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 "compiler/nir/glsl_to_nir.h"
27 #include "compiler/nir/nir_builder.h"
28 #include "program/prog_to_nir.h"
31 is_input(nir_intrinsic_instr
*intrin
)
33 return intrin
->intrinsic
== nir_intrinsic_load_input
||
34 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_input
;
38 is_output(nir_intrinsic_instr
*intrin
)
40 return intrin
->intrinsic
== nir_intrinsic_load_output
||
41 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_output
||
42 intrin
->intrinsic
== nir_intrinsic_store_output
||
43 intrin
->intrinsic
== nir_intrinsic_store_per_vertex_output
;
47 * In many cases, we just add the base and offset together, so there's no
48 * reason to keep them separate. Sometimes, combining them is essential:
49 * if a shader only accesses part of a compound variable (such as a matrix
50 * or array), the variable's base may not actually exist in the VUE map.
52 * This pass adds constant offsets to instr->const_index[0], and resets
53 * the offset source to 0. Non-constant offsets remain unchanged - since
54 * we don't know what part of a compound variable is accessed, we allocate
55 * storage for the entire thing.
57 struct add_const_offset_to_base_params
{
59 nir_variable_mode mode
;
63 add_const_offset_to_base_block(nir_block
*block
, void *closure
)
65 struct add_const_offset_to_base_params
*params
= closure
;
66 nir_builder
*b
= ¶ms
->b
;
68 nir_foreach_instr_safe(block
, instr
) {
69 if (instr
->type
!= nir_instr_type_intrinsic
)
72 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
74 if ((params
->mode
== nir_var_shader_in
&& is_input(intrin
)) ||
75 (params
->mode
== nir_var_shader_out
&& is_output(intrin
))) {
76 nir_src
*offset
= nir_get_io_offset_src(intrin
);
77 nir_const_value
*const_offset
= nir_src_as_const_value(*offset
);
80 intrin
->const_index
[0] += const_offset
->u
[0];
81 b
->cursor
= nir_before_instr(&intrin
->instr
);
82 nir_instr_rewrite_src(&intrin
->instr
, offset
,
83 nir_src_for_ssa(nir_imm_int(b
, 0)));
91 add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
)
93 struct add_const_offset_to_base_params params
= { .mode
= mode
};
95 nir_foreach_function(nir
, f
) {
97 nir_builder_init(¶ms
.b
, f
->impl
);
98 nir_foreach_block(f
->impl
, add_const_offset_to_base_block
, ¶ms
);
104 remap_vs_attrs(nir_block
*block
, void *closure
)
106 GLbitfield64 inputs_read
= *((GLbitfield64
*) closure
);
108 nir_foreach_instr(block
, instr
) {
109 if (instr
->type
!= nir_instr_type_intrinsic
)
112 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
114 if (intrin
->intrinsic
== nir_intrinsic_load_input
) {
115 /* Attributes come in a contiguous block, ordered by their
116 * gl_vert_attrib value. That means we can compute the slot
117 * number for an attribute by masking out the enabled attributes
118 * before it and counting the bits.
