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/shaderapi.h"
37 #include "main/uniforms.h"
39 #include "st_context.h"
40 #include "st_program.h"
42 #include "compiler/nir/nir.h"
43 #include "compiler/glsl_types.h"
44 #include "compiler/glsl/glsl_to_nir.h"
45 #include "compiler/glsl/ir.h"
46 #include "compiler/glsl/string_to_uint_map.h"
50 type_size(const struct glsl_type
*type
)
52 return type
->count_attribute_slots(false);
55 /* Depending on PIPE_CAP_TGSI_TEXCOORD (st->needs_texcoord_semantic) we
56 * may need to fix up varying slots so the glsl->nir path is aligned
57 * with the anything->tgsi->nir path.
60 st_nir_fixup_varying_slots(struct st_context
*st
, struct exec_list
*var_list
)
62 if (st
->needs_texcoord_semantic
)
65 nir_foreach_variable(var
, var_list
) {
66 if (var
->data
.location
>= VARYING_SLOT_VAR0
) {
67 var
->data
.location
+= 9;
68 } else if ((var
->data
.location
>= VARYING_SLOT_TEX0
) &&
69 (var
->data
.location
<= VARYING_SLOT_TEX7
)) {
70 var
->data
.location
+= VARYING_SLOT_VAR0
- VARYING_SLOT_TEX0
;
75 /* input location assignment for VS inputs must be handled specially, so
76 * that it is aligned w/ st's vbo state.
77 * (This isn't the case with, for ex, FS inputs, which only need to agree
78 * on varying-slot w/ the VS outputs)
81 st_nir_assign_vs_in_locations(struct gl_program
*prog
, nir_shader
*nir
)
83 unsigned attr
, num_inputs
= 0;
84 unsigned input_to_index
[VERT_ATTRIB_MAX
] = {0};
86 /* TODO de-duplicate w/ similar code in st_translate_vertex_program()? */
87 for (attr
= 0; attr
< VERT_ATTRIB_MAX
; attr
++) {
88 if ((prog
->info
.inputs_read
& BITFIELD64_BIT(attr
)) != 0) {
89 input_to_index
[attr
] = num_inputs
;
91 if ((prog
->info
.double_inputs_read
& BITFIELD64_BIT(attr
)) != 0) {
92 /* add placeholder for second part of a double attribute */
96 input_to_index
[attr
] = ~0;
100 /* bit of a hack, mirroring st_translate_vertex_program */
101 input_to_index
[VERT_ATTRIB_EDGEFLAG
] = num_inputs
;
104 nir_foreach_variable_safe(var
, &nir
->inputs
) {
105 attr
= var
->data
.location
;
106 assert(attr
< ARRAY_SIZE(input_to_index
));
108 if (input_to_index
[attr
] != ~0u) {
109 var
->data
.driver_location
= input_to_index
[attr
];
112 /* Move unused input variables to the globals list (with no
113 * initialization), to avoid confusing drivers looking through the
114 * inputs array and expecting to find inputs with a driver_location
117 exec_node_remove(&var
->node
);
118 var
->data
.mode
= nir_var_global
;
119 exec_list_push_tail(&nir
->globals
, &var
->node
);
125 st_nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
)
127 unsigned location
= 0;
128 unsigned assigned_locations
[VARYING_SLOT_MAX
];
129 uint64_t processed_locs
= 0;
131 nir_foreach_variable(var
, var_list
) {
132 /* Because component packing allows varyings to share the same location
133 * we may have already have processed this location.
135 if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
136 processed_locs
& ((uint64_t)1 << var
->data
.location
)) {
137 var
->data
.driver_location
= assigned_locations
[var
->data
.location
];
138 *size
+= type_size(var
->type
);
142 assigned_locations
[var
->data
.location
] = location
;
143 var
->data
.driver_location
= location
;
144 location
+= type_size(var
->type
);
146 processed_locs
|= ((uint64_t)1 << var
->data
.location
);
153 st_nir_lookup_parameter_index(const struct gl_program_parameter_list
*params
,
156 int loc
= _mesa_lookup_parameter_index(params
, name
);
158 /* is there a better way to do this? If we have something like:
166 * Then what we get in prog->Parameters looks like:
168 * 0: Name=color.f, Type=6, DataType=1406, Size=1
169 * 1: Name=color.v, Type=6, DataType=8b52, Size=4
171 * So the name doesn't match up and _mesa_lookup_parameter_index()
172 * fails. In this case just find the first matching "color.*"..
