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 "nvc0/nvc0_context.h"
27 #include "codegen/nv50_ir_driver.h"
28 #include "nvc0/nve4_compute.h"
30 /* NOTE: Using a[0x270] in FP may cause an error even if we're using less than
31 * 124 scalar varying values.
34 nvc0_shader_input_address(unsigned sn
, unsigned si
, unsigned ubase
)
37 case TGSI_SEMANTIC_TESSOUTER
: return 0x000 + si
* 0x4;
38 case TGSI_SEMANTIC_TESSINNER
: return 0x010 + si
* 0x4;
39 case TGSI_SEMANTIC_PATCH
: return 0x020 + si
* 0x10;
40 case TGSI_SEMANTIC_PRIMID
: return 0x060;
41 case TGSI_SEMANTIC_LAYER
: return 0x064;
42 case TGSI_SEMANTIC_VIEWPORT_INDEX
:return 0x068;
43 case TGSI_SEMANTIC_PSIZE
: return 0x06c;
44 case TGSI_SEMANTIC_POSITION
: return 0x070;
45 case TGSI_SEMANTIC_GENERIC
: return ubase
+ si
* 0x10;
46 case TGSI_SEMANTIC_FOG
: return 0x2e8;
47 case TGSI_SEMANTIC_COLOR
: return 0x280 + si
* 0x10;
48 case TGSI_SEMANTIC_BCOLOR
: return 0x2a0 + si
* 0x10;
49 case TGSI_SEMANTIC_CLIPDIST
: return 0x2c0 + si
* 0x10;
50 case TGSI_SEMANTIC_CLIPVERTEX
: return 0x270;
51 case TGSI_SEMANTIC_PCOORD
: return 0x2e0;
52 case TGSI_SEMANTIC_TESSCOORD
: return 0x2f0;
53 case TGSI_SEMANTIC_INSTANCEID
: return 0x2f8;
54 case TGSI_SEMANTIC_VERTEXID
: return 0x2fc;
55 case TGSI_SEMANTIC_TEXCOORD
: return 0x300 + si
* 0x10;
56 case TGSI_SEMANTIC_FACE
: return 0x3fc;
58 assert(!"invalid TGSI input semantic");
64 nvc0_shader_output_address(unsigned sn
, unsigned si
, unsigned ubase
)
67 case TGSI_SEMANTIC_TESSOUTER
: return 0x000 + si
* 0x4;
68 case TGSI_SEMANTIC_TESSINNER
: return 0x010 + si
* 0x4;
69 case TGSI_SEMANTIC_PATCH
: return 0x020 + si
* 0x10;
70 case TGSI_SEMANTIC_PRIMID
: return 0x060;
71 case TGSI_SEMANTIC_LAYER
: return 0x064;
72 case TGSI_SEMANTIC_VIEWPORT_INDEX
:return 0x068;
73 case TGSI_SEMANTIC_PSIZE
: return 0x06c;
74 case TGSI_SEMANTIC_POSITION
: return 0x070;
75 case TGSI_SEMANTIC_GENERIC
: return ubase
+ si
* 0x10;
76 case TGSI_SEMANTIC_FOG
: return 0x2e8;
77 case TGSI_SEMANTIC_COLOR
: return 0x280 + si
* 0x10;
78 case TGSI_SEMANTIC_BCOLOR
: return 0x2a0 + si
* 0x10;
79 case TGSI_SEMANTIC_CLIPDIST
: return 0x2c0 + si
* 0x10;
80 case TGSI_SEMANTIC_CLIPVERTEX
: return 0x270;
81 case TGSI_SEMANTIC_TEXCOORD
: return 0x300 + si
* 0x10;
82 case TGSI_SEMANTIC_EDGEFLAG
: return ~0;
84 assert(!"invalid TGSI output semantic");
90 nvc0_vp_assign_input_slots(struct nv50_ir_prog_info
*info
)
94 for (n
= 0, i
= 0; i
< info
->numInputs
; ++i
) {
95 switch (info
->in
[i
].sn
) {
96 case TGSI_SEMANTIC_INSTANCEID
: /* for SM4 only, in TGSI they're SVs */
97 case TGSI_SEMANTIC_VERTEXID
:
98 info
->in
[i
].mask
= 0x1;
