2 * Copyright © 2018 Valve Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
26 #include <unordered_map>
29 #include "nir_control_flow.h"
30 #include "vulkan/radv_shader.h"
31 #include "vulkan/radv_descriptor_set.h"
32 #include "vulkan/radv_shader_args.h"
34 #include "ac_exp_param.h"
35 #include "ac_shader_util.h"
37 #include "util/u_math.h"
39 #define MAX_INLINE_PUSH_CONSTS 8
43 struct shader_io_state
{
44 uint8_t mask
[VARYING_SLOT_MAX
];
45 Temp temps
[VARYING_SLOT_MAX
* 4u];
48 memset(mask
, 0, sizeof(mask
));
49 std::fill_n(temps
, VARYING_SLOT_MAX
* 4u, Temp(0, RegClass::v1
));
54 const struct radv_nir_compiler_options
*options
;
55 struct radv_shader_args
*args
;
58 uint32_t constant_data_offset
;
61 std::unique_ptr
<Temp
[]> allocated
;
62 std::unordered_map
<unsigned, std::array
<Temp
,NIR_MAX_VEC_COMPONENTS
>> allocated_vec
;
63 Stage stage
; /* Stage */
64 bool has_gfx10_wave64_bpermute
= false;
67 uint16_t loop_nest_depth
= 0;
71 bool has_divergent_continue
= false;
72 bool has_divergent_branch
= false;
75 bool is_divergent
= false;
77 bool exec_potentially_empty_discard
= false; /* set to false when loop_nest_depth==0 && parent_if.is_divergent==false */
78 uint16_t exec_potentially_empty_break_depth
= UINT16_MAX
;
79 /* Set to false when loop_nest_depth==exec_potentially_empty_break_depth
80 * and parent_if.is_divergent==false. Called _break but it's also used for
82 bool exec_potentially_empty_break
= false;
83 std::unique_ptr
<unsigned[]> nir_to_aco
; /* NIR block index to ACO block index */
86 Temp arg_temps
[AC_MAX_ARGS
];
89 Temp persp_centroid
, linear_centroid
;
94 /* gathered information */
95 uint64_t input_masks
[MESA_SHADER_COMPUTE
];
96 uint64_t output_masks
[MESA_SHADER_COMPUTE
];
98 /* VS output information */
99 bool export_clip_dists
;
100 unsigned num_clip_distances
;
101 unsigned num_cull_distances
;
103 /* tessellation information */
104 unsigned tcs_tess_lvl_out_loc
;
105 unsigned tcs_tess_lvl_in_loc
;
106 uint64_t tcs_temp_only_inputs
;
107 uint32_t tcs_num_inputs
;
108 uint32_t tcs_num_patches
;
109 bool tcs_in_out_eq
= false;
111 /* I/O information */
112 shader_io_state inputs
;
113 shader_io_state outputs
;
116 Temp
get_arg(isel_context
*ctx
, struct ac_arg arg
)
119 return ctx
->arg_temps
[arg
.arg_index
];
122 unsigned get_interp_input(nir_intrinsic_op intrin
, enum glsl_interp_mode interp
)
125 case INTERP_MODE_SMOOTH
:
126 case INTERP_MODE_NONE
:
127 if (intrin
== nir_intrinsic_load_barycentric_pixel
||
128 intrin
== nir_intrinsic_load_barycentric_at_sample
||
129 intrin
== nir_intrinsic_load_barycentric_at_offset
)
130 return S_0286CC_PERSP_CENTER_ENA(1);
131 else if (intrin
== nir_intrinsic_load_barycentric_centroid
)
132 return S_0286CC_PERSP_CENTROID_ENA(1);
133 else if (intrin
== nir_intrinsic_load_barycentric_sample
)
134 return S_0286CC_PERSP_SAMPLE_ENA(1);
136 case INTERP_MODE_NOPERSPECTIVE
:
137 if (intrin
== nir_intrinsic_load_barycentric_pixel
)
138 return S_0286CC_LINEAR_CENTER_ENA(1);
139 else if (intrin
== nir_intrinsic_load_barycentric_centroid
)
140 return S_0286CC_LINEAR_CENTROID_ENA(1);
141 else if (intrin
== nir_intrinsic_load_barycentric_sample
)
142 return S_0286CC_LINEAR_SAMPLE_ENA(1);
150 /* If one side of a divergent IF ends in a branch and the other doesn't, we
151 * might have to emit the contents of the side without the branch at the merge
152 * block instead. This is so that we can use any SGPR live-out of the side
153 * without the branch without creating a linear phi in the invert or merge block. */
155 sanitize_if(nir_function_impl
*impl
, bool *divergent
, nir_if
*nif
)
157 if (!divergent
[nif
->condition
.ssa
->index
])
160 nir_block
*then_block
= nir_if_last_then_block(nif
);
161 nir_block
*else_block
= nir_if_last_else_block(nif
);
162 bool then_jump
= nir_block_ends_in_jump(then_block
) || nir_block_is_unreachable(then_block
);
163 bool else_jump
= nir_block_ends_in_jump(else_block
) || nir_block_is_unreachable(else_block
);
164 if (then_jump
== else_jump
)
167 /* If the continue from block is empty then return as there is nothing to
170 if (nir_cf_list_is_empty_block(else_jump
? &nif
->then_list
: &nif
->else_list
))
173 /* Even though this if statement has a jump on one side, we may still have
174 * phis afterwards. Single-source phis can be produced by loop unrolling
175 * or dead control-flow passes and are perfectly legal. Run a quick phi
176 * removal on the block after the if to clean up any such phis.
178 nir_opt_remove_phis_block(nir_cf_node_as_block(nir_cf_node_next(&nif
->cf_node
)));
180 /* Finally, move the continue from branch after the if-statement. */
181 nir_block
*last_continue_from_blk
= else_jump
? then_block
: else_block
;
182 nir_block
*first_continue_from_blk
= else_jump
?
