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 "st_context.h"
40 #include "st_glsl_types.h"
41 #include "st_program.h"
43 #include "compiler/nir/nir.h"
44 #include "compiler/glsl_types.h"
45 #include "compiler/glsl/glsl_to_nir.h"
46 #include "compiler/glsl/gl_nir.h"
47 #include "compiler/glsl/ir.h"
48 #include "compiler/glsl/ir_optimization.h"
49 #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(struct gl_program
*prog
, nir_shader
*nir
)
86 unsigned attr
, num_inputs
= 0;
87 unsigned input_to_index
[VERT_ATTRIB_MAX
] = {0};
89 /* TODO de-duplicate w/ similar code in st_translate_vertex_program()? */
90 for (attr
= 0; attr
< VERT_ATTRIB_MAX
; attr
++) {
91 if ((prog
->info
.inputs_read
& BITFIELD64_BIT(attr
)) != 0) {
92 input_to_index
[attr
] = num_inputs
;
94 if ((prog
->info
.vs
.double_inputs_read
& BITFIELD64_BIT(attr
)) != 0) {
95 /* add placeholder for second part of a double attribute */
99 input_to_index
[attr
] = ~0;
103 /* bit of a hack, mirroring st_translate_vertex_program */
104 input_to_index
[VERT_ATTRIB_EDGEFLAG
] = num_inputs
;
107 nir_foreach_variable_safe(var
, &nir
->inputs
) {
108 attr
= var
->data
.location
;
109 assert(attr
< ARRAY_SIZE(input_to_index
));
111 if (input_to_index
[attr
] != ~0u) {
112 var
->data
.driver_location
= input_to_index
[attr
];
115 /* Move unused input variables to the globals list (with no
116 * initialization), to avoid confusing drivers looking through the
117 * inputs array and expecting to find inputs with a driver_location
120 exec_node_remove(&var
->node
);
121 var
->data
.mode
= nir_var_global
;
122 exec_list_push_tail(&nir
->globals
, &var
->node
);
128 st_nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
129 gl_shader_stage stage
)
131 unsigned location
= 0;
132 unsigned assigned_locations
[VARYING_SLOT_TESS_MAX
];
133 uint64_t processed_locs
[2] = {0};
135 const int base
= stage
== MESA_SHADER_FRAGMENT
?
136 (int) FRAG_RESULT_DATA0
: (int) VARYING_SLOT_VAR0
;
138 int UNUSED last_loc
= 0;
139 nir_foreach_variable(var
, var_list
) {
141 const struct glsl_type
*type
= var
->type
;
142 if (nir_is_per_vertex_io(var
, stage
)) {
143 assert(glsl_type_is_array(type
));
144 type
= glsl_get_array_element(type
);
147 unsigned var_size
= type_size(type
);
149 /* Builtins don't allow component packing so we only need to worry about
150 * user defined varyings sharing the same location.
152 bool processed
= false;
153 if (var
->data
.location
>= base
) {
154 unsigned glsl_location
= var
->data
.location
- base
;
156 for (unsigned i
= 0; i
< var_size
; i
++) {
157 if (processed_locs
[var
->data
.index
] &
158 ((uint64_t)1 << (glsl_location
+ i
)))
161 processed_locs
[var
->data
.index
] |=
162 ((uint64_t)1 << (glsl_location
+ i
));
166 /* Because component packing allows varyings to share the same location
167 * we may have already have processed this location.
170 unsigned driver_location
= assigned_locations
[var
->data
.location
];
171 var
->data
.driver_location
= driver_location
;
172 *size
+= type_size(type
);
174 /* An array may be packed such that is crosses multiple other arrays
175 * or variables, we need to make sure we have allocated the elements
176 * consecutively if the previously proccessed var was shorter than
177 * the current array we are processing.
179 * NOTE: The code below assumes the var list is ordered in ascending
182 assert(last_loc
<= var
->data
.location
);
183 last_loc
= var
->data
.location
;
184 unsigned last_slot_location
= driver_location
+ var_size
;
185 if (last_slot_location
> location
) {
186 unsigned num_unallocated_slots
= last_slot_location
- location
;
187 unsigned first_unallocated_slot
= var_size
- num_unallocated_slots
;
188 for (unsigned i
= first_unallocated_slot
; i
< num_unallocated_slots
; i
++) {
189 assigned_locations
[var
->data
.location
+ i
] = location
;
196 for (unsigned i
= 0; i
< var_size
; i
++) {
197 assigned_locations
[var
->data
.location
+ i
] = location
+ i
;
200 var
->data
.driver_location
= location
;
201 location
+= var_size
;
208 st_nir_lookup_parameter_index(const struct gl_program_parameter_list
*params
,
211 int loc
= _mesa_lookup_parameter_index(params
, name
);
213 /* is there a better way to do this? If we have something like:
221 * Then what we get in prog->Parameters looks like:
223 * 0: Name=color.f, Type=6, DataType=1406, Size=1
224 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
226 * So the name doesn't match up and _mesa_lookup_parameter_index()
227 * fails. In this case just find the first matching "color.*"..
