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/glspirv.h"
37 #include "main/shaderapi.h"
38 #include "main/uniforms.h"
40 #include "main/shaderobj.h"
41 #include "st_context.h"
42 #include "st_glsl_types.h"
43 #include "st_program.h"
45 #include "compiler/nir/nir.h"
46 #include "compiler/glsl_types.h"
47 #include "compiler/glsl/glsl_to_nir.h"
48 #include "compiler/glsl/gl_nir.h"
49 #include "compiler/glsl/gl_nir_linker.h"
50 #include "compiler/glsl/ir.h"
51 #include "compiler/glsl/ir_optimization.h"
52 #include "compiler/glsl/string_to_uint_map.h"
55 type_size(const struct glsl_type
*type
)
57 return type
->count_attribute_slots(false);
60 /* Depending on PIPE_CAP_TGSI_TEXCOORD (st->needs_texcoord_semantic) we
61 * may need to fix up varying slots so the glsl->nir path is aligned
62 * with the anything->tgsi->nir path.
65 st_nir_fixup_varying_slots(struct st_context
*st
, struct exec_list
*var_list
)
67 if (st
->needs_texcoord_semantic
)
70 nir_foreach_variable(var
, var_list
) {
71 if (var
->data
.location
>= VARYING_SLOT_VAR0
) {
72 var
->data
.location
+= 9;
73 } else if ((var
->data
.location
>= VARYING_SLOT_TEX0
) &&
74 (var
->data
.location
<= VARYING_SLOT_TEX7
)) {
75 var
->data
.location
+= VARYING_SLOT_VAR0
- VARYING_SLOT_TEX0
;
80 /* input location assignment for VS inputs must be handled specially, so
81 * that it is aligned w/ st's vbo state.
82 * (This isn't the case with, for ex, FS inputs, which only need to agree
83 * on varying-slot w/ the VS outputs)
86 st_nir_assign_vs_in_locations(nir_shader
*nir
)
88 nir
->num_inputs
= util_bitcount64(nir
->info
.inputs_read
);
89 nir_foreach_variable_safe(var
, &nir
->inputs
) {
90 /* NIR already assigns dual-slot inputs to two locations so all we have
91 * to do is compact everything down.
93 if (var
->data
.location
== VERT_ATTRIB_EDGEFLAG
) {
94 /* bit of a hack, mirroring st_translate_vertex_program */
95 var
->data
.driver_location
= nir
->num_inputs
++;
96 } else if (nir
->info
.inputs_read
& BITFIELD64_BIT(var
->data
.location
)) {
97 var
->data
.driver_location
=
98 util_bitcount64(nir
->info
.inputs_read
&
99 BITFIELD64_MASK(var
->data
.location
));
101 /* Move unused input variables to the globals list (with no
102 * initialization), to avoid confusing drivers looking through the
103 * inputs array and expecting to find inputs with a driver_location
106 exec_node_remove(&var
->node
);
107 var
->data
.mode
= nir_var_shader_temp
;
108 exec_list_push_tail(&nir
->globals
, &var
->node
);
114 st_nir_lookup_parameter_index(struct gl_program
*prog
, nir_variable
*var
)
116 struct gl_program_parameter_list
*params
= prog
->Parameters
;
118 /* Lookup the first parameter that the uniform storage that match the
121 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
122 int index
= params
->Parameters
[i
].MainUniformStorageIndex
;
123 if (index
== var
->data
.location
)
127 /* TODO: Handle this fallback for SPIR-V. We need this for GLSL e.g. in
128 * dEQP-GLES2.functional.uniform_api.random.3
131 /* is there a better way to do this? If we have something like:
139 * Then what we get in prog->Parameters looks like:
141 * 0: Name=color.f, Type=6, DataType=1406, Size=1
142 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
144 * So the name doesn't match up and _mesa_lookup_parameter_index()
145 * fails. In this case just find the first matching "color.*"..
147 * Note for arrays you could end up w/ color[n].f, for example.
