2 * Copyright 2017 Advanced Micro Devices, Inc.
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 #include "si_shader_internal.h"
27 #include "ac_nir_to_llvm.h"
29 #include "tgsi/tgsi_from_mesa.h"
31 #include "compiler/nir/nir.h"
32 #include "compiler/nir_types.h"
35 static void scan_instruction(struct tgsi_shader_info
*info
,
38 if (instr
->type
== nir_instr_type_alu
) {
39 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
44 case nir_op_fddx_fine
:
45 case nir_op_fddy_fine
:
46 case nir_op_fddx_coarse
:
47 case nir_op_fddy_coarse
:
48 info
->uses_derivatives
= true;
53 } else if (instr
->type
== nir_instr_type_tex
) {
54 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
57 info
->samplers_declared
|=
58 u_bit_consecutive(tex
->sampler_index
, 1);
65 info
->uses_derivatives
= true;
70 } else if (instr
->type
== nir_instr_type_intrinsic
) {
71 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
73 switch (intr
->intrinsic
) {
74 case nir_intrinsic_load_front_face
:
75 info
->uses_frontface
= 1;
77 case nir_intrinsic_load_instance_id
:
78 info
->uses_instanceid
= 1;
80 case nir_intrinsic_load_invocation_id
:
81 info
->uses_invocationid
= true;
83 case nir_intrinsic_load_num_work_groups
:
84 info
->uses_grid_size
= true;
86 case nir_intrinsic_load_local_group_size
:
87 /* The block size is translated to IMM with a fixed block size. */
88 if (info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] == 0)
89 info
->uses_block_size
= true;
91 case nir_intrinsic_load_local_invocation_id
:
92 case nir_intrinsic_load_work_group_id
: {
93 unsigned mask
= nir_ssa_def_components_read(&intr
->dest
.ssa
);
95 unsigned i
= u_bit_scan(&mask
);
97 if (intr
->intrinsic
== nir_intrinsic_load_work_group_id
)
98 info
->uses_block_id
[i
] = true;
100 info
->uses_thread_id
[i
] = true;
104 case nir_intrinsic_load_vertex_id
:
105 info
->uses_vertexid
= 1;
107 case nir_intrinsic_load_vertex_id_zero_base
:
108 info
->uses_vertexid_nobase
= 1;
110 case nir_intrinsic_load_base_vertex
:
111 info
->uses_basevertex
= 1;
113 case nir_intrinsic_load_primitive_id
:
114 info
->uses_primid
= 1;
116 case nir_intrinsic_load_sample_mask_in
:
117 info
->reads_samplemask
= true;
119 case nir_intrinsic_load_tess_level_inner
:
120 case nir_intrinsic_load_tess_level_outer
:
121 info
->reads_tess_factors
= true;
123 case nir_intrinsic_image_var_store
:
124 case nir_intrinsic_image_var_atomic_add
:
125 case nir_intrinsic_image_var_atomic_min
:
126 case nir_intrinsic_image_var_atomic_max
:
127 case nir_intrinsic_image_var_atomic_and
:
128 case nir_intrinsic_image_var_atomic_or
:
129 case nir_intrinsic_image_var_atomic_xor
:
130 case nir_intrinsic_image_var_atomic_exchange
:
131 case nir_intrinsic_image_var_atomic_comp_swap
:
132 case nir_intrinsic_store_ssbo
:
133 case nir_intrinsic_ssbo_atomic_add
:
134 case nir_intrinsic_ssbo_atomic_imin
:
135 case nir_intrinsic_ssbo_atomic_umin
:
136 case nir_intrinsic_ssbo_atomic_imax
:
137 case nir_intrinsic_ssbo_atomic_umax
:
138 case nir_intrinsic_ssbo_atomic_and
:
139 case nir_intrinsic_ssbo_atomic_or
:
140 case nir_intrinsic_ssbo_atomic_xor
:
141 case nir_intrinsic_ssbo_atomic_exchange
:
