projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
lima: remove lima_ctx_buff_va submit flags (v2)
[mesa.git] / src / gallium / state_trackers / nine / nine_state.c
1 /*
2 * Copyright 2011 Joakim Sindholt <opensource@zhasha.com>
3 * Copyright 2013 Christoph Bumiller
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE. */
23
24 #define NINE_STATE
25
26 #include "device9.h"
27 #include "swapchain9.h"
28 #include "basetexture9.h"
29 #include "buffer9.h"
30 #include "indexbuffer9.h"
31 #include "surface9.h"
32 #include "vertexbuffer9.h"
33 #include "vertexdeclaration9.h"
34 #include "vertexshader9.h"
35 #include "pixelshader9.h"
36 #include "nine_pipe.h"
37 #include "nine_ff.h"
38 #include "nine_limits.h"
39 #include "pipe/p_context.h"
40 #include "pipe/p_state.h"
41 #include "cso_cache/cso_context.h"
42 #include "util/u_atomic.h"
43 #include "util/u_upload_mgr.h"
44 #include "util/u_math.h"
45 #include "util/u_box.h"
46 #include "util/u_simple_shaders.h"
47 #include "util/u_gen_mipmap.h"
48
49 /* CSMT headers */
50 #include "nine_queue.h"
51 #include "nine_csmt_helper.h"
52 #include "os/os_thread.h"
53
54 #define DBG_CHANNEL DBG_DEVICE
55
56 /* Nine CSMT */
57
58 struct csmt_instruction {
59 int (* func)(struct NineDevice9 *This, struct csmt_instruction *instr);
60 };
61
62 struct csmt_context {
63 thrd_t worker;
64 struct nine_queue_pool* pool;
65 BOOL terminate;
66 cnd_t event_processed;
67 mtx_t mutex_processed;
68 struct NineDevice9 *device;
69 BOOL processed;
70 BOOL toPause;
71 BOOL hasPaused;
72 mtx_t thread_running;
73 mtx_t thread_resume;
74 };
75
76 /* Wait for instruction to be processed.
77 * Caller has to ensure that only one thread waits at time.
78 */
79 static void
80 nine_csmt_wait_processed(struct csmt_context *ctx)
81 {
82 mtx_lock(&ctx->mutex_processed);
83 while (!p_atomic_read(&ctx->processed)) {
84 cnd_wait(&ctx->event_processed, &ctx->mutex_processed);
85 }
86 mtx_unlock(&ctx->mutex_processed);
87 }
88
89 /* CSMT worker thread */
90 static
91 int
92 nine_csmt_worker(void *arg)
93 {
94 struct csmt_context *ctx = arg;
95 struct csmt_instruction *instr;
96 DBG("CSMT worker spawned\n");
97
98 u_thread_setname("CSMT-Worker");
99
100 while (1) {
101 nine_queue_wait_flush(ctx->pool);
102 mtx_lock(&ctx->thread_running);
103
104 /* Get instruction. NULL on empty cmdbuf. */
105 while (!p_atomic_read(&ctx->terminate) &&
106 (instr = (struct csmt_instruction *)nine_queue_get(ctx->pool))) {
107
108 /* decode */
109 if (instr->func(ctx->device, instr)) {
110 mtx_lock(&ctx->mutex_processed);
111 p_atomic_set(&ctx->processed, TRUE);
112 cnd_signal(&ctx->event_processed);
113 mtx_unlock(&ctx->mutex_processed);
114 }
115 if (p_atomic_read(&ctx->toPause)) {
116 mtx_unlock(&ctx->thread_running);
117 /* will wait here the thread can be resumed */
118 mtx_lock(&ctx->thread_resume);
119 mtx_lock(&ctx->thread_running);
120 mtx_unlock(&ctx->thread_resume);
121 }
122 }
123
124 mtx_unlock(&ctx->thread_running);
125 if (p_atomic_read(&ctx->terminate)) {
126 mtx_lock(&ctx->mutex_processed);
127 p_atomic_set(&ctx->processed, TRUE);
128 cnd_signal(&ctx->event_processed);
129 mtx_unlock(&ctx->mutex_processed);
130 break;
131 }
132 }
133
134 DBG("CSMT worker destroyed\n");
135 return 0;
136 }
137
138 /* Create a CSMT context.
139 * Spawns a worker thread.
140 */
141 struct csmt_context *
142 nine_csmt_create( struct NineDevice9 *This )
143 {
144 struct csmt_context *ctx;
145
146 ctx = CALLOC_STRUCT(csmt_context);
147 if (!ctx)
148 return NULL;
149
150 ctx->pool = nine_queue_create();
151 if (!ctx->pool) {
152 FREE(ctx);
153 return NULL;
154 }
155 cnd_init(&ctx->event_processed);
156 (void) mtx_init(&ctx->mutex_processed, mtx_plain);
157 (void) mtx_init(&ctx->thread_running, mtx_plain);
158 (void) mtx_init(&ctx->thread_resume, mtx_plain);
159
160 #if defined(DEBUG) || !defined(NDEBUG)
161 u_thread_setname("Main thread");
162 #endif
163
164 ctx->device = This;
165
166 ctx->worker = u_thread_create(nine_csmt_worker, ctx);
167 if (!ctx->worker) {
168 nine_queue_delete(ctx->pool);
169 FREE(ctx);
170 return NULL;
171 }
172
173 DBG("Returning context %p\n", ctx);
174
175 return ctx;
176 }
177
178 static int
179 nop_func( struct NineDevice9 *This, struct csmt_instruction *instr )
180 {
181 (void) This;
182 (void) instr;
183
184 return 1;
185 }
186
187 /* Push nop instruction and flush the queue.
188 * Waits for the worker to complete. */
189 void
190 nine_csmt_process( struct NineDevice9 *device )
191 {
192 struct csmt_instruction* instr;
193 struct csmt_context *ctx = device->csmt_ctx;
194
195 if (!device->csmt_active)
196 return;
197
198 if (nine_queue_isempty(ctx->pool))
199 return;
200
201 DBG("device=%p\n", device);
202
203 /* NOP */
204 instr = nine_queue_alloc(ctx->pool, sizeof(struct csmt_instruction));
205 assert(instr);
206 instr->func = nop_func;
207
208 p_atomic_set(&ctx->processed, FALSE);
209 nine_queue_flush(ctx->pool);
210
211 nine_csmt_wait_processed(ctx);
212 }
213
214 /* Destroys a CSMT context.
