2 * Copyright © 2010 - 2015 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
24 #ifndef BRW_COMPILER_H
25 #define BRW_COMPILER_H
28 #include "common/gen_device_info.h"
29 #include "main/macros.h"
30 #include "util/ralloc.h"
41 const struct gen_device_info
*devinfo
;
47 * Array of the ra classes for the unaligned contiguous register
53 * Mapping for register-allocated objects in *regs to the first
54 * GRF for that object.
56 uint8_t *ra_reg_to_grf
;
63 * Array of the ra classes for the unaligned contiguous register
64 * block sizes used, indexed by register size.
69 * Mapping from classes to ra_reg ranges. Each of the per-size
70 * classes corresponds to a range of ra_reg nodes. This array stores
71 * those ranges in the form of first ra_reg in each class and the
72 * total number of ra_reg elements in the last array element. This
73 * way the range of the i'th class is given by:
74 * [ class_to_ra_reg_range[i], class_to_ra_reg_range[i+1] )
76 int class_to_ra_reg_range
[17];
79 * Mapping for register-allocated objects in *regs to the first
80 * GRF for that object.
82 uint8_t *ra_reg_to_grf
;
85 * ra class for the aligned pairs we use for PLN, which doesn't
88 int aligned_pairs_class
;
91 void (*shader_debug_log
)(void *, const char *str
, ...) PRINTFLIKE(2, 3);
92 void (*shader_perf_log
)(void *, const char *str
, ...) PRINTFLIKE(2, 3);
94 bool scalar_stage
[MESA_SHADER_STAGES
];
95 struct gl_shader_compiler_options glsl_compiler_options
[MESA_SHADER_STAGES
];
98 * Apply workarounds for SIN and COS output range problems.
99 * This can negatively impact performance.
104 * Is 3DSTATE_CONSTANT_*'s Constant Buffer 0 relative to Dynamic State
105 * Base Address? (If not, it's a normal GPU address.)
107 bool constant_buffer_0_is_relative
;
110 * Whether or not the driver supports pull constants. If not, the compiler
111 * will attempt to push everything.
113 bool supports_pull_constants
;
118 * Program key structures.
120 * When drawing, we look for the currently bound shaders in the program
121 * cache. This is essentially a hash table lookup, and these are the keys.
123 * Sometimes OpenGL features specified as state need to be simulated via
124 * shader code, due to a mismatch between the API and the hardware. This
125 * is often referred to as "non-orthagonal state" or "NOS". We store NOS
126 * in the program key so it's considered when searching for a program. If
127 * we haven't seen a particular combination before, we have to recompile a
128 * new specialized version.
130 * Shader compilation should not look up state in gl_context directly, but
131 * instead use the copy in the program key. This guarantees recompiles will
137 enum PACKED gen6_gather_sampler_wa
{
138 WA_SIGN
= 1, /* whether we need to sign extend */
139 WA_8BIT
= 2, /* if we have an 8bit format needing wa */
140 WA_16BIT
= 4, /* if we have a 16bit format needing wa */
144 * Sampler information needed by VS, WM, and GS program cache keys.
146 struct brw_sampler_prog_key_data
{
148 * EXT_texture_swizzle and DEPTH_TEXTURE_MODE swizzles.
150 uint16_t swizzles
[MAX_SAMPLERS
];
152 uint32_t gl_clamp_mask
[3];
155 * For RG32F, gather4's channel select is broken.
157 uint32_t gather_channel_quirk_mask
;
160 * Whether this sampler uses the compressed multisample surface layout.
162 uint32_t compressed_multisample_layout_mask
;
165 * Whether this sampler is using 16x multisampling. If so fetching from
166 * this sampler will be handled with a different instruction, ld2dms_w
172 * For Sandybridge, which shader w/a we need for gather quirks.
174 enum gen6_gather_sampler_wa gen6_gather_wa
[MAX_SAMPLERS
];
177 * Texture units that have a YUV image bound.
179 uint32_t y_u_v_image_mask
;
180 uint32_t y_uv_image_mask
;
181 uint32_t yx_xuxv_image_mask
;
182 uint32_t xy_uxvx_image_mask
;
186 * The VF can't natively handle certain types of attributes, such as GL_FIXED
187 * or most 10_10_10_2 types. These flags enable various VS workarounds to
188 * "fix" attributes at the beginning of shaders.
190 #define BRW_ATTRIB_WA_COMPONENT_MASK 7 /* mask for GL_FIXED scale channel count */
191 #define BRW_ATTRIB_WA_NORMALIZE 8 /* normalize in shader */
192 #define BRW_ATTRIB_WA_BGRA 16 /* swap r/b channels in shader */
193 #define BRW_ATTRIB_WA_SIGN 32 /* interpret as signed in shader */
194 #define BRW_ATTRIB_WA_SCALE 64 /* interpret as scaled in shader */
197 * OpenGL attribute slots fall in [0, VERT_ATTRIB_MAX - 1] with the range
198 * [VERT_ATTRIB_GENERIC0, VERT_ATTRIB_MAX - 1] reserved for up to 16 user
199 * input vertex attributes. In Vulkan, we expose up to 28 user vertex input
200 * attributes that are mapped to slots also starting at VERT_ATTRIB_GENERIC0.
202 #define MAX_GL_VERT_ATTRIB VERT_ATTRIB_MAX
203 #define MAX_VK_VERT_ATTRIB (VERT_ATTRIB_GENERIC0 + 28)
205 /** The program key for Vertex Shaders. */
206 struct brw_vs_prog_key
{
207 unsigned program_string_id
;
210 * Per-attribute workaround flags
212 * For each attribute, a combination of BRW_ATTRIB_WA_*.
