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
, unsigned *size
)
249 nir_foreach_variable(uniform
, uniform_list
) {
253 * UBO's have their own address spaces, so don't count them towards the
254 * number of global uniforms
256 if (uniform
->data
.mode
== nir_var_mem_ubo
|| uniform
->data
.mode
== nir_var_mem_ssbo
)
259 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
260 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
261 if (type
->is_sampler()) {
263 shaderidx
+= type_size(uniform
->type
);
266 imageidx
+= type_size(uniform
->type
);
268 } else if (strncmp(uniform
->name
, "gl_", 3) == 0) {
269 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
270 /* This state reference has already been setup by ir_to_mesa, but we'll
271 * get the same index back here.
275 if (glsl_type_is_struct_or_ifc(type
)) {
278 comps
= glsl_get_vector_elements(type
);
281 if (ctx
->Const
.PackedDriverUniformStorage
) {
282 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
283 stateTokens
, comps
, false);
284 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
286 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
289 loc
= st_nir_lookup_parameter_index(prog
->Parameters
, uniform
->name
);
291 if (ctx
->Const
.PackedDriverUniformStorage
) {
292 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
296 uniform
->data
.driver_location
= loc
;
298 max
= MAX2(max
, loc
+ type_size(uniform
->type
));
304 st_nir_opts(nir_shader
*nir
, bool scalar
)
310 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
313 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
);
314 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
317 NIR_PASS_V(nir
, nir_lower_alu
);
318 NIR_PASS_V(nir
, nir_lower_pack
);
319 NIR_PASS(progress
, nir
, nir_copy_prop
);
320 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
321 NIR_PASS(progress
, nir
, nir_opt_dce
);
322 if (nir_opt_trivial_continues(nir
)) {
324 NIR_PASS(progress
, nir
, nir_copy_prop
);
325 NIR_PASS(progress
, nir
, nir_opt_dce
);
327 NIR_PASS(progress
, nir
, nir_opt_if
);
328 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
329 NIR_PASS(progress
, nir
, nir_opt_cse
);
330 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
332 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
333 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
335 NIR_PASS(progress
, nir
, nir_opt_undef
);
336 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
337 if (nir
->options
->max_unroll_iterations
) {
338 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
343 /* First third of converting glsl_to_nir.. this leaves things in a pre-
344 * nir_lower_io state, so that shader variants can more easily insert/
345 * replace variables, etc.
348 st_glsl_to_nir(struct st_context
*st
, struct gl_program
*prog
,
349 struct gl_shader_program
*shader_program
,
350 gl_shader_stage stage
)
352 const nir_shader_compiler_options
*options
=
353 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
354 enum pipe_shader_type type
= pipe_shader_type_from_mesa(stage
);
355 struct pipe_screen
*screen
= st
->pipe
->screen
;
356 bool is_scalar
= screen
->get_shader_param(screen
, type
, PIPE_SHADER_CAP_SCALAR_ISA
);
359 options
->lower_int64_options
|| options
->lower_doubles_options
;
364 nir_shader
*nir
= glsl_to_nir(shader_program
, stage
, options
);
366 /* Set the next shader stage hint for VS and TES. */
367 if (!nir
->info
.separate_shader
&&
368 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
369 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
371 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
372 unsigned stages_mask
=
373 ~prev_stages
& shader_program
->data
->linked_stages
;
375 nir
->info
.next_stage
= stages_mask
?
