2 * Copyright (C) 2015 Rob Clark <robclark@freedesktop.org>
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 * Rob Clark <robclark@freedesktop.org>
28 #include "util/debug.h"
29 #include "util/u_math.h"
32 #include "ir3_compiler.h"
33 #include "ir3_shader.h"
35 static const nir_shader_compiler_options options
= {
46 .lower_uadd_carry
= true,
47 .lower_usub_borrow
= true,
48 .lower_mul_high
= true,
49 .lower_mul_2x32_64
= true,
51 .vertex_id_zero_based
= true,
52 .lower_extract_byte
= true,
53 .lower_extract_word
= true,
54 .lower_all_io_to_elements
= true,
55 .lower_helper_invocation
= true,
56 .lower_bitfield_insert_to_shifts
= true,
57 .lower_bitfield_extract_to_shifts
= true,
58 .lower_pack_half_2x16
= true,
59 .lower_pack_snorm_4x8
= true,
60 .lower_pack_snorm_2x16
= true,
61 .lower_pack_unorm_4x8
= true,
62 .lower_pack_unorm_2x16
= true,
63 .lower_unpack_half_2x16
= true,
64 .lower_unpack_snorm_4x8
= true,
65 .lower_unpack_snorm_2x16
= true,
66 .lower_unpack_unorm_4x8
= true,
67 .lower_unpack_unorm_2x16
= true,
68 .lower_pack_split
= true,
69 .use_interpolated_input_intrinsics
= true,
71 .lower_to_scalar
= true,
73 .lower_wpos_pntc
= true,
76 /* we don't want to lower vertex_id to _zero_based on newer gpus: */
77 static const nir_shader_compiler_options options_a6xx
= {
88 .lower_uadd_carry
= true,
89 .lower_usub_borrow
= true,
90 .lower_mul_high
= true,
91 .lower_mul_2x32_64
= true,
93 .vertex_id_zero_based
= false,
94 .lower_extract_byte
= true,
95 .lower_extract_word
= true,
96 .lower_all_io_to_elements
= true,
97 .lower_helper_invocation
= true,
98 .lower_bitfield_insert_to_shifts
= true,
99 .lower_bitfield_extract_to_shifts
= true,
100 .lower_pack_half_2x16
= true,
101 .lower_pack_snorm_4x8
= true,
102 .lower_pack_snorm_2x16
= true,
103 .lower_pack_unorm_4x8
= true,
104 .lower_pack_unorm_2x16
= true,
105 .lower_unpack_half_2x16
= true,
106 .lower_unpack_snorm_4x8
= true,
107 .lower_unpack_snorm_2x16
= true,
108 .lower_unpack_unorm_4x8
= true,
109 .lower_unpack_unorm_2x16
= true,
110 .lower_pack_split
= true,
111 .use_interpolated_input_intrinsics
= true,
112 .lower_rotate
= true,
113 .vectorize_io
= true,
114 .lower_to_scalar
= true,
116 .max_unroll_iterations
= 32,
117 .lower_wpos_pntc
= true,
120 const nir_shader_compiler_options
*
121 ir3_get_compiler_options(struct ir3_compiler
*compiler
)
123 if (compiler
->gpu_id
>= 600)
124 return &options_a6xx
;
128 #define OPT(nir, pass, ...) ({ \
129 bool this_progress = false; \
130 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
134 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
137 ir3_optimize_loop(nir_shader
*s
)
140 unsigned lower_flrp
=
141 (s
->options
->lower_flrp16
? 16 : 0) |
142 (s
->options
->lower_flrp32
? 32 : 0) |
143 (s
->options
->lower_flrp64
? 64 : 0);
148 OPT_V(s
, nir_lower_vars_to_ssa
);
149 progress
|= OPT(s
, nir_opt_copy_prop_vars
);
150 progress
|= OPT(s
, nir_opt_dead_write_vars
);
151 progress
|= OPT(s
, nir_lower_alu_to_scalar
, NULL
, NULL
);
152 progress
|= OPT(s
, nir_lower_phis_to_scalar
);
154 progress
|= OPT(s
, nir_copy_prop
);
155 progress
|= OPT(s
, nir_opt_dce
);
156 progress
|= OPT(s
, nir_opt_cse
);
159 gcm
= env_var_as_unsigned("GCM", 0);
161 progress
|= OPT(s
, nir_opt_gcm
, true);
163 progress
|= OPT(s
, nir_opt_gcm
, false);
164 progress
|= OPT(s
, nir_opt_peephole_select
, 16, true, true);
165 progress
|= OPT(s
, nir_opt_intrinsics
);
166 progress
|= OPT(s
, nir_opt_algebraic
);
167 progress
|= OPT(s
, nir_lower_alu
);
168 progress
|= OPT(s
, nir_lower_pack
);
169 progress
|= OPT(s
, nir_opt_constant_folding
);
171 if (lower_flrp
!= 0) {
172 if (OPT(s
, nir_lower_flrp
,
174 false /* always_precise */,
175 s
->options
->lower_ffma
)) {
176 OPT(s
, nir_opt_constant_folding
);
180 /* Nothing should rematerialize any flrps, so we only
181 * need to do this lowering once.