120 int attr
= intrin
->const_index
[0];
121 int slot
= _mesa_bitcount_64(inputs_read
& BITFIELD64_MASK(attr
));
123 intrin
->const_index
[0] = 4 * slot
;
130 remap_inputs_with_vue_map(nir_block
*block
, void *closure
)
132 const struct brw_vue_map
*vue_map
= closure
;
134 nir_foreach_instr(block
, instr
) {
135 if (instr
->type
!= nir_instr_type_intrinsic
)
138 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
140 if (intrin
->intrinsic
== nir_intrinsic_load_input
||
141 intrin
->intrinsic
== nir_intrinsic_load_per_vertex_input
) {
142 int vue_slot
= vue_map
->varying_to_slot
[intrin
->const_index
[0]];
143 assert(vue_slot
!= -1);
144 intrin
->const_index
[0] = vue_slot
;
150 struct remap_patch_urb_offsets_state
{
152 const struct brw_vue_map
*vue_map
;
156 remap_patch_urb_offsets(nir_block
*block
, void *closure
)
158 struct remap_patch_urb_offsets_state
*state
= closure
;
160 nir_foreach_instr_safe(block
, instr
) {
161 if (instr
->type
!= nir_instr_type_intrinsic
)
164 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
166 gl_shader_stage stage
= state
->b
.shader
->stage
;
168 if ((stage
== MESA_SHADER_TESS_CTRL
&& is_output(intrin
)) ||
169 (stage
== MESA_SHADER_TESS_EVAL
&& is_input(intrin
))) {
170 int vue_slot
= state
->vue_map
->varying_to_slot
[intrin
->const_index
[0]];
171 assert(vue_slot
!= -1);
172 intrin
->const_index
[0] = vue_slot
;
174 nir_src
*vertex
= nir_get_io_vertex_index_src(intrin
);
176 nir_const_value
*const_vertex
= nir_src_as_const_value(*vertex
);
178 intrin
->const_index
[0] += const_vertex
->u
[0] *
179 state
->vue_map
->num_per_vertex_slots
;
181 state
->b
.cursor
= nir_before_instr(&intrin
->instr
);
183 /* Multiply by the number of per-vertex slots. */
184 nir_ssa_def
*vertex_offset
=
186 nir_ssa_for_src(&state
->b
, *vertex
, 1),
187 nir_imm_int(&state
->b
,
188 state
->vue_map
->num_per_vertex_slots
));
190 /* Add it to the existing offset */
191 nir_src
*offset
= nir_get_io_offset_src(intrin
);
192 nir_ssa_def
*total_offset
=
193 nir_iadd(&state
->b
, vertex_offset
,
194 nir_ssa_for_src(&state
->b
, *offset
, 1));
196 nir_instr_rewrite_src(&intrin
->instr
, offset
,
197 nir_src_for_ssa(total_offset
));
206 brw_nir_lower_vs_inputs(nir_shader
*nir
,
207 const struct brw_device_info
*devinfo
,
209 bool use_legacy_snorm_formula
,
210 const uint8_t *vs_attrib_wa_flags
)
212 /* Start with the location of the variable's base. */
213 foreach_list_typed(nir_variable
, var
, node
, &nir
->inputs
) {
214 var
->data
.driver_location
= var
->data
.location
;
217 /* Now use nir_lower_io to walk dereference chains. Attribute arrays
218 * are loaded as one vec4 per element (or matrix column), so we use
219 * type_size_vec4 here.
221 nir_lower_io(nir
, nir_var_shader_in
, type_size_vec4
);
223 /* This pass needs actual constants */
224 nir_opt_constant_folding(nir
);
226 add_const_offset_to_base(nir
, nir_var_shader_in
);
228 brw_nir_apply_attribute_workarounds(nir
, use_legacy_snorm_formula
,
232 /* Finally, translate VERT_ATTRIB_* values into the actual registers.
234 * Note that we can use nir->info.inputs_read instead of
235 * key->inputs_read since the two are identical aside from Gen4-5
236 * edge flag differences.
238 GLbitfield64 inputs_read
= nir
->info
.inputs_read
;
240 nir_foreach_function(nir
, function
) {
241 if (function
->impl
) {
242 nir_foreach_block(function
->impl
, remap_vs_attrs
, &inputs_read
);
249 brw_nir_lower_vue_inputs(nir_shader
*nir
,
250 const struct brw_device_info
*devinfo
,
253 if (!is_scalar
&& nir
->stage
== MESA_SHADER_GEOMETRY
) {
254 foreach_list_typed(nir_variable
, var
, node
, &nir
->inputs
) {
255 var
->data
.driver_location
= var
->data
.location
;
257 nir_lower_io(nir
, nir_var_shader_in
, type_size_vec4
);
259 /* The GLSL linker will have already matched up GS inputs and
260 * the outputs of prior stages. The driver does extend VS outputs
261 * in some cases, but only for legacy OpenGL or Gen4-5 hardware,
262 * neither of which offer geometry shader support. So we can
263 * safely ignore that.
