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
, 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_PNTC
) {
75 var
->data
.location
= VARYING_SLOT_VAR8
;
76 } else if ((var
->data
.location
>= VARYING_SLOT_TEX0
) &&
77 (var
->data
.location
<= VARYING_SLOT_TEX7
)) {
78 var
->data
.location
+= VARYING_SLOT_VAR0
- VARYING_SLOT_TEX0
;
83 /* input location assignment for VS inputs must be handled specially, so
84 * that it is aligned w/ st's vbo state.
85 * (This isn't the case with, for ex, FS inputs, which only need to agree
86 * on varying-slot w/ the VS outputs)
89 st_nir_assign_vs_in_locations(struct nir_shader
*nir
)
91 if (nir
->info
.stage
!= MESA_SHADER_VERTEX
)
94 bool removed_inputs
= false;
96 nir
->num_inputs
= util_bitcount64(nir
->info
.inputs_read
);
97 nir_foreach_variable_safe(var
, &nir
->inputs
) {
98 /* NIR already assigns dual-slot inputs to two locations so all we have
99 * to do is compact everything down.
101 if (nir
->info
.inputs_read
& BITFIELD64_BIT(var
->data
.location
)) {
102 var
->data
.driver_location
=
103 util_bitcount64(nir
->info
.inputs_read
&
104 BITFIELD64_MASK(var
->data
.location
));
106 /* Move unused input variables to the globals list (with no
107 * initialization), to avoid confusing drivers looking through the
108 * inputs array and expecting to find inputs with a driver_location
111 exec_node_remove(&var
->node
);
112 var
->data
.mode
= nir_var_shader_temp
;
113 exec_list_push_tail(&nir
->globals
, &var
->node
);
114 removed_inputs
= true;
118 /* Re-lower global vars, to deal with any dead VS inputs. */
120 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
124 st_nir_lookup_parameter_index(struct gl_program
*prog
, nir_variable
*var
)
126 struct gl_program_parameter_list
*params
= prog
->Parameters
;
128 /* Lookup the first parameter that the uniform storage that match the
131 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
132 int index
= params
->Parameters
[i
].MainUniformStorageIndex
;
133 if (index
== var
->data
.location
)
137 /* TODO: Handle this fallback for SPIR-V. We need this for GLSL e.g. in
138 * dEQP-GLES2.functional.uniform_api.random.3
141 /* is there a better way to do this? If we have something like:
149 * Then what we get in prog->Parameters looks like:
151 * 0: Name=color.f, Type=6, DataType=1406, Size=1
152 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
154 * So the name doesn't match up and _mesa_lookup_parameter_index()
155 * fails. In this case just find the first matching "color.*"..
157 * Note for arrays you could end up w/ color[n].f, for example.
159 * glsl_to_tgsi works slightly differently in this regard. It is
160 * emitting something more low level, so it just translates the
161 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
162 * it just calculates the additional offset of struct field members
163 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
164 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
165 * needs to work backwards to get base var loc from the param-list
166 * which already has them separated out.
168 if (!prog
->sh
.data
->spirv
) {
169 int namelen
= strlen(var
->name
);
170 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
171 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
172 if ((strncmp(p
->Name
, var
->name
, namelen
) == 0) &&
173 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
183 st_nir_assign_uniform_locations(struct gl_context
*ctx
,
184 struct gl_program
*prog
,
185 struct exec_list
*uniform_list
)
190 nir_foreach_variable(uniform
, uniform_list
) {
194 * UBO's have their own address spaces, so don't count them towards the
195 * number of global uniforms
197 if (uniform
->data
.mode
== nir_var_mem_ubo
|| uniform
->data
.mode
== nir_var_mem_ssbo
)
200 const struct glsl_type
*type
= glsl_without_array(uniform
->type
);
201 if (!uniform
->data
.bindless
&& (type
->is_sampler() || type
->is_image())) {
202 if (type
->is_sampler()) {
204 shaderidx
+= type_size(uniform
->type
);
207 imageidx
+= type_size(uniform
->type
);
209 } else if (uniform
->state_slots
) {
210 const gl_state_index16
*const stateTokens
= uniform
->state_slots
[0].tokens
;
211 /* This state reference has already been setup by ir_to_mesa, but we'll
212 * get the same index back here.
