2 * Copyright 2010 Christoph Bumiller
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 in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * 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
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
23 #include "pipe/p_defines.h"
25 #include "tgsi/tgsi_ureg.h"
27 #include "nvc0/nvc0_context.h"
29 #include "codegen/nv50_ir_driver.h"
30 #include "nvc0/nve4_compute.h"
32 /* NOTE: Using a[0x270] in FP may cause an error even if we're using less than
33 * 124 scalar varying values.
36 nvc0_shader_input_address(unsigned sn
, unsigned si
)
39 case TGSI_SEMANTIC_TESSOUTER
: return 0x000 + si
* 0x4;
40 case TGSI_SEMANTIC_TESSINNER
: return 0x010 + si
* 0x4;
41 case TGSI_SEMANTIC_PATCH
: return 0x020 + si
* 0x10;
42 case TGSI_SEMANTIC_PRIMID
: return 0x060;
43 case TGSI_SEMANTIC_LAYER
: return 0x064;
44 case TGSI_SEMANTIC_VIEWPORT_INDEX
:return 0x068;
45 case TGSI_SEMANTIC_PSIZE
: return 0x06c;
46 case TGSI_SEMANTIC_POSITION
: return 0x070;
47 case TGSI_SEMANTIC_GENERIC
: return 0x080 + si
* 0x10;
48 case TGSI_SEMANTIC_FOG
: return 0x2e8;
49 case TGSI_SEMANTIC_COLOR
: return 0x280 + si
* 0x10;
50 case TGSI_SEMANTIC_BCOLOR
: return 0x2a0 + si
* 0x10;
51 case TGSI_SEMANTIC_CLIPDIST
: return 0x2c0 + si
* 0x10;
52 case TGSI_SEMANTIC_CLIPVERTEX
: return 0x270;
53 case TGSI_SEMANTIC_PCOORD
: return 0x2e0;
54 case TGSI_SEMANTIC_TESSCOORD
: return 0x2f0;
55 case TGSI_SEMANTIC_INSTANCEID
: return 0x2f8;
56 case TGSI_SEMANTIC_VERTEXID
: return 0x2fc;
57 case TGSI_SEMANTIC_TEXCOORD
: return 0x300 + si
* 0x10;
58 case TGSI_SEMANTIC_FACE
: return 0x3fc;
60 assert(!"invalid TGSI input semantic");
66 nvc0_shader_output_address(unsigned sn
, unsigned si
)
69 case TGSI_SEMANTIC_TESSOUTER
: return 0x000 + si
* 0x4;
70 case TGSI_SEMANTIC_TESSINNER
: return 0x010 + si
* 0x4;
71 case TGSI_SEMANTIC_PATCH
: return 0x020 + si
* 0x10;
72 case TGSI_SEMANTIC_PRIMID
: return 0x060;
73 case TGSI_SEMANTIC_LAYER
: return 0x064;
74 case TGSI_SEMANTIC_VIEWPORT_INDEX
:return 0x068;
75 case TGSI_SEMANTIC_PSIZE
: return 0x06c;
76 case TGSI_SEMANTIC_POSITION
: return 0x070;
77 case TGSI_SEMANTIC_GENERIC
: return 0x080 + si
* 0x10;
78 case TGSI_SEMANTIC_FOG
: return 0x2e8;
79 case TGSI_SEMANTIC_COLOR
: return 0x280 + si
* 0x10;
80 case TGSI_SEMANTIC_BCOLOR
: return 0x2a0 + si
* 0x10;
81 case TGSI_SEMANTIC_CLIPDIST
: return 0x2c0 + si
* 0x10;
82 case TGSI_SEMANTIC_CLIPVERTEX
: return 0x270;
83 case TGSI_SEMANTIC_TEXCOORD
: return 0x300 + si
* 0x10;
84 case TGSI_SEMANTIC_EDGEFLAG
: return ~0;
86 assert(!"invalid TGSI output semantic");
92 nvc0_vp_assign_input_slots(struct nv50_ir_prog_info
*info
)
96 for (n
= 0, i
= 0; i
< info
->numInputs
; ++i
) {
97 switch (info
->in
[i
].sn
) {
98 case TGSI_SEMANTIC_INSTANCEID
: /* for SM4 only, in TGSI they're SVs */
99 case TGSI_SEMANTIC_VERTEXID
:
100 info
->in
[i
].mask
= 0x1;
101 info
->in
[i
].slot
[0] =
102 nvc0_shader_input_address(info
->in
[i
].sn
, 0) / 4;
107 for (c
= 0; c
< 4; ++c
)
108 info
->in
[i
].slot
[c
] = (0x80 + n
* 0x10 + c
* 0x4) / 4;