376 (gl_shader_stage
) u_bit_scan(&stages_mask
) : MESA_SHADER_FRAGMENT
;
378 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
381 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
382 if (nir
->info
.uses_64bit
&&
383 (options
->lower_doubles_options
& nir_lower_fp64_full_software
) != 0) {
384 nir_shader
*fp64
= glsl_float64_funcs_to_nir(st
->ctx
, options
);
385 ralloc_steal(ralloc_parent(nir
), fp64
);
386 exec_list_append(&nir
->functions
, &fp64
->functions
);
389 nir_variable_mode mask
=
390 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
391 nir_remove_dead_variables(nir
, mask
);
393 if (options
->lower_all_io_to_temps
||
394 nir
->info
.stage
== MESA_SHADER_VERTEX
||
395 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
396 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
397 nir_shader_get_entrypoint(nir
),
399 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
400 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
401 nir_shader_get_entrypoint(nir
),
405 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
406 NIR_PASS_V(nir
, nir_split_var_copies
);
407 NIR_PASS_V(nir
, nir_lower_var_copies
);
410 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
);
414 bool lowered_64bit_ops
= false;
415 bool progress
= false;
417 NIR_PASS_V(nir
, nir_opt_algebraic
);
421 if (options
->lower_int64_options
) {
422 NIR_PASS(progress
, nir
, nir_lower_int64
,
423 options
->lower_int64_options
);
425 if (options
->lower_doubles_options
) {
426 NIR_PASS(progress
, nir
, nir_lower_doubles
,
427 options
->lower_doubles_options
);
429 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
430 lowered_64bit_ops
|= progress
;
433 if (lowered_64bit_ops
) {
434 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_function_temp
);
435 NIR_PASS_V(nir
, nir_lower_returns
);
436 NIR_PASS_V(nir
, nir_inline_functions
);
437 NIR_PASS_V(nir
, nir_opt_deref
);
440 const nir_function
*entry_point
= nir_shader_get_entrypoint(nir
)->function
;
441 foreach_list_typed_safe(nir_function
, func
, node
, &nir
->functions
) {
442 if (func
!= entry_point
) {
443 exec_node_remove(&func
->node
);
446 assert(exec_list_length(&nir
->functions
) == 1);
449 st_nir_opts(nir
, is_scalar
);
454 /* Second third of converting glsl_to_nir. This creates uniforms, gathers
455 * info on varyings, etc after NIR link time opts have been applied.
458 st_glsl_to_nir_post_opts(struct st_context
*st
, struct gl_program
*prog
,
459 struct gl_shader_program
*shader_program
)
461 nir_shader
*nir
= prog
->nir
;
463 /* Make a pass over the IR to add state references for any built-in
464 * uniforms that are used. This has to be done now (during linking).
465 * Code generation doesn't happen until the first time this shader is
466 * used for rendering. Waiting until then to generate the parameters is
467 * too late. At that point, the values for the built-in uniforms won't
468 * get sent to the shader.
470 nir_foreach_variable(var
, &nir
->uniforms
) {
471 if (strncmp(var
->name
, "gl_", 3) == 0) {
472 const nir_state_slot
*const slots
= var
->state_slots
;
473 assert(var
->state_slots
!= NULL
);
475 const struct glsl_type
*type
= glsl_without_array(var
->type
);
476 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
478 if (glsl_type_is_struct_or_ifc(type
)) {
479 /* Builtin struct require specical handling for now we just
480 * make all members vec4. See st_nir_lower_builtin.
484 comps
= glsl_get_vector_elements(type
);
487 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
488 _mesa_add_sized_state_reference(prog
->Parameters
,
492 _mesa_add_state_reference(prog
->Parameters
,
499 /* Avoid reallocation of the program parameter list, because the uniform
500 * storage is only associated with the original parameter list.
501 * This should be enough for Bitmap and DrawPixels constants.
503 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
505 /* This has to be done last. Any operation the can cause
506 * prog->ParameterValues to get reallocated (e.g., anything that adds a
507 * program constant) has to happen before creating this linkage.
