2 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 * Rob Clark <robclark@freedesktop.org>
32 #include "compiler/shader_enums.h"
33 #include "compiler/nir/nir.h"
34 #include "util/bitscan.h"
40 /* driver param indices: */
41 enum ir3_driver_param
{
42 /* compute shader driver params: */
43 IR3_DP_NUM_WORK_GROUPS_X
= 0,
44 IR3_DP_NUM_WORK_GROUPS_Y
= 1,
45 IR3_DP_NUM_WORK_GROUPS_Z
= 2,
46 IR3_DP_LOCAL_GROUP_SIZE_X
= 4,
47 IR3_DP_LOCAL_GROUP_SIZE_Y
= 5,
48 IR3_DP_LOCAL_GROUP_SIZE_Z
= 6,
49 /* NOTE: gl_NumWorkGroups should be vec4 aligned because
50 * glDispatchComputeIndirect() needs to load these from
51 * the info->indirect buffer. Keep that in mind when/if
52 * adding any addition CS driver params.
54 IR3_DP_CS_COUNT
= 8, /* must be aligned to vec4 */
56 /* vertex shader driver params: */
57 IR3_DP_VTXID_BASE
= 0,
58 IR3_DP_VTXCNT_MAX
= 1,
59 /* user-clip-plane components, up to 8x vec4's: */
63 IR3_DP_VS_COUNT
= 36 /* must be aligned to vec4 */
66 #define IR3_MAX_SHADER_BUFFERS 32
67 #define IR3_MAX_SHADER_IMAGES 32
68 #define IR3_MAX_SO_BUFFERS 4
69 #define IR3_MAX_SO_OUTPUTS 64
72 * For consts needed to pass internal values to shader which may or may not
73 * be required, rather than allocating worst-case const space, we scan the
74 * shader and allocate consts as-needed:
76 * + SSBO sizes: only needed if shader has a get_buffer_size intrinsic
79 * + Image dimensions: needed to calculate pixel offset, but only for
80 * images that have a image_store intrinsic
82 struct ir3_driver_const_layout
{
84 uint32_t mask
; /* bitmask of SSBOs that have get_buffer_size */
85 uint32_t count
; /* number of consts allocated */
86 /* one const allocated per SSBO which has get_buffer_size,
87 * ssbo_sizes.off[ssbo_id] is offset from start of ssbo_sizes
90 uint32_t off
[IR3_MAX_SHADER_BUFFERS
];
94 uint32_t mask
; /* bitmask of images that have image_store */
95 uint32_t count
; /* number of consts allocated */
96 /* three const allocated per image which has image_store:
97 * + cpp (bytes per pixel)
99 * + array_pitch (z pitch)
101 uint32_t off
[IR3_MAX_SHADER_IMAGES
];
106 * A single output for vertex transform feedback.
108 struct ir3_stream_output
{
109 unsigned register_index
:6; /**< 0 to 63 (OUT index) */
110 unsigned start_component
:2; /** 0 to 3 */
111 unsigned num_components
:3; /** 1 to 4 */
112 unsigned output_buffer
:3; /**< 0 to PIPE_MAX_SO_BUFFERS */
113 unsigned dst_offset
:16; /**< offset into the buffer in dwords */
114 unsigned stream
:2; /**< 0 to 3 */
118 * Stream output for vertex transform feedback.
120 struct ir3_stream_output_info
{
121 unsigned num_outputs
;
122 /** stride for an entire vertex for each buffer in dwords */
123 uint16_t stride
[IR3_MAX_SO_BUFFERS
];
126 * Array of stream outputs, in the order they are to be written in.
127 * Selected components are tightly packed into the output buffer.
129 struct ir3_stream_output output
[IR3_MAX_SO_OUTPUTS
];
132 /* Configuration key used to identify a shader variant.. different
133 * shader variants can be used to implement features not supported
134 * in hw (two sided color), binning-pass vertex shader, etc.
