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"
44 #include "st_shader_cache.h"
46 #include "compiler/nir/nir.h"
47 #include "compiler/glsl_types.h"
48 #include "compiler/glsl/glsl_to_nir.h"
49 #include "compiler/glsl/gl_nir.h"
50 #include "compiler/glsl/gl_nir_linker.h"
51 #include "compiler/glsl/ir.h"
52 #include "compiler/glsl/ir_optimization.h"
53 #include "compiler/glsl/string_to_uint_map.h"
56 type_size(const struct glsl_type
*type
)
58 return type
->count_attribute_slots(false);
61 /* Depending on PIPE_CAP_TGSI_TEXCOORD (st->needs_texcoord_semantic) we
62 * may need to fix up varying slots so the glsl->nir path is aligned
63 * with the anything->tgsi->nir path.
66 st_nir_fixup_varying_slots(struct st_context
*st
, struct exec_list
*var_list
)
68 if (st
->needs_texcoord_semantic
)
71 nir_foreach_variable(var
, var_list
) {
72 if (var
->data
.location
>= VARYING_SLOT_VAR0
) {
73 var
->data
.location
+= 9;
74 } else if ((var
->data
.location
>= VARYING_SLOT_TEX0
) &&
75 (var
->data
.location
<= VARYING_SLOT_TEX7
)) {
76 var
->data
.location
+= VARYING_SLOT_VAR0
- VARYING_SLOT_TEX0
;
81 /* input location assignment for VS inputs must be handled specially, so
82 * that it is aligned w/ st's vbo state.
83 * (This isn't the case with, for ex, FS inputs, which only need to agree
84 * on varying-slot w/ the VS outputs)
87 st_nir_assign_vs_in_locations(nir_shader
*nir
)
89 nir
->num_inputs
= util_bitcount64(nir
->info
.inputs_read
);
90 nir_foreach_variable_safe(var
, &nir
->inputs
) {
91 /* NIR already assigns dual-slot inputs to two locations so all we have
92 * to do is compact everything down.
94 if (var
->data
.location
== VERT_ATTRIB_EDGEFLAG
) {
95 /* bit of a hack, mirroring st_translate_vertex_program */
96 var
->data
.driver_location
= nir
->num_inputs
++;
97 } else if (nir
->info
.inputs_read
& BITFIELD64_BIT(var
->data
.location
)) {
98 var
->data
.driver_location
=
99 util_bitcount64(nir
->info
.inputs_read
&
100 BITFIELD64_MASK(var
->data
.location
));
102 /* Move unused input variables to the globals list (with no
103 * initialization), to avoid confusing drivers looking through the
104 * inputs array and expecting to find inputs with a driver_location
107 exec_node_remove(&var
->node
);
108 var
->data
.mode
= nir_var_shader_temp
;
109 exec_list_push_tail(&nir
->globals
, &var
->node
);
115 st_nir_lookup_parameter_index(struct gl_program
*prog
, nir_variable
*var
)
117 struct gl_program_parameter_list
*params
= prog
->Parameters
;
119 /* Lookup the first parameter that the uniform storage that match the
122 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
123 int index
= params
->Parameters
[i
].MainUniformStorageIndex
;
124 if (index
== var
->data
.location
)
128 /* TODO: Handle this fallback for SPIR-V. We need this for GLSL e.g. in
129 * dEQP-GLES2.functional.uniform_api.random.3
132 /* is there a better way to do this? If we have something like:
140 * Then what we get in prog->Parameters looks like:
142 * 0: Name=color.f, Type=6, DataType=1406, Size=1
143 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
145 * So the name doesn't match up and _mesa_lookup_parameter_index()
146 * fails. In this case just find the first matching "color.*"..
148 * Note for arrays you could end up w/ color[n].f, for example.
150 * glsl_to_tgsi works slightly differently in this regard. It is
151 * emitting something more low level, so it just translates the
152 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
153 * it just calculates the additional offset of struct field members
154 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
155 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
156 * needs to work backwards to get base var loc from the param-list
157 * which already has them separated out.
