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 * Implements indirect uniform loads through the TMU general memory access
115 ntq_emit_tmu_general(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
117 uint32_t tmu_op
= GENERAL_TMU_READ_OP_READ
;
118 bool has_index
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
119 int offset_src
= 0 + has_index
;
122 if (instr
->intrinsic
== nir_intrinsic_load_uniform
) {
123 offset
= vir_uniform(c
, QUNIFORM_UBO_ADDR
, 0);
125 /* Find what variable in the default uniform block this
126 * uniform load is coming from.
128 uint32_t base
= nir_intrinsic_base(instr
);
130 struct v3d_ubo_range
*range
= NULL
;
131 for (i
= 0; i
< c
->num_ubo_ranges
; i
++) {
132 range
= &c
->ubo_ranges
[i
];
133 if (base
>= range
->src_offset
&&
134 base
< range
->src_offset
+ range
->size
) {
138 /* The driver-location-based offset always has to be within a
139 * declared uniform range.
141 assert(i
!= c
->num_ubo_ranges
);
142 if (!c
->ubo_range_used
[i
]) {
143 c
->ubo_range_used
[i
] = true;
144 range
->dst_offset
= c
->next_ubo_dst_offset
;
145 c
->next_ubo_dst_offset
+= range
->size
;
148 base
= base
- range
->src_offset
+ range
->dst_offset
;
151 offset
= vir_ADD(c
, offset
, vir_uniform_ui(c
, base
));
153 /* Note that QUNIFORM_UBO_ADDR takes a UBO index shifted up by
154 * 1 (0 is gallium's constant buffer 0).
156 offset
= vir_uniform(c
, QUNIFORM_UBO_ADDR
,
157 nir_src_as_uint(instr
->src
[0]) + 1);
160 uint32_t config
= (0xffffff00 |
162 GENERAL_TMU_LOOKUP_PER_PIXEL
);
163 if (instr
->num_components
== 1) {
164 config
|= GENERAL_TMU_LOOKUP_TYPE_32BIT_UI
;
166 config
|= (GENERAL_TMU_LOOKUP_TYPE_VEC2
+
167 instr
->num_components
- 2);
172 dest
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUA
);
174 dest
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUAU
);
177 if (nir_src_is_const(instr
->src
[offset_src
]) &&
178 nir_src_as_uint(instr
->src
[offset_src
]) == 0) {
179 tmu
= vir_MOV_dest(c
, dest
, offset
);
181 tmu
= vir_ADD_dest(c
, dest
,
183 ntq_get_src(c
, instr
->src
[offset_src
], 0));
187 tmu
->src
[vir_get_implicit_uniform_src(tmu
)] =
188 vir_uniform_ui(c
, config
);
193 for (int i
= 0; i
< nir_intrinsic_dest_components(instr
); i
++)
194 ntq_store_dest(c
, &instr
->dest
, i
, vir_MOV(c
, vir_LDTMU(c
)));
198 ntq_init_ssa_def(struct v3d_compile
*c
, nir_ssa_def
*def
)
200 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
201 def
->num_components
);
202 _mesa_hash_table_insert(c
->def_ht
, def
, qregs
);
207 * This function is responsible for getting VIR results into the associated
208 * storage for a NIR instruction.
210 * If it's a NIR SSA def, then we just set the associated hash table entry to
213 * If it's a NIR reg, then we need to update the existing qreg assigned to the
214 * NIR destination with the incoming value. To do that without introducing
215 * new MOVs, we require that the incoming qreg either be a uniform, or be
216 * SSA-defined by the previous VIR instruction in the block and rewritable by
217 * this function. That lets us sneak ahead and insert the SF flag beforehand
218 * (knowing that the previous instruction doesn't depend on flags) and rewrite
219 * its destination to be the NIR reg's destination
222 ntq_store_dest(struct v3d_compile
*c
, nir_dest
*dest
, int chan
,
225 struct qinst
*last_inst
= NULL
;
226 if (!list_empty(&c
->cur_block
->instructions
))
227 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
229 assert(result
.file
== QFILE_UNIF
||
230 (result
.file
== QFILE_TEMP
&&
231 last_inst
&& last_inst
== c
->defs
[result
.index
]));
234 assert(chan
< dest
->ssa
.num_components
);
237 struct hash_entry
*entry
=
238 _mesa_hash_table_search(c
->def_ht
, &dest
->ssa
);
243 qregs
= ntq_init_ssa_def(c
, &dest
->ssa
);
245 qregs
[chan
] = result
;
247 nir_register
*reg
= dest
->reg
.reg
;
248 assert(dest
->reg
.base_offset
== 0);
249 assert(reg
->num_array_elems
== 0);
250 struct hash_entry
*entry
=
251 _mesa_hash_table_search(c
->def_ht
, reg
);
252 struct qreg
*qregs
= entry
->data
;
254 /* Insert a MOV if the source wasn't an SSA def in the
255 * previous instruction.
257 if (result
.file
== QFILE_UNIF
) {
258 result
= vir_MOV(c
, result
);
259 last_inst
= c
->defs
[result
.index
];
262 /* We know they're both temps, so just rewrite index. */
263 c
->defs
[last_inst
->dst
.index
] = NULL
;
264 last_inst
->dst
.index
= qregs
[chan
].index
;
266 /* If we're in control flow, then make this update of the reg
267 * conditional on the execution mask.
269 if (c
->execute
.file
!= QFILE_NULL
) {
270 last_inst
->dst
.index
= qregs
[chan
].index
;
272 /* Set the flags to the current exec mask.
274 c
->cursor
= vir_before_inst(last_inst
);
275 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
276 c
->cursor
= vir_after_inst(last_inst
);
278 vir_set_cond(last_inst
, V3D_QPU_COND_IFA
);
279 last_inst
->cond_is_exec_mask
= true;
285 ntq_get_src(struct v3d_compile
*c
, nir_src src
, int i
)
287 struct hash_entry
*entry
;
289 entry
= _mesa_hash_table_search(c
->def_ht
, src
.ssa
);
290 assert(i
< src
.ssa
->num_components
);
292 nir_register
*reg
= src
.reg
.reg
;
293 entry
= _mesa_hash_table_search(c
->def_ht
, reg
);
294 assert(reg
->num_array_elems
== 0);
295 assert(src
.reg
.base_offset
== 0);
296 assert(i
< reg
->num_components
);
299 struct qreg
*qregs
= entry
->data
;
304 ntq_get_alu_src(struct v3d_compile
*c
, nir_alu_instr
*instr
,
307 assert(util_is_power_of_two_or_zero(instr
->dest
.write_mask
));
308 unsigned chan
= ffs(instr
->dest
.write_mask
) - 1;
309 struct qreg r
= ntq_get_src(c
, instr
->src
[src
].src
,
310 instr
->src
[src
].swizzle
[chan
]);
312 assert(!instr
->src
[src
].abs
);
313 assert(!instr
->src
[src
].negate
);
319 ntq_minify(struct v3d_compile
*c
, struct qreg size
, struct qreg level
)
321 return vir_MAX(c
, vir_SHR(c
, size
, level
), vir_uniform_ui(c
, 1));
325 ntq_emit_txs(struct v3d_compile
*c
, nir_tex_instr
*instr
)
327 unsigned unit
= instr
->texture_index
;
328 int lod_index
= nir_tex_instr_src_index(instr
, nir_tex_src_lod
);
329 int dest_size
= nir_tex_instr_dest_size(instr
);
331 struct qreg lod
= c
->undef
;
333 lod
= ntq_get_src(c
, instr
->src
[lod_index
].src
, 0);
335 for (int i
= 0; i
< dest_size
; i
++) {
337 enum quniform_contents contents
;
339 if (instr
->is_array
&& i
== dest_size
- 1)
340 contents
= QUNIFORM_TEXTURE_ARRAY_SIZE
;
342 contents
= QUNIFORM_TEXTURE_WIDTH
+ i
;
344 struct qreg size
= vir_uniform(c
, contents
, unit
);
346 switch (instr
->sampler_dim
) {
347 case GLSL_SAMPLER_DIM_1D
:
348 case GLSL_SAMPLER_DIM_2D
:
349 case GLSL_SAMPLER_DIM_MS
:
350 case GLSL_SAMPLER_DIM_3D
:
351 case GLSL_SAMPLER_DIM_CUBE
:
352 /* Don't minify the array size. */
353 if (!(instr
->is_array
&& i
== dest_size
- 1)) {
354 size
= ntq_minify(c
, size
, lod
);
358 case GLSL_SAMPLER_DIM_RECT
:
359 /* There's no LOD field for rects */
363 unreachable("Bad sampler type");
366 ntq_store_dest(c
, &instr
->dest
, i
, size
);
371 ntq_emit_tex(struct v3d_compile
*c
, nir_tex_instr
*instr
)
373 unsigned unit
= instr
->texture_index
;
375 /* Since each texture sampling op requires uploading uniforms to
376 * reference the texture, there's no HW support for texture size and
377 * you just upload uniforms containing the size.
