2 * Copyright © 2015 Red Hat
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
26 #include "pipe/p_defines.h"
27 #include "pipe/p_screen.h"
28 #include "pipe/p_context.h"
30 #include "program/program.h"
31 #include "program/prog_statevars.h"
32 #include "program/prog_parameter.h"
33 #include "program/ir_to_mesa.h"
34 #include "main/mtypes.h"
35 #include "main/errors.h"
36 #include "main/shaderapi.h"
37 #include "main/uniforms.h"
39 #include "main/shaderobj.h"
40 #include "st_context.h"
41 #include "st_glsl_types.h"
42 #include "st_program.h"
44 #include "compiler/nir/nir.h"
45 #include "compiler/glsl_types.h"
46 #include "compiler/glsl/glsl_to_nir.h"
47 #include "compiler/glsl/gl_nir.h"
48 #include "compiler/glsl/ir.h"
49 #include "compiler/glsl/ir_optimization.h"
50 #include "compiler/glsl/string_to_uint_map.h"
53 type_size(const struct glsl_type
*type
)
55 return type
->count_attribute_slots(false);
58 /* Depending on PIPE_CAP_TGSI_TEXCOORD (st->needs_texcoord_semantic) we
59 * may need to fix up varying slots so the glsl->nir path is aligned
60 * with the anything->tgsi->nir path.
63 st_nir_fixup_varying_slots(struct st_context
*st
, struct exec_list
*var_list
)
65 if (st
->needs_texcoord_semantic
)
68 nir_foreach_variable(var
, var_list
) {
69 if (var
->data
.location
>= VARYING_SLOT_VAR0
) {
70 var
->data
.location
+= 9;
71 } else if ((var
->data
.location
>= VARYING_SLOT_TEX0
) &&
72 (var
->data
.location
<= VARYING_SLOT_TEX7
)) {
73 var
->data
.location
+= VARYING_SLOT_VAR0
- VARYING_SLOT_TEX0
;
78 /* input location assignment for VS inputs must be handled specially, so
79 * that it is aligned w/ st's vbo state.
80 * (This isn't the case with, for ex, FS inputs, which only need to agree
81 * on varying-slot w/ the VS outputs)
84 st_nir_assign_vs_in_locations(nir_shader
*nir
)
87 nir_foreach_variable_safe(var
, &nir
->inputs
) {
88 /* NIR already assigns dual-slot inputs to two locations so all we have
89 * to do is compact everything down.
91 if (var
->data
.location
== VERT_ATTRIB_EDGEFLAG
) {
92 /* bit of a hack, mirroring st_translate_vertex_program */
93 var
->data
.driver_location
= util_bitcount64(nir
->info
.inputs_read
);
94 } else if (nir
->info
.inputs_read
& BITFIELD64_BIT(var
->data
.location
)) {
95 var
->data
.driver_location
=
96 util_bitcount64(nir
->info
.inputs_read
&
97 BITFIELD64_MASK(var
->data
.location
));
100 /* Move unused input variables to the globals list (with no
101 * initialization), to avoid confusing drivers looking through the
102 * inputs array and expecting to find inputs with a driver_location
105 exec_node_remove(&var
->node
);
106 var
->data
.mode
= nir_var_shader_temp
;
107 exec_list_push_tail(&nir
->globals
, &var
->node
);
113 st_nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
114 gl_shader_stage stage
)
116 unsigned location
= 0;
117 unsigned assigned_locations
[VARYING_SLOT_TESS_MAX
];
118 uint64_t processed_locs
[2] = {0};
120 const int base
= stage
== MESA_SHADER_FRAGMENT
?
121 (int) FRAG_RESULT_DATA0
: (int) VARYING_SLOT_VAR0
;
123 int UNUSED last_loc
= 0;
124 nir_foreach_variable(var
, var_list
) {
126 const struct glsl_type
*type
= var
->type
;
127 if (nir_is_per_vertex_io(var
, stage
)) {
128 assert(glsl_type_is_array(type
));
129 type
= glsl_get_array_element(type
);
132 unsigned var_size
= type_size(type
);
134 /* Builtins don't allow component packing so we only need to worry about
135 * user defined varyings sharing the same location.
