2 * Copyright (C) 2020 Collabora Ltd.
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * Authors (Collabora):
24 * Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
27 #include "main/mtypes.h"
28 #include "compiler/glsl/glsl_to_nir.h"
29 #include "compiler/nir_types.h"
30 #include "compiler/nir/nir_builder.h"
32 #include "disassemble.h"
33 #include "bifrost_compile.h"
34 #include "bifrost_nir.h"
36 #include "bi_quirks.h"
39 static bi_block
*emit_cf_list(bi_context
*ctx
, struct exec_list
*list
);
40 static bi_instruction
*bi_emit_branch(bi_context
*ctx
);
41 static void bi_schedule_barrier(bi_context
*ctx
);
44 emit_jump(bi_context
*ctx
, nir_jump_instr
*instr
)
46 bi_instruction
*branch
= bi_emit_branch(ctx
);
48 switch (instr
->type
) {
50 branch
->branch
.target
= ctx
->break_block
;
52 case nir_jump_continue
:
53 branch
->branch
.target
= ctx
->continue_block
;
56 unreachable("Unhandled jump type");
59 pan_block_add_successor(&ctx
->current_block
->base
, &branch
->branch
.target
->base
);
62 /* Gets a bytemask for a complete vecN write */
64 bi_mask_for_channels_32(unsigned i
)
66 return (1 << (4 * i
)) - 1;
70 bi_load(enum bi_class T
, nir_intrinsic_instr
*instr
)
72 bi_instruction load
= {
74 .writemask
= bi_mask_for_channels_32(instr
->num_components
),
75 .src
= { BIR_INDEX_CONSTANT
},
76 .src_types
= { nir_type_uint32
},
77 .constant
= { .u64
= nir_intrinsic_base(instr
) },
80 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
83 load
.dest
= bir_dest_index(&instr
->dest
);
85 if (info
->has_dest
&& info
->index_map
[NIR_INTRINSIC_TYPE
] > 0)
86 load
.dest_type
= nir_intrinsic_type(instr
);
88 nir_src
*offset
= nir_get_io_offset_src(instr
);
90 if (nir_src_is_const(*offset
))
91 load
.constant
.u64
+= nir_src_as_uint(*offset
);
93 load
.src
[0] = bir_src_index(offset
);
99 bi_emit_ld_vary(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
101 bi_instruction ins
= bi_load(BI_LOAD_VAR
, instr
);
102 ins
.load_vary
.interp_mode
= BIFROST_INTERP_DEFAULT
; /* TODO */
103 ins
.load_vary
.reuse
= false; /* TODO */
104 ins
.load_vary
.flat
= instr
->intrinsic
!= nir_intrinsic_load_interpolated_input
;
105 ins
.dest_type
= nir_type_float
| nir_dest_bit_size(instr
->dest
);
107 if (nir_src_is_const(*nir_get_io_offset_src(instr
))) {
108 /* Zero it out for direct */
109 ins
.src
[1] = BIR_INDEX_ZERO
;
111 /* R61 contains sample mask stuff, TODO RA XXX */
112 ins
.src
[1] = BIR_INDEX_REGISTER
| 61;
119 bi_emit_frag_out(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
121 if (!ctx
->emitted_atest
) {
122 bi_instruction ins
= {
125 BIR_INDEX_REGISTER
| 60 /* TODO: RA */,
126 bir_src_index(&instr
->src
[0])
134 { 3, 0 } /* swizzle out the alpha */
136 .dest
= BIR_INDEX_REGISTER
| 60 /* TODO: RA */,
137 .dest_type
= nir_type_uint32
,
142 bi_schedule_barrier(ctx
);
143 ctx
->emitted_atest
= true;
146 bi_instruction blend
= {
148 .blend_location
= nir_intrinsic_base(instr
),
150 bir_src_index(&instr
->src
[0]),
151 BIR_INDEX_REGISTER
| 60 /* Can this be arbitrary? */,
161 .dest
= BIR_INDEX_REGISTER
| 48 /* Looks like magic */,
162 .dest_type
= nir_type_uint32
,
167 bi_schedule_barrier(ctx
);
170 static bi_instruction
171 bi_load_with_r61(enum bi_class T
, nir_intrinsic_instr
*instr
)
173 bi_instruction ld
= bi_load(T
, instr
);
174 ld
.src
[1] = BIR_INDEX_REGISTER
| 61; /* TODO: RA */
175 ld
.src
[2] = BIR_INDEX_REGISTER
| 62;
177 ld
.src_types
[1] = nir_type_uint32
;
178 ld
