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"
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_block_add_successor(bi_block
*block
, bi_block
*successor
);
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 bi_block_add_successor(ctx
->current_block
, branch
->branch
.target
);
64 bi_emit_ld_vary(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
66 bi_instruction ins
= {
70 .location
= nir_intrinsic_base(instr
),
71 .channels
= instr
->num_components
,
73 .interp_mode
= BIFROST_INTERP_DEFAULT
, /* TODO */
74 .reuse
= false, /* TODO */
75 .flat
= instr
->intrinsic
!= nir_intrinsic_load_interpolated_input
77 .dest
= bir_dest_index(&instr
->dest
),
78 .dest_type
= nir_type_float
| nir_dest_bit_size(instr
->dest
),
81 nir_src
*offset
= nir_get_io_offset_src(instr
);
83 if (nir_src_is_const(*offset
))
84 ins
.load_vary
.load
.location
+= nir_src_as_uint(*offset
);
86 ins
.src
[0] = bir_src_index(offset
);
92 bi_emit_frag_out(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
94 if (!ctx
->emitted_atest
) {
95 bi_instruction ins
= {
100 bi_schedule_barrier(ctx
);
101 ctx
->emitted_atest
= true;
104 bi_instruction blend
= {
106 .blend_location
= nir_intrinsic_base(instr
),
108 bir_src_index(&instr
->src
[0])
116 bi_schedule_barrier(ctx
);
119 static struct bi_load
120 bi_direct_load_for_instr(nir_intrinsic_instr
*instr
)
122 nir_src
*offset
= nir_get_io_offset_src(instr
);
123 assert(nir_src_is_const(*offset
)); /* no indirects */
125 struct bi_load load
= {
126 .location
= nir_intrinsic_base(instr
) + nir_src_as_uint(*offset
),
127 .channels
= instr
->num_components
134 bi_emit_ld_attr(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
136 bi_instruction load
= {
137 .type
= BI_LOAD_ATTR
,
138 .load
= bi_direct_load_for_instr(instr
),
139 .dest
= bir_dest_index(&instr
->dest
),
140 .dest_type
= nir_intrinsic_type(instr
)
147 bi_emit_st_vary(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
149 nir_src
*offset
= nir_get_io_offset_src(instr
);
150 assert(nir_src_is_const(*offset
)); /* no indirects */
152 bi_instruction address
= {
153 .type
= BI_LOAD_VAR_ADDRESS
,
154 .load
= bi_direct_load_for_instr(instr
),
155 .dest_type
= nir_intrinsic_type(instr
),
156 .dest
= bi_make_temp(ctx
)
159 bi_instruction st
= {
160 .type
= BI_STORE_VAR
,
163 bir_src_index(&instr
->src
[0])
170 bi_emit(ctx
, address
);
175 bi_emit_ld_uniform(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
177 /* TODO: Indirect access */
179 bi_instruction ld
= {
180 .type
= BI_LOAD_UNIFORM
,
181 .load
= bi_direct_load_for_instr(instr
),
182 .dest
= bir_dest_index(&instr
->dest
),
183 .dest_type
= nir_intrinsic_type(instr
),
185 BIR_INDEX_ZERO
/* TODO: UBOs */
193 emit_intrinsic(bi_context
*ctx
, nir_intrinsic_instr
*instr
)
196 switch (instr
->intrinsic
) {
197 case nir_intrinsic_load_barycentric_pixel
:
200 case nir_intrinsic_load_interpolated_input
:
201 case nir_intrinsic_load_input
:
202 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
203 bi_emit_ld_vary(ctx
, instr
);
204 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
205 bi_emit_ld_attr(ctx
, instr
);
207 unreachable("Unsupported shader stage");
211 case nir_intrinsic_store_output
:
212 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
213 bi_emit_frag_out(ctx
, instr
);
214 else if (ctx
->stage
== MESA_SHADER_VERTEX
)
215 bi_emit_st_vary(ctx
, instr
);
217 unreachable("Unsupported shader stage");
220 case nir_intrinsic_load_uniform
:
221 bi_emit_ld_uniform(ctx
, instr
);
231 emit_load_const(bi_context
*ctx
, nir_load_const_instr
*instr
)
233 /* Make sure we've been lowered */
234 assert(instr
->def
.num_components
== 1);
236 bi_instruction move
= {
238 .dest
= bir_ssa_index(&instr
->def
),
239 .dest_type
= instr
->def
.bit_size
| nir_type_uint
,
244 .u64
= nir_const_value_as_uint(instr
->value
[0], instr
->def
.bit_size
)
252 emit_instr(bi_context
*ctx
, struct nir_instr
*instr
)
254 switch (instr
->type
) {
255 case nir_instr_type_load_const
:
256 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
259 case nir_instr_type_intrinsic
:
260 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
264 case nir_instr_type_alu
:
265 emit_alu(ctx
, nir_instr_as_alu(instr
));
268 case nir_instr_type_tex
:
269 emit_tex(ctx
, nir_instr_as_tex(instr
));
273 case nir_instr_type_jump