216 if (glsl_type_is_struct_or_ifc(type
)) {
219 comps
= glsl_get_vector_elements(type
);
222 if (ctx
->Const
.PackedDriverUniformStorage
) {
223 loc
= _mesa_add_sized_state_reference(prog
->Parameters
,
224 stateTokens
, comps
, false);
225 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
227 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
230 loc
= st_nir_lookup_parameter_index(prog
, uniform
);
232 /* We need to check that loc is not -1 here before accessing the
233 * array. It can be negative for example when we have a struct that
234 * only contains opaque types.
236 if (loc
>= 0 && ctx
->Const
.PackedDriverUniformStorage
) {
237 loc
= prog
->Parameters
->ParameterValueOffset
[loc
];
241 uniform
->data
.driver_location
= loc
;
246 st_nir_opts(nir_shader
*nir
)
253 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
255 /* Linking deals with unused inputs/outputs, but here we can remove
256 * things local to the shader in the hopes that we can cleanup other
257 * things. This pass will also remove variables with only stores, so we
258 * might be able to make progress after it.
260 NIR_PASS(progress
, nir
, nir_remove_dead_variables
,
261 (nir_variable_mode
)(nir_var_function_temp
|
262 nir_var_shader_temp
|
266 NIR_PASS(progress
, nir
, nir_opt_copy_prop_vars
);
267 NIR_PASS(progress
, nir
, nir_opt_dead_write_vars
);
269 if (nir
->options
->lower_to_scalar
) {
270 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
271 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
274 NIR_PASS_V(nir
, nir_lower_alu
);
275 NIR_PASS_V(nir
, nir_lower_pack
);
276 NIR_PASS(progress
, nir
, nir_copy_prop
);
277 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
278 NIR_PASS(progress
, nir
, nir_opt_dce
);
279 if (nir_opt_trivial_continues(nir
)) {
281 NIR_PASS(progress
, nir
, nir_copy_prop
);
282 NIR_PASS(progress
, nir
, nir_opt_dce
);
284 NIR_PASS(progress
, nir
, nir_opt_if
, false);
285 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
286 NIR_PASS(progress
, nir
, nir_opt_cse
);
287 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
289 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
290 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
292 if (!nir
->info
.flrp_lowered
) {
293 unsigned lower_flrp
=
294 (nir
->options
->lower_flrp16
? 16 : 0) |
295 (nir
->options
->lower_flrp32
? 32 : 0) |
296 (nir
->options
->lower_flrp64
? 64 : 0);
299 bool lower_flrp_progress
= false;
301 NIR_PASS(lower_flrp_progress
, nir
, nir_lower_flrp
,
303 false /* always_precise */,
304 nir
->options
->lower_ffma
);
305 if (lower_flrp_progress
) {
306 NIR_PASS(progress
, nir
,
307 nir_opt_constant_folding
);
312 /* Nothing should rematerialize any flrps, so we only need to do this
315 nir
->info
.flrp_lowered
= true;
318 NIR_PASS(progress
, nir
, nir_opt_undef
);
319 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
320 if (nir
->options
->max_unroll_iterations
) {
321 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
327 shared_type_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
329 assert(glsl_type_is_vector_or_scalar(type
));
331 uint32_t comp_size
= glsl_type_is_boolean(type
)
332 ? 4 : glsl_get_bit_size(type
) / 8;
333 unsigned length
= glsl_get_vector_elements(type
);
334 *size
= comp_size
* length
,
335 *align
= comp_size
* (length
== 3 ? 4 : length
);
338 /* First third of converting glsl_to_nir.. this leaves things in a pre-
339 * nir_lower_io state, so that shader variants can more easily insert/
340 * replace variables, etc.
343 st_nir_preprocess(struct st_context
*st
, struct gl_program
*prog
,
344 struct gl_shader_program
*shader_program
,
345 gl_shader_stage stage
)
347 struct pipe_screen
*screen
= st
->pipe
->screen
;
348 const nir_shader_compiler_options
*options
=
349 st
->ctx
->Const
.ShaderCompilerOptions
[prog
->info
.stage
].NirOptions
;
351 nir_shader
*nir
= prog
->nir
;
353 /* Set the next shader stage hint for VS and TES. */
354 if (!nir
->info
.separate_shader
&&
355 (nir
->info
.stage
== MESA_SHADER_VERTEX
||
356 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
)) {
358 unsigned prev_stages
= (1 << (prog
->info
.stage
+ 1)) - 1;
359 unsigned stages_mask
=
360 ~prev_stages
& shader_program
->data
->linked_stages
;
362 nir
->info
.next_stage
= stages_mask
?
