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
)
86 nir
->num_inputs
= util_bitcount64(nir
->info
.inputs_read
);
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
= nir
->num_inputs
++;
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
));
99 /* Move unused input variables to the globals list (with no
100 * initialization), to avoid confusing drivers looking through the
101 * inputs array and expecting to find inputs with a driver_location
104 exec_node_remove(&var
->node
);
105 var
->data
.mode
= nir_var_shader_temp
;
106 exec_list_push_tail(&nir
->globals
, &var
->node
);
112 st_nir_lookup_parameter_index(const struct gl_program_parameter_list
*params
,
115 int loc
= _mesa_lookup_parameter_index(params
, name
);
117 /* is there a better way to do this? If we have something like:
125 * Then what we get in prog->Parameters looks like:
127 * 0: Name=color.f, Type=6, DataType=1406, Size=1
128 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
130 * So the name doesn't match up and _mesa_lookup_parameter_index()
131 * fails. In this case just find the first matching "color.*"..
133 * Note for arrays you could end up w/ color[n].f, for example.
135 * glsl_to_tgsi works slightly differently in this regard. It is
136 * emitting something more low level, so it just translates the
137 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
138 * it just calculates the additional offset of struct field members
139 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
140 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
141 * needs to work backwards to get base var loc from the param-list
142 * which already has them separated out.
145 int namelen
= strlen(name
);
146 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
147 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
148 if ((strncmp(p
->Name
, name
, namelen
) == 0) &&
149 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
160 st_nir_assign_uniform_locations(struct gl_context
*ctx
,
161 struct gl_program
*prog
,
162 struct exec_list
*uniform_list
)
167 nir_foreach_variable(uniform
, uniform_list
) {
171 * UBO's have their own address spaces, so don't count them towards the
172 * number of global uniforms
174 if (uniform
->data
.mode
== nir_var_mem_ubo
|| uniform
->data
.mode
== nir_var_mem_ssbo
)
177 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
178 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
179 if (type
->is_sampler()) {
181 shaderidx
+= type_size(uniform
->type
);
184 imageidx
+= type_size(uniform
->type
);
186 } else if (strncmp(uniform
->name
, "gl_", 3) == 0) {
187 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
188 /* This state reference has already been setup by ir_to_mesa, but we'll
189 * get the same index back here.
193 if (glsl_type_is_struct_or_ifc(type
)) {
196 comps
= glsl_get_vector_elements(type
);
199 if (ctx
->Const
.PackedDriverUniformStorage
) {
200 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
201 stateTokens
, comps
, false);
202 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
204 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
207 loc
= st_nir_lookup_parameter_index(prog
->Parameters
, uniform
->name
);
209 /* We need to check that loc is not -1 here before accessing the
210 * array. It can be negative for example when we have a struct that
211 * only contains opaque types.
213 if (loc
>= 0 && ctx
->Const
.PackedDriverUniformStorage
) {
214 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
218 uniform
->data
.driver_location
= loc
;
223 st_nir_opts(nir_shader
*nir
, bool scalar
)
226 unsigned lower_flrp
=
227 (nir
->options
->lower_flrp16
? 16 : 0) |
228 (nir
->options
->lower_flrp32
? 32 : 0) |
229 (nir
->options
->lower_flrp64
? 64 : 0);
234 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
236 /* Linking deals with unused inputs/outputs, but here we can remove
237 * things local to the shader in the hopes that we can cleanup other
238 * things. This pass will also remove variables with only stores, so we
239 * might be able to make progress after it.
