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_glsl_types.h"
44 #include "st_program.h"
45 #include "st_shader_cache.h"
47 #include "compiler/nir/nir.h"
48 #include "compiler/glsl_types.h"
49 #include "compiler/glsl/glsl_to_nir.h"
50 #include "compiler/glsl/gl_nir.h"
51 #include "compiler/glsl/gl_nir_linker.h"
52 #include "compiler/glsl/ir.h"
53 #include "compiler/glsl/ir_optimization.h"
54 #include "compiler/glsl/string_to_uint_map.h"
57 type_size(const struct glsl_type
*type
)
59 return type
->count_attribute_slots(false);
62 /* Depending on PIPE_CAP_TGSI_TEXCOORD (st->needs_texcoord_semantic) we
63 * may need to fix up varying slots so the glsl->nir path is aligned
64 * with the anything->tgsi->nir path.
67 st_nir_fixup_varying_slots(struct st_context
*st
, struct exec_list
*var_list
)
69 if (st
->needs_texcoord_semantic
)
72 nir_foreach_variable(var
, var_list
) {
73 if (var
->data
.location
>= VARYING_SLOT_VAR0
) {
74 var
->data
.location
+= 9;
75 } else if ((var
->data
.location
>= VARYING_SLOT_TEX0
) &&
76 (var
->data
.location
<= VARYING_SLOT_TEX7
)) {
77 var
->data
.location
+= VARYING_SLOT_VAR0
- VARYING_SLOT_TEX0
;
82 /* input location assignment for VS inputs must be handled specially, so
83 * that it is aligned w/ st's vbo state.
84 * (This isn't the case with, for ex, FS inputs, which only need to agree
85 * on varying-slot w/ the VS outputs)
88 st_nir_assign_vs_in_locations(struct nir_shader
*nir
)
90 if (nir
->info
.stage
!= MESA_SHADER_VERTEX
)
93 bool removed_inputs
= false;
95 nir
->num_inputs
= util_bitcount64(nir
->info
.inputs_read
);
96 nir_foreach_variable_safe(var
, &nir
->inputs
) {
97 /* NIR already assigns dual-slot inputs to two locations so all we have
98 * to do is compact everything down.
100 if (nir
->info
.inputs_read
& BITFIELD64_BIT(var
->data
.location
)) {
101 var
->data
.driver_location
=
102 util_bitcount64(nir
->info
.inputs_read
&
103 BITFIELD64_MASK(var
->data
.location
));
105 /* Move unused input variables to the globals list (with no
106 * initialization), to avoid confusing drivers looking through the
107 * inputs array and expecting to find inputs with a driver_location
110 exec_node_remove(&var
->node
);
111 var
->data
.mode
= nir_var_shader_temp
;
112 exec_list_push_tail(&nir
->globals
, &var
->node
);
113 removed_inputs
= true;
117 /* Re-lower global vars, to deal with any dead VS inputs. */
119 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
123 st_nir_lookup_parameter_index(struct gl_program
*prog
, nir_variable
*var
)
125 struct gl_program_parameter_list
*params
= prog
->Parameters
;
127 /* Lookup the first parameter that the uniform storage that match the
130 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
131 int index
= params
->Parameters
[i
].MainUniformStorageIndex
;
132 if (index
== var
->data
.location
)
136 /* TODO: Handle this fallback for SPIR-V. We need this for GLSL e.g. in
137 * dEQP-GLES2.functional.uniform_api.random.3
140 /* is there a better way to do this? If we have something like:
148 * Then what we get in prog->Parameters looks like:
150 * 0: Name=color.f, Type=6, DataType=1406, Size=1
151 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
153 * So the name doesn't match up and _mesa_lookup_parameter_index()
154 * fails. In this case just find the first matching "color.*"..
156 * Note for arrays you could end up w/ color[n].f, for example.
158 * glsl_to_tgsi works slightly differently in this regard. It is
159 * emitting something more low level, so it just translates the
160 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
161 * it just calculates the additional offset of struct field members
162 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
163 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
164 * needs to work backwards to get base var loc from the param-list
165 * which already has them separated out.