186 progress
|= OPT(s
, nir_opt_dead_cf
);
187 if (OPT(s
, nir_opt_trivial_continues
)) {
189 /* If nir_opt_trivial_continues makes progress, then we need to clean
190 * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
193 OPT(s
, nir_copy_prop
);
196 progress
|= OPT(s
, nir_opt_if
, false);
197 progress
|= OPT(s
, nir_opt_remove_phis
);
198 progress
|= OPT(s
, nir_opt_undef
);
203 should_split_wrmask(const nir_instr
*instr
, const void *data
)
205 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
207 switch (intr
->intrinsic
) {
208 case nir_intrinsic_store_ssbo
:
209 case nir_intrinsic_store_shared
:
210 case nir_intrinsic_store_global
:
218 ir3_finalize_nir(struct ir3_compiler
*compiler
, nir_shader
*s
)
220 struct nir_lower_tex_options tex_options
= {
222 .lower_tg4_offsets
= true,
225 if (compiler
->gpu_id
>= 400) {
226 /* a4xx seems to have *no* sam.p */
227 tex_options
.lower_txp
= ~0; /* lower all txp */
229 /* a3xx just needs to avoid sam.p for 3d tex */
230 tex_options
.lower_txp
= (1 << GLSL_SAMPLER_DIM_3D
);
233 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
234 debug_printf("----------------------\n");
235 nir_print_shader(s
, stdout
);
236 debug_printf("----------------------\n");
239 if (s
->info
.stage
== MESA_SHADER_GEOMETRY
)
240 NIR_PASS_V(s
, ir3_nir_lower_gs
);
242 NIR_PASS_V(s
, nir_lower_io_arrays_to_elements_no_indirects
, false);
244 NIR_PASS_V(s
, nir_lower_amul
, ir3_glsl_type_size
);
246 OPT_V(s
, nir_lower_regs_to_ssa
);
247 OPT_V(s
, nir_lower_wrmasks
, should_split_wrmask
, s
);
249 OPT_V(s
, nir_lower_tex
, &tex_options
);
250 OPT_V(s
, nir_lower_load_const_to_scalar
);
251 if (compiler
->gpu_id
< 500)
252 OPT_V(s
, ir3_nir_lower_tg4_to_tex
);
254 ir3_optimize_loop(s
);
256 /* do idiv lowering after first opt loop to get a chance to propagate
257 * constants for divide by immed power-of-two:
259 const bool idiv_progress
= OPT(s
, nir_lower_idiv
, nir_lower_idiv_fast
);
262 ir3_optimize_loop(s
);
264 OPT_V(s
, nir_remove_dead_variables
, nir_var_function_temp
, NULL
);
266 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
267 debug_printf("----------------------\n");
268 nir_print_shader(s
, stdout
);
269 debug_printf("----------------------\n");
276 * Late passes that need to be done after pscreen->finalize_nir()
279 ir3_nir_post_finalize(struct ir3_compiler
*compiler
, nir_shader
*s
)
281 NIR_PASS_V(s
, nir_lower_io
, nir_var_shader_in
| nir_var_shader_out
,
282 ir3_glsl_type_size
, (nir_lower_io_options
)0);
284 if (s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
285 /* NOTE: lower load_barycentric_at_sample first, since it
286 * produces load_barycentric_at_offset:
288 NIR_PASS_V(s
, ir3_nir_lower_load_barycentric_at_sample
);
289 NIR_PASS_V(s
, ir3_nir_lower_load_barycentric_at_offset
);
290 NIR_PASS_V(s
, ir3_nir_move_varying_inputs
);
291 NIR_PASS_V(s
, nir_lower_fb_read
);
294 if (compiler
->gpu_id
>= 600 &&
295 s
->info
.stage
== MESA_SHADER_FRAGMENT
&&
296 !(ir3_shader_debug
& IR3_DBG_NOFP16
)) {
297 NIR_PASS_V(s
, nir_lower_mediump_outputs
);
300 /* we cannot ensure that ir3_finalize_nir() is only called once, so
301 * we also need to do trig workarounds here:
303 OPT_V(s
, ir3_nir_apply_trig_workarounds
);
305 ir3_optimize_loop(s
);
309 ir3_nir_lower_layer_id(nir_shader
*nir
)
311 unsigned layer_id_loc
= ~0;
312 nir_foreach_variable(var
, &nir
->inputs
) {
313 if (var
->data
.location
== VARYING_SLOT_LAYER
) {
314 layer_id_loc
= var
->data
.driver_location
;