:
274 emit_jump(ctx
, nir_instr_as_jump(instr
));
277 case nir_instr_type_ssa_undef
:
282 //unreachable("Unhandled instruction type");
290 create_empty_block(bi_context
*ctx
)
292 bi_block
*blk
= rzalloc(ctx
, bi_block
);
294 blk
->predecessors
= _mesa_set_create(blk
,
296 _mesa_key_pointer_equal
);
298 blk
->name
= ctx
->block_name_count
++;
304 bi_block_add_successor(bi_block
*block
, bi_block
*successor
)
309 for (unsigned i
= 0; i
< ARRAY_SIZE(block
->successors
); ++i
) {
310 if (block
->successors
[i
]) {
311 if (block
->successors
[i
] == successor
)
317 block
->successors
[i
] = successor
;
318 _mesa_set_add(successor
->predecessors
, block
);
322 unreachable("Too many successors");
326 bi_schedule_barrier(bi_context
*ctx
)
328 bi_block
*temp
= ctx
->after_block
;
329 ctx
->after_block
= create_empty_block(ctx
);
330 list_addtail(&ctx
->after_block
->link
, &ctx
->blocks
);
331 list_inithead(&ctx
->after_block
->instructions
);
332 bi_block_add_successor(ctx
->current_block
, ctx
->after_block
);
333 ctx
->current_block
= ctx
->after_block
;
334 ctx
->after_block
= temp
;
338 emit_block(bi_context
*ctx
, nir_block
*block
)
340 if (ctx
->after_block
) {
341 ctx
->current_block
= ctx
->after_block
;
342 ctx
->after_block
= NULL
;
344 ctx
->current_block
= create_empty_block(ctx
);
347 list_addtail(&ctx
->current_block
->link
, &ctx
->blocks
);
348 list_inithead(&ctx
->current_block
->instructions
);
350 nir_foreach_instr(instr
, block
) {
351 emit_instr(ctx
, instr
);
352 ++ctx
->instruction_count
;
355 return ctx
->current_block
;
358 /* Emits an unconditional branch to the end of the current block, returning a
359 * pointer so the user can fill in details */
361 static bi_instruction
*
362 bi_emit_branch(bi_context
*ctx
)
364 bi_instruction branch
= {
367 .cond
= BI_COND_ALWAYS
371 return bi_emit(ctx
, branch
);
374 /* Sets a condition for a branch by examing the NIR condition. If we're
375 * familiar with the condition, we unwrap it to fold it into the branch
376 * instruction. Otherwise, we consume the condition directly. We
377 * generally use 1-bit booleans which allows us to use small types for
382 bi_set_branch_cond(bi_instruction
*branch
, nir_src
*cond
, bool invert
)
384 /* TODO: Try to unwrap instead of always bailing */
385 branch
->src
[0] = bir_src_index(cond
);
386 branch
->src
[1] = BIR_INDEX_ZERO
;
387 branch
->src_types
[0] = branch
->src_types
[1] = nir_type_uint16
;
388 branch
->branch
.cond
= invert
? BI_COND_EQ
: BI_COND_NE
;
392 emit_if(bi_context
*ctx
, nir_if
*nif
)
394 bi_block
*before_block
= ctx
->current_block
;
396 /* Speculatively emit the branch, but we can't fill it in until later */
397 bi_instruction
*then_branch
= bi_emit_branch(ctx
);
398 bi_set_branch_cond(then_branch
, &nif
->condition
, true);
400 /* Emit the two subblocks. */
401 bi_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
402 bi_block
*end_then_block
= ctx
->current_block
;
404 /* Emit a jump from the end of the then block to the end of the else */
405 bi_instruction
*then_exit
= bi_emit_branch(ctx
);
407 /* Emit second block, and check if it's empty */
409 int count_in
= ctx
->instruction_count
;
410 bi_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
411 bi_block
*end_else_block
= ctx
->current_block
;
412 ctx
->after_block
= create_empty_block(ctx
);
414 /* Now that we have the subblocks emitted, fix up the branches */
419 if (ctx
->instruction_count
== count_in
) {
420 /* The else block is empty, so don't emit an exit jump */
421 bi_remove_instruction(then_exit
);
422 then_branch
->branch
.target
= ctx
->after_block
;
424 then_branch
->branch
.target
= else_block
;
425 then_exit
->branch
.target
= ctx
->after_block
;
426 bi_block_add_successor(end_then_block
, then_exit
->branch
.target
);
429 /* Wire up the successors */
431 bi_block_add_successor(before_block
, then_branch
->branch
.target
); /* then_branch */
433 bi_block_add_successor(before_block
, then_block
); /* fallthrough */
434 bi_block_add_successor(end_else_block
, ctx
->after_block
); /* fallthrough */
438 emit_loop(bi_context
*ctx
, nir_loop
*nloop
)
440 /* Remember where we are */
441 bi_block
*start_block
= ctx
->current_block
;
443 bi_block
*saved_break
= ctx
->break_block
;
444 bi_block
*saved_continue
= ctx
->continue_block
;
446 ctx