136 struct ir3_shader_key
{
140 * Combined Vertex/Fragment shader parameters:
142 unsigned ucp_enables
: 8;
144 /* do we need to check {v,f}saturate_{s,t,r}? */
145 unsigned has_per_samp
: 1;
148 * Vertex shader variant parameters:
150 unsigned vclamp_color
: 1;
153 * Fragment shader variant parameters:
155 unsigned color_two_side
: 1;
156 unsigned half_precision
: 1;
157 /* used when shader needs to handle flat varyings (a4xx)
158 * for front/back color inputs to frag shader:
160 unsigned rasterflat
: 1;
161 unsigned fclamp_color
: 1;
166 /* bitmask of sampler which needs coords clamped for vertex
169 uint16_t vsaturate_s
, vsaturate_t
, vsaturate_r
;
171 /* bitmask of sampler which needs coords clamped for frag
174 uint16_t fsaturate_s
, fsaturate_t
, fsaturate_r
;
176 /* bitmask of ms shifts */
177 uint32_t vsamples
, fsamples
;
179 /* bitmask of samplers which need astc srgb workaround: */
180 uint16_t vastc_srgb
, fastc_srgb
;
184 ir3_shader_key_equal(struct ir3_shader_key
*a
, struct ir3_shader_key
*b
)
186 /* slow-path if we need to check {v,f}saturate_{s,t,r} */
187 if (a
->has_per_samp
|| b
->has_per_samp
)
188 return memcmp(a
, b
, sizeof(struct ir3_shader_key
)) == 0;
189 return a
->global
== b
->global
;
192 /* will the two keys produce different lowering for a fragment shader? */
194 ir3_shader_key_changes_fs(struct ir3_shader_key
*key
, struct ir3_shader_key
*last_key
)
196 if (last_key
->has_per_samp
|| key
->has_per_samp
) {
197 if ((last_key
->fsaturate_s
!= key
->fsaturate_s
) ||
198 (last_key
->fsaturate_t
!= key
->fsaturate_t
) ||
199 (last_key
->fsaturate_r
!= key
->fsaturate_r
) ||
200 (last_key
->fsamples
!= key
->fsamples
) ||
201 (last_key
->fastc_srgb
!= key
->fastc_srgb
))
205 if (last_key
->fclamp_color
!= key
->fclamp_color
)
208 if (last_key
->color_two_side
!= key
->color_two_side
)
211 if (last_key
->half_precision
!= key
->half_precision
)
214 if (last_key
->rasterflat
!= key
->rasterflat
)
217 if (last_key
->ucp_enables
!= key
->ucp_enables
)
223 /* will the two keys produce different lowering for a vertex shader? */
225 ir3_shader_key_changes_vs(struct ir3_shader_key
*key
, struct ir3_shader_key
*last_key
)
227 if (last_key
->has_per_samp
|| key
->has_per_samp
) {
228 if ((last_key
->vsaturate_s
!= key
->vsaturate_s
) ||
229 (last_key
->vsaturate_t
!= key
->vsaturate_t
) ||
230 (last_key
->vsaturate_r
!= key
->vsaturate_r
) ||
231 (last_key
->vsamples
!= key
->vsamples
) ||
232 (last_key
->vastc_srgb
!= key
->vastc_srgb
))
236 if (last_key
->vclamp_color
!= key
->vclamp_color
)
239 if (last_key
->ucp_enables
!= key
->ucp_enables
)
245 /* clears shader-key flags which don't apply to the given shader
249 ir3_normalize_key(struct ir3_shader_key
*key
, gl_shader_stage type
)
252 case MESA_SHADER_FRAGMENT
:
253 if (key
->has_per_samp
) {
254 key
->vsaturate_s
= 0;
255 key
->vsaturate_t
= 0;
256 key
->vsaturate_r
= 0;
261 case MESA_SHADER_VERTEX
:
262 key
->color_two_side
= false;
263 key
->half_precision
= false;
264 key
->rasterflat
= false;
265 if (key
->has_per_samp
) {
266 key
->fsaturate_s
= 0;
267 key
->fsaturate_t
= 0;
268 key
->fsaturate_r
= 0;
280 struct ir3_shader_variant
{
283 /* variant id (for debug) */
286 struct ir3_shader_key key
;
288 /* vertex shaders can have an extra version for hwbinning pass,
289 * which is pointed to by so->binning:
292 struct ir3_shader_variant
*binning
;
294 struct ir3_driver_const_layout const_layout
;
295 struct ir3_info info
;
298 /* the instructions length is in units of instruction groups
299 * (4 instructions for a3xx, 16 instructions for a4xx.. each
300 * instruction is 2 dwords):
304 /* the constants length is in units of vec4's, and is the sum of
305 * the uniforms and the built-in compiler constants
309 /* number of uniforms (in vec4), not including built-in compiler
312 unsigned num_uniforms
;
317 * + Let the frag shader determine the position/compmask for the
318 * varyings, since it is the place where we know if the varying
319 * is actually used, and if so, which components are used. So
320 * what the hw calls "outloc" is taken from the "inloc" of the
322 * + From the vert shader, we only need the output regid
325 bool frag_coord
, frag_face
, color0_mrt
;
327 /* NOTE: for input/outputs, slot is:
328 * gl_vert_attrib - for VS inputs
329 * gl_varying_slot - for VS output / FS input
330 * gl_frag_result - for FS output
333 /* varyings/outputs: */
334 unsigned outputs_count
;
338 } outputs
[16 + 2]; /* +POSITION +PSIZE */
339 bool writes_pos
, writes_psize
;
341 /* attributes (VS) / varyings (FS):
342 * Note that sysval's should come *after* normal inputs.