.src_types
[2] = nir_type_uint32
;
179 ld
.src_types
[3] = nir_intrinsic_type(instr
);
184 bi_emit_st_vary(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
186 bi_instruction address
= bi_load_with_r61(BI_LOAD_VAR_ADDRESS
, instr
);
187 address
.dest
= bi_make_temp(ctx
);
188 address
.dest_type
= nir_type_uint32
;
189 address
.writemask
= (1 << 12) - 1;
191 unsigned nr
= nir_intrinsic_src_components(instr
, 0);
192 assert(nir_intrinsic_write_mask(instr
) == ((1 << nr
) - 1));
194 bi_instruction st
= {
195 .type
= BI_STORE_VAR
,
197 bir_src_index(&instr
->src
[0]),
198 address
.dest
, address
.dest
, address
.dest
,
202 nir_type_uint32
, nir_type_uint32
, nir_type_uint32
,
208 .store_channels
= nr
,
211 for (unsigned i
= 0; i
< nr
; ++i
)
212 st
.swizzle
[0][i
] = i
;
214 bi_emit(ctx
, address
);
219 bi_emit_ld_uniform(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
221 bi_instruction ld
= bi_load(BI_LOAD_UNIFORM
, instr
);
222 ld
.src
[1] = BIR_INDEX_ZERO
; /* TODO: UBO index */
224 /* TODO: Indirect access, since we need to multiply by the element
225 * size. I believe we can get this lowering automatically via
226 * nir_lower_io (as mul instructions) with the proper options, but this
228 assert(ld
.src
[0] & BIR_INDEX_CONSTANT
);
229 ld
.constant
.u64
+= ctx
->sysvals
.sysval_count
;
230 ld
.constant
.u64
*= 16;
236 bi_emit_sysval(bi_context
*ctx
, nir_instr
*instr
,
237 unsigned nr_components
, unsigned offset
)
241 /* Figure out which uniform this is */
242 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
243 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
245 /* Sysvals are prefix uniforms */
246 unsigned uniform
= ((uintptr_t) val
) - 1;
248 /* Emit the read itself -- this is never indirect */
250 bi_instruction load
= {
251 .type
= BI_LOAD_UNIFORM
,
252 .writemask
= (1 << (nr_components
* 4)) - 1,
253 .src
= { BIR_INDEX_CONSTANT
, BIR_INDEX_ZERO
},
254 .src_types
= { nir_type_uint32
, nir_type_uint32
},
255 .constant
= { (uniform
* 16) + offset
},
256 .dest
= bir_dest_index(&nir_dest
),
257 .dest_type
= nir_type_uint32
, /* TODO */
264 emit_intrinsic(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
267 switch (instr
->intrinsic
) {
268 case nir_intrinsic_load_barycentric_pixel
:
271 case nir_intrinsic_load_interpolated_input
:
272 case nir_intrinsic_load_input
:
273 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
274 bi_emit_ld_vary(ctx
, instr
);
275 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
276 bi_emit(ctx
, bi_load_with_r61(BI_LOAD_ATTR
, instr
));
278 unreachable("Unsupported shader stage");
282 case nir_intrinsic_store_output
:
283 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
284 bi_emit_frag_out(ctx
, instr
);
285 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
286 bi_emit_st_vary(ctx
, instr
);
288 unreachable("Unsupported shader stage");
291 case nir_intrinsic_load_uniform
:
292 bi_emit_ld_uniform(ctx
, instr
);
295 case nir_intrinsic_load_ssbo_address
:
296 bi_emit_sysval(ctx
, &instr
->instr
, 1, 0);
299 case nir_intrinsic_get_buffer_size
:
300 bi_emit_sysval(ctx
, &instr
->instr
, 1, 8);
303 case nir_intrinsic_load_viewport_scale
:
304 case nir_intrinsic_load_viewport_offset
:
305 case nir_intrinsic_load_num_work_groups
:
306 case nir_intrinsic_load_sampler_lod_parameters_pan
:
307 bi_emit_sysval(ctx
, &instr
->instr
, 3, 0);
317 emit_load_const(bi_context
*ctx
, nir_load_const_instr
*instr
)
319 /* Make sure we've been lowered */
