2 * Copyright 2017 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 #include "si_shader_internal.h"
28 #include "ac_nir_to_llvm.h"
30 #include "tgsi/tgsi_from_mesa.h"
32 #include "compiler/nir/nir.h"
33 #include "compiler/nir_types.h"
34 #include "compiler/nir/nir_builder.h"
36 static nir_variable
* tex_get_texture_var(nir_tex_instr
*instr
)
38 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
39 switch (instr
->src
[i
].src_type
) {
40 case nir_tex_src_texture_deref
:
41 return nir_deref_instr_get_variable(nir_src_as_deref(instr
->src
[i
].src
));
50 static nir_variable
* intrinsic_get_var(nir_intrinsic_instr
*instr
)
52 return nir_deref_instr_get_variable(nir_src_as_deref(instr
->src
[0]));
55 static void gather_intrinsic_load_deref_input_info(const nir_shader
*nir
,
56 const nir_intrinsic_instr
*instr
,
58 struct tgsi_shader_info
*info
)
60 assert(var
&& var
->data
.mode
== nir_var_shader_in
);
62 switch (nir
->info
.stage
) {
63 case MESA_SHADER_VERTEX
: {
64 unsigned i
= var
->data
.driver_location
;
65 unsigned attrib_count
= glsl_count_attribute_slots(var
->type
, false);
67 for (unsigned j
= 0; j
< attrib_count
; j
++, i
++) {
68 if (glsl_type_is_64bit(glsl_without_array(var
->type
))) {
69 /* TODO: set usage mask more accurately for doubles */
70 info
->input_usage_mask
[i
] = TGSI_WRITEMASK_XYZW
;
72 uint8_t mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
73 info
->input_usage_mask
[i
] |= mask
<< var
->data
.location_frac
;
78 case MESA_SHADER_FRAGMENT
:
79 if (var
->data
.location
== VARYING_SLOT_COL0
||
80 var
->data
.location
== VARYING_SLOT_COL1
) {
81 unsigned index
= var
->data
.location
== VARYING_SLOT_COL1
;
82 uint8_t mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
83 info
->colors_read
|= mask
<< (index
* 4);
90 static void gather_intrinsic_load_deref_output_info(const nir_shader
*nir
,
91 const nir_intrinsic_instr
*instr
,
93 struct tgsi_shader_info
*info
)
95 assert(var
&& var
->data
.mode
== nir_var_shader_out
);
97 switch (nir
->info
.stage
) {
98 case MESA_SHADER_FRAGMENT
:
99 if (var
->data
.fb_fetch_output
)
100 info
->uses_fbfetch
= true;
106 static void scan_instruction(const struct nir_shader
*nir
,
107 struct tgsi_shader_info
*info
,
110 if (instr
->type
== nir_instr_type_alu
) {
111 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
116 case nir_op_fddx_fine
:
117 case nir_op_fddy_fine
:
118 case nir_op_fddx_coarse
:
119 case nir_op_fddy_coarse
:
120 info
->uses_derivatives
= true;
125 } else if (instr
->type
== nir_instr_type_tex
) {
126 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
127 nir_variable
*texture
= tex_get_texture_var(tex
);
130 info
->samplers_declared
|=
131 u_bit_consecutive(tex
->sampler_index
, 1);
133 if (texture
->data
.bindless
)
134 info
->uses_bindless_samplers
= true;
141 info
->uses_derivatives
= true;
146 } else if (instr
->type
== nir_instr_type_intrinsic
) {
147 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
149 switch (intr
->intrinsic
) {
150 case nir_intrinsic_load_front_face
:
151 info
->uses_frontface
= 1;
153 case nir_intrinsic_load_instance_id
:
154 info
->uses_instanceid
= 1;
156 case nir_intrinsic_load_invocation_id
:
157 info
->uses_invocationid
= true;
159 case nir_intrinsic_load_num_work_groups
:
160 info
->uses_grid_size
= true;
162 case nir_intrinsic_load_local_group_size
:
163 /* The block size is translated to IMM with a fixed block size. */
164 if (info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] == 0)
165 info
->uses_block_size
= true;
167 case nir_intrinsic_load_local_invocation_id
:
168 case nir_intrinsic_load_work_group_id
: {
169 unsigned mask
= nir_ssa_def_components_read(&intr
->dest
.ssa
);
171 unsigned i
= u_bit_scan(&mask
);
173 if (intr
->intrinsic
== nir_intrinsic_load_work_group_id
)
174 info
->uses_block_id
[i
] = true;
176 info
->uses_thread_id
[i
] = true;
180 case nir_intrinsic_load_vertex_id
:
181 info
->uses_vertexid
= 1;
183 case nir_intrinsic_load_vertex_id_zero_base
:
184 info
->uses_vertexid_nobase
= 1;
186 case nir_intrinsic_load_base_vertex
:
187 info
->uses_basevertex
= 1;
189 case nir_intrinsic_load_draw_id
:
190 info
->uses_drawid
= 1;
192 case nir_intrinsic_load_primitive_id
:
193 info
->uses_primid
= 1;
195 case nir_intrinsic_load_sample_mask_in
:
196 info
->reads_samplemask
= true;
198 case nir_intrinsic_load_tess_level_inner
:
199 case nir_intrinsic_load_tess_level_outer
