aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/gpu/drm/nouveau/nv50_grctx.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/gpu/drm/nouveau/nv50_grctx.c')
-rw-r--r--drivers/gpu/drm/nouveau/nv50_grctx.c2383
1 files changed, 2383 insertions, 0 deletions
diff --git a/drivers/gpu/drm/nouveau/nv50_grctx.c b/drivers/gpu/drm/nouveau/nv50_grctx.c
new file mode 100644
index 000000000000..42a8fb20c1e6
--- /dev/null
+++ b/drivers/gpu/drm/nouveau/nv50_grctx.c
@@ -0,0 +1,2383 @@
1/*
2 * Copyright 2009 Marcin Kościelnicki
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 COPYRIGHT HOLDER(S) OR AUTHOR(S) 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#define CP_FLAG_CLEAR 0
24#define CP_FLAG_SET 1
25#define CP_FLAG_SWAP_DIRECTION ((0 * 32) + 0)
26#define CP_FLAG_SWAP_DIRECTION_LOAD 0
27#define CP_FLAG_SWAP_DIRECTION_SAVE 1
28#define CP_FLAG_UNK01 ((0 * 32) + 1)
29#define CP_FLAG_UNK01_CLEAR 0
30#define CP_FLAG_UNK01_SET 1
31#define CP_FLAG_UNK03 ((0 * 32) + 3)
32#define CP_FLAG_UNK03_CLEAR 0
33#define CP_FLAG_UNK03_SET 1
34#define CP_FLAG_USER_SAVE ((0 * 32) + 5)
35#define CP_FLAG_USER_SAVE_NOT_PENDING 0
36#define CP_FLAG_USER_SAVE_PENDING 1
37#define CP_FLAG_USER_LOAD ((0 * 32) + 6)
38#define CP_FLAG_USER_LOAD_NOT_PENDING 0
39#define CP_FLAG_USER_LOAD_PENDING 1
40#define CP_FLAG_UNK0B ((0 * 32) + 0xb)
41#define CP_FLAG_UNK0B_CLEAR 0
42#define CP_FLAG_UNK0B_SET 1
43#define CP_FLAG_UNK1D ((0 * 32) + 0x1d)
44#define CP_FLAG_UNK1D_CLEAR 0
45#define CP_FLAG_UNK1D_SET 1
46#define CP_FLAG_UNK20 ((1 * 32) + 0)
47#define CP_FLAG_UNK20_CLEAR 0
48#define CP_FLAG_UNK20_SET 1
49#define CP_FLAG_STATUS ((2 * 32) + 0)
50#define CP_FLAG_STATUS_BUSY 0
51#define CP_FLAG_STATUS_IDLE 1
52#define CP_FLAG_AUTO_SAVE ((2 * 32) + 4)
53#define CP_FLAG_AUTO_SAVE_NOT_PENDING 0
54#define CP_FLAG_AUTO_SAVE_PENDING 1
55#define CP_FLAG_AUTO_LOAD ((2 * 32) + 5)
56#define CP_FLAG_AUTO_LOAD_NOT_PENDING 0
57#define CP_FLAG_AUTO_LOAD_PENDING 1
58#define CP_FLAG_NEWCTX ((2 * 32) + 10)
59#define CP_FLAG_NEWCTX_BUSY 0
60#define CP_FLAG_NEWCTX_DONE 1
61#define CP_FLAG_XFER ((2 * 32) + 11)
62#define CP_FLAG_XFER_IDLE 0
63#define CP_FLAG_XFER_BUSY 1
64#define CP_FLAG_ALWAYS ((2 * 32) + 13)
65#define CP_FLAG_ALWAYS_FALSE 0
66#define CP_FLAG_ALWAYS_TRUE 1
67#define CP_FLAG_INTR ((2 * 32) + 15)
68#define CP_FLAG_INTR_NOT_PENDING 0
69#define CP_FLAG_INTR_PENDING 1
70
71#define CP_CTX 0x00100000
72#define CP_CTX_COUNT 0x000f0000
73#define CP_CTX_COUNT_SHIFT 16
74#define CP_CTX_REG 0x00003fff
75#define CP_LOAD_SR 0x00200000
76#define CP_LOAD_SR_VALUE 0x000fffff
77#define CP_BRA 0x00400000
78#define CP_BRA_IP 0x0001ff00
79#define CP_BRA_IP_SHIFT 8
80#define CP_BRA_IF_CLEAR 0x00000080
81#define CP_BRA_FLAG 0x0000007f
82#define CP_WAIT 0x00500000
83#define CP_WAIT_SET 0x00000080
84#define CP_WAIT_FLAG 0x0000007f
85#define CP_SET 0x00700000
86#define CP_SET_1 0x00000080
87#define CP_SET_FLAG 0x0000007f
88#define CP_NEWCTX 0x00600004
89#define CP_NEXT_TO_SWAP 0x00600005
90#define CP_SET_CONTEXT_POINTER 0x00600006
91#define CP_SET_XFER_POINTER 0x00600007
92#define CP_ENABLE 0x00600009
93#define CP_END 0x0060000c
94#define CP_NEXT_TO_CURRENT 0x0060000d
95#define CP_DISABLE1 0x0090ffff
96#define CP_DISABLE2 0x0091ffff
97#define CP_XFER_1 0x008000ff
98#define CP_XFER_2 0x008800ff
99#define CP_SEEK_1 0x00c000ff
100#define CP_SEEK_2 0x00c800ff
101
102#include "drmP.h"
103#include "nouveau_drv.h"
104#include "nouveau_grctx.h"
105
106/*
107 * This code deals with PGRAPH contexts on NV50 family cards. Like NV40, it's
108 * the GPU itself that does context-switching, but it needs a special
109 * microcode to do it. And it's the driver's task to supply this microcode,
110 * further known as ctxprog, as well as the initial context values, known
111 * as ctxvals.
112 *
113 * Without ctxprog, you cannot switch contexts. Not even in software, since
114 * the majority of context [xfer strands] isn't accessible directly. You're
115 * stuck with a single channel, and you also suffer all the problems resulting
116 * from missing ctxvals, since you cannot load them.
117 *
118 * Without ctxvals, you're stuck with PGRAPH's default context. It's enough to
119 * run 2d operations, but trying to utilise 3d or CUDA will just lock you up,
120 * since you don't have... some sort of needed setup.
121 *
122 * Nouveau will just disable acceleration if not given ctxprog + ctxvals, since
123 * it's too much hassle to handle no-ctxprog as a special case.
124 */
125
126/*
127 * How ctxprogs work.
128 *
129 * The ctxprog is written in its own kind of microcode, with very small and
130 * crappy set of available commands. You upload it to a small [512 insns]
131 * area of memory on PGRAPH, and it'll be run when PFIFO wants PGRAPH to
132 * switch channel. or when the driver explicitely requests it. Stuff visible
133 * to ctxprog consists of: PGRAPH MMIO registers, PGRAPH context strands,
134 * the per-channel context save area in VRAM [known as ctxvals or grctx],
135 * 4 flags registers, a scratch register, two grctx pointers, plus many
136 * random poorly-understood details.
137 *
138 * When ctxprog runs, it's supposed to check what operations are asked of it,
139 * save old context if requested, optionally reset PGRAPH and switch to the
140 * new channel, and load the new context. Context consists of three major
141 * parts: subset of MMIO registers and two "xfer areas".
142 */
143
144/* TODO:
145 * - document unimplemented bits compared to nvidia
146 * - NVAx: make a TP subroutine, use it.
147 * - use 0x4008fc instead of 0x1540?
148 */
149
150enum cp_label {
151 cp_check_load = 1,
152 cp_setup_auto_load,
153 cp_setup_load,
154 cp_setup_save,
155 cp_swap_state,
156 cp_prepare_exit,
157 cp_exit,
158};
159
160static void nv50_graph_construct_mmio(struct nouveau_grctx *ctx);
161static void nv50_graph_construct_xfer1(struct nouveau_grctx *ctx);
162static void nv50_graph_construct_xfer2(struct nouveau_grctx *ctx);
163
164/* Main function: construct the ctxprog skeleton, call the other functions. */
165
166int
167nv50_grctx_init(struct nouveau_grctx *ctx)
168{
169 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
170
171 switch (dev_priv->chipset) {
172 case 0x50:
173 case 0x84:
174 case 0x86:
175 case 0x92:
176 case 0x94:
177 case 0x96:
178 case 0x98:
179 case 0xa0:
180 case 0xa3:
181 case 0xa5:
182 case 0xa8:
183 case 0xaa:
184 case 0xac:
185 break;
186 default:
187 NV_ERROR(ctx->dev, "I don't know how to make a ctxprog for "
188 "your NV%x card.\n", dev_priv->chipset);
189 NV_ERROR(ctx->dev, "Disabling acceleration. Please contact "
190 "the devs.\n");
191 return -ENOSYS;
192 }
193 /* decide whether we're loading/unloading the context */
194 cp_bra (ctx, AUTO_SAVE, PENDING, cp_setup_save);
195 cp_bra (ctx, USER_SAVE, PENDING, cp_setup_save);
196
197 cp_name(ctx, cp_check_load);
198 cp_bra (ctx, AUTO_LOAD, PENDING, cp_setup_auto_load);
199 cp_bra (ctx, USER_LOAD, PENDING, cp_setup_load);
200 cp_bra (ctx, ALWAYS, TRUE, cp_exit);
201
202 /* setup for context load */
203 cp_name(ctx, cp_setup_auto_load);
204 cp_out (ctx, CP_DISABLE1);
205 cp_out (ctx, CP_DISABLE2);
206 cp_out (ctx, CP_ENABLE);
207 cp_out (ctx, CP_NEXT_TO_SWAP);
208 cp_set (ctx, UNK01, SET);
209 cp_name(ctx, cp_setup_load);
210 cp_out (ctx, CP_NEWCTX);
211 cp_wait(ctx, NEWCTX, BUSY);
212 cp_set (ctx, UNK1D, CLEAR);
213 cp_set (ctx, SWAP_DIRECTION, LOAD);
214 cp_bra (ctx, UNK0B, SET, cp_prepare_exit);
215 cp_bra (ctx, ALWAYS, TRUE, cp_swap_state);
216
217 /* setup for context save */
218 cp_name(ctx, cp_setup_save);
219 cp_set (ctx, UNK1D, SET);
220 cp_wait(ctx, STATUS, BUSY);
221 cp_wait(ctx, INTR, PENDING);
222 cp_bra (ctx, STATUS, BUSY, cp_setup_save);
223 cp_set (ctx, UNK01, SET);
224 cp_set (ctx, SWAP_DIRECTION, SAVE);
225
226 /* general PGRAPH state */
227 cp_name(ctx, cp_swap_state);
228 cp_set (ctx, UNK03, SET);
229 cp_pos (ctx, 0x00004/4);
230 cp_ctx (ctx, 0x400828, 1); /* needed. otherwise, flickering happens. */
231 cp_pos (ctx, 0x00100/4);
232 nv50_graph_construct_mmio(ctx);
233 nv50_graph_construct_xfer1(ctx);
234 nv50_graph_construct_xfer2(ctx);
235
236 cp_bra (ctx, SWAP_DIRECTION, SAVE, cp_check_load);
237
238 cp_set (ctx, UNK20, SET);
239 cp_set (ctx, SWAP_DIRECTION, SAVE); /* no idea why this is needed, but fixes at least one lockup. */
240 cp_lsr (ctx, ctx->ctxvals_base);
241 cp_out (ctx, CP_SET_XFER_POINTER);
242 cp_lsr (ctx, 4);
243 cp_out (ctx, CP_SEEK_1);
244 cp_out (ctx, CP_XFER_1);
245 cp_wait(ctx, XFER, BUSY);
246
247 /* pre-exit state updates */
248 cp_name(ctx, cp_prepare_exit);
249 cp_set (ctx, UNK01, CLEAR);
250 cp_set (ctx, UNK03, CLEAR);
251 cp_set (ctx, UNK1D, CLEAR);
252
253 cp_bra (ctx, USER_SAVE, PENDING, cp_exit);
254 cp_out (ctx, CP_NEXT_TO_CURRENT);
255
256 cp_name(ctx, cp_exit);
257 cp_set (ctx, USER_SAVE, NOT_PENDING);
258 cp_set (ctx, USER_LOAD, NOT_PENDING);
259 cp_out (ctx, CP_END);
260 ctx->ctxvals_pos += 0x400; /* padding... no idea why you need it */
261
262 return 0;
263}
264
265/*
266 * Constructs MMIO part of ctxprog and ctxvals. Just a matter of knowing which
267 * registers to save/restore and the default values for them.