363 (gl_shader_stage
) u_bit_scan(&stages_mask
) : MESA_SHADER_FRAGMENT
;
365 nir
->info
.next_stage
= MESA_SHADER_FRAGMENT
;
368 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
369 if (!st
->ctx
->SoftFP64
&& nir
->info
.uses_64bit
&&
370 (options
->lower_doubles_options
& nir_lower_fp64_full_software
) != 0) {
371 st
->ctx
->SoftFP64
= glsl_float64_funcs_to_nir(st
->ctx
, options
);
374 /* ES has strict SSO validation rules for shader IO matching so we can't
375 * remove dead IO until the resource list has been built. Here we skip
376 * removing them until later. This will potentially make the IO lowering
377 * calls below do a little extra work but should otherwise have no impact.
379 if (!_mesa_is_gles(st
->ctx
) || !nir
->info
.separate_shader
) {
380 nir_variable_mode mask
=
381 (nir_variable_mode
) (nir_var_shader_in
| nir_var_shader_out
);
382 nir_remove_dead_variables(nir
, mask
, NULL
);
385 if (options
->lower_all_io_to_temps
||
386 nir
->info
.stage
== MESA_SHADER_VERTEX
||
387 nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
388 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
389 nir_shader_get_entrypoint(nir
),
391 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
||
392 !screen
->get_param(screen
, PIPE_CAP_TGSI_CAN_READ_OUTPUTS
)) {
393 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
394 nir_shader_get_entrypoint(nir
),
398 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
399 NIR_PASS_V(nir
, nir_split_var_copies
);
400 NIR_PASS_V(nir
, nir_lower_var_copies
);
402 if (options
->lower_to_scalar
) {
403 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
, NULL
);
406 /* before buffers and vars_to_ssa */
407 NIR_PASS_V(nir
, gl_nir_lower_images
, true);
409 /* TODO: Change GLSL to not lower shared memory. */
410 if (prog
->nir
->info
.stage
== MESA_SHADER_COMPUTE
&&
411 shader_program
->data
->spirv
) {
412 NIR_PASS_V(prog
->nir
, nir_lower_vars_to_explicit_types
,
413 nir_var_mem_shared
, shared_type_info
);
414 NIR_PASS_V(prog
->nir
, nir_lower_explicit_io
,
415 nir_var_mem_shared
, nir_address_format_32bit_offset
);
418 /* Do a round of constant folding to clean up address calculations */
419 NIR_PASS_V(nir
, nir_opt_constant_folding
);
422 /* Second third of converting glsl_to_nir. This creates uniforms, gathers
423 * info on varyings, etc after NIR link time opts have been applied.
426 st_glsl_to_nir_post_opts(struct st_context
*st
, struct gl_program
*prog
,
427 struct gl_shader_program
*shader_program
)
429 nir_shader
*nir
= prog
->nir
;
431 /* Make a pass over the IR to add state references for any built-in
432 * uniforms that are used. This has to be done now (during linking).
433 * Code generation doesn't happen until the first time this shader is
434 * used for rendering. Waiting until then to generate the parameters is
435 * too late. At that point, the values for the built-in uniforms won't
436 * get sent to the shader.
438 nir_foreach_variable(var
, &nir
->uniforms
) {
439 const nir_state_slot
*const slots
= var
->state_slots
;
441 const struct glsl_type
*type
= glsl_without_array(var
->type
);
442 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
444 if (glsl_type_is_struct_or_ifc(type
)) {
445 /* Builtin struct require specical handling for now we just
446 * make all members vec4. See st_nir_lower_builtin.
450 comps
= glsl_get_vector_elements(type
);
453 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
454 _mesa_add_sized_state_reference(prog
->Parameters
,
458 _mesa_add_state_reference(prog
->Parameters
,
465 /* Avoid reallocation of the program parameter list, because the uniform
466 * storage is only associated with the original parameter list.
467 * This should be enough for Bitmap and DrawPixels constants.
469 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
471 /* This has to be done last. Any operation the can cause
472 * prog->ParameterValues to get reallocated (e.g., anything that adds a
473 * program constant) has to happen before creating this linkage.
475 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
);
477 st_set_prog_affected_state_flags(prog
);
479 /* None of the builtins being lowered here can be produced by SPIR-V. See
480 * _mesa_builtin_uniform_desc.