215 * Waits for the worker thread to terminate.
216 */
217 void
218 nine_csmt_destroy( struct NineDevice9 *device, struct csmt_context *ctx )
219 {
220 struct csmt_instruction* instr;
221 thrd_t render_thread = ctx->worker;
222
223 DBG("device=%p ctx=%p\n", device, ctx);
224
225 /* Push nop and flush the queue. */
226 instr = nine_queue_alloc(ctx->pool, sizeof(struct csmt_instruction));
227 assert(instr);
228 instr->func = nop_func;
229
230 p_atomic_set(&ctx->processed, FALSE);
231 /* Signal worker to terminate. */
232 p_atomic_set(&ctx->terminate, TRUE);
233 nine_queue_flush(ctx->pool);
234
235 nine_csmt_wait_processed(ctx);
236 nine_queue_delete(ctx->pool);
237
238 mtx_destroy(&ctx->thread_resume);
239 mtx_destroy(&ctx->thread_running);
240
241 mtx_destroy(&ctx->mutex_processed);
242 cnd_destroy(&ctx->event_processed);
243
244 FREE(ctx);
245
246 thrd_join(render_thread, NULL);
247 }
248
249 static void
250 nine_csmt_pause( struct NineDevice9 *device )
251 {
252 struct csmt_context *ctx = device->csmt_ctx;
253
254 if (!device->csmt_active)
255 return;
256
257 /* No need to pause the thread */
258 if (nine_queue_no_flushed_work(ctx->pool))
259 return;
260
261 mtx_lock(&ctx->thread_resume);
262 p_atomic_set(&ctx->toPause, TRUE);
263
264 /* Wait the thread is paused */
265 mtx_lock(&ctx->thread_running);
266 ctx->hasPaused = TRUE;
267 p_atomic_set(&ctx->toPause, FALSE);
268 }
269
270 static void
271 nine_csmt_resume( struct NineDevice9 *device )
272 {
273 struct csmt_context *ctx = device->csmt_ctx;
274
275 if (!device->csmt_active)
276 return;
277
278 if (!ctx->hasPaused)
279 return;
280
281 ctx->hasPaused = FALSE;
282 mtx_unlock(&ctx->thread_running);
283 mtx_unlock(&ctx->thread_resume);
284 }
285
286 struct pipe_context *
287 nine_context_get_pipe( struct NineDevice9 *device )
288 {
289 nine_csmt_process(device);
290 return device->context.pipe;
291 }
292
293 struct pipe_context *
294 nine_context_get_pipe_multithread( struct NineDevice9 *device )
295 {
296 struct csmt_context *ctx = device->csmt_ctx;
297
298 if (!device->csmt_active)
299 return device->context.pipe;
300
301 if (!u_thread_is_self(ctx->worker))
302 nine_csmt_process(device);
303
304 return device->context.pipe;
305 }
306
307 struct pipe_context *
308 nine_context_get_pipe_acquire( struct NineDevice9 *device )
309 {
310 nine_csmt_pause(device);
311 return device->context.pipe;
312 }
313
314 void
315 nine_context_get_pipe_release( struct NineDevice9 *device )
316 {
317 nine_csmt_resume(device);
318 }
319
320 /* Nine state functions */
321
322 /* Check if some states need to be set dirty */
323
324 static inline DWORD
325 check_multisample(struct NineDevice9 *device)
326 {
327 DWORD *rs = device->context.rs;
328 DWORD new_value = (rs[D3DRS_ZENABLE] || rs[D3DRS_STENCILENABLE]) &&
329 device->context.rt[0]->desc.MultiSampleType >= 1 &&
330 rs[D3DRS_MULTISAMPLEANTIALIAS];
331 if (rs[NINED3DRS_MULTISAMPLE] != new_value) {
332 rs[NINED3DRS_MULTISAMPLE] = new_value;
333 return NINE_STATE_RASTERIZER;
334 }
335 return 0;
336 }
337
338 /* State preparation only */
339
340 static inline void
341 prepare_blend(struct NineDevice9 *device)
342 {
343 nine_convert_blend_state(&device->context.pipe_data.blend, device->context.rs);
344 device->context.commit |= NINE_STATE_COMMIT_BLEND;
345 }
346
347 static inline void
348 prepare_dsa(struct NineDevice9 *device)
349 {
350 nine_convert_dsa_state(&device->context.pipe_data.dsa, device->context.rs);
351 device->context.commit |= NINE_STATE_COMMIT_DSA;
352 }
353
354 static inline void
355 prepare_rasterizer(struct NineDevice9 *device)
356 {
357 nine_convert_rasterizer_state(device, &device->context.pipe_data.rast, device->context.rs);
358 device->context.commit |= NINE_STATE_COMMIT_RASTERIZER;
359 }
360
361 static void
362 prepare_vs_constants_userbuf_swvp(struct NineDevice9 *device)
363 {
364 struct nine_context *context = &device->context;
365
366 if (context->changed.vs_const_f || context->changed.group & NINE_STATE_SWVP) {
367 struct pipe_constant_buffer cb;
368
369 cb.buffer_offset = 0;
370 cb.buffer_size = 4096 * sizeof(float[4]);
371 cb.user_buffer = context->vs_const_f_swvp;
372
373 if (context->vs->lconstf.ranges) {
374 const struct nine_lconstf *lconstf = &(context->vs->lconstf);
375 const struct nine_range *r = lconstf->ranges;
376 unsigned n = 0;
377 float *dst = context->vs_lconstf_temp;
378 float *src = (float *)cb.user_buffer;
379 memcpy(dst, src, cb.buffer_size);
380 while (r) {
381 unsigned p = r->bgn;
382 unsigned c = r->end - r->bgn;
383 memcpy(&dst[p * 4], &lconstf->data[n * 4], c * 4 * sizeof(float));
384 n += c;
385 r = r->next;
386 }
387 cb.user_buffer = dst;
388 }
389
390 context->pipe_data.cb0_swvp.buffer_offset = cb.buffer_offset;
391 context->pipe_data.cb0_swvp.buffer_size = cb.buffer_size;
392 context->pipe_data.cb0_swvp.user_buffer = cb.user_buffer;
393
394 cb.user_buffer = (char *)cb.user_buffer + 4096 * sizeof(float[4]);
395 context->pipe_data.cb1_swvp.buffer_offset = cb.buffer_offset;
396 context->pipe_data.cb1_swvp.buffer_size = cb.buffer_size;
397 context->pipe_data.cb1_swvp.user_buffer = cb.user_buffer;
398
399 context->changed.vs_const_f = 0;
400 }
401