:
200 info
->reads_tess_factors
= true;
202 case nir_intrinsic_bindless_image_load
:
203 info
->uses_bindless_images
= true;
205 if (nir_intrinsic_image_dim(intr
) == GLSL_SAMPLER_DIM_BUF
)
206 info
->uses_bindless_buffer_load
= true;
208 info
->uses_bindless_image_load
= true;
210 case nir_intrinsic_bindless_image_size
:
211 case nir_intrinsic_bindless_image_samples
:
212 info
->uses_bindless_images
= true;
214 case nir_intrinsic_bindless_image_store
:
215 info
->uses_bindless_images
= true;
217 if (nir_intrinsic_image_dim(intr
) == GLSL_SAMPLER_DIM_BUF
)
218 info
->uses_bindless_buffer_store
= true;
220 info
->uses_bindless_image_store
= true;
222 info
->writes_memory
= true;
224 case nir_intrinsic_image_deref_store
:
225 info
->writes_memory
= true;
227 case nir_intrinsic_bindless_image_atomic_add
:
228 case nir_intrinsic_bindless_image_atomic_min
:
229 case nir_intrinsic_bindless_image_atomic_max
:
230 case nir_intrinsic_bindless_image_atomic_and
:
231 case nir_intrinsic_bindless_image_atomic_or
:
232 case nir_intrinsic_bindless_image_atomic_xor
:
233 case nir_intrinsic_bindless_image_atomic_exchange
:
234 case nir_intrinsic_bindless_image_atomic_comp_swap
:
235 info
->uses_bindless_images
= true;
237 if (nir_intrinsic_image_dim(intr
) == GLSL_SAMPLER_DIM_BUF
)
238 info
->uses_bindless_buffer_atomic
= true;
240 info
->uses_bindless_image_atomic
= true;
242 info
->writes_memory
= true;
244 case nir_intrinsic_image_deref_atomic_add
:
245 case nir_intrinsic_image_deref_atomic_min
:
246 case nir_intrinsic_image_deref_atomic_max
:
247 case nir_intrinsic_image_deref_atomic_and
:
248 case nir_intrinsic_image_deref_atomic_or
:
249 case nir_intrinsic_image_deref_atomic_xor
:
250 case nir_intrinsic_image_deref_atomic_exchange
:
251 case nir_intrinsic_image_deref_atomic_comp_swap
:
252 info
->writes_memory
= true;
254 case nir_intrinsic_store_ssbo
:
255 case nir_intrinsic_ssbo_atomic_add
:
256 case nir_intrinsic_ssbo_atomic_imin
:
257 case nir_intrinsic_ssbo_atomic_umin
:
258 case nir_intrinsic_ssbo_atomic_imax
:
259 case nir_intrinsic_ssbo_atomic_umax
:
260 case nir_intrinsic_ssbo_atomic_and
:
261 case nir_intrinsic_ssbo_atomic_or
:
262 case nir_intrinsic_ssbo_atomic_xor
:
263 case nir_intrinsic_ssbo_atomic_exchange
:
264 case nir_intrinsic_ssbo_atomic_comp_swap
:
265 info
->writes_memory
= true;
267 case nir_intrinsic_load_deref
: {
268 nir_variable
*var
= intrinsic_get_var(intr
);
269 nir_variable_mode mode
= var
->data
.mode
;
270 enum glsl_base_type base_type
=
271 glsl_get_base_type(glsl_without_array(var
->type
));
273 if (mode
== nir_var_shader_in
) {
274 gather_intrinsic_load_deref_input_info(nir
, intr
, var
, info
);
276 switch (var
->data
.interpolation
) {
277 case INTERP_MODE_NONE
:
278 if (glsl_base_type_is_integer(base_type
))
282 case INTERP_MODE_SMOOTH
:
283 if (var
->data
.sample
)
284 info
->uses_persp_sample
= true;
285 else if (var
->data
.centroid
)
286 info
->uses_persp_centroid
= true;
288 info
->uses_persp_center
= true;
291 case INTERP_MODE_NOPERSPECTIVE
:
292 if (var
->data
.sample
)
293 info
->uses_linear_sample
= true;
294 else if (var
->data
.centroid
)
295 info
->uses_linear_centroid
= true;
297 info
->uses_linear_center
= true;
300 } else if (mode
== nir_var_shader_out
) {
301 gather_intrinsic_load_deref_output_info(nir
, intr
, var
, info
);
305 case nir_intrinsic_interp_deref_at_centroid
:
306 case nir_intrinsic_interp_deref_at_sample
:
307 case nir_intrinsic_interp_deref_at_offset
: {
308 enum glsl_interp_mode interp
= intrinsic_get_var(intr
)->data
.interpolation
;
310 case INTERP_MODE_SMOOTH
:
311 case INTERP_MODE_NONE
:
312 if (intr
->intrinsic
== nir_intrinsic_interp_deref_at_centroid
)
313 info
->uses_persp_opcode_interp_centroid
= true;
314 else if (intr
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
315 info
->uses_persp_opcode_interp_sample
= true;
317 info
->uses_persp_opcode_interp_offset
= true;
319 case INTERP_MODE_NOPERSPECTIVE
:
320 if (intr
->intrinsic
== nir_intrinsic_interp_deref_at_centroid
)
321 info
->uses_linear_opcode_interp_centroid
= true;
322 else if (intr
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
323 info
->uses_linear_opcode_interp_sample
= true;
325 info
->uses_linear_opcode_interp_offset
= true;
327 case INTERP_MODE_FLAT
:
330 unreachable("Unsupported interpoation type");
340 void si_nir_scan_tess_ctrl(const struct nir_shader