116 nvc0_sp_assign_input_slots(struct nv50_ir_prog_info
*info
)
121 for (i
= 0; i
< info
->numInputs
; ++i
) {
122 offset
= nvc0_shader_input_address(info
->in
[i
].sn
, info
->in
[i
].si
);
124 for (c
= 0; c
< 4; ++c
)
125 info
->in
[i
].slot
[c
] = (offset
+ c
* 0x4) / 4;
132 nvc0_fp_assign_output_slots(struct nv50_ir_prog_info
*info
)
134 unsigned count
= info
->prop
.fp
.numColourResults
* 4;
137 for (i
= 0; i
< info
->numOutputs
; ++i
)
138 if (info
->out
[i
].sn
== TGSI_SEMANTIC_COLOR
)
139 for (c
= 0; c
< 4; ++c
)
140 info
->out
[i
].slot
[c
] = info
->out
[i
].si
* 4 + c
;
142 if (info
->io
.sampleMask
< PIPE_MAX_SHADER_OUTPUTS
)
143 info
->out
[info
->io
.sampleMask
].slot
[0] = count
++;
145 if (info
->target
>= 0xe0)
146 count
++; /* on Kepler, depth is always last colour reg + 2 */
148 if (info
->io
.fragDepth
< PIPE_MAX_SHADER_OUTPUTS
)
149 info
->out
[info
->io
.fragDepth
].slot
[2] = count
;
155 nvc0_sp_assign_output_slots(struct nv50_ir_prog_info
*info
)
160 for (i
= 0; i
< info
->numOutputs
; ++i
) {
161 offset
= nvc0_shader_output_address(info
->out
[i
].sn
, info
->out
[i
].si
);
163 for (c
= 0; c
< 4; ++c
)
164 info
->out
[i
].slot
[c
] = (offset
+ c
* 0x4) / 4;
171 nvc0_program_assign_varying_slots(struct nv50_ir_prog_info
*info
)
175 if (info
->type
== PIPE_SHADER_VERTEX
)
176 ret
= nvc0_vp_assign_input_slots(info
);
178 ret
= nvc0_sp_assign_input_slots(info
);
182 if (info
->type
== PIPE_SHADER_FRAGMENT
)
183 ret
= nvc0_fp_assign_output_slots(info
);
185 ret
= nvc0_sp_assign_output_slots(info
);
190 nvc0_vtgp_hdr_update_oread(struct nvc0_program
*vp
, uint8_t slot
)
192 uint8_t min
= (vp
->hdr
[4] >> 12) & 0xff;
193 uint8_t max
= (vp
->hdr
[4] >> 24);
195 min
= MIN2(min
, slot
);
196 max
= MAX2(max
, slot
);
198 vp
->hdr
[4] = (max
<< 24) | (min
<< 12);
201 /* Common part of header generation for VP, TCP, TEP and GP. */
203 nvc0_vtgp_gen_header(struct nvc0_program
*vp
, struct nv50_ir_prog_info
*info
)
207 for (i
= 0; i
< info
->numInputs
; ++i
) {
208 if (info
->in
[i
].patch
)
210 for (c
= 0; c
< 4; ++c
) {
211 a
= info
->in
[i
].slot
[c
];
212 if (info
->in
[i
].mask
& (1 << c
))
213 vp
->hdr
[5 + a
/ 32] |= 1 << (a
% 32);
217 for (i
= 0; i
< info
->numOutputs
; ++i
) {
218 if (info
->out
[i
].patch
)
220 for (c
= 0; c
< 4; ++c
) {
221 if (!(info
->out
[i
].mask
& (1 << c
)))
223 assert(info
->out
[i
].slot
[c
] >= 0x40 / 4);
224 a
= info
->out
[i
].slot
[c
] - 0x40 / 4;
225 vp
->hdr
[13 + a
/ 32] |= 1 << (a
% 32);
226 if (info
->out
[i
].oread
)
227 nvc0_vtgp_hdr_update_oread(vp
, info
->out
[i
].slot
[c
]);
231 for (i
= 0; i
< info
->numSysVals
; ++i
) {
232 switch (info
->sv
[i
].sn
) {
233 case TGSI_SEMANTIC_PRIMID
:
234 vp
->hdr
[5] |= 1 << 24;
236 case TGSI_SEMANTIC_INSTANCEID
:
237 vp
->hdr
[10] |= 1 << 30;
239 case TGSI_SEMANTIC_VERTEXID
:
240 vp
->hdr
[10] |= 1 << 31;
242 case TGSI_SEMANTIC_TESSCOORD
:
243 /* We don't have the mask, nor the slots populated. While this could
244 * be achieved, the vast majority of the time if either of the coords
245 * are read, then both will be read.
