2 * Copyright © 2014 Intel 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
25 #include "brw_shader.h"
26 #include "dev/gen_debug.h"
27 #include "compiler/glsl_types.h"
28 #include "compiler/nir/nir_builder.h"
29 #include "util/u_math.h"
32 remap_tess_levels(nir_builder
*b
, nir_intrinsic_instr
*intr
,
33 GLenum primitive_mode
)
35 const int location
= nir_intrinsic_base(intr
);
36 const unsigned component
= nir_intrinsic_component(intr
);
39 if (location
== VARYING_SLOT_TESS_LEVEL_INNER
) {
40 switch (primitive_mode
) {
42 /* gl_TessLevelInner[0..1] lives at DWords 3-2 (reversed). */
43 nir_intrinsic_set_base(intr
, 0);
44 nir_intrinsic_set_component(intr
, 3 - component
);
45 out_of_bounds
= false;
48 /* gl_TessLevelInner[0] lives at DWord 4. */
49 nir_intrinsic_set_base(intr
, 1);
50 out_of_bounds
= component
> 0;
56 unreachable("Bogus tessellation domain");
58 } else if (location
== VARYING_SLOT_TESS_LEVEL_OUTER
) {
59 if (primitive_mode
== GL_ISOLINES
) {
60 /* gl_TessLevelOuter[0..1] lives at DWords 6-7 (in order). */
61 nir_intrinsic_set_base(intr
, 1);
62 nir_intrinsic_set_component(intr
, 2 + nir_intrinsic_component(intr
));
63 out_of_bounds
= component
> 1;
65 /* Triangles use DWords 7-5 (reversed); Quads use 7-4 (reversed) */
66 nir_intrinsic_set_base(intr
, 1);
67 nir_intrinsic_set_component(intr
, 3 - nir_intrinsic_component(intr
));
68 out_of_bounds
= component
== 3 && primitive_mode
== GL_TRIANGLES
;
75 if (nir_intrinsic_infos
[intr
->intrinsic
].has_dest
) {
76 b
->cursor
= nir_before_instr(&intr
->instr
);
77 nir_ssa_def
*undef
= nir_ssa_undef(b
, 1, 32);
78 nir_ssa_def_rewrite_uses(&intr
->dest
.ssa
, nir_src_for_ssa(undef
));
80 nir_instr_remove(&intr
->instr
);
87 is_input(nir_intrinsic_instr
*intrin
)
89 return intrin
->intrinsic
== nir_intrinsic_load_input
||
90 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_input
||
91 intrin
->intrinsic
== nir_intrinsic_load_interpolated_input
;
95 is_output(nir_intrinsic_instr
*intrin
)
97 return intrin
->intrinsic
== nir_intrinsic_load_output
||
98 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_output
||
99 intrin
->intrinsic
== nir_intrinsic_store_output
||
100 intrin
->intrinsic
== nir_intrinsic_store_per_vertex_output
;
105 remap_patch_urb_offsets(nir_block
*block
, nir_builder
*b
,
106 const struct brw_vue_map
*vue_map
,
107 GLenum tes_primitive_mode
)
109 const bool is_passthrough_tcs
= b
->shader
->info
.name
&&
110 strcmp(b
->shader
->info
.name
, "passthrough") == 0;
112 nir_foreach_instr_safe(instr
, block
) {
113 if (instr
->type
!= nir_instr_type_intrinsic
)
116 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
118 gl_shader_stage stage
= b
->shader
->info
.stage
;
120 if ((stage
== MESA_SHADER_TESS_CTRL
&& is_output(intrin
)) ||
121 (stage
== MESA_SHADER_TESS_EVAL
&& is_input(intrin
))) {
123 if (!is_passthrough_tcs
&&
124 remap_tess_levels(b
, intrin
, tes_primitive_mode
))
127 int vue_slot
= vue_map
->varying_to_slot
[intrin
->const_index
[0]];
128 assert(vue_slot
!= -1);
129 intrin
->const_index
[0] = vue_slot
;
131 nir_src
*vertex
= nir_get_io_vertex_index_src(intrin
);
133 if (nir_src_is_const(*vertex
)) {
134 intrin
->const_index
[0] += nir_src_as_uint(*vertex
) *
135 vue_map
->num_per_vertex_slots
;
137 b
->cursor
= nir_before_instr(&intrin
->instr
);
139 /* Multiply by the number of per-vertex slots. */
140 nir_ssa_def
*vertex_offset
=
142 nir_ssa_for_src(b
, *vertex
, 1),
144 vue_map
->num_per_vertex_slots
));
146 /* Add it to the existing offset */
147 nir_src
*offset
= nir_get_io_offset_src(intrin
);
148 nir_ssa_def
*total_offset
=
149 nir_iadd(b
, vertex_offset
,
150 nir_ssa_for_src(b
, *offset
, 1));
152 nir_instr_rewrite_src(&intrin
->instr
, offset
,
153 nir_src_for_ssa(total_offset
));
162 brw_nir_lower_vs_inputs(nir_shader
*nir
,
163 const uint8_t *vs_attrib_wa_flags
)
165 /* Start with the location of the variable's base. */
166 foreach_list_typed(nir_variable
, var
, node
, &nir
->inputs
) {
167 var
->data
.driver_location
= var
->data
.location
;
170 /* Now use nir_lower_io to walk dereference chains. Attribute arrays are
171 * loaded as one vec4 or dvec4 per element (or matrix column), depending on
172 * whether it is a double-precision type or not.
174 nir_lower_io(nir
, nir_var_shader_in
, type_size_vec4
,
175 nir_lower_io_lower_64bit_to_32
);
177 /* This pass needs actual constants */
178 nir_opt_constant_folding(nir
);
180 nir_io_add_const_offset_to_base(nir
, nir_var_shader_in
);
182 brw_nir_apply_attribute_workarounds(nir
, vs_attrib_wa_flags
);
184 /* The last step is to remap VERT_ATTRIB_* to actual registers */
186 /* Whether or not we have any system generated values. gl_DrawID is not
187 * included here as it lives in its own vec4.
