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 "v3d_compiler.h"
34 /* We don't do any address packing. */
35 #define __gen_user_data void
36 #define __gen_address_type uint32_t
37 #define __gen_address_offset(reloc) (*reloc)
38 #define __gen_emit_reloc(cl, reloc)
39 #include "cle/v3d_packet_v33_pack.h"
42 ntq_get_src(struct v3d_compile
*c
, nir_src src
, int i
);
44 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
47 resize_qreg_array(struct v3d_compile
*c
,
52 if (*size
>= decl_size
)
55 uint32_t old_size
= *size
;
56 *size
= MAX2(*size
* 2, decl_size
);
57 *regs
= reralloc(c
, *regs
, struct qreg
, *size
);
59 fprintf(stderr
, "Malloc failure\n");
63 for (uint32_t i
= old_size
; i
< *size
; i
++)
64 (*regs
)[i
] = c
->undef
;
68 vir_SFU(struct v3d_compile
*c
, int waddr
, struct qreg src
)
70 vir_FMOV_dest(c
, vir_reg(QFILE_MAGIC
, waddr
), src
);
71 return vir_FMOV(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R4
));
75 vir_LDTMU(struct v3d_compile
*c
)
77 vir_NOP(c
)->qpu
.sig
.ldtmu
= true;
78 return vir_MOV(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R4
));
82 indirect_uniform_load(struct v3d_compile
*c
, nir_intrinsic_instr
*intr
)
84 struct qreg indirect_offset
= ntq_get_src(c
, intr
->src
[0], 0);
85 uint32_t offset
= nir_intrinsic_base(intr
);
86 struct v3d_ubo_range
*range
= NULL
;
89 for (i
= 0; i
< c
->num_ubo_ranges
; i
++) {
90 range
= &c
->ubo_ranges
[i
];
91 if (offset
>= range
->src_offset
&&
92 offset
< range
->src_offset
+ range
->size
) {
96 /* The driver-location-based offset always has to be within a declared
99 assert(i
!= c
->num_ubo_ranges
);
100 if (!c
->ubo_range_used
[i
]) {
101 c
->ubo_range_used
[i
] = true;
102 range
->dst_offset
= c
->next_ubo_dst_offset
;
103 c
->next_ubo_dst_offset
+= range
->size
;
106 offset
-= range
->src_offset
;
108 if (range
->dst_offset
+ offset
!= 0) {
109 indirect_offset
= vir_ADD(c
, indirect_offset
,
110 vir_uniform_ui(c
, range
->dst_offset
+
114 /* Adjust for where we stored the TGSI register base. */
116 vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUA
),
117 vir_uniform(c
, QUNIFORM_UBO_ADDR
, 0),
124 ntq_init_ssa_def(struct v3d_compile
*c
, nir_ssa_def
*def
)
126 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
127 def
->num_components
);
128 _mesa_hash_table_insert(c
->def_ht
, def
, qregs
);
133 * This function is responsible for getting VIR results into the associated
134 * storage for a NIR instruction.
136 * If it's a NIR SSA def, then we just set the associated hash table entry to
139 * If it's a NIR reg, then we need to update the existing qreg assigned to the
140 * NIR destination with the incoming value. To do that without introducing
141 * new MOVs, we require that the incoming qreg either be a uniform, or be
142 * SSA-defined by the previous VIR instruction in the block and rewritable by
143 * this function. That lets us sneak ahead and insert the SF flag beforehand
144 * (knowing that the previous instruction doesn't depend on flags) and rewrite
145 * its destination to be the NIR reg's destination
148 ntq_store_dest(struct v3d_compile
*c
, nir_dest
*dest
, int chan
,
151 struct qinst
*last_inst
= NULL
;
152 if (!list_empty(&c
->cur_block
->instructions
))
153 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
155 assert(result
.file
== QFILE_UNIF
||
156 (result
.file
== QFILE_TEMP
&&
157 last_inst
&& last_inst
== c
->defs
[result
.index
]));
160 assert(chan
< dest
->ssa
.num_components
);
163 struct hash_entry
*entry
=
164 _mesa_hash_table_search(c
->def_ht
, &dest
->ssa
);
169 qregs
= ntq_init_ssa_def(c
, &dest
->ssa
);
171 qregs
[chan
] = result
;
173 nir_register
*reg
= dest
->reg
.reg
;
174 assert(dest
->reg
.base_offset
== 0);
175 assert(reg
->num_array_elems
== 0);
176 struct hash_entry
*entry
=
177 _mesa_hash_table_search(c
->def_ht
, reg
);
178 struct qreg
*qregs
= entry
->data
;
180 /* Insert a MOV if the source wasn't an SSA def in the
181 * previous instruction.
183 if (result
.file
== QFILE_UNIF
) {
184 result
= vir_MOV(c
, result
);
185 last_inst
= c
->defs
[result
.index
];
188 /* We know they're both temps, so just rewrite index. */
189 c
->defs
[last_inst
->dst
.index
] = NULL
;
190 last_inst
->dst
.index
= qregs
[chan
].index
;
192 /* If we're in control flow, then make this update of the reg
193 * conditional on the execution mask.
195 if (c
->execute
.file
!= QFILE_NULL
) {
196 last_inst
->dst
.index
= qregs
[chan
].index
;
198 /* Set the flags to the current exec mask. To insert
199 * the flags push, we temporarily remove our SSA
202 list_del(&last_inst
->link
);
203 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
204 list_addtail(&last_inst
->link
,
205 &c
->cur_block
->instructions
);
207 vir_set_cond(last_inst
, V3D_QPU_COND_IFA
);
208 last_inst
->cond_is_exec_mask
= true;
214 ntq_get_src(struct v3d_compile
*c
, nir_src src
, int i
)
216 struct hash_entry
*entry
;
218 entry
= _mesa_hash_table_search(c
->def_ht
, src
.ssa
);
219 assert(i
< src
.ssa
->num_components
);
221 nir_register
*reg
= src
.reg
.reg
;
222 entry
= _mesa_hash_table_search(c
->def_ht
, reg
);
223 assert(reg
->num_array_elems
== 0);
224 assert(src
.reg
.base_offset
== 0);
225 assert(i
< reg
->num_components
);
228 struct qreg
*qregs
= entry
->data
;
233 ntq_get_alu_src(struct v3d_compile
*c
, nir_alu_instr
*instr
,
236 assert(util_is_power_of_two(instr
->dest
.write_mask
));
237 unsigned chan
= ffs(instr
->dest
.write_mask
) - 1;
238 struct qreg r
= ntq_get_src(c
, instr
->src
[src
].src
,
239 instr
->src
[src
].swizzle
[chan
]);
241 assert(!instr
->src
[src
].abs
);
242 assert(!instr
->src
[src
].negate
);
247 static inline struct qreg
248 vir_SAT(struct v3d_compile
*c
, struct qreg val
)
251 vir_FMIN(c
, val
, vir_uniform_f(c
, 1.0)),
252 vir_uniform_f(c
, 0.0));
256 ntq_umul(struct v3d_compile
*c
, struct qreg src0
, struct qreg src1
)
258 vir_MULTOP(c
, src0
, src1
);
259 return vir_UMUL24(c
, src0
, src1
);
263 ntq_minify(struct v3d_compile
*c
, struct qreg size
, struct qreg level
)
265 return vir_MAX(c
, vir_SHR(c
, size
, level
), vir_uniform_ui(c
, 1));
269 ntq_emit_txs(struct v3d_compile
*c
, nir_tex_instr
*instr
)
271 unsigned unit
= instr
->texture_index
;
272 int lod_index
= nir_tex_instr_src_index(instr
, nir_tex_src_lod
);
273 int dest_size
= nir_tex_instr_dest_size(instr
);
275 struct qreg lod
= c
->undef
;
277 lod
= ntq_get_src(c
, instr
->src
[lod_index
].src
, 0);
279 for (int i
= 0; i
< dest_size
; i
++) {
281 enum quniform_contents contents
;
283 if (instr
->is_array
&& i
== dest_size
- 1)
284 contents
= QUNIFORM_TEXTURE_ARRAY_SIZE
;
286 contents
= QUNIFORM_TEXTURE_WIDTH
+ i
;
288 struct qreg size
= vir_uniform(c
, contents
, unit
);
290 switch (instr
->sampler_dim
) {
291 case GLSL_SAMPLER_DIM_1D
:
292 case GLSL_SAMPLER_DIM_2D
:
293 case GLSL_SAMPLER_DIM_3D
:
294 case GLSL_SAMPLER_DIM_CUBE
:
295 /* Don't minify the array size. */
296 if (!(instr
->is_array
&& i
== dest_size
- 1)) {
297 size
= ntq_minify(c
, size
, lod
);
301 case GLSL_SAMPLER_DIM_RECT
:
302 /* There's no LOD field for rects */
306 unreachable("Bad sampler type");
309 ntq_store_dest(c
, &instr
->dest
, i
, size
);
314 ntq_emit_tex(struct v3d_compile
*c
, nir_tex_instr
*instr
)
316 unsigned unit
= instr
->texture_index
;
318 /* Since each texture sampling op requires uploading uniforms to
319 * reference the texture, there's no HW support for texture size and
320 * you just upload uniforms containing the size.