247 nvc0_vtgp_hdr_update_oread(vp
, 0x2f0 / 4);
248 nvc0_vtgp_hdr_update_oread(vp
, 0x2f4 / 4);
256 (1 << (info
->io
.clipDistances
+ info
->io
.cullDistances
)) - 1;
257 for (i
= 0; i
< info
->io
.cullDistances
; ++i
)
258 vp
->vp
.clip_mode
|= 1 << ((info
->io
.clipDistances
+ i
) * 4);
260 if (info
->io
.genUserClip
< 0)
261 vp
->vp
.num_ucps
= PIPE_MAX_CLIP_PLANES
+ 1; /* prevent rebuilding */
267 nvc0_vp_gen_header(struct nvc0_program
*vp
, struct nv50_ir_prog_info
*info
)
269 vp
->hdr
[0] = 0x20061 | (1 << 10);
270 vp
->hdr
[4] = 0xff000;
272 return nvc0_vtgp_gen_header(vp
, info
);
276 nvc0_tp_get_tess_mode(struct nvc0_program
*tp
, struct nv50_ir_prog_info
*info
)
278 if (info
->prop
.tp
.outputPrim
== PIPE_PRIM_MAX
) {
279 tp
->tp
.tess_mode
= ~0;
282 switch (info
->prop
.tp
.domain
) {
283 case PIPE_PRIM_LINES
:
284 tp
->tp
.tess_mode
= NVC0_3D_TESS_MODE_PRIM_ISOLINES
;
286 case PIPE_PRIM_TRIANGLES
:
287 tp
->tp
.tess_mode
= NVC0_3D_TESS_MODE_PRIM_TRIANGLES
;
288 if (info
->prop
.tp
.winding
> 0)
289 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_CW
;
291 case PIPE_PRIM_QUADS
:
292 tp
->tp
.tess_mode
= NVC0_3D_TESS_MODE_PRIM_QUADS
;
295 tp
->tp
.tess_mode
= ~0;
298 if (info
->prop
.tp
.outputPrim
!= PIPE_PRIM_POINTS
)
299 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_CONNECTED
;
301 switch (info
->prop
.tp
.partitioning
) {
302 case PIPE_TESS_SPACING_EQUAL
:
303 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_SPACING_EQUAL
;
305 case PIPE_TESS_SPACING_FRACTIONAL_ODD
:
306 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_SPACING_FRACTIONAL_ODD
;
308 case PIPE_TESS_SPACING_FRACTIONAL_EVEN
:
309 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_SPACING_FRACTIONAL_EVEN
;
312 assert(!"invalid tessellator partitioning");
318 nvc0_tcp_gen_header(struct nvc0_program
*tcp
, struct nv50_ir_prog_info
*info
)
320 unsigned opcs
= 6; /* output patch constants (at least the TessFactors) */
322 tcp
->tp
.input_patch_size
= info
->prop
.tp
.inputPatchSize
;
324 if (info
->numPatchConstants
)
325 opcs
= 8 + info
->numPatchConstants
* 4;
327 tcp
->hdr
[0] = 0x20061 | (2 << 10);
329 tcp
->hdr
[1] = opcs
<< 24;
330 tcp
->hdr
[2] = info
->prop
.tp
.outputPatchSize
<< 24;
332 tcp
->hdr
[4] = 0xff000; /* initial min/max parallel output read address */
334 nvc0_vtgp_gen_header(tcp
, info
);
336 nvc0_tp_get_tess_mode(tcp
, info
);
342 nvc0_tep_gen_header(struct nvc0_program
*tep
, struct nv50_ir_prog_info
*info
)
344 tep
->tp
.input_patch_size
= ~0;
346 tep
->hdr
[0] = 0x20061 | (3 << 10);
347 tep
->hdr
[4] = 0xff000;
349 nvc0_vtgp_gen_header(tep
, info
);
351 nvc0_tp_get_tess_mode(tep
