2 * Copyright © 2017 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 shall be included
12 * in all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20 * DEALINGS IN THE SOFTWARE.
24 * @file iris_program.c
26 * This file contains the driver interface for compiling shaders.
28 * See iris_program_cache.c for the in-memory program cache where the
29 * compiled shaders are stored.
34 #include "pipe/p_defines.h"
35 #include "pipe/p_state.h"
36 #include "pipe/p_context.h"
37 #include "pipe/p_screen.h"
38 #include "util/u_atomic.h"
39 #include "util/u_upload_mgr.h"
40 #include "util/debug.h"
41 #include "compiler/nir/nir.h"
42 #include "compiler/nir/nir_builder.h"
43 #include "compiler/nir/nir_serialize.h"
44 #include "intel/compiler/brw_compiler.h"
45 #include "intel/compiler/brw_nir.h"
46 #include "iris_context.h"
47 #include "nir/tgsi_to_nir.h"
49 #define KEY_ID(prefix) .prefix.program_string_id = ish->program_id
50 #define BRW_KEY_INIT(gen, prog_id) \
51 .base.program_string_id = prog_id, \
52 .base.subgroup_size_type = BRW_SUBGROUP_SIZE_UNIFORM, \
53 .base.tex.swizzles[0 ... MAX_SAMPLERS - 1] = 0x688, \
54 .base.tex.compressed_multisample_layout_mask = ~0, \
55 .base.tex.msaa_16 = (gen >= 9 ? ~0 : 0)
58 get_new_program_id(struct iris_screen
*screen
)
60 return p_atomic_inc_return(&screen
->program_id
);
63 static struct brw_vs_prog_key
64 iris_to_brw_vs_key(const struct gen_device_info
*devinfo
,
65 const struct iris_vs_prog_key
*key
)
67 return (struct brw_vs_prog_key
) {
68 BRW_KEY_INIT(devinfo
->gen
, key
->vue
.base
.program_string_id
),
70 /* Don't tell the backend about our clip plane constants, we've
71 * already lowered them in NIR and don't want it doing it again.
73 .nr_userclip_plane_consts
= 0,
77 static struct brw_tcs_prog_key
78 iris_to_brw_tcs_key(const struct gen_device_info
*devinfo
,
79 const struct iris_tcs_prog_key
*key
)
81 return (struct brw_tcs_prog_key
) {
82 BRW_KEY_INIT(devinfo
->gen
, key
->vue
.base
.program_string_id
),
83 .tes_primitive_mode
= key
->tes_primitive_mode
,
84 .input_vertices
= key
->input_vertices
,
85 .patch_outputs_written
= key
->patch_outputs_written
,
86 .outputs_written
= key
->outputs_written
,
87 .quads_workaround
= key
->quads_workaround
,
91 static struct brw_tes_prog_key
92 iris_to_brw_tes_key(const struct gen_device_info
*devinfo
,
93 const struct iris_tes_prog_key
*key
)
95 return (struct brw_tes_prog_key
) {
96 BRW_KEY_INIT(devinfo
->gen
, key
->vue
.base
.program_string_id
),
97 .patch_inputs_read
= key
->patch_inputs_read
,
98 .inputs_read
= key
->inputs_read
,
102 static struct brw_gs_prog_key
103 iris_to_brw_gs_key(const struct gen_device_info
*devinfo
,
104 const struct iris_gs_prog_key
*key
)
106 return (struct brw_gs_prog_key
) {
107 BRW_KEY_INIT(devinfo
->gen
, key
->vue
.base
.program_string_id
),
111 static struct brw_wm_prog_key
112 iris_to_brw_fs_key(const struct gen_device_info
*devinfo
,
113 const struct iris_fs_prog_key
*key
)
115 return (struct brw_wm_prog_key
) {
116 BRW_KEY_INIT(devinfo
->gen
, key
->base
.program_string_id
),
117 .nr_color_regions
= key
->nr_color_regions
,
118 .flat_shade
= key
->flat_shade
,
119 .alpha_test_replicate_alpha
= key
->alpha_test_replicate_alpha
,
120 .alpha_to_coverage
= key
->alpha_to_coverage
,
121 .clamp_fragment_color
= key
->clamp_fragment_color
,
122 .persample_interp
= key
->persample_interp
,
123 .multisample_fbo
= key
->multisample_fbo
,
124 .force_dual_color_blend
= key
->force_dual_color_blend
,
125 .coherent_fb_fetch
= key
->coherent_fb_fetch
,
126 .color_outputs_valid
= key
->color_outputs_valid
,
127 .input_slots_valid
= key
->input_slots_valid
,
131 static struct brw_cs_prog_key
132 iris_to_brw_cs_key(const struct gen_device_info
*devinfo
,
133 const struct iris_cs_prog_key
*key
)
135 return (struct brw_cs_prog_key
) {
136 BRW_KEY_INIT(devinfo
->gen
, key
->base
.program_string_id
),
141 upload_state(struct u_upload_mgr
*uploader
,
142 struct iris_state_ref
*ref
,
147 u_upload_alloc(uploader
, 0, size
, alignment
, &ref
->offset
, &ref
->res
, &p
);
152 iris_upload_ubo_ssbo_surf_state(struct iris_context
*ice
,
153 struct pipe_shader_buffer
*buf
,
154 struct iris_state_ref
*surf_state
,
157 struct pipe_context
*ctx
= &ice
->ctx
;
158 struct iris_screen
*screen
= (struct iris_screen
*) ctx
->screen
;
161 upload_state(ice
->state
.surface_uploader
, surf_state
,
162 screen
->isl_dev
.ss
.size
, 64);
163 if (!unlikely(map
)) {
164 surf_state
->res
= NULL
;
168 struct iris_resource
*res
= (void *) buf
->buffer
;
169 struct iris_bo
*surf_bo
= iris_resource_bo(surf_state
->res
);
170 surf_state
->offset
+= iris_bo_offset_from_base_address(surf_bo
);
172 isl_buffer_fill_state(&screen
->isl_dev
, map
,
173 .address
= res
->bo
->gtt_offset
+ res
->offset
+
175 .size_B
= buf
->buffer_size
- res
->offset
,
176 .format
= ssbo
? ISL_FORMAT_RAW
177 : ISL_FORMAT_R32G32B32A32_FLOAT
,
178 .swizzle
= ISL_SWIZZLE_IDENTITY
,
180 .mocs
= iris_mocs(res
->bo
, &screen
->isl_dev
));
184 get_aoa_deref_offset(nir_builder
*b
,
185 nir_deref_instr
*deref
,
188 unsigned array_size
= elem_size
;
189 nir_ssa_def
*offset
= nir_imm_int(b
, 0);
191 while (deref
->deref_type
!= nir_deref_type_var
) {
192 assert(deref
->deref_type
== nir_deref_type_array
);
194 /* This level's element size is the previous level's array size */
195 nir_ssa_def
*index
= nir_ssa_for_src(b
, deref
->arr
.index
, 1);
196 assert(deref
->arr
.index
.ssa
);
197 offset
= nir_iadd(b
, offset
,
198 nir_imul(b
, index
, nir_imm_int(b
, array_size
)));
200 deref
= nir_deref_instr_parent(deref
);
201 assert(glsl_type_is_array(deref
->type
));
202 array_size
*= glsl_get_length(deref
->type
);
205 /* Accessing an invalid surface index with the dataport can result in a
206 * hang. According to the spec "if the index used to select an individual
207 * element is negative or greater than or equal to the size of the array,
208 * the results of the operation are undefined but may not lead to
209 * termination" -- which is one of the possible outcomes of the hang.
210 * Clamp the index to prevent access outside of the array bounds.
212 return nir_umin(b
, offset
, nir_imm_int(b
, array_size
- elem_size
));
216 iris_lower_storage_image_derefs(nir_shader
*nir
)
218 nir_function_impl
*impl
= nir_shader_get_entrypoint(nir
);
221 nir_builder_init(&b
, impl
);
223 nir_foreach_block(block
, impl
) {
224 nir_foreach_instr_safe(instr
, block
) {
225 if (instr
->type
!= nir_instr_type_intrinsic
)
228 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
229 switch (intrin
->intrinsic
) {
230 case nir_intrinsic_image_deref_load
:
231 case nir_intrinsic_image_deref_store
:
232 case nir_intrinsic_image_deref_atomic_add
:
233 case nir_intrinsic_image_deref_atomic_imin
:
234 case nir_intrinsic_image_deref_atomic_umin
:
235 case nir_intrinsic_image_deref_atomic_imax
:
236 case nir_intrinsic_image_deref_atomic_umax
:
237 case nir_intrinsic_image_deref_atomic_and
:
238 case nir_intrinsic_image_deref_atomic_or
:
239 case nir_intrinsic_image_deref_atomic_xor
:
240 case nir_intrinsic_image_deref_atomic_exchange
:
241 case nir_intrinsic_image_deref_atomic_comp_swap
:
242 case nir_intrinsic_image_deref_size
:
243 case nir_intrinsic_image_deref_samples
:
244 case nir_intrinsic_image_deref_load_raw_intel
:
245 case nir_intrinsic_image_deref_store_raw_intel
: {
246 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
247 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
249 b
.cursor
= nir_before_instr(&intrin
->instr
);
251 nir_iadd(&b
, nir_imm_int(&b
, var
->data
.driver_location
),
252 get_aoa_deref_offset(&b
, deref
, 1));
253 nir_rewrite_image_intrinsic(intrin
, index
, false);
265 * Undo nir_lower_passthrough_edgeflags but keep the inputs_read flag.
268 iris_fix_edge_flags(nir_shader
*nir
)
270 if (nir
->info
.stage
!= MESA_SHADER_VERTEX
)
273 nir_variable
*var
= NULL
;
274 nir_foreach_variable(v
, &nir
->outputs
) {
275 if (v
->data
.location
== VARYING_SLOT_EDGE
) {
284 exec_node_remove(&var
->node
);
285 var
->data
.mode
= nir_var_shader_temp
;
286 exec_list_push_tail(&nir
->globals
, &var
->node
);
287 nir
->info
.outputs_written
&= ~VARYING_BIT_EDGE
;
288 nir
->info
.inputs_read
&= ~VERT_BIT_EDGEFLAG
;
289 nir_fixup_deref_modes(nir
);
291 nir_foreach_function(f
, nir
) {
293 nir_metadata_preserve(f
->impl
, nir_metadata_block_index
|
294 nir_metadata_dominance
|
295 nir_metadata_live_ssa_defs
|
296 nir_metadata_loop_analysis
);
304 * Fix an uncompiled shader's stream output info.
306 * Core Gallium stores output->register_index as a "slot" number, where
307 * slots are assigned consecutively to all outputs in info->outputs_written.
308 * This naive packing of outputs doesn't work for us - we too have slots,
309 * but the layout is defined by the VUE map, which we won't have until we
310 * compile a specific shader variant. So, we remap these and simply store
311 * VARYING_SLOT_* in our copy's output->register_index fields.
313 * We also fix up VARYING_SLOT_{LAYER,VIEWPORT,PSIZ} to select the Y/Z/W
314 * components of our VUE header. See brw_vue_map.c for the layout.
317 update_so_info(struct pipe_stream_output_info
*so_info
,
318 uint64_t outputs_written
)
320 uint8_t reverse_map
[64] = {};
322 while (outputs_written
) {
323 reverse_map
[slot
++] = u_bit_scan64(&outputs_written
);
326 for (unsigned i
= 0; i
< so_info
->num_outputs
; i
++) {
327 struct pipe_stream_output
*output
= &so_info
->output
[i
];
329 /* Map Gallium's condensed "slots" back to real VARYING_SLOT_* enums */
330 output
->register_index
= reverse_map
[output
->register_index
];
332 /* The VUE header contains three scalar fields packed together:
333 * - gl_PointSize is stored in VARYING_SLOT_PSIZ.w
334 * - gl_Layer is stored in VARYING_SLOT_PSIZ.y
335 * - gl_ViewportIndex is stored in VARYING_SLOT_PSIZ.z
337 switch (output
->register_index
) {
338 case VARYING_SLOT_LAYER
:
339 assert(output
->num_components
== 1);
340 output
->register_index
= VARYING_SLOT_PSIZ
;
341 output
->start_component
= 1;
343 case VARYING_SLOT_VIEWPORT
:
344 assert(output
->num_components
== 1);
345 output
->register_index
= VARYING_SLOT_PSIZ
;
346 output
->start_component
= 2;
348 case VARYING_SLOT_PSIZ
:
349 assert(output
->num_components
== 1);
350 output
->start_component
= 3;
354 //info->outputs_written |= 1ull << output->register_index;
359 setup_vec4_image_sysval(uint32_t *sysvals
, uint32_t idx
,
360 unsigned offset
, unsigned n
)
362 assert(offset
% sizeof(uint32_t) == 0);
364 for (unsigned i
= 0; i
< n
; ++i
)
365 sysvals
[i
] = BRW_PARAM_IMAGE(idx
, offset
/ sizeof(uint32_t) + i
);
367 for (unsigned i
= n
; i
< 4; ++i
)
368 sysvals
[i
] = BRW_PARAM_BUILTIN_ZERO
;
372 * Associate NIR uniform variables with the prog_data->param[] mechanism
373 * used by the backend. Also, decide which UBOs we'd like to push in an
374 * ideal situation (though the backend can reduce this).