149 * glsl_to_tgsi works slightly differently in this regard. It is
150 * emitting something more low level, so it just translates the
151 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
152 * it just calculates the additional offset of struct field members
153 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
154 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
155 * needs to work backwards to get base var loc from the param-list
156 * which already has them separated out.
158 if (!prog
->sh
.data
->spirv
) {
159 int namelen
= strlen(var
->name
);
160 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
161 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
162 if ((strncmp(p
->Name
, var
->name
, namelen
) == 0) &&
163 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
173 st_nir_assign_uniform_locations(struct gl_context
*ctx
,
174 struct gl_program
*prog
,
175 struct exec_list
*uniform_list
)
180 nir_foreach_variable(uniform
, uniform_list
) {
184 * UBO's have their own address spaces, so don't count them towards the
185 * number of global uniforms
187 if (uniform
->data
.mode
== nir_var_mem_ubo
|| uniform
->data
.mode
== nir_var_mem_ssbo
)
190 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
191 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
192 if (type
->is_sampler()) {
194 shaderidx
+= type_size(uniform
->type
);
197 imageidx
+= type_size(uniform
->type
);
199 } else if (uniform
->state_slots
) {
200 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
201 /* This state reference has already been setup by ir_to_mesa, but we'll
202 * get the same index back here.
206 if (glsl_type_is_struct_or_ifc(type
)) {
209 comps
= glsl_get_vector_elements(type
);
212 if (ctx
->Const
.PackedDriverUniformStorage
) {
213 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
214 stateTokens
, comps
, false);
215 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
217 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
220 loc
= st_nir_lookup_parameter_index(prog
, uniform
);
222 /* We need to check that loc is not -1 here before accessing the
223 * array. It can be negative for example when we have a struct that
224 * only contains opaque types.
226 if (loc
>= 0 && ctx
->Const
.PackedDriverUniformStorage
) {
227 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
231 uniform
->data
.driver_location
= loc
;
236 st_nir_opts(nir_shader
*nir
)
239 unsigned lower_flrp
=
240 (nir
->options
->lower_flrp16
? 16 : 0) |
241 (nir
->options
->lower_flrp32
? 32 : 0) |
242 (nir
->options
->lower_flrp64
? 64 : 0);
247 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
249 /* Linking deals with unused inputs/outputs, but here we can remove
250 * things local to the shader in the hopes that we can cleanup other
251 * things. This pass will also remove variables with only stores, so we
252 * might be able to make progress after it.
254 NIR_PASS(progress
, nir
, nir_remove_dead_variables
,
255 (nir_variable_mode
)(nir_var_function_temp
|
256 nir_var_shader_temp
|
257 nir_var_mem_shared
));
259 NIR_PASS(progress
, nir
, nir_opt_copy_prop_vars
);
260 NIR_PASS(progress
, nir
, nir_opt_dead_write_vars
);
262 if (nir
->options
->lower_to_scalar
) {
263 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
264 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
267 NIR_PASS_V(nir
, nir_lower_alu
);
268 NIR_PASS_V(nir
, nir_lower_pack
);
269 NIR_PASS(progress
, nir
, nir_copy_prop
);
270 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
271 NIR_PASS(progress
, nir
, nir_opt_dce
);
272 if (nir_opt_trivial_continues(nir
)) {
274 NIR_PASS(progress
, nir
, nir_copy_prop
);
275 NIR_PASS(progress
, nir
, nir_opt_dce
);
277 NIR_PASS(progress
, nir
, nir_opt_if
, false);
278 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
279 NIR_PASS(progress
, nir
, nir_opt_cse
);
280 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
282 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
283 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
285 if (lower_flrp
!= 0) {
286 bool lower_flrp_progress
= false;
288 NIR_PASS(lower_flrp_progress
, nir
, nir_lower_flrp
,
290 false /* always_precise */,
291 nir
->options
->lower_ffma
);
292 if (lower_flrp_progress
) {
293 NIR_PASS(progress
, nir
,
294 nir_opt_constant_folding
);
298 /* Nothing should rematerialize any flrps, so we only need to do this
304 NIR_PASS(progress
, nir
, nir_opt_access
);
306 NIR_PASS(progress
, nir
, nir_opt_undef
);
307 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
308 if (nir
->options
->max_unroll_iterations
) {
309 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
315 shared_type_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
317 assert(glsl_type_is_vector_or_scalar(type
));
319 uint32_t comp_size
= glsl_type_is_boolean(type
)
320 ? 4 : glsl_get_bit_size(type
) / 8;
321 unsigned length
= glsl_get_vector_elements(type
);
322 *size
= comp_size
* length
,
323 *align
= comp_size
* (length
== 3 ? 4 : length
);
326 /* First third of converting glsl_to_nir.. this leaves things in a pre-
327 * nir_lower_io state, so that shader variants can more easily insert/
328 * replace variables, etc.