189 const bool has_sgvs
=
190 nir
->info
.system_values_read
&
191 (BITFIELD64_BIT(SYSTEM_VALUE_FIRST_VERTEX
) |
192 BITFIELD64_BIT(SYSTEM_VALUE_BASE_INSTANCE
) |
193 BITFIELD64_BIT(SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
) |
194 BITFIELD64_BIT(SYSTEM_VALUE_INSTANCE_ID
));
196 const unsigned num_inputs
= util_bitcount64(nir
->info
.inputs_read
);
198 nir_foreach_function(function
, nir
) {
203 nir_builder_init(&b
, function
->impl
);
205 nir_foreach_block(block
, function
->impl
) {
206 nir_foreach_instr_safe(instr
, block
) {
207 if (instr
->type
!= nir_instr_type_intrinsic
)
210 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
212 switch (intrin
->intrinsic
) {
213 case nir_intrinsic_load_first_vertex
:
214 case nir_intrinsic_load_base_instance
:
215 case nir_intrinsic_load_vertex_id_zero_base
:
216 case nir_intrinsic_load_instance_id
:
217 case nir_intrinsic_load_is_indexed_draw
:
218 case nir_intrinsic_load_draw_id
: {
219 b
.cursor
= nir_after_instr(&intrin
->instr
);
221 /* gl_VertexID and friends are stored by the VF as the last
222 * vertex element. We convert them to load_input intrinsics at
223 * the right location.
225 nir_intrinsic_instr
*load
=
226 nir_intrinsic_instr_create(nir
, nir_intrinsic_load_input
);
227 load
->src
[0] = nir_src_for_ssa(nir_imm_int(&b
, 0));
229 nir_intrinsic_set_base(load
, num_inputs
);
230 switch (intrin
->intrinsic
) {
231 case nir_intrinsic_load_first_vertex
:
232 nir_intrinsic_set_component(load
, 0);
234 case nir_intrinsic_load_base_instance
:
235 nir_intrinsic_set_component(load
, 1);
237 case nir_intrinsic_load_vertex_id_zero_base
:
238 nir_intrinsic_set_component(load
, 2);
240 case nir_intrinsic_load_instance_id
:
241 nir_intrinsic_set_component(load
, 3);
243 case nir_intrinsic_load_draw_id
:
244 case nir_intrinsic_load_is_indexed_draw
:
245 /* gl_DrawID and IsIndexedDraw are stored right after
246 * gl_VertexID and friends if any of them exist.
248 nir_intrinsic_set_base(load
, num_inputs
+ has_sgvs
);
249 if (intrin
->intrinsic
== nir_intrinsic_load_draw_id
)
250 nir_intrinsic_set_component(load
, 0);
252 nir_intrinsic_set_component(load
, 1);
255 unreachable("Invalid system value intrinsic");
258 load
->num_components
= 1;
259 nir_ssa_dest_init(&load
->instr
, &load
->dest
, 1, 32, NULL
);
260 nir_builder_instr_insert(&b
, &load
->instr
);
262 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
263 nir_src_for_ssa(&load
->dest
.ssa
));
264 nir_instr_remove(&intrin
->instr
);
268 case nir_intrinsic_load_input
: {
269 /* Attributes come in a contiguous block, ordered by their
270 * gl_vert_attrib value. That means we can compute the slot
271 * number for an attribute by masking out the enabled attributes
272 * before it and counting the bits.
274 int attr
= nir_intrinsic_base(intrin
);
275 int slot
= util_bitcount64(nir
->info
.inputs_read
&
276 BITFIELD64_MASK(attr
));
277 nir_intrinsic_set_base(intrin
, slot
);
282 break; /* Nothing to do */
290 brw_nir_lower_vue_inputs(nir_shader
*nir
,
291 const struct brw_vue_map
*vue_map
)
293 foreach_list_typed(nir_variable
, var
, node
, &nir
->inputs
) {
294 var
->data
.driver_location
= var
->data
.location
;
297 /* Inputs are stored in vec4 slots, so use type_size_vec4(). */
298 nir_lower_io(nir
, nir_var_shader_in
, type_size_vec4
,
299 nir_lower_io_lower_64bit_to_32
);
301 /* This pass needs actual constants */
302 nir_opt_constant_folding(nir
);
304 nir_io_add_const_offset_to_base(nir
, nir_var_shader_in
);
306 nir_foreach_function(function
, nir
) {
310 nir_foreach_block(block
, function
->impl
) {
311 nir_foreach_instr(instr
, block
) {
312 if (instr
->type
!= nir_instr_type_intrinsic
)
315 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
317 if (intrin
->intrinsic
== nir_intrinsic_load_input
||
318 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_input
) {
319 /* Offset 0 is the VUE header, which contains
320 * VARYING_SLOT_LAYER [.y], VARYING_SLOT_VIEWPORT [.z], and
321 * VARYING_SLOT_PSIZ [.w].
323 int varying
= nir_intrinsic_base(intrin
);
326 case VARYING_SLOT_PSIZ
:
327 nir_intrinsic_set_base(intrin
, 0);
328 nir_intrinsic_set_component(intrin
, 3);
332 vue_slot
= vue_map
->varying_to_slot
[varying
];
333 assert(vue_slot
!= -1);
334 nir_intrinsic_set_base(intrin
, vue_slot
);
344 brw_nir_lower_tes_inputs(nir_shader
*nir
, const struct brw_vue_map
*vue_map
)
346 foreach_list_typed(nir_variable
, var
, node
, &nir
->inputs
) {
347 var
->data
.driver_location
= var
->data
.location
;
350 nir_lower_io(nir
, nir_var_shader_in
, type_size_vec4
,
351 nir_lower_io_lower_64bit_to_32
);
353 /* This pass needs actual constants */
354 nir_opt_constant_folding(nir
);
356 nir_io_add_const_offset_to_base(nir
, nir_var_shader_in
);
358 nir_foreach_function(function
, nir
) {
359 if (function
->impl
) {
361 nir_builder_init(&b
, function
->impl
);
362 nir_foreach_block(block
, function
->impl
) {
363 remap_patch_urb_offsets(block
, &b
, vue_map
,
364 nir
->info
.tess
.primitive_mode
);
371 brw_nir_lower_fs_inputs(nir_shader
*nir
,
372 const struct gen_device_info
*devinfo
,
373 const struct brw_wm_prog_key
*key
)
375 foreach_list_typed(nir_variable
, var
, node
, &nir
->inputs
) {
376 var
->data
.driver_location
= var
->data
.location
;
378 /* Apply default interpolation mode.
