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/context.h"
35 #include "main/mtypes.h"
36 #include "main/errors.h"
37 #include "main/glspirv.h"
38 #include "main/shaderapi.h"
39 #include "main/uniforms.h"
41 #include "main/shaderobj.h"
42 #include "st_context.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
, nir_shader
*shader
,
67 nir_variable_mode mode
)
69 if (st
->needs_texcoord_semantic
)
72 nir_foreach_variable_with_modes(var
, shader
, mode
) {
73 if (var
->data
.location
>= VARYING_SLOT_VAR0
) {
74 var
->data
.location
+= 9;
75 } else if (var
->data
.location
== VARYING_SLOT_PNTC
) {
76 var
->data
.location
= VARYING_SLOT_VAR8
;
77 } else if ((var
->data
.location
>= VARYING_SLOT_TEX0
) &&
78 (var
->data
.location
<= VARYING_SLOT_TEX7
)) {
79 var
->data
.location
+= VARYING_SLOT_VAR0
- VARYING_SLOT_TEX0
;
85 st_shader_gather_info(nir_shader
*nir
, struct gl_program
*prog
)
87 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
89 /* Copy the info we just generated back into the gl_program */
90 const char *prog_name
= prog
->info
.name
;
91 const char *prog_label
= prog
->info
.label
;
92 prog
->info
= nir
->info
;
93 prog
->info
.name
= prog_name
;
94 prog
->info
.label
= prog_label
;
97 /* input location assignment for VS inputs must be handled specially, so
98 * that it is aligned w/ st's vbo state.
99 * (This isn't the case with, for ex, FS inputs, which only need to agree
100 * on varying-slot w/ the VS outputs)
103 st_nir_assign_vs_in_locations(struct nir_shader
*nir
)
105 if (nir
->info
.stage
!= MESA_SHADER_VERTEX
|| nir
->info
.io_lowered
)
108 nir
->num_inputs
= util_bitcount64(nir
->info
.inputs_read
);
110 bool removed_inputs
= false;
112 nir_foreach_shader_in_variable_safe(var
, nir
) {
113 /* NIR already assigns dual-slot inputs to two locations so all we have
114 * to do is compact everything down.
116 if (nir
->info
.inputs_read
& BITFIELD64_BIT(var
->data
.location
)) {
117 var
->data
.driver_location
=
118 util_bitcount64(nir
->info
.inputs_read
&
119 BITFIELD64_MASK(var
->data
.location
));
121 /* Convert unused input variables to shader_temp (with no
122 * initialization), to avoid confusing drivers looking through the
123 * inputs array and expecting to find inputs with a driver_location
126 var
->data
.mode
= nir_var_shader_temp
;
127 removed_inputs
= true;
131 /* Re-lower global vars, to deal with any dead VS inputs. */
133 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
137 st_nir_lookup_parameter_index(struct gl_program
*prog
, nir_variable
*var
)
139 struct gl_program_parameter_list
*params
= prog
->Parameters
;
141 /* Lookup the first parameter that the uniform storage that match the
144 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
145 int index
= params
->Parameters
[i
].MainUniformStorageIndex
;
146 if (index
== var
->data
.location
)
150 /* TODO: Handle this fallback for SPIR-V. We need this for GLSL e.g. in
151 * dEQP-GLES2.functional.uniform_api.random.3
154 /* is there a better way to do this? If we have something like:
162 * Then what we get in prog->Parameters looks like:
164 * 0: Name=color.f, Type=6, DataType=1406, Size=1
165 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
167 * So the name doesn't match up and _mesa_lookup_parameter_index()
168 * fails. In this case just find the first matching "color.*"..
170 * Note for arrays you could end up w/ color[n].f, for example.
172 * glsl_to_tgsi works slightly differently in this regard. It is
173 * emitting something more low level, so it just translates the
174 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
175 * it just calculates the additional offset of struct field members
176 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
177 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
178 * needs to work backwards to get base var loc from the param-list
179 * which already has them separated out.
