1 /*
2 * Copyright 2010 Christoph Bumiller
3 *
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:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
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.
21 */
22
23 #include "pipe/p_defines.h"
24
25 #include "tgsi/tgsi_ureg.h"
26
27 #include "nvc0/nvc0_context.h"
28
29 #include "codegen/nv50_ir_driver.h"
30 #include "nvc0/nve4_compute.h"
31
32 /* NOTE: Using a[0x270] in FP may cause an error even if we're using less than
33 * 124 scalar varying values.
34 */
35 static uint32_t
nvc0_shader_input_address(unsigned sn,unsigned si)36 nvc0_shader_input_address(unsigned sn, unsigned si)
37 {
38 switch (sn) {
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 default:
59 assert(!"invalid TGSI input semantic");
60 return ~0;
61 }
62 }
63
64 static uint32_t
nvc0_shader_output_address(unsigned sn,unsigned si)65 nvc0_shader_output_address(unsigned sn, unsigned si)
66 {
67 switch (sn) {
68 case TGSI_SEMANTIC_TESSOUTER: return 0x000 + si * 0x4;
69 case TGSI_SEMANTIC_TESSINNER: return 0x010 + si * 0x4;
70 case TGSI_SEMANTIC_PATCH: return 0x020 + si * 0x10;
71 case TGSI_SEMANTIC_PRIMID: return 0x060;
72 case TGSI_SEMANTIC_LAYER: return 0x064;
73 case TGSI_SEMANTIC_VIEWPORT_INDEX:return 0x068;
74 case TGSI_SEMANTIC_PSIZE: return 0x06c;
75 case TGSI_SEMANTIC_POSITION: return 0x070;
76 case TGSI_SEMANTIC_GENERIC: return 0x080 + si * 0x10;
77 case TGSI_SEMANTIC_FOG: return 0x2e8;
78 case TGSI_SEMANTIC_COLOR: return 0x280 + si * 0x10;
79 case TGSI_SEMANTIC_BCOLOR: return 0x2a0 + si * 0x10;
80 case TGSI_SEMANTIC_CLIPDIST: return 0x2c0 + si * 0x10;
81 case TGSI_SEMANTIC_CLIPVERTEX: return 0x270;
82 case TGSI_SEMANTIC_TEXCOORD: return 0x300 + si * 0x10;
83 /* case TGSI_SEMANTIC_VIEWPORT_MASK: return 0x3a0; */
84 case TGSI_SEMANTIC_EDGEFLAG: return ~0;
85 default:
86 assert(!"invalid TGSI output semantic");
87 return ~0;
88 }
89 }
90
91 static int
nvc0_vp_assign_input_slots(struct nv50_ir_prog_info * info)92 nvc0_vp_assign_input_slots(struct nv50_ir_prog_info *info)
93 {
94 unsigned i, c, n;
95
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;
103 continue;
104 default:
105 break;
106 }
107 for (c = 0; c < 4; ++c)
108 info->in[i].slot[c] = (0x80 + n * 0x10 + c * 0x4) / 4;
109 ++n;
110 }
111
112 return 0;
113 }
114
115 static int
nvc0_sp_assign_input_slots(struct nv50_ir_prog_info * info)116 nvc0_sp_assign_input_slots(struct nv50_ir_prog_info *info)
117 {
118 unsigned offset;
119 unsigned i, c;
120
121 for (i = 0; i < info->numInputs; ++i) {
122 offset = nvc0_shader_input_address(info->in[i].sn, info->in[i].si);
123
124 for (c = 0; c < 4; ++c)
125 info->in[i].slot[c] = (offset + c * 0x4) / 4;
126 }
127
128 return 0;
129 }
130
131 static int
nvc0_fp_assign_output_slots(struct nv50_ir_prog_info * info)132 nvc0_fp_assign_output_slots(struct nv50_ir_prog_info *info)
133 {
134 unsigned count = info->prop.fp.numColourResults * 4;
135 unsigned i, c;
136
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;
141
142 if (info->io.sampleMask < PIPE_MAX_SHADER_OUTPUTS)
143 info->out[info->io.sampleMask].slot[0] = count++;
144 else
145 if (info->target >= 0xe0)
146 count++; /* on Kepler, depth is always last colour reg + 2 */
147
148 if (info->io.fragDepth < PIPE_MAX_SHADER_OUTPUTS)
149 info->out[info->io.fragDepth].slot[2] = count;
150
151 return 0;
152 }
153
154 static int
nvc0_sp_assign_output_slots(struct nv50_ir_prog_info * info)155 nvc0_sp_assign_output_slots(struct nv50_ir_prog_info *info)
156 {
157 unsigned offset;
158 unsigned i, c;
159
160 for (i = 0; i < info->numOutputs; ++i) {
161 offset = nvc0_shader_output_address(info->out[i].sn, info->out[i].si);