167 if (!prog
->sh
.data
->spirv
) {
168 int namelen
= strlen(var
->name
);
169 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
170 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
171 if ((strncmp(p
->Name
, var
->name
, namelen
) == 0) &&
172 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
182 st_nir_assign_uniform_locations(struct gl_context
*ctx
,
183 struct gl_program
*prog
,
184 struct exec_list
*uniform_list
)
189 nir_foreach_variable(uniform
, uniform_list
) {
193 * UBO's have their own address spaces, so don't count them towards the
194 * number of global uniforms
196 if (uniform
->data
.mode
== nir_var_mem_ubo
|| uniform
->data
.mode
== nir_var_mem_ssbo
)
199 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
200 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
201 if (type
->is_sampler()) {
203 shaderidx
+= type_size(uniform
->type
);
206 imageidx
+= type_size(uniform
->type
);
208 } else if (uniform
->state_slots
) {
209 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
210 /* This state reference has already been setup by ir_to_mesa, but we'll
211 * get the same index back here.
215 if (glsl_type_is_struct_or_ifc(type
)) {
218 comps
= glsl_get_vector_elements(type
);
221 if (ctx
->Const
.PackedDriverUniformStorage
) {
222 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
223 stateTokens
, comps
, false);
224 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
226 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
229 loc
= st_nir_lookup_parameter_index(prog
, uniform
);
231 /* We need to check that loc is not -1 here before accessing the
232 * array. It can be negative for example when we have a struct that
233 * only contains opaque types.
235 if (loc
>= 0 && ctx
->Const
.PackedDriverUniformStorage
) {
236 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
240 uniform
->data
.driver_location
= loc
;
245 st_nir_opts(nir_shader
*nir
)
252 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
254 /* Linking deals with unused inputs/outputs, but here we can remove
255 * things local to the shader in the hopes that we can cleanup other
256 * things. This pass will also remove variables with only stores, so we
257 * might be able to make progress after it.
259 NIR_PASS(progress
, nir
, nir_remove_dead_variables
,
260 (nir_variable_mode
)(nir_var_function_temp
|
261 nir_var_shader_temp
|
262 nir_var_mem_shared
));
264 NIR_PASS(progress
, nir
, nir_opt_copy_prop_vars
);
265 NIR_PASS(progress
, nir
, nir_opt_dead_write_vars
);
267 if (nir
->options
->lower_to_scalar
) {
268 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
269 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
272 NIR_PASS_V(nir
, nir_lower_alu
);
273 NIR_PASS_V(nir
, nir_lower_pack
);
274 NIR_PASS(progress
, nir
, nir_copy_prop
);
275 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
276 NIR_PASS(progress
, nir
, nir_opt_dce
);
277 if (nir_opt_trivial_continues(nir
)) {
279 NIR_PASS(progress
, nir
, nir_copy_prop
);
280 NIR_PASS(progress
, nir
, nir_opt_dce
);
282 NIR_PASS(progress
, nir
, nir_opt_if
, false);
283 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
284 NIR_PASS(progress
, nir
, nir_opt_cse
);
285 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
287 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
288 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
290 if (!nir
->info
.flrp_lowered
) {
291 unsigned lower_flrp
=
292 (nir
->options
->lower_flrp16
? 16 : 0) |
293 (nir
->options
->lower_flrp32
? 32 : 0) |
294 (nir
->options
->lower_flrp64
? 64 : 0);
297 bool lower_flrp_progress
= false;
299 NIR_PASS(lower_flrp_progress
, nir
, nir_lower_flrp
,
301 false /* always_precise */,
302 nir
->options
->lower_ffma
);
303 if (lower_flrp_progress
) {
304 NIR_PASS(progress
, nir
,
305 nir_opt_constant_folding
);
310 /* Nothing should rematerialize any flrps, so we only need to do this
313 nir
->info
.flrp_lowered
= true;
316 NIR_PASS(progress
, nir
, nir_opt_undef
);
317 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
318 if (nir
->options
->max_unroll_iterations
) {
319 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
325 shared_type_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
327 assert(glsl_type_is_vector_or_scalar(type
));
329 uint32_t comp_size
= glsl_type_is_boolean(type
)
330 ? 4 : glsl_get_bit_size(type
) / 8;
331 unsigned length
= glsl_get_vector_elements(type
);
332 *size
= comp_size
* length
,
333 *align
= comp_size
* (length
== 3 ? 4 : length
);
336 /* First third of converting glsl_to_nir.. this leaves things in a pre-
337 * nir_lower_io state, so that shader variants can more easily insert/
338 * replace variables, etc.