319 assert(layer_id_loc
!= ~0);
321 bool progress
= false;
324 nir_foreach_function(func
, nir
) {
325 nir_builder_init(&b
, func
->impl
);
327 nir_foreach_block(block
, func
->impl
) {
328 nir_foreach_instr_safe(instr
, block
) {
329 if (instr
->type
!= nir_instr_type_intrinsic
)
332 nir_intrinsic_instr
*intrin
=
333 nir_instr_as_intrinsic(instr
);
335 if (intrin
->intrinsic
!= nir_intrinsic_load_input
)
338 unsigned base
= nir_intrinsic_base(intrin
);
339 if (base
!= layer_id_loc
)
342 b
.cursor
= nir_before_instr(&intrin
->instr
);
343 nir_ssa_def
*zero
= nir_imm_int(&b
, 0);
344 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
345 nir_src_for_ssa(zero
));
346 nir_instr_remove(&intrin
->instr
);
352 nir_metadata_preserve(func
->impl
,
353 nir_metadata_block_index
|
354 nir_metadata_dominance
);
356 nir_metadata_preserve(func
->impl
, nir_metadata_all
);
364 ir3_nir_lower_variant(struct ir3_shader_variant
*so
, nir_shader
*s
)
366 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
367 debug_printf("----------------------\n");
368 nir_print_shader(s
, stdout
);
369 debug_printf("----------------------\n");
372 bool progress
= false;
374 if (so
->key
.has_gs
|| so
->key
.tessellation
) {
375 switch (so
->shader
->type
) {
376 case MESA_SHADER_VERTEX
:
377 NIR_PASS_V(s
, ir3_nir_lower_to_explicit_output
, so
, so
->key
.tessellation
);
380 case MESA_SHADER_TESS_CTRL
:
381 NIR_PASS_V(s
, ir3_nir_lower_tess_ctrl
, so
, so
->key
.tessellation
);
382 NIR_PASS_V(s
, ir3_nir_lower_to_explicit_input
);
385 case MESA_SHADER_TESS_EVAL
:
386 NIR_PASS_V(s
, ir3_nir_lower_tess_eval
, so
->key
.tessellation
);
388 NIR_PASS_V(s
, ir3_nir_lower_to_explicit_output
, so
, so
->key
.tessellation
);
391 case MESA_SHADER_GEOMETRY
:
392 NIR_PASS_V(s
, ir3_nir_lower_to_explicit_input
);
400 if (s
->info
.stage
== MESA_SHADER_VERTEX
) {
401 if (so
->key
.ucp_enables
)
402 progress
|= OPT(s
, nir_lower_clip_vs
, so
->key
.ucp_enables
, false, false, NULL
);
403 if (so
->key
.vclamp_color
)
404 progress
|= OPT(s
, nir_lower_clamp_color_outputs
);
405 } else if (s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
406 if (so
->key
.ucp_enables
)
407 progress
|= OPT(s
, nir_lower_clip_fs
, so
->key
.ucp_enables
, false);
408 if (so
->key
.fclamp_color
)
409 progress
|= OPT(s
, nir_lower_clamp_color_outputs
);
410 if (so
->key
.layer_zero
&& (s
->info
.inputs_read
& VARYING_BIT_LAYER
))
411 progress
|= OPT(s
, ir3_nir_lower_layer_id
);
413 if (so
->key
.color_two_side
) {
414 OPT_V(s
, nir_lower_two_sided_color
);
418 struct nir_lower_tex_options tex_options
= { };
420 switch (so
->shader
->type
) {
421 case MESA_SHADER_FRAGMENT
:
422 tex_options
.saturate_s
= so
->key
.fsaturate_s
;
423 tex_options
.saturate_t
= so
->key
.fsaturate_t
;
424 tex_options
.saturate_r
= so
->key
.fsaturate_r
;
426 case MESA_SHADER_VERTEX
:
427 tex_options
.saturate_s
= so
->key
.vsaturate_s
;
428 tex_options
.saturate_t
= so
->key
.vsaturate_t
;
429 tex_options
.saturate_r
= so
->key
.vsaturate_r
;
436 if (tex_options
.saturate_s
|| tex_options
.saturate_t
||
437 tex_options
.saturate_r
) {
438 progress
|= OPT(s
, nir_lower_tex
, &tex_options
);
441 if (!so
->binning_pass
)
442 OPT_V(s
, ir3_nir_analyze_ubo_ranges
, so
);
444 progress
|= OPT(s
, ir3_nir_lower_ubo_loads
, so
);
446 /* UBO offset lowering has to come after we've decided what will
447 * be left as load_ubo
449 OPT_V(s
, ir3_nir_lower_io_offsets
, so
->shader
->compiler
->gpu_id
);
452 ir3_optimize_loop(s
);
