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
);
63 bi_load(enum bi_class T
, nir_intrinsic_instr
*instr
)
65 bi_instruction load
= {
67 .vector_channels
= instr
->num_components
,
68 .src
= { BIR_INDEX_CONSTANT
},
69 .src_types
= { nir_type_uint32
},
70 .constant
= { .u64
= nir_intrinsic_base(instr
) },
73 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
76 load
.dest
= pan_dest_index(&instr
->dest
);
78 if (info
->has_dest
&& info
->index_map
[NIR_INTRINSIC_TYPE
] > 0)
79 load
.dest_type
= nir_intrinsic_type(instr
);
81 nir_src
*offset
= nir_get_io_offset_src(instr
);
83 if (nir_src_is_const(*offset
))
84 load
.constant
.u64
+= nir_src_as_uint(*offset
);
86 load
.src
[0] = pan_src_index(offset
);
92 bi_emit_ld_vary(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
94 bi_instruction ins
= bi_load(BI_LOAD_VAR
, instr
);
95 ins
.load_vary
.interp_mode
= BIFROST_INTERP_DEFAULT
; /* TODO */
96 ins
.load_vary
.reuse
= false; /* TODO */
97 ins
.load_vary
.flat
= instr
->intrinsic
!= nir_intrinsic_load_interpolated_input
;
98 ins
.dest_type
= nir_type_float
| nir_dest_bit_size(instr
->dest
);
100 if (nir_src_is_const(*nir_get_io_offset_src(instr
))) {
101 /* Zero it out for direct */
102 ins
.src
[1] = BIR_INDEX_ZERO
;
104 /* R61 contains sample mask stuff, TODO RA XXX */
105 ins
.src
[1] = BIR_INDEX_REGISTER
| 61;
112 bi_emit_frag_out(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
114 if (!ctx
->emitted_atest
) {
115 bi_instruction ins
= {
118 BIR_INDEX_REGISTER
| 60 /* TODO: RA */,
119 pan_src_index(&instr
->src
[0])
123 nir_intrinsic_type(instr
)
127 { 3, 0 } /* swizzle out the alpha */
129 .dest
= BIR_INDEX_REGISTER
| 60 /* TODO: RA */,
130 .dest_type
= nir_type_uint32
,
134 bi_schedule_barrier(ctx
);
135 ctx
->emitted_atest
= true;
138 bi_instruction blend
= {
140 .blend_location
= nir_intrinsic_base(instr
),
142 pan_src_index(&instr
->src
[0]),
143 BIR_INDEX_REGISTER
| 60 /* Can this be arbitrary? */,
146 nir_intrinsic_type(instr
),
153 .dest
= BIR_INDEX_REGISTER
| 48 /* Looks like magic */,
154 .dest_type
= nir_type_uint32
,
158 assert(blend
.blend_location
< BIFROST_MAX_RENDER_TARGET_COUNT
);
159 assert(ctx
->blend_types
);
160 assert(blend
.src_types
[0]);
161 ctx
->blend_types
[blend
.blend_location
] = blend
.src_types
[0];
164 bi_schedule_barrier(ctx
);
167 static bi_instruction
168 bi_load_with_r61(enum bi_class T
, nir_intrinsic_instr
*instr
)
170 bi_instruction ld
= bi_load(T
, instr
);
171 ld
.src
[1] = BIR_INDEX_REGISTER
| 61; /* TODO: RA */
172 ld
.src
[2] = BIR_INDEX_REGISTER
| 62;
174 ld
.src_types
[1] = nir_type_uint32
;
175 ld
.src_types
[2] = nir_type_uint32
;
176 ld
.src_types
[3] = nir_intrinsic_type(instr
);
181 bi_emit_st_vary(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
183 bi_instruction address
= bi_load_with_r61(BI_LOAD_VAR_ADDRESS
, instr
);
184 address
.dest
= bi_make_temp(ctx
);
185 address
.dest_type
= nir_type_uint32
;
186 address
.vector_channels
= 3;
188 unsigned nr
= nir_intrinsic_src_components(instr
, 0);
189 assert(nir_intrinsic_write_mask(instr
) == ((1 << nr
) - 1));
191 bi_instruction st
= {
192 .type
= BI_STORE_VAR
,
194 pan_src_index(&instr
->src
[0]),
195 address
.dest
, address
.dest
, address
.dest
,
199 nir_type_uint32
, nir_type_uint32
, nir_type_uint32
,
205 .vector_channels
= nr
,
208 for (unsigned i
= 0; i
< nr
; ++i