159 if (!prog
->sh
.data
->spirv
) {
160 int namelen
= strlen(var
->name
);
161 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
162 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
163 if ((strncmp(p
->Name
, var
->name
, namelen
) == 0) &&
164 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
174 st_nir_assign_uniform_locations(struct gl_context
*ctx
,
175 struct gl_program
*prog
,
176 struct exec_list
*uniform_list
)
181 nir_foreach_variable(uniform
, uniform_list
) {
185 * UBO's have their own address spaces, so don't count them towards the
186 * number of global uniforms
188 if (uniform
->data
.mode
== nir_var_mem_ubo
|| uniform
->data
.mode
== nir_var_mem_ssbo
)
191 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
192 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
193 if (type
->is_sampler()) {
195 shaderidx
+= type_size(uniform
->type
);
198 imageidx
+= type_size(uniform
->type
);
200 } else if (uniform
->state_slots
) {
201 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
202 /* This state reference has already been setup by ir_to_mesa, but we'll
203 * get the same index back here.
207 if (glsl_type_is_struct_or_ifc(type
)) {
210 comps
= glsl_get_vector_elements(type
);
213 if (ctx
->Const
.PackedDriverUniformStorage
) {
214 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
215 stateTokens
, comps
, false);
216 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
218 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
221 loc
= st_nir_lookup_parameter_index(prog
, uniform
);
223 /* We need to check that loc is not -1 here before accessing the
224 * array. It can be negative for example when we have a struct that
225 * only contains opaque types.
227 if (loc
>= 0 && ctx
->Const
.PackedDriverUniformStorage
) {
228 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
232 uniform
->data
.driver_location
= loc
;
237 st_nir_opts(nir_shader
*nir
)
240 unsigned lower_flrp
=
241 (nir
->options
->lower_flrp16
? 16 : 0) |
242 (nir
->options
->lower_flrp32
? 32 : 0) |
243 (nir
->options
->lower_flrp64
? 64 : 0);
248 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
250 /* Linking deals with unused inputs/outputs, but here we can remove
251 * things local to the shader in the hopes that we can cleanup other
252 * things. This pass will also remove variables with only stores, so we
253 * might be able to make progress after it.
255 NIR_PASS(progress
, nir
, nir_remove_dead_variables
,
256 (nir_variable_mode
)(nir_var_function_temp
|
257 nir_var_shader_temp
|
258 nir_var_mem_shared
));
260 NIR_PASS(progress
, nir
, nir_opt_copy_prop_vars
);
261 NIR_PASS(progress
, nir
, nir_opt_dead_write_vars
);
263 if (nir
->options
->lower_to_scalar
) {
264 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
265 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
268 NIR_PASS_V(nir
, nir_lower_alu
);
269 NIR_PASS_V(nir
, nir_lower_pack
);
270 NIR_PASS(progress
, nir
, nir_copy_prop
);
271 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
272 NIR_PASS(progress
, nir
, nir_opt_dce
);
273 if (nir_opt_trivial_continues(nir
)) {
275 NIR_PASS(progress
, nir
, nir_copy_prop
);
276 NIR_PASS(progress
, nir
, nir_opt_dce
);
278 NIR_PASS(progress
, nir
, nir_opt_if
, false);
279 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
280 NIR_PASS(progress
, nir
, nir_opt_cse
);
281 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
283 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
284 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
286 if (lower_flrp
!= 0) {
287 bool lower_flrp_progress
= false;
289 NIR_PASS(lower_flrp_progress
, nir
, nir_lower_flrp
,
291 false /* always_precise */,
292 nir
->options
->lower_ffma
);
293 if (lower_flrp_progress
) {
294 NIR_PASS(progress
, nir
,
295 nir_opt_constant_folding
);
299 /* Nothing should rematerialize any flrps, so we only need to do this
305 NIR_PASS(progress
, nir
, nir_opt_access
);
307 NIR_PASS(progress
, nir
, nir_opt_undef
);
308 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
309 if (nir
->options
->max_unroll_iterations
) {
310 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
316 shared_type_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
318 assert(glsl_type_is_vector_or_scalar(type
));
320 uint32_t comp_size
= glsl_type_is_boolean(type
)
321 ? 4 : glsl_get_bit_size(type
) / 8;
322 unsigned length
= glsl_get_vector_elements(type
);
323 *size
= comp_size
* length
,
324 *align
= comp_size
* (length
== 3 ? 4 : length
);
327 /* First third of converting glsl_to_nir.. this leaves things in a pre-
328 * nir_lower_io state, so that shader variants can more easily insert/
329 * replace variables, etc.