380 case nir_texop_query_levels
:
381 ntq_store_dest(c
, &instr
->dest
, 0,
382 vir_uniform(c
, QUNIFORM_TEXTURE_LEVELS
, unit
));
385 ntq_emit_txs(c
, instr
);
391 if (c
->devinfo
->ver
>= 40)
392 v3d40_vir_emit_tex(c
, instr
);
394 v3d33_vir_emit_tex(c
, instr
);
398 ntq_fsincos(struct v3d_compile
*c
, struct qreg src
, bool is_cos
)
400 struct qreg input
= vir_FMUL(c
, src
, vir_uniform_f(c
, 1.0f
/ M_PI
));
402 input
= vir_FADD(c
, input
, vir_uniform_f(c
, 0.5));
404 struct qreg periods
= vir_FROUND(c
, input
);
405 struct qreg sin_output
= vir_SIN(c
, vir_FSUB(c
, input
, periods
));
406 return vir_XOR(c
, sin_output
, vir_SHL(c
,
407 vir_FTOIN(c
, periods
),
408 vir_uniform_ui(c
, -1)));
412 ntq_fsign(struct v3d_compile
*c
, struct qreg src
)
414 struct qreg t
= vir_get_temp(c
);
416 vir_MOV_dest(c
, t
, vir_uniform_f(c
, 0.0));
417 vir_PF(c
, vir_FMOV(c
, src
), V3D_QPU_PF_PUSHZ
);
418 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_f(c
, 1.0));
419 vir_PF(c
, vir_FMOV(c
, src
), V3D_QPU_PF_PUSHN
);
420 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_f(c
, -1.0));
421 return vir_MOV(c
, t
);
425 ntq_isign(struct v3d_compile
*c
, struct qreg src
)
427 struct qreg t
= vir_get_temp(c
);
429 vir_MOV_dest(c
, t
, vir_uniform_ui(c
, 0));
430 vir_PF(c
, vir_MOV(c
, src
), V3D_QPU_PF_PUSHZ
);
431 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_ui(c
, 1));
432 vir_PF(c
, vir_MOV(c
, src
), V3D_QPU_PF_PUSHN
);
433 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_ui(c
, -1));
434 return vir_MOV(c
, t
);
438 emit_fragcoord_input(struct v3d_compile
*c
, int attr
)
440 c
->inputs
[attr
* 4 + 0] = vir_FXCD(c
);
441 c
->inputs
[attr
* 4 + 1] = vir_FYCD(c
);
442 c
->inputs
[attr
* 4 + 2] = c
->payload_z
;
443 c
->inputs
[attr
* 4 + 3] = vir_RECIP(c
, c
->payload_w
);
447 emit_fragment_varying(struct v3d_compile
*c
, nir_variable
*var
,
450 struct qreg r3
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R3
);
451 struct qreg r5
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R5
);
454 if (c
->devinfo
->ver
>= 41) {
455 struct qinst
*ldvary
= vir_add_inst(V3D_QPU_A_NOP
, c
->undef
,
457 ldvary
->qpu
.sig
.ldvary
= true;
458 vary
= vir_emit_def(c
, ldvary
);
460 vir_NOP(c
)->qpu
.sig
.ldvary
= true;
464 /* For gl_PointCoord input or distance along a line, we'll be called
465 * with no nir_variable, and we don't count toward VPM size so we
466 * don't track an input slot.
469 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
472 int i
= c
->num_inputs
++;
473 c
->input_slots
[i
] = v3d_slot_from_slot_and_component(var
->data
.location
,
476 switch (var
->data
.interpolation
) {
477 case INTERP_MODE_NONE
:
478 /* If a gl_FrontColor or gl_BackColor input has no interp
479 * qualifier, then if we're using glShadeModel(GL_FLAT) it
480 * needs to be flat shaded.
482 switch (var
->data
.location
) {
483 case VARYING_SLOT_COL0
:
484 case VARYING_SLOT_COL1
:
485 case VARYING_SLOT_BFC0
:
486 case VARYING_SLOT_BFC1
:
487 if (c
->fs_key
->shade_model_flat
) {
488 BITSET_SET(c
->flat_shade_flags
, i
);
489 vir_MOV_dest(c
, c
->undef
, vary
);
490 return vir_MOV(c
, r5
);
492 return vir_FADD(c
, vir_FMUL(c
, vary
,
499 case INTERP_MODE_SMOOTH
:
500 if (var
->data
.centroid
) {
501 BITSET_SET(c
->centroid_flags
, i
);
502 return vir_FADD(c
, vir_FMUL(c
, vary
,
503 c
->payload_w_centroid
), r5
);
505 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
507 case INTERP_MODE_NOPERSPECTIVE
:
508 BITSET_SET(c
->noperspective_flags
, i
);
509 return vir_FADD(c
, vir_MOV(c
, vary
), r5
);
510 case INTERP_MODE_FLAT
:
511 BITSET_SET(c
->flat_shade_flags
, i
);
512 vir_MOV_dest(c
, c
->undef
, vary
);
513 return vir_MOV(c
, r5
);
515 unreachable("Bad interp mode");
520 emit_fragment_input(struct v3d_compile
*c
, int attr
, nir_variable
*var
)
522 for (int i
= 0; i
< glsl_get_vector_elements(var
->type
); i
++) {
523 int chan
= var
->data
.location_frac
+ i
;
524 c
->inputs
[attr
* 4 + chan
] =
525 emit_fragment_varying(c
, var
, chan
);
530 add_output(struct v3d_compile
*c
,
531 uint32_t decl_offset
,
535 uint32_t old_array_size
= c
->outputs_array_size
;
536 resize_qreg_array(c
, &c
->outputs
, &c
->outputs_array_size
,
539 if (old_array_size
!= c
->outputs_array_size
) {
540 c
->output_slots
= reralloc(c
,
542 struct v3d_varying_slot
,
543 c
->outputs_array_size
);
546 c
->output_slots
[decl_offset
] =
547 v3d_slot_from_slot_and_component(slot
, swizzle
);
551 declare_uniform_range(struct v3d_compile
*c
, uint32_t start
, uint32_t size
)
553 unsigned array_id
= c
->num_ubo_ranges
++;
554 if (array_id
>= c
->ubo_ranges_array_size
) {
555 c
->ubo_ranges_array_size
= MAX2(c
->ubo_ranges_array_size
* 2,
557 c
->ubo_ranges
= reralloc(c
, c
->ubo_ranges
,
558 struct v3d_ubo_range
,
559 c
->ubo_ranges_array_size
);
560 c
->ubo_range_used
= reralloc(c
, c
->ubo_range_used
,
562 c
->ubo_ranges_array_size
);
565 c
->ubo_ranges
[array_id
].dst_offset
= 0;
566 c
->ubo_ranges
[array_id
].src_offset
= start
;
567 c
->ubo_ranges
[array_id
].size
= size
;
568 c
->ubo_range_used
[array_id
] = false;
572 * If compare_instr is a valid comparison instruction, emits the
573 * compare_instr's comparison and returns the sel_instr's return value based
574 * on the compare_instr's result.
577 ntq_emit_comparison(struct v3d_compile
*c
,
578 nir_alu_instr
*compare_instr
,
579 enum v3d_qpu_cond
*out_cond
)
581 struct qreg src0
= ntq_get_alu_src(c
, compare_instr
, 0);
583 if (nir_op_infos
[compare_instr
->op
].num_inputs
> 1)
584 src1
= ntq_get_alu_src(c
, compare_instr
, 1);
585 bool cond_invert
= false;
586 struct qreg nop
= vir_reg(QFILE_NULL
, 0);
588 switch (compare_instr
->op
) {
591 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
594 vir_set_pf(vir_XOR_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
599 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
603 vir_set_pf(vir_XOR_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
609 vir_set_pf(vir_FCMP_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
612 vir_set_pf(vir_MIN_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
616 vir_set_pf(vir_SUB_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHC
);
622 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHN
);
625 vir_set_pf(vir_MIN_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
628 vir_set_pf(vir_SUB_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHC
);
635 *out_cond
= cond_invert
? V3D_QPU_COND_IFNA
: V3D_QPU_COND_IFA
;
640 /* Finds an ALU instruction that generates our src value that could
641 * (potentially) be greedily emitted in the consuming instruction.