377 iris_setup_uniforms(const struct brw_compiler
*compiler
,
380 struct brw_stage_prog_data
*prog_data
,
381 enum brw_param_builtin
**out_system_values
,
382 unsigned *out_num_system_values
,
383 unsigned *out_num_cbufs
)
385 UNUSED
const struct gen_device_info
*devinfo
= compiler
->devinfo
;
387 const unsigned IRIS_MAX_SYSTEM_VALUES
=
388 PIPE_MAX_SHADER_IMAGES
* BRW_IMAGE_PARAM_SIZE
;
389 enum brw_param_builtin
*system_values
=
390 rzalloc_array(mem_ctx
, enum brw_param_builtin
, IRIS_MAX_SYSTEM_VALUES
);
391 unsigned num_system_values
= 0;
393 unsigned patch_vert_idx
= -1;
394 unsigned ucp_idx
[IRIS_MAX_CLIP_PLANES
];
395 unsigned img_idx
[PIPE_MAX_SHADER_IMAGES
];
396 memset(ucp_idx
, -1, sizeof(ucp_idx
));
397 memset(img_idx
, -1, sizeof(img_idx
));
399 nir_function_impl
*impl
= nir_shader_get_entrypoint(nir
);
402 nir_builder_init(&b
, impl
);
404 b
.cursor
= nir_before_block(nir_start_block(impl
));
405 nir_ssa_def
*temp_ubo_name
= nir_ssa_undef(&b
, 1, 32);
406 nir_ssa_def
*temp_const_ubo_name
= NULL
;
408 /* Turn system value intrinsics into uniforms */
409 nir_foreach_block(block
, impl
) {
410 nir_foreach_instr_safe(instr
, block
) {
411 if (instr
->type
!= nir_instr_type_intrinsic
)
414 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
417 switch (intrin
->intrinsic
) {
418 case nir_intrinsic_load_constant
: {
419 /* This one is special because it reads from the shader constant
420 * data and not cbuf0 which gallium uploads for us.
422 b
.cursor
= nir_before_instr(instr
);
423 nir_ssa_def
*offset
=
424 nir_iadd_imm(&b
, nir_ssa_for_src(&b
, intrin
->src
[0], 1),
425 nir_intrinsic_base(intrin
));
427 if (temp_const_ubo_name
== NULL
)
428 temp_const_ubo_name
= nir_imm_int(&b
, 0);
430 nir_intrinsic_instr
*load_ubo
=
431 nir_intrinsic_instr_create(b
.shader
, nir_intrinsic_load_ubo
);
432 load_ubo
->num_components
= intrin
->num_components
;
433 load_ubo
->src
[0] = nir_src_for_ssa(temp_const_ubo_name
);
434 load_ubo
->src
[1] = nir_src_for_ssa(offset
);
435 nir_intrinsic_set_align(load_ubo
,
436 nir_intrinsic_align_mul(intrin
),
437 nir_intrinsic_align_offset(intrin
));
438 nir_ssa_dest_init(&load_ubo
->instr
, &load_ubo
->dest
,
439 intrin
->dest
.ssa
.num_components
,
440 intrin
->dest
.ssa
.bit_size
,
441 intrin
->dest
.ssa
.name
);
442 nir_builder_instr_insert(&b
, &load_ubo
->instr
);
444 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
445 nir_src_for_ssa(&load_ubo
->dest
.ssa
));
446 nir_instr_remove(&intrin
->instr
);
449 case nir_intrinsic_load_user_clip_plane
: {
450 unsigned ucp
= nir_intrinsic_ucp_id(intrin
);
452 if (ucp_idx
[ucp
] == -1) {
453 ucp_idx
[ucp
] = num_system_values
;
454 num_system_values
+= 4;
457 for (int i
= 0; i
< 4; i
++) {
458 system_values
[ucp_idx
[ucp
] + i
] =
459 BRW_PARAM_BUILTIN_CLIP_PLANE(ucp
, i
);
462 b
.cursor
= nir_before_instr(instr
);
463 offset
= nir_imm_int(&b
, ucp_idx
[ucp
] * sizeof(uint32_t));
466 case nir_intrinsic_load_patch_vertices_in
:
467 if (patch_vert_idx
== -1)
468 patch_vert_idx
= num_system_values
++;
470 system_values
[patch_vert_idx
] =
471 BRW_PARAM_BUILTIN_PATCH_VERTICES_IN
;
473 b
.cursor
= nir_before_instr(instr
);
474 offset
= nir_imm_int(&b
, patch_vert_idx
* sizeof(uint32_t));
476 case nir_intrinsic_image_deref_load_param_intel
: {
477 assert(devinfo
->gen
< 9);
478 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
479 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
481 if (img_idx
[var
->data
.binding
] == -1) {
482 /* GL only allows arrays of arrays of images. */
483 assert(glsl_type_is_image(glsl_without_array(var
->type
)));
484 unsigned num_images
= MAX2(1, glsl_get_aoa_size(var
->type
));
486 for (int i
= 0; i
< num_images
; i
++) {
487 const unsigned img
= var
->data
.binding
+ i
;
489 img_idx
[img
] = num_system_values
;
490 num_system_values
+= BRW_IMAGE_PARAM_SIZE
;
492 uint32_t *img_sv
= &system_values
[img_idx
[img
]];
494 setup_vec4_image_sysval(
495 img_sv
+ BRW_IMAGE_PARAM_OFFSET_OFFSET
, img
,
496 offsetof(struct brw_image_param
, offset
), 2);
497 setup_vec4_image_sysval(
498 img_sv
+ BRW_IMAGE_PARAM_SIZE_OFFSET
, img
,
499 offsetof(struct brw_image_param
, size
), 3);
500 setup_vec4_image_sysval(
501 img_sv
+ BRW_IMAGE_PARAM_STRIDE_OFFSET
, img
,
502 offsetof(struct brw_image_param
, stride
), 4);
503 setup_vec4_image_sysval(
504 img_sv
+ BRW_IMAGE_PARAM_TILING_OFFSET
, img
,
505 offsetof(struct brw_image_param
, tiling
), 3);
506 setup_vec4_image_sysval(
507 img_sv
+ BRW_IMAGE_PARAM_SWIZZLING_OFFSET
, img
,
508 offsetof(struct brw_image_param
, swizzling
), 2);
512 b
.cursor
= nir_before_instr(instr
);
513 offset
= nir_iadd(&b
,
514 get_aoa_deref_offset(&b
, deref
, BRW_IMAGE_PARAM_SIZE
* 4),
515 nir_imm_int(&b
, img_idx
[var
->data
.binding
] * 4 +
516 nir_intrinsic_base(intrin
) * 16));
523 unsigned comps
= nir_intrinsic_dest_components(intrin
);
525 nir_intrinsic_instr
*load
=
526 nir_intrinsic_instr_create(nir
, nir_intrinsic_load_ubo
);
527 load
->num_components
= comps
;
528 load
->src
[0] = nir_src_for_ssa(temp_ubo_name
);
529 load
->src
[1] = nir_src_for_ssa(offset
);
530 nir_intrinsic_set_align(load
, 4, 0);
531 nir_ssa_dest_init(&load
->instr
, &load
->dest
, comps
, 32, NULL
);
532 nir_builder_instr_insert(&b
, &load
->instr
);
533 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
534 nir_src_for_ssa(&load
->dest
.ssa
));
535 nir_instr_remove(instr
);
539 nir_validate_shader(nir
, "before remapping");
541 /* Uniforms are stored in constant buffer 0, the
542 * user-facing UBOs are indexed by one. So if any constant buffer is
543 * needed, the constant buffer 0 will be needed, so account for it.
545 unsigned num_cbufs
= nir
->info
.num_ubos
;
546 if (num_cbufs
|| nir
->num_uniforms
)
549 /* Place the new params in a new cbuf. */
550 if (num_system_values
> 0) {
551 unsigned sysval_cbuf_index
= num_cbufs
;
554 system_values
= reralloc(mem_ctx
, system_values
, enum brw_param_builtin
,
557 nir_foreach_block(block
, impl
) {
558 nir_foreach_instr_safe(instr
, block
) {
559 if (instr
->type
!= nir_instr_type_intrinsic
)
562 nir_intrinsic_instr
*load
= nir_instr_as_intrinsic(instr
);
564 if (load
->intrinsic
!= nir_intrinsic_load_ubo
)
567 b
.cursor
= nir_before_instr(instr
);
569 assert(load
->src
[0].is_ssa
);
571 if (load
->src
[0].ssa
== temp_ubo_name
) {
572 nir_ssa_def
*imm
= nir_imm_int(&b
, sysval_cbuf_index
);
573 nir_instr_rewrite_src(instr
, &load
->src
[0],
574 nir_src_for_ssa(imm
));
579 /* We need to fold the new iadds for brw_nir_analyze_ubo_ranges */
580 nir_opt_constant_folding(nir
);
582 ralloc_free(system_values
);
583 system_values
= NULL
;
586 assert(num_cbufs
< PIPE_MAX_CONSTANT_BUFFERS
);
587 nir_validate_shader(nir
, "after remap");
589 /* We don't use params[] but gallium leaves num_uniforms set. We use this
590 * to detect when cbuf0 exists but we don't need it anymore when we get
591 * here. Instead, zero it out so that the back-end doesn't get confused
592 * when nr_params * 4 != num_uniforms != nr_params * 4.
594 nir
->num_uniforms
= 0;
596 /* Constant loads (if any) need to go at the end of the constant buffers so
597 * we need to know num_cbufs before we can lower to them.
599 if (temp_const_ubo_name
!= NULL
) {
600 nir_load_const_instr
*const_ubo_index
=
601 nir_instr_as_load_const(temp_const_ubo_name
->parent_instr
);
602 assert(const_ubo_index
->def
.bit_size
== 32);
603 const_ubo_index
->value
[0].u32
= num_cbufs
;
606 *out_system_values
= system_values
;
607 *out_num_system_values
= num_system_values
;
608 *out_num_cbufs
= num_cbufs
;
611 static const char *surface_group_names
[] = {
612 [IRIS_SURFACE_GROUP_RENDER_TARGET
] = "render target",
613 [IRIS_SURFACE_GROUP_RENDER_TARGET_READ
] = "non-coherent render target read",
614 [IRIS_SURFACE_GROUP_CS_WORK_GROUPS
] = "CS work groups",
615 [IRIS_SURFACE_GROUP_TEXTURE
] = "texture",
616 [IRIS_SURFACE_GROUP_UBO
] = "ubo",
617 [IRIS_SURFACE_GROUP_SSBO
] = "ssbo",
618 [IRIS_SURFACE_GROUP_IMAGE
] = "image",
622 iris_print_binding_table(FILE *fp
, const char *name
,
623 const struct iris_binding_table
*bt
)
625 STATIC_ASSERT(ARRAY_SIZE(surface_group_names
) == IRIS_SURFACE_GROUP_COUNT
);
628 uint32_t compacted
= 0;
630 for (int i
= 0; i
< IRIS_SURFACE_GROUP_COUNT
; i
++) {
631 uint32_t size
= bt
->sizes
[i
];
634 compacted
+= util_bitcount64(bt
->used_mask
[i
]);
638 fprintf(fp
, "Binding table for %s is empty\n\n", name
);
642 if (total
!= compacted
) {
643 fprintf(fp
, "Binding table for %s "
644 "(compacted to %u entries from %u entries)\n",
645 name
, compacted
, total
);
647 fprintf(fp
, "Binding table for %s (%u entries)\n", name
, total
);
651 for (int i
= 0; i
< IRIS_SURFACE_GROUP_COUNT
; i
++) {
652 uint64_t mask
= bt
->used_mask
[i
];
654 int index
= u_bit_scan64(&mask
);
655 fprintf(fp
, " [%u] %s #%d\n", entry
++, surface_group_names
[i
], index
);
662 /* Max elements in a surface group. */
663 SURFACE_GROUP_MAX_ELEMENTS
= 64,
667 * Map a <group, index> pair to a binding table index.
669 * For example: <UBO, 5> => binding table index 12
672 iris_group_index_to_bti(const struct iris_binding_table
*bt
,
673 enum iris_surface_group group
, uint32_t index
)
675 assert(index
< bt
->sizes
[group
]);
676 uint64_t mask
= bt
->used_mask
[group
];
677 uint64_t bit
= 1ull << index
;
679 return bt
->offsets
[group
] + util_bitcount64((bit
- 1) & mask
);
681 return IRIS_SURFACE_NOT_USED
;
686 * Map a binding table index back to a <group, index> pair.
