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 bool removed_inputs
= false;
91 nir
->num_inputs
= util_bitcount64(nir
->info
.inputs_read
);
92 nir_foreach_variable_safe(var
, &nir
->inputs
) {
93 /* NIR already assigns dual-slot inputs to two locations so all we have
94 * to do is compact everything down.
96 if (var
->data
.location
== VERT_ATTRIB_EDGEFLAG
) {
97 /* bit of a hack, mirroring st_translate_vertex_program */
98 var
->data
.driver_location
= nir
->num_inputs
++;
99 } else if (nir
->info
.inputs_read
& BITFIELD64_BIT(var
->data
.location
)) {
100 var
->data
.driver_location
=
101 util_bitcount64(nir
->info
.inputs_read
&
102 BITFIELD64_MASK(var
->data
.location
));
104 /* Move unused input variables to the globals list (with no
105 * initialization), to avoid confusing drivers looking through the
106 * inputs array and expecting to find inputs with a driver_location
109 exec_node_remove(&var
->node
);
110 var
->data
.mode
= nir_var_shader_temp
;
111 exec_list_push_tail(&nir
->globals
, &var
->node
);
112 removed_inputs
= true;
116 /* Re-lower global vars, to deal with any dead VS inputs. */
118 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
122 st_nir_lookup_parameter_index(struct gl_program
*prog
, nir_variable
*var
)
124 struct gl_program_parameter_list
*params
= prog
->Parameters
;
126 /* Lookup the first parameter that the uniform storage that match the
129 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
130 int index
= params
->Parameters
[i
].MainUniformStorageIndex
;
131 if (index
== var
->data
.location
)
135 /* TODO: Handle this fallback for SPIR-V. We need this for GLSL e.g. in
136 * dEQP-GLES2.functional.uniform_api.random.3
139 /* is there a better way to do this? If we have something like:
147 * Then what we get in prog->Parameters looks like:
149 * 0: Name=color.f, Type=6, DataType=1406, Size=1
150 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
152 * So the name doesn't match up and _mesa_lookup_parameter_index()
153 * fails. In this case just find the first matching "color.*"..
155 * Note for arrays you could end up w/ color[n].f, for example.
157 * glsl_to_tgsi works slightly differently in this regard. It is
158 * emitting something more low level, so it just translates the
159 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
160 * it just calculates the additional offset of struct field members
161 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
162 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
163 * needs to work backwards to get base var loc from the param-list
164 * which already has them separated out.
166 if (!prog
->sh
.data
->spirv
) {
167 int namelen
= strlen(var
->name
);
168 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
169 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
170 if ((strncmp(p
->Name
, var
->name
, namelen
) == 0) &&
171 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
181 st_nir_assign_uniform_locations(struct gl_context
*ctx
,
182 struct gl_program
*prog
,
183 struct exec_list
*uniform_list
)
188 nir_foreach_variable(uniform
, uniform_list
) {
192 * UBO's have their own address spaces, so don't count them towards the
193 * number of global uniforms
195 if (uniform
->data
.mode
== nir_var_mem_ubo
|| uniform
->data
.mode
== nir_var_mem_ssbo
)
198 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
199 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
200 if (type
->is_sampler()) {
202 shaderidx
+= type_size(uniform
->type
);
205 imageidx
+= type_size(uniform
->type
);
207 } else if (uniform
->state_slots
) {
208 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
209 /* This state reference has already been setup by ir_to_mesa, but we'll
210 * get the same index back here.
214 if (glsl_type_is_struct_or_ifc(type
)) {
217 comps
= glsl_get_vector_elements(type
);
220 if (ctx
->Const
.PackedDriverUniformStorage
) {
221 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
222 stateTokens
, comps
, false);
223 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
225 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
228 loc
= st_nir_lookup_parameter_index(prog
, uniform
);
230 /* We need to check that loc is not -1 here before accessing the
231 * array. It can be negative for example when we have a struct that
232 * only contains opaque types.