174 * Note for arrays you could end up w/ color[n].f, for example.
176 * glsl_to_tgsi works slightly differently in this regard. It is
177 * emitting something more low level, so it just translates the
178 * params list 1:1 to CONST[] regs. Going from GLSL IR to TGSI,
179 * it just calculates the additional offset of struct field members
180 * in glsl_to_tgsi_visitor::visit(ir_dereference_record *ir) or
181 * glsl_to_tgsi_visitor::visit(ir_dereference_array *ir). It never
182 * needs to work backwards to get base var loc from the param-list
183 * which already has them separated out.
186 int namelen
= strlen(name
);
187 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
188 struct gl_program_parameter
*p
= ¶ms
->Parameters
[i
];
189 if ((strncmp(p
->Name
, name
, namelen
) == 0) &&
190 ((p
->Name
[namelen
] == '.') || (p
->Name
[namelen
] == '['))) {
201 st_nir_assign_uniform_locations(struct gl_program
*prog
,
202 struct gl_shader_program
*shader_program
,
203 struct exec_list
*uniform_list
, unsigned *size
)
209 nir_foreach_variable(uniform
, uniform_list
) {
213 * UBO's have their own address spaces, so don't count them towards the
214 * number of global uniforms
216 if ((uniform
->data
.mode
== nir_var_uniform
|| uniform
->data
.mode
== nir_var_shader_storage
) &&
217 uniform
->interface_type
!= NULL
)
220 if (uniform
->type
->is_sampler() || uniform
->type
->is_image()) {
222 bool found
= shader_program
->UniformHash
->get(val
, uniform
->name
);
223 if (uniform
->type
->is_sampler())
228 (void) found
; /* silence unused var warning */
229 /* this ensure that nir_lower_samplers looks at the correct
230 * shader_program->UniformStorage[location]:
232 uniform
->data
.location
= val
;
233 } else if (strncmp(uniform
->name
, "gl_", 3) == 0) {
234 const gl_state_index
*const stateTokens
= (gl_state_index
*)uniform
->state_slots
[0].tokens
;
235 /* This state reference has already been setup by ir_to_mesa, but we'll
236 * get the same index back here.
238 loc
= _mesa_add_state_reference(prog
->Parameters
, stateTokens
);
240 loc
= st_nir_lookup_parameter_index(prog
->Parameters
, uniform
->name
);
243 uniform
->data
.driver_location
= loc
;
245 max
= MAX2(max
, loc
+ type_size(uniform
->type
));
251 st_nir_opts(nir_shader
*nir
)
257 NIR_PASS(progress
, nir
, nir_copy_prop
);
258 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
259 NIR_PASS(progress
, nir
, nir_opt_dce
);
260 if (nir_opt_trivial_continues(nir
)) {
262 NIR_PASS(progress
, nir
, nir_copy_prop
);
263 NIR_PASS(progress
, nir
, nir_opt_dce
);
265 NIR_PASS(progress
, nir
, nir_opt_if
);
266 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
267 NIR_PASS(progress
, nir
, nir_opt_cse
);
268 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8);
270 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
271 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
273 NIR_PASS(progress
, nir
, nir_opt_undef
);
274 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
275 if (nir
->options
->max_unroll_iterations
) {
276 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, (nir_variable_mode
)0);
281 /* First third of converting glsl_to_nir.. this leaves things in a pre-
282 * nir_lower_io state, so that shader variants can more easily insert/
283 * replace variables, etc.
286 st_glsl_to_nir(struct st_context
*st
, struct gl_program
*prog
,
287 struct gl_shader_program
*shader_program
,
288 gl_shader_stage stage
)
290 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
291 enum pipe_shader_type ptarget
= pipe_shader_type_from_mesa(stage
);
292 const nir_shader_compiler_options
*options
;
294 assert(pscreen
->get_compiler_options
); /* drivers using NIR must implement this */
296 options
= (const nir_shader_compiler_options
*)
297 pscreen
->get_compiler_options(pscreen
, PIPE_SHADER_IR_NIR
, ptarget
);
303 nir_shader
*nir
= glsl_to_nir(shader_program
, stage
, options
);
310 /* Second third of converting glsl_to_nir. This creates uniforms, gathers
311 * info on varyings, etc after NIR link time opts have been applied.