, info
);
353 tep
->hdr
[18] |= 0x3 << 12; /* ? */
359 nvc0_gp_gen_header(struct nvc0_program
*gp
, struct nv50_ir_prog_info
*info
)
361 gp
->hdr
[0] = 0x20061 | (4 << 10);
363 gp
->hdr
[2] = MIN2(info
->prop
.gp
.instanceCount
, 32) << 24;
365 switch (info
->prop
.gp
.outputPrim
) {
366 case PIPE_PRIM_POINTS
:
367 gp
->hdr
[3] = 0x01000000;
368 gp
->hdr
[0] |= 0xf0000000;
370 case PIPE_PRIM_LINE_STRIP
:
371 gp
->hdr
[3] = 0x06000000;
372 gp
->hdr
[0] |= 0x10000000;
374 case PIPE_PRIM_TRIANGLE_STRIP
:
375 gp
->hdr
[3] = 0x07000000;
376 gp
->hdr
[0] |= 0x10000000;
383 gp
->hdr
[4] = MIN2(info
->prop
.gp
.maxVertices
, 1024);
385 return nvc0_vtgp_gen_header(gp
, info
);
388 #define NVC0_INTERP_FLAT (1 << 0)
389 #define NVC0_INTERP_PERSPECTIVE (2 << 0)
390 #define NVC0_INTERP_LINEAR (3 << 0)
391 #define NVC0_INTERP_CENTROID (1 << 2)
394 nvc0_hdr_interp_mode(const struct nv50_ir_varying
*var
)
397 return NVC0_INTERP_LINEAR
;
399 return NVC0_INTERP_FLAT
;
400 return NVC0_INTERP_PERSPECTIVE
;
404 nvc0_fp_gen_header(struct nvc0_program
*fp
, struct nv50_ir_prog_info
*info
)
408 /* just 00062 on Kepler */
409 fp
->hdr
[0] = 0x20062 | (5 << 10);
410 fp
->hdr
[5] = 0x80000000; /* getting a trap if FRAG_COORD_UMASK.w = 0 */
412 if (info
->prop
.fp
.usesDiscard
)
413 fp
->hdr
[0] |= 0x8000;
414 if (info
->prop
.fp
.numColourResults
> 1)
415 fp
->hdr
[0] |= 0x4000;
416 if (info
->io
.sampleMask
< PIPE_MAX_SHADER_OUTPUTS
)
418 if (info
->prop
.fp
.writesDepth
) {
420 fp
->flags
[0] = 0x11; /* deactivate ZCULL */
423 for (i
= 0; i
< info
->numInputs
; ++i
) {
424 m
= nvc0_hdr_interp_mode(&info
->in
[i
]);
425 if (info
->in
[i
].sn
== TGSI_SEMANTIC_COLOR
) {
426 fp
->fp
.colors
|= 1 << info
->in
[i
].si
;
428 fp
->fp
.color_interp
[info
->in
[i
].si
] = m
| (info
->in
[i
].mask
<< 4);
430 for (c
= 0; c
< 4; ++c
) {
431 if (!(info
->in
[i
].mask
& (1 << c
)))
433 a
= info
->in
[i
].slot
[c
];
434 if (info
->in
[i
].slot
[0] >= (0x060 / 4) &&
435 info
->in
[i
].slot
[0] <= (0x07c / 4)) {
436 fp
->hdr
[5] |= 1 << (24 + (a
- 0x060 / 4));
438 if (info
->in
[i
].slot
[0] >= (0x2c0 / 4) &&
439 info
->in
[i
].slot
[0] <= (0x2fc / 4)) {
440 fp
->hdr
[14] |= (1 << (a
- 0x280 / 4)) & 0x07ff0000;
442 if (info
->in
[i
].slot
[c
] < (0x040 / 4) ||
443 info
->in
[i
].slot
[c
] > (0x380 / 4))
446 if (info
->in
[i
].slot
[0] >= (0x300 / 4))
448 fp
->hdr
[4 + a
/ 32] |= m
<< (a
% 32);
453 for (i
= 0; i
< info
->numOutputs
; ++i
) {
454 if (info
->out
[i
].sn
== TGSI_SEMANTIC_COLOR
)
455 fp
->hdr
[18] |= 0xf << info
->out
[i
].slot
[0];
458 fp
->fp
.early_z
= info
->prop
.fp
.earlyFragTests
;