->continue_block
= create_empty_block(ctx
);
447 ctx
->break_block
= create_empty_block(ctx
);
448 ctx
->after_block
= ctx
->continue_block
;
450 /* Emit the body itself */
451 emit_cf_list(ctx
, &nloop
->body
);
453 /* Branch back to loop back */
454 bi_instruction
*br_back
= bi_emit_branch(ctx
);
455 br_back
->branch
.target
= ctx
->continue_block
;
456 bi_block_add_successor(start_block
, ctx
->continue_block
);
457 bi_block_add_successor(ctx
->current_block
, ctx
->continue_block
);
459 ctx
->after_block
= ctx
->break_block
;
462 ctx
->break_block
= saved_break
;
463 ctx
->continue_block
= saved_continue
;
468 emit_cf_list(bi_context
*ctx
, struct exec_list
*list
)
470 bi_block
*start_block
= NULL
;
472 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
473 switch (node
->type
) {
474 case nir_cf_node_block
: {
475 bi_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
484 emit_if(ctx
, nir_cf_node_as_if(node
));
487 case nir_cf_node_loop
:
488 emit_loop(ctx
, nir_cf_node_as_loop(node
));
492 unreachable("Unknown control flow");
500 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
502 return glsl_count_attribute_slots(type
, false);
506 bi_optimize_nir(nir_shader
*nir
)
509 unsigned lower_flrp
= 16 | 32 | 64;
511 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
512 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
514 nir_lower_tex_options lower_tex_options
= {
515 .lower_txs_lod
= true,
517 .lower_tex_without_implicit_lod
= true,
521 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
522 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
523 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
528 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
529 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
531 NIR_PASS(progress
, nir
, nir_copy_prop
);
532 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
533 NIR_PASS(progress
, nir
, nir_opt_dce
);
534 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
535 NIR_PASS(progress
, nir
, nir_opt_cse
);
536 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
537 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
538 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
540 if (lower_flrp
!= 0) {
541 bool lower_flrp_progress
= false;
542 NIR_PASS(lower_flrp_progress
,
546 false /* always_precise */,
547 nir
->options
->lower_ffma
);
548 if (lower_flrp_progress
) {
549 NIR_PASS(progress
, nir
,
550 nir_opt_constant_folding
);
554 /* Nothing should rematerialize any flrps, so we only
555 * need to do this lowering once.
560 NIR_PASS(progress
, nir
, nir_opt_undef
);
561 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
564 nir_var_function_temp
);
567 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
568 NIR_PASS(progress
, nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
569 NIR_PASS(progress
, nir
, nir_lower_load_const_to_scalar
);
571 /* Take us out of SSA */
572 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
573 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
577 bifrost_compile_shader_nir(nir_shader
*nir
, bifrost_program
*program
, unsigned product_id
)
579 bi_context
*ctx
= rzalloc(NULL
, bi_context
);
581 ctx
->stage
= nir
->info
.stage
;
582 ctx
->quirks
= bifrost_get_quirks(product_id
);
583 list_inithead(&ctx
->blocks
);
585 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
586 * (so we don't accidentally duplicate the epilogue since mesa/st has
587 * messed with our I/O quite a bit already) */
589 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
591 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
592 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
593 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
596 NIR_PASS_V(nir
, nir_split_var_copies
);
597 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
598 NIR_PASS_V(nir
, nir_lower_var_copies
);
599 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
600 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
601 NIR_PASS_V(nir
, nir_lower_ssbo
);
603 bi_optimize_nir(nir
);
604 nir_print_shader(nir
, stdout
);
606 nir_foreach_function(func
, nir
) {
610 ctx
->impl
= func
->impl
;
611 emit_cf_list(ctx
, &func
->impl
->body
);
612 break; /* TODO: Multi-function shaders */
615 bi_print_shader(ctx
, stdout
);
617 bi_print_shader(ctx
, stdout
);