482 if (!shader_program
->data
->spirv
)
483 NIR_PASS_V(nir
, st_nir_lower_builtin
);
485 NIR_PASS_V(nir
, gl_nir_lower_atomics
, shader_program
, true);
486 NIR_PASS_V(nir
, nir_opt_intrinsics
);
488 /* Lower 64-bit ops. */
489 if (nir
->options
->lower_int64_options
||
490 nir
->options
->lower_doubles_options
) {
491 bool lowered_64bit_ops
= false;
492 if (nir
->options
->lower_doubles_options
) {
493 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_doubles
,
494 st
->ctx
->SoftFP64
, nir
->options
->lower_doubles_options
);
496 if (nir
->options
->lower_int64_options
) {
497 NIR_PASS(lowered_64bit_ops
, nir
, nir_lower_int64
,
498 nir
->options
->lower_int64_options
);
501 if (lowered_64bit_ops
)
505 nir_variable_mode mask
= (nir_variable_mode
)
506 (nir_var_shader_in
| nir_var_shader_out
| nir_var_function_temp
);
507 nir_remove_dead_variables(nir
, mask
, NULL
);
509 if (!st
->has_hw_atomics
)
510 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
);
512 st_finalize_nir_before_variants(nir
);
514 if (st
->allow_st_finalize_nir_twice
)
515 st_finalize_nir(st
, prog
, shader_program
, nir
, true);
517 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
519 _mesa_log("NIR IR for linked %s program %d:\n",
520 _mesa_shader_stage_to_string(prog
->info
.stage
),
521 shader_program
->Name
);
522 nir_print_shader(nir
, _mesa_get_log_file());
528 st_nir_vectorize_io(nir_shader
*producer
, nir_shader
*consumer
)
530 NIR_PASS_V(producer
, nir_lower_io_to_vector
, nir_var_shader_out
);
531 NIR_PASS_V(producer
, nir_opt_combine_stores
, nir_var_shader_out
);
532 NIR_PASS_V(consumer
, nir_lower_io_to_vector
, nir_var_shader_in
);
534 if ((producer
)->info
.stage
!= MESA_SHADER_TESS_CTRL
) {
535 /* Calling lower_io_to_vector creates output variable writes with
536 * write-masks. We only support these for TCS outputs, so for other
537 * stages, we need to call nir_lower_io_to_temporaries to get rid of
538 * them. This, in turn, creates temporary variables and extra
539 * copy_deref intrinsics that we need to clean up.
541 NIR_PASS_V(producer
, nir_lower_io_to_temporaries
,
542 nir_shader_get_entrypoint(producer
), true, false);
543 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
544 NIR_PASS_V(producer
, nir_split_var_copies
);
545 NIR_PASS_V(producer
, nir_lower_var_copies
);
550 st_nir_link_shaders(nir_shader
*producer
, nir_shader
*consumer
)
552 if (producer
->options
->lower_to_scalar
) {
553 NIR_PASS_V(producer
, nir_lower_io_to_scalar_early
, nir_var_shader_out
);
554 NIR_PASS_V(consumer
, nir_lower_io_to_scalar_early
, nir_var_shader_in
);
557 nir_lower_io_arrays_to_elements(producer
, consumer
);
559 st_nir_opts(producer
);
560 st_nir_opts(consumer
);
562 if (nir_link_opt_varyings(producer
, consumer
))
563 st_nir_opts(consumer
);
565 NIR_PASS_V(producer
, nir_remove_dead_variables
, nir_var_shader_out
, NULL
);
566 NIR_PASS_V(consumer
, nir_remove_dead_variables
, nir_var_shader_in
, NULL
);
568 if (nir_remove_unused_varyings(producer
, consumer
)) {
569 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
570 NIR_PASS_V(consumer
, nir_lower_global_vars_to_local
);
572 st_nir_opts(producer
);
573 st_nir_opts(consumer
);
575 /* Optimizations can cause varyings to become unused.
576 * nir_compact_varyings() depends on all dead varyings being removed so
577 * we need to call nir_remove_dead_variables() again here.