341 st_nir_preprocess(struct st_context
*st
, struct gl_program
*prog
,
342 struct gl_shader_program
*shader_program
,
343 gl_shader_stage stage
)
345 struct pipe_screen
*screen
= st
->pipe
->screen
;
346 const nir_shader_compiler_options
*options
=
347 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
349 nir_shader
*nir
= prog
->nir
;
351 /* Set the next shader stage hint for VS and TES. */
352 if (!nir
->info
.separate_shader
&&
353 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
354 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
356 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
357 unsigned stages_mask
=
358 ~prev_stages
& shader_program
->data
->linked_stages
;
360 nir
->info
.next_stage
= stages_mask
?
361 (gl_shader_stage
) u_bit_scan(&stages_mask
) : MESA_SHADER_FRAGMENT
;
363 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
366 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
367 if (!st
->ctx
->SoftFP64
&& nir
->info
.uses_64bit
&&
368 (options
->lower_doubles_options
& nir_lower_fp64_full_software
) != 0) {
369 st
->ctx
->SoftFP64
= glsl_float64_funcs_to_nir(st
->ctx
, options
);
372 /* ES has strict SSO validation rules for shader IO matching so we can't
373 * remove dead IO until the resource list has been built. Here we skip
374 * removing them until later. This will potentially make the IO lowering
375 * calls below do a little extra work but should otherwise have no impact.
377 if (!_mesa_is_gles(st
->ctx
) || !nir
->info
.separate_shader
) {
378 nir_variable_mode mask
=
379 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
380 nir_remove_dead_variables(nir
, mask
);
383 if (options
->lower_all_io_to_temps
||
384 nir
->info
.stage
== MESA_SHADER_VERTEX
||
385 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
386 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
387 nir_shader_get_entrypoint(nir
),
389 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
||
390 !screen
->get_param(screen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
)) {
391 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
392 nir_shader_get_entrypoint(nir
),
396 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
397 NIR_PASS_V(nir
, nir_split_var_copies
);
398 NIR_PASS_V(nir
, nir_lower_var_copies
);
400 if (options
->lower_to_scalar
) {
401 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
404 /* before buffers and vars_to_ssa */
405 NIR_PASS_V(nir
, gl_nir_lower_bindless_images
);
407 /* TODO: Change GLSL to not lower shared memory. */
408 if (prog
->nir
->info
.stage
== MESA_SHADER_COMPUTE
&&
409 shader_program
->data
->spirv
) {
410 NIR_PASS_V(prog
->nir
, nir_lower_vars_to_explicit_types
,
411 nir_var_mem_shared
, shared_type_info
);
412 NIR_PASS_V(prog
->nir
, nir_lower_explicit_io
,
413 nir_var_mem_shared
, nir_address_format_32bit_offset
);
416 /* Do a round of constant folding to clean up address calculations */
417 NIR_PASS_V(nir
, nir_opt_constant_folding
);
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 /* Lower 64-bit ops. */
487 if (nir
->options
->lower_int64_options
||
488 nir
->options
->lower_doubles_options
) {
489 bool lowered_64bit_ops
= false;
490 if (nir
->options
->lower_doubles_options
) {
491 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_doubles
,
492 st
->ctx
->SoftFP64
, nir
->options
->lower_doubles_options
);
494 if (nir
->options
->lower_int64_options
) {
495 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_int64
,
496 nir
->options
->lower_int64_options
);
499 if (lowered_64bit_ops
)
503 nir_variable_mode mask
= (nir_variable_mode
)
504 (nir_var_shader_in
| nir_var_shader_out
| nir_var_function_temp
);
505 nir_remove_dead_variables(nir
, mask
);
507 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
,
508 st
->ctx
->Const
.Program
[nir
->info
.stage
].MaxAtomicBuffers
);
510 st_finalize_nir_before_variants(nir
);
512 if (st
->allow_st_finalize_nir_twice
)
513 st_finalize_nir(st
, prog
, shader_program
, nir
, true);
515 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
517 _mesa_log("NIR IR for linked %s program %d:\n",
518 _mesa_shader_stage_to_string(prog
->info
.stage
),
519 shader_program
->Name
);
520 nir_print_shader(nir
, _mesa_get_log_file());
526 st_nir_vectorize_io(nir_shader
*producer
, nir_shader
*consumer
)
528 NIR_PASS_V(producer
, nir_lower_io_to_vector
, nir_var_shader_out
);
529 NIR_PASS_V(producer
, nir_opt_combine_stores
, nir_var_shader_out
);
530 NIR_PASS_V(consumer
, nir_lower_io_to_vector
, nir_var_shader_in
);
532 if ((producer
)->info
.stage
!= MESA_SHADER_TESS_CTRL
) {
533 /* Calling lower_io_to_vector creates output variable writes with
534 * write-masks. We only support these for TCS outputs, so for other
535 * stages, we need to call nir_lower_io_to_temporaries to get rid of
536 * them. This, in turn, creates temporary variables and extra
537 * copy_deref intrinsics that we need to clean up.