320 assert(instr
->def
.num_components
== 1);
322 bi_instruction move
= {
324 .dest
= bir_ssa_index(&instr
->def
),
325 .dest_type
= instr
->def
.bit_size
| nir_type_uint
,
326 .writemask
= (1 << (instr
->def
.bit_size
/ 8)) - 1,
331 instr
->def
.bit_size
| nir_type_uint
,
334 .u64
= nir_const_value_as_uint(instr
->value
[0], instr
->def
.bit_size
)
341 #define BI_CASE_CMP(op) \
347 bi_class_for_nir_alu(nir_op op
)
357 BI_CASE_CMP(nir_op_flt
)
358 BI_CASE_CMP(nir_op_fge
)
359 BI_CASE_CMP(nir_op_feq
)
360 BI_CASE_CMP(nir_op_fne
)
361 BI_CASE_CMP(nir_op_ilt
)
362 BI_CASE_CMP(nir_op_ige
)
363 BI_CASE_CMP(nir_op_ieq
)
364 BI_CASE_CMP(nir_op_ine
)
405 unreachable("should've been lowered");
426 case nir_op_fround_even
:
437 unreachable("Unknown ALU op");
441 /* Gets a bi_cond for a given NIR comparison opcode. In soft mode, it will
442 * return BI_COND_ALWAYS as a sentinel if it fails to do so (when used for
443 * optimizations). Otherwise it will bail (when used for primary code
447 bi_cond_for_nir(nir_op op
, bool soft
)
450 BI_CASE_CMP(nir_op_flt
)
451 BI_CASE_CMP(nir_op_ilt
)
454 BI_CASE_CMP(nir_op_fge
)
455 BI_CASE_CMP(nir_op_ige
)
458 BI_CASE_CMP(nir_op_feq
)
459 BI_CASE_CMP(nir_op_ieq
)
462 BI_CASE_CMP(nir_op_fne
)
463 BI_CASE_CMP(nir_op_ine
)
467 return BI_COND_ALWAYS
;
469 unreachable("Invalid compare");
474 bi_copy_src(bi_instruction
*alu
, nir_alu_instr
*instr
, unsigned i
, unsigned to
,
475 unsigned *constants_left
, unsigned *constant_shift
)
477 unsigned bits
= nir_src_bit_size(instr
->src
[i
].src
);
478 unsigned dest_bits
= nir_dest_bit_size(instr
->dest
.dest
);
480 alu
->src_types
[to
] = nir_op_infos
[instr
->op
].input_types
[i
]
483 /* Try to inline a constant */
484 if (nir_src_is_const(instr
->src
[i
].src
) && *constants_left
&& (dest_bits
== bits
)) {
485 uint64_t mask
= (1ull << dest_bits
) - 1;
486 uint64_t cons
= nir_src_as_uint(instr
->src
[i
].src
);
488 /* Try to reuse a constant */
489 for (unsigned i
= 0; i
< (*constant_shift
); i
+= dest_bits
) {
490 if (((alu
->constant
.u64
>> i
) & mask
) == cons
) {
491 alu
->src
[to
] = BIR_INDEX_CONSTANT
| i
;
496 alu
->constant
.u64
|= cons
<< *constant_shift
;
497 alu
->src
[to
] = BIR_INDEX_CONSTANT
| (*constant_shift
);
499 (*constant_shift
) += MAX2(dest_bits
, 32); /* lo/hi */
503 alu
->src
[to
] = bir_src_index(&instr
->src
[i
].src
);
505 /* We assert scalarization above */
506 alu
->swizzle
[to
][0] = instr
->src
[i
].swizzle
[0];
510 bi_fuse_csel_cond(bi_instruction
*csel
, nir_alu_src cond
,
511 unsigned *constants_left
, unsigned *constant_shift
)
513 /* Bail for vector weirdness */
514 if (cond
.swizzle
[0] != 0)
517 if (!cond
.src
.is_ssa
)
520 nir_ssa_def
*def
= cond
.src
.ssa
;
521 nir_instr
*parent
= def
->parent_instr
;
523 if (parent
->type
!= nir_instr_type_alu
)
526 nir_alu_instr
*alu
= nir_instr_as_alu(parent
);
528 /* Try to match a condition */
529 enum bi_cond bcond
= bi_cond_for_nir(alu
->op
, true);
531 if (bcond
== BI_COND_ALWAYS
)
534 /* We found one, let's fuse it in */
535 csel
->csel_cond
= bcond
;
536 bi_copy_src(csel
, alu
, 0, 0, constants_left
, constant_shift
);
537 bi_copy_src(csel
, alu
, 1, 1, constants_left
, constant_shift
);
541 emit_alu(bi_context
*ctx
, nir_alu_instr
*instr
)
543 /* Try some special functions */
546 bi_emit_fexp2(ctx
, instr
);
549 bi_emit_flog2(ctx
, instr
);
555 /* Otherwise, assume it's something we can handle normally */
556 bi_instruction alu