181 if (!prog
->sh
.data
->spirv
) {
182 int namelen
= strlen(var
->name
);
183 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
184 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
185 if ((strncmp(p
->Name
, var
->name
, namelen
) == 0) &&
186 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
196 st_nir_assign_uniform_locations(struct gl_context
*ctx
,
197 struct gl_program
*prog
,
203 nir_foreach_uniform_variable(uniform
, nir
) {
206 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
207 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
208 if (type
->is_sampler()) {
210 shaderidx
+= type_size(uniform
->type
);
213 imageidx
+= type_size(uniform
->type
);
215 } else if (uniform
->state_slots
) {
216 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
217 /* This state reference has already been setup by ir_to_mesa, but we'll
218 * get the same index back here.
222 if (glsl_type_is_struct_or_ifc(type
)) {
225 comps
= glsl_get_vector_elements(type
);
228 if (ctx
->Const
.PackedDriverUniformStorage
) {
229 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
230 stateTokens
, comps
, false);
231 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
233 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
236 loc
= st_nir_lookup_parameter_index(prog
, uniform
);
238 /* We need to check that loc is not -1 here before accessing the
239 * array. It can be negative for example when we have a struct that
240 * only contains opaque types.
242 if (loc
>= 0 && ctx
->Const
.PackedDriverUniformStorage
) {
243 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
247 uniform
->data
.driver_location
= loc
;
252 st_nir_opts(nir_shader
*nir
)
259 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
261 /* Linking deals with unused inputs/outputs, but here we can remove
262 * things local to the shader in the hopes that we can cleanup other
263 * things. This pass will also remove variables with only stores, so we
264 * might be able to make progress after it.
266 NIR_PASS(progress
, nir
, nir_remove_dead_variables
,
267 (nir_variable_mode
)(nir_var_function_temp
|
268 nir_var_shader_temp
|
272 NIR_PASS(progress
, nir
, nir_opt_copy_prop_vars
);
273 NIR_PASS(progress
, nir
, nir_opt_dead_write_vars
);
275 if (nir
->options
->lower_to_scalar
) {
276 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
277 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
280 NIR_PASS_V(nir
, nir_lower_alu
);
281 NIR_PASS_V(nir
, nir_lower_pack
);
282 NIR_PASS(progress
, nir
, nir_copy_prop
);
283 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
284 NIR_PASS(progress
, nir
, nir_opt_dce
);
285 if (nir_opt_trivial_continues(nir
)) {
287 NIR_PASS(progress
, nir
, nir_copy_prop
);
288 NIR_PASS(progress
, nir
, nir_opt_dce
);
290 NIR_PASS(progress
, nir
, nir_opt_if
, false);
291 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
292 NIR_PASS(progress
, nir
, nir_opt_cse
);
293 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
295 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
296 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
298 if (!nir
->info
.flrp_lowered
) {
299 unsigned lower_flrp
=
300 (nir
->options
->lower_flrp16
? 16 : 0) |
301 (nir
->options
->lower_flrp32
? 32 : 0) |
302 (nir
->options
->lower_flrp64
? 64 : 0);
305 bool lower_flrp_progress
= false;
307 NIR_PASS(lower_flrp_progress
, nir
, nir_lower_flrp
,
309 false /* always_precise */,
310 nir
->options
->lower_ffma
);
311 if (lower_flrp_progress
) {
312 NIR_PASS(progress
, nir
,
313 nir_opt_constant_folding
);
318 /* Nothing should rematerialize any flrps, so we only need to do this
321 nir
->info
.flrp_lowered
= true;
324 NIR_PASS(progress
, nir
, nir_opt_undef
);
325 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
326 if (nir
->options
->max_unroll_iterations
) {
327 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
333 shared_type_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
335 assert(glsl_type_is_vector_or_scalar(type
));
337 uint32_t comp_size
= glsl_type_is_boolean(type
)
338 ? 4 : glsl_get_bit_size(type
) / 8;
339 unsigned length
= glsl_get_vector_elements(type
);
340 *size
= comp_size
* length
,
341 *align
= comp_size
* (length
== 3 ? 4 : length
);
344 /* First third of converting glsl_to_nir.. this leaves things in a pre-
345 * nir_lower_io state, so that shader variants can more easily insert/
346 * replace variables, etc.