265 * For SSO pipelines, we use a fixed VUE map layout based on variable
266 * locations, so we can rely on rendezvous-by-location to make this
269 * However, we need to ignore VARYING_SLOT_PRIMITIVE_ID, as it's not
270 * written by previous stages and shows up via payload magic.
272 struct brw_vue_map input_vue_map
;
273 GLbitfield64 inputs_read
=
274 nir
->info
.inputs_read
& ~VARYING_BIT_PRIMITIVE_ID
;
275 brw_compute_vue_map(devinfo
, &input_vue_map
, inputs_read
,
276 nir
->info
.separate_shader
||
277 nir
->stage
== MESA_SHADER_TESS_CTRL
);
279 foreach_list_typed(nir_variable
, var
, node
, &nir
->inputs
) {
280 var
->data
.driver_location
= var
->data
.location
;
283 /* Inputs are stored in vec4 slots, so use type_size_vec4(). */
284 nir_lower_io(nir
, nir_var_shader_in
, type_size_vec4
);
286 /* This pass needs actual constants */
287 nir_opt_constant_folding(nir
);
289 add_const_offset_to_base(nir
, nir_var_shader_in
);
291 nir_foreach_function(nir
, function
) {
292 if (function
->impl
) {
293 nir_foreach_block(function
->impl
, remap_inputs_with_vue_map
,
301 brw_nir_lower_tes_inputs(nir_shader
*nir
, const struct brw_vue_map
*vue_map
)
303 struct remap_patch_urb_offsets_state state
;
304 state
.vue_map
= vue_map
;
306 foreach_list_typed(nir_variable
, var
, node
, &nir
->inputs
) {
307 var
->data
.driver_location
= var
->data
.location
;
310 nir_lower_io(nir
, nir_var_shader_in
, type_size_vec4
);
312 /* This pass needs actual constants */
313 nir_opt_constant_folding(nir
);
315 add_const_offset_to_base(nir
, nir_var_shader_in
);
317 nir_foreach_function(nir
, function
) {
318 if (function
->impl
) {
319 nir_builder_init(&state
.b
, function
->impl
);
320 nir_foreach_block(function
->impl
, remap_patch_urb_offsets
, &state
);
326 brw_nir_lower_fs_inputs(nir_shader
*nir
)
328 nir_assign_var_locations(&nir
->inputs
, &nir
->num_inputs
, type_size_scalar
);
329 nir_lower_io(nir
, nir_var_shader_in
, type_size_scalar
);
333 brw_nir_lower_vue_outputs(nir_shader
*nir
,
337 nir_assign_var_locations(&nir
->outputs
, &nir
->num_outputs
,
338 type_size_vec4_times_4
);
339 nir_lower_io(nir
, nir_var_shader_out
, type_size_vec4_times_4
);
341 nir_foreach_variable(var
, &nir
->outputs
)
342 var
->data
.driver_location
= var
->data
.location
;
343 nir_lower_io(nir
, nir_var_shader_out
, type_size_vec4
);
348 brw_nir_lower_tcs_outputs(nir_shader
*nir
, const struct brw_vue_map
*vue_map
)
350 struct remap_patch_urb_offsets_state state
;
351 state
.vue_map
= vue_map
;
353 nir_foreach_variable(var
, &nir
->outputs
) {
354 var
->data
.driver_location
= var
->data
.location
;
357 nir_lower_io(nir
, nir_var_shader_out
, type_size_vec4
);
359 /* This pass needs actual constants */
360 nir_opt_constant_folding(nir
);
362 add_const_offset_to_base(nir
, nir_var_shader_out
);
364 nir_foreach_function(nir
, function
) {
365 if (function
->impl
) {
366 nir_builder_init(&state
.b
, function
->impl
);
367 nir_foreach_block(function
->impl
, remap_patch_urb_offsets
, &state
);
373 brw_nir_lower_fs_outputs(nir_shader
*nir
)
375 nir_assign_var_locations(&nir
->outputs
, &nir
->num_outputs
,
377 nir_lower_io(nir
, nir_var_shader_out
, type_size_scalar
);
381 type_size_scalar_bytes(const struct glsl_type
*type
)
383 return type_size_scalar(type
) * 4;
387 type_size_vec4_bytes(const struct glsl_type
*type
)
389 return type_size_vec4(type
) * 16;
393 brw_nir_lower_uniforms(nir_shader
*nir
, bool is_scalar
)
396 nir_assign_var_locations(&nir