*nir
,
341 struct tgsi_tessctrl_info
*out
)
343 memset(out
, 0, sizeof(*out
));
345 if (nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
)
348 out
->tessfactors_are_def_in_all_invocs
=
349 ac_are_tessfactors_def_in_all_invocs(nir
);
352 void si_nir_scan_shader(const struct nir_shader
*nir
,
353 struct tgsi_shader_info
*info
)
358 info
->processor
= pipe_shader_type_from_mesa(nir
->info
.stage
);
359 info
->num_tokens
= 2; /* indicate that the shader is non-empty */
360 info
->num_instructions
= 2;
362 info
->properties
[TGSI_PROPERTY_NEXT_SHADER
] =
363 pipe_shader_type_from_mesa(nir
->info
.next_stage
);
365 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
366 info
->properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
] =
367 nir
->info
.vs
.window_space_position
;
370 if (nir
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
371 info
->properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
] =
372 nir
->info
.tess
.tcs_vertices_out
;
375 if (nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
376 if (nir
->info
.tess
.primitive_mode
== GL_ISOLINES
)
377 info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] = PIPE_PRIM_LINES
;
379 info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] = nir
->info
.tess
.primitive_mode
;
381 STATIC_ASSERT((TESS_SPACING_EQUAL
+ 1) % 3 == PIPE_TESS_SPACING_EQUAL
);
382 STATIC_ASSERT((TESS_SPACING_FRACTIONAL_ODD
+ 1) % 3 ==
383 PIPE_TESS_SPACING_FRACTIONAL_ODD
);
384 STATIC_ASSERT((TESS_SPACING_FRACTIONAL_EVEN
+ 1) % 3 ==
385 PIPE_TESS_SPACING_FRACTIONAL_EVEN
);
387 info
->properties
[TGSI_PROPERTY_TES_SPACING
] = (nir
->info
.tess
.spacing
+ 1) % 3;
388 info
->properties
[TGSI_PROPERTY_TES_VERTEX_ORDER_CW
] = !nir
->info
.tess
.ccw
;
389 info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
] = nir
->info
.tess
.point_mode
;
392 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
393 info
->properties
[TGSI_PROPERTY_GS_INPUT_PRIM
] = nir
->info
.gs
.input_primitive
;
394 info
->properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
] = nir
->info
.gs
.output_primitive
;
395 info
->properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
] = nir
->info
.gs
.vertices_out
;
396 info
->properties
[TGSI_PROPERTY_GS_INVOCATIONS
] = nir
->info
.gs
.invocations
;
399 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
400 info
->properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
] =
401 nir
->info
.fs
.early_fragment_tests
| nir
->info
.fs
.post_depth_coverage
;
402 info
->properties
[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE
] = nir
->info
.fs
.post_depth_coverage
;
404 if (nir
->info
.fs
.pixel_center_integer
) {
405 info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] =
406 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
;
409 if (nir
->info
.fs
.depth_layout
!= FRAG_DEPTH_LAYOUT_NONE
) {
410 switch (nir
->info
.fs
.depth_layout
) {
411 case FRAG_DEPTH_LAYOUT_ANY
:
412 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_ANY
;
414 case FRAG_DEPTH_LAYOUT_GREATER
:
415 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_GREATER
;
417 case FRAG_DEPTH_LAYOUT_LESS
:
418 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_LESS
;
420 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
421 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_UNCHANGED
;
424 unreachable("Unknow depth layout");
429 if (gl_shader_stage_is_compute(nir
->info
.stage
)) {
430 info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] = nir
->info
.cs
.local_size
[0];
431 info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] = nir
->info
.cs
.local_size
[1];
432 info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
] = nir
->info
.cs
.local_size
[2];
436 uint64_t processed_inputs
= 0;
437 unsigned num_inputs
= 0;
438 nir_foreach_variable(variable
, &nir
->inputs
) {
439 unsigned semantic_name
, semantic_index
;
441 const struct glsl_type
*type
= variable
->type
;
442 if (nir_is_per_vertex_io(variable
, nir
->info
.stage
)) {
443 assert(glsl_type_is_array(type
));
444 type
= glsl_get_array_element(type
);
447 unsigned attrib_count
= glsl_count_attribute_slots(type
,
448 nir
->info
.stage
== MESA_SHADER_VERTEX
);
450 i
= variable
->data
.driver_location
;
452 /* Vertex shader inputs don't have semantics. The state
453 * tracker has already mapped them to attributes via
454 * variable->data.driver_location.