183 nir_if_first_then_block(nif
) : nir_if_first_else_block(nif
);
186 nir_cf_extract(&tmp
, nir_before_block(first_continue_from_blk
),
187 nir_after_block(last_continue_from_blk
));
188 nir_cf_reinsert(&tmp
, nir_after_cf_node(&nif
->cf_node
));
190 /* nir_cf_extract() invalidates dominance metadata, but it should still be
191 * correct because of the specific type of transformation we did. Block
192 * indices are not valid except for block_0's, which is all we care about for
193 * nir_block_is_unreachable(). */
194 impl
->valid_metadata
=
195 (nir_metadata
)(impl
->valid_metadata
| nir_metadata_dominance
| nir_metadata_block_index
);
201 sanitize_cf_list(nir_function_impl
*impl
, bool *divergent
, struct exec_list
*cf_list
)
203 bool progress
= false;
204 foreach_list_typed(nir_cf_node
, cf_node
, node
, cf_list
) {
205 switch (cf_node
->type
) {
206 case nir_cf_node_block
:
208 case nir_cf_node_if
: {
209 nir_if
*nif
= nir_cf_node_as_if(cf_node
);
210 progress
|= sanitize_cf_list(impl
, divergent
, &nif
->then_list
);
211 progress
|= sanitize_cf_list(impl
, divergent
, &nif
->else_list
);
212 progress
|= sanitize_if(impl
, divergent
, nif
);
215 case nir_cf_node_loop
: {
216 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
217 progress
|= sanitize_cf_list(impl
, divergent
, &loop
->body
);
220 case nir_cf_node_function
:
221 unreachable("Invalid cf type");
228 void init_context(isel_context
*ctx
, nir_shader
*shader
)
230 nir_function_impl
*impl
= nir_shader_get_entrypoint(shader
);
231 unsigned lane_mask_size
= ctx
->program
->lane_mask
.size();
233 ctx
->shader
= shader
;
234 ctx
->divergent_vals
= nir_divergence_analysis(shader
, nir_divergence_view_index_uniform
);
236 /* sanitize control flow */
237 nir_metadata_require(impl
, nir_metadata_dominance
);
238 sanitize_cf_list(impl
, ctx
->divergent_vals
, &impl
->body
);
239 nir_metadata_preserve(impl
, (nir_metadata
)~nir_metadata_block_index
);
241 /* we'll need this for isel */
242 nir_metadata_require(impl
, nir_metadata_block_index
);
244 if (!(ctx
->stage
& sw_gs_copy
) && ctx
->options
->dump_preoptir
) {
245 fprintf(stderr
, "NIR shader before instruction selection:\n");
246 nir_print_shader(shader
, stderr
);
249 std::unique_ptr
<Temp
[]> allocated
{new Temp
[impl
->ssa_alloc
]()};
251 unsigned spi_ps_inputs
= 0;
253 std::unique_ptr
<unsigned[]> nir_to_aco
{new unsigned[impl
->num_blocks
]()};
258 nir_foreach_block(block
, impl
) {
259 nir_foreach_instr(instr
, block
) {
260 switch(instr
->type
) {
261 case nir_instr_type_alu
: {
262 nir_alu_instr
*alu_instr
= nir_instr_as_alu(instr
);
263 unsigned size
= alu_instr
->dest
.dest
.ssa
.num_components
;
264 if (alu_instr
->dest
.dest
.ssa
.bit_size
== 64)
266 RegType type
= RegType::sgpr
;
267 switch(alu_instr
->op
) {
289 case nir_op_fround_even
:
298 case nir_op_pack_half_2x16
:
299 case nir_op_unpack_half_2x16_split_x
:
300 case nir_op_unpack_half_2x16_split_y
:
303 case nir_op_fddx_fine
:
304 case nir_op_fddy_fine
:
305 case nir_op_fddx_coarse
:
306 case nir_op_fddy_coarse
:
307 case nir_op_fquantize2f16
:
309 case nir_op_frexp_sig
:
310 case nir_op_frexp_exp
:
311 case nir_op_cube_face_index
:
312 case nir_op_cube_face_coord
:
313 type
= RegType::vgpr
;
327 size
= lane_mask_size
;
336 type
= ctx
->divergent_vals
[alu_instr
->dest
.dest
.ssa
.index
] ? RegType::vgpr
: RegType::sgpr
;
339 if (alu_instr
->dest
.dest
.ssa
.bit_size
== 1) {
340 size
= lane_mask_size
;
342 if (ctx
->divergent_vals
[alu_instr
->dest
.dest
.ssa
.index
]) {
343 type
= RegType::vgpr
;
345 if (allocated
[alu_instr
->src
[1].src
.ssa
->index
].type() == RegType::vgpr
||
346 allocated
[alu_instr
->src
[2].src
.ssa
->index
].type() == RegType::vgpr
) {
347 type
= RegType::vgpr
;
350 if (alu_instr
->src
[1].src
.ssa
->num_components
== 1 && alu_instr
->src
[2].src
.ssa
->num_components
== 1) {
351 assert(allocated
[alu_instr
->src
[1].src
.ssa
->index
].size() == allocated
[alu_instr
->src
[2].src
.ssa
->index
].size());
352 size
= allocated
[alu_instr
->src
[1].src
.ssa
->index
].size();
357 if (alu_instr
->dest
.dest
.ssa
.bit_size
== 1) {
358 size
= lane_mask_size
;
360 type
= ctx
->divergent_vals
[alu_instr
->dest
.dest
.ssa
.index
] ? RegType::vgpr
: RegType::sgpr
;
364 if (alu_instr
->dest
.dest
.ssa
.bit_size
== 1) {
365 size
= lane_mask_size
;
367 for (unsigned i
= 0; i
< nir_op_infos
[alu_instr
->op
].num_inputs
; i
++) {
368 if (allocated
[alu_instr
->src
[i
].src
.ssa
->index
].type() == RegType::vgpr
)
369 type
= RegType::vgpr
;
374 allocated
[alu_instr
->dest
.dest
.ssa
.index
] = Temp(0, RegClass(type
, size
));
377 case nir_instr_type_load_const
: {
378 unsigned size
= nir_instr_as_load_const(instr
)->def
.num_components
;
379 if (nir_instr_as_load_const(instr
)->def
.bit_size
== 64)
381 else if (nir_instr_as_load_const(instr
)->def
.bit_size
== 1)
382 size
*= lane_mask_size
;
383 allocated
[nir_instr_as_load_const(instr
)->def
.index
] = Temp(0, RegClass(RegType::sgpr
, size
));
386 case nir_instr_type_intrinsic
: {
387 nir_intrinsic_instr
*intrinsic
= nir_instr_as_intrinsic(instr
);
388 if (!nir_intrinsic_infos
[intrinsic
->intrinsic
].has_dest
)
390 unsigned size
= intrinsic
->dest
.ssa
.num_components
;
391 if (intrinsic
->dest
.ssa
.bit_size
== 64)
393 RegType type
= RegType::sgpr
;
394 switch(intrinsic
->intrinsic
) {
395 case nir_intrinsic_load_push_constant
:
396 case nir_intrinsic_load_work_group_id
:
397 case nir_intrinsic_load_num_work_groups
:
398 case nir_intrinsic_load_subgroup_id
:
399 case nir_intrinsic_load_num_subgroups
:
400 case nir_intrinsic_load_first_vertex
:
401 case nir_intrinsic_load_base_instance
:
402 case nir_intrinsic_get_buffer_size
:
403 case nir_intrinsic_vote_all
:
404 case nir_intrinsic_vote_any
:
405 case nir_intrinsic_read_first_invocation
:
406 case nir_intrinsic_read_invocation
:
407 case nir_intrinsic_first_invocation
:
408 type
= RegType::sgpr
;
409 if (intrinsic
->dest
.ssa
.bit_size
== 1)
410 size
= lane_mask_size
;
412 case nir_intrinsic_ballot
:
413 type
= RegType::sgpr
;
415 case nir_intrinsic_load_sample_id
:
416 case nir_intrinsic_load_sample_mask_in
:
417 case nir_intrinsic_load_input
:
418 case nir_intrinsic_load_output
:
419 case nir_intrinsic_load_input_vertex
:
420 case nir_intrinsic_load_per_vertex_input
:
421 case nir_intrinsic_load_per_vertex_output
:
422 case nir_intrinsic_load_vertex_id
:
423 case nir_intrinsic_load_vertex_id_zero_base
:
424 case nir_intrinsic_load_barycentric_sample
:
425 case nir_intrinsic_load_barycentric_pixel
:
426 case nir_intrinsic_load_barycentric_model
:
427 case nir_intrinsic_load_barycentric_centroid
:
428 case nir_intrinsic_load_barycentric_at_sample
:
429 case nir_intrinsic_load_barycentric_at_offset
:
430 case nir_intrinsic_load_interpolated_input
:
431 case nir_intrinsic_load_frag_coord
:
432 case nir_intrinsic_load_sample_pos
:
433 case nir_intrinsic_load_layer_id
:
434 case nir_intrinsic_load_local_invocation_id
:
435 case nir_intrinsic_load_local_invocation_index
:
436 case nir_intrinsic_load_subgroup_invocation
:
437 case nir_intrinsic_load_tess_coord
:
438 case nir_intrinsic_write_invocation_amd
:
439 case nir_intrinsic_mbcnt_amd
:
440 case nir_intrinsic_load_instance_id
:
441 case nir_intrinsic_ssbo_atomic_add
:
442 case nir_intrinsic_ssbo_atomic_imin
:
443 case nir_intrinsic_ssbo_atomic_umin
:
444 case nir_intrinsic_ssbo_atomic_imax
:
445 case nir_intrinsic_ssbo_atomic_umax
:
446 case nir_intrinsic_ssbo_atomic_and
:
447 case nir_intrinsic_ssbo_atomic_or
:
448 case nir_intrinsic_ssbo_atomic_xor
:
449 case nir_intrinsic_ssbo_atomic_exchange
:
450 case nir_intrinsic_ssbo_atomic_comp_swap
:
451 case nir_intrinsic_global_atomic_add
:
452 case nir_intrinsic_global_atomic_imin
:
453 case nir_intrinsic_global_atomic_umin
:
454 case nir_intrinsic_global_atomic_imax
:
455 case nir_intrinsic_global_atomic_umax
:
456 case nir_intrinsic_global_atomic_and
:
457 case nir_intrinsic_global_atomic_or
:
458 case nir_intrinsic_global_atomic_xor
:
459 case nir_intrinsic_global_atomic_exchange
:
460 case nir_intrinsic_global_atomic_comp_swap
:
461 case nir_intrinsic_image_deref_atomic_add
:
462 case nir_intrinsic_image_deref_atomic_umin
:
463 case nir_intrinsic_image_deref_atomic_imin
:
464 case nir_intrinsic_image_deref_atomic_umax
:
465 case nir_intrinsic_image_deref_atomic_imax
:
466 case nir_intrinsic_image_deref_atomic_and
:
467 case nir_intrinsic_image_deref_atomic_or
:
468 case nir_intrinsic_image_deref_atomic_xor
:
469 case nir_intrinsic_image_deref_atomic_exchange
:
470 case nir_intrinsic_image_deref_atomic_comp_swap
:
471 case nir_intrinsic_image_deref_size
:
472 case nir_intrinsic_shared_atomic_add
:
473 case nir_intrinsic_shared_atomic_imin
:
474 case nir_intrinsic_shared_atomic_umin
:
475 case nir_intrinsic_shared_atomic_imax
:
476 case nir_intrinsic_shared_atomic_umax
:
477 case nir_intrinsic_shared_atomic_and
:
478 case nir_intrinsic_shared_atomic_or
:
479 case nir_intrinsic_shared_atomic_xor
:
480 case nir_intrinsic_shared_atomic_exchange
:
481 case nir_intrinsic_shared_atomic_comp_swap
:
482 case nir_intrinsic_load_scratch
:
483 case nir_intrinsic_load_invocation_id
:
484 case nir_intrinsic_load_primitive_id
:
485 type
= RegType::vgpr
;
487 case nir_intrinsic_shuffle
:
488 case nir_intrinsic_quad_broadcast
:
489 case nir_intrinsic_quad_swap_horizontal
:
490 case nir_intrinsic_quad_swap_vertical
:
491 case nir_intrinsic_quad_swap_diagonal
:
492 case nir_intrinsic_quad_swizzle_amd
:
493 case nir_intrinsic_masked_swizzle_amd
:
494 case nir_intrinsic_inclusive_scan
:
495 case nir_intrinsic_exclusive_scan
:
496 if (intrinsic
->dest
.ssa
.bit_size
== 1) {
497 size
= lane_mask_size
;
498 type
= RegType::sgpr
;
499 } else if (!ctx
->divergent_vals
[intrinsic
->dest
.ssa
.index
]) {
500 type
= RegType::sgpr
;
502 type
= RegType::vgpr
;
505 case nir_intrinsic_load_view_index
:
506 type
= ctx
->stage
== fragment_fs
? RegType::vgpr
: RegType::sgpr
;
508 case nir_intrinsic_load_front_face
:
509 case nir_intrinsic_load_helper_invocation
:
510 case nir_intrinsic_is_helper_invocation
:
511 type
= RegType::sgpr
;
512 size
= lane_mask_size
;
514 case nir_intrinsic_reduce
:
515 if (intrinsic
->dest
.ssa
.bit_size
== 1) {
516 size
= lane_mask_size
;
517 type
= RegType::sgpr
;
518 } else if (!ctx
->divergent_vals
[intrinsic
->dest
.ssa
.index
]) {
519 type
= RegType::sgpr
;
521 type
= RegType::vgpr
;
524 case nir_intrinsic_load_ubo
:
525 case nir_intrinsic_load_ssbo
:
526 case nir_intrinsic_load_global
:
527 case nir_intrinsic_vulkan_resource_index
:
528 type
= ctx
->divergent_vals
[intrinsic
->dest
.ssa
.index
] ? RegType::vgpr
: RegType::sgpr
;
530 /* due to copy propagation, the swizzled imov is removed if num dest components == 1 */
531 case nir_intrinsic_load_shared
:
532 if (ctx
->divergent_vals
[intrinsic
->dest
.ssa
.index
])
533 type
= RegType::vgpr
;
535 type
= RegType::sgpr
;
538 for (unsigned i
= 0; i
< nir_intrinsic_infos
[intrinsic
->intrinsic
].num_srcs
; i
++) {
539 if (allocated
[intrinsic
->src
[i
].ssa
->index
].type() == RegType::vgpr
)
540 type
= RegType::vgpr
;
544 allocated
[intrinsic
->dest
.ssa
.index
] = Temp(0, RegClass(type
, size
));
546 switch(intrinsic
->intrinsic
) {
547 case nir_intrinsic_load_barycentric_sample
:
548 case nir_intrinsic_load_barycentric_pixel
:
549 case nir_intrinsic_load_barycentric_centroid
:
550 case nir_intrinsic_load_barycentric_at_sample
:
551 case nir_intrinsic_load_barycentric_at_offset
: {
552 glsl_interp_mode mode
= (glsl_interp_mode
)nir_intrinsic_interp_mode(intrinsic
);
553 spi_ps_inputs
|= get_interp_input(intrinsic
->intrinsic
, mode
);
