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.h"
25 #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"
36 type_size(const struct glsl_type
*type
)
38 return glsl_count_attribute_slots(type
, false);
41 static void scan_instruction(struct tgsi_shader_info
*info
,
44 if (instr
->type
== nir_instr_type_alu
) {
45 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
50 case nir_op_fddx_fine
:
51 case nir_op_fddy_fine
:
52 case nir_op_fddx_coarse
:
53 case nir_op_fddy_coarse
:
54 info
->uses_derivatives
= true;
59 } else if (instr
->type
== nir_instr_type_tex
) {
60 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
63 info
->samplers_declared
|=
64 u_bit_consecutive(tex
->sampler_index
, 1);
71 info
->uses_derivatives
= true;
76 } else if (instr
->type
== nir_instr_type_intrinsic
) {
77 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
79 switch (intr
->intrinsic
) {
80 case nir_intrinsic_load_front_face
:
81 info
->uses_frontface
= 1;
83 case nir_intrinsic_load_instance_id
:
84 info
->uses_instanceid
= 1;
86 case nir_intrinsic_load_vertex_id
:
87 info
->uses_vertexid
= 1;
89 case nir_intrinsic_load_vertex_id_zero_base
:
90 info
->uses_vertexid_nobase
= 1;
92 case nir_intrinsic_load_base_vertex
:
93 info
->uses_basevertex
= 1;
95 case nir_intrinsic_load_primitive_id
:
96 info
->uses_primid
= 1;
98 case nir_intrinsic_image_store
:
99 case nir_intrinsic_image_atomic_add
:
100 case nir_intrinsic_image_atomic_min
:
101 case nir_intrinsic_image_atomic_max
:
102 case nir_intrinsic_image_atomic_and
:
103 case nir_intrinsic_image_atomic_or
:
104 case nir_intrinsic_image_atomic_xor
:
105 case nir_intrinsic_image_atomic_exchange
:
106 case nir_intrinsic_image_atomic_comp_swap
:
107 case nir_intrinsic_store_ssbo
:
108 case nir_intrinsic_ssbo_atomic_add
:
109 case nir_intrinsic_ssbo_atomic_imin
:
110 case nir_intrinsic_ssbo_atomic_umin
:
111 case nir_intrinsic_ssbo_atomic_imax
:
112 case nir_intrinsic_ssbo_atomic_umax
:
113 case nir_intrinsic_ssbo_atomic_and
:
114 case nir_intrinsic_ssbo_atomic_or
:
115 case nir_intrinsic_ssbo_atomic_xor
:
116 case nir_intrinsic_ssbo_atomic_exchange
:
117 case nir_intrinsic_ssbo_atomic_comp_swap
:
118 info
->writes_memory
= true;
126 void si_nir_scan_shader(const struct nir_shader
*nir
,
127 struct tgsi_shader_info
*info
)
132 assert(nir
->info
.stage
== MESA_SHADER_VERTEX
||
133 nir
->info
.stage
== MESA_SHADER_FRAGMENT
);
135 info
->processor
= pipe_shader_type_from_mesa(nir
->info
.stage
);
136 info
->num_tokens
= 2; /* indicate that the shader is non-empty */
137 info
->num_instructions
= 2;
139 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
) {
140 info
->properties
[TGSI_PROPERTY_GS_INPUT_PRIM
] = nir
->info
.gs
.input_primitive
;
141 info
->properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
] = nir
->info
.gs
.output_primitive
;
142 info
->properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
] = nir
->info
.gs
.vertices_out
;
143 info
->properties
[TGSI_PROPERTY_GS_INVOCATIONS
] = nir
->info
.gs
.invocations
;
147 uint64_t processed_inputs
= 0;
148 unsigned num_inputs
= 0;
149 nir_foreach_variable(variable
, &nir
->inputs
) {
150 unsigned semantic_name
, semantic_index
;
151 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
,
152 nir
->info
.stage
== MESA_SHADER_VERTEX
);
154 assert(attrib_count
== 1 && "not implemented");
156 /* Vertex shader inputs don't have semantics. The state
157 * tracker has already mapped them to attributes via
158 * variable->data.driver_location.
