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"
35 static nir_variable
* tex_get_texture_var(nir_tex_instr
*instr
)
37 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
38 switch (instr
->src
[i
].src_type
) {
39 case nir_tex_src_texture_deref
:
40 return nir_deref_instr_get_variable(nir_src_as_deref(instr
->src
[i
].src
));
49 static nir_variable
* intrinsic_get_var(nir_intrinsic_instr
*instr
)
51 return nir_deref_instr_get_variable(nir_src_as_deref(instr
->src
[0]));
54 static void gather_intrinsic_load_deref_info(const nir_shader
*nir
,
55 const nir_intrinsic_instr
*instr
,
57 struct tgsi_shader_info
*info
)
59 assert(var
&& var
->data
.mode
== nir_var_shader_in
);
61 switch (nir
->info
.stage
) {
62 case MESA_SHADER_VERTEX
: {
63 unsigned i
= var
->data
.driver_location
;
64 unsigned attrib_count
= glsl_count_attribute_slots(var
->type
, false);
66 for (unsigned j
= 0; j
< attrib_count
; j
++, i
++) {
67 if (glsl_type_is_64bit(glsl_without_array(var
->type
))) {
68 /* TODO: set usage mask more accurately for doubles */
69 info
->input_usage_mask
[i
] = TGSI_WRITEMASK_XYZW
;
71 uint8_t mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
72 info
->input_usage_mask
[i
] |= mask
<< var
->data
.location_frac
;
78 unsigned semantic_name
, semantic_index
;
79 tgsi_get_gl_varying_semantic(var
->data
.location
, true,
80 &semantic_name
, &semantic_index
);
82 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
83 uint8_t mask
= nir_ssa_def_components_read(&instr
->dest
.ssa
);
84 info
->colors_read
|= mask
<< (semantic_index
* 4);
91 static void scan_instruction(const struct nir_shader
*nir
,
92 struct tgsi_shader_info
*info
,
95 if (instr
->type
== nir_instr_type_alu
) {
96 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
101 case nir_op_fddx_fine
:
102 case nir_op_fddy_fine
:
103 case nir_op_fddx_coarse
:
104 case nir_op_fddy_coarse
:
105 info
->uses_derivatives
= true;
110 } else if (instr
->type
== nir_instr_type_tex
) {
111 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
112 nir_variable
*texture
= tex_get_texture_var(tex
);
115 info
->samplers_declared
|=
116 u_bit_consecutive(tex
->sampler_index
, 1);
118 if (texture
->data
.bindless
)
119 info
->uses_bindless_samplers
= true;
126 info
->uses_derivatives
= true;
131 } else if (instr
->type
== nir_instr_type_intrinsic
) {
132 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
134 switch (intr
->intrinsic
) {
135 case nir_intrinsic_load_front_face
:
136 info
->uses_frontface
= 1;
138 case nir_intrinsic_load_instance_id
:
139 info
->uses_instanceid
= 1;
141 case nir_intrinsic_load_invocation_id
:
142 info
->uses_invocationid
= true;
144 case nir_intrinsic_load_num_work_groups
:
145 info
->uses_grid_size
= true;
147 case nir_intrinsic_load_local_group_size
:
148 /* The block size is translated to IMM with a fixed block size. */
149 if (info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] == 0)
150 info
->uses_block_size
= true;
152 case nir_intrinsic_load_local_invocation_id
:
153 case nir_intrinsic_load_work_group_id
: {
154 unsigned mask
= nir_ssa_def_components_read(&intr
->dest
.ssa
);
156 unsigned i
= u_bit_scan(&mask
);
158 if (intr
->intrinsic
== nir_intrinsic_load_work_group_id
)
159 info
->uses_block_id
[i
] = true;
161 info
->uses_thread_id
[i
] = true;
165 case nir_intrinsic_load_vertex_id
:
166 info
->uses_vertexid
= 1;
168 case nir_intrinsic_load_vertex_id_zero_base
:
169 info
->uses_vertexid_nobase
= 1;
171 case nir_intrinsic_load_base_vertex
:
172 info
->uses_basevertex
= 1;
174 case nir_intrinsic_load_primitive_id
:
175 info
->uses_primid
= 1;
177 case nir_intrinsic_load_sample_mask_in
:
178 info
->reads_samplemask
= true;
180 case nir_intrinsic_load_tess_level_inner
:
181 case nir_intrinsic_load_tess_level_outer
:
182 info
->reads_tess_factors
= true;
184 case nir_intrinsic_image_deref_load
: {
185 nir_variable
*var
= intrinsic_get_var(intr
);
186 if (var
->data
.bindless
) {
187 info
->uses_bindless_images
= true;
189 if (glsl_get_sampler_dim(var
->type
) == GLSL_SAMPLER_DIM_BUF
)