241 NIR_PASS(progress
, nir
, nir_remove_dead_variables
,
242 (nir_variable_mode
)(nir_var_function_temp
|
243 nir_var_shader_temp
|
244 nir_var_mem_shared
));
246 NIR_PASS(progress
, nir
, nir_opt_copy_prop_vars
);
247 NIR_PASS(progress
, nir
, nir_opt_dead_write_vars
);
250 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
251 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
254 NIR_PASS_V(nir
, nir_lower_alu
);
255 NIR_PASS_V(nir
, nir_lower_pack
);
256 NIR_PASS(progress
, nir
, nir_copy_prop
);
257 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
258 NIR_PASS(progress
, nir
, nir_opt_dce
);
259 if (nir_opt_trivial_continues(nir
)) {
261 NIR_PASS(progress
, nir
, nir_copy_prop
);
262 NIR_PASS(progress
, nir
, nir_opt_dce
);
264 NIR_PASS(progress
, nir
, nir_opt_if
, false);
265 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
266 NIR_PASS(progress
, nir
, nir_opt_cse
);
267 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
269 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
270 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
272 if (lower_flrp
!= 0) {
273 bool lower_flrp_progress
= false;
275 NIR_PASS(lower_flrp_progress
, nir
, nir_lower_flrp
,
277 false /* always_precise */,
278 nir
->options
->lower_ffma
);
279 if (lower_flrp_progress
) {
280 NIR_PASS(progress
, nir
,
281 nir_opt_constant_folding
);
285 /* Nothing should rematerialize any flrps, so we only need to do this
291 NIR_PASS(progress
, nir
, gl_nir_opt_access
);
293 NIR_PASS(progress
, nir
, nir_opt_undef
);
294 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
295 if (nir
->options
->max_unroll_iterations
) {
296 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
301 /* First third of converting glsl_to_nir.. this leaves things in a pre-
302 * nir_lower_io state, so that shader variants can more easily insert/
303 * replace variables, etc.
306 st_glsl_to_nir(struct st_context
*st
, struct gl_program
*prog
,
307 struct gl_shader_program
*shader_program
,
308 gl_shader_stage stage
)
310 const nir_shader_compiler_options
*options
=
311 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
312 enum pipe_shader_type type
= pipe_shader_type_from_mesa(stage
);
313 struct pipe_screen
*screen
= st
->pipe
->screen
;
314 bool is_scalar
= screen
->get_shader_param(screen
, type
, PIPE_SHADER_CAP_SCALAR_ISA
);
317 options
->lower_int64_options
|| options
->lower_doubles_options
;
322 nir_shader
*nir
= glsl_to_nir(st
->ctx
, shader_program
, stage
, options
);
324 /* Set the next shader stage hint for VS and TES. */
325 if (!nir
->info
.separate_shader
&&
326 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
327 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
329 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
330 unsigned stages_mask
=
331 ~prev_stages
& shader_program
->data
->linked_stages
;
333 nir
->info
.next_stage
= stages_mask
?
334 (gl_shader_stage
) u_bit_scan(&stages_mask
) : MESA_SHADER_FRAGMENT
;
336 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
339 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
340 if (!st
->ctx
->SoftFP64
&& nir
->info
.uses_64bit
&&
341 (options
->lower_doubles_options
& nir_lower_fp64_full_software
) != 0) {
342 st
->ctx
->SoftFP64
= glsl_float64_funcs_to_nir(st
->ctx
, options
);
345 nir_variable_mode mask
=
346 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
347 nir_remove_dead_variables(nir
, mask
);
349 if (options
->lower_all_io_to_temps
||
350 nir
->info
.stage
== MESA_SHADER_VERTEX
||
351 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
352 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
353 nir_shader_get_entrypoint(nir
),
355 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
356 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
357 nir_shader_get_entrypoint(nir
),
361 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
362 NIR_PASS_V(nir
, nir_split_var_copies
);
363 NIR_PASS_V(nir
, nir_lower_var_copies
);
366 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
369 /* before buffers and vars_to_ssa */
370 NIR_PASS_V(nir
, gl_nir_lower_bindless_images
);
371 st_nir_opts(nir
, is_scalar
);
373 NIR_PASS_V(nir
, gl_nir_lower_buffers
, shader_program
);
374 /* Do a round of constant folding to clean up address calculations */
375 NIR_PASS_V(nir
, nir_opt_constant_folding
);
378 bool lowered_64bit_ops
= false;
379 if (options
->lower_doubles_options
) {
380 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_doubles
,
381 st
->ctx
->SoftFP64
, options
->lower_doubles_options
);
383 if (options
->lower_int64_options
) {
384 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_int64
,
385 options
->lower_int64_options
);
388 if (lowered_64bit_ops
)
389 st_nir_opts(nir
, is_scalar
);
395 /* Second third of converting glsl_to_nir. This creates uniforms, gathers
396 * info on varyings, etc after NIR link time opts have been applied.