100 nvc0_shader_input_address(info
->in
[i
].sn
, 0, 0) / 4;
105 for (c
= 0; c
< 4; ++c
)
106 info
->in
[i
].slot
[c
] = (0x80 + n
* 0x10 + c
* 0x4) / 4;
114 nvc0_sp_assign_input_slots(struct nv50_ir_prog_info
*info
)
116 unsigned ubase
= MAX2(0x80, 0x20 + info
->numPatchConstants
* 0x10);
120 for (i
= 0; i
< info
->numInputs
; ++i
) {
121 offset
= nvc0_shader_input_address(info
->in
[i
].sn
,
122 info
->in
[i
].si
, ubase
);
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
)
157 unsigned ubase
= MAX2(0x80, 0x20 + info
->numPatchConstants
* 0x10);
161 for (i
= 0; i
< info
->numOutputs
; ++i
) {
162 offset
= nvc0_shader_output_address(info
->out
[i
].sn
,
163 info
->out
[i
].si
, ubase
);
165 for (c
= 0; c
< 4; ++c
)
166 info
->out
[i
].slot
[c
] = (offset
+ c
* 0x4) / 4;
173 nvc0_program_assign_varying_slots(struct nv50_ir_prog_info
*info
)
177 if (info
->type
== PIPE_SHADER_VERTEX
)
178 ret
= nvc0_vp_assign_input_slots(info
);
180 ret
= nvc0_sp_assign_input_slots(info
);
184 if (info
->type
== PIPE_SHADER_FRAGMENT
)
185 ret
= nvc0_fp_assign_output_slots(info
);
187 ret
= nvc0_sp_assign_output_slots(info
);
192 nvc0_vtgp_hdr_update_oread(struct nvc0_program
*vp
, uint8_t slot
)
194 uint8_t min
= (vp
->hdr
[4] >> 12) & 0xff;
195 uint8_t max
= (vp
->hdr
[4] >> 24);
197 min
= MIN2(min
, slot
);
198 max
= MAX2(max
, slot
);
200 vp
->hdr
[4] = (max
<< 24) | (min
<< 12);
203 /* Common part of header generation for VP, TCP, TEP and GP. */
205 nvc0_vtgp_gen_header(struct nvc0_program
*vp
, struct nv50_ir_prog_info
*info
)
209 for (i
= 0; i
< info
->numInputs
; ++i
) {
210 if (info
->in
[i
].patch
)
212 for (c
= 0; c
< 4; ++c
) {
213 a
= info
->in
[i
].slot
[c
];
214 if (info
->in
[i
].mask
& (1 << c
))
215 vp
->hdr
[5 + a
/ 32] |= 1 << (a
% 32);
219 for (i
= 0; i
< info
->numOutputs
; ++i
) {
220 if (info
->out
[i
].patch
)
222 for (c
= 0; c
< 4; ++c
) {
223 if (!(info
->out
[i
].mask
& (1 << c
)))
225 assert(info
->out
[i
].slot
[c
] >= 0x40 / 4);
226 a
= info
->out
[i
].slot
[c
] - 0x40 / 4;
227 vp
->hdr
[13 + a
/ 32] |= 1 << (a
% 32);
228 if (info
->out
[i
].oread
)
229 nvc0_vtgp_hdr_update_oread(vp
, info
->out
[i
].slot
[c
]);
233 for (i
= 0; i
< info
->numSysVals
; ++i
) {
234 switch (info
->sv
[i
].sn
) {
235 case TGSI_SEMANTIC_PRIMID
:
236 vp
->hdr
[5] |= 1 << 24;
238 case TGSI_SEMANTIC_INSTANCEID
:
239 vp
->hdr
[10] |= 1 << 30;
241 case TGSI_SEMANTIC_VERTEXID
:
242 vp
->hdr
[10] |= 1 << 31;
244 case TGSI_SEMANTIC_TESSCOORD
:
245 /* We don't have the mask, nor the slots populated. While this could
246 * be achieved, the vast majority of the time if either of the coords
247 * are read, then both will be read.