190 info
->uses_bindless_buffer_load
= true;
192 info
->uses_bindless_image_load
= true;
196 case nir_intrinsic_image_deref_size
:
197 case nir_intrinsic_image_deref_samples
: {
198 nir_variable
*var
= intrinsic_get_var(intr
);
199 if (var
->data
.bindless
)
200 info
->uses_bindless_images
= true;
203 case nir_intrinsic_image_deref_store
: {
204 const nir_deref_instr
*image_deref
= nir_instr_as_deref(intr
->src
[0].ssa
->parent_instr
);
205 nir_variable
*var
= intrinsic_get_var(intr
);
206 if (var
->data
.bindless
) {
207 info
->uses_bindless_images
= true;
209 if (glsl_get_sampler_dim(image_deref
->type
) == GLSL_SAMPLER_DIM_BUF
)
210 info
->uses_bindless_buffer_store
= true;
212 info
->uses_bindless_image_store
= true;
214 info
->writes_memory
= true;
217 case nir_intrinsic_image_deref_atomic_add
:
218 case nir_intrinsic_image_deref_atomic_min
:
219 case nir_intrinsic_image_deref_atomic_max
:
220 case nir_intrinsic_image_deref_atomic_and
:
221 case nir_intrinsic_image_deref_atomic_or
:
222 case nir_intrinsic_image_deref_atomic_xor
:
223 case nir_intrinsic_image_deref_atomic_exchange
:
224 case nir_intrinsic_image_deref_atomic_comp_swap
: {
225 nir_variable
*var
= intrinsic_get_var(intr
);
226 if (var
->data
.bindless
) {
227 info
->uses_bindless_images
= true;
229 if (glsl_get_sampler_dim(var
->type
) == GLSL_SAMPLER_DIM_BUF
)
230 info
->uses_bindless_buffer_atomic
= true;
232 info
->uses_bindless_image_atomic
= true;
234 info
->writes_memory
= true;
237 case nir_intrinsic_store_ssbo
:
238 case nir_intrinsic_ssbo_atomic_add
:
239 case nir_intrinsic_ssbo_atomic_imin
:
240 case nir_intrinsic_ssbo_atomic_umin
:
241 case nir_intrinsic_ssbo_atomic_imax
:
242 case nir_intrinsic_ssbo_atomic_umax
:
243 case nir_intrinsic_ssbo_atomic_and
:
244 case nir_intrinsic_ssbo_atomic_or
:
245 case nir_intrinsic_ssbo_atomic_xor
:
246 case nir_intrinsic_ssbo_atomic_exchange
:
247 case nir_intrinsic_ssbo_atomic_comp_swap
:
248 info
->writes_memory
= true;
250 case nir_intrinsic_load_deref
: {
251 nir_variable
*var
= intrinsic_get_var(intr
);
252 nir_variable_mode mode
= var
->data
.mode
;
253 enum glsl_base_type base_type
=
254 glsl_get_base_type(glsl_without_array(var
->type
));
256 if (mode
== nir_var_shader_in
) {
257 gather_intrinsic_load_deref_info(nir
, intr
, var
, info
);
259 switch (var
->data
.interpolation
) {
260 case INTERP_MODE_NONE
:
261 if (glsl_base_type_is_integer(base_type
))
265 case INTERP_MODE_SMOOTH
:
266 if (var
->data
.sample
)
267 info
->uses_persp_sample
= true;
268 else if (var
->data
.centroid
)
269 info
->uses_persp_centroid
= true;
271 info
->uses_persp_center
= true;
274 case INTERP_MODE_NOPERSPECTIVE
:
275 if (var
->data
.sample
)
276 info
->uses_linear_sample
= true;
277 else if (var
->data
.centroid
)
278 info
->uses_linear_centroid
= true;
280 info
->uses_linear_center
= true;
286 case nir_intrinsic_interp_deref_at_centroid
:
287 case nir_intrinsic_interp_deref_at_sample
:
288 case nir_intrinsic_interp_deref_at_offset
: {
289 enum glsl_interp_mode interp
= intrinsic_get_var(intr
)->data
.interpolation
;
291 case INTERP_MODE_SMOOTH
:
292 case INTERP_MODE_NONE
:
293 if (intr
->intrinsic
== nir_intrinsic_interp_deref_at_centroid
)
294 info
->uses_persp_opcode_interp_centroid
= true;
295 else if (intr
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
296 info
->uses_persp_opcode_interp_sample
= true;
298 info
->uses_persp_opcode_interp_offset
= true;
300 case INTERP_MODE_NOPERSPECTIVE
:
301 if (intr
->intrinsic
== nir_intrinsic_interp_deref_at_centroid
)
302 info
->uses_linear_opcode_interp_centroid
= true;
303 else if (intr
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
304 info
->uses_linear_opcode_interp_sample
= true;
306 info
->uses_linear_opcode_interp_offset
= true;
308 case INTERP_MODE_FLAT
:
311 unreachable("Unsupported interpoation type");
321 void si_nir_scan_tess_ctrl(const struct nir_shader
*nir
,
322 struct tgsi_tessctrl_info
*out
)
324 memset(out
, 0, sizeof(*out
));
326 if (nir
->info
.stage
!= MESA_SHADER_TESS_CTRL
)
329 out
->tessfactors_are_def_in_all_invocs