323 case nir_texop_query_levels
:
324 ntq_store_dest(c
, &instr
->dest
, 0,
325 vir_uniform(c
, QUNIFORM_TEXTURE_LEVELS
, unit
));
328 ntq_emit_txs(c
, instr
);
334 struct V3D33_TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1 p0_unpacked
= {
335 V3D33_TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1_header
,
337 .fetch_sample_mode
= instr
->op
== nir_texop_txf
,
340 switch (instr
->sampler_dim
) {
341 case GLSL_SAMPLER_DIM_1D
:
343 p0_unpacked
.lookup_type
= TEXTURE_1D_ARRAY
;
345 p0_unpacked
.lookup_type
= TEXTURE_1D
;
347 case GLSL_SAMPLER_DIM_2D
:
348 case GLSL_SAMPLER_DIM_RECT
:
350 p0_unpacked
.lookup_type
= TEXTURE_2D_ARRAY
;
352 p0_unpacked
.lookup_type
= TEXTURE_2D
;
354 case GLSL_SAMPLER_DIM_3D
:
355 p0_unpacked
.lookup_type
= TEXTURE_3D
;
357 case GLSL_SAMPLER_DIM_CUBE
:
358 p0_unpacked
.lookup_type
= TEXTURE_CUBE_MAP
;
361 unreachable("Bad sampler type");
364 struct qreg coords
[5];
366 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
367 switch (instr
->src
[i
].src_type
) {
368 case nir_tex_src_coord
:
369 for (int j
= 0; j
< instr
->coord_components
; j
++) {
370 coords
[next_coord
++] =
371 ntq_get_src(c
, instr
->src
[i
].src
, j
);
373 if (instr
->coord_components
< 2)
374 coords
[next_coord
++] = vir_uniform_f(c
, 0.5);
376 case nir_tex_src_bias
:
377 coords
[next_coord
++] =
378 ntq_get_src(c
, instr
->src
[i
].src
, 0);
380 p0_unpacked
.bias_supplied
= true;
382 case nir_tex_src_lod
:
383 /* XXX: Needs base level addition */
384 coords
[next_coord
++] =
385 ntq_get_src(c
, instr
->src
[i
].src
, 0);
387 if (instr
->op
!= nir_texop_txf
&&
388 instr
->op
!= nir_texop_tg4
) {
389 p0_unpacked
.disable_autolod_use_bias_only
= true;
392 case nir_tex_src_comparator
:
393 coords
[next_coord
++] =
394 ntq_get_src(c
, instr
->src
[i
].src
, 0);
396 p0_unpacked
.shadow
= true;
399 case nir_tex_src_offset
: {
400 nir_const_value
*offset
=
401 nir_src_as_const_value(instr
->src
[i
].src
);
402 p0_unpacked
.texel_offset_for_s_coordinate
=
405 if (instr
->coord_components
>= 2)
406 p0_unpacked
.texel_offset_for_t_coordinate
=
409 if (instr
->coord_components
>= 3)
410 p0_unpacked
.texel_offset_for_r_coordinate
=
416 unreachable("unknown texture source");
421 V3D33_TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1_pack(NULL
,
422 (uint8_t *)&p0_packed
,
425 /* There is no native support for GL texture rectangle coordinates, so
426 * we have to rescale from ([0, width], [0, height]) to ([0, 1], [0,
429 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
) {
430 coords
[0] = vir_FMUL(c
, coords
[0],
431 vir_uniform(c
, QUNIFORM_TEXRECT_SCALE_X
,
433 coords
[1] = vir_FMUL(c
, coords
[1],
434 vir_uniform(c
, QUNIFORM_TEXRECT_SCALE_Y
,
438 struct qreg texture_u
[] = {
439 vir_uniform(c
, QUNIFORM_TEXTURE_CONFIG_P0_0
+ unit
, p0_packed
),
440 vir_uniform(c
, QUNIFORM_TEXTURE_CONFIG_P1
, unit
),
442 uint32_t next_texture_u
= 0;
444 for (int i
= 0; i
< next_coord
; i
++) {
447 if (i
== next_coord
- 1)
448 dst
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUL
);
450 dst
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMU
);
452 struct qinst
*tmu
= vir_MOV_dest(c
, dst
, coords
[i
]);
455 tmu
->has_implicit_uniform
= true;
456 tmu
->src
[vir_get_implicit_uniform_src(tmu
)] =
457 texture_u
[next_texture_u
++];
461 bool return_16
= (c
->key
->tex
[unit
].return_size
== 16 ||
464 struct qreg return_values
[4];
465 for (int i
= 0; i
< c
->key
->tex
[unit
].return_channels
; i
++)
466 return_values
[i
] = vir_LDTMU(c
);
467 /* Swizzling .zw of an RG texture should give undefined results, not
468 * crash the compiler.