688 * For example: binding table index 12 => <UBO, 5>
691 iris_bti_to_group_index(const struct iris_binding_table
*bt
,
692 enum iris_surface_group group
, uint32_t bti
)
694 uint64_t used_mask
= bt
->used_mask
[group
];
695 assert(bti
>= bt
->offsets
[group
]);
697 uint32_t c
= bti
- bt
->offsets
[group
];
699 int i
= u_bit_scan64(&used_mask
);
705 return IRIS_SURFACE_NOT_USED
;
709 rewrite_src_with_bti(nir_builder
*b
, struct iris_binding_table
*bt
,
710 nir_instr
*instr
, nir_src
*src
,
711 enum iris_surface_group group
)
713 assert(bt
->sizes
[group
] > 0);
715 b
->cursor
= nir_before_instr(instr
);
717 if (nir_src_is_const(*src
)) {
718 uint32_t index
= nir_src_as_uint(*src
);
719 bti
= nir_imm_intN_t(b
, iris_group_index_to_bti(bt
, group
, index
),
722 /* Indirect usage makes all the surfaces of the group to be available,
723 * so we can just add the base.
725 assert(bt
->used_mask
[group
] == BITFIELD64_MASK(bt
->sizes
[group
]));
726 bti
= nir_iadd_imm(b
, src
->ssa
, bt
->offsets
[group
]);
728 nir_instr_rewrite_src(instr
, src
, nir_src_for_ssa(bti
));
732 mark_used_with_src(struct iris_binding_table
*bt
, nir_src
*src
,
733 enum iris_surface_group group
)
735 assert(bt
->sizes
[group
] > 0);
737 if (nir_src_is_const(*src
)) {
738 uint64_t index
= nir_src_as_uint(*src
);
739 assert(index
< bt
->sizes
[group
]);
740 bt
->used_mask
[group
] |= 1ull << index
;
742 /* There's an indirect usage, we need all the surfaces. */
743 bt
->used_mask
[group
] = BITFIELD64_MASK(bt
->sizes
[group
]);
748 skip_compacting_binding_tables(void)
750 static int skip
= -1;
752 skip
= env_var_as_boolean("INTEL_DISABLE_COMPACT_BINDING_TABLE", false);
757 * Set up the binding table indices and apply to the shader.
760 iris_setup_binding_table(const struct gen_device_info
*devinfo
,
761 struct nir_shader
*nir
,
762 struct iris_binding_table
*bt
,
763 unsigned num_render_targets
,
764 unsigned num_system_values
,
767 const struct shader_info
*info
= &nir
->info
;
769 memset(bt
, 0, sizeof(*bt
));
771 /* Set the sizes for each surface group. For some groups, we already know
772 * upfront how many will be used, so mark them.
774 if (info
->stage
== MESA_SHADER_FRAGMENT
) {
775 bt
->sizes
[IRIS_SURFACE_GROUP_RENDER_TARGET
] = num_render_targets
;
776 /* All render targets used. */
777 bt
->used_mask
[IRIS_SURFACE_GROUP_RENDER_TARGET
] =
778 BITFIELD64_MASK(num_render_targets
);
780 /* Setup render target read surface group inorder to support non-coherent
781 * framebuffer fetch on Gen8
783 if (devinfo
->gen
== 8 && info
->outputs_read
) {
784 bt
->sizes
[IRIS_SURFACE_GROUP_RENDER_TARGET_READ
] = num_render_targets
;
785 bt
->used_mask
[IRIS_SURFACE_GROUP_RENDER_TARGET_READ
] =
786 BITFIELD64_MASK(num_render_targets
);
788 } else if (info
->stage
== MESA_SHADER_COMPUTE
) {
789 bt
->sizes
[IRIS_SURFACE_GROUP_CS_WORK_GROUPS
] = 1;
792 bt
->sizes
[IRIS_SURFACE_GROUP_TEXTURE
] = util_last_bit(info
->textures_used
);
793 bt
->used_mask
[IRIS_SURFACE_GROUP_TEXTURE
] = info
->textures_used
;
795 bt
->sizes
[IRIS_SURFACE_GROUP_IMAGE
] = info
->num_images
;
797 /* Allocate an extra slot in the UBO section for NIR constants.
798 * Binding table compaction will remove it if unnecessary.
800 * We don't include them in iris_compiled_shader::num_cbufs because
801 * they are uploaded separately from shs->constbuf[], but from a shader
802 * point of view, they're another UBO (at the end of the section).
804 bt
->sizes
[IRIS_SURFACE_GROUP_UBO
] = num_cbufs
+ 1;
806 bt
->sizes
[IRIS_SURFACE_GROUP_SSBO
] = info
->num_ssbos
;
808 for (int i
= 0; i
< IRIS_SURFACE_GROUP_COUNT
; i
++)
809 assert(bt
->sizes
[i
] <= SURFACE_GROUP_MAX_ELEMENTS
);
811 /* Mark surfaces used for the cases we don't have the information available
814 nir_function_impl
*impl
= nir_shader_get_entrypoint(nir
);
815 nir_foreach_block (block
, impl
) {
816 nir_foreach_instr (instr
, block
) {
817 if (instr
->type
!= nir_instr_type_intrinsic
)
820 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
821 switch (intrin
->intrinsic
) {
822 case nir_intrinsic_load_num_work_groups
:
823 bt
->used_mask
[IRIS_SURFACE_GROUP_CS_WORK_GROUPS
] = 1;
826 case nir_intrinsic_load_output
:
827 if (devinfo
->gen
== 8) {
828 mark_used_with_src(bt
, &intrin
->src
[0],
829 IRIS_SURFACE_GROUP_RENDER_TARGET_READ
);
833 case nir_intrinsic_image_size
:
834 case nir_intrinsic_image_load
:
835 case nir_intrinsic_image_store
:
836 case nir_intrinsic_image_atomic_add
:
837 case nir_intrinsic_image_atomic_imin
:
838 case nir_intrinsic_image_atomic_umin
:
839 case nir_intrinsic_image_atomic_imax
:
840 case nir_intrinsic_image_atomic_umax
:
841 case nir_intrinsic_image_atomic_and
:
842 case nir_intrinsic_image_atomic_or
:
843 case nir_intrinsic_image_atomic_xor
:
844 case nir_intrinsic_image_atomic_exchange
:
845 case nir_intrinsic_image_atomic_comp_swap
:
846 case nir_intrinsic_image_load_raw_intel
:
847 case nir_intrinsic_image_store_raw_intel
:
848 mark_used_with_src(bt
, &intrin
->src
[0], IRIS_SURFACE_GROUP_IMAGE
);
851 case nir_intrinsic_load_ubo
:
852 mark_used_with_src(bt
, &intrin
->src
[0], IRIS_SURFACE_GROUP_UBO
);
855 case nir_intrinsic_store_ssbo
:
856 mark_used_with_src(bt
, &intrin
->src
[1], IRIS_SURFACE_GROUP_SSBO
);
859 case nir_intrinsic_get_buffer_size
:
860 case nir_intrinsic_ssbo_atomic_add
:
861 case nir_intrinsic_ssbo_atomic_imin
:
862 case nir_intrinsic_ssbo_atomic_umin
:
863 case nir_intrinsic_ssbo_atomic_imax
:
864 case nir_intrinsic_ssbo_atomic_umax
:
865 case nir_intrinsic_ssbo_atomic_and
:
866 case nir_intrinsic_ssbo_atomic_or
:
867 case nir_intrinsic_ssbo_atomic_xor
:
868 case nir_intrinsic_ssbo_atomic_exchange
:
869 case nir_intrinsic_ssbo_atomic_comp_swap
:
870 case nir_intrinsic_ssbo_atomic_fmin
:
871 case nir_intrinsic_ssbo_atomic_fmax
:
872 case nir_intrinsic_ssbo_atomic_fcomp_swap
:
873 case nir_intrinsic_load_ssbo
:
874 mark_used_with_src(bt
, &intrin
->src
[0], IRIS_SURFACE_GROUP_SSBO
);
883 /* When disable we just mark everything as used. */
884 if (unlikely(skip_compacting_binding_tables())) {
885 for (int i
= 0; i
< IRIS_SURFACE_GROUP_COUNT
; i
++)
886 bt
->used_mask
[i
] = BITFIELD64_MASK(bt
->sizes
[i
]);
889 /* Calculate the offsets and the binding table size based on the used
890 * surfaces. After this point, the functions to go between "group indices"
891 * and binding table indices can be used.
894 for (int i
= 0; i
< IRIS_SURFACE_GROUP_COUNT
; i
++) {
895 if (bt
->used_mask
[i
] != 0) {
896 bt
->offsets
[i
] = next
;
897 next
+= util_bitcount64(bt
->used_mask
[i
]);
900 bt
->size_bytes
= next
* 4;
902 if (unlikely(INTEL_DEBUG
& DEBUG_BT
)) {
903 iris_print_binding_table(stderr
, gl_shader_stage_name(info
->stage
), bt
);
906 /* Apply the binding table indices. The backend compiler is not expected
907 * to change those, as we haven't set any of the *_start entries in brw
911 nir_builder_init(&b
, impl
);
913 nir_foreach_block (block
, impl
) {
914 nir_foreach_instr (instr
, block
) {
915 if (instr
->type
== nir_instr_type_tex
) {
916 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
918 iris_group_index_to_bti(bt
, IRIS_SURFACE_GROUP_TEXTURE
,
923 if (instr
->type
!= nir_instr_type_intrinsic
)
926 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
927 switch (intrin
->intrinsic
) {
928 case nir_intrinsic_image_size
:
929 case nir_intrinsic_image_load
:
930 case nir_intrinsic_image_store
:
931 case nir_intrinsic_image_atomic_add
:
932 case nir_intrinsic_image_atomic_imin
:
933 case nir_intrinsic_image_atomic_umin
:
934 case nir_intrinsic_image_atomic_imax
:
935 case nir_intrinsic_image_atomic_umax
:
936 case nir_intrinsic_image_atomic_and
:
937 case nir_intrinsic_image_atomic_or
:
938 case nir_intrinsic_image_atomic_xor
:
939 case nir_intrinsic_image_atomic_exchange
:
940 case nir_intrinsic_image_atomic_comp_swap
:
941 case nir_intrinsic_image_load_raw_intel
:
942 case nir_intrinsic_image_store_raw_intel
:
943 rewrite_src_with_bti(&b
, bt
, instr
, &intrin
->src
[0],
944 IRIS_SURFACE_GROUP_IMAGE
);
947 case nir_intrinsic_load_ubo
:
948 rewrite_src_with_bti(&b
, bt
, instr
, &intrin
->src
[0],
949 IRIS_SURFACE_GROUP_UBO
);
952 case nir_intrinsic_store_ssbo
:
953 rewrite_src_with_bti(&b
, bt
, instr
, &intrin
->src
[1],
954 IRIS_SURFACE_GROUP_SSBO
);
957 case nir_intrinsic_load_output
:
958 if (devinfo
->gen
== 8) {
959 rewrite_src_with_bti(&b
, bt
, instr
, &intrin
->src
[0],
960 IRIS_SURFACE_GROUP_RENDER_TARGET_READ
);
964 case nir_intrinsic_get_buffer_size
:
965 case nir_intrinsic_ssbo_atomic_add
:
966 case nir_intrinsic_ssbo_atomic_imin
:
967 case nir_intrinsic_ssbo_atomic_umin
:
968 case nir_intrinsic_ssbo_atomic_imax
:
969 case nir_intrinsic_ssbo_atomic_umax
:
970 case nir_intrinsic_ssbo_atomic_and
:
971 case nir_intrinsic_ssbo_atomic_or
:
972 case nir_intrinsic_ssbo_atomic_xor
:
973 case nir_intrinsic_ssbo_atomic_exchange
:
974 case nir_intrinsic_ssbo_atomic_comp_swap
:
975 case nir_intrinsic_ssbo_atomic_fmin
:
976 case nir_intrinsic_ssbo_atomic_fmax
:
977 case nir_intrinsic_ssbo_atomic_fcomp_swap
:
978 case nir_intrinsic_load_ssbo
:
979 rewrite_src_with_bti(&b
, bt
, instr
, &intrin
->src
[0],
980 IRIS_SURFACE_GROUP_SSBO
);
991 iris_debug_recompile(struct iris_context
*ice
,
992 struct shader_info
*info
,
993 const struct brw_base_prog_key
*key
)
995 struct iris_screen
*screen
= (struct iris_screen
*) ice
->ctx
.screen
;
996 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
997 const struct brw_compiler
*c
= screen
->compiler
;
1002 c
->shader_perf_log(&ice
->dbg
, "Recompiling %s shader for program %s: %s\n",
1003 _mesa_shader_stage_to_string(info
->stage
),
1004 info
->name
? info
->name
: "(no identifier)",
1005 info
->label
? info
->label
: "");
1007 const void *old_iris_key
=
1008 iris_find_previous_compile(ice
, info
->stage
, key
->program_string_id
);
1010 union brw_any_prog_key old_key
;
1012 switch (info
->stage
) {
1013 case MESA_SHADER_VERTEX
:
1014 old_key
.vs
= iris_to_brw_vs_key(devinfo
, old_iris_key
);
1016 case MESA_SHADER_TESS_CTRL
:
1017 old_key
.tcs
= iris_to_brw_tcs_key(devinfo
, old_iris_key
);
1019 case MESA_SHADER_TESS_EVAL
:
1020 old_key
.tes
= iris_to_brw_tes_key(devinfo
, old_iris_key
);
1022 case MESA_SHADER_GEOMETRY
:
1023 old_key
.gs
= iris_to_brw_gs_key(devinfo
, old_iris_key
);
1025 case MESA_SHADER_FRAGMENT
:
1026 old_key
.wm
= iris_to_brw_fs_key(devinfo
, old_iris_key
);
1028 case MESA_SHADER_COMPUTE
:
1029 old_key
.cs
= iris_to_brw_cs_key(devinfo
, old_iris_key
);
1032 unreachable("invalid shader stage");
1035 brw_debug_key_recompile(c
, &ice
->dbg
, info
->stage
, &old_key
.base
, key
);
1039 * Get the shader for the last enabled geometry stage.