142 case nir_intrinsic_ssbo_atomic_comp_swap
:
143 info
->writes_memory
= true;
145 case nir_intrinsic_load_var
: {
146 nir_variable
*var
= intr
->variables
[0]->var
;
147 nir_variable_mode mode
= var
->data
.mode
;
148 enum glsl_base_type base_type
=
149 glsl_get_base_type(glsl_without_array(var
->type
));
151 if (mode
== nir_var_shader_in
) {
152 switch (var
->data
.interpolation
) {
153 case INTERP_MODE_NONE
:
154 if (glsl_base_type_is_integer(base_type
))
158 case INTERP_MODE_SMOOTH
:
159 if (var
->data
.sample
)
160 info
->uses_persp_sample
= true;
161 else if (var
->data
.centroid
)
162 info
->uses_persp_centroid
= true;
164 info
->uses_persp_center
= true;
167 case INTERP_MODE_NOPERSPECTIVE
:
168 if (var
->data
.sample
)
169 info
->uses_linear_sample
= true;
170 else if (var
->data
.centroid
)
171 info
->uses_linear_centroid
= true;
173 info
->uses_linear_center
= true;
179 case nir_intrinsic_interp_var_at_centroid
:
180 case nir_intrinsic_interp_var_at_sample
:
181 case nir_intrinsic_interp_var_at_offset
: {
182 enum glsl_interp_mode interp
=
183 intr
->variables
[0]->var
->data
.interpolation
;
185 case INTERP_MODE_SMOOTH
:
186 case INTERP_MODE_NONE
:
187 if (intr
->intrinsic
== nir_intrinsic_interp_var_at_centroid
)
188 info
->uses_persp_opcode_interp_centroid
= true;
189 else if (intr
->intrinsic
== nir_intrinsic_interp_var_at_sample
)
190 info
->uses_persp_opcode_interp_sample
= true;
192 info
->uses_persp_opcode_interp_offset
= true;
194 case INTERP_MODE_NOPERSPECTIVE
:
195 if (intr
->intrinsic
== nir_intrinsic_interp_var_at_centroid
)
196 info
->uses_linear_opcode_interp_centroid
= true;
197 else if (intr
->intrinsic
== nir_intrinsic_interp_var_at_sample
)
198 info
->uses_linear_opcode_interp_sample
= true;
200 info
->uses_linear_opcode_interp_offset
= true;
202 case INTERP_MODE_FLAT
:
205 unreachable("Unsupported interpoation type");
215 void si_nir_scan_tess_ctrl(const struct nir_shader
*nir
,
216 const struct tgsi_shader_info
*info
,
217 struct tgsi_tessctrl_info
*out
)
219 memset(out
, 0, sizeof(*out
));
221 if (nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
)
224 /* Initial value = true. Here the pass will accumulate results from
225 * multiple segments surrounded by barriers. If tess factors aren't
226 * written at all, it's a shader bug and we don't care if this will be
229 out
->tessfactors_are_def_in_all_invocs
= true;
231 /* TODO: Implement scanning of tess factors, see tgsi backend. */
234 void si_nir_scan_shader(const struct nir_shader
*nir
,
235 struct tgsi_shader_info
*info
)
240 info
->processor
= pipe_shader_type_from_mesa(nir
->info
.stage
);
241 info
->num_tokens
= 2; /* indicate that the shader is non-empty */
242 info
->num_instructions
= 2;
244 if (nir
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
245 info
->properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
] =
246 nir
->info
.tess
.tcs_vertices_out
;
249 if (nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
250 if (nir
->info
.tess
.primitive_mode
== GL_ISOLINES
)
251 info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] = PIPE_PRIM_LINES
;
253 info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] = nir