162
163 for (c = 0; c < 4; ++c)
164 info->out[i].slot[c] = (offset + c * 0x4) / 4;
165 }
166
167 return 0;
168 }
169
170 static int
nvc0_program_assign_varying_slots(struct nv50_ir_prog_info * info)171 nvc0_program_assign_varying_slots(struct nv50_ir_prog_info *info)
172 {
173 int ret;
174
175 if (info->type == PIPE_SHADER_VERTEX)
176 ret = nvc0_vp_assign_input_slots(info);
177 else
178 ret = nvc0_sp_assign_input_slots(info);
179 if (ret)
180 return ret;
181
182 if (info->type == PIPE_SHADER_FRAGMENT)
183 ret = nvc0_fp_assign_output_slots(info);
184 else
185 ret = nvc0_sp_assign_output_slots(info);
186 return ret;
187 }
188
189 static inline void
nvc0_vtgp_hdr_update_oread(struct nvc0_program * vp,uint8_t slot)190 nvc0_vtgp_hdr_update_oread(struct nvc0_program *vp, uint8_t slot)
191 {
192 uint8_t min = (vp->hdr[4] >> 12) & 0xff;
193 uint8_t max = (vp->hdr[4] >> 24);
194
195 min = MIN2(min, slot);
196 max = MAX2(max, slot);
197
198 vp->hdr[4] = (max << 24) | (min << 12);
199 }
200
201 /* Common part of header generation for VP, TCP, TEP and GP. */
202 static int
nvc0_vtgp_gen_header(struct nvc0_program * vp,struct nv50_ir_prog_info * info)203 nvc0_vtgp_gen_header(struct nvc0_program *vp, struct nv50_ir_prog_info *info)
204 {
205 unsigned i, c, a;
206
207 for (i = 0; i < info->numInputs; ++i) {
208 if (info->in[i].patch)
209 continue;
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);
214 }
215 }
216
217 for (i = 0; i < info->numOutputs; ++i) {
218 if (info->out[i].patch)
219 continue;
220 for (c = 0; c < 4; ++c) {
221 if (!(info->out[i].mask & (1 << c)))
222 continue;
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]);
228 }
229 }
230
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;
235 break;
236 case TGSI_SEMANTIC_INSTANCEID:
237 vp->hdr[10] |= 1 << 30;
238 break;
239 case TGSI_SEMANTIC_VERTEXID:
240 vp->hdr[10] |= 1 << 31;
241 break;
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.
246 */
247 nvc0_vtgp_hdr_update_oread(vp, 0x2f0 / 4);
248 nvc0_vtgp_hdr_update_oread(vp, 0x2f4 / 4);
249 break;
250 default:
251 break;
252 }
253 }
254
255 vp->vp.clip_enable = (1 << info->io.clipDistances) - 1;
256 vp->vp.cull_enable =
257 ((1 << info->io.cullDistances) - 1) << info->io.clipDistances;
258 for (i = 0; i < info->io.cullDistances; ++i)
259 vp->vp.clip_mode |= 1 << ((info->io.clipDistances + i) * 4);
260
261 if (info->io.genUserClip < 0)
262 vp->vp.num_ucps = PIPE_MAX_CLIP_PLANES + 1; /* prevent rebuilding */
263
264 return 0;
265 }
266
267 static int
nvc0_vp_gen_header(struct nvc0_program * vp,struct nv50_ir_prog_info * info)268 nvc0_vp_gen_header(struct nvc0_program *vp, struct nv50_ir_prog_info *info)
269 {
270 vp->hdr[0] = 0x20061 | (1 << 10);
271 vp->hdr[4] = 0xff000;
272
273 return nvc0_vtgp_gen_header(vp, info);
274 }
275
276 static void
nvc0_tp_get_tess_mode(struct nvc0_program * tp,struct nv50_ir_prog_info * info)277 nvc0_tp_get_tess_mode(struct nvc0_program *tp, struct nv50_ir_prog_info *info)
278 {
279 if (info->prop.tp.outputPrim == PIPE_PRIM_MAX) {
280 tp->tp.tess_mode = ~0;
281 return;
282 }
283 switch (info->prop.tp.domain) {
284 case PIPE_PRIM_LINES:
285 tp->tp.tess_mode = NVC0_3D_TESS_MODE_PRIM_ISOLINES;
286 break;
287 case PIPE_PRIM_TRIANGLES:
288 tp->tp.tess_mode = NVC0_3D_TESS_MODE_PRIM_TRIANGLES;
289 break;
290 case PIPE_PRIM_QUADS:
291 tp->tp.tess_mode = NVC0_3D_TESS_MODE_PRIM_QUADS;
292 break;
293 default:
294 tp->tp.tess_mode = ~0;
295 return;
296 }
297
298 /* It seems like lines want the "CW" bit to indicate they're connected, and
299 * spit out errors in dmesg when the "CONNECTED" bit is set.