=
330 ac_are_tessfactors_def_in_all_invocs(nir
);
333 void si_nir_scan_shader(const struct nir_shader
*nir
,
334 struct tgsi_shader_info
*info
)
339 info
->processor
= pipe_shader_type_from_mesa(nir
->info
.stage
);
340 info
->num_tokens
= 2; /* indicate that the shader is non-empty */
341 info
->num_instructions
= 2;
343 info
->properties
[TGSI_PROPERTY_NEXT_SHADER
] =
344 pipe_shader_type_from_mesa(nir
->info
.next_stage
);
346 if (nir
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
347 info
->properties
[TGSI_PROPERTY_TCS_VERTICES_OUT
] =
348 nir
->info
.tess
.tcs_vertices_out
;
351 if (nir
->info
.stage
== MESA_SHADER_TESS_EVAL
) {
352 if (nir
->info
.tess
.primitive_mode
== GL_ISOLINES
)
353 info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] = PIPE_PRIM_LINES
;
355 info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
] = nir
->info
.tess
.primitive_mode
;
357 STATIC_ASSERT((TESS_SPACING_EQUAL
+ 1) % 3 == PIPE_TESS_SPACING_EQUAL
);
358 STATIC_ASSERT((TESS_SPACING_FRACTIONAL_ODD
+ 1) % 3 ==
359 PIPE_TESS_SPACING_FRACTIONAL_ODD
);
360 STATIC_ASSERT((TESS_SPACING_FRACTIONAL_EVEN
+ 1) % 3 ==
361 PIPE_TESS_SPACING_FRACTIONAL_EVEN
);
363 info
->properties
[TGSI_PROPERTY_TES_SPACING
] = (nir
->info
.tess
.spacing
+ 1) % 3;
364 info
->properties
[TGSI_PROPERTY_TES_VERTEX_ORDER_CW
] = !nir
->info
.tess
.ccw
;
365 info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
] = nir
->info
.tess
.point_mode
;
368 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
369 info
->properties
[TGSI_PROPERTY_GS_INPUT_PRIM
] = nir
->info
.gs
.input_primitive
;
370 info
->properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
] = nir
->info
.gs
.output_primitive
;
371 info
->properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
] = nir
->info
.gs
.vertices_out
;
372 info
->properties
[TGSI_PROPERTY_GS_INVOCATIONS
] = nir
->info
.gs
.invocations
;
375 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
376 info
->properties
[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL
] =
377 nir
->info
.fs
.early_fragment_tests
| nir
->info
.fs
.post_depth_coverage
;
378 info
->properties
[TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE
] = nir
->info
.fs
.post_depth_coverage
;
380 if (nir
->info
.fs
.pixel_center_integer
) {
381 info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] =
382 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
;
385 if (nir
->info
.fs
.depth_layout
!= FRAG_DEPTH_LAYOUT_NONE
) {
386 switch (nir
->info
.fs
.depth_layout
) {
387 case FRAG_DEPTH_LAYOUT_ANY
:
388 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_ANY
;
390 case FRAG_DEPTH_LAYOUT_GREATER
:
391 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_GREATER
;
393 case FRAG_DEPTH_LAYOUT_LESS
:
394 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_LESS
;
396 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
397 info
->properties
[TGSI_PROPERTY_FS_DEPTH_LAYOUT
] = TGSI_FS_DEPTH_LAYOUT_UNCHANGED
;
400 unreachable("Unknow depth layout");
405 if (gl_shader_stage_is_compute(nir
->info
.stage
)) {
406 info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH
] = nir
->info
.cs
.local_size
[0];
407 info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT
] = nir
->info
.cs
.local_size
[1];
408 info
->properties
[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH
] = nir
->info
.cs
.local_size
[2];
412 uint64_t processed_inputs
= 0;
413 unsigned num_inputs
= 0;
414 nir_foreach_variable(variable
, &nir
->inputs
) {
415 unsigned semantic_name
, semantic_index
;
417 const struct glsl_type
*type
= variable
->type
;
418 if (nir_is_per_vertex_io(variable
, nir
->info
.stage
)) {
419 assert(glsl_type_is_array(type
));
420 type
= glsl_get_array_element(type
);
423 unsigned attrib_count
= glsl_count_attribute_slots(type
,
424 nir
->info
.stage
== MESA_SHADER_VERTEX
);
426 i
= variable
->data
.driver_location
;
428 /* Vertex shader inputs don't have semantics. The state
429 * tracker has already mapped them to attributes via
430 * variable->data.driver_location.