->info
.tess
.primitive_mode
;
255 STATIC_ASSERT((TESS_SPACING_EQUAL
+ 1) % 3 == PIPE_TESS_SPACING_EQUAL
);
256 STATIC_ASSERT((TESS_SPACING_FRACTIONAL_ODD
+ 1) % 3 ==
257 PIPE_TESS_SPACING_FRACTIONAL_ODD
);
258 STATIC_ASSERT((TESS_SPACING_FRACTIONAL_EVEN
+ 1) % 3 ==
259 PIPE_TESS_SPACING_FRACTIONAL_EVEN
);
261 info
->properties
[TGSI_PROPERTY_TES_SPACING
] = (nir
->info
.tess
.spacing
+ 1) % 3;
262 info
->properties
[TGSI_PROPERTY_TES_VERTEX_ORDER_CW
] = !nir
->info
.tess
.ccw
;
263 info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
] = nir
->info
.tess
.point_mode
;
266 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
267 info
->properties
[TGSI_PROPERTY_GS_INPUT_PRIM
] = nir
->info
.gs
.input_primitive
;
268 info
->properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
] = nir
->info
.gs
.output_primitive
;
269 info
->properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
] = nir
->info
.gs
.vertices_out
;
270 info
->properties
[TGSI_PROPERTY_GS_INVOCATIONS
] = nir
->info
.gs
.invocations
;
273 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
274 info
->properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
] =
275 nir
->info
.fs
.early_fragment_tests
| nir
->info
.fs
.post_depth_coverage
;
276 info
->properties
[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE
] = nir
->info
.fs
.post_depth_coverage
;
278 if (nir
->info
.fs
.pixel_center_integer
) {
279 info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] =
280 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
;
283 if (nir
->info
.fs
.depth_layout
!= FRAG_DEPTH_LAYOUT_NONE
) {
284 switch (nir
->info
.fs
.depth_layout
) {
285 case FRAG_DEPTH_LAYOUT_ANY
:
286 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_ANY
;
288 case FRAG_DEPTH_LAYOUT_GREATER
:
289 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_GREATER
;
291 case FRAG_DEPTH_LAYOUT_LESS
:
292 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_LESS
;
294 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
295 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_UNCHANGED
;
298 unreachable("Unknow depth layout");
303 if (nir
->info
.stage
== MESA_SHADER_COMPUTE
) {
304 info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] = nir
->info
.cs
.local_size
[0];
305 info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] = nir
->info
.cs
.local_size
[1];
306 info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
] = nir
->info
.cs
.local_size
[2];
310 uint64_t processed_inputs
= 0;
311 unsigned num_inputs
= 0;
312 nir_foreach_variable(variable
, &nir
->inputs
) {
313 unsigned semantic_name
, semantic_index
;
315 const struct glsl_type
*type
= variable
->type
;
316 if (nir_is_per_vertex_io(variable
, nir
->info
.stage
)) {
317 assert(glsl_type_is_array(type
));
318 type
= glsl_get_array_element(type
);
321 unsigned attrib_count
= glsl_count_attribute_slots(type
,
322 nir
->info
.stage
== MESA_SHADER_VERTEX
);
324 i
= variable
->data
.driver_location
;
326 /* Vertex shader inputs don't have semantics. The state
327 * tracker has already mapped them to attributes via
328 * variable->data.driver_location.