234 if (loc
>= 0 && ctx
->Const
.PackedDriverUniformStorage
) {
235 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
239 uniform
->data
.driver_location
= loc
;
244 st_nir_opts(nir_shader
*nir
)
247 unsigned lower_flrp
=
248 (nir
->options
->lower_flrp16
? 16 : 0) |
249 (nir
->options
->lower_flrp32
? 32 : 0) |
250 (nir
->options
->lower_flrp64
? 64 : 0);
255 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
257 /* Linking deals with unused inputs/outputs, but here we can remove
258 * things local to the shader in the hopes that we can cleanup other
259 * things. This pass will also remove variables with only stores, so we
260 * might be able to make progress after it.
262 NIR_PASS(progress
, nir
, nir_remove_dead_variables
,
263 (nir_variable_mode
)(nir_var_function_temp
|
264 nir_var_shader_temp
|
265 nir_var_mem_shared
));
267 NIR_PASS(progress
, nir
, nir_opt_copy_prop_vars
);
268 NIR_PASS(progress
, nir
, nir_opt_dead_write_vars
);
270 if (nir
->options
->lower_to_scalar
) {
271 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
272 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
275 NIR_PASS_V(nir
, nir_lower_alu
);
276 NIR_PASS_V(nir
, nir_lower_pack
);
277 NIR_PASS(progress
, nir
, nir_copy_prop
);
278 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
279 NIR_PASS(progress
, nir
, nir_opt_dce
);
280 if (nir_opt_trivial_continues(nir
)) {
282 NIR_PASS(progress
, nir
, nir_copy_prop
);
283 NIR_PASS(progress
, nir
, nir_opt_dce
);
285 NIR_PASS(progress
, nir
, nir_opt_if
, false);
286 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
287 NIR_PASS(progress
, nir
, nir_opt_cse
);
288 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
290 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
291 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
293 if (lower_flrp
!= 0) {
294 bool lower_flrp_progress
= false;
296 NIR_PASS(lower_flrp_progress
, nir
, nir_lower_flrp
,
298 false /* always_precise */,
299 nir
->options
->lower_ffma
);
300 if (lower_flrp_progress
) {
301 NIR_PASS(progress
, nir
,
302 nir_opt_constant_folding
);
306 /* Nothing should rematerialize any flrps, so we only need to do this
312 NIR_PASS(progress
, nir
, nir_opt_access
);
314 NIR_PASS(progress
, nir
, nir_opt_undef
);
315 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
316 if (nir
->options
->max_unroll_iterations
) {
317 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
323 shared_type_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
325 assert(glsl_type_is_vector_or_scalar(type
));
327 uint32_t comp_size
= glsl_type_is_boolean(type
)
328 ? 4 : glsl_get_bit_size(type
) / 8;
329 unsigned length
= glsl_get_vector_elements(type
);
330 *size
= comp_size
* length
,
331 *align
= comp_size
* (length
== 3 ? 4 : length
);
334 /* First third of converting glsl_to_nir.. this leaves things in a pre-
335 * nir_lower_io state, so that shader variants can more easily insert/
336 * replace variables, etc.
339 st_nir_preprocess(struct st_context
*st
, struct gl_program
*prog
,
340 struct gl_shader_program
*shader_program
,
341 gl_shader_stage stage
)
343 const nir_shader_compiler_options
*options
=
344 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
346 nir_shader
*nir
= prog
->nir
;
348 /* Set the next shader stage hint for VS and TES. */
349 if (!nir
->info
.separate_shader
&&
350 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
351 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
353 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
354 unsigned stages_mask
=
355 ~prev_stages
& shader_program
->data
->linked_stages
;
357 nir
->info
.next_stage
= stages_mask
?