)
209 st
.swizzle
[0][i
] = i
;
211 bi_emit(ctx
, address
);
216 bi_emit_ld_uniform(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
218 bi_instruction ld
= bi_load(BI_LOAD_UNIFORM
, instr
);
219 ld
.src
[1] = BIR_INDEX_ZERO
; /* TODO: UBO index */
221 /* TODO: Indirect access, since we need to multiply by the element
222 * size. I believe we can get this lowering automatically via
223 * nir_lower_io (as mul instructions) with the proper options, but this
225 assert(ld
.src
[0] & BIR_INDEX_CONSTANT
);
226 ld
.constant
.u64
+= ctx
->sysvals
.sysval_count
;
227 ld
.constant
.u64
*= 16;
233 bi_emit_sysval(bi_context
*ctx
, nir_instr
*instr
,
234 unsigned nr_components
, unsigned offset
)
238 /* Figure out which uniform this is */
239 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
240 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
242 /* Sysvals are prefix uniforms */
243 unsigned uniform
= ((uintptr_t) val
) - 1;
245 /* Emit the read itself -- this is never indirect */
247 bi_instruction load
= {
248 .type
= BI_LOAD_UNIFORM
,
249 .vector_channels
= nr_components
,
250 .src
= { BIR_INDEX_CONSTANT
, BIR_INDEX_ZERO
},
251 .src_types
= { nir_type_uint32
, nir_type_uint32
},
252 .constant
= { (uniform
* 16) + offset
},
253 .dest
= pan_dest_index(&nir_dest
),
254 .dest_type
= nir_type_uint32
, /* TODO */
261 emit_intrinsic(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
264 switch (instr
->intrinsic
) {
265 case nir_intrinsic_load_barycentric_pixel
:
268 case nir_intrinsic_load_interpolated_input
:
269 case nir_intrinsic_load_input
:
270 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
271 bi_emit_ld_vary(ctx
, instr
);
272 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
273 bi_emit(ctx
, bi_load_with_r61(BI_LOAD_ATTR
, instr
));
275 unreachable("Unsupported shader stage");
279 case nir_intrinsic_store_output
:
280 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
281 bi_emit_frag_out(ctx
, instr
);
282 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
283 bi_emit_st_vary(ctx
, instr
);
285 unreachable("Unsupported shader stage");
288 case nir_intrinsic_load_uniform
:
289 bi_emit_ld_uniform(ctx
, instr
);
292 case nir_intrinsic_load_ssbo_address
:
293 bi_emit_sysval(ctx
, &instr
->instr
, 1, 0);
296 case nir_intrinsic_get_buffer_size
:
297 bi_emit_sysval(ctx
, &instr
->instr
, 1, 8);
300 case nir_intrinsic_load_viewport_scale
:
301 case nir_intrinsic_load_viewport_offset
:
302 case nir_intrinsic_load_num_work_groups
:
303 case nir_intrinsic_load_sampler_lod_parameters_pan
:
304 bi_emit_sysval(ctx
, &instr
->instr
, 3, 0);
314 emit_load_const(bi_context
*ctx
, nir_load_const_instr
*instr
)
316 /* Make sure we've been lowered */
317 assert(instr
->def
.num_components
== 1);
319 bi_instruction move
= {
321 .dest
= pan_ssa_index(&instr
->def
),
322 .dest_type
= instr
->def
.bit_size
| nir_type_uint
,
327 instr
->def
.bit_size
| nir_type_uint
,
330 .u64
= nir_const_value_as_uint(instr
->value
[0], instr
->def
.bit_size
)
337 #define BI_CASE_CMP(op) \
343 bi_class_for_nir_alu(nir_op op
)
358 BI_CASE_CMP(nir_op_flt
)
359 BI_CASE_CMP(nir_op_fge
)
360 BI_CASE_CMP(nir_op_feq
)
361 BI_CASE_CMP(nir_op_fne
)
362 BI_CASE_CMP(nir_op_ilt
)
363 BI_CASE_CMP(nir_op_ige
)
364 BI_CASE_CMP(nir_op_ieq
)
365 BI_CASE_CMP(nir_op_ine
)
406 unreachable("should've been lowered");