509 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
, true);
511 st_set_prog_affected_state_flags(prog
);
513 NIR_PASS_V(nir
, st_nir_lower_builtin
);
514 NIR_PASS_V(nir
, gl_nir_lower_atomics
, shader_program
, true);
516 nir_variable_mode mask
= nir_var_function_temp
;
517 nir_remove_dead_variables(nir
, mask
);
519 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
521 _mesa_log("NIR IR for linked %s program %d:\n",
522 _mesa_shader_stage_to_string(prog
->info
.stage
),
523 shader_program
->Name
);
524 nir_print_shader(nir
, _mesa_get_log_file());
529 /* TODO any better helper somewhere to sort a list? */
532 insert_sorted(struct exec_list
*var_list
, nir_variable
*new_var
)
534 nir_foreach_variable(var
, var_list
) {
535 if (var
->data
.location
> new_var
->data
.location
) {
536 exec_node_insert_node_before(&var
->node
, &new_var
->node
);
540 exec_list_push_tail(var_list
, &new_var
->node
);
544 sort_varyings(struct exec_list
*var_list
)
546 struct exec_list new_list
;
547 exec_list_make_empty(&new_list
);
548 nir_foreach_variable_safe(var
, var_list
) {
549 exec_node_remove(&var
->node
);
550 insert_sorted(&new_list
, var
);
552 exec_list_move_nodes_to(&new_list
, var_list
);
556 set_st_program(struct gl_program
*prog
,
557 struct gl_shader_program
*shader_program
,
560 struct st_vertex_program
*stvp
;
561 struct st_common_program
*stp
;
562 struct st_fragment_program
*stfp
;
563 struct st_compute_program
*stcp
;
565 switch (prog
->info
.stage
) {
566 case MESA_SHADER_VERTEX
:
567 stvp
= (struct st_vertex_program
*)prog
;
568 stvp
->shader_program
= shader_program
;
569 stvp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
570 stvp
->tgsi
.ir
.nir
= nir
;
572 case MESA_SHADER_GEOMETRY
:
573 case MESA_SHADER_TESS_CTRL
:
574 case MESA_SHADER_TESS_EVAL
:
575 stp
= (struct st_common_program
*)prog
;
576 stp
->shader_program
= shader_program
;
577 stp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
578 stp
->tgsi
.ir
.nir
= nir
;
580 case MESA_SHADER_FRAGMENT
:
581 stfp
= (struct st_fragment_program
*)prog
;
582 stfp
->shader_program
= shader_program
;
583 stfp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
584 stfp
->tgsi
.ir
.nir
= nir
;
586 case MESA_SHADER_COMPUTE
:
587 stcp
= (struct st_compute_program
*)prog
;
588 stcp
->shader_program
= shader_program
;
589 stcp
->tgsi
.ir_type
= PIPE_SHADER_IR_NIR
;
590 stcp
->tgsi
.prog
= nir
;
593 unreachable("unknown shader stage");
598 st_nir_get_mesa_program(struct gl_context
*ctx
,
599 struct gl_shader_program
*shader_program
,
600 struct gl_linked_shader
*shader
)
602 struct st_context
*st
= st_context(ctx
);
603 struct pipe_screen
*pscreen
= ctx
->st
->pipe
->screen
;
604 struct gl_program
*prog
;
606 validate_ir_tree(shader
->ir
);
608 prog
= shader
->Program
;
610 prog
->Parameters
= _mesa_new_parameter_list();
612 _mesa_copy_linked_program_data(shader_program
, shader
);
613 _mesa_generate_parameters_list_for_uniforms(ctx
, shader_program
, shader
,
616 /* Remove reads from output registers. */
617 if (!pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
))
618 lower_output_reads(shader
->Stage
, shader
->ir
);
620 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
622 _mesa_log("GLSL IR for linked %s program %d:\n",
623 _mesa_shader_stage_to_string(shader
->Stage
),
624 shader_program
->Name
);
625 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
629 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
630 _mesa_update_shader_textures_used(shader_program
, prog
);
632 nir_shader
*nir
= st_glsl_to_nir(st
, prog
, shader_program
, shader
->Stage
);
634 set_st_program(prog
, shader_program
, nir
);
639 st_nir_link_shaders(nir_shader
**producer
, nir_shader
**consumer
, bool scalar
)
642 NIR_PASS_V(*producer
, nir_lower_io_to_scalar_early
, nir_var_shader_out
);
643 NIR_PASS_V(*consumer
, nir_lower_io_to_scalar_early
, nir_var_shader_in
);
646 nir_lower_io_arrays_to_elements(*producer
, *consumer
);
648 st_nir_opts(*producer
, scalar
);
649 st_nir_opts(*consumer
, scalar
);
651 if (nir_link_opt_varyings(*producer
, *consumer
))
652 st_nir_opts(*consumer
, scalar
);
654 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
655 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
657 if (nir_remove_unused_varyings(*producer
, *consumer
)) {
658 NIR_PASS_V(*producer
, nir_lower_global_vars_to_local
);
659 NIR_PASS_V(*consumer
, nir_lower_global_vars_to_local
);
661 /* The backend might not be able to handle indirects on
662 * temporaries so we need to lower indirects on any of the
663 * varyings we have demoted here.