137 bool processed
= false;
138 if (var
->data
.location
>= base
) {
139 unsigned glsl_location
= var
->data
.location
- base
;
141 for (unsigned i
= 0; i
< var_size
; i
++) {
142 if (processed_locs
[var
->data
.index
] &
143 ((uint64_t)1 << (glsl_location
+ i
)))
146 processed_locs
[var
->data
.index
] |=
147 ((uint64_t)1 << (glsl_location
+ i
));
151 /* Because component packing allows varyings to share the same location
152 * we may have already have processed this location.
155 unsigned driver_location
= assigned_locations
[var
->data
.location
];
156 var
->data
.driver_location
= driver_location
;
157 *size
+= type_size(type
);
159 /* An array may be packed such that is crosses multiple other arrays
160 * or variables, we need to make sure we have allocated the elements
161 * consecutively if the previously proccessed var was shorter than
162 * the current array we are processing.
164 * NOTE: The code below assumes the var list is ordered in ascending
167 assert(last_loc
<= var
->data
.location
);
168 last_loc
= var
->data
.location
;
169 unsigned last_slot_location
= driver_location
+ var_size
;
170 if (last_slot_location
> location
) {
171 unsigned num_unallocated_slots
= last_slot_location
- location
;
172 unsigned first_unallocated_slot
= var_size
- num_unallocated_slots
;
173 for (unsigned i
= first_unallocated_slot
; i
< num_unallocated_slots
; i
++) {
174 assigned_locations
[var
->data
.location
+ i
] = location
;
181 for (unsigned i
= 0; i
< var_size
; i
++) {
182 assigned_locations
[var
->data
.location
+ i
] = location
+ i
;
185 var
->data
.driver_location
= location
;
186 location
+= var_size
;
193 st_nir_lookup_parameter_index(const struct gl_program_parameter_list
*params
,
196 int loc
= _mesa_lookup_parameter_index(params
, name
);
198 /* is there a better way to do this? If we have something like:
206 * Then what we get in prog->Parameters looks like:
208 * 0: Name=color.f, Type=6, DataType=1406, Size=1
209 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
211 * So the name doesn't match up and _mesa_lookup_parameter_index()
212 * fails. In this case just find the first matching "color.*"..
214 * Note for arrays you could end up w/ color[n].f, for example.
216 * glsl_to_tgsi works slightly differently in this regard. It is
217 * emitting something more low level, so it just translates the
218 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
219 * it just calculates the additional offset of struct field members
220 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
221 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
222 * needs to work backwards to get base var loc from the param-list
223 * which already has them separated out.
226 int namelen
= strlen(name
);
227 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
228 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
229 if ((strncmp(p
->Name
, name
, namelen
) == 0) &&
230 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
241 st_nir_assign_uniform_locations(struct gl_context
*ctx
,
242 struct gl_program
*prog
,
243 struct exec_list
*uniform_list
)
248 nir_foreach_variable(uniform
, uniform_list
) {
252 * UBO's have their own address spaces, so don't count them towards the
253 * number of global uniforms
255 if (uniform
->data
.mode
== nir_var_mem_ubo
|| uniform
->data
.mode
== nir_var_mem_ssbo
)
258 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
259 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
260 if (type
->is_sampler()) {
262 shaderidx
+= type_size(uniform
->type
);
265 imageidx
+= type_size(uniform
->type
);
267 } else if (strncmp(uniform
->name
, "gl_", 3) == 0) {
268 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
269 /* This state reference has already been setup by ir_to_mesa, but we'll
270 * get the same index back here.
274 if (glsl_type_is_struct_or_ifc(type
)) {
277 comps
= glsl_get_vector_elements(type
);
280 if (ctx
->Const
.PackedDriverUniformStorage
) {
281 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
282 stateTokens
, comps
, false);
283 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
285 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
288 loc
= st_nir_lookup_parameter_index(prog
->Parameters
, uniform
->name
);
290 /* We need to check that loc is not -1 here before accessing the
291 * array. It can be negative for example when we have a struct that
292 * only contains opaque types.