= {
557 .type
= bi_class_for_nir_alu(instr
->op
),
558 .dest
= bir_dest_index(&instr
->dest
.dest
),
559 .dest_type
= nir_op_infos
[instr
->op
].output_type
560 | nir_dest_bit_size(instr
->dest
.dest
),
563 /* TODO: Implement lowering of special functions for older Bifrost */
564 assert((alu
.type
!= BI_SPECIAL
) || !(ctx
->quirks
& BIFROST_NO_FAST_OP
));
566 if (instr
->dest
.dest
.is_ssa
) {
567 /* Construct a writemask */
568 unsigned bits_per_comp
= instr
->dest
.dest
.ssa
.bit_size
;
569 unsigned comps
= instr
->dest
.dest
.ssa
.num_components
;
571 if (alu
.type
!= BI_COMBINE
)
574 unsigned bits
= bits_per_comp
* comps
;
575 unsigned bytes
= bits
/ 8;
576 alu
.writemask
= (1 << bytes
) - 1;
578 unsigned comp_mask
= instr
->dest
.write_mask
;
580 alu
.writemask
= pan_to_bytemask(nir_dest_bit_size(instr
->dest
.dest
),
584 /* We inline constants as we go. This tracks how many constants have
585 * been inlined, since we're limited to 64-bits of constants per
588 unsigned dest_bits
= nir_dest_bit_size(instr
->dest
.dest
);
589 unsigned constants_left
= (64 / dest_bits
);
590 unsigned constant_shift
= 0;
592 if (alu
.type
== BI_COMBINE
)
597 unsigned num_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
598 assert(num_inputs
<= ARRAY_SIZE(alu
.src
));
600 for (unsigned i
= 0; i
< num_inputs
; ++i
) {
603 if (i
&& alu
.type
== BI_CSEL
)
606 bi_copy_src(&alu
, instr
, i
, i
+ f
, &constants_left
, &constant_shift
);
609 /* Op-specific fixup */
612 alu
.src
[2] = BIR_INDEX_ZERO
; /* FMA */
613 alu
.src_types
[2] = alu
.src_types
[1];
616 alu
.outmod
= BIFROST_SAT
; /* FMOV */
619 alu
.src_neg
[0] = true; /* FMOV */
622 alu
.src_abs
[0] = true; /* FMOV */
625 alu
.src_neg
[1] = true; /* FADD */
630 alu
.op
.minmax
= BI_MINMAX_MAX
; /* MINMAX */
633 alu
.op
.special
= BI_SPECIAL_FRCP
;
636 alu
.op
.special
= BI_SPECIAL_FRSQ
;
638 BI_CASE_CMP(nir_op_flt
)
639 BI_CASE_CMP(nir_op_ilt
)
640 BI_CASE_CMP(nir_op_fge
)
641 BI_CASE_CMP(nir_op_ige
)
642 BI_CASE_CMP(nir_op_feq
)
643 BI_CASE_CMP(nir_op_ieq
)
644 BI_CASE_CMP(nir_op_fne
)
645 BI_CASE_CMP(nir_op_ine
)
646 alu
.op
.compare
= bi_cond_for_nir(instr
->op
, false);
648 case nir_op_fround_even
:
649 alu
.op
.round
= BI_ROUND_MODE
;
650 alu
.roundmode
= BIFROST_RTE
;
653 alu
.op
.round
= BI_ROUND_MODE
;
654 alu
.roundmode
= BIFROST_RTP
;
657 alu
.op
.round
= BI_ROUND_MODE
;
658 alu
.roundmode
= BIFROST_RTN
;
661 alu
.op
.round
= BI_ROUND_MODE
;
662 alu
.roundmode
= BIFROST_RTZ
;
668 if (alu
.type
== BI_CSEL
) {
669 /* Default to csel3 */
670 alu
.csel_cond
= BI_COND_NE
;
671 alu
.src
[1] = BIR_INDEX_ZERO
;
672 alu
.src_types
[1] = alu
.src_types
[0];
674 bi_fuse_csel_cond(&alu
, instr
->src
[0],
675 &constants_left
, &constant_shift
);
681 /* TEX_COMPACT instructions assume normal 2D f32 operation but are more
682 * space-efficient and with simpler RA/scheduling requirements*/
685 emit_tex_compact(bi_context
*ctx
, nir_tex_instr
*instr
)
687 bi_instruction tex
= {
689 .op
= { .texture
= BI_TEX_COMPACT
},
690 .dest
= bir_dest_index(&instr
->dest
),
691 .dest_type
= instr
->dest_type
,
692 .src_types
= { nir_type_float32
, nir_type_float32
},
693 .writemask
= instr
->dest_type
== nir_type_float32
?
697 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
698 int index
= bir_src_index(&instr
->src
[i
].src
);
699 assert (instr
->src
[i
].src_type
== nir_tex_src_coord
);
703 tex
.swizzle
[0][0] = 0;
704 tex
.swizzle
[1][0] = 1;