470 for (int i
= c
->key
->tex
[unit
].return_channels
; i
< 4; i
++)
471 return_values
[i
] = c
->undef
;
473 for (int i
= 0; i
< nir_tex_instr_dest_size(instr
); i
++) {
477 STATIC_ASSERT(PIPE_SWIZZLE_X
== 0);
478 chan
= return_values
[i
/ 2];
480 enum v3d_qpu_input_unpack unpack
;
482 unpack
= V3D_QPU_UNPACK_H
;
484 unpack
= V3D_QPU_UNPACK_L
;
486 chan
= vir_FMOV(c
, chan
);
487 vir_set_unpack(c
->defs
[chan
.index
], 0, unpack
);
489 chan
= vir_MOV(c
, return_values
[i
]);
491 ntq_store_dest(c
, &instr
->dest
, i
, chan
);
496 ntq_fsincos(struct v3d_compile
*c
, struct qreg src
, bool is_cos
)
498 struct qreg input
= vir_FMUL(c
, src
, vir_uniform_f(c
, 1.0f
/ M_PI
));
500 input
= vir_FADD(c
, input
, vir_uniform_f(c
, 0.5));
502 struct qreg periods
= vir_FROUND(c
, input
);
503 struct qreg sin_output
= vir_SFU(c
, V3D_QPU_WADDR_SIN
,
504 vir_FSUB(c
, input
, periods
));
505 return vir_XOR(c
, sin_output
, vir_SHL(c
,
506 vir_FTOIN(c
, periods
),
507 vir_uniform_ui(c
, -1)));
511 ntq_fsign(struct v3d_compile
*c
, struct qreg src
)
513 struct qreg t
= vir_get_temp(c
);
515 vir_MOV_dest(c
, t
, vir_uniform_f(c
, 0.0));
516 vir_PF(c
, vir_FMOV(c
, src
), V3D_QPU_PF_PUSHZ
);
517 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_f(c
, 1.0));
518 vir_PF(c
, vir_FMOV(c
, src
), V3D_QPU_PF_PUSHN
);
519 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_f(c
, -1.0));
520 return vir_MOV(c
, t
);
524 ntq_isign(struct v3d_compile
*c
, struct qreg src
)
526 struct qreg t
= vir_get_temp(c
);
528 vir_MOV_dest(c
, t
, vir_uniform_ui(c
, 0));
529 vir_PF(c
, vir_MOV(c
, src
), V3D_QPU_PF_PUSHZ
);
530 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_ui(c
, 1));
531 vir_PF(c
, vir_MOV(c
, src
), V3D_QPU_PF_PUSHN
);
532 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_ui(c
, -1));
533 return vir_MOV(c
, t
);
537 emit_fragcoord_input(struct v3d_compile
*c
, int attr
)
539 c
->inputs
[attr
* 4 + 0] = vir_FXCD(c
);
540 c
->inputs
[attr
* 4 + 1] = vir_FYCD(c
);
541 c
->inputs
[attr
* 4 + 2] = c
->payload_z
;
542 c
->inputs
[attr
* 4 + 3] = vir_SFU(c
, V3D_QPU_WADDR_RECIP
,
547 emit_fragment_varying(struct v3d_compile
*c
, nir_variable
*var
,
550 struct qreg vary
= vir_reg(QFILE_VARY
, ~0);
551 struct qreg r5
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R5
);
553 /* For gl_PointCoord input or distance along a line, we'll be called
554 * with no nir_variable, and we don't count toward VPM size so we
555 * don't track an input slot.
558 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
561 int i
= c
->num_inputs
++;
562 c
->input_slots
[i
] = v3d_slot_from_slot_and_component(var
->data
.location
,
565 switch (var
->data
.interpolation
) {
566 case INTERP_MODE_NONE
:
567 case INTERP_MODE_SMOOTH
:
568 if (var
->data
.centroid
) {
569 return vir_FADD(c
, vir_FMUL(c
, vary
,
570 c
->payload_w_centroid
), r5
);
572 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
574 case INTERP_MODE_NOPERSPECTIVE
:
575 /* C appears after the mov from the varying.
576 XXX: improve ldvary setup.
578 return vir_FADD(c
, vir_MOV(c
, vary
), r5
);
579 case INTERP_MODE_FLAT
:
580 BITSET_SET(c
->flat_shade_flags
, i
);
581 vir_MOV_dest(c
, c
->undef
, vary
);
582 return vir_MOV(c
, r5
);
584 unreachable("Bad interp mode");
589 emit_fragment_input(struct v3d_compile
*c
, int attr
, nir_variable
*var
)
591 for (int i
= 0; i
< glsl_get_vector_elements(var
->type
); i
++) {
592 c
->inputs
[attr
* 4 + i
] =
593 emit_fragment_varying(c
, var
, i
);
598 add_output(struct v3d_compile
*c
,
599 uint32_t decl_offset
,
603 uint32_t old_array_size
= c
->outputs_array_size
;
604 resize_qreg_array(c
, &c
->outputs
, &c
->outputs_array_size
,
607 if (old_array_size
!= c
->outputs_array_size
) {
608 c
->output_slots
= reralloc(c
,
610 struct v3d_varying_slot
,
611 c
->outputs_array_size
);
614 c
->output_slots
[decl_offset
] =
615 v3d_slot_from_slot_and_component(slot
, swizzle
);
619 declare_uniform_range(struct v3d_compile
*c
, uint32_t start
, uint32_t size
)
621 unsigned array_id
= c
->num_ubo_ranges
++;
622 if (array_id
>= c
->ubo_ranges_array_size
) {
623 c
->ubo_ranges_array_size
= MAX2(c
->ubo_ranges_array_size
* 2,
625 c
->ubo_ranges
= reralloc(c
, c
->ubo_ranges
,
626 struct v3d_ubo_range
,
627 c
->ubo_ranges_array_size
);
628 c
->ubo_range_used
= reralloc(c
, c
->ubo_range_used
,
630 c
->ubo_ranges_array_size
);
633 c
->ubo_ranges
[array_id
].dst_offset
= 0;
634 c
->ubo_ranges
[array_id
].src_offset
= start
;
635 c
->ubo_ranges
[array_id
].size
= size
;
636 c
->ubo_range_used
[array_id
] = false;
640 * If compare_instr is a valid comparison instruction, emits the
641 * compare_instr's comparison and returns the sel_instr's return value based
642 * on the compare_instr's result.
645 ntq_emit_comparison(struct v3d_compile
*c
, struct qreg
*dest
,
646 nir_alu_instr
*compare_instr
,
647 nir_alu_instr
*sel_instr
)
649 struct qreg src0
= ntq_get_alu_src(c
, compare_instr
, 0);
650 struct qreg src1
= ntq_get_alu_src(c
, compare_instr
, 1);
651 bool cond_invert
= false;
653 switch (compare_instr
->op
) {
656 vir_PF(c
, vir_FCMP(c
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
659 vir_PF(c
, vir_XOR(c
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
664 vir_PF(c
, vir_FCMP(c
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
668 vir_PF(c
, vir_XOR(c
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
674 vir_PF(c
, vir_FCMP(c
, src1
, src0
), V3D_QPU_PF_PUSHC
);
677 vir_PF(c
, vir_MIN(c
, src1
, src0
), V3D_QPU_PF_PUSHC
);
681 vir_PF(c
, vir_SUB(c
, src0
, src1
), V3D_QPU_PF_PUSHC
);
687 vir_PF(c
, vir_FCMP(c
, src0
, src1
), V3D_QPU_PF_PUSHN
);
690 vir_PF(c
, vir_MIN(c
, src1
, src0
), V3D_QPU_PF_PUSHC
);
693 vir_PF(c
, vir_SUB(c
, src0
, src1
), V3D_QPU_PF_PUSHC
);
700 enum v3d_qpu_cond cond
= (cond_invert
?
704 switch (sel_instr
->op
) {
709 *dest
= vir_SEL(c
, cond
,
710 vir_uniform_f(c
, 1.0), vir_uniform_f(c
, 0.0));
714 *dest
= vir_SEL(c
, cond
,
715 ntq_get_alu_src(c
, sel_instr
, 1),
716 ntq_get_alu_src(c
, sel_instr
, 2));
720 *dest
= vir_SEL(c
, cond
,
721 vir_uniform_ui(c
, ~0), vir_uniform_ui(c
, 0));
725 /* Make the temporary for nir_store_dest(). */
726 *dest
= vir_MOV(c
, *dest
);
732 * Attempts to fold a comparison generating a boolean result into the
733 * condition code for selecting between two values, instead of comparing the
734 * boolean result against 0 to generate the condition code.