294 if (loc
>= 0 && ctx
->Const
.PackedDriverUniformStorage
) {
295 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
299 uniform
->data
.driver_location
= loc
;
304 st_nir_opts(nir_shader
*nir
, bool scalar
)
307 unsigned lower_flrp
=
308 (nir
->options
->lower_flrp16
? 16 : 0) |
309 (nir
->options
->lower_flrp32
? 32 : 0) |
310 (nir
->options
->lower_flrp64
? 64 : 0);
315 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
317 NIR_PASS(progress
, nir
, nir_opt_copy_prop_vars
);
318 NIR_PASS(progress
, nir
, nir_opt_dead_write_vars
);
321 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
);
322 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
325 NIR_PASS_V(nir
, nir_lower_alu
);
326 NIR_PASS_V(nir
, nir_lower_pack
);
327 NIR_PASS(progress
, nir
, nir_copy_prop
);
328 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
329 NIR_PASS(progress
, nir
, nir_opt_dce
);
330 if (nir_opt_trivial_continues(nir
)) {
332 NIR_PASS(progress
, nir
, nir_copy_prop
);
333 NIR_PASS(progress
, nir
, nir_opt_dce
);
335 NIR_PASS(progress
, nir
, nir_opt_if
, false);
336 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
337 NIR_PASS(progress
, nir
, nir_opt_cse
);
338 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
340 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
341 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
343 if (lower_flrp
!= 0) {
344 bool lower_flrp_progress
= false;
346 NIR_PASS(lower_flrp_progress
, nir
, nir_lower_flrp
,
348 false /* always_precise */,
349 nir
->options
->lower_ffma
);
350 if (lower_flrp_progress
) {
351 NIR_PASS(progress
, nir
,
352 nir_opt_constant_folding
);
356 /* Nothing should rematerialize any flrps, so we only need to do this
362 NIR_PASS(progress
, nir
, gl_nir_opt_access
);
364 NIR_PASS(progress
, nir
, nir_opt_undef
);
365 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
366 if (nir
->options
->max_unroll_iterations
) {
367 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
372 /* First third of converting glsl_to_nir.. this leaves things in a pre-
373 * nir_lower_io state, so that shader variants can more easily insert/
374 * replace variables, etc.
377 st_glsl_to_nir(struct st_context
*st
, struct gl_program
*prog
,
378 struct gl_shader_program
*shader_program
,
379 gl_shader_stage stage
)
381 const nir_shader_compiler_options
*options
=
382 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
383 enum pipe_shader_type type
= pipe_shader_type_from_mesa(stage
);
384 struct pipe_screen
*screen
= st
->pipe
->screen
;
385 bool is_scalar
= screen
->get_shader_param(screen
, type
, PIPE_SHADER_CAP_SCALAR_ISA
);
388 options
->lower_int64_options
|| options
->lower_doubles_options
;
393 nir_shader
*nir
= glsl_to_nir(st
->ctx
, shader_program
, stage
, options
);
395 /* Set the next shader stage hint for VS and TES. */
396 if (!nir
->info
.separate_shader
&&
397 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
398 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
400 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
401 unsigned stages_mask
=
402 ~prev_stages
& shader_program
->data
->linked_stages
;
404 nir
->info
.next_stage
= stages_mask
?
405 (gl_shader_stage
) u_bit_scan(&stages_mask
) : MESA_SHADER_FRAGMENT
;
407 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
410 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
411 nir_shader
*softfp64
= NULL
;
412 if (nir
->info
.uses_64bit
&&
413 (options
->lower_doubles_options
& nir_lower_fp64_full_software
) != 0) {
414 softfp64
= glsl_float64_funcs_to_nir(st
->ctx
, options
);
415 ralloc_steal(ralloc_parent(nir
), softfp64
);
418 nir_variable_mode mask
=
419 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
420 nir_remove_dead_variables(nir
, mask
);
422 if (options
->lower_all_io_to_temps
||
423 nir
->info
.stage
== MESA_SHADER_VERTEX
||
424 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
425 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
426 nir_shader_get_entrypoint(nir
),
428 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
429 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
430 nir_shader_get_entrypoint(nir
),
434 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
435 NIR_PASS_V(nir
, nir_split_var_copies
);
436 NIR_PASS_V(nir
, nir_lower_var_copies
);
439 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
);
442 /* before buffers and vars_to_ssa */
443 NIR_PASS_V(nir
, gl_nir_lower_bindless_images
);
444 st_nir_opts(nir
, is_scalar
);
446 NIR_PASS_V(nir
, gl_nir_lower_buffers
, shader_program
);
447 /* Do a round of constant folding to clean up address calculations */
448 NIR_PASS_V(nir
, nir_opt_constant_folding
);
451 bool lowered_64bit_ops
= false;
452 bool progress
= false;
454 NIR_PASS_V(nir
, nir_opt_algebraic
);
458 if (options
->lower_int64_options
) {
459 NIR_PASS(progress
, nir
, nir_lower_int64
,
460 options
->lower_int64_options
);
462 if (options
->lower_doubles_options
) {
463 NIR_PASS(progress
, nir
, nir_lower_doubles
,
464 softfp64
, options
->lower_doubles_options
);
466 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
467 lowered_64bit_ops
|= progress
;
470 if (lowered_64bit_ops
)
471 st_nir_opts(nir
, is_scalar
);
477 /* Second third of converting glsl_to_nir. This creates uniforms, gathers
478 * info on varyings, etc after NIR link time opts have been applied.