330 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
331 /* TODO: gather the actual input useage and remove this. */
332 info
->input_usage_mask
[i
] = TGSI_WRITEMASK_XYZW
;
334 if (glsl_type_is_dual_slot(variable
->type
)) {
337 /* TODO: gather the actual input useage and remove this. */
338 info
->input_usage_mask
[i
+1] = TGSI_WRITEMASK_XYZW
;
344 /* Fragment shader position is a system value. */
345 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&&
346 variable
->data
.location
== VARYING_SLOT_POS
) {
347 if (variable
->data
.pixel_center_integer
)
348 info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] =
349 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
;
355 for (unsigned j
= 0; j
< attrib_count
; j
++, i
++) {
357 if (processed_inputs
& ((uint64_t)1 << i
))
360 processed_inputs
|= ((uint64_t)1 << i
);
363 tgsi_get_gl_varying_semantic(variable
->data
.location
+ j
, true,
364 &semantic_name
, &semantic_index
);
366 info
->input_semantic_name
[i
] = semantic_name
;
367 info
->input_semantic_index
[i
] = semantic_index
;
369 if (semantic_name
== TGSI_SEMANTIC_PRIMID
)
370 info
->uses_primid
= true;
372 if (variable
->data
.sample
)
373 info
->input_interpolate_loc
[i
] = TGSI_INTERPOLATE_LOC_SAMPLE
;
374 else if (variable
->data
.centroid
)
375 info
->input_interpolate_loc
[i
] = TGSI_INTERPOLATE_LOC_CENTROID
;
377 info
->input_interpolate_loc
[i
] = TGSI_INTERPOLATE_LOC_CENTER
;
379 enum glsl_base_type base_type
=
380 glsl_get_base_type(glsl_without_array(variable
->type
));
382 switch (variable
->data
.interpolation
) {
383 case INTERP_MODE_NONE
:
384 if (glsl_base_type_is_integer(base_type
)) {
385 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_CONSTANT
;
389 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
390 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_COLOR
;
395 case INTERP_MODE_SMOOTH
:
396 assert(!glsl_base_type_is_integer(base_type
));
398 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_PERSPECTIVE
;
401 case INTERP_MODE_NOPERSPECTIVE
:
402 assert(!glsl_base_type_is_integer(base_type
));
404 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_LINEAR
;
407 case INTERP_MODE_FLAT
:
408 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_CONSTANT
;
412 /* TODO make this more precise */
413 if (variable
->data
.location
== VARYING_SLOT_COL0
)
414 info
->colors_read
|= 0x0f;
415 else if (variable
->data
.location
== VARYING_SLOT_COL1
)
416 info
->colors_read
|= 0xf0;
420 info
->num_inputs
= num_inputs
;
424 uint64_t processed_outputs
= 0;
425 unsigned num_outputs
= 0;
426 nir_foreach_variable(variable
, &nir
->outputs
) {
427 unsigned semantic_name
, semantic_index
;
429 i
= variable
->data
.driver_location
;
431 const struct glsl_type
*type
= variable
->type
;
432 if (nir_is_per_vertex_io(variable
, nir
->info
.stage
)) {
433 assert(glsl_type_is_array(type
));
434 type
= glsl_get_array_element(type
);
437 unsigned attrib_count
= glsl_count_attribute_slots(type
, false);
438 for (unsigned k
= 0; k
< attrib_count
; k
++, i
++) {
440 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
441 tgsi_get_gl_frag_result_semantic(variable
->data
.location
+ k
,
442 &semantic_name
, &semantic_index
);
444 /* Adjust for dual source blending */
445 if (variable
->data
.index
> 0) {
449 tgsi_get_gl_varying_semantic(variable
->data
.location
+ k
, true,
450 &semantic_name
, &semantic_index
);
453 unsigned num_components
= 4;
454 unsigned vector_elements
= glsl_get_vector_elements(glsl_without_array(variable
->type
));
456 num_components
= vector_elements
;
458 unsigned component
= variable
->data
.location_frac
;
459 if (glsl_type_is_64bit(glsl_without_array(variable
->type
))) {
460 if (glsl_type_is_dual_slot(glsl_without_array(variable
->type
)) && k
% 2) {