380 * Everything defaults to smooth except for the legacy GL color
381 * built-in variables, which might be flat depending on API state.
383 if (var
->data
.interpolation
== INTERP_MODE_NONE
) {
384 const bool flat
= key
->flat_shade
&&
385 (var
->data
.location
== VARYING_SLOT_COL0
||
386 var
->data
.location
== VARYING_SLOT_COL1
);
388 var
->data
.interpolation
= flat
? INTERP_MODE_FLAT
389 : INTERP_MODE_SMOOTH
;
392 /* On Ironlake and below, there is only one interpolation mode.
393 * Centroid interpolation doesn't mean anything on this hardware --
394 * there is no multisampling.
396 if (devinfo
->gen
< 6) {
397 var
->data
.centroid
= false;
398 var
->data
.sample
= false;
402 nir_lower_io_options lower_io_options
= nir_lower_io_lower_64bit_to_32
;
403 if (key
->persample_interp
)
404 lower_io_options
|= nir_lower_io_force_sample_interpolation
;
406 nir_lower_io(nir
, nir_var_shader_in
, type_size_vec4
, lower_io_options
);
407 if (devinfo
->gen
>= 11)
408 nir_lower_interpolation(nir
, ~0);
410 /* This pass needs actual constants */
411 nir_opt_constant_folding(nir
);
413 nir_io_add_const_offset_to_base(nir
, nir_var_shader_in
);
417 brw_nir_lower_vue_outputs(nir_shader
*nir
)
419 nir_foreach_variable(var
, &nir
->outputs
) {
420 var
->data
.driver_location
= var
->data
.location
;
423 nir_lower_io(nir
, nir_var_shader_out
, type_size_vec4
,
424 nir_lower_io_lower_64bit_to_32
);
428 brw_nir_lower_tcs_outputs(nir_shader
*nir
, const struct brw_vue_map
*vue_map
,
429 GLenum tes_primitive_mode
)
431 nir_foreach_variable(var
, &nir
->outputs
) {
432 var
->data
.driver_location
= var
->data
.location
;
435 nir_lower_io(nir
, nir_var_shader_out
, type_size_vec4
,
436 nir_lower_io_lower_64bit_to_32
);
438 /* This pass needs actual constants */
439 nir_opt_constant_folding(nir
);
441 nir_io_add_const_offset_to_base(nir
, nir_var_shader_out
);
443 nir_foreach_function(function
, nir
) {
444 if (function
->impl
) {
446 nir_builder_init(&b
, function
->impl
);
447 nir_foreach_block(block
, function
->impl
) {
448 remap_patch_urb_offsets(block
, &b
, vue_map
, tes_primitive_mode
);
455 brw_nir_lower_fs_outputs(nir_shader
*nir
)
457 nir_foreach_variable(var
, &nir
->outputs
) {
458 var
->data
.driver_location
=
459 SET_FIELD(var
->data
.index
, BRW_NIR_FRAG_OUTPUT_INDEX
) |
460 SET_FIELD(var
->data
.location
, BRW_NIR_FRAG_OUTPUT_LOCATION
);
463 nir_lower_io(nir
, nir_var_shader_out
, type_size_dvec4
, 0);
466 #define OPT(pass, ...) ({ \
467 bool this_progress = false; \
468 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
474 static nir_variable_mode
475 brw_nir_no_indirect_mask(const struct brw_compiler
*compiler
,
476 gl_shader_stage stage
)
478 nir_variable_mode indirect_mask
= 0;
480 if (compiler
->glsl_compiler_options
[stage
].EmitNoIndirectInput
)
481 indirect_mask
|= nir_var_shader_in
;
482 if (compiler
->glsl_compiler_options
[stage
].EmitNoIndirectOutput
)
483 indirect_mask
|= nir_var_shader_out
;
484 if (compiler
->glsl_compiler_options
[stage
].EmitNoIndirectTemp
)
485 indirect_mask
|= nir_var_function_temp
;
487 return indirect_mask
;
491 brw_nir_optimize(nir_shader
*nir
, const struct brw_compiler
*compiler
,
492 bool is_scalar
, bool allow_copies
)
494 nir_variable_mode indirect_mask
=
495 brw_nir_no_indirect_mask(compiler
, nir
->info
.stage
);
498 unsigned lower_flrp
=
499 (nir
->options
->lower_flrp16
? 16 : 0) |
500 (nir
->options
->lower_flrp32
? 32 : 0) |
501 (nir
->options
->lower_flrp64
? 64 : 0);
505 OPT(nir_split_array_vars
, nir_var_function_temp
);
506 OPT(nir_shrink_vec_array_vars
, nir_var_function_temp
);
508 OPT(nir_lower_vars_to_ssa
);
510 /* Only run this pass in the first call to brw_nir_optimize. Later
511 * calls assume that we've lowered away any copy_deref instructions
512 * and we don't want to introduce any more.
514 OPT(nir_opt_find_array_copies
);
516 OPT(nir_opt_copy_prop_vars
);
517 OPT(nir_opt_dead_write_vars
);
518 OPT(nir_opt_combine_stores
, nir_var_all
);
521 OPT(nir_lower_alu_to_scalar
, NULL
, NULL
);
527 OPT(nir_lower_phis_to_scalar
);
533 OPT(nir_opt_combine_stores
, nir_var_all
);
535 /* Passing 0 to the peephole select pass causes it to convert
536 * if-statements that contain only move instructions in the branches
537 * regardless of the count.