539 NIR_PASS_V(producer
, nir_lower_io_to_temporaries
,
540 nir_shader_get_entrypoint(producer
), true, false);
541 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
542 NIR_PASS_V(producer
, nir_split_var_copies
);
543 NIR_PASS_V(producer
, nir_lower_var_copies
);
548 st_nir_link_shaders(nir_shader
*producer
, nir_shader
*consumer
)
550 if (producer
->options
->lower_to_scalar
) {
551 NIR_PASS_V(producer
, nir_lower_io_to_scalar_early
, nir_var_shader_out
);
552 NIR_PASS_V(consumer
, nir_lower_io_to_scalar_early
, nir_var_shader_in
);
555 nir_lower_io_arrays_to_elements(producer
, consumer
);
557 st_nir_opts(producer
);
558 st_nir_opts(consumer
);
560 if (nir_link_opt_varyings(producer
, consumer
))
561 st_nir_opts(consumer
);
563 NIR_PASS_V(producer
, nir_remove_dead_variables
, nir_var_shader_out
);
564 NIR_PASS_V(consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
566 if (nir_remove_unused_varyings(producer
, consumer
)) {
567 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
568 NIR_PASS_V(consumer
, nir_lower_global_vars_to_local
);
570 st_nir_opts(producer
);
571 st_nir_opts(consumer
);
573 /* Optimizations can cause varyings to become unused.
574 * nir_compact_varyings() depends on all dead varyings being removed so
575 * we need to call nir_remove_dead_variables() again here.
577 NIR_PASS_V(producer
, nir_remove_dead_variables
, nir_var_shader_out
);
578 NIR_PASS_V(consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
583 st_lower_patch_vertices_in(struct gl_shader_program
*shader_prog
)
585 struct gl_linked_shader
*linked_tcs
=
586 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_CTRL
];
587 struct gl_linked_shader
*linked_tes
=
588 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_EVAL
];
590 /* If we have a TCS and TES linked together, lower TES patch vertices. */
591 if (linked_tcs
&& linked_tes
) {
592 nir_shader
*tcs_nir
= linked_tcs
->Program
->nir
;
593 nir_shader
*tes_nir
= linked_tes
->Program
->nir
;
595 /* The TES input vertex count is the TCS output vertex count,
596 * lower TES gl_PatchVerticesIn to a constant.