736 static struct qreg
ntq_emit_bcsel(struct v3d_compile
*c
, nir_alu_instr
*instr
,
739 if (!instr
->src
[0].src
.is_ssa
)
741 if (instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
743 nir_alu_instr
*compare
=
744 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
749 if (ntq_emit_comparison(c
, &dest
, compare
, instr
))
753 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
754 return vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
, src
[1], src
[2]));
759 ntq_emit_alu(struct v3d_compile
*c
, nir_alu_instr
*instr
)
761 /* This should always be lowered to ALU operations for V3D. */
762 assert(!instr
->dest
.saturate
);
764 /* Vectors are special in that they have non-scalarized writemasks,
765 * and just take the first swizzle channel for each argument in order
766 * into each writemask channel.
768 if (instr
->op
== nir_op_vec2
||
769 instr
->op
== nir_op_vec3
||
770 instr
->op
== nir_op_vec4
) {
772 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
773 srcs
[i
] = ntq_get_src(c
, instr
->src
[i
].src
,
774 instr
->src
[i
].swizzle
[0]);
775 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
776 ntq_store_dest(c
, &instr
->dest
.dest
, i
,
777 vir_MOV(c
, srcs
[i
]));
781 /* General case: We can just grab the one used channel per src. */
782 struct qreg src
[nir_op_infos
[instr
->op
].num_inputs
];
783 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
784 src
[i
] = ntq_get_alu_src(c
, instr
, i
);
792 result
= vir_MOV(c
, src
[0]);
795 result
= vir_FMUL(c
, src
[0], src
[1]);
798 result
= vir_FADD(c
, src
[0], src
[1]);
801 result
= vir_FSUB(c
, src
[0], src
[1]);
804 result
= vir_FMIN(c
, src
[0], src
[1]);
807 result
= vir_FMAX(c
, src
[0], src
[1]);
811 result
= vir_FTOIZ(c
, src
[0]);
814 result
= vir_FTOUZ(c
, src
[0]);
817 result
= vir_ITOF(c
, src
[0]);
820 result
= vir_UTOF(c
, src
[0]);
823 result
= vir_AND(c
, src
[0], vir_uniform_f(c
, 1.0));
826 result
= vir_AND(c
, src
[0], vir_uniform_ui(c
, 1));
830 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
831 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
832 vir_uniform_ui(c
, ~0),
833 vir_uniform_ui(c
, 0)));
837 result
= vir_ADD(c
, src
[0], src
[1]);
840 result
= vir_SHR(c
, src
[0], src
[1]);
843 result
= vir_SUB(c
, src
[0], src
[1]);
846 result
= vir_ASR(c
, src
[0], src
[1]);
849 result
= vir_SHL(c
, src
[0], src
[1]);
852 result
= vir_MIN(c
, src
[0], src
[1]);
855 result
= vir_UMIN(c
, src
[0], src
[1]);
858 result
= vir_MAX(c
, src
[0], src
[1]);
861 result
= vir_UMAX(c
, src
[0], src
[1]);
864 result
= vir_AND(c
, src
[0], src
[1]);
867 result
= vir_OR(c
, src
[0], src
[1]);
870 result
= vir_XOR(c
, src
[0], src
[1]);
873 result
= vir_NOT(c
, src
[0]);
877 result
= ntq_umul(c
, src
[0], src
[1]);
894 if (!ntq_emit_comparison(c
, &result
, instr
, instr
)) {
895 fprintf(stderr
, "Bad comparison instruction\n");
900 result
= ntq_emit_bcsel(c
, instr
, src
);
903 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
904 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
909 result
= vir_SFU(c
, V3D_QPU_WADDR_RECIP
, src
[0]);
912 result
= vir_SFU(c
, V3D_QPU_WADDR_RSQRT
, src
[0]);
915 result
= vir_SFU(c
, V3D_QPU_WADDR_EXP
, src
[0]);
918 result
= vir_SFU(c
, V3D_QPU_WADDR_LOG
, src
[0]);
922 result
= vir_FCEIL(c
, src
[0]);
925 result
= vir_FFLOOR(c
, src
[0]);
927 case nir_op_fround_even
:
928 result
= vir_FROUND(c
, src
[0]);
931 result
= vir_FTRUNC(c
, src
[0]);
934 result
= vir_FSUB(c
, src
[0], vir_FFLOOR(c
, src
[0]));
938 result
= ntq_fsincos(c
, src
[0], false);
941 result
= ntq_fsincos(c
, src
[0], true);
945 result
= ntq_fsign(c
, src
[0]);
948 result
= ntq_isign(c
, src
[0]);
952 result
= vir_FMOV(c
, src
[0]);
953 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_ABS
);
958 result
= vir_MAX(c
, src
[0],
959 vir_SUB(c
, vir_uniform_ui(c
, 0), src
[0]));
963 case nir_op_fddx_coarse
:
964 case nir_op_fddx_fine
:
965 result
= vir_FDX(c
, src
[0]);
969 case nir_op_fddy_coarse
:
970 case nir_op_fddy_fine
:
971 result
= vir_FDY(c
, src
[0]);
975 fprintf(stderr
, "unknown NIR ALU inst: ");
976 nir_print_instr(&instr
->instr
, stderr
);
977 fprintf(stderr
, "\n");
981 /* We have a scalar result, so the instruction should only have a
982 * single channel written to.
984 assert(util_is_power_of_two(instr
->dest
.write_mask
));
985 ntq_store_dest(c
, &instr
->dest
.dest
,
986 ffs(instr
->dest
.write_mask
) - 1, result
);
989 /* Each TLB read/write setup (a render target or depth buffer) takes an 8-bit
990 * specifier. They come from a register that's preloaded with 0xffffffff
991 * (0xff gets you normal vec4 f16 RT0 writes), and when one is neaded the low
992 * 8 bits are shifted off the bottom and 0xff shifted in from the top.