300 */
301 if (info->prop.tp.outputPrim != PIPE_PRIM_POINTS) {
302 if (info->prop.tp.domain == PIPE_PRIM_LINES)
303 tp->tp.tess_mode |= NVC0_3D_TESS_MODE_CW;
304 else
305 tp->tp.tess_mode |= NVC0_3D_TESS_MODE_CONNECTED;
306 }
307
308 /* Winding only matters for triangles/quads, not lines. */
309 if (info->prop.tp.domain != PIPE_PRIM_LINES &&
310 info->prop.tp.outputPrim != PIPE_PRIM_POINTS &&
311 info->prop.tp.winding > 0)
312 tp->tp.tess_mode |= NVC0_3D_TESS_MODE_CW;
313
314 switch (info->prop.tp.partitioning) {
315 case PIPE_TESS_SPACING_EQUAL:
316 tp->tp.tess_mode |= NVC0_3D_TESS_MODE_SPACING_EQUAL;
317 break;
318 case PIPE_TESS_SPACING_FRACTIONAL_ODD:
319 tp->tp.tess_mode |= NVC0_3D_TESS_MODE_SPACING_FRACTIONAL_ODD;
320 break;
321 case PIPE_TESS_SPACING_FRACTIONAL_EVEN:
322 tp->tp.tess_mode |= NVC0_3D_TESS_MODE_SPACING_FRACTIONAL_EVEN;
323 break;
324 default:
325 assert(!"invalid tessellator partitioning");
326 break;
327 }
328 }
329
330 static int
nvc0_tcp_gen_header(struct nvc0_program * tcp,struct nv50_ir_prog_info * info)331 nvc0_tcp_gen_header(struct nvc0_program *tcp, struct nv50_ir_prog_info *info)
332 {
333 unsigned opcs = 6; /* output patch constants (at least the TessFactors) */
334
335 tcp->tp.input_patch_size = info->prop.tp.inputPatchSize;
336
337 if (info->numPatchConstants)
338 opcs = 8 + info->numPatchConstants * 4;
339
340 tcp->hdr[0] = 0x20061 | (2 << 10);
341
342 tcp->hdr[1] = opcs << 24;
343 tcp->hdr[2] = info->prop.tp.outputPatchSize << 24;
344
345 tcp->hdr[4] = 0xff000; /* initial min/max parallel output read address */
346
347 nvc0_vtgp_gen_header(tcp, info);
348
349 if (info->target >= NVISA_GM107_CHIPSET) {
350 /* On GM107+, the number of output patch components has moved in the TCP
351 * header, but it seems like blob still also uses the old position.
352 * Also, the high 8-bits are located inbetween the min/max parallel
353 * field and has to be set after updating the outputs. */
354 tcp->hdr[3] = (opcs & 0x0f) << 28;
355 tcp->hdr[4] |= (opcs & 0xf0) << 16;
356 }
357
358 nvc0_tp_get_tess_mode(tcp, info);
359
360 return 0;
361 }
362
363 static int
nvc0_tep_gen_header(struct nvc0_program * tep,struct nv50_ir_prog_info * info)364 nvc0_tep_gen_header(struct nvc0_program *tep, struct nv50_ir_prog_info *info)
365 {
366 tep->tp.input_patch_size = ~0;
367
368 tep->hdr[0] = 0x20061 | (3 << 10);
369 tep->hdr[4] = 0xff000;
370
371 nvc0_vtgp_gen_header(tep, info);
372
373 nvc0_tp_get_tess_mode(tep, info);
374
375 tep->hdr[18] |= 0x3 << 12; /* ? */
376
377 return 0;
378 }
379
380 static int
nvc0_gp_gen_header(struct nvc0_program * gp,struct nv50_ir_prog_info * info)381 nvc0_gp_gen_header(struct nvc0_program *gp, struct nv50_ir_prog_info *info)
382 {
383 gp->hdr[0] = 0x20061 | (4 << 10);
384
385 gp->hdr[2] = MIN2(info->prop.gp.instanceCount, 32) << 24;
386
387 switch (info->prop.gp.outputPrim) {
388 case PIPE_PRIM_POINTS:
389 gp->hdr[3] = 0x01000000;
390 gp->hdr[0] |= 0xf0000000;
391 break;
392 case PIPE_PRIM_LINE_STRIP:
393 gp->hdr[3] = 0x06000000;
394 gp->hdr[0] |= 0x10000000;
395 break;
396 case PIPE_PRIM_TRIANGLE_STRIP:
397 gp->hdr[3] = 0x07000000;
398 gp->hdr[0] |= 0x10000000;
399 break;
400 default:
401 assert(0);
402 break;
403 }
404
405 gp->hdr[4] = CLAMP(info->prop.gp.maxVertices, 1, 1024);
406
407 return nvc0_vtgp_gen_header(gp, info);
408 }
409