481 st_glsl_to_nir_post_opts(struct st_context
*st
, struct gl_program
*prog
,
482 struct gl_shader_program
*shader_program
)
484 nir_shader
*nir
= prog
->nir
;
486 /* Make a pass over the IR to add state references for any built-in
487 * uniforms that are used. This has to be done now (during linking).
488 * Code generation doesn't happen until the first time this shader is
489 * used for rendering. Waiting until then to generate the parameters is
490 * too late. At that point, the values for the built-in uniforms won't
491 * get sent to the shader.
493 nir_foreach_variable(var
, &nir
->uniforms
) {
494 if (strncmp(var
->name
, "gl_", 3) == 0) {
495 const nir_state_slot
*const slots
= var
->state_slots
;
496 assert(var
->state_slots
!= NULL
);
498 const struct glsl_type
*type
= glsl_without_array(var
->type
);
499 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
501 if (glsl_type_is_struct_or_ifc(type
)) {
502 /* Builtin struct require specical handling for now we just
503 * make all members vec4. See st_nir_lower_builtin.
507 comps
= glsl_get_vector_elements(type
);
510 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
511 _mesa_add_sized_state_reference(prog
->Parameters
,
515 _mesa_add_state_reference(prog
->Parameters
,
522 /* Avoid reallocation of the program parameter list, because the uniform
523 * storage is only associated with the original parameter list.
524 * This should be enough for Bitmap and DrawPixels constants.
526 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
528 /* This has to be done last. Any operation the can cause
529 * prog->ParameterValues to get reallocated (e.g., anything that adds a
530 * program constant) has to happen before creating this linkage.
532 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
);
534 st_set_prog_affected_state_flags(prog
);
536 NIR_PASS_V(nir
, st_nir_lower_builtin
);
537 NIR_PASS_V(nir
, gl_nir_lower_atomics
, shader_program
, true);
538 NIR_PASS_V(nir
, nir_opt_intrinsics
);
540 nir_variable_mode mask
= nir_var_function_temp
;
541 nir_remove_dead_variables(nir
, mask
);
543 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
545 _mesa_log("NIR IR for linked %s program %d:\n",
546 _mesa_shader_stage_to_string(prog
->info
.stage
),
547 shader_program
->Name
);
548 nir_print_shader(nir
, _mesa_get_log_file());
553 /* TODO any better helper somewhere to sort a list? */
556 insert_sorted(struct exec_list
*var_list
, nir_variable
*new_var
)
558 nir_foreach_variable(var
, var_list
) {
559 if (var
->data
.location
> new_var
->data
.location
) {
560 exec_node_insert_node_before(&var
->node
, &new_var
->node
);
564 exec_list_push_tail(var_list
, &new_var
->node
);
568 sort_varyings(struct exec_list
*var_list
)
570 struct exec_list new_list
;
571 exec_list_make_empty(&new_list
);
572 nir_foreach_variable_safe(var
, var_list
) {
573 exec_node_remove(&var
->node
);
574 insert_sorted(&new_list
, var
);
576 exec_list_move_nodes_to(&new_list
, var_list
);
580 set_st_program(struct gl_program
*prog
,
581 struct gl_shader_program
*shader_program
,
584 struct st_vertex_program
*stvp
;
585 struct st_common_program
*stp
;
586 struct st_fragment_program
*stfp
;
587 struct st_compute_program
*stcp
;
589 switch (prog
->info
.stage
) {
590 case MESA_SHADER_VERTEX
:
591 stvp
= (struct st_vertex_program
*)prog
;
592 stvp
->shader_program
= shader_program
;
593 stvp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
594 stvp
->tgsi
.ir
.nir
= nir
;
596 case MESA_SHADER_GEOMETRY
:
597 case MESA_SHADER_TESS_CTRL
:
598 case MESA_SHADER_TESS_EVAL
:
599 stp
= (struct st_common_program
*)prog
;
600 stp
->shader_program
= shader_program
;
601 stp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
602 stp
->tgsi
.ir
.nir