214 * For OpenGL, where we expose a maximum of 16 user input atttributes
215 * we only need up to VERT_ATTRIB_MAX slots, however, in Vulkan
216 * slots preceding VERT_ATTRIB_GENERIC0 are unused and we can
217 * expose up to 28 user input vertex attributes that are mapped to slots
218 * starting at VERT_ATTRIB_GENERIC0, so this array needs to be large
219 * enough to hold this many slots.
221 uint8_t gl_attrib_wa_flags
[MAX2(MAX_GL_VERT_ATTRIB
, MAX_VK_VERT_ATTRIB
)];
223 bool copy_edgeflag
:1;
225 bool clamp_vertex_color
:1;
228 * How many user clipping planes are being uploaded to the vertex shader as
231 * These are used for lowering legacy gl_ClipVertex/gl_Position clipping to
234 unsigned nr_userclip_plane_consts
:4;
237 * For pre-Gen6 hardware, a bitfield indicating which texture coordinates
238 * are going to be replaced with point coordinates (as a consequence of a
239 * call to glTexEnvi(GL_POINT_SPRITE, GL_COORD_REPLACE, GL_TRUE)). Because
240 * our SF thread requires exact matching between VS outputs and FS inputs,
241 * these texture coordinates will need to be unconditionally included in
242 * the VUE, even if they aren't written by the vertex shader.
244 uint8_t point_coord_replace
;
246 struct brw_sampler_prog_key_data tex
;
249 /** The program key for Tessellation Control Shaders. */
250 struct brw_tcs_prog_key
252 unsigned program_string_id
;
254 GLenum tes_primitive_mode
;
256 unsigned input_vertices
;
258 /** A bitfield of per-patch outputs written. */
259 uint32_t patch_outputs_written
;
261 /** A bitfield of per-vertex outputs written. */
262 uint64_t outputs_written
;
264 bool quads_workaround
;
266 struct brw_sampler_prog_key_data tex
;
269 /** The program key for Tessellation Evaluation Shaders. */
270 struct brw_tes_prog_key
272 unsigned program_string_id
;
274 /** A bitfield of per-patch inputs read. */
275 uint32_t patch_inputs_read
;
277 /** A bitfield of per-vertex inputs read. */
278 uint64_t inputs_read
;
280 struct brw_sampler_prog_key_data tex
;
283 /** The program key for Geometry Shaders. */
284 struct brw_gs_prog_key
286 unsigned program_string_id
;
288 struct brw_sampler_prog_key_data tex
;
291 enum brw_sf_primitive
{
292 BRW_SF_PRIM_POINTS
= 0,
293 BRW_SF_PRIM_LINES
= 1,
294 BRW_SF_PRIM_TRIANGLES
= 2,
295 BRW_SF_PRIM_UNFILLED_TRIS
= 3,
298 struct brw_sf_prog_key
{
300 bool contains_flat_varying
;
301 unsigned char interp_mode
[65]; /* BRW_VARYING_SLOT_COUNT */
302 uint8_t point_sprite_coord_replace
;
303 enum brw_sf_primitive primitive
:2;
304 bool do_twoside_color
:1;
305 bool frontface_ccw
:1;
306 bool do_point_sprite
:1;
307 bool do_point_coord
:1;
308 bool sprite_origin_lower_left
:1;
309 bool userclip_active
:1;
313 BRW_CLIP_MODE_NORMAL
= 0,
314 BRW_CLIP_MODE_CLIP_ALL
= 1,
315 BRW_CLIP_MODE_CLIP_NON_REJECTED
= 2,
316 BRW_CLIP_MODE_REJECT_ALL
= 3,
317 BRW_CLIP_MODE_ACCEPT_ALL
= 4,
318 BRW_CLIP_MODE_KERNEL_CLIP
= 5,
321 enum brw_clip_fill_mode
{
322 BRW_CLIP_FILL_MODE_LINE
= 0,
323 BRW_CLIP_FILL_MODE_POINT
= 1,
324 BRW_CLIP_FILL_MODE_FILL
= 2,
325 BRW_CLIP_FILL_MODE_CULL
= 3,
328 /* Note that if unfilled primitives are being emitted, we have to fix
329 * up polygon offset and flatshading at this point:
331 struct brw_clip_prog_key
{
333 bool contains_flat_varying
;
334 bool contains_noperspective_varying
;
335 unsigned char interp_mode
[65]; /* BRW_VARYING_SLOT_COUNT */
336 unsigned primitive
:4;
337 unsigned nr_userclip
:4;
340 enum brw_clip_fill_mode fill_cw
:2; /* includes cull information */
341 enum brw_clip_fill_mode fill_ccw
:2; /* includes cull information */
346 enum brw_clip_mode clip_mode
:3;
353 /* A big lookup table is used to figure out which and how many
354 * additional regs will inserted before the main payload in the WM
355 * program execution. These mainly relate to depth and stencil
356 * processing and the early-depth-test optimization.
