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"
36 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
39 resize_qreg_array(struct v3d_compile
*c
,
44 if (*size
>= decl_size
)
47 uint32_t old_size
= *size
;
48 *size
= MAX2(*size
* 2, decl_size
);
49 *regs
= reralloc(c
, *regs
, struct qreg
, *size
);
51 fprintf(stderr
, "Malloc failure\n");
55 for (uint32_t i
= old_size
; i
< *size
; i
++)
56 (*regs
)[i
] = c
->undef
;
60 vir_emit_thrsw(struct v3d_compile
*c
)
65 /* Always thread switch after each texture operation for now.
67 * We could do better by batching a bunch of texture fetches up and
68 * then doing one thread switch and collecting all their results
71 c
->last_thrsw
= vir_NOP(c
);
72 c
->last_thrsw
->qpu
.sig
.thrsw
= true;
73 c
->last_thrsw_at_top_level
= (c
->execute
.file
== QFILE_NULL
);
77 vir_SFU(struct v3d_compile
*c
, int waddr
, struct qreg src
)
79 vir_FMOV_dest(c
, vir_reg(QFILE_MAGIC
, waddr
), src
);
80 return vir_FMOV(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R4
));
84 indirect_uniform_load(struct v3d_compile
*c
, nir_intrinsic_instr
*intr
)
86 struct qreg indirect_offset
= ntq_get_src(c
, intr
->src
[0], 0);
87 uint32_t offset
= nir_intrinsic_base(intr
);
88 struct v3d_ubo_range
*range
= NULL
;
91 for (i
= 0; i
< c
->num_ubo_ranges
; i
++) {
92 range
= &c
->ubo_ranges
[i
];
93 if (offset
>= range
->src_offset
&&
94 offset
< range
->src_offset
+ range
->size
) {
98 /* The driver-location-based offset always has to be within a declared
101 assert(i
!= c
->num_ubo_ranges
);
102 if (!c
->ubo_range_used
[i
]) {
103 c
->ubo_range_used
[i
] = true;
104 range
->dst_offset
= c
->next_ubo_dst_offset
;
105 c
->next_ubo_dst_offset
+= range
->size
;
108 offset
-= range
->src_offset
;
110 if (range
->dst_offset
+ offset
!= 0) {
111 indirect_offset
= vir_ADD(c
, indirect_offset
,
112 vir_uniform_ui(c
, range
->dst_offset
+
116 /* Adjust for where we stored the TGSI register base. */
118 vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUA
),
119 vir_uniform(c
, QUNIFORM_UBO_ADDR
, 0),
127 ntq_init_ssa_def(struct v3d_compile
*c
, nir_ssa_def
*def
)
129 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
130 def
->num_components
);
131 _mesa_hash_table_insert(c
->def_ht
, def
, qregs
);
136 * This function is responsible for getting VIR results into the associated
137 * storage for a NIR instruction.
139 * If it's a NIR SSA def, then we just set the associated hash table entry to
142 * If it's a NIR reg, then we need to update the existing qreg assigned to the
143 * NIR destination with the incoming value. To do that without introducing
144 * new MOVs, we require that the incoming qreg either be a uniform, or be
145 * SSA-defined by the previous VIR instruction in the block and rewritable by
146 * this function. That lets us sneak ahead and insert the SF flag beforehand
147 * (knowing that the previous instruction doesn't depend on flags) and rewrite
148 * its destination to be the NIR reg's destination
151 ntq_store_dest(struct v3d_compile
*c
, nir_dest
*dest
, int chan
,
154 struct qinst
*last_inst
= NULL
;
155 if (!list_empty(&c
->cur_block
->instructions
))
156 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
158 assert(result
.file
== QFILE_UNIF
||
159 (result
.file
== QFILE_TEMP
&&
160 last_inst
&& last_inst
== c
->defs
[result
.index
]));
163 assert(chan
< dest
->ssa
.num_components
);
166 struct hash_entry
*entry
=
167 _mesa_hash_table_search(c
->def_ht
, &dest
->ssa
);
172 qregs
= ntq_init_ssa_def(c
, &dest
->ssa
);
174 qregs
[chan
] = result
;
176 nir_register
*reg
= dest
->reg
.reg
;
177 assert(dest
->reg
.base_offset
== 0);
178 assert(reg
->num_array_elems
== 0);
179 struct hash_entry
*entry
=
180 _mesa_hash_table_search(c
->def_ht
, reg
);
181 struct qreg
*qregs
= entry
->data
;
183 /* Insert a MOV if the source wasn't an SSA def in the
184 * previous instruction.
186 if (result
.file
== QFILE_UNIF
) {
187 result
= vir_MOV(c
, result
);
188 last_inst
= c
->defs
[result
.index
];
191 /* We know they're both temps, so just rewrite index. */
192 c
->defs
[last_inst
->dst
.index
] = NULL
;
193 last_inst
->dst
.index
= qregs
[chan
].index
;
195 /* If we're in control flow, then make this update of the reg
196 * conditional on the execution mask.
198 if (c
->execute
.file
!= QFILE_NULL
) {
199 last_inst
->dst
.index
= qregs
[chan
].index
;
201 /* Set the flags to the current exec mask.
203 c
->cursor
= vir_before_inst(last_inst
);
204 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
205 c
->cursor
= vir_after_inst(last_inst
);
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_or_zero(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_minify(struct v3d_compile
*c
, struct qreg size
, struct qreg level
)
258 return vir_MAX(c
, vir_SHR(c
, size
, level
), vir_uniform_ui(c
, 1));
262 ntq_emit_txs(struct v3d_compile
*c
, nir_tex_instr
*instr
)
264 unsigned unit
= instr
->texture_index
;
265 int lod_index
= nir_tex_instr_src_index(instr
, nir_tex_src_lod
);
266 int dest_size
= nir_tex_instr_dest_size(instr
);
268 struct qreg lod
= c
->undef
;
270 lod
= ntq_get_src(c
, instr
->src
[lod_index
].src
, 0);
272 for (int i
= 0; i
< dest_size
; i
++) {
274 enum quniform_contents contents
;
276 if (instr
->is_array
&& i
== dest_size
- 1)
277 contents
= QUNIFORM_TEXTURE_ARRAY_SIZE
;
279 contents
= QUNIFORM_TEXTURE_WIDTH
+ i
;
281 struct qreg size
= vir_uniform(c
, contents
, unit
);
283 switch (instr
->sampler_dim
) {
284 case GLSL_SAMPLER_DIM_1D
:
285 case GLSL_SAMPLER_DIM_2D
:
286 case GLSL_SAMPLER_DIM_3D
:
287 case GLSL_SAMPLER_DIM_CUBE
:
288 /* Don't minify the array size. */
289 if (!(instr
->is_array
&& i
== dest_size
- 1)) {
290 size
= ntq_minify(c
, size
, lod
);
294 case GLSL_SAMPLER_DIM_RECT
:
295 /* There's no LOD field for rects */
299 unreachable("Bad sampler type");
302 ntq_store_dest(c
, &instr
->dest
, i
, size
);
307 ntq_emit_tex(struct v3d_compile
*c
, nir_tex_instr
*instr
)
309 unsigned unit
= instr
->texture_index
;
311 /* Since each texture sampling op requires uploading uniforms to
312 * reference the texture, there's no HW support for texture size and
313 * you just upload uniforms containing the size.