= nir
;
604 case MESA_SHADER_FRAGMENT
:
605 stfp
= (struct st_fragment_program
*)prog
;
606 stfp
->shader_program
= shader_program
;
607 stfp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
608 stfp
->tgsi
.ir
.nir
= nir
;
610 case MESA_SHADER_COMPUTE
:
611 stcp
= (struct st_compute_program
*)prog
;
612 stcp
->shader_program
= shader_program
;
613 stcp
->tgsi
.ir_type
= PIPE_SHADER_IR_NIR
;
614 stcp
->tgsi
.prog
= nir
;
617 unreachable("unknown shader stage");
622 st_nir_get_mesa_program(struct gl_context
*ctx
,
623 struct gl_shader_program
*shader_program
,
624 struct gl_linked_shader
*shader
)
626 struct st_context
*st
= st_context(ctx
);
627 struct pipe_screen
*pscreen
= ctx
->st
->pipe
->screen
;
628 struct gl_program
*prog
;
630 validate_ir_tree(shader
->ir
);
632 prog
= shader
->Program
;
634 prog
->Parameters
= _mesa_new_parameter_list();
636 _mesa_copy_linked_program_data(shader_program
, shader
);
637 _mesa_generate_parameters_list_for_uniforms(ctx
, shader_program
, shader
,
640 /* Remove reads from output registers. */
641 if (!pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
))
642 lower_output_reads(shader
->Stage
, shader
->ir
);
644 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
646 _mesa_log("GLSL IR for linked %s program %d:\n",
647 _mesa_shader_stage_to_string(shader
->Stage
),
648 shader_program
->Name
);
649 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
653 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
654 _mesa_update_shader_textures_used(shader_program
, prog
);
656 nir_shader
*nir
= st_glsl_to_nir(st
, prog
, shader_program
, shader
->Stage
);
658 set_st_program(prog
, shader_program
, nir
);
663 st_nir_vectorize_io(nir_shader
*producer
, nir_shader
*consumer
)
665 NIR_PASS_V(producer
, nir_lower_io_to_vector
, nir_var_shader_out
);
666 NIR_PASS_V(producer
, nir_opt_combine_stores
, nir_var_shader_out
);
667 NIR_PASS_V(consumer
, nir_lower_io_to_vector
, nir_var_shader_in
);
669 if ((producer
)->info
.stage
!= MESA_SHADER_TESS_CTRL
) {
670 /* Calling lower_io_to_vector creates output variable writes with
671 * write-masks. We only support these for TCS outputs, so for other
672 * stages, we need to call nir_lower_io_to_temporaries to get rid of
673 * them. This, in turn, creates temporary variables and extra
674 * copy_deref intrinsics that we need to clean up.
676 NIR_PASS_V(producer
, nir_lower_io_to_temporaries
,
677 nir_shader_get_entrypoint(producer
), true, false);
678 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
679 NIR_PASS_V(producer
, nir_split_var_copies
);
680 NIR_PASS_V(producer
, nir_lower_var_copies
);
685 st_nir_link_shaders(nir_shader
**producer
, nir_shader
**consumer
, bool scalar
)
688 NIR_PASS_V(*producer
, nir_lower_io_to_scalar_early
, nir_var_shader_out
);
689 NIR_PASS_V(*consumer
, nir_lower_io_to_scalar_early
, nir_var_shader_in
);
692 nir_lower_io_arrays_to_elements(*producer
, *consumer
);
694 st_nir_opts(*producer
, scalar
);
695 st_nir_opts(*consumer
, scalar
);
697 if (nir_link_opt_varyings(*producer
, *consumer
))
698 st_nir_opts(*consumer
, scalar
);
700 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
701 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
703 if (nir_remove_unused_varyings(*producer
, *consumer
)) {
704 NIR_PASS_V(*producer
, nir_lower_global_vars_to_local
);
705 NIR_PASS_V(*consumer
, nir_lower_global_vars_to_local
);
707 /* The backend might not be able to handle indirects on
708 * temporaries so we need to lower indirects on any of the
709 * varyings we have demoted here.
711 * TODO: radeonsi shouldn't need to do this, however LLVM isn't
712 * currently smart enough to handle indirects without causing excess
713 * spilling causing the gpu to hang.