358 enum brw_wm_iz_bits
{
359 BRW_WM_IZ_PS_KILL_ALPHATEST_BIT
= 0x1,
360 BRW_WM_IZ_PS_COMPUTES_DEPTH_BIT
= 0x2,
361 BRW_WM_IZ_DEPTH_WRITE_ENABLE_BIT
= 0x4,
362 BRW_WM_IZ_DEPTH_TEST_ENABLE_BIT
= 0x8,
363 BRW_WM_IZ_STENCIL_WRITE_ENABLE_BIT
= 0x10,
364 BRW_WM_IZ_STENCIL_TEST_ENABLE_BIT
= 0x20,
365 BRW_WM_IZ_BIT_MAX
= 0x40
368 enum brw_wm_aa_enable
{
374 /** The program key for Fragment/Pixel Shaders. */
375 struct brw_wm_prog_key
{
376 /* Some collection of BRW_WM_IZ_* */
380 unsigned nr_color_regions
:5;
381 bool replicate_alpha
:1;
382 bool clamp_fragment_color
:1;
383 bool persample_interp
:1;
384 bool multisample_fbo
:1;
385 bool frag_coord_adds_sample_pos
:1;
386 enum brw_wm_aa_enable line_aa
:2;
387 bool high_quality_derivatives
:1;
388 bool force_dual_color_blend
:1;
389 bool coherent_fb_fetch
:1;
391 uint64_t input_slots_valid
;
392 unsigned program_string_id
;
393 GLenum alpha_test_func
; /* < For Gen4/5 MRT alpha test */
394 float alpha_test_ref
;
396 struct brw_sampler_prog_key_data tex
;
399 struct brw_cs_prog_key
{
400 uint32_t program_string_id
;
401 struct brw_sampler_prog_key_data tex
;
404 /* brw_any_prog_key is any of the keys that map to an API stage */
405 union brw_any_prog_key
{
406 struct brw_vs_prog_key vs
;
407 struct brw_tcs_prog_key tcs
;
408 struct brw_tes_prog_key tes
;
409 struct brw_gs_prog_key gs
;
410 struct brw_wm_prog_key wm
;
411 struct brw_cs_prog_key cs
;
415 * Image metadata structure as laid out in the shader parameter
416 * buffer. Entries have to be 16B-aligned for the vec4 back-end to be
417 * able to use them. That's okay because the padding and any unused
418 * entries [most of them except when we're doing untyped surface
419 * access] will be removed by the uniform packing pass.
421 #define BRW_IMAGE_PARAM_SURFACE_IDX_OFFSET 0
422 #define BRW_IMAGE_PARAM_OFFSET_OFFSET 4
423 #define BRW_IMAGE_PARAM_SIZE_OFFSET 8
424 #define BRW_IMAGE_PARAM_STRIDE_OFFSET 12
425 #define BRW_IMAGE_PARAM_TILING_OFFSET 16
426 #define BRW_IMAGE_PARAM_SWIZZLING_OFFSET 20
427 #define BRW_IMAGE_PARAM_SIZE 24
429 struct brw_image_param
{
430 /** Surface binding table index. */
431 uint32_t surface_idx
;
433 /** Offset applied to the X and Y surface coordinates. */
436 /** Surface X, Y and Z dimensions. */
439 /** X-stride in bytes, Y-stride in pixels, horizontal slice stride in
440 * pixels, vertical slice stride in pixels.
444 /** Log2 of the tiling modulus in the X, Y and Z dimension. */
448 * Right shift to apply for bit 6 address swizzling. Two different
449 * swizzles can be specified and will be applied one after the other. The
450 * resulting address will be:
452 * addr' = addr ^ ((1 << 6) & ((addr >> swizzling[0]) ^
453 * (addr >> swizzling[1])))
455 * Use \c 0xff if any of the swizzles is not required.
457 uint32_t swizzling
[2];
460 /** Max number of render targets in a shader */
461 #define BRW_MAX_DRAW_BUFFERS 8
464 * Max number of binding table entries used for stream output.
466 * From the OpenGL 3.0 spec, table 6.44 (Transform Feedback State), the
467 * minimum value of MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS is 64.
469 * On Gen6, the size of transform feedback data is limited not by the number
470 * of components but by the number of binding table entries we set aside. We
471 * use one binding table entry for a float, one entry for a vector, and one
472 * entry per matrix column. Since the only way we can communicate our
473 * transform feedback capabilities to the client is via
474 * MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS, we need to plan for the
475 * worst case, in which all the varyings are floats, so we use up one binding
476 * table entry per component. Therefore we need to set aside at least 64
477 * binding table entries for use by transform feedback.
479 * Note: since we don't currently pack varyings, it is currently impossible
480 * for the client to actually use up all of these binding table entries--if
481 * all of their varyings were floats, they would run out of varying slots and
482 * fail to link. But that's a bug, so it seems prudent to go ahead and
483 * allocate the number of binding table entries we will need once the bug is
486 #define BRW_MAX_SOL_BINDINGS 64
489 * Binding table index for the first gen6 SOL binding.
491 #define BRW_GEN6_SOL_BINDING_START 0
494 * Stride in bytes between shader_time entries.