463 static struct nvc0_transform_feedback_state
*
464 nvc0_program_create_tfb_state(const struct nv50_ir_prog_info
*info
,
465 const struct pipe_stream_output_info
*pso
)
467 struct nvc0_transform_feedback_state
*tfb
;
470 tfb
= MALLOC_STRUCT(nvc0_transform_feedback_state
);
473 for (b
= 0; b
< 4; ++b
) {
474 tfb
->stride
[b
] = pso
->stride
[b
] * 4;
475 tfb
->varying_count
[b
] = 0;
477 memset(tfb
->varying_index
, 0xff, sizeof(tfb
->varying_index
)); /* = skip */
479 for (i
= 0; i
< pso
->num_outputs
; ++i
) {
480 unsigned s
= pso
->output
[i
].start_component
;
481 unsigned p
= pso
->output
[i
].dst_offset
;
482 b
= pso
->output
[i
].output_buffer
;
484 for (c
= 0; c
< pso
->output
[i
].num_components
; ++c
)
485 tfb
->varying_index
[b
][p
++] =
486 info
->out
[pso
->output
[i
].register_index
].slot
[s
+ c
];
488 tfb
->varying_count
[b
] = MAX2(tfb
->varying_count
[b
], p
);
489 tfb
->stream
[b
] = pso
->output
[i
].stream
;
491 for (b
= 0; b
< 4; ++b
) // zero unused indices (looks nicer)
492 for (c
= tfb
->varying_count
[b
]; c
& 3; ++c
)
493 tfb
->varying_index
[b
][c
] = 0;
500 nvc0_program_dump(struct nvc0_program
*prog
)
504 if (prog
->type
!= PIPE_SHADER_COMPUTE
) {
505 for (pos
= 0; pos
< ARRAY_SIZE(prog
->hdr
); ++pos
)
506 debug_printf("HDR[%02"PRIxPTR
"] = 0x%08x\n",
507 pos
* sizeof(prog
->hdr
[0]), prog
->hdr
[pos
]);
509 debug_printf("shader binary code (0x%x bytes):", prog
->code_size
);
510 for (pos
= 0; pos
< prog
->code_size
/ 4; ++pos
) {
513 debug_printf("%08x ", prog
->code
[pos
]);
520 nvc0_program_translate(struct nvc0_program
*prog
, uint16_t chipset
,
521 struct pipe_debug_callback
*debug
)
523 struct nv50_ir_prog_info
*info
;
526 info
= CALLOC_STRUCT(nv50_ir_prog_info
);
530 info
->type
= prog
->type
;
531 info
->target
= chipset
;
532 info
->bin
.sourceRep
= NV50_PROGRAM_IR_TGSI
;
533 info
->bin
.source
= (void *)prog
->pipe
.tokens
;
535 info
->io
.genUserClip
= prog
->vp
.num_ucps
;
536 info
->io
.ucpBase
= 256;
537 info
->io
.ucpCBSlot
= 15;
539 if (prog
->type
== PIPE_SHADER_COMPUTE
) {
540 if (chipset
>= NVISA_GK104_CHIPSET
) {
541 info
->io
.resInfoCBSlot
= 0;
542 info
->io
.texBindBase
= NVE4_CP_INPUT_TEX(0);
543 info
->io
.suInfoBase
= NVE4_CP_INPUT_SUF(0);
544 info
->prop
.cp
.gridInfoBase
= NVE4_CP_INPUT_GRID_INFO(0);
546 info
->io
.msInfoCBSlot
= 0;
547 info
->io
.msInfoBase
= NVE4_CP_INPUT_MS_OFFSETS
;
549 if (chipset
>= NVISA_GK104_CHIPSET
) {
550 info
->io
.texBindBase
= 0x20;
551 info
->io
.suInfoBase
= 0; /* TODO */
553 info
->io
.resInfoCBSlot
= 15;
554 info
->io
.sampleInfoBase
= 256 + 128;
555 info
->io
.msInfoCBSlot
= 15;
556 info
->io
.msInfoBase