402 if (context->changed.vs_const_i || context->changed.group & NINE_STATE_SWVP) {
403 struct pipe_constant_buffer cb;
404
405 cb.buffer_offset = 0;
406 cb.buffer_size = 2048 * sizeof(float[4]);
407 cb.user_buffer = context->vs_const_i;
408
409 context->pipe_data.cb2_swvp.buffer_offset = cb.buffer_offset;
410 context->pipe_data.cb2_swvp.buffer_size = cb.buffer_size;
411 context->pipe_data.cb2_swvp.user_buffer = cb.user_buffer;
412 context->changed.vs_const_i = 0;
413 }
414
415 if (context->changed.vs_const_b || context->changed.group & NINE_STATE_SWVP) {
416 struct pipe_constant_buffer cb;
417
418 cb.buffer_offset = 0;
419 cb.buffer_size = 512 * sizeof(float[4]);
420 cb.user_buffer = context->vs_const_b;
421
422 context->pipe_data.cb3_swvp.buffer_offset = cb.buffer_offset;
423 context->pipe_data.cb3_swvp.buffer_size = cb.buffer_size;
424 context->pipe_data.cb3_swvp.user_buffer = cb.user_buffer;
425 context->changed.vs_const_b = 0;
426 }
427
428 context->changed.group &= ~NINE_STATE_VS_CONST;
429 context->commit |= NINE_STATE_COMMIT_CONST_VS;
430 }
431
432 static void
433 prepare_vs_constants_userbuf(struct NineDevice9 *device)
434 {
435 struct nine_context *context = &device->context;
436 uint8_t *upload_ptr = NULL;
437 struct pipe_constant_buffer cb;
438 cb.buffer = NULL;
439 cb.buffer_offset = 0;
440 cb.buffer_size = context->cso_shader.vs_const_used_size;
441 cb.user_buffer = context->vs_const_f;
442
443 if (context->swvp) {
444 prepare_vs_constants_userbuf_swvp(device);
445 return;
446 }
447
448 if (context->changed.vs_const_i || context->changed.group & NINE_STATE_SWVP) {
449 int *idst = (int *)&context->vs_const_f[4 * device->max_vs_const_f];
450 memcpy(idst, context->vs_const_i, NINE_MAX_CONST_I * sizeof(int[4]));
451 context->changed.vs_const_i = 0;
452 }
453
454 if (context->changed.vs_const_b || context->changed.group & NINE_STATE_SWVP) {
455 int *idst = (int *)&context->vs_const_f[4 * device->max_vs_const_f];
456 uint32_t *bdst = (uint32_t *)&idst[4 * NINE_MAX_CONST_I];
457 memcpy(bdst, context->vs_const_b, NINE_MAX_CONST_B * sizeof(BOOL));
458 context->changed.vs_const_b = 0;
459 }
460
461 if (!cb.buffer_size)
462 return;
463
464 if (context->vs->lconstf.ranges) {
465 /* TODO: Can we make it so that we don't have to copy everything ? */
466 const struct nine_lconstf *lconstf = &(context->vs->lconstf);
467 const struct nine_range *r = lconstf->ranges;
468 unsigned n = 0;
469 float *dst = context->vs_lconstf_temp;
470 float *src = (float *)cb.user_buffer;
471 memcpy(dst, src, cb.buffer_size);
472 while (r) {
473 unsigned p = r->bgn;
474 unsigned c = r->end - r->bgn;
475 memcpy(&dst[p * 4], &lconstf->data[n * 4], c * 4 * sizeof(float));
476 n += c;
477 r = r->next;
478 }
479 cb.user_buffer = dst;
480 }
481
482 /* Note: We probably don't want to do separate memcpy to
483 * upload_ptr directly, if we have to copy some constants
484 * at random locations (context->vs->lconstf.ranges),
485 * to have efficient WC. Thus for this case we really want
486 * that intermediate buffer. */
487
488 u_upload_alloc(context->pipe->const_uploader,
489 0,
490 cb.buffer_size,
491 256, /* Be conservative about alignment */
492 &(cb.buffer_offset),
493 &(cb.buffer),
494 (void**)&upload_ptr);
495
496 assert(cb.buffer && upload_ptr);
497
498 if (!context->cso_shader.vs_const_ranges) {
499 memcpy(upload_ptr, cb.user_buffer, cb.buffer_size);
500 } else {
501 unsigned i = 0;
502 unsigned offset = 0;
503 while (context->cso_shader.vs_const_ranges[i*2+1] != 0) {
504 memcpy(upload_ptr+offset,
505 &((float*)cb.user_buffer)[4*context->cso_shader.vs_const_ranges[i*2]],
506 context->cso_shader.vs_const_ranges[i*2+1] * sizeof(float[4]));
507 offset += context->cso_shader.vs_const_ranges[i*2+1] * sizeof(float[4]);
508 i++;
509 }
510 }
511
512 u_upload_unmap(context->pipe->const_uploader);
513 cb.user_buffer = NULL;
514
515 /* Free previous resource */
516 pipe_resource_reference(&context->pipe_data.cb_vs.buffer, NULL);
517
518 context->pipe_data.cb_vs = cb;
519 context->changed.vs_const_f = 0;
520
521 context->changed.group &= ~NINE_STATE_VS_CONST;
522 context->commit |= NINE_STATE_COMMIT_CONST_VS;
523 }
524
525 static void
526 prepare_ps_constants_userbuf(struct NineDevice9 *device)
527 {
528 struct nine_context *context = &device->context;
529 uint8_t *upload_ptr = NULL;
530 struct pipe_constant_buffer cb;
531 cb.buffer = NULL;
532 cb.buffer_offset = 0;
533 cb.buffer_size = context->cso_shader.ps_const_used_size;
534 cb.user_buffer = context->ps_const_f;
535
536 if (context->changed.ps_const_i) {
537 int *idst = (int *)&context->ps_const_f[4 * device->max_ps_const_f];
538 memcpy(idst, context->ps_const_i, sizeof(context->ps_const_i));
539 context->changed.ps_const_i = 0;
540 }
541 if (context->changed.ps_const_b) {
542 int *idst = (int *)&context->ps_const_f[4 * device->max_ps_const_f];
543 uint32_t *bdst = (uint32_t *)&idst[4 * NINE_MAX_CONST_I];
544 memcpy(bdst, context->ps_const_b, sizeof(context->ps_const_b));
545 context->changed.ps_const_b = 0;
546 }
547
548 /* Upload special constants needed to implement PS1.x instructions like TEXBEM,TEXBEML and BEM */
549 if (context->ps->bumpenvmat_needed) {
550 memcpy(context->ps_lconstf_temp, cb.user_buffer, 8 * sizeof(float[4]));