344 unsigned inputs_count
;
350 /* location of input (ie. offset passed to bary.f, etc). This
351 * matches the SP_VS_VPC_DST_REG.OUTLOCn value (a3xx and a4xx
352 * have the OUTLOCn value offset by 8, presumably to account
353 * for gl_Position/gl_PointSize)
356 /* vertex shader specific: */
357 bool sysval
: 1; /* slot is a gl_system_value */
358 /* fragment shader specific: */
359 bool bary
: 1; /* fetched varying (vs one loaded into reg) */
360 bool rasterflat
: 1; /* special handling for emit->rasterflat */
361 enum glsl_interp_mode interpolate
;
362 } inputs
[16 + 2]; /* +POSITION +FACE */
364 /* sum of input components (scalar). For frag shaders, it only counts
365 * the varying inputs:
369 /* For frag shaders, the total number of inputs (not scalar,
370 * ie. SP_VS_PARAM_REG.TOTALVSOUTVAR)
374 /* number of samplers/textures (which are currently 1:1): */
377 /* do we have one or more SSBO instructions: */
380 /* do we have kill instructions: */
383 /* Layout of constant registers, each section (in vec4). Pointer size
384 * is 32b (a3xx, a4xx), or 64b (a5xx+), which effects the size of the
385 * UBO and stream-out consts.
388 /* user const start at zero */
390 /* NOTE that a3xx might need a section for SSBO addresses too */
393 unsigned driver_param
;
398 unsigned immediates_count
;
399 unsigned immediates_size
;
404 /* for astc srgb workaround, the number/base of additional
405 * alpha tex states we need, and index of original tex states
408 unsigned base
, count
;
409 unsigned orig_idx
[16];
412 /* shader variants form a linked list: */
413 struct ir3_shader_variant
*next
;
415 /* replicated here to avoid passing extra ptrs everywhere: */
416 gl_shader_stage type
;
417 struct ir3_shader
*shader
;
421 gl_shader_stage type
;
423 /* shader id (for debug): */
425 uint32_t variant_count
;
427 /* so we know when we can disable TGSI related hacks: */
430 struct ir3_compiler
*compiler
;
432 struct nir_shader
*nir
;
433 struct ir3_stream_output_info stream_output
;
435 struct ir3_shader_variant
*variants
;
438 void * ir3_shader_assemble(struct ir3_shader_variant
*v
, uint32_t gpu_id
);
439 struct ir3_shader_variant
* ir3_shader_get_variant(struct ir3_shader
*shader
,
440 struct ir3_shader_key
*key
, bool binning_pass
, bool *created
);
441 struct ir3_shader
* ir3_shader_from_nir(struct ir3_compiler
*compiler
, nir_shader
*nir
);
442 void ir3_shader_destroy(struct ir3_shader
*shader
);
443 void ir3_shader_disasm(struct ir3_shader_variant
*so
, uint32_t *bin
, FILE *out
);
444 uint64_t ir3_shader_outputs(const struct ir3_shader
*so
);
447 ir3_glsl_type_size(const struct glsl_type
*type
);
449 static inline const char *
450 ir3_shader_stage(struct ir3_shader
*shader
)
452 switch (shader
->type
) {
453 case MESA_SHADER_VERTEX
: return "VERT";