229 * Note for arrays you could end up w/ color[n].f, for example.
231 * glsl_to_tgsi works slightly differently in this regard. It is
232 * emitting something more low level, so it just translates the
233 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
234 * it just calculates the additional offset of struct field members
235 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
236 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
237 * needs to work backwards to get base var loc from the param-list
238 * which already has them separated out.
241 int namelen
= strlen(name
);
242 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
243 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
244 if ((strncmp(p
->Name
, name
, namelen
) == 0) &&
245 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
256 st_nir_assign_uniform_locations(struct gl_context
*ctx
,
257 struct gl_program
*prog
,
258 struct gl_shader_program
*shader_program
,
259 struct exec_list
*uniform_list
, unsigned *size
)
265 nir_foreach_variable(uniform
, uniform_list
) {
269 * UBO's have their own address spaces, so don't count them towards the
270 * number of global uniforms
272 if ((uniform
->data
.mode
== nir_var_uniform
|| uniform
->data
.mode
== nir_var_shader_storage
) &&
273 uniform
->interface_type
!= NULL
)
276 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
277 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
278 if (type
->is_sampler()) {
280 shaderidx
+= type_size(uniform
->type
);
283 imageidx
+= type_size(uniform
->type
);
285 } else if (strncmp(uniform
->name
, "gl_", 3) == 0) {
286 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
287 /* This state reference has already been setup by ir_to_mesa, but we'll
288 * get the same index back here.
292 if (glsl_type_is_struct(type
)) {
295 comps
= glsl_get_vector_elements(type
);
298 if (ctx
->Const
.PackedDriverUniformStorage
) {
299 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
300 stateTokens
, comps
, false);
301 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
303 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
306 loc
= st_nir_lookup_parameter_index(prog
->Parameters
, uniform
->name
);
308 if (ctx
->Const
.PackedDriverUniformStorage
) {
309 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
313 uniform
->data
.driver_location
= loc
;
315 max
= MAX2(max
, loc
+ type_size(uniform
->type
));
321 st_nir_opts(nir_shader
*nir
, bool scalar
)
327 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
330 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
);
331 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
334 NIR_PASS_V(nir
, nir_lower_alu
);
335 NIR_PASS_V(nir
, nir_lower_pack
);
336 NIR_PASS(progress
, nir
, nir_copy_prop
);
337 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
338 NIR_PASS(progress
, nir
, nir_opt_dce
);
339 if (nir_opt_trivial_continues(nir
)) {
341 NIR_PASS(progress
, nir
, nir_copy_prop
);
342 NIR_PASS(progress
, nir
, nir_opt_dce
);
344 NIR_PASS(progress
, nir
, nir_opt_if
);
345 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
346 NIR_PASS(progress
, nir
, nir_opt_cse
);
347 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8);
349 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
350 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
352 NIR_PASS(progress
, nir
, nir_opt_undef
);
353 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
354 if (nir
->options
->max_unroll_iterations
) {
355 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
360 /* First third of converting glsl_to_nir.. this leaves things in a pre-
361 * nir_lower_io state, so that shader variants can more easily insert/
362 * replace variables, etc.
365 st_glsl_to_nir(struct st_context
*st
, struct gl_program
*prog
,
366 struct gl_shader_program
*shader_program
,
367 gl_shader_stage stage
)
369 const nir_shader_compiler_options
*options
=
370 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
371 enum pipe_shader_type type
= pipe_shader_type_from_mesa(stage
);
372 struct pipe_screen
*screen
= st
->pipe
->screen
;
373 bool is_scalar
= screen
->get_shader_param(screen
, type
, PIPE_SHADER_CAP_SCALAR_ISA
);
379 nir_shader
*nir
= glsl_to_nir(shader_program
, stage
, options
);
381 /* Set the next shader stage hint for VS and TES. */
382 if (!nir
->info
.separate_shader
&&
383 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
384 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
386 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
387 unsigned stages_mask
=
388 ~prev_stages
& shader_program
->data
->linked_stages
;
390 nir
->info
.next_stage
= stages_mask
?