349 st_nir_preprocess(struct st_context
*st
, struct gl_program
*prog
,
350 struct gl_shader_program
*shader_program
,
351 gl_shader_stage stage
)
353 struct pipe_screen
*screen
= st
->pipe
->screen
;
354 const nir_shader_compiler_options
*options
=
355 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
357 nir_shader
*nir
= prog
->nir
;
359 /* Set the next shader stage hint for VS and TES. */
360 if (!nir
->info
.separate_shader
&&
361 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
362 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
364 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
365 unsigned stages_mask
=
366 ~prev_stages
& shader_program
->data
->linked_stages
;
368 nir
->info
.next_stage
= stages_mask
?
369 (gl_shader_stage
) u_bit_scan(&stages_mask
) : MESA_SHADER_FRAGMENT
;
371 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
374 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
375 if (!st
->ctx
->SoftFP64
&& nir
->info
.uses_64bit
&&
376 (options
->lower_doubles_options
& nir_lower_fp64_full_software
) != 0) {
377 st
->ctx
->SoftFP64
= glsl_float64_funcs_to_nir(st
->ctx
, options
);
380 /* ES has strict SSO validation rules for shader IO matching so we can't
381 * remove dead IO until the resource list has been built. Here we skip
382 * removing them until later. This will potentially make the IO lowering
383 * calls below do a little extra work but should otherwise have no impact.
385 if (!_mesa_is_gles(st
->ctx
) || !nir
->info
.separate_shader
) {
386 nir_variable_mode mask
=
387 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
388 nir_remove_dead_variables(nir
, mask
, NULL
);
391 if (options
->lower_all_io_to_temps
||
392 nir
->info
.stage
== MESA_SHADER_VERTEX
||
393 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
394 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
395 nir_shader_get_entrypoint(nir
),
397 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
||
398 !screen
->get_param(screen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
)) {
399 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
400 nir_shader_get_entrypoint(nir
),
404 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
405 NIR_PASS_V(nir
, nir_split_var_copies
);
406 NIR_PASS_V(nir
, nir_lower_var_copies
);
408 if (options
->lower_to_scalar
) {
409 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
412 /* before buffers and vars_to_ssa */
413 NIR_PASS_V(nir
, gl_nir_lower_images
, true);
415 /* TODO: Change GLSL to not lower shared memory. */
416 if (prog
->nir
->info
.stage
== MESA_SHADER_COMPUTE
&&
417 shader_program
->data
->spirv
) {
418 NIR_PASS_V(prog
->nir
, nir_lower_vars_to_explicit_types
,
419 nir_var_mem_shared
, shared_type_info
);
420 NIR_PASS_V(prog
->nir
, nir_lower_explicit_io
,
421 nir_var_mem_shared
, nir_address_format_32bit_offset
);
424 /* Do a round of constant folding to clean up address calculations */
425 NIR_PASS_V(nir
, nir_opt_constant_folding
);
428 /* Second third of converting glsl_to_nir. This creates uniforms, gathers
429 * info on varyings, etc after NIR link time opts have been applied.
432 st_glsl_to_nir_post_opts(struct st_context
*st
, struct gl_program
*prog
,
433 struct gl_shader_program
*shader_program
)
435 nir_shader
*nir
= prog
->nir
;
436 struct pipe_screen
*screen
= st
->pipe
->screen
;
438 /* Make a pass over the IR to add state references for any built-in
439 * uniforms that are used. This has to be done now (during linking).