665 * TODO: radeonsi shouldn't need to do this, however LLVM isn't
666 * currently smart enough to handle indirects without causing excess
667 * spilling causing the gpu to hang.
669 * See the following thread for more details of the problem:
670 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
672 nir_variable_mode indirect_mask
= nir_var_function_temp
;
674 NIR_PASS_V(*producer
, nir_lower_indirect_derefs
, indirect_mask
);
675 NIR_PASS_V(*consumer
, nir_lower_indirect_derefs
, indirect_mask
);
677 st_nir_opts(*producer
, scalar
);
678 st_nir_opts(*consumer
, scalar
);
680 /* Lowering indirects can cause varying to become unused.
681 * nir_compact_varyings() depends on all dead varyings being removed so
682 * we need to call nir_remove_dead_variables() again here.
684 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
685 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
690 st_lower_patch_vertices_in(struct gl_shader_program
*shader_prog
)
692 struct gl_linked_shader
*linked_tcs
=
693 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_CTRL
];
694 struct gl_linked_shader
*linked_tes
=
695 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_EVAL
];
697 /* If we have a TCS and TES linked together, lower TES patch vertices. */
698 if (linked_tcs
&& linked_tes
) {
699 nir_shader
*tcs_nir
= linked_tcs
->Program
->nir
;
700 nir_shader
*tes_nir
= linked_tes
->Program
->nir
;
702 /* The TES input vertex count is the TCS output vertex count,
703 * lower TES gl_PatchVerticesIn to a constant.
705 uint32_t tes_patch_verts
= tcs_nir
->info
.tess
.tcs_vertices_out
;
706 NIR_PASS_V(tes_nir
, nir_lower_patch_vertices
, tes_patch_verts
, NULL
);
713 st_nir_lower_wpos_ytransform(struct nir_shader
*nir
,
714 struct gl_program
*prog
,
715 struct pipe_screen
*pscreen
)
717 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
720 static const gl_state_index16 wposTransformState
[STATE_LENGTH
] = {
721 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
723 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
725 memcpy(wpos_options
.state_tokens
, wposTransformState
,
726 sizeof(wpos_options
.state_tokens
));
727 wpos_options
.fs_coord_origin_upper_left
=
728 pscreen
->get_param(pscreen
,
729 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
730 wpos_options
.fs_coord_origin_lower_left
=
731 pscreen
->get_param(pscreen
,
732 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
733 wpos_options
.fs_coord_pixel_center_integer
=
734 pscreen
->get_param(pscreen
,
735 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
736 wpos_options
.fs_coord_pixel_center_half_integer
=
737 pscreen
->get_param(pscreen
,
738 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
740 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
741 nir_validate_shader(nir
, "after nir_lower_wpos_ytransform");
742 _mesa_add_state_reference(prog
->Parameters
, wposTransformState
);
747 st_link_nir(struct gl_context
*ctx
,
748 struct gl_shader_program
*shader_program
)
750 struct st_context
*st
= st_context(ctx
);
751 struct pipe_screen
*screen
= st
->pipe
->screen
;
752 bool is_scalar
[MESA_SHADER_STAGES
];
754 unsigned last_stage
= 0;
755 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
756 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
760 /* Determine scalar property of each shader stage */
761 enum pipe_shader_type type
= pipe_shader_type_from_mesa(shader
->Stage
);
762 is_scalar
[i
] = screen
->get_shader_param(screen
, type
,
763 PIPE_SHADER_CAP_SCALAR_ISA
);
765 st_nir_get_mesa_program(ctx
, shader_program
, shader
);
769 NIR_PASS_V(shader
->Program
->nir
, nir_lower_load_const_to_scalar
);
773 /* Linking the stages in the opposite order (from fragment to vertex)
774 * ensures that inter-shader outputs written to in an earlier stage
775 * are eliminated if they are (transitively) not used in a later
778 int next
= last_stage
;
779 for (int i
= next
- 1; i
>= 0; i
--) {
780 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
784 st_nir_link_shaders(&shader
->Program
->nir
,
785 &shader_program
->_LinkedShaders
[next
]->Program
->nir
,
791 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
792 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
796 nir_shader
*nir
= shader
->Program
->nir
;
798 NIR_PASS_V(nir
, st_nir_lower_wpos_ytransform
, shader
->Program
,
801 NIR_PASS_V(nir
, nir_lower_system_values
);
802 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
804 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
805 shader
->Program
->info
= nir
->info
;
806 if (i
== MESA_SHADER_VERTEX
) {
807 /* NIR expands dual-slot inputs out to two locations. We need to
808 * compact things back down GL-style single-slot inputs to avoid
809 * confusing the state tracker.