598 uint32_t tes_patch_verts
= tcs_nir
->info
.tess
.tcs_vertices_out
;
599 NIR_PASS_V(tes_nir
, nir_lower_patch_vertices
, tes_patch_verts
, NULL
);
606 st_nir_lower_wpos_ytransform(struct nir_shader
*nir
,
607 struct gl_program
*prog
,
608 struct pipe_screen
*pscreen
)
610 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
613 static const gl_state_index16 wposTransformState
[STATE_LENGTH
] = {
614 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
616 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
618 memcpy(wpos_options
.state_tokens
, wposTransformState
,
619 sizeof(wpos_options
.state_tokens
));
620 wpos_options
.fs_coord_origin_upper_left
=
621 pscreen
->get_param(pscreen
,
622 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
623 wpos_options
.fs_coord_origin_lower_left
=
624 pscreen
->get_param(pscreen
,
625 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
626 wpos_options
.fs_coord_pixel_center_integer
=
627 pscreen
->get_param(pscreen
,
628 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
629 wpos_options
.fs_coord_pixel_center_half_integer
=
630 pscreen
->get_param(pscreen
,
631 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
633 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
634 nir_validate_shader(nir
, "after nir_lower_wpos_ytransform");
635 _mesa_add_state_reference(prog
->Parameters
, wposTransformState
);
640 st_link_nir(struct gl_context
*ctx
,
641 struct gl_shader_program
*shader_program
)
643 struct st_context
*st
= st_context(ctx
);
644 struct gl_linked_shader
*linked_shader
[MESA_SHADER_STAGES
];
645 unsigned num_shaders
= 0;
647 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
648 if (shader_program
->_LinkedShaders
[i
])
649 linked_shader
[num_shaders
++] = shader_program
->_LinkedShaders
[i
];
652 for (unsigned i
= 0; i
< num_shaders
; i
++) {
653 struct gl_linked_shader
*shader
= linked_shader
[i
];
654 const nir_shader_compiler_options
*options
=
655 st
->ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
;
656 struct gl_program
*prog
= shader
->Program
;
657 struct st_program
*stp
= (struct st_program
*)prog
;
659 _mesa_copy_linked_program_data(shader_program
, shader
);
662 stp
->shader_program
= shader_program
;
663 stp
->state
.type
= PIPE_SHADER_IR_NIR
;
665 if (shader_program
->data
->spirv
) {
666 prog
->Parameters
= _mesa_new_parameter_list();
667 /* Parameters will be filled during NIR linking. */
669 prog
->nir
= _mesa_spirv_to_nir(ctx
, shader_program
, shader
->Stage
, options
);
671 validate_ir_tree(shader
->ir
);
673 prog
->Parameters
= _mesa_new_parameter_list();
674 _mesa_generate_parameters_list_for_uniforms(ctx
, shader_program
, shader
,
677 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
679 _mesa_log("GLSL IR for linked %s program %d:\n",
680 _mesa_shader_stage_to_string(shader
->Stage
),
681 shader_program
->Name
);
682 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
686 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
687 _mesa_update_shader_textures_used(shader_program
, prog
);
689 prog
->nir
= glsl_to_nir(st
->ctx
, shader_program
, shader
->Stage
, options
);
690 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
693 if (options
->lower_to_scalar
) {
694 NIR_PASS_V(shader
->Program
->nir
, nir_lower_load_const_to_scalar
);
698 st_lower_patch_vertices_in(shader_program
);
700 /* For SPIR-V, we have to perform the NIR linking before applying
703 if (shader_program
->data
->spirv
) {
704 static const gl_nir_linker_options opts
= {
705 true /*fill_parameters */
707 if (!gl_nir_link(ctx
, shader_program
, &opts
))
710 nir_build_program_resource_list(ctx
, shader_program
, true);
712 for (unsigned i
= 0; i
< num_shaders
; i
++) {
713 struct gl_linked_shader
*shader
= linked_shader
[i
];
714 struct gl_program
*prog
= shader
->Program
;
716 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
717 _mesa_update_shader_textures_used(shader_program
, prog
);
718 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
722 /* Linking the stages in the opposite order (from fragment to vertex)
723 * ensures that inter-shader outputs written to in an earlier stage
724 * are eliminated if they are (transitively) not used in a later
727 for (int i
= num_shaders
- 2; i
>= 0; i
--) {
728 st_nir_link_shaders(linked_shader
[i
]->Program
->nir
,
729 linked_shader
[i
+ 1]->Program
->nir
);
732 if (!shader_program
->data
->spirv
)
733 nir_build_program_resource_list(ctx
, shader_program
, false);
735 for (unsigned i
= 0; i
< num_shaders
; i
++) {
736 struct gl_linked_shader
*shader
= linked_shader
[i
];
737 nir_shader
*nir
= shader
->Program
->nir
;
739 NIR_PASS_V(nir
, gl_nir_lower_buffers
, shader_program
);
741 /* Linked shaders are optimized in st_nir_link_shaders. Separate shaders
742 * and shaders with a fixed-func VS or FS are optimized here.
744 if (num_shaders
== 1)
747 /* Remap the locations to slots so those requiring two slots will occupy
748 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
749 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
750 * locations/slots 0 and 1, and attr1 will use location/slot 2
752 if (nir
->info
.stage
== MESA_SHADER_VERTEX
&& !shader_program
->data
->spirv
)
753 nir_remap_dual_slot_attributes(nir
, &shader
->Program
->DualSlotInputs
);
755 NIR_PASS_V(nir
, st_nir_lower_wpos_ytransform
, shader
->Program
,
758 NIR_PASS_V(nir
, nir_lower_system_values
);
759 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
761 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
762 shader
->Program
->info
= nir
->info
;
763 if (shader
->Stage
== MESA_SHADER_VERTEX
) {
764 /* NIR expands dual-slot inputs out to two locations. We need to
765 * compact things back down GL-style single-slot inputs to avoid
766 * confusing the state tracker.