249 nvc0_vtgp_hdr_update_oread(vp
, 0x2f0 / 4);
250 nvc0_vtgp_hdr_update_oread(vp
, 0x2f4 / 4);
257 vp
->vp
.clip_enable
= info
->io
.clipDistanceMask
;
258 for (i
= 0; i
< 8; ++i
)
259 if (info
->io
.cullDistanceMask
& (1 << i
))
260 vp
->vp
.clip_mode
|= 1 << (i
* 4);
262 if (info
->io
.genUserClip
< 0)
263 vp
->vp
.num_ucps
= PIPE_MAX_CLIP_PLANES
+ 1; /* prevent rebuilding */
269 nvc0_vp_gen_header(struct nvc0_program
*vp
, struct nv50_ir_prog_info
*info
)
271 vp
->hdr
[0] = 0x20061 | (1 << 10);
272 vp
->hdr
[4] = 0xff000;
274 vp
->hdr
[18] = info
->io
.clipDistanceMask
;
276 return nvc0_vtgp_gen_header(vp
, info
);
280 nvc0_tp_get_tess_mode(struct nvc0_program
*tp
, struct nv50_ir_prog_info
*info
)
282 if (info
->prop
.tp
.outputPrim
== PIPE_PRIM_MAX
) {
283 tp
->tp
.tess_mode
= ~0;
286 switch (info
->prop
.tp
.domain
) {
287 case PIPE_PRIM_LINES
:
288 tp
->tp
.tess_mode
= NVC0_3D_TESS_MODE_PRIM_ISOLINES
;
290 case PIPE_PRIM_TRIANGLES
:
291 tp
->tp
.tess_mode
= NVC0_3D_TESS_MODE_PRIM_TRIANGLES
;
292 if (info
->prop
.tp
.winding
> 0)
293 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_CW
;
295 case PIPE_PRIM_QUADS
:
296 tp
->tp
.tess_mode
= NVC0_3D_TESS_MODE_PRIM_QUADS
;
299 tp
->tp
.tess_mode
= ~0;
302 if (info
->prop
.tp
.outputPrim
!= PIPE_PRIM_POINTS
)
303 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_CONNECTED
;
305 switch (info
->prop
.tp
.partitioning
) {
306 case PIPE_TESS_SPACING_EQUAL
:
307 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_SPACING_EQUAL
;
309 case PIPE_TESS_SPACING_FRACTIONAL_ODD
:
310 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_SPACING_FRACTIONAL_ODD
;
312 case PIPE_TESS_SPACING_FRACTIONAL_EVEN
:
313 tp
->tp
.tess_mode
|= NVC0_3D_TESS_MODE_SPACING_FRACTIONAL_EVEN
;
316 assert(!"invalid tessellator partitioning");
322 nvc0_tcp_gen_header(struct nvc0_program
*tcp
, struct nv50_ir_prog_info
*info
)
324 unsigned opcs
= 6; /* output patch constants (at least the TessFactors) */
326 tcp
->tp
.input_patch_size
= info
->prop
.tp
.inputPatchSize
;
328 if (info
->numPatchConstants
)
329 opcs
= 8 + info
->numPatchConstants
* 4;
331 tcp
->hdr
[0] = 0x20061 | (2 << 10);
333 tcp
->hdr
[1] = opcs
<< 24;
334 tcp
->hdr
[2] = info
->prop
.tp
.outputPatchSize
<< 24;
336 tcp
->hdr
[4] = 0xff000; /* initial min/max parallel output read address */
338 nvc0_vtgp_gen_header(tcp
, info
);
340 nvc0_tp_get_tess_mode(tcp
, info
);
346 nvc0_tep_gen_header(struct nvc0_program
*tep
, struct nv50_ir_prog_info
*info
)
348 tep
->tp
.input_patch_size
= ~0;
350 tep
->hdr
[0] = 0x20061 | (3 << 10);
351 tep
->hdr
[4] = 0xff000;
353 nvc0_vtgp_gen_header(tep
, info
);
355 nvc0_tp_get_tess_mode(tep
, info
);
357 tep
->hdr
[18] |= 0x3 << 12; /* ? */
363 nvc0_gp_gen_header(struct nvc0_program
*gp
, struct nv50_ir_prog_info
*info