551 memcpy(&context->ps_lconstf_temp[4 * 8], &device->context.bumpmap_vars, sizeof(device->context.bumpmap_vars));
552
553 cb.user_buffer = context->ps_lconstf_temp;
554 }
555
556 if (context->ps->byte_code.version < 0x30 &&
557 context->rs[D3DRS_FOGENABLE]) {
558 float *dst = &context->ps_lconstf_temp[4 * 32];
559 if (cb.user_buffer != context->ps_lconstf_temp) {
560 memcpy(context->ps_lconstf_temp, cb.user_buffer, 32 * sizeof(float[4]));
561 cb.user_buffer = context->ps_lconstf_temp;
562 }
563
564 d3dcolor_to_rgba(dst, context->rs[D3DRS_FOGCOLOR]);
565 if (context->rs[D3DRS_FOGTABLEMODE] == D3DFOG_LINEAR) {
566 dst[4] = asfloat(context->rs[D3DRS_FOGEND]);
567 dst[5] = 1.0f / (asfloat(context->rs[D3DRS_FOGEND]) - asfloat(context->rs[D3DRS_FOGSTART]));
568 } else if (context->rs[D3DRS_FOGTABLEMODE] != D3DFOG_NONE) {
569 dst[4] = asfloat(context->rs[D3DRS_FOGDENSITY]);
570 }
571 }
572
573 if (!cb.buffer_size)
574 return;
575
576 u_upload_alloc(context->pipe->const_uploader,
577 0,
578 cb.buffer_size,
579 256, /* Be conservative about alignment */
580 &(cb.buffer_offset),
581 &(cb.buffer),
582 (void**)&upload_ptr);
583
584 assert(cb.buffer && upload_ptr);
585
586 if (!context->cso_shader.ps_const_ranges) {
587 memcpy(upload_ptr, cb.user_buffer, cb.buffer_size);
588 } else {
589 unsigned i = 0;
590 unsigned offset = 0;
591 while (context->cso_shader.ps_const_ranges[i*2+1] != 0) {
592 memcpy(upload_ptr+offset,
593 &((float*)cb.user_buffer)[4*context->cso_shader.ps_const_ranges[i*2]],
594 context->cso_shader.ps_const_ranges[i*2+1] * sizeof(float[4]));
595 offset += context->cso_shader.ps_const_ranges[i*2+1] * sizeof(float[4]);
596 i++;
597 }
598 }
599
600 u_upload_unmap(context->pipe->const_uploader);
601 cb.user_buffer = NULL;
602
603 /* Free previous resource */
604 pipe_resource_reference(&context->pipe_data.cb_ps.buffer, NULL);
605
606 context->pipe_data.cb_ps = cb;
607 context->changed.ps_const_f = 0;
608
609 context->changed.group &= ~NINE_STATE_PS_CONST;
610 context->commit |= NINE_STATE_COMMIT_CONST_PS;
611 }
612
613 static inline uint32_t
614 prepare_vs(struct NineDevice9 *device, uint8_t shader_changed)
615 {
616 struct nine_context *context = &device->context;
617 struct NineVertexShader9 *vs = context->vs;
618 uint32_t changed_group = 0;
619 int has_key_changed = 0;
620
621 if (likely(context->programmable_vs))
622 has_key_changed = NineVertexShader9_UpdateKey(vs, device);
623
624 if (!shader_changed && !has_key_changed)
625 return 0;
626
627 /* likely because we dislike FF */
628 if (likely(context->programmable_vs)) {
629 context->cso_shader.vs = NineVertexShader9_GetVariant(vs,
630 &context->cso_shader.vs_const_ranges,
631 &context->cso_shader.vs_const_used_size);
632 } else {
633 vs = device->ff.vs;
634 context->cso_shader.vs = vs->ff_cso;
635 }
636
637 if (context->rs[NINED3DRS_VSPOINTSIZE] != vs->point_size) {
638 context->rs[NINED3DRS_VSPOINTSIZE] = vs->point_size;
639 changed_group |= NINE_STATE_RASTERIZER;
640 }
641
642 if ((context->bound_samplers_mask_vs & vs->sampler_mask) != vs->sampler_mask)
643 /* Bound dummy sampler. */
644 changed_group |= NINE_STATE_SAMPLER;
645
646 context->commit |= NINE_STATE_COMMIT_VS;
647 return changed_group;
648 }
649
650 static inline uint32_t
651 prepare_ps(struct NineDevice9 *device, uint8_t shader_changed)
652 {
653 struct nine_context *context = &device->context;
654 struct NinePixelShader9 *ps = context->ps;
655 uint32_t changed_group = 0;
656 int has_key_changed = 0;
657
658 if (likely(ps))
659 has_key_changed = NinePixelShader9_UpdateKey(ps, context);
660
661 if (!shader_changed && !has_key_changed)
662 return 0;
663
664 if (likely(ps)) {
665 context->cso_shader.ps = NinePixelShader9_GetVariant(ps,
666 &context->cso_shader.ps_const_ranges,
667 &context->cso_shader.ps_const_used_size);
668 } else {
669 ps = device->ff.ps;
670 context->cso_shader.ps = ps->ff_cso;
671 }
672
673 if ((context->bound_samplers_mask_ps & ps->sampler_mask) != ps->sampler_mask)
674 /* Bound dummy sampler. */
675 changed_group |= NINE_STATE_SAMPLER;
676
677 context->commit |= NINE_STATE_COMMIT_PS;
678 return changed_group;
679 }
680
681 /* State preparation incremental */
682
683 /* State preparation + State commit */
684
685 static void
686 update_framebuffer(struct NineDevice9 *device, bool is_clear)
687 {
688 struct nine_context *context = &device->context;
689 struct pipe_context *pipe = context->pipe;
690 struct pipe_framebuffer_state *fb = &context->pipe_data.fb;
691 unsigned i;
692 struct NineSurface9 *rt0 = context->rt[0];
693 unsigned w = rt0->desc.Width;
694 unsigned h = rt0->desc.Height;
695 unsigned nr_samples = rt0->base.info.nr_samples;
696 unsigned ps_mask = context->ps ? context->ps->rt_mask : 1;
697 unsigned mask = is_clear ? 0xf : ps_mask;
698 const int sRGB = context->rs[D3DRS_SRGBWRITEENABLE] ? 1 : 0;
699
700 DBG("\n");
701
702 context->rt_mask = 0x0;
703 fb->nr_cbufs = 0;
704
705 /* all render targets must have the same size and the depth buffer must be
706 * bigger. Multisample has to match, according to spec. But some apps do
707 * things wrong there, and no error is returned. The behaviour they get
708 * apparently is that depth buffer is disabled if it doesn't match.