496 * We separate entries by a cacheline to reduce traffic between EUs writing to
499 #define BRW_SHADER_TIME_STRIDE 64
508 /* We reserve the first 2^16 values for builtins */
509 #define BRW_PARAM_IS_BUILTIN(param) (((param) & 0xffff0000) == 0)
511 enum brw_param_builtin
{
512 BRW_PARAM_BUILTIN_ZERO
,
514 BRW_PARAM_BUILTIN_CLIP_PLANE_0_X
,
515 BRW_PARAM_BUILTIN_CLIP_PLANE_0_Y
,
516 BRW_PARAM_BUILTIN_CLIP_PLANE_0_Z
,
517 BRW_PARAM_BUILTIN_CLIP_PLANE_0_W
,
518 BRW_PARAM_BUILTIN_CLIP_PLANE_1_X
,
519 BRW_PARAM_BUILTIN_CLIP_PLANE_1_Y
,
520 BRW_PARAM_BUILTIN_CLIP_PLANE_1_Z
,
521 BRW_PARAM_BUILTIN_CLIP_PLANE_1_W
,
522 BRW_PARAM_BUILTIN_CLIP_PLANE_2_X
,
523 BRW_PARAM_BUILTIN_CLIP_PLANE_2_Y
,
524 BRW_PARAM_BUILTIN_CLIP_PLANE_2_Z
,
525 BRW_PARAM_BUILTIN_CLIP_PLANE_2_W
,
526 BRW_PARAM_BUILTIN_CLIP_PLANE_3_X
,
527 BRW_PARAM_BUILTIN_CLIP_PLANE_3_Y
,
528 BRW_PARAM_BUILTIN_CLIP_PLANE_3_Z
,
529 BRW_PARAM_BUILTIN_CLIP_PLANE_3_W
,
530 BRW_PARAM_BUILTIN_CLIP_PLANE_4_X
,
531 BRW_PARAM_BUILTIN_CLIP_PLANE_4_Y
,
532 BRW_PARAM_BUILTIN_CLIP_PLANE_4_Z
,
533 BRW_PARAM_BUILTIN_CLIP_PLANE_4_W
,
534 BRW_PARAM_BUILTIN_CLIP_PLANE_5_X
,
535 BRW_PARAM_BUILTIN_CLIP_PLANE_5_Y
,
536 BRW_PARAM_BUILTIN_CLIP_PLANE_5_Z
,
537 BRW_PARAM_BUILTIN_CLIP_PLANE_5_W
,
538 BRW_PARAM_BUILTIN_CLIP_PLANE_6_X
,
539 BRW_PARAM_BUILTIN_CLIP_PLANE_6_Y
,
540 BRW_PARAM_BUILTIN_CLIP_PLANE_6_Z
,
541 BRW_PARAM_BUILTIN_CLIP_PLANE_6_W
,
542 BRW_PARAM_BUILTIN_CLIP_PLANE_7_X
,
543 BRW_PARAM_BUILTIN_CLIP_PLANE_7_Y
,
544 BRW_PARAM_BUILTIN_CLIP_PLANE_7_Z
,
545 BRW_PARAM_BUILTIN_CLIP_PLANE_7_W
,
547 BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_X
,
548 BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_Y
,
549 BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_Z
,
550 BRW_PARAM_BUILTIN_TESS_LEVEL_OUTER_W
,
551 BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_X
,
552 BRW_PARAM_BUILTIN_TESS_LEVEL_INNER_Y
,
554 BRW_PARAM_BUILTIN_SUBGROUP_ID
,
557 #define BRW_PARAM_BUILTIN_CLIP_PLANE(idx, comp) \
558 (BRW_PARAM_BUILTIN_CLIP_PLANE_0_X + ((idx) << 2) + (comp))
560 #define BRW_PARAM_BUILTIN_IS_CLIP_PLANE(param) \
561 ((param) >= BRW_PARAM_BUILTIN_CLIP_PLANE_0_X && \
562 (param) <= BRW_PARAM_BUILTIN_CLIP_PLANE_7_W)
564 #define BRW_PARAM_BUILTIN_CLIP_PLANE_IDX(param) \
565 (((param) - BRW_PARAM_BUILTIN_CLIP_PLANE_0_X) >> 2)
567 #define BRW_PARAM_BUILTIN_CLIP_PLANE_COMP(param) \
568 (((param) - BRW_PARAM_BUILTIN_CLIP_PLANE_0_X) & 0x3)
570 struct brw_stage_prog_data
{
572 /** size of our binding table. */
576 * surface indices for the various groups of surfaces
578 uint32_t pull_constants_start
;
579 uint32_t texture_start
;
580 uint32_t gather_texture_start
;
583 uint32_t image_start
;
584 uint32_t shader_time_start
;
585 uint32_t plane_start
[3];
589 struct brw_ubo_range ubo_ranges
[4];
591 GLuint nr_params
; /**< number of float params/constants */
592 GLuint nr_pull_params
;
594 unsigned curb_read_length
;
595 unsigned total_scratch
;
596 unsigned total_shared
;
598 unsigned program_size
;
601 * Register where the thread expects to find input data from the URB
602 * (typically uniforms, followed by vertex or fragment attributes).
604 unsigned dispatch_grf_start_reg
;
606 bool use_alt_mode
; /**< Use ALT floating point mode? Otherwise, IEEE. */
608 /* 32-bit identifiers for all push/pull parameters. These can be anything
609 * the driver wishes them to be; the core of the back-end compiler simply
610 * re-arranges them. The one restriction is that the bottom 2^16 values
611 * are reserved for builtins defined in the brw_param_builtin enum defined
615 uint32_t *pull_param
;
618 static inline uint32_t *
619 brw_stage_prog_data_add_params(struct brw_stage_prog_data
*prog_data
,
620 unsigned nr_new_params
)
622 unsigned old_nr_params
= prog_data
->nr_params
;
623 prog_data
->nr_params
+= nr_new_params
;
624 prog_data
->param
= reralloc(ralloc_parent(prog_data
->param
),
625 prog_data
->param
, uint32_t,
626 prog_data
->nr_params
);
627 return prog_data
->param
+ old_nr_params
;
631 brw_mark_surface_used(struct brw_stage_prog_data
*prog_data
,
634 /* A binding table index is 8 bits and the top 3 values are reserved for
635 * special things (stateless and SLM).