410 #define NVC0_INTERP_FLAT (1 << 0)
411 #define NVC0_INTERP_PERSPECTIVE (2 << 0)
412 #define NVC0_INTERP_LINEAR (3 << 0)
413 #define NVC0_INTERP_CENTROID (1 << 2)
414
415 static uint8_t
nvc0_hdr_interp_mode(const struct nv50_ir_varying * var)416 nvc0_hdr_interp_mode(const struct nv50_ir_varying *var)
417 {
418 if (var->linear)
419 return NVC0_INTERP_LINEAR;
420 if (var->flat)
421 return NVC0_INTERP_FLAT;
422 return NVC0_INTERP_PERSPECTIVE;
423 }
424
425 static int
nvc0_fp_gen_header(struct nvc0_program * fp,struct nv50_ir_prog_info * info)426 nvc0_fp_gen_header(struct nvc0_program *fp, struct nv50_ir_prog_info *info)
427 {
428 unsigned i, c, a, m;
429
430 /* just 00062 on Kepler */
431 fp->hdr[0] = 0x20062 | (5 << 10);
432 fp->hdr[5] = 0x80000000; /* getting a trap if FRAG_COORD_UMASK.w = 0 */
433
434 if (info->prop.fp.usesDiscard)
435 fp->hdr[0] |= 0x8000;
436 if (info->prop.fp.numColourResults > 1)
437 fp->hdr[0] |= 0x4000;
438 if (info->io.sampleMask < PIPE_MAX_SHADER_OUTPUTS)
439 fp->hdr[19] |= 0x1;
440 if (info->prop.fp.writesDepth) {
441 fp->hdr[19] |= 0x2;
442 fp->flags[0] = 0x11; /* deactivate ZCULL */
443 }
444
445 for (i = 0; i < info->numInputs; ++i) {
446 m = nvc0_hdr_interp_mode(&info->in[i]);
447 if (info->in[i].sn == TGSI_SEMANTIC_COLOR) {
448 fp->fp.colors |= 1 << info->in[i].si;
449 if (info->in[i].sc)
450 fp->fp.color_interp[info->in[i].si] = m | (info->in[i].mask << 4);
451 }
452 for (c = 0; c < 4; ++c) {
453 if (!(info->in[i].mask & (1 << c)))
454 continue;
455 a = info->in[i].slot[c];
456 if (info->in[i].slot[0] >= (0x060 / 4) &&
457 info->in[i].slot[0] <= (0x07c / 4)) {
458 fp->hdr[5] |= 1 << (24 + (a - 0x060 / 4));
459 } else
460 if (info->in[i].slot[0] >= (0x2c0 / 4) &&
461 info->in[i].slot[0] <= (0x2fc / 4)) {
462 fp->hdr[14] |= (1 << (a - 0x280 / 4)) & 0x07ff0000;
463 } else {
464 if (info->in[i].slot[c] < (0x040 / 4) ||
465 info->in[i].slot[c] > (0x380 / 4))
466 continue;
467 a *= 2;
468 if (info->in[i].slot[0] >= (0x300 / 4))
469 a -= 32;
470 fp->hdr[4 + a / 32] |= m << (a % 32);
471 }
472 }
473 }
474
475 for (i = 0; i < info->numOutputs; ++i) {
476 if (info->out[i].sn == TGSI_SEMANTIC_COLOR)
477 fp->hdr[18] |= 0xf << info->out[i].slot[0];
478 }
479
480 /* There are no "regular" attachments, but the shader still needs to be
481 * executed. It seems like it wants to think that it has some color
482 * outputs in order to actually run.
483 */
484 if (info->prop.fp.numColourResults == 0 && !info->prop.fp.writesDepth)
485 fp->hdr[18] |= 0xf;
486
487 fp->fp.early_z = info->prop.fp.earlyFragTests;
488 fp->fp.sample_mask_in = info->prop.fp.usesSampleMaskIn;
489 fp->fp.reads_framebuffer = info->prop.fp.readsFramebuffer;
490 fp->fp.post_depth_coverage = info->prop.fp.postDepthCoverage;
491
492 /* Mark position xy and layer as read */
493 if (fp->fp.reads_framebuffer)
494 fp->hdr[5] |= 0x32000000;
495
496 return 0;
497 }
498
499 static struct nvc0_transform_feedback_state *
nvc0_program_create_tfb_state(const struct nv50_ir_prog_info * info,const struct pipe_stream_output_info * pso)500 nvc0_program_create_tfb_state(const struct nv50_ir_prog_info *info,
501 const struct pipe_stream_output_info *pso)
502 {
503 struct nvc0_transform_feedback_state *tfb;
504 unsigned b, i, c;
505
506 tfb = MALLOC_STRUCT(nvc0_transform_feedback_state);
507 if (!tfb)
508 return NULL;
509 for (b = 0; b < 4; ++b) {