160 if (nir
->info
.stage
== MESA_SHADER_VERTEX
)
163 /* Fragment shader position is a system value. */
164 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
&&
165 variable
->data
.location
== VARYING_SLOT_POS
) {
166 if (variable
->data
.pixel_center_integer
)
167 info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] =
168 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
;
174 i
= variable
->data
.driver_location
;
175 if (processed_inputs
& ((uint64_t)1 << i
))
178 processed_inputs
|= ((uint64_t)1 << i
);
181 tgsi_get_gl_varying_semantic(variable
->data
.location
, true,
182 &semantic_name
, &semantic_index
);
184 info
->input_semantic_name
[i
] = semantic_name
;
185 info
->input_semantic_index
[i
] = semantic_index
;
187 if (variable
->data
.sample
)
188 info
->input_interpolate_loc
[i
] = TGSI_INTERPOLATE_LOC_SAMPLE
;
189 else if (variable
->data
.centroid
)
190 info
->input_interpolate_loc
[i
] = TGSI_INTERPOLATE_LOC_CENTROID
;
192 info
->input_interpolate_loc
[i
] = TGSI_INTERPOLATE_LOC_CENTER
;
194 enum glsl_base_type base_type
=
195 glsl_get_base_type(glsl_without_array(variable
->type
));
197 switch (variable
->data
.interpolation
) {
198 case INTERP_MODE_NONE
:
199 if (glsl_base_type_is_integer(base_type
)) {
200 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_CONSTANT
;
204 if (semantic_name
== TGSI_SEMANTIC_COLOR
) {
205 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_COLOR
;
206 goto persp_locations
;
209 case INTERP_MODE_SMOOTH
:
210 assert(!glsl_base_type_is_integer(base_type
));
212 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_PERSPECTIVE
;
215 if (variable
->data
.sample
)
216 info
->uses_persp_sample
= true;
217 else if (variable
->data
.centroid
)
218 info
->uses_persp_centroid
= true;
220 info
->uses_persp_center
= true;
223 case INTERP_MODE_NOPERSPECTIVE
:
224 assert(!glsl_base_type_is_integer(base_type
));
226 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_LINEAR
;
228 if (variable
->data
.sample
)
229 info
->uses_linear_sample
= true;
230 else if (variable
->data
.centroid
)
231 info
->uses_linear_centroid
= true;
233 info
->uses_linear_center
= true;
236 case INTERP_MODE_FLAT
:
237 info
->input_interpolate
[i
] = TGSI_INTERPOLATE_CONSTANT
;
241 /* TODO make this more precise */
242 if (variable
->data
.location
== VARYING_SLOT_COL0
)
243 info
->colors_read
|= 0x0f;
244 else if (variable
->data
.location
== VARYING_SLOT_COL1
)
245 info
->colors_read
|= 0xf0;
248 if (nir
->info
.stage
!= MESA_SHADER_VERTEX
)
249 info
->num_inputs
= num_inputs
;
251 info
->num_inputs
= nir
->num_inputs
;
254 uint64_t processed_outputs
= 0;
255 unsigned num_outputs
= 0;
256 nir_foreach_variable(variable
, &nir
->outputs
) {
257 unsigned semantic_name
, semantic_index
;
259 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
260 tgsi_get_gl_frag_result_semantic(variable
->data
.location
,
261 &semantic_name
, &semantic_index
);
263 tgsi_get_gl_varying_semantic(variable
->data
.location
, true,
264 &semantic_name
, &semantic_index
);
267 i
= variable
->data
.driver_location
;
268 if (processed_outputs
& ((uint64_t)1 << i
))
271 processed_outputs
|= ((uint64_t)1 << i
);
274 info
->output_semantic_name
[i
] = semantic_name
;
275 info
->output_semantic_index
[i
] = semantic_index
;
276 info
->output_usagemask
[i
] = TGSI_WRITEMASK_XYZW
;
278 unsigned num_components
= 4;
279 unsigned vector_elements
= glsl_get_vector_elements(glsl_without_array(variable
->type
));
281 num_components
= vector_elements
;
283 unsigned gs_out_streams
;
284 if (variable