427 case nir_op_fround_even
:
438 unreachable("Unknown ALU op");
442 /* Gets a bi_cond for a given NIR comparison opcode. In soft mode, it will
443 * return BI_COND_ALWAYS as a sentinel if it fails to do so (when used for
444 * optimizations). Otherwise it will bail (when used for primary code
448 bi_cond_for_nir(nir_op op
, bool soft
)
451 BI_CASE_CMP(nir_op_flt
)
452 BI_CASE_CMP(nir_op_ilt
)
455 BI_CASE_CMP(nir_op_fge
)
456 BI_CASE_CMP(nir_op_ige
)
459 BI_CASE_CMP(nir_op_feq
)
460 BI_CASE_CMP(nir_op_ieq
)
463 BI_CASE_CMP(nir_op_fne
)
464 BI_CASE_CMP(nir_op_ine
)
468 return BI_COND_ALWAYS
;
470 unreachable("Invalid compare");
475 bi_copy_src(bi_instruction
*alu
, nir_alu_instr
*instr
, unsigned i
, unsigned to
,
476 unsigned *constants_left
, unsigned *constant_shift
, unsigned comps
)
478 unsigned bits
= nir_src_bit_size(instr
->src
[i
].src
);
479 unsigned dest_bits
= nir_dest_bit_size(instr
->dest
.dest
);
481 alu
->src_types
[to
] = nir_op_infos
[instr
->op
].input_types
[i
]
484 /* Try to inline a constant */
485 if (nir_src_is_const(instr
->src
[i
].src
) && *constants_left
&& (dest_bits
== bits
)) {
486 uint64_t mask
= (1ull << dest_bits
) - 1;
487 uint64_t cons
= nir_src_as_uint(instr
->src
[i
].src
);
489 /* Try to reuse a constant */
490 for (unsigned i
= 0; i
< (*constant_shift
); i
+= dest_bits
) {
491 if (((alu
->constant
.u64
>> i
) & mask
) == cons
) {
492 alu
->src
[to
] = BIR_INDEX_CONSTANT
| i
;
497 alu
->constant
.u64
|= cons
<< *constant_shift
;
498 alu
->src
[to
] = BIR_INDEX_CONSTANT
| (*constant_shift
);
500 (*constant_shift
) += MAX2(dest_bits
, 32); /* lo/hi */
504 alu
->src
[to
] = pan_src_index(&instr
->src
[i
].src
);
506 /* Copy swizzle for all vectored components, replicating last component
507 * to fill undersized */
509 unsigned vec
= alu
->type
== BI_COMBINE
? 1 :
510 MAX2(1, 32 / dest_bits
);
512 for (unsigned j
= 0; j
< vec
; ++j
)
513 alu
->swizzle
[to
][j
] = instr
->src
[i
].swizzle
[MIN2(j
, comps
- 1)];
517 bi_fuse_csel_cond(bi_instruction
*csel
, nir_alu_src cond
,
518 unsigned *constants_left
, unsigned *constant_shift
, unsigned comps
)
520 /* Bail for vector weirdness */
521 if (cond
.swizzle
[0] != 0)
524 if (!cond
.src
.is_ssa
)
527 nir_ssa_def
*def
= cond
.src
.ssa
;
528 nir_instr
*parent
= def
->parent_instr
;
530 if (parent
->type
!= nir_instr_type_alu
)
533 nir_alu_instr
*alu
= nir_instr_as_alu(parent
);
535 /* Try to match a condition */
536 enum bi_cond bcond
= bi_cond_for_nir(alu
->op
, true);
538 if (bcond
== BI_COND_ALWAYS
)
541 /* We found one, let's fuse it in */
543 bi_copy_src(csel
, alu
, 0, 0, constants_left
, constant_shift
, comps
);
544 bi_copy_src(csel
, alu
, 1, 1, constants_left
, constant_shift
, comps
);
548 emit_alu(bi_context
*ctx
, nir_alu_instr
*instr
)
550 /* Try some special functions */
553 bi_emit_fexp2(ctx
, instr
);
556 bi_emit_flog2(ctx
, instr
);
562 /* Otherwise, assume it's something we can handle normally */
563 bi_instruction alu
= {
564 .type
= bi_class_for_nir_alu(instr
->op
),
565 .dest
= pan_dest_index(&instr
->dest
.dest
),
566 .dest_type
= nir_op_infos
[instr
->op
].output_type
567 | nir_dest_bit_size(instr
->dest
.dest
),
570 /* TODO: Implement lowering of special functions for older Bifrost */
571 assert((alu
.type
!= BI_SPECIAL
) || !(ctx
->quirks
& BIFROST_NO_FAST_OP
));
573 unsigned comps
= nir_dest_num_components(instr