539 * Passing 1 to the peephole select pass causes it to convert
540 * if-statements that contain at most a single ALU instruction (total)
541 * in both branches. Before Gen6, some math instructions were
542 * prohibitively expensive and the results of compare operations need an
543 * extra resolve step. For these reasons, this pass is more harmful
544 * than good on those platforms.
546 * For indirect loads of uniforms (push constants), we assume that array
547 * indices will nearly always be in bounds and the cost of the load is
548 * low. Therefore there shouldn't be a performance benefit to avoid it.
549 * However, in vec4 tessellation shaders, these loads operate by
550 * actually pulling from memory.
552 const bool is_vec4_tessellation
= !is_scalar
&&
553 (nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
554 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
);
555 OPT(nir_opt_peephole_select
, 0, !is_vec4_tessellation
, false);
556 OPT(nir_opt_peephole_select
, 8, !is_vec4_tessellation
,
557 compiler
->devinfo
->gen
>= 6);
559 OPT(nir_opt_intrinsics
);
560 OPT(nir_opt_idiv_const
, 32);
561 OPT(nir_opt_algebraic
);
562 OPT(nir_opt_constant_folding
);
564 if (lower_flrp
!= 0) {
565 if (OPT(nir_lower_flrp
,
567 false /* always_precise */,
568 compiler
->devinfo
->gen
>= 6)) {
569 OPT(nir_opt_constant_folding
);
572 /* Nothing should rematerialize any flrps, so we only need to do this
578 OPT(nir_opt_dead_cf
);
579 if (OPT(nir_opt_trivial_continues
)) {
580 /* If nir_opt_trivial_continues makes progress, then we need to clean
581 * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
587 OPT(nir_opt_if
, false);
588 OPT(nir_opt_conditional_discard
);
589 if (nir
->options
->max_unroll_iterations
!= 0) {
590 OPT(nir_opt_loop_unroll
, indirect_mask
);
592 OPT(nir_opt_remove_phis
);
597 /* Workaround Gfxbench unused local sampler variable which will trigger an
598 * assert in the opt_large_constants pass.
600 OPT(nir_remove_dead_variables
, nir_var_function_temp
);
604 lower_bit_size_callback(const nir_alu_instr
*alu
, UNUSED
void *data
)
606 assert(alu
->dest
.dest
.is_ssa
);
607 if (alu
->dest
.dest
.ssa
.bit_size
>= 32)
610 const struct brw_compiler
*compiler
= (const struct brw_compiler
*) data
;
621 case nir_op_fround_even
:
632 return compiler
->devinfo
->gen
< 9 ? 32 : 0;
638 /* Does some simple lowering and runs the standard suite of optimizations
640 * This is intended to be called more-or-less directly after you get the
641 * shader out of GLSL or some other source. While it is geared towards i965,
642 * it is not at all generator-specific except for the is_scalar flag. Even
643 * there, it is safe to call with is_scalar = false for a shader that is
644 * intended for the FS backend as long as nir_optimize is called again with
645 * is_scalar = true to scalarize everything prior to code gen.
648 brw_preprocess_nir(const struct brw_compiler
*compiler
, nir_shader
*nir
,
649 const nir_shader
*softfp64
)
651 const struct gen_device_info
*devinfo
= compiler
->devinfo
;
652 UNUSED
bool progress
; /* Written by OPT */
654 const bool is_scalar
= compiler
->scalar_stage
[nir
->info
.stage
];
657 OPT(nir_lower_alu_to_scalar
, NULL
, NULL
);
660 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
)
661 OPT(nir_lower_gs_intrinsics
, false);
663 /* See also brw_nir_trig_workarounds.py */
664 if (compiler
->precise_trig
&&
665 !(devinfo
->gen
>= 10 || devinfo
->is_kabylake
))
666 OPT(brw_nir_apply_trig_workarounds
);
668 static const nir_lower_tex_options tex_options
= {
670 .lower_txf_offset
= true,
671 .lower_rect_offset
= true,
672 .lower_tex_without_implicit_lod
= true,
673 .lower_txd_cube_map
= true,
674 .lower_txb_shadow_clamp
= true,
675 .lower_txd_shadow_clamp
= true,
676 .lower_txd_offset_clamp
= true,
677 .lower_tg4_offsets
= true,
680 OPT(nir_lower_tex
, &tex_options
);
681 OPT(nir_normalize_cubemap_coords
);
683 OPT(nir_lower_global_vars_to_local
);
685 OPT(nir_split_var_copies
);
686 OPT(nir_split_struct_vars
, nir_var_function_temp
);
688 brw_nir_optimize(nir
, compiler
, is_scalar
, true);
690 OPT(nir_lower_doubles
, softfp64
, nir
->options
->lower_doubles_options
);
691 OPT(nir_lower_int64
, nir
->options
->lower_int64_options
);
693 OPT(nir_lower_bit_size
, lower_bit_size_callback
, (void *)compiler
);
696 OPT(nir_lower_load_const_to_scalar
);
699 /* Lower a bunch of stuff */
700 OPT(nir_lower_var_copies
);
702 /* This needs to be run after the first optimization pass but before we
703 * lower indirect derefs away
705 if (compiler
->supports_shader_constants
) {
706 OPT(nir_opt_large_constants
, NULL
, 32);
709 OPT(nir_lower_system_values
);
711 const nir_lower_subgroups_options subgroups_options
= {
712 .ballot_bit_size
= 32,
713 .lower_to_scalar
= true,
714 .lower_vote_trivial
= !is_scalar
,
715 .lower_shuffle
= true,
716 .lower_quad_broadcast_dynamic
= true,
718 OPT(nir_lower_subgroups
, &subgroups_options
);
720 OPT(nir_lower_clip_cull_distance_arrays
);
722 if ((devinfo
->gen
>= 8 || devinfo
->is_haswell
) && is_scalar
) {
723 /* TODO: Yes, we could in theory do this on gen6 and earlier. However,
724 * that would require plumbing through support for these indirect
725 * scratch read/write messages with message registers and that's just a
726 * pain. Also, the primary benefit of this is for compute shaders which
727 * won't run on gen6 and earlier anyway.