510 tfb->stride[b] = pso->stride[b] * 4;
511 tfb->varying_count[b] = 0;
512 }
513 memset(tfb->varying_index, 0xff, sizeof(tfb->varying_index)); /* = skip */
514
515 for (i = 0; i < pso->num_outputs; ++i) {
516 unsigned s = pso->output[i].start_component;
517 unsigned p = pso->output[i].dst_offset;
518 const unsigned r = pso->output[i].register_index;
519 b = pso->output[i].output_buffer;
520
521 if (r >= info->numOutputs)
522 continue;
523
524 for (c = 0; c < pso->output[i].num_components; ++c)
525 tfb->varying_index[b][p++] = info->out[r].slot[s + c];
526
527 tfb->varying_count[b] = MAX2(tfb->varying_count[b], p);
528 tfb->stream[b] = pso->output[i].stream;
529 }
530 for (b = 0; b < 4; ++b) // zero unused indices (looks nicer)
531 for (c = tfb->varying_count[b]; c & 3; ++c)
532 tfb->varying_index[b][c] = 0;
533
534 return tfb;
535 }
536
537 #ifdef DEBUG
538 static void
nvc0_program_dump(struct nvc0_program * prog)539 nvc0_program_dump(struct nvc0_program *prog)
540 {
541 unsigned pos;
542
543 if (prog->type != PIPE_SHADER_COMPUTE) {
544 for (pos = 0; pos < ARRAY_SIZE(prog->hdr); ++pos)
545 debug_printf("HDR[%02"PRIxPTR"] = 0x%08x\n",
546 pos * sizeof(prog->hdr[0]), prog->hdr[pos]);
547 }
548 debug_printf("shader binary code (0x%x bytes):", prog->code_size);
549 for (pos = 0; pos < prog->code_size / 4; ++pos) {
550 if ((pos % 8) == 0)
551 debug_printf("\n");
552 debug_printf("%08x ", prog->code[pos]);
553 }
554 debug_printf("\n");
555 }
556 #endif
557
558 bool
nvc0_program_translate(struct nvc0_program * prog,uint16_t chipset,struct pipe_debug_callback * debug)559 nvc0_program_translate(struct nvc0_program *prog, uint16_t chipset,
560 struct pipe_debug_callback *debug)
561 {
562 struct nv50_ir_prog_info *info;
563 int ret;
564
565 info = CALLOC_STRUCT(nv50_ir_prog_info);
566 if (!info)
567 return false;
568
569 info->type = prog->type;
570 info->target = chipset;
571 info->bin.sourceRep = PIPE_SHADER_IR_TGSI;
572 info->bin.source = (void *)prog->pipe.tokens;
573
574 #ifdef DEBUG
575 info->target = debug_get_num_option("NV50_PROG_CHIPSET", chipset);
576 info->optLevel = debug_get_num_option("NV50_PROG_OPTIMIZE", 3);
577 info->dbgFlags = debug_get_num_option("NV50_PROG_DEBUG", 0);
578 info->omitLineNum = debug_get_num_option("NV50_PROG_DEBUG_OMIT_LINENUM", 0);
579 #else
580 info->optLevel = 3;
581 #endif
582
583 info->bin.smemSize = prog->cp.smem_size;
584 info->io.genUserClip = prog->vp.num_ucps;
585 info->io.auxCBSlot = 15;
586 info->io.msInfoCBSlot = 15;
587 info->io.ucpBase = NVC0_CB_AUX_UCP_INFO;
588 info->io.drawInfoBase = NVC0_CB_AUX_DRAW_INFO;
589 info->io.msInfoBase = NVC0_CB_AUX_MS_INFO;
590 info->io.bufInfoBase = NVC0_CB_AUX_BUF_INFO(0);
591 info->io.suInfoBase = NVC0_CB_AUX_SU_INFO(0);
592 if (info->target >= NVISA_GK104_CHIPSET) {
593 info->io.texBindBase = NVC0_CB_AUX_TEX_INFO(0);
594 info->io.fbtexBindBase = NVC0_CB_AUX_FB_TEX_INFO;
595 info->io.bindlessBase = NVC0_CB_AUX_BINDLESS_INFO(0);
596 }
597
598 if (prog->type == PIPE_SHADER_COMPUTE) {
599 if (info->target >= NVISA_GK104_CHIPSET) {
600 info->io.auxCBSlot = 7;
601 info->io.msInfoCBSlot = 7;
602 info->io.uboInfoBase = NVC0_CB_AUX_UBO_INFO(0);
603 }
604 info->prop.cp.gridInfoBase = NVC0_CB_AUX_GRID_INFO(0);
605 } else {
606 info->io.sampleInfoBase = NVC0_CB_AUX_SAMPLE_INFO;
607 }
608
609 info->assignSlots = nvc0_program_assign_varying_slots;