->dest
.dest
);
575 if (alu
.type
!= BI_COMBINE
)
576 assert(comps
<= MAX2(1, 32 / comps
));
578 if (!instr
->dest
.dest
.is_ssa
) {
579 for (unsigned i
= 0; i
< comps
; ++i
)
580 assert(instr
->dest
.write_mask
);
583 /* We inline constants as we go. This tracks how many constants have
584 * been inlined, since we're limited to 64-bits of constants per
587 unsigned dest_bits
= nir_dest_bit_size(instr
->dest
.dest
);
588 unsigned constants_left
= (64 / dest_bits
);
589 unsigned constant_shift
= 0;
591 if (alu
.type
== BI_COMBINE
)
596 unsigned num_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
597 assert(num_inputs
<= ARRAY_SIZE(alu
.src
));
599 for (unsigned i
= 0; i
< num_inputs
; ++i
) {
602 if (i
&& alu
.type
== BI_CSEL
)
605 bi_copy_src(&alu
, instr
, i
, i
+ f
, &constants_left
, &constant_shift
, comps
);
608 /* Op-specific fixup */
611 alu
.src
[2] = BIR_INDEX_ZERO
; /* FMA */
612 alu
.src_types
[2] = alu
.src_types
[1];
615 alu
.outmod
= BIFROST_SAT
; /* FMOV */
618 alu
.src_neg
[0] = true; /* FMOV */
621 alu
.src_abs
[0] = true; /* FMOV */
624 alu
.src_neg
[1] = true; /* FADD */
629 alu
.op
.minmax
= BI_MINMAX_MAX
; /* MINMAX */
632 alu
.op
.special
= BI_SPECIAL_FRCP
;
635 alu
.op
.special
= BI_SPECIAL_FRSQ
;
637 BI_CASE_CMP(nir_op_flt
)
638 BI_CASE_CMP(nir_op_ilt
)
639 BI_CASE_CMP(nir_op_fge
)
640 BI_CASE_CMP(nir_op_ige
)
641 BI_CASE_CMP(nir_op_feq
)
642 BI_CASE_CMP(nir_op_ieq
)
643 BI_CASE_CMP(nir_op_fne
)
644 BI_CASE_CMP(nir_op_ine
)
645 alu
.cond
= bi_cond_for_nir(instr
->op
, false);
647 case nir_op_fround_even
:
648 alu
.roundmode
= BIFROST_RTE
;
651 alu
.roundmode
= BIFROST_RTP
;
654 alu
.roundmode
= BIFROST_RTN
;
657 alu
.roundmode
= BIFROST_RTZ
;
660 alu
.op
.bitwise
= BI_BITWISE_AND
;
663 alu
.op
.bitwise
= BI_BITWISE_OR
;
666 alu
.op
.bitwise
= BI_BITWISE_XOR
;
672 if (alu
.type
== BI_CSEL
) {
673 /* Default to csel3 */
674 alu
.cond
= BI_COND_NE
;
675 alu
.src
[1] = BIR_INDEX_ZERO
;
676 alu
.src_types
[1] = alu
.src_types
[0];
678 bi_fuse_csel_cond(&alu
, instr
->src
[0],
679 &constants_left
, &constant_shift
, comps
);
680 } else if (alu
.type
== BI_BITWISE
) {
681 /* Implicit shift argument... at some point we should fold */
682 alu
.src
[2] = BIR_INDEX_ZERO
;
683 alu
.src_types
[2] = alu
.src_types
[1];
689 /* TEX_COMPACT instructions assume normal 2D f32 operation but are more
690 * space-efficient and with simpler RA/scheduling requirements*/
693 emit_tex_compact(bi_context
*ctx
, nir_tex_instr
*instr
)
695 bi_instruction tex
= {
697 .op
= { .texture
= BI_TEX_COMPACT
},
699 .texture_index
= instr
->texture_index
,
700 .sampler_index
= instr
->sampler_index
,
702 .dest
= pan_dest_index(&instr
->dest
),
703 .dest_type
= instr
->dest_type
,
704 .src_types
= { nir_type_float32
, nir_type_float32
},
708 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
709 int index
= pan_src_index(&instr
->src
[i
].src
);
710 assert (instr
->src
[i
].src_type
== nir_tex_src_coord
);
714 tex
.swizzle
[0][0] = 0;
715 tex
.swizzle
[1][0] = 1;
722 emit_tex_full(bi_context
*ctx
, nir_tex_instr
*instr
)
728 emit_tex(bi_context
*ctx
, nir_tex_instr
*instr
)
730 nir_alu_type base
= nir_alu_type_get_base_type(instr
->dest_type
);
731 unsigned sz
= nir_dest_bit_size(instr
->dest
);
732 instr
->dest_type
= base
| sz
;
734 bool is_normal
= instr
->op
== nir_texop_tex
;
735 bool is_2d