)
365 gp
->hdr
[0] = 0x20061 | (4 << 10);
367 gp
->hdr
[2] = MIN2(info
->prop
.gp
.instanceCount
, 32) << 24;
369 switch (info
->prop
.gp
.outputPrim
) {
370 case PIPE_PRIM_POINTS
:
371 gp
->hdr
[3] = 0x01000000;
372 gp
->hdr
[0] |= 0xf0000000;
374 case PIPE_PRIM_LINE_STRIP
:
375 gp
->hdr
[3] = 0x06000000;
376 gp
->hdr
[0] |= 0x10000000;
378 case PIPE_PRIM_TRIANGLE_STRIP
:
379 gp
->hdr
[3] = 0x07000000;
380 gp
->hdr
[0] |= 0x10000000;
387 gp
->hdr
[4] = MIN2(info
->prop
.gp
.maxVertices
, 1024);
389 return nvc0_vtgp_gen_header(gp
, info
);
392 #define NVC0_INTERP_FLAT (1 << 0)
393 #define NVC0_INTERP_PERSPECTIVE (2 << 0)
394 #define NVC0_INTERP_LINEAR (3 << 0)
395 #define NVC0_INTERP_CENTROID (1 << 2)
398 nvc0_hdr_interp_mode(const struct nv50_ir_varying
*var
)
401 return NVC0_INTERP_LINEAR
;
403 return NVC0_INTERP_FLAT
;
404 return NVC0_INTERP_PERSPECTIVE
;
408 nvc0_fp_gen_header(struct nvc0_program
*fp
, struct nv50_ir_prog_info
*info
)
412 /* just 00062 on Kepler */
413 fp
->hdr
[0] = 0x20062 | (5 << 10);
414 fp
->hdr
[5] = 0x80000000; /* getting a trap if FRAG_COORD_UMASK.w = 0 */
416 if (info
->prop
.fp
.usesDiscard
)
417 fp
->hdr
[0] |= 0x8000;
418 if (info
->prop
.fp
.numColourResults
> 1)
419 fp
->hdr
[0] |= 0x4000;
420 if (info
->io
.sampleMask
< PIPE_MAX_SHADER_OUTPUTS
)
422 if (info
->prop
.fp
.writesDepth
) {
424 fp
->flags
[0] = 0x11; /* deactivate ZCULL */
427 for (i
= 0; i
< info
->numInputs
; ++i
) {
428 m
= nvc0_hdr_interp_mode(&info
->in
[i
]);
429 for (c
= 0; c
< 4; ++c
) {
430 if (!(info
->in
[i
].mask
& (1 << c
)))
432 a
= info
->in
[i
].slot
[c
];
433 if (info
->in
[i
].slot
[0] >= (0x060 / 4) &&
434 info
->in
[i
].slot
[0] <= (0x07c / 4)) {
435 fp
->hdr
[5] |= 1 << (24 + (a
- 0x060 / 4));
437 if (info
->in
[i
].slot
[0] >= (0x2c0 / 4) &&
438 info
->in
[i
].slot
[0] <= (0x2fc / 4)) {
439 fp
->hdr
[14] |= (1 << (a
- 0x280 / 4)) & 0x07ff0000;
441 if (info
->in
[i
].slot
[c
] < (0x040 / 4) ||
442 info
->in
[i
].slot
[c
] > (0x380 / 4))
445 if (info
->in
[i
].slot
[0] >= (0x300 / 4))
447 fp
->hdr
[4 + a
/ 32] |= m
<< (a
% 32);
452 for (i
= 0; i
< info
->numOutputs
; ++i
) {
453 if (info
->out
[i
].sn
== TGSI_SEMANTIC_COLOR
)
454 fp
->hdr
[18] |= info
->out
[i
].mask
<< info
->out
[i
].slot
[0];
457 fp
->fp
.early_z
= info
->prop
.fp
.earlyFragTests
;
462 static struct nvc0_transform_feedback_state
*
463 nvc0_program_create_tfb_state(const struct nv50_ir_prog_info
*info
,
464 const struct pipe_stream_output_info
*pso
)
466 struct nvc0_transform_feedback_state
*tfb
;
469 tfb
= MALLOC_STRUCT(nvc0_transform_feedback_state
);
472 for (b
= 0; b
< 4; ++b
) {
473 tfb
->stride
[b
] = pso
->stride
[b
] * 4;
474 tfb
->varying_count
[b
] = 0;
476 memset(tfb
->varying_index
, 0xff, sizeof(tfb
->varying_index
)); /* = skip */
478 for (i
= 0; i