268 */
269
270static void
271nv50_graph_construct_mmio(struct nouveau_grctx *ctx)
272{
273 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
274 int i, j;
275 int offset, base;
276 uint32_t units = nv_rd32 (ctx->dev, 0x1540);
277
278 /* 0800: DISPATCH */
279 cp_ctx(ctx, 0x400808, 7);
280 gr_def(ctx, 0x400814, 0x00000030);
281 cp_ctx(ctx, 0x400834, 0x32);
282 if (dev_priv->chipset == 0x50) {
283 gr_def(ctx, 0x400834, 0xff400040);
284 gr_def(ctx, 0x400838, 0xfff00080);
285 gr_def(ctx, 0x40083c, 0xfff70090);
286 gr_def(ctx, 0x400840, 0xffe806a8);
287 }
288 gr_def(ctx, 0x400844, 0x00000002);
289 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
290 gr_def(ctx, 0x400894, 0x00001000);
291 gr_def(ctx, 0x4008e8, 0x00000003);
292 gr_def(ctx, 0x4008ec, 0x00001000);
293 if (dev_priv->chipset == 0x50)
294 cp_ctx(ctx, 0x400908, 0xb);
295 else if (dev_priv->chipset < 0xa0)
296 cp_ctx(ctx, 0x400908, 0xc);
297 else
298 cp_ctx(ctx, 0x400908, 0xe);
299
300 if (dev_priv->chipset >= 0xa0)
301 cp_ctx(ctx, 0x400b00, 0x1);
302 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
303 cp_ctx(ctx, 0x400b10, 0x1);
304 gr_def(ctx, 0x400b10, 0x0001629d);
305 cp_ctx(ctx, 0x400b20, 0x1);
306 gr_def(ctx, 0x400b20, 0x0001629d);
307 }
308
309 /* 0C00: VFETCH */
310 cp_ctx(ctx, 0x400c08, 0x2);
311 gr_def(ctx, 0x400c08, 0x0000fe0c);
312
313 /* 1000 */
314 if (dev_priv->chipset < 0xa0) {
315 cp_ctx(ctx, 0x401008, 0x4);
316 gr_def(ctx, 0x401014, 0x00001000);
317 } else if (dev_priv->chipset == 0xa0 || dev_priv->chipset >= 0xaa) {
318 cp_ctx(ctx, 0x401008, 0x5);
319 gr_def(ctx, 0x401018, 0x00001000);
320 } else {
321 cp_ctx(ctx, 0x401008, 0x5);
322 gr_def(ctx, 0x401018, 0x00004000);
323 }
324
325 /* 1400 */
326 cp_ctx(ctx, 0x401400, 0x8);
327 cp_ctx(ctx, 0x401424, 0x3);
328 if (dev_priv->chipset == 0x50)
329 gr_def(ctx, 0x40142c, 0x0001fd87);
330 else
331 gr_def(ctx, 0x40142c, 0x00000187);
332 cp_ctx(ctx, 0x401540, 0x5);
333 gr_def(ctx, 0x401550, 0x00001018);
334
335 /* 1800: STREAMOUT */
336 cp_ctx(ctx, 0x401814, 0x1);
337 gr_def(ctx, 0x401814, 0x000000ff);
338 if (dev_priv->chipset == 0x50) {
339 cp_ctx(ctx, 0x40181c, 0xe);
340 gr_def(ctx, 0x401850, 0x00000004);
341 } else if (dev_priv->chipset < 0xa0) {
342 cp_ctx(ctx, 0x40181c, 0xf);
343 gr_def(ctx, 0x401854, 0x00000004);
344 } else {
345 cp_ctx(ctx, 0x40181c, 0x13);
346 gr_def(ctx, 0x401864, 0x00000004);
347 }
348
349 /* 1C00 */
350 cp_ctx(ctx, 0x401c00, 0x1);
351 switch (dev_priv->chipset) {
352 case 0x50:
353 gr_def(ctx, 0x401c00, 0x0001005f);
354 break;
355 case 0x84:
356 case 0x86:
357 case 0x94:
358 gr_def(ctx, 0x401c00, 0x044d00df);
359 break;
360 case 0x92:
361 case 0x96:
362 case 0x98:
363 case 0xa0:
364 case 0xaa:
365 case 0xac:
366 gr_def(ctx, 0x401c00, 0x042500df);
367 break;
368 case 0xa3:
369 case 0xa5:
370 case 0xa8:
371 gr_def(ctx, 0x401c00, 0x142500df);
372 break;
373 }
374
375 /* 2400 */
376 cp_ctx(ctx, 0x402400, 0x1);
377 if (dev_priv->chipset == 0x50)
378 cp_ctx(ctx, 0x402408, 0x1);
379 else
380 cp_ctx(ctx, 0x402408, 0x2);
381 gr_def(ctx, 0x402408, 0x00000600);
382
383 /* 2800 */
384 cp_ctx(ctx, 0x402800, 0x1);
385 if (dev_priv->chipset == 0x50)
386 gr_def(ctx, 0x402800, 0x00000006);
387
388 /* 2C00 */
389 cp_ctx(ctx, 0x402c08, 0x6);
390 if (dev_priv->chipset != 0x50)
391 gr_def(ctx, 0x402c14, 0x01000000);
392 gr_def(ctx, 0x402c18, 0x000000ff);
393 if (dev_priv->chipset == 0x50)
394 cp_ctx(ctx, 0x402ca0, 0x1);
395 else
396 cp_ctx(ctx, 0x402ca0, 0x2);
397 if (dev_priv->chipset < 0xa0)
398 gr_def(ctx, 0x402ca0, 0x00000400);
399 else if (dev_priv->chipset == 0xa0 || dev_priv->chipset >= 0xaa)
400 gr_def(ctx, 0x402ca0, 0x00000800);
401 else
402 gr_def(ctx, 0x402ca0, 0x00000400);
403 cp_ctx(ctx, 0x402cac, 0x4);
404
405 /* 3000 */
406 cp_ctx(ctx, 0x403004, 0x1);
407 gr_def(ctx, 0x403004, 0x00000001);
408
409 /* 3404 */
410 if (dev_priv->chipset >= 0xa0) {
411 cp_ctx(ctx, 0x403404, 0x1);
412 gr_def(ctx, 0x403404, 0x00000001);
413 }
414
415 /* 5000 */
416 cp_ctx(ctx, 0x405000, 0x1);
417 switch (dev_priv->chipset) {
418 case 0x50:
419 gr_def(ctx, 0x405000, 0x00300080);
420 break;
421 case 0x84:
422 case 0xa0:
423 case 0xa3:
424 case 0xa5:
425 case 0xa8:
426 case 0xaa:
427 case 0xac:
428 gr_def(ctx, 0x405000, 0x000e0080);
429 break;
430 case 0x86:
431 case 0x92:
432 case 0x94:
433 case 0x96:
434 case 0x98:
435 gr_def(ctx, 0x405000, 0x00000080);
436 break;
437 }
438 cp_ctx(ctx, 0x405014, 0x1);
439 gr_def(ctx, 0x405014, 0x00000004);
440 cp_ctx(ctx, 0x40501c, 0x1);
441 cp_ctx(ctx, 0x405024, 0x1);
442 cp_ctx(ctx, 0x40502c, 0x1);
443
444 /* 5400 or maybe 4800 */
445 if (dev_priv->chipset == 0x50) {
446 offset = 0x405400;
447 cp_ctx(ctx, 0x405400, 0xea);
448 } else if (dev_priv->chipset < 0x94) {
449 offset = 0x405400;
450 cp_ctx(ctx, 0x405400, 0xcb);
451 } else if (dev_priv->chipset < 0xa0) {
452 offset = 0x405400;
453 cp_ctx(ctx, 0x405400, 0xcc);
454 } else if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
455 offset = 0x404800;
456 cp_ctx(ctx, 0x404800, 0xda);
457 } else {
458 offset = 0x405400;
459 cp_ctx(ctx, 0x405400, 0xd4);
460 }
461 gr_def(ctx, offset + 0x0c, 0x00000002);
462 gr_def(ctx, offset + 0x10, 0x00000001);
463 if (dev_priv->chipset >= 0x94)
464 offset += 4;
465 gr_def(ctx, offset + 0x1c, 0x00000001);
466 gr_def(ctx, offset + 0x20, 0x00000100);
467 gr_def(ctx, offset + 0x38, 0x00000002);
468 gr_def(ctx, offset + 0x3c, 0x00000001);
469 gr_def(ctx, offset + 0x40, 0x00000001);
470 gr_def(ctx, offset + 0x50, 0x00000001);
471 gr_def(ctx, offset + 0x54, 0x003fffff);
472 gr_def(ctx, offset + 0x58, 0x00001fff);
473 gr_def(ctx, offset + 0x60, 0x00000001);
474 gr_def(ctx, offset + 0x64, 0x00000001);
475 gr_def(ctx, offset + 0x6c, 0x00000001);
476 gr_def(ctx, offset + 0x70, 0x00000001);
477 gr_def(ctx, offset + 0x74, 0x00000001);
478 gr_def(ctx, offset + 0x78, 0x00000004);
479 gr_def(ctx, offset + 0x7c, 0x00000001);
480 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
481 offset += 4;
482 gr_def(ctx, offset + 0x80, 0x00000001);
483 gr_def(ctx, offset + 0x84, 0x00000001);
484 gr_def(ctx, offset + 0x88, 0x00000007);
485 gr_def(ctx, offset + 0x8c, 0x00000001);
486 gr_def(ctx, offset + 0x90, 0x00000007);
487 gr_def(ctx, offset + 0x94, 0x00000001);
488 gr_def(ctx, offset + 0x98, 0x00000001);
489 gr_def(ctx, offset + 0x9c, 0x00000001);
490 if (dev_priv->chipset == 0x50) {
491 gr_def(ctx, offset + 0xb0, 0x00000001);
492 gr_def(ctx, offset + 0xb4, 0x00000001);
493 gr_def(ctx, offset + 0xbc, 0x00000001);
494 gr_def(ctx, offset + 0xc0, 0x0000000a);
495 gr_def(ctx, offset + 0xd0, 0x00000040);
496 gr_def(ctx, offset + 0xd8, 0x00000002);
497 gr_def(ctx, offset + 0xdc, 0x00000100);
498 gr_def(ctx, offset + 0xe0, 0x00000001);
499 gr_def(ctx, offset + 0xe4, 0x00000100);
500 gr_def(ctx, offset + 0x100, 0x00000001);
501 gr_def(ctx, offset + 0x124, 0x00000004);
502 gr_def(ctx, offset + 0x13c, 0x00000001);
503 gr_def(ctx, offset + 0x140, 0x00000100);
504 gr_def(ctx, offset + 0x148, 0x00000001);
505 gr_def(ctx, offset + 0x154, 0x00000100);
506 gr_def(ctx, offset + 0x158, 0x00000001);
507 gr_def(ctx, offset + 0x15c, 0x00000100);
508 gr_def(ctx, offset + 0x164, 0x00000001);
509 gr_def(ctx, offset + 0x170, 0x00000100);
510 gr_def(ctx, offset + 0x174, 0x00000001);
511 gr_def(ctx, offset + 0x17c, 0x00000001);
512 gr_def(ctx, offset + 0x188, 0x00000002);
513 gr_def(ctx, offset + 0x190, 0x00000001);
514 gr_def(ctx, offset + 0x198, 0x00000001);
515 gr_def(ctx, offset + 0x1ac, 0x00000003);
516 offset += 0xd0;
517 } else {
518 gr_def(ctx, offset + 0xb0, 0x00000001);
519 gr_def(ctx, offset + 0xb4, 0x00000100);
520 gr_def(ctx, offset + 0xbc, 0x00000001);
521 gr_def(ctx, offset + 0xc8, 0x00000100);
522 gr_def(ctx, offset + 0xcc, 0x00000001);
523 gr_def(ctx, offset + 0xd0, 0x00000100);
524 gr_def(ctx, offset + 0xd8, 0x00000001);
525 gr_def(ctx, offset + 0xe4, 0x00000100);
526 }
527 gr_def(ctx, offset + 0xf8, 0x00000004);
528 gr_def(ctx, offset + 0xfc, 0x00000070);
529 gr_def(ctx, offset + 0x100, 0x00000080);
530 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
531 offset += 4;
532 gr_def(ctx, offset + 0x114, 0x0000000c);
533 if (dev_priv->chipset == 0x50)
534 offset -= 4;
535 gr_def(ctx, offset + 0x11c, 0x00000008);
536 gr_def(ctx, offset + 0x120, 0x00000014);
537 if (dev_priv->chipset == 0x50) {
538 gr_def(ctx, offset + 0x124, 0x00000026);
539 offset -= 0x18;
540 } else {
541 gr_def(ctx, offset + 0x128, 0x00000029);
542 gr_def(ctx, offset + 0x12c, 0x00000027);
543 gr_def(ctx, offset + 0x130, 0x00000026);
544 gr_def(ctx, offset + 0x134, 0x00000008);
545 gr_def(ctx, offset + 0x138, 0x00000004);
546 gr_def(ctx, offset + 0x13c, 0x00000027);
547 }
548 gr_def(ctx, offset + 0x148, 0x00000001);
549 gr_def(ctx, offset + 0x14c, 0x00000002);
550 gr_def(ctx, offset + 0x150, 0x00000003);
551 gr_def(ctx, offset + 0x154, 0x00000004);
552 gr_def(ctx, offset + 0x158, 0x00000005);
553 gr_def(ctx, offset + 0x15c, 0x00000006);
554 gr_def(ctx, offset + 0x160, 0x00000007);
555 gr_def(ctx, offset + 0x164, 0x00000001);
556 gr_def(ctx, offset + 0x1a8, 0x000000cf);
557 if (dev_priv->chipset == 0x50)
558 offset -= 4;
559 gr_def(ctx, offset + 0x1d8, 0x00000080);
560 gr_def(ctx, offset + 0x1dc, 0x00000004);
561 gr_def(ctx, offset + 0x1e0, 0x00000004);
562 if (dev_priv->chipset == 0x50)
563 offset -= 4;
564 else
565 gr_def(ctx, offset + 0x1e4, 0x00000003);
566 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
567 gr_def(ctx, offset + 0x1ec, 0x00000003);
568 offset += 8;
569 }
570 gr_def(ctx, offset + 0x1e8, 0x00000001);
571 if (dev_priv->chipset == 0x50)
572 offset -= 4;
573 gr_def(ctx, offset + 0x1f4, 0x00000012);
574 gr_def(ctx, offset + 0x1f8, 0x00000010);
575 gr_def(ctx, offset + 0x1fc, 0x0000000c);
576 gr_def(ctx, offset + 0x200, 0x00000001);
577 gr_def(ctx, offset + 0x210, 0x00000004);
578 gr_def(ctx, offset + 0x214, 0x00000002);
579 gr_def(ctx, offset + 0x218, 0x00000004);
580 if (dev_priv->chipset >= 0xa0)
581 offset += 4;
582 gr_def(ctx, offset + 0x224, 0x003fffff);
583 gr_def(ctx, offset + 0x228, 0x00001fff);
584 if (dev_priv->chipset == 0x50)
585 offset -= 0x20;
586 else if (dev_priv->chipset >= 0xa0) {
587 gr_def(ctx, offset + 0x250, 0x00000001);
588 gr_def(ctx, offset + 0x254, 0x00000001);
589 gr_def(ctx, offset + 0x258, 0x00000002);
590 offset += 0x10;
591 }
592 gr_def(ctx, offset + 0x250, 0x00000004);
593 gr_def(ctx, offset + 0x254, 0x00000014);
594 gr_def(ctx, offset + 0x258, 0x00000001);
595 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
596 offset += 4;
597 gr_def(ctx, offset + 0x264, 0x00000002);
598 if (dev_priv->chipset >= 0xa0)
599 offset += 8;
600 gr_def(ctx, offset + 0x270, 0x00000001);
601 gr_def(ctx, offset + 0x278, 0x00000002);
602 gr_def(ctx, offset + 0x27c, 0x00001000);
603 if (dev_priv->chipset == 0x50)
604 offset -= 0xc;
605 else {
606 gr_def(ctx, offset + 0x280, 0x00000e00);