994 #define TLB_TYPE_F16_COLOR (3 << 6)
995 #define TLB_TYPE_I32_COLOR (1 << 6)
996 #define TLB_TYPE_F32_COLOR (0 << 6)
997 #define TLB_RENDER_TARGET_SHIFT 3 /* Reversed! 7 = RT 0, 0 = RT 7. */
998 #define TLB_SAMPLE_MODE_PER_SAMPLE (0 << 2)
999 #define TLB_SAMPLE_MODE_PER_PIXEL (1 << 2)
1000 #define TLB_F16_SWAP_HI_LO (1 << 1)
1001 #define TLB_VEC_SIZE_4_F16 (1 << 0)
1002 #define TLB_VEC_SIZE_2_F16 (0 << 0)
1003 #define TLB_VEC_SIZE_MINUS_1_SHIFT 0
1005 /* Triggers Z/Stencil testing, used when the shader state's "FS modifies Z"
1008 #define TLB_TYPE_DEPTH ((2 << 6) | (0 << 4))
1009 #define TLB_DEPTH_TYPE_INVARIANT (0 << 2) /* Unmodified sideband input used */
1010 #define TLB_DEPTH_TYPE_PER_PIXEL (1 << 2) /* QPU result used */
1012 /* Stencil is a single 32-bit write. */
1013 #define TLB_TYPE_STENCIL_ALPHA ((2 << 6) | (1 << 4))
1016 emit_frag_end(struct v3d_compile
*c
)
1018 uint32_t discard_cond
= V3D_QPU_COND_NONE
;
1019 if (c
->s
->info
.fs
.uses_discard
) {
1020 vir_PF(c
, vir_MOV(c
, c
->discard
), V3D_QPU_PF_PUSHZ
);
1021 discard_cond
= V3D_QPU_COND_IFA
;
1025 if (c->output_sample_mask_index != -1) {
1026 vir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
1030 if (c
->output_position_index
!= -1) {
1031 struct qinst
*inst
= vir_MOV_dest(c
,
1032 vir_reg(QFILE_TLBU
, 0),
1033 c
->outputs
[c
->output_position_index
]);
1034 vir_set_cond(inst
, discard_cond
);
1036 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1039 TLB_DEPTH_TYPE_PER_PIXEL
|
1041 } else if (c
->s
->info
.fs
.uses_discard
) {
1042 struct qinst
*inst
= vir_MOV_dest(c
,
1043 vir_reg(QFILE_TLBU
, 0),
1044 vir_reg(QFILE_NULL
, 0));
1045 vir_set_cond(inst
, discard_cond
);
1047 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1050 TLB_DEPTH_TYPE_INVARIANT
|
1054 /* XXX: Performance improvement: Merge Z write and color writes TLB
1058 if (c
->output_color_var
) {
1059 nir_variable
*var
= c
->output_color_var
;
1060 struct qreg
*color
= &c
->outputs
[var
->data
.driver_location
* 4];
1061 int num_components
= glsl_get_vector_elements(var
->type
);
1065 assert(num_components
!= 0);
1066 switch (glsl_get_base_type(var
->type
)) {
1067 case GLSL_TYPE_UINT
:
1069 conf
= (TLB_TYPE_I32_COLOR
|
1070 TLB_SAMPLE_MODE_PER_PIXEL
|
1071 ((7 - 0) << TLB_RENDER_TARGET_SHIFT
) |
1072 ((num_components
- 1) <<
1073 TLB_VEC_SIZE_MINUS_1_SHIFT
) |
1077 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLBU
, 0), color
[0]);
1078 vir_set_cond(inst
, discard_cond
);
1079 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1080 vir_uniform_ui(c
, conf
);
1082 for (int i
= 1; i
< num_components
; i
++) {
1083 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0),
1085 vir_set_cond(inst
, discard_cond
);
1090 struct qreg r
= color
[0];
1091 struct qreg g
= color
[1];
1092 struct qreg b
= color
[2];
1093 struct qreg a
= color
[3];
1095 if (c
->fs_key
->swap_color_rb
) {
1100 inst
= vir_VFPACK_dest(c
, vir_reg(QFILE_TLB
, 0), r
, g
);
1101 vir_set_cond(inst
, discard_cond
);
1102 inst
= vir_VFPACK_dest(c
, vir_reg(QFILE_TLB
, 0), b
, a
);
1103 vir_set_cond(inst
, discard_cond
);
1111 emit_scaled_viewport_write(struct v3d_compile
*c
, struct qreg rcp_w
)
1113 for (int i
= 0; i
< 2; i
++) {
1114 struct qreg coord
= c
->outputs
[c
->output_position_index
+ i
];
1115 coord
= vir_FMUL(c
, coord
,
1116 vir_uniform(c
, QUNIFORM_VIEWPORT_X_SCALE
+ i
,
1118 coord
= vir_FMUL(c
, coord
, rcp_w
);
1119 vir_FTOIN_dest(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_VPM
),
1126 emit_zs_write(struct v3d_compile
*c
, struct qreg rcp_w
)
1128 struct qreg zscale
= vir_uniform(c
, QUNIFORM_VIEWPORT_Z_SCALE
, 0);
1129 struct qreg zoffset
= vir_uniform(c
, QUNIFORM_VIEWPORT_Z_OFFSET
, 0);
1131 vir_FADD_dest(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_VPM
),
1132 vir_FMUL(c
, vir_FMUL(c
,
1133 c
->outputs
[c
->output_position_index
+ 2],
1140 emit_rcp_wc_write(struct v3d_compile
*c
, struct qreg rcp_w
)
1142 vir_VPM_WRITE(c
, rcp_w
);
1146 emit_point_size_write(struct v3d_compile
*c
)
1148 struct qreg point_size
;
1150 if (c
->output_point_size_index
!= -1)
1151 point_size
= c
->outputs
[c
->output_point_size_index
];
1153 point_size
= vir_uniform_f(c
, 1.0);
1155 /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
1158 point_size
= vir_FMAX(c
, point_size
, vir_uniform_f(c
, .125));
1160 vir_VPM_WRITE(c
, point_size
);
1164 emit_vpm_write_setup(struct v3d_compile
*c
)
1167 struct V3D33_VPM_GENERIC_BLOCK_WRITE_SETUP unpacked
= {
1168 V3D33_VPM_GENERIC_BLOCK_WRITE_SETUP_header
,
1174 .size
= VPM_SETUP_SIZE_32_BIT
,
1178 V3D33_VPM_GENERIC_BLOCK_WRITE_SETUP_pack(NULL
,
1181 vir_VPMSETUP(c
, vir_uniform_ui(c
, packed
));
1185 emit_vert_end(struct v3d_compile
*c
)
1187 struct qreg rcp_w
= vir_SFU(c
, V3D_QPU_WADDR_RECIP
,
1188 c
->outputs
[c
->output_position_index
+ 3]);
1190 emit_vpm_write_setup(c
);
1192 if (c
->vs_key
->is_coord
) {
1193 for (int i
= 0; i
< 4; i
++)
1194 vir_VPM_WRITE(c
, c
->outputs
[c
->output_position_index
+ i
]);
1195 emit_scaled_viewport_write(c
, rcp_w
);
1196 if (c
->vs_key
->per_vertex_point_size
) {
1197 emit_point_size_write(c
);
1198 /* emit_rcp_wc_write(c, rcp_w); */
1200 /* XXX: Z-only rendering */
1202 emit_zs_write(c
, rcp_w
);
1204 emit_scaled_viewport_write(c
, rcp_w
);
1205 emit_zs_write(c
, rcp_w
);
1206 emit_rcp_wc_write(c
, rcp_w
);
1207 if (c
->vs_key
->per_vertex_point_size
)
1208 emit_point_size_write(c
);
1211 for (int i
= 0; i
< c
->vs_key
->num_fs_inputs
; i
++) {
1212 struct v3d_varying_slot input
= c
->vs_key
->fs_inputs
[i
];
1215 for (j
= 0; j
< c
->num_outputs
; j
++) {
1216 struct v3d_varying_slot output
= c
->output_slots
[j
];
1218 if (!memcmp(&input
, &output
, sizeof(input
))) {
1219 vir_VPM_WRITE(c
, c
->outputs
[j
]);
1223 /* Emit padding if we didn't find a declared VS output for
1226 if (j
== c
->num_outputs
)
1227 vir_VPM_WRITE(c
, vir_uniform_f(c
, 0.0));
1232 v3d_optimize_nir(struct nir_shader
*s
)
1239 NIR_PASS_V(s
, nir_lower_vars_to_ssa
);
1240 NIR_PASS(progress
, s
, nir_lower_alu_to_scalar
);
1241 NIR_PASS(progress
, s
, nir_lower_phis_to_scalar
);