391 (gl_shader_stage
) ffs(stages_mask
) : MESA_SHADER_FRAGMENT
;
393 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
396 nir_variable_mode mask
=
397 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
398 nir_remove_dead_variables(nir
, mask
);
400 if (options
->lower_all_io_to_temps
||
401 nir
->info
.stage
== MESA_SHADER_VERTEX
||
402 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
403 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
404 nir_shader_get_entrypoint(nir
),
406 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
407 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
408 nir_shader_get_entrypoint(nir
),
412 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
413 NIR_PASS_V(nir
, nir_split_var_copies
);
414 NIR_PASS_V(nir
, nir_lower_var_copies
);
416 st_nir_opts(nir
, is_scalar
);
421 /* Second third of converting glsl_to_nir. This creates uniforms, gathers
422 * info on varyings, etc after NIR link time opts have been applied.
425 st_glsl_to_nir_post_opts(struct st_context
*st
, struct gl_program
*prog
,
426 struct gl_shader_program
*shader_program
)
428 nir_shader
*nir
= prog
->nir
;
430 /* Make a pass over the IR to add state references for any built-in
431 * uniforms that are used. This has to be done now (during linking).
432 * Code generation doesn't happen until the first time this shader is
433 * used for rendering. Waiting until then to generate the parameters is
434 * too late. At that point, the values for the built-in uniforms won't
435 * get sent to the shader.
437 nir_foreach_variable(var
, &nir
->uniforms
) {
438 if (strncmp(var
->name
, "gl_", 3) == 0) {
439 const nir_state_slot
*const slots
= var
->state_slots
;
440 assert(var
->state_slots
!= NULL
);
442 const struct glsl_type
*type
= glsl_without_array(var
->type
);
443 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
445 if (glsl_type_is_struct(type
)) {
446 /* Builtin struct require specical handling for now we just
447 * make all members vec4. See st_nir_lower_builtin.
451 comps
= glsl_get_vector_elements(type
);
454 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
455 _mesa_add_sized_state_reference(prog
->Parameters
,
459 _mesa_add_state_reference(prog
->Parameters
,
466 /* Avoid reallocation of the program parameter list, because the uniform
467 * storage is only associated with the original parameter list.
468 * This should be enough for Bitmap and DrawPixels constants.
470 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
472 /* This has to be done last. Any operation the can cause
473 * prog->ParameterValues to get reallocated (e.g., anything that adds a
474 * program constant) has to happen before creating this linkage.
476 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
, true);
478 st_set_prog_affected_state_flags(prog
);
480 NIR_PASS_V(nir
, st_nir_lower_builtin
);
481 NIR_PASS_V(nir
, gl_nir_lower_atomics
, shader_program
, true);
483 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
485 _mesa_log("NIR IR for linked %s program %d:\n",
486 _mesa_shader_stage_to_string(prog
->info
.stage
),
487 shader_program
->Name
);
488 nir_print_shader(nir
, _mesa_get_log_file());
493 /* TODO any better helper somewhere to sort a list? */
496 insert_sorted(struct exec_list
*var_list
, nir_variable
*new_var
)
498 nir_foreach_variable(var
, var_list
) {
499 if (var
->data
.location
> new_var
->data
.location
) {
500 exec_node_insert_node_before(&var
->node
, &new_var
->node
);
504 exec_list_push_tail(var_list
, &new_var
->node
);
508 sort_varyings(struct exec_list
*var_list
)
510 struct exec_list new_list
;
511 exec_list_make_empty(&new_list
);
512 nir_foreach_variable_safe(var
, var_list
) {
513 exec_node_remove(&var
->node
);
514 insert_sorted(&new_list
, var
);
516 exec_list_move_nodes_to(&new_list
, var_list
);
520 set_st_program(struct gl_program
*prog
,
521 struct gl_shader_program
*shader_program
,
524 struct st_vertex_program
*stvp
;
525 struct st_common_program
*stp
;
526 struct st_fragment_program
*stfp
;
527 struct st_compute_program
*stcp
;
529 switch (prog
->info
.stage
) {
530 case MESA_SHADER_VERTEX
:
531 stvp
= (struct st_vertex_program
*)prog
;
532 stvp
->shader_program
= shader_program
;
533 stvp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
534 stvp
->tgsi
.ir
.nir
= nir
;
536 case MESA_SHADER_GEOMETRY
:
537 case MESA_SHADER_TESS_CTRL
:
538 case MESA_SHADER_TESS_EVAL
:
539 stp
= (struct st_common_program
*)prog
;
540 stp
->shader_program
= shader_program
;
541 stp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
542 stp
->tgsi
.ir
.nir
= nir
;
544 case MESA_SHADER_FRAGMENT
:
545 stfp
= (struct st_fragment_program
*)prog
;
546 stfp
->shader_program
= shader_program
;
547 stfp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