637 assert(surf_index
<= 252);
639 prog_data
->binding_table
.size_bytes
=
640 MAX2(prog_data
->binding_table
.size_bytes
, (surf_index
+ 1) * 4);
643 enum brw_barycentric_mode
{
644 BRW_BARYCENTRIC_PERSPECTIVE_PIXEL
= 0,
645 BRW_BARYCENTRIC_PERSPECTIVE_CENTROID
= 1,
646 BRW_BARYCENTRIC_PERSPECTIVE_SAMPLE
= 2,
647 BRW_BARYCENTRIC_NONPERSPECTIVE_PIXEL
= 3,
648 BRW_BARYCENTRIC_NONPERSPECTIVE_CENTROID
= 4,
649 BRW_BARYCENTRIC_NONPERSPECTIVE_SAMPLE
= 5,
650 BRW_BARYCENTRIC_MODE_COUNT
= 6
652 #define BRW_BARYCENTRIC_NONPERSPECTIVE_BITS \
653 ((1 << BRW_BARYCENTRIC_NONPERSPECTIVE_PIXEL) | \
654 (1 << BRW_BARYCENTRIC_NONPERSPECTIVE_CENTROID) | \
655 (1 << BRW_BARYCENTRIC_NONPERSPECTIVE_SAMPLE))
657 enum brw_pixel_shader_computed_depth_mode
{
658 BRW_PSCDEPTH_OFF
= 0, /* PS does not compute depth */
659 BRW_PSCDEPTH_ON
= 1, /* PS computes depth; no guarantee about value */
660 BRW_PSCDEPTH_ON_GE
= 2, /* PS guarantees output depth >= source depth */
661 BRW_PSCDEPTH_ON_LE
= 3, /* PS guarantees output depth <= source depth */
664 /* Data about a particular attempt to compile a program. Note that
665 * there can be many of these, each in a different GL state
666 * corresponding to a different brw_wm_prog_key struct, with different
669 struct brw_wm_prog_data
{
670 struct brw_stage_prog_data base
;
672 GLuint num_varying_inputs
;
674 uint8_t reg_blocks_0
;
675 uint8_t reg_blocks_2
;
677 uint8_t dispatch_grf_start_reg_2
;
678 uint32_t prog_offset_2
;
682 * surface indices the WM-specific surfaces
684 uint32_t render_target_read_start
;
688 uint8_t computed_depth_mode
;
689 bool computed_stencil
;
691 bool early_fragment_tests
;
692 bool post_depth_coverage
;
697 bool persample_dispatch
;
698 bool uses_pos_offset
;
703 bool uses_sample_mask
;
704 bool has_render_target_reads
;
705 bool has_side_effects
;
708 bool contains_flat_varying
;
709 bool contains_noperspective_varying
;
712 * Mask of which interpolation modes are required by the fragment shader.
713 * Used in hardware setup on gen6+.
715 uint32_t barycentric_interp_modes
;
718 * Mask of which FS inputs are marked flat by the shader source. This is
719 * needed for setting up 3DSTATE_SF/SBE.
721 uint32_t flat_inputs
;
723 /* Mapping of VUE slots to interpolation modes.
724 * Used by the Gen4-5 clip/sf/wm stages.
726 unsigned char interp_mode
[65]; /* BRW_VARYING_SLOT_COUNT */
729 * Map from gl_varying_slot to the position within the FS setup data
730 * payload where the varying's attribute vertex deltas should be delivered.
731 * For varying slots that are not used by the FS, the value is -1.
733 int urb_setup
[VARYING_SLOT_MAX
];
736 struct brw_push_const_block
{
737 unsigned dwords
; /* Dword count, not reg aligned */
739 unsigned size
; /* Bytes, register aligned */
742 struct brw_cs_prog_data
{
743 struct brw_stage_prog_data base
;
745 unsigned local_size
[3];
749 bool uses_num_work_groups
;
752 struct brw_push_const_block cross_thread
;
753 struct brw_push_const_block per_thread
;
754 struct brw_push_const_block total
;
759 * surface indices the CS-specific surfaces
761 uint32_t work_groups_start
;
767 * Enum representing the i965-specific vertex results that don't correspond
768 * exactly to any element of gl_varying_slot. The values of this enum are
769 * assigned such that they don't conflict with gl_varying_slot.
773 BRW_VARYING_SLOT_NDC
= VARYING_SLOT_MAX
,
774 BRW_VARYING_SLOT_PAD
,
776 * Technically this is not a varying but just a placeholder that
777 * compile_sf_prog() inserts into its VUE map to cause the gl_PointCoord
778 * builtin variable to be compiled correctly. see compile_sf_prog() for
781 BRW_VARYING_SLOT_PNTC
,
782 BRW_VARYING_SLOT_COUNT
786 * We always program SF to start reading at an offset of 1 (2 varying slots)
787 * from the start of the vertex URB entry. This causes it to skip:
788 * - VARYING_SLOT_PSIZ and BRW_VARYING_SLOT_NDC on gen4-5
789 * - VARYING_SLOT_PSIZ and VARYING_SLOT_POS on gen6+
791 #define BRW_SF_URB_ENTRY_READ_OFFSET 1
794 * Bitmask indicating which fragment shader inputs represent varyings (and
795 * hence have to be delivered to the fragment shader by the SF/SBE stage).
797 #define BRW_FS_VARYING_INPUT_MASK \
798 (BITFIELD64_RANGE(0, VARYING_SLOT_MAX) & \
799 ~VARYING_BIT_POS & ~VARYING_BIT_FACE)
802 * Data structure recording the relationship between the gl_varying_slot enum
803 * and "slots" within the vertex URB entry (VUE). A "slot" is defined as a
804 * single octaword within the VUE (128 bits).
806 * Note that each BRW register contains 256 bits (2 octawords), so when
807 * accessing the VUE in URB_NOSWIZZLE mode, each register corresponds to two
808 * consecutive VUE slots. When accessing the VUE in URB_INTERLEAVED mode (as
809 * in a vertex shader), each register corresponds to a single VUE slot, since
810 * it contains data for two separate vertices.
814 * Bitfield representing all varying slots that are (a) stored in this VUE
815 * map, and (b) actually written by the shader. Does not include any of
816 * the additional varying slots defined in brw_varying_slot.