454 /* Do late algebraic optimization to turn add(a, neg(b)) back into
455 * subs, then the mandatory cleanup after algebraic. Note that it may
456 * produce fnegs, and if so then we need to keep running to squash
459 bool more_late_algebraic
= true;
460 while (more_late_algebraic
) {
461 more_late_algebraic
= OPT(s
, nir_opt_algebraic_late
);
462 OPT_V(s
, nir_opt_constant_folding
);
463 OPT_V(s
, nir_copy_prop
);
464 OPT_V(s
, nir_opt_dce
);
465 OPT_V(s
, nir_opt_cse
);
468 OPT_V(s
, nir_opt_sink
, nir_move_const_undef
);
470 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
471 debug_printf("----------------------\n");
472 nir_print_shader(s
, stdout
);
473 debug_printf("----------------------\n");
478 /* Binning pass variants re-use the const_state of the corresponding
479 * draw pass shader, so that same const emit can be re-used for both
482 if (!so
->binning_pass
)
483 ir3_setup_const_state(s
, so
, ir3_const_state(so
));
487 ir3_nir_scan_driver_consts(nir_shader
*shader
,
488 struct ir3_const_state
*layout
)
490 nir_foreach_function (function
, shader
) {
494 nir_foreach_block (block
, function
->impl
) {
495 nir_foreach_instr (instr
, block
) {
496 if (instr
->type
!= nir_instr_type_intrinsic
)
499 nir_intrinsic_instr
*intr
=
500 nir_instr_as_intrinsic(instr
);
503 switch (intr
->intrinsic
) {
504 case nir_intrinsic_get_buffer_size
:
505 idx
= nir_src_as_uint(intr
->src
[0]);
506 if (layout
->ssbo_size
.mask
& (1 << idx
))
508 layout
->ssbo_size
.mask
|= (1 << idx
);
509 layout
->ssbo_size
.off
[idx
] =
510 layout
->ssbo_size
.count
;
511 layout
->ssbo_size
.count
+= 1; /* one const per */
513 case nir_intrinsic_image_atomic_add
:
514 case nir_intrinsic_image_atomic_imin
:
515 case nir_intrinsic_image_atomic_umin
:
516 case nir_intrinsic_image_atomic_imax
:
517 case nir_intrinsic_image_atomic_umax
:
518 case nir_intrinsic_image_atomic_and
:
519 case nir_intrinsic_image_atomic_or
:
520 case nir_intrinsic_image_atomic_xor
:
521 case nir_intrinsic_image_atomic_exchange
:
522 case nir_intrinsic_image_atomic_comp_swap
:
523 case nir_intrinsic_image_store
:
524 case nir_intrinsic_image_size
:
525 idx
= nir_src_as_uint(intr
->src
[0]);
526 if (layout
->image_dims
.mask
& (1 << idx
))
528 layout
->image_dims
.mask
|= (1 << idx
);
529 layout
->image_dims
.off
[idx
] =
530 layout
->image_dims
.count
;
531 layout
->image_dims
.count
+= 3; /* three const per */
533 case nir_intrinsic_load_base_vertex
:
534 case nir_intrinsic_load_first_vertex
:
535 layout
->num_driver_params
=
536 MAX2(layout
->num_driver_params
, IR3_DP_VTXID_BASE
+ 1);
538 case nir_intrinsic_load_base_instance
:
539 layout
->num_driver_params
=
540 MAX2(layout
->num_driver_params
, IR3_DP_INSTID_BASE
+ 1);
542 case nir_intrinsic_load_user_clip_plane
:
543 idx
= nir_intrinsic_ucp_id(intr
);
544 layout
->num_driver_params
=
545 MAX2(layout
->num_driver_params
, IR3_DP_UCP0_X
+ (idx
+ 1) * 4);
547 case nir_intrinsic_load_num_work_groups
:
548 layout
->num_driver_params
=
549 MAX2(layout
->num_driver_params
, IR3_DP_NUM_WORK_GROUPS_Z
+ 1);
551 case nir_intrinsic_load_local_group_size
:
552 layout
->num_driver_params
=
553 MAX2(layout
->num_driver_params
, IR3_DP_LOCAL_GROUP_SIZE_Z
+ 1);
563 /* Sets up the variant-dependent constant state for the ir3_shader. Note
564 * that it is also used from ir3_nir_analyze_ubo_ranges() to figure out the
565 * maximum number of driver params that would eventually be used, to leave
566 * space for this function to allocate the driver params.