456 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
457 if (glsl_type_is_dual_slot(glsl_without_array(variable
->type
)))
464 for (unsigned j
= 0; j
< attrib_count
; j
++, i
++) {
466 if (processed_inputs
& ((uint64_t)1 << i
))
469 processed_inputs
|= ((uint64_t)1 << i
);
472 tgsi_get_gl_varying_semantic(variable
->data
.location
+ j
, true,
473 &semantic_name
, &semantic_index
);
475 info
->input_semantic_name
[i
] = semantic_name
;
476 info
->input_semantic_index
[i
] = semantic_index
;
478 if (semantic_name
== TGSI_SEMANTIC_PRIMID
)
479 info
->uses_primid
= true;
481 enum glsl_base_type base_type
=
482 glsl_get_base_type(glsl_without_array(variable
->type
));
484 switch (variable
->data
.interpolation
) {
485 case INTERP_MODE_NONE
:
486 if (glsl_base_type_is_integer(base_type
)) {
487 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_CONSTANT
;
491 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
492 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_COLOR
;
497 case INTERP_MODE_SMOOTH
:
498 assert(!glsl_base_type_is_integer(base_type
));
500 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_PERSPECTIVE
;
503 case INTERP_MODE_NOPERSPECTIVE
:
504 assert(!glsl_base_type_is_integer(base_type
));
506 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_LINEAR
;
509 case INTERP_MODE_FLAT
:
510 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_CONSTANT
;
516 info
->num_inputs
= num_inputs
;
519 uint64_t processed_outputs
= 0;
520 unsigned num_outputs
= 0;
521 nir_foreach_variable(variable
, &nir
->outputs
) {
522 unsigned semantic_name
, semantic_index
;
524 i
= variable
->data
.driver_location
;
526 const struct glsl_type
*type
= variable
->type
;
527 if (nir_is_per_vertex_io(variable
, nir
->info
.stage
)) {
528 assert(glsl_type_is_array(type
));
529 type
= glsl_get_array_element(type
);
532 unsigned attrib_count
= glsl_count_attribute_slots(type
, false);
533 for (unsigned k
= 0; k
< attrib_count
; k
++, i
++) {
535 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
536 tgsi_get_gl_frag_result_semantic(variable
->data
.location
+ k
,
537 &semantic_name
, &semantic_index
);
539 /* Adjust for dual source blending */
540 if (variable
->data
.index
> 0) {
544 tgsi_get_gl_varying_semantic(variable
->data
.location
+ k
, true,
545 &semantic_name
, &semantic_index
);
548 unsigned num_components
= 4;
549 unsigned vector_elements
= glsl_get_vector_elements(glsl_without_array(variable
->type
));
551 num_components
= vector_elements
;
553 unsigned component
= variable
->data
.location_frac
;
554 if (glsl_type_is_64bit(glsl_without_array(variable
->type
))) {
555 if (glsl_type_is_dual_slot(glsl_without_array(variable
->type
)) && k
% 2) {
556 num_components
= (num_components
* 2) - 4;
559 num_components
= MIN2(num_components
* 2, 4);
564 for (unsigned j
= component
; j
< num_components
+ component
; j
++) {
567 usagemask
|= TGSI_WRITEMASK_X
;
570 usagemask
|= TGSI_WRITEMASK_Y
;
573 usagemask
|= TGSI_WRITEMASK_Z
;
576 usagemask
|= TGSI_WRITEMASK_W
;
579 unreachable("error calculating component index");
583 unsigned gs_out_streams
;
584 if (variable
->data
.stream
& (1u << 31)) {
585 gs_out_streams
= variable
->data
.stream
& ~(1u << 31);
587 assert(variable
->data
.stream
< 4);
589 for (unsigned j
= 0; j
< num_components
; ++j
)
590 gs_out_streams
|= variable
->data
.stream
<< (2 * (component
+ j
));
593 unsigned streamx
= gs_out_streams
& 3;
594 unsigned streamy
= (gs_out_streams
>> 2) & 3;
595 unsigned streamz
= (gs_out_streams
>> 4) & 3;
596 unsigned streamw
= (gs_out_streams
>> 6) & 3;
598 if (usagemask
& TGSI_WRITEMASK_X
) {
599 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_X
;
600 info
->output_streams
[i
] |= streamx
;
601 info
->num_stream_output_components
[streamx
]++;
603 if (usagemask
& TGSI_WRITEMASK_Y
) {
604 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_Y
;
605 info
->output_streams
[i
] |= streamy
<< 2;
606 info
->num_stream_output_components
[streamy
]++;
608 if (usagemask
& TGSI_WRITEMASK_Z
) {
609 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_Z
;
610 info
->output_streams
[i
] |= streamz
<< 4;
611 info
->num_stream_output_components
[streamz
]++;
613 if (usagemask
& TGSI_WRITEMASK_W
) {
614 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_W
;
615 info
->output_streams
[i
] |= streamw
<< 6;
616 info
->num_stream_output_components
[streamw
]++;
619 /* make sure we only count this location once against
620 * the num_outputs counter.