->uniforms
, &nir
->num_uniforms
,
397 type_size_scalar_bytes
);
398 nir_lower_io(nir
, nir_var_uniform
, type_size_scalar_bytes
);
400 nir_assign_var_locations(&nir
->uniforms
, &nir
->num_uniforms
,
401 type_size_vec4_bytes
);
402 nir_lower_io(nir
, nir_var_uniform
, type_size_vec4_bytes
);
406 #define OPT(pass, ...) ({ \
407 bool this_progress = false; \
408 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
414 #define OPT_V(pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
417 nir_optimize(nir_shader
*nir
, bool is_scalar
)
422 OPT_V(nir_lower_vars_to_ssa
);
425 OPT_V(nir_lower_alu_to_scalar
);
431 OPT_V(nir_lower_phis_to_scalar
);
437 OPT(nir_opt_peephole_select
);
438 OPT(nir_opt_algebraic
);
439 OPT(nir_opt_constant_folding
);
440 OPT(nir_opt_dead_cf
);
441 OPT(nir_opt_remove_phis
);
448 /* Does some simple lowering and runs the standard suite of optimizations
450 * This is intended to be called more-or-less directly after you get the
451 * shader out of GLSL or some other source. While it is geared towards i965,
452 * it is not at all generator-specific except for the is_scalar flag. Even
453 * there, it is safe to call with is_scalar = false for a shader that is
454 * intended for the FS backend as long as nir_optimize is called again with
455 * is_scalar = true to scalarize everything prior to code gen.
458 brw_preprocess_nir(nir_shader
*nir
, bool is_scalar
)
460 bool progress
; /* Written by OPT and OPT_V */
463 if (nir
->stage
== MESA_SHADER_GEOMETRY
)
464 OPT(nir_lower_gs_intrinsics
);
466 static const nir_lower_tex_options tex_options
= {
470 OPT(nir_lower_tex
, &tex_options
);
471 OPT(nir_normalize_cubemap_coords
);
473 OPT(nir_lower_global_vars_to_local
);
475 OPT(nir_split_var_copies
);
477 nir
= nir_optimize(nir
, is_scalar
);
480 OPT_V(nir_lower_load_const_to_scalar
);
483 /* Lower a bunch of stuff */
484 OPT_V(nir_lower_var_copies
);
486 /* Get rid of split copies */
487 nir
= nir_optimize(nir
, is_scalar
);
489 OPT(nir_remove_dead_variables
);
494 /* Prepare the given shader for codegen
496 * This function is intended to be called right before going into the actual
497 * backend and is highly backend-specific. Also, once this function has been
498 * called on a shader, it will no longer be in SSA form so most optimizations
502 brw_postprocess_nir(nir_shader
*nir
,
503 const struct brw_device_info
*devinfo
,
507 (INTEL_DEBUG
& intel_debug_flag_for_shader_stage(nir
->stage
));
509 bool progress
; /* Written by OPT and OPT_V */
512 nir
= nir_optimize(nir
, is_scalar
);
514 if (devinfo
->gen
>= 6) {
515 /* Try and fuse multiply-adds */
516 OPT(brw_nir_opt_peephole_ffma
);
519 OPT(nir_opt_algebraic_late
);
521 OPT(nir_lower_locals_to_regs
);
523 OPT_V(nir_lower_to_source_mods
);
527 if (unlikely(debug_enabled
)) {
528 /* Re-index SSA defs so we print more sensible numbers. */
529 nir_foreach_function(nir
, function
) {
531 nir_index_ssa_defs(function
->impl
);
534 fprintf(stderr
, "NIR (SSA form) for %s shader:\n",
535 _mesa_shader_stage_to_string(nir
->stage
));
536 nir_print_shader(nir
, stderr
);
539 OPT_V(nir_convert_from_ssa
, true);
542 OPT_V(nir_move_vec_src_uses_to_dest
);
543 OPT(nir_lower_vec_to_movs
);
546 /* This is the last pass we run before we start emitting stuff. It
547 * determines when we need to insert boolean resolves on Gen <= 5. We
548 * run it last because it stashes data in instr->pass_flags and we don't
549 * want that to be squashed by other NIR passes.