331 st_nir_preprocess(struct st_context
*st
, struct gl_program
*prog
,
332 struct gl_shader_program
*shader_program
,
333 gl_shader_stage stage
)
335 const nir_shader_compiler_options
*options
=
336 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
339 options
->lower_int64_options
|| options
->lower_doubles_options
;
340 nir_shader
*nir
= prog
->nir
;
342 /* Set the next shader stage hint for VS and TES. */
343 if (!nir
->info
.separate_shader
&&
344 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
345 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
347 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
348 unsigned stages_mask
=
349 ~prev_stages
& shader_program
->data
->linked_stages
;
351 nir
->info
.next_stage
= stages_mask
?
352 (gl_shader_stage
) u_bit_scan(&stages_mask
) : MESA_SHADER_FRAGMENT
;
354 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
357 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
358 if (!st
->ctx
->SoftFP64
&& nir
->info
.uses_64bit
&&
359 (options
->lower_doubles_options
& nir_lower_fp64_full_software
) != 0) {
360 st
->ctx
->SoftFP64
= glsl_float64_funcs_to_nir(st
->ctx
, options
);
363 nir_variable_mode mask
=
364 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
365 nir_remove_dead_variables(nir
, mask
);
367 if (options
->lower_all_io_to_temps
||
368 nir
->info
.stage
== MESA_SHADER_VERTEX
||
369 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
370 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
371 nir_shader_get_entrypoint(nir
),
373 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
374 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
375 nir_shader_get_entrypoint(nir
),
379 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
380 NIR_PASS_V(nir
, nir_split_var_copies
);
381 NIR_PASS_V(nir
, nir_lower_var_copies
);
383 if (options
->lower_to_scalar
) {
384 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
387 /* before buffers and vars_to_ssa */
388 NIR_PASS_V(nir
, gl_nir_lower_bindless_images
);
391 /* TODO: Change GLSL to not lower shared memory. */
392 if (prog
->nir
->info
.stage
== MESA_SHADER_COMPUTE
&&
393 shader_program
->data
->spirv
) {
394 NIR_PASS_V(prog
->nir
, nir_lower_vars_to_explicit_types
,
395 nir_var_mem_shared
, shared_type_info
);
396 NIR_PASS_V(prog
->nir
, nir_lower_explicit_io
,
397 nir_var_mem_shared
, nir_address_format_32bit_offset
);
400 NIR_PASS_V(nir
, gl_nir_lower_buffers
, shader_program
);
401 /* Do a round of constant folding to clean up address calculations */
402 NIR_PASS_V(nir
, nir_opt_constant_folding
);
405 bool lowered_64bit_ops
= false;
406 if (options
->lower_doubles_options
) {
407 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_doubles
,
408 st
->ctx
->SoftFP64
, options
->lower_doubles_options
);
410 if (options
->lower_int64_options
) {
411 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_int64
,
412 options
->lower_int64_options
);
415 if (lowered_64bit_ops
)
420 /* Second third of converting glsl_to_nir. This creates uniforms, gathers
421 * info on varyings, etc after NIR link time opts have been applied.