454 case MESA_SHADER_FRAGMENT
: return "FRAG";
455 case MESA_SHADER_COMPUTE
: return "CL";
457 unreachable("invalid type");
467 ir3_find_output(const struct ir3_shader_variant
*so
, gl_varying_slot slot
)
471 for (j
= 0; j
< so
->outputs_count
; j
++)
472 if (so
->outputs
[j
].slot
== slot
)
475 /* it seems optional to have a OUT.BCOLOR[n] for each OUT.COLOR[n]
476 * in the vertex shader.. but the fragment shader doesn't know this
477 * so it will always have both IN.COLOR[n] and IN.BCOLOR[n]. So
478 * at link time if there is no matching OUT.BCOLOR[n], we must map
479 * OUT.COLOR[n] to IN.BCOLOR[n]. And visa versa if there is only
480 * a OUT.BCOLOR[n] but no matching OUT.COLOR[n]
482 if (slot
== VARYING_SLOT_BFC0
) {
483 slot
= VARYING_SLOT_COL0
;
484 } else if (slot
== VARYING_SLOT_BFC1
) {
485 slot
= VARYING_SLOT_COL1
;
486 } else if (slot
== VARYING_SLOT_COL0
) {
487 slot
= VARYING_SLOT_BFC0
;
488 } else if (slot
== VARYING_SLOT_COL1
) {
489 slot
= VARYING_SLOT_BFC1
;
494 for (j
= 0; j
< so
->outputs_count
; j
++)
495 if (so
->outputs
[j
].slot
== slot
)
504 ir3_next_varying(const struct ir3_shader_variant
*so
, int i
)
506 while (++i
< so
->inputs_count
)
507 if (so
->inputs
[i
].compmask
&& so
->inputs
[i
].bary
)
512 struct ir3_shader_linkage
{
523 ir3_link_add(struct ir3_shader_linkage
*l
, uint8_t regid
, uint8_t compmask
, uint8_t loc
)
527 debug_assert(i
< ARRAY_SIZE(l
->var
));
529 l
->var
[i
].regid
= regid
;
530 l
->var
[i
].compmask
= compmask
;
532 l
->max_loc
= MAX2(l
->max_loc
, loc
+ util_last_bit(compmask
));
536 ir3_link_shaders(struct ir3_shader_linkage
*l
,
537 const struct ir3_shader_variant
*vs
,
538 const struct ir3_shader_variant
*fs
)
542 while (l
->cnt
< ARRAY_SIZE(l
->var
)) {
543 j
= ir3_next_varying(fs
, j
);
545 if (j
>= fs
->inputs_count
)
548 if (fs
->inputs
[j
].inloc
>= fs
->total_in
)
551 k
= ir3_find_output(vs
, fs
->inputs
[j
].slot
);
553 ir3_link_add(l
, vs
->outputs
[k
].regid
,
554 fs
->inputs
[j
].compmask
, fs
->inputs
[j
].inloc
);
558 static inline uint32_t
559 ir3_find_output_regid(const struct ir3_shader_variant
*so
, unsigned slot
)
562 for (j
= 0; j
< so
->outputs_count
; j
++)
563 if (so
->outputs
[j
].slot
== slot
)
564 return so
->outputs
[j
].regid
;
568 static inline uint32_t
569 ir3_find_sysval_regid(const struct ir3_shader_variant
*so
, unsigned slot
)
572 for (j
= 0; j
< so
->inputs_count
; j
++)
573 if (so
->inputs
[j
].sysval
&& (so
->inputs
[j
].slot
== slot
))
574 return so
->inputs
[j
].regid
;
578 /* calculate register footprint in terms of half-regs (ie. one full
579 * reg counts as two half-regs).
581 static inline uint32_t
582 ir3_shader_halfregs(const struct ir3_shader_variant
*v
)
584 return (2 * (v
->info
.max_reg
+ 1)) + (v
->info
.max_half_reg
+ 1);
587 #endif /* IR3_SHADER_H_ */