622 if (processed_outputs
& ((uint64_t)1 << i
))
625 processed_outputs
|= ((uint64_t)1 << i
);
628 info
->output_semantic_name
[i
] = semantic_name
;
629 info
->output_semantic_index
[i
] = semantic_index
;
631 switch (semantic_name
) {
632 case TGSI_SEMANTIC_PRIMID
:
633 info
->writes_primid
= true;
635 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
636 info
->writes_viewport_index
= true;
638 case TGSI_SEMANTIC_LAYER
:
639 info
->writes_layer
= true;
641 case TGSI_SEMANTIC_PSIZE
:
642 info
->writes_psize
= true;
644 case TGSI_SEMANTIC_CLIPVERTEX
:
645 info
->writes_clipvertex
= true;
647 case TGSI_SEMANTIC_COLOR
:
648 info
->colors_written
|= 1 << semantic_index
;
650 case TGSI_SEMANTIC_STENCIL
:
651 info
->writes_stencil
= true;
653 case TGSI_SEMANTIC_SAMPLEMASK
:
654 info
->writes_samplemask
= true;
656 case TGSI_SEMANTIC_EDGEFLAG
:
657 info
->writes_edgeflag
= true;
659 case TGSI_SEMANTIC_POSITION
:
660 if (info
->processor
== PIPE_SHADER_FRAGMENT
)
661 info
->writes_z
= true;
663 info
->writes_position
= true;
667 if (nir
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
668 switch (semantic_name
) {
669 case TGSI_SEMANTIC_PATCH
:
670 info
->reads_perpatch_outputs
= true;
672 case TGSI_SEMANTIC_TESSINNER
:
673 case TGSI_SEMANTIC_TESSOUTER
:
674 info
->reads_tessfactor_outputs
= true;
677 info
->reads_pervertex_outputs
= true;
682 unsigned loc
= variable
->data
.location
;
683 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&&
684 loc
== FRAG_RESULT_COLOR
&&
685 nir
->info
.outputs_written
& (1ull << loc
)) {
686 assert(attrib_count
== 1);
687 info
->properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
] = true;
691 info
->num_outputs
= num_outputs
;
693 struct set
*ubo_set
= _mesa_set_create(NULL
, _mesa_hash_pointer
,
694 _mesa_key_pointer_equal
);
695 struct set
*ssbo_set
= _mesa_set_create(NULL
, _mesa_hash_pointer
,
696 _mesa_key_pointer_equal
);
698 /* Intialise const_file_max[0] */
699 info
->const_file_max
[0] = -1;
701 /* The first 8 are reserved for atomic counters using ssbo */
702 unsigned ssbo_idx
= 8;
704 unsigned ubo_idx
= 1;
705 nir_foreach_variable(variable
, &nir
->uniforms
) {
706 const struct glsl_type
*type
= variable
->type
;
707 enum glsl_base_type base_type
=
708 glsl_get_base_type(glsl_without_array(type
));
709 unsigned aoa_size
= MAX2(1, glsl_get_aoa_size(type
));
710 unsigned loc
= variable
->data
.driver_location
/ 4;
711 int slot_count
= glsl_count_attribute_slots(type
, false);
712 int max_slot
= MAX2(info
->const_file_max
[0], (int) loc
) + slot_count
;
714 /* Gather buffers declared bitmasks. Note: radeonsi doesn't
715 * really use the mask (other than ubo_idx == 1 for regular
716 * uniforms) its really only used for getting the buffer count
717 * so we don't need to worry about the ordering.
719 if (variable
->interface_type
!= NULL
) {
720 if (variable
->data
.mode
== nir_var_uniform
||
721 variable
->data
.mode
== nir_var_mem_ubo
||
722 variable
->data
.mode
== nir_var_mem_ssbo
) {
724 struct set
*buf_set
= variable
->data
.mode
== nir_var_mem_ssbo
?
727 unsigned block_count
;
728 if (base_type
!= GLSL_TYPE_INTERFACE
) {
729 struct set_entry
*entry
=
730 _mesa_set_search(buf_set
, variable
->interface_type
);
732 /* Check if we have already processed
733 * a member from this ubo.
740 block_count
= aoa_size
;
743 if (variable
->data
.mode
== nir_var_uniform
||
744 variable
->data
.mode
== nir_var_mem_ubo
) {
745 info
->const_buffers_declared
|= u_bit_consecutive(ubo_idx
, block_count
);
746 ubo_idx
+= block_count
;
748 assert(variable
->data
.mode
== nir_var_mem_ssbo
);
750 info
->shader_buffers_declared
|= u_bit_consecutive(ssbo_idx
, block_count
);
751 ssbo_idx
+= block_count
;
754 _mesa_set_add(buf_set
, variable
->interface_type
);
760 /* We rely on the fact that nir_lower_samplers_as_deref has
761 * eliminated struct dereferences.