715 * See the following thread for more details of the problem:
716 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
718 nir_variable_mode indirect_mask
= nir_var_function_temp
;
720 NIR_PASS_V(*producer
, nir_lower_indirect_derefs
, indirect_mask
);
721 NIR_PASS_V(*consumer
, nir_lower_indirect_derefs
, indirect_mask
);
723 st_nir_opts(*producer
, scalar
);
724 st_nir_opts(*consumer
, scalar
);
726 /* Lowering indirects can cause varying to become unused.
727 * nir_compact_varyings() depends on all dead varyings being removed so
728 * we need to call nir_remove_dead_variables() again here.
730 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
731 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
736 st_lower_patch_vertices_in(struct gl_shader_program
*shader_prog
)
738 struct gl_linked_shader
*linked_tcs
=
739 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_CTRL
];
740 struct gl_linked_shader
*linked_tes
=
741 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_EVAL
];
743 /* If we have a TCS and TES linked together, lower TES patch vertices. */
744 if (linked_tcs
&& linked_tes
) {
745 nir_shader
*tcs_nir
= linked_tcs
->Program
->nir
;
746 nir_shader
*tes_nir
= linked_tes
->Program
->nir
;
748 /* The TES input vertex count is the TCS output vertex count,
749 * lower TES gl_PatchVerticesIn to a constant.
751 uint32_t tes_patch_verts
= tcs_nir
->info
.tess
.tcs_vertices_out
;
752 NIR_PASS_V(tes_nir
, nir_lower_patch_vertices
, tes_patch_verts
, NULL
);
759 st_nir_lower_wpos_ytransform(struct nir_shader
*nir
,
760 struct gl_program
*prog
,
761 struct pipe_screen
*pscreen
)
763 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
766 static const gl_state_index16 wposTransformState
[STATE_LENGTH
] = {
767 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
769 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
771 memcpy(wpos_options
.state_tokens
, wposTransformState
,
772 sizeof(wpos_options
.state_tokens
));
773 wpos_options
.fs_coord_origin_upper_left
=
774 pscreen
->get_param(pscreen
,
775 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
776 wpos_options
.fs_coord_origin_lower_left
=
777 pscreen
->get_param(pscreen
,
778 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
779 wpos_options
.fs_coord_pixel_center_integer
=
780 pscreen
->get_param(pscreen
,
781 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
782 wpos_options
.fs_coord_pixel_center_half_integer
=
783 pscreen
->get_param(pscreen
,
784 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
786 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
787 nir_validate_shader(nir
, "after nir_lower_wpos_ytransform");
788 _mesa_add_state_reference(prog
->Parameters
, wposTransformState
);
793 st_link_nir(struct gl_context
*ctx
,
794 struct gl_shader_program
*shader_program
)
796 struct st_context
*st
= st_context(ctx
);
797 struct pipe_screen
*screen
= st
->pipe
->screen
;
798 bool is_scalar
[MESA_SHADER_STAGES
];
800 unsigned last_stage
= 0;
801 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
802 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
806 /* Determine scalar property of each shader stage */
807 enum pipe_shader_type type
= pipe_shader_type_from_mesa(shader
->Stage
);
808 is_scalar
[i
] = screen
->get_shader_param(screen
, type
,
809 PIPE_SHADER_CAP_SCALAR_ISA
);
811 st_nir_get_mesa_program(ctx
, shader_program
, shader
);
815 NIR_PASS_V(shader
->Program
->nir
, nir_lower_load_const_to_scalar
);
819 /* Linking the stages in the opposite order (from fragment to vertex)
820 * ensures that inter-shader outputs written to in an earlier stage
821 * are eliminated if they are (transitively) not used in a later
824 int next
= last_stage
;
825 for (int i
= next
- 1; i
>= 0; i
--) {
826 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
830 st_nir_link_shaders(&shader
->Program
->nir
,
831 &shader_program
->_LinkedShaders
[next
]->Program
->nir
,
837 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
838 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
842 nir_shader
*nir
= shader
->Program
->nir
;
844 NIR_PASS_V(nir
, st_nir_lower_wpos_ytransform
, shader
->Program
,
847 NIR_PASS_V(nir
, nir_lower_system_values
);
848 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
850 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
851 shader
->Program
->info
= nir
->info
;
852 if (i
== MESA_SHADER_VERTEX
) {
853 /* NIR expands dual-slot inputs out to two locations. We need to
854 * compact things back down GL-style single-slot inputs to avoid
855 * confusing the state tracker.