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
||
736 instr
->sampler_dim
== GLSL_SAMPLER_DIM_EXTERNAL
;
737 bool is_f
= base
== nir_type_float
&& (sz
== 16 || sz
== 32);
739 bool is_compact
= is_normal
&& is_2d
&& is_f
&& !instr
->is_shadow
;
742 emit_tex_compact(ctx
, instr
);
744 emit_tex_full(ctx
, instr
);
748 emit_instr(bi_context
*ctx
, struct nir_instr
*instr
)
750 switch (instr
->type
) {
751 case nir_instr_type_load_const
:
752 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
755 case nir_instr_type_intrinsic
:
756 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
759 case nir_instr_type_alu
:
760 emit_alu(ctx
, nir_instr_as_alu(instr
));
763 case nir_instr_type_tex
:
764 emit_tex(ctx
, nir_instr_as_tex(instr
));
767 case nir_instr_type_jump
:
768 emit_jump(ctx
, nir_instr_as_jump(instr
));
771 case nir_instr_type_ssa_undef
:
776 unreachable("Unhandled instruction type");
784 create_empty_block(bi_context
*ctx
)
786 bi_block
*blk
= rzalloc(ctx
, bi_block
);
788 blk
->base
.predecessors
= _mesa_set_create(blk
,
790 _mesa_key_pointer_equal
);
792 blk
->base
.name
= ctx
->block_name_count
++;
798 bi_schedule_barrier(bi_context
*ctx
)
800 bi_block
*temp
= ctx
->after_block
;
801 ctx
->after_block
= create_empty_block(ctx
);
802 list_addtail(&ctx
->after_block
->base
.link
, &ctx
->blocks
);
803 list_inithead(&ctx
->after_block
->base
.instructions
);
804 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->after_block
->base
);
805 ctx
->current_block
= ctx
->after_block
;
806 ctx
->after_block
= temp
;
810 emit_block(bi_context
*ctx
, nir_block
*block
)
812 if (ctx
->after_block
) {
813 ctx
->current_block
= ctx
->after_block
;
814 ctx
->after_block
= NULL
;
816 ctx
->current_block
= create_empty_block(ctx
);
819 list_addtail(&ctx
->current_block
->base
.link
, &ctx
->blocks
);
820 list_inithead(&ctx
->current_block
->base
.instructions
);
822 nir_foreach_instr(instr
, block
) {
823 emit_instr(ctx
, instr
);
824 ++ctx
->instruction_count
;
827 return ctx
->current_block
;
830 /* Emits an unconditional branch to the end of the current block, returning a
831 * pointer so the user can fill in details */
833 static bi_instruction
*
834 bi_emit_branch(bi_context
*ctx
)
836 bi_instruction branch
= {
839 .cond
= BI_COND_ALWAYS
843 return bi_emit(ctx
, branch
);
846 /* Sets a condition for a branch by examing the NIR condition. If we're
847 * familiar with the condition, we unwrap it to fold it into the branch
848 * instruction. Otherwise, we consume the condition directly. We
849 * generally use 1-bit booleans which allows us to use small types for
854 bi_set_branch_cond(bi_instruction
*branch
, nir_src
*cond
, bool invert
)
856 /* TODO: Try to unwrap instead of always bailing */
857 branch
->src
[0] = pan_src_index(cond
);
858 branch
->src
[1] = BIR_INDEX_ZERO
;
859 branch
->src_types
[0] = branch
->src_types
[1] = nir_type_uint16
;
860 branch
->branch
.cond
= invert
? BI_COND_EQ
: BI_COND_NE
;
864 emit_if(bi_context
*ctx
, nir_if
*nif
)
866 bi_block
*before_block
= ctx
->current_block
;
868 /* Speculatively emit the branch, but we can't fill it in until later */
869 bi_instruction
*then_branch
= bi_emit_branch(ctx
);
870 bi_set_branch_cond(then_branch
, &nif
->condition
, true);
872 /* Emit the two subblocks. */
873 bi_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
874 bi_block
