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 "main/imports.h"
31 #include "compiler/nir/nir_builder.h"
33 #include "disassemble.h"
34 #include "bifrost_compile.h"
35 #include "bifrost_nir.h"
37 #include "bi_quirks.h"
40 static bi_block
*emit_cf_list(bi_context
*ctx
, struct exec_list
*list
);
41 static bi_instruction
*bi_emit_branch(bi_context
*ctx
);
42 static void bi_schedule_barrier(bi_context
*ctx
);
45 emit_jump(bi_context
*ctx
, nir_jump_instr
*instr
)
47 bi_instruction
*branch
= bi_emit_branch(ctx
);
49 switch (instr
->type
) {
51 branch
->branch
.target
= ctx
->break_block
;
53 case nir_jump_continue
:
54 branch
->branch
.target
= ctx
->continue_block
;
57 unreachable("Unhandled jump type");
60 pan_block_add_successor(&ctx
->current_block
->base
, &branch
->branch
.target
->base
);
63 /* Gets a bytemask for a complete vecN write */
65 bi_mask_for_channels_32(unsigned i
)
67 return (1 << (4 * i
)) - 1;
71 bi_load(enum bi_class T
, nir_intrinsic_instr
*instr
)
73 bi_instruction load
= {
75 .writemask
= bi_mask_for_channels_32(instr
->num_components
),
76 .src
= { BIR_INDEX_CONSTANT
},
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
),
110 bi_emit_frag_out(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
112 if (!ctx
->emitted_atest
) {
113 bi_instruction ins
= {
118 bi_schedule_barrier(ctx
);
119 ctx
->emitted_atest
= true;
122 bi_instruction blend
= {
124 .blend_location
= nir_intrinsic_base(instr
),
126 bir_src_index(&instr
->src
[0])
134 bi_schedule_barrier(ctx
);
138 bi_emit_st_vary(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
140 bi_instruction address
= bi_load(BI_LOAD_VAR_ADDRESS
, instr
);
141 address
.dest
= bi_make_temp(ctx
);
142 address
.dest_type
= nir_type_uint64
;
143 address
.writemask
= (1 << 8) - 1;
145 bi_instruction st
= {
146 .type
= BI_STORE_VAR
,
149 bir_src_index(&instr
->src
[0])
156 bi_emit(ctx
, address
);
161 bi_emit_ld_uniform(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
163 bi_instruction ld
= bi_load(BI_LOAD_UNIFORM
, instr
);
164 ld
.src
[1] = BIR_INDEX_ZERO
; /* TODO: UBO index */
166 /* TODO: Indirect access, since we need to multiply by the element
167 * size. I believe we can get this lowering automatically via
168 * nir_lower_io (as mul instructions) with the proper options, but this
170 assert(ld
.src
[0] & BIR_INDEX_CONSTANT
);
171 ld
.constant
.u64
+= ctx
->sysvals
.sysval_count
;
172 ld
.constant
.u64
*= 16;
178 bi_emit_sysval(bi_context
*ctx
, nir_instr
*instr
,
179 unsigned nr_components
, unsigned offset
)
183 /* Figure out which uniform this is */
184 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
185 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
187 /* Sysvals are prefix uniforms */
188 unsigned uniform
= ((uintptr_t) val
) - 1;
190 /* Emit the read itself -- this is never indirect */
192 bi_instruction load
= {
193 .type
= BI_LOAD_UNIFORM
,
194 .writemask
= (1 << (nr_components
* 4)) - 1,
195 .src
= { BIR_INDEX_CONSTANT
},
196 .constant
= { (uniform
* 16) + offset
},
197 .dest
= bir_dest_index(&nir_dest
),
198 .dest_type
= nir_type_uint32
, /* TODO */
205 emit_intrinsic(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
208 switch (instr
->intrinsic
) {
209 case nir_intrinsic_load_barycentric_pixel
:
212 case nir_intrinsic_load_interpolated_input
:
213 case nir_intrinsic_load_input
:
214 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
215 bi_emit_ld_vary(ctx
, instr
);
216 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
217 bi_emit(ctx
, bi_load(BI_LOAD_ATTR
, instr
));
219 unreachable("Unsupported shader stage");
223 case nir_intrinsic_store_output
:
224 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
225 bi_emit_frag_out(ctx
, instr
);
226 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
227 bi_emit_st_vary(ctx
, instr
);
229 unreachable("Unsupported shader stage");
232 case nir_intrinsic_load_uniform
:
233 bi_emit_ld_uniform(ctx
, instr
);
236 case nir_intrinsic_load_ssbo_address
:
237 bi_emit_sysval(ctx
, &instr
->instr
, 1, 0);
240 case nir_intrinsic_get_buffer_size
:
241 bi_emit_sysval(ctx
, &instr
->instr
, 1, 8);
244 case nir_intrinsic_load_viewport_scale
:
245 case nir_intrinsic_load_viewport_offset
:
246 case nir_intrinsic_load_num_work_groups
:
247 case nir_intrinsic_load_sampler_lod_parameters_pan
:
248 bi_emit_sysval(ctx
, &instr
->instr
, 3, 0);
258 emit_load_const(bi_context
*ctx
, nir_load_const_instr
*instr
)
260 /* Make sure we've been lowered */
261 assert(instr
->def
.num_components
== 1);
263 bi_instruction move
= {
265 .dest
= bir_ssa_index(&instr
->def
),
266 .dest_type
= instr
->def
.bit_size
| nir_type_uint
,
267 .writemask
= (1 << (instr
->def
.bit_size
/ 8)) - 1,
272 .u64
= nir_const_value_as_uint(instr
->value
[0], instr
->def
.bit_size
)
280 bi_class_for_nir_alu(nir_op op
)
349 unreachable("Unknown ALU op");
354 bi_cond_for_nir(nir_op op
)
370 unreachable("Invalid compare");
375 emit_alu(bi_context
*ctx
, nir_alu_instr
*instr
)
377 /* Assume it's something we can handle normally */
378 bi_instruction alu
= {
379 .type
= bi_class_for_nir_alu(instr
->op
),
380 .dest
= bir_dest_index(&instr
->dest
.dest
),
381 .dest_type
= nir_op_infos
[instr
->op
].output_type
382 | nir_dest_bit_size(instr
->dest
.dest
),
385 /* TODO: Implement lowering of special functions for older Bifrost */
386 assert((alu
.type
!= BI_SPECIAL
) || !(ctx
->quirks
& BIFROST_NO_FAST_OP
));
388 if (instr
->dest
.dest
.is_ssa
) {
389 /* Construct a writemask */
390 unsigned bits_per_comp
= instr
->dest
.dest
.ssa
.bit_size
;
391 unsigned comps
= instr
->dest
.dest
.ssa
.num_components
;
393 unsigned bits
= bits_per_comp
* comps
;
394 unsigned bytes
= MAX2(bits
/ 8, 1);
395 alu
.writemask
= (1 << bytes
) - 1;
397 unsigned comp_mask
= instr
->dest
.write_mask
;
399 alu
.writemask
= pan_to_bytemask(nir_dest_bit_size(instr
->dest
.dest
),
403 /* We inline constants as we go. This tracks how many constants have
404 * been inlined, since we're limited to 64-bits of constants per
407 unsigned dest_bits
= nir_dest_bit_size(instr
->dest
.dest
);
408 unsigned constants_left
= (64 / dest_bits
);
409 unsigned constant_shift
= 0;
413 unsigned num_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
414 assert(num_inputs
<= ARRAY_SIZE(alu
.src
));
416 for (unsigned i
= 0; i
< num_inputs
; ++i
) {
417 unsigned bits
= nir_src_bit_size(instr
->src
[i
].src
);
418 alu
.src_types
[i
] = nir_op_infos
[instr
->op
].input_types
[i
]
421 /* Try to inline a constant */
422 if (nir_src_is_const(instr
->src
[i
].src
) && constants_left
&& (dest_bits
== bits
)) {
424 (nir_src_as_uint(instr
->src
[i
].src
)) << constant_shift
;
426 alu
.src
[i
] = BIR_INDEX_CONSTANT
| constant_shift
;
428 constant_shift
+= dest_bits
;
432 alu
.src
[i
] = bir_src_index(&instr
->src
[i
].src
);
434 /* We assert scalarization above */
435 alu
.swizzle
[i
][0] = instr
->src
[i
].swizzle
[0];
438 /* Op-specific fixup */
441 alu
.src
[2] = BIR_INDEX_ZERO
; /* FMA */
444 alu
.outmod
= BIFROST_SAT
; /* MOV */
447 alu
.src_neg
[0] = true; /* MOV */
450 alu
.src_abs
[0] = true; /* MOV */
453 alu
.src_neg
[1] = true; /* ADD */
458 alu
.op
.minmax
= BI_MINMAX_MAX
; /* MINMAX */
461 alu
.op
.special
= BI_SPECIAL_FRCP
;
464 alu
.op
.special
= BI_SPECIAL_FRSQ
;
467 alu
.op
.special
= BI_SPECIAL_FSIN
;
470 alu
.op
.special
= BI_SPECIAL_FCOS
;
480 alu
.op
.compare
= bi_cond_for_nir(instr
->op
);
490 emit_instr(bi_context
*ctx
, struct nir_instr
*instr
)
492 switch (instr
->type
) {
493 case nir_instr_type_load_const
:
494 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
497 case nir_instr_type_intrinsic
:
498 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
501 case nir_instr_type_alu
:
502 emit_alu(ctx
, nir_instr_as_alu(instr
));
506 case nir_instr_type_tex
:
507 emit_tex(ctx
, nir_instr_as_tex(instr
));
511 case nir_instr_type_jump
:
512 emit_jump(ctx
, nir_instr_as_jump(instr
));
515 case nir_instr_type_ssa_undef
:
520 //unreachable("Unhandled instruction type");
528 create_empty_block(bi_context
*ctx
)
530 bi_block
*blk
= rzalloc(ctx
, bi_block
);
532 blk
->base
.predecessors
= _mesa_set_create(blk
,
534 _mesa_key_pointer_equal
);
536 blk
->base
.name
= ctx
->block_name_count
++;
542 bi_schedule_barrier(bi_context
*ctx
)
544 bi_block
*temp
= ctx
->after_block
;
545 ctx
->after_block
= create_empty_block(ctx
);
546 list_addtail(&ctx