314 st_glsl_to_nir_post_opts(struct st_context
*st
, struct gl_program
*prog
,
315 struct gl_shader_program
*shader_program
)
317 nir_shader
*nir
= prog
->nir
;
319 /* Make a pass over the IR to add state references for any built-in
320 * uniforms that are used. This has to be done now (during linking).
321 * Code generation doesn't happen until the first time this shader is
322 * used for rendering. Waiting until then to generate the parameters is
323 * too late. At that point, the values for the built-in uniforms won't
324 * get sent to the shader.
326 nir_foreach_variable(var
, &nir
->uniforms
) {
327 if (strncmp(var
->name
, "gl_", 3) == 0) {
328 const nir_state_slot
*const slots
= var
->state_slots
;
329 assert(var
->state_slots
!= NULL
);
331 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
332 _mesa_add_state_reference(prog
->Parameters
,
333 (gl_state_index
*)slots
[i
].tokens
);
338 /* Avoid reallocation of the program parameter list, because the uniform
339 * storage is only associated with the original parameter list.
340 * This should be enough for Bitmap and DrawPixels constants.
342 _mesa_reserve_parameter_storage(prog
->Parameters
, 8);
344 /* This has to be done last. Any operation the can cause
345 * prog->ParameterValues to get reallocated (e.g., anything that adds a
346 * program constant) has to happen before creating this linkage.
348 _mesa_associate_uniform_storage(st
->ctx
, shader_program
, prog
, true);
350 st_set_prog_affected_state_flags(prog
);
352 NIR_PASS_V(nir
, st_nir_lower_builtin
);
353 NIR_PASS_V(nir
, nir_lower_atomics
, shader_program
);
355 if (st
->ctx
->_Shader
->Flags
& GLSL_DUMP
) {
357 _mesa_log("NIR IR for linked %s program %d:\n",
358 _mesa_shader_stage_to_string(prog
->info
.stage
),
359 shader_program
->Name
);
360 nir_print_shader(nir
, _mesa_get_log_file());
365 /* TODO any better helper somewhere to sort a list? */
368 insert_sorted(struct exec_list
*var_list
, nir_variable
*new_var
)
370 nir_foreach_variable(var
, var_list
) {
371 if (var
->data
.location
> new_var
->data
.location
) {
372 exec_node_insert_node_before(&var
->node
, &new_var
->node
);
376 exec_list_push_tail(var_list
, &new_var
->node
);
380 sort_varyings(struct exec_list
*var_list
)
382 struct exec_list new_list
;
383 exec_list_make_empty(&new_list
);
384 nir_foreach_variable_safe(var
, var_list
) {
385 exec_node_remove(&var
->node
);
386 insert_sorted(&new_list
, var
);
388 exec_list_move_nodes_to(&new_list
, var_list
);
392 set_st_program(struct gl_program
*prog
,
393 struct gl_shader_program
*shader_program
,
396 struct st_vertex_program
*stvp
;
397 struct st_common_program
*stp
;
398 struct st_fragment_program
*stfp
;
399 struct st_compute_program
*stcp
;
401 switch (prog
->info
.stage
) {
402 case MESA_SHADER_VERTEX
:
403 stvp
= (struct st_vertex_program
*)prog
;
404 stvp
->shader_program
= shader_program