432 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
433 if (glsl_type_is_dual_slot(glsl_without_array(variable
->type
)))
440 /* Fragment shader position is a system value. */
441 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&&
442 variable
->data
.location
== VARYING_SLOT_POS
) {
443 if (variable
->data
.pixel_center_integer
)
444 info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] =
445 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
;
451 for (unsigned j
= 0; j
< attrib_count
; j
++, i
++) {
453 if (processed_inputs
& ((uint64_t)1 << i
))
456 processed_inputs
|= ((uint64_t)1 << i
);
459 tgsi_get_gl_varying_semantic(variable
->data
.location
+ j
, true,
460 &semantic_name
, &semantic_index
);
462 info
->input_semantic_name
[i
] = semantic_name
;
463 info
->input_semantic_index
[i
] = semantic_index
;
465 if (semantic_name
== TGSI_SEMANTIC_PRIMID
)
466 info
->uses_primid
= true;
468 if (variable
->data
.sample
)
469 info
->input_interpolate_loc
[i
] = TGSI_INTERPOLATE_LOC_SAMPLE
;
470 else if (variable
->data
.centroid
)
471 info
->input_interpolate_loc
[i
] = TGSI_INTERPOLATE_LOC_CENTROID
;
473 info
->input_interpolate_loc
[i
] = TGSI_INTERPOLATE_LOC_CENTER
;
475 enum glsl_base_type base_type
=
476 glsl_get_base_type(glsl_without_array(variable
->type
));
478 switch (variable
->data
.interpolation
) {
479 case INTERP_MODE_NONE
:
480 if (glsl_base_type_is_integer(base_type
)) {
481 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_CONSTANT
;
485 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
486 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_COLOR
;
491 case INTERP_MODE_SMOOTH
:
492 assert(!glsl_base_type_is_integer(base_type
));
494 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_PERSPECTIVE
;
497 case INTERP_MODE_NOPERSPECTIVE
:
498 assert(!glsl_base_type_is_integer(base_type
));
500 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_LINEAR
;
503 case INTERP_MODE_FLAT
:
504 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_CONSTANT
;
510 info
->num_inputs
= num_inputs
;
514 uint64_t processed_outputs
= 0;
515 unsigned num_outputs
= 0;
516 nir_foreach_variable(variable
, &nir
->outputs
) {
517 unsigned semantic_name
, semantic_index
;
519 i
= variable
->data
.driver_location
;
521 const struct glsl_type
*type
= variable
->type
;
522 if (nir_is_per_vertex_io(variable
, nir
->info
.stage
)) {
523 assert(glsl_type_is_array(type
));
524 type
= glsl_get_array_element(type
);
527 unsigned attrib_count
= glsl_count_attribute_slots(type
, false);
528 for (unsigned k
= 0; k
< attrib_count
; k
++, i
++) {
530 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
531 tgsi_get_gl_frag_result_semantic(variable
->data
.location
+ k
,
532 &semantic_name
, &semantic_index
);
534 /* Adjust for dual source blending */
535 if (variable
->data
.index
> 0) {
539 tgsi_get_gl_varying_semantic(variable
->data
.location
+ k
, true,
540 &semantic_name
, &semantic_index
);
543 unsigned num_components
= 4;
544 unsigned vector_elements
= glsl_get_vector_elements(glsl_without_array(variable
->type
));
546 num_components
= vector_elements
;
548 unsigned component
= variable
->data
.location_frac
;
549 if (glsl_type_is_64bit(glsl_without_array(variable
->type
))) {
550 if (glsl_type_is_dual_slot(glsl_without_array(variable
->type
)) && k
% 2) {
551 num_components
= (num_components
* 2) - 4;
554 num_components
= MIN2(num_components
* 2, 4);
559 for (unsigned j
= component
; j
< num_components
+ component
; j
++) {
562 usagemask
|= TGSI_WRITEMASK_X
;
565 usagemask
|= TGSI_WRITEMASK_Y
;
568 usagemask
|= TGSI_WRITEMASK_Z
;
571 usagemask
|= TGSI_WRITEMASK_W
;
574 unreachable("error calculating component index");
578 unsigned gs_out_streams
;
579 if (variable
->data
.stream
& (1u << 31)) {
580 gs_out_streams
= variable
->data
.stream
& ~(1u << 31);
582 assert(variable
->data
.stream
< 4);
584 for (unsigned j
= 0; j
< num_components
; ++j
)
585 gs_out_streams
|= variable
->data
.stream
<< (2 * (component
+ j
));
588 unsigned streamx
= gs_out_streams
& 3;
589 unsigned streamy
= (gs_out_streams
>> 2) & 3;
590 unsigned streamz
= (gs_out_streams
>> 4) & 3;
591 unsigned streamw
= (gs_out_streams
>> 6) & 3;
593 if (usagemask
& TGSI_WRITEMASK_X
) {
594 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_X
;
595 info
->output_streams
[i
] |= streamx
;
596 info
->num_stream_output_components
[streamx
]++;
598 if (usagemask
& TGSI_WRITEMASK_Y
) {
599 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_Y
;
600 info
->output_streams
[i
] |= streamy
<< 2;
601 info
->num_stream_output_components
[streamy
]++;
603 if (usagemask
& TGSI_WRITEMASK_Z
) {
604 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_Z
;
605 info
->output_streams
[i
] |= streamz
<< 4;
606 info
->num_stream_output_components
[streamz
]++;
608 if (usagemask
& TGSI_WRITEMASK_W
) {
609 info
->output_usagemask
[i
] |= TGSI_WRITEMASK_W
;
610 info
->output_streams
[i
] |= streamw
<< 6;
611 info
->num_stream_output_components
[streamw
]++;
614 /* make sure we only count this location once against
615 * the num_outputs counter.