551 if (devinfo
->gen
<= 5)
552 brw_nir_analyze_boolean_resolves(nir
);
556 if (unlikely(debug_enabled
)) {
557 fprintf(stderr
, "NIR (final form) for %s shader:\n",
558 _mesa_shader_stage_to_string(nir
->stage
));
559 nir_print_shader(nir
, stderr
);
566 brw_create_nir(struct brw_context
*brw
,
567 const struct gl_shader_program
*shader_prog
,
568 const struct gl_program
*prog
,
569 gl_shader_stage stage
,
572 struct gl_context
*ctx
= &brw
->ctx
;
573 const nir_shader_compiler_options
*options
=
574 ctx
->Const
.ShaderCompilerOptions
[stage
].NirOptions
;
578 /* First, lower the GLSL IR or Mesa IR to NIR */
580 nir
= glsl_to_nir(shader_prog
, stage
, options
);
582 nir
= prog_to_nir(prog
, options
);
583 OPT_V(nir_convert_to_ssa
); /* turn registers into SSA */
585 nir_validate_shader(nir
);
589 nir
= brw_preprocess_nir(nir
, is_scalar
);
591 OPT(nir_lower_system_values
);
592 OPT_V(brw_nir_lower_uniforms
, is_scalar
);
595 OPT_V(nir_lower_samplers
, shader_prog
);
596 OPT_V(nir_lower_atomics
, shader_prog
);
603 brw_nir_apply_sampler_key(nir_shader
*nir
,
604 const struct brw_device_info
*devinfo
,
605 const struct brw_sampler_prog_key_data
*key_tex
,
608 nir_lower_tex_options tex_options
= { 0 };
610 /* Iron Lake and prior require lowering of all rectangle textures */
611 if (devinfo
->gen
< 6)
612 tex_options
.lower_rect
= true;
614 /* Prior to Broadwell, our hardware can't actually do GL_CLAMP */
615 if (devinfo
->gen
< 8) {
616 tex_options
.saturate_s
= key_tex
->gl_clamp_mask
[0];
617 tex_options
.saturate_t
= key_tex
->gl_clamp_mask
[1];
618 tex_options
.saturate_r
= key_tex
->gl_clamp_mask
[2];
621 /* Prior to Haswell, we have to fake texture swizzle */
622 for (unsigned s
= 0; s
< MAX_SAMPLERS
; s
++) {
623 if (key_tex
->swizzles
[s
] == SWIZZLE_NOOP
)
626 tex_options
.swizzle_result
|= (1 << s
);
627 for (unsigned c
= 0; c
< 4; c
++)
628 tex_options
.swizzles
[s
][c
] = GET_SWZ(key_tex
->swizzles
[s
], c
);
631 if (nir_lower_tex(nir
, &tex_options
)) {
632 nir_validate_shader(nir
);
633 nir
= nir_optimize(nir
, is_scalar
);
640 brw_type_for_nir_type(nir_alu_type type
)
644 return BRW_REGISTER_TYPE_UD
;
647 return BRW_REGISTER_TYPE_D
;
649 return BRW_REGISTER_TYPE_F
;
651 unreachable("unknown type");
654 return BRW_REGISTER_TYPE_F
;
657 /* Returns the glsl_base_type corresponding to a nir_alu_type.
658 * This is used by both brw_vec4_nir and brw_fs_nir.
661 brw_glsl_base_type_for_nir_type(nir_alu_type type
)
665 return GLSL_TYPE_FLOAT
;
668 return GLSL_TYPE_INT
;
671 return GLSL_TYPE_UINT
;
674 unreachable("bad type");