399 st_glsl_to_nir_post_opts(struct st_context
*st
, struct gl_program
*prog
,
400 struct gl_shader_program
*shader_program
)
402 nir_shader
*nir
= prog
->nir
;
404 /* Make a pass over the IR to add state references for any built-in
405 * uniforms that are used. This has to be done now (during linking).
406 * Code generation doesn't happen until the first time this shader is
407 * used for rendering. Waiting until then to generate the parameters is
408 * too late. At that point, the values for the built-in uniforms won't
409 * get sent to the shader.
411 nir_foreach_variable(var
, &nir
->uniforms
) {
412 if (strncmp(var
->name
, "gl_", 3) == 0) {
413 const nir_state_slot
*const slots
= var
->state_slots
;
414 assert(var
->state_slots
!= NULL
);
416 const struct glsl_type
*type
= glsl_without_array(var
->type
);
417 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
419 if (glsl_type_is_struct_or_ifc(type
)) {
420 /* Builtin struct require specical handling for now we just
421 * make all members vec4. See st_nir_lower_builtin.
425 comps
= glsl_get_vector_elements(type
);
428 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
429 _mesa_add_sized_state_reference(prog
->Parameters
,
433 _mesa_add_state_reference(prog
->Parameters
,
440 /* Avoid reallocation of the program parameter list, because the uniform
441 * storage is only associated with the original parameter list.
442 * This should be enough for Bitmap and DrawPixels constants.
444 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
446 /* This has to be done last. Any operation the can cause
447 * prog->ParameterValues to get reallocated (e.g., anything that adds a
448 * program constant) has to happen before creating this linkage.
450 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
);
452 st_set_prog_affected_state_flags(prog
);
454 NIR_PASS_V(nir
, st_nir_lower_builtin
);
455 NIR_PASS_V(nir
, gl_nir_lower_atomics
, shader_program
, true);
456 NIR_PASS_V(nir
, nir_opt_intrinsics
);
458 nir_variable_mode mask
= nir_var_function_temp
;
459 nir_remove_dead_variables(nir
, mask
);
461 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
463 _mesa_log("NIR IR for linked %s program %d:\n",
464 _mesa_shader_stage_to_string(prog
->info
.stage
),
465 shader_program
->Name
);
466 nir_print_shader(nir
, _mesa_get_log_file());
472 set_st_program(struct gl_program
*prog
,
473 struct gl_shader_program
*shader_program
,
476 struct st_vertex_program
*stvp
;
477 struct st_common_program
*stp
;
478 struct st_fragment_program
*stfp
;
479 struct st_compute_program
*stcp
;
481 switch (prog
->info
.stage
) {
482 case MESA_SHADER_VERTEX
:
483 stvp
= (struct st_vertex_program
*)prog
;
484 stvp
->shader_program
= shader_program
;
485 stvp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
486 stvp
->tgsi
.ir
.nir
= nir
;
488 case MESA_SHADER_GEOMETRY
:
489 case MESA_SHADER_TESS_CTRL
:
490 case MESA_SHADER_TESS_EVAL
:
491 stp
= (struct st_common_program
*)prog
;
492 stp
->shader_program
= shader_program
;
493 stp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
494 stp
->tgsi
.ir
.nir
= nir
;
496 case MESA_SHADER_FRAGMENT
:
497 stfp
= (struct st_fragment_program
*)prog
;
498 stfp
->shader_program
= shader_program
;
499 stfp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
500 stfp
->tgsi
.ir
.nir
= nir
;
502 case MESA_SHADER_COMPUTE
:
503 stcp
= (struct st_compute_program