316 case nir_texop_query_levels
:
317 ntq_store_dest(c
, &instr
->dest
, 0,
318 vir_uniform(c
, QUNIFORM_TEXTURE_LEVELS
, unit
));
321 ntq_emit_txs(c
, instr
);
327 if (c
->devinfo
->ver
>= 40)
328 v3d40_vir_emit_tex(c
, instr
);
330 v3d33_vir_emit_tex(c
, instr
);
334 ntq_fsincos(struct v3d_compile
*c
, struct qreg src
, bool is_cos
)
336 struct qreg input
= vir_FMUL(c
, src
, vir_uniform_f(c
, 1.0f
/ M_PI
));
338 input
= vir_FADD(c
, input
, vir_uniform_f(c
, 0.5));
340 struct qreg periods
= vir_FROUND(c
, input
);
341 struct qreg sin_output
= vir_SFU(c
, V3D_QPU_WADDR_SIN
,
342 vir_FSUB(c
, input
, periods
));
343 return vir_XOR(c
, sin_output
, vir_SHL(c
,
344 vir_FTOIN(c
, periods
),
345 vir_uniform_ui(c
, -1)));
349 ntq_fsign(struct v3d_compile
*c
, struct qreg src
)
351 struct qreg t
= vir_get_temp(c
);
353 vir_MOV_dest(c
, t
, vir_uniform_f(c
, 0.0));
354 vir_PF(c
, vir_FMOV(c
, src
), V3D_QPU_PF_PUSHZ
);
355 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_f(c
, 1.0));
356 vir_PF(c
, vir_FMOV(c
, src
), V3D_QPU_PF_PUSHN
);
357 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_f(c
, -1.0));
358 return vir_MOV(c
, t
);
362 ntq_isign(struct v3d_compile
*c
, struct qreg src
)
364 struct qreg t
= vir_get_temp(c
);
366 vir_MOV_dest(c
, t
, vir_uniform_ui(c
, 0));
367 vir_PF(c
, vir_MOV(c
, src
), V3D_QPU_PF_PUSHZ
);
368 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_ui(c
, 1));
369 vir_PF(c
, vir_MOV(c
, src
), V3D_QPU_PF_PUSHN
);
370 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_ui(c
, -1));
371 return vir_MOV(c
, t
);
375 emit_fragcoord_input(struct v3d_compile
*c
, int attr
)
377 c
->inputs
[attr
* 4 + 0] = vir_FXCD(c
);
378 c
->inputs
[attr
* 4 + 1] = vir_FYCD(c
);
379 c
->inputs
[attr
* 4 + 2] = c
->payload_z
;
380 c
->inputs
[attr
* 4 + 3] = vir_SFU(c
, V3D_QPU_WADDR_RECIP
,
385 emit_fragment_varying(struct v3d_compile
*c
, nir_variable
*var
,
388 struct qreg r3
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R3
);
389 struct qreg r5
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R5
);
392 if (c
->devinfo
->ver
>= 41) {
393 struct qinst
*ldvary
= vir_add_inst(V3D_QPU_A_NOP
, c
->undef
,
395 ldvary
->qpu
.sig
.ldvary
= true;
396 vary
= vir_emit_def(c
, ldvary
);
398 vir_NOP(c
)->qpu
.sig
.ldvary
= true;
402 /* For gl_PointCoord input or distance along a line, we'll be called
403 * with no nir_variable, and we don't count toward VPM size so we
404 * don't track an input slot.
407 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
410 int i
= c
->num_inputs
++;
411 c
->input_slots
[i
] = v3d_slot_from_slot_and_component(var
->data
.location
,
414 switch (var
->data
.interpolation
) {
415 case INTERP_MODE_NONE
:
416 /* If a gl_FrontColor or gl_BackColor input has no interp
417 * qualifier, then if we're using glShadeModel(GL_FLAT) it
418 * needs to be flat shaded.
420 switch (var
->data
.location
) {
421 case VARYING_SLOT_COL0
:
422 case VARYING_SLOT_COL1
:
423 case VARYING_SLOT_BFC0
:
424 case VARYING_SLOT_BFC1
:
425 if (c
->fs_key
->shade_model_flat
) {
426 BITSET_SET(c
->flat_shade_flags
, i
);
427 vir_MOV_dest(c
, c
->undef
, vary
);
428 return vir_MOV(c
, r5
);
430 return vir_FADD(c
, vir_FMUL(c
, vary
,
437 case INTERP_MODE_SMOOTH
:
438 if (var
->data
.centroid
) {
439 BITSET_SET(c
->centroid_flags
, i
);
440 return vir_FADD(c
, vir_FMUL(c
, vary
,
441 c
->payload_w_centroid
), r5
);
443 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
445 case INTERP_MODE_NOPERSPECTIVE
:
446 BITSET_SET(c
->noperspective_flags
, i
);
447 return vir_FADD(c
, vir_MOV(c
, vary
), r5
);
448 case INTERP_MODE_FLAT
:
449 BITSET_SET(c
->flat_shade_flags
, i
);
450 vir_MOV_dest(c
, c
->undef
, vary
);
451 return vir_MOV(c
, r5
);
453 unreachable("Bad interp mode");
458 emit_fragment_input(struct v3d_compile
*c
, int attr
, nir_variable
*var
)
460 for (int i
= 0; i
< glsl_get_vector_elements(var
->type
); i
++) {
461 int chan
= var
->data
.location_frac
+ i
;
462 c
->inputs
[attr
* 4 + chan
] =
463 emit_fragment_varying(c
, var
, chan
);
468 add_output(struct v3d_compile
*c
,
469 uint32_t decl_offset
,
473 uint32_t old_array_size
= c
->outputs_array_size
;
474 resize_qreg_array(c
, &c
->outputs
, &c
->outputs_array_size
,
477 if (old_array_size
!= c
->outputs_array_size
) {
478 c
->output_slots
= reralloc(c
,
480 struct v3d_varying_slot
,
481 c
->outputs_array_size
);
484 c
->output_slots
[decl_offset
] =
485 v3d_slot_from_slot_and_component(slot
, swizzle
);
489 declare_uniform_range(struct v3d_compile
*c
, uint32_t start
, uint32_t size
)
491 unsigned array_id
= c
->num_ubo_ranges
++;
492 if (array_id
>= c
->ubo_ranges_array_size
) {
493 c
->ubo_ranges_array_size
= MAX2(c
->ubo_ranges_array_size
* 2,
495 c
->ubo_ranges
= reralloc(c
, c
->ubo_ranges
,
496 struct v3d_ubo_range
,
497 c
->ubo_ranges_array_size
);
498 c
->ubo_range_used
= reralloc(c
, c
->ubo_range_used
,
500 c
->ubo_ranges_array_size
);
503 c
->ubo_ranges
[array_id
].dst_offset
= 0;
504 c
->ubo_ranges
[array_id
].src_offset
= start
;
505 c
->ubo_ranges
[array_id
].size
= size
;
506 c
->ubo_range_used
[array_id
] = false;
510 * If compare_instr is a valid comparison instruction, emits the
511 * compare_instr's comparison and returns the sel_instr's return value based
512 * on the compare_instr's result.
515 ntq_emit_comparison(struct v3d_compile
*c
, struct qreg
*dest
,
516 nir_alu_instr
*compare_instr
,
517 nir_alu_instr
*sel_instr
)
519 struct qreg src0
= ntq_get_alu_src(c
, compare_instr
, 0);
521 if (nir_op_infos
[compare_instr
->op
].num_inputs
> 1)
522 src1
= ntq_get_alu_src(c
, compare_instr
, 1);
523 bool cond_invert
= false;
525 switch (compare_instr
->op
) {
528 vir_PF(c
, vir_FCMP(c
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
531 vir_PF(c
, vir_XOR(c
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
536 vir_PF(c
, vir_FCMP(c
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
540 vir_PF(c
, vir_XOR(c
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
546 vir_PF(c
, vir_FCMP(c
, src1
, src0
), V3D_QPU_PF_PUSHC
);
549 vir_PF(c
, vir_MIN(c
, src1
, src0
), V3D_QPU_PF_PUSHC
);
553 vir_PF(c
, vir_SUB(c
, src0
, src1
), V3D_QPU_PF_PUSHC
);
559 vir_PF(c
, vir_FCMP(c
, src0
, src1
), V3D_QPU_PF_PUSHN
);
562 vir_PF(c
, vir_MIN(c
, src1
, src0
), V3D_QPU_PF_PUSHC
);
565 vir_PF(c
, vir_SUB(c
, src0
, src1
), V3D_QPU_PF_PUSHC
);
572 enum v3d_qpu_cond cond
= (cond_invert
?