1041 * This stage is the one which will feed stream output and the rasterizer.
1043 static gl_shader_stage
1044 last_vue_stage(struct iris_context
*ice
)
1046 if (ice
->shaders
.uncompiled
[MESA_SHADER_GEOMETRY
])
1047 return MESA_SHADER_GEOMETRY
;
1049 if (ice
->shaders
.uncompiled
[MESA_SHADER_TESS_EVAL
])
1050 return MESA_SHADER_TESS_EVAL
;
1052 return MESA_SHADER_VERTEX
;
1056 * Compile a vertex shader, and upload the assembly.
1058 static struct iris_compiled_shader
*
1059 iris_compile_vs(struct iris_context
*ice
,
1060 struct iris_uncompiled_shader
*ish
,
1061 const struct iris_vs_prog_key
*key
)
1063 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1064 const struct brw_compiler
*compiler
= screen
->compiler
;
1065 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1066 void *mem_ctx
= ralloc_context(NULL
);
1067 struct brw_vs_prog_data
*vs_prog_data
=
1068 rzalloc(mem_ctx
, struct brw_vs_prog_data
);
1069 struct brw_vue_prog_data
*vue_prog_data
= &vs_prog_data
->base
;
1070 struct brw_stage_prog_data
*prog_data
= &vue_prog_data
->base
;
1071 enum brw_param_builtin
*system_values
;
1072 unsigned num_system_values
;
1075 nir_shader
*nir
= nir_shader_clone(mem_ctx
, ish
->nir
);
1077 if (key
->vue
.nr_userclip_plane_consts
) {
1078 nir_function_impl
*impl
= nir_shader_get_entrypoint(nir
);
1079 nir_lower_clip_vs(nir
, (1 << key
->vue
.nr_userclip_plane_consts
) - 1,
1081 nir_lower_io_to_temporaries(nir
, impl
, true, false);
1082 nir_lower_global_vars_to_local(nir
);
1083 nir_lower_vars_to_ssa(nir
);
1084 nir_shader_gather_info(nir
, impl
);
1087 prog_data
->use_alt_mode
= ish
->use_alt_mode
;
1089 iris_setup_uniforms(compiler
, mem_ctx
, nir
, prog_data
, &system_values
,
1090 &num_system_values
, &num_cbufs
);
1092 struct iris_binding_table bt
;
1093 iris_setup_binding_table(devinfo
, nir
, &bt
, /* num_render_targets */ 0,
1094 num_system_values
, num_cbufs
);
1096 brw_nir_analyze_ubo_ranges(compiler
, nir
, NULL
, prog_data
->ubo_ranges
);
1098 brw_compute_vue_map(devinfo
,
1099 &vue_prog_data
->vue_map
, nir
->info
.outputs_written
,
1100 nir
->info
.separate_shader
, /* pos_slots */ 1);
1102 struct brw_vs_prog_key brw_key
= iris_to_brw_vs_key(devinfo
, key
);
1104 char *error_str
= NULL
;
1105 const unsigned *program
=
1106 brw_compile_vs(compiler
, &ice
->dbg
, mem_ctx
, &brw_key
, vs_prog_data
,
1107 nir
, -1, NULL
, &error_str
);
1108 if (program
== NULL
) {
1109 dbg_printf("Failed to compile vertex shader: %s\n", error_str
);
1110 ralloc_free(mem_ctx
);
1114 if (ish
->compiled_once
) {
1115 iris_debug_recompile(ice
, &nir
->info
, &brw_key
.base
);
1117 ish
->compiled_once
= true;
1120 uint32_t *so_decls
=
1121 screen
->vtbl
.create_so_decl_list(&ish
->stream_output
,
1122 &vue_prog_data
->vue_map
);
1124 struct iris_compiled_shader
*shader
=
1125 iris_upload_shader(ice
, IRIS_CACHE_VS
, sizeof(*key
), key
, program
,
1126 prog_data
, so_decls
, system_values
, num_system_values
,
1129 iris_disk_cache_store(screen
->disk_cache
, ish
, shader
, key
, sizeof(*key
));
1131 ralloc_free(mem_ctx
);
1136 * Update the current vertex shader variant.
1138 * Fill out the key, look in the cache, compile and bind if needed.
1141 iris_update_compiled_vs(struct iris_context
*ice
)
1143 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1144 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_VERTEX
];
1145 struct iris_uncompiled_shader
*ish
=
1146 ice
->shaders
.uncompiled
[MESA_SHADER_VERTEX
];
1148 struct iris_vs_prog_key key
= { KEY_ID(vue
.base
) };
1149 screen
->vtbl
.populate_vs_key(ice
, &ish
->nir
->info
, last_vue_stage(ice
), &key
);
1151 struct iris_compiled_shader
*old
= ice
->shaders
.prog
[IRIS_CACHE_VS
];
1152 struct iris_compiled_shader
*shader
=
1153 iris_find_cached_shader(ice
, IRIS_CACHE_VS
, sizeof(key
), &key
);
1156 shader
= iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
));
1159 shader
= iris_compile_vs(ice
, ish
, &key
);
1161 if (old
!= shader
) {
1162 ice
->shaders
.prog
[IRIS_CACHE_VS
] = shader
;
1163 ice
->state
.dirty
|= IRIS_DIRTY_VS
|
1164 IRIS_DIRTY_BINDINGS_VS
|
1165 IRIS_DIRTY_CONSTANTS_VS
|
1167 shs
->sysvals_need_upload
= true;
1169 const struct brw_vs_prog_data
*vs_prog_data
=
1170 (void *) shader
->prog_data
;
1171 const bool uses_draw_params
= vs_prog_data
->uses_firstvertex
||
1172 vs_prog_data
->uses_baseinstance
;
1173 const bool uses_derived_draw_params
= vs_prog_data
->uses_drawid
||
1174 vs_prog_data
->uses_is_indexed_draw
;
1175 const bool needs_sgvs_element
= uses_draw_params
||
1176 vs_prog_data
->uses_instanceid
||
1177 vs_prog_data
->uses_vertexid
;
1179 if (ice
->state
.vs_uses_draw_params
!= uses_draw_params
||
1180 ice
->state
.vs_uses_derived_draw_params
!= uses_derived_draw_params
||
1181 ice
->state
.vs_needs_edge_flag
!= ish
->needs_edge_flag
) {
1182 ice
->state
.dirty
|= IRIS_DIRTY_VERTEX_BUFFERS
|
1183 IRIS_DIRTY_VERTEX_ELEMENTS
;
1185 ice
->state
.vs_uses_draw_params
= uses_draw_params
;
1186 ice
->state
.vs_uses_derived_draw_params
= uses_derived_draw_params
;
1187 ice
->state
.vs_needs_sgvs_element
= needs_sgvs_element
;
1188 ice
->state
.vs_needs_edge_flag
= ish
->needs_edge_flag
;
1193 * Get the shader_info for a given stage, or NULL if the stage is disabled.
1195 const struct shader_info
*
1196 iris_get_shader_info(const struct iris_context
*ice
, gl_shader_stage stage
)
1198 const struct iris_uncompiled_shader
*ish
= ice
->shaders
.uncompiled
[stage
];
1203 const nir_shader
*nir
= ish
->nir
;
1208 * Get the union of TCS output and TES input slots.
1210 * TCS and TES need to agree on a common URB entry layout. In particular,
1211 * the data for all patch vertices is stored in a single URB entry (unlike
1212 * GS which has one entry per input vertex). This means that per-vertex
1213 * array indexing needs a stride.
1215 * SSO requires locations to match, but doesn't require the number of
1216 * outputs/inputs to match (in fact, the TCS often has extra outputs).
1217 * So, we need to take the extra step of unifying these on the fly.
1220 get_unified_tess_slots(const struct iris_context
*ice
,
1221 uint64_t *per_vertex_slots
,
1222 uint32_t *per_patch_slots
)
1224 const struct shader_info
*tcs
=
1225 iris_get_shader_info(ice
, MESA_SHADER_TESS_CTRL
);
1226 const struct shader_info
*tes
=
1227 iris_get_shader_info(ice
, MESA_SHADER_TESS_EVAL
);
1229 *per_vertex_slots
= tes
->inputs_read
;
1230 *per_patch_slots
= tes
->patch_inputs_read
;
1233 *per_vertex_slots
|= tcs
->outputs_written
;
1234 *per_patch_slots
|= tcs
->patch_outputs_written
;
1239 * Compile a tessellation control shader, and upload the assembly.
1241 static struct iris_compiled_shader
*
1242 iris_compile_tcs(struct iris_context
*ice
,
1243 struct iris_uncompiled_shader
*ish
,
1244 const struct iris_tcs_prog_key
*key
)
1246 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1247 const struct brw_compiler
*compiler
= screen
->compiler
;
1248 const struct nir_shader_compiler_options
*options
=
1249 compiler
->glsl_compiler_options
[MESA_SHADER_TESS_CTRL
].NirOptions
;
1250 void *mem_ctx
= ralloc_context(NULL
);
1251 struct brw_tcs_prog_data
*tcs_prog_data
=
1252 rzalloc(mem_ctx
, struct brw_tcs_prog_data
);
1253 struct brw_vue_prog_data
*vue_prog_data
= &tcs_prog_data
->base
;
1254 struct brw_stage_prog_data
*prog_data
= &vue_prog_data
->base
;
1255 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1256 enum brw_param_builtin
*system_values
= NULL
;
1257 unsigned num_system_values
= 0;
1258 unsigned num_cbufs
= 0;
1262 struct iris_binding_table bt
;
1264 struct brw_tcs_prog_key brw_key
= iris_to_brw_tcs_key(devinfo
, key
);
1267 nir
= nir_shader_clone(mem_ctx
, ish
->nir
);
1269 iris_setup_uniforms(compiler
, mem_ctx
, nir
, prog_data
, &system_values
,
1270 &num_system_values
, &num_cbufs
);
1271 iris_setup_binding_table(devinfo
, nir
, &bt
, /* num_render_targets */ 0,
1272 num_system_values
, num_cbufs
);
1273 brw_nir_analyze_ubo_ranges(compiler
, nir
, NULL
, prog_data
->ubo_ranges
);
1276 brw_nir_create_passthrough_tcs(mem_ctx
, compiler
, options
, &brw_key
);
1278 /* Reserve space for passing the default tess levels as constants. */
1280 num_system_values
= 8;
1282 rzalloc_array(mem_ctx
, enum brw_param_builtin
, num_system_values
);
1283 prog_data
->param
= rzalloc_array(mem_ctx
, uint32_t, num_system_values
);
1284 prog_data
->nr_params
= num_system_values
;
1286 if (key
->tes_primitive_mode
== GL_QUADS
) {
1287 for (int i
= 0; i
< 4; i
++)
1288 system_values
[7 - i
] = BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X
+ i
;
1290 system_values
[3] = BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_X
;
1291 system_values
[2] = BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_Y
;
1292 } else if (key
->tes_primitive_mode
== GL_TRIANGLES
) {
1293 for (int i
= 0; i
< 3; i
++)
1294 system_values
[7 - i
] = BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X
+ i
;
1296 system_values
[4] = BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_X
;
1298 assert(key
->tes_primitive_mode
== GL_ISOLINES
);
1299 system_values
[7] = BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_Y
;
1300 system_values
[6] = BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X
;
1303 /* Manually setup the TCS binding table. */
1304 memset(&bt
, 0, sizeof(bt
));
1305 bt
.sizes
[IRIS_SURFACE_GROUP_UBO
] = 1;
1306 bt
.used_mask
[IRIS_SURFACE_GROUP_UBO
] = 1;
1309 prog_data
->ubo_ranges
[0].length
= 1;
1312 char *error_str
= NULL
;
1313 const unsigned *program
=
1314 brw_compile_tcs(compiler
, &ice
->dbg
, mem_ctx
, &brw_key
, tcs_prog_data
,
1315 nir
, -1, NULL
, &error_str
);
1316 if (program
== NULL
) {
1317 dbg_printf("Failed to compile control shader: %s\n", error_str
);
1318 ralloc_free(mem_ctx
);
1323 if (ish
->compiled_once
) {
1324 iris_debug_recompile(ice
, &nir
->info
, &brw_key
.base
);
1326 ish
->compiled_once
= true;
1330 struct iris_compiled_shader
*shader
=
1331 iris_upload_shader(ice
, IRIS_CACHE_TCS
, sizeof(*key
), key
, program
,
1332 prog_data
, NULL
, system_values
, num_system_values
,
1336 iris_disk_cache_store(screen
->disk_cache
, ish
, shader
, key
, sizeof(*key
));
1338 ralloc_free(mem_ctx
);
1343 * Update the current tessellation control shader variant.