607 gr_def(ctx, offset + 0x284, 0x00001000);
608 gr_def(ctx, offset + 0x288, 0x00001e00);
609 }
610 gr_def(ctx, offset + 0x290, 0x00000001);
611 gr_def(ctx, offset + 0x294, 0x00000001);
612 gr_def(ctx, offset + 0x298, 0x00000001);
613 gr_def(ctx, offset + 0x29c, 0x00000001);
614 gr_def(ctx, offset + 0x2a0, 0x00000001);
615 gr_def(ctx, offset + 0x2b0, 0x00000200);
616 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
617 gr_def(ctx, offset + 0x2b4, 0x00000200);
618 offset += 4;
619 }
620 if (dev_priv->chipset < 0xa0) {
621 gr_def(ctx, offset + 0x2b8, 0x00000001);
622 gr_def(ctx, offset + 0x2bc, 0x00000070);
623 gr_def(ctx, offset + 0x2c0, 0x00000080);
624 gr_def(ctx, offset + 0x2cc, 0x00000001);
625 gr_def(ctx, offset + 0x2d0, 0x00000070);
626 gr_def(ctx, offset + 0x2d4, 0x00000080);
627 } else {
628 gr_def(ctx, offset + 0x2b8, 0x00000001);
629 gr_def(ctx, offset + 0x2bc, 0x000000f0);
630 gr_def(ctx, offset + 0x2c0, 0x000000ff);
631 gr_def(ctx, offset + 0x2cc, 0x00000001);
632 gr_def(ctx, offset + 0x2d0, 0x000000f0);
633 gr_def(ctx, offset + 0x2d4, 0x000000ff);
634 gr_def(ctx, offset + 0x2dc, 0x00000009);
635 offset += 4;
636 }
637 gr_def(ctx, offset + 0x2e4, 0x00000001);
638 gr_def(ctx, offset + 0x2e8, 0x000000cf);
639 gr_def(ctx, offset + 0x2f0, 0x00000001);
640 gr_def(ctx, offset + 0x300, 0x000000cf);
641 gr_def(ctx, offset + 0x308, 0x00000002);
642 gr_def(ctx, offset + 0x310, 0x00000001);
643 gr_def(ctx, offset + 0x318, 0x00000001);
644 gr_def(ctx, offset + 0x320, 0x000000cf);
645 gr_def(ctx, offset + 0x324, 0x000000cf);
646 gr_def(ctx, offset + 0x328, 0x00000001);
647
648 /* 6000? */
649 if (dev_priv->chipset == 0x50)
650 cp_ctx(ctx, 0x4063e0, 0x1);
651
652 /* 6800: M2MF */
653 if (dev_priv->chipset < 0x90) {
654 cp_ctx(ctx, 0x406814, 0x2b);
655 gr_def(ctx, 0x406818, 0x00000f80);
656 gr_def(ctx, 0x406860, 0x007f0080);
657 gr_def(ctx, 0x40689c, 0x007f0080);
658 } else {
659 cp_ctx(ctx, 0x406814, 0x4);
660 if (dev_priv->chipset == 0x98)
661 gr_def(ctx, 0x406818, 0x00000f80);
662 else
663 gr_def(ctx, 0x406818, 0x00001f80);
664 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
665 gr_def(ctx, 0x40681c, 0x00000030);
666 cp_ctx(ctx, 0x406830, 0x3);
667 }
668
669 /* 7000: per-ROP group state */
670 for (i = 0; i < 8; i++) {
671 if (units & (1<<(i+16))) {
672 cp_ctx(ctx, 0x407000 + (i<<8), 3);
673 if (dev_priv->chipset == 0x50)
674 gr_def(ctx, 0x407000 + (i<<8), 0x1b74f820);
675 else if (dev_priv->chipset != 0xa5)
676 gr_def(ctx, 0x407000 + (i<<8), 0x3b74f821);
677 else
678 gr_def(ctx, 0x407000 + (i<<8), 0x7b74f821);
679 gr_def(ctx, 0x407004 + (i<<8), 0x89058001);
680
681 if (dev_priv->chipset == 0x50) {
682 cp_ctx(ctx, 0x407010 + (i<<8), 1);
683 } else if (dev_priv->chipset < 0xa0) {
684 cp_ctx(ctx, 0x407010 + (i<<8), 2);
685 gr_def(ctx, 0x407010 + (i<<8), 0x00001000);
686 gr_def(ctx, 0x407014 + (i<<8), 0x0000001f);
687 } else {
688 cp_ctx(ctx, 0x407010 + (i<<8), 3);
689 gr_def(ctx, 0x407010 + (i<<8), 0x00001000);
690 if (dev_priv->chipset != 0xa5)
691 gr_def(ctx, 0x407014 + (i<<8), 0x000000ff);
692 else
693 gr_def(ctx, 0x407014 + (i<<8), 0x000001ff);
694 }
695
696 cp_ctx(ctx, 0x407080 + (i<<8), 4);
697 if (dev_priv->chipset != 0xa5)
698 gr_def(ctx, 0x407080 + (i<<8), 0x027c10fa);
699 else
700 gr_def(ctx, 0x407080 + (i<<8), 0x827c10fa);
701 if (dev_priv->chipset == 0x50)
702 gr_def(ctx, 0x407084 + (i<<8), 0x000000c0);
703 else
704 gr_def(ctx, 0x407084 + (i<<8), 0x400000c0);
705 gr_def(ctx, 0x407088 + (i<<8), 0xb7892080);
706
707 if (dev_priv->chipset < 0xa0)
708 cp_ctx(ctx, 0x407094 + (i<<8), 1);
709 else if (dev_priv->chipset <= 0xa0 || dev_priv->chipset >= 0xaa)
710 cp_ctx(ctx, 0x407094 + (i<<8), 3);
711 else {
712 cp_ctx(ctx, 0x407094 + (i<<8), 4);
713 gr_def(ctx, 0x4070a0 + (i<<8), 1);
714 }
715 }
716 }
717
718 cp_ctx(ctx, 0x407c00, 0x3);
719 if (dev_priv->chipset < 0x90)
720 gr_def(ctx, 0x407c00, 0x00010040);
721 else if (dev_priv->chipset < 0xa0)
722 gr_def(ctx, 0x407c00, 0x00390040);
723 else
724 gr_def(ctx, 0x407c00, 0x003d0040);
725 gr_def(ctx, 0x407c08, 0x00000022);
726 if (dev_priv->chipset >= 0xa0) {
727 cp_ctx(ctx, 0x407c10, 0x3);
728 cp_ctx(ctx, 0x407c20, 0x1);
729 cp_ctx(ctx, 0x407c2c, 0x1);
730 }
731
732 if (dev_priv->chipset < 0xa0) {
733 cp_ctx(ctx, 0x407d00, 0x9);
734 } else {
735 cp_ctx(ctx, 0x407d00, 0x15);
736 }
737 if (dev_priv->chipset == 0x98)
738 gr_def(ctx, 0x407d08, 0x00380040);
739 else {
740 if (dev_priv->chipset < 0x90)
741 gr_def(ctx, 0x407d08, 0x00010040);
742 else if (dev_priv->chipset < 0xa0)
743 gr_def(ctx, 0x407d08, 0x00390040);
744 else
745 gr_def(ctx, 0x407d08, 0x003d0040);
746 gr_def(ctx, 0x407d0c, 0x00000022);
747 }
748
749 /* 8000+: per-TP state */
750 for (i = 0; i < 10; i++) {
751 if (units & (1<<i)) {
752 if (dev_priv->chipset < 0xa0)
753 base = 0x408000 + (i<<12);
754 else
755 base = 0x408000 + (i<<11);
756 if (dev_priv->chipset < 0xa0)
757 offset = base + 0xc00;
758 else
759 offset = base + 0x80;
760 cp_ctx(ctx, offset + 0x00, 1);
761 gr_def(ctx, offset + 0x00, 0x0000ff0a);
762 cp_ctx(ctx, offset + 0x08, 1);
763
764 /* per-MP state */
765 for (j = 0; j < (dev_priv->chipset < 0xa0 ? 2 : 4); j++) {
766 if (!(units & (1 << (j+24)))) continue;
767 if (dev_priv->chipset < 0xa0)
768 offset = base + 0x200 + (j<<7);
769 else
770 offset = base + 0x100 + (j<<7);
771 cp_ctx(ctx, offset, 0x20);
772 gr_def(ctx, offset + 0x00, 0x01800000);
773 gr_def(ctx, offset + 0x04, 0x00160000);
774 gr_def(ctx, offset + 0x08, 0x01800000);
775 gr_def(ctx, offset + 0x18, 0x0003ffff);
776 switch (dev_priv->chipset) {
777 case 0x50:
778 gr_def(ctx, offset + 0x1c, 0x00080000);
779 break;
780 case 0x84:
781 gr_def(ctx, offset + 0x1c, 0x00880000);
782 break;
783 case 0x86:
784 gr_def(ctx, offset + 0x1c, 0x008c0000);
785 break;
786 case 0x92:
787 case 0x96:
788 case 0x98:
789 gr_def(ctx, offset + 0x1c, 0x118c0000);
790 break;
791 case 0x94:
792 gr_def(ctx, offset + 0x1c, 0x10880000);
793 break;
794 case 0xa0:
795 case 0xa5:
796 gr_def(ctx, offset + 0x1c, 0x310c0000);
797 break;
798 case 0xa3:
799 case 0xa8:
800 case 0xaa:
801 case 0xac:
802 gr_def(ctx, offset + 0x1c, 0x300c0000);
803 break;
804 }
805 gr_def(ctx, offset + 0x40, 0x00010401);
806 if (dev_priv->chipset == 0x50)
807 gr_def(ctx, offset + 0x48, 0x00000040);
808 else
809 gr_def(ctx, offset + 0x48, 0x00000078);
810 gr_def(ctx, offset + 0x50, 0x000000bf);
811 gr_def(ctx, offset + 0x58, 0x00001210);
812 if (dev_priv->chipset == 0x50)
813 gr_def(ctx, offset + 0x5c, 0x00000080);
814 else
815 gr_def(ctx, offset + 0x5c, 0x08000080);
816 if (dev_priv->chipset >= 0xa0)
817 gr_def(ctx, offset + 0x68, 0x0000003e);
818 }
819
820 if (dev_priv->chipset < 0xa0)
821 cp_ctx(ctx, base + 0x300, 0x4);
822 else
823 cp_ctx(ctx, base + 0x300, 0x5);
824 if (dev_priv->chipset == 0x50)
825 gr_def(ctx, base + 0x304, 0x00007070);
826 else if (dev_priv->chipset < 0xa0)
827 gr_def(ctx, base + 0x304, 0x00027070);
828 else if (dev_priv->chipset <= 0xa0 || dev_priv->chipset >= 0xaa)
829 gr_def(ctx, base + 0x304, 0x01127070);
830 else
831 gr_def(ctx, base + 0x304, 0x05127070);
832
833 if (dev_priv->chipset < 0xa0)
834 cp_ctx(ctx, base + 0x318, 1);
835 else
836 cp_ctx(ctx, base + 0x320, 1);
837 if (dev_priv->chipset == 0x50)
838 gr_def(ctx, base + 0x318, 0x0003ffff);
839 else if (dev_priv->chipset < 0xa0)
840 gr_def(ctx, base + 0x318, 0x03ffffff);
841 else
842 gr_def(ctx, base + 0x320, 0x07ffffff);
843
844 if (dev_priv->chipset < 0xa0)
845 cp_ctx(ctx, base + 0x324, 5);
846 else
847 cp_ctx(ctx, base + 0x328, 4);
848
849 if (dev_priv->chipset < 0xa0) {
850 cp_ctx(ctx, base + 0x340, 9);
851 offset = base + 0x340;
852 } else if (dev_priv->chipset <= 0xa0 || dev_priv->chipset >= 0xaa) {
853 cp_ctx(ctx, base + 0x33c, 0xb);
854 offset = base + 0x344;
855 } else {
856 cp_ctx(ctx, base + 0x33c, 0xd);
857 offset = base + 0x344;
858 }
859 gr_def(ctx, offset + 0x0, 0x00120407);
860 gr_def(ctx, offset + 0x4, 0x05091507);
861 if (dev_priv->chipset == 0x84)
862 gr_def(ctx, offset + 0x8, 0x05100202);
863 else
864 gr_def(ctx, offset + 0x8, 0x05010202);
865 gr_def(ctx, offset + 0xc, 0x00030201);
866 if (dev_priv->chipset == 0xa3)
867 cp_ctx(ctx, base + 0x36c, 1);
868
869 cp_ctx(ctx, base + 0x400, 2);
870 gr_def(ctx, base + 0x404, 0x00000040);
871 cp_ctx(ctx, base + 0x40c, 2);
872 gr_def(ctx, base + 0x40c, 0x0d0c0b0a);
873 gr_def(ctx, base + 0x410, 0x00141210);
874
875 if (dev_priv->chipset < 0xa0)
876 offset = base + 0x800;
877 else
878 offset = base + 0x500;
879 cp_ctx(ctx, offset, 6);
880 gr_def(ctx, offset + 0x0, 0x000001f0);
881 gr_def(ctx, offset + 0x4, 0x00000001);
882 gr_def(ctx, offset + 0x8, 0x00000003);
883 if (dev_priv->chipset == 0x50 || dev_priv->chipset >= 0xaa)
884 gr_def(ctx, offset + 0xc, 0x00008000);
885 gr_def(ctx, offset + 0x14, 0x00039e00);
886 cp_ctx(ctx, offset + 0x1c, 2);
887 if (dev_priv->chipset == 0x50)
888 gr_def(ctx, offset + 0x1c, 0x00000040);
889 else
890 gr_def(ctx, offset + 0x1c, 0x00000100);
891 gr_def(ctx, offset + 0x20, 0x00003800);
892
893 if (dev_priv->chipset >= 0xa0) {
894 cp_ctx(ctx, base + 0x54c, 2);
895 if (dev_priv->chipset <= 0xa0 || dev_priv->chipset >= 0xaa)
896 gr_def(ctx, base + 0x54c, 0x003fe006);
897 else
898 gr_def(ctx, base + 0x54c, 0x003fe007);
899 gr_def(ctx, base + 0x550, 0x003fe000);
900 }
901
902 if (dev_priv->chipset < 0xa0)
903 offset = base + 0xa00;
904 else
905 offset = base + 0x680;
906 cp_ctx(ctx, offset, 1);
907 gr_def(ctx, offset, 0x00404040);
908
909 if (dev_priv->chipset < 0xa0)
910 offset = base + 0xe00;
911 else
912 offset = base + 0x700;
913 cp_ctx(ctx, offset, 2);
914 if (dev_priv->chipset < 0xa0)
915 gr_def(ctx, offset, 0x0077f005);
916 else if (dev_priv->chipset == 0xa5)
917 gr_def(ctx, offset, 0x6cf7f007);
918 else if (dev_priv->chipset == 0xa8)
919 gr_def(ctx, offset, 0x6cfff007);
920 else if (dev_priv->chipset == 0xac)
921 gr_def(ctx, offset, 0x0cfff007);
922 else
923 gr_def(ctx, offset, 0x0cf7f007);
924 if (dev_priv->chipset == 0x50)
925 gr_def(ctx, offset + 0x4, 0x00007fff);
926 else if (dev_priv->chipset < 0xa0)
927 gr_def(ctx, offset + 0x4, 0x003f7fff);
928 else
929 gr_def(ctx, offset + 0x4, 0x02bf7fff);
930 cp_ctx(ctx, offset + 0x2c, 1);
931 if (dev_priv->chipset == 0x50) {
932 cp_ctx(ctx, offset + 0x50, 9);
933 gr_def(ctx, offset + 0x54, 0x000003ff);
934 gr_def(ctx, offset + 0x58, 0x00000003);
935 gr_def(ctx, offset + 0x5c, 0x00000003);
936 gr_def(ctx, offset + 0x60, 0x000001ff);
937 gr_def(ctx, offset + 0x64, 0x0000001f);
938 gr_def(ctx, offset + 0x68, 0x0000000f);
939 gr_def(ctx, offset + 0x6c, 0x0000000f);
940 } else if(dev_priv->chipset < 0xa0) {
941 cp_ctx(ctx, offset + 0x50, 1);
942 cp_ctx(ctx, offset + 0x70, 1);
943 } else {
944 cp_ctx(ctx, offset + 0x50, 1);
945 cp_ctx(ctx, offset + 0x60, 5);
946 }
947 }
948 }
949}
950
951/*
952 * xfer areas. These are a pain.