576 switch (sel_instr
->op
) {
581 *dest
= vir_SEL(c
, cond
,
582 vir_uniform_f(c
, 1.0), vir_uniform_f(c
, 0.0));
586 *dest
= vir_SEL(c
, cond
,
587 ntq_get_alu_src(c
, sel_instr
, 1),
588 ntq_get_alu_src(c
, sel_instr
, 2));
592 *dest
= vir_SEL(c
, cond
,
593 vir_uniform_ui(c
, ~0), vir_uniform_ui(c
, 0));
597 /* Make the temporary for nir_store_dest(). */
598 *dest
= vir_MOV(c
, *dest
);
604 * Attempts to fold a comparison generating a boolean result into the
605 * condition code for selecting between two values, instead of comparing the
606 * boolean result against 0 to generate the condition code.
608 static struct qreg
ntq_emit_bcsel(struct v3d_compile
*c
, nir_alu_instr
*instr
,
611 if (!instr
->src
[0].src
.is_ssa
)
613 if (instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
615 nir_alu_instr
*compare
=
616 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
621 if (ntq_emit_comparison(c
, &dest
, compare
, instr
))
625 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
626 return vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
, src
[1], src
[2]));
631 ntq_emit_alu(struct v3d_compile
*c
, nir_alu_instr
*instr
)
633 /* This should always be lowered to ALU operations for V3D. */
634 assert(!instr
->dest
.saturate
);
636 /* Vectors are special in that they have non-scalarized writemasks,
637 * and just take the first swizzle channel for each argument in order
638 * into each writemask channel.
640 if (instr
->op
== nir_op_vec2
||
641 instr
->op
== nir_op_vec3
||
642 instr
->op
== nir_op_vec4
) {
644 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
645 srcs
[i
] = ntq_get_src(c
, instr
->src
[i
].src
,
646 instr
->src
[i
].swizzle
[0]);
647 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
648 ntq_store_dest(c
, &instr
->dest
.dest
, i
,
649 vir_MOV(c
, srcs
[i
]));
653 /* General case: We can just grab the one used channel per src. */
654 struct qreg src
[nir_op_infos
[instr
->op
].num_inputs
];
655 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
656 src
[i
] = ntq_get_alu_src(c
, instr
, i
);
664 result
= vir_MOV(c
, src
[0]);
668 result
= vir_XOR(c
, src
[0], vir_uniform_ui(c
, 1 << 31));
671 result
= vir_NEG(c
, src
[0]);
675 result
= vir_FMUL(c
, src
[0], src
[1]);
678 result
= vir_FADD(c
, src
[0], src
[1]);
681 result
= vir_FSUB(c
, src
[0], src
[1]);
684 result
= vir_FMIN(c
, src
[0], src
[1]);
687 result
= vir_FMAX(c
, src
[0], src
[1]);
691 result
= vir_FTOIZ(c
, src
[0]);
694 result
= vir_FTOUZ(c
, src
[0]);
697 result
= vir_ITOF(c
, src
[0]);
700 result
= vir_UTOF(c
, src
[0]);
703 result
= vir_AND(c
, src
[0], vir_uniform_f(c
, 1.0));
706 result
= vir_AND(c
, src
[0], vir_uniform_ui(c
, 1));
710 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
711 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
712 vir_uniform_ui(c
, ~0),
713 vir_uniform_ui(c
, 0)));
717 result
= vir_ADD(c
, src
[0], src
[1]);
720 result
= vir_SHR(c
, src
[0], src
[1]);
723 result
= vir_SUB(c
, src
[0], src
[1]);
726 result
= vir_ASR(c
, src
[0], src
[1]);
729 result
= vir_SHL(c
, src
[0], src
[1]);
732 result
= vir_MIN(c
, src
[0], src
[1]);
735 result
= vir_UMIN(c
, src
[0], src
[1]);
738 result
= vir_MAX(c
, src
[0], src
[1]);
741 result
= vir_UMAX(c
, src
[0], src
[1]);
744 result
= vir_AND(c
, src
[0], src
[1]);
747 result
= vir_OR(c
, src
[0], src
[1]);
750 result
= vir_XOR(c
, src
[0], src
[1]);
753 result
= vir_NOT(c
, src
[0]);
756 case nir_op_ufind_msb
:
757 result
= vir_SUB(c
, vir_uniform_ui(c
, 31), vir_CLZ(c
, src
[0]));
761 result
= vir_UMUL(c
, src
[0], src
[1]);
778 if (!ntq_emit_comparison(c
, &result
, instr
, instr
)) {
779 fprintf(stderr
, "Bad comparison instruction\n");
784 result
= ntq_emit_bcsel(c
, instr
, src
);
787 vir_PF(c
, src
[0], V3D_QPU_PF_PUSHZ
);
788 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
793 result
= vir_SFU(c
, V3D_QPU_WADDR_RECIP
, src
[0]);
796 result
= vir_SFU(c
, V3D_QPU_WADDR_RSQRT
, src
[0]);
799 result
= vir_SFU(c
, V3D_QPU_WADDR_EXP
, src
[0]);
802 result
= vir_SFU(c
, V3D_QPU_WADDR_LOG
, src
[0]);
806 result
= vir_FCEIL(c
, src
[0]);
809 result
= vir_FFLOOR(c
, src
[0]);
811 case nir_op_fround_even
:
812 result
= vir_FROUND(c
, src
[0]);
815 result
= vir_FTRUNC(c
, src
[0]);
818 result
= vir_FSUB(c
, src
[0], vir_FFLOOR(c
, src
[0]));
822 result
= ntq_fsincos(c
, src
[0], false);
825 result
= ntq_fsincos(c
, src
[0], true);
829 result
= ntq_fsign(c
, src
[0]);
832 result
= ntq_isign(c
, src
[0]);
836 result
= vir_FMOV(c
, src
[0]);
837 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_ABS
);
842 result
= vir_MAX(c
, src
[0],
843 vir_SUB(c
, vir_uniform_ui(c
, 0), src
[0]));
847 case nir_op_fddx_coarse
:
848 case nir_op_fddx_fine
:
849 result
= vir_FDX(c
, src
[0]);
853 case nir_op_fddy_coarse
:
854 case nir_op_fddy_fine
:
855 result
= vir_FDY(c
, src
[0]);
858 case nir_op_uadd_carry
:
859 vir_PF(c
, vir_ADD(c
, src
[0], src
[1]), V3D_QPU_PF_PUSHC
);
860 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFA
,
861 vir_uniform_ui(c
, ~0),
862 vir_uniform_ui(c
, 0)));
866 fprintf(stderr
, "unknown NIR ALU inst: ");
867 nir_print_instr(&instr
->instr
, stderr
);
868 fprintf(stderr
, "\n");
872 /* We have a scalar result, so the instruction should only have a
873 * single channel written to.
875 assert(util_is_power_of_two_or_zero(instr
->dest
.write_mask
));
876 ntq_store_dest(c
, &instr
->dest
.dest
,
877 ffs(instr
->dest
.write_mask
) - 1, result
);
880 /* Each TLB read/write setup (a render target or depth buffer) takes an 8-bit
881 * specifier. They come from a register that's preloaded with 0xffffffff
882 * (0xff gets you normal vec4 f16 RT0 writes), and when one is neaded the low
883 * 8 bits are shifted off the bottom and 0xff shifted in from the top.