548 stfp
->tgsi
.ir
.nir
= nir
;
550 case MESA_SHADER_COMPUTE
:
551 stcp
= (struct st_compute_program
*)prog
;
552 stcp
->shader_program
= shader_program
;
553 stcp
->tgsi
.ir_type
= PIPE_SHADER_IR_NIR
;
554 stcp
->tgsi
.prog
= nir
;
557 unreachable("unknown shader stage");
562 st_nir_get_mesa_program(struct gl_context
*ctx
,
563 struct gl_shader_program
*shader_program
,
564 struct gl_linked_shader
*shader
)
566 struct st_context
*st
= st_context(ctx
);
567 struct pipe_screen
*pscreen
= ctx
->st
->pipe
->screen
;
568 struct gl_program
*prog
;
570 validate_ir_tree(shader
->ir
);
572 prog
= shader
->Program
;
574 prog
->Parameters
= _mesa_new_parameter_list();
576 _mesa_copy_linked_program_data(shader_program
, shader
);
577 _mesa_generate_parameters_list_for_uniforms(ctx
, shader_program
, shader
,
580 /* Remove reads from output registers. */
581 if (!pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
))
582 lower_output_reads(shader
->Stage
, shader
->ir
);
584 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
586 _mesa_log("GLSL IR for linked %s program %d:\n",
587 _mesa_shader_stage_to_string(shader
->Stage
),
588 shader_program
->Name
);
589 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
593 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
594 _mesa_update_shader_textures_used(shader_program
, prog
);
596 nir_shader
*nir
= st_glsl_to_nir(st
, prog
, shader_program
, shader
->Stage
);
598 set_st_program(prog
, shader_program
, nir
);
603 st_nir_link_shaders(nir_shader
**producer
, nir_shader
**consumer
, bool scalar
)
605 nir_lower_io_arrays_to_elements(*producer
, *consumer
);
607 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
608 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
610 if (nir_remove_unused_varyings(*producer
, *consumer
)) {
611 NIR_PASS_V(*producer
, nir_lower_global_vars_to_local
);
612 NIR_PASS_V(*consumer
, nir_lower_global_vars_to_local
);
614 /* The backend might not be able to handle indirects on
615 * temporaries so we need to lower indirects on any of the
616 * varyings we have demoted here.
618 * TODO: radeonsi shouldn't need to do this, however LLVM isn't
619 * currently smart enough to handle indirects without causing excess
620 * spilling causing the gpu to hang.
622 * See the following thread for more details of the problem:
623 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
625 nir_variable_mode indirect_mask
= nir_var_local
;
627 NIR_PASS_V(*producer
, nir_lower_indirect_derefs
, indirect_mask
);
628 NIR_PASS_V(*consumer
, nir_lower_indirect_derefs
, indirect_mask
);
630 st_nir_opts(*producer
, scalar
);
631 st_nir_opts(*consumer
, scalar
);
638 st_link_nir(struct gl_context
*ctx
,
639 struct gl_shader_program
*shader_program
)
641 struct st_context
*st
= st_context(ctx
);
642 struct pipe_screen
*screen
= st
->pipe
->screen
;
643 bool is_scalar
[MESA_SHADER_STAGES
];
645 /* Determine scalar property of each shader stage */
646 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
647 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
648 enum pipe_shader_type type
;
653 type
= pipe_shader_type_from_mesa(shader
->Stage
);
654 is_scalar
[i
] = screen
->get_shader_param(screen
, type
, PIPE_SHADER_CAP_SCALAR_ISA
);
657 /* Determine first and last stage. */
658 unsigned first
= MESA_SHADER_STAGES
;
660 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
661 if (!shader_program
->_LinkedShaders
[i
])
663 if (first
== MESA_SHADER_STAGES
)
668 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
669 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
673 st_nir_get_mesa_program(ctx
, shader_program
, shader
);
675 nir_variable_mode mask
= (nir_variable_mode
) 0;
677 mask
= (nir_variable_mode
)(mask
| nir_var_shader_in
);
680 mask
= (nir_variable_mode
)(mask
| nir_var_shader_out
);
682 nir_shader
*nir
= shader
->Program
->nir
;
683 NIR_PASS_V(nir
, nir_lower_io_to_scalar_early
, mask
);
684 st_nir_opts(nir
, is_scalar
[i
]);
687 /* Linking the stages in the opposite order (from fragment to vertex)
688 * ensures that inter-shader outputs written to in an earlier stage
689 * are eliminated if they are (transitively) not used in a later
693 for (int i
= next
- 1; i
>= 0; i
--) {
694 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
698 st_nir_link_shaders(&shader
->Program
->nir
,
699 &shader_program
->_LinkedShaders
[next
]->Program
->nir
,
705 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
706 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
710 nir_shader
*nir
= shader
->Program
->nir
;
712 /* fragment shaders may need : */
713 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
714 static const gl_state_index16 wposTransformState