579 NIR_PASS_V(producer
, nir_remove_dead_variables
, nir_var_shader_out
,
581 NIR_PASS_V(consumer
, nir_remove_dead_variables
, nir_var_shader_in
,
587 st_lower_patch_vertices_in(struct gl_shader_program
*shader_prog
)
589 struct gl_linked_shader
*linked_tcs
=
590 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_CTRL
];
591 struct gl_linked_shader
*linked_tes
=
592 shader_prog
->_LinkedShaders
[MESA_SHADER_TESS_EVAL
];
594 /* If we have a TCS and TES linked together, lower TES patch vertices. */
595 if (linked_tcs
&& linked_tes
) {
596 nir_shader
*tcs_nir
= linked_tcs
->Program
->nir
;
597 nir_shader
*tes_nir
= linked_tes
->Program
->nir
;
599 /* The TES input vertex count is the TCS output vertex count,
600 * lower TES gl_PatchVerticesIn to a constant.
602 uint32_t tes_patch_verts
= tcs_nir
->info
.tess
.tcs_vertices_out
;
603 NIR_PASS_V(tes_nir
, nir_lower_patch_vertices
, tes_patch_verts
, NULL
);
610 st_nir_lower_wpos_ytransform(struct nir_shader
*nir
,
611 struct gl_program
*prog
,
612 struct pipe_screen
*pscreen
)
614 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
617 static const gl_state_index16 wposTransformState
[STATE_LENGTH
] = {
618 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
620 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
622 memcpy(wpos_options
.state_tokens
, wposTransformState
,
623 sizeof(wpos_options
.state_tokens
));
624 wpos_options
.fs_coord_origin_upper_left
=
625 pscreen
->get_param(pscreen
,
626 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
627 wpos_options
.fs_coord_origin_lower_left
=
628 pscreen
->get_param(pscreen
,
629 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
630 wpos_options
.fs_coord_pixel_center_integer
=
631 pscreen
->get_param(pscreen
,
632 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
633 wpos_options
.fs_coord_pixel_center_half_integer
=
634 pscreen
->get_param(pscreen
,
635 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
637 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
638 nir_validate_shader(nir
, "after nir_lower_wpos_ytransform");
639 _mesa_add_state_reference(prog
->Parameters
, wposTransformState
);
644 st_link_nir(struct gl_context
*ctx
,
645 struct gl_shader_program
*shader_program
)
647 struct st_context
*st
= st_context(ctx
);
648 struct gl_linked_shader
*linked_shader
[MESA_SHADER_STAGES
];
649 unsigned num_shaders
= 0;
651 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
652 if (shader_program
->_LinkedShaders
[i
])
653 linked_shader
[num_shaders
++] = shader_program
->_LinkedShaders
[i
];
656 for (unsigned i
= 0; i
< num_shaders
; i
++) {
657 struct gl_linked_shader
*shader
= linked_shader
[i
];
658 const nir_shader_compiler_options
*options
=
659 st
->ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
;
660 struct gl_program
*prog
= shader
->Program
;
661 struct st_program
*stp
= (struct st_program
*)prog
;
663 _mesa_copy_linked_program_data(shader_program
, shader
);
666 stp
->shader_program
= shader_program
;
667 stp
->state
.type
= PIPE_SHADER_IR_NIR
;
669 /* Parameters will be filled during NIR linking. */
670 prog
->Parameters
= _mesa_new_parameter_list();
672 if (shader_program
->data
->spirv
) {
673 prog
->nir
= _mesa_spirv_to_nir(ctx
, shader_program
, shader
->Stage
, options
);
675 validate_ir_tree(shader
->ir
);
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
->nir
= glsl_to_nir(st
->ctx
, shader_program
, shader
->Stage
, options
);
687 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
690 if (options
->lower_to_scalar
) {
691 NIR_PASS_V(shader
->Program
->nir
, nir_lower_load_const_to_scalar
);
695 st_lower_patch_vertices_in(shader_program
);
697 /* For SPIR-V, we have to perform the NIR linking before applying
700 if (shader_program
->data
->spirv
) {
701 static const gl_nir_linker_options opts
= {
702 true /*fill_parameters */
704 if (!gl_nir_link_spirv(ctx
, shader_program
, &opts
))
707 nir_build_program_resource_list(ctx
, shader_program
, true);
709 for (unsigned i
= 0; i