709 * Surely the same for render targets. */
710
711 /* Special case: D3DFMT_NULL is used to bound no real render target,
712 * but render to depth buffer. We have to not take into account the render
713 * target info. TODO: know what should happen when there are several render targers
714 * and the first one is D3DFMT_NULL */
715 if (rt0->desc.Format == D3DFMT_NULL && context->ds) {
716 w = context->ds->desc.Width;
717 h = context->ds->desc.Height;
718 nr_samples = context->ds->base.info.nr_samples;
719 }
720
721 for (i = 0; i < device->caps.NumSimultaneousRTs; ++i) {
722 struct NineSurface9 *rt = context->rt[i];
723
724 if (rt && rt->desc.Format != D3DFMT_NULL && (mask & (1 << i)) &&
725 rt->desc.Width == w && rt->desc.Height == h &&
726 rt->base.info.nr_samples == nr_samples) {
727 fb->cbufs[i] = NineSurface9_GetSurface(rt, sRGB);
728 context->rt_mask |= 1 << i;
729 fb->nr_cbufs = i + 1;
730 } else {
731 /* Color outputs must match RT slot,
732 * drivers will have to handle NULL entries for GL, too.
733 */
734 fb->cbufs[i] = NULL;
735 }
736 }
737
738 if (context->ds && context->ds->desc.Width >= w &&
739 context->ds->desc.Height >= h &&
740 context->ds->base.info.nr_samples == nr_samples) {
741 fb->zsbuf = NineSurface9_GetSurface(context->ds, 0);
742 } else {
743 fb->zsbuf = NULL;
744 }
745
746 fb->width = w;
747 fb->height = h;
748
749 pipe->set_framebuffer_state(pipe, fb); /* XXX: cso ? */
750
751 if (is_clear && context->rt_mask == ps_mask)
752 context->changed.group &= ~NINE_STATE_FB;
753 }
754
755 static void
756 update_viewport(struct NineDevice9 *device)
757 {
758 struct nine_context *context = &device->context;
759 const D3DVIEWPORT9 *vport = &context->viewport;
760 struct pipe_viewport_state pvport;
761
762 /* D3D coordinates are:
763 * -1 .. +1 for X,Y and
764 * 0 .. +1 for Z (we use pipe_rasterizer_state.clip_halfz)
765 */
766 pvport.scale[0] = (float)vport->Width * 0.5f;
767 pvport.scale[1] = (float)vport->Height * -0.5f;
768 pvport.scale[2] = vport->MaxZ - vport->MinZ;
769 pvport.translate[0] = (float)vport->Width * 0.5f + (float)vport->X;
770 pvport.translate[1] = (float)vport->Height * 0.5f + (float)vport->Y;
771 pvport.translate[2] = vport->MinZ;
772
773 /* We found R600 and SI cards have some imprecision
774 * on the barycentric coordinates used for interpolation.
775 * Some shaders rely on having something precise.
776 * We found that the proprietary driver has the imprecision issue,
777 * except when the render target width and height are powers of two.
778 * It is using some sort of workaround for these cases
779 * which covers likely all the cases the applications rely
780 * on something precise.
781 * We haven't found the workaround, but it seems like it's better
782 * for applications if the imprecision is biased towards infinity
783 * instead of -infinity (which is what measured). So shift slightly
784 * the viewport: not enough to change rasterization result (in particular
785 * for multisampling), but enough to make the imprecision biased
786 * towards infinity. We do this shift only if render target width and
787 * height are powers of two.
788 * Solves 'red shadows' bug on UE3 games.
789 */
790 if (device->driver_bugs.buggy_barycentrics &&
791 ((vport->Width & (vport->Width-1)) == 0) &&
792 ((vport->Height & (vport->Height-1)) == 0)) {
793 pvport.translate[0] -= 1.0f / 128.0f;
794 pvport.translate[1] -= 1.0f / 128.0f;
795 }
796
797 cso_set_viewport(context->cso, &pvport);
798 }
799
800 /* Loop through VS inputs and pick the vertex elements with the declared
801 * usage from the vertex declaration, then insert the instance divisor from
802 * the stream source frequency setting.