;
405 stvp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
406 stvp
->tgsi
.ir
.nir
= nir
;
408 case MESA_SHADER_GEOMETRY
:
409 case MESA_SHADER_TESS_CTRL
:
410 case MESA_SHADER_TESS_EVAL
:
411 stp
= (struct st_common_program
*)prog
;
412 stp
->shader_program
= shader_program
;
413 stp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
414 stp
->tgsi
.ir
.nir
= nir
;
416 case MESA_SHADER_FRAGMENT
:
417 stfp
= (struct st_fragment_program
*)prog
;
418 stfp
->shader_program
= shader_program
;
419 stfp
->tgsi
.type
= PIPE_SHADER_IR_NIR
;
420 stfp
->tgsi
.ir
.nir
= nir
;
422 case MESA_SHADER_COMPUTE
:
423 stcp
= (struct st_compute_program
*)prog
;
424 stcp
->shader_program
= shader_program
;
425 stcp
->tgsi
.ir_type
= PIPE_SHADER_IR_NIR
;
426 stcp
->tgsi
.prog
= nir_shader_clone(NULL
, nir
);
429 unreachable("unknown shader stage");
434 st_nir_get_mesa_program(struct gl_context
*ctx
,
435 struct gl_shader_program
*shader_program
,
436 struct gl_linked_shader
*shader
)
438 struct st_context
*st
= st_context(ctx
);
439 struct gl_program
*prog
;
441 validate_ir_tree(shader
->ir
);
443 prog
= shader
->Program
;
445 prog
->Parameters
= _mesa_new_parameter_list();
447 _mesa_copy_linked_program_data(shader_program
, shader
);
448 _mesa_generate_parameters_list_for_uniforms(ctx
, shader_program
, shader
,
451 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
453 _mesa_log("GLSL IR for linked %s program %d:\n",
454 _mesa_shader_stage_to_string(shader
->Stage
),
455 shader_program
->Name
);
456 _mesa_print_ir(_mesa_get_log_file(), shader
->ir
, NULL
);
460 prog
->ExternalSamplersUsed
= gl_external_samplers(prog
);
461 _mesa_update_shader_textures_used(shader_program
, prog
);
463 nir_shader
*nir
= st_glsl_to_nir(st
, prog
, shader_program
, shader
->Stage
);
465 set_st_program(prog
, shader_program
, nir
);
468 NIR_PASS_V(nir
, nir_lower_io_to_temporaries
,
469 nir_shader_get_entrypoint(nir
),
471 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
472 NIR_PASS_V(nir
, nir_split_var_copies
);
473 NIR_PASS_V(nir
, nir_lower_var_copies
);
477 st_nir_link_shaders(nir_shader
**producer
, nir_shader
**consumer
)
479 nir_lower_io_arrays_to_elements(*producer
, *consumer
);
481 NIR_PASS_V(*producer
, nir_remove_dead_variables
, nir_var_shader_out
);
482 NIR_PASS_V(*consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
484 if (nir_remove_unused_varyings(*producer
, *consumer
)) {
485 NIR_PASS_V(*producer
, nir_lower_global_vars_to_local
);
486 NIR_PASS_V(*consumer
, nir_lower_global_vars_to_local
);
488 /* The backend might not be able to handle indirects on
489 * temporaries so we need to lower indirects on any of the
490 * varyings we have demoted here.
492 * TODO: radeonsi shouldn't need to do this, however LLVM isn't
493 * currently smart enough to handle indirects without causing excess
494 * spilling causing the gpu to hang.