885 #define TLB_TYPE_F16_COLOR (3 << 6)
886 #define TLB_TYPE_I32_COLOR (1 << 6)
887 #define TLB_TYPE_F32_COLOR (0 << 6)
888 #define TLB_RENDER_TARGET_SHIFT 3 /* Reversed! 7 = RT 0, 0 = RT 7. */
889 #define TLB_SAMPLE_MODE_PER_SAMPLE (0 << 2)
890 #define TLB_SAMPLE_MODE_PER_PIXEL (1 << 2)
891 #define TLB_F16_SWAP_HI_LO (1 << 1)
892 #define TLB_VEC_SIZE_4_F16 (1 << 0)
893 #define TLB_VEC_SIZE_2_F16 (0 << 0)
894 #define TLB_VEC_SIZE_MINUS_1_SHIFT 0
896 /* Triggers Z/Stencil testing, used when the shader state's "FS modifies Z"
899 #define TLB_TYPE_DEPTH ((2 << 6) | (0 << 4))
900 #define TLB_DEPTH_TYPE_INVARIANT (0 << 2) /* Unmodified sideband input used */
901 #define TLB_DEPTH_TYPE_PER_PIXEL (1 << 2) /* QPU result used */
903 /* Stencil is a single 32-bit write. */
904 #define TLB_TYPE_STENCIL_ALPHA ((2 << 6) | (1 << 4))
907 emit_frag_end(struct v3d_compile
*c
)
910 if (c->output_sample_mask_index != -1) {
911 vir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
915 bool has_any_tlb_color_write
= false;
916 for (int rt
= 0; rt
< c
->fs_key
->nr_cbufs
; rt
++) {
917 if (c
->output_color_var
[rt
])
918 has_any_tlb_color_write
= true;
921 if (c
->fs_key
->sample_alpha_to_coverage
&& c
->output_color_var
[0]) {
922 struct nir_variable
*var
= c
->output_color_var
[0];
923 struct qreg
*color
= &c
->outputs
[var
->data
.driver_location
* 4];
925 vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
928 vir_FTOC(c
, color
[3])));
931 if (c
->output_position_index
!= -1) {
932 struct qinst
*inst
= vir_MOV_dest(c
,
933 vir_reg(QFILE_TLBU
, 0),
934 c
->outputs
[c
->output_position_index
]);
936 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
939 TLB_DEPTH_TYPE_PER_PIXEL
|
941 } else if (c
->s
->info
.fs
.uses_discard
||
942 c
->fs_key
->sample_alpha_to_coverage
||
943 !has_any_tlb_color_write
) {
944 /* Emit passthrough Z if it needed to be delayed until shader
945 * end due to potential discards.
947 * Since (single-threaded) fragment shaders always need a TLB
948 * write, emit passthrouh Z if we didn't have any color
949 * buffers and flag us as potentially discarding, so that we
950 * can use Z as the TLB write.
952 c
->s
->info
.fs
.uses_discard
= true;
954 struct qinst
*inst
= vir_MOV_dest(c
,
955 vir_reg(QFILE_TLBU
, 0),
956 vir_reg(QFILE_NULL
, 0));
958 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
961 TLB_DEPTH_TYPE_INVARIANT
|
965 /* XXX: Performance improvement: Merge Z write and color writes TLB
969 for (int rt
= 0; rt
< c
->fs_key
->nr_cbufs
; rt
++) {
970 if (!c
->output_color_var
[rt
])
973 nir_variable
*var
= c
->output_color_var
[rt
];
974 struct qreg
*color
= &c
->outputs
[var
->data
.driver_location
* 4];
975 int num_components
= glsl_get_vector_elements(var
->type
);
976 uint32_t conf
= 0xffffff00;
979 conf
|= TLB_SAMPLE_MODE_PER_PIXEL
;
980 conf
|= (7 - rt
) << TLB_RENDER_TARGET_SHIFT
;
982 if (c
->fs_key
->swap_color_rb
& (1 << rt
))
983 num_components
= MAX2(num_components
, 3);
985 assert(num_components
!= 0);
986 switch (glsl_get_base_type(var
->type
)) {
989 /* The F32 vs I32 distinction was dropped in 4.2. */
990 if (c
->devinfo
->ver
< 42)
991 conf
|= TLB_TYPE_I32_COLOR
;
993 conf
|= TLB_TYPE_F32_COLOR
;
994 conf
|= ((num_components
- 1) <<
995 TLB_VEC_SIZE_MINUS_1_SHIFT
);
997 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLBU
, 0), color
[0]);
998 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
999 vir_uniform_ui(c
, conf
);
1001 for (int i
= 1; i
< num_components
; i
++) {
1002 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0),
1008 struct qreg r
= color
[0];
1009 struct qreg g
= color
[1];
1010 struct qreg b
= color
[2];
1011 struct qreg a
= color
[3];
1013 if (c
->fs_key
->f32_color_rb
& (1 << rt
)) {
1014 conf
|= TLB_TYPE_F32_COLOR
;
1015 conf
|= ((num_components
- 1) <<
1016 TLB_VEC_SIZE_MINUS_1_SHIFT
);
1018 conf
|= TLB_TYPE_F16_COLOR
;
1019 conf
|= TLB_F16_SWAP_HI_LO
;
1020 if (num_components
>= 3)
1021 conf
|= TLB_VEC_SIZE_4_F16
;
1023 conf
|= TLB_VEC_SIZE_2_F16
;
1026 if (c
->fs_key
->swap_color_rb
& (1 << rt
)) {
1031 if (c
->fs_key
->sample_alpha_to_one
)
1032 a
= vir_uniform_f(c
, 1.0);
1034 if (c
->fs_key
->f32_color_rb
& (1 << rt
)) {
1035 inst
= vir_MOV_dest(c
, vir_reg(QFILE_TLBU
, 0), r
);
1036 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1037 vir_uniform_ui(c
, conf
);
1039 if (num_components
>= 2)
1040 vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0), g
);
1041 if (num_components
>= 3)
1042 vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0), b
);
1043 if (num_components
>= 4)
1044 vir_MOV_dest(c
, vir_reg(QFILE_TLB
, 0), a
);
1046 inst
= vir_VFPACK_dest(c
, vir_reg(QFILE_TLB
, 0), r
, g
);
1048 inst
->dst
.file
= QFILE_TLBU
;
1049 inst
->src
[vir_get_implicit_uniform_src(inst
)] =
1050 vir_uniform_ui(c
, conf
);
1053 if (num_components
>= 3)
1054 inst
= vir_VFPACK_dest(c
, vir_reg(QFILE_TLB
, 0), b
, a
);
1063 vir_VPM_WRITE(struct v3d_compile
*c
, struct qreg val
, uint32_t *vpm_index
)
1065 if (c
->devinfo
->ver
>= 40) {
1066 vir_STVPMV(c
, vir_uniform_ui(c
, *vpm_index
), val
);
1067 *vpm_index
= *vpm_index
+ 1;
1069 vir_MOV_dest(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_VPM
), val
);
1072 c
->num_vpm_writes
++;
1076 emit_scaled_viewport_write(struct v3d_compile
*c
, struct qreg rcp_w
,
1077 uint32_t *vpm_index
)
1079 for (int i
= 0; i
< 2; i
++) {
1080 struct qreg coord
= c
->outputs
[c
->output_position_index
+ i
];
1081 coord
= vir_FMUL(c
, coord
,
1082 vir_uniform(c
, QUNIFORM_VIEWPORT_X_SCALE
+ i
,
1084 coord
= vir_FMUL(c
, coord
, rcp_w
);
1085 vir_VPM_WRITE(c
, vir_FTOIN(c
, coord
), vpm_index
);
1091 emit_zs_write(struct v3d_compile
*c
, struct qreg rcp_w
, uint32_t *vpm_index
)
1093 struct qreg zscale
= vir_uniform(c
, QUNIFORM_VIEWPORT_Z_SCALE
, 0);
1094 struct qreg zoffset
= vir_uniform(c
, QUNIFORM_VIEWPORT_Z_OFFSET
