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])
157 bi_emit(ctx
, address
);
162 bi_emit_ld_uniform(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
164 bi_instruction ld
= bi_load(BI_LOAD_UNIFORM
, instr
);
165 ld
.src
[1] = BIR_INDEX_ZERO
; /* TODO: UBO index */
167 /* TODO: Indirect access, since we need to multiply by the element
168 * size. I believe we can get this lowering automatically via
169 * nir_lower_io (as mul instructions) with the proper options, but this
171 assert(ld
.src
[0] & BIR_INDEX_CONSTANT
);
172 ld
.constant
.u64
+= ctx
->sysvals
.sysval_count
;
173 ld
.constant
.u64
*= 16;
179 bi_emit_sysval(bi_context
*ctx
, nir_instr
*instr
,
180 unsigned nr_components
, unsigned offset
)
184 /* Figure out which uniform this is */
185 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
186 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
188 /* Sysvals are prefix uniforms */
189 unsigned uniform
= ((uintptr_t) val
) - 1;
191 /* Emit the read itself -- this is never indirect */
193 bi_instruction load
= {
194 .type
= BI_LOAD_UNIFORM
,
195 .writemask
= (1 << (nr_components
* 4)) - 1,
196 .src
= { BIR_INDEX_CONSTANT
},
197 .constant
= { (uniform
* 16) + offset
},
198 .dest
= bir_dest_index(&nir_dest
),
199 .dest_type
= nir_type_uint32
, /* TODO */
206 emit_intrinsic(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
209 switch (instr
->intrinsic
) {
210 case nir_intrinsic_load_barycentric_pixel
:
213 case nir_intrinsic_load_interpolated_input
:
214 case nir_intrinsic_load_input
:
215 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
216 bi_emit_ld_vary(ctx
, instr
);
217 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
218 bi_emit(ctx
, bi_load(BI_LOAD_ATTR
, instr
));
220 unreachable("Unsupported shader stage");
224 case nir_intrinsic_store_output
:
225 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
226 bi_emit_frag_out(ctx
, instr
);
227 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
228 bi_emit_st_vary(ctx
, instr
);
230 unreachable("Unsupported shader stage");
233 case nir_intrinsic_load_uniform
:
234 bi_emit_ld_uniform(ctx
, instr
);
237 case nir_intrinsic_load_ssbo_address
:
238 bi_emit_sysval(ctx
, &instr
->instr
, 1, 0);
241 case nir_intrinsic_get_buffer_size
:
242 bi_emit_sysval(ctx
, &instr
->instr
, 1, 8);
245 case nir_intrinsic_load_viewport_scale
:
246 case nir_intrinsic_load_viewport_offset
:
247 case nir_intrinsic_load_num_work_groups
:
248 case nir_intrinsic_load_sampler_lod_parameters_pan
:
249 bi_emit_sysval(ctx
, &instr
->instr
, 3, 0);
259 emit_load_const(bi_context
*ctx
, nir_load_const_instr
*instr
)
261 /* Make sure we've been lowered */
262 assert(instr
->def
.num_components
== 1);
264 bi_instruction move
= {
266 .dest
= bir_ssa_index(&instr
->def
),
267 .dest_type
= instr
->def
.bit_size
| nir_type_uint
,
268 .writemask
= (1 << (instr
->def
.bit_size
/ 8)) - 1,
273 .u64
= nir_const_value_as_uint(instr
->value
[0], instr
->def
.bit_size
)
281 bi_class_for_nir_alu(nir_op op
)
351 unreachable("Unknown ALU op");
356 bi_cond_for_nir(nir_op op
)
372 unreachable("Invalid compare");
377 emit_alu(bi_context
*ctx
, nir_alu_instr
*instr
)
379 /* Assume it's something we can handle normally */
380 bi_instruction alu
= {
381 .type
= bi_class_for_nir_alu(instr
->op
),
382 .dest
= bir_dest_index(&instr
->dest
.dest
),
383 .dest_type
= nir_op_infos
[instr
->op
].output_type
384 | nir_dest_bit_size(instr
->dest
.dest
),
387 /* TODO: Implement lowering of special functions for older Bifrost */
388 assert((alu
.type
!= BI_SPECIAL
) || !(ctx
->quirks
& BIFROST_NO_FAST_OP
));
390 if (instr
->dest
.dest
.is_ssa
) {
391 /* Construct a writemask */
392 unsigned bits_per_comp
= instr
->dest
.dest
.ssa
.bit_size
;
393 unsigned comps
= instr
->dest
.dest