556 case nir_intrinsic_load_barycentric_model
:
557 spi_ps_inputs
|= S_0286CC_PERSP_PULL_MODEL_ENA(1);
559 case nir_intrinsic_load_front_face
:
560 spi_ps_inputs
|= S_0286CC_FRONT_FACE_ENA(1);
562 case nir_intrinsic_load_frag_coord
:
563 case nir_intrinsic_load_sample_pos
: {
564 uint8_t mask
= nir_ssa_def_components_read(&intrinsic
->dest
.ssa
);
565 for (unsigned i
= 0; i
< 4; i
++) {
567 spi_ps_inputs
|= S_0286CC_POS_X_FLOAT_ENA(1) << i
;
572 case nir_intrinsic_load_sample_id
:
573 spi_ps_inputs
|= S_0286CC_ANCILLARY_ENA(1);
575 case nir_intrinsic_load_sample_mask_in
:
576 spi_ps_inputs
|= S_0286CC_ANCILLARY_ENA(1);
577 spi_ps_inputs
|= S_0286CC_SAMPLE_COVERAGE_ENA(1);
584 case nir_instr_type_tex
: {
585 nir_tex_instr
* tex
= nir_instr_as_tex(instr
);
586 unsigned size
= tex
->dest
.ssa
.num_components
;
588 if (tex
->dest
.ssa
.bit_size
== 64)
590 if (tex
->op
== nir_texop_texture_samples
)
591 assert(!ctx
->divergent_vals
[tex
->dest
.ssa
.index
]);
592 if (ctx
->divergent_vals
[tex
->dest
.ssa
.index
])
593 allocated
[tex
->dest
.ssa
.index
] = Temp(0, RegClass(RegType::vgpr
, size
));
595 allocated
[tex
->dest
.ssa
.index
] = Temp(0, RegClass(RegType::sgpr
, size
));
598 case nir_instr_type_parallel_copy
: {
599 nir_foreach_parallel_copy_entry(entry
, nir_instr_as_parallel_copy(instr
)) {
600 allocated
[entry
->dest
.ssa
.index
] = allocated
[entry
->src
.ssa
->index
];
604 case nir_instr_type_ssa_undef
: {
605 unsigned size
= nir_instr_as_ssa_undef(instr
)->def
.num_components
;
606 if (nir_instr_as_ssa_undef(instr
)->def
.bit_size
== 64)
608 else if (nir_instr_as_ssa_undef(instr
)->def
.bit_size
== 1)
609 size
*= lane_mask_size
;
610 allocated
[nir_instr_as_ssa_undef(instr
)->def
.index
] = Temp(0, RegClass(RegType::sgpr
, size
));
613 case nir_instr_type_phi
: {
614 nir_phi_instr
* phi
= nir_instr_as_phi(instr
);
616 unsigned size
= phi
->dest
.ssa
.num_components
;
618 if (phi
->dest
.ssa
.bit_size
== 1) {
619 assert(size
== 1 && "multiple components not yet supported on boolean phis.");
620 type
= RegType::sgpr
;
621 size
*= lane_mask_size
;
622 allocated
[phi
->dest
.ssa
.index
] = Temp(0, RegClass(type
, size
));
626 if (ctx
->divergent_vals
[phi
->dest
.ssa
.index
]) {
627 type
= RegType::vgpr
;
629 type
= RegType::sgpr
;
630 nir_foreach_phi_src (src
, phi
) {
631 if (allocated
[src
->src
.ssa
->index
].type() == RegType::vgpr
)
632 type
= RegType::vgpr
;
633 if (allocated
[src
->src
.ssa
->index
].type() == RegType::none
)
638 size
*= phi
->dest
.ssa
.bit_size
== 64 ? 2 : 1;
639 RegClass rc
= RegClass(type
, size
);
640 if (rc
!= allocated
[phi
->dest
.ssa
.index
].regClass()) {
643 nir_foreach_phi_src(src
, phi
)
644 assert(allocated
[src
->src
.ssa
->index
].size() == rc
.size());
646 allocated
[phi
->dest
.ssa
.index
] = Temp(0, rc
);
656 if (G_0286CC_POS_W_FLOAT_ENA(spi_ps_inputs
)) {
657 /* If POS_W_FLOAT (11) is enabled, at least one of PERSP_* must be enabled too */
658 spi_ps_inputs
|= S_0286CC_PERSP_CENTER_ENA(1);
661 if (!(spi_ps_inputs
& 0x7F)) {
662 /* At least one of PERSP_* (0xF) or LINEAR_* (0x70) must be enabled */
663 spi_ps_inputs
|= S_0286CC_PERSP_CENTER_ENA(1);
666 ctx
->program
->config
->spi_ps_input_ena
= spi_ps_inputs
;
667 ctx
->program
->config
->spi_ps_input_addr
= spi_ps_inputs
;
669 for (unsigned i
= 0; i
< impl
->ssa_alloc
; i
++)
670 allocated
[i
] = Temp(ctx
->program
->allocateId(), allocated
[i
].regClass());
672 ctx
->allocated
.reset(allocated
.release());
673 ctx
->cf_info
.nir_to_aco
.reset(nir_to_aco
.release());
676 Pseudo_instruction
*add_startpgm(struct isel_context
*ctx
)
678 unsigned arg_count
= ctx
->args
->ac
.arg_count
;
679 if (ctx
->stage
== fragment_fs
) {
680 /* LLVM optimizes away unused FS inputs and computes spi_ps_input_addr
681 * itself and then communicates the results back via the ELF binary.
682 * Mirror what LLVM does by re-mapping the VGPR arguments here.
684 * TODO: If we made the FS input scanning code into a separate pass that
685 * could run before argument setup, then this wouldn't be necessary
688 struct ac_shader_args
*args
= &ctx
->args
->ac
;
690 for (unsigned i
= 0, vgpr_arg
= 0, vgpr_reg
= 0; i
< args
->arg_count
; i
++) {
691 if (args
->args
[i
].file
!= AC_ARG_VGPR
) {
696 if (!(ctx
->program
->config
->spi_ps_input_addr
& (1 << vgpr_arg
))) {
697 args
->args
[i
].skip
= true;
699 args
->args
[i
].offset
= vgpr_reg
;
700 vgpr_reg
+= args
->args
[i
].size
;
707 aco_ptr
<Pseudo_instruction
> startpgm
{create_instruction
<Pseudo_instruction
>(aco_opcode::p_startpgm
, Format::PSEUDO
, 0, arg_count
+ 1)};
708 for (unsigned i
= 0, arg
= 0; i
< ctx
->args
->ac
.arg_count
; i
++) {
709 if (ctx
->args
->ac
.args
[i
].skip
)
712 enum ac_arg_regfile file
= ctx
->args
->ac
.args
[i
].file
;
713 unsigned size
= ctx
->args
->ac
.args
[i
].size
;
714 unsigned reg
= ctx
->args
->ac
.args
[i
].offset
;
715 RegClass type
= RegClass(file
== AC_ARG_SGPR
? RegType::sgpr
: RegType::vgpr
, size
);
716 Temp dst
= Temp
{ctx
->program
->allocateId(), type
};
717 ctx
->arg_temps
[i
] = dst
;
718 startpgm
->definitions
[arg
] = Definition(dst
);
719 startpgm
->definitions
[arg
].setFixed(PhysReg
{file
== AC_ARG_SGPR
? reg
: reg
+ 256});
722 startpgm
->definitions
[arg_count
] = Definition
{ctx
->program
->allocateId(), exec
, ctx
->program
->lane_mask
};
723 Pseudo_instruction
*instr
= startpgm
.get();
724 ctx
->block
->instructions
.push_back(std::move(startpgm
));
726 /* Stash these in the program so that they can be accessed later when
729 ctx
->program
->private_segment_buffer
= get_arg(ctx
, ctx
->args
->ring_offsets
);
730 ctx
->program
->scratch_offset
= get_arg(ctx
, ctx
->args
->scratch_offset
);
736 type_size(const struct glsl_type
*type
, bool bindless
)
738 // TODO: don't we need type->std430_base_alignment() here?