1345 * Fill out the key, look in the cache, compile and bind if needed.
1348 iris_update_compiled_tcs(struct iris_context
*ice
)
1350 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_TESS_CTRL
];
1351 struct iris_uncompiled_shader
*tcs
=
1352 ice
->shaders
.uncompiled
[MESA_SHADER_TESS_CTRL
];
1353 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1354 const struct brw_compiler
*compiler
= screen
->compiler
;
1355 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1357 const struct shader_info
*tes_info
=
1358 iris_get_shader_info(ice
, MESA_SHADER_TESS_EVAL
);
1359 struct iris_tcs_prog_key key
= {
1360 .vue
.base
.program_string_id
= tcs
? tcs
->program_id
: 0,
1361 .tes_primitive_mode
= tes_info
->tess
.primitive_mode
,
1363 !tcs
|| compiler
->use_tcs_8_patch
? ice
->state
.vertices_per_patch
: 0,
1364 .quads_workaround
= devinfo
->gen
< 9 &&
1365 tes_info
->tess
.primitive_mode
== GL_QUADS
&&
1366 tes_info
->tess
.spacing
== TESS_SPACING_EQUAL
,
1368 get_unified_tess_slots(ice
, &key
.outputs_written
,
1369 &key
.patch_outputs_written
);
1370 screen
->vtbl
.populate_tcs_key(ice
, &key
);
1372 struct iris_compiled_shader
*old
= ice
->shaders
.prog
[IRIS_CACHE_TCS
];
1373 struct iris_compiled_shader
*shader
=
1374 iris_find_cached_shader(ice
, IRIS_CACHE_TCS
, sizeof(key
), &key
);
1377 shader
= iris_disk_cache_retrieve(ice
, tcs
, &key
, sizeof(key
));
1380 shader
= iris_compile_tcs(ice
, tcs
, &key
);
1382 if (old
!= shader
) {
1383 ice
->shaders
.prog
[IRIS_CACHE_TCS
] = shader
;
1384 ice
->state
.dirty
|= IRIS_DIRTY_TCS
|
1385 IRIS_DIRTY_BINDINGS_TCS
|
1386 IRIS_DIRTY_CONSTANTS_TCS
;
1387 shs
->sysvals_need_upload
= true;
1392 * Compile a tessellation evaluation shader, and upload the assembly.
1394 static struct iris_compiled_shader
*
1395 iris_compile_tes(struct iris_context
*ice
,
1396 struct iris_uncompiled_shader
*ish
,
1397 const struct iris_tes_prog_key
*key
)
1399 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1400 const struct brw_compiler
*compiler
= screen
->compiler
;
1401 void *mem_ctx
= ralloc_context(NULL
);
1402 struct brw_tes_prog_data
*tes_prog_data
=
1403 rzalloc(mem_ctx
, struct brw_tes_prog_data
);
1404 struct brw_vue_prog_data
*vue_prog_data
= &tes_prog_data
->base
;
1405 struct brw_stage_prog_data
*prog_data
= &vue_prog_data
->base
;
1406 enum brw_param_builtin
*system_values
;
1407 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1408 unsigned num_system_values
;
1411 nir_shader
*nir
= nir_shader_clone(mem_ctx
, ish
->nir
);
1413 if (key
->vue
.nr_userclip_plane_consts
) {
1414 nir_function_impl
*impl
= nir_shader_get_entrypoint(nir
);
1415 nir_lower_clip_vs(nir
, (1 << key
->vue
.nr_userclip_plane_consts
) - 1,
1417 nir_lower_io_to_temporaries(nir
, impl
, true, false);
1418 nir_lower_global_vars_to_local(nir
);
1419 nir_lower_vars_to_ssa(nir
);
1420 nir_shader_gather_info(nir
, impl
);
1423 iris_setup_uniforms(compiler
, mem_ctx
, nir
, prog_data
, &system_values
,
1424 &num_system_values
, &num_cbufs
);
1426 struct iris_binding_table bt
;
1427 iris_setup_binding_table(devinfo
, nir
, &bt
, /* num_render_targets */ 0,
1428 num_system_values
, num_cbufs
);
1430 brw_nir_analyze_ubo_ranges(compiler
, nir
, NULL
, prog_data
->ubo_ranges
);
1432 struct brw_vue_map input_vue_map
;
1433 brw_compute_tess_vue_map(&input_vue_map
, key
->inputs_read
,
1434 key
->patch_inputs_read
);
1436 struct brw_tes_prog_key brw_key
= iris_to_brw_tes_key(devinfo
, key
);
1438 char *error_str
= NULL
;
1439 const unsigned *program
=
1440 brw_compile_tes(compiler
, &ice
->dbg
, mem_ctx
, &brw_key
, &input_vue_map
,
1441 tes_prog_data
, nir
, -1, NULL
, &error_str
);
1442 if (program
== NULL
) {
1443 dbg_printf("Failed to compile evaluation shader: %s\n", error_str
);
1444 ralloc_free(mem_ctx
);
1448 if (ish
->compiled_once
) {
1449 iris_debug_recompile(ice
, &nir
->info
, &brw_key
.base
);
1451 ish
->compiled_once
= true;
1454 uint32_t *so_decls
=
1455 screen
->vtbl
.create_so_decl_list(&ish
->stream_output
,
1456 &vue_prog_data
->vue_map
);
1459 struct iris_compiled_shader
*shader
=
1460 iris_upload_shader(ice
, IRIS_CACHE_TES
, sizeof(*key
), key
, program
,
1461 prog_data
, so_decls
, system_values
, num_system_values
,
1464 iris_disk_cache_store(screen
->disk_cache
, ish
, shader
, key
, sizeof(*key
));
1466 ralloc_free(mem_ctx
);
1471 * Update the current tessellation evaluation shader variant.
1473 * Fill out the key, look in the cache, compile and bind if needed.
1476 iris_update_compiled_tes(struct iris_context
*ice
)
1478 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1479 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_TESS_EVAL
];
1480 struct iris_uncompiled_shader
*ish
=
1481 ice
->shaders
.uncompiled
[MESA_SHADER_TESS_EVAL
];
1483 struct iris_tes_prog_key key
= { KEY_ID(vue
.base
) };
1484 get_unified_tess_slots(ice
, &key
.inputs_read
, &key
.patch_inputs_read
);
1485 screen
->vtbl
.populate_tes_key(ice
, &ish
->nir
->info
, last_vue_stage(ice
), &key
);
1487 struct iris_compiled_shader
*old
= ice
->shaders
.prog
[IRIS_CACHE_TES
];
1488 struct iris_compiled_shader
*shader
=
1489 iris_find_cached_shader(ice
, IRIS_CACHE_TES
, sizeof(key
), &key
);
1492 shader
= iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
));
1495 shader
= iris_compile_tes(ice
, ish
, &key
);
1497 if (old
!= shader
) {
1498 ice
->shaders
.prog
[IRIS_CACHE_TES
] = shader
;
1499 ice
->state
.dirty
|= IRIS_DIRTY_TES
|
1500 IRIS_DIRTY_BINDINGS_TES
|
1501 IRIS_DIRTY_CONSTANTS_TES
;
1502 shs
->sysvals_need_upload
= true;
1505 /* TODO: Could compare and avoid flagging this. */
1506 const struct shader_info
*tes_info
= &ish
->nir
->info
;
1507 if (tes_info
->system_values_read
& (1ull << SYSTEM_VALUE_VERTICES_IN
)) {
1508 ice
->state
.dirty
|= IRIS_DIRTY_CONSTANTS_TES
;
1509 ice
->state
.shaders
[MESA_SHADER_TESS_EVAL
].sysvals_need_upload
= true;
1514 * Compile a geometry shader, and upload the assembly.
1516 static struct iris_compiled_shader
*
1517 iris_compile_gs(struct iris_context
*ice
,
1518 struct iris_uncompiled_shader
*ish
,
1519 const struct iris_gs_prog_key
*key
)
1521 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1522 const struct brw_compiler
*compiler
= screen
->compiler
;
1523 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1524 void *mem_ctx
= ralloc_context(NULL
);
1525 struct brw_gs_prog_data
*gs_prog_data
=
1526 rzalloc(mem_ctx
, struct brw_gs_prog_data
);
1527 struct brw_vue_prog_data
*vue_prog_data
= &gs_prog_data
->base
;
1528 struct brw_stage_prog_data
*prog_data
= &vue_prog_data
->base
;
1529 enum brw_param_builtin
*system_values
;
1530 unsigned num_system_values
;
1533 nir_shader
*nir
= nir_shader_clone(mem_ctx
, ish
->nir
);
1535 if (key
->vue
.nr_userclip_plane_consts
) {
1536 nir_function_impl
*impl
= nir_shader_get_entrypoint(nir
);
1537 nir_lower_clip_gs(nir
, (1 << key
->vue
.nr_userclip_plane_consts
) - 1,
1539 nir_lower_io_to_temporaries(nir
, impl
, true, false);
1540 nir_lower_global_vars_to_local(nir
);
1541 nir_lower_vars_to_ssa(nir
);
1542 nir_shader_gather_info(nir
, impl
);
1545 iris_setup_uniforms(compiler
, mem_ctx
, nir
, prog_data
, &system_values
,
1546 &num_system_values
, &num_cbufs
);
1548 struct iris_binding_table bt
;
1549 iris_setup_binding_table(devinfo
, nir
, &bt
, /* num_render_targets */ 0,
1550 num_system_values
, num_cbufs
);
1552 brw_nir_analyze_ubo_ranges(compiler
, nir
, NULL
, prog_data
->ubo_ranges
);
1554 brw_compute_vue_map(devinfo
,
1555 &vue_prog_data
->vue_map
, nir
->info
.outputs_written
,
1556 nir
->info
.separate_shader
, /* pos_slots */ 1);
1558 struct brw_gs_prog_key brw_key
= iris_to_brw_gs_key(devinfo
, key
);
1560 char *error_str
= NULL
;
1561 const unsigned *program
=
1562 brw_compile_gs(compiler
, &ice
->dbg
, mem_ctx
, &brw_key
, gs_prog_data
,
1563 nir
, NULL
, -1, NULL
, &error_str
);
1564 if (program
== NULL
) {
1565 dbg_printf("Failed to compile geometry shader: %s\n", error_str
);
1566 ralloc_free(mem_ctx
);
1570 if (ish
->compiled_once
) {
1571 iris_debug_recompile(ice
, &nir
->info
, &brw_key
.base
);
1573 ish
->compiled_once
= true;
1576 uint32_t *so_decls
=
1577 screen
->vtbl
.create_so_decl_list(&ish
->stream_output
,
1578 &vue_prog_data
->vue_map
);
1580 struct iris_compiled_shader
*shader
=
1581 iris_upload_shader(ice
, IRIS_CACHE_GS
, sizeof(*key
), key
, program
,
1582 prog_data
, so_decls
, system_values
, num_system_values
,
1585 iris_disk_cache_store(screen
->disk_cache
, ish
, shader
, key
, sizeof(*key
));
1587 ralloc_free(mem_ctx
);
1592 * Update the current geometry shader variant.
1594 * Fill out the key, look in the cache, compile and bind if needed.
1597 iris_update_compiled_gs(struct iris_context
*ice
)
1599 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_GEOMETRY
];
1600 struct iris_uncompiled_shader
*ish
=
1601 ice
->shaders
.uncompiled
[MESA_SHADER_GEOMETRY
];
1602 struct iris_compiled_shader
*old
= ice
->shaders
.prog
[IRIS_CACHE_GS
];
1603 struct iris_compiled_shader
*shader
= NULL
;
1604 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1607 struct iris_gs_prog_key key
= { KEY_ID(vue
.base
) };
1608 screen
->vtbl
.populate_gs_key(ice
, &ish
->nir
->info
, last_vue_stage(ice
), &key
);
1611 iris_find_cached_shader(ice
, IRIS_CACHE_GS
, sizeof(key
), &key
);
1614 shader
= iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
));
1617 shader
= iris_compile_gs(ice
, ish
, &key
);
1620 if (old
!= shader
) {
1621 ice
->shaders
.prog
[IRIS_CACHE_GS
] = shader
;
1622 ice
->state
.dirty
|= IRIS_DIRTY_GS
|
1623 IRIS_DIRTY_BINDINGS_GS
|
1624 IRIS_DIRTY_CONSTANTS_GS
;
1625 shs
->sysvals_need_upload
= true;
1630 * Compile a fragment (pixel) shader, and upload the assembly.