.ssa
.num_components
;
395 unsigned bits
= bits_per_comp
* comps
;
396 unsigned bytes
= MAX2(bits
/ 8, 1);
397 alu
.writemask
= (1 << bytes
) - 1;
399 unsigned comp_mask
= instr
->dest
.write_mask
;
401 alu
.writemask
= pan_to_bytemask(nir_dest_bit_size(instr
->dest
.dest
),
405 /* We inline constants as we go. This tracks how many constants have
406 * been inlined, since we're limited to 64-bits of constants per
409 unsigned dest_bits
= nir_dest_bit_size(instr
->dest
.dest
);
410 unsigned constants_left
= (64 / dest_bits
);
411 unsigned constant_shift
= 0;
415 unsigned num_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
416 assert(num_inputs
<= ARRAY_SIZE(alu
.src
));
418 for (unsigned i
= 0; i
< num_inputs
; ++i
) {
419 unsigned bits
= nir_src_bit_size(instr
->src
[i
].src
);
420 alu
.src_types
[i
] = nir_op_infos
[instr
->op
].input_types
[i
]
423 /* Try to inline a constant */
424 if (nir_src_is_const(instr
->src
[i
].src
) && constants_left
&& (dest_bits
== bits
)) {
426 (nir_src_as_uint(instr
->src
[i
].src
)) << constant_shift
;
428 alu
.src
[i
] = BIR_INDEX_CONSTANT
| constant_shift
;
430 constant_shift
+= dest_bits
;
434 alu
.src
[i
] = bir_src_index(&instr
->src
[i
].src
);
436 /* We assert scalarization above */
437 alu
.swizzle
[i
][0] = instr
->src
[i
].swizzle
[0];
440 /* Op-specific fixup */
443 alu
.src
[2] = BIR_INDEX_ZERO
; /* FMA */
446 alu
.outmod
= BIFROST_SAT
; /* MOV */
449 alu
.src_neg
[0] = true; /* MOV */
452 alu
.src_abs
[0] = true; /* MOV */
455 alu
.src_neg
[1] = true; /* ADD */
460 alu
.op
.minmax
= BI_MINMAX_MAX
; /* MINMAX */
463 alu
.op
.special
= BI_SPECIAL_FRCP
;
466 alu
.op
.special
= BI_SPECIAL_FRSQ
;
469 alu
.op
.special
= BI_SPECIAL_FSIN
;
472 alu
.op
.special
= BI_SPECIAL_FCOS
;
482 alu
.op
.compare
= bi_cond_for_nir(instr
->op
);
492 emit_instr(bi_context
*ctx
, struct nir_instr
*instr
)
494 switch (instr
->type
) {
495 case nir_instr_type_load_const
:
496 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
499 case nir_instr_type_intrinsic
:
500 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
503 case nir_instr_type_alu
:
504 emit_alu(ctx
, nir_instr_as_alu(instr
));
508 case nir_instr_type_tex
:
509 emit_tex(ctx
, nir_instr_as_tex(instr
));
513 case nir_instr_type_jump
:
514 emit_jump(ctx
, nir_instr_as_jump(instr
));
517 case nir_instr_type_ssa_undef
:
522 //unreachable("Unhandled instruction type");
530 create_empty_block(bi_context
*ctx
)
532 bi_block
*blk
= rzalloc(ctx
, bi_block
);
534 blk
->base
.predecessors
= _mesa_set_create(blk
,
536 _mesa_key_pointer_equal
);
538 blk
->base
.name
= ctx
->block_name_count
++;
544 bi_schedule_barrier(bi_context
*ctx
)
546 bi_block
*temp
= ctx
->after_block
;
547 ctx
->after_block
= create_empty_block(ctx
);
548 list_addtail(&ctx
->after_block
->base
.link
, &ctx
->blocks
);
549 list_inithead(&ctx
->after_block
->base
.instructions
);
550 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->after_block
->base
);
551 ctx
->current_block
= ctx
->after_block
;
552 ctx
->after_block
= temp
;
556 emit_block(bi_context
*ctx
, nir_block
*block
)
558 if (ctx
->after_block
) {
559 ctx
->current_block
= ctx
->after_block
;
560 ctx
->after_block
= NULL
;
562 ctx
->current_block
= create_empty_block(ctx
);
565 list_addtail(&ctx
->current_block
->base
.link
, &ctx
->blocks
);
566 list_inithead(&ctx
->current_block
->base
.instructions
);
568 nir_foreach_instr(instr
, block
) {
569 emit_instr(ctx
, instr
);
570 ++ctx
->instruction_count
;
573 return ctx
->current_block
;
576 /* Emits an unconditional branch to the end of the current block, returning a
577 * pointer so the user can fill in details */
579 static bi_instruction
*
580 bi_emit_branch(bi_context
*ctx
)
582 bi_instruction branch
= {
585 .cond
= BI_COND_ALWAYS