< num_shaders
; i
++) {
710 struct gl_linked_shader
*shader
= linked_shader
[i
];
711 struct gl_program
*prog
= shader
->Program
;
713 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
714 _mesa_update_shader_textures_used(shader_program
, prog
);
715 st_nir_preprocess(st
, prog
, shader_program
, shader
->Stage
);
719 /* Linking the stages in the opposite order (from fragment to vertex)
720 * ensures that inter-shader outputs written to in an earlier stage
721 * are eliminated if they are (transitively) not used in a later
724 for (int i
= num_shaders
- 2; i
>= 0; i
--) {
725 st_nir_link_shaders(linked_shader
[i
]->Program
->nir
,
726 linked_shader
[i
+ 1]->Program
->nir
);
728 /* Linking shaders also optimizes them. Separate shaders, compute shaders
729 * and shaders with a fixed-func VS or FS that don't need linking are
732 if (num_shaders
== 1)
733 st_nir_opts(linked_shader
[0]->Program
->nir
);
735 if (!shader_program
->data
->spirv
) {
736 if (!gl_nir_link_glsl(ctx
, shader_program
))
739 for (unsigned i
= 0; i
< num_shaders
; i
++) {
740 struct gl_program
*prog
= linked_shader
[i
]->Program
;
741 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
742 _mesa_update_shader_textures_used(shader_program
, prog
);
745 nir_build_program_resource_list(ctx
, shader_program
, false);
748 for (unsigned i
= 0; i
< num_shaders
; i
++) {
749 struct gl_linked_shader
*shader
= linked_shader
[i
];
750 nir_shader
*nir
= shader
->Program
->nir
;
752 /* This needs to run after the initial pass of nir_lower_vars_to_ssa, so
753 * that the buffer indices are constants in nir where they where
754 * constants in GLSL. */
755 NIR_PASS_V(nir
, gl_nir_lower_buffers
, shader_program
);
757 /* Remap the locations to slots so those requiring two slots will occupy
758 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
759 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
760 * locations/slots 0 and 1, and attr1 will use location/slot 2
762 if (nir
->info
.stage
== MESA_SHADER_VERTEX
&& !shader_program
->data
->spirv
)
763 nir_remap_dual_slot_attributes(nir
, &shader
->Program
->DualSlotInputs
);
765 NIR_PASS_V(nir
, st_nir_lower_wpos_ytransform
, shader
->Program
,
768 NIR_PASS_V(nir
, nir_lower_system_values
);
769 NIR_PASS_V(nir
, nir_lower_clip_cull_distance_arrays
);
771 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
772 shader
->Program
->info
= nir
->info
;
773 if (shader
->Stage
== MESA_SHADER_VERTEX
) {
774 /* NIR expands dual-slot inputs out to two locations. We need to
775 * compact things back down GL-style single-slot inputs to avoid
776 * confusing the state tracker.
778 shader
->Program
->info
.inputs_read
=
779 nir_get_single_slot_attribs_mask(nir
->info
.inputs_read
,
780 shader
->Program
->DualSlotInputs
);
784 struct gl_program
*prev_shader
= linked_shader
[i
- 1]->Program
;
786 /* We can't use nir_compact_varyings with transform feedback, since
787 * the pipe_stream_output->output_register field is based on the
788 * pre-compacted driver_locations.
790 if (!(prev_shader
->sh
.LinkedTransformFeedback
&&
791 prev_shader
->sh
.LinkedTransformFeedback
->NumVarying
> 0))
792 nir_compact_varyings(prev_shader
->nir
,
793 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
795 if (ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
->vectorize_io
)
796 st_nir_vectorize_io(prev_shader
->nir
, nir
);
800 struct shader_info
*prev_info
= NULL
;
802 for (unsigned i
= 0; i
< num_shaders
; i
++) {
803 struct gl_linked_shader
*shader
= linked_shader
[i
];
804 struct shader_info
*info
= &shader
->Program
->nir
->info
;
807 ctx
->Const
.ShaderCompilerOptions
[shader
->Stage
].NirOptions
->unify_interfaces
) {
808 prev_info
->outputs_written
|= info
->inputs_read
&
809 ~(VARYING_BIT_TESS_LEVEL_INNER
| VARYING_BIT_TESS_LEVEL_OUTER
);
810 info
->inputs_read
|= prev_info
->outputs_written
&
811 ~(VARYING_BIT_TESS_LEVEL_INNER
| VARYING_BIT_TESS_LEVEL_OUTER
);
813 prev_info
->patch_outputs_written
|= info
->patch_inputs_read
;
814 info