818 uint64_t slots_valid
;
821 * Is this VUE map for a separate shader pipeline?
823 * Separable programs (GL_ARB_separate_shader_objects) can be mixed and matched
824 * without the linker having a chance to dead code eliminate unused varyings.
826 * This means that we have to use a fixed slot layout, based on the output's
827 * location field, rather than assigning slots in a compact contiguous block.
832 * Map from gl_varying_slot value to VUE slot. For gl_varying_slots that are
833 * not stored in a slot (because they are not written, or because
834 * additional processing is applied before storing them in the VUE), the
837 signed char varying_to_slot
[VARYING_SLOT_TESS_MAX
];
840 * Map from VUE slot to gl_varying_slot value. For slots that do not
841 * directly correspond to a gl_varying_slot, the value comes from
844 * For slots that are not in use, the value is BRW_VARYING_SLOT_PAD.
846 signed char slot_to_varying
[VARYING_SLOT_TESS_MAX
];
849 * Total number of VUE slots in use
854 * Number of per-patch VUE slots. Only valid for tessellation control
855 * shader outputs and tessellation evaluation shader inputs.
857 int num_per_patch_slots
;
860 * Number of per-vertex VUE slots. Only valid for tessellation control
861 * shader outputs and tessellation evaluation shader inputs.
863 int num_per_vertex_slots
;
866 void brw_print_vue_map(FILE *fp
, const struct brw_vue_map
*vue_map
);
869 * Convert a VUE slot number into a byte offset within the VUE.
871 static inline GLuint
brw_vue_slot_to_offset(GLuint slot
)
877 * Convert a vertex output (brw_varying_slot) into a byte offset within the
881 GLuint
brw_varying_to_offset(const struct brw_vue_map
*vue_map
, GLuint varying
)
883 return brw_vue_slot_to_offset(vue_map
->varying_to_slot
[varying
]);
886 void brw_compute_vue_map(const struct gen_device_info
*devinfo
,
887 struct brw_vue_map
*vue_map
,
888 uint64_t slots_valid
,
889 bool separate_shader
);
891 void brw_compute_tess_vue_map(struct brw_vue_map
*const vue_map
,
892 uint64_t slots_valid
,
895 /* brw_interpolation_map.c */
896 void brw_setup_vue_interpolation(struct brw_vue_map
*vue_map
,
897 struct nir_shader
*nir
,
898 struct brw_wm_prog_data
*prog_data
,
899 const struct gen_device_info
*devinfo
);
901 enum shader_dispatch_mode
{
902 DISPATCH_MODE_4X1_SINGLE
= 0,
903 DISPATCH_MODE_4X2_DUAL_INSTANCE
= 1,
904 DISPATCH_MODE_4X2_DUAL_OBJECT
= 2,
905 DISPATCH_MODE_SIMD8
= 3,
909 * @defgroup Tessellator parameter enumerations.
911 * These correspond to the hardware values in 3DSTATE_TE, and are provided
912 * as part of the tessellation evaluation shader.
916 enum brw_tess_partitioning
{
917 BRW_TESS_PARTITIONING_INTEGER
= 0,
918 BRW_TESS_PARTITIONING_ODD_FRACTIONAL
= 1,
919 BRW_TESS_PARTITIONING_EVEN_FRACTIONAL
= 2,
922 enum brw_tess_output_topology
{
923 BRW_TESS_OUTPUT_TOPOLOGY_POINT
= 0,
924 BRW_TESS_OUTPUT_TOPOLOGY_LINE
= 1,
925 BRW_TESS_OUTPUT_TOPOLOGY_TRI_CW
= 2,
926 BRW_TESS_OUTPUT_TOPOLOGY_TRI_CCW
= 3,
929 enum brw_tess_domain
{
930 BRW_TESS_DOMAIN_QUAD
= 0,
931 BRW_TESS_DOMAIN_TRI
= 1,
932 BRW_TESS_DOMAIN_ISOLINE
= 2,
936 struct brw_vue_prog_data
{
937 struct brw_stage_prog_data base
;
938 struct brw_vue_map vue_map
;
940 /** Should the hardware deliver input VUE handles for URB pull loads? */
941 bool include_vue_handles
;
943 GLuint urb_read_length
;
946 uint32_t clip_distance_mask
;
947 uint32_t cull_distance_mask
;
949 /* Used for calculating urb partitions. In the VS, this is the size of the
950 * URB entry used for both input and output to the thread. In the GS, this
951 * is the size of the URB entry used for output.
953 GLuint urb_entry_size
;
955 enum shader_dispatch_mode dispatch_mode
;
958 struct brw_vs_prog_data
{
959 struct brw_vue_prog_data base
;
961 GLbitfield64 inputs_read
;
962 GLbitfield64 double_inputs_read
;
964 unsigned nr_attribute_slots
;
967 bool uses_instanceid
;
968 bool uses_basevertex
;
969 bool uses_baseinstance
;
973 struct brw_tcs_prog_data
975 struct brw_vue_prog_data base
;
977 /** Number vertices in output patch */
982 struct brw_tes_prog_data
984 struct brw_vue_prog_data base
;
986 enum brw_tess_partitioning partitioning
;
987 enum brw_tess_output_topology output_topology
;
988 enum brw_tess_domain domain
;
991 struct brw_gs_prog_data
993 struct brw_vue_prog_data base
;
995 unsigned vertices_in
;
998 * Size of an output vertex, measured in HWORDS (32 bytes).
1000 unsigned output_vertex_size_hwords
;
1002 unsigned output_topology
;
1005 * Size of the control data (cut bits or StreamID bits), in hwords (32
1006 * bytes). 0 if there is no control data.