= 0; /* TODO */
559 info
->assignSlots
= nvc0_program_assign_varying_slots
;
562 info
->optLevel
= debug_get_num_option("NV50_PROG_OPTIMIZE", 3);
563 info
->dbgFlags
= debug_get_num_option("NV50_PROG_DEBUG", 0);
568 ret
= nv50_ir_generate_code(info
);
570 NOUVEAU_ERR("shader translation failed: %i\n", ret
);
573 if (prog
->type
!= PIPE_SHADER_COMPUTE
)
574 FREE(info
->bin
.syms
);
576 prog
->code
= info
->bin
.code
;
577 prog
->code_size
= info
->bin
.codeSize
;
578 prog
->immd_data
= info
->immd
.buf
;
579 prog
->immd_size
= info
->immd
.bufSize
;
580 prog
->relocs
= info
->bin
.relocData
;
581 prog
->interps
= info
->bin
.interpData
;
582 prog
->num_gprs
= MAX2(4, (info
->bin
.maxGPR
+ 1));
583 prog
->num_barriers
= info
->numBarriers
;
585 prog
->vp
.need_vertex_id
= info
->io
.vertexId
< PIPE_MAX_SHADER_INPUTS
;
587 if (info
->io
.edgeFlagOut
< PIPE_MAX_ATTRIBS
)
588 info
->out
[info
->io
.edgeFlagOut
].mask
= 0; /* for headergen */
589 prog
->vp
.edgeflag
= info
->io
.edgeFlagIn
;
591 switch (prog
->type
) {
592 case PIPE_SHADER_VERTEX
:
593 ret
= nvc0_vp_gen_header(prog
, info
);
595 case PIPE_SHADER_TESS_CTRL
:
596 ret
= nvc0_tcp_gen_header(prog
, info
);
598 case PIPE_SHADER_TESS_EVAL
:
599 ret
= nvc0_tep_gen_header(prog
, info
);
601 case PIPE_SHADER_GEOMETRY
:
602 ret
= nvc0_gp_gen_header(prog
, info
);
604 case PIPE_SHADER_FRAGMENT
:
605 ret
= nvc0_fp_gen_header(prog
, info
);
607 case PIPE_SHADER_COMPUTE
:
608 prog
->cp
.syms
= info
->bin
.syms
;
609 prog
->cp
.num_syms
= info
->bin
.numSyms
;
613 NOUVEAU_ERR("unknown program type: %u\n", prog
->type
);
619 if (info
->bin
.tlsSpace
) {
620 assert(info
->bin
.tlsSpace
< (1 << 24));
621 prog
->hdr
[0] |= 1 << 26;
622 prog
->hdr
[1] |= align(info
->bin
.tlsSpace
, 0x10); /* l[] size */
623 prog
->need_tls
= true;
625 /* TODO: factor 2 only needed where joinat/precont is used,
626 * and we only have to count non-uniform branches
629 if ((info->maxCFDepth * 2) > 16) {
630 prog->hdr[2] |= (((info->maxCFDepth * 2) + 47) / 48) * 0x200;
631 prog->need_tls = true;
634 if (info
->io
.globalAccess
)
635 prog
->hdr
[0] |= 1 << 16;
637 prog
->hdr
[0] |= 1 << 27;
639 if (prog
->pipe
.stream_output
.num_outputs
)
640 prog
->tfb
= nvc0_program_create_tfb_state(info
,
641 &prog
->pipe
.stream_output
);
643 pipe_debug_message(debug
, SHADER_INFO
,
644 "type: %d, local: %d, gpr: %d, inst: %d, bytes: %d",
645 prog
->type
, info
->bin
.tlsSpace
, prog
->num_gprs
,
646 info
->bin
.instructions
, info
->bin
.codeSize
);
654 nvc0_program_upload_code(struct nvc0_context
*nvc0
, struct nvc0_program
*prog
)
656 struct nvc0_screen
*screen
= nvc0
->screen