643 static struct nir_alu_instr
*
644 ntq_get_alu_parent(nir_src src
)
646 if (!src
.is_ssa
|| src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
648 nir_alu_instr
*instr
= nir_instr_as_alu(src
.ssa
->parent_instr
);
652 /* If the ALU instr's srcs are non-SSA, then we would have to avoid
653 * moving emission of the ALU instr down past another write of the
656 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
657 if (!instr
->src
[i
].src
.is_ssa
)
665 * Attempts to fold a comparison generating a boolean result into the
666 * condition code for selecting between two values, instead of comparing the
667 * boolean result against 0 to generate the condition code.
669 static struct qreg
ntq_emit_bcsel(struct v3d_compile
*c
, nir_alu_instr
*instr
,
672 nir_alu_instr
*compare
= ntq_get_alu_parent(instr
->src
[0].src
);
676 enum v3d_qpu_cond cond
;
677 if (ntq_emit_comparison(c
, compare
, &cond
))
678 return vir_MOV(c
, vir_SEL(c
, cond
, src
[1], src
[2]));
681 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
682 return vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
, src
[1], src
[2]));
687 ntq_emit_alu(struct v3d_compile
*c
, nir_alu_instr
*instr
)
689 /* This should always be lowered to ALU operations for V3D. */
690 assert(!instr
->dest
.saturate
);
692 /* Vectors are special in that they have non-scalarized writemasks,
693 * and just take the first swizzle channel for each argument in order
694 * into each writemask channel.
696 if (instr
->op
== nir_op_vec2
||
697 instr
->op
== nir_op_vec3
||
698 instr
->op
== nir_op_vec4
) {
700 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
701 srcs
[i
] = ntq_get_src(c
, instr
->src
[i
].src
,
702 instr
->src
[i
].swizzle
[0]);
703 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
704 ntq_store_dest(c
, &instr
->dest
.dest
, i
,
705 vir_MOV(c
, srcs
[i
]));
709 /* General case: We can just grab the one used channel per src. */
710 struct qreg src
[nir_op_infos
[instr
->op
].num_inputs
];
711 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
712 src
[i
] = ntq_get_alu_src(c
, instr
, i
);
720 result
= vir_MOV(c
, src
[0]);
724 result
= vir_XOR(c
, src
[0], vir_uniform_ui(c
, 1 << 31));
727 result
= vir_NEG(c
, src
[0]);
731 result
= vir_FMUL(c
, src
[0], src
[1]);
734 result
= vir_FADD(c
, src
[0], src
[1]);
737 result
= vir_FSUB(c
, src
[0], src
[1]);
740 result
= vir_FMIN(c
, src
[0], src
[1]);
743 result
= vir_FMAX(c
, src
[0], src
[1]);
747 result
= vir_FTOIZ(c
, src
[0]);
750 result
= vir_FTOUZ(c
, src
[0]);
753 result
= vir_ITOF(c
, src
[0]);
756 result
= vir_UTOF(c
, src
[0]);
759 result
= vir_AND(c
, src
[0], vir_uniform_f(c
, 1.0));
762 result
= vir_AND(c
, src
[0], vir_uniform_ui(c
, 1));
766 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
767 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
768 vir_uniform_ui(c
, ~0),
769 vir_uniform_ui(c
, 0)));
773 result
= vir_ADD(c
, src
[0], src
[1]);
776 result
= vir_SHR(c
, src
[0], src
[1]);
779 result
= vir_SUB(c
, src
[0], src
[1]);
782 result
= vir_ASR(c
, src
[0], src
[1]);
785 result
= vir_SHL(c
, src
[0], src
[1]);
788 result
= vir_MIN(c
, src
[0], src
[1]);
791 result
= vir_UMIN(c
, src
[0], src
[1]);
794 result
= vir_MAX(c
, src
[0], src
[1]);
797 result
= vir_UMAX(c
, src
[0], src
[1]);
800 result
= vir_AND(c
, src
[0], src
[1]);
803 result
= vir_OR(c
, src
[0], src
[1]);
806 result
= vir_XOR(c
, src
[0], src
[1]);
809 result
= vir_NOT(c
, src
[0]);
812 case nir_op_ufind_msb
:
813 result
= vir_SUB(c
, vir_uniform_ui(c
, 31), vir_CLZ(c
, src
[0]));
817 result
= vir_UMUL(c
, src
[0], src
[1]);
824 enum v3d_qpu_cond cond
;
825 MAYBE_UNUSED
bool ok
= ntq_emit_comparison(c
, instr
, &cond
);
827 result
= vir_MOV(c
, vir_SEL(c
, cond
,
828 vir_uniform_f(c
, 1.0),
829 vir_uniform_f(c
, 0.0)));
843 enum v3d_qpu_cond cond
;
844 MAYBE_UNUSED
bool ok
= ntq_emit_comparison(c
, instr
, &cond
);
846 result
= vir_MOV(c
, vir_SEL(c
, cond
,
847 vir_uniform_ui(c
, ~0),
848 vir_uniform_ui(c
, 0)));
853 result
= ntq_emit_bcsel(c
, instr
, src
);
856 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
857 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
862 result
= vir_RECIP(c
, src
[0]);
865 result
= vir_RSQRT(c
, src
[0]);
868 result
= vir_EXP(c
, src
[0]);
871 result
= vir_LOG(c
, src
[0]);
875 result
= vir_FCEIL(c
, src
[0]);
878 result
= vir_FFLOOR(c
, src
[0]);
880 case nir_op_fround_even
:
881 result
= vir_FROUND(c
, src
[0]);
884 result
= vir_FTRUNC(c
, src
[0]);
887 result
= vir_FSUB(c
, src
[0], vir_FFLOOR(c
, src
[0]));
891 result
= ntq_fsincos(c
, src
[0], false);
894 result
= ntq_fsincos(c
, src
[0], true);
898 result
= ntq_fsign(c
, src
[0]);
901 result
= ntq_isign(c
, src
[0]);
905 result
= vir_FMOV(c
, src
[0]);
906 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_ABS
);
911 result
= vir_MAX(c
, src
[0],
912 vir_SUB(c
, vir_uniform_ui(c
, 0), src
[0]));
916 case nir_op_fddx_coarse
:
917 case nir_op_fddx_fine
:
918 result
= vir_FDX(c
, src
[0]);
922 case nir_op_fddy_coarse
:
923 case nir_op_fddy_fine
:
924 result
= vir_FDY(c
, src
[0]);
927 case nir_op_uadd_carry
:
928 vir_PF(c
, vir_ADD(c
, src
[0], src
[1]), V3D_QPU_PF_PUSHC
);
929 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFA
,
930 vir_uniform_ui(c
, ~0),
931 vir_uniform_ui(c
, 0)));
934 case nir_op_pack_half_2x16_split
:
935 result
= vir_VFPACK(c
, src
[0], src
[1]);
938 case nir_op_unpack_half_2x16_split_x
:
939 /* XXX perf: It would be good to be able to merge this unpack
940 * with whatever uses our result.
942 result
= vir_FMOV(c
, src
[0]);
943 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_L
);
946 case nir_op_unpack_half_2x16_split_y
:
947 result
= vir_FMOV(c
, src
[0]);
948 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_H
);
952 fprintf(stderr
, "unknown NIR ALU inst: ");
953 nir_print_instr(&instr
->instr
, stderr
);
954 fprintf(stderr
, "\n");
958 /* We have a scalar result, so the instruction should only have a
959 * single channel written to.
961 assert(util_is_power_of_two_or_zero(instr
->dest
.write_mask
));
962 ntq_store_dest(c
, &instr
->dest
.dest
,
963 ffs(instr
->dest
.write_mask
) - 1, result
);
966 /* Each TLB read/write setup (a render target or depth buffer) takes an 8-bit
967 * specifier. They come from a register that's preloaded with 0xffffffff
968 * (0xff gets you normal vec4 f16 RT0 writes), and when one is neaded the low
969 * 8 bits are shifted off the bottom and 0xff shifted in from the top.