617 if (processed_outputs
& ((uint64_t)1 << i
))
620 processed_outputs
|= ((uint64_t)1 << i
);
623 info
->output_semantic_name
[i
] = semantic_name
;
624 info
->output_semantic_index
[i
] = semantic_index
;
626 switch (semantic_name
) {
627 case TGSI_SEMANTIC_PRIMID
:
628 info
->writes_primid
= true;
630 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
631 info
->writes_viewport_index
= true;
633 case TGSI_SEMANTIC_LAYER
:
634 info
->writes_layer
= true;
636 case TGSI_SEMANTIC_PSIZE
:
637 info
->writes_psize
= true;
639 case TGSI_SEMANTIC_CLIPVERTEX
:
640 info
->writes_clipvertex
= true;
642 case TGSI_SEMANTIC_COLOR
:
643 info
->colors_written
|= 1 << semantic_index
;
645 case TGSI_SEMANTIC_STENCIL
:
646 info
->writes_stencil
= true;
648 case TGSI_SEMANTIC_SAMPLEMASK
:
649 info
->writes_samplemask
= true;
651 case TGSI_SEMANTIC_EDGEFLAG
:
652 info
->writes_edgeflag
= true;
654 case TGSI_SEMANTIC_POSITION
:
655 if (info
->processor
== PIPE_SHADER_FRAGMENT
)
656 info
->writes_z
= true;
658 info
->writes_position
= true;
662 if (nir
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
663 switch (semantic_name
) {
664 case TGSI_SEMANTIC_PATCH
:
665 info
->reads_perpatch_outputs
= true;
667 case TGSI_SEMANTIC_TESSINNER
:
668 case TGSI_SEMANTIC_TESSOUTER
:
669 info
->reads_tessfactor_outputs
= true;
672 info
->reads_pervertex_outputs
= true;
677 unsigned loc
= variable
->data
.location
;
678 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&&
679 loc
== FRAG_RESULT_COLOR
&&
680 nir
->info
.outputs_written
& (1ull << loc
)) {
681 assert(attrib_count
== 1);
682 info
->properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
] = true;
686 info
->num_outputs
= num_outputs
;
688 struct set
*ubo_set
= _mesa_set_create(NULL
, _mesa_hash_pointer
,
689 _mesa_key_pointer_equal
);
691 /* Intialise const_file_max[0] */
692 info
->const_file_max
[0] = -1;
694 unsigned ubo_idx
= 1;
695 nir_foreach_variable(variable
, &nir
->uniforms
) {
696 const struct glsl_type
*type
= variable
->type
;
697 enum glsl_base_type base_type
=
698 glsl_get_base_type(glsl_without_array(type
));
699 unsigned aoa_size
= MAX2(1, glsl_get_aoa_size(type
));
700 unsigned loc
= variable
->data
.location
;
701 int slot_count
= glsl_count_attribute_slots(type
, false);
702 int max_slot
= MAX2(info
->const_file_max
[0], (int) loc
) + slot_count
;
704 /* Gather buffers declared bitmasks. Note: radeonsi doesn't
705 * really use the mask (other than ubo_idx == 1 for regular
706 * uniforms) its really only used for getting the buffer count
707 * so we don't need to worry about the ordering.
709 if (variable
->interface_type
!= NULL
) {
710 if (variable
->data
.mode
== nir_var_uniform
||
711 variable
->data
.mode
== nir_var_mem_ubo
) {
713 unsigned block_count
;
714 if (base_type
!= GLSL_TYPE_INTERFACE
) {
715 struct set_entry
*entry
=
716 _mesa_set_search(ubo_set
, variable
->interface_type
);
718 /* Check if we have already processed
719 * a member from this ubo.
726 block_count
= aoa_size
;
729 info
->const_buffers_declared
|= u_bit_consecutive(ubo_idx
, block_count
);
730 ubo_idx
+= block_count
;
732 _mesa_set_add(ubo_set
, variable
->interface_type
);
735 if (variable
->data
.mode
== nir_var_mem_ssbo
) {
736 /* TODO: make this more accurate */
737 info
->shader_buffers_declared
=
738 u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS
);
744 /* We rely on the fact that nir_lower_samplers_as_deref has
745 * eliminated struct dereferences.