440 * Code generation doesn't happen until the first time this shader is
441 * used for rendering. Waiting until then to generate the parameters is
442 * too late. At that point, the values for the built-in uniforms won't
443 * get sent to the shader.
445 nir_foreach_uniform_variable(var
, nir
) {
446 const nir_state_slot
*const slots
= var
->state_slots
;
448 const struct glsl_type
*type
= glsl_without_array(var
->type
);
449 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
451 if (glsl_type_is_struct_or_ifc(type
)) {
452 comps
= _mesa_program_state_value_size(slots
[i
].tokens
);
454 comps
= glsl_get_vector_elements(type
);
457 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
458 _mesa_add_sized_state_reference(prog
->Parameters
,
462 _mesa_add_state_reference(prog
->Parameters
,
469 /* Avoid reallocation of the program parameter list, because the uniform
470 * storage is only associated with the original parameter list.
471 * This should be enough for Bitmap and DrawPixels constants.
473 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
475 /* This has to be done last. Any operation the can cause
476 * prog->ParameterValues to get reallocated (e.g., anything that adds a
477 * program constant) has to happen before creating this linkage.
479 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
);
481 st_set_prog_affected_state_flags(prog
);
483 /* None of the builtins being lowered here can be produced by SPIR-V. See
484 * _mesa_builtin_uniform_desc. Also drivers that support packed uniform
485 * storage don't need to lower builtins.
487 if (!shader_program
->data
->spirv
&&
488 !st
->ctx
->Const
.PackedDriverUniformStorage
)
489 NIR_PASS_V(nir
, st_nir_lower_builtin
);
491 if (!screen
->get_param(screen
, PIPE_CAP_NIR_ATOMICS_AS_DEREF
))
492 NIR_PASS_V(nir
, gl_nir_lower_atomics
, shader_program
, true);
494 NIR_PASS_V(nir
, nir_opt_intrinsics
);
496 /* Lower 64-bit ops. */
497 if (nir
->options
->lower_int64_options
||
498 nir
->options
->lower_doubles_options
) {
499 bool lowered_64bit_ops
= false;
500 if (nir
->options
->lower_doubles_options
) {
501 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_doubles
,
502 st
->ctx
->SoftFP64
, nir
->options
->lower_doubles_options
);
504 if (nir
->options
->lower_int64_options
)
505 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_int64
);
507 if (lowered_64bit_ops
)
511 nir_variable_mode mask
= (nir_variable_mode
)
512 (nir_var_shader_in
| nir_var_shader_out
| nir_var_function_temp
);
513 nir_remove_dead_variables(nir
, mask
, NULL
);
515 if (!st
->has_hw_atomics
&& !screen
->get_param(screen
, PIPE_CAP_NIR_ATOMICS_AS_DEREF
))
516 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
);
518 st_finalize_nir_before_variants(nir
);
520 if (st
->allow_st_finalize_nir_twice
)
521 st_finalize_nir(st
, prog
, shader_program
, nir
, true);
523 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
525 _mesa_log("NIR IR for linked %s program %d:\n",
526 _mesa_shader_stage_to_string(prog
->info
.stage
),
527 shader_program
->Name
);
528 nir_print_shader(nir
, _mesa_get_log_file());
534 st_nir_vectorize_io(nir_shader
*producer
, nir_shader
*consumer
)
536 NIR_PASS_V(producer
, nir_lower_io_to_vector
, nir_var_shader_out
);
537 NIR_PASS_V(producer
, nir_opt_combine_stores
, nir_var_shader_out
);
538 NIR_PASS_V(consumer
, nir_lower_io_to_vector
, nir_var_shader_in
);
540 if ((producer
)->info
.stage
!= MESA_SHADER_TESS_CTRL
) {
541 /* Calling lower_io_to_vector creates output variable writes with
542 * write-masks. We only support these for TCS outputs, so for other
543 * stages, we need to call nir_lower_io_to_temporaries to get rid of
544 * them. This, in turn, creates temporary variables and extra
545 * copy_deref intrinsics that we need to clean up.