->patch_inputs_read
|= prev_info
->patch_outputs_written
;
819 for (unsigned i
= 0; i
< num_shaders
; i
++) {
820 struct gl_linked_shader
*shader
= linked_shader
[i
];
821 struct gl_program
*prog
= shader
->Program
;
822 struct st_program
*stp
= st_program(prog
);
823 st_glsl_to_nir_post_opts(st
, prog
, shader_program
);
825 /* Initialize st_vertex_program members. */
826 if (shader
->Stage
== MESA_SHADER_VERTEX
)
827 st_prepare_vertex_program(stp
);
829 /* Get pipe_stream_output_info. */
830 if (shader
->Stage
== MESA_SHADER_VERTEX
||
831 shader
->Stage
== MESA_SHADER_TESS_EVAL
||
832 shader
->Stage
== MESA_SHADER_GEOMETRY
)
833 st_translate_stream_output_info(prog
);
835 st_store_ir_in_disk_cache(st
, prog
, true);
837 st_release_variants(st
, stp
);
838 st_finalize_program(st
, prog
);
840 /* The GLSL IR won't be needed anymore. */
841 ralloc_free(shader
->ir
);
849 st_nir_assign_varying_locations(struct st_context
*st
, nir_shader
*nir
)
851 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
852 nir_assign_io_var_locations(&nir
->outputs
,
855 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
856 } else if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
||
857 nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
858 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
859 nir_assign_io_var_locations(&nir
->inputs
,
862 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
864 nir_assign_io_var_locations(&nir
->outputs
,
867 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
868 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
869 nir_assign_io_var_locations(&nir
->inputs
,
872 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
873 nir_assign_io_var_locations(&nir
->outputs
,
876 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
879 unreachable("invalid shader type");
884 st_nir_lower_samplers(struct pipe_screen
*screen
, nir_shader
*nir
,
885 struct gl_shader_program
*shader_program
,
886 struct gl_program
*prog
)
888 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
889 NIR_PASS_V(nir
, gl_nir_lower_samplers_as_deref
, shader_program
);
891 NIR_PASS_V(nir
, gl_nir_lower_samplers
, shader_program
);
894 prog
->info
.textures_used
= nir
->info
.textures_used
;
895 prog
->info
.textures_used_by_txf
= nir
->info
.textures_used_by_txf
;
896 prog
->info
.images_used
= nir
->info
.images_used
;
901 st_packed_uniforms_type_size(const struct glsl_type
*type
, bool bindless
)
903 return glsl_count_dword_slots(type
, bindless
);
907 st_unpacked_uniforms_type_size(const struct glsl_type
*type
, bool bindless
)
909 return glsl_count_vec4_slots(type
, false, bindless
);
913 st_nir_lower_uniforms(struct st_context
*st
, nir_shader
*nir
)
915 if (st
->ctx
->Const
.PackedDriverUniformStorage
) {
916 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
,
917 st_packed_uniforms_type_size
,
918 (nir_lower_io_options
)0);
919 NIR_PASS_V(nir
, nir_lower_uniforms_to_ubo
, 4);
921 NIR_PASS_V(nir
, nir_lower_io
, nir_var_uniform
,
922 st_unpacked_uniforms_type_size
,
923 (nir_lower_io_options
)0);
927 /* Last third of preparing nir from glsl, which happens after shader
931 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
932 struct gl_shader_program
*shader_program
,
933 nir_shader
*nir
, bool finalize_by_driver
)
935 struct pipe_screen
*screen
= st
->pipe
->screen
;
937 NIR_PASS_V(nir
, nir_split_var_copies
);
938 NIR_PASS_V(nir
, nir_lower_var_copies
);
940 st_nir_assign_varying_locations(st
, nir
);
941 st_nir_assign_uniform_locations(st
->ctx
, prog
,
944 /* Set num_uniforms in number of attribute slots (vec4s) */
945 nir
->num_uniforms
= DIV_ROUND_UP(prog
->Parameters
->NumParameterValues
, 4);
947 st_nir_lower_uniforms(st
, nir
);
948 st_nir_lower_samplers(screen
, nir
, shader_program
, prog
);
949 if (!screen
->get_param(screen
, PIPE_CAP_NIR_IMAGES_AS_DEREF
))
950 NIR_PASS_V(nir
, gl_nir_lower_images
, false);
952 if (finalize_by_driver
&& screen
->finalize_nir
)
953 screen
->finalize_nir(screen
, nir
, false);