610
611 ret = nv50_ir_generate_code(info);
612 if (ret) {
613 NOUVEAU_ERR("shader translation failed: %i\n", ret);
614 goto out;
615 }
616 if (prog->type != PIPE_SHADER_COMPUTE)
617 FREE(info->bin.syms);
618
619 prog->code = info->bin.code;
620 prog->code_size = info->bin.codeSize;
621 prog->relocs = info->bin.relocData;
622 prog->fixups = info->bin.fixupData;
623 prog->num_gprs = MAX2(4, (info->bin.maxGPR + 1));
624 prog->cp.smem_size = info->bin.smemSize;
625 prog->num_barriers = info->numBarriers;
626
627 prog->vp.need_vertex_id = info->io.vertexId < PIPE_MAX_SHADER_INPUTS;
628 prog->vp.need_draw_parameters = info->prop.vp.usesDrawParameters;
629
630 if (info->io.edgeFlagOut < PIPE_MAX_ATTRIBS)
631 info->out[info->io.edgeFlagOut].mask = 0; /* for headergen */
632 prog->vp.edgeflag = info->io.edgeFlagIn;
633
634 switch (prog->type) {
635 case PIPE_SHADER_VERTEX:
636 ret = nvc0_vp_gen_header(prog, info);
637 break;
638 case PIPE_SHADER_TESS_CTRL:
639 ret = nvc0_tcp_gen_header(prog, info);
640 break;
641 case PIPE_SHADER_TESS_EVAL:
642 ret = nvc0_tep_gen_header(prog, info);
643 break;
644 case PIPE_SHADER_GEOMETRY:
645 ret = nvc0_gp_gen_header(prog, info);
646 break;
647 case PIPE_SHADER_FRAGMENT:
648 ret = nvc0_fp_gen_header(prog, info);
649 break;
650 case PIPE_SHADER_COMPUTE:
651 prog->cp.syms = info->bin.syms;
652 prog->cp.num_syms = info->bin.numSyms;
653 break;
654 default:
655 ret = -1;
656 NOUVEAU_ERR("unknown program type: %u\n", prog->type);
657 break;
658 }
659 if (ret)
660 goto out;
661
662 if (info->bin.tlsSpace) {
663 assert(info->bin.tlsSpace < (1 << 24));
664 prog->hdr[0] |= 1 << 26;
665 prog->hdr[1] |= align(info->bin.tlsSpace, 0x10); /* l[] size */
666 prog->need_tls = true;
667 }
668 /* TODO: factor 2 only needed where joinat/precont is used,
669 * and we only have to count non-uniform branches
670 */
671 /*
672 if ((info->maxCFDepth * 2) > 16) {
673 prog->hdr[2] |= (((info->maxCFDepth * 2) + 47) / 48) * 0x200;
674 prog->need_tls = true;
675 }
676 */
677 if (info->io.globalAccess)
678 prog->hdr[0] |= 1 << 26;
679 if (info->io.globalAccess & 0x2)
680 prog->hdr[0] |= 1 << 16;
681 if (info->io.fp64)
682 prog->hdr[0] |= 1 << 27;
683
684 if (prog->pipe.stream_output.num_outputs)
685 prog->tfb = nvc0_program_create_tfb_state(info,
686 &prog->pipe.stream_output);
687
688 pipe_debug_message(debug, SHADER_INFO,
689 "type: %d, local: %d, shared: %d, gpr: %d, inst: %d, bytes: %d",
690 prog->type, info->bin.tlsSpace, info->bin.smemSize,
691 prog->num_gprs, info->bin.instructions,
692 info->bin.codeSize);
693
694 #ifdef DEBUG
695 if (debug_get_option("NV50_PROG_CHIPSET", NULL) && info->dbgFlags)
696 nvc0_program_dump(prog);
697 #endif
698
699 out:
700 FREE(info);
701 return !ret;
702 }
703
704 static inline int
nvc0_program_alloc_code(struct nvc0_context * nvc0,struct nvc0_program * prog)705 nvc0_program_alloc_code(struct nvc0_context *nvc0, struct nvc0_program *prog)
706 {
707 struct nvc0_screen *screen = nvc0->screen;
708 const bool is_cp = prog->type == PIPE_SHADER_COMPUTE;
709 int ret;
710 uint32_t size = prog->code_size + (is_cp ? 0 : NVC0_SHADER_HEADER_SIZE);
711
712 /* On Fermi, SP_START_ID must be aligned to 0x40.
713 * On Kepler, the first instruction must be aligned to 0x80 because
714 * latency information is expected only at certain positions.