->data
.stream
& (1u << 31)) {
285 gs_out_streams
= variable
->data
.stream
& ~(1u << 31);
287 assert(variable
->data
.stream
< 4);
289 for (unsigned j
= 0; j
< num_components
; ++j
)
290 gs_out_streams
|= variable
->data
.stream
<< (2 * (variable
->data
.location_frac
+ j
));
293 unsigned streamx
= gs_out_streams
& 3;
294 unsigned streamy
= (gs_out_streams
>> 2) & 3;
295 unsigned streamz
= (gs_out_streams
>> 4) & 3;
296 unsigned streamw
= (gs_out_streams
>> 6) & 3;
298 if (info
->output_usagemask
[i
] & TGSI_WRITEMASK_X
) {
299 info
->output_streams
[i
] |= streamx
;
300 info
->num_stream_output_components
[streamx
]++;
302 if (info
->output_usagemask
[i
] & TGSI_WRITEMASK_Y
) {
303 info
->output_streams
[i
] |= streamy
<< 2;
304 info
->num_stream_output_components
[streamy
]++;
306 if (info
->output_usagemask
[i
] & TGSI_WRITEMASK_Z
) {
307 info
->output_streams
[i
] |= streamz
<< 4;
308 info
->num_stream_output_components
[streamz
]++;
310 if (info
->output_usagemask
[i
] & TGSI_WRITEMASK_W
) {
311 info
->output_streams
[i
] |= streamw
<< 6;
312 info
->num_stream_output_components
[streamw
]++;
315 switch (semantic_name
) {
316 case TGSI_SEMANTIC_PRIMID
:
317 info
->writes_primid
= true;
319 case TGSI_SEMANTIC_VIEWPORT_INDEX
:
320 info
->writes_viewport_index
= true;
322 case TGSI_SEMANTIC_LAYER
:
323 info
->writes_layer
= true;
325 case TGSI_SEMANTIC_PSIZE
:
326 info
->writes_psize
= true;
328 case TGSI_SEMANTIC_CLIPVERTEX
:
329 info
->writes_clipvertex
= true;
331 case TGSI_SEMANTIC_COLOR
:
332 info
->colors_written
|= 1 << semantic_index
;
334 case TGSI_SEMANTIC_STENCIL
:
335 info
->writes_stencil
= true;
337 case TGSI_SEMANTIC_SAMPLEMASK
:
338 info
->writes_samplemask
= true;
340 case TGSI_SEMANTIC_EDGEFLAG
:
341 info
->writes_edgeflag
= true;
343 case TGSI_SEMANTIC_POSITION
:
344 if (info
->processor
== PIPE_SHADER_FRAGMENT
)
345 info
->writes_z
= true;
347 info
->writes_position
= true;
352 info
->num_outputs
= num_outputs
;
354 nir_foreach_variable(variable
, &nir
->uniforms
) {
355 const struct glsl_type
*type
= variable
->type
;
356 enum glsl_base_type base_type
=
357 glsl_get_base_type(glsl_without_array(type
));
358 unsigned aoa_size
= MAX2(1, glsl_get_aoa_size(type
));
360 /* We rely on the fact that nir_lower_samplers_as_deref has
361 * eliminated struct dereferences.
363 if (base_type
== GLSL_TYPE_SAMPLER
)
364 info
->samplers_declared
|=
365 u_bit_consecutive(variable
->data
.binding
, aoa_size
);
366 else if (base_type
== GLSL_TYPE_IMAGE
)
367 info
->images_declared
|=
368 u_bit_consecutive(variable
->data
.binding
, aoa_size
);
371 info
->num_written_clipdistance
= nir
->info
.clip_distance_array_size
;
372 info
->num_written_culldistance
= nir
->info
.cull_distance_array_size
;
373 info
->clipdist_writemask
= u_bit_consecutive(0, info
->num_written_clipdistance
);
374 info
->culldist_writemask
= u_bit_consecutive(0, info
->num_written_culldistance
);
376 if (info
->processor
== PIPE_SHADER_FRAGMENT
)
377 info
->uses_kill
= nir
->info
.fs
.uses_discard
;
379 /* TODO make this more accurate */
380 info
->const_buffers_declared
= u_bit_consecutive(0, SI_NUM_CONST_BUFFERS
);
381 info
->shader_buffers_declared
= u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS
);
383 func
= (struct nir_function
*)exec_list_get_head_const(&nir
->functions
);
384 nir_foreach_block(block
, func
->impl
) {
385 nir_foreach_instr(instr
, block
)
386 scan_instruction(info
, instr
);
391 * Perform "lowering" operations on the NIR that are run once when the shader
392 * selector is created.