971 #define TLB_TYPE_F16_COLOR (3 << 6)
972 #define TLB_TYPE_I32_COLOR (1 << 6)
973 #define TLB_TYPE_F32_COLOR (0 << 6)
974 #define TLB_RENDER_TARGET_SHIFT 3 /* Reversed! 7 = RT 0, 0 = RT 7. */
975 #define TLB_SAMPLE_MODE_PER_SAMPLE (0 << 2)
976 #define TLB_SAMPLE_MODE_PER_PIXEL (1 << 2)
977 #define TLB_F16_SWAP_HI_LO (1 << 1)
978 #define TLB_VEC_SIZE_4_F16 (1 << 0)
979 #define TLB_VEC_SIZE_2_F16 (0 << 0)
980 #define TLB_VEC_SIZE_MINUS_1_SHIFT 0
982 /* Triggers Z/Stencil testing, used when the shader state's "FS modifies Z"
985 #define TLB_TYPE_DEPTH ((2 << 6) | (0 << 4))
986 #define TLB_DEPTH_TYPE_INVARIANT (0 << 2) /* Unmodified sideband input used */
987 #define TLB_DEPTH_TYPE_PER_PIXEL (1 << 2) /* QPU result used */
988 #define TLB_V42_DEPTH_TYPE_INVARIANT (0 << 3) /* Unmodified sideband input used */
989 #define TLB_V42_DEPTH_TYPE_PER_PIXEL (1 << 3) /* QPU result used */
991 /* Stencil is a single 32-bit write. */
992 #define TLB_TYPE_STENCIL_ALPHA ((2 << 6) | (1 << 4))
995 emit_frag_end(struct v3d_compile
*c
)
998 if (c->output_sample_mask_index != -1) {
999 vir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
1003 bool has_any_tlb_color_write
= false;
1004 for (int rt
= 0; rt
< c
->fs_key
->nr_cbufs
; rt
++) {
1005 if (c
->output_color_var
[rt
])
1006 has_any_tlb_color_write
= true;
1009 if (c
->fs_key
->sample_alpha_to_coverage
&& c
->output_color_var
[0]) {
1010 struct nir_variable
*var
= c
->output_color_var
[0];
1011 struct qreg
*color
= &c
->outputs
[var
->data
.driver_location
* 4];
1013 vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1016 vir_FTOC(c
, color
[3])));
1019 if (c
->output_position_index
!= -1) {
1020 struct qinst
*inst
= vir_MOV_dest(c
,
1021 vir_reg(QFILE_TLBU
, 0),
1022 c
->outputs
[c
->output_position_index
]);
1023 uint8_t tlb_specifier
= TLB_TYPE_DEPTH
;
1025 if (c
->devinfo
->ver
>= 42) {
1026 tlb_specifier
|= (TLB_V42_DEPTH_TYPE_PER_PIXEL
|
1027 TLB_SAMPLE_MODE_PER_PIXEL
);
1029 tlb_specifier
|= TLB_DEPTH_TYPE_PER_PIXEL
;
1031 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1032 vir_uniform_ui(c
, tlb_specifier
| 0xffffff00);
1033 } else if (c
->s
->info
.fs
.uses_discard
||
1034 c
->fs_key
->sample_alpha_to_coverage
||
1035 !has_any_tlb_color_write
) {
1036 /* Emit passthrough Z if it needed to be delayed until shader
1037 * end due to potential discards.
1039 * Since (single-threaded) fragment shaders always need a TLB
1040 * write, emit passthrouh Z if we didn't have any color
1041 * buffers and flag us as potentially discarding, so that we
1042 * can use Z as the TLB write.
1044 c
->s
->info
.fs
.uses_discard
= true;
1046 struct qinst
*inst
= vir_MOV_dest(c
,
1047 vir_reg(QFILE_TLBU
, 0),
1048 vir_reg(QFILE_NULL
, 0));
1049 uint8_t tlb_specifier
= TLB_TYPE_DEPTH
;
1051 if (c
->devinfo
->ver
>= 42) {
1052 /* The spec says the PER_PIXEL flag is ignored for
1053 * invariant writes, but the simulator demands it.
1055 tlb_specifier
|= (TLB_V42_DEPTH_TYPE_INVARIANT
|
1056 TLB_SAMPLE_MODE_PER_PIXEL
);
1058 tlb_specifier
|= TLB_DEPTH_TYPE_INVARIANT
;
1061 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1062 vir_uniform_ui(c
, tlb_specifier
| 0xffffff00);
1065 /* XXX: Performance improvement: Merge Z write and color writes TLB
1069 for (int rt
= 0; rt
< c
->fs_key
->nr_cbufs
; rt
++) {
1070 if (!c
->output_color_var
[rt
])
1073 nir_variable
*var
= c
->output_color_var
[rt
];
1074 struct qreg
*color
= &c
->outputs
[var
->data
.driver_location
* 4];
1075 int num_components
= glsl_get_vector_elements(var
->type
);
1076 uint32_t conf
= 0xffffff00;
1079 conf
|= TLB_SAMPLE_MODE_PER_PIXEL
;
1080 conf
|= (7 - rt
) << TLB_RENDER_TARGET_SHIFT
;
1082 if (c
->fs_key
->swap_color_rb
& (1 << rt
))
1083 num_components
= MAX2(num_components
, 3);
1085 assert(num_components
!= 0);
1086 switch (glsl_get_base_type(var
->type
)) {
1087 case GLSL_TYPE_UINT
:
1089 /* The F32 vs I32 distinction was dropped in 4.2. */
1090 if (c
->devinfo
->ver
< 42)
1091 conf
|= TLB_TYPE_I32_COLOR
;
1093 conf
|= TLB_TYPE_F32_COLOR
;
1094 conf
|= ((num_components
- 1) <<
1095 TLB_VEC_SIZE_MINUS_1_SHIFT
);
1097 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLBU
, 0), color
[0]);
1098 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1099 vir_uniform_ui(c
, conf
);
1101 for (int i
= 1; i
< num_components
; i
++) {
1102 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0),
1108 struct qreg r
= color
[0];
1109 struct qreg g
= color
[1];
1110 struct qreg b
= color
[2];
1111 struct qreg a
= color
[3];
1113 if (c
->fs_key
->f32_color_rb
& (1 << rt
)) {
1114 conf
|= TLB_TYPE_F32_COLOR
;
1115 conf
|= ((num_components
- 1) <<
1116 TLB_VEC_SIZE_MINUS_1_SHIFT
);
1118 conf
|= TLB_TYPE_F16_COLOR
;
1119 conf
|= TLB_F16_SWAP_HI_LO
;
1120 if (num_components
>= 3)
1121 conf
|= TLB_VEC_SIZE_4_F16
;
1123 conf
|= TLB_VEC_SIZE_2_F16
;
1126 if (c
->fs_key
->swap_color_rb
& (1 << rt
)) {
1131 if (c
->fs_key
->sample_alpha_to_one
)
1132 a
= vir_uniform_f(c
, 1.0);
1134 if (c
->fs_key
->f32_color_rb
& (1 << rt
)) {
1135 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLBU
, 0), r
);
1136 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1137 vir_uniform_ui(c
, conf
);
1139 if (num_components
>= 2)
1140 vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0), g
);
1141 if (num_components
>= 3)
1142 vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0), b
);
1143 if (num_components
>= 4)
1144 vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0), a
);
1146 inst
= vir_VFPACK_dest(c
, vir_reg(QFILE_TLB
, 0), r
, g
);
1148 inst
->dst
.file
= QFILE_TLBU
;
1149 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1150 vir_uniform_ui(c
, conf
);
1153 if (num_components
>= 3)
1154 inst
= vir_VFPACK_dest(c
, vir_reg(QFILE_TLB
, 0), b
, a
);
1163 vir_VPM_WRITE(struct v3d_compile
*c
, struct qreg val
, uint32_t *vpm_index
)
1165 if (c
->devinfo
->ver
>= 40) {
1166 vir_STVPMV(c
, vir_uniform_ui(c
, *vpm_index
), val
);
1167 *vpm_index
= *vpm_index
+ 1;
1169 vir_MOV_dest(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_VPM
), val
);
1172 c
->num_vpm_writes
++;
1176 emit_scaled_viewport_write(struct v3d_compile
*c
, struct qreg rcp_w
,
1177 uint32_t *vpm_index
)
1179 for (int i
= 0; i
< 2; i
++) {
1180 struct qreg coord
= c
->outputs
[c
->output_position_index
+ i
];
1181 coord
= vir_FMUL(c
, coord
,
1182 vir_uniform(c
, QUNIFORM_VIEWPORT_X_SCALE
+ i
,
1184 coord
= vir_FMUL(c
, coord
, rcp_w
);
1185 vir_VPM_WRITE(c
, vir_FTOIN(c
, coord
), vpm_index
);
1191 emit_zs_write(struct v3d_compile
*c
, struct qreg rcp_w
, uint32_t *vpm_index
)
1193 struct qreg zscale
= vir_uniform(c
, QUNIFORM_VIEWPORT_Z_SCALE
, 0);
1194 struct qreg zoffset
= vir_uniform(c
, QUNIFORM_VIEWPORT_Z_OFFSET
, 0);
1196 struct qreg z
= c
->outputs
[c
->output_position_index
+ 2];
1197 z
= vir_FMUL(c
, z
, zscale
);
1198 z
= vir_FMUL(c
, z
, rcp_w
);
1199 z
= vir_FADD(c
, z
, zoffset
);
1200 vir_VPM_WRITE(c
, z
, vpm_index
);
1204 emit_rcp_wc_write(struct v3d_compile
*c
, struct qreg rcp_w
, uint32_t *vpm_index
)
1206 vir_VPM_WRITE(c
, rcp_w
, vpm_index
);
1210 emit_point_size_write(struct v3d_compile
*c
, uint32_t *vpm_index
)
1212 struct qreg point_size
;
1214 if (c
->output_point_size_index
!= -1)
1215 point_size
= c
->outputs
[c
->output_point_size_index
];
1217 point_size
= vir_uniform_f(c
, 1.0);
1219 /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
1222 point_size
= vir_FMAX(c
, point_size
, vir_uniform_f(c
, .125));
1224 vir_VPM_WRITE(c
, point_size
, vpm_index
);
1228 emit_vpm_write_setup(struct v3d_compile
*c
)
1230 if (c
->devinfo
->ver
>= 40)
1233 v3d33_vir_vpm_write_setup(c
);
1237 * Sets up c->outputs[c->output_position_index] for the vertex shader
1238 * epilogue, if an output vertex position wasn't specified in the user's
1239 * shader. This may be the case for transform feedback with rasterizer
1243 setup_default_position(struct v3d_compile
*c
)
1245 if (c
->output_position_index
!= -1)
1248 c