711 emit_tex_full(bi_context
*ctx
, nir_tex_instr
*instr
)
717 emit_tex(bi_context
*ctx
, nir_tex_instr
*instr
)
719 nir_alu_type base
= nir_alu_type_get_base_type(instr
->dest_type
);
720 unsigned sz
= nir_dest_bit_size(instr
->dest
);
721 instr
->dest_type
= base
| sz
;
723 bool is_normal
= instr
->op
== nir_texop_tex
;
724 bool is_2d
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
||
725 instr
->sampler_dim
== GLSL_SAMPLER_DIM_EXTERNAL
;
726 bool is_f
= base
== nir_type_float
&& (sz
== 16 || sz
== 32);
728 bool is_compact
= is_normal
&& is_2d
&& is_f
&& !instr
->is_shadow
;
731 emit_tex_compact(ctx
, instr
);
733 emit_tex_full(ctx
, instr
);
737 emit_instr(bi_context
*ctx
, struct nir_instr
*instr
)
739 switch (instr
->type
) {
740 case nir_instr_type_load_const
:
741 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
744 case nir_instr_type_intrinsic
:
745 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
748 case nir_instr_type_alu
:
749 emit_alu(ctx
, nir_instr_as_alu(instr
));
752 case nir_instr_type_tex
:
753 emit_tex(ctx
, nir_instr_as_tex(instr
));
756 case nir_instr_type_jump
:
757 emit_jump(ctx
, nir_instr_as_jump(instr
));
760 case nir_instr_type_ssa_undef
:
765 unreachable("Unhandled instruction type");
773 create_empty_block(bi_context
*ctx
)
775 bi_block
*blk
= rzalloc(ctx
, bi_block
);
777 blk
->base
.predecessors
= _mesa_set_create(blk
,
779 _mesa_key_pointer_equal
);
781 blk
->base
.name
= ctx
->block_name_count
++;
787 bi_schedule_barrier(bi_context
*ctx
)
789 bi_block
*temp
= ctx
->after_block
;
790 ctx
->after_block
= create_empty_block(ctx
);
791 list_addtail(&ctx
->after_block
->base
.link
, &ctx
->blocks
);
792 list_inithead(&ctx
->after_block
->base
.instructions
);
793 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->after_block
->base
);
794 ctx
->current_block
= ctx
->after_block
;
795 ctx
->after_block
= temp
;
799 emit_block(bi_context
*ctx
, nir_block
*block
)
801 if (ctx
->after_block
) {
802 ctx
->current_block
= ctx
->after_block
;
803 ctx
->after_block
= NULL
;
805 ctx
->current_block
= create_empty_block(ctx
);
808 list_addtail(&ctx
->current_block
->base
.link
, &ctx
->blocks
);
809 list_inithead(&ctx
->current_block
->base
.instructions
);
811 nir_foreach_instr(instr
, block
) {
812 emit_instr(ctx
, instr
);
813 ++ctx
->instruction_count
;
816 return ctx
->current_block
;
819 /* Emits an unconditional branch to the end of the current block, returning a
820 * pointer so the user can fill in details */
822 static bi_instruction
*
823 bi_emit_branch(bi_context
*ctx
)
825 bi_instruction branch
= {
828 .cond
= BI_COND_ALWAYS
832 return bi_emit(ctx
, branch
);
835 /* Sets a condition for a branch by examing the NIR condition. If we're
836 * familiar with the condition, we unwrap it to fold it into the branch
837 * instruction. Otherwise, we consume the condition directly. We
838 * generally use 1-bit booleans which allows us to use small types for
843 bi_set_branch_cond(bi_instruction
*branch
, nir_src
*cond
, bool invert
)
845 /* TODO: Try to unwrap instead of always bailing */
846 branch
->src
[0] = bir_src_index(cond
);
847 branch
->src
[1] = BIR_INDEX_ZERO
;
848 branch
->src_types
[0] = branch
->src_types
[1] = nir_type_uint16
;
849 branch
->branch
.cond
= invert
? BI_COND_EQ
: BI_COND_NE
;
853 emit_if(bi_context
*ctx
, nir_if
*nif
)
855 bi_block
*before_block
= ctx
->current_block
;
857 /* Speculatively emit the branch, but we can't fill it in until later */