1008 unsigned control_data_header_size_hwords
;
1011 * Format of the control data (either GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
1012 * if the control data is StreamID bits, or
1013 * GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT if the control data is cut bits).
1014 * Ignored if control_data_header_size is 0.
1016 unsigned control_data_format
;
1018 bool include_primitive_id
;
1021 * The number of vertices emitted, if constant - otherwise -1.
1023 int static_vertex_count
;
1028 * Gen6: Provoking vertex convention for odd-numbered triangles
1034 * Gen6: Number of varyings that are output to transform feedback.
1036 GLuint num_transform_feedback_bindings
:7; /* 0-BRW_MAX_SOL_BINDINGS */
1039 * Gen6: Map from the index of a transform feedback binding table entry to the
1040 * gl_varying_slot that should be streamed out through that binding table
1043 unsigned char transform_feedback_bindings
[64 /* BRW_MAX_SOL_BINDINGS */];
1046 * Gen6: Map from the index of a transform feedback binding table entry to the
1047 * swizzles that should be used when streaming out data through that
1048 * binding table entry.
1050 unsigned char transform_feedback_swizzles
[64 /* BRW_MAX_SOL_BINDINGS */];
1053 struct brw_sf_prog_data
{
1054 uint32_t urb_read_length
;
1057 /* Each vertex may have upto 12 attributes, 4 components each,
1058 * except WPOS which requires only 2. (11*4 + 2) == 44 ==> 11
1061 * Actually we use 4 for each, so call it 12 rows.
1063 unsigned urb_entry_size
;
1066 struct brw_clip_prog_data
{
1067 uint32_t curb_read_length
; /* user planes? */
1069 uint32_t urb_read_length
;
1073 /* brw_any_prog_data is prog_data for any stage that maps to an API stage */
1074 union brw_any_prog_data
{
1075 struct brw_stage_prog_data base
;
1076 struct brw_vue_prog_data vue
;
1077 struct brw_vs_prog_data vs
;
1078 struct brw_tcs_prog_data tcs
;
1079 struct brw_tes_prog_data tes
;
1080 struct brw_gs_prog_data gs
;
1081 struct brw_wm_prog_data wm
;
1082 struct brw_cs_prog_data cs
;
1085 #define DEFINE_PROG_DATA_DOWNCAST(stage) \
1086 static inline struct brw_##stage##_prog_data * \
1087 brw_##stage##_prog_data(struct brw_stage_prog_data *prog_data) \
1089 return (struct brw_##stage##_prog_data *) prog_data; \
1091 DEFINE_PROG_DATA_DOWNCAST(vue
)
1092 DEFINE_PROG_DATA_DOWNCAST(vs
)
1093 DEFINE_PROG_DATA_DOWNCAST(tcs
)
1094 DEFINE_PROG_DATA_DOWNCAST(tes
)
1095 DEFINE_PROG_DATA_DOWNCAST(gs
)
1096 DEFINE_PROG_DATA_DOWNCAST(wm
)
1097 DEFINE_PROG_DATA_DOWNCAST(cs
)
1098 DEFINE_PROG_DATA_DOWNCAST(ff_gs
)
1099 DEFINE_PROG_DATA_DOWNCAST(clip
)
1100 DEFINE_PROG_DATA_DOWNCAST(sf
)
1101 #undef DEFINE_PROG_DATA_DOWNCAST
1105 struct brw_compiler
*
1106 brw_compiler_create(void *mem_ctx
, const struct gen_device_info
*devinfo
);
1109 brw_prog_data_size(gl_shader_stage stage
);
1112 brw_prog_key_size(gl_shader_stage stage
);
1115 * Compile a vertex shader.
1117 * Returns the final assembly and the program's size.
1120 brw_compile_vs(const struct brw_compiler
*compiler
, void *log_data
,
1122 const struct brw_vs_prog_key
*key
,
1123 struct brw_vs_prog_data
*prog_data
,
1124 const struct nir_shader
*shader
,
1125 int shader_time_index
,
1129 * Compile a tessellation control shader.
1131 * Returns the final assembly and the program's size.
1134 brw_compile_tcs(const struct brw_compiler
*compiler
,
1137 const struct brw_tcs_prog_key
*key
,
1138 struct brw_tcs_prog_data
*prog_data
,
1139 const struct nir_shader
*nir
,
1140 int shader_time_index
,
1144 * Compile a tessellation evaluation shader.
1146 * Returns the final assembly and the program's size.
1149 brw_compile_tes(const struct brw_compiler
*compiler
, void *log_data
,
1151 const struct brw_tes_prog_key
*key
,
1152 const struct brw_vue_map
*input_vue_map
,
1153 struct brw_tes_prog_data
*prog_data
,
1154 const struct nir_shader
*shader
,
1155 struct gl_program
*prog
,
1156 int shader_time_index
,
1160 * Compile a vertex shader.
1162 * Returns the final assembly and the program's size.
1165 brw_compile_gs(const struct brw_compiler
*compiler
, void *log_data
,
1167 const struct brw_gs_prog_key
*key
,
1168 struct brw_gs_prog_data
*prog_data
,
1169 const struct nir_shader
*shader
,
1170 struct gl_program
*prog
,
1171 int shader_time_index
,
1175 * Compile a strips and fans shader.
1177 * This is a fixed-function shader determined entirely by the shader key and
1180 * Returns the final assembly and the program's size.
1183 brw_compile_sf(const struct brw_compiler
*compiler
,
1185 const struct brw_sf_prog_key
*key
,
1186 struct brw_sf_prog_data
*prog_data
,
1187 struct brw_vue_map
*vue_map
,
1188 unsigned *final_assembly_size
);
1191 * Compile a clipper shader.