747 if (base_type
== GLSL_TYPE_SAMPLER
) {
748 if (variable
->data
.bindless
) {
749 info
->const_buffers_declared
|= 1;
750 info
->const_file_max
[0] = max_slot
;
752 info
->samplers_declared
|=
753 u_bit_consecutive(variable
->data
.binding
, aoa_size
);
755 } else if (base_type
== GLSL_TYPE_IMAGE
) {
756 if (variable
->data
.bindless
) {
757 info
->const_buffers_declared
|= 1;
758 info
->const_file_max
[0] = max_slot
;
760 info
->images_declared
|=
761 u_bit_consecutive(variable
->data
.binding
, aoa_size
);
763 } else if (base_type
!= GLSL_TYPE_ATOMIC_UINT
) {
764 if (strncmp(variable
->name
, "state.", 6) == 0 ||
765 strncmp(variable
->name
, "gl_", 3) == 0) {
766 /* FIXME: figure out why piglit tests with builtin
767 * uniforms are failing without this.
769 info
->const_buffers_declared
=
770 u_bit_consecutive(0, SI_NUM_CONST_BUFFERS
);
772 info
->const_buffers_declared
|= 1;
773 info
->const_file_max
[0] = max_slot
;
778 _mesa_set_destroy(ubo_set
, NULL
);
780 info
->num_written_clipdistance
= nir
->info
.clip_distance_array_size
;
781 info
->num_written_culldistance
= nir
->info
.cull_distance_array_size
;
782 info
->clipdist_writemask
= u_bit_consecutive(0, info
->num_written_clipdistance
);
783 info
->culldist_writemask
= u_bit_consecutive(0, info
->num_written_culldistance
);
785 if (info
->processor
== PIPE_SHADER_FRAGMENT
)
786 info
->uses_kill
= nir
->info
.fs
.uses_discard
;
788 func
= (struct nir_function
*)exec_list_get_head_const(&nir
->functions
);
789 nir_foreach_block(block
, func
->impl
) {
790 nir_foreach_instr(instr
, block
)
791 scan_instruction(nir
, info
, instr
);
796 * Perform "lowering" operations on the NIR that are run once when the shader
797 * selector is created.
800 si_lower_nir(struct si_shader_selector
* sel
)
802 /* Adjust the driver location of inputs and outputs. The state tracker
803 * interprets them as slots, while the ac/nir backend interprets them
804 * as individual components.
806 nir_foreach_variable(variable
, &sel
->nir
->inputs
)
807 variable
->data
.driver_location
*= 4;
809 nir_foreach_variable(variable
, &sel
->nir
->outputs
) {
810 variable
->data
.driver_location
*= 4;
812 if (sel
->nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
813 if (variable
->data
.location
== FRAG_RESULT_DEPTH
)
814 variable
->data
.driver_location
+= 2;
815 else if (variable
->data
.location
== FRAG_RESULT_STENCIL
)
816 variable
->data
.driver_location
+= 1;
820 /* Perform lowerings (and optimizations) of code.
822 * Performance considerations aside, we must:
823 * - lower certain ALU operations
824 * - ensure constant offsets for texture instructions are folded
825 * and copy-propagated
827 NIR_PASS_V(sel
->nir
, nir_lower_returns
);
828 NIR_PASS_V(sel
->nir
, nir_lower_vars_to_ssa
);
829 NIR_PASS_V(sel
->nir
, nir_lower_alu_to_scalar
);
830 NIR_PASS_V(sel
->nir
, nir_lower_phis_to_scalar
);
832 static const struct nir_lower_tex_options lower_tex_options
= {
835 NIR_PASS_V(sel
->nir
, nir_lower_tex
, &lower_tex_options
);
837 const nir_lower_subgroups_options subgroups_options
= {
839 .ballot_bit_size
= 64,
840 .lower_to_scalar
= true,
841 .lower_subgroup_masks
= true,
842 .lower_vote_trivial
= false,
843 .lower_vote_eq_to_ballot
= true,
845 NIR_PASS_V(sel
->nir
, nir_lower_subgroups
, &subgroups_options
);
847 ac_lower_indirect_derefs(sel
->nir
, sel
->screen
->info
.chip_class
);
853 /* (Constant) copy propagation is needed for txf with offsets. */
854 NIR_PASS(progress
, sel
->nir
, nir_copy_prop
);
855 NIR_PASS(progress
, sel
->nir
, nir_opt_remove_phis
);
856 NIR_PASS(progress
, sel
->nir
, nir_opt_dce
);
857 if (nir_opt_trivial_continues(sel
->nir
)) {
859 NIR_PASS(progress
, sel
->nir
, nir_copy_prop
);
860 NIR_PASS(progress
, sel
->nir
, nir_opt_dce
);
862 NIR_PASS(progress
, sel
->nir
, nir_opt_if
);
863 NIR_PASS(progress
, sel
->nir