*end_then_block
= ctx
->current_block
;
876 /* Emit a jump from the end of the then block to the end of the else */
877 bi_instruction
*then_exit
= bi_emit_branch(ctx
);
879 /* Emit second block, and check if it's empty */
881 int count_in
= ctx
->instruction_count
;
882 bi_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
883 bi_block
*end_else_block
= ctx
->current_block
;
884 ctx
->after_block
= create_empty_block(ctx
);
886 /* Now that we have the subblocks emitted, fix up the branches */
891 if (ctx
->instruction_count
== count_in
) {
892 /* The else block is empty, so don't emit an exit jump */
893 bi_remove_instruction(then_exit
);
894 then_branch
->branch
.target
= ctx
->after_block
;
896 then_branch
->branch
.target
= else_block
;
897 then_exit
->branch
.target
= ctx
->after_block
;
898 pan_block_add_successor(&end_then_block
->base
, &then_exit
->branch
.target
->base
);
901 /* Wire up the successors */
903 pan_block_add_successor(&before_block
->base
, &then_branch
->branch
.target
->base
); /* then_branch */
905 pan_block_add_successor(&before_block
->base
, &then_block
->base
); /* fallthrough */
906 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
); /* fallthrough */
910 emit_loop(bi_context
*ctx
, nir_loop
*nloop
)
912 /* Remember where we are */
913 bi_block
*start_block
= ctx
->current_block
;
915 bi_block
*saved_break
= ctx
->break_block
;
916 bi_block
*saved_continue
= ctx
->continue_block
;
918 ctx
->continue_block
= create_empty_block(ctx
);
919 ctx
->break_block
= create_empty_block(ctx
);
920 ctx
->after_block
= ctx
->continue_block
;
922 /* Emit the body itself */
923 emit_cf_list(ctx
, &nloop
->body
);
925 /* Branch back to loop back */
926 bi_instruction
*br_back
= bi_emit_branch(ctx
);
927 br_back
->branch
.target
= ctx
->continue_block
;
928 pan_block_add_successor(&start_block
->base
, &ctx
->continue_block
->base
);
929 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->continue_block
->base
);
931 ctx
->after_block
= ctx
->break_block
;
934 ctx
->break_block
= saved_break
;
935 ctx
->continue_block
= saved_continue
;
940 emit_cf_list(bi_context
*ctx
, struct exec_list
*list
)
942 bi_block
*start_block
= NULL
;
944 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
945 switch (node
->type
) {
946 case nir_cf_node_block
: {
947 bi_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
956 emit_if(ctx
, nir_cf_node_as_if(node
));
959 case nir_cf_node_loop
:
960 emit_loop(ctx
, nir_cf_node_as_loop(node
));
964 unreachable("Unknown control flow");
972 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
974 return glsl_count_attribute_slots(type
, false);
978 bi_optimize_nir(nir_shader
*nir
)
981 unsigned lower_flrp
= 16 | 32 | 64;
983 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
984 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
986 nir_lower_tex_options lower_tex_options
= {
987 .lower_txs_lod
= true,
989 .lower_tex_without_implicit_lod
= true,
993 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
994 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
995 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
1000 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
1001 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
1003 NIR_PASS(progress
, nir
, nir_copy_prop
);
1004 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