953 *
954 * There are 2 xfer areas: the first one is big and contains all sorts of
955 * stuff, the second is small and contains some per-TP context.
956 *
957 * Each area is split into 8 "strands". The areas, when saved to grctx,
958 * are made of 8-word blocks. Each block contains a single word from
959 * each strand. The strands are independent of each other, their
960 * addresses are unrelated to each other, and data in them is closely
961 * packed together. The strand layout varies a bit between cards: here
962 * and there, a single word is thrown out in the middle and the whole
963 * strand is offset by a bit from corresponding one on another chipset.
964 * For this reason, addresses of stuff in strands are almost useless.
965 * Knowing sequence of stuff and size of gaps between them is much more
966 * useful, and that's how we build the strands in our generator.
967 *
968 * NVA0 takes this mess to a whole new level by cutting the old strands
969 * into a few dozen pieces [known as genes], rearranging them randomly,
970 * and putting them back together to make new strands. Hopefully these
971 * genes correspond more or less directly to the same PGRAPH subunits
972 * as in 400040 register.
973 *
974 * The most common value in default context is 0, and when the genes
975 * are separated by 0's, gene bounduaries are quite speculative...
976 * some of them can be clearly deduced, others can be guessed, and yet
977 * others won't be resolved without figuring out the real meaning of
978 * given ctxval. For the same reason, ending point of each strand
979 * is unknown. Except for strand 0, which is the longest strand and
980 * its end corresponds to end of the whole xfer.
981 *
982 * An unsolved mystery is the seek instruction: it takes an argument
983 * in bits 8-18, and that argument is clearly the place in strands to
984 * seek to... but the offsets don't seem to correspond to offsets as
985 * seen in grctx. Perhaps there's another, real, not randomly-changing
986 * addressing in strands, and the xfer insn just happens to skip over
987 * the unused bits? NV10-NV30 PIPE comes to mind...
988 *
989 * As far as I know, there's no way to access the xfer areas directly
990 * without the help of ctxprog.
991 */
992
993static inline void
994xf_emit(struct nouveau_grctx *ctx, int num, uint32_t val) {
995 int i;
996 if (val && ctx->mode == NOUVEAU_GRCTX_VALS)
997 for (i = 0; i < num; i++)
998 nv_wo32(ctx->dev, ctx->data, ctx->ctxvals_pos + (i << 3), val);
999 ctx->ctxvals_pos += num << 3;
1000}
1001
1002/* Gene declarations... */
1003
1004static void nv50_graph_construct_gene_m2mf(struct nouveau_grctx *ctx);
1005static void nv50_graph_construct_gene_unk1(struct nouveau_grctx *ctx);
1006static void nv50_graph_construct_gene_unk2(struct nouveau_grctx *ctx);
1007static void nv50_graph_construct_gene_unk3(struct nouveau_grctx *ctx);
1008static void nv50_graph_construct_gene_unk4(struct nouveau_grctx *ctx);
1009static void nv50_graph_construct_gene_unk5(struct nouveau_grctx *ctx);
1010static void nv50_graph_construct_gene_unk6(struct nouveau_grctx *ctx);
1011static void nv50_graph_construct_gene_unk7(struct nouveau_grctx *ctx);
1012static void nv50_graph_construct_gene_unk8(struct nouveau_grctx *ctx);
1013static void nv50_graph_construct_gene_unk9(struct nouveau_grctx *ctx);
1014static void nv50_graph_construct_gene_unk10(struct nouveau_grctx *ctx);
1015static void nv50_graph_construct_gene_ropc(struct nouveau_grctx *ctx);
1016static void nv50_graph_construct_xfer_tp(struct nouveau_grctx *ctx);
1017
1018static void
1019nv50_graph_construct_xfer1(struct nouveau_grctx *ctx)
1020{
1021 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1022 int i;
1023 int offset;
1024 int size = 0;
1025 uint32_t units = nv_rd32 (ctx->dev, 0x1540);
1026
1027 offset = (ctx->ctxvals_pos+0x3f)&~0x3f;
1028 ctx->ctxvals_base = offset;
1029
1030 if (dev_priv->chipset < 0xa0) {
1031 /* Strand 0 */
1032 ctx->ctxvals_pos = offset;
1033 switch (dev_priv->chipset) {
1034 case 0x50:
1035 xf_emit(ctx, 0x99, 0);
1036 break;
1037 case 0x84:
1038 case 0x86:
1039 xf_emit(ctx, 0x384, 0);
1040 break;
1041 case 0x92:
1042 case 0x94:
1043 case 0x96:
1044 case 0x98:
1045 xf_emit(ctx, 0x380, 0);
1046 break;
1047 }
1048 nv50_graph_construct_gene_m2mf (ctx);
1049 switch (dev_priv->chipset) {
1050 case 0x50:
1051 case 0x84:
1052 case 0x86:
1053 case 0x98:
1054 xf_emit(ctx, 0x4c4, 0);
1055 break;
1056 case 0x92:
1057 case 0x94:
1058 case 0x96:
1059 xf_emit(ctx, 0x984, 0);
1060 break;
1061 }
1062 nv50_graph_construct_gene_unk5(ctx);
1063 if (dev_priv->chipset == 0x50)
1064 xf_emit(ctx, 0xa, 0);
1065 else
1066 xf_emit(ctx, 0xb, 0);
1067 nv50_graph_construct_gene_unk4(ctx);
1068 nv50_graph_construct_gene_unk3(ctx);
1069 if ((ctx->ctxvals_pos-offset)/8 > size)
1070 size = (ctx->ctxvals_pos-offset)/8;
1071
1072 /* Strand 1 */
1073 ctx->ctxvals_pos = offset + 0x1;
1074 nv50_graph_construct_gene_unk6(ctx);
1075 nv50_graph_construct_gene_unk7(ctx);
1076 nv50_graph_construct_gene_unk8(ctx);
1077 switch (dev_priv->chipset) {
1078 case 0x50:
1079 case 0x92:
1080 xf_emit(ctx, 0xfb, 0);
1081 break;
1082 case 0x84:
1083 xf_emit(ctx, 0xd3, 0);
1084 break;
1085 case 0x94:
1086 case 0x96:
1087 xf_emit(ctx, 0xab, 0);
1088 break;
1089 case 0x86:
1090 case 0x98:
1091 xf_emit(ctx, 0x6b, 0);
1092 break;
1093 }
1094 xf_emit(ctx, 2, 0x4e3bfdf);
1095 xf_emit(ctx, 4, 0);
1096 xf_emit(ctx, 1, 0x0fac6881);
1097 xf_emit(ctx, 0xb, 0);
1098 xf_emit(ctx, 2, 0x4e3bfdf);
1099 if ((ctx->ctxvals_pos-offset)/8 > size)
1100 size = (ctx->ctxvals_pos-offset)/8;
1101
1102 /* Strand 2 */
1103 ctx->ctxvals_pos = offset + 0x2;
1104 switch (dev_priv->chipset) {
1105 case 0x50:
1106 case 0x92:
1107 xf_emit(ctx, 0xa80, 0);
1108 break;
1109 case 0x84:
1110 xf_emit(ctx, 0xa7e, 0);
1111 break;
1112 case 0x94:
1113 case 0x96:
1114 xf_emit(ctx, 0xa7c, 0);
1115 break;
1116 case 0x86:
1117 case 0x98:
1118 xf_emit(ctx, 0xa7a, 0);
1119 break;
1120 }
1121 xf_emit(ctx, 1, 0x3fffff);
1122 xf_emit(ctx, 2, 0);
1123 xf_emit(ctx, 1, 0x1fff);
1124 xf_emit(ctx, 0xe, 0);
1125 nv50_graph_construct_gene_unk9(ctx);
1126 nv50_graph_construct_gene_unk2(ctx);
1127 nv50_graph_construct_gene_unk1(ctx);
1128 nv50_graph_construct_gene_unk10(ctx);
1129 if ((ctx->ctxvals_pos-offset)/8 > size)
1130 size = (ctx->ctxvals_pos-offset)/8;
1131
1132 /* Strand 3: per-ROP group state */
1133 ctx->ctxvals_pos = offset + 3;
1134 for (i = 0; i < 6; i++)
1135 if (units & (1 << (i + 16)))
1136 nv50_graph_construct_gene_ropc(ctx);
1137 if ((ctx->ctxvals_pos-offset)/8 > size)
1138 size = (ctx->ctxvals_pos-offset)/8;
1139
1140 /* Strands 4-7: per-TP state */
1141 for (i = 0; i < 4; i++) {
1142 ctx->ctxvals_pos = offset + 4 + i;
1143 if (units & (1 << (2 * i)))
1144 nv50_graph_construct_xfer_tp(ctx);
1145 if (units & (1 << (2 * i + 1)))
1146 nv50_graph_construct_xfer_tp(ctx);
1147 if ((ctx->ctxvals_pos-offset)/8 > size)
1148 size = (ctx->ctxvals_pos-offset)/8;
1149 }
1150 } else {
1151 /* Strand 0 */
1152 ctx->ctxvals_pos = offset;
1153 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1154 xf_emit(ctx, 0x385, 0);
1155 else
1156 xf_emit(ctx, 0x384, 0);
1157 nv50_graph_construct_gene_m2mf(ctx);
1158 xf_emit(ctx, 0x950, 0);
1159 nv50_graph_construct_gene_unk10(ctx);
1160 xf_emit(ctx, 1, 0x0fac6881);
1161 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
1162 xf_emit(ctx, 1, 1);
1163 xf_emit(ctx, 3, 0);
1164 }
1165 nv50_graph_construct_gene_unk8(ctx);
1166 if (dev_priv->chipset == 0xa0)
1167 xf_emit(ctx, 0x189, 0);
1168 else if (dev_priv->chipset == 0xa3)
1169 xf_emit(ctx, 0xd5, 0);
1170 else if (dev_priv->chipset == 0xa5)
1171 xf_emit(ctx, 0x99, 0);
1172 else if (dev_priv->chipset == 0xaa)
1173 xf_emit(ctx, 0x65, 0);
1174 else
1175 xf_emit(ctx, 0x6d, 0);
1176 nv50_graph_construct_gene_unk9(ctx);
1177 if ((ctx->ctxvals_pos-offset)/8 > size)
1178 size = (ctx->ctxvals_pos-offset)/8;
1179
1180 /* Strand 1 */
1181 ctx->ctxvals_pos = offset + 1;
1182 nv50_graph_construct_gene_unk1(ctx);
1183 if ((ctx->ctxvals_pos-offset)/8 > size)
1184 size = (ctx->ctxvals_pos-offset)/8;
1185
1186 /* Strand 2 */
1187 ctx->ctxvals_pos = offset + 2;
1188 if (dev_priv->chipset == 0xa0) {
1189 nv50_graph_construct_gene_unk2(ctx);
1190 }
1191 xf_emit(ctx, 0x36, 0);
1192 nv50_graph_construct_gene_unk5(ctx);
1193 if ((ctx->ctxvals_pos-offset)/8 > size)
1194 size = (ctx->ctxvals_pos-offset)/8;
1195
1196 /* Strand 3 */
1197 ctx->ctxvals_pos = offset + 3;
1198 xf_emit(ctx, 1, 0);
1199 xf_emit(ctx, 1, 1);
1200 nv50_graph_construct_gene_unk6(ctx);
1201 if ((ctx->ctxvals_pos-offset)/8 > size)