, nir_opt_dead_cf
);
864 NIR_PASS(progress
, sel
->nir
, nir_opt_cse
);
865 NIR_PASS(progress
, sel
->nir
, nir_opt_peephole_select
, 8, true, true);
867 /* Needed for algebraic lowering */
868 NIR_PASS(progress
, sel
->nir
, nir_opt_algebraic
);
869 NIR_PASS(progress
, sel
->nir
, nir_opt_constant_folding
);
871 NIR_PASS(progress
, sel
->nir
, nir_opt_undef
);
872 NIR_PASS(progress
, sel
->nir
, nir_opt_conditional_discard
);
873 if (sel
->nir
->options
->max_unroll_iterations
) {
874 NIR_PASS(progress
, sel
->nir
, nir_opt_loop_unroll
, 0);
878 NIR_PASS_V(sel
->nir
, nir_lower_bool_to_int32
);
881 static void declare_nir_input_vs(struct si_shader_context
*ctx
,
882 struct nir_variable
*variable
,
883 unsigned input_index
,
886 si_llvm_load_input_vs(ctx
, input_index
, out
);
889 static void declare_nir_input_fs(struct si_shader_context
*ctx
,
890 struct nir_variable
*variable
,
891 unsigned input_index
,
894 unsigned slot
= variable
->data
.location
;
895 if (slot
== VARYING_SLOT_POS
) {
896 out
[0] = LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
);
897 out
[1] = LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
);
898 out
[2] = LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Z_FLOAT
);
899 out
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
900 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_W_FLOAT
));
904 si_llvm_load_input_fs(ctx
, input_index
, out
);
908 si_nir_lookup_interp_param(struct ac_shader_abi
*abi
,
909 enum glsl_interp_mode interp
, unsigned location
)
911 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
912 int interp_param_idx
= -1;
915 case INTERP_MODE_FLAT
:
917 case INTERP_MODE_SMOOTH
:
918 case INTERP_MODE_NONE
:
919 if (location
== INTERP_CENTER
)
920 interp_param_idx
= SI_PARAM_PERSP_CENTER
;
921 else if (location
== INTERP_CENTROID
)
922 interp_param_idx
= SI_PARAM_PERSP_CENTROID
;
923 else if (location
== INTERP_SAMPLE
)
924 interp_param_idx
= SI_PARAM_PERSP_SAMPLE
;
926 case INTERP_MODE_NOPERSPECTIVE
:
927 if (location
== INTERP_CENTER
)
928 interp_param_idx
= SI_PARAM_LINEAR_CENTER
;
929 else if (location
== INTERP_CENTROID
)
930 interp_param_idx
= SI_PARAM_LINEAR_CENTROID
;
931 else if (location
== INTERP_SAMPLE
)
932 interp_param_idx
= SI_PARAM_LINEAR_SAMPLE
;
935 assert(!"Unhandled interpolation mode.");
939 return interp_param_idx
!= -1 ?
940 LLVMGetParam(ctx
->main_fn
, interp_param_idx
) : NULL
;
944 si_nir_load_sampler_desc(struct ac_shader_abi
*abi
,
945 unsigned descriptor_set
, unsigned base_index
,
946 unsigned constant_index
, LLVMValueRef dynamic_index
,
947 enum ac_descriptor_type desc_type
, bool image
,
948 bool write
, bool bindless
)
950 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
951 const struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
952 LLVMBuilderRef builder
= ctx
->ac
.builder
;
953 unsigned const_index
= base_index
+ constant_index
;
954 bool dcc_off
= write
;
956 /* TODO: images_store and images_atomic are not set */
957 if (!dynamic_index
&& image
&&
958 (info
->images_store
| info
->images_atomic
) & (1 << const_index
))
961 assert(!descriptor_set
);
962 assert(!image
|| desc_type
== AC_DESC_IMAGE
|| desc_type
== AC_DESC_BUFFER
);
966 LLVMGetParam(ctx
->main_fn
, ctx
->param_bindless_samplers_and_images
);
968 /* dynamic_index is the bindless handle */
970 /* For simplicity, bindless image descriptors use fixed
971 * 16-dword slots for now.
973 dynamic_index
= LLVMBuildMul(ctx
->ac
.builder
, dynamic_index
,
974 LLVMConstInt(ctx
->i32
, 2, 0), "");
976 return si_load_image_desc(ctx
, list
, dynamic_index
, desc_type
,
980 /* Since bindless handle arithmetic can contain an unsigned integer
981 * wraparound and si_load_sampler_desc assumes there isn't any,
982 * use GEP without "inbounds" (inside ac_build_pointer_add)
983 * to prevent incorrect code generation and hangs.