1242 NIR_PASS(progress
, s
, nir_copy_prop
);
1243 NIR_PASS(progress
, s
, nir_opt_remove_phis
);
1244 NIR_PASS(progress
, s
, nir_opt_dce
);
1245 NIR_PASS(progress
, s
, nir_opt_dead_cf
);
1246 NIR_PASS(progress
, s
, nir_opt_cse
);
1247 NIR_PASS(progress
, s
, nir_opt_peephole_select
, 8);
1248 NIR_PASS(progress
, s
, nir_opt_algebraic
);
1249 NIR_PASS(progress
, s
, nir_opt_constant_folding
);
1250 NIR_PASS(progress
, s
, nir_opt_undef
);
1255 driver_location_compare(const void *in_a
, const void *in_b
)
1257 const nir_variable
*const *a
= in_a
;
1258 const nir_variable
*const *b
= in_b
;
1260 return (*a
)->data
.driver_location
- (*b
)->data
.driver_location
;
1264 ntq_emit_vpm_read(struct v3d_compile
*c
,
1265 uint32_t *num_components_queued
,
1266 uint32_t *remaining
,
1269 struct qreg vpm
= vir_reg(QFILE_VPM
, vpm_index
);
1271 if (*num_components_queued
!= 0) {
1272 (*num_components_queued
)--;
1274 return vir_MOV(c
, vpm
);
1277 uint32_t num_components
= MIN2(*remaining
, 32);
1279 struct V3D33_VPM_GENERIC_BLOCK_READ_SETUP unpacked
= {
1280 V3D33_VPM_GENERIC_BLOCK_READ_SETUP_header
,
1284 /* If the field is 0, that means a read count of 32. */
1285 .num
= num_components
& 31,
1288 .size
= VPM_SETUP_SIZE_32_BIT
,
1289 .addr
= c
->num_inputs
,
1293 V3D33_VPM_GENERIC_BLOCK_READ_SETUP_pack(NULL
,
1296 vir_VPMSETUP(c
, vir_uniform_ui(c
, packed
));
1298 *num_components_queued
= num_components
- 1;
1299 *remaining
-= num_components
;
1302 return vir_MOV(c
, vpm
);
1306 ntq_setup_inputs(struct v3d_compile
*c
)
1308 unsigned num_entries
= 0;
1309 unsigned num_components
= 0;
1310 nir_foreach_variable(var
, &c
->s
->inputs
) {
1312 num_components
+= glsl_get_components(var
->type
);
1315 nir_variable
*vars
[num_entries
];
1318 nir_foreach_variable(var
, &c
->s
->inputs
)
1321 /* Sort the variables so that we emit the input setup in
1322 * driver_location order. This is required for VPM reads, whose data
1323 * is fetched into the VPM in driver_location (TGSI register index)
1326 qsort(&vars
, num_entries
, sizeof(*vars
), driver_location_compare
);
1328 uint32_t vpm_components_queued
= 0;
1329 if (c
->s
->stage
== MESA_SHADER_VERTEX
) {
1330 bool uses_iid
= c
->s
->info
.system_values_read
&
1331 (1ull << SYSTEM_VALUE_INSTANCE_ID
);
1332 bool uses_vid
= c
->s
->info
.system_values_read
&
1333 (1ull << SYSTEM_VALUE_VERTEX_ID
);
1335 num_components
+= uses_iid
;
1336 num_components
+= uses_vid
;
1339 c
->iid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1340 &num_components
, ~0);
1344 c
->vid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1345 &num_components
, ~0);
1349 for (unsigned i
= 0; i
< num_entries
; i
++) {
1350 nir_variable
*var
= vars
[i
];
1351 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1352 unsigned loc
= var
->data
.driver_location
;
1354 assert(array_len
== 1);
1356 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1359 if (c
->s
->stage
== MESA_SHADER_FRAGMENT
) {
1360 if (var
->data
.location
== VARYING_SLOT_POS
) {
1361 emit_fragcoord_input(c
, loc
);
1362 } else if (var
->data
.location
== VARYING_SLOT_PNTC
||
1363 (var
->data
.location
>= VARYING_SLOT_VAR0
&&
1364 (c
->fs_key
->point_sprite_mask
&
1365 (1 << (var
->data
.location
-
1366 VARYING_SLOT_VAR0
))))) {
1367 c
->inputs
[loc
* 4 + 0] = c
->point_x
;
1368 c
->inputs
[loc
* 4 + 1] = c
->point_y
;
1370 emit_fragment_input(c
, loc
, var
);
1373 int var_components
= glsl_get_components(var
->type
);
1375 for (int i
= 0; i
< var_components
; i
++) {
1376 c
->inputs
[loc
* 4 + i
] =
1377 ntq_emit_vpm_read(c
,
1378 &vpm_components_queued
,
1383 c
->vattr_sizes
[loc
] = var_components
;
1387 if (c
->s
->stage
== MESA_SHADER_VERTEX
) {
1388 assert(vpm_components_queued
== 0);
1389 assert(num_components
== 0);
1394 ntq_setup_outputs(struct v3d_compile
*c
)
1396 nir_foreach_variable(var
, &c
->s
->outputs
) {
1397 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1398 unsigned loc
= var
->data
.driver_location
* 4;
1400 assert(array_len
== 1);
1403 for (int i
= 0; i
< 4; i
++)
1404 add_output(c
, loc
+ i
, var
->data
.location
, i
);
1406 if (c
->s
->stage
== MESA_SHADER_FRAGMENT
) {
1407 switch (var
->data
.location
) {
1408 case FRAG_RESULT_COLOR
:
1409 case FRAG_RESULT_DATA0
:
1410 c
->output_color_var
= var
;
1412 case FRAG_RESULT_DEPTH
:
1413 c
->output_position_index
= loc
;
1415 case FRAG_RESULT_SAMPLE_MASK
:
1416 c
->output_sample_mask_index
= loc
;
1420 switch (var
->data
.location
) {
1421 case VARYING_SLOT_POS
:
1422 c
->output_position_index
= loc
;
1424 case VARYING_SLOT_PSIZ
:
1425 c
->output_point_size_index
= loc
;
1433 ntq_setup_uniforms(struct v3d_compile
*c
)
1435 nir_foreach_variable(var
, &c
->s
->uniforms
) {
1436 uint32_t vec4_count
= glsl_count_attribute_slots(var
->type
,
1438 unsigned vec4_size
= 4 * sizeof(float);
1440 declare_uniform_range(c
, var
->data
.driver_location
* vec4_size
,
1441 vec4_count
* vec4_size
);
1447 * Sets up the mapping from nir_register to struct qreg *.
1449 * Each nir_register gets a struct qreg per 32-bit component being stored.
1452 ntq_setup_registers(struct v3d_compile
*c
, struct exec_list
*list
)
1454 foreach_list_typed(nir_register
, nir_reg
, node
, list
) {
1455 unsigned array_len
= MAX2(nir_reg
->num_array_elems
, 1);
1456 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
1458 nir_reg
->num_components
);
1460 _mesa_hash_table_insert(c
->def_ht
, nir_reg
, qregs
);
1462 for (int i
= 0; i
< array_len
* nir_reg
->num_components
; i
++)
1463 qregs
[i
] = vir_get_temp(c
);
1468 ntq_emit_load_const(struct v3d_compile
*c
, nir_load_const_instr
*instr
)
1470 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1471 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1472 qregs
[i
] = vir_uniform_ui(c
, instr
->value
.u32
[i
]);
1474 _mesa_hash_table_insert(c
->def_ht
, &instr
->def
, qregs
);
1478 ntq_emit_ssa_undef(struct v3d_compile
*c
, nir_ssa_undef_instr
*instr
)
1480 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1482 /* VIR needs there to be *some* value, so pick 0 (same as for
1483 * ntq_setup_registers().