->output_position_index
= c
->outputs_array_size
;
1249 for (int i
= 0; i
< 4; i
++) {
1251 c
->output_position_index
+ i
,
1252 VARYING_SLOT_POS
, i
);
1257 emit_vert_end(struct v3d_compile
*c
)
1259 setup_default_position(c
);
1261 uint32_t vpm_index
= 0;
1262 struct qreg rcp_w
= vir_RECIP(c
,
1263 c
->outputs
[c
->output_position_index
+ 3]);
1265 emit_vpm_write_setup(c
);
1267 if (c
->vs_key
->is_coord
) {
1268 for (int i
= 0; i
< 4; i
++)
1269 vir_VPM_WRITE(c
, c
->outputs
[c
->output_position_index
+ i
],
1271 emit_scaled_viewport_write(c
, rcp_w
, &vpm_index
);
1272 if (c
->vs_key
->per_vertex_point_size
) {
1273 emit_point_size_write(c
, &vpm_index
);
1274 /* emit_rcp_wc_write(c, rcp_w); */
1276 /* XXX: Z-only rendering */
1278 emit_zs_write(c
, rcp_w
, &vpm_index
);
1280 emit_scaled_viewport_write(c
, rcp_w
, &vpm_index
);
1281 emit_zs_write(c
, rcp_w
, &vpm_index
);
1282 emit_rcp_wc_write(c
, rcp_w
, &vpm_index
);
1283 if (c
->vs_key
->per_vertex_point_size
)
1284 emit_point_size_write(c
, &vpm_index
);
1287 for (int i
= 0; i
< c
->vs_key
->num_fs_inputs
; i
++) {
1288 struct v3d_varying_slot input
= c
->vs_key
->fs_inputs
[i
];
1291 for (j
= 0; j
< c
->num_outputs
; j
++) {
1292 struct v3d_varying_slot output
= c
->output_slots
[j
];
1294 if (!memcmp(&input
, &output
, sizeof(input
))) {
1295 vir_VPM_WRITE(c
, c
->outputs
[j
],
1300 /* Emit padding if we didn't find a declared VS output for
1303 if (j
== c
->num_outputs
)
1304 vir_VPM_WRITE(c
, vir_uniform_f(c
, 0.0),
1308 /* GFXH-1684: VPM writes need to be complete by the end of the shader.
1310 if (c
->devinfo
->ver
>= 40 && c
->devinfo
->ver
<= 42)
1315 v3d_optimize_nir(struct nir_shader
*s
)
1322 NIR_PASS_V(s
, nir_lower_vars_to_ssa
);
1323 NIR_PASS(progress
, s
, nir_lower_alu_to_scalar
);
1324 NIR_PASS(progress
, s
, nir_lower_phis_to_scalar
);
1325 NIR_PASS(progress
, s
, nir_copy_prop
);
1326 NIR_PASS(progress
, s
, nir_opt_remove_phis
);
1327 NIR_PASS(progress
, s
, nir_opt_dce
);
1328 NIR_PASS(progress
, s
, nir_opt_dead_cf
);
1329 NIR_PASS(progress
, s
, nir_opt_cse
);
1330 NIR_PASS(progress
, s
, nir_opt_peephole_select
, 8, true, true);
1331 NIR_PASS(progress
, s
, nir_opt_algebraic
);
1332 NIR_PASS(progress
, s
, nir_opt_constant_folding
);
1333 NIR_PASS(progress
, s
, nir_opt_undef
);
1336 NIR_PASS(progress
, s
, nir_opt_move_load_ubo
);
1340 driver_location_compare(const void *in_a
, const void *in_b
)
1342 const nir_variable
*const *a
= in_a
;
1343 const nir_variable
*const *b
= in_b
;
1345 return (*a
)->data
.driver_location
- (*b
)->data
.driver_location
;
1349 ntq_emit_vpm_read(struct v3d_compile
*c
,
1350 uint32_t *num_components_queued
,
1351 uint32_t *remaining
,
1354 struct qreg vpm
= vir_reg(QFILE_VPM
, vpm_index
);
1356 if (c
->devinfo
->ver
>= 40 ) {
1357 return vir_LDVPMV_IN(c
,
1359 (*num_components_queued
)++));
1362 if (*num_components_queued
!= 0) {
1363 (*num_components_queued
)--;
1365 return vir_MOV(c
, vpm
);
1368 uint32_t num_components
= MIN2(*remaining
, 32);
1370 v3d33_vir_vpm_read_setup(c
, num_components
);
1372 *num_components_queued
= num_components
- 1;
1373 *remaining
-= num_components
;
1376 return vir_MOV(c
, vpm
);
1380 ntq_setup_vpm_inputs(struct v3d_compile
*c
)
1382 /* Figure out how many components of each vertex attribute the shader
1383 * uses. Each variable should have been split to individual
1384 * components and unused ones DCEed. The vertex fetcher will load
1385 * from the start of the attribute to the number of components we
1386 * declare we need in c->vattr_sizes[].
1388 nir_foreach_variable(var
, &c
->s
->inputs
) {
1389 /* No VS attribute array support. */
1390 assert(MAX2(glsl_get_length(var
->type
), 1) == 1);
1392 unsigned loc
= var
->data
.driver_location
;
1393 int start_component
= var
->data
.location_frac
;
1394 int num_components
= glsl_get_components(var
->type
);
1396 c
->vattr_sizes
[loc
] = MAX2(c
->vattr_sizes
[loc
],
1397 start_component
+ num_components
);
1400 unsigned num_components
= 0;
1401 uint32_t vpm_components_queued
= 0;
1402 bool uses_iid
= c
->s
->info
.system_values_read
&
1403 (1ull << SYSTEM_VALUE_INSTANCE_ID
);
1404 bool uses_vid
= c
->s
->info
.system_values_read
&
1405 (1ull << SYSTEM_VALUE_VERTEX_ID
);
1406 num_components
+= uses_iid
;
1407 num_components
+= uses_vid
;
1409 for (int i
= 0; i
< ARRAY_SIZE(c
->vattr_sizes
); i
++)
1410 num_components
+= c
->vattr_sizes
[i
];
1413 c
->iid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1414 &num_components
, ~0);
1418 c
->vid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1419 &num_components
, ~0);
1422 for (int loc
= 0; loc
< ARRAY_SIZE(c
->vattr_sizes
); loc
++) {
1423 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1426 for (int i
= 0; i
< c
->vattr_sizes
[loc
]; i
++) {
1427 c
->inputs
[loc
* 4 + i
] =
1428 ntq_emit_vpm_read(c
,
1429 &vpm_components_queued
,
1436 if (c
->devinfo
->ver
>= 40) {
1437 assert(vpm_components_queued
== num_components
);
1439 assert(vpm_components_queued
== 0);
1440 assert(num_components
== 0);
1445 ntq_setup_fs_inputs(struct v3d_compile
*c
)
1447 unsigned num_entries
= 0;
1448 unsigned num_components
= 0;
1449 nir_foreach_variable(var
, &c
->s
->inputs
) {
1451 num_components
+= glsl_get_components(var
->type
);
1454 nir_variable
*vars
[num_entries
];
1457 nir_foreach_variable(var
, &c
->s
->inputs
)
1460 /* Sort the variables so that we emit the input setup in
1461 * driver_location order. This is required for VPM reads, whose data
1462 * is fetched into the VPM in driver_location (TGSI register index)
1465 qsort(&vars
, num_entries
, sizeof(*vars
), driver_location_compare
);
1467 for (unsigned i
= 0; i
< num_entries
; i
++) {
1468 nir_variable
*var
= vars
[i
];
1469 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1470 unsigned loc
= var
->data
.driver_location
;
1472 assert(array_len
== 1);
1474 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1477 if (var
->data
.location
== VARYING_SLOT_POS
) {
1478 emit_fragcoord_input(c
, loc
);
1479 } else if (var
->data
.location
== VARYING_SLOT_PNTC
||
1480 (var
->data
.location
>= VARYING_SLOT_VAR0
&&
1481 (c
->fs_key
->point_sprite_mask
&
1482 (1 << (var
->data
.location
-
1483 VARYING_SLOT_VAR0
))))) {
1484 c
->inputs
[loc
* 4 + 0] = c
->point_x
;
1485 c
->inputs
[loc
* 4 + 1] = c
->point_y
;
1487 emit_fragment_input(c
, loc
, var
);
1493 ntq_setup_outputs(struct v3d_compile
*c
)
1495 nir_foreach_variable(var
, &c
->s
->outputs
) {
1496 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1497 unsigned loc
= var
->data
.driver_location
* 4;
1499 assert(array_len
== 1);
1502 for (int i
= 0; i
< 4 - var
->data
.location_frac
; i
++) {
1503 add_output(c
, loc
+ var
->data
.location_frac
+ i
,
1505 var
->data
.location_frac
+ i
);
1508 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1509 switch (var
->data
.location
) {
1510 case FRAG_RESULT_COLOR
:
1511 c
->output_color_var
[0] = var
;
1512 c
->output_color_var
[1] = var
;
1513 c
->output_color_var
[2] = var
;
1514 c
->output_color_var
[3] = var
;
1516 case FRAG_RESULT_DATA0
:
1517 case FRAG_RESULT_DATA1
:
1518 case FRAG_RESULT_DATA2
:
1519 case FRAG_RESULT_DATA3
:
1520 c
->output_color_var
[var
->data
.location
-
1521 FRAG_RESULT_DATA0
] = var
;
1523 case FRAG_RESULT_DEPTH
:
1524 c
->output_position_index
= loc
;
1526 case FRAG_RESULT_SAMPLE_MASK
:
1527 c
->output_sample_mask_index
= loc
;
1531 switch (var
->data
.location
) {
1532 case VARYING_SLOT_POS
:
1533 c
->output_position_index
= loc
;
1535 case VARYING_SLOT_PSIZ
:
1536 c
->output_point_size_index
= loc
;
1544 ntq_setup_uniforms(struct v3d_compile
*c
)
1546 nir_foreach_variable(var
, &c
->s
->uniforms
) {
1547 uint32_t vec4_count
= glsl_count_attribute_slots(var
->type
,
1549 unsigned vec4_size
= 4 * sizeof(float);
1551 declare_uniform_range(c
, var
->data
.driver_location
* vec4_size
,
1552 vec4_count
* vec4_size
);
1558 * Sets up the mapping from nir_register to struct qreg *.