1193 * This is a fixed-function shader determined entirely by the shader key and
1196 * Returns the final assembly and the program's size.
1199 brw_compile_clip(const struct brw_compiler
*compiler
,
1201 const struct brw_clip_prog_key
*key
,
1202 struct brw_clip_prog_data
*prog_data
,
1203 struct brw_vue_map
*vue_map
,
1204 unsigned *final_assembly_size
);
1207 * Compile a fragment shader.
1209 * Returns the final assembly and the program's size.
1212 brw_compile_fs(const struct brw_compiler
*compiler
, void *log_data
,
1214 const struct brw_wm_prog_key
*key
,
1215 struct brw_wm_prog_data
*prog_data
,
1216 const struct nir_shader
*shader
,
1217 struct gl_program
*prog
,
1218 int shader_time_index8
,
1219 int shader_time_index16
,
1220 bool allow_spilling
,
1221 bool use_rep_send
, struct brw_vue_map
*vue_map
,
1225 * Compile a compute shader.
1227 * Returns the final assembly and the program's size.
1230 brw_compile_cs(const struct brw_compiler
*compiler
, void *log_data
,
1232 const struct brw_cs_prog_key
*key
,
1233 struct brw_cs_prog_data
*prog_data
,
1234 const struct nir_shader
*shader
,
1235 int shader_time_index
,
1238 static inline uint32_t
1239 encode_slm_size(unsigned gen
, uint32_t bytes
)
1241 uint32_t slm_size
= 0;
1243 /* Shared Local Memory is specified as powers of two, and encoded in
1244 * INTERFACE_DESCRIPTOR_DATA with the following representations:
1246 * Size | 0 kB | 1 kB | 2 kB | 4 kB | 8 kB | 16 kB | 32 kB | 64 kB |
1247 * -------------------------------------------------------------------
1248 * Gen7-8 | 0 | none | none | 1 | 2 | 4 | 8 | 16 |
1249 * -------------------------------------------------------------------
1250 * Gen9+ | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
1252 assert(bytes
<= 64 * 1024);
1255 /* Shared Local Memory Size is specified as powers of two. */
1256 slm_size
= util_next_power_of_two(bytes
);
1259 /* Use a minimum of 1kB; turn an exponent of 10 (1024 kB) into 1. */
1260 slm_size
= ffs(MAX2(slm_size
, 1024)) - 10;
1262 /* Use a minimum of 4kB; convert to the pre-Gen9 representation. */
1263 slm_size
= MAX2(slm_size
, 4096) / 4096;
1271 * Return true if the given shader stage is dispatched contiguously by the
1272 * relevant fixed function starting from channel 0 of the SIMD thread, which
1273 * implies that the dispatch mask of a thread can be assumed to have the form
1274 * '2^n - 1' for some n.
1277 brw_stage_has_packed_dispatch(MAYBE_UNUSED
const struct gen_device_info
*devinfo
,
1278 gl_shader_stage stage
,
1279 const struct brw_stage_prog_data
*prog_data
)
1281 /* The code below makes assumptions about the hardware's thread dispatch
1282 * behavior that could be proven wrong in future generations -- Make sure
1283 * to do a full test run with brw_fs_test_dispatch_packing() hooked up to
1284 * the NIR front-end before changing this assertion.
1286 assert(devinfo
->gen
<= 10);
1289 case MESA_SHADER_FRAGMENT
: {
1290 /* The PSD discards subspans coming in with no lit samples, which in the
1291 * per-pixel shading case implies that each subspan will either be fully
1292 * lit (due to the VMask being used to allow derivative computations),
1293 * or not dispatched at all. In per-sample dispatch mode individual
1294 * samples from the same subspan have a fixed relative location within
1295 * the SIMD thread, so dispatch of unlit samples cannot be avoided in
1296 * general and we should return false.
1298 const struct brw_wm_prog_data
*wm_prog_data
=
1299 (const struct brw_wm_prog_data
*)prog_data
;
1300 return !wm_prog_data
->persample_dispatch
;
1302 case MESA_SHADER_COMPUTE
:
1303 /* Compute shaders will be spawned with either a fully enabled dispatch
1304 * mask or with whatever bottom/right execution mask was given to the
1305 * GPGPU walker command to be used along the workgroup edges -- In both
1306 * cases the dispatch mask is required to be tightly packed for our
1307 * invocation index calculations to work.
1311 /* Most remaining fixed functions are limited to use a packed dispatch
1312 * mask due to the hardware representation of the dispatch mask as a
1313 * single counter representing the number of enabled channels.
1320 * Computes the first varying slot in the URB produced by the previous stage
1321 * that is used in the next stage. We do this by testing the varying slots in
1322 * the previous stage's vue map against the inputs read in the next stage.
1326 * - Each URB offset contains two varying slots and we can only skip a
1327 * full offset if both slots are unused, so the value we return here is always
1328 * rounded down to the closest multiple of two.
1330 * - gl_Layer and gl_ViewportIndex don't have their own varying slots, they are
1331 * part of the vue header, so if these are read we can't skip anything.
1334 brw_compute_first_urb_slot_required(uint64_t inputs_read
,
1335 const struct brw_vue_map
*prev_stage_vue_map
)
1337 if ((inputs_read
& (VARYING_BIT_LAYER
| VARYING_BIT_VIEWPORT
)) == 0) {
1338 for (int i
= 0; i
< prev_stage_vue_map
->num_slots
; i
++) {
1339 int varying
= prev_stage_vue_map
->slot_to_varying
[i
];
1340 if (varying
> 0 && (inputs_read
& BITFIELD64_BIT(varying
)) != 0)
1341 return ROUND_DOWN_TO(i
, 2);
1352 #endif /* BRW_COMPILER_H */