1485 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1486 qregs
[i
] = vir_uniform_ui(c
, 0);
1490 ntq_emit_intrinsic(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1492 nir_const_value
*const_offset
;
1495 switch (instr
->intrinsic
) {
1496 case nir_intrinsic_load_uniform
:
1497 assert(instr
->num_components
== 1);
1498 const_offset
= nir_src_as_const_value(instr
->src
[0]);
1500 offset
= nir_intrinsic_base(instr
) + const_offset
->u32
[0];
1501 assert(offset
% 4 == 0);
1502 /* We need dwords */
1503 offset
= offset
/ 4;
1504 ntq_store_dest(c
, &instr
->dest
, 0,
1505 vir_uniform(c
, QUNIFORM_UNIFORM
,
1508 ntq_store_dest(c
, &instr
->dest
, 0,
1509 indirect_uniform_load(c
, instr
));
1513 case nir_intrinsic_load_ubo
:
1514 for (int i
= 0; i
< instr
->num_components
; i
++) {
1515 int ubo
= nir_src_as_const_value(instr
->src
[0])->u32
[0];
1517 /* Adjust for where we stored the TGSI register base. */
1519 vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUA
),
1520 vir_uniform(c
, QUNIFORM_UBO_ADDR
, 1 + ubo
),
1522 ntq_get_src(c
, instr
->src
[1], 0),
1523 vir_uniform_ui(c
, i
* 4)));
1525 ntq_store_dest(c
, &instr
->dest
, i
, vir_LDTMU(c
));
1529 const_offset
= nir_src_as_const_value(instr
->src
[0]);
1531 offset
= nir_intrinsic_base(instr
) + const_offset
->u32
[0];
1532 assert(offset
% 4 == 0);
1533 /* We need dwords */
1534 offset
= offset
/ 4;
1535 ntq_store_dest(c
, &instr
->dest
, 0,
1536 vir_uniform(c
, QUNIFORM_UNIFORM
,
1539 ntq_store_dest(c
, &instr
->dest
, 0,
1540 indirect_uniform_load(c
, instr
));
1544 case nir_intrinsic_load_user_clip_plane
:
1545 for (int i
= 0; i
< instr
->num_components
; i
++) {
1546 ntq_store_dest(c
, &instr
->dest
, i
,
1547 vir_uniform(c
, QUNIFORM_USER_CLIP_PLANE
,
1548 nir_intrinsic_ucp_id(instr
) *
1553 case nir_intrinsic_load_alpha_ref_float
:
1554 ntq_store_dest(c
, &instr
->dest
, 0,
1555 vir_uniform(c
, QUNIFORM_ALPHA_REF
, 0));
1558 case nir_intrinsic_load_sample_mask_in
:
1559 ntq_store_dest(c
, &instr
->dest
, 0,
1560 vir_uniform(c
, QUNIFORM_SAMPLE_MASK
, 0));
1563 case nir_intrinsic_load_front_face
:
1564 /* The register contains 0 (front) or 1 (back), and we need to
1565 * turn it into a NIR bool where true means front.
1567 ntq_store_dest(c
, &instr
->dest
, 0,
1569 vir_uniform_ui(c
, -1),
1573 case nir_intrinsic_load_instance_id
:
1574 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->iid
));
1577 case nir_intrinsic_load_vertex_id
:
1578 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->vid
));
1581 case nir_intrinsic_load_input
:
1582 const_offset
= nir_src_as_const_value(instr
->src
[0]);
1583 assert(const_offset
&& "v3d doesn't support indirect inputs");
1584 for (int i
= 0; i
< instr
->num_components
; i
++) {
1585 offset
= nir_intrinsic_base(instr
) + const_offset
->u32
[0];
1586 int comp
= nir_intrinsic_component(instr
) + i
;
1587 ntq_store_dest(c
, &instr
->dest
, i
,
1588 vir_MOV(c
, c
->inputs
[offset
* 4 + comp
]));
1592 case nir_intrinsic_store_output
:
1593 const_offset
= nir_src_as_const_value(instr
->src
[1]);
1594 assert(const_offset
&& "v3d doesn't support indirect outputs");
1595 offset
= ((nir_intrinsic_base(instr
) +
1596 const_offset
->u32
[0]) * 4 +
1597 nir_intrinsic_component(instr
));
1599 for (int i
= 0; i
< instr
->num_components
; i
++) {
1600 c
->outputs
[offset
+ i
] =
1601 vir_MOV(c
, ntq_get_src(c
, instr
->src
[0], i
));
1603 c
->num_outputs
= MAX2(c
->num_outputs
,
1604 offset
+ instr
->num_components
);
1607 case nir_intrinsic_discard
:
1608 if (c
->execute
.file
!= QFILE_NULL
) {
1609 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1610 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->discard
,
1611 vir_uniform_ui(c
, ~0));
1613 vir_MOV_dest(c
, c
->discard
, vir_uniform_ui(c
, ~0));
1617 case nir_intrinsic_discard_if
: {
1618 /* true (~0) if we're discarding */
1619 struct qreg cond
= ntq_get_src(c
, instr
->src
[0], 0);
1621 if (c
->execute
.file
!= QFILE_NULL
) {
1622 /* execute == 0 means the channel is active. Invert
1623 * the condition so that we can use zero as "executing
1626 vir_PF(c
, vir_AND(c
, c
->execute
, vir_NOT(c
, cond
)),
1628 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->discard
, cond
);
1630 vir_OR_dest(c
, c
->discard
, c
->discard
, cond
);
1637 fprintf(stderr
, "Unknown intrinsic: ");
1638 nir_print_instr(&instr
->instr
, stderr
);
1639 fprintf(stderr
, "\n");
1644 /* Clears (activates) the execute flags for any channels whose jump target
1645 * matches this block.
1648 ntq_activate_execute_for_block(struct v3d_compile
*c
)
1650 vir_PF(c
, vir_SUB(c
, c
->execute
, vir_uniform_ui(c
, c
->cur_block
->index
)),
1653 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
1657 ntq_emit_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
1659 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
1660 bool empty_else_block
=
1661 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
1662 exec_list_is_empty(&nir_else_block
->instr_list
));
1664 struct qblock
*then_block
= vir_new_block(c
);
1665 struct qblock
*after_block
= vir_new_block(c
);
1666 struct qblock
*else_block
;
1667 if (empty_else_block
)
1668 else_block
= after_block
;
1670 else_block
= vir_new_block(c
);
1672 bool was_top_level
= false;
1673 if (c
->execute
.file
== QFILE_NULL
) {
1674 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
1675 was_top_level
= true;
1678 /* Set A for executing (execute == 0) and jumping (if->condition ==
1679 * 0) channels, and then update execute flags for those to point to
1684 ntq_get_src(c
, if_stmt
->condition
, 0)),
1686 vir_MOV_cond(c
, V3D_QPU_COND_IFA
,
1688 vir_uniform_ui(c
, else_block
->index
));
1690 /* Jump to ELSE if nothing is active for THEN, otherwise fall
1693 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1694 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLNA
);
1695 vir_link_blocks(c
->cur_block
, else_block
);
1696 vir_link_blocks(c
->cur_block
, then_block
);
1698 /* Process the THEN block. */
1699 vir_set_emit_block(c
, then_block
);
1700 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
1702 if (!empty_else_block
) {
1703 /* Handle the end of the THEN block. First, all currently
1704 * active channels update their execute flags to point to
1707 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1708 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
1709 vir_uniform_ui(c
, after_block
->index
));
1711 /* If everything points at ENDIF, then jump there immediately. */
1712 vir_PF(c
, vir_SUB(c
, c
->execute
,
1713 vir_uniform_ui(c
, after_block
->index
)),
1715 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLA
);