395 si_lower_nir(struct si_shader_selector
* sel
)
397 /* Adjust the driver location of inputs and outputs. The state tracker
398 * interprets them as slots, while the ac/nir backend interprets them
399 * as individual components.
401 nir_foreach_variable(variable
, &sel
->nir
->inputs
)
402 variable
->data
.driver_location
*= 4;
404 nir_foreach_variable(variable
, &sel
->nir
->outputs
) {
405 variable
->data
.driver_location
*= 4;
407 if (sel
->nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
408 if (variable
->data
.location
== FRAG_RESULT_DEPTH
)
409 variable
->data
.driver_location
+= 2;
410 else if (variable
->data
.location
== FRAG_RESULT_STENCIL
)
411 variable
->data
.driver_location
+= 1;
415 /* Perform lowerings (and optimizations) of code.
417 * Performance considerations aside, we must:
418 * - lower certain ALU operations
419 * - ensure constant offsets for texture instructions are folded
420 * and copy-propagated
422 NIR_PASS_V(sel
->nir
, nir_lower_io
, nir_var_uniform
, type_size
,
423 (nir_lower_io_options
)0);
424 NIR_PASS_V(sel
->nir
, nir_lower_uniforms_to_ubo
);
426 NIR_PASS_V(sel
->nir
, nir_lower_returns
);
427 NIR_PASS_V(sel
->nir
, nir_lower_vars_to_ssa
);
428 NIR_PASS_V(sel
->nir
, nir_lower_alu_to_scalar
);
429 NIR_PASS_V(sel
->nir
, nir_lower_phis_to_scalar
);
431 static const struct nir_lower_tex_options lower_tex_options
= {
434 NIR_PASS_V(sel
->nir
, nir_lower_tex
, &lower_tex_options
);
440 /* (Constant) copy propagation is needed for txf with offsets. */
441 NIR_PASS(progress
, sel
->nir
, nir_copy_prop
);
442 NIR_PASS(progress
, sel
->nir
, nir_opt_remove_phis
);
443 NIR_PASS(progress
, sel
->nir
, nir_opt_dce
);
444 if (nir_opt_trivial_continues(sel
->nir
)) {
446 NIR_PASS(progress
, sel
->nir
, nir_copy_prop
);
447 NIR_PASS(progress
, sel
->nir
, nir_opt_dce
);
449 NIR_PASS(progress
, sel
->nir
, nir_opt_if
);
450 NIR_PASS(progress
, sel
->nir
, nir_opt_dead_cf
);
451 NIR_PASS(progress
, sel
->nir
, nir_opt_cse
);
452 NIR_PASS(progress
, sel
->nir
, nir_opt_peephole_select
, 8);
454 /* Needed for algebraic lowering */
455 NIR_PASS(progress
, sel
->nir
, nir_opt_algebraic
);
456 NIR_PASS(progress
, sel
->nir
, nir_opt_constant_folding
);
458 NIR_PASS(progress
, sel
->nir
, nir_opt_undef
);
459 NIR_PASS(progress
, sel
->nir
, nir_opt_conditional_discard
);
460 if (sel
->nir
->options
->max_unroll_iterations
) {
461 NIR_PASS(progress
, sel
->nir
, nir_opt_loop_unroll
, 0);
466 static void declare_nir_input_vs(struct si_shader_context
*ctx
,
467 struct nir_variable
*variable
,
470 si_llvm_load_input_vs(ctx
, variable
->data
.driver_location
/ 4, out
);
473 static void declare_nir_input_fs(struct si_shader_context
*ctx
,
474 struct nir_variable
*variable
,
475 unsigned input_index
,
478 unsigned slot
= variable
->data
.location
;
479 if (slot
== VARYING_SLOT_POS
) {
480 out
[0] = LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_X_FLOAT
);
481 out
[1] = LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Y_FLOAT
);
482 out
[2] = LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_Z_FLOAT