1632 static struct iris_compiled_shader
*
1633 iris_compile_fs(struct iris_context
*ice
,
1634 struct iris_uncompiled_shader
*ish
,
1635 const struct iris_fs_prog_key
*key
,
1636 struct brw_vue_map
*vue_map
)
1638 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1639 const struct brw_compiler
*compiler
= screen
->compiler
;
1640 void *mem_ctx
= ralloc_context(NULL
);
1641 struct brw_wm_prog_data
*fs_prog_data
=
1642 rzalloc(mem_ctx
, struct brw_wm_prog_data
);
1643 struct brw_stage_prog_data
*prog_data
= &fs_prog_data
->base
;
1644 enum brw_param_builtin
*system_values
;
1645 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1646 unsigned num_system_values
;
1649 nir_shader
*nir
= nir_shader_clone(mem_ctx
, ish
->nir
);
1651 prog_data
->use_alt_mode
= ish
->use_alt_mode
;
1653 iris_setup_uniforms(compiler
, mem_ctx
, nir
, prog_data
, &system_values
,
1654 &num_system_values
, &num_cbufs
);
1656 /* Lower output variables to load_output intrinsics before setting up
1657 * binding tables, so iris_setup_binding_table can map any load_output
1658 * intrinsics to IRIS_SURFACE_GROUP_RENDER_TARGET_READ on Gen8 for
1659 * non-coherent framebuffer fetches.
1661 brw_nir_lower_fs_outputs(nir
);
1663 /* On Gen11+, shader RT write messages have a "Null Render Target" bit
1664 * and do not need a binding table entry with a null surface. Earlier
1665 * generations need an entry for a null surface.
1667 int null_rts
= devinfo
->gen
< 11 ? 1 : 0;
1669 struct iris_binding_table bt
;
1670 iris_setup_binding_table(devinfo
, nir
, &bt
,
1671 MAX2(key
->nr_color_regions
, null_rts
),
1672 num_system_values
, num_cbufs
);
1674 brw_nir_analyze_ubo_ranges(compiler
, nir
, NULL
, prog_data
->ubo_ranges
);
1676 struct brw_wm_prog_key brw_key
= iris_to_brw_fs_key(devinfo
, key
);
1678 char *error_str
= NULL
;
1679 const unsigned *program
=
1680 brw_compile_fs(compiler
, &ice
->dbg
, mem_ctx
, &brw_key
, fs_prog_data
,
1681 nir
, -1, -1, -1, true, false, vue_map
,
1683 if (program
== NULL
) {
1684 dbg_printf("Failed to compile fragment shader: %s\n", error_str
);
1685 ralloc_free(mem_ctx
);
1689 if (ish
->compiled_once
) {
1690 iris_debug_recompile(ice
, &nir
->info
, &brw_key
.base
);
1692 ish
->compiled_once
= true;
1695 struct iris_compiled_shader
*shader
=
1696 iris_upload_shader(ice
, IRIS_CACHE_FS
, sizeof(*key
), key
, program
,
1697 prog_data
, NULL
, system_values
, num_system_values
,
1700 iris_disk_cache_store(screen
->disk_cache
, ish
, shader
, key
, sizeof(*key
));
1702 ralloc_free(mem_ctx
);
1707 * Update the current fragment shader variant.
1709 * Fill out the key, look in the cache, compile and bind if needed.
1712 iris_update_compiled_fs(struct iris_context
*ice
)
1714 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_FRAGMENT
];
1715 struct iris_uncompiled_shader
*ish
=
1716 ice
->shaders
.uncompiled
[MESA_SHADER_FRAGMENT
];
1717 struct iris_fs_prog_key key
= { KEY_ID(base
) };
1718 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1719 screen
->vtbl
.populate_fs_key(ice
, &ish
->nir
->info
, &key
);
1721 if (ish
->nos
& (1ull << IRIS_NOS_LAST_VUE_MAP
))
1722 key
.input_slots_valid
= ice
->shaders
.last_vue_map
->slots_valid
;
1724 struct iris_compiled_shader
*old
= ice
->shaders
.prog
[IRIS_CACHE_FS
];
1725 struct iris_compiled_shader
*shader
=
1726 iris_find_cached_shader(ice
, IRIS_CACHE_FS
, sizeof(key
), &key
);
1729 shader
= iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
));
1732 shader
= iris_compile_fs(ice
, ish
, &key
, ice
->shaders
.last_vue_map
);
1734 if (old
!= shader
) {
1735 // XXX: only need to flag CLIP if barycentric has NONPERSPECTIVE
1736 // toggles. might be able to avoid flagging SBE too.
1737 ice
->shaders
.prog
[IRIS_CACHE_FS
] = shader
;
1738 ice
->state
.dirty
|= IRIS_DIRTY_FS
|
1739 IRIS_DIRTY_BINDINGS_FS
|
1740 IRIS_DIRTY_CONSTANTS_FS
|
1744 shs
->sysvals_need_upload
= true;
1749 * Update the last enabled stage's VUE map.
1751 * When the shader feeding the rasterizer's output interface changes, we
1752 * need to re-emit various packets.
1755 update_last_vue_map(struct iris_context
*ice
,
1756 struct brw_stage_prog_data
*prog_data
)
1758 struct brw_vue_prog_data
*vue_prog_data
= (void *) prog_data
;
1759 struct brw_vue_map
*vue_map
= &vue_prog_data
->vue_map
;
1760 struct brw_vue_map
*old_map
= ice
->shaders
.last_vue_map
;
1761 const uint64_t changed_slots
=
1762 (old_map
? old_map
->slots_valid
: 0ull) ^ vue_map
->slots_valid
;
1764 if (changed_slots
& VARYING_BIT_VIEWPORT
) {
1765 ice
->state
.num_viewports
=
1766 (vue_map
->slots_valid
& VARYING_BIT_VIEWPORT
) ? IRIS_MAX_VIEWPORTS
: 1;
1767 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
|
1768 IRIS_DIRTY_SF_CL_VIEWPORT
|
1769 IRIS_DIRTY_CC_VIEWPORT
|
1770 IRIS_DIRTY_SCISSOR_RECT
|
1771 IRIS_DIRTY_UNCOMPILED_FS
|
1772 ice
->state
.dirty_for_nos
[IRIS_NOS_LAST_VUE_MAP
];
1775 if (changed_slots
|| (old_map
&& old_map
->separate
!= vue_map
->separate
)) {
1776 ice
->state
.dirty
|= IRIS_DIRTY_SBE
;
1779 ice
->shaders
.last_vue_map
= &vue_prog_data
->vue_map
;
1783 iris_update_pull_constant_descriptors(struct iris_context
*ice
,
1784 gl_shader_stage stage
)
1786 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[stage
];
1788 if (!shader
|| !shader
->prog_data
->has_ubo_pull
)
1791 struct iris_shader_state
*shs
= &ice
->state
.shaders
[stage
];
1792 bool any_new_descriptors
=
1793 shader
->num_system_values
> 0 && shs
->sysvals_need_upload
;
1795 unsigned bound_cbufs
= shs
->bound_cbufs
;
1797 while (bound_cbufs
) {
1798 const int i
= u_bit_scan(&bound_cbufs
);
1799 struct pipe_shader_buffer
*cbuf
= &shs
->constbuf
[i
];
1800 struct iris_state_ref
*surf_state
= &shs
->constbuf_surf_state
[i
];
1801 if (!surf_state
->res
&& cbuf
->buffer
) {
1802 iris_upload_ubo_ssbo_surf_state(ice
, cbuf
, surf_state
, false);
1803 any_new_descriptors
= true;
1807 if (any_new_descriptors
)
1808 ice
->state
.dirty
|= IRIS_DIRTY_BINDINGS_VS
<< stage
;
1812 * Get the prog_data for a given stage, or NULL if the stage is disabled.
1814 static struct brw_vue_prog_data
*
1815 get_vue_prog_data(struct iris_context
*ice
, gl_shader_stage stage
)
1817 if (!ice
->shaders
.prog
[stage
])
1820 return (void *) ice
->shaders
.prog
[stage
]->prog_data
;
1823 // XXX: iris_compiled_shaders are space-leaking :(
1824 // XXX: do remember to unbind them if deleting them.
1827 * Update the current shader variants for the given state.
1829 * This should be called on every draw call to ensure that the correct
1830 * shaders are bound. It will also flag any dirty state triggered by
1831 * swapping out those shaders.
1834 iris_update_compiled_shaders(struct iris_context
*ice
)
1836 const uint64_t dirty
= ice
->state
.dirty
;
1838 struct brw_vue_prog_data
*old_prog_datas
[4];
1839 if (!(dirty
& IRIS_DIRTY_URB
)) {
1840 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++)
1841 old_prog_datas
[i
] = get_vue_prog_data(ice
, i
);
1844 if (dirty
& (IRIS_DIRTY_UNCOMPILED_TCS
| IRIS_DIRTY_UNCOMPILED_TES
)) {
1845 struct iris_uncompiled_shader
*tes
=
1846 ice
->shaders
.uncompiled
[MESA_SHADER_TESS_EVAL
];
1848 iris_update_compiled_tcs(ice
);
1849 iris_update_compiled_tes(ice
);
1851 ice
->shaders
.prog
[IRIS_CACHE_TCS
] = NULL
;
1852 ice
->shaders
.prog
[IRIS_CACHE_TES
] = NULL
;
1854 IRIS_DIRTY_TCS
| IRIS_DIRTY_TES
|
1855 IRIS_DIRTY_BINDINGS_TCS
| IRIS_DIRTY_BINDINGS_TES
|
1856 IRIS_DIRTY_CONSTANTS_TCS
| IRIS_DIRTY_CONSTANTS_TES
;
1860 if (dirty
& IRIS_DIRTY_UNCOMPILED_VS
)
1861 iris_update_compiled_vs(ice
);
1862 if (dirty
& IRIS_DIRTY_UNCOMPILED_GS
)
1863 iris_update_compiled_gs(ice
);
1865 if (dirty
& (IRIS_DIRTY_UNCOMPILED_GS
| IRIS_DIRTY_UNCOMPILED_TES
)) {
1866 const struct iris_compiled_shader
*gs
=
1867 ice
->shaders
.prog
[MESA_SHADER_GEOMETRY
];
1868 const struct iris_compiled_shader
*tes
=
1869 ice
->shaders
.prog
[MESA_SHADER_TESS_EVAL
];
1871 bool points_or_lines
= false;
1874 const struct brw_gs_prog_data
*gs_prog_data
= (void *) gs
->prog_data
;
1876 gs_prog_data
->output_topology
== _3DPRIM_POINTLIST
||
1877 gs_prog_data
->output_topology
== _3DPRIM_LINESTRIP
;
1879 const struct brw_tes_prog_data
*tes_data
= (void *) tes
->prog_data
;
1881 tes_data
->output_topology
== BRW_TESS_OUTPUT_TOPOLOGY_LINE
||
1882 tes_data
->output_topology
== BRW_TESS_OUTPUT_TOPOLOGY_POINT
;
1885 if (ice
->shaders
.output_topology_is_points_or_lines
!= points_or_lines
) {
1886 /* Outbound to XY Clip enables */
1887 ice
->shaders
.output_topology_is_points_or_lines
= points_or_lines
;
1888 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
;
1892 gl_shader_stage last_stage
= last_vue_stage(ice
);
1893 struct iris_compiled_shader
*shader
= ice
->shaders
.prog
[last_stage
];
1894 struct iris_uncompiled_shader
*ish
= ice
->shaders
.uncompiled
[last_stage
];
1895 update_last_vue_map(ice
, shader
->prog_data
);
1896 if (ice
->state
.streamout
!= shader
->streamout
) {
1897 ice
->state
.streamout
= shader
->streamout
;
1898 ice
->state
.dirty
|= IRIS_DIRTY_SO_DECL_LIST
| IRIS_DIRTY_STREAMOUT
;
1901 if (ice
->state
.streamout_active
) {
1902 for (int i
= 0; i
< PIPE_MAX_SO_BUFFERS
; i
++) {
1903 struct iris_stream_output_target
*so
=
1904 (void *) ice
->state
.so_target
[i
];
1906 so
->stride
= ish
->stream_output
.stride
[i
] * sizeof(uint32_t);
1910 if (dirty
& IRIS_DIRTY_UNCOMPILED_FS
)
1911 iris_update_compiled_fs(ice
);
1913 /* Changing shader interfaces may require a URB configuration. */
1914 if (!(dirty
& IRIS_DIRTY_URB
)) {
1915 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_GEOMETRY
; i
++) {
1916 struct brw_vue_prog_data
*old
= old_prog_datas
[i
];
1917 struct brw_vue_prog_data
*new = get_vue_prog_data(ice
, i
);
1918 if (!!old
!= !!new ||
1919 (new && new->urb_entry_size
!= old
->urb_entry_size
)) {
1920 ice
->state
.dirty
|= IRIS_DIRTY_URB
;
1926 for (int i
= MESA_SHADER_VERTEX
; i
<= MESA_SHADER_FRAGMENT
; i
++) {
1927 if (ice
->state
.dirty
& (IRIS_DIRTY_CONSTANTS_VS
<< i
))
1928 iris_update_pull_constant_descriptors(ice
, i
);
1932 static struct iris_compiled_shader
*
1933 iris_compile_cs(struct iris_context
*ice
,
1934 struct iris_uncompiled_shader
*ish
,
1935 const struct iris_cs_prog_key