729 * On gen7 and earlier the scratch space size is limited to 12kB.
730 * By enabling this optimization we may easily exceed this limit without
731 * having any fallback.
733 * The threshold of 128B was chosen semi-arbitrarily. The idea is that
734 * 128B per channel on a SIMD8 program is 32 registers or 25% of the
735 * register file. Any array that large is likely to cause pressure
736 * issues. Also, this value is sufficiently high that the benchmarks
737 * known to suffer from large temporary array issues are helped but
738 * nothing else in shader-db is hurt except for maybe that one kerbal
739 * space program shader.
741 OPT(nir_lower_vars_to_scratch
, nir_var_function_temp
, 128,
742 glsl_get_natural_size_align_bytes
);
745 nir_variable_mode indirect_mask
=
746 brw_nir_no_indirect_mask(compiler
, nir
->info
.stage
);
747 OPT(nir_lower_indirect_derefs
, indirect_mask
);
749 /* Lower array derefs of vectors for SSBO and UBO loads. For both UBOs and
750 * SSBOs, our back-end is capable of loading an entire vec4 at a time and
751 * we would like to take advantage of that whenever possible regardless of
752 * whether or not the app gives us full loads. This should allow the
753 * optimizer to combine UBO and SSBO load operations and save us some send
756 OPT(nir_lower_array_deref_of_vec
,
757 nir_var_mem_ubo
| nir_var_mem_ssbo
,
758 nir_lower_direct_array_deref_of_vec_load
);
760 /* Get rid of split copies */
761 brw_nir_optimize(nir
, compiler
, is_scalar
, false);
765 brw_nir_link_shaders(const struct brw_compiler
*compiler
,
766 nir_shader
*producer
, nir_shader
*consumer
)
768 nir_lower_io_arrays_to_elements(producer
, consumer
);
769 nir_validate_shader(producer
, "after nir_lower_io_arrays_to_elements");
770 nir_validate_shader(consumer
, "after nir_lower_io_arrays_to_elements");
772 const bool p_is_scalar
= compiler
->scalar_stage
[producer
->info
.stage
];
773 const bool c_is_scalar
= compiler
->scalar_stage
[consumer
->info
.stage
];
775 if (p_is_scalar
&& c_is_scalar
) {
776 NIR_PASS_V(producer
, nir_lower_io_to_scalar_early
, nir_var_shader_out
);
777 NIR_PASS_V(consumer
, nir_lower_io_to_scalar_early
, nir_var_shader_in
);
778 brw_nir_optimize(producer
, compiler
, p_is_scalar
, false);
779 brw_nir_optimize(consumer
, compiler
, c_is_scalar
, false);
782 if (nir_link_opt_varyings(producer
, consumer
))
783 brw_nir_optimize(consumer
, compiler
, c_is_scalar
, false);
785 NIR_PASS_V(producer
, nir_remove_dead_variables
, nir_var_shader_out
);
786 NIR_PASS_V(consumer
, nir_remove_dead_variables
, nir_var_shader_in
);
788 if (nir_remove_unused_varyings(producer
, consumer
)) {
789 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
790 NIR_PASS_V(consumer
, nir_lower_global_vars_to_local
);
792 /* The backend might not be able to handle indirects on
793 * temporaries so we need to lower indirects on any of the
794 * varyings we have demoted here.
796 NIR_PASS_V(producer
, nir_lower_indirect_derefs
,
797 brw_nir_no_indirect_mask(compiler
, producer
->info
.stage
));
798 NIR_PASS_V(consumer
, nir_lower_indirect_derefs
,
799 brw_nir_no_indirect_mask(compiler
, consumer
->info
.stage
));
801 brw_nir_optimize(producer
, compiler
, p_is_scalar
, false);
802 brw_nir_optimize(consumer
, compiler
, c_is_scalar
, false);
805 NIR_PASS_V(producer
, nir_lower_io_to_vector
, nir_var_shader_out
);
806 NIR_PASS_V(producer
, nir_opt_combine_stores
, nir_var_shader_out
);
807 NIR_PASS_V(consumer
, nir_lower_io_to_vector
, nir_var_shader_in
);
809 if (producer
->info
.stage
!= MESA_SHADER_TESS_CTRL
) {
810 /* Calling lower_io_to_vector creates output variable writes with
811 * write-masks. On non-TCS outputs, the back-end can't handle it and we
812 * need to call nir_lower_io_to_temporaries to get rid of them. This,
813 * in turn, creates temporary variables and extra copy_deref intrinsics
814 * that we need to clean up.
816 NIR_PASS_V(producer
, nir_lower_io_to_temporaries
,
817 nir_shader_get_entrypoint(producer
), true, false);
818 NIR_PASS_V(producer
, nir_lower_global_vars_to_local
);
819 NIR_PASS_V(producer
, nir_split_var_copies
);
820 NIR_PASS_V(producer
, nir_lower_var_copies
);
824 /* Prepare the given shader for codegen
826 * This function is intended to be called right before going into the actual
827 * backend and is highly backend-specific. Also, once this function has been
828 * called on a shader, it will no longer be in SSA form so most optimizations
832 brw_postprocess_nir(nir_shader
*nir
, const struct brw_compiler
*compiler
,
835 const struct gen_device_info
*devinfo
= compiler
->devinfo
;
837 (INTEL_DEBUG
& intel_debug_flag_for_shader_stage(nir
->info
.stage
));
839 UNUSED
bool progress
; /* Written by OPT */
841 OPT(brw_nir_lower_mem_access_bit_sizes
, devinfo
);
845 OPT(nir_opt_algebraic_before_ffma
);
848 brw_nir_optimize(nir
, compiler
, is_scalar
, false);
850 if (OPT(nir_lower_int64
, nir
->options
->lower_int64_options
))
851 brw_nir_optimize(nir
, compiler
, is_scalar
, false);
853 if (devinfo
->gen
>= 6) {
854 /* Try and fuse multiply-adds */
855 OPT(brw_nir_opt_peephole_ffma
);
858 if (OPT(nir_opt_comparison_pre
)) {
863 /* Do the select peepehole again. nir_opt_comparison_pre (combined with
864 * the other optimization passes) will have removed at least one
865 * instruction from one of the branches of the if-statement, so now it
866 * might be under the threshold of conversion to bcsel.