424 st_glsl_to_nir_post_opts(struct st_context
*st
, struct gl_program
*prog
,
425 struct gl_shader_program
*shader_program
)
427 nir_shader
*nir
= prog
->nir
;
429 /* Make a pass over the IR to add state references for any built-in
430 * uniforms that are used. This has to be done now (during linking).
431 * Code generation doesn't happen until the first time this shader is
432 * used for rendering. Waiting until then to generate the parameters is
433 * too late. At that point, the values for the built-in uniforms won't
434 * get sent to the shader.
436 nir_foreach_variable(var
, &nir
->uniforms
) {
437 const nir_state_slot
*const slots
= var
->state_slots
;
439 const struct glsl_type
*type
= glsl_without_array(var
->type
);
440 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
442 if (glsl_type_is_struct_or_ifc(type
)) {
443 /* Builtin struct require specical handling for now we just
444 * make all members vec4. See st_nir_lower_builtin.
448 comps
= glsl_get_vector_elements(type
);
451 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
452 _mesa_add_sized_state_reference(prog
->Parameters
,
456 _mesa_add_state_reference(prog
->Parameters
,
463 /* Avoid reallocation of the program parameter list, because the uniform
464 * storage is only associated with the original parameter list.
465 * This should be enough for Bitmap and DrawPixels constants.
467 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
469 /* This has to be done last. Any operation the can cause
470 * prog->ParameterValues to get reallocated (e.g., anything that adds a
471 * program constant) has to happen before creating this linkage.
473 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
);
475 st_set_prog_affected_state_flags(prog
);
477 /* None of the builtins being lowered here can be produced by SPIR-V. See
478 * _mesa_builtin_uniform_desc.
480 if (!shader_program
->data
->spirv
)
481 NIR_PASS_V(nir
, st_nir_lower_builtin
);
483 NIR_PASS_V(nir
, gl_nir_lower_atomics
, shader_program
, true);
484 NIR_PASS_V(nir
, nir_opt_intrinsics
);
486 nir_variable_mode mask
= nir_var_function_temp
;
487 nir_remove_dead_variables(nir
, mask
);
489 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
491 _mesa_log("NIR IR for linked %s program %d:\n",
492 _mesa_shader_stage_to_string(prog
->info
.stage
),
493 shader_program
->Name
);
494 nir_print_shader(nir
, _mesa_get_log_file());
500 set_st_program(struct gl_program
*prog
,
501 struct gl_shader_program
*shader_program
,
504 struct st_vertex_program
*stvp
;
505 struct st_common_program
*stp
;
506 struct st_fragment_program
*stfp
;
507 struct st_compute_program
*stcp
;
509 switch (prog
->info
.stage
) {
510 case MESA_SHADER_VERTEX
:
511 stvp
= (struct st_vertex_program
*)prog
;
512 stvp
->shader_program
= shader_program
;
513 stvp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
514 stvp
->tgsi
.ir
.nir
= nir
;
516 case MESA_SHADER_GEOMETRY
:
517 case MESA_SHADER_TESS_CTRL
:
518 case MESA_SHADER_TESS_EVAL
:
519 stp
= (struct st_common_program
*)prog
;
520 stp
->shader_program
= shader_program
;
521 stp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
522 stp
->tgsi
.ir
.nir
= nir
;
524 case MESA_SHADER_FRAGMENT
:
525 stfp
= (struct st_fragment_program
*)prog
;
526 stfp
->shader_program
= shader_program
;
527 stfp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
528 stfp
->tgsi
.ir
.nir
= nir
;
530 case MESA_SHADER_COMPUTE
:
531 stcp
= (struct st_compute_program
*)prog
;
532 stcp
->shader_program
= shader_program
;
533 stcp
->tgsi
.ir_type
= PIPE_SHADER_IR_NIR
;
534 stcp
->tgsi
.prog
= nir
;
537 unreachable("unknown shader stage");
542 st_nir_vectorize_io(nir_shader
*producer
, nir_shader
*consumer
)
544 NIR_PASS_V(producer
, nir_lower_io_to_vector
, nir_var_shader_out
);
545 NIR_PASS_V(producer
, nir_opt_combine_stores
, nir_var_shader_out
);
546 NIR_PASS_V(consumer
, nir_lower_io_to_vector
, nir_var_shader_in
);
548 if ((producer
)->info
.stage
!= MESA_SHADER_TESS_CTRL
) {
549 /* Calling lower_io_to_vector creates output variable writes with
550 * write-masks. We only support these for TCS outputs, so for other
551 * stages, we need to call nir_lower_io_to_temporaries to get rid of
552 * them. This, in turn, creates temporary variables and extra
553 * copy_deref intrinsics that we need to clean up.