1560 * Each nir_register gets a struct qreg per 32-bit component being stored.
1563 ntq_setup_registers(struct v3d_compile
*c
, struct exec_list
*list
)
1565 foreach_list_typed(nir_register
, nir_reg
, node
, list
) {
1566 unsigned array_len
= MAX2(nir_reg
->num_array_elems
, 1);
1567 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
1569 nir_reg
->num_components
);
1571 _mesa_hash_table_insert(c
->def_ht
, nir_reg
, qregs
);
1573 for (int i
= 0; i
< array_len
* nir_reg
->num_components
; i
++)
1574 qregs
[i
] = vir_get_temp(c
);
1579 ntq_emit_load_const(struct v3d_compile
*c
, nir_load_const_instr
*instr
)
1581 /* XXX perf: Experiment with using immediate loads to avoid having
1582 * these end up in the uniform stream. Watch out for breaking the
1583 * small immediates optimization in the process!
1585 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1586 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1587 qregs
[i
] = vir_uniform_ui(c
, instr
->value
.u32
[i
]);
1589 _mesa_hash_table_insert(c
->def_ht
, &instr
->def
, qregs
);
1593 ntq_emit_ssa_undef(struct v3d_compile
*c
, nir_ssa_undef_instr
*instr
)
1595 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1597 /* VIR needs there to be *some* value, so pick 0 (same as for
1598 * ntq_setup_registers().
1600 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1601 qregs
[i
] = vir_uniform_ui(c
, 0);
1605 ntq_emit_intrinsic(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1609 switch (instr
->intrinsic
) {
1610 case nir_intrinsic_load_uniform
:
1611 if (nir_src_is_const(instr
->src
[0])) {
1612 int offset
= (nir_intrinsic_base(instr
) +
1613 nir_src_as_uint(instr
->src
[0]));
1614 assert(offset
% 4 == 0);
1615 /* We need dwords */
1616 offset
= offset
/ 4;
1617 for (int i
= 0; i
< instr
->num_components
; i
++) {
1618 ntq_store_dest(c
, &instr
->dest
, i
,
1619 vir_uniform(c
, QUNIFORM_UNIFORM
,
1623 ntq_emit_tmu_general(c
, instr
);
1627 case nir_intrinsic_load_ubo
:
1628 ntq_emit_tmu_general(c
, instr
);
1631 case nir_intrinsic_load_user_clip_plane
:
1632 for (int i
= 0; i
< instr
->num_components
; i
++) {
1633 ntq_store_dest(c
, &instr
->dest
, i
,
1634 vir_uniform(c
, QUNIFORM_USER_CLIP_PLANE
,
1635 nir_intrinsic_ucp_id(instr
) *
1640 case nir_intrinsic_load_alpha_ref_float
:
1641 ntq_store_dest(c
, &instr
->dest
, 0,
1642 vir_uniform(c
, QUNIFORM_ALPHA_REF
, 0));
1645 case nir_intrinsic_load_sample_mask_in
:
1646 ntq_store_dest(c
, &instr
->dest
, 0, vir_MSF(c
));
1649 case nir_intrinsic_load_helper_invocation
:
1650 vir_PF(c
, vir_MSF(c
), V3D_QPU_PF_PUSHZ
);
1651 ntq_store_dest(c
, &instr
->dest
, 0,
1652 vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFA
,
1653 vir_uniform_ui(c
, ~0),
1654 vir_uniform_ui(c
, 0))));
1657 case nir_intrinsic_load_front_face
:
1658 /* The register contains 0 (front) or 1 (back), and we need to
1659 * turn it into a NIR bool where true means front.
1661 ntq_store_dest(c
, &instr
->dest
, 0,
1663 vir_uniform_ui(c
, -1),
1667 case nir_intrinsic_load_instance_id
:
1668 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->iid
));
1671 case nir_intrinsic_load_vertex_id
:
1672 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->vid
));
1675 case nir_intrinsic_load_input
:
1676 for (int i
= 0; i
< instr
->num_components
; i
++) {
1677 offset
= (nir_intrinsic_base(instr
) +
1678 nir_src_as_uint(instr
->src
[0]));
1679 int comp
= nir_intrinsic_component(instr
) + i
;
1680 ntq_store_dest(c
, &instr
->dest
, i
,
1681 vir_MOV(c
, c
->inputs
[offset
* 4 + comp
]));
1685 case nir_intrinsic_store_output
:
1686 offset
= ((nir_intrinsic_base(instr
) +
1687 nir_src_as_uint(instr
->src
[1])) * 4 +
1688 nir_intrinsic_component(instr
));
1690 for (int i
= 0; i
< instr
->num_components
; i
++) {
1691 c
->outputs
[offset
+ i
] =
1692 vir_MOV(c
, ntq_get_src(c
, instr
->src
[0], i
));
1694 c
->num_outputs
= MAX2(c
->num_outputs
,
1695 offset
+ instr
->num_components
);
1698 case nir_intrinsic_discard
:
1699 if (c
->execute
.file
!= QFILE_NULL
) {
1700 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1701 vir_set_cond(vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1702 vir_uniform_ui(c
, 0)),
1705 vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1706 vir_uniform_ui(c
, 0));
1710 case nir_intrinsic_discard_if
: {
1711 /* true (~0) if we're discarding */
1712 struct qreg cond
= ntq_get_src(c
, instr
->src
[0], 0);
1714 if (c
->execute
.file
!= QFILE_NULL
) {
1715 /* execute == 0 means the channel is active. Invert
1716 * the condition so that we can use zero as "executing
1719 vir_PF(c
, vir_OR(c
, c
->execute
, vir_NOT(c
, cond
)),
1721 vir_set_cond(vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1722 vir_uniform_ui(c
, 0)),
1725 vir_PF(c
, cond
, V3D_QPU_PF_PUSHZ
);
1726 vir_set_cond(vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1727 vir_uniform_ui(c
, 0)),
1735 fprintf(stderr
, "Unknown intrinsic: ");
1736 nir_print_instr(&instr
->instr
, stderr
);
1737 fprintf(stderr
, "\n");
1742 /* Clears (activates) the execute flags for any channels whose jump target
1743 * matches this block.