461 num_components
= (num_components
* 2) - 4;
464 num_components
= MIN2(num_components
* 2, 4);
469 for (unsigned j
= component
; j
< num_components
+ component
; j
++) {
472 usagemask
|= TGSI_WRITEMASK_X
;
475 usagemask
|= TGSI_WRITEMASK_Y
;
478 usagemask
|= TGSI_WRITEMASK_Z
;
481 usagemask
|= TGSI_WRITEMASK_W
;
484 unreachable("error calculating component index");
488 unsigned gs_out_streams
;
489 if (variable
->data
.stream
& (1u << 31)) {
490 gs_out_streams
= variable
->data
.stream
& ~(1u << 31);
492 assert(variable
->data
.stream
< 4);
494 for (unsigned j
= 0; j
< num_components
; ++j
)
495 gs_out_streams
|= variable
->data
.stream
<< (2 * (component
+ j
));
498 unsigned streamx
= gs_out_streams
& 3;
499 unsigned streamy
= (gs_out_streams
>> 2) & 3;
500 unsigned streamz
= (gs_out_streams
>> 4) & 3;
501 unsigned streamw
= (gs_out_streams
>> 6) & 3;
503 if (usagemask
& TGSI_WRITEMASK_X
) {
504 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_X
;
505 info
->output_streams
[i
] |= streamx
;
506 info
->num_stream_output_components
[streamx
]++;
508 if (usagemask
& TGSI_WRITEMASK_Y
) {
509 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_Y
;
510 info
->output_streams
[i
] |= streamy
<< 2;
511 info
->num_stream_output_components
[streamy
]++;
513 if (usagemask
& TGSI_WRITEMASK_Z
) {
514 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_Z
;
515 info
->output_streams
[i
] |= streamz
<< 4;
516 info
->num_stream_output_components
[streamz
]++;
518 if (usagemask
& TGSI_WRITEMASK_W
) {
519 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_W
;
520 info
->output_streams
[i
] |= streamw
<< 6;
521 info
->num_stream_output_components
[streamw
]++;
524 /* make sure we only count this location once against
525 * the num_outputs counter.
527 if (processed_outputs
& ((uint64_t)1 << i
))
530 processed_outputs
|= ((uint64_t)1 << i
);
533 info
->output_semantic_name
[i
] = semantic_name
;
534 info
->output_semantic_index
[i
] = semantic_index
;
536 switch (semantic_name
) {
537 case TGSI_SEMANTIC_PRIMID
:
538 info
->writes_primid
= true;
540 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
541 info
->writes_viewport_index
= true;
543 case TGSI_SEMANTIC_LAYER
:
544 info
->writes_layer
= true;
546 case TGSI_SEMANTIC_PSIZE
:
547 info
->writes_psize
= true;
549 case TGSI_SEMANTIC_CLIPVERTEX
:
550 info
->writes_clipvertex
= true;
552 case TGSI_SEMANTIC_COLOR
:
553 info
->colors_written
|= 1 << semantic_index
;
555 case TGSI_SEMANTIC_STENCIL
:
556 info
->writes_stencil
= true;
558 case TGSI_SEMANTIC_SAMPLEMASK
:
559 info
->writes_samplemask
= true;
561 case TGSI_SEMANTIC_EDGEFLAG
:
562 info
->writes_edgeflag
= true;
564 case TGSI_SEMANTIC_POSITION
:
565 if (info
->processor
== PIPE_SHADER_FRAGMENT
)
566 info
->writes_z
= true;
568 info
->writes_position
= true;
572 if (nir
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
573 switch (semantic_name
) {
574 case TGSI_SEMANTIC_PATCH
:
575 info
->reads_perpatch_outputs
= true;
577 case TGSI_SEMANTIC_TESSINNER
:
578 case TGSI_SEMANTIC_TESSOUTER
:
579 info
->reads_tessfactor_outputs
= true;
582 info
->reads_pervertex_outputs
= true;
587 unsigned loc
= variable
->data
.location
;
588 if (loc
== FRAG_RESULT_COLOR
&&
589 nir
->info
.outputs_written
& (1ull << loc
)) {
590 assert(attrib_count
== 1);
591 info
->properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
] = true;
595 info
->num_outputs
= num_outputs
;
597 nir_foreach_variable(variable
, &nir
->uniforms
) {
598 const struct glsl_type
*type
= variable
->type
;
599 enum glsl_base_type base_type
=
600 glsl_get_base_type(glsl_without_array(type
));
601 unsigned aoa_size
= MAX2(1, glsl_get_aoa_size(type
));
603 /* We rely on the fact that nir_lower_samplers_as_deref has
604 * eliminated struct dereferences.