858 bi_instruction
*then_branch
= bi_emit_branch(ctx
);
859 bi_set_branch_cond(then_branch
, &nif
->condition
, true);
861 /* Emit the two subblocks. */
862 bi_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
863 bi_block
*end_then_block
= ctx
->current_block
;
865 /* Emit a jump from the end of the then block to the end of the else */
866 bi_instruction
*then_exit
= bi_emit_branch(ctx
);
868 /* Emit second block, and check if it's empty */
870 int count_in
= ctx
->instruction_count
;
871 bi_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
872 bi_block
*end_else_block
= ctx
->current_block
;
873 ctx
->after_block
= create_empty_block(ctx
);
875 /* Now that we have the subblocks emitted, fix up the branches */
880 if (ctx
->instruction_count
== count_in
) {
881 /* The else block is empty, so don't emit an exit jump */
882 bi_remove_instruction(then_exit
);
883 then_branch
->branch
.target
= ctx
->after_block
;
885 then_branch
->branch
.target
= else_block
;
886 then_exit
->branch
.target
= ctx
->after_block
;
887 pan_block_add_successor(&end_then_block
->base
, &then_exit
->branch
.target
->base
);
890 /* Wire up the successors */
892 pan_block_add_successor(&before_block
->base
, &then_branch
->branch
.target
->base
); /* then_branch */
894 pan_block_add_successor(&before_block
->base
, &then_block
->base
); /* fallthrough */
895 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
); /* fallthrough */
899 emit_loop(bi_context
*ctx
, nir_loop
*nloop
)
901 /* Remember where we are */
902 bi_block
*start_block
= ctx
->current_block
;
904 bi_block
*saved_break
= ctx
->break_block
;
905 bi_block
*saved_continue
= ctx
->continue_block
;
907 ctx
->continue_block
= create_empty_block(ctx
);
908 ctx
->break_block
= create_empty_block(ctx
);
909 ctx
->after_block
= ctx
->continue_block
;
911 /* Emit the body itself */
912 emit_cf_list(ctx
, &nloop
->body
);
914 /* Branch back to loop back */
915 bi_instruction
*br_back
= bi_emit_branch(ctx
);
916 br_back
->branch
.target
= ctx
->continue_block
;
917 pan_block_add_successor(&start_block
->base
, &ctx
->continue_block
->base
);
918 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->continue_block
->base
);
920 ctx
->after_block
= ctx
->break_block
;
923 ctx
->break_block
= saved_break
;
924 ctx
->continue_block
= saved_continue
;
929 emit_cf_list(bi_context
*ctx
, struct exec_list
*list
)
931 bi_block
*start_block
= NULL
;
933 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
934 switch (node
->type
) {
935 case nir_cf_node_block
: {
936 bi_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
945 emit_if(ctx
, nir_cf_node_as_if(node
));
948 case nir_cf_node_loop
:
949 emit_loop(ctx
, nir_cf_node_as_loop(node
));
953 unreachable("Unknown control flow");
961 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
963 return glsl_count_attribute_slots(type
, false);
967 bi_optimize_nir(nir_shader
*nir
)
970 unsigned lower_flrp
= 16 | 32 | 64;
972 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
973 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
975 nir_lower_tex_options lower_tex_options
= {
976 .lower_txs_lod
= true,
978 .lower_tex_without_implicit_lod
= true,
982 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
983 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
984 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
989 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