< pso
->num_outputs
; ++i
) {
479 unsigned s
= pso
->output
[i
].start_component
;
480 unsigned p
= pso
->output
[i
].dst_offset
;
481 b
= pso
->output
[i
].output_buffer
;
483 for (c
= 0; c
< pso
->output
[i
].num_components
; ++c
)
484 tfb
->varying_index
[b
][p
++] =
485 info
->out
[pso
->output
[i
].register_index
].slot
[s
+ c
];
487 tfb
->varying_count
[b
] = MAX2(tfb
->varying_count
[b
], p
);
488 tfb
->stream
[b
] = pso
->output
[i
].stream
;
490 for (b
= 0; b
< 4; ++b
) // zero unused indices (looks nicer)
491 for (c
= tfb
->varying_count
[b
]; c
& 3; ++c
)
492 tfb
->varying_index
[b
][c
] = 0;
499 nvc0_program_dump(struct nvc0_program
*prog
)
503 if (prog
->type
!= PIPE_SHADER_COMPUTE
) {
504 for (pos
= 0; pos
< sizeof(prog
->hdr
) / sizeof(prog
->hdr
[0]); ++pos
)
505 debug_printf("HDR[%02"PRIxPTR
"] = 0x%08x\n",
506 pos
* sizeof(prog
->hdr
[0]), prog
->hdr
[pos
]);
508 debug_printf("shader binary code (0x%x bytes):", prog
->code_size
);
509 for (pos
= 0; pos
< prog
->code_size
/ 4; ++pos
) {
512 debug_printf("%08x ", prog
->code
[pos
]);
519 nvc0_program_translate(struct nvc0_program
*prog
, uint16_t chipset
)
521 struct nv50_ir_prog_info
*info
;
524 info
= CALLOC_STRUCT(nv50_ir_prog_info
);
528 info
->type
= prog
->type
;
529 info
->target
= chipset
;
530 info
->bin
.sourceRep
= NV50_PROGRAM_IR_TGSI
;
531 info
->bin
.source
= (void *)prog
->pipe
.tokens
;
533 info
->io
.genUserClip
= prog
->vp
.num_ucps
;
534 info
->io
.ucpBase
= 256;
535 info
->io
.ucpCBSlot
= 15;
536 info
->io
.sampleInterp
= prog
->fp
.sample_interp
;
538 if (prog
->type
== PIPE_SHADER_COMPUTE
) {
539 if (chipset
>= NVISA_GK104_CHIPSET
) {
540 info
->io
.resInfoCBSlot
= 0;
541 info
->io
.texBindBase
= NVE4_CP_INPUT_TEX(0);
542 info
->io
.suInfoBase
= NVE4_CP_INPUT_SUF(0);
543 info
->prop
.cp
.gridInfoBase
= NVE4_CP_INPUT_GRID_INFO(0);
545 info
->io
.msInfoCBSlot
= 0;
546 info
->io
.msInfoBase
= NVE4_CP_INPUT_MS_OFFSETS
;
548 if (chipset
>= NVISA_GK104_CHIPSET
) {
549 info
->io
.texBindBase
= 0x20;
550 info
->io
.suInfoBase
= 0; /* TODO */
552 info
->io
.resInfoCBSlot
= 15;
553 info
->io
.sampleInfoBase
= 256 + 128;
554 info
->io
.msInfoCBSlot
= 15;
555 info
->io
.msInfoBase
= 0; /* TODO */
558 info
->assignSlots
= nvc0_program_assign_varying_slots
;
561 info
->optLevel
= debug_get_num_option("NV50_PROG_OPTIMIZE", 3);
562 info
->dbgFlags
= debug_get_num_option("NV50_PROG_DEBUG", 0);
567 ret
= nv50_ir_generate_code(info
);
569 NOUVEAU_ERR("shader translation failed: %i\n", ret
);
572 if (prog
->type
!= PIPE_SHADER_COMPUTE
)
573 FREE(info
->bin
.syms
);
575 prog
->code
= info
->bin
.code
;
576 prog
->code_size
= info
->bin
.codeSize
;
577 prog
->immd_data
= info
->immd
.buf
;
578 prog
->immd_size
= info
->immd
.bufSize
;
579 prog
->relocs
= info
->bin
.relocData
;