1745 * XXX perf: Could we be using flpush/flpop somehow for our execution channel
1748 * XXX perf: For uniform control flow, we should be able to skip c->execute
1749 * handling entirely.
1752 ntq_activate_execute_for_block(struct v3d_compile
*c
)
1754 vir_set_pf(vir_XOR_dest(c
, vir_reg(QFILE_NULL
, 0),
1755 c
->execute
, vir_uniform_ui(c
, c
->cur_block
->index
)),
1758 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
1762 ntq_emit_uniform_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
1764 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
1765 bool empty_else_block
=
1766 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
1767 exec_list_is_empty(&nir_else_block
->instr_list
));
1769 struct qblock
*then_block
= vir_new_block(c
);
1770 struct qblock
*after_block
= vir_new_block(c
);
1771 struct qblock
*else_block
;
1772 if (empty_else_block
)
1773 else_block
= after_block
;
1775 else_block
= vir_new_block(c
);
1777 /* Set up the flags for the IF condition (taking the THEN branch). */
1778 nir_alu_instr
*if_condition_alu
= ntq_get_alu_parent(if_stmt
->condition
);
1779 enum v3d_qpu_cond cond
;
1780 if (!if_condition_alu
||
1781 !ntq_emit_comparison(c
, if_condition_alu
, &cond
)) {
1782 vir_PF(c
, ntq_get_src(c
, if_stmt
->condition
, 0),
1784 cond
= V3D_QPU_COND_IFNA
;
1788 vir_BRANCH(c
, cond
== V3D_QPU_COND_IFA
?
1789 V3D_QPU_BRANCH_COND_ALLNA
:
1790 V3D_QPU_BRANCH_COND_ALLA
);
1791 vir_link_blocks(c
->cur_block
, else_block
);
1792 vir_link_blocks(c
->cur_block
, then_block
);
1794 /* Process the THEN block. */
1795 vir_set_emit_block(c
, then_block
);
1796 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
1798 if (!empty_else_block
) {
1799 /* At the end of the THEN block, jump to ENDIF */
1800 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALWAYS
);
1801 vir_link_blocks(c
->cur_block
, after_block
);
1803 /* Emit the else block. */
1804 vir_set_emit_block(c
, else_block
);
1805 ntq_activate_execute_for_block(c
);
1806 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
1809 vir_link_blocks(c
->cur_block
, after_block
);
1811 vir_set_emit_block(c
, after_block
);
1815 ntq_emit_nonuniform_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
1817 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
1818 bool empty_else_block
=
1819 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
1820 exec_list_is_empty(&nir_else_block
->instr_list
));
1822 struct qblock
*then_block
= vir_new_block(c
);
1823 struct qblock
*after_block
= vir_new_block(c
);
1824 struct qblock
*else_block
;
1825 if (empty_else_block
)
1826 else_block
= after_block
;
1828 else_block
= vir_new_block(c
);
1830 bool was_top_level
= false;
1831 if (c
->execute
.file
== QFILE_NULL
) {
1832 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
1833 was_top_level
= true;
1836 /* Set up the flags for the IF condition (taking the THEN branch). */
1837 nir_alu_instr
*if_condition_alu
= ntq_get_alu_parent(if_stmt
->condition
);
1838 enum v3d_qpu_cond cond
;
1839 if (!if_condition_alu
||
1840 !ntq_emit_comparison(c
, if_condition_alu
, &cond
)) {
1841 vir_PF(c
, ntq_get_src(c
, if_stmt
->condition
, 0),
1843 cond
= V3D_QPU_COND_IFNA
;
1846 /* Update the flags+cond to mean "Taking the ELSE branch (!cond) and
1847 * was previously active (execute Z) for updating the exec flags.
1849 if (was_top_level
) {
1850 cond
= v3d_qpu_cond_invert(cond
);
1852 struct qinst
*inst
= vir_MOV_dest(c
, vir_reg(QFILE_NULL
, 0),
1854 if (cond
== V3D_QPU_COND_IFA
) {
1855 vir_set_uf(inst
, V3D_QPU_UF_NORNZ
);
1857 vir_set_uf(inst
, V3D_QPU_UF_ANDZ
);
1858 cond
= V3D_QPU_COND_IFA
;
1862 vir_MOV_cond(c
, cond
,
1864 vir_uniform_ui(c
, else_block
->index
));
1866 /* Jump to ELSE if nothing is active for THEN, otherwise fall
1869 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1870 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLNA
);
1871 vir_link_blocks(c
->cur_block
, else_block
);
1872 vir_link_blocks(c
->cur_block
, then_block
);
1874 /* Process the THEN block. */
1875 vir_set_emit_block(c
, then_block
);
1876 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
1878 if (!empty_else_block
) {
1879 /* Handle the end of the THEN block. First, all currently
1880 * active channels update their execute flags to point to
1883 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1884 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
1885 vir_uniform_ui(c
, after_block
->index
));
1887 /* If everything points at ENDIF, then jump there immediately. */
1888 vir_PF(c
, vir_XOR(c
, c
->execute
,
1889 vir_uniform_ui(c
, after_block
->index
)),
1891 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLA
);
1892 vir_link_blocks(c
->cur_block
, after_block
);
1893 vir_link_blocks(c
->cur_block
, else_block
);
1895 vir_set_emit_block(c
, else_block
);
1896 ntq_activate_execute_for_block(c
);
1897 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
1900 vir_link_blocks(c
->cur_block
, after_block
);
1902 vir_set_emit_block(c
, after_block
);
1904 c
->execute
= c
->undef
;
1906 ntq_activate_execute_for_block(c
);
1910 ntq_emit_if(struct v3d_compile
*c
, nir_if
*nif
)
1912 if (c
->execute
.file
== QFILE_NULL
&&
1913 nir_src_is_dynamically_uniform(nif
->condition
)) {
1914 ntq_emit_uniform_if(c
, nif
);
1916 ntq_emit_nonuniform_if(c
, nif
);
1921 ntq_emit_jump(struct v3d_compile
*c
, nir_jump_instr
*jump
)
1923 switch (jump
->type
) {
1924 case nir_jump_break
:
1925 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1926 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
1927 vir_uniform_ui(c
, c
->loop_break_block
->index
));
1930 case nir_jump_continue
:
1931 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1932 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
1933 vir_uniform_ui(c
, c
->loop_cont_block
->index
));
1936 case nir_jump_return
:
1937 unreachable("All returns shouold be lowered\n");
1942 ntq_emit_instr(struct v3d_compile
*c
, nir_instr
*instr
)
1944 switch (instr
->type
) {
1945 case nir_instr_type_alu
:
1946 ntq_emit_alu(c
, nir_instr_as_alu(instr
));
1949 case nir_instr_type_intrinsic
:
1950 ntq_emit_intrinsic(c
, nir_instr_as_intrinsic(instr
));
1953 case nir_instr_type_load_const
:
1954 ntq_emit_load_const(c
, nir_instr_as_load_const(instr
));
1957 case nir_instr_type_ssa_undef
:
1958 ntq_emit_ssa_undef(c
, nir_instr_as_ssa_undef(instr
));
1961 case nir_instr_type_tex
:
1962 ntq_emit_tex(c
, nir_instr_as_tex(instr
));
1965 case nir_instr_type_jump
:
1966 ntq_emit_jump(c
, nir_instr_as_jump(instr
));
1970 fprintf(stderr
, "Unknown NIR instr type: ");
1971 nir_print_instr(instr
, stderr
);
1972 fprintf(stderr
, "\n");
1978 ntq_emit_block(struct v3d_compile
*c
, nir_block
*block
)
1980 nir_foreach_instr(instr
, block
) {
1981 ntq_emit_instr(c
, instr
);
1985 static void ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
1988 ntq_emit_loop(struct v3d_compile
*c
, nir_loop
*loop
)
1990 bool was_top_level
= false;
1991 if (c
->execute
.file
== QFILE_NULL
) {
1992 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
1993 was_top_level
= true;
1996 struct qblock
*save_loop_cont_block
= c
->loop_cont_block
;
1997 struct qblock
*save_loop_break_block
= c
->loop_break_block
;
1999 c
->loop_cont_block
= vir_new_block(c
);
2000 c
->loop_break_block
= vir_new_block(c
);
2002 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
2003 vir_set_emit_block(c
, c
->loop_cont_block
);
2004 ntq_activate_execute_for_block(c
);
2006 ntq_emit_cf_list(c
, &loop
->body
);
2008 /* Re-enable any previous continues now, so our ANYA check below
2011 * XXX: Use the .ORZ flags update, instead.
2013 vir_PF(c
, vir_XOR(c
,
2015 vir_uniform_ui(c
, c
->loop_cont_block
->index
)),
2017 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
2019 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
2021 struct qinst
*branch
= vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ANYA
);
2022 /* Pixels that were not dispatched or have been discarded should not
2023 * contribute to looping again.