768 shader
->Program
->info
.inputs_read
=
769 nir_get_single_slot_attribs_mask(nir
->info
.inputs_read
,
770 shader
->Program
->DualSlotInputs
);
774 struct gl_program
*prev_shader
= linked_shader
[i
- 1]->Program
;
776 /* We can't use nir_compact_varyings with transform feedback, since
777 * the pipe_stream_output->output_register field is based on the
778 * pre-compacted driver_locations.
780 if (!(prev_shader
->sh
.LinkedTransformFeedback
&&
781 prev_shader
->sh
.LinkedTransformFeedback
->NumVarying
> 0))
782 nir_compact_varyings(prev_shader
->nir
,
783 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
785 if (ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
->vectorize_io
)
786 st_nir_vectorize_io(prev_shader
->nir
, nir
);
790 for (unsigned i
= 0; i
< num_shaders
; i
++) {
791 struct gl_linked_shader
*shader
= linked_shader
[i
];
792 struct gl_program
*prog
= shader
->Program
;
793 struct st_program
*stp
= st_program(prog
);
794 st_glsl_to_nir_post_opts(st
, prog
, shader_program
);
796 /* Initialize st_vertex_program members. */
797 if (shader
->Stage
== MESA_SHADER_VERTEX
)
798 st_prepare_vertex_program(stp
);
800 /* Get pipe_stream_output_info. */
801 if (shader
->Stage
== MESA_SHADER_VERTEX
||
802 shader
->Stage
== MESA_SHADER_TESS_EVAL
||
803 shader
->Stage
== MESA_SHADER_GEOMETRY
)
804 st_translate_stream_output_info(prog
);
806 st_store_ir_in_disk_cache(st
, prog
, true);
808 st_release_variants(st
, stp
);
809 st_finalize_program(st
, prog
);
811 /* The GLSL IR won't be needed anymore. */
812 ralloc_free(shader
->ir
);
816 struct shader_info
*prev_info
= NULL
;
818 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
819 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
823 struct shader_info
*info
= &shader
->Program
->nir
->info
;
826 ctx
->Const
.ShaderCompilerOptions
[i
].NirOptions
->unify_interfaces
) {
827 prev_info
->outputs_written
|= info
->inputs_read
&
828 ~(VARYING_BIT_TESS_LEVEL_INNER
| VARYING_BIT_TESS_LEVEL_OUTER
);
829 info
->inputs_read
|= prev_info
->outputs_written
&
830 ~(VARYING_BIT_TESS_LEVEL_INNER
| VARYING_BIT_TESS_LEVEL_OUTER
);
832 prev_info
->patch_outputs_written
|= info
->patch_inputs_read
;
833 info
->patch_inputs_read
|= prev_info
->patch_outputs_written
;
842 st_nir_assign_varying_locations(struct st_context
*st
, nir_shader
*nir
)
844 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
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
,
898 nir_shader
*nir
, bool finalize_by_driver
)
900 struct pipe_screen
*screen
= st
->pipe
->screen
;
902 NIR_PASS_V(nir
, nir_split_var_copies
);
903 NIR_PASS_V(nir
, nir_lower_var_copies
);
905 st_nir_assign_varying_locations(st
, nir
);
906 st_nir_assign_uniform_locations(st
->ctx
, prog
,
909 /* Set num_uniforms in number of attribute slots (vec4s) */
910 nir
->num_uniforms
= DIV_ROUND_UP(prog
->Parameters
->NumParameterValues
, 4);
912 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
913 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_type_dword_size
,
914 (nir_lower_io_options
)0);
915 NIR_PASS_V(nir
, nir_lower_uniforms_to_ubo
, 4);
917 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_uniforms_type_size
,
918 (nir_lower_io_options
)0);
921 st_nir_lower_samplers(screen
, nir
, shader_program
, prog
);
923 if (finalize_by_driver
&& screen
->finalize_nir
)
924 screen
->finalize_nir(screen
, nir
, false);