nir_instr
*instr
)
2112 switch (instr
->type
) {
2113 case nir_instr_type_deref
:
2114 /* ignored, will be walked by the intrinsic using it. */
2117 case nir_instr_type_alu
:
2118 ntq_emit_alu(c
, nir_instr_as_alu(instr
));
2121 case nir_instr_type_intrinsic
:
2122 ntq_emit_intrinsic(c
, nir_instr_as_intrinsic(instr
));
2125 case nir_instr_type_load_const
:
2126 ntq_emit_load_const(c
, nir_instr_as_load_const(instr
));
2129 case nir_instr_type_ssa_undef
:
2130 ntq_emit_ssa_undef(c
, nir_instr_as_ssa_undef(instr
));
2133 case nir_instr_type_tex
:
2134 ntq_emit_tex(c
, nir_instr_as_tex(instr
));
2137 case nir_instr_type_jump
:
2138 ntq_emit_jump(c
, nir_instr_as_jump(instr
));
2142 fprintf(stderr
, "Unknown NIR instr type: ");
2143 nir_print_instr(instr
, stderr
);
2144 fprintf(stderr
, "\n");
2150 ntq_emit_block(struct v3d_compile
*c
, nir_block
*block
)
2152 nir_foreach_instr(instr
, block
) {
2153 ntq_emit_instr(c
, instr
);
2157 static void ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
2160 ntq_emit_loop(struct v3d_compile
*c
, nir_loop
*loop
)
2162 bool was_top_level
= false;
2163 if (c
->execute
.file
== QFILE_NULL
) {
2164 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
2165 was_top_level
= true;
2168 struct qblock
*save_loop_cont_block
= c
->loop_cont_block
;
2169 struct qblock
*save_loop_break_block
= c
->loop_break_block
;
2171 c
->loop_cont_block
= vir_new_block(c
);
2172 c
->loop_break_block
= vir_new_block(c
);
2174 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
2175 vir_set_emit_block(c
, c
->loop_cont_block
);
2176 ntq_activate_execute_for_block(c
);
2178 ntq_emit_cf_list(c
, &loop
->body
);
2180 /* Re-enable any previous continues now, so our ANYA check below
2183 * XXX: Use the .ORZ flags update, instead.
2185 vir_PF(c
, vir_XOR(c
,
2187 vir_uniform_ui(c
, c
->loop_cont_block
->index
)),
2189 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
2191 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
2193 struct qinst
*branch
= vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ANYA
);
2194 /* Pixels that were not dispatched or have been discarded should not
2195 * contribute to looping again.
2197 branch
->qpu
.branch
.msfign
= V3D_QPU_MSFIGN_P
;
2198 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
2199 vir_link_blocks(c
->cur_block
, c
->loop_break_block
);
2201 vir_set_emit_block(c
, c
->loop_break_block
);
2203 c
->execute
= c
->undef
;
2205 ntq_activate_execute_for_block(c
);
2207 c
->loop_break_block
= save_loop_break_block
;
2208 c
->loop_cont_block
= save_loop_cont_block
;
2214 ntq_emit_function(struct v3d_compile
*c
, nir_function_impl
*func
)
2216 fprintf(stderr
, "FUNCTIONS not handled.\n");
2221 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
)
2223 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2224 switch (node
->type
) {
2225 case nir_cf_node_block
:
2226 ntq_emit_block(c
, nir_cf_node_as_block(node
));
2229 case nir_cf_node_if
:
2230 ntq_emit_if(c
, nir_cf_node_as_if(node
));
2233 case nir_cf_node_loop
:
2234 ntq_emit_loop(c
, nir_cf_node_as_loop(node
));
2237 case nir_cf_node_function
:
2238 ntq_emit_function(c
, nir_cf_node_as_function(node
));
2242 fprintf(stderr
, "Unknown NIR node type\n");
2249 ntq_emit_impl(struct v3d_compile
*c
, nir_function_impl
*impl
)
2251 ntq_setup_registers(c
, &impl
->registers
);
2252 ntq_emit_cf_list(c
, &impl
->body
);
2256 nir_to_vir(struct v3d_compile
*c
)
2258 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
2259 c
->payload_w
= vir_MOV(c
, vir_reg(QFILE_REG
, 0));
2260 c
->payload_w_centroid
= vir_MOV(c
, vir_reg(QFILE_REG
, 1));
2261 c
->payload_z
= vir_MOV(c
, vir_reg(QFILE_REG
, 2));
2263 /* XXX perf: We could set the "disable implicit point/line
2264 * varyings" field in the shader record and not emit these, if
2265 * they're not going to be used.