1005 NIR_PASS(progress
, nir
, nir_opt_dce
);
1006 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
1007 NIR_PASS(progress
, nir
, nir_opt_cse
);
1008 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
1009 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
1010 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
1012 if (lower_flrp
!= 0) {
1013 bool lower_flrp_progress
= false;
1014 NIR_PASS(lower_flrp_progress
,
1018 false /* always_precise */,
1019 nir
->options
->lower_ffma
);
1020 if (lower_flrp_progress
) {
1021 NIR_PASS(progress
, nir
,
1022 nir_opt_constant_folding
);
1026 /* Nothing should rematerialize any flrps, so we only
1027 * need to do this lowering once.
1032 NIR_PASS(progress
, nir
, nir_opt_undef
);
1033 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
1035 nir_var_shader_out
|
1036 nir_var_function_temp
);
1039 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
1040 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
1041 NIR_PASS(progress
, nir
, bifrost_nir_lower_algebraic_late
);
1042 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
1043 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
1045 /* Take us out of SSA */
1046 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
1047 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
1048 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
1052 bifrost_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, unsigned product_id
)
1054 bi_context
*ctx
= rzalloc(NULL
, bi_context
);
1056 ctx
->stage
= nir
->info
.stage
;
1057 ctx
->quirks
= bifrost_get_quirks(product_id
);
1058 list_inithead(&ctx
->blocks
);
1060 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
1061 * (so we don't accidentally duplicate the epilogue since mesa/st has
1062 * messed with our I/O quite a bit already) */
1064 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
1066 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1067 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
1068 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
1071 NIR_PASS_V(nir
, nir_split_var_copies
);
1072 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
1073 NIR_PASS_V(nir
, nir_lower_var_copies
);
1074 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
1075 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
1076 NIR_PASS_V(nir
, nir_lower_ssbo
);
1078 bi_optimize_nir(nir
);
1079 nir_print_shader(nir
, stdout
);
1081 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
1082 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
1083 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
1084 ctx
->blend_types
= program
->blend_types
;
1086 nir_foreach_function(func
, nir
) {
1090 ctx
->impl
= func
->impl
;
1091 emit_cf_list(ctx
, &func
->impl
->body
);
1092 break; /* TODO: Multi-function shaders */
1095 bi_foreach_block(ctx
, _block
) {
1096 bi_block
*block
= (bi_block
*) _block
;
1097 bi_lower_combine(ctx
, block
);
1100 bool progress
= false;
1105 bi_foreach_block(ctx
, _block
) {
1106 bi_block
*block
= (bi_block
*) _block
;
1107 progress
|= bi_opt_dead_code_eliminate(ctx
, block
);
1111 bi_print_shader(ctx
, stdout
);
1113 bi_register_allocate(ctx
);
1114 bi_print_shader(ctx
, stdout
);
1115 bi_pack(ctx
, &program
->compiled
);
1116 disassemble_bifrost(stdout
, program
->compiled
.data
, program
->compiled
.size
, true);