358 (gl_shader_stage
) u_bit_scan(&stages_mask
) : MESA_SHADER_FRAGMENT
;
360 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
363 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
364 if (!st
->ctx
->SoftFP64
&& nir
->info
.uses_64bit
&&
365 (options
->lower_doubles_options
& nir_lower_fp64_full_software
) != 0) {
366 st
->ctx
->SoftFP64
= glsl_float64_funcs_to_nir(st
->ctx
, options
);
369 nir_variable_mode mask
=
370 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
371 nir_remove_dead_variables(nir
, mask
);
373 if (options
->lower_all_io_to_temps
||
374 nir
->info
.stage
== MESA_SHADER_VERTEX
||
375 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
376 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
377 nir_shader_get_entrypoint(nir
),
379 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
380 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
381 nir_shader_get_entrypoint(nir
),
385 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
386 NIR_PASS_V(nir
, nir_split_var_copies
);
387 NIR_PASS_V(nir
, nir_lower_var_copies
);
389 if (options
->lower_to_scalar
) {
390 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
393 /* before buffers and vars_to_ssa */
394 NIR_PASS_V(nir
, gl_nir_lower_bindless_images
);
396 /* TODO: Change GLSL to not lower shared memory. */
397 if (prog
->nir
->info
.stage
== MESA_SHADER_COMPUTE
&&
398 shader_program
->data
->spirv
) {
399 NIR_PASS_V(prog
->nir
, nir_lower_vars_to_explicit_types
,
400 nir_var_mem_shared
, shared_type_info
);
401 NIR_PASS_V(prog
->nir
, nir_lower_explicit_io
,
402 nir_var_mem_shared
, nir_address_format_32bit_offset
);
405 NIR_PASS_V(nir
, gl_nir_lower_buffers
, shader_program
);
406 /* Do a round of constant folding to clean up address calculations */
407 NIR_PASS_V(nir
, nir_opt_constant_folding
);
410 /* Second third of converting glsl_to_nir. This creates uniforms, gathers
411 * info on varyings, etc after NIR link time opts have been applied.
414 st_glsl_to_nir_post_opts(struct st_context
*st
, struct gl_program
*prog
,
415 struct gl_shader_program
*shader_program
)
417 nir_shader
*nir
= prog
->nir
;
419 /* Make a pass over the IR to add state references for any built-in
420 * uniforms that are used. This has to be done now (during linking).
421 * Code generation doesn't happen until the first time this shader is
422 * used for rendering. Waiting until then to generate the parameters is
423 * too late. At that point, the values for the built-in uniforms won't
424 * get sent to the shader.
426 nir_foreach_variable(var
, &nir
->uniforms
) {
427 const nir_state_slot
*const slots
= var
->state_slots
;
429 const struct glsl_type
*type
= glsl_without_array(var
->type
);
430 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
432 if (glsl_type_is_struct_or_ifc(type
)) {
433 /* Builtin struct require specical handling for now we just
434 * make all members vec4. See st_nir_lower_builtin.
438 comps
= glsl_get_vector_elements(type
);
441 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
442 _mesa_add_sized_state_reference(prog
->Parameters
,
446 _mesa_add_state_reference(prog
->Parameters
,
453 /* Avoid reallocation of the program parameter list, because the uniform
454 * storage is only associated with the original parameter list.
455 * This should be enough for Bitmap and DrawPixels constants.
457 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
459 /* This has to be done last. Any operation the can cause
460 * prog->ParameterValues to get reallocated (e.g., anything that adds a
461 * program constant) has to happen before creating this linkage.
463 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
);
465 st_set_prog_affected_state_flags(prog
);
467 /* None of the builtins being lowered here can be produced by SPIR-V. See
468 * _mesa_builtin_uniform_desc.
470 if (!shader_program
->data
->spirv
)
471 NIR_PASS_V(nir
, st_nir_lower_builtin
);
473 NIR_PASS_V(nir
, gl_nir_lower_atomics
, shader_program
, true);
474 NIR_PASS_V(nir
, nir_opt_intrinsics
);
476 /* Lower 64-bit ops. */
477 if (nir
->options
->lower_int64_options
||
478 nir
->options
->lower_doubles_options
) {
479 bool lowered_64bit_ops
= false;
480 if (nir
->options
->lower_doubles_options
) {
481 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_doubles
,
482 st
->ctx
->SoftFP64
, nir
->options
->lower_doubles_options
);
484 if (nir
->options
->lower_int64_options
) {
485 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_int64
,
486 nir
->options
->lower_int64_options
);
489 if (lowered_64bit_ops
)
493 nir_variable_mode mask
= nir_var_function_temp
;
494 nir_remove_dead_variables(nir
, mask
);
496 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
,
497 st
->ctx
->Const
.Program
[nir
->info
.stage
].MaxAtomicBuffers