1202 size = (ctx->ctxvals_pos-offset)/8;
1203
1204 /* Strand 4 */
1205 ctx->ctxvals_pos = offset + 4;
1206 if (dev_priv->chipset == 0xa0)
1207 xf_emit(ctx, 0xa80, 0);
1208 else if (dev_priv->chipset == 0xa3)
1209 xf_emit(ctx, 0xa7c, 0);
1210 else
1211 xf_emit(ctx, 0xa7a, 0);
1212 xf_emit(ctx, 1, 0x3fffff);
1213 xf_emit(ctx, 2, 0);
1214 xf_emit(ctx, 1, 0x1fff);
1215 if ((ctx->ctxvals_pos-offset)/8 > size)
1216 size = (ctx->ctxvals_pos-offset)/8;
1217
1218 /* Strand 5 */
1219 ctx->ctxvals_pos = offset + 5;
1220 xf_emit(ctx, 1, 0);
1221 xf_emit(ctx, 1, 0x0fac6881);
1222 xf_emit(ctx, 0xb, 0);
1223 xf_emit(ctx, 2, 0x4e3bfdf);
1224 xf_emit(ctx, 3, 0);
1225 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1226 xf_emit(ctx, 1, 0x11);
1227 xf_emit(ctx, 1, 0);
1228 xf_emit(ctx, 2, 0x4e3bfdf);
1229 xf_emit(ctx, 2, 0);
1230 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1231 xf_emit(ctx, 1, 0x11);
1232 xf_emit(ctx, 1, 0);
1233 for (i = 0; i < 8; i++)
1234 if (units & (1<<(i+16)))
1235 nv50_graph_construct_gene_ropc(ctx);
1236 if ((ctx->ctxvals_pos-offset)/8 > size)
1237 size = (ctx->ctxvals_pos-offset)/8;
1238
1239 /* Strand 6 */
1240 ctx->ctxvals_pos = offset + 6;
1241 nv50_graph_construct_gene_unk3(ctx);
1242 xf_emit(ctx, 0xb, 0);
1243 nv50_graph_construct_gene_unk4(ctx);
1244 nv50_graph_construct_gene_unk7(ctx);
1245 if (units & (1 << 0))
1246 nv50_graph_construct_xfer_tp(ctx);
1247 if (units & (1 << 1))
1248 nv50_graph_construct_xfer_tp(ctx);
1249 if (units & (1 << 2))
1250 nv50_graph_construct_xfer_tp(ctx);
1251 if (units & (1 << 3))
1252 nv50_graph_construct_xfer_tp(ctx);
1253 if ((ctx->ctxvals_pos-offset)/8 > size)
1254 size = (ctx->ctxvals_pos-offset)/8;
1255
1256 /* Strand 7 */
1257 ctx->ctxvals_pos = offset + 7;
1258 if (dev_priv->chipset == 0xa0) {
1259 if (units & (1 << 4))
1260 nv50_graph_construct_xfer_tp(ctx);
1261 if (units & (1 << 5))
1262 nv50_graph_construct_xfer_tp(ctx);
1263 if (units & (1 << 6))
1264 nv50_graph_construct_xfer_tp(ctx);
1265 if (units & (1 << 7))
1266 nv50_graph_construct_xfer_tp(ctx);
1267 if (units & (1 << 8))
1268 nv50_graph_construct_xfer_tp(ctx);
1269 if (units & (1 << 9))
1270 nv50_graph_construct_xfer_tp(ctx);
1271 } else {
1272 nv50_graph_construct_gene_unk2(ctx);
1273 }
1274 if ((ctx->ctxvals_pos-offset)/8 > size)
1275 size = (ctx->ctxvals_pos-offset)/8;
1276 }
1277
1278 ctx->ctxvals_pos = offset + size * 8;
1279 ctx->ctxvals_pos = (ctx->ctxvals_pos+0x3f)&~0x3f;
1280 cp_lsr (ctx, offset);
1281 cp_out (ctx, CP_SET_XFER_POINTER);
1282 cp_lsr (ctx, size);
1283 cp_out (ctx, CP_SEEK_1);
1284 cp_out (ctx, CP_XFER_1);
1285 cp_wait(ctx, XFER, BUSY);
1286}
1287
1288/*
1289 * non-trivial demagiced parts of ctx init go here
1290 */
1291
1292static void
1293nv50_graph_construct_gene_m2mf(struct nouveau_grctx *ctx)
1294{
1295 /* m2mf state */
1296 xf_emit (ctx, 1, 0); /* DMA_NOTIFY instance >> 4 */
1297 xf_emit (ctx, 1, 0); /* DMA_BUFFER_IN instance >> 4 */
1298 xf_emit (ctx, 1, 0); /* DMA_BUFFER_OUT instance >> 4 */
1299 xf_emit (ctx, 1, 0); /* OFFSET_IN */
1300 xf_emit (ctx, 1, 0); /* OFFSET_OUT */
1301 xf_emit (ctx, 1, 0); /* PITCH_IN */
1302 xf_emit (ctx, 1, 0); /* PITCH_OUT */
1303 xf_emit (ctx, 1, 0); /* LINE_LENGTH */
1304 xf_emit (ctx, 1, 0); /* LINE_COUNT */
1305 xf_emit (ctx, 1, 0x21); /* FORMAT: bits 0-4 INPUT_INC, bits 5-9 OUTPUT_INC */
1306 xf_emit (ctx, 1, 1); /* LINEAR_IN */
1307 xf_emit (ctx, 1, 0x2); /* TILING_MODE_IN: bits 0-2 y tiling, bits 3-5 z tiling */
1308 xf_emit (ctx, 1, 0x100); /* TILING_PITCH_IN */
1309 xf_emit (ctx, 1, 0x100); /* TILING_HEIGHT_IN */
1310 xf_emit (ctx, 1, 1); /* TILING_DEPTH_IN */
1311 xf_emit (ctx, 1, 0); /* TILING_POSITION_IN_Z */
1312 xf_emit (ctx, 1, 0); /* TILING_POSITION_IN */
1313 xf_emit (ctx, 1, 1); /* LINEAR_OUT */
1314 xf_emit (ctx, 1, 0x2); /* TILING_MODE_OUT: bits 0-2 y tiling, bits 3-5 z tiling */
1315 xf_emit (ctx, 1, 0x100); /* TILING_PITCH_OUT */
1316 xf_emit (ctx, 1, 0x100); /* TILING_HEIGHT_OUT */
1317 xf_emit (ctx, 1, 1); /* TILING_DEPTH_OUT */
1318 xf_emit (ctx, 1, 0); /* TILING_POSITION_OUT_Z */
1319 xf_emit (ctx, 1, 0); /* TILING_POSITION_OUT */
1320 xf_emit (ctx, 1, 0); /* OFFSET_IN_HIGH */
1321 xf_emit (ctx, 1, 0); /* OFFSET_OUT_HIGH */
1322}
1323
1324static void
1325nv50_graph_construct_gene_unk1(struct nouveau_grctx *ctx)
1326{
1327 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1328 /* end of area 2 on pre-NVA0, area 1 on NVAx */
1329 xf_emit(ctx, 2, 4);
1330 xf_emit(ctx, 1, 0);
1331 xf_emit(ctx, 1, 0x80);
1332 xf_emit(ctx, 1, 4);
1333 xf_emit(ctx, 1, 0x80c14);
1334 xf_emit(ctx, 1, 0);
1335 if (dev_priv->chipset == 0x50)
1336 xf_emit(ctx, 1, 0x3ff);
1337 else
1338 xf_emit(ctx, 1, 0x7ff);
1339 switch (dev_priv->chipset) {
1340 case 0x50:
1341 case 0x86:
1342 case 0x98:
1343 case 0xaa:
1344 case 0xac:
1345 xf_emit(ctx, 0x542, 0);
1346 break;
1347 case 0x84:
1348 case 0x92:
1349 case 0x94:
1350 case 0x96:
1351 xf_emit(ctx, 0x942, 0);
1352 break;
1353 case 0xa0:
1354 case 0xa3:
1355 xf_emit(ctx, 0x2042, 0);
1356 break;
1357 case 0xa5:
1358 case 0xa8:
1359 xf_emit(ctx, 0x842, 0);
1360 break;
1361 }
1362 xf_emit(ctx, 2, 4);
1363 xf_emit(ctx, 1, 0);
1364 xf_emit(ctx, 1, 0x80);
1365 xf_emit(ctx, 1, 4);
1366 xf_emit(ctx, 1, 1);
1367 xf_emit(ctx, 1, 0);
1368 xf_emit(ctx, 1, 0x27);
1369 xf_emit(ctx, 1, 0);
1370 xf_emit(ctx, 1, 0x26);
1371 xf_emit(ctx, 3, 0);
1372}
1373
1374static void
1375nv50_graph_construct_gene_unk10(struct nouveau_grctx *ctx)
1376{
1377 /* end of area 2 on pre-NVA0, area 1 on NVAx */
1378 xf_emit(ctx, 0x10, 0x04000000);
1379 xf_emit(ctx, 0x24, 0);
1380 xf_emit(ctx, 2, 0x04e3bfdf);
1381 xf_emit(ctx, 2, 0);
1382 xf_emit(ctx, 1, 0x1fe21);
1383}
1384
1385static void
1386nv50_graph_construct_gene_unk2(struct nouveau_grctx *ctx)
1387{
1388 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1389 /* middle of area 2 on pre-NVA0, beginning of area 2 on NVA0, area 7 on >NVA0 */
1390 if (dev_priv->chipset != 0x50) {
1391 xf_emit(ctx, 5, 0);
1392 xf_emit(ctx, 1, 0x80c14);
1393 xf_emit(ctx, 2, 0);
1394 xf_emit(ctx, 1, 0x804);
1395 xf_emit(ctx, 1, 0);
1396 xf_emit(ctx, 2, 4);
1397 xf_emit(ctx, 1, 0x8100c12);
1398 }
1399 xf_emit(ctx, 1, 0);
1400 xf_emit(ctx, 2, 4);
1401 xf_emit(ctx, 1, 0);
1402 xf_emit(ctx, 1, 0x10);
1403 if (dev_priv->chipset == 0x50)
1404 xf_emit(ctx, 3, 0);
1405 else
1406 xf_emit(ctx, 4, 0);
1407 xf_emit(ctx, 1, 0x804);
1408 xf_emit(ctx, 1, 1);
1409 xf_emit(ctx, 1, 0x1a);
1410 if (dev_priv->chipset != 0x50)
1411 xf_emit(ctx, 1, 0x7f);
1412 xf_emit(ctx, 1, 0);
1413 xf_emit(ctx, 1, 1);
1414 xf_emit(ctx, 1, 0x80c14);
1415 xf_emit(ctx, 1, 0);
1416 xf_emit(ctx, 1, 0x8100c12);
1417 xf_emit(ctx, 2, 4);
1418 xf_emit(ctx, 1, 0);
1419 xf_emit(ctx, 1, 0x10);
1420 xf_emit(ctx, 3, 0);
1421 xf_emit(ctx, 1, 1);
1422 xf_emit(ctx, 1, 0x8100c12);
1423 xf_emit(ctx, 6, 0);
1424 if (dev_priv->chipset == 0x50)
1425 xf_emit(ctx, 1, 0x3ff);
1426 else
1427 xf_emit(ctx, 1, 0x7ff);
1428 xf_emit(ctx, 1, 0x80c14);
1429 xf_emit(ctx, 0x38, 0);
1430 xf_emit(ctx, 1, 1);
1431 xf_emit(ctx, 2, 0);
1432 xf_emit(ctx, 1, 0x10);
1433 xf_emit(ctx, 0x38, 0);
1434 xf_emit(ctx, 2, 0x88);
1435 xf_emit(ctx, 2, 0);
1436 xf_emit(ctx, 1, 4);
1437 xf_emit(ctx, 0x16, 0);
1438 xf_emit(ctx, 1, 0x26);
1439 xf_emit(ctx, 2, 0);
1440 xf_emit(ctx, 1, 0x3f800000);
1441 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1442 xf_emit(ctx, 4, 0);
1443 else
1444 xf_emit(ctx, 3, 0);
1445 xf_emit(ctx, 1, 0x1a);
1446 xf_emit(ctx, 1, 0x10);
1447 if (dev_priv->chipset != 0x50)
1448 xf_emit(ctx, 0x28, 0);
1449 else
1450 xf_emit(ctx, 0x25, 0);
1451 xf_emit(ctx, 1, 0x52);
1452 xf_emit(ctx, 1, 0);
1453 xf_emit(ctx, 1, 0x26);
1454 xf_emit(ctx, 1, 0);
1455 xf_emit(ctx, 2, 4);
1456 xf_emit(ctx, 1, 0);
1457 xf_emit(ctx, 1, 0x1a);
1458 xf_emit(ctx, 2, 0);
1459 xf_emit(ctx, 1, 0x00ffff00);
1460 xf_emit(ctx, 1, 0);
1461}
1462
1463static void
1464nv50_graph_construct_gene_unk3(struct nouveau_grctx *ctx)
1465{
1466 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1467 /* end of area 0 on pre-NVA0, beginning of area 6 on NVAx */
1468 xf_emit(ctx, 1, 0x3f);
1469 xf_emit(ctx, 0xa, 0);
1470 xf_emit(ctx, 1, 2);
1471 xf_emit(ctx, 2, 0x04000000);
1472 xf_emit(ctx, 8, 0);
1473 xf_emit(ctx, 1, 4);
1474 xf_emit(ctx, 3, 0);
1475 xf_emit(ctx, 1, 4);
1476 if (dev_priv->chipset == 0x50)
1477 xf_emit(ctx, 0x10, 0);
1478 else
1479 xf_emit(ctx, 0x11, 0);
1480 xf_emit(ctx, 1, 1);
1481 xf_emit(ctx, 1, 0x1001);
1482 xf_emit(ctx, 4, 0xffff);
1483 xf_emit(ctx, 0x20, 0);
1484 xf_emit(ctx, 0x10, 0x3f800000);
1485 xf_emit(ctx, 1, 0x10);