857 shader
->Program
->info
.inputs_read
=
858 nir_get_single_slot_attribs_mask(nir
->info
.inputs_read
,
859 shader
->Program
->DualSlotInputs
);
863 struct gl_program
*prev_shader
=
864 shader_program
->_LinkedShaders
[prev
]->Program
;
866 /* We can't use nir_compact_varyings with transform feedback, since
867 * the pipe_stream_output->output_register field is based on the
868 * pre-compacted driver_locations.
870 if (!(prev_shader
->sh
.LinkedTransformFeedback
&&
871 prev_shader
->sh
.LinkedTransformFeedback
->NumVarying
> 0))
872 nir_compact_varyings(shader_program
->_LinkedShaders
[prev
]->Program
->nir
,
873 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
875 if (ctx
->Const
.ShaderCompilerOptions
[i
].NirOptions
->vectorize_io
)
876 st_nir_vectorize_io(prev_shader
->nir
, nir
);
881 st_lower_patch_vertices_in(shader_program
);
883 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
884 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
888 st_glsl_to_nir_post_opts(st
, shader
->Program
, shader_program
);
890 assert(shader
->Program
);
891 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
892 _mesa_shader_stage_to_program(i
),
894 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
898 nir_sweep(shader
->Program
->nir
);
900 /* The GLSL IR won't be needed anymore. */
901 ralloc_free(shader
->ir
);
909 st_nir_assign_varying_locations(struct st_context
*st
, nir_shader
*nir
)
911 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
912 /* Needs special handling so drvloc matches the vbo state: */
913 st_nir_assign_vs_in_locations(nir
);
914 /* Re-lower global vars, to deal with any dead VS inputs. */
915 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
917 sort_varyings(&nir
->outputs
);
918 st_nir_assign_var_locations(&nir
->outputs
,
921 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
922 } else if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
923 nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
924 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
925 sort_varyings(&nir
->inputs
);
926 st_nir_assign_var_locations(&nir
->inputs
,
929 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
931 sort_varyings(&nir
->outputs
);
932 st_nir_assign_var_locations(&nir
->outputs
,
935 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
936 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
937 sort_varyings(&nir
->inputs
);
938 st_nir_assign_var_locations(&nir
->inputs
,
941 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
942 st_nir_assign_var_locations(&nir
->outputs
,
945 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
948 unreachable("invalid shader type");
953 st_nir_lower_samplers(struct pipe_screen
*screen
, nir_shader
*nir
,
954 struct gl_shader_program
*shader_program
,
955 struct gl_program
*prog
)
957 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
958 NIR_PASS_V(nir
, gl_nir_lower_samplers_as_deref
, shader_program
);
960 NIR_PASS_V(nir
, gl_nir_lower_samplers
, shader_program
);
963 prog
->info
.textures_used
= nir
->info
.textures_used
;
964 prog
->info
.textures_used_by_txf
= nir
->info
.textures_used_by_txf
;
968 /* Last third of preparing nir from glsl, which happens after shader
972 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
973 struct gl_shader_program
*shader_program
, nir_shader
*nir
)
975 struct pipe_screen
*screen
= st
->pipe
->screen
;
976 const nir_shader_compiler_options
*options
=
977 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
979 NIR_PASS_V(nir
, nir_split_var_copies
);
980 NIR_PASS_V(nir
, nir_lower_var_copies
);
981 if (options
->lower_all_io_to_temps
||
982 options
->lower_all_io_to_elements
||
983 nir
->info
.stage
== MESA_SHADER_VERTEX
||
984 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
985 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, false);
986 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
987 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, true);
990 st_nir_assign_varying_locations(st
, nir
);
992 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
,
993 st
->ctx
->Const
.Program
[nir
->info
.stage
].MaxAtomicBuffers
);
995 st_nir_assign_uniform_locations(st
->ctx
, prog
,
998 /* Set num_uniforms in number of attribute slots (vec4s) */
999 nir
->num_uniforms
= DIV_ROUND_UP(prog
->Parameters
->NumParameterValues
, 4);
1001 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
1002 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_type_dword_size
,
1003 (nir_lower_io_options
)0);
1004 NIR_PASS_V(nir
, nir_lower_uniforms_to_ubo
, 4);
1006 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_uniforms_type_size
,
1007 (nir_lower_io_options
)0);
1010 st_nir_lower_samplers(screen
, nir
, shader_program
, prog
);