547 NIR_PASS_V(producer
, nir_lower_io_to_temporaries
,
548 nir_shader_get_entrypoint(producer
), true, false);
549 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
550 NIR_PASS_V(producer
, nir_split_var_copies
);
551 NIR_PASS_V(producer
, nir_lower_var_copies
);
556 st_nir_link_shaders(nir_shader
*producer
, nir_shader
*consumer
)
558 if (producer
->options
->lower_to_scalar
) {
559 NIR_PASS_V(producer
, nir_lower_io_to_scalar_early
, nir_var_shader_out
);
560 NIR_PASS_V(consumer
, nir_lower_io_to_scalar_early
, nir_var_shader_in
);
563 nir_lower_io_arrays_to_elements(producer
, consumer
);
565 st_nir_opts(producer
);
566 st_nir_opts(consumer
);
568 if (nir_link_opt_varyings(producer
, consumer
))
569 st_nir_opts(consumer
);
571 NIR_PASS_V(producer
, nir_remove_dead_variables
, nir_var_shader_out
, NULL
);
572 NIR_PASS_V(consumer
, nir_remove_dead_variables
, nir_var_shader_in
, NULL
);
574 if (nir_remove_unused_varyings(producer
, consumer
)) {
575 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
576 NIR_PASS_V(consumer
, nir_lower_global_vars_to_local
);
578 st_nir_opts(producer
);
579 st_nir_opts(consumer
);
581 /* Optimizations can cause varyings to become unused.
582 * nir_compact_varyings() depends on all dead varyings being removed so
583 * we need to call nir_remove_dead_variables() again here.
585 NIR_PASS_V(producer
, nir_remove_dead_variables
, nir_var_shader_out
,
587 NIR_PASS_V(consumer
, nir_remove_dead_variables
, nir_var_shader_in
,
593 st_lower_patch_vertices_in(struct gl_shader_program
*shader_prog
)
595 struct gl_linked_shader
*linked_tcs
=
596 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_CTRL
];
597 struct gl_linked_shader
*linked_tes
=
598 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_EVAL
];
600 /* If we have a TCS and TES linked together, lower TES patch vertices. */
601 if (linked_tcs
&& linked_tes
) {
602 nir_shader
*tcs_nir
= linked_tcs
->Program
->nir
;
603 nir_shader
*tes_nir
= linked_tes
->Program
->nir
;
605 /* The TES input vertex count is the TCS output vertex count,
606 * lower TES gl_PatchVerticesIn to a constant.