606 if (base_type
== GLSL_TYPE_SAMPLER
)
607 info
->samplers_declared
|=
608 u_bit_consecutive(variable
->data
.binding
, aoa_size
);
609 else if (base_type
== GLSL_TYPE_IMAGE
)
610 info
->images_declared
|=
611 u_bit_consecutive(variable
->data
.binding
, aoa_size
);
614 info
->num_written_clipdistance
= nir
->info
.clip_distance_array_size
;
615 info
->num_written_culldistance
= nir
->info
.cull_distance_array_size
;
616 info
->clipdist_writemask
= u_bit_consecutive(0, info
->num_written_clipdistance
);
617 info
->culldist_writemask
= u_bit_consecutive(0, info
->num_written_culldistance
);
619 if (info
->processor
== PIPE_SHADER_FRAGMENT
)
620 info
->uses_kill
= nir
->info
.fs
.uses_discard
;
622 /* TODO make this more accurate */
623 info
->const_buffers_declared
= u_bit_consecutive(0, SI_NUM_CONST_BUFFERS
);
624 info
->shader_buffers_declared
= u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS
);
626 func
= (struct nir_function
*)exec_list_get_head_const(&nir
->functions
);
627 nir_foreach_block(block
, func
->impl
) {
628 nir_foreach_instr(instr
, block
)
629 scan_instruction(info
, instr
);
634 * Perform "lowering" operations on the NIR that are run once when the shader
635 * selector is created.
638 si_lower_nir(struct si_shader_selector
* sel
)
640 /* Adjust the driver location of inputs and outputs. The state tracker
641 * interprets them as slots, while the ac/nir backend interprets them
642 * as individual components.
644 nir_foreach_variable(variable
, &sel
->nir
->inputs
)
645 variable
->data
.driver_location
*= 4;
647 nir_foreach_variable(variable
, &sel
->nir
->outputs
) {
648 variable
->data
.driver_location
*= 4;
650 if (sel
->nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
651 if (variable
->data
.location
== FRAG_RESULT_DEPTH
)
652 variable
->data
.driver_location
+= 2;
653 else if (variable
->data
.location
== FRAG_RESULT_STENCIL
)
654 variable
->data
.driver_location
+= 1;
658 /* Perform lowerings (and optimizations) of code.
660 * Performance considerations aside, we must:
661 * - lower certain ALU operations
662 * - ensure constant offsets for texture instructions are folded
663 * and copy-propagated
665 NIR_PASS_V(sel
->nir
, nir_lower_returns
);
666 NIR_PASS_V(sel
->nir
, nir_lower_vars_to_ssa
);
667 NIR_PASS_V(sel
->nir
, nir_lower_alu_to_scalar
);
668 NIR_PASS_V(sel
->nir
, nir_lower_phis_to_scalar
);
670 static const struct nir_lower_tex_options lower_tex_options
= {
673 NIR_PASS_V(sel
->nir
, nir_lower_tex
, &lower_tex_options
);
675 const nir_lower_subgroups_options subgroups_options
= {
677 .ballot_bit_size
= 64,
678 .lower_to_scalar
= true,
679 .lower_subgroup_masks
= true,
680 .lower_vote_trivial
= false,
681 .lower_vote_eq_to_ballot
= true,
683 NIR_PASS_V(sel
->nir
, nir_lower_subgroups
, &subgroups_options
);