, 0);
1096 struct qreg z
= c
->outputs
[c
->output_position_index
+ 2];
1097 z
= vir_FMUL(c
, z
, zscale
);
1098 z
= vir_FMUL(c
, z
, rcp_w
);
1099 z
= vir_FADD(c
, z
, zoffset
);
1100 vir_VPM_WRITE(c
, z
, vpm_index
);
1104 emit_rcp_wc_write(struct v3d_compile
*c
, struct qreg rcp_w
, uint32_t *vpm_index
)
1106 vir_VPM_WRITE(c
, rcp_w
, vpm_index
);
1110 emit_point_size_write(struct v3d_compile
*c
, uint32_t *vpm_index
)
1112 struct qreg point_size
;
1114 if (c
->output_point_size_index
!= -1)
1115 point_size
= c
->outputs
[c
->output_point_size_index
];
1117 point_size
= vir_uniform_f(c
, 1.0);
1119 /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
1122 point_size
= vir_FMAX(c
, point_size
, vir_uniform_f(c
, .125));
1124 vir_VPM_WRITE(c
, point_size
, vpm_index
);
1128 emit_vpm_write_setup(struct v3d_compile
*c
)
1130 if (c
->devinfo
->ver
>= 40)
1133 v3d33_vir_vpm_write_setup(c
);
1137 * Sets up c->outputs[c->output_position_index] for the vertex shader
1138 * epilogue, if an output vertex position wasn't specified in the user's
1139 * shader. This may be the case for transform feedback with rasterizer
1143 setup_default_position(struct v3d_compile
*c
)
1145 if (c
->output_position_index
!= -1)
1148 c
->output_position_index
= c
->outputs_array_size
;
1149 for (int i
= 0; i
< 4; i
++) {
1151 c
->output_position_index
+ i
,
1152 VARYING_SLOT_POS
, i
);
1157 emit_vert_end(struct v3d_compile
*c
)
1159 setup_default_position(c
);
1161 uint32_t vpm_index
= 0;
1162 struct qreg rcp_w
= vir_SFU(c
, V3D_QPU_WADDR_RECIP
,
1163 c
->outputs
[c
->output_position_index
+ 3]);
1165 emit_vpm_write_setup(c
);
1167 if (c
->vs_key
->is_coord
) {
1168 for (int i
= 0; i
< 4; i
++)
1169 vir_VPM_WRITE(c
, c
->outputs
[c
->output_position_index
+ i
],
1171 emit_scaled_viewport_write(c
, rcp_w
, &vpm_index
);
1172 if (c
->vs_key
->per_vertex_point_size
) {
1173 emit_point_size_write(c
, &vpm_index
);
1174 /* emit_rcp_wc_write(c, rcp_w); */
1176 /* XXX: Z-only rendering */
1178 emit_zs_write(c
, rcp_w
, &vpm_index
);
1180 emit_scaled_viewport_write(c
, rcp_w
, &vpm_index
);
1181 emit_zs_write(c
, rcp_w
, &vpm_index
);
1182 emit_rcp_wc_write(c
, rcp_w
, &vpm_index
);
1183 if (c
->vs_key
->per_vertex_point_size
)
1184 emit_point_size_write(c
, &vpm_index
);
1187 for (int i
= 0; i
< c
->vs_key
->num_fs_inputs
; i
++) {
1188 struct v3d_varying_slot input
= c
->vs_key
->fs_inputs
[i
];
1191 for (j
= 0; j
< c
->num_outputs
; j
++) {
1192 struct v3d_varying_slot output
= c
->output_slots
[j
];
1194 if (!memcmp(&input
, &output
, sizeof(input
))) {
1195 vir_VPM_WRITE(c
, c
->outputs
[j
],
1200 /* Emit padding if we didn't find a declared VS output for
1203 if (j
== c
->num_outputs
)
1204 vir_VPM_WRITE(c
, vir_uniform_f(c
, 0.0),
1208 /* GFXH-1684: VPM writes need to be complete by the end of the shader.
1210 if (c
->devinfo
->ver
>= 40 && c
->devinfo
->ver
<= 42)
1215 v3d_optimize_nir(struct nir_shader
*s
)
1222 NIR_PASS_V(s
, nir_lower_vars_to_ssa
);
1223 NIR_PASS(progress
, s
, nir_lower_alu_to_scalar
);
1224 NIR_PASS(progress
, s
, nir_lower_phis_to_scalar
);
1225 NIR_PASS(progress
, s
, nir_copy_prop
);
1226 NIR_PASS(progress
, s
, nir_opt_remove_phis
);
1227 NIR_PASS(progress
, s
, nir_opt_dce
);
1228 NIR_PASS(progress
, s
, nir_opt_dead_cf
);
1229 NIR_PASS(progress
, s
, nir_opt_cse
);
1230 NIR_PASS(progress
, s
, nir_opt_peephole_select
, 8);
1231 NIR_PASS(progress
, s
, nir_opt_algebraic
);
1232 NIR_PASS(progress
, s
, nir_opt_constant_folding
);
1233 NIR_PASS(progress
, s
, nir_opt_undef
);
1236 NIR_PASS(progress
, s
, nir_opt_move_load_ubo
);
1240 driver_location_compare(const void *in_a
, const void *in_b
)
1242 const nir_variable
*const *a
= in_a
;
1243 const nir_variable
*const *b
= in_b
;
1245 return (*a
)->data
.driver_location
- (*b
)->data
.driver_location
;
1249 ntq_emit_vpm_read(struct v3d_compile
*c
,
1250 uint32_t *num_components_queued
,
1251 uint32_t *remaining
,
1254 struct qreg vpm
= vir_reg(QFILE_VPM
, vpm_index
);
1256 if (c
->devinfo
->ver
>= 40 ) {
1257 return vir_LDVPMV_IN(c
,
1259 (*num_components_queued
)++));
1262 if (*num_components_queued
!= 0) {
1263 (*num_components_queued
)--;
1265 return vir_MOV(c
, vpm
);
1268 uint32_t num_components
= MIN2(*remaining
, 32);
1270 v3d33_vir_vpm_read_setup(c
, num_components
);
1272 *num_components_queued
= num_components
- 1;
1273 *remaining
-= num_components
;
1276 return vir_MOV(c
, vpm
);
1280 ntq_setup_inputs(struct v3d_compile
*c
)
1282 unsigned num_entries
= 0;
1283 unsigned num_components
= 0;
1284 nir_foreach_variable(var
, &c
->s
->inputs
) {
1286 num_components
+= glsl_get_components(var
->type
);
1289 nir_variable
*vars
[num_entries
];
1292 nir_foreach_variable(var
, &c
->s
->inputs
)
1295 /* Sort the variables so that we emit the input setup in
1296 * driver_location order. This is required for VPM reads, whose data
1297 * is fetched into the VPM in driver_location (TGSI register index)
1300 qsort(&vars
, num_entries
, sizeof(*vars
), driver_location_compare
);
1302 uint32_t vpm_components_queued
= 0;
1303 if (c
->s
->info
.stage
== MESA_SHADER_VERTEX
) {
1304 bool uses_iid
= c
->s
->info
.system_values_read
&
1305 (1ull << SYSTEM_VALUE_INSTANCE_ID
);
1306 bool uses_vid
= c
->s
->info
.system_values_read
&
1307 (1ull << SYSTEM_VALUE_VERTEX_ID
);
1309 num_components
+= uses_iid
;
1310 num_components
+= uses_vid
;
1313 c
->iid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1314 &num_components
, ~0);
1318 c
->vid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1319 &num_components
, ~0);
1323 for (unsigned i
= 0; i
< num_entries
; i
++) {
1324 nir_variable
*var
= vars
[i
];
1325 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1326 unsigned loc
= var
->data
.driver_location
;
1328 assert(array_len
== 1);