990 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
992 NIR_PASS(progress
, nir
, nir_copy_prop
);
993 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
994 NIR_PASS(progress
, nir
, nir_opt_dce
);
995 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
996 NIR_PASS(progress
, nir
, nir_opt_cse
);
997 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
998 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
999 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
1001 if (lower_flrp
!= 0) {
1002 bool lower_flrp_progress
= false;
1003 NIR_PASS(lower_flrp_progress
,
1007 false /* always_precise */,
1008 nir
->options
->lower_ffma
);
1009 if (lower_flrp_progress
) {
1010 NIR_PASS(progress
, nir
,
1011 nir_opt_constant_folding
);
1015 /* Nothing should rematerialize any flrps, so we only
1016 * need to do this lowering once.
1021 NIR_PASS(progress
, nir
, nir_opt_undef
);
1022 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
1024 nir_var_shader_out
|
1025 nir_var_function_temp
);
1028 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
1029 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
1030 NIR_PASS(progress
, nir
, bifrost_nir_lower_algebraic_late
);
1031 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
1032 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
1034 /* Take us out of SSA */
1035 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
1036 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
1037 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
1041 bifrost_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, unsigned product_id
)
1043 bi_context
*ctx
= rzalloc(NULL
, bi_context
);
1045 ctx
->stage
= nir
->info
.stage
;
1046 ctx
->quirks
= bifrost_get_quirks(product_id
);
1047 list_inithead(&ctx
->blocks
);
1049 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
1050 * (so we don't accidentally duplicate the epilogue since mesa/st has
1051 * messed with our I/O quite a bit already) */
1053 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
1055 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1056 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
1057 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
1060 NIR_PASS_V(nir
, nir_split_var_copies
);
1061 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
1062 NIR_PASS_V(nir
, nir_lower_var_copies
);
1063 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
1064 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
1065 NIR_PASS_V(nir
, nir_lower_ssbo
);
1067 bi_optimize_nir(nir
);
1068 nir_print_shader(nir
, stdout
);
1070 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
1071 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
1072 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
1074 nir_foreach_function(func
, nir
) {
1078 ctx
->impl
= func
->impl
;
1079 emit_cf_list(ctx
, &func
->impl
->body
);
1080 break; /* TODO: Multi-function shaders */
1083 bi_foreach_block(ctx
, _block
) {
1084 bi_block
*block
= (bi_block
*) _block
;
1085 bi_lower_combine(ctx
, block
);
1088 bool progress
= false;
1093 bi_foreach_block(ctx
, _block
) {
1094 bi_block
*block
= (bi_block
*) _block
;
1095 progress
|= bi_opt_dead_code_eliminate(ctx
, block
);
1099 bi_print_shader(ctx
, stdout
);
1101 bi_register_allocate(ctx
);
1102 bi_print_shader(ctx
, stdout
);
1103 bi_pack(ctx
, &program
->compiled
);
1104 disassemble_bifrost(stdout
, program
->compiled
.data
, program
->compiled
.size
, true);