685 ac_lower_indirect_derefs(sel
->nir
, sel
->screen
->info
.chip_class
);
691 /* (Constant) copy propagation is needed for txf with offsets. */
692 NIR_PASS(progress
, sel
->nir
, nir_copy_prop
);
693 NIR_PASS(progress
, sel
->nir
, nir_opt_remove_phis
);
694 NIR_PASS(progress
, sel
->nir
, nir_opt_dce
);
695 if (nir_opt_trivial_continues(sel
->nir
)) {
697 NIR_PASS(progress
, sel
->nir
, nir_copy_prop
);
698 NIR_PASS(progress
, sel
->nir
, nir_opt_dce
);
700 NIR_PASS(progress
, sel
->nir
, nir_opt_if
);
701 NIR_PASS(progress
, sel
->nir
, nir_opt_dead_cf
);
702 NIR_PASS(progress
, sel
->nir
, nir_opt_cse
);
703 NIR_PASS(progress
, sel
->nir
, nir_opt_peephole_select
, 8);
705 /* Needed for algebraic lowering */
706 NIR_PASS(progress
, sel
->nir
, nir_opt_algebraic
);
707 NIR_PASS(progress
, sel
->nir
, nir_opt_constant_folding
);
709 NIR_PASS(progress
, sel
->nir
, nir_opt_undef
);
710 NIR_PASS(progress
, sel
->nir
, nir_opt_conditional_discard
);
711 if (sel
->nir
->options
->max_unroll_iterations
) {
712 NIR_PASS(progress
, sel
->nir
, nir_opt_loop_unroll
, 0);
717 static void declare_nir_input_vs(struct si_shader_context
*ctx
,
718 struct nir_variable
*variable
,
719 unsigned input_index
,
722 si_llvm_load_input_vs(ctx
, input_index
, out
);
725 static void declare_nir_input_fs(struct si_shader_context
*ctx
,
726 struct nir_variable
*variable
,
727 unsigned input_index
,
730 unsigned slot
= variable
->data
.location
;
731 if (slot
== VARYING_SLOT_POS
) {
732 out
[0] = LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
);
733 out
[1] = LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
);
734 out
[2] = LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Z_FLOAT
);
735 out
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
736 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_W_FLOAT
));
740 si_llvm_load_input_fs(ctx
, input_index
, out
);
744 si_nir_lookup_interp_param(struct ac_shader_abi
*abi
,
745 enum glsl_interp_mode interp
, unsigned location
)
747 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
748 int interp_param_idx
= -1;
751 case INTERP_MODE_FLAT
:
753 case INTERP_MODE_SMOOTH
:
754 case INTERP_MODE_NONE
:
755 if (location
== INTERP_CENTER
)
756 interp_param_idx
= SI_PARAM_PERSP_CENTER
;
757 else if (location
== INTERP_CENTROID
)
758 interp_param_idx
= SI_PARAM_PERSP_CENTROID
;
759 else if (location
== INTERP_SAMPLE
)
760 interp_param_idx
= SI_PARAM_PERSP_SAMPLE
;
762 case INTERP_MODE_NOPERSPECTIVE
:
763 if (location
== INTERP_CENTER
)
764 interp_param_idx
= SI_PARAM_LINEAR_CENTER
;
765 else if (location
== INTERP_CENTROID
)
766 interp_param_idx
= SI_PARAM_LINEAR_CENTROID
;
767 else if (location
== INTERP_SAMPLE
)
768 interp_param_idx
= SI_PARAM_LINEAR_SAMPLE
;
771 assert(!"Unhandled interpolation mode.");
775 return interp_param_idx
!= -1 ?