[STATE_LENGTH
] = {
715 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
717 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
718 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
720 memcpy(wpos_options
.state_tokens
, wposTransformState
,
721 sizeof(wpos_options
.state_tokens
));
722 wpos_options
.fs_coord_origin_upper_left
=
723 pscreen
->get_param(pscreen
,
724 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
725 wpos_options
.fs_coord_origin_lower_left
=
726 pscreen
->get_param(pscreen
,
727 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
728 wpos_options
.fs_coord_pixel_center_integer
=
729 pscreen
->get_param(pscreen
,
730 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
731 wpos_options
.fs_coord_pixel_center_half_integer
=
732 pscreen
->get_param(pscreen
,
733 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
735 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
736 nir_validate_shader(nir
);
737 _mesa_add_state_reference(shader
->Program
->Parameters
,
742 NIR_PASS_V(nir
, nir_lower_system_values
);
744 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
745 shader
->Program
->info
= nir
->info
;
748 struct gl_program
*prev_shader
=
749 shader_program
->_LinkedShaders
[prev
]->Program
;
751 /* We can't use nir_compact_varyings with transform feedback, since
752 * the pipe_stream_output->output_register field is based on the
753 * pre-compacted driver_locations.
755 if (!prev_shader
->sh
.LinkedTransformFeedback
)
756 nir_compact_varyings(shader_program
->_LinkedShaders
[prev
]->Program
->nir
,
757 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
762 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
763 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
767 st_glsl_to_nir_post_opts(st
, shader
->Program
, shader_program
);
769 assert(shader
->Program
);
770 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
771 _mesa_shader_stage_to_program(i
),
773 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
777 nir_sweep(shader
->Program
->nir
);
783 /* Last third of preparing nir from glsl, which happens after shader
787 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
788 struct gl_shader_program
*shader_program
, nir_shader
*nir
)
790 struct pipe_screen
*screen
= st
->pipe
->screen
;
791 const nir_shader_compiler_options
*options
=
792 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
794 NIR_PASS_V(nir
, nir_split_var_copies
);
795 NIR_PASS_V(nir
, nir_lower_var_copies
);
796 if (options
->lower_all_io_to_temps
||
797 nir
->info
.stage
== MESA_SHADER_VERTEX
||
798 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
799 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, false);
800 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
801 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, true);
804 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
805 /* Needs special handling so drvloc matches the vbo state: */
806 st_nir_assign_vs_in_locations(prog
, nir
);
807 /* Re-lower global vars, to deal with any dead VS inputs. */
808 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
810 sort_varyings(&nir
->outputs
);
811 st_nir_assign_var_locations(&nir
->outputs
,
814 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
815 } else if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
816 nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
817 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
818 sort_varyings(&nir
->inputs
);
819 st_nir_assign_var_locations(&nir
->inputs
,
822 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
824 sort_varyings(&nir
->outputs
);
825 st_nir_assign_var_locations(&nir
->outputs
,
828 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
829 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
830 sort_varyings(&nir
->inputs
);
831 st_nir_assign_var_locations(&nir
->inputs
,
834 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
835 st_nir_assign_var_locations(&nir
->outputs
,
838 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
841 unreachable("invalid shader type for tgsi bypass\n");
844 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
,
845 st
->ctx
->Const
.Program
[nir
->info
.stage
].MaxAtomicBuffers
);
847 st_nir_assign_uniform_locations(st
->ctx
, prog
, shader_program
,
848 &nir
->uniforms
, &nir
->num_uniforms
);
850 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
851 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_type_dword_size
,
852 (nir_lower_io_options
)0);
853 NIR_PASS_V(nir
, st_nir_lower_uniforms_to_ubo
);
856 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
857 NIR_PASS_V(nir
, gl_nir_lower_samplers_as_deref
, shader_program
);
859 NIR_PASS_V(nir
, gl_nir_lower_samplers
, shader_program
);