332 st_nir_preprocess(struct st_context
*st
, struct gl_program
*prog
,
333 struct gl_shader_program
*shader_program
,
334 gl_shader_stage stage
)
336 const nir_shader_compiler_options
*options
=
337 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
340 options
->lower_int64_options
|| options
->lower_doubles_options
;
341 nir_shader
*nir
= prog
->nir
;
343 /* Set the next shader stage hint for VS and TES. */
344 if (!nir
->info
.separate_shader
&&
345 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
346 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
348 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
349 unsigned stages_mask
=
350 ~prev_stages
& shader_program
->data
->linked_stages
;
352 nir
->info
.next_stage
= stages_mask
?
353 (gl_shader_stage
) u_bit_scan(&stages_mask
) : MESA_SHADER_FRAGMENT
;
355 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
358 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
359 if (!st
->ctx
->SoftFP64
&& nir
->info
.uses_64bit
&&
360 (options
->lower_doubles_options
& nir_lower_fp64_full_software
) != 0) {
361 st
->ctx
->SoftFP64
= glsl_float64_funcs_to_nir(st
->ctx
, options
);
364 nir_variable_mode mask
=
365 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
366 nir_remove_dead_variables(nir
, mask
);
368 if (options
->lower_all_io_to_temps
||
369 nir
->info
.stage
== MESA_SHADER_VERTEX
||
370 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
371 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
372 nir_shader_get_entrypoint(nir
),
374 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
375 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
376 nir_shader_get_entrypoint(nir
),
380 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
381 NIR_PASS_V(nir
, nir_split_var_copies
);
382 NIR_PASS_V(nir
, nir_lower_var_copies
);
384 if (options
->lower_to_scalar
) {
385 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
388 /* before buffers and vars_to_ssa */
389 NIR_PASS_V(nir
, gl_nir_lower_bindless_images
);
392 /* TODO: Change GLSL to not lower shared memory. */
393 if (prog
->nir
->info
.stage
== MESA_SHADER_COMPUTE
&&
394 shader_program
->data
->spirv
) {
395 NIR_PASS_V(prog
->nir
, nir_lower_vars_to_explicit_types
,
396 nir_var_mem_shared
, shared_type_info
);
397 NIR_PASS_V(prog
->nir
, nir_lower_explicit_io
,
398 nir_var_mem_shared
, nir_address_format_32bit_offset
);
401 NIR_PASS_V(nir
, gl_nir_lower_buffers
, shader_program
);
402 /* Do a round of constant folding to clean up address calculations */
403 NIR_PASS_V(nir
, nir_opt_constant_folding
);
406 bool lowered_64bit_ops
= false;
407 if (options
->lower_doubles_options
) {
408 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_doubles
,
409 st
->ctx
->SoftFP64
, options
->lower_doubles_options
);
411 if (options
->lower_int64_options
) {
412 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_int64
,
413 options
->lower_int64_options
);
416 if (lowered_64bit_ops
)
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 const nir_state_slot
*const slots
= var
->state_slots
;
440 const struct glsl_type
*type
= glsl_without_array(var
->type
);
441 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
443 if (glsl_type_is_struct_or_ifc(type
)) {
444 /* Builtin struct require specical handling for now we just
445 * make all members vec4. See st_nir_lower_builtin.
449 comps
= glsl_get_vector_elements(type
);
452 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
453 _mesa_add_sized_state_reference(prog
->Parameters
,
457 _mesa_add_state_reference(prog
->Parameters
,
464 /* Avoid reallocation of the program parameter list, because the uniform
465 * storage is only associated with the original parameter list.
466 * This should be enough for Bitmap and DrawPixels constants.
468 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
470 /* This has to be done last. Any operation the can cause
471 * prog->ParameterValues to get reallocated (e.g., anything that adds a
472 * program constant) has to happen before creating this linkage.
474 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
);
476 st_set_prog_affected_state_flags(prog
);
478 /* None of the builtins being lowered here can be produced by SPIR-V. See
479 * _mesa_builtin_uniform_desc.