868 * See brw_nir_optimize for the explanation of is_vec4_tessellation.
870 const bool is_vec4_tessellation
= !is_scalar
&&
871 (nir
->info
.stage
== MESA_SHADER_TESS_CTRL
||
872 nir
->info
.stage
== MESA_SHADER_TESS_EVAL
);
873 OPT(nir_opt_peephole_select
, 0, is_vec4_tessellation
, false);
874 OPT(nir_opt_peephole_select
, 1, is_vec4_tessellation
,
875 compiler
->devinfo
->gen
>= 6);
880 if (OPT(nir_opt_algebraic_late
)) {
881 /* At this late stage, anything that makes more constants will wreak
882 * havok on the vec4 backend. The handling of constants in the vec4
883 * backend is not good.
886 OPT(nir_opt_constant_folding
);
895 OPT(brw_nir_lower_conversions
);
898 OPT(nir_lower_alu_to_scalar
, NULL
, NULL
);
899 OPT(nir_lower_to_source_mods
, nir_lower_all_source_mods
);
902 OPT(nir_opt_move
, nir_move_comparisons
);
904 OPT(nir_lower_bool_to_int32
);
906 OPT(nir_lower_locals_to_regs
);
908 if (unlikely(debug_enabled
)) {
909 /* Re-index SSA defs so we print more sensible numbers. */
910 nir_foreach_function(function
, nir
) {
912 nir_index_ssa_defs(function
->impl
);
915 fprintf(stderr
, "NIR (SSA form) for %s shader:\n",
916 _mesa_shader_stage_to_string(nir
->info
.stage
));
917 nir_print_shader(nir
, stderr
);
920 OPT(nir_convert_from_ssa
, true);
923 OPT(nir_move_vec_src_uses_to_dest
);
924 OPT(nir_lower_vec_to_movs
);
929 if (OPT(nir_opt_rematerialize_compares
))
932 /* This is the last pass we run before we start emitting stuff. It
933 * determines when we need to insert boolean resolves on Gen <= 5. We
934 * run it last because it stashes data in instr->pass_flags and we don't
935 * want that to be squashed by other NIR passes.
937 if (devinfo
->gen
<= 5)
938 brw_nir_analyze_boolean_resolves(nir
);
942 if (unlikely(debug_enabled
)) {
943 fprintf(stderr
, "NIR (final form) for %s shader:\n",
944 _mesa_shader_stage_to_string(nir
->info
.stage
));
945 nir_print_shader(nir
, stderr
);
950 brw_nir_apply_sampler_key(nir_shader
*nir
,
951 const struct brw_compiler
*compiler
,
952 const struct brw_sampler_prog_key_data
*key_tex
)
954 const struct gen_device_info
*devinfo
= compiler
->devinfo
;
955 nir_lower_tex_options tex_options
= {
956 .lower_txd_clamp_bindless_sampler
= true,
957 .lower_txd_clamp_if_sampler_index_not_lt_16
= true,
960 /* Iron Lake and prior require lowering of all rectangle textures */
961 if (devinfo
->gen
< 6)
962 tex_options
.lower_rect
= true;
964 /* Prior to Broadwell, our hardware can't actually do GL_CLAMP */
965 if (devinfo
->gen
< 8) {
966 tex_options
.saturate_s
= key_tex
->gl_clamp_mask
[0];
967 tex_options
.saturate_t
= key_tex
->gl_clamp_mask
[1];
968 tex_options
.saturate_r
= key_tex
->gl_clamp_mask
[2];
971 /* Prior to Haswell, we have to fake texture swizzle */
972 for (unsigned s
= 0; s
< MAX_SAMPLERS
; s
++) {
973 if (key_tex
->swizzles
[s
] == SWIZZLE_NOOP
)
976 tex_options
.swizzle_result
|= (1 << s
);
977 for (unsigned c
= 0; c
< 4; c
++)
978 tex_options
.swizzles
[s
][c
] = GET_SWZ(key_tex
->swizzles
[s
], c
);
981 /* Prior to Haswell, we have to lower gradients on shadow samplers */
982 tex_options
.lower_txd_shadow
= devinfo
->gen
< 8 && !devinfo
->is_haswell
;
984 tex_options
.lower_y_uv_external
= key_tex
->y_uv_image_mask
;
985 tex_options
.lower_y_u_v_external
= key_tex
->y_u_v_image_mask
;
986 tex_options
.lower_yx_xuxv_external
= key_tex
->yx_xuxv_image_mask
;
987 tex_options
.lower_xy_uxvx_external
= key_tex
->xy_uxvx_image_mask
;
988 tex_options
.lower_ayuv_external
= key_tex
->ayuv_image_mask
;
989 tex_options
.lower_xyuv_external
= key_tex
->xyuv_image_mask
;
991 /* Setup array of scaling factors for each texture. */
992 memcpy(&tex_options
.scale_factors
, &key_tex
->scale_factors
,
993 sizeof(tex_options
.scale_factors
));
995 return nir_lower_tex(nir
, &tex_options
);
999 get_subgroup_size(gl_shader_stage stage
,
1000 const struct brw_base_prog_key
*key
,
1001 unsigned max_subgroup_size
)
1003 switch (key
->subgroup_size_type
) {
1004 case BRW_SUBGROUP_SIZE_API_CONSTANT
:
1005 /* We have to use the global constant size. */
1006 return BRW_SUBGROUP_SIZE
;
1008 case BRW_SUBGROUP_SIZE_UNIFORM
:
1009 /* It has to be uniform across all invocations but can vary per stage
1010 * if we want. This gives us a bit more freedom.