*)prog
;
504 stcp
->shader_program
= shader_program
;
505 stcp
->tgsi
.ir_type
= PIPE_SHADER_IR_NIR
;
506 stcp
->tgsi
.prog
= nir
;
509 unreachable("unknown shader stage");
514 st_nir_get_mesa_program(struct gl_context
*ctx
,
515 struct gl_shader_program
*shader_program
,
516 struct gl_linked_shader
*shader
)
518 struct st_context
*st
= st_context(ctx
);
519 struct pipe_screen
*pscreen
= ctx
->st
->pipe
->screen
;
520 struct gl_program
*prog
;
522 validate_ir_tree(shader
->ir
);
524 prog
= shader
->Program
;
526 prog
->Parameters
= _mesa_new_parameter_list();
528 _mesa_copy_linked_program_data(shader_program
, shader
);
529 _mesa_generate_parameters_list_for_uniforms(ctx
, shader_program
, shader
,
532 /* Remove reads from output registers. */
533 if (!pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
))
534 lower_output_reads(shader
->Stage
, shader
->ir
);
536 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
538 _mesa_log("GLSL IR for linked %s program %d:\n",
539 _mesa_shader_stage_to_string(shader
->Stage
),
540 shader_program
->Name
);
541 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
545 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
546 _mesa_update_shader_textures_used(shader_program
, prog
);
548 nir_shader
*nir
= st_glsl_to_nir(st
, prog
, shader_program
, shader
->Stage
);
550 set_st_program(prog
, shader_program
, nir
);
555 st_nir_vectorize_io(nir_shader
*producer
, nir_shader
*consumer
)
557 NIR_PASS_V(producer
, nir_lower_io_to_vector
, nir_var_shader_out
);
558 NIR_PASS_V(producer
, nir_opt_combine_stores
, nir_var_shader_out
);
559 NIR_PASS_V(consumer
, nir_lower_io_to_vector
, nir_var_shader_in
);
561 if ((producer
)->info
.stage
!= MESA_SHADER_TESS_CTRL
) {
562 /* Calling lower_io_to_vector creates output variable writes with
563 * write-masks. We only support these for TCS outputs, so for other
564 * stages, we need to call nir_lower_io_to_temporaries to get rid of
565 * them. This, in turn, creates temporary variables and extra
566 * copy_deref intrinsics that we need to clean up.
568 NIR_PASS_V(producer
, nir_lower_io_to_temporaries
,
569 nir_shader_get_entrypoint(producer
), true, false);
570 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
571 NIR_PASS_V(producer
, nir_split_var_copies
);
572 NIR_PASS_V(producer
, nir_lower_var_copies
);
577 st_nir_link_shaders(nir_shader
**producer
, nir_shader
**consumer
, bool scalar
)
580 NIR_PASS_V(*producer
, nir_lower_io_to_scalar_early
, nir_var_shader_out
);
581 NIR_PASS_V(*consumer
, nir_lower_io_to_scalar_early
, nir_var_shader_in
);
584 nir_lower_io_arrays_to_elements(*producer
, *consumer
);
586 st_nir_opts(*producer
, scalar
);
587 st_nir_opts(*consumer
, scalar
);
589 if (nir_link_opt_varyings(*producer
, *consumer
))
590 st_nir_opts(*consumer
, scalar
);
592 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
593 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
595 if (nir_remove_unused_varyings(*producer
, *consumer
)) {
596 NIR_PASS_V(*producer
, nir_lower_global_vars_to_local
);
597 NIR_PASS_V(*consumer
, nir_lower_global_vars_to_local
);
599 st_nir_opts(*producer
, scalar
);
600 st_nir_opts(*consumer
, scalar
);
602 /* Optimizations can cause varyings to become unused.
603 * nir_compact_varyings() depends on all dead varyings being removed so
604 * we need to call nir_remove_dead_variables() again here.