608 uint32_t tes_patch_verts
= tcs_nir
->info
.tess
.tcs_vertices_out
;
609 NIR_PASS_V(tes_nir
, nir_lower_patch_vertices
, tes_patch_verts
, NULL
);
616 st_nir_lower_wpos_ytransform(struct nir_shader
*nir
,
617 struct gl_program
*prog
,
618 struct pipe_screen
*pscreen
)
620 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
623 static const gl_state_index16 wposTransformState
[STATE_LENGTH
] = {
624 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
626 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
628 memcpy(wpos_options
.state_tokens
, wposTransformState
,
629 sizeof(wpos_options
.state_tokens
));
630 wpos_options
.fs_coord_origin_upper_left
=
631 pscreen
->get_param(pscreen
,
632 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
633 wpos_options
.fs_coord_origin_lower_left
=
634 pscreen
->get_param(pscreen
,
635 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
636 wpos_options
.fs_coord_pixel_center_integer
=
637 pscreen
->get_param(pscreen
,
638 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
639 wpos_options
.fs_coord_pixel_center_half_integer
=
640 pscreen
->get_param(pscreen
,
641 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
643 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
644 nir_validate_shader(nir
, "after nir_lower_wpos_ytransform");
645 _mesa_add_state_reference(prog
->Parameters
, wposTransformState
);
650 st_link_nir(struct gl_context
*ctx
,
651 struct gl_shader_program
*shader_program
)
653 struct st_context
*st
= st_context(ctx
);
654 struct gl_linked_shader
*linked_shader
[MESA_SHADER_STAGES
];
655 unsigned num_shaders
= 0;
657 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
658 if (shader_program
->_LinkedShaders
[i
])
659 linked_shader
[num_shaders
++] = shader_program
->_LinkedShaders
[i
];
662 for (unsigned i
= 0; i
< num_shaders
; i
++) {
663 struct gl_linked_shader
*shader
= linked_shader
[i
];
664 const nir_shader_compiler_options
*options
=
665 st
->ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
;
666 struct gl_program
*prog
= shader
->Program
;
667 struct st_program
*stp
= (struct st_program
*)prog
;
669 _mesa_copy_linked_program_data(shader_program
, shader
);
672 stp
->shader_program
= shader_program
;
673 stp
->state
.type
= PIPE_SHADER_IR_NIR
;
675 /* Parameters will be filled during NIR linking. */
676 prog
->Parameters
= _mesa_new_parameter_list();
678 if (shader_program
->data
->spirv
) {
679 prog
->nir
= _mesa_spirv_to_nir(ctx
, shader_program
, shader
->Stage
, options
);
681 validate_ir_tree(shader
->ir
);
683 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
685 _mesa_log("GLSL IR for linked %s program %d:\n",
686 _mesa_shader_stage_to_string(shader
->Stage
),
687 shader_program
->Name
);
688 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
692 prog
->nir
= glsl_to_nir(st
->ctx
, shader_program
, shader
->Stage
, options
);
693 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
696 if (options
->lower_to_scalar
) {
697 NIR_PASS_V(shader
->Program
->nir
, nir_lower_load_const_to_scalar
);
701 st_lower_patch_vertices_in(shader_program
);
703 /* For SPIR-V, we have to perform the NIR linking before applying
706 if (shader_program
->data
->spirv
) {
707 static const gl_nir_linker_options opts
= {
708 true /*fill_parameters */
710 if (!gl_nir_link_spirv(ctx
, shader_program
, &opts
))
713 nir_build_program_resource_list(ctx
, shader_program
, true);
715 for (unsigned i
= 0; i
< num_shaders
; i
++) {
716 struct gl_linked_shader
*shader
= linked_shader
[i
];
717 struct gl_program
*prog
= shader
->Program
;
719 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
720 _mesa_update_shader_textures_used(shader_program
, prog
);
721 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
725 /* Linking the stages in the opposite order (from fragment to vertex)