496 * See the following thread for more details of the problem:
497 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
499 nir_variable_mode indirect_mask
= nir_var_local
;
501 NIR_PASS_V(*producer
, nir_lower_indirect_derefs
, indirect_mask
);
502 NIR_PASS_V(*consumer
, nir_lower_indirect_derefs
, indirect_mask
);
504 st_nir_opts(*producer
);
505 st_nir_opts(*consumer
);
512 st_link_nir(struct gl_context
*ctx
,
513 struct gl_shader_program
*shader_program
)
515 struct st_context
*st
= st_context(ctx
);
517 /* Determine first and last stage. */
518 unsigned first
= MESA_SHADER_STAGES
;
520 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
521 if (!shader_program
->_LinkedShaders
[i
])
523 if (first
== MESA_SHADER_STAGES
)
528 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
529 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
533 st_nir_get_mesa_program(ctx
, shader_program
, shader
);
535 nir_variable_mode mask
= (nir_variable_mode
) 0;
537 mask
= (nir_variable_mode
)(mask
| nir_var_shader_in
);
540 mask
= (nir_variable_mode
)(mask
| nir_var_shader_out
);
542 nir_shader
*nir
= shader
->Program
->nir
;
543 nir_lower_io_to_scalar_early(nir
, mask
);
547 /* Linking the stages in the opposite order (from fragment to vertex)
548 * ensures that inter-shader outputs written to in an earlier stage
549 * are eliminated if they are (transitively) not used in a later
553 for (int i
= next
- 1; i
>= 0; i
--) {
554 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
558 st_nir_link_shaders(&shader
->Program
->nir
,
559 &shader_program
->_LinkedShaders
[next
]->Program
->nir
);
564 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
565 struct gl_linked_shader
*shader
= shader_program
->_LinkedShaders
[i
];
569 nir_shader
*nir
= shader
->Program
->nir
;
571 /* fragment shaders may need : */
572 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
573 static const gl_state_index wposTransformState
[STATE_LENGTH
] = {
574 STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
576 nir_lower_wpos_ytransform_options wpos_options
= { { 0 } };
577 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
579 memcpy(wpos_options
.state_tokens
, wposTransformState
,
580 sizeof(wpos_options
.state_tokens
));
581 wpos_options
.fs_coord_origin_upper_left
=
582 pscreen
->get_param(pscreen
,
583 PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
);
584 wpos_options
.fs_coord_origin_lower_left
=
585 pscreen
->get_param(pscreen
,
586 PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
587 wpos_options
.fs_coord_pixel_center_integer
=
588 pscreen
->get_param(pscreen
,
589 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
590 wpos_options
.fs_coord_pixel_center_half_integer
=
591 pscreen
->get_param(pscreen
,
592 PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
);
594 if (nir_lower_wpos_ytransform(nir
, &wpos_options
)) {
595 nir_validate_shader(nir
);
596 _mesa_add_state_reference(shader
->Program
->Parameters
,
601 NIR_PASS_V(nir
, nir_lower_system_values
);
603 nir_shader_gather_info(nir
, nir_shader_get_entrypoint(nir
));
604 shader
->Program
->info
= nir
->info
;
607 nir_compact_varyings(shader_program
->_LinkedShaders
[prev
]->Program
->nir
,
608 nir
, ctx
->API
!= API_OPENGL_COMPAT
);
612 st_glsl_to_nir_post_opts(st
, shader
->Program
, shader_program
);
614 assert(shader
->Program
);
615 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
616 _mesa_shader_stage_to_program(i
),
618 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
626 /* Last third of preparing nir from glsl, which happens after shader
630 st_finalize_nir(struct st_context
*st
, struct gl_program
*prog
,
631 struct gl_shader_program
*shader_program
, nir_shader
*nir
)
633 struct pipe_screen
*screen
= st
->pipe
->screen
;
635 NIR_PASS_V(nir
, nir_split_var_copies
);
636 NIR_PASS_V(nir
, nir_lower_var_copies
);
637 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
);
639 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
640 /* Needs special handling so drvloc matches the vbo state: */
641 st_nir_assign_vs_in_locations(prog
, nir
);
642 /* Re-lower global vars, to deal with any dead VS inputs. */
643 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
645 sort_varyings(&nir
->outputs
);
646 st_nir_assign_var_locations(&nir
->outputs
,
648 st_nir_fixup_varying_slots(st
, &nir
->outputs
);
649 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
650 sort_varyings(&nir
->inputs
);
651 st_nir_assign_var_locations(&nir
->inputs
,
653 st_nir_fixup_varying_slots(st
, &nir
->inputs
);
654 st_nir_assign_var_locations(&nir
->outputs
,
656 } else if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
659 unreachable("invalid shader type for tgsi bypass\n");
662 NIR_PASS_V(nir
, nir_lower_atomics_to_ssbo
,
663 st
->ctx
->Const
.Program
[nir
->info
.stage
].MaxAtomicBuffers
);
665 st_nir_assign_uniform_locations(prog
, shader_program
,
666 &nir
->uniforms
, &nir
->num_uniforms
);
668 if (screen
->get_param(screen
, PIPE_CAP_NIR_SAMPLERS_AS_DEREF
))
669 NIR_PASS_V(nir
, nir_lower_samplers_as_deref
, shader_program
);
671 NIR_PASS_V(nir
, nir_lower_samplers
, shader_program
);