2025 branch
->qpu
.branch
.msfign
= V3D_QPU_MSFIGN_P
;
2026 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
2027 vir_link_blocks(c
->cur_block
, c
->loop_break_block
);
2029 vir_set_emit_block(c
, c
->loop_break_block
);
2031 c
->execute
= c
->undef
;
2033 ntq_activate_execute_for_block(c
);
2035 c
->loop_break_block
= save_loop_break_block
;
2036 c
->loop_cont_block
= save_loop_cont_block
;
2042 ntq_emit_function(struct v3d_compile
*c
, nir_function_impl
*func
)
2044 fprintf(stderr
, "FUNCTIONS not handled.\n");
2049 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
)
2051 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2052 switch (node
->type
) {
2053 case nir_cf_node_block
:
2054 ntq_emit_block(c
, nir_cf_node_as_block(node
));
2057 case nir_cf_node_if
:
2058 ntq_emit_if(c
, nir_cf_node_as_if(node
));
2061 case nir_cf_node_loop
:
2062 ntq_emit_loop(c
, nir_cf_node_as_loop(node
));
2065 case nir_cf_node_function
:
2066 ntq_emit_function(c
, nir_cf_node_as_function(node
));
2070 fprintf(stderr
, "Unknown NIR node type\n");
2077 ntq_emit_impl(struct v3d_compile
*c
, nir_function_impl
*impl
)
2079 ntq_setup_registers(c
, &impl
->registers
);
2080 ntq_emit_cf_list(c
, &impl
->body
);
2084 nir_to_vir(struct v3d_compile
*c
)
2086 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
2087 c
->payload_w
= vir_MOV(c
, vir_reg(QFILE_REG
, 0));
2088 c
->payload_w_centroid
= vir_MOV(c
, vir_reg(QFILE_REG
, 1));
2089 c
->payload_z
= vir_MOV(c
, vir_reg(QFILE_REG
, 2));
2091 /* XXX perf: We could set the "disable implicit point/line
2092 * varyings" field in the shader record and not emit these, if
2093 * they're not going to be used.
2095 if (c
->fs_key
->is_points
) {
2096 c
->point_x
= emit_fragment_varying(c
, NULL
, 0);
2097 c
->point_y
= emit_fragment_varying(c
, NULL
, 0);
2098 } else if (c
->fs_key
->is_lines
) {
2099 c
->line_x
= emit_fragment_varying(c
, NULL
, 0);
2103 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
)
2104 ntq_setup_fs_inputs(c
);
2106 ntq_setup_vpm_inputs(c
);
2108 ntq_setup_outputs(c
);
2109 ntq_setup_uniforms(c
);
2110 ntq_setup_registers(c
, &c
->s
->registers
);
2112 /* Find the main function and emit the body. */
2113 nir_foreach_function(function
, c
->s
) {
2114 assert(strcmp(function
->name
, "main") == 0);
2115 assert(function
->impl
);
2116 ntq_emit_impl(c
, function
->impl
);
2120 const nir_shader_compiler_options v3d_nir_options
= {
2121 .lower_all_io_to_temps
= true,
2122 .lower_extract_byte
= true,
2123 .lower_extract_word
= true,
2125 .lower_bitfield_insert_to_shifts
= true,
2126 .lower_bitfield_extract_to_shifts
= true,
2127 .lower_bitfield_reverse
= true,
2128 .lower_bit_count
= true,
2129 .lower_pack_unorm_2x16
= true,
2130 .lower_pack_snorm_2x16
= true,
2131 .lower_pack_unorm_4x8
= true,
2132 .lower_pack_snorm_4x8
= true,
2133 .lower_unpack_unorm_4x8
= true,
2134 .lower_unpack_snorm_4x8
= true,
2135 .lower_pack_half_2x16
= true,
2136 .lower_unpack_half_2x16
= true,
2138 .lower_find_lsb
= true,
2140 .lower_flrp32
= true,
2143 .lower_fsqrt
= true,
2144 .lower_ifind_msb
= true,
2145 .lower_ldexp
= true,
2146 .lower_mul_high
= true,
2147 .lower_wpos_pntc
= true,
2148 .native_integers
= true,
2152 * When demoting a shader down to single-threaded, removes the THRSW
2153 * instructions (one will still be inserted at v3d_vir_to_qpu() for the
2157 vir_remove_thrsw(struct v3d_compile
*c
)
2159 vir_for_each_block(block
, c
) {
2160 vir_for_each_inst_safe(inst
, block
) {
2161 if (inst
->qpu
.sig
.thrsw
)
2162 vir_remove_instruction(c
, inst
);
2166 c
->last_thrsw
= NULL
;
2170 vir_emit_last_thrsw(struct v3d_compile
*c
)
2172 /* On V3D before 4.1, we need a TMU op to be outstanding when thread
2173 * switching, so disable threads if we didn't do any TMU ops (each of
2174 * which would have emitted a THRSW).
2176 if (!c
->last_thrsw_at_top_level
&& c
->devinfo
->ver
< 41) {
2179 vir_remove_thrsw(c
);
2183 /* If we're threaded and the last THRSW was in conditional code, then
2184 * we need to emit another one so that we can flag it as the last
2187 if (c
->last_thrsw
&& !c
->last_thrsw_at_top_level
) {
2188 assert(c
->devinfo
->ver
>= 41);
2192 /* If we're threaded, then we need to mark the last THRSW instruction
2193 * so we can emit a pair of them at QPU emit time.
2195 * For V3D 4.x, we can spawn the non-fragment shaders already in the
2196 * post-last-THRSW state, so we can skip this.
2198 if (!c
->last_thrsw
&& c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
2199 assert(c
->devinfo
->ver
>= 41);
2204 c
->last_thrsw
->is_last_thrsw
= true;
2207 /* There's a flag in the shader for "center W is needed for reasons other than
2208 * non-centroid varyings", so we just walk the program after VIR optimization
2209 * to see if it's used. It should be harmless to set even if we only use
2210 * center W for varyings.
2213 vir_check_payload_w(struct v3d_compile
*c
)
2215 if (c
->s
->info
.stage
!= MESA_SHADER_FRAGMENT
)
2218 vir_for_each_inst_inorder(inst
, c
) {
2219 for (int i
= 0; i
< vir_get_nsrc(inst
); i
++) {
2220 if (inst
->src
[i
].file
== QFILE_REG
&&
2221 inst
->src
[i
].index
== 0) {
2222 c
->uses_center_w
= true;
2231 v3d_nir_to_vir(struct v3d_compile
*c
)
2233 if (V3D_DEBUG
& (V3D_DEBUG_NIR
|
2234 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2235 fprintf(stderr
, "%s prog %d/%d NIR:\n",
2236 vir_get_stage_name(c
),
2237 c
->program_id
, c
->variant_id
);
2238 nir_print_shader(c
->s
, stderr
);
2243 /* Emit the last THRSW before STVPM and TLB writes. */
2244 vir_emit_last_thrsw(c
);
2246 switch (c
->s
->info
.stage
) {
2247 case MESA_SHADER_FRAGMENT
:
2250 case MESA_SHADER_VERTEX
:
2254 unreachable("bad stage");
2257 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
2258 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2259 fprintf(stderr
, "%s prog %d/%d pre-opt VIR:\n",
2260 vir_get_stage_name(c
),
2261 c
->program_id
, c
->variant_id
);
2263 fprintf(stderr
, "\n");
2267 vir_lower_uniforms(c
);
2269 vir_check_payload_w(c
);
2271 /* XXX perf: On VC4, we do a VIR-level instruction scheduling here.
2272 * We used that on that platform to pipeline TMU writes and reduce the
2273 * number of thread switches, as well as try (mostly successfully) to
2274 * reduce maximum register pressure to allow more threads. We should
2275 * do something of that sort for V3D -- either instruction scheduling
2276 * here, or delay the the THRSW and LDTMUs from our texture
2277 * instructions until the results are needed.
2280 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
2281 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2282 fprintf(stderr
, "%s prog %d/%d VIR:\n",
2283 vir_get_stage_name(c
),
2284 c
->program_id
, c
->variant_id
);
2286 fprintf(stderr
, "\n");
2289 /* Attempt to allocate registers for the temporaries. If we fail,
2290 * reduce thread count and try again.
2292 int min_threads
= (c
->devinfo
->ver
>= 41) ? 2 : 1;
2293 struct qpu_reg
*temp_registers
;
2296 temp_registers
= v3d_register_allocate(c
, &spilled
);
2303 if (c
->threads
== min_threads
) {
2304 fprintf(stderr
, "Failed to register allocate at %d threads:\n",
2313 if (c
->threads
== 1)
2314 vir_remove_thrsw(c
);
2317 v3d_vir_to_qpu(c
, temp_registers
);