569 ir3_setup_const_state(nir_shader
*nir
, struct ir3_shader_variant
*v
,
570 struct ir3_const_state
*const_state
)
572 struct ir3_compiler
*compiler
= v
->shader
->compiler
;
574 memset(&const_state
->offsets
, ~0, sizeof(const_state
->offsets
));
576 ir3_nir_scan_driver_consts(nir
, const_state
);
578 if ((compiler
->gpu_id
< 500) &&
579 (v
->shader
->stream_output
.num_outputs
> 0)) {
580 const_state
->num_driver_params
=
581 MAX2(const_state
->num_driver_params
, IR3_DP_VTXCNT_MAX
+ 1);
584 const_state
->num_ubos
= nir
->info
.num_ubos
;
586 /* num_driver_params is scalar, align to vec4: */
587 const_state
->num_driver_params
= align(const_state
->num_driver_params
, 4);
589 debug_assert((const_state
->ubo_state
.size
% 16) == 0);
590 unsigned constoff
= const_state
->ubo_state
.size
/ 16;
591 unsigned ptrsz
= ir3_pointer_size(compiler
);
593 if (const_state
->num_ubos
> 0) {
594 const_state
->offsets
.ubo
= constoff
;
595 constoff
+= align(const_state
->num_ubos
* ptrsz
, 4) / 4;
598 if (const_state
->ssbo_size
.count
> 0) {
599 unsigned cnt
= const_state
->ssbo_size
.count
;
600 const_state
->offsets
.ssbo_sizes
= constoff
;
601 constoff
+= align(cnt
, 4) / 4;
604 if (const_state
->image_dims
.count
> 0) {
605 unsigned cnt
= const_state
->image_dims
.count
;
606 const_state
->offsets
.image_dims
= constoff
;
607 constoff
+= align(cnt
, 4) / 4;
610 if (const_state
->num_driver_params
> 0) {
611 /* offset cannot be 0 for vs params loaded by CP_DRAW_INDIRECT_MULTI */
612 if (v
->type
== MESA_SHADER_VERTEX
&& compiler
->gpu_id
>= 600)
613 constoff
= MAX2(constoff
, 1);
614 const_state
->offsets
.driver_param
= constoff
;
616 constoff
+= const_state
->num_driver_params
/ 4;
618 if ((v
->type
== MESA_SHADER_VERTEX
) &&
619 (compiler
->gpu_id
< 500) &&
620 v
->shader
->stream_output
.num_outputs
> 0) {
621 const_state
->offsets
.tfbo
= constoff
;
622 constoff
+= align(IR3_MAX_SO_BUFFERS
* ptrsz
, 4) / 4;
626 case MESA_SHADER_VERTEX
:
627 const_state
->offsets
.primitive_param
= constoff
;
630 case MESA_SHADER_TESS_CTRL
:
631 case MESA_SHADER_TESS_EVAL
:
632 constoff
= align(constoff
- 1, 4) + 3;
633 const_state
->offsets
.primitive_param
= constoff
;
634 const_state
->offsets
.primitive_map
= constoff
+ 5;
635 constoff
+= 5 + DIV_ROUND_UP(nir
->num_inputs
, 4);
637 case MESA_SHADER_GEOMETRY
:
638 const_state
->offsets
.primitive_param
= constoff
;
639 const_state
->offsets
.primitive_map
= constoff
+ 1;
640 constoff
+= 1 + DIV_ROUND_UP(nir
->num_inputs
, 4);
646 const_state
->offsets
.immediate
= constoff
;
648 assert(constoff
<= ir3_max_const(v
));