);
499 st_finalize_nir_before_variants(nir
);
501 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
503 _mesa_log("NIR IR for linked %s program %d:\n",
504 _mesa_shader_stage_to_string(prog
->info
.stage
),
505 shader_program
->Name
);
506 nir_print_shader(nir
, _mesa_get_log_file());
512 set_st_program(struct gl_program
*prog
,
513 struct gl_shader_program
*shader_program
,
516 struct st_vertex_program
*stvp
;
517 struct st_common_program
*stp
;
519 switch (prog
->info
.stage
) {
520 case MESA_SHADER_VERTEX
:
521 stvp
= (struct st_vertex_program
*)prog
;
522 stvp
->shader_program
= shader_program
;
523 stvp
->state
.type
= PIPE_SHADER_IR_NIR
;
524 stvp
->state
.ir
.nir
= nir
;
526 case MESA_SHADER_GEOMETRY
:
527 case MESA_SHADER_TESS_CTRL
:
528 case MESA_SHADER_TESS_EVAL
:
529 case MESA_SHADER_COMPUTE
:
530 case MESA_SHADER_FRAGMENT
:
531 stp
= (struct st_common_program
*)prog
;
532 stp
->shader_program
= shader_program
;
533 stp
->state
.type
= PIPE_SHADER_IR_NIR
;
534 stp
->state
.ir
.nir
= 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
;
661 unsigned num_linked_shaders
= 0;
663 unsigned last_stage
= 0;
664 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
665 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
669 num_linked_shaders
++;
671 const nir_shader_compiler_options
*options
=
672 st
->ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
;
673 struct gl_program
*prog
= shader
->Program
;
674 _mesa_copy_linked_program_data(shader_program
, shader
);
678 if (shader_program
->data
->spirv
) {
679 prog
->Parameters
= _mesa_new_parameter_list();
680 /* Parameters will be filled during NIR linking. */
682 prog
->nir
= _mesa_spirv_to_nir(ctx
, shader_program
, shader
->Stage
, options
);
683 set_st_program(prog
, shader_program
, prog
->nir
);
685 validate_ir_tree(shader
->ir
);
687 prog
->Parameters
= _mesa_new_parameter_list();
688 _mesa_generate_parameters_list_for_uniforms(ctx
, shader_program
, shader
,
691 /* Remove reads from output registers. */
692 if (!screen
->get_param(screen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
))
693 lower_output_reads(shader
->Stage
, shader
->ir
);
695 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
697 _mesa_log("GLSL IR for linked %s program %d:\n",
698 _mesa_shader_stage_to_string(shader
->Stage
),
699 shader_program
->Name
);
700 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
704 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
705 _mesa_update_shader_textures_used(shader_program
, prog
);
707 prog
->nir
= glsl_to_nir(st
->ctx
, shader_program
, shader
->Stage
, options
);
708 set_st_program(prog
, shader_program
, prog
->nir
);
709 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
714 if (options
->lower_to_scalar
) {
715 NIR_PASS_V(shader
->Program
->nir
, nir_lower_load_const_to_scalar
);
719 /* For SPIR-V, we have to perform the NIR linking before applying
722 if (shader_program
->data
->spirv
) {
723 static const gl_nir_linker_options opts
= {
724 true /*fill_parameters */
726 if (!gl_nir_link(ctx
, shader_program
, &opts
))
729 nir_build_program_resource_list(ctx
, shader_program
);
731 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
732 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
736 struct gl_program
*prog
= shader
->Program
;
737 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
738 _mesa_update_shader_textures_used(shader_program
, prog
);
740 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
744 /* Linking the stages in the opposite order (from fragment to vertex)
745 * ensures that inter-shader outputs written to in an earlier stage
746 * are eliminated if they are (transitively) not used in a later
749 int next
= last_stage
;
750 for (int i
= next
- 1; i
>= 0; i
--) {
751 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
755 st_nir_link_shaders(&shader
->Program
->nir
,
756 &shader_program
->_LinkedShaders
[next
]->Program
->nir
);
761 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
762 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
766 nir_shader
*nir
= shader
->Program
->nir
;
768 /* Linked shaders are optimized in st_nir_link_shaders. Separate shaders
769 * and shaders with a fixed-func VS or FS are optimized here.
771 if (num_linked_shaders
== 1)
774 NIR_PASS_V(nir
, st_nir_lower_wpos_ytransform
, shader
->Program
,
777 NIR_PASS_V(nir
, nir_lower_system_values
);
778 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
780 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
781 shader
->Program
->info
= nir
->info
;
782 if (i
== MESA_SHADER_VERTEX
) {
783 /* NIR expands dual-slot inputs out to two locations. We need to
784 * compact things back down GL-style single-slot inputs to avoid
785 * confusing the state tracker.