763 if (base_type
== GLSL_TYPE_SAMPLER
&& !variable
->data
.bindless
) {
764 info
->samplers_declared
|=
765 u_bit_consecutive(variable
->data
.binding
, aoa_size
);
766 } else if (base_type
== GLSL_TYPE_IMAGE
&& !variable
->data
.bindless
) {
767 info
->images_declared
|=
768 u_bit_consecutive(variable
->data
.binding
, aoa_size
);
769 } else if (base_type
!= GLSL_TYPE_ATOMIC_UINT
) {
770 info
->const_buffers_declared
|= 1;
771 info
->const_file_max
[0] = max_slot
;
775 _mesa_set_destroy(ubo_set
, NULL
);
776 _mesa_set_destroy(ssbo_set
, NULL
);
778 info
->num_written_clipdistance
= nir
->info
.clip_distance_array_size
;
779 info
->num_written_culldistance
= nir
->info
.cull_distance_array_size
;
780 info
->clipdist_writemask
= u_bit_consecutive(0, info
->num_written_clipdistance
);
781 info
->culldist_writemask
= u_bit_consecutive(0, info
->num_written_culldistance
);
783 if (info
->processor
== PIPE_SHADER_FRAGMENT
)
784 info
->uses_kill
= nir
->info
.fs
.uses_discard
;
786 func
= (struct nir_function
*)exec_list_get_head_const(&nir
->functions
);
787 nir_foreach_block(block
, func
->impl
) {
788 nir_foreach_instr(instr
, block
)
789 scan_instruction(nir
, info
, instr
);
794 si_nir_opts(struct nir_shader
*nir
)
797 unsigned lower_flrp
=
798 (nir
->options
->lower_flrp16
? 16 : 0) |
799 (nir
->options
->lower_flrp32
? 32 : 0) |
800 (nir
->options
->lower_flrp64
? 64 : 0);
805 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
807 NIR_PASS(progress
, nir
, nir_opt_copy_prop_vars
);
808 NIR_PASS(progress
, nir
, nir_opt_dead_write_vars
);
810 NIR_PASS_V(nir
, nir_lower_alu_to_scalar
, NULL
);
811 NIR_PASS_V(nir
, nir_lower_phis_to_scalar
);
813 /* (Constant) copy propagation is needed for txf with offsets. */
814 NIR_PASS(progress
, nir
, nir_copy_prop
);
815 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
816 NIR_PASS(progress
, nir
, nir_opt_dce
);
817 if (nir_opt_trivial_continues(nir
)) {
819 NIR_PASS(progress
, nir
, nir_copy_prop
);
820 NIR_PASS(progress
, nir
, nir_opt_dce
);
822 NIR_PASS(progress
, nir
, nir_opt_if
, true);
823 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
824 NIR_PASS(progress
, nir
, nir_opt_cse
);
825 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 8, true, true);
827 /* Needed for algebraic lowering */
828 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
829 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
831 if (lower_flrp
!= 0) {
832 bool lower_flrp_progress
= false;
834 NIR_PASS(lower_flrp_progress
, nir
, nir_lower_flrp
,
836 false /* always_precise */,
837 nir
->options
->lower_ffma
);
838 if (lower_flrp_progress
) {
839 NIR_PASS(progress
, nir
,
840 nir_opt_constant_folding
);
844 /* Nothing should rematerialize any flrps, so we only
845 * need to do this lowering once.
850 NIR_PASS(progress
, nir
, nir_opt_undef
);
851 NIR_PASS(progress
, nir
, nir_opt_conditional_discard
);
852 if (nir
->options
->max_unroll_iterations
) {
853 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
, 0);
859 type_size_vec4(const struct glsl_type
*type
, bool bindless
)
861 return glsl_count_attribute_slots(type
, false);
865 si_nir_lower_color(nir_shader
*nir
)
867 nir_function_impl
*entrypoint
= nir_shader_get_entrypoint(nir
);
870 nir_builder_init(&b
, entrypoint
);
872 nir_foreach_block(block
, entrypoint
) {
873 nir_foreach_instr_safe(instr
, block
) {
874 if (instr
->type
!= nir_instr_type_intrinsic
)
877 nir_intrinsic_instr
*intrin
=
878 nir_instr_as_intrinsic(instr
);
880 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
)
883 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
884 if (deref
->mode
!= nir_var_shader_in
)
887 b
.cursor
= nir_before_instr(instr
);
888 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
891 if (var
->data
.location
== VARYING_SLOT_COL0
) {
892 def
= nir_load_color0(&b
);
893 } else if (var
->data
.location
== VARYING_SLOT_COL1
) {
894 def
= nir_load_color1(&b
);
899 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
, nir_src_for_ssa(def
));
900 nir_instr_remove(instr
);
906 * Perform "lowering" operations on the NIR that are run once when the shader
907 * selector is created.
910 si_lower_nir(struct si_shader_selector
* sel
, unsigned wave_size
)
912 /* Adjust the driver location of inputs and outputs. The state tracker
913 * interprets them as slots, while the ac/nir backend interprets them
914 * as individual components.
916 if (sel
->nir
->info
.stage
!= MESA_SHADER_FRAGMENT
) {
917 nir_foreach_variable(variable
, &sel
->nir
->inputs
)
918 variable
->data
.driver_location
*= 4;
920 NIR_PASS_V(sel
->nir
, nir_lower_io_to_temporaries
,
921 nir_shader_get_entrypoint(sel
->nir
), false, true);
923 /* Since we're doing nir_lower_io_to_temporaries late, we need
924 * to lower all the copy_deref's introduced by
925 * lower_io_to_temporaries before calling nir_lower_io.