1012 * For compute, brw_nir_apply_key is called per-dispatch-width so this
1013 * is the actual subgroup size and not a maximum. However, we only
1014 * invoke one size of any given compute shader so it's still guaranteed
1015 * to be uniform across invocations.
1017 return max_subgroup_size
;
1019 case BRW_SUBGROUP_SIZE_VARYING
:
1020 /* The subgroup size is allowed to be fully varying. For geometry
1021 * stages, we know it's always 8 which is max_subgroup_size so we can
1022 * return that. For compute, brw_nir_apply_key is called once per
1023 * dispatch-width so max_subgroup_size is the real subgroup size.
1025 * For fragment, we return 0 and let it fall through to the back-end
1026 * compiler. This means we can't optimize based on subgroup size but
1027 * that's a risk the client took when it asked for a varying subgroup
1030 return stage
== MESA_SHADER_FRAGMENT
? 0 : max_subgroup_size
;
1032 case BRW_SUBGROUP_SIZE_REQUIRE_8
:
1033 case BRW_SUBGROUP_SIZE_REQUIRE_16
:
1034 case BRW_SUBGROUP_SIZE_REQUIRE_32
:
1035 assert(stage
== MESA_SHADER_COMPUTE
);
1036 /* These enum values are expressly chosen to be equal to the subgroup
1037 * size that they require.
1039 return key
->subgroup_size_type
;
1042 unreachable("Invalid subgroup size type");
1046 brw_nir_apply_key(nir_shader
*nir
,
1047 const struct brw_compiler
*compiler
,
1048 const struct brw_base_prog_key
*key
,
1049 unsigned max_subgroup_size
,
1052 bool progress
= false;
1054 OPT(brw_nir_apply_sampler_key
, compiler
, &key
->tex
);
1056 const nir_lower_subgroups_options subgroups_options
= {
1057 .subgroup_size
= get_subgroup_size(nir
->info
.stage
, key
,
1059 .ballot_bit_size
= 32,
1060 .lower_subgroup_masks
= true,
1062 OPT(nir_lower_subgroups
, &subgroups_options
);
1065 brw_nir_optimize(nir
, compiler
, is_scalar
, false);
1068 enum brw_conditional_mod
1069 brw_cmod_for_nir_comparison(nir_op op
)
1078 return BRW_CONDITIONAL_L
;
1086 return BRW_CONDITIONAL_GE
;
1092 case nir_op_b32all_fequal2
:
1093 case nir_op_b32all_iequal2
:
1094 case nir_op_b32all_fequal3
:
1095 case nir_op_b32all_iequal3
:
1096 case nir_op_b32all_fequal4
:
1097 case nir_op_b32all_iequal4
:
1098 return BRW_CONDITIONAL_Z
;
1104 case nir_op_b32any_fnequal2
:
1105 case nir_op_b32any_inequal2
:
1106 case nir_op_b32any_fnequal3
:
1107 case nir_op_b32any_inequal3
:
1108 case nir_op_b32any_fnequal4
:
1109 case nir_op_b32any_inequal4
:
1110 return BRW_CONDITIONAL_NZ
;
1113 unreachable("Unsupported NIR comparison op");
1118 brw_aop_for_nir_intrinsic(const nir_intrinsic_instr
*atomic
)
1120 switch (atomic
->intrinsic
) {
1121 #define AOP_CASE(atom) \
1122 case nir_intrinsic_image_atomic_##atom: \
1123 case nir_intrinsic_bindless_image_atomic_##atom: \
1124 case nir_intrinsic_ssbo_atomic_##atom: \
1125 case nir_intrinsic_shared_atomic_##atom: \
1126 case nir_intrinsic_global_atomic_##atom
1130 switch (atomic
->intrinsic
) {
1131 case nir_intrinsic_image_atomic_add
:
1132 case nir_intrinsic_bindless_image_atomic_add
:
1135 case nir_intrinsic_ssbo_atomic_add
:
1138 case nir_intrinsic_shared_atomic_add
:
1139 case nir_intrinsic_global_atomic_add
:
1143 unreachable("Invalid add atomic opcode");
1146 if (nir_src_is_const(atomic
->src
[src_idx
])) {
1147 int64_t add_val
= nir_src_as_int(atomic
->src
[src_idx
]);
1150 else if (add_val
== -1)
1156 AOP_CASE(imin
): return BRW_AOP_IMIN
;
1157 AOP_CASE(umin
): return BRW_AOP_UMIN
;
1158 AOP_CASE(imax
): return BRW_AOP_IMAX
;
1159 AOP_CASE(umax
): return BRW_AOP_UMAX
;
1160 AOP_CASE(and): return BRW_AOP_AND
;
1161 AOP_CASE(or): return BRW_AOP_OR
;
1162 AOP_CASE(xor): return BRW_AOP_XOR
;
1163 AOP_CASE(exchange
): return BRW_AOP_MOV
;
1164 AOP_CASE(comp_swap
): return BRW_AOP_CMPWR
;
1167 #define AOP_CASE(atom) \
1168 case nir_intrinsic_ssbo_atomic_##atom: \
1169 case nir_intrinsic_shared_atomic_##atom: \
1170 case nir_intrinsic_global_atomic_##atom
1172 AOP_CASE(fmin
): return BRW_AOP_FMIN
;
1173 AOP_CASE(fmax
): return BRW_AOP_FMAX
;
1174 AOP_CASE(fcomp_swap
): return BRW_AOP_FCMPWR
;
1179 unreachable("Unsupported NIR atomic intrinsic");
1184 brw_type_for_nir_type(const struct gen_device_info
*devinfo
, nir_alu_type type
)
1188 case nir_type_uint32
:
1189 return BRW_REGISTER_TYPE_UD
;
1192 case nir_type_bool32
:
1193 case nir_type_int32
:
1194 return BRW_REGISTER_TYPE_D
;
1195 case nir_type_float
:
1196 case nir_type_float32
:
1197 return BRW_REGISTER_TYPE_F
;
1198 case nir_type_float16
:
1199 return BRW_REGISTER_TYPE_HF
;
1200 case nir_type_float64
:
1201 return BRW_REGISTER_TYPE_DF
;
1202 case nir_type_int64
:
1203 return devinfo
->gen
< 8 ? BRW_REGISTER_TYPE_DF
: BRW_REGISTER_TYPE_Q
;
1204 case nir_type_uint64
:
1205 return devinfo
->gen
< 8 ? BRW_REGISTER_TYPE_DF
: BRW_REGISTER_TYPE_UQ
;
1206 case nir_type_int16
:
1207 return BRW_REGISTER_TYPE_W
;
1208 case nir_type_uint16
:
1209 return BRW_REGISTER_TYPE_UW
;
1211 return BRW_REGISTER_TYPE_B
;
1212 case nir_type_uint8
:
1213 return BRW_REGISTER_TYPE_UB
;
1215 unreachable("unknown type");
1218 return BRW_REGISTER_TYPE_F
;
1221 /* Returns the glsl_base_type corresponding to a nir_alu_type.