606 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
607 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
612 st_lower_patch_vertices_in(struct gl_shader_program
*shader_prog
)
614 struct gl_linked_shader
*linked_tcs
=
615 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_CTRL
];
616 struct gl_linked_shader
*linked_tes
=
617 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_EVAL
];
619 /* If we have a TCS and TES linked together, lower TES patch vertices. */
620 if (linked_tcs
&& linked_tes
) {
621 nir_shader
*tcs_nir
= linked_tcs
->Program
->nir
;
622 nir_shader
*tes_nir
= linked_tes
->Program
->nir
;
624 /* The TES input vertex count is the TCS output vertex count,
625 * lower TES gl_PatchVerticesIn to a constant.
627 uint32_t tes_patch_verts
= tcs_nir
->info
.tess
.tcs_vertices_out
;
628 NIR_PASS_V(tes_nir
, nir_lower_patch_vertices
, tes_patch_verts
, NULL
);
635 st_nir_lower_wpos_ytransform(struct nir_shader
*nir
,
636 struct gl_program
*prog
,
637 struct pipe_screen
*pscreen
)
639 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
642 static const gl_state_index16 wposTransformState
[STATE_LENGTH
] = {
643 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
645 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
647 memcpy(wpos_options
.state_tokens
, wposTransformState
,
648 sizeof(wpos_options
.state_tokens
));
649 wpos_options
.fs_coord_origin_upper_left
=
650 pscreen
->get_param(pscreen
,
651 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
652 wpos_options
.fs_coord_origin_lower_left
=
653 pscreen
->get_param(pscreen
,
654 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
655 wpos_options
.fs_coord_pixel_center_integer
=
656 pscreen
->get_param(pscreen
,
657 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
658 wpos_options
.fs_coord_pixel_center_half_integer
=
659 pscreen
->get_param(pscreen
,
660 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
662 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
663 nir_validate_shader(nir
, "after nir_lower_wpos_ytransform");
664 _mesa_add_state_reference(prog
->Parameters
, wposTransformState
);
669 st_link_nir(struct gl_context
*ctx
,
670 struct gl_shader_program
*shader_program
)
672 struct st_context
*st
= st_context(ctx
);
673 struct pipe_screen
*screen
= st
->pipe
->screen
;
674 bool is_scalar
[MESA_SHADER_STAGES
];
676 unsigned last_stage
= 0;
677 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
678 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
682 /* Determine scalar property of each shader stage */
683 enum pipe_shader_type type
= pipe_shader_type_from_mesa(shader
->Stage
);
684 is_scalar
[i
] = screen
->get_shader_param(screen
, type
,
685 PIPE_SHADER_CAP_SCALAR_ISA
);
687 st_nir_get_mesa_program(ctx
, shader_program
, shader
);
691 NIR_PASS_V(shader
->Program
->nir
, nir_lower_load_const_to_scalar
);
695 /* Linking the stages in the opposite order (from fragment to vertex)
696 * ensures that inter-shader outputs written to in an earlier stage
697 * are eliminated if they are (transitively) not used in a later
700 int next
= last_stage
;
701 for (int i
= next
- 1; i
>= 0; i
--) {
702 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
706 st_nir_link_shaders(&shader
->Program
->nir
,
707 &shader_program
->_LinkedShaders
[next
]->Program
->nir
,
713 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
714 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
718 nir_shader
*nir
= shader
->Program
->nir
;
720 NIR_PASS_V(nir
, st_nir_lower_wpos_ytransform
, shader
->Program
,
723 NIR_PASS_V(nir
, nir_lower_system_values
);
724 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
726 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
727 shader
->Program
->info
= nir
->info
;
728 if (i
== MESA_SHADER_VERTEX
) {
729 /* NIR expands dual-slot inputs out to two locations. We need to
730 * compact things back down GL-style single-slot inputs to avoid
731 * confusing the state tracker.