;
657 const bool is_cp
= prog
->type
== PIPE_SHADER_COMPUTE
;
659 uint32_t size
= prog
->code_size
+ (is_cp
? 0 : NVC0_SHADER_HEADER_SIZE
);
660 uint32_t lib_pos
= screen
->lib_code
->start
;
663 /* c[] bindings need to be aligned to 0x100, but we could use relocations
665 if (prog
->immd_size
) {
666 prog
->immd_base
= size
;
667 size
= align(size
, 0x40);
668 size
+= prog
->immd_size
+ 0xc0; /* add 0xc0 for align 0x40 -> 0x100 */
670 /* On Fermi, SP_START_ID must be aligned to 0x40.
671 * On Kepler, the first instruction must be aligned to 0x80 because
672 * latency information is expected only at certain positions.
674 if (screen
->base
.class_3d
>= NVE4_3D_CLASS
)
675 size
= size
+ (is_cp
? 0x40 : 0x70);
676 size
= align(size
, 0x40);
678 ret
= nouveau_heap_alloc(screen
->text_heap
, size
, prog
, &prog
->mem
);
680 struct nouveau_heap
*heap
= screen
->text_heap
;
681 /* Note that the code library, which is allocated before anything else,
682 * does not have a priv pointer. We can stop once we hit it.
684 while (heap
->next
&& heap
->next
->priv
) {
685 struct nvc0_program
*evict
= heap
->next
->priv
;
686 nouveau_heap_free(&evict
->mem
);
688 debug_printf("WARNING: out of code space, evicting all shaders.\n");
689 ret
= nouveau_heap_alloc(heap
, size
, prog
, &prog
->mem
);
691 NOUVEAU_ERR("shader too large (0x%x) to fit in code space ?\n", size
);
694 IMMED_NVC0(nvc0
->base
.pushbuf
, NVC0_3D(SERIALIZE
), 0);
696 prog
->code_base
= prog
->mem
->start
;
697 prog
->immd_base
= align(prog
->mem
->start
+ prog
->immd_base
, 0x100);
698 assert((prog
->immd_size
== 0) || (prog
->immd_base
+ prog
->immd_size
<=
699 prog
->mem
->start
+ prog
->mem
->size
));
702 if (screen
->base
.class_3d
>= NVE4_3D_CLASS
) {
703 switch (prog
->mem
->start
& 0xff) {
704 case 0x40: prog
->code_base
+= 0x70; break;
705 case 0x80: prog
->code_base
+= 0x30; break;
706 case 0xc0: prog
->code_base
+= 0x70; break;
708 prog
->code_base
+= 0x30;
709 assert((prog
->mem
->start
& 0xff) == 0x00);
713 code_pos
= prog
->code_base
+ NVC0_SHADER_HEADER_SIZE
;
715 if (screen
->base
.class_3d
>= NVE4_3D_CLASS
) {
716 if (prog
->mem
->start
& 0x40)
717 prog
->code_base
+= 0x40;
718 assert((prog
->code_base
& 0x7f) == 0x00);
720 code_pos
= prog
->code_base
;
724 nv50_ir_relocate_code(prog
->relocs
, prog
->code
, code_pos
, lib_pos
, 0);
726 nv50_ir_change_interp(prog
->interps
, prog
->code
,
727 prog
->fp
.force_persample_interp
,
729 for (int i
= 0; i
< 2; i
++) {
730 unsigned mask
= prog
->fp
.color_interp
[i
] >> 4;
731 unsigned interp
= prog
->fp
.color_interp
[i
] & 3;
734 prog
->hdr
[14] &= ~(0xff << (8 * i
));