);
483 out
[3] = ac_build_fdiv(&ctx
->ac
, ctx
->ac
.f32_1
,
484 LLVMGetParam(ctx
->main_fn
, SI_PARAM_POS_W_FLOAT
));
488 si_llvm_load_input_fs(ctx
, input_index
, out
);
492 si_nir_load_sampler_desc(struct ac_shader_abi
*abi
,
493 unsigned descriptor_set
, unsigned base_index
,
494 unsigned constant_index
, LLVMValueRef dynamic_index
,
495 enum ac_descriptor_type desc_type
, bool image
,
498 struct si_shader_context
*ctx
= si_shader_context_from_abi(abi
);
499 LLVMBuilderRef builder
= ctx
->ac
.builder
;
500 LLVMValueRef list
= LLVMGetParam(ctx
->main_fn
, ctx
->param_samplers_and_images
);
501 LLVMValueRef index
= dynamic_index
;
503 assert(!descriptor_set
);
506 index
= ctx
->ac
.i32_0
;
508 index
= LLVMBuildAdd(builder
, index
,
509 LLVMConstInt(ctx
->ac
.i32
, base_index
+ constant_index
, false),
513 assert(desc_type
== AC_DESC_IMAGE
|| desc_type
== AC_DESC_BUFFER
);
514 assert(base_index
+ constant_index
< ctx
->num_images
);
517 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_images
);
519 index
= LLVMBuildSub(ctx
->gallivm
.builder
,
520 LLVMConstInt(ctx
->i32
, SI_NUM_IMAGES
- 1, 0),
523 /* TODO: be smarter about when we use dcc_off */
524 return si_load_image_desc(ctx
, list
, index
, desc_type
, write
);
527 assert(base_index
+ constant_index
< ctx
->num_samplers
);
530 index
= si_llvm_bound_index(ctx
, index
, ctx
->num_samplers
);
532 index
= LLVMBuildAdd(ctx
->gallivm
.builder
, index
,
533 LLVMConstInt(ctx
->i32
, SI_NUM_IMAGES
/ 2, 0), "");
535 return si_load_sampler_desc(ctx
, list
, index
, desc_type
);
538 bool si_nir_build_llvm(struct si_shader_context
*ctx
, struct nir_shader
*nir
)
540 struct tgsi_shader_info
*info
= &ctx
->shader
->selector
->info
;
542 uint64_t processed_inputs
= 0;
543 nir_foreach_variable(variable
, &nir
->inputs
) {
544 unsigned attrib_count
= glsl_count_attribute_slots(variable
->type
,
545 nir
->info
.stage
== MESA_SHADER_VERTEX
);
546 unsigned input_idx
= variable
->data
.driver_location
;
548 assert(attrib_count
== 1);
550 LLVMValueRef data
[4];
551 unsigned loc
= variable
->data
.location
;
553 /* Packed components share the same location so skip
554 * them if we have already processed the location.
556 if (processed_inputs
& ((uint64_t)1 << loc
))
559 if (nir
->info
.stage
== MESA_SHADER_VERTEX
)
560 declare_nir_input_vs(ctx
, variable
, data
);
561 else if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
)
562 declare_nir_input_fs(ctx
, variable
, input_idx
/ 4, data
);
564 for (unsigned chan
= 0; chan
< 4; chan
++) {
565 ctx
->inputs
[input_idx
+ chan
] =
566 LLVMBuildBitCast(ctx
->ac
.builder
, data
[chan
], ctx
->ac
.i32
, "");
568 processed_inputs
|= ((uint64_t)1 << loc
);
571 ctx
->abi
.inputs
= &ctx
->inputs
[0];
572 ctx
->abi
.load_sampler_desc
= si_nir_load_sampler_desc
;
573 ctx
->abi
.clamp_shadow_reference
= true;
575 ctx
->num_samplers
= util_last_bit(info
->samplers_declared
);
576 ctx
->num_images
= util_last_bit(info
->images_declared
);
578 ac_nir_translate(&ctx
->ac
, &ctx
->abi
, nir
, NULL
);