927 NIR_PASS_V(sel
->nir
, nir_split_var_copies
);
928 NIR_PASS_V(sel
->nir
, nir_lower_var_copies
);
929 NIR_PASS_V(sel
->nir
, nir_lower_global_vars_to_local
);
931 si_nir_lower_color(sel
->nir
);
932 NIR_PASS_V(sel
->nir
, nir_lower_io
, nir_var_shader_in
, type_size_vec4
, 0);
934 /* This pass needs actual constants */
935 NIR_PASS_V(sel
->nir
, nir_opt_constant_folding
);
936 NIR_PASS_V(sel
->nir
, nir_io_add_const_offset_to_base
,
940 nir_foreach_variable(variable
, &sel
->nir
->outputs
) {
941 variable
->data
.driver_location
*= 4;
943 if (sel
->nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
944 if (variable
->data
.location
== FRAG_RESULT_DEPTH
)
945 variable
->data
.driver_location
+= 2;
946 else if (variable
->data
.location
== FRAG_RESULT_STENCIL
)
947 variable
->data
.driver_location
+= 1;
951 /* Perform lowerings (and optimizations) of code.
953 * Performance considerations aside, we must:
954 * - lower certain ALU operations
955 * - ensure constant offsets for texture instructions are folded
956 * and copy-propagated
959 static const struct nir_lower_tex_options lower_tex_options
= {
962 NIR_PASS_V(sel
->nir
, nir_lower_tex
, &lower_tex_options
);
964 const nir_lower_subgroups_options subgroups_options
= {
965 .subgroup_size
= wave_size
,
966 .ballot_bit_size
= wave_size
,
967 .lower_to_scalar
= true,
968 .lower_subgroup_masks
= true,
969 .lower_vote_trivial
= false,
970 .lower_vote_eq_to_ballot
= true,
972 NIR_PASS_V(sel
->nir
, nir_lower_subgroups
, &subgroups_options
);
974 ac_lower_indirect_derefs(sel
->nir
, sel
->screen
->info
.chip_class
);
976 si_nir_opts(sel
->nir
);
978 NIR_PASS_V(sel
->nir
, nir_lower_bool_to_int32
);
980 /* Strip the resulting shader so that the shader cache is more likely
981 * to hit from other similar shaders.
986 static void declare_nir_input_vs(struct si_shader_context
*ctx
,
987 struct nir_variable
*variable
,
988 unsigned input_index
,
991 si_llvm_load_input_vs(ctx
, input_index
, out
);
995 si_nir_lookup_interp_param(struct ac_shader_abi
*abi
,
996 enum glsl_interp_mode interp
, unsigned location
)
998 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
999 int interp_param_idx
= -1;
1002 case INTERP_MODE_FLAT
:
1004 case INTERP_MODE_SMOOTH
:
1005 case INTERP_MODE_NONE
:
1006 if (location
== INTERP_CENTER
)
1007 interp_param_idx
= SI_PARAM_PERSP_CENTER
;
1008 else if (location
== INTERP_CENTROID
)
1009 interp_param_idx
= SI_PARAM_PERSP_CENTROID
;
1010 else if (location
== INTERP_SAMPLE
)
1011 interp_param_idx
= SI_PARAM_PERSP_SAMPLE
;
1013 case INTERP_MODE_NOPERSPECTIVE
:
1014 if (location
== INTERP_CENTER
)
1015 interp_param_idx
= SI_PARAM_LINEAR_CENTER
;
1016 else if (location
== INTERP_CENTROID
)
1017 interp_param_idx
= SI_PARAM_LINEAR_CENTROID
;
1018 else if (location
== INTERP_SAMPLE
)
1019 interp_param_idx
= SI_PARAM_LINEAR_SAMPLE
;
1022 assert(!"Unhandled interpolation mode.");
1026 return interp_param_idx
!= -1 ?
1027 LLVMGetParam(ctx
->main_fn
, interp_param_idx
) : NULL
;
1031 si_nir_load_sampler_desc(struct ac_shader_abi
*abi
,
1032 unsigned descriptor_set
, unsigned base_index
,
1033 unsigned constant_index
, LLVMValueRef dynamic_index
,
1034 enum ac_descriptor_type desc_type
, bool image
,
1035 bool write
, bool bindless
)
1037 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
1038 LLVMBuilderRef builder
= ctx
->ac
.builder
;
1039 unsigned const_index
= base_index
+ constant_index
;
1041 assert(!descriptor_set
);
1042 assert(!image
|| desc_type
== AC_DESC_IMAGE
|| desc_type
== AC_DESC_BUFFER
);
1046 LLVMGetParam(ctx
->main_fn
, ctx
->param_bindless_samplers_and_images
);
1048 /* dynamic_index is the bindless handle */
1050 /* For simplicity, bindless image descriptors use fixed
1051 * 16-dword slots for now.
1053 dynamic_index
= LLVMBuildMul(ctx
->ac
.builder
, dynamic_index
,
1054 LLVMConstInt(ctx
->i64
, 2, 0), "");
1056 return si_load_image_desc(ctx
, list
, dynamic_index
, desc_type
,
1060 /* Since bindless handle arithmetic can contain an unsigned integer
1061 * wraparound and si_load_sampler_desc assumes there isn't any,
1062 * use GEP without "inbounds" (inside ac_build_pointer_add)
1063 * to prevent incorrect code generation and hangs.