715 */
716 if (screen->base.class_3d >= NVE4_3D_CLASS)
717 size = size + (is_cp ? 0x40 : 0x70);
718 size = align(size, 0x40);
719
720 ret = nouveau_heap_alloc(screen->text_heap, size, prog, &prog->mem);
721 if (ret)
722 return ret;
723 prog->code_base = prog->mem->start;
724
725 if (!is_cp) {
726 if (screen->base.class_3d >= NVE4_3D_CLASS) {
727 switch (prog->mem->start & 0xff) {
728 case 0x40: prog->code_base += 0x70; break;
729 case 0x80: prog->code_base += 0x30; break;
730 case 0xc0: prog->code_base += 0x70; break;
731 default:
732 prog->code_base += 0x30;
733 assert((prog->mem->start & 0xff) == 0x00);
734 break;
735 }
736 }
737 } else {
738 if (screen->base.class_3d >= NVE4_3D_CLASS) {
739 if (prog->mem->start & 0x40)
740 prog->code_base += 0x40;
741 assert((prog->code_base & 0x7f) == 0x00);
742 }
743 }
744
745 return 0;
746 }
747
748 static inline void
nvc0_program_upload_code(struct nvc0_context * nvc0,struct nvc0_program * prog)749 nvc0_program_upload_code(struct nvc0_context *nvc0, struct nvc0_program *prog)
750 {
751 struct nvc0_screen *screen = nvc0->screen;
752 const bool is_cp = prog->type == PIPE_SHADER_COMPUTE;
753 uint32_t code_pos = prog->code_base + (is_cp ? 0 : NVC0_SHADER_HEADER_SIZE);
754
755 if (prog->relocs)
756 nv50_ir_relocate_code(prog->relocs, prog->code, code_pos,
757 screen->lib_code->start, 0);
758 if (prog->fixups) {
759 nv50_ir_apply_fixups(prog->fixups, prog->code,
760 prog->fp.force_persample_interp,
761 prog->fp.flatshade,
762 0 /* alphatest */);
763 for (int i = 0; i < 2; i++) {
764 unsigned mask = prog->fp.color_interp[i] >> 4;
765 unsigned interp = prog->fp.color_interp[i] & 3;
766 if (!mask)
767 continue;
768 prog->hdr[14] &= ~(0xff << (8 * i));
769 if (prog->fp.flatshade)
770 interp = NVC0_INTERP_FLAT;
771 for (int c = 0; c < 4; c++)
772 if (mask & (1 << c))
773 prog->hdr[14] |= interp << (2 * (4 * i + c));
774 }
775 }
776
777 if (!is_cp)
778 nvc0->base.push_data(&nvc0->base, screen->text, prog->code_base,
779 NV_VRAM_DOMAIN(&screen->base),
780 NVC0_SHADER_HEADER_SIZE, prog->hdr);
781
782 nvc0->base.push_data(&nvc0->base, screen->text, code_pos,
783 NV_VRAM_DOMAIN(&screen->base), prog->code_size,
784 prog->code);
785 }
786
787 bool
nvc0_program_upload(struct nvc0_context * nvc0,struct nvc0_program * prog)788 nvc0_program_upload(struct nvc0_context *nvc0, struct nvc0_program *prog)
789 {
790 struct nvc0_screen *screen = nvc0->screen;
791 const bool is_cp = prog->type == PIPE_SHADER_COMPUTE;
792 int ret;
793 uint32_t size = prog->code_size + (is_cp ? 0 : NVC0_SHADER_HEADER_SIZE);
794
795 ret = nvc0_program_alloc_code(nvc0, prog);
796 if (ret) {
797 struct nouveau_heap *heap = screen->text_heap;
798 struct nvc0_program *progs[] = { /* Sorted accordingly to SP_START_ID */
799 nvc0->compprog, nvc0->vertprog, nvc0->tctlprog,
800 nvc0->tevlprog, nvc0->gmtyprog, nvc0->fragprog
801 };
802
803 /* Note that the code library, which is allocated before anything else,
804 * does not have a priv pointer. We can stop once we hit it.
805 */
806 while (heap->next && heap->next->priv) {
807 struct nvc0_program *evict = heap->next->priv;
808 nouveau_heap_free(&evict->mem);
809 }
810 debug_printf("WARNING: out of code space, evicting all shaders.\n");
811
812 /* Make sure to synchronize before deleting the code segment. */
813 IMMED_NVC0(nvc0->base.pushbuf, NVC0_3D(SERIALIZE), 0);
814
815 if ((screen->text->size << 1) <= (1 << 23)) {
816 ret = nvc0_screen_resize_text_area(screen, screen->text->size << 1);
817 if (ret) {
818 NOUVEAU_ERR("Error allocating TEXT area: %d\n", ret);
819 return false;
820 }
821 nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_3D_TEXT);
822 BCTX_REFN_bo(nvc0->bufctx_3d, 3D_TEXT,
823 NV_VRAM_DOMAIN(&screen->base) | NOUVEAU_BO_RD,
824 screen->text);
825 if (screen->compute) {
826 nouveau_bufctx_reset(nvc0->bufctx_cp, NVC0_BIND_CP_TEXT);
827 BCTX_REFN_bo(nvc0->bufctx_cp, CP_TEXT,
828 NV_VRAM_DOMAIN(&screen->base) | NOUVEAU_BO_RD,