803 */
804 static void
805 update_vertex_elements(struct NineDevice9 *device)
806 {
807 struct nine_context *context = &device->context;
808 const struct NineVertexDeclaration9 *vdecl = device->context.vdecl;
809 const struct NineVertexShader9 *vs;
810 unsigned n, b, i;
811 int index;
812 char vdecl_index_map[16]; /* vs->num_inputs <= 16 */
813 char used_streams[device->caps.MaxStreams];
814 int dummy_vbo_stream = -1;
815 BOOL need_dummy_vbo = FALSE;
816 struct pipe_vertex_element ve[PIPE_MAX_ATTRIBS];
817
818 context->stream_usage_mask = 0;
819 memset(vdecl_index_map, -1, 16);
820 memset(used_streams, 0, device->caps.MaxStreams);
821 vs = context->programmable_vs ? context->vs : device->ff.vs;
822
823 if (vdecl) {
824 for (n = 0; n < vs->num_inputs; ++n) {
825 DBG("looking up input %u (usage %u) from vdecl(%p)\n",
826 n, vs->input_map[n].ndecl, vdecl);
827
828 for (i = 0; i < vdecl->nelems; i++) {
829 if (vdecl->usage_map[i] == vs->input_map[n].ndecl) {
830 vdecl_index_map[n] = i;
831 used_streams[vdecl->elems[i].vertex_buffer_index] = 1;
832 break;
833 }
834 }
835 if (vdecl_index_map[n] < 0)
836 need_dummy_vbo = TRUE;
837 }
838 } else {
839 /* No vertex declaration. Likely will never happen in practice,
840 * but we need not crash on this */
841 need_dummy_vbo = TRUE;
842 }
843
844 if (need_dummy_vbo) {
845 for (i = 0; i < device->caps.MaxStreams; i++ ) {
846 if (!used_streams[i]) {
847 dummy_vbo_stream = i;
848 break;
849 }
850 }
851 }
852 /* there are less vertex shader inputs than stream slots,
853 * so if we need a slot for the dummy vbo, we should have found one */
854 assert (!need_dummy_vbo || dummy_vbo_stream != -1);
855
856 for (n = 0; n < vs->num_inputs; ++n) {
857 index = vdecl_index_map[n];
858 if (index >= 0) {
859 ve[n] = vdecl->elems[index];
860 b = ve[n].vertex_buffer_index;
861 context->stream_usage_mask |= 1 << b;
862 /* XXX wine just uses 1 here: */
863 if (context->stream_freq[b] & D3DSTREAMSOURCE_INSTANCEDATA)
864 ve[n].instance_divisor = context->stream_freq[b] & 0x7FFFFF;
865 } else {
866 /* if the vertex declaration is incomplete compared to what the
867 * vertex shader needs, we bind a dummy vbo with 0 0 0 0.
868 * This is not precised by the spec, but is the behaviour
869 * tested on win */
870 ve[n].vertex_buffer_index = dummy_vbo_stream;
871 ve[n].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
872 ve[n].src_offset = 0;
873 ve[n].instance_divisor = 0;
874 }
875 }
876
877 if (context->dummy_vbo_bound_at != dummy_vbo_stream) {
878 if (context->dummy_vbo_bound_at >= 0)
879 context->changed.vtxbuf |= 1 << context->dummy_vbo_bound_at;
880 if (dummy_vbo_stream >= 0) {
881 context->changed.vtxbuf |= 1 << dummy_vbo_stream;
882 context->vbo_bound_done = FALSE;
883 }
884 context->dummy_vbo_bound_at = dummy_vbo_stream;
885 }
886
887 cso_set_vertex_elements(context->cso, vs->num_inputs, ve);
888 }
889
890 static void
891 update_vertex_buffers(struct NineDevice9 *device)
892 {
893 struct nine_context *context = &device->context;
894 struct pipe_context *pipe = context->pipe;
895 struct pipe_vertex_buffer dummy_vtxbuf;
896 uint32_t mask = context->changed.vtxbuf;
897 unsigned i;
898
899 DBG("mask=%x\n", mask);
900
901 if (context->dummy_vbo_bound_at >= 0) {
902 if (!context->vbo_bound_done) {
903 dummy_vtxbuf.buffer.resource = device->dummy_vbo;
904 dummy_vtxbuf.stride = 0;
905 dummy_vtxbuf.is_user_buffer = false;
906 dummy_vtxbuf.buffer_offset = 0;
907 pipe->set_vertex_buffers(pipe, context->dummy_vbo_bound_at,
908 1, &dummy_vtxbuf);
909 context->vbo_bound_done = TRUE;
910 }
911 mask &= ~(1 << context->dummy_vbo_bound_at);
912 }
913
914 for (i = 0; mask; mask >>= 1, ++i) {
915 if (mask & 1) {
916 if (context->vtxbuf[i].buffer.resource)
917 pipe->set_vertex_buffers(pipe, i, 1, &context->vtxbuf[i]);
918 else
919 pipe->set_vertex_buffers(pipe, i, 1, NULL);
920 }
921 }
922
923 context->changed.vtxbuf = 0;
924 }
925
926 static inline boolean
927 update_sampler_derived(struct nine_context *context, unsigned s)
928 {
929 boolean changed = FALSE;
930
931 if (context->samp[s][NINED3DSAMP_SHADOW] != context->texture[s].shadow) {
932 changed = TRUE;
933 context->samp[s][NINED3DSAMP_SHADOW] = context->texture[s].shadow;
934 }
935
936 if (context->samp[s][NINED3DSAMP_CUBETEX] !=
937 (context->texture[s].type == D3DRTYPE_CUBETEXTURE)) {
938 changed = TRUE;
939 context->samp[s][NINED3DSAMP_CUBETEX] =
940 context->texture[s].type == D3DRTYPE_CUBETEXTURE;
941 }
942
943 if (context->samp[s][D3DSAMP_MIPFILTER] != D3DTEXF_NONE) {
944 int lod = context->samp[s][D3DSAMP_MAXMIPLEVEL] - context->texture[s].lod;
945 if (lod < 0)
946 lod = 0;
947 if (context->samp[s][NINED3DSAMP_MINLOD] != lod) {
948 changed = TRUE;
949 context->samp[s][NINED3DSAMP_MINLOD] = lod;
950 }
951 } else {
952 context->changed.sampler[s] &= ~0x300; /* lod changes irrelevant */
953 }
954
955 return changed;
956 }
957
958 /* TODO: add sRGB override to pipe_sampler_state ? */
959 static void
960 update_textures_and_samplers(struct NineDevice9 *device)
961 {
962 struct nine_context *context = &device->context;
963 struct pipe_sampler_view *view[NINE_MAX_SAMPLERS];
964 unsigned num_textures;
965 unsigned i;
966 boolean commit_samplers;
967 uint16_t sampler_mask = context->ps ? context->ps->sampler_mask :
968 device->ff.ps->sampler_mask;
969
970 /* TODO: Can we reduce iterations here ? */