555 NIR_PASS_V(producer
, nir_lower_io_to_temporaries
,
556 nir_shader_get_entrypoint(producer
), true, false);
557 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
558 NIR_PASS_V(producer
, nir_split_var_copies
);
559 NIR_PASS_V(producer
, nir_lower_var_copies
);
564 st_nir_link_shaders(nir_shader
**producer
, nir_shader
**consumer
)
566 if ((*producer
)->options
->lower_to_scalar
) {
567 NIR_PASS_V(*producer
, nir_lower_io_to_scalar_early
, nir_var_shader_out
);
568 NIR_PASS_V(*consumer
, nir_lower_io_to_scalar_early
, nir_var_shader_in
);
571 nir_lower_io_arrays_to_elements(*producer
, *consumer
);
573 st_nir_opts(*producer
);
574 st_nir_opts(*consumer
);
576 if (nir_link_opt_varyings(*producer
, *consumer
))
577 st_nir_opts(*consumer
);
579 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
580 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
582 if (nir_remove_unused_varyings(*producer
, *consumer
)) {
583 NIR_PASS_V(*producer
, nir_lower_global_vars_to_local
);
584 NIR_PASS_V(*consumer
, nir_lower_global_vars_to_local
);
586 st_nir_opts(*producer
);
587 st_nir_opts(*consumer
);
589 /* Optimizations can cause varyings to become unused.
590 * nir_compact_varyings() depends on all dead varyings being removed so
591 * we need to call nir_remove_dead_variables() again here.
593 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
594 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
599 st_lower_patch_vertices_in(struct gl_shader_program
*shader_prog
)
601 struct gl_linked_shader
*linked_tcs
=
602 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_CTRL
];
603 struct gl_linked_shader
*linked_tes
=
604 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_EVAL
];
606 /* If we have a TCS and TES linked together, lower TES patch vertices. */
607 if (linked_tcs
&& linked_tes
) {
608 nir_shader
*tcs_nir
= linked_tcs
->Program
->nir
;
609 nir_shader
*tes_nir
= linked_tes
->Program
->nir
;
611 /* The TES input vertex count is the TCS output vertex count,
612 * lower TES gl_PatchVerticesIn to a constant.