739 return glsl_count_attribute_slots(type
, false);
743 shared_var_info(const struct glsl_type
*type
, unsigned *size
, unsigned *align
)
745 assert(glsl_type_is_vector_or_scalar(type
));
747 uint32_t comp_size
= glsl_type_is_boolean(type
)
748 ? 4 : glsl_get_bit_size(type
) / 8;
749 unsigned length
= glsl_get_vector_elements(type
);
750 *size
= comp_size
* length
,
755 mem_vectorize_callback(unsigned align
, unsigned bit_size
,
756 unsigned num_components
, unsigned high_offset
,
757 nir_intrinsic_instr
*low
, nir_intrinsic_instr
*high
)
759 if ((bit_size
!= 32 && bit_size
!= 64) || num_components
> 4)
762 /* >128 bit loads are split except with SMEM */
763 if (bit_size
* num_components
> 128)
766 switch (low
->intrinsic
) {
767 case nir_intrinsic_load_ubo
:
768 case nir_intrinsic_load_ssbo
:
769 case nir_intrinsic_store_ssbo
:
770 case nir_intrinsic_load_push_constant
:
771 return align
% 4 == 0;
772 case nir_intrinsic_load_deref
:
773 case nir_intrinsic_store_deref
:
774 assert(nir_src_as_deref(low
->src
[0])->mode
== nir_var_mem_shared
);
776 case nir_intrinsic_load_shared
:
777 case nir_intrinsic_store_shared
:
778 if (bit_size
* num_components
> 64) /* 96 and 128 bit loads require 128 bit alignment and are split otherwise */
779 return align
% 16 == 0;
781 return align
% 4 == 0;
789 setup_vs_output_info(isel_context
*ctx
, nir_shader
*nir
,
790 bool export_prim_id
, bool export_clip_dists
,
791 radv_vs_output_info
*outinfo
)
793 memset(outinfo
->vs_output_param_offset
, AC_EXP_PARAM_UNDEFINED
,
794 sizeof(outinfo
->vs_output_param_offset
));
796 outinfo
->param_exports
= 0;
797 int pos_written
= 0x1;
798 if (outinfo
->writes_pointsize
|| outinfo
->writes_viewport_index
|| outinfo
->writes_layer
)
799 pos_written
|= 1 << 1;
801 uint64_t mask
= ctx
->output_masks
[nir
->info
.stage
];
803 int idx
= u_bit_scan64(&mask
);
804 if (idx
>= VARYING_SLOT_VAR0
|| idx
== VARYING_SLOT_LAYER
|| idx
== VARYING_SLOT_PRIMITIVE_ID
||
805 ((idx
== VARYING_SLOT_CLIP_DIST0
|| idx
== VARYING_SLOT_CLIP_DIST1
) && export_clip_dists
)) {
806 if (outinfo
->vs_output_param_offset
[idx
] == AC_EXP_PARAM_UNDEFINED
)
807 outinfo
->vs_output_param_offset
[idx
] = outinfo
->param_exports
++;
810 if (outinfo
->writes_layer
&&
811 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] == AC_EXP_PARAM_UNDEFINED
) {
812 /* when ctx->options->key.has_multiview_view_index = true, the layer
813 * variable isn't declared in NIR and it's isel's job to get the layer */
814 outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
] = outinfo
->param_exports
++;
817 if (export_prim_id
) {
818 assert(outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] == AC_EXP_PARAM_UNDEFINED
);
819 outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
] = outinfo
->param_exports
++;
822 ctx
->export_clip_dists
= export_clip_dists
;
823 ctx
->num_clip_distances
= util_bitcount(outinfo
->clip_dist_mask
);
824 ctx
->num_cull_distances
= util_bitcount(outinfo
->cull_dist_mask
);
826 assert(ctx
->num_clip_distances
+ ctx
->num_cull_distances
<= 8);
828 if (ctx
->num_clip_distances
+ ctx
->num_cull_distances
> 0)
829 pos_written
|= 1 << 2;
830 if (ctx
->num_clip_distances
+ ctx
->num_cull_distances
> 4)
831 pos_written
|= 1 << 3;
833 outinfo
->pos_exports
= util_bitcount(pos_written
);
837 setup_vs_variables(isel_context
*ctx
, nir_shader
*nir
)
839 nir_foreach_variable(variable
, &nir
->inputs
)
841 variable
->data
.driver_location
= variable
->data
.location
* 4;
843 nir_foreach_variable(variable
, &nir
->outputs
)
845 if (ctx
->stage
== vertex_geometry_gs
)
846 variable
->data
.driver_location
= util_bitcount64(ctx
->output_masks
[nir
->info
.stage
] & ((1ull << variable
->data
.location
) - 1ull)) * 4;
847 else if (ctx
->stage
== vertex_es
||
848 ctx
->stage
== vertex_ls
||
849 ctx
->stage
== vertex_tess_control_hs
)
850 // TODO: make this more compact
851 variable
->data
.driver_location
= shader_io_get_unique_index((gl_varying_slot
) variable
->data
.location
) * 4;
852 else if (ctx
->stage
== vertex_vs
)
853 variable
->data
.driver_location
= variable
->data
.location
* 4;
855 unreachable("Unsupported VS stage");
858 if (ctx
->stage
== vertex_vs
) {
859 radv_vs_output_info
*outinfo
= &ctx
->program
->info
->vs
.outinfo
;
860 setup_vs_output_info(ctx
, nir
, outinfo
->export_prim_id
,
861 ctx
->options
->key
.vs_common_out
.export_clip_dists
, outinfo
);
862 } else if (ctx
->stage
== vertex_geometry_gs
|| ctx
->stage
== vertex_es
) {
863 /* TODO: radv_nir_shader_info_pass() already sets this but it's larger
864 * than it needs to be in order to set it better, we have to improve
865 * radv_nir_shader_info_pass() because gfx9_get_gs_info() uses
866 * esgs_itemsize and has to be done before compilation
868 /* radv_es_output_info *outinfo = &ctx->program->info->vs.es_info;
869 outinfo->esgs_itemsize = util_bitcount64(ctx->output_masks[nir->info.stage]) * 16u; */
873 void setup_gs_variables(isel_context
*ctx
, nir_shader
*nir
)
875 if (ctx
->stage
== vertex_geometry_gs
|| ctx
->stage
== tess_eval_geometry_gs
) {
876 nir_foreach_variable(variable
, &nir
->inputs
) {
877 variable
->data
.driver_location
= util_bitcount64(ctx
->input_masks
[nir
->info
.stage
] & ((1ull << variable
->data
.location
) - 1ull)) * 4;
879 } else if (ctx
->stage
== geometry_gs
) {
880 //TODO: make this more compact
881 nir_foreach_variable(variable
, &nir
->inputs
) {
882 variable
->data
.driver_location
= shader_io_get_unique_index((gl_varying_slot
)variable
->data
.location
) * 4;
885 unreachable("Unsupported GS stage.");
888 nir_foreach_variable(variable
, &nir
->outputs
) {
889 variable
->data
.driver_location
= variable
->data
.location
* 4;
892 if (ctx
->stage
== vertex_geometry_gs
)
893 ctx
->program
->info
->gs
.es_type
= MESA_SHADER_VERTEX
;
894 else if (ctx
->stage
== tess_eval_geometry_gs
)
895 ctx
->program
->info
->gs
.es_type
= MESA_SHADER_TESS_EVAL
;
899 setup_tcs_info(isel_context
*ctx
, nir_shader
*nir
)
901 /* When the number of TCS input and output vertices are the same (typically 3):
902 * - There is an equal amount of LS and HS invocations
903 * - In case of merged LSHS shaders, the LS and HS halves of the shader
904 * always process the exact same vertex. We can use this knowledge to optimize them.