481 if (!shader_program
->data
->spirv
)
482 NIR_PASS_V(nir
, st_nir_lower_builtin
);
484 NIR_PASS_V(nir
, gl_nir_lower_atomics
, shader_program
, true);
485 NIR_PASS_V(nir
, nir_opt_intrinsics
);
487 nir_variable_mode mask
= nir_var_function_temp
;
488 nir_remove_dead_variables(nir
, mask
);
490 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
492 _mesa_log("NIR IR for linked %s program %d:\n",
493 _mesa_shader_stage_to_string(prog
->info
.stage
),
494 shader_program
->Name
);
495 nir_print_shader(nir
, _mesa_get_log_file());
501 set_st_program(struct gl_program
*prog
,
502 struct gl_shader_program
*shader_program
,
505 struct st_vertex_program
*stvp
;
506 struct st_common_program
*stp
;
507 struct st_fragment_program
*stfp
;
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 case MESA_SHADER_COMPUTE
:
520 stp
= (struct st_common_program
*)prog
;
521 stp
->shader_program
= shader_program
;
522 stp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
523 stp
->tgsi
.ir
.nir
= nir
;
525 case MESA_SHADER_FRAGMENT
:
526 stfp
= (struct st_fragment_program
*)prog
;
527 stfp
->shader_program
= shader_program
;
528 stfp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
529 stfp
->tgsi
.ir
.nir
= nir
;
532 unreachable("unknown shader stage");
537 st_nir_vectorize_io(nir_shader
*producer
, nir_shader
*consumer
)
539 NIR_PASS_V(producer
, nir_lower_io_to_vector
, nir_var_shader_out
);
540 NIR_PASS_V(producer
, nir_opt_combine_stores
, nir_var_shader_out
);
541 NIR_PASS_V(consumer
, nir_lower_io_to_vector
, nir_var_shader_in
);
543 if ((producer
)->info
.stage
!= MESA_SHADER_TESS_CTRL
) {
544 /* Calling lower_io_to_vector creates output variable writes with
545 * write-masks. We only support these for TCS outputs, so for other
546 * stages, we need to call nir_lower_io_to_temporaries to get rid of
547 * them. This, in turn, creates temporary variables and extra
548 * copy_deref intrinsics that we need to clean up.
550 NIR_PASS_V(producer
, nir_lower_io_to_temporaries
,
551 nir_shader_get_entrypoint(producer
), true, false);
552 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
553 NIR_PASS_V(producer
, nir_split_var_copies
);
554 NIR_PASS_V(producer
, nir_lower_var_copies
);
559 st_nir_link_shaders(nir_shader
**producer
, nir_shader
**consumer
)
561 if ((*producer
)->options
->lower_to_scalar
) {
562 NIR_PASS_V(*producer
, nir_lower_io_to_scalar_early
, nir_var_shader_out
);
563 NIR_PASS_V(*consumer
, nir_lower_io_to_scalar_early
, nir_var_shader_in
);
566 nir_lower_io_arrays_to_elements(*producer
, *consumer
);
568 st_nir_opts(*producer
);
569 st_nir_opts(*consumer
);
571 if (nir_link_opt_varyings(*producer
, *consumer
))
572 st_nir_opts(*consumer
);
574 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
575 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
577 if (nir_remove_unused_varyings(*producer
, *consumer
)) {
578 NIR_PASS_V(*producer
, nir_lower_global_vars_to_local
);
579 NIR_PASS_V(*consumer
, nir_lower_global_vars_to_local
);
581 st_nir_opts(*producer
);
582 st_nir_opts(*consumer
);
584 /* Optimizations can cause varyings to become unused.
585 * nir_compact_varyings() depends on all dead varyings being removed so
586 * we need to call nir_remove_dead_variables() again here.
588 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
589 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
594 st_lower_patch_vertices_in(struct gl_shader_program
*shader_prog
)
596 struct gl_linked_shader
*linked_tcs
=
597 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_CTRL
];
598 struct gl_linked_shader
*linked_tes
=
599 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_EVAL
];
601 /* If we have a TCS and TES linked together, lower TES patch vertices. */
602 if (linked_tcs
&& linked_tes
) {
603 nir_shader
*tcs_nir
= linked_tcs
->Program
->nir
;
604 nir_shader
*tes_nir
= linked_tes
->Program
->nir
;
606 /* The TES input vertex count is the TCS output vertex count,
607 * lower TES gl_PatchVerticesIn to a constant.