1330 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1333 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1334 if (var
->data
.location
== VARYING_SLOT_POS
) {
1335 emit_fragcoord_input(c
, loc
);
1336 } else if (var
->data
.location
== VARYING_SLOT_PNTC
||
1337 (var
->data
.location
>= VARYING_SLOT_VAR0
&&
1338 (c
->fs_key
->point_sprite_mask
&
1339 (1 << (var
->data
.location
-
1340 VARYING_SLOT_VAR0
))))) {
1341 c
->inputs
[loc
* 4 + 0] = c
->point_x
;
1342 c
->inputs
[loc
* 4 + 1] = c
->point_y
;
1344 emit_fragment_input(c
, loc
, var
);
1347 int var_components
= glsl_get_components(var
->type
);
1349 for (int i
= 0; i
< var_components
; i
++) {
1350 c
->inputs
[loc
* 4 + i
] =
1351 ntq_emit_vpm_read(c
,
1352 &vpm_components_queued
,
1357 c
->vattr_sizes
[loc
] = var_components
;
1361 if (c
->s
->info
.stage
== MESA_SHADER_VERTEX
) {
1362 if (c
->devinfo
->ver
>= 40) {
1363 assert(vpm_components_queued
== num_components
);
1365 assert(vpm_components_queued
== 0);
1366 assert(num_components
== 0);
1372 ntq_setup_outputs(struct v3d_compile
*c
)
1374 nir_foreach_variable(var
, &c
->s
->outputs
) {
1375 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1376 unsigned loc
= var
->data
.driver_location
* 4;
1378 assert(array_len
== 1);
1381 for (int i
= 0; i
< 4; i
++) {
1382 add_output(c
, loc
+ var
->data
.location_frac
+ i
,
1384 var
->data
.location_frac
+ i
);
1387 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1388 switch (var
->data
.location
) {
1389 case FRAG_RESULT_COLOR
:
1390 c
->output_color_var
[0] = var
;
1391 c
->output_color_var
[1] = var
;
1392 c
->output_color_var
[2] = var
;
1393 c
->output_color_var
[3] = var
;
1395 case FRAG_RESULT_DATA0
:
1396 case FRAG_RESULT_DATA1
:
1397 case FRAG_RESULT_DATA2
:
1398 case FRAG_RESULT_DATA3
:
1399 c
->output_color_var
[var
->data
.location
-
1400 FRAG_RESULT_DATA0
] = var
;
1402 case FRAG_RESULT_DEPTH
:
1403 c
->output_position_index
= loc
;
1405 case FRAG_RESULT_SAMPLE_MASK
:
1406 c
->output_sample_mask_index
= loc
;
1410 switch (var
->data
.location
) {
1411 case VARYING_SLOT_POS
:
1412 c
->output_position_index
= loc
;
1414 case VARYING_SLOT_PSIZ
:
1415 c
->output_point_size_index
= loc
;
1423 ntq_setup_uniforms(struct v3d_compile
*c
)
1425 nir_foreach_variable(var
, &c
->s
->uniforms
) {
1426 uint32_t vec4_count
= glsl_count_attribute_slots(var
->type
,
1428 unsigned vec4_size
= 4 * sizeof(float);
1430 declare_uniform_range(c
, var
->data
.driver_location
* vec4_size
,
1431 vec4_count
* vec4_size
);
1437 * Sets up the mapping from nir_register to struct qreg *.
1439 * Each nir_register gets a struct qreg per 32-bit component being stored.
1442 ntq_setup_registers(struct v3d_compile
*c
, struct exec_list
*list
)
1444 foreach_list_typed(nir_register
, nir_reg
, node
, list
) {
1445 unsigned array_len
= MAX2(nir_reg
->num_array_elems
, 1);
1446 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
1448 nir_reg
->num_components
);
1450 _mesa_hash_table_insert(c
->def_ht
, nir_reg
, qregs
);
1452 for (int i
= 0; i
< array_len
* nir_reg
->num_components
; i
++)
1453 qregs
[i
] = vir_get_temp(c
);
1458 ntq_emit_load_const(struct v3d_compile
*c
, nir_load_const_instr
*instr
)
1460 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1461 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1462 qregs
[i
] = vir_uniform_ui(c
, instr
->value
.u32
[i
]);
1464 _mesa_hash_table_insert(c
->def_ht
, &instr
->def
, qregs
);
1468 ntq_emit_ssa_undef(struct v3d_compile
*c
, nir_ssa_undef_instr
*instr
)
1470 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1472 /* VIR needs there to be *some* value, so pick 0 (same as for
1473 * ntq_setup_registers().
1475 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1476 qregs
[i
] = vir_uniform_ui(c
, 0);
1480 ntq_emit_intrinsic(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1482 nir_const_value
*const_offset
;
1485 switch (instr
->intrinsic
) {
1486 case nir_intrinsic_load_uniform
:
1487 assert(instr
->num_components
== 1);
1488 const_offset
= nir_src_as_const_value(instr
->src
[0]);
1490 offset
= nir_intrinsic_base(instr
) + const_offset
->u32
[0];
1491 assert(offset
% 4 == 0);
1492 /* We need dwords */
1493 offset
= offset
/ 4;
1494 ntq_store_dest(c
, &instr
->dest
, 0,
1495 vir_uniform(c
, QUNIFORM_UNIFORM
,
1498 ntq_store_dest(c
, &instr
->dest
, 0,
1499 indirect_uniform_load(c
, instr
));
1503 case nir_intrinsic_load_ubo
:
1504 for (int i
= 0; i
< instr
->num_components
; i
++) {
1505 int ubo
= nir_src_as_const_value(instr
->src
[0])->u32
[0];
1507 /* Adjust for where we stored the TGSI register base. */
1509 vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUA
),
1510 vir_uniform(c
, QUNIFORM_UBO_ADDR
, 1 + ubo
),
1512 ntq_get_src(c
, instr
->src
[1], 0),
1513 vir_uniform_ui(c
, i
* 4)));
1517 ntq_store_dest(c
, &instr
->dest
, i
, vir_LDTMU(c
));
1521 const_offset
= nir_src_as_const_value(instr
->src
[0]);
1523 offset
= nir_intrinsic_base(instr
) + const_offset
->u32
[0];
1524 assert(offset
% 4 == 0);
1525 /* We need dwords */
1526 offset
= offset
/ 4;
1527 ntq_store_dest(c
, &instr
->dest
, 0,
1528 vir_uniform(c
, QUNIFORM_UNIFORM
,
1531 ntq_store_dest(c
, &instr
->dest
, 0,
1532 indirect_uniform_load(c
, instr
));
1536 case nir_intrinsic_load_user_clip_plane
:
1537 for (int i
= 0; i
< instr
->num_components
; i
++) {
1538 ntq_store_dest(c
, &instr
->dest
, i
,
1539 vir_uniform(c
, QUNIFORM_USER_CLIP_PLANE
,
1540 nir_intrinsic_ucp_id(instr
) *
1545 case nir_intrinsic_load_alpha_ref_float
:
1546 ntq_store_dest(c
, &instr
->dest
, 0,
1547 vir_uniform(c
, QUNIFORM_ALPHA_REF
, 0));
1550 case nir_intrinsic_load_sample_mask_in
:
1551 ntq_store_dest(c
, &instr
->dest
, 0,
1552 vir_uniform(c
, QUNIFORM_SAMPLE_MASK
, 0));
1555 case nir_intrinsic_load_front_face
:
1556 /* The register contains 0 (front) or 1 (back), and we need to
1557 * turn it into a NIR bool where true means front.