907 ctx
->stage
== vertex_tess_control_hs
&&
908 ctx
->args
->options
->key
.tcs
.input_vertices
== nir
->info
.tess
.tcs_vertices_out
;
910 if (ctx
->stage
== tess_control_hs
) {
911 ctx
->tcs_num_inputs
= ctx
->args
->options
->key
.tcs
.num_inputs
;
912 } else if (ctx
->stage
== vertex_tess_control_hs
) {
913 ctx
->tcs_num_inputs
= util_last_bit64(ctx
->args
->shader_info
->vs
.ls_outputs_written
);
915 if (ctx
->tcs_in_out_eq
) {
916 ctx
->tcs_temp_only_inputs
= ~nir
->info
.tess
.tcs_cross_invocation_inputs_read
&
917 ~nir
->info
.inputs_read_indirectly
&
918 nir
->info
.inputs_read
;
921 unreachable("Unsupported TCS shader stage");
924 ctx
->tcs_num_patches
= get_tcs_num_patches(
925 ctx
->args
->options
->key
.tcs
.input_vertices
,
926 nir
->info
.tess
.tcs_vertices_out
,
928 ctx
->args
->shader_info
->tcs
.outputs_written
,
929 ctx
->args
->shader_info
->tcs
.patch_outputs_written
,
930 ctx
->args
->options
->tess_offchip_block_dw_size
,
931 ctx
->args
->options
->chip_class
,
932 ctx
->args
->options
->family
);
933 unsigned lds_size
= calculate_tess_lds_size(
934 ctx
->args
->options
->key
.tcs
.input_vertices
,
935 nir
->info
.tess
.tcs_vertices_out
,
937 ctx
->tcs_num_patches
,
938 ctx
->args
->shader_info
->tcs
.outputs_written
,
939 ctx
->args
->shader_info
->tcs
.patch_outputs_written
);
941 ctx
->args
->shader_info
->tcs
.num_patches
= ctx
->tcs_num_patches
;
942 ctx
->args
->shader_info
->tcs
.lds_size
= lds_size
;
943 ctx
->program
->config
->lds_size
= (lds_size
+ ctx
->program
->lds_alloc_granule
- 1) /
944 ctx
->program
->lds_alloc_granule
;
948 setup_tcs_variables(isel_context
*ctx
, nir_shader
*nir
)
950 nir_foreach_variable(variable
, &nir
->inputs
) {
951 variable
->data
.driver_location
= shader_io_get_unique_index((gl_varying_slot
) variable
->data
.location
) * 4;
954 nir_foreach_variable(variable
, &nir
->outputs
) {
955 variable
->data
.driver_location
= shader_io_get_unique_index((gl_varying_slot
) variable
->data
.location
) * 4;
958 ctx
->tcs_tess_lvl_out_loc
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER
) * 16u;
959 ctx
->tcs_tess_lvl_in_loc
= shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER
) * 16u;
963 setup_tes_variables(isel_context
*ctx
, nir_shader
*nir
)
965 ctx
->tcs_num_patches
= ctx
->args
->options
->key
.tes
.num_patches
;
967 nir_foreach_variable(variable
, &nir
->inputs
) {
968 variable
->data
.driver_location
= shader_io_get_unique_index((gl_varying_slot
) variable
->data
.location
) * 4;
971 nir_foreach_variable(variable
, &nir
->outputs
) {
972 if (ctx
->stage
== tess_eval_vs
)
973 variable
->data
.driver_location
= variable
->data
.location
* 4;
974 else if (ctx
->stage
== tess_eval_es
)
975 variable
->data
.driver_location
= shader_io_get_unique_index((gl_varying_slot
) variable
->data
.location
) * 4;
976 else if (ctx
->stage
== tess_eval_geometry_gs
)
977 variable
->data
.driver_location
= util_bitcount64(ctx
->output_masks
[nir
->info
.stage
] & ((1ull << variable
->data
.location
) - 1ull)) * 4;
979 unreachable("Unsupported TES shader stage");
982 if (ctx
->stage
== tess_eval_vs
) {
983 radv_vs_output_info
*outinfo
= &ctx
->program
->info
->tes
.outinfo
;
984 setup_vs_output_info(ctx
, nir
, outinfo
->export_prim_id
,
985 ctx
->options
->key
.vs_common_out
.export_clip_dists
, outinfo
);
990 setup_variables(isel_context
*ctx
, nir_shader
*nir
)
992 switch (nir
->info
.stage
) {
993 case MESA_SHADER_FRAGMENT
: {
994 nir_foreach_variable(variable
, &nir
->outputs
)
996 int idx
= variable
->data
.location
+ variable
->data
.index
;
997 variable
->data
.driver_location
= idx
* 4;
1001 case MESA_SHADER_COMPUTE
: {
1002 ctx
->program
->config
->lds_size
= (nir
->info
.cs
.shared_size
+ ctx
->program
->lds_alloc_granule
- 1) /
1003 ctx
->program
->lds_alloc_granule
;
1006 case MESA_SHADER_VERTEX
: {
1007 setup_vs_variables(ctx
, nir
);
1010 case MESA_SHADER_GEOMETRY
: {
1011 setup_gs_variables(ctx
, nir
);
1014 case MESA_SHADER_TESS_CTRL
: {
1015 setup_tcs_variables(ctx
, nir
);
1018 case MESA_SHADER_TESS_EVAL
: {
1019 setup_tes_variables(ctx
, nir
);
1023 unreachable("Unhandled shader stage.");
1028 get_io_masks(isel_context
*ctx
, unsigned shader_count
, struct nir_shader
*const *shaders
)
1030 for (unsigned i
= 0; i
< shader_count
; i
++) {
1031 nir_shader
*nir
= shaders
[i
];
1032 if (nir
->info
.stage
== MESA_SHADER_COMPUTE
)
1035 uint64_t output_mask
= 0;
1036 nir_foreach_variable(variable
, &nir
->outputs
) {
1037 const glsl_type
*type
= variable
->type
;
1038 if (nir_is_per_vertex_io(variable
, nir
->info
.stage
))
1039 type
= type
->fields
.array
;
1040 unsigned slots
= type
->count_attribute_slots(false);
1041 if (variable
->data
.compact
) {
1042 unsigned component_count
= variable
->data
.location_frac
+ type
->length
;
1043 slots
= (component_count
+ 3) / 4;
1045 output_mask
|= ((1ull << slots
) - 1) << variable
->data
.location
;
1048 uint64_t input_mask
= 0;
1049 nir_foreach_variable(variable
, &nir
->inputs
) {
1050 const glsl_type
*type
= variable
->type
;
1051 if (nir_is_per_vertex_io(variable
, nir
->info
.stage
))
1052 type
= type
->fields
.array
;
1053 unsigned slots
= type
->count_attribute_slots(false);
1054 if (variable
->data
.compact
) {
1055 unsigned component_count