580 prog
->num_gprs
= MAX2(4, (info
->bin
.maxGPR
+ 1));
581 prog
->num_barriers
= info
->numBarriers
;
583 prog
->vp
.need_vertex_id
= info
->io
.vertexId
< PIPE_MAX_SHADER_INPUTS
;
585 if (info
->io
.edgeFlagOut
< PIPE_MAX_ATTRIBS
)
586 info
->out
[info
->io
.edgeFlagOut
].mask
= 0; /* for headergen */
587 prog
->vp
.edgeflag
= info
->io
.edgeFlagIn
;
589 switch (prog
->type
) {
590 case PIPE_SHADER_VERTEX
:
591 ret
= nvc0_vp_gen_header(prog
, info
);
593 case PIPE_SHADER_TESS_CTRL
:
594 ret
= nvc0_tcp_gen_header(prog
, info
);
596 case PIPE_SHADER_TESS_EVAL
:
597 ret
= nvc0_tep_gen_header(prog
, info
);
599 case PIPE_SHADER_GEOMETRY
:
600 ret
= nvc0_gp_gen_header(prog
, info
);
602 case PIPE_SHADER_FRAGMENT
:
603 ret
= nvc0_fp_gen_header(prog
, info
);
605 case PIPE_SHADER_COMPUTE
:
606 prog
->cp
.syms
= info
->bin
.syms
;
607 prog
->cp
.num_syms
= info
->bin
.numSyms
;
611 NOUVEAU_ERR("unknown program type: %u\n", prog
->type
);
617 if (info
->bin
.tlsSpace
) {
618 assert(info
->bin
.tlsSpace
< (1 << 24));
619 prog
->hdr
[0] |= 1 << 26;
620 prog
->hdr
[1] |= align(info
->bin
.tlsSpace
, 0x10); /* l[] size */
621 prog
->need_tls
= true;
623 /* TODO: factor 2 only needed where joinat/precont is used,
624 * and we only have to count non-uniform branches
627 if ((info->maxCFDepth * 2) > 16) {
628 prog->hdr[2] |= (((info->maxCFDepth * 2) + 47) / 48) * 0x200;
629 prog->need_tls = true;
632 if (info
->io
.globalAccess
)
633 prog
->hdr
[0] |= 1 << 16;
635 prog
->hdr
[0] |= 1 << 27;
637 if (prog
->pipe
.stream_output
.num_outputs
)
638 prog
->tfb
= nvc0_program_create_tfb_state(info
,
639 &prog
->pipe
.stream_output
);
647 nvc0_program_upload_code(struct nvc0_context
*nvc0
, struct nvc0_program
*prog
)
649 struct nvc0_screen
*screen
= nvc0
->screen
;
650 const bool is_cp
= prog
->type
== PIPE_SHADER_COMPUTE
;
652 uint32_t size
= prog
->code_size
+ (is_cp
? 0 : NVC0_SHADER_HEADER_SIZE
);
653 uint32_t lib_pos
= screen
->lib_code
->start
;
656 /* c[] bindings need to be aligned to 0x100, but we could use relocations
658 if (prog
->immd_size
) {
659 prog
->immd_base
= size
;
660 size
= align(size
, 0x40);
661 size
+= prog
->immd_size
+ 0xc0; /* add 0xc0 for align 0x40 -> 0x100 */
663 /* On Fermi, SP_START_ID must be aligned to 0x40.
664 * On Kepler, the first instruction must be aligned to 0x80 because
665 * latency information is expected only at certain positions.
667 if (screen
->base
.class_3d
>= NVE4_3D_CLASS
)
668 size
= size
+ (is_cp
? 0x40 : 0x70);
669 size
= align(size
, 0x40);
671 ret
= nouveau_heap_alloc(screen
->text_heap
, size
, prog
, &prog
->mem
);
673 struct nouveau_heap
*heap
= screen
->text_heap
;
674 /* Note that the code library, which is allocated before anything else,
675 * does not have a priv pointer. We can stop once we hit it.