589 return bi_emit(ctx
, branch
);
592 /* Sets a condition for a branch by examing the NIR condition. If we're
593 * familiar with the condition, we unwrap it to fold it into the branch
594 * instruction. Otherwise, we consume the condition directly. We
595 * generally use 1-bit booleans which allows us to use small types for
600 bi_set_branch_cond(bi_instruction
*branch
, nir_src
*cond
, bool invert
)
602 /* TODO: Try to unwrap instead of always bailing */
603 branch
->src
[0] = bir_src_index(cond
);
604 branch
->src
[1] = BIR_INDEX_ZERO
;
605 branch
->src_types
[0] = branch
->src_types
[1] = nir_type_uint16
;
606 branch
->branch
.cond
= invert
? BI_COND_EQ
: BI_COND_NE
;
610 emit_if(bi_context
*ctx
, nir_if
*nif
)
612 bi_block
*before_block
= ctx
->current_block
;
614 /* Speculatively emit the branch, but we can't fill it in until later */
615 bi_instruction
*then_branch
= bi_emit_branch(ctx
);
616 bi_set_branch_cond(then_branch
, &nif
->condition
, true);
618 /* Emit the two subblocks. */
619 bi_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
620 bi_block
*end_then_block
= ctx
->current_block
;
622 /* Emit a jump from the end of the then block to the end of the else */
623 bi_instruction
*then_exit
= bi_emit_branch(ctx
);
625 /* Emit second block, and check if it's empty */
627 int count_in
= ctx
->instruction_count
;
628 bi_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
629 bi_block
*end_else_block
= ctx
->current_block
;
630 ctx
->after_block
= create_empty_block(ctx
);
632 /* Now that we have the subblocks emitted, fix up the branches */
637 if (ctx
->instruction_count
== count_in
) {
638 /* The else block is empty, so don't emit an exit jump */
639 bi_remove_instruction(then_exit
);
640 then_branch
->branch
.target
= ctx
->after_block
;
642 then_branch
->branch
.target
= else_block
;
643 then_exit
->branch
.target
= ctx
->after_block
;
644 pan_block_add_successor(&end_then_block
->base
, &then_exit
->branch
.target
->base
);
647 /* Wire up the successors */
649 pan_block_add_successor(&before_block
->base
, &then_branch
->branch
.target
->base
); /* then_branch */
651 pan_block_add_successor(&before_block
->base
, &then_block
->base
); /* fallthrough */
652 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
); /* fallthrough */
656 emit_loop(bi_context
*ctx
, nir_loop
*nloop
)
658 /* Remember where we are */
659 bi_block
*start_block
= ctx
->current_block
;
661 bi_block
*saved_break
= ctx
->break_block
;
662 bi_block
*saved_continue
= ctx
->continue_block
;
664 ctx
->continue_block
= create_empty_block(ctx
);
665 ctx
->break_block
= create_empty_block(ctx
);
666 ctx
->after_block
= ctx
->continue_block
;
668 /* Emit the body itself */
669 emit_cf_list(ctx
, &nloop
->body
);
671 /* Branch back to loop back */
672 bi_instruction
*br_back
= bi_emit_branch(ctx
);
673 br_back
->branch
.target
= ctx
->continue_block
;
674 pan_block_add_successor(&start_block
->base
, &ctx
->continue_block
->base
);
675 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->continue_block
->base
);
677 ctx
->after_block
= ctx
->break_block
;
680 ctx
->break_block
= saved_break
;
681 ctx
->continue_block
= saved_continue
;
686 emit_cf_list(bi_context
*ctx
, struct exec_list
*list
)
688 bi_block
*start_block
= NULL
;
690 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
691 switch (node
->type
) {
692 case nir_cf_node_block
: {
693 bi_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
702 emit_if(ctx
, nir_cf_node_as_if(node
));
705 case nir_cf_node_loop
:
706 emit_loop(ctx
, nir_cf_node_as_loop(node
));
710 unreachable("Unknown control flow");
718 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
720 return glsl_count_attribute_slots(type
, false);
724 bi_optimize_nir(nir_shader
*nir
)