= variable
->data
.location_frac
+ type
->length
;
1056 slots
= (component_count
+ 3) / 4;
1058 input_mask
|= ((1ull << slots
) - 1) << variable
->data
.location
;
1061 ctx
->output_masks
[nir
->info
.stage
] |= output_mask
;
1062 if (i
+ 1 < shader_count
)
1063 ctx
->input_masks
[shaders
[i
+ 1]->info
.stage
] |= output_mask
;
1065 ctx
->input_masks
[nir
->info
.stage
] |= input_mask
;
1067 ctx
->output_masks
[shaders
[i
- 1]->info
.stage
] |= input_mask
;
1072 setup_nir(isel_context
*ctx
, nir_shader
*nir
)
1074 Program
*program
= ctx
->program
;
1076 /* align and copy constant data */
1077 while (program
->constant_data
.size() % 4u)
1078 program
->constant_data
.push_back(0);
1079 ctx
->constant_data_offset
= program
->constant_data
.size();
1080 program
->constant_data
.insert(program
->constant_data
.end(),
1081 (uint8_t*)nir
->constant_data
,
1082 (uint8_t*)nir
->constant_data
+ nir
->constant_data_size
);
1084 /* the variable setup has to be done before lower_io / CSE */
1085 setup_variables(ctx
, nir
);
1087 /* optimize and lower memory operations */
1088 bool lower_to_scalar
= false;
1089 bool lower_pack
= false;
1090 if (nir_opt_load_store_vectorize(nir
,
1091 (nir_variable_mode
)(nir_var_mem_ssbo
| nir_var_mem_ubo
|
1092 nir_var_mem_push_const
| nir_var_mem_shared
),
1093 mem_vectorize_callback
)) {
1094 lower_to_scalar
= true;
1097 if (nir
->info
.stage
!= MESA_SHADER_COMPUTE
)
1098 nir_lower_io(nir
, (nir_variable_mode
)(nir_var_shader_in
| nir_var_shader_out
), type_size
, (nir_lower_io_options
)0);
1099 nir_lower_explicit_io(nir
, nir_var_mem_global
, nir_address_format_64bit_global
);
1101 if (lower_to_scalar
)
1102 nir_lower_alu_to_scalar(nir
, NULL
, NULL
);
1104 nir_lower_pack(nir
);
1106 /* lower ALU operations */
1107 // TODO: implement logic64 in aco, it's more effective for sgprs
1108 nir_lower_int64(nir
, nir
->options
->lower_int64_options
);
1110 nir_opt_idiv_const(nir
, 32);
1111 nir_lower_idiv(nir
, nir_lower_idiv_precise
);
1113 /* optimize the lowered ALU operations */
1114 bool more_algebraic
= true;
1115 while (more_algebraic
) {
1116 more_algebraic
= false;
1117 NIR_PASS_V(nir
, nir_copy_prop
);
1118 NIR_PASS_V(nir
, nir_opt_dce
);
1119 NIR_PASS_V(nir
, nir_opt_constant_folding
);
1120 NIR_PASS(more_algebraic
, nir
, nir_opt_algebraic
);
1123 /* Do late algebraic optimization to turn add(a, neg(b)) back into
1124 * subs, then the mandatory cleanup after algebraic. Note that it may
1125 * produce fnegs, and if so then we need to keep running to squash
1128 bool more_late_algebraic
= true;
1129 while (more_late_algebraic
) {
1130 more_late_algebraic
= false;
1131 NIR_PASS(more_late_algebraic
, nir
, nir_opt_algebraic_late
);
1132 NIR_PASS_V(nir
, nir_opt_constant_folding
);
1133 NIR_PASS_V(nir
, nir_copy_prop
);
1134 NIR_PASS_V(nir
, nir_opt_dce
);
1135 NIR_PASS_V(nir
, nir_opt_cse
);
1138 /* cleanup passes */
1139 nir_lower_load_const_to_scalar(nir
);
1140 nir_opt_shrink_load(nir
);
1141 nir_move_options move_opts
= (nir_move_options
)(
1142 nir_move_const_undef
| nir_move_load_ubo
| nir_move_load_input
|
1143 nir_move_comparisons
| nir_move_copies
);
1144 nir_opt_sink(nir
, move_opts
);
1145 nir_opt_move(nir
, move_opts
);
1146 nir_convert_to_lcssa(nir
, true, false);
1147 nir_lower_phis_to_scalar(nir
);
1149 nir_function_impl
*func
= nir_shader_get_entrypoint(nir
);
1150 nir_index_ssa_defs(func
);
1154 setup_xnack(Program
*program
)
1156 switch (program
->family
) {
1164 program
->xnack_enabled
= true;
1172 setup_isel_context(Program
* program
,
1173 unsigned shader_count
,
1174 struct nir_shader
*const *shaders
,
1175 ac_shader_config
* config
,
1176 struct radv_shader_args
*args
,
1177 bool is_gs_copy_shader
)
1180 for (unsigned i
= 0; i
< shader_count
; i
++) {
1181 switch (shaders
[i
]->info
.stage
) {
1182 case MESA_SHADER_VERTEX
:
1183 program
->stage
|= sw_vs
;
1185 case MESA_SHADER_TESS_CTRL
:
1186 program
->stage
|= sw_tcs
;
1188 case MESA_SHADER_TESS_EVAL
:
1189 program
->stage
|= sw_tes
;
1191 case MESA_SHADER_GEOMETRY
:
1192 program
->stage
|= is_gs_copy_shader
? sw_gs_copy
: sw_gs
;
1194 case MESA_SHADER_FRAGMENT
:
1195 program
->stage
|= sw_fs
;
1197 case MESA_SHADER_COMPUTE
:
1198 program
->stage
|= sw_cs
;
1201 unreachable("Shader stage not implemented");
1204 bool gfx9_plus
= args
->options
->chip_class
>= GFX9
;
1205 bool ngg
= args
->shader_info
->is_ngg
&& args
->options
->chip_class
>= GFX10
;
1206 if (program
->stage
== sw_vs
&& args
->shader_info
->vs
.as_es
)
1207 program
->stage
|= hw_es
;
1208 else if (program
->stage
== sw_vs
&& !args
->shader_info
->vs
.as_ls
)
1209 program
->stage
|= hw_vs
;
1210 else if (program
->stage
== sw_gs
)
1211 program
->stage
|= hw_gs
;
1212 else if (program
->stage
== sw_fs
)
1213 program
->stage
|= hw_fs
;
1214 else if (program
->stage
== sw_cs
)
1215 program
->stage
|= hw_cs
;
1216 else if (program
->stage
== sw_gs_copy
)
1217 program
->stage
|= hw_vs
;
1218 else if (program
->stage
== (sw_vs
| sw_gs
) && gfx9_plus
&& !ngg
)
1219 program
->stage
|= hw_gs
;
1220 else if (program
->stage
== sw_vs
&& args
->shader_info
->vs
.as_ls
)
1221 program
->stage
|= hw_ls
; /* GFX6-8: VS is a Local Shader, when tessellation is used */
1222 else if (program
->stage