735 if (prog
->fp
.flatshade
)
736 interp
= NVC0_INTERP_FLAT
;
737 for (int c
= 0; c
< 4; c
++)
739 prog
->hdr
[14] |= interp
<< (2 * (4 * i
+ c
));
744 if (debug_get_bool_option("NV50_PROG_DEBUG", false))
745 nvc0_program_dump(prog
);
749 nvc0
->base
.push_data(&nvc0
->base
, screen
->text
, prog
->code_base
,
750 NV_VRAM_DOMAIN(&screen
->base
), NVC0_SHADER_HEADER_SIZE
, prog
->hdr
);
751 nvc0
->base
.push_data(&nvc0
->base
, screen
->text
, code_pos
,
752 NV_VRAM_DOMAIN(&screen
->base
), prog
->code_size
, prog
->code
);
754 nvc0
->base
.push_data(&nvc0
->base
,
755 screen
->text
, prog
->immd_base
, NV_VRAM_DOMAIN(&screen
->base
),
756 prog
->immd_size
, prog
->immd_data
);
758 BEGIN_NVC0(nvc0
->base
.pushbuf
, NVC0_3D(MEM_BARRIER
), 1);
759 PUSH_DATA (nvc0
->base
.pushbuf
, 0x1011);
764 /* Upload code for builtin functions like integer division emulation. */
766 nvc0_program_library_upload(struct nvc0_context
*nvc0
)
768 struct nvc0_screen
*screen
= nvc0
->screen
;
771 const uint32_t *code
;
773 if (screen
->lib_code
)
776 nv50_ir_get_target_library(screen
->base
.device
->chipset
, &code
, &size
);
780 ret
= nouveau_heap_alloc(screen
->text_heap
, align(size
, 0x100), NULL
,
785 nvc0
->base
.push_data(&nvc0
->base
,
786 screen
->text
, screen
->lib_code
->start
, NV_VRAM_DOMAIN(&screen
->base
),
788 /* no need for a memory barrier, will be emitted with first program */
792 nvc0_program_destroy(struct nvc0_context
*nvc0
, struct nvc0_program
*prog
)
794 const struct pipe_shader_state pipe
= prog
->pipe
;
795 const ubyte type
= prog
->type
;
798 nouveau_heap_free(&prog
->mem
);
799 FREE(prog
->code
); /* may be 0 for hardcoded shaders */
800 FREE(prog
->immd_data
);
803 if (prog
->type
== PIPE_SHADER_COMPUTE
&& prog
->cp
.syms
)
806 if (nvc0
->state
.tfb
== prog
->tfb
)
807 nvc0
->state
.tfb
= NULL
;
811 memset(prog
, 0, sizeof(*prog
));
818 nvc0_program_symbol_offset(const struct nvc0_program
*prog
, uint32_t label
)
820 const struct nv50_ir_prog_symbol
*syms
=
821 (const struct nv50_ir_prog_symbol
*)prog
->cp
.syms
;
824 if (prog
->type
!= PIPE_SHADER_COMPUTE
)
825 base
= NVC0_SHADER_HEADER_SIZE
;
826 for (i
= 0; i
< prog
->cp
.num_syms
; ++i
)
827 if (syms
[i
].label
== label
)
828 return prog
->code_base
+ base
+ syms
[i
].offset
;
829 return prog
->code_base
; /* no symbols or symbol not found */
833 nvc0_program_init_tcp_empty(struct nvc0_context
*nvc0
)
835 struct ureg_program
*ureg
;
837 ureg
= ureg_create(TGSI_PROCESSOR_TESS_CTRL
);
841 ureg_property(ureg
, TGSI_PROPERTY_TCS_VERTICES_OUT
, 1);
844 nvc0
->tcp_empty
= ureg_create_shader_and_destroy(ureg
, &nvc0
->base
.pipe
);