614 uint32_t tes_patch_verts
= tcs_nir
->info
.tess
.tcs_vertices_out
;
615 NIR_PASS_V(tes_nir
, nir_lower_patch_vertices
, tes_patch_verts
, NULL
);
622 st_nir_lower_wpos_ytransform(struct nir_shader
*nir
,
623 struct gl_program
*prog
,
624 struct pipe_screen
*pscreen
)
626 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
629 static const gl_state_index16 wposTransformState
[STATE_LENGTH
] = {
630 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
632 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
634 memcpy(wpos_options
.state_tokens
, wposTransformState
,
635 sizeof(wpos_options
.state_tokens
));
636 wpos_options
.fs_coord_origin_upper_left
=
637 pscreen
->get_param(pscreen
,
638 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
639 wpos_options
.fs_coord_origin_lower_left
=
640 pscreen
->get_param(pscreen
,
641 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
642 wpos_options
.fs_coord_pixel_center_integer
=
643 pscreen
->get_param(pscreen
,
644 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
645 wpos_options
.fs_coord_pixel_center_half_integer
=
646 pscreen
->get_param(pscreen
,
647 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
649 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
650 nir_validate_shader(nir
, "after nir_lower_wpos_ytransform");
651 _mesa_add_state_reference(prog
->Parameters
, wposTransformState
);
656 st_link_nir(struct gl_context
*ctx
,
657 struct gl_shader_program
*shader_program
)
659 struct st_context
*st
= st_context(ctx
);
660 struct pipe_screen
*screen
= st
->pipe
->screen
;
662 unsigned last_stage
= 0;
663 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
664 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
668 const nir_shader_compiler_options
*options
=
669 st
->ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
;
670 struct gl_program
*prog
= shader
->Program
;
671 _mesa_copy_linked_program_data(shader_program
, shader
);
675 if (shader_program
->data
->spirv
) {
676 prog
->Parameters
= _mesa_new_parameter_list();
677 /* Parameters will be filled during NIR linking. */
679 prog
->nir
= _mesa_spirv_to_nir(ctx
, shader_program
, shader
->Stage
, options
);
680 set_st_program(prog
, shader_program
, prog
->nir
);
682 validate_ir_tree(shader
->ir
);
684 prog
->Parameters
= _mesa_new_parameter_list();
685 _mesa_generate_parameters_list_for_uniforms(ctx
, shader_program
, shader
,
688 /* Remove reads from output registers. */
689 if (!screen
->get_param(screen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
))
690 lower_output_reads(shader
->Stage
, shader
->ir
);
692 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
694 _mesa_log("GLSL IR for linked %s program %d:\n",
695 _mesa_shader_stage_to_string(shader
->Stage
),
696 shader_program
->Name
);
697 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
701 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
702 _mesa_update_shader_textures_used(shader_program
, prog
);
704 prog
->nir
= glsl_to_nir(st
->ctx
, shader_program
, shader
->Stage
, options
);
705 set_st_program(prog
, shader_program
, prog
->nir
);
706 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
711 if (options
->lower_to_scalar
) {
712 NIR_PASS_V(shader
->Program
->nir
, nir_lower_load_const_to_scalar
);
716 /* For SPIR-V, we have to perform the NIR linking before applying
719 if (shader_program
->data
->spirv
) {
720 static const gl_nir_linker_options opts
= {
721 true /*fill_parameters */
723 if (!gl_nir_link(ctx
, shader_program
, &opts
))
726 nir_build_program_resource_list(ctx
, shader_program
);
728 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
729 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
733 struct gl_program
*prog
= shader
->Program
;
734 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
735 _mesa_update_shader_textures_used(shader_program
, prog
);
737 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
741 /* Linking the stages in the opposite order (from fragment to vertex)
742 * ensures that inter-shader outputs written to in an earlier stage
743 * are eliminated if they are (transitively) not used in a later
746 int next
= last_stage
;
747 for (int i
= next
- 1; i
>= 0; i
--) {
748 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
752 st_nir_link_shaders(&shader
->Program
->nir
,
753 &shader_program
->_LinkedShaders
[next
]->Program
->nir
);
758 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
759 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
763 nir_shader
*nir
= shader
->Program
->nir
;
765 NIR_PASS_V(nir
, st_nir_lower_wpos_ytransform
, shader
->Program
,
768 NIR_PASS_V(nir
, nir_lower_system_values
);
769 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
771 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
772 shader
->Program
->info
= nir
->info
;
773 if (i
== MESA_SHADER_VERTEX
) {
774 /* NIR expands dual-slot inputs out to two locations. We need to
775 * compact things back down GL-style single-slot inputs to avoid
776 * confusing the state tracker.