*key
)
1937 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1938 const struct brw_compiler
*compiler
= screen
->compiler
;
1939 void *mem_ctx
= ralloc_context(NULL
);
1940 struct brw_cs_prog_data
*cs_prog_data
=
1941 rzalloc(mem_ctx
, struct brw_cs_prog_data
);
1942 struct brw_stage_prog_data
*prog_data
= &cs_prog_data
->base
;
1943 enum brw_param_builtin
*system_values
;
1944 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
1945 unsigned num_system_values
;
1948 nir_shader
*nir
= nir_shader_clone(mem_ctx
, ish
->nir
);
1950 iris_setup_uniforms(compiler
, mem_ctx
, nir
, prog_data
, &system_values
,
1951 &num_system_values
, &num_cbufs
);
1953 struct iris_binding_table bt
;
1954 iris_setup_binding_table(devinfo
, nir
, &bt
, /* num_render_targets */ 0,
1955 num_system_values
, num_cbufs
);
1957 struct brw_cs_prog_key brw_key
= iris_to_brw_cs_key(devinfo
, key
);
1959 char *error_str
= NULL
;
1960 const unsigned *program
=
1961 brw_compile_cs(compiler
, &ice
->dbg
, mem_ctx
, &brw_key
, cs_prog_data
,
1962 nir
, -1, NULL
, &error_str
);
1963 if (program
== NULL
) {
1964 dbg_printf("Failed to compile compute shader: %s\n", error_str
);
1965 ralloc_free(mem_ctx
);
1969 if (ish
->compiled_once
) {
1970 iris_debug_recompile(ice
, &nir
->info
, &brw_key
.base
);
1972 ish
->compiled_once
= true;
1975 struct iris_compiled_shader
*shader
=
1976 iris_upload_shader(ice
, IRIS_CACHE_CS
, sizeof(*key
), key
, program
,
1977 prog_data
, NULL
, system_values
, num_system_values
,
1980 iris_disk_cache_store(screen
->disk_cache
, ish
, shader
, key
, sizeof(*key
));
1982 ralloc_free(mem_ctx
);
1987 iris_update_compiled_cs(struct iris_context
*ice
)
1989 struct iris_shader_state
*shs
= &ice
->state
.shaders
[MESA_SHADER_COMPUTE
];
1990 struct iris_uncompiled_shader
*ish
=
1991 ice
->shaders
.uncompiled
[MESA_SHADER_COMPUTE
];
1993 struct iris_cs_prog_key key
= { KEY_ID(base
) };
1994 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
1995 screen
->vtbl
.populate_cs_key(ice
, &key
);
1997 struct iris_compiled_shader
*old
= ice
->shaders
.prog
[IRIS_CACHE_CS
];
1998 struct iris_compiled_shader
*shader
=
1999 iris_find_cached_shader(ice
, IRIS_CACHE_CS
, sizeof(key
), &key
);
2002 shader
= iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
));
2005 shader
= iris_compile_cs(ice
, ish
, &key
);
2007 if (old
!= shader
) {
2008 ice
->shaders
.prog
[IRIS_CACHE_CS
] = shader
;
2009 ice
->state
.dirty
|= IRIS_DIRTY_CS
|
2010 IRIS_DIRTY_BINDINGS_CS
|
2011 IRIS_DIRTY_CONSTANTS_CS
;
2012 shs
->sysvals_need_upload
= true;
2017 iris_update_compiled_compute_shader(struct iris_context
*ice
)
2019 if (ice
->state
.dirty
& IRIS_DIRTY_UNCOMPILED_CS
)
2020 iris_update_compiled_cs(ice
);
2022 if (ice
->state
.dirty
& IRIS_DIRTY_CONSTANTS_CS
)
2023 iris_update_pull_constant_descriptors(ice
, MESA_SHADER_COMPUTE
);
2027 iris_fill_cs_push_const_buffer(struct brw_cs_prog_data
*cs_prog_data
,
2031 assert(brw_cs_push_const_total_size(cs_prog_data
, threads
) > 0);
2032 assert(cs_prog_data
->push
.cross_thread
.size
== 0);
2033 assert(cs_prog_data
->push
.per_thread
.dwords
== 1);
2034 assert(cs_prog_data
->base
.param
[0] == BRW_PARAM_BUILTIN_SUBGROUP_ID
);
2035 for (unsigned t
= 0; t
< threads
; t
++)
2040 * Allocate scratch BOs as needed for the given per-thread size and stage.
2043 iris_get_scratch_space(struct iris_context
*ice
,
2044 unsigned per_thread_scratch
,
2045 gl_shader_stage stage
)
2047 struct iris_screen
*screen
= (struct iris_screen
*)ice
->ctx
.screen
;
2048 struct iris_bufmgr
*bufmgr
= screen
->bufmgr
;
2049 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
2051 unsigned encoded_size
= ffs(per_thread_scratch
) - 11;
2052 assert(encoded_size
< (1 << 16));
2054 struct iris_bo
**bop
= &ice
->shaders
.scratch_bos
[encoded_size
][stage
];
2056 /* The documentation for 3DSTATE_PS "Scratch Space Base Pointer" says:
2058 * "Scratch Space per slice is computed based on 4 sub-slices. SW
2059 * must allocate scratch space enough so that each slice has 4
2062 * According to the other driver team, this applies to compute shaders
2063 * as well. This is not currently documented at all.
2065 * This hack is no longer necessary on Gen11+.
2067 * For, Gen11+, scratch space allocation is based on the number of threads
2068 * in the base configuration.
2070 unsigned subslice_total
= screen
->subslice_total
;
2071 if (devinfo
->gen
>= 12)
2072 subslice_total
= devinfo
->num_subslices
[0];
2073 else if (devinfo
->gen
== 11)
2075 else if (devinfo
->gen
< 11)
2076 subslice_total
= 4 * devinfo
->num_slices
;
2077 assert(subslice_total
>= screen
->subslice_total
);
2080 unsigned scratch_ids_per_subslice
= devinfo
->max_cs_threads
;
2082 if (devinfo
->gen
>= 12) {
2083 /* Same as ICL below, but with 16 EUs. */
2084 scratch_ids_per_subslice
= 16 * 8;
2085 } else if (devinfo
->gen
== 11) {
2086 /* The MEDIA_VFE_STATE docs say:
2088 * "Starting with this configuration, the Maximum Number of
2089 * Threads must be set to (#EU * 8) for GPGPU dispatches.
2091 * Although there are only 7 threads per EU in the configuration,
2092 * the FFTID is calculated as if there are 8 threads per EU,
2093 * which in turn requires a larger amount of Scratch Space to be
2094 * allocated by the driver."
2096 scratch_ids_per_subslice
= 8 * 8;
2099 uint32_t max_threads
[] = {
2100 [MESA_SHADER_VERTEX
] = devinfo
->max_vs_threads
,
2101 [MESA_SHADER_TESS_CTRL
] = devinfo
->max_tcs_threads
,
2102 [MESA_SHADER_TESS_EVAL
] = devinfo
->max_tes_threads
,
2103 [MESA_SHADER_GEOMETRY
] = devinfo
->max_gs_threads
,
2104 [MESA_SHADER_FRAGMENT
] = devinfo
->max_wm_threads
,
2105 [MESA_SHADER_COMPUTE
] = scratch_ids_per_subslice
* subslice_total
,
2108 uint32_t size
= per_thread_scratch
* max_threads
[stage
];
2110 *bop
= iris_bo_alloc(bufmgr
, "scratch", size
, IRIS_MEMZONE_SHADER
);
2116 /* ------------------------------------------------------------------- */
2119 * The pipe->create_[stage]_state() driver hooks.
2121 * Performs basic NIR preprocessing, records any state dependencies, and
2122 * returns an iris_uncompiled_shader as the Gallium CSO.
2124 * Actual shader compilation to assembly happens later, at first use.
2127 iris_create_uncompiled_shader(struct pipe_context
*ctx
,
2129 const struct pipe_stream_output_info
*so_info
)
2131 struct iris_context
*ice
= (void *)ctx
;
2132 struct iris_screen
*screen
= (struct iris_screen
*)ctx
->screen
;
2133 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
2135 struct iris_uncompiled_shader
*ish
=
2136 calloc(1, sizeof(struct iris_uncompiled_shader
));
2140 NIR_PASS(ish
->needs_edge_flag
, nir
, iris_fix_edge_flags
);
2142 brw_preprocess_nir(screen
->compiler
, nir
, NULL
);
2144 NIR_PASS_V(nir
, brw_nir_lower_image_load_store
, devinfo
,
2145 &ish
->uses_atomic_load_store
);
2146 NIR_PASS_V(nir
, iris_lower_storage_image_derefs
);
2150 if (nir
->constant_data_size
> 0) {
2151 unsigned data_offset
;
2152 u_upload_data(ice
->shaders
.uploader
, 0, nir
->constant_data_size
,
2153 32, nir
->constant_data
, &data_offset
, &ish
->const_data
);
2155 struct pipe_shader_buffer psb
= {
2156 .buffer
= ish
->const_data
,
2157 .buffer_offset
= data_offset
,
2158 .buffer_size
= nir
->constant_data_size
,
2160 iris_upload_ubo_ssbo_surf_state(ice
, &psb
, &ish
->const_data_state
, false);
2163 ish
->program_id
= get_new_program_id(screen
);
2166 memcpy(&ish
->stream_output
, so_info
, sizeof(*so_info
));
2167 update_so_info(&ish
->stream_output
, nir
->info
.outputs_written
);
2170 /* Save this now before potentially dropping nir->info.name */
2171 if (nir
->info
.name
&& strncmp(nir
->info
.name
, "ARB", 3) == 0)
2172 ish
->use_alt_mode
= true;
2174 if (screen
->disk_cache
) {
2175 /* Serialize the NIR to a binary blob that we can hash for the disk
2176 * cache. Drop unnecessary information (like variable names)
2177 * so the serialized NIR is smaller, and also to let us detect more
2178 * isomorphic shaders when hashing, increasing cache hits.
2182 nir_serialize(&blob
, nir
, true);
2183 _mesa_sha1_compute(blob
.data
, blob
.size
, ish
->nir_sha1
);
2190 static struct iris_uncompiled_shader
*
2191 iris_create_shader_state(struct pipe_context
*ctx
,
2192 const struct pipe_shader_state
*state
)
2194 struct nir_shader
*nir
;
2196 if (state
->type
== PIPE_SHADER_IR_TGSI
)
2197 nir
= tgsi_to_nir(state
->tokens
, ctx
->screen
);
2199 nir
= state
->ir
.nir
;
2201 return iris_create_uncompiled_shader(ctx
, nir
, &state
->stream_output
);
2205 iris_create_vs_state(struct pipe_context
*ctx
,
2206 const struct pipe_shader_state
*state
)
2208 struct iris_context
*ice
= (void *) ctx
;
2209 struct iris_screen
*screen
= (void *) ctx
->screen
;
2210 struct iris_uncompiled_shader
*ish
= iris_create_shader_state(ctx
, state
);
2212 /* User clip planes */
2213 if (ish
->nir
->info
.clip_distance_array_size
== 0)
2214 ish
->nos
|= (1ull << IRIS_NOS_RASTERIZER
);
2216 if (screen
->precompile
) {
2217 struct iris_vs_prog_key key
= { KEY_ID(vue
.base
) };
2219 if (!iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
)))
2220 iris_compile_vs(ice
, ish
, &key
);
2227 iris_create_tcs_state(struct pipe_context
*ctx
,
2228 const struct pipe_shader_state
*state
)
2230 struct iris_context
*ice
= (void *) ctx
;
2231 struct iris_screen
*screen
= (void *) ctx
->screen
;
2232 const struct brw_compiler
*compiler
= screen
->compiler
;
2233 struct iris_uncompiled_shader
*ish
= iris_create_shader_state(ctx
, state
);
2234 struct shader_info
*info
= &ish
->nir
->info
;
2236 if (screen
->precompile
) {
2237 const unsigned _GL_TRIANGLES
= 0x0004;
2238 struct iris_tcs_prog_key key
= {
2240 // XXX: make sure the linker fills this out from the TES...
2241 .tes_primitive_mode
=
2242 info
->tess
.primitive_mode
? info
->tess
.primitive_mode
2244 .outputs_written
= info
->outputs_written
,
2245 .patch_outputs_written
= info
->patch_outputs_written
,
2248 /* 8_PATCH mode needs the key to contain the input patch dimensionality.
2249 * We don't have that information, so we randomly guess that the input
2250 * and output patches are the same size. This is a bad guess, but we
2251 * can't do much better.