1486 if (dev_priv->chipset == 0x50)
1487 xf_emit(ctx, 1, 0);
1488 else
1489 xf_emit(ctx, 2, 0);
1490 xf_emit(ctx, 1, 3);
1491 xf_emit(ctx, 2, 0);
1492}
1493
1494static void
1495nv50_graph_construct_gene_unk4(struct nouveau_grctx *ctx)
1496{
1497 /* middle of area 0 on pre-NVA0, middle of area 6 on NVAx */
1498 xf_emit(ctx, 2, 0x04000000);
1499 xf_emit(ctx, 1, 0);
1500 xf_emit(ctx, 1, 0x80);
1501 xf_emit(ctx, 3, 0);
1502 xf_emit(ctx, 1, 0x80);
1503 xf_emit(ctx, 1, 0);
1504}
1505
1506static void
1507nv50_graph_construct_gene_unk5(struct nouveau_grctx *ctx)
1508{
1509 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1510 /* middle of area 0 on pre-NVA0 [after m2mf], end of area 2 on NVAx */
1511 xf_emit(ctx, 2, 4);
1512 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1513 xf_emit(ctx, 0x1c4d, 0);
1514 else
1515 xf_emit(ctx, 0x1c4b, 0);
1516 xf_emit(ctx, 2, 4);
1517 xf_emit(ctx, 1, 0x8100c12);
1518 if (dev_priv->chipset != 0x50)
1519 xf_emit(ctx, 1, 3);
1520 xf_emit(ctx, 1, 0);
1521 xf_emit(ctx, 1, 0x8100c12);
1522 xf_emit(ctx, 1, 0);
1523 xf_emit(ctx, 1, 0x80c14);
1524 xf_emit(ctx, 1, 1);
1525 if (dev_priv->chipset >= 0xa0)
1526 xf_emit(ctx, 2, 4);
1527 xf_emit(ctx, 1, 0x80c14);
1528 xf_emit(ctx, 2, 0);
1529 xf_emit(ctx, 1, 0x8100c12);
1530 xf_emit(ctx, 1, 0x27);
1531 xf_emit(ctx, 2, 0);
1532 xf_emit(ctx, 1, 1);
1533 xf_emit(ctx, 0x3c1, 0);
1534 xf_emit(ctx, 1, 1);
1535 xf_emit(ctx, 0x16, 0);
1536 xf_emit(ctx, 1, 0x8100c12);
1537 xf_emit(ctx, 1, 0);
1538}
1539
1540static void
1541nv50_graph_construct_gene_unk6(struct nouveau_grctx *ctx)
1542{
1543 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1544 /* beginning of area 1 on pre-NVA0 [after m2mf], area 3 on NVAx */
1545 xf_emit(ctx, 4, 0);
1546 xf_emit(ctx, 1, 0xf);
1547 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1548 xf_emit(ctx, 8, 0);
1549 else
1550 xf_emit(ctx, 4, 0);
1551 xf_emit(ctx, 1, 0x20);
1552 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1553 xf_emit(ctx, 0x11, 0);
1554 else if (dev_priv->chipset >= 0xa0)
1555 xf_emit(ctx, 0xf, 0);
1556 else
1557 xf_emit(ctx, 0xe, 0);
1558 xf_emit(ctx, 1, 0x1a);
1559 xf_emit(ctx, 0xd, 0);
1560 xf_emit(ctx, 2, 4);
1561 xf_emit(ctx, 1, 0);
1562 xf_emit(ctx, 1, 4);
1563 xf_emit(ctx, 1, 8);
1564 xf_emit(ctx, 1, 0);
1565 if (dev_priv->chipset == 0x50)
1566 xf_emit(ctx, 1, 0x3ff);
1567 else
1568 xf_emit(ctx, 1, 0x7ff);
1569 if (dev_priv->chipset == 0xa8)
1570 xf_emit(ctx, 1, 0x1e00);
1571 xf_emit(ctx, 0xc, 0);
1572 xf_emit(ctx, 1, 0xf);
1573 if (dev_priv->chipset == 0x50)
1574 xf_emit(ctx, 0x125, 0);
1575 else if (dev_priv->chipset < 0xa0)
1576 xf_emit(ctx, 0x126, 0);
1577 else if (dev_priv->chipset == 0xa0 || dev_priv->chipset >= 0xaa)
1578 xf_emit(ctx, 0x124, 0);
1579 else
1580 xf_emit(ctx, 0x1f7, 0);
1581 xf_emit(ctx, 1, 0xf);
1582 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1583 xf_emit(ctx, 3, 0);
1584 else
1585 xf_emit(ctx, 1, 0);
1586 xf_emit(ctx, 1, 1);
1587 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1588 xf_emit(ctx, 0xa1, 0);
1589 else
1590 xf_emit(ctx, 0x5a, 0);
1591 xf_emit(ctx, 1, 0xf);
1592 if (dev_priv->chipset < 0xa0)
1593 xf_emit(ctx, 0x834, 0);
1594 else if (dev_priv->chipset == 0xa0)
1595 xf_emit(ctx, 0x1873, 0);
1596 else if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1597 xf_emit(ctx, 0x8ba, 0);
1598 else
1599 xf_emit(ctx, 0x833, 0);
1600 xf_emit(ctx, 1, 0xf);
1601 xf_emit(ctx, 0xf, 0);
1602}
1603
1604static void
1605nv50_graph_construct_gene_unk7(struct nouveau_grctx *ctx)
1606{
1607 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1608 /* middle of area 1 on pre-NVA0 [after m2mf], middle of area 6 on NVAx */
1609 xf_emit(ctx, 2, 0);
1610 if (dev_priv->chipset == 0x50)
1611 xf_emit(ctx, 2, 1);
1612 else
1613 xf_emit(ctx, 2, 0);
1614 xf_emit(ctx, 1, 0);
1615 xf_emit(ctx, 1, 1);
1616 xf_emit(ctx, 2, 0x100);
1617 xf_emit(ctx, 1, 0x11);
1618 xf_emit(ctx, 1, 0);
1619 xf_emit(ctx, 1, 8);
1620 xf_emit(ctx, 5, 0);
1621 xf_emit(ctx, 1, 1);
1622 xf_emit(ctx, 1, 0);
1623 xf_emit(ctx, 3, 1);
1624 xf_emit(ctx, 1, 0xcf);
1625 xf_emit(ctx, 1, 2);
1626 xf_emit(ctx, 6, 0);
1627 xf_emit(ctx, 1, 1);
1628 xf_emit(ctx, 1, 0);
1629 xf_emit(ctx, 3, 1);
1630 xf_emit(ctx, 4, 0);
1631 xf_emit(ctx, 1, 4);
1632 xf_emit(ctx, 1, 0);
1633 xf_emit(ctx, 1, 1);
1634 xf_emit(ctx, 1, 0x15);
1635 xf_emit(ctx, 3, 0);
1636 xf_emit(ctx, 1, 0x4444480);
1637 xf_emit(ctx, 0x37, 0);
1638}
1639
1640static void
1641nv50_graph_construct_gene_unk8(struct nouveau_grctx *ctx)
1642{
1643 /* middle of area 1 on pre-NVA0 [after m2mf], middle of area 0 on NVAx */
1644 xf_emit(ctx, 4, 0);
1645 xf_emit(ctx, 1, 0x8100c12);
1646 xf_emit(ctx, 4, 0);
1647 xf_emit(ctx, 1, 0x100);
1648 xf_emit(ctx, 2, 0);
1649 xf_emit(ctx, 1, 0x10001);
1650 xf_emit(ctx, 1, 0);
1651 xf_emit(ctx, 1, 0x10001);
1652 xf_emit(ctx, 1, 1);
1653 xf_emit(ctx, 1, 0x10001);
1654 xf_emit(ctx, 1, 1);
1655 xf_emit(ctx, 1, 4);
1656 xf_emit(ctx, 1, 2);
1657}
1658
1659static void
1660nv50_graph_construct_gene_unk9(struct nouveau_grctx *ctx)
1661{
1662 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1663 /* middle of area 2 on pre-NVA0 [after m2mf], end of area 0 on NVAx */
1664 xf_emit(ctx, 1, 0x3f800000);
1665 xf_emit(ctx, 6, 0);
1666 xf_emit(ctx, 1, 4);
1667 xf_emit(ctx, 1, 0x1a);
1668 xf_emit(ctx, 2, 0);
1669 xf_emit(ctx, 1, 1);
1670 xf_emit(ctx, 0x12, 0);
1671 xf_emit(ctx, 1, 0x00ffff00);
1672 xf_emit(ctx, 6, 0);
1673 xf_emit(ctx, 1, 0xf);
1674 xf_emit(ctx, 7, 0);
1675 xf_emit(ctx, 1, 0x0fac6881);
1676 xf_emit(ctx, 1, 0x11);
1677 xf_emit(ctx, 0xf, 0);
1678 xf_emit(ctx, 1, 4);
1679 xf_emit(ctx, 2, 0);
1680 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1681 xf_emit(ctx, 1, 3);
1682 else if (dev_priv->chipset >= 0xa0)
1683 xf_emit(ctx, 1, 1);
1684 xf_emit(ctx, 2, 0);
1685 xf_emit(ctx, 1, 2);
1686 xf_emit(ctx, 2, 0x04000000);
1687 xf_emit(ctx, 3, 0);
1688 xf_emit(ctx, 1, 5);
1689 xf_emit(ctx, 1, 0x52);
1690 if (dev_priv->chipset == 0x50) {
1691 xf_emit(ctx, 0x13, 0);
1692 } else {
1693 xf_emit(ctx, 4, 0);
1694 xf_emit(ctx, 1, 1);
1695 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1696 xf_emit(ctx, 0x11, 0);
1697 else
1698 xf_emit(ctx, 0x10, 0);
1699 }
1700 xf_emit(ctx, 0x10, 0x3f800000);
1701 xf_emit(ctx, 1, 0x10);
1702 xf_emit(ctx, 0x26, 0);
1703 xf_emit(ctx, 1, 0x8100c12);
1704 xf_emit(ctx, 1, 5);
1705 xf_emit(ctx, 2, 0);
1706 xf_emit(ctx, 1, 1);
1707 xf_emit(ctx, 1, 0);
1708 xf_emit(ctx, 4, 0xffff);
1709 if (dev_priv->chipset != 0x50)
1710 xf_emit(ctx, 1, 3);
1711 if (dev_priv->chipset < 0xa0)
1712 xf_emit(ctx, 0x1f, 0);
1713 else if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1714 xf_emit(ctx, 0xc, 0);
1715 else
1716 xf_emit(ctx, 3, 0);
1717 xf_emit(ctx, 1, 0x00ffff00);
1718 xf_emit(ctx, 1, 0x1a);
1719 if (dev_priv->chipset != 0x50) {
1720 xf_emit(ctx, 1, 0);
1721 xf_emit(ctx, 1, 3);
1722 }
1723 if (dev_priv->chipset < 0xa0)
1724 xf_emit(ctx, 0x26, 0);
1725 else
1726 xf_emit(ctx, 0x3c, 0);
1727 xf_emit(ctx, 1, 0x102);
1728 xf_emit(ctx, 1, 0);
1729 xf_emit(ctx, 4, 4);
1730 if (dev_priv->chipset >= 0xa0)
1731 xf_emit(ctx, 8, 0);
1732 xf_emit(ctx, 2, 4);
1733 xf_emit(ctx, 1, 0);
1734 if (dev_priv->chipset == 0x50)
1735 xf_emit(ctx, 1, 0x3ff);
1736 else
1737 xf_emit(ctx, 1, 0x7ff);
1738 xf_emit(ctx, 1, 0);
1739 xf_emit(ctx, 1, 0x102);
1740 xf_emit(ctx, 9, 0);
1741 xf_emit(ctx, 4, 4);
1742 xf_emit(ctx, 0x2c, 0);
1743}
1744
1745static void
1746nv50_graph_construct_gene_ropc(struct nouveau_grctx *ctx)
1747{
1748 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1749 int magic2;
1750 if (dev_priv->chipset == 0x50) {
1751 magic2 = 0x00003e60;
1752 } else if (dev_priv->chipset <= 0xa0 || dev_priv->chipset >= 0xaa) {
1753 magic2 = 0x001ffe67;
1754 } else {
1755 magic2 = 0x00087e67;
1756 }
1757 xf_emit(ctx, 8, 0);
1758 xf_emit(ctx, 1, 2);
1759 xf_emit(ctx, 1, 0);
1760 xf_emit(ctx, 1, magic2);
1761 xf_emit(ctx, 4, 0);
1762 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1763 xf_emit(ctx, 1, 1);
1764 xf_emit(ctx, 7, 0);
1765 if (dev_priv->chipset >= 0xa0 && dev_priv->chipset < 0xaa)
1766 xf_emit(ctx, 1, 0x15);
1767 xf_emit(ctx, 1, 0);
1768 xf_emit(ctx, 1, 1);
1769 xf_emit(ctx, 1, 0x10);
1770 xf_emit(ctx, 2, 0);
1771 xf_emit(ctx, 1, 1);
1772 xf_emit(ctx, 4, 0);
1773 if (dev_priv->chipset == 0x86 || dev_priv->chipset == 0x92 || dev_priv->chipset == 0x98 || dev_priv->chipset >= 0xa0) {
1774 xf_emit(ctx, 1, 4);
1775 xf_emit(ctx, 1, 0x400);
1776 xf_emit(ctx, 1, 0x300);
1777 xf_emit(ctx, 1, 0x1001);
1778 if (dev_priv->chipset != 0xa0) {
1779 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1780 xf_emit(ctx, 1, 0);