609 uint32_t tes_patch_verts
= tcs_nir
->info
.tess
.tcs_vertices_out
;
610 NIR_PASS_V(tes_nir
, nir_lower_patch_vertices
, tes_patch_verts
, NULL
);
617 st_nir_lower_wpos_ytransform(struct nir_shader
*nir
,
618 struct gl_program
*prog
,
619 struct pipe_screen
*pscreen
)
621 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
624 static const gl_state_index16 wposTransformState
[STATE_LENGTH
] = {
625 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
627 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
629 memcpy(wpos_options
.state_tokens
, wposTransformState
,
630 sizeof(wpos_options
.state_tokens
));
631 wpos_options
.fs_coord_origin_upper_left
=
632 pscreen
->get_param(pscreen
,
633 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
634 wpos_options
.fs_coord_origin_lower_left
=
635 pscreen
->get_param(pscreen
,
636 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
637 wpos_options
.fs_coord_pixel_center_integer
=
638 pscreen
->get_param(pscreen
,
639 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
640 wpos_options
.fs_coord_pixel_center_half_integer
=
641 pscreen
->get_param(pscreen
,
642 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
644 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
645 nir_validate_shader(nir
, "after nir_lower_wpos_ytransform");
646 _mesa_add_state_reference(prog
->Parameters
, wposTransformState
);
651 st_link_nir(struct gl_context
*ctx
,
652 struct gl_shader_program
*shader_program
)
654 struct st_context
*st
= st_context(ctx
);
655 struct pipe_screen
*screen
= st
->pipe
->screen
;
657 unsigned last_stage
= 0;
658 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
659 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
663 const nir_shader_compiler_options
*options
=
664 st
->ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
;
665 struct gl_program
*prog
= shader
->Program
;
666 _mesa_copy_linked_program_data(shader_program
, shader
);
670 if (shader_program
->data
->spirv
) {
671 prog
->Parameters
= _mesa_new_parameter_list();
672 /* Parameters will be filled during NIR linking. */
674 prog
->nir
= _mesa_spirv_to_nir(ctx
, shader_program
, shader
->Stage
, options
);
675 set_st_program(prog
, shader_program
, prog
->nir
);
677 validate_ir_tree(shader
->ir
);
679 prog
->Parameters
= _mesa_new_parameter_list();
680 _mesa_generate_parameters_list_for_uniforms(ctx
, shader_program
, shader
,
683 /* Remove reads from output registers. */
684 if (!screen
->get_param(screen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
))
685 lower_output_reads(shader
->Stage
, shader
->ir
);
687 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
689 _mesa_log("GLSL IR for linked %s program %d:\n",
690 _mesa_shader_stage_to_string(shader
->Stage
),
691 shader_program
->Name
);
692 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
696 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
697 _mesa_update_shader_textures_used(shader_program
, prog
);
699 prog
->nir
= glsl_to_nir(st
->ctx
, shader_program
, shader
->Stage
, options
);
700 set_st_program(prog
, shader_program
, prog
->nir
);
701 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
706 if (options
->lower_to_scalar
) {
707 NIR_PASS_V(shader
->Program
->nir
, nir_lower_load_const_to_scalar
);
711 /* For SPIR-V, we have to perform the NIR linking before applying
714 if (shader_program
->data
->spirv
) {
715 static const gl_nir_linker_options opts
= {
716 true /*fill_parameters */
718 if (!gl_nir_link(ctx
, shader_program
, &opts
))
721 nir_build_program_resource_list(ctx
, shader_program
);
723 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
724 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
728 struct gl_program
*prog
= shader
->Program
;
729 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
730 _mesa_update_shader_textures_used(shader_program
, prog
);
732 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
736 /* Linking the stages in the opposite order (from fragment to vertex)
737 * ensures that inter-shader outputs written to in an earlier stage
738 * are eliminated if they are (transitively) not used in a later
741 int next
= last_stage
;
742 for (int i
= next
- 1; i
>= 0; i
--) {
743 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
747 st_nir_link_shaders(&shader
->Program
->nir
,
748 &shader_program
->_LinkedShaders
[next
]->Program
->nir
);
753 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
754 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
758 nir_shader
*nir
= shader
->Program
->nir
;
760 NIR_PASS_V(nir
, st_nir_lower_wpos_ytransform
, shader
->Program
,
763 NIR_PASS_V(nir
, nir_lower_system_values
);
764 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
766 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
767 shader
->Program
->info
= nir
->info
;
768 if (i
== MESA_SHADER_VERTEX
) {
769 /* NIR expands dual-slot inputs out to two locations. We need to
770 * compact things back down GL-style single-slot inputs to avoid
771 * confusing the state tracker.
773 shader
->Program
->info
.inputs_read
=
774 nir_get_single_slot_attribs_mask(nir
->info
.inputs_read
,
775 shader
->Program
->DualSlotInputs
);
779 struct gl_program
*prev_shader
=
780 shader_program
->_LinkedShaders
[prev
]->Program
;
782 /* We can't use nir_compact_varyings with transform feedback, since
783 * the pipe_stream_output->output_register field is based on the
784 * pre-compacted driver_locations.