1222 * This is used by both brw_vec4_nir and brw_fs_nir.
1225 brw_glsl_base_type_for_nir_type(nir_alu_type type
)
1228 case nir_type_float
:
1229 case nir_type_float32
:
1230 return GLSL_TYPE_FLOAT
;
1232 case nir_type_float16
:
1233 return GLSL_TYPE_FLOAT16
;
1235 case nir_type_float64
:
1236 return GLSL_TYPE_DOUBLE
;
1239 case nir_type_int32
:
1240 return GLSL_TYPE_INT
;
1243 case nir_type_uint32
:
1244 return GLSL_TYPE_UINT
;
1246 case nir_type_int16
:
1247 return GLSL_TYPE_INT16
;
1249 case nir_type_uint16
:
1250 return GLSL_TYPE_UINT16
;
1253 unreachable("bad type");
1258 brw_nir_create_passthrough_tcs(void *mem_ctx
, const struct brw_compiler
*compiler
,
1259 const nir_shader_compiler_options
*options
,
1260 const struct brw_tcs_prog_key
*key
)
1263 nir_builder_init_simple_shader(&b
, mem_ctx
, MESA_SHADER_TESS_CTRL
,
1265 nir_shader
*nir
= b
.shader
;
1267 nir_intrinsic_instr
*load
;
1268 nir_intrinsic_instr
*store
;
1269 nir_ssa_def
*zero
= nir_imm_int(&b
, 0);
1270 nir_ssa_def
*invoc_id
= nir_load_invocation_id(&b
);
1272 nir
->info
.inputs_read
= key
->outputs_written
&
1273 ~(VARYING_BIT_TESS_LEVEL_INNER
| VARYING_BIT_TESS_LEVEL_OUTER
);
1274 nir
->info
.outputs_written
= key
->outputs_written
;
1275 nir
->info
.tess
.tcs_vertices_out
= key
->input_vertices
;
1276 nir
->info
.name
= ralloc_strdup(nir
, "passthrough");
1277 nir
->num_uniforms
= 8 * sizeof(uint32_t);
1279 var
= nir_variable_create(nir
, nir_var_uniform
, glsl_vec4_type(), "hdr_0");
1280 var
->data
.location
= 0;
1281 var
= nir_variable_create(nir
, nir_var_uniform
, glsl_vec4_type(), "hdr_1");
1282 var
->data
.location
= 1;
1284 /* Write the patch URB header. */
1285 for (int i
= 0; i
<= 1; i
++) {
1286 load
= nir_intrinsic_instr_create(nir
, nir_intrinsic_load_uniform
);
1287 load
->num_components
= 4;
1288 load
->src
[0] = nir_src_for_ssa(zero
);
1289 nir_ssa_dest_init(&load
->instr
, &load
->dest
, 4, 32, NULL
);
1290 nir_intrinsic_set_base(load
, i
* 4 * sizeof(uint32_t));
1291 nir_builder_instr_insert(&b
, &load
->instr
);
1293 store
= nir_intrinsic_instr_create(nir
, nir_intrinsic_store_output
);
1294 store
->num_components
= 4;
1295 store
->src
[0] = nir_src_for_ssa(&load
->dest
.ssa
);
1296 store
->src
[1] = nir_src_for_ssa(zero
);
1297 nir_intrinsic_set_base(store
, VARYING_SLOT_TESS_LEVEL_INNER
- i
);
1298 nir_intrinsic_set_write_mask(store
, WRITEMASK_XYZW
);
1299 nir_builder_instr_insert(&b
, &store
->instr
);
1302 /* Copy inputs to outputs. */
1303 uint64_t varyings
= nir
->info
.inputs_read
;
1305 while (varyings
!= 0) {
1306 const int varying
= ffsll(varyings
) - 1;
1308 load
= nir_intrinsic_instr_create(nir
,
1309 nir_intrinsic_load_per_vertex_input
);
1310 load
->num_components
= 4;
1311 load
->src
[0] = nir_src_for_ssa(invoc_id
);
1312 load
->src
[1] = nir_src_for_ssa(zero
);
1313 nir_ssa_dest_init(&load
->instr
, &load
->dest
, 4, 32, NULL
);
1314 nir_intrinsic_set_base(load
, varying
);
1315 nir_builder_instr_insert(&b
, &load
->instr
);
1317 store
= nir_intrinsic_instr_create(nir
,
1318 nir_intrinsic_store_per_vertex_output
);
1319 store
->num_components
= 4;
1320 store
->src
[0] = nir_src_for_ssa(&load
->dest
.ssa
);
1321 store
->src
[1] = nir_src_for_ssa(invoc_id
);
1322 store
->src
[2] = nir_src_for_ssa(zero
);
1323 nir_intrinsic_set_base(store
, varying
);
1324 nir_intrinsic_set_write_mask(store
, WRITEMASK_XYZW
);
1325 nir_builder_instr_insert(&b
, &store
->instr
);
1327 varyings
&= ~BITFIELD64_BIT(varying
);
1330 nir_validate_shader(nir
, "in brw_nir_create_passthrough_tcs");
1332 brw_preprocess_nir(compiler
, nir
, NULL
);