->after_block
->base
.link
, &ctx
->blocks
);
547 list_inithead(&ctx
->after_block
->base
.instructions
);
548 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->after_block
->base
);
549 ctx
->current_block
= ctx
->after_block
;
550 ctx
->after_block
= temp
;
554 emit_block(bi_context
*ctx
, nir_block
*block
)
556 if (ctx
->after_block
) {
557 ctx
->current_block
= ctx
->after_block
;
558 ctx
->after_block
= NULL
;
560 ctx
->current_block
= create_empty_block(ctx
);
563 list_addtail(&ctx
->current_block
->base
.link
, &ctx
->blocks
);
564 list_inithead(&ctx
->current_block
->base
.instructions
);
566 nir_foreach_instr(instr
, block
) {
567 emit_instr(ctx
, instr
);
568 ++ctx
->instruction_count
;
571 return ctx
->current_block
;
574 /* Emits an unconditional branch to the end of the current block, returning a
575 * pointer so the user can fill in details */
577 static bi_instruction
*
578 bi_emit_branch(bi_context
*ctx
)
580 bi_instruction branch
= {
583 .cond
= BI_COND_ALWAYS
587 return bi_emit(ctx
, branch
);
590 /* Sets a condition for a branch by examing the NIR condition. If we're
591 * familiar with the condition, we unwrap it to fold it into the branch
592 * instruction. Otherwise, we consume the condition directly. We
593 * generally use 1-bit booleans which allows us to use small types for
598 bi_set_branch_cond(bi_instruction
*branch
, nir_src
*cond
, bool invert
)
600 /* TODO: Try to unwrap instead of always bailing */
601 branch
->src
[0] = bir_src_index(cond
);
602 branch
->src
[1] = BIR_INDEX_ZERO
;
603 branch
->src_types
[0] = branch
->src_types
[1] = nir_type_uint16
;
604 branch
->branch
.cond
= invert
? BI_COND_EQ
: BI_COND_NE
;
608 emit_if(bi_context
*ctx
, nir_if
*nif
)
610 bi_block
*before_block
= ctx
->current_block
;
612 /* Speculatively emit the branch, but we can't fill it in until later */
613 bi_instruction
*then_branch
= bi_emit_branch(ctx
);
614 bi_set_branch_cond(then_branch
, &nif
->condition
, true);
616 /* Emit the two subblocks. */
617 bi_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
618 bi_block
*end_then_block
= ctx
->current_block
;
620 /* Emit a jump from the end of the then block to the end of the else */
621 bi_instruction
*then_exit
= bi_emit_branch(ctx
);
623 /* Emit second block, and check if it's empty */
625 int count_in
= ctx
->instruction_count
;
626 bi_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
627 bi_block
*end_else_block
= ctx
->current_block
;
628 ctx
->after_block
= create_empty_block(ctx
);
630 /* Now that we have the subblocks emitted, fix up the branches */
635 if (ctx
->instruction_count
== count_in
) {
636 /* The else block is empty, so don't emit an exit jump */
637 bi_remove_instruction(then_exit
);
638 then_branch
->branch
.target
= ctx
->after_block
;
640 then_branch
->branch
.target
= else_block
;
641 then_exit
->branch
.target
= ctx
->after_block
;
642 pan_block_add_successor(&end_then_block
->base
, &then_exit
->branch
.target
->base
);
645 /* Wire up the successors */
647 pan_block_add_successor(&before_block
->base
, &then_branch
->branch
.target
->base
); /* then_branch */
649 pan_block_add_successor(&before_block
->base
, &then_block
->base
); /* fallthrough */
650 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
); /* fallthrough */
654 emit_loop(bi_context
*ctx
, nir_loop
*nloop
)
656 /* Remember where we are */
657 bi_block
*start_block
= ctx
->current_block
;
659 bi_block
*saved_break
= ctx
->break_block
;
660 bi_block
*saved_continue
= ctx
->continue_block
;
662 ctx
->continue_block
= create_empty_block(ctx
);
663 ctx
->break_block
= create_empty_block(ctx
);
664 ctx
->after_block
= ctx
->continue_block
;
666 /* Emit the body itself */
667 emit_cf_list(ctx
, &nloop
->body
);
669 /* Branch back to loop back */
670 bi_instruction
*br_back
= bi_emit_branch(ctx
);
671 br_back
->branch
.target
= ctx
->continue_block
;
672 pan_block_add_successor(&start_block
->base
, &ctx
->continue_block
->base
);
673 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->continue_block
->base
);
675 ctx
->after_block
= ctx
->break_block
;
678 ctx
->break_block
= saved_break
;
679 ctx
->continue_block
= saved_continue
;
684 emit_cf_list(bi_context