727 unsigned lower_flrp
= 16 | 32 | 64;
729 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
730 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
732 nir_lower_tex_options lower_tex_options
= {
733 .lower_txs_lod
= true,
735 .lower_tex_without_implicit_lod
= true,
739 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
740 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
741 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
746 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
747 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
749 NIR_PASS(progress
, nir
, nir_copy_prop
);
750 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
751 NIR_PASS(progress
, nir
, nir_opt_dce
);
752 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
753 NIR_PASS(progress
, nir
, nir_opt_cse
);
754 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
755 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
756 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
758 if (lower_flrp
!= 0) {
759 bool lower_flrp_progress
= false;
760 NIR_PASS(lower_flrp_progress
,
764 false /* always_precise */,
765 nir
->options
->lower_ffma
);
766 if (lower_flrp_progress
) {
767 NIR_PASS(progress
, nir
,
768 nir_opt_constant_folding
);
772 /* Nothing should rematerialize any flrps, so we only
773 * need to do this lowering once.
778 NIR_PASS(progress
, nir
, nir_opt_undef
);
779 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
782 nir_var_function_temp
);
785 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
786 NIR_PASS(progress
, nir
, bifrost_nir_lower_algebraic_late
);
787 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
788 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
790 /* Take us out of SSA */
791 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
792 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
794 /* We're a primary scalar architecture but there's enough vector that
795 * we use a vector IR so let's not also deal with scalar hacks on top
796 * of the vector hacks */
798 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
799 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
800 NIR_PASS(progress
, nir
, nir_opt_dce
);
804 bifrost_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, unsigned product_id
)
806 bi_context
*ctx
= rzalloc(NULL
, bi_context
);
808 ctx
->stage
= nir
->info
.stage
;
809 ctx
->quirks
= bifrost_get_quirks(product_id
);
810 list_inithead(&ctx
->blocks
);
812 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
813 * (so we don't accidentally duplicate the epilogue since mesa/st has
814 * messed with our I/O quite a bit already) */
816 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
818 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
819 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
820 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
823 NIR_PASS_V(nir
, nir_split_var_copies
);
824 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
825 NIR_PASS_V(nir
, nir_lower_var_copies
);
826 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
827 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
828 NIR_PASS_V(nir
, nir_lower_ssbo
);
830 bi_optimize_nir(nir
);
831 nir_print_shader(nir
, stdout
);
833 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
834 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
835 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
837 nir_foreach_function(func
, nir
) {
841 ctx
->impl
= func
->impl
;
842 emit_cf_list(ctx
, &func
->impl
->body
);
843 break; /* TODO: Multi-function shaders */
846 bool progress
= false;
851 bi_foreach_block(ctx
, _block
) {
852 bi_block
*block
= (bi_block
*) _block
;
853 progress
|= bi_opt_dead_code_eliminate(ctx
, block
);
857 bi_print_shader(ctx
, stdout
);