811 shader
->Program
->info
.inputs_read
=
812 nir_get_single_slot_attribs_mask(nir
->info
.inputs_read
,
813 shader
->Program
->DualSlotInputs
);
817 struct gl_program
*prev_shader
=
818 shader_program
->_LinkedShaders
[prev
]->Program
;
820 /* We can't use nir_compact_varyings with transform feedback, since
821 * the pipe_stream_output->output_register field is based on the
822 * pre-compacted driver_locations.
824 if (!(prev_shader
->sh
.LinkedTransformFeedback
&&
825 prev_shader
->sh
.LinkedTransformFeedback
->NumVarying
> 0))
826 nir_compact_varyings(shader_program
->_LinkedShaders
[prev
]->Program
->nir
,
827 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
832 st_lower_patch_vertices_in(shader_program
);
834 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
835 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
839 st_glsl_to_nir_post_opts(st
, shader
->Program
, shader_program
);
841 assert(shader
->Program
);
842 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
843 _mesa_shader_stage_to_program(i
),
845 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
849 nir_sweep(shader
->Program
->nir
);
856 st_nir_assign_varying_locations(struct st_context
*st
, nir_shader
*nir
)
858 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
859 /* Needs special handling so drvloc matches the vbo state: */
860 st_nir_assign_vs_in_locations(nir
);
861 /* Re-lower global vars, to deal with any dead VS inputs. */
862 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
864 sort_varyings(&nir
->outputs
);
865 st_nir_assign_var_locations(&nir
->outputs
,
868 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
869 } else if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
870 nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
871 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
872 sort_varyings(&nir
->inputs
);
873 st_nir_assign_var_locations(&nir
->inputs
,
876 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
878 sort_varyings(&nir
->outputs
);
879 st_nir_assign_var_locations(&nir
->outputs
,
882 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
883 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
884 sort_varyings(&nir
->inputs
);
885 st_nir_assign_var_locations(&nir
->inputs
,
888 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
889 st_nir_assign_var_locations(&nir
->outputs
,
892 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
895 unreachable("invalid shader type");
900 st_nir_lower_samplers(struct pipe_screen
*screen
, nir_shader
*nir
,
901 struct gl_shader_program
*shader_program
,
902 struct gl_program
*prog
)
904 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
905 NIR_PASS_V(nir
, gl_nir_lower_samplers_as_deref
, shader_program
);
907 NIR_PASS_V(nir
, gl_nir_lower_samplers
, shader_program
);
910 prog
->info
.textures_used
= nir
->info
.textures_used
;
911 prog
->info
.textures_used_by_txf
= nir
->info
.textures_used_by_txf
;
915 /* Last third of preparing nir from glsl, which happens after shader
919 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
920 struct gl_shader_program
*shader_program
, nir_shader
*nir
)
922 struct pipe_screen
*screen
= st
->pipe
->screen
;
923 const nir_shader_compiler_options
*options
=
924 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
926 NIR_PASS_V(nir
, nir_split_var_copies
);
927 NIR_PASS_V(nir
, nir_lower_var_copies
);
928 if (options
->lower_all_io_to_temps
||
929 nir
->info
.stage
== MESA_SHADER_VERTEX
||
930 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
931 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, false);
932 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
933 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, true);
936 st_nir_assign_varying_locations(st
, nir
);
938 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
,
939 st
->ctx
->Const
.Program
[nir
->info
.stage
].MaxAtomicBuffers
);
941 st_nir_assign_uniform_locations(st
->ctx
, prog
,
942 &nir
->uniforms
, &nir
->num_uniforms
);
944 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
945 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_type_dword_size
,
946 (nir_lower_io_options
)0);
947 NIR_PASS_V(nir
, nir_lower_uniforms_to_ubo
, 4);
950 st_nir_lower_samplers(screen
, nir
, shader_program
, prog
);