829 screen->text);
830 }
831
832 /* Re-upload the builtin function into the new code segment. */
833 nvc0_program_library_upload(nvc0);
834 }
835
836 ret = nvc0_program_alloc_code(nvc0, prog);
837 if (ret) {
838 NOUVEAU_ERR("shader too large (0x%x) to fit in code space ?\n", size);
839 return false;
840 }
841
842 /* All currently bound shaders have to be reuploaded. */
843 for (int i = 0; i < ARRAY_SIZE(progs); i++) {
844 if (!progs[i] || progs[i] == prog)
845 continue;
846
847 ret = nvc0_program_alloc_code(nvc0, progs[i]);
848 if (ret) {
849 NOUVEAU_ERR("failed to re-upload a shader after code eviction.\n");
850 return false;
851 }
852 nvc0_program_upload_code(nvc0, progs[i]);
853
854 if (progs[i]->type == PIPE_SHADER_COMPUTE) {
855 /* Caches have to be invalidated but the CP_START_ID will be
856 * updated in the launch_grid functions. */
857 BEGIN_NVC0(nvc0->base.pushbuf, NVC0_CP(FLUSH), 1);
858 PUSH_DATA (nvc0->base.pushbuf, NVC0_COMPUTE_FLUSH_CODE);
859 } else {
860 BEGIN_NVC0(nvc0->base.pushbuf, NVC0_3D(SP_START_ID(i)), 1);
861 PUSH_DATA (nvc0->base.pushbuf, progs[i]->code_base);
862 }
863 }
864 }
865
866 nvc0_program_upload_code(nvc0, prog);
867
868 #ifdef DEBUG
869 if (debug_get_bool_option("NV50_PROG_DEBUG", false))
870 nvc0_program_dump(prog);
871 #endif
872
873 BEGIN_NVC0(nvc0->base.pushbuf, NVC0_3D(MEM_BARRIER), 1);
874 PUSH_DATA (nvc0->base.pushbuf, 0x1011);
875
876 return true;
877 }
878
879 /* Upload code for builtin functions like integer division emulation. */
880 void
nvc0_program_library_upload(struct nvc0_context * nvc0)881 nvc0_program_library_upload(struct nvc0_context *nvc0)
882 {
883 struct nvc0_screen *screen = nvc0->screen;
884 int ret;
885 uint32_t size;
886 const uint32_t *code;
887
888 if (screen->lib_code)
889 return;
890
891 nv50_ir_get_target_library(screen->base.device->chipset, &code, &size);
892 if (!size)
893 return;
894
895 ret = nouveau_heap_alloc(screen->text_heap, align(size, 0x100), NULL,
896 &screen->lib_code);
897 if (ret)
898 return;
899
900 nvc0->base.push_data(&nvc0->base,
901 screen->text, screen->lib_code->start, NV_VRAM_DOMAIN(&screen->base),
902 size, code);
903 /* no need for a memory barrier, will be emitted with first program */
904 }
905
906 void
nvc0_program_destroy(struct nvc0_context * nvc0,struct nvc0_program * prog)907 nvc0_program_destroy(struct nvc0_context *nvc0, struct nvc0_program *prog)
908 {
909 const struct pipe_shader_state pipe = prog->pipe;
910 const ubyte type = prog->type;
911
912 if (prog->mem)
913 nouveau_heap_free(&prog->mem);
914 FREE(prog->code); /* may be 0 for hardcoded shaders */
915 FREE(prog->relocs);
916 FREE(prog->fixups);
917 if (prog->type == PIPE_SHADER_COMPUTE && prog->cp.syms)
918 FREE(prog->cp.syms);
919 if (prog->tfb) {
920 if (nvc0->state.tfb == prog->tfb)
921 nvc0->state.tfb = NULL;
922 FREE(prog->tfb);
923 }
924
925 memset(prog, 0, sizeof(*prog));
926
927 prog->pipe = pipe;
928 prog->type = type;
929 }
930
931 uint32_t
nvc0_program_symbol_offset(const struct nvc0_program * prog,uint32_t label)932 nvc0_program_symbol_offset(const struct nvc0_program *prog, uint32_t label)
933 {
934 const struct nv50_ir_prog_symbol *syms =
935 (const struct nv50_ir_prog_symbol *)prog->cp.syms;
936 unsigned base = 0;
937 unsigned i;
938 if (prog->type != PIPE_SHADER_COMPUTE)
939 base = NVC0_SHADER_HEADER_SIZE;
940 for (i = 0; i < prog->cp.num_syms; ++i)
941 if (syms[i].label == label)
942 return prog->code_base + base + syms[i].offset;
943 return prog->code_base; /* no symbols or symbol not found */
944 }
945
946 void
nvc0_program_init_tcp_empty(struct nvc0_context * nvc0)947 nvc0_program_init_tcp_empty(struct nvc0_context *nvc0)
948 {
949 struct ureg_program *ureg;
950
951 ureg = ureg_create(PIPE_SHADER_TESS_CTRL);
952 if (!ureg)
953 return;
954
955 ureg_property(ureg, TGSI_PROPERTY_TCS_VERTICES_OUT, 1);
956 ureg_END(ureg);
957
958 nvc0->tcp_empty = ureg_create_shader_and_destroy(ureg, &nvc0->base.pipe);
959 }
960