2267 if (c
->fs_key
->is_points
) {
2268 c
->point_x
= emit_fragment_varying(c
, NULL
, 0, 0);
2269 c
->point_y
= emit_fragment_varying(c
, NULL
, 0, 0);
2270 } else if (c
->fs_key
->is_lines
) {
2271 c
->line_x
= emit_fragment_varying(c
, NULL
, 0, 0);
2275 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
)
2276 ntq_setup_fs_inputs(c
);
2278 ntq_setup_vpm_inputs(c
);
2280 ntq_setup_outputs(c
);
2281 ntq_setup_uniforms(c
);
2282 ntq_setup_registers(c
, &c
->s
->registers
);
2284 /* Find the main function and emit the body. */
2285 nir_foreach_function(function
, c
->s
) {
2286 assert(strcmp(function
->name
, "main") == 0);
2287 assert(function
->impl
);
2288 ntq_emit_impl(c
, function
->impl
);
2292 const nir_shader_compiler_options v3d_nir_options
= {
2293 .lower_all_io_to_temps
= true,
2294 .lower_extract_byte
= true,
2295 .lower_extract_word
= true,
2297 .lower_bitfield_insert_to_shifts
= true,
2298 .lower_bitfield_extract_to_shifts
= true,
2299 .lower_bitfield_reverse
= true,
2300 .lower_bit_count
= true,
2301 .lower_pack_unorm_2x16
= true,
2302 .lower_pack_snorm_2x16
= true,
2303 .lower_pack_unorm_4x8
= true,
2304 .lower_pack_snorm_4x8
= true,
2305 .lower_unpack_unorm_4x8
= true,
2306 .lower_unpack_snorm_4x8
= true,
2307 .lower_pack_half_2x16
= true,
2308 .lower_unpack_half_2x16
= true,
2310 .lower_find_lsb
= true,
2312 .lower_flrp32
= true,
2315 .lower_fsqrt
= true,
2316 .lower_ifind_msb
= true,
2317 .lower_ldexp
= true,
2318 .lower_mul_high
= true,
2319 .lower_wpos_pntc
= true,
2320 .native_integers
= true,
2324 * When demoting a shader down to single-threaded, removes the THRSW
2325 * instructions (one will still be inserted at v3d_vir_to_qpu() for the
2329 vir_remove_thrsw(struct v3d_compile
*c
)
2331 vir_for_each_block(block
, c
) {
2332 vir_for_each_inst_safe(inst
, block
) {
2333 if (inst
->qpu
.sig
.thrsw
)
2334 vir_remove_instruction(c
, inst
);
2338 c
->last_thrsw
= NULL
;
2342 vir_emit_last_thrsw(struct v3d_compile
*c
)
2344 /* On V3D before 4.1, we need a TMU op to be outstanding when thread
2345 * switching, so disable threads if we didn't do any TMU ops (each of
2346 * which would have emitted a THRSW).
2348 if (!c
->last_thrsw_at_top_level
&& c
->devinfo
->ver
< 41) {
2351 vir_remove_thrsw(c
);
2355 /* If we're threaded and the last THRSW was in conditional code, then
2356 * we need to emit another one so that we can flag it as the last
2359 if (c
->last_thrsw
&& !c
->last_thrsw_at_top_level
) {
2360 assert(c
->devinfo
->ver
>= 41);
2364 /* If we're threaded, then we need to mark the last THRSW instruction
2365 * so we can emit a pair of them at QPU emit time.
2367 * For V3D 4.x, we can spawn the non-fragment shaders already in the
2368 * post-last-THRSW state, so we can skip this.
2370 if (!c
->last_thrsw
&& c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
2371 assert(c
->devinfo
->ver
>= 41);
2376 c
->last_thrsw
->is_last_thrsw
= true;
2379 /* There's a flag in the shader for "center W is needed for reasons other than
2380 * non-centroid varyings", so we just walk the program after VIR optimization
2381 * to see if it's used. It should be harmless to set even if we only use
2382 * center W for varyings.
2385 vir_check_payload_w(struct v3d_compile
*c
)
2387 if (c
->s
->info
.stage
!= MESA_SHADER_FRAGMENT
)
2390 vir_for_each_inst_inorder(inst
, c
) {
2391 for (int i
= 0; i
< vir_get_nsrc(inst
); i
++) {
2392 if (inst
->src
[i
].file
== QFILE_REG
&&
2393 inst
->src
[i
].index
== 0) {
2394 c
->uses_center_w
= true;
2403 v3d_nir_to_vir(struct v3d_compile
*c
)
2405 if (V3D_DEBUG
& (V3D_DEBUG_NIR
|
2406 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2407 fprintf(stderr
, "%s prog %d/%d NIR:\n",
2408 vir_get_stage_name(c
),
2409 c
->program_id
, c
->variant_id
);
2410 nir_print_shader(c
->s
, stderr
);
2415 /* Emit the last THRSW before STVPM and TLB writes. */
2416 vir_emit_last_thrsw(c
);
2418 switch (c
->s
->info
.stage
) {
2419 case MESA_SHADER_FRAGMENT
:
2422 case MESA_SHADER_VERTEX
:
2426 unreachable("bad stage");
2429 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
2430 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2431 fprintf(stderr
, "%s prog %d/%d pre-opt VIR:\n",
2432 vir_get_stage_name(c
),
2433 c
->program_id
, c
->variant_id
);
2435 fprintf(stderr
, "\n");
2439 vir_lower_uniforms(c
);
2441 vir_check_payload_w(c
);
2443 /* XXX perf: On VC4, we do a VIR-level instruction scheduling here.
2444 * We used that on that platform to pipeline TMU writes and reduce the
2445 * number of thread switches, as well as try (mostly successfully) to
2446 * reduce maximum register pressure to allow more threads. We should
2447 * do something of that sort for V3D -- either instruction scheduling
2448 * here, or delay the the THRSW and LDTMUs from our texture
2449 * instructions until the results are needed.
2452 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
2453 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2454 fprintf(stderr
, "%s prog %d/%d VIR:\n",
2455 vir_get_stage_name(c
),
2456 c
->program_id
, c
->variant_id
);
2458 fprintf(stderr
, "\n");
2461 /* Attempt to allocate registers for the temporaries. If we fail,
2462 * reduce thread count and try again.
2464 int min_threads
= (c
->devinfo
->ver
>= 41) ? 2 : 1;
2465 struct qpu_reg
*temp_registers
;
2468 temp_registers
= v3d_register_allocate(c
, &spilled
);
2475 if (c
->threads
== min_threads
) {
2476 fprintf(stderr
, "Failed to register allocate at %d threads:\n",
2485 if (c
->threads
== 1)
2486 vir_remove_thrsw(c
);
2489 v3d_vir_to_qpu(c
, temp_registers
);