786 if (!(prev_shader
->sh
.LinkedTransformFeedback
&&
787 prev_shader
->sh
.LinkedTransformFeedback
->NumVarying
> 0))
788 nir_compact_varyings(shader_program
->_LinkedShaders
[prev
]->Program
->nir
,
789 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
791 if (ctx
->Const
.ShaderCompilerOptions
[i
].NirOptions
->vectorize_io
)
792 st_nir_vectorize_io(prev_shader
->nir
, nir
);
797 st_lower_patch_vertices_in(shader_program
);
799 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
800 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
804 struct gl_program
*prog
= shader
->Program
;
805 st_glsl_to_nir_post_opts(st
, prog
, shader_program
);
807 /* Initialize st_vertex_program members. */
808 if (i
== MESA_SHADER_VERTEX
)
809 st_prepare_vertex_program(st_vertex_program(prog
));
811 /* Get pipe_stream_output_info. */
812 if (i
== MESA_SHADER_VERTEX
||
813 i
== MESA_SHADER_TESS_EVAL
||
814 i
== MESA_SHADER_GEOMETRY
)
815 st_translate_stream_output_info(prog
);
817 st_store_ir_in_disk_cache(st
, prog
, true);
819 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
820 _mesa_shader_stage_to_program(i
),
822 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
826 nir_sweep(prog
->nir
);
828 /* The GLSL IR won't be needed anymore. */
829 ralloc_free(shader
->ir
);
837 st_nir_assign_varying_locations(struct st_context
*st
, nir_shader
*nir
)
839 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
840 /* Needs special handling so drvloc matches the vbo state: */
841 st_nir_assign_vs_in_locations(nir
);
842 /* Re-lower global vars, to deal with any dead VS inputs. */
843 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
845 nir_assign_io_var_locations(&nir
->outputs
,
848 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
849 } else if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
850 nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
851 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
852 nir_assign_io_var_locations(&nir
->inputs
,
855 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
857 nir_assign_io_var_locations(&nir
->outputs
,
860 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
861 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
862 nir_assign_io_var_locations(&nir
->inputs
,
865 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
866 nir_assign_io_var_locations(&nir
->outputs
,
869 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
872 unreachable("invalid shader type");
877 st_nir_lower_samplers(struct pipe_screen
*screen
, nir_shader
*nir
,
878 struct gl_shader_program
*shader_program
,
879 struct gl_program
*prog
)
881 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
882 NIR_PASS_V(nir
, gl_nir_lower_samplers_as_deref
, shader_program
);
884 NIR_PASS_V(nir
, gl_nir_lower_samplers
, shader_program
);
887 prog
->info
.textures_used
= nir
->info
.textures_used
;
888 prog
->info
.textures_used_by_txf
= nir
->info
.textures_used_by_txf
;
892 /* Last third of preparing nir from glsl, which happens after shader
896 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
897 struct gl_shader_program
*shader_program
, nir_shader
*nir
)
899 struct pipe_screen
*screen
= st
->pipe
->screen
;
900 const nir_shader_compiler_options
*options
=
901 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
903 NIR_PASS_V(nir
, nir_split_var_copies
);
904 NIR_PASS_V(nir
, nir_lower_var_copies
);
905 if (options
->lower_all_io_to_temps
||
906 options
->lower_all_io_to_elements
||
907 nir
->info
.stage
== MESA_SHADER_VERTEX
||
908 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
909 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, false);
910 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
911 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, true);
914 st_nir_assign_varying_locations(st
, nir
);
916 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
,
917 st
->ctx
->Const
.Program
[nir
->info
.stage
].MaxAtomicBuffers
);
919 st_nir_assign_uniform_locations(st
->ctx
, prog
,
922 /* Set num_uniforms in number of attribute slots (vec4s) */
923 nir
->num_uniforms
= DIV_ROUND_UP(prog
->Parameters
->NumParameterValues
, 4);
925 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
926 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_type_dword_size
,
927 (nir_lower_io_options
)0);
928 NIR_PASS_V(nir
, nir_lower_uniforms_to_ubo
, 4);
930 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_uniforms_type_size
,
931 (nir_lower_io_options
)0);
934 st_nir_lower_samplers(screen
, nir
, shader_program
, prog
);