1781 else
1782 xf_emit(ctx, 1, 0x15);
1783 }
1784 xf_emit(ctx, 3, 0);
1785 }
1786 xf_emit(ctx, 2, 0);
1787 xf_emit(ctx, 1, 2);
1788 xf_emit(ctx, 8, 0);
1789 xf_emit(ctx, 1, 1);
1790 xf_emit(ctx, 1, 0x10);
1791 xf_emit(ctx, 1, 0);
1792 xf_emit(ctx, 1, 1);
1793 xf_emit(ctx, 0x13, 0);
1794 xf_emit(ctx, 1, 0x10);
1795 xf_emit(ctx, 0x10, 0);
1796 xf_emit(ctx, 0x10, 0x3f800000);
1797 xf_emit(ctx, 0x19, 0);
1798 xf_emit(ctx, 1, 0x10);
1799 xf_emit(ctx, 1, 0);
1800 xf_emit(ctx, 1, 0x3f);
1801 xf_emit(ctx, 6, 0);
1802 xf_emit(ctx, 1, 1);
1803 xf_emit(ctx, 1, 0);
1804 xf_emit(ctx, 1, 1);
1805 xf_emit(ctx, 1, 0);
1806 xf_emit(ctx, 1, 1);
1807 if (dev_priv->chipset >= 0xa0) {
1808 xf_emit(ctx, 2, 0);
1809 xf_emit(ctx, 1, 0x1001);
1810 xf_emit(ctx, 0xb, 0);
1811 } else {
1812 xf_emit(ctx, 0xc, 0);
1813 }
1814 xf_emit(ctx, 1, 0x11);
1815 xf_emit(ctx, 7, 0);
1816 xf_emit(ctx, 1, 0xf);
1817 xf_emit(ctx, 7, 0);
1818 xf_emit(ctx, 1, 0x11);
1819 if (dev_priv->chipset == 0x50)
1820 xf_emit(ctx, 4, 0);
1821 else
1822 xf_emit(ctx, 6, 0);
1823 xf_emit(ctx, 3, 1);
1824 xf_emit(ctx, 1, 2);
1825 xf_emit(ctx, 1, 1);
1826 xf_emit(ctx, 1, 2);
1827 xf_emit(ctx, 1, 1);
1828 xf_emit(ctx, 1, 0);
1829 xf_emit(ctx, 1, magic2);
1830 xf_emit(ctx, 1, 0);
1831 xf_emit(ctx, 1, 0x0fac6881);
1832 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
1833 xf_emit(ctx, 1, 0);
1834 xf_emit(ctx, 0x18, 1);
1835 xf_emit(ctx, 8, 2);
1836 xf_emit(ctx, 8, 1);
1837 xf_emit(ctx, 8, 2);
1838 xf_emit(ctx, 8, 1);
1839 xf_emit(ctx, 3, 0);
1840 xf_emit(ctx, 1, 1);
1841 xf_emit(ctx, 5, 0);
1842 xf_emit(ctx, 1, 1);
1843 xf_emit(ctx, 0x16, 0);
1844 } else {
1845 if (dev_priv->chipset >= 0xa0)
1846 xf_emit(ctx, 0x1b, 0);
1847 else
1848 xf_emit(ctx, 0x15, 0);
1849 }
1850 xf_emit(ctx, 1, 1);
1851 xf_emit(ctx, 1, 2);
1852 xf_emit(ctx, 2, 1);
1853 xf_emit(ctx, 1, 2);
1854 xf_emit(ctx, 2, 1);
1855 if (dev_priv->chipset >= 0xa0)
1856 xf_emit(ctx, 4, 0);
1857 else
1858 xf_emit(ctx, 3, 0);
1859 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
1860 xf_emit(ctx, 0x10, 1);
1861 xf_emit(ctx, 8, 2);
1862 xf_emit(ctx, 0x10, 1);
1863 xf_emit(ctx, 8, 2);
1864 xf_emit(ctx, 8, 1);
1865 xf_emit(ctx, 3, 0);
1866 }
1867 xf_emit(ctx, 1, 0x11);
1868 xf_emit(ctx, 1, 1);
1869 xf_emit(ctx, 0x5b, 0);
1870}
1871
1872static void
1873nv50_graph_construct_xfer_tp_x1(struct nouveau_grctx *ctx)
1874{
1875 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1876 int magic3;
1877 if (dev_priv->chipset == 0x50)
1878 magic3 = 0x1000;
1879 else if (dev_priv->chipset == 0x86 || dev_priv->chipset == 0x98 || dev_priv->chipset >= 0xa8)
1880 magic3 = 0x1e00;
1881 else
1882 magic3 = 0;
1883 xf_emit(ctx, 1, 0);
1884 xf_emit(ctx, 1, 4);
1885 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1886 xf_emit(ctx, 0x24, 0);
1887 else if (dev_priv->chipset >= 0xa0)
1888 xf_emit(ctx, 0x14, 0);
1889 else
1890 xf_emit(ctx, 0x15, 0);
1891 xf_emit(ctx, 2, 4);
1892 if (dev_priv->chipset >= 0xa0)
1893 xf_emit(ctx, 1, 0x03020100);
1894 else
1895 xf_emit(ctx, 1, 0x00608080);
1896 xf_emit(ctx, 4, 0);
1897 xf_emit(ctx, 1, 4);
1898 xf_emit(ctx, 2, 0);
1899 xf_emit(ctx, 2, 4);
1900 xf_emit(ctx, 1, 0x80);
1901 if (magic3)
1902 xf_emit(ctx, 1, magic3);
1903 xf_emit(ctx, 1, 4);
1904 xf_emit(ctx, 0x24, 0);
1905 xf_emit(ctx, 1, 4);
1906 xf_emit(ctx, 1, 0x80);
1907 xf_emit(ctx, 1, 4);
1908 xf_emit(ctx, 1, 0x03020100);
1909 xf_emit(ctx, 1, 3);
1910 if (magic3)
1911 xf_emit(ctx, 1, magic3);
1912 xf_emit(ctx, 1, 4);
1913 xf_emit(ctx, 4, 0);
1914 xf_emit(ctx, 1, 4);
1915 xf_emit(ctx, 1, 3);
1916 xf_emit(ctx, 3, 0);
1917 xf_emit(ctx, 1, 4);
1918 if (dev_priv->chipset == 0x94 || dev_priv->chipset == 0x96)
1919 xf_emit(ctx, 0x1024, 0);
1920 else if (dev_priv->chipset < 0xa0)
1921 xf_emit(ctx, 0xa24, 0);
1922 else if (dev_priv->chipset == 0xa0 || dev_priv->chipset >= 0xaa)
1923 xf_emit(ctx, 0x214, 0);
1924 else
1925 xf_emit(ctx, 0x414, 0);
1926 xf_emit(ctx, 1, 4);
1927 xf_emit(ctx, 1, 3);
1928 xf_emit(ctx, 2, 0);
1929}
1930
1931static void
1932nv50_graph_construct_xfer_tp_x2(struct nouveau_grctx *ctx)
1933{
1934 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
1935 int magic1, magic2;
1936 if (dev_priv->chipset == 0x50) {
1937 magic1 = 0x3ff;
1938 magic2 = 0x00003e60;
1939 } else if (dev_priv->chipset <= 0xa0 || dev_priv->chipset >= 0xaa) {
1940 magic1 = 0x7ff;
1941 magic2 = 0x001ffe67;
1942 } else {
1943 magic1 = 0x7ff;
1944 magic2 = 0x00087e67;
1945 }
1946 xf_emit(ctx, 3, 0);
1947 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1948 xf_emit(ctx, 1, 1);
1949 xf_emit(ctx, 0xc, 0);
1950 xf_emit(ctx, 1, 0xf);
1951 xf_emit(ctx, 0xb, 0);
1952 xf_emit(ctx, 1, 4);
1953 xf_emit(ctx, 4, 0xffff);
1954 xf_emit(ctx, 8, 0);
1955 xf_emit(ctx, 1, 1);
1956 xf_emit(ctx, 3, 0);
1957 xf_emit(ctx, 1, 1);
1958 xf_emit(ctx, 5, 0);
1959 xf_emit(ctx, 1, 1);
1960 xf_emit(ctx, 2, 0);
1961 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
1962 xf_emit(ctx, 1, 3);
1963 xf_emit(ctx, 1, 0);
1964 } else if (dev_priv->chipset >= 0xa0)
1965 xf_emit(ctx, 1, 1);
1966 xf_emit(ctx, 0xa, 0);
1967 xf_emit(ctx, 2, 1);
1968 xf_emit(ctx, 1, 2);
1969 xf_emit(ctx, 2, 1);
1970 xf_emit(ctx, 1, 2);
1971 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
1972 xf_emit(ctx, 1, 0);
1973 xf_emit(ctx, 0x18, 1);
1974 xf_emit(ctx, 8, 2);
1975 xf_emit(ctx, 8, 1);
1976 xf_emit(ctx, 8, 2);
1977 xf_emit(ctx, 8, 1);
1978 xf_emit(ctx, 1, 0);
1979 }
1980 xf_emit(ctx, 1, 1);
1981 xf_emit(ctx, 1, 0);
1982 xf_emit(ctx, 1, 0x11);
1983 xf_emit(ctx, 7, 0);
1984 xf_emit(ctx, 1, 0x0fac6881);
1985 xf_emit(ctx, 2, 0);
1986 xf_emit(ctx, 1, 4);
1987 xf_emit(ctx, 3, 0);
1988 xf_emit(ctx, 1, 0x11);
1989 xf_emit(ctx, 1, 1);
1990 xf_emit(ctx, 1, 0);
1991 xf_emit(ctx, 3, 0xcf);
1992 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
1993 xf_emit(ctx, 1, 1);
1994 xf_emit(ctx, 0xa, 0);
1995 xf_emit(ctx, 2, 1);
1996 xf_emit(ctx, 1, 2);
1997 xf_emit(ctx, 2, 1);
1998 xf_emit(ctx, 1, 2);
1999 xf_emit(ctx, 1, 1);
2000 xf_emit(ctx, 1, 0);
2001 xf_emit(ctx, 8, 1);
2002 xf_emit(ctx, 1, 0x11);
2003 xf_emit(ctx, 7, 0);
2004 xf_emit(ctx, 1, 0x0fac6881);
2005 xf_emit(ctx, 1, 0xf);
2006 xf_emit(ctx, 7, 0);
2007 xf_emit(ctx, 1, magic2);
2008 xf_emit(ctx, 2, 0);
2009 xf_emit(ctx, 1, 0x11);
2010 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2011 xf_emit(ctx, 2, 1);
2012 else
2013 xf_emit(ctx, 1, 1);
2014 if(dev_priv->chipset == 0x50)
2015 xf_emit(ctx, 1, 0);
2016 else
2017 xf_emit(ctx, 3, 0);
2018 xf_emit(ctx, 1, 4);
2019 xf_emit(ctx, 5, 0);
2020 xf_emit(ctx, 1, 1);
2021 xf_emit(ctx, 4, 0);
2022 xf_emit(ctx, 1, 0x11);
2023 xf_emit(ctx, 7, 0);
2024 xf_emit(ctx, 1, 0x0fac6881);
2025 xf_emit(ctx, 3, 0);
2026 xf_emit(ctx, 1, 0x11);
2027 xf_emit(ctx, 1, 1);
2028 xf_emit(ctx, 1, 0);
2029 xf_emit(ctx, 1, 1);
2030 xf_emit(ctx, 1, 0);
2031 xf_emit(ctx, 1, 1);
2032 xf_emit(ctx, 1, 0);
2033 xf_emit(ctx, 1, magic1);
2034 xf_emit(ctx, 1, 0);
2035 xf_emit(ctx, 1, 1);
2036 xf_emit(ctx, 1, 0);
2037 xf_emit(ctx, 1, 1);
2038 xf_emit(ctx, 2, 0);
2039 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2040 xf_emit(ctx, 1, 1);
2041 xf_emit(ctx, 0x28, 0);
2042 xf_emit(ctx, 8, 8);
2043 xf_emit(ctx, 1, 0x11);
2044 xf_emit(ctx, 7, 0);
2045 xf_emit(ctx, 1, 0x0fac6881);
2046 xf_emit(ctx, 8, 0x400);
2047 xf_emit(ctx, 8, 0x300);
2048 xf_emit(ctx, 1, 1);
2049 xf_emit(ctx, 1, 0xf);
2050 xf_emit(ctx, 7, 0);
2051 xf_emit(ctx, 1, 0x20);
2052 xf_emit(ctx, 1, 0x11);
2053 xf_emit(ctx, 1, 0x100);
2054 xf_emit(ctx, 1, 0);
2055 xf_emit(ctx, 1, 1);
2056 xf_emit(ctx, 2, 0);
2057 xf_emit(ctx, 1, 0x40);
2058 xf_emit(ctx, 1, 0x100);
2059 xf_emit(ctx, 1, 0);
2060 xf_emit(ctx, 1, 3);
2061 xf_emit(ctx, 4, 0);
2062 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2063 xf_emit(ctx, 1, 1);
2064 xf_emit(ctx, 1, magic2);
2065 xf_emit(ctx, 3, 0);
2066 xf_emit(ctx, 1, 2);
2067 xf_emit(ctx, 1, 0x0fac6881);
2068 xf_emit(ctx, 9, 0);
2069 xf_emit(ctx, 1, 1);
2070 xf_emit(ctx, 4, 0);
2071 xf_emit(ctx, 1, 4);
2072 xf_emit(ctx, 1, 0);
2073 xf_emit(ctx, 1, 1);
2074 xf_emit(ctx, 1, 0x400);
2075 xf_emit(ctx, 1, 0x300);
2076 xf_emit(ctx, 1, 0x1001);
2077 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2078 xf_emit(ctx, 4, 0);
2079 else
2080 xf_emit(ctx, 3, 0);
2081 xf_emit(ctx, 1, 0x11);
2082 xf_emit(ctx, 7, 0);
2083 xf_emit(ctx, 1, 0x0fac6881);
2084 xf_emit(ctx, 1, 0xf);
2085 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
2086 xf_emit(ctx, 0x15, 0);
2087 xf_emit(ctx, 1, 1);
2088 xf_emit(ctx, 3, 0);
2089 } else
2090 xf_emit(ctx, 0x17, 0);
2091 if (dev_priv->chipset >= 0xa0)
2092 xf_emit(ctx, 1, 0x0fac6881);
2093 xf_emit(ctx, 1, magic2);