726 * ensures that inter-shader outputs written to in an earlier stage
727 * are eliminated if they are (transitively) not used in a later
730 for (int i
= num_shaders
- 2; i
>= 0; i
--) {
731 st_nir_link_shaders(linked_shader
[i
]->Program
->nir
,
732 linked_shader
[i
+ 1]->Program
->nir
);
734 /* Linking shaders also optimizes them. Separate shaders, compute shaders
735 * and shaders with a fixed-func VS or FS that don't need linking are
738 if (num_shaders
== 1)
739 st_nir_opts(linked_shader
[0]->Program
->nir
);
741 if (!shader_program
->data
->spirv
) {
742 if (!gl_nir_link_glsl(ctx
, shader_program
))
745 for (unsigned i
= 0; i
< num_shaders
; i
++) {
746 struct gl_program
*prog
= linked_shader
[i
]->Program
;
747 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
748 _mesa_update_shader_textures_used(shader_program
, prog
);
751 nir_build_program_resource_list(ctx
, shader_program
, false);
754 for (unsigned i
= 0; i
< num_shaders
; i
++) {
755 struct gl_linked_shader
*shader
= linked_shader
[i
];
756 nir_shader
*nir
= shader
->Program
->nir
;
758 /* This needs to run after the initial pass of nir_lower_vars_to_ssa, so
759 * that the buffer indices are constants in nir where they where
760 * constants in GLSL. */
761 NIR_PASS_V(nir
, gl_nir_lower_buffers
, shader_program
);
763 /* Remap the locations to slots so those requiring two slots will occupy
764 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
765 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
766 * locations/slots 0 and 1, and attr1 will use location/slot 2
768 if (nir
->info
.stage
== MESA_SHADER_VERTEX
&& !shader_program
->data
->spirv
)
769 nir_remap_dual_slot_attributes(nir
, &shader
->Program
->DualSlotInputs
);
771 NIR_PASS_V(nir
, st_nir_lower_wpos_ytransform
, shader
->Program
,
774 NIR_PASS_V(nir
, nir_lower_system_values
);
775 NIR_PASS_V(nir
, nir_lower_compute_system_values
, NULL
);
777 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
779 st_shader_gather_info(nir
, shader
->Program
);
780 if (shader
->Stage
== MESA_SHADER_VERTEX
) {
781 /* NIR expands dual-slot inputs out to two locations. We need to
782 * compact things back down GL-style single-slot inputs to avoid
783 * confusing the state tracker.
785 shader
->Program
->info
.inputs_read
=
786 nir_get_single_slot_attribs_mask(nir
->info
.inputs_read
,
787 shader
->Program
->DualSlotInputs
);
791 struct gl_program
*prev_shader
= linked_shader
[i
- 1]->Program
;
793 /* We can't use nir_compact_varyings with transform feedback, since
794 * the pipe_stream_output->output_register field is based on the
795 * pre-compacted driver_locations.
797 if (!(prev_shader
->sh
.LinkedTransformFeedback
&&
798 prev_shader
->sh
.LinkedTransformFeedback
->NumVarying
> 0))
799 nir_compact_varyings(prev_shader
->nir
,
800 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
802 if (ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
->vectorize_io
)
803 st_nir_vectorize_io(prev_shader
->nir
, nir
);
807 struct shader_info
*prev_info
= NULL
;
809 for (unsigned i
= 0; i
< num_shaders
; i
++) {
810 struct gl_linked_shader
*shader
= linked_shader
[i
];
811 struct shader_info
*info
= &shader
->Program
->nir
->info
;
814 ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
->unify_interfaces
) {
815 prev_info
->outputs_written
|= info
->inputs_read
&
816 ~(VARYING_BIT_TESS_LEVEL_INNER
| VARYING_BIT_TESS_LEVEL_OUTER
);
817 info
->inputs_read
|= prev_info
->outputs_written
&
818 ~(VARYING_BIT_TESS_LEVEL_INNER
| VARYING_BIT_TESS_LEVEL_OUTER
);
820 prev_info
->patch_outputs_written
|= info
->patch_inputs_read
;
821 info
->patch_inputs_read
|= prev_info
->patch_outputs_written
;
826 for (unsigned i
= 0; i
< num_shaders
; i
++) {
827 struct gl_linked_shader
*shader
= linked_shader
[i
];
828 struct gl_program
*prog
= shader
->Program
;
829 struct st_program
*stp