776 LLVMGetParam(ctx
->main_fn
, interp_param_idx
) : NULL
;
780 si_nir_load_sampler_desc(struct ac_shader_abi
*abi
,
781 unsigned descriptor_set
, unsigned base_index
,
782 unsigned constant_index
, LLVMValueRef dynamic_index
,
783 enum ac_descriptor_type desc_type
, bool image
,
786 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
787 LLVMBuilderRef builder
= ctx
->ac
.builder
;
788 LLVMValueRef list
= LLVMGetParam(ctx
->main_fn
, ctx
->param_samplers_and_images
);
789 LLVMValueRef index
= dynamic_index
;
791 assert(!descriptor_set
);
794 index
= ctx
->ac
.i32_0
;
796 index
= LLVMBuildAdd(builder
, index
,
797 LLVMConstInt(ctx
->ac
.i32
, base_index
+ constant_index
, false),
801 assert(desc_type
== AC_DESC_IMAGE
|| desc_type
== AC_DESC_BUFFER
);
802 assert(base_index
+ constant_index
< ctx
->num_images
);
805 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_images
);
807 index
= LLVMBuildSub(ctx
->gallivm
.builder
,
808 LLVMConstInt(ctx
->i32
, SI_NUM_IMAGES
- 1, 0),
811 /* TODO: be smarter about when we use dcc_off */
812 return si_load_image_desc(ctx
, list
, index
, desc_type
, write
);
815 assert(base_index
+ constant_index
< ctx
->num_samplers
);
818 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_samplers
);
820 index
= LLVMBuildAdd(ctx
->gallivm
.builder
, index
,
821 LLVMConstInt(ctx
->i32
, SI_NUM_IMAGES
/ 2, 0), "");
823 return si_load_sampler_desc(ctx
, list
, index
, desc_type
);
826 static void bitcast_inputs(struct si_shader_context
*ctx
,
827 LLVMValueRef data
[4],
830 for (unsigned chan
= 0; chan
< 4; chan
++) {
831 ctx
->inputs
[input_idx
+ chan
] =
832 LLVMBuildBitCast(ctx
->ac
.builder
, data
[chan
], ctx
->ac
.i32
, "");
836 bool si_nir_build_llvm(struct si_shader_context
*ctx
, struct nir_shader
*nir
)
838 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
840 if (nir
->info
.stage
== MESA_SHADER_VERTEX
||
841 nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
842 uint64_t processed_inputs
= 0;
843 nir_foreach_variable(variable
, &nir
->inputs
) {
844 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
,
845 nir
->info
.stage
== MESA_SHADER_VERTEX
);
846 unsigned input_idx
= variable
->data
.driver_location
;
848 LLVMValueRef data
[4];
849 unsigned loc
= variable
->data
.location
;
851 for (unsigned i
= 0; i
< attrib_count
; i
++) {
852 /* Packed components share the same location so skip
853 * them if we have already processed the location.
855 if (processed_inputs
& ((uint64_t)1 << (loc
+ i
))) {
860 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
861 declare_nir_input_vs(ctx
, variable
, input_idx
/ 4, data
);
862 bitcast_inputs(ctx
, data
, input_idx
);
863 if (glsl_type_is_dual_slot(variable
->type
)) {
865 declare_nir_input_vs(ctx
, variable
, input_idx
/ 4, data
);
866 bitcast_inputs(ctx
, data
, input_idx
);
868 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
869 declare_nir_input_fs(ctx
, variable
, input_idx
/ 4, data
);
870 bitcast_inputs(ctx
, data
, input_idx
);
873 processed_inputs
|= ((uint64_t)1 << (loc
+ i
));
879 ctx
->abi
.inputs
= &ctx
->inputs
[0];
880 ctx
->abi
.load_sampler_desc
= si_nir_load_sampler_desc
;
881 ctx
->abi
.clamp_shadow_reference
= true;
883 ctx
->num_samplers
= util_last_bit(info
->samplers_declared
);
884 ctx
->num_images
= util_last_bit(info
->images_declared
);
886 if (ctx
->shader
->selector
->local_size
) {
887 assert(nir
->info
.stage
== MESA_SHADER_COMPUTE
);
888 si_declare_compute_memory(ctx
);
890 ac_nir_translate(&ctx
->ac
, &ctx
->abi
, nir
);