971
972 commit_samplers = FALSE;
973 context->bound_samplers_mask_ps = 0;
974 for (num_textures = 0, i = 0; i < NINE_MAX_SAMPLERS_PS; ++i) {
975 const unsigned s = NINE_SAMPLER_PS(i);
976 int sRGB;
977
978 if (!context->texture[s].enabled && !(sampler_mask & (1 << i))) {
979 view[i] = NULL;
980 continue;
981 }
982
983 if (context->texture[s].enabled) {
984 sRGB = context->samp[s][D3DSAMP_SRGBTEXTURE] ? 1 : 0;
985
986 view[i] = context->texture[s].view[sRGB];
987 num_textures = i + 1;
988
989 if (update_sampler_derived(context, s) || (context->changed.sampler[s] & 0x05fe)) {
990 context->changed.sampler[s] = 0;
991 commit_samplers = TRUE;
992 nine_convert_sampler_state(context->cso, s, context->samp[s]);
993 }
994 } else {
995 /* Bind dummy sampler. We do not bind dummy sampler when
996 * it is not needed because it could add overhead. The
997 * dummy sampler should have r=g=b=0 and a=1. We do not
998 * unbind dummy sampler directly when they are not needed
999 * anymore, but they're going to be removed as long as texture
1000 * or sampler states are changed. */
1001 view[i] = device->dummy_sampler_view;
1002 num_textures = i + 1;
1003
1004 cso_single_sampler(context->cso, PIPE_SHADER_FRAGMENT,
1005 s - NINE_SAMPLER_PS(0), &device->dummy_sampler_state);
1006
1007 commit_samplers = TRUE;
1008 context->changed.sampler[s] = ~0;
1009 }
1010
1011 context->bound_samplers_mask_ps |= (1 << s);
1012 }
1013
1014 cso_set_sampler_views(context->cso, PIPE_SHADER_FRAGMENT, num_textures, view);
1015
1016 if (commit_samplers)
1017 cso_single_sampler_done(context->cso, PIPE_SHADER_FRAGMENT);
1018
1019 commit_samplers = FALSE;
1020 sampler_mask = context->programmable_vs ? context->vs->sampler_mask : 0;
1021 context->bound_samplers_mask_vs = 0;
1022 for (num_textures = 0, i = 0; i < NINE_MAX_SAMPLERS_VS; ++i) {
1023 const unsigned s = NINE_SAMPLER_VS(i);
1024 int sRGB;
1025
1026 if (!context->texture[s].enabled && !(sampler_mask & (1 << i))) {
1027 view[i] = NULL;
1028 continue;
1029 }
1030
1031 if (context->texture[s].enabled) {
1032 sRGB = context->samp[s][D3DSAMP_SRGBTEXTURE] ? 1 : 0;
1033
1034 view[i] = context->texture[s].view[sRGB];
1035 num_textures = i + 1;
1036
1037 if (update_sampler_derived(context, s) || (context->changed.sampler[s] & 0x05fe)) {
1038 context->changed.sampler[s] = 0;
1039 commit_samplers = TRUE;
1040 nine_convert_sampler_state(context->cso, s, context->samp[s]);
1041 }
1042 } else {
1043 /* Bind dummy sampler. We do not bind dummy sampler when
1044 * it is not needed because it could add overhead. The
1045 * dummy sampler should have r=g=b=0 and a=1. We do not
1046 * unbind dummy sampler directly when they are not needed
1047 * anymore, but they're going to be removed as long as texture
1048 * or sampler states are changed. */
1049 view[i] = device->dummy_sampler_view;
1050 num_textures = i + 1;
1051
1052 cso_single_sampler(context->cso, PIPE_SHADER_VERTEX,
1053 s - NINE_SAMPLER_VS(0), &device->dummy_sampler_state);
1054
1055 commit_samplers = TRUE;
1056 context->changed.sampler[s] = ~0;
1057 }
1058
1059 context->bound_samplers_mask_vs |= (1 << i);
1060 }
1061
1062 cso_set_sampler_views(context->cso, PIPE_SHADER_VERTEX, num_textures, view);
1063
1064 if (commit_samplers)
1065 cso_single_sampler_done(context->cso, PIPE_SHADER_VERTEX);
1066 }
1067
1068 /* State commit only */
1069
1070 static inline void
1071 commit_blend(struct NineDevice9 *device)
1072 {
1073 struct nine_context *context = &device->context;
1074
1075 cso_set_blend(context->cso, &context->pipe_data.blend);
1076 }
1077
1078 static inline void
1079 commit_dsa(struct NineDevice9 *device)
1080 {
1081 struct nine_context *context = &device->context;
1082
1083 cso_set_depth_stencil_alpha(context->cso, &context->pipe_data.dsa);
1084 }
1085
1086 static inline void
1087 commit_scissor(struct NineDevice9 *device)
1088 {
1089 struct nine_context *context = &device->context;
1090 struct pipe_context *pipe = context->pipe;
1091
1092 pipe->set_scissor_states(pipe, 0, 1, &context->scissor);
1093 }
1094
1095 static inline void
1096 commit_rasterizer(struct NineDevice9 *device)
1097 {
1098 struct nine_context *context = &device->context;
1099
1100 cso_set_rasterizer(context->cso, &context->pipe_data.rast);
1101 }
1102
1103 static inline void
1104 commit_vs_constants(struct NineDevice9 *device)
1105 {
1106 struct nine_context *context = &device->context;
1107 struct pipe_context *pipe = context->pipe;
1108
1109 if (unlikely(!context->programmable_vs))
1110 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &context->pipe_data.cb_vs_ff);
1111 else {
1112 if (context->swvp) {
1113 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &context->pipe_data.cb0_swvp);
1114 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 1, &context->pipe_data.cb1_swvp);
1115 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 2, &context->pipe_data.cb2_swvp);
1116 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 3, &context->pipe_data.cb3_swvp);
1117 } else {
1118 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &context->pipe_data.cb_vs);
1119 }
1120 }
1121 }
1122
1123 static inline void
1124 commit_ps_constants(struct NineDevice9 *device)
1125 {
1126 struct nine_context *context = &device->context;
1127 struct pipe_context *pipe = context->pipe;
1128
1129 if (unlikely(!context->ps))