733 shader
->Program
->info
.inputs_read
=
734 nir_get_single_slot_attribs_mask(nir
->info
.inputs_read
,
735 shader
->Program
->DualSlotInputs
);
739 struct gl_program
*prev_shader
=
740 shader_program
->_LinkedShaders
[prev
]->Program
;
742 /* We can't use nir_compact_varyings with transform feedback, since
743 * the pipe_stream_output->output_register field is based on the
744 * pre-compacted driver_locations.
746 if (!(prev_shader
->sh
.LinkedTransformFeedback
&&
747 prev_shader
->sh
.LinkedTransformFeedback
->NumVarying
> 0))
748 nir_compact_varyings(shader_program
->_LinkedShaders
[prev
]->Program
->nir
,
749 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
751 if (ctx
->Const
.ShaderCompilerOptions
[i
].NirOptions
->vectorize_io
)
752 st_nir_vectorize_io(prev_shader
->nir
, nir
);
757 st_lower_patch_vertices_in(shader_program
);
759 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
760 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
764 st_glsl_to_nir_post_opts(st
, shader
->Program
, shader_program
);
766 assert(shader
->Program
);
767 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
768 _mesa_shader_stage_to_program(i
),
770 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
774 nir_sweep(shader
->Program
->nir
);
776 /* The GLSL IR won't be needed anymore. */
777 ralloc_free(shader
->ir
);
785 st_nir_assign_varying_locations(struct st_context
*st
, nir_shader
*nir
)
787 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
788 /* Needs special handling so drvloc matches the vbo state: */
789 st_nir_assign_vs_in_locations(nir
);
790 /* Re-lower global vars, to deal with any dead VS inputs. */
791 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
793 nir_assign_io_var_locations(&nir
->outputs
,
796 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
797 } else if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
798 nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
799 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
800 nir_assign_io_var_locations(&nir
->inputs
,
803 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
805 nir_assign_io_var_locations(&nir
->outputs
,
808 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
809 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
810 nir_assign_io_var_locations(&nir
->inputs
,
813 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
814 nir_assign_io_var_locations(&nir
->outputs
,
817 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
820 unreachable("invalid shader type");
825 st_nir_lower_samplers(struct pipe_screen
*screen
, nir_shader
*nir
,
826 struct gl_shader_program
*shader_program
,
827 struct gl_program
*prog
)
829 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
830 NIR_PASS_V(nir
, gl_nir_lower_samplers_as_deref
, shader_program
);
832 NIR_PASS_V(nir
, gl_nir_lower_samplers
, shader_program
);
835 prog
->info
.textures_used
= nir
->info
.textures_used
;
836 prog
->info
.textures_used_by_txf
= nir
->info
.textures_used_by_txf
;
840 /* Last third of preparing nir from glsl, which happens after shader
844 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
845 struct gl_shader_program
*shader_program
, nir_shader
*nir
)
847 struct pipe_screen
*screen
= st
->pipe
->screen
;
848 const nir_shader_compiler_options
*options
=
849 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
851 NIR_PASS_V(nir
, nir_split_var_copies
);
852 NIR_PASS_V(nir
, nir_lower_var_copies
);
853 if (options
->lower_all_io_to_temps
||
854 options
->lower_all_io_to_elements
||
855 nir
->info
.stage
== MESA_SHADER_VERTEX
||
856 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
857 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, false);
858 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
859 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, true);
862 st_nir_assign_varying_locations(st
, nir
);
864 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
,
865 st
->ctx
->Const
.Program
[nir
->info
.stage
].MaxAtomicBuffers
);
867 st_nir_assign_uniform_locations(st
->ctx
, prog
,
870 /* Set num_uniforms in number of attribute slots (vec4s) */
871 nir
->num_uniforms
= DIV_ROUND_UP(prog
->Parameters
->NumParameterValues
, 4);
873 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
874 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_type_dword_size
,
875 (nir_lower_io_options
)0);
876 NIR_PASS_V(nir
, nir_lower_uniforms_to_ubo
, 4);
878 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_uniforms_type_size
,
879 (nir_lower_io_options
)0);
882 st_nir_lower_samplers(screen
, nir
, shader_program
, prog
);