1065 dynamic_index
= LLVMBuildMul(ctx
->ac
.builder
, dynamic_index
,
1066 LLVMConstInt(ctx
->i64
, 2, 0), "");
1067 list
= ac_build_pointer_add(&ctx
->ac
, list
, dynamic_index
);
1068 return si_load_sampler_desc(ctx
, list
, ctx
->i32_0
, desc_type
);
1071 unsigned num_slots
= image
? ctx
->num_images
: ctx
->num_samplers
;
1072 assert(const_index
< num_slots
);
1074 LLVMValueRef list
= LLVMGetParam(ctx
->main_fn
, ctx
->param_samplers_and_images
);
1075 LLVMValueRef index
= LLVMConstInt(ctx
->ac
.i32
, const_index
, false);
1077 if (dynamic_index
) {
1078 index
= LLVMBuildAdd(builder
, index
, dynamic_index
, "");
1080 /* From the GL_ARB_shader_image_load_store extension spec:
1082 * If a shader performs an image load, store, or atomic
1083 * operation using an image variable declared as an array,
1084 * and if the index used to select an individual element is
1085 * negative or greater than or equal to the size of the
1086 * array, the results of the operation are undefined but may
1087 * not lead to termination.
1089 index
= si_llvm_bound_index(ctx
, index
, num_slots
);
1093 index
= LLVMBuildSub(ctx
->ac
.builder
,
1094 LLVMConstInt(ctx
->i32
, SI_NUM_IMAGES
- 1, 0),
1096 return si_load_image_desc(ctx
, list
, index
, desc_type
, write
, false);
1099 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
1100 LLVMConstInt(ctx
->i32
, SI_NUM_IMAGES
/ 2, 0), "");
1101 return si_load_sampler_desc(ctx
, list
, index
, desc_type
);
1104 static void bitcast_inputs(struct si_shader_context
*ctx
,
1105 LLVMValueRef data
[4],
1108 for (unsigned chan
= 0; chan
< 4; chan
++) {
1109 ctx
->inputs
[input_idx
+ chan
] =
1110 LLVMBuildBitCast(ctx
->ac
.builder
, data
[chan
], ctx
->ac
.i32
, "");
1114 bool si_nir_build_llvm(struct si_shader_context
*ctx
, struct nir_shader
*nir
)
1116 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1118 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
1119 uint64_t processed_inputs
= 0;
1120 nir_foreach_variable(variable
, &nir
->inputs
) {
1121 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
,
1123 unsigned input_idx
= variable
->data
.driver_location
;
1125 LLVMValueRef data
[4];
1126 unsigned loc
= variable
->data
.location
;
1128 for (unsigned i
= 0; i
< attrib_count
; i
++) {
1129 /* Packed components share the same location so skip
1130 * them if we have already processed the location.
1132 if (processed_inputs
& ((uint64_t)1 << (loc
+ i
))) {
1137 declare_nir_input_vs(ctx
, variable
, input_idx
/ 4, data
);
1138 bitcast_inputs(ctx
, data
, input_idx
);
1139 if (glsl_type_is_dual_slot(variable
->type
)) {
1141 declare_nir_input_vs(ctx
, variable
, input_idx
/ 4, data
);
1142 bitcast_inputs(ctx
, data
, input_idx
);
1145 processed_inputs
|= ((uint64_t)1 << (loc
+ i
));
1149 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1150 unsigned colors_read
=
1151 ctx
->shader
->selector
->info
.colors_read
;
1152 LLVMValueRef main_fn
= ctx
->main_fn
;
1154 LLVMValueRef undef
= LLVMGetUndef(ctx
->f32
);
1156 unsigned offset
= SI_PARAM_POS_FIXED_PT
+ 1;
1158 if (colors_read
& 0x0f) {
1159 unsigned mask
= colors_read
& 0x0f;
1160 LLVMValueRef values
[4];
1161 values
[0] = mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1162 values
[1] = mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1163 values
[2] = mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1164 values
[3] = mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1166 ac_to_integer(&ctx
->ac
,
1167 ac_build_gather_values(&ctx
->ac
, values
, 4));
1169 if (colors_read
& 0xf0) {
1170 unsigned mask
= (colors_read
& 0xf0) >> 4;
1171 LLVMValueRef values
[4];
1172 values
[0] = mask
& 0x1 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1173 values
[1] = mask
& 0x2 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1174 values
[2] = mask
& 0x4 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1175 values
[3] = mask
& 0x8 ? LLVMGetParam(main_fn
, offset
++) : undef
;
1177 ac_to_integer(&ctx
->ac
,
1178 ac_build_gather_values(&ctx
->ac
, values
, 4));
1182 ctx
->abi
.inputs
= &ctx
->inputs
[0];
1183 ctx
->abi
.load_sampler_desc
= si_nir_load_sampler_desc
;
1184 ctx
->abi
.clamp_shadow_reference
= true;
1186 ctx
->num_samplers
= util_last_bit(info
->samplers_declared
);
1187 ctx
->num_images
= util_last_bit(info
->images_declared
);
1189 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_CS_LOCAL_SIZE
]) {
1190 assert(gl_shader_stage_is_compute(nir
->info
.stage
));
1191 si_declare_compute_memory(ctx
);
1193 ac_nir_translate(&ctx
->ac
, &ctx
->abi
, nir
);