1559 ntq_store_dest(c
, &instr
->dest
, 0,
1561 vir_uniform_ui(c
, -1),
1565 case nir_intrinsic_load_instance_id
:
1566 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->iid
));
1569 case nir_intrinsic_load_vertex_id
:
1570 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->vid
));
1573 case nir_intrinsic_load_input
:
1574 const_offset
= nir_src_as_const_value(instr
->src
[0]);
1575 assert(const_offset
&& "v3d doesn't support indirect inputs");
1576 for (int i
= 0; i
< instr
->num_components
; i
++) {
1577 offset
= nir_intrinsic_base(instr
) + const_offset
->u32
[0];
1578 int comp
= nir_intrinsic_component(instr
) + i
;
1579 ntq_store_dest(c
, &instr
->dest
, i
,
1580 vir_MOV(c
, c
->inputs
[offset
* 4 + comp
]));
1584 case nir_intrinsic_store_output
:
1585 const_offset
= nir_src_as_const_value(instr
->src
[1]);
1586 assert(const_offset
&& "v3d doesn't support indirect outputs");
1587 offset
= ((nir_intrinsic_base(instr
) +
1588 const_offset
->u32
[0]) * 4 +
1589 nir_intrinsic_component(instr
));
1591 for (int i
= 0; i
< instr
->num_components
; i
++) {
1592 c
->outputs
[offset
+ i
] =
1593 vir_MOV(c
, ntq_get_src(c
, instr
->src
[0], i
));
1595 c
->num_outputs
= MAX2(c
->num_outputs
,
1596 offset
+ instr
->num_components
);
1599 case nir_intrinsic_discard
:
1600 if (c
->execute
.file
!= QFILE_NULL
) {
1601 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1602 vir_set_cond(vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1603 vir_uniform_ui(c
, 0)),
1606 vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1607 vir_uniform_ui(c
, 0));
1611 case nir_intrinsic_discard_if
: {
1612 /* true (~0) if we're discarding */
1613 struct qreg cond
= ntq_get_src(c
, instr
->src
[0], 0);
1615 if (c
->execute
.file
!= QFILE_NULL
) {
1616 /* execute == 0 means the channel is active. Invert
1617 * the condition so that we can use zero as "executing
1620 vir_PF(c
, vir_OR(c
, c
->execute
, vir_NOT(c
, cond
)),
1622 vir_set_cond(vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1623 vir_uniform_ui(c
, 0)),
1626 vir_PF(c
, cond
, V3D_QPU_PF_PUSHZ
);
1627 vir_set_cond(vir_SETMSF_dest(c
, vir_reg(QFILE_NULL
, 0),
1628 vir_uniform_ui(c
, 0)),
1636 fprintf(stderr
, "Unknown intrinsic: ");
1637 nir_print_instr(&instr
->instr
, stderr
);
1638 fprintf(stderr
, "\n");
1643 /* Clears (activates) the execute flags for any channels whose jump target
1644 * matches this block.
1647 ntq_activate_execute_for_block(struct v3d_compile
*c
)
1649 vir_PF(c
, vir_XOR(c
, c
->execute
, vir_uniform_ui(c
, c
->cur_block
->index
)),
1652 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
1656 ntq_emit_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
1658 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
1659 bool empty_else_block
=
1660 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
1661 exec_list_is_empty(&nir_else_block
->instr_list
));
1663 struct qblock
*then_block
= vir_new_block(c
);
1664 struct qblock
*after_block
= vir_new_block(c
);
1665 struct qblock
*else_block
;
1666 if (empty_else_block
)
1667 else_block
= after_block
;
1669 else_block
= vir_new_block(c
);
1671 bool was_top_level
= false;
1672 if (c
->execute
.file
== QFILE_NULL
) {
1673 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
1674 was_top_level
= true;
1677 /* Set A for executing (execute == 0) and jumping (if->condition ==
1678 * 0) channels, and then update execute flags for those to point to
1683 ntq_get_src(c
, if_stmt
->condition
, 0)),
1685 vir_MOV_cond(c
, V3D_QPU_COND_IFA
,
1687 vir_uniform_ui(c
, else_block
->index
));
1689 /* Jump to ELSE if nothing is active for THEN, otherwise fall
1692 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1693 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLNA
);
1694 vir_link_blocks(c
->cur_block
, else_block
);
1695 vir_link_blocks(c
->cur_block
, then_block
);
1697 /* Process the THEN block. */
1698 vir_set_emit_block(c
, then_block
);
1699 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
1701 if (!empty_else_block
) {
1702 /* Handle the end of the THEN block. First, all currently
1703 * active channels update their execute flags to point to
1706 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1707 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
1708 vir_uniform_ui(c
, after_block
->index
));
1710 /* If everything points at ENDIF, then jump there immediately. */
1711 vir_PF(c
, vir_XOR(c
, c
->execute
,
1712 vir_uniform_ui(c
, after_block
->index
)),
1714 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLA
);
1715 vir_link_blocks(c
->cur_block
, after_block
);
1716 vir_link_blocks(c
->cur_block
, else_block
);
1718 vir_set_emit_block(c
, else_block
);
1719 ntq_activate_execute_for_block(c
);
1720 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
1723 vir_link_blocks(c
->cur_block
, after_block
);
1725 vir_set_emit_block(c
, after_block
);
1727 c
->execute
= c
->undef
;
1729 ntq_activate_execute_for_block(c
);
1733 ntq_emit_jump(struct v3d_compile
*c
, nir_jump_instr
*jump
)
1735 switch (jump
->type
) {
1736 case nir_jump_break
:
1737 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1738 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
1739 vir_uniform_ui(c
, c
->loop_break_block
->index
));
1742 case nir_jump_continue
:
1743 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1744 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
1745 vir_uniform_ui(c
, c
->loop_cont_block
->index
));
1748 case nir_jump_return
:
1749 unreachable("All returns shouold be lowered\n");
1754 ntq_emit_instr(struct v3d_compile
*c
, nir_instr
*instr
)
1756 switch (instr
->type
) {
1757 case nir_instr_type_alu
:
1758 ntq_emit_alu(c
, nir_instr_as_alu(instr
));
1761 case nir_instr_type_intrinsic
:
1762 ntq_emit_intrinsic(c
, nir_instr_as_intrinsic(instr
));
1765 case nir_instr_type_load_const
:
1766 ntq_emit_load_const(c
, nir_instr_as_load_const(instr
));
1769 case nir_instr_type_ssa_undef
:
1770 ntq_emit_ssa_undef(c
, nir_instr_as_ssa_undef(instr
));
1773 case nir_instr_type_tex
:
1774 ntq_emit_tex(c
, nir_instr_as_tex(instr
));
1777 case nir_instr_type_jump
:
1778 ntq_emit_jump(c
, nir_instr_as_jump(instr
));
1782 fprintf(stderr
, "Unknown NIR instr type: ");
1783 nir_print_instr(instr
, stderr
);
1784 fprintf(stderr
, "\n");
1790 ntq_emit_block(struct v3d_compile
*c
, nir_block
*block
)
1792 nir_foreach_instr(instr
, block
) {
1793 ntq_emit_instr(c
, instr
);
1797 static void ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
1800 ntq_emit_loop(struct v3d_compile
*c
, nir_loop
*loop
)
1802 bool was_top_level
= false;
1803 if (c
->execute
.file
== QFILE_NULL
) {
1804 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
1805 was_top_level
= true;
1808 struct qblock
*save_loop_cont_block
= c
->loop_cont_block
;
1809 struct qblock
*save_loop_break_block
= c
->loop_break_block
;
1811 c
->loop_cont_block
= vir_new_block(c
);
1812 c
->loop_break_block
= vir_new_block(c
);
1814 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
1815 vir_set_emit_block(c
, c
->loop_cont_block
);
1816 ntq_activate_execute_for_block(c
);
1818 ntq_emit_cf_list(c
, &loop
->body
);
1820 /* Re-enable any previous continues now, so our ANYA check below
1823 * XXX: Use the .ORZ flags update, instead.