2253 if (compiler
->use_tcs_8_patch
)
2254 key
.input_vertices
= info
->tess
.tcs_vertices_out
;
2256 if (!iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
)))
2257 iris_compile_tcs(ice
, ish
, &key
);
2264 iris_create_tes_state(struct pipe_context
*ctx
,
2265 const struct pipe_shader_state
*state
)
2267 struct iris_context
*ice
= (void *) ctx
;
2268 struct iris_screen
*screen
= (void *) ctx
->screen
;
2269 struct iris_uncompiled_shader
*ish
= iris_create_shader_state(ctx
, state
);
2270 struct shader_info
*info
= &ish
->nir
->info
;
2272 /* User clip planes */
2273 if (ish
->nir
->info
.clip_distance_array_size
== 0)
2274 ish
->nos
|= (1ull << IRIS_NOS_RASTERIZER
);
2276 if (screen
->precompile
) {
2277 struct iris_tes_prog_key key
= {
2279 // XXX: not ideal, need TCS output/TES input unification
2280 .inputs_read
= info
->inputs_read
,
2281 .patch_inputs_read
= info
->patch_inputs_read
,
2284 if (!iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
)))
2285 iris_compile_tes(ice
, ish
, &key
);
2292 iris_create_gs_state(struct pipe_context
*ctx
,
2293 const struct pipe_shader_state
*state
)
2295 struct iris_context
*ice
= (void *) ctx
;
2296 struct iris_screen
*screen
= (void *) ctx
->screen
;
2297 struct iris_uncompiled_shader
*ish
= iris_create_shader_state(ctx
, state
);
2299 /* User clip planes */
2300 if (ish
->nir
->info
.clip_distance_array_size
== 0)
2301 ish
->nos
|= (1ull << IRIS_NOS_RASTERIZER
);
2303 if (screen
->precompile
) {
2304 struct iris_gs_prog_key key
= { KEY_ID(vue
.base
) };
2306 if (!iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
)))
2307 iris_compile_gs(ice
, ish
, &key
);
2314 iris_create_fs_state(struct pipe_context
*ctx
,
2315 const struct pipe_shader_state
*state
)
2317 struct iris_context
*ice
= (void *) ctx
;
2318 struct iris_screen
*screen
= (void *) ctx
->screen
;
2319 struct iris_uncompiled_shader
*ish
= iris_create_shader_state(ctx
, state
);
2320 struct shader_info
*info
= &ish
->nir
->info
;
2322 ish
->nos
|= (1ull << IRIS_NOS_FRAMEBUFFER
) |
2323 (1ull << IRIS_NOS_DEPTH_STENCIL_ALPHA
) |
2324 (1ull << IRIS_NOS_RASTERIZER
) |
2325 (1ull << IRIS_NOS_BLEND
);
2327 /* The program key needs the VUE map if there are > 16 inputs */
2328 if (util_bitcount64(ish
->nir
->info
.inputs_read
&
2329 BRW_FS_VARYING_INPUT_MASK
) > 16) {
2330 ish
->nos
|= (1ull << IRIS_NOS_LAST_VUE_MAP
);
2333 if (screen
->precompile
) {
2334 const uint64_t color_outputs
= info
->outputs_written
&
2335 ~(BITFIELD64_BIT(FRAG_RESULT_DEPTH
) |
2336 BITFIELD64_BIT(FRAG_RESULT_STENCIL
) |
2337 BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK
));
2339 bool can_rearrange_varyings
=
2340 util_bitcount64(info
->inputs_read
& BRW_FS_VARYING_INPUT_MASK
) <= 16;
2342 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
2343 struct iris_fs_prog_key key
= {
2345 .nr_color_regions
= util_bitcount(color_outputs
),
2346 .coherent_fb_fetch
= devinfo
->gen
>= 9,
2347 .input_slots_valid
=
2348 can_rearrange_varyings
? 0 : info
->inputs_read
| VARYING_BIT_POS
,
2351 if (!iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
)))
2352 iris_compile_fs(ice
, ish
, &key
, NULL
);
2359 iris_create_compute_state(struct pipe_context
*ctx
,
2360 const struct pipe_compute_state
*state
)
2362 assert(state
->ir_type
== PIPE_SHADER_IR_NIR
);
2364 struct iris_context
*ice
= (void *) ctx
;
2365 struct iris_screen
*screen
= (void *) ctx
->screen
;
2366 struct iris_uncompiled_shader
*ish
=
2367 iris_create_uncompiled_shader(ctx
, (void *) state
->prog
, NULL
);
2369 // XXX: disallow more than 64KB of shared variables
2371 if (screen
->precompile
) {
2372 struct iris_cs_prog_key key
= { KEY_ID(base
) };
2374 if (!iris_disk_cache_retrieve(ice
, ish
, &key
, sizeof(key
)))
2375 iris_compile_cs(ice
, ish
, &key
);
2382 * The pipe->delete_[stage]_state() driver hooks.
2384 * Frees the iris_uncompiled_shader.
2387 iris_delete_shader_state(struct pipe_context
*ctx
, void *state
, gl_shader_stage stage
)
2389 struct iris_uncompiled_shader
*ish
= state
;
2390 struct iris_context
*ice
= (void *) ctx
;
2392 if (ice
->shaders
.uncompiled
[stage
] == ish
) {
2393 ice
->shaders
.uncompiled
[stage
] = NULL
;
2394 ice
->state
.dirty
|= IRIS_DIRTY_UNCOMPILED_VS
<< stage
;
2397 if (ish
->const_data
) {
2398 pipe_resource_reference(&ish
->const_data
, NULL
);
2399 pipe_resource_reference(&ish
->const_data_state
.res
, NULL
);
2402 ralloc_free(ish
->nir
);
2407 iris_delete_vs_state(struct pipe_context
*ctx
, void *state
)
2409 iris_delete_shader_state(ctx
, state
, MESA_SHADER_VERTEX
);
2413 iris_delete_tcs_state(struct pipe_context
*ctx
, void *state
)
2415 iris_delete_shader_state(ctx
, state
, MESA_SHADER_TESS_CTRL
);
2419 iris_delete_tes_state(struct pipe_context
*ctx
, void *state
)
2421 iris_delete_shader_state(ctx
, state
, MESA_SHADER_TESS_EVAL
);
2425 iris_delete_gs_state(struct pipe_context
*ctx
, void *state
)
2427 iris_delete_shader_state(ctx
, state
, MESA_SHADER_GEOMETRY
);
2431 iris_delete_fs_state(struct pipe_context
*ctx
, void *state
)
2433 iris_delete_shader_state(ctx
, state
, MESA_SHADER_FRAGMENT
);
2437 iris_delete_cs_state(struct pipe_context
*ctx
, void *state
)
2439 iris_delete_shader_state(ctx
, state
, MESA_SHADER_COMPUTE
);
2443 * The pipe->bind_[stage]_state() driver hook.
2445 * Binds an uncompiled shader as the current one for a particular stage.
2446 * Updates dirty tracking to account for the shader's NOS.
2449 bind_shader_state(struct iris_context
*ice
,
2450 struct iris_uncompiled_shader
*ish
,
2451 gl_shader_stage stage
)
2453 uint64_t dirty_bit
= IRIS_DIRTY_UNCOMPILED_VS
<< stage
;
2454 const uint64_t nos
= ish
? ish
->nos
: 0;
2456 const struct shader_info
*old_info
= iris_get_shader_info(ice
, stage
);
2457 const struct shader_info
*new_info
= ish
? &ish
->nir
->info
: NULL
;
2459 if ((old_info
? util_last_bit(old_info
->textures_used
) : 0) !=
2460 (new_info
? util_last_bit(new_info
->textures_used
) : 0)) {
2461 ice
->state
.dirty
|= IRIS_DIRTY_SAMPLER_STATES_VS
<< stage
;
2464 ice
->shaders
.uncompiled
[stage
] = ish
;
2465 ice
->state
.dirty
|= dirty_bit
;
2467 /* Record that CSOs need to mark IRIS_DIRTY_UNCOMPILED_XS when they change
2468 * (or that they no longer need to do so).
2470 for (int i
= 0; i
< IRIS_NOS_COUNT
; i
++) {
2472 ice
->state
.dirty_for_nos
[i
] |= dirty_bit
;
2474 ice
->state
.dirty_for_nos
[i
] &= ~dirty_bit
;
2479 iris_bind_vs_state(struct pipe_context
*ctx
, void *state
)
2481 struct iris_context
*ice
= (struct iris_context
*)ctx
;
2482 struct iris_uncompiled_shader
*new_ish
= state
;
2485 ice
->state
.window_space_position
!=
2486 new_ish
->nir
->info
.vs
.window_space_position
) {
2487 ice
->state
.window_space_position
=
2488 new_ish
->nir
->info
.vs
.window_space_position
;
2490 ice
->state
.dirty
|= IRIS_DIRTY_CLIP
|
2492 IRIS_DIRTY_CC_VIEWPORT
;
2495 bind_shader_state((void *) ctx
, state
, MESA_SHADER_VERTEX
);
2499 iris_bind_tcs_state(struct pipe_context
*ctx
, void *state
)
2501 bind_shader_state((void *) ctx
, state
, MESA_SHADER_TESS_CTRL
);
2505 iris_bind_tes_state(struct pipe_context
*ctx
, void *state
)
2507 struct iris_context
*ice
= (struct iris_context
*)ctx
;
2509 /* Enabling/disabling optional stages requires a URB reconfiguration. */
2510 if (!!state
!= !!ice
->shaders
.uncompiled
[MESA_SHADER_TESS_EVAL
])
2511 ice
->state
.dirty
|= IRIS_DIRTY_URB
;
2513 bind_shader_state((void *) ctx
, state
, MESA_SHADER_TESS_EVAL
);
2517 iris_bind_gs_state(struct pipe_context
*ctx
, void *state
)
2519 struct iris_context
*ice
= (struct iris_context
*)ctx
;
2521 /* Enabling/disabling optional stages requires a URB reconfiguration. */
2522 if (!!state
!= !!ice
->shaders
.uncompiled
[MESA_SHADER_GEOMETRY
])
2523 ice
->state
.dirty
|= IRIS_DIRTY_URB
;
2525 bind_shader_state((void *) ctx
, state
, MESA_SHADER_GEOMETRY
);
2529 iris_bind_fs_state(struct pipe_context
*ctx
, void *state
)
2531 struct iris_context
*ice
= (struct iris_context
*) ctx
;
2532 struct iris_screen
*screen
= (struct iris_screen
*) ctx
->screen
;
2533 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
2534 struct iris_uncompiled_shader
*old_ish
=
2535 ice
->shaders
.uncompiled
[MESA_SHADER_FRAGMENT
];
2536 struct iris_uncompiled_shader
*new_ish
= state
;
2538 const unsigned color_bits
=
2539 BITFIELD64_BIT(FRAG_RESULT_COLOR
) |
2540 BITFIELD64_RANGE(FRAG_RESULT_DATA0
, BRW_MAX_DRAW_BUFFERS
);
2542 /* Fragment shader outputs influence HasWriteableRT */
2543 if (!old_ish
|| !new_ish
||
2544 (old_ish
->nir
->info
.outputs_written
& color_bits
) !=
2545 (new_ish
->nir
->info
.outputs_written
& color_bits
))
2546 ice
->state
.dirty
|= IRIS_DIRTY_PS_BLEND
;
2548 if (devinfo
->gen
== 8)
2549 ice
->state
.dirty
|= IRIS_DIRTY_PMA_FIX
;
2551 bind_shader_state((void *) ctx
, state
, MESA_SHADER_FRAGMENT
);
2555 iris_bind_cs_state(struct pipe_context
*ctx
, void *state
)
2557 bind_shader_state((void *) ctx
, state
, MESA_SHADER_COMPUTE
);
2561 iris_init_program_functions(struct pipe_context
*ctx
)
2563 ctx
->create_vs_state
= iris_create_vs_state
;
2564 ctx
->create_tcs_state
= iris_create_tcs_state
;
2565 ctx
->create_tes_state
= iris_create_tes_state
;
2566 ctx
->create_gs_state
= iris_create_gs_state
;
2567 ctx
->create_fs_state
= iris_create_fs_state
;
2568 ctx
->create_compute_state
= iris_create_compute_state
;
2570 ctx
->delete_vs_state
= iris_delete_vs_state
;
2571 ctx
->delete_tcs_state
= iris_delete_tcs_state
;
2572 ctx
->delete_tes_state
= iris_delete_tes_state
;
2573 ctx
->delete_gs_state
= iris_delete_gs_state
;
2574 ctx
->delete_fs_state
= iris_delete_fs_state
;
2575 ctx
->delete_compute_state
= iris_delete_cs_state
;
2577 ctx
->bind_vs_state
= iris_bind_vs_state
;
2578 ctx
->bind_tcs_state
= iris_bind_tcs_state
;
2579 ctx
->bind_tes_state
= iris_bind_tes_state
;
2580 ctx
->bind_gs_state
= iris_bind_gs_state
;
2581 ctx
->bind_fs_state
= iris_bind_fs_state
;
2582 ctx
->bind_compute_state
= iris_bind_cs_state
;