1716 vir_link_blocks(c
->cur_block
, after_block
);
1717 vir_link_blocks(c
->cur_block
, else_block
);
1719 vir_set_emit_block(c
, else_block
);
1720 ntq_activate_execute_for_block(c
);
1721 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
1724 vir_link_blocks(c
->cur_block
, after_block
);
1726 vir_set_emit_block(c
, after_block
);
1728 c
->execute
= c
->undef
;
1730 ntq_activate_execute_for_block(c
);
1734 ntq_emit_jump(struct v3d_compile
*c
, nir_jump_instr
*jump
)
1736 switch (jump
->type
) {
1737 case nir_jump_break
:
1738 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1739 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
1740 vir_uniform_ui(c
, c
->loop_break_block
->index
));
1743 case nir_jump_continue
:
1744 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1745 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
1746 vir_uniform_ui(c
, c
->loop_cont_block
->index
));
1749 case nir_jump_return
:
1750 unreachable("All returns shouold be lowered\n");
1755 ntq_emit_instr(struct v3d_compile
*c
, nir_instr
*instr
)
1757 switch (instr
->type
) {
1758 case nir_instr_type_alu
:
1759 ntq_emit_alu(c
, nir_instr_as_alu(instr
));
1762 case nir_instr_type_intrinsic
:
1763 ntq_emit_intrinsic(c
, nir_instr_as_intrinsic(instr
));
1766 case nir_instr_type_load_const
:
1767 ntq_emit_load_const(c
, nir_instr_as_load_const(instr
));
1770 case nir_instr_type_ssa_undef
:
1771 ntq_emit_ssa_undef(c
, nir_instr_as_ssa_undef(instr
));
1774 case nir_instr_type_tex
:
1775 ntq_emit_tex(c
, nir_instr_as_tex(instr
));
1778 case nir_instr_type_jump
:
1779 ntq_emit_jump(c
, nir_instr_as_jump(instr
));
1783 fprintf(stderr
, "Unknown NIR instr type: ");
1784 nir_print_instr(instr
, stderr
);
1785 fprintf(stderr
, "\n");
1791 ntq_emit_block(struct v3d_compile
*c
, nir_block
*block
)
1793 nir_foreach_instr(instr
, block
) {
1794 ntq_emit_instr(c
, instr
);
1798 static void ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
1801 ntq_emit_loop(struct v3d_compile
*c
, nir_loop
*loop
)
1803 bool was_top_level
= false;
1804 if (c
->execute
.file
== QFILE_NULL
) {
1805 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
1806 was_top_level
= true;
1809 struct qblock
*save_loop_cont_block
= c
->loop_cont_block
;
1810 struct qblock
*save_loop_break_block
= c
->loop_break_block
;
1812 c
->loop_cont_block
= vir_new_block(c
);
1813 c
->loop_break_block
= vir_new_block(c
);
1815 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
1816 vir_set_emit_block(c
, c
->loop_cont_block
);
1817 ntq_activate_execute_for_block(c
);
1819 ntq_emit_cf_list(c
, &loop
->body
);
1821 /* Re-enable any previous continues now, so our ANYA check below
1824 * XXX: Use the .ORZ flags update, instead.
1826 vir_PF(c
, vir_SUB(c
,
1828 vir_uniform_ui(c
, c
->loop_cont_block
->index
)),
1830 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
1832 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1834 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ANYA
);
1835 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
1836 vir_link_blocks(c
->cur_block
, c
->loop_break_block
);
1838 vir_set_emit_block(c
, c
->loop_break_block
);
1840 c
->execute
= c
->undef
;
1842 ntq_activate_execute_for_block(c
);
1844 c
->loop_break_block
= save_loop_break_block
;
1845 c
->loop_cont_block
= save_loop_cont_block
;
1849 ntq_emit_function(struct v3d_compile
*c
, nir_function_impl
*func
)
1851 fprintf(stderr
, "FUNCTIONS not handled.\n");
1856 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
)
1858 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
1859 switch (node
->type
) {
1860 case nir_cf_node_block
:
1861 ntq_emit_block(c
, nir_cf_node_as_block(node
));
1864 case nir_cf_node_if
:
1865 ntq_emit_if(c
, nir_cf_node_as_if(node
));
1868 case nir_cf_node_loop
:
1869 ntq_emit_loop(c
, nir_cf_node_as_loop(node
));
1872 case nir_cf_node_function
:
1873 ntq_emit_function(c
, nir_cf_node_as_function(node
));
1877 fprintf(stderr
, "Unknown NIR node type\n");
1884 ntq_emit_impl(struct v3d_compile
*c
, nir_function_impl
*impl
)
1886 ntq_setup_registers(c
, &impl
->registers
);
1887 ntq_emit_cf_list(c
, &impl
->body
);
1891 nir_to_vir(struct v3d_compile
*c
)
1893 if (c
->s
->stage
== MESA_SHADER_FRAGMENT
) {
1894 c
->payload_w
= vir_MOV(c
, vir_reg(QFILE_REG
, 0));
1895 c
->payload_w_centroid
= vir_MOV(c
, vir_reg(QFILE_REG
, 1));
1896 c
->payload_z
= vir_MOV(c
, vir_reg(QFILE_REG
, 2));
1898 if (c
->s
->info
.fs
.uses_discard
)
1899 c
->discard
= vir_MOV(c
, vir_uniform_ui(c
, 0));
1901 if (c
->fs_key
->is_points
) {
1902 c
->point_x
= emit_fragment_varying(c
, NULL
, 0);
1903 c
->point_y
= emit_fragment_varying(c
, NULL
, 0);
1904 } else if (c
->fs_key
->is_lines
) {
1905 c
->line_x
= emit_fragment_varying(c
, NULL
, 0);
1909 ntq_setup_inputs(c
);
1910 ntq_setup_outputs(c
);
1911 ntq_setup_uniforms(c
);
1912 ntq_setup_registers(c
, &c
->s
->registers
);
1914 /* Find the main function and emit the body. */
1915 nir_foreach_function(function
, c
->s
) {
1916 assert(strcmp(function
->name
, "main") == 0);
1917 assert(function
->impl
);
1918 ntq_emit_impl(c
, function
->impl
);
1922 const nir_shader_compiler_options v3d_nir_options
= {
1923 .lower_extract_byte
= true,
1924 .lower_extract_word
= true,
1925 .lower_bitfield_insert
= true,
1926 .lower_bitfield_extract
= true,
1928 .lower_flrp32
= true,
1931 .lower_fsqrt
= true,
1932 .lower_negate
= true,
1933 .native_integers
= true,
1939 count_nir_instrs(nir_shader
*nir
)
1942 nir_foreach_function(function
, nir
) {
1943 if (!function
->impl
)
1945 nir_foreach_block(block
, function
->impl
) {
1946 nir_foreach_instr(instr
, block
)
1955 v3d_nir_to_vir(struct v3d_compile
*c
)
1957 if (V3D_DEBUG
& (V3D_DEBUG_NIR
|
1958 v3d_debug_flag_for_shader_stage(c
->s
->stage
))) {
1959 fprintf(stderr
, "%s prog %d/%d NIR:\n",
1960 vir_get_stage_name(c
),
1961 c
->program_id
, c
->variant_id
);
1962 nir_print_shader(c
->s
, stderr
);
1967 switch (c
->s
->stage
) {
1968 case MESA_SHADER_FRAGMENT
:
1971 case MESA_SHADER_VERTEX
:
1975 unreachable("bad stage");
1978 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
1979 v3d_debug_flag_for_shader_stage(c
->s
->stage
))) {
1980 fprintf(stderr
, "%s prog %d/%d pre-opt VIR:\n",
1981 vir_get_stage_name(c
),
1982 c
->program_id
, c
->variant_id
);
1984 fprintf(stderr
, "\n");
1988 vir_lower_uniforms(c
);
1990 /* XXX: vir_schedule_instructions(c); */
1992 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
1993 v3d_debug_flag_for_shader_stage(c
->s
->stage
))) {
1994 fprintf(stderr
, "%s prog %d/%d VIR:\n",
1995 vir_get_stage_name(c
),
1996 c
->program_id
, c
->variant_id
);
1998 fprintf(stderr
, "\n");