== sw_tcs
)
1223 program
->stage
|= hw_hs
; /* GFX6-8: TCS is a Hull Shader */
1224 else if (program
->stage
== (sw_vs
| sw_tcs
))
1225 program
->stage
|= hw_hs
; /* GFX9-10: VS+TCS merged into a Hull Shader */
1226 else if (program
->stage
== sw_tes
&& !args
->shader_info
->tes
.as_es
&& !ngg
)
1227 program
->stage
|= hw_vs
; /* GFX6-9: TES without GS uses the HW VS stage (and GFX10/legacy) */
1228 else if (program
->stage
== sw_tes
&& args
->shader_info
->tes
.as_es
&& !ngg
)
1229 program
->stage
|= hw_es
; /* GFX6-8: TES is an Export Shader */
1230 else if (program
->stage
== (sw_tes
| sw_gs
) && gfx9_plus
&& !ngg
)
1231 program
->stage
|= hw_gs
; /* GFX9: TES+GS merged into a GS (and GFX10/legacy) */
1233 unreachable("Shader stage not implemented");
1235 program
->config
= config
;
1236 program
->info
= args
->shader_info
;
1237 program
->chip_class
= args
->options
->chip_class
;
1238 program
->family
= args
->options
->family
;
1239 program
->wave_size
= args
->shader_info
->wave_size
;
1240 program
->lane_mask
= program
->wave_size
== 32 ? s1
: s2
;
1242 program
->lds_alloc_granule
= args
->options
->chip_class
>= GFX7
? 512 : 256;
1243 program
->lds_limit
= args
->options
->chip_class
>= GFX7
? 65536 : 32768;
1244 /* apparently gfx702 also has 16-bank LDS but I can't find a family for that */
1245 program
->has_16bank_lds
= args
->options
->family
== CHIP_KABINI
|| args
->options
->family
== CHIP_STONEY
;
1247 program
->vgpr_limit
= 256;
1248 program
->vgpr_alloc_granule
= 3;
1250 if (args
->options
->chip_class
>= GFX10
) {
1251 program
->physical_sgprs
= 2560; /* doesn't matter as long as it's at least 128 * 20 */
1252 program
->sgpr_alloc_granule
= 127;
1253 program
->sgpr_limit
= 106;
1254 program
->vgpr_alloc_granule
= program
->wave_size
== 32 ? 7 : 3;
1255 } else if (program
->chip_class
>= GFX8
) {
1256 program
->physical_sgprs
= 800;
1257 program
->sgpr_alloc_granule
= 15;
1258 if (args
->options
->family
== CHIP_TONGA
|| args
->options
->family
== CHIP_ICELAND
)
1259 program
->sgpr_limit
= 94; /* workaround hardware bug */
1261 program
->sgpr_limit
= 102;
1263 program
->physical_sgprs
= 512;
1264 program
->sgpr_alloc_granule
= 7;
1265 program
->sgpr_limit
= 104;
1268 isel_context ctx
= {};
1269 ctx
.program
= program
;
1271 ctx
.options
= args
->options
;
1272 ctx
.stage
= program
->stage
;
1274 /* TODO: Check if we need to adjust min_waves for unknown workgroup sizes. */
1275 if (program
->stage
& (hw_vs
| hw_fs
)) {
1276 /* PS and legacy VS have separate waves, no workgroups */
1277 program
->workgroup_size
= program
->wave_size
;
1278 } else if (program
->stage
== compute_cs
) {
1279 /* CS sets the workgroup size explicitly */
1280 unsigned* bsize
= program
->info
->cs
.block_size
;
1281 program
->workgroup_size
= bsize
[0] * bsize
[1] * bsize
[2];
1282 } else if ((program
->stage
& hw_es
) || program
->stage
== geometry_gs
) {
1283 /* Unmerged ESGS operate in workgroups if on-chip GS (LDS rings) are enabled on GFX7-8 (not implemented in Mesa) */
1284 program
->workgroup_size
= program
->wave_size
;
1285 } else if (program
->stage
& hw_gs
) {
1286 /* If on-chip GS (LDS rings) are enabled on GFX9 or later, merged GS operates in workgroups */
1287 program
->workgroup_size
= UINT_MAX
; /* TODO: set by VGT_GS_ONCHIP_CNTL, which is not plumbed to ACO */
1288 } else if (program
->stage
== vertex_ls
) {
1289 /* Unmerged LS operates in workgroups */
1290 program
->workgroup_size
= UINT_MAX
; /* TODO: probably tcs_num_patches * tcs_vertices_in, but those are not plumbed to ACO for LS */
1291 } else if (program
->stage
== tess_control_hs
) {
1292 /* Unmerged HS operates in workgroups, size is determined by the output vertices */
1293 setup_tcs_info(&ctx
, shaders
[0]);
1294 program
->workgroup_size
= ctx
.tcs_num_patches
* shaders
[0]->info
.tess
.tcs_vertices_out
;
1295 } else if (program
->stage
== vertex_tess_control_hs
) {
1296 /* Merged LSHS operates in workgroups, but can still have a different number of LS and HS invocations */
1297 setup_tcs_info(&ctx
, shaders
[1]);
1298 program
->workgroup_size
= ctx
.tcs_num_patches
* MAX2(shaders
[1]->info
.tess
.tcs_vertices_out
, ctx
.args
->options
->key
.tcs
.input_vertices
);
1300 unreachable("Unsupported shader stage.");
1303 calc_min_waves(program
);
1304 program
->vgpr_limit
= get_addr_vgpr_from_waves(program
, program
->min_waves
);
1305 program
->sgpr_limit
= get_addr_sgpr_from_waves(program
, program
->min_waves
);
1307 get_io_masks(&ctx
, shader_count
, shaders
);
1309 unsigned scratch_size
= 0;
1310 if (program
->stage
== gs_copy_vs
) {
1311 assert(shader_count
== 1);
1312 setup_vs_output_info(&ctx
, shaders
[0], false, true, &args
->shader_info
->vs
.outinfo
);
1314 for (unsigned i
= 0; i
< shader_count
; i
++) {
1315 nir_shader
*nir
= shaders
[i
];
1316 setup_nir(&ctx
, nir
);
1319 for (unsigned i
= 0; i
< shader_count
; i
++)
1320 scratch_size
= std::max(scratch_size
, shaders
[i
]->scratch_size
);
1323 ctx
.program
->config
->scratch_bytes_per_wave
= align(scratch_size
* ctx
.program
->wave_size
, 1024);
1325 ctx
.block
= ctx
.program
->create_and_insert_block();
1326 ctx
.block
->loop_nest_depth
= 0;
1327 ctx
.block
->kind
= block_kind_top_level
;
1329 setup_xnack(program
);