985 dynamic_index
= LLVMBuildMul(ctx
->ac
.builder
, dynamic_index
,
986 LLVMConstInt(ctx
->i32
, 2, 0), "");
987 list
= ac_build_pointer_add(&ctx
->ac
, list
, dynamic_index
);
988 return si_load_sampler_desc(ctx
, list
, ctx
->i32_0
, desc_type
);
991 unsigned num_slots
= image
? ctx
->num_images
: ctx
->num_samplers
;
992 assert(const_index
< num_slots
);
994 LLVMValueRef list
= LLVMGetParam(ctx
->main_fn
, ctx
->param_samplers_and_images
);
995 LLVMValueRef index
= LLVMConstInt(ctx
->ac
.i32
, const_index
, false);
998 index
= LLVMBuildAdd(builder
, index
, dynamic_index
, "");
1000 /* From the GL_ARB_shader_image_load_store extension spec:
1002 * If a shader performs an image load, store, or atomic
1003 * operation using an image variable declared as an array,
1004 * and if the index used to select an individual element is
1005 * negative or greater than or equal to the size of the
1006 * array, the results of the operation are undefined but may
1007 * not lead to termination.
1009 index
= si_llvm_bound_index(ctx
, index
, num_slots
);
1013 index
= LLVMBuildSub(ctx
->ac
.builder
,
1014 LLVMConstInt(ctx
->i32
, SI_NUM_IMAGES
- 1, 0),
1016 return si_load_image_desc(ctx
, list
, index
, desc_type
, dcc_off
, false);
1019 index
= LLVMBuildAdd(ctx
->ac
.builder
, index
,
1020 LLVMConstInt(ctx
->i32
, SI_NUM_IMAGES
/ 2, 0), "");
1021 return si_load_sampler_desc(ctx
, list
, index
, desc_type
);
1024 static void bitcast_inputs(struct si_shader_context
*ctx
,
1025 LLVMValueRef data
[4],
1028 for (unsigned chan
= 0; chan
< 4; chan
++) {
1029 ctx
->inputs
[input_idx
+ chan
] =
1030 LLVMBuildBitCast(ctx
->ac
.builder
, data
[chan
], ctx
->ac
.i32
, "");
1034 bool si_nir_build_llvm(struct si_shader_context
*ctx
, struct nir_shader
*nir
)
1036 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
1038 if (nir
->info
.stage
== MESA_SHADER_VERTEX
||
1039 nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1040 uint64_t processed_inputs
= 0;
1041 nir_foreach_variable(variable
, &nir
->inputs
) {
1042 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
,
1043 nir
->info
.stage
== MESA_SHADER_VERTEX
);
1044 unsigned input_idx
= variable
->data
.driver_location
;
1046 LLVMValueRef data
[4];
1047 unsigned loc
= variable
->data
.location
;
1049 if (loc
>= VARYING_SLOT_VAR0
&& nir
->info
.stage
== MESA_SHADER_FRAGMENT
)
1050 ctx
->abi
.fs_input_attr_indices
[loc
- VARYING_SLOT_VAR0
] = input_idx
/ 4;
1052 for (unsigned i
= 0; i
< attrib_count
; i
++) {
1053 /* Packed components share the same location so skip
1054 * them if we have already processed the location.
1056 if (processed_inputs
& ((uint64_t)1 << (loc
+ i
))) {
1061 if (nir
->info
.stage
== MESA_SHADER_VERTEX
) {
1062 declare_nir_input_vs(ctx
, variable
, input_idx
/ 4, data
);
1063 bitcast_inputs(ctx
, data
, input_idx
);
1064 if (glsl_type_is_dual_slot(variable
->type
)) {
1066 declare_nir_input_vs(ctx
, variable
, input_idx
/ 4, data
);
1067 bitcast_inputs(ctx
, data
, input_idx
);
1069 } else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1070 declare_nir_input_fs(ctx
, variable
, input_idx
/ 4, data
);
1071 bitcast_inputs(ctx
, data
, input_idx
);
1074 processed_inputs
|= ((uint64_t)1 << (loc
+ i
));
1080 ctx
->abi
.inputs
= &ctx
->inputs
[0];
1081 ctx
->abi
.load_sampler_desc
= si_nir_load_sampler_desc
;
1082 ctx
->abi
.clamp_shadow_reference
= true;
1084 ctx
->num_samplers
= util_last_bit(info
->samplers_declared
);
1085 ctx
->num_images
= util_last_bit(info
->images_declared
);
1087 if (ctx
->shader
->selector
->info
.properties
[TGSI_PROPERTY_CS_LOCAL_SIZE
]) {
1088 assert(gl_shader_stage_is_compute(nir
->info
.stage
));
1089 si_declare_compute_memory(ctx
);
1091 ac_nir_translate(&ctx
->ac
, &ctx
->abi
, nir
);