677 while (heap
->next
&& heap
->next
->priv
) {
678 struct nvc0_program
*evict
= heap
->next
->priv
;
679 nouveau_heap_free(&evict
->mem
);
681 debug_printf("WARNING: out of code space, evicting all shaders.\n");
682 ret
= nouveau_heap_alloc(heap
, size
, prog
, &prog
->mem
);
684 NOUVEAU_ERR("shader too large (0x%x) to fit in code space ?\n", size
);
687 IMMED_NVC0(nvc0
->base
.pushbuf
, NVC0_3D(SERIALIZE
), 0);
689 prog
->code_base
= prog
->mem
->start
;
690 prog
->immd_base
= align(prog
->mem
->start
+ prog
->immd_base
, 0x100);
691 assert((prog
->immd_size
== 0) || (prog
->immd_base
+ prog
->immd_size
<=
692 prog
->mem
->start
+ prog
->mem
->size
));
695 if (screen
->base
.class_3d
>= NVE4_3D_CLASS
) {
696 switch (prog
->mem
->start
& 0xff) {
697 case 0x40: prog
->code_base
+= 0x70; break;
698 case 0x80: prog
->code_base
+= 0x30; break;
699 case 0xc0: prog
->code_base
+= 0x70; break;
701 prog
->code_base
+= 0x30;
702 assert((prog
->mem
->start
& 0xff) == 0x00);
706 code_pos
= prog
->code_base
+ NVC0_SHADER_HEADER_SIZE
;
708 if (screen
->base
.class_3d
>= NVE4_3D_CLASS
) {
709 if (prog
->mem
->start
& 0x40)
710 prog
->code_base
+= 0x40;
711 assert((prog
->code_base
& 0x7f) == 0x00);
713 code_pos
= prog
->code_base
;
717 nv50_ir_relocate_code(prog
->relocs
, prog
->code
, code_pos
, lib_pos
, 0);
720 if (debug_get_bool_option("NV50_PROG_DEBUG", false))
721 nvc0_program_dump(prog
);
725 nvc0
->base
.push_data(&nvc0
->base
, screen
->text
, prog
->code_base
,
726 NV_VRAM_DOMAIN(&screen
->base
), NVC0_SHADER_HEADER_SIZE
, prog
->hdr
);
727 nvc0
->base
.push_data(&nvc0
->base
, screen
->text
, code_pos
,
728 NV_VRAM_DOMAIN(&screen
->base
), prog
->code_size
, prog
->code
);
730 nvc0
->base
.push_data(&nvc0
->base
,
731 screen
->text
, prog
->immd_base
, NV_VRAM_DOMAIN(&screen
->base
),
732 prog
->immd_size
, prog
->immd_data
);
734 BEGIN_NVC0(nvc0
->base
.pushbuf
, NVC0_3D(MEM_BARRIER
), 1);
735 PUSH_DATA (nvc0
->base
.pushbuf
, 0x1011);
740 /* Upload code for builtin functions like integer division emulation. */
742 nvc0_program_library_upload(struct nvc0_context
*nvc0
)
744 struct nvc0_screen
*screen
= nvc0
->screen
;
747 const uint32_t *code
;
749 if (screen
->lib_code
)
752 nv50_ir_get_target_library(screen
->base
.device
->chipset
, &code
, &size
);
756 ret
= nouveau_heap_alloc(screen
->text_heap
, align(size
, 0x100), NULL
,
761 nvc0
->base
.push_data(&nvc0
->base
,
762 screen
->text
, screen
->lib_code
->start
, NV_VRAM_DOMAIN(&screen
->base
),
764 /* no need for a memory barrier, will be emitted with first program */
768 nvc0_program_destroy(struct nvc0_context
*nvc0
, struct nvc0_program
*prog
)
770 const struct pipe_shader_state pipe
= prog
->pipe
;
771 const ubyte type
= prog
->type
;
774 nouveau_heap_free(&prog
->mem
);
775 FREE(prog
->code
); /* may be 0 for hardcoded shaders */
776 FREE(prog
->immd_data
);
778 if (prog
->type
== PIPE_SHADER_COMPUTE
&& prog
->cp
.syms
)
781 if (nvc0
->state
.tfb
== prog
->tfb
)
782 nvc0
->state
.tfb
= NULL
;
786 memset(prog
, 0, sizeof(*prog
));
793 nvc0_program_symbol_offset(const struct nvc0_program
*prog
, uint32_t label
)
795 const struct nv50_ir_prog_symbol
*syms
=
796 (const struct nv50_ir_prog_symbol
*)prog
->cp
.syms
;
799 if (prog
->type
!= PIPE_SHADER_COMPUTE
)
800 base
= NVC0_SHADER_HEADER_SIZE
;
801 for (i
= 0; i
< prog
->cp
.num_syms
; ++i
)
802 if (syms
[i
].label
== label
)
803 return prog
->code_base
+ base
+ syms
[i
].offset
;
804 return prog
->code_base
; /* no symbols or symbol not found */