2094 xf_emit(ctx, 3, 0);
2095 xf_emit(ctx, 1, 0x11);
2096 xf_emit(ctx, 2, 0);
2097 xf_emit(ctx, 1, 4);
2098 xf_emit(ctx, 1, 0);
2099 xf_emit(ctx, 2, 1);
2100 xf_emit(ctx, 3, 0);
2101 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2102 xf_emit(ctx, 2, 1);
2103 else
2104 xf_emit(ctx, 1, 1);
2105 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2106 xf_emit(ctx, 2, 0);
2107 else if (dev_priv->chipset != 0x50)
2108 xf_emit(ctx, 1, 0);
2109}
2110
2111static void
2112nv50_graph_construct_xfer_tp_x3(struct nouveau_grctx *ctx)
2113{
2114 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
2115 xf_emit(ctx, 3, 0);
2116 xf_emit(ctx, 1, 1);
2117 xf_emit(ctx, 1, 0);
2118 xf_emit(ctx, 1, 1);
2119 if (dev_priv->chipset == 0x50)
2120 xf_emit(ctx, 2, 0);
2121 else
2122 xf_emit(ctx, 3, 0);
2123 xf_emit(ctx, 1, 0x2a712488);
2124 xf_emit(ctx, 1, 0);
2125 xf_emit(ctx, 1, 0x4085c000);
2126 xf_emit(ctx, 1, 0x40);
2127 xf_emit(ctx, 1, 0x100);
2128 xf_emit(ctx, 1, 0x10100);
2129 xf_emit(ctx, 1, 0x02800000);
2130}
2131
2132static void
2133nv50_graph_construct_xfer_tp_x4(struct nouveau_grctx *ctx)
2134{
2135 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
2136 xf_emit(ctx, 2, 0x04e3bfdf);
2137 xf_emit(ctx, 1, 1);
2138 xf_emit(ctx, 1, 0);
2139 xf_emit(ctx, 1, 0x00ffff00);
2140 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2141 xf_emit(ctx, 2, 1);
2142 else
2143 xf_emit(ctx, 1, 1);
2144 xf_emit(ctx, 2, 0);
2145 xf_emit(ctx, 1, 0x00ffff00);
2146 xf_emit(ctx, 8, 0);
2147 xf_emit(ctx, 1, 1);
2148 xf_emit(ctx, 1, 0);
2149 xf_emit(ctx, 1, 1);
2150 xf_emit(ctx, 1, 0x30201000);
2151 xf_emit(ctx, 1, 0x70605040);
2152 xf_emit(ctx, 1, 0xb8a89888);
2153 xf_emit(ctx, 1, 0xf8e8d8c8);
2154 xf_emit(ctx, 1, 0);
2155 xf_emit(ctx, 1, 0x1a);
2156}
2157
2158static void
2159nv50_graph_construct_xfer_tp_x5(struct nouveau_grctx *ctx)
2160{
2161 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
2162 xf_emit(ctx, 3, 0);
2163 xf_emit(ctx, 1, 0xfac6881);
2164 xf_emit(ctx, 4, 0);
2165 xf_emit(ctx, 1, 4);
2166 xf_emit(ctx, 1, 0);
2167 xf_emit(ctx, 2, 1);
2168 xf_emit(ctx, 2, 0);
2169 xf_emit(ctx, 1, 1);
2170 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2171 xf_emit(ctx, 0xb, 0);
2172 else
2173 xf_emit(ctx, 0xa, 0);
2174 xf_emit(ctx, 8, 1);
2175 xf_emit(ctx, 1, 0x11);
2176 xf_emit(ctx, 7, 0);
2177 xf_emit(ctx, 1, 0xfac6881);
2178 xf_emit(ctx, 1, 0xf);
2179 xf_emit(ctx, 7, 0);
2180 xf_emit(ctx, 1, 0x11);
2181 xf_emit(ctx, 1, 1);
2182 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
2183 xf_emit(ctx, 6, 0);
2184 xf_emit(ctx, 1, 1);
2185 xf_emit(ctx, 6, 0);
2186 } else {
2187 xf_emit(ctx, 0xb, 0);
2188 }
2189}
2190
2191static void
2192nv50_graph_construct_xfer_tp(struct nouveau_grctx *ctx)
2193{
2194 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
2195 if (dev_priv->chipset < 0xa0) {
2196 nv50_graph_construct_xfer_tp_x1(ctx);
2197 nv50_graph_construct_xfer_tp_x2(ctx);
2198 nv50_graph_construct_xfer_tp_x3(ctx);
2199 if (dev_priv->chipset == 0x50)
2200 xf_emit(ctx, 0xf, 0);
2201 else
2202 xf_emit(ctx, 0x12, 0);
2203 nv50_graph_construct_xfer_tp_x4(ctx);
2204 } else {
2205 nv50_graph_construct_xfer_tp_x3(ctx);
2206 if (dev_priv->chipset < 0xaa)
2207 xf_emit(ctx, 0xc, 0);
2208 else
2209 xf_emit(ctx, 0xa, 0);
2210 nv50_graph_construct_xfer_tp_x2(ctx);
2211 nv50_graph_construct_xfer_tp_x5(ctx);
2212 nv50_graph_construct_xfer_tp_x4(ctx);
2213 nv50_graph_construct_xfer_tp_x1(ctx);
2214 }
2215}
2216
2217static void
2218nv50_graph_construct_xfer_tp2(struct nouveau_grctx *ctx)
2219{
2220 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
2221 int i, mpcnt;
2222 if (dev_priv->chipset == 0x98 || dev_priv->chipset == 0xaa)
2223 mpcnt = 1;
2224 else if (dev_priv->chipset < 0xa0 || dev_priv->chipset >= 0xa8)
2225 mpcnt = 2;
2226 else
2227 mpcnt = 3;
2228 for (i = 0; i < mpcnt; i++) {
2229 xf_emit(ctx, 1, 0);
2230 xf_emit(ctx, 1, 0x80);
2231 xf_emit(ctx, 1, 0x80007004);
2232 xf_emit(ctx, 1, 0x04000400);
2233 if (dev_priv->chipset >= 0xa0)
2234 xf_emit(ctx, 1, 0xc0);
2235 xf_emit(ctx, 1, 0x1000);
2236 xf_emit(ctx, 2, 0);
2237 if (dev_priv->chipset == 0x86 || dev_priv->chipset == 0x98 || dev_priv->chipset >= 0xa8) {
2238 xf_emit(ctx, 1, 0xe00);
2239 xf_emit(ctx, 1, 0x1e00);
2240 }
2241 xf_emit(ctx, 1, 1);
2242 xf_emit(ctx, 2, 0);
2243 if (dev_priv->chipset == 0x50)
2244 xf_emit(ctx, 2, 0x1000);
2245 xf_emit(ctx, 1, 1);
2246 xf_emit(ctx, 1, 0);
2247 xf_emit(ctx, 1, 4);
2248 xf_emit(ctx, 1, 2);
2249 if (dev_priv->chipset >= 0xaa)
2250 xf_emit(ctx, 0xb, 0);
2251 else if (dev_priv->chipset >= 0xa0)
2252 xf_emit(ctx, 0xc, 0);
2253 else
2254 xf_emit(ctx, 0xa, 0);
2255 }
2256 xf_emit(ctx, 1, 0x08100c12);
2257 xf_emit(ctx, 1, 0);
2258 if (dev_priv->chipset >= 0xa0) {
2259 xf_emit(ctx, 1, 0x1fe21);
2260 }
2261 xf_emit(ctx, 5, 0);
2262 xf_emit(ctx, 4, 0xffff);
2263 xf_emit(ctx, 1, 1);
2264 xf_emit(ctx, 2, 0x10001);
2265 xf_emit(ctx, 1, 1);
2266 xf_emit(ctx, 1, 0);
2267 xf_emit(ctx, 1, 0x1fe21);
2268 xf_emit(ctx, 1, 0);
2269 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2270 xf_emit(ctx, 1, 1);
2271 xf_emit(ctx, 4, 0);
2272 xf_emit(ctx, 1, 0x08100c12);
2273 xf_emit(ctx, 1, 4);
2274 xf_emit(ctx, 1, 0);
2275 xf_emit(ctx, 1, 2);
2276 xf_emit(ctx, 1, 0x11);
2277 xf_emit(ctx, 8, 0);
2278 xf_emit(ctx, 1, 0xfac6881);
2279 xf_emit(ctx, 1, 0);
2280 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa)
2281 xf_emit(ctx, 1, 3);
2282 xf_emit(ctx, 3, 0);
2283 xf_emit(ctx, 1, 4);
2284 xf_emit(ctx, 9, 0);
2285 xf_emit(ctx, 1, 2);
2286 xf_emit(ctx, 2, 1);
2287 xf_emit(ctx, 1, 2);
2288 xf_emit(ctx, 3, 1);
2289 xf_emit(ctx, 1, 0);
2290 if (dev_priv->chipset > 0xa0 && dev_priv->chipset < 0xaa) {
2291 xf_emit(ctx, 8, 2);
2292 xf_emit(ctx, 0x10, 1);
2293 xf_emit(ctx, 8, 2);
2294 xf_emit(ctx, 0x18, 1);
2295 xf_emit(ctx, 3, 0);
2296 }
2297 xf_emit(ctx, 1, 4);
2298 if (dev_priv->chipset == 0x50)
2299 xf_emit(ctx, 0x3a0, 0);
2300 else if (dev_priv->chipset < 0x94)
2301 xf_emit(ctx, 0x3a2, 0);
2302 else if (dev_priv->chipset == 0x98 || dev_priv->chipset == 0xaa)
2303 xf_emit(ctx, 0x39f, 0);
2304 else
2305 xf_emit(ctx, 0x3a3, 0);
2306 xf_emit(ctx, 1, 0x11);
2307 xf_emit(ctx, 1, 0);
2308 xf_emit(ctx, 1, 1);
2309 xf_emit(ctx, 0x2d, 0);
2310}
2311
2312static void
2313nv50_graph_construct_xfer2(struct nouveau_grctx *ctx)
2314{
2315 struct drm_nouveau_private *dev_priv = ctx->dev->dev_private;
2316 int i;
2317 uint32_t offset;
2318 uint32_t units = nv_rd32 (ctx->dev, 0x1540);
2319 int size = 0;
2320
2321 offset = (ctx->ctxvals_pos+0x3f)&~0x3f;
2322
2323 if (dev_priv->chipset < 0xa0) {
2324 for (i = 0; i < 8; i++) {
2325 ctx->ctxvals_pos = offset + i;
2326 if (i == 0)
2327 xf_emit(ctx, 1, 0x08100c12);
2328 if (units & (1 << i))
2329 nv50_graph_construct_xfer_tp2(ctx);
2330 if ((ctx->ctxvals_pos-offset)/8 > size)
2331 size = (ctx->ctxvals_pos-offset)/8;
2332 }
2333 } else {
2334 /* Strand 0: TPs 0, 1 */
2335 ctx->ctxvals_pos = offset;
2336 xf_emit(ctx, 1, 0x08100c12);
2337 if (units & (1 << 0))
2338 nv50_graph_construct_xfer_tp2(ctx);
2339 if (units & (1 << 1))
2340 nv50_graph_construct_xfer_tp2(ctx);
2341 if ((ctx->ctxvals_pos-offset)/8 > size)
2342 size = (ctx->ctxvals_pos-offset)/8;
2343
2344 /* Strand 0: TPs 2, 3 */
2345 ctx->ctxvals_pos = offset + 1;
2346 if (units & (1 << 2))
2347 nv50_graph_construct_xfer_tp2(ctx);
2348 if (units & (1 << 3))
2349 nv50_graph_construct_xfer_tp2(ctx);
2350 if ((ctx->ctxvals_pos-offset)/8 > size)
2351 size = (ctx->ctxvals_pos-offset)/8;
2352
2353 /* Strand 0: TPs 4, 5, 6 */
2354 ctx->ctxvals_pos = offset + 2;
2355 if (units & (1 << 4))
2356 nv50_graph_construct_xfer_tp2(ctx);
2357 if (units & (1 << 5))
2358 nv50_graph_construct_xfer_tp2(ctx);
2359 if (units & (1 << 6))
2360 nv50_graph_construct_xfer_tp2(ctx);
2361 if ((ctx->ctxvals_pos-offset)/8 > size)
2362 size = (ctx->ctxvals_pos-offset)/8;
2363
2364 /* Strand 0: TPs 7, 8, 9 */
2365 ctx->ctxvals_pos = offset + 3;
2366 if (units & (1 << 7))
2367 nv50_graph_construct_xfer_tp2(ctx);
2368 if (units & (1 << 8))
2369 nv50_graph_construct_xfer_tp2(ctx);
2370 if (units & (1 << 9))
2371 nv50_graph_construct_xfer_tp2(ctx);
2372 if ((ctx->ctxvals_pos-offset)/8 > size)
2373 size = (ctx->ctxvals_pos-offset)/8;
2374 }
2375 ctx->ctxvals_pos = offset + size * 8;
2376 ctx->ctxvals_pos = (ctx->ctxvals_pos+0x3f)&~0x3f;
2377 cp_lsr (ctx, offset);
2378 cp_out (ctx, CP_SET_XFER_POINTER);
2379 cp_lsr (ctx, size);
2380 cp_out (ctx, CP_SEEK_2);
2381 cp_out (ctx, CP_XFER_2);
2382 cp_wait(ctx, XFER, BUSY);
2383}
generated by cgit v1.2.2 (git 2.25.0) at 2025-07-24 17:46:24 -0400