= st_program(prog
);
830 st_glsl_to_nir_post_opts(st
, prog
, shader_program
);
832 /* Initialize st_vertex_program members. */
833 if (shader
->Stage
== MESA_SHADER_VERTEX
)
834 st_prepare_vertex_program(stp
);
836 /* Get pipe_stream_output_info. */
837 if (shader
->Stage
== MESA_SHADER_VERTEX
||
838 shader
->Stage
== MESA_SHADER_TESS_EVAL
||
839 shader
->Stage
== MESA_SHADER_GEOMETRY
)
840 st_translate_stream_output_info(prog
);
842 st_store_ir_in_disk_cache(st
, prog
, true);
844 st_release_variants(st
, stp
);
845 st_finalize_program(st
, prog
);
847 /* The GLSL IR won't be needed anymore. */
848 ralloc_free(shader
->ir
);
856 st_nir_assign_varying_locations(struct st_context
*st
, nir_shader
*nir
)
858 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
859 nir_assign_io_var_locations(nir
, nir_var_shader_out
,
862 st_nir_fixup_varying_slots(st
, nir
, nir_var_shader_out
);
863 } else if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
864 nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
865 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
866 nir_assign_io_var_locations(nir
, nir_var_shader_in
,
869 st_nir_fixup_varying_slots(st
, nir
, nir_var_shader_in
);
871 nir_assign_io_var_locations(nir
, nir_var_shader_out
,
874 st_nir_fixup_varying_slots(st
, nir
, nir_var_shader_out
);
875 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
876 nir_assign_io_var_locations(nir
, nir_var_shader_in
,
879 st_nir_fixup_varying_slots(st
, nir
, nir_var_shader_in
);
880 nir_assign_io_var_locations(nir
, nir_var_shader_out
,
883 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
886 unreachable("invalid shader type");
891 st_nir_lower_samplers(struct pipe_screen
*screen
, nir_shader
*nir
,
892 struct gl_shader_program
*shader_program
,
893 struct gl_program
*prog
)
895 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
896 NIR_PASS_V(nir
, gl_nir_lower_samplers_as_deref
, shader_program
);
898 NIR_PASS_V(nir
, gl_nir_lower_samplers
, shader_program
);
901 prog
->info
.textures_used
= nir
->info
.textures_used
;
902 prog
->info
.textures_used_by_txf
= nir
->info
.textures_used_by_txf
;
903 prog
->info
.images_used
= nir
->info
.images_used
;
908 st_packed_uniforms_type_size(const struct glsl_type
*type
, bool bindless
)
910 return glsl_count_dword_slots(type
, bindless
);
914 st_unpacked_uniforms_type_size(const struct glsl_type
*type
, bool bindless
)
916 return glsl_count_vec4_slots(type
, false, bindless
);
920 st_nir_lower_uniforms(struct st_context
*st
, nir_shader
*nir
)
922 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
923 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
,
924 st_packed_uniforms_type_size
,
925 (nir_lower_io_options
)0);
926 NIR_PASS_V(nir
, nir_lower_uniforms_to_ubo
, 4);
928 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
,
929 st_unpacked_uniforms_type_size
,
930 (nir_lower_io_options
)0);
934 /* Last third of preparing nir from glsl, which happens after shader
938 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
939 struct gl_shader_program
*shader_program
,
940 nir_shader
*nir
, bool finalize_by_driver
)
942 struct pipe_screen
*screen
= st
->pipe
->screen
;
944 NIR_PASS_V(nir
, nir_split_var_copies
);
945 NIR_PASS_V(nir
, nir_lower_var_copies
);
947 st_nir_assign_varying_locations(st
, nir
);
948 st_nir_assign_uniform_locations(st
->ctx
, prog
, nir
);
950 /* Set num_uniforms in number of attribute slots (vec4s) */
951 nir
->num_uniforms
= DIV_ROUND_UP(prog
->Parameters
->NumParameterValues
, 4);
953 st_nir_lower_uniforms(st
, nir
);
954 st_nir_lower_samplers(screen
, nir
, shader_program
, prog
);
955 if (!screen
->get_param(screen
, PIPE_CAP_NIR_IMAGES_AS_DEREF
))
956 NIR_PASS_V(nir
, gl_nir_lower_images
, false);
958 if (finalize_by_driver
&& screen
->finalize_nir
)
959 screen
->finalize_nir(screen
, nir
, false);