1825 vir_PF(c
, vir_XOR(c
,
1827 vir_uniform_ui(c
, c
->loop_cont_block
->index
)),
1829 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
1831 vir_PF(c
, c
->execute
, V3D_QPU_PF_PUSHZ
);
1833 struct qinst
*branch
= vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ANYA
);
1834 /* Pixels that were not dispatched or have been discarded should not
1835 * contribute to looping again.
1837 branch
->qpu
.branch
.msfign
= V3D_QPU_MSFIGN_P
;
1838 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
1839 vir_link_blocks(c
->cur_block
, c
->loop_break_block
);
1841 vir_set_emit_block(c
, c
->loop_break_block
);
1843 c
->execute
= c
->undef
;
1845 ntq_activate_execute_for_block(c
);
1847 c
->loop_break_block
= save_loop_break_block
;
1848 c
->loop_cont_block
= save_loop_cont_block
;
1852 ntq_emit_function(struct v3d_compile
*c
, nir_function_impl
*func
)
1854 fprintf(stderr
, "FUNCTIONS not handled.\n");
1859 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
)
1861 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
1862 switch (node
->type
) {
1863 case nir_cf_node_block
:
1864 ntq_emit_block(c
, nir_cf_node_as_block(node
));
1867 case nir_cf_node_if
:
1868 ntq_emit_if(c
, nir_cf_node_as_if(node
));
1871 case nir_cf_node_loop
:
1872 ntq_emit_loop(c
, nir_cf_node_as_loop(node
));
1875 case nir_cf_node_function
:
1876 ntq_emit_function(c
, nir_cf_node_as_function(node
));
1880 fprintf(stderr
, "Unknown NIR node type\n");
1887 ntq_emit_impl(struct v3d_compile
*c
, nir_function_impl
*impl
)
1889 ntq_setup_registers(c
, &impl
->registers
);
1890 ntq_emit_cf_list(c
, &impl
->body
);
1894 nir_to_vir(struct v3d_compile
*c
)
1896 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1897 c
->payload_w
= vir_MOV(c
, vir_reg(QFILE_REG
, 0));
1898 c
->payload_w_centroid
= vir_MOV(c
, vir_reg(QFILE_REG
, 1));
1899 c
->payload_z
= vir_MOV(c
, vir_reg(QFILE_REG
, 2));
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_all_io_to_temps
= true,
1924 .lower_extract_byte
= true,
1925 .lower_extract_word
= true,
1927 .lower_bitfield_insert_to_shifts
= true,
1928 .lower_bitfield_extract_to_shifts
= true,
1929 .lower_bitfield_reverse
= true,
1930 .lower_bit_count
= true,
1931 .lower_pack_unorm_2x16
= true,
1932 .lower_pack_snorm_2x16
= true,
1933 .lower_pack_unorm_4x8
= true,
1934 .lower_pack_snorm_4x8
= true,
1935 .lower_unpack_unorm_4x8
= true,
1936 .lower_unpack_snorm_4x8
= true,
1938 .lower_find_lsb
= true,
1940 .lower_flrp32
= true,
1943 .lower_fsqrt
= true,
1944 .lower_ifind_msb
= true,
1945 .lower_ldexp
= true,
1946 .lower_mul_high
= true,
1947 .native_integers
= true,
1953 count_nir_instrs(nir_shader
*nir
)
1956 nir_foreach_function(function
, nir
) {
1957 if (!function
->impl
)
1959 nir_foreach_block(block
, function
->impl
) {
1960 nir_foreach_instr(instr
, block
)
1969 * When demoting a shader down to single-threaded, removes the THRSW
1970 * instructions (one will still be inserted at v3d_vir_to_qpu() for the
1974 vir_remove_thrsw(struct v3d_compile
*c
)
1976 vir_for_each_block(block
, c
) {
1977 vir_for_each_inst_safe(inst
, block
) {
1978 if (inst
->qpu
.sig
.thrsw
)
1979 vir_remove_instruction(c
, inst
);
1983 c
->last_thrsw
= NULL
;
1987 vir_emit_last_thrsw(struct v3d_compile
*c
)
1989 /* On V3D before 4.1, we need a TMU op to be outstanding when thread
1990 * switching, so disable threads if we didn't do any TMU ops (each of
1991 * which would have emitted a THRSW).
1993 if (!c
->last_thrsw_at_top_level
&& c
->devinfo
->ver
< 41) {
1996 vir_remove_thrsw(c
);
2000 /* If we're threaded and the last THRSW was in conditional code, then
2001 * we need to emit another one so that we can flag it as the last
2004 if (c
->last_thrsw
&& !c
->last_thrsw_at_top_level
) {
2005 assert(c
->devinfo
->ver
>= 41);
2009 /* If we're threaded, then we need to mark the last THRSW instruction
2010 * so we can emit a pair of them at QPU emit time.
2012 * For V3D 4.x, we can spawn the non-fragment shaders already in the
2013 * post-last-THRSW state, so we can skip this.
2015 if (!c
->last_thrsw
&& c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
2016 assert(c
->devinfo
->ver
>= 41);
2021 c
->last_thrsw
->is_last_thrsw
= true;
2024 /* There's a flag in the shader for "center W is needed for reasons other than
2025 * non-centroid varyings", so we just walk the program after VIR optimization
2026 * to see if it's used. It should be harmless to set even if we only use
2027 * center W for varyings.
2030 vir_check_payload_w(struct v3d_compile
*c
)
2032 if (c
->s
->info
.stage
!= MESA_SHADER_FRAGMENT
)
2035 vir_for_each_inst_inorder(inst
, c
) {
2036 for (int i
= 0; i
< vir_get_nsrc(inst
); i
++) {
2037 if (inst
->src
[i
].file
== QFILE_REG
&&
2038 inst
->src
[i
].index
== 0) {
2039 c
->uses_center_w
= true;
2048 v3d_nir_to_vir(struct v3d_compile
*c
)
2050 if (V3D_DEBUG
& (V3D_DEBUG_NIR
|
2051 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2052 fprintf(stderr
, "%s prog %d/%d NIR:\n",
2053 vir_get_stage_name(c
),
2054 c
->program_id
, c
->variant_id
);
2055 nir_print_shader(c
->s
, stderr
);
2060 /* Emit the last THRSW before STVPM and TLB writes. */
2061 vir_emit_last_thrsw(c
);
2063 switch (c
->s
->info
.stage
) {
2064 case MESA_SHADER_FRAGMENT
:
2067 case MESA_SHADER_VERTEX
:
2071 unreachable("bad stage");
2074 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
2075 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2076 fprintf(stderr
, "%s prog %d/%d pre-opt VIR:\n",
2077 vir_get_stage_name(c
),
2078 c
->program_id
, c
->variant_id
);
2080 fprintf(stderr
, "\n");
2084 vir_lower_uniforms(c
);
2086 vir_check_payload_w(c
);
2088 /* XXX: vir_schedule_instructions(c); */
2090 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
2091 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2092 fprintf(stderr
, "%s prog %d/%d VIR:\n",
2093 vir_get_stage_name(c
),
2094 c
->program_id
, c
->variant_id
);
2096 fprintf(stderr
, "\n");
2099 /* Attempt to allocate registers for the temporaries. If we fail,
2100 * reduce thread count and try again.
2102 int min_threads
= (c
->devinfo
->ver
>= 41) ? 2 : 1;
2103 struct qpu_reg
*temp_registers
;
2106 temp_registers
= v3d_register_allocate(c
, &spilled
);
2113 if (c
->threads
== min_threads
) {
2114 fprintf(stderr
, "Failed to register allocate at %d threads:\n",
2123 if (c
->threads
== 1)
2124 vir_remove_thrsw(c
);
2127 v3d_vir_to_qpu(c
, temp_registers
);