*ctx
, struct exec_list
*list
)
686 bi_block
*start_block
= NULL
;
688 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
689 switch (node
->type
) {
690 case nir_cf_node_block
: {
691 bi_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
700 emit_if(ctx
, nir_cf_node_as_if(node
));
703 case nir_cf_node_loop
:
704 emit_loop(ctx
, nir_cf_node_as_loop(node
));
708 unreachable("Unknown control flow");
716 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
718 return glsl_count_attribute_slots(type
, false);
722 bi_optimize_nir(nir_shader
*nir
)
725 unsigned lower_flrp
= 16 | 32 | 64;
727 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
728 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
730 nir_lower_tex_options lower_tex_options
= {
731 .lower_txs_lod
= true,
733 .lower_tex_without_implicit_lod
= true,
737 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
738 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
739 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
744 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
745 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
747 NIR_PASS(progress
, nir
, nir_copy_prop
);
748 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
749 NIR_PASS(progress
, nir
, nir_opt_dce
);
750 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
751 NIR_PASS(progress
, nir
, nir_opt_cse
);
752 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
753 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
754 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
756 if (lower_flrp
!= 0) {
757 bool lower_flrp_progress
= false;
758 NIR_PASS(lower_flrp_progress
,
762 false /* always_precise */,
763 nir
->options
->lower_ffma
);
764 if (lower_flrp_progress
) {
765 NIR_PASS(progress
, nir
,
766 nir_opt_constant_folding
);
770 /* Nothing should rematerialize any flrps, so we only
771 * need to do this lowering once.
776 NIR_PASS(progress
, nir
, nir_opt_undef
);
777 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
780 nir_var_function_temp
);
783 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
784 NIR_PASS(progress
, nir
, bifrost_nir_lower_algebraic_late
);
785 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
786 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
788 /* Take us out of SSA */
789 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
790 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
792 /* We're a primary scalar architecture but there's enough vector that
793 * we use a vector IR so let's not also deal with scalar hacks on top
794 * of the vector hacks */
796 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
797 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
798 NIR_PASS(progress
, nir
, nir_opt_dce
);
802 bifrost_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, unsigned product_id
)
804 bi_context
*ctx
= rzalloc(NULL
, bi_context
);
806 ctx
->stage
= nir
->info
.stage
;
807 ctx
->quirks
= bifrost_get_quirks(product_id
);
808 list_inithead(&ctx
->blocks
);
810 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
811 * (so we don't accidentally duplicate the epilogue since mesa/st has
812 * messed with our I/O quite a bit already) */
814 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
816 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
817 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
818 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
821 NIR_PASS_V(nir
, nir_split_var_copies
);
822 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
823 NIR_PASS_V(nir
, nir_lower_var_copies
);
824 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
825 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
826 NIR_PASS_V(nir
, nir_lower_ssbo
);
828 bi_optimize_nir(nir
);
829 nir_print_shader(nir
, stdout
);
831 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
832 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
833 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
835 nir_foreach_function(func
, nir
) {
839 ctx
->impl
= func
->impl
;
840 emit_cf_list(ctx
, &func
->impl
->body
);
841 break; /* TODO: Multi-function shaders */
844 bi_print_shader(ctx
, stdout
);