778 shader
->Program
->info
.inputs_read
=
779 nir_get_single_slot_attribs_mask(nir
->info
.inputs_read
,
780 shader
->Program
->DualSlotInputs
);
784 struct gl_program
*prev_shader
=
785 shader_program
->_LinkedShaders
[prev
]->Program
;
787 /* We can't use nir_compact_varyings with transform feedback, since
788 * the pipe_stream_output->output_register field is based on the
789 * pre-compacted driver_locations.
791 if (!(prev_shader
->sh
.LinkedTransformFeedback
&&
792 prev_shader
->sh
.LinkedTransformFeedback
->NumVarying
> 0))
793 nir_compact_varyings(shader_program
->_LinkedShaders
[prev
]->Program
->nir
,
794 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
796 if (ctx
->Const
.ShaderCompilerOptions
[i
].NirOptions
->vectorize_io
)
797 st_nir_vectorize_io(prev_shader
->nir
, nir
);
802 st_lower_patch_vertices_in(shader_program
);
804 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
805 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
809 st_glsl_to_nir_post_opts(st
, shader
->Program
, shader_program
);
811 assert(shader
->Program
);
812 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
813 _mesa_shader_stage_to_program(i
),
815 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
819 nir_sweep(shader
->Program
->nir
);
821 /* The GLSL IR won't be needed anymore. */
822 ralloc_free(shader
->ir
);
830 st_nir_assign_varying_locations(struct st_context
*st
, nir_shader
*nir
)
832 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
833 /* Needs special handling so drvloc matches the vbo state: */
834 st_nir_assign_vs_in_locations(nir
);
835 /* Re-lower global vars, to deal with any dead VS inputs. */
836 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
838 nir_assign_io_var_locations(&nir
->outputs
,
841 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
842 } else if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
843 nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
844 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
845 nir_assign_io_var_locations(&nir
->inputs
,
848 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
850 nir_assign_io_var_locations(&nir
->outputs
,
853 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
854 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
855 nir_assign_io_var_locations(&nir
->inputs
,
858 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
859 nir_assign_io_var_locations(&nir
->outputs
,
862 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
865 unreachable("invalid shader type");
870 st_nir_lower_samplers(struct pipe_screen
*screen
, nir_shader
*nir
,
871 struct gl_shader_program
*shader_program
,
872 struct gl_program
*prog
)
874 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
875 NIR_PASS_V(nir
, gl_nir_lower_samplers_as_deref
, shader_program
);
877 NIR_PASS_V(nir
, gl_nir_lower_samplers
, shader_program
);
880 prog
->info
.textures_used
= nir
->info
.textures_used
;
881 prog
->info
.textures_used_by_txf
= nir
->info
.textures_used_by_txf
;
885 /* Last third of preparing nir from glsl, which happens after shader
889 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
890 struct gl_shader_program
*shader_program
, nir_shader
*nir
)
892 struct pipe_screen
*screen
= st
->pipe
->screen
;
893 const nir_shader_compiler_options
*options
=
894 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
896 NIR_PASS_V(nir
, nir_split_var_copies
);
897 NIR_PASS_V(nir
, nir_lower_var_copies
);
898 if (options
->lower_all_io_to_temps
||
899 options
->lower_all_io_to_elements
||
900 nir
->info
.stage
== MESA_SHADER_VERTEX
||
901 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
902 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, false);
903 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
904 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, true);
907 st_nir_assign_varying_locations(st
, nir
);
909 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
,
910 st
->ctx
->Const
.Program
[nir
->info
.stage
].MaxAtomicBuffers
);
912 st_nir_assign_uniform_locations(st
->ctx
, prog
,
915 /* Set num_uniforms in number of attribute slots (vec4s) */
916 nir
->num_uniforms
= DIV_ROUND_UP(prog
->Parameters
->NumParameterValues
, 4);
918 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
919 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_type_dword_size
,
920 (nir_lower_io_options
)0);
921 NIR_PASS_V(nir
, nir_lower_uniforms_to_ubo
, 4);
923 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_uniforms_type_size
,
924 (nir_lower_io_options
)0);
927 st_nir_lower_samplers(screen
, nir
, shader_program
, prog
);