787 shader
->Program
->info
.inputs_read
=
788 nir_get_single_slot_attribs_mask(nir
->info
.inputs_read
,
789 shader
->Program
->DualSlotInputs
);
793 struct gl_program
*prev_shader
=
794 shader_program
->_LinkedShaders
[prev
]->Program
;
796 /* We can't use nir_compact_varyings with transform feedback, since
797 * the pipe_stream_output->output_register field is based on the
798 * pre-compacted driver_locations.
800 if (!(prev_shader
->sh
.LinkedTransformFeedback
&&
801 prev_shader
->sh
.LinkedTransformFeedback
->NumVarying
> 0))
802 nir_compact_varyings(shader_program
->_LinkedShaders
[prev
]->Program
->nir
,
803 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
805 if (ctx
->Const
.ShaderCompilerOptions
[i
].NirOptions
->vectorize_io
)
806 st_nir_vectorize_io(prev_shader
->nir
, nir
);
811 st_lower_patch_vertices_in(shader_program
);
813 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
814 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
818 struct gl_program
*prog
= shader
->Program
;
819 st_glsl_to_nir_post_opts(st
, prog
, shader_program
);
821 /* Initialize st_vertex_program members. */
822 if (i
== MESA_SHADER_VERTEX
)
823 st_prepare_vertex_program(st_vertex_program(prog
));
825 /* Get pipe_stream_output_info. */
826 if (i
== MESA_SHADER_VERTEX
||
827 i
== MESA_SHADER_TESS_EVAL
||
828 i
== MESA_SHADER_GEOMETRY
)
829 st_translate_stream_output_info(prog
);
831 st_store_ir_in_disk_cache(st
, prog
, true);
833 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
834 _mesa_shader_stage_to_program(i
),
836 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
840 nir_sweep(prog
->nir
);
842 /* The GLSL IR won't be needed anymore. */
843 ralloc_free(shader
->ir
);
851 st_nir_assign_varying_locations(struct st_context
*st
, nir_shader
*nir
)
853 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
854 /* Needs special handling so drvloc matches the vbo state: */
855 st_nir_assign_vs_in_locations(nir
);
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_GEOMETRY
||
862 nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
863 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
864 nir_assign_io_var_locations(&nir
->inputs
,
867 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
869 nir_assign_io_var_locations(&nir
->outputs
,
872 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
873 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
874 nir_assign_io_var_locations(&nir
->inputs
,
877 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
878 nir_assign_io_var_locations(&nir
->outputs
,
881 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
884 unreachable("invalid shader type");
889 st_nir_lower_samplers(struct pipe_screen
*screen
, nir_shader
*nir
,
890 struct gl_shader_program
*shader_program
,
891 struct gl_program
*prog
)
893 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
894 NIR_PASS_V(nir
, gl_nir_lower_samplers_as_deref
, shader_program
);
896 NIR_PASS_V(nir
, gl_nir_lower_samplers
, shader_program
);
899 prog
->info
.textures_used
= nir
->info
.textures_used
;
900 prog
->info
.textures_used_by_txf
= nir
->info
.textures_used_by_txf
;
904 /* Last third of preparing nir from glsl, which happens after shader
908 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
909 struct gl_shader_program
*shader_program
, nir_shader
*nir
)
911 struct pipe_screen
*screen
= st
->pipe
->screen
;
913 NIR_PASS_V(nir
, nir_split_var_copies
);
914 NIR_PASS_V(nir
, nir_lower_var_copies
);
916 st_nir_assign_varying_locations(st
, nir
);
917 st_nir_assign_uniform_locations(st
->ctx
, prog
,
920 /* Set num_uniforms in number of attribute slots (vec4s) */
921 nir
->num_uniforms
= DIV_ROUND_UP(prog
->Parameters
->NumParameterValues
, 4);
923 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
924 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
, st_glsl_type_dword_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
, st_glsl_uniforms_type_size
,
929 (nir_lower_io_options
)0);
932 st_nir_lower_samplers(screen
, nir
, shader_program
, prog
);