1130 pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, 0, &context->pipe_data.cb_ps_ff);
1131 else
1132 pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, 0, &context->pipe_data.cb_ps);
1133 }
1134
1135 static inline void
1136 commit_vs(struct NineDevice9 *device)
1137 {
1138 struct nine_context *context = &device->context;
1139
1140 context->pipe->bind_vs_state(context->pipe, context->cso_shader.vs);
1141 }
1142
1143
1144 static inline void
1145 commit_ps(struct NineDevice9 *device)
1146 {
1147 struct nine_context *context = &device->context;
1148
1149 context->pipe->bind_fs_state(context->pipe, context->cso_shader.ps);
1150 }
1151 /* State Update */
1152
1153 #define NINE_STATE_SHADER_CHANGE_VS \
1154 (NINE_STATE_VS | \
1155 NINE_STATE_TEXTURE | \
1156 NINE_STATE_VS_PARAMS_MISC | \
1157 NINE_STATE_SWVP)
1158
1159 #define NINE_STATE_SHADER_CHANGE_PS \
1160 (NINE_STATE_PS | \
1161 NINE_STATE_TEXTURE | \
1162 NINE_STATE_PS_PARAMS_MISC)
1163
1164 #define NINE_STATE_FREQUENT \
1165 (NINE_STATE_RASTERIZER | \
1166 NINE_STATE_TEXTURE | \
1167 NINE_STATE_SAMPLER | \
1168 NINE_STATE_VS_CONST | \
1169 NINE_STATE_PS_CONST | \
1170 NINE_STATE_MULTISAMPLE)
1171
1172 #define NINE_STATE_COMMON \
1173 (NINE_STATE_FB | \
1174 NINE_STATE_BLEND | \
1175 NINE_STATE_DSA | \
1176 NINE_STATE_VIEWPORT | \
1177 NINE_STATE_VDECL | \
1178 NINE_STATE_IDXBUF | \
1179 NINE_STATE_STREAMFREQ)
1180
1181 #define NINE_STATE_RARE \
1182 (NINE_STATE_SCISSOR | \
1183 NINE_STATE_BLEND_COLOR | \
1184 NINE_STATE_STENCIL_REF | \
1185 NINE_STATE_SAMPLE_MASK)
1186
1187 static void
1188 nine_update_state(struct NineDevice9 *device)
1189 {
1190 struct nine_context *context = &device->context;
1191 struct pipe_context *pipe = context->pipe;
1192 uint32_t group;
1193
1194 DBG("changed state groups: %x\n", context->changed.group);
1195
1196 /* NOTE: We may want to use the cso cache for everything, or let
1197 * NineDevice9.RestoreNonCSOState actually set the states, then we wouldn't
1198 * have to care about state being clobbered here and could merge this back
1199 * into update_textures. Except, we also need to re-validate textures that
1200 * may be dirty anyway, even if no texture bindings changed.
1201 */
1202
1203 /* ff_update may change VS/PS dirty bits */
1204 if (unlikely(!context->programmable_vs || !context->ps))
1205 nine_ff_update(device);
1206 group = context->changed.group;
1207
1208 if (group & (NINE_STATE_SHADER_CHANGE_VS | NINE_STATE_SHADER_CHANGE_PS)) {
1209 if (group & NINE_STATE_SHADER_CHANGE_VS)
1210 group |= prepare_vs(device, (group & NINE_STATE_VS) != 0); /* may set NINE_STATE_RASTERIZER and NINE_STATE_SAMPLER*/
1211 if (group & NINE_STATE_SHADER_CHANGE_PS)
1212 group |= prepare_ps(device, (group & NINE_STATE_PS) != 0);
1213 }
1214
1215 if (group & (NINE_STATE_COMMON | NINE_STATE_VS)) {
1216 if (group & NINE_STATE_FB)
1217 update_framebuffer(device, FALSE);
1218 if (group & NINE_STATE_BLEND)
1219 prepare_blend(device);
1220 if (group & NINE_STATE_DSA)
1221 prepare_dsa(device);
1222 if (group & NINE_STATE_VIEWPORT)
1223 update_viewport(device);
1224 if (group & (NINE_STATE_VDECL | NINE_STATE_VS | NINE_STATE_STREAMFREQ))
1225 update_vertex_elements(device);
1226 }
1227
1228 if (likely(group & (NINE_STATE_FREQUENT | NINE_STATE_VS | NINE_STATE_PS | NINE_STATE_SWVP))) {
1229 if (group & NINE_STATE_MULTISAMPLE)
1230 group |= check_multisample(device);
1231 if (group & NINE_STATE_RASTERIZER)
1232 prepare_rasterizer(device);
1233 if (group & (NINE_STATE_TEXTURE | NINE_STATE_SAMPLER))
1234 update_textures_and_samplers(device);
1235 if ((group & (NINE_STATE_VS_CONST | NINE_STATE_VS | NINE_STATE_SWVP)) && context->programmable_vs)
1236 prepare_vs_constants_userbuf(device);
1237 if ((group & (NINE_STATE_PS_CONST | NINE_STATE_PS)) && context->ps)
1238 prepare_ps_constants_userbuf(device);
1239 }
1240
1241 if (context->changed.vtxbuf)
1242 update_vertex_buffers(device);
1243
1244 if (context->commit & NINE_STATE_COMMIT_BLEND)
1245 commit_blend(device);
1246 if (context->commit & NINE_STATE_COMMIT_DSA)
1247 commit_dsa(device);
1248 if (context->commit & NINE_STATE_COMMIT_RASTERIZER)
1249 commit_rasterizer(device);
1250 if (context->commit & NINE_STATE_COMMIT_CONST_VS)
1251 commit_vs_constants(device);
1252 if (context->commit & NINE_STATE_COMMIT_CONST_PS)
1253 commit_ps_constants(device);
1254 if (context->commit & NINE_STATE_COMMIT_VS)
1255 commit_vs(device);
1256 if (context->commit & NINE_STATE_COMMIT_PS)
1257 commit_ps(device);
1258
1259 context->commit = 0;
1260
1261 if (unlikely(context->changed.ucp)) {
1262 pipe->set_clip_state(pipe, &context->clip);
1263 context->changed.ucp = FALSE;
1264 }
1265
1266 if (unlikely(group & NINE_STATE_RARE)) {
1267 if (group & NINE_STATE_SCISSOR)
1268 commit_scissor(device);
1269 if (group & NINE_STATE_BLEND_COLOR) {
1270 struct pipe_blend_color color;
1271 d3dcolor_to_rgba(&color.color[0], context->rs[D3DRS_BLENDFACTOR]);
1272 pipe->set_blend_color(pipe, &color);
1273 }
1274 if (group & NINE_STATE_SAMPLE_MASK) {
1275 if (context->rt[0]->desc.MultiSampleType <= D3DMULTISAMPLE_NONMASKABLE) {
1276 pipe->set_sample_mask(pipe, ~0);
1277 } else {
1278 pipe->set_sample_mask(pipe, context->rs[D3DRS_MULTISAMPLEMASK]);
1279 }
1280 }
1281 if (group & NINE_STATE_STENCIL_REF) {
1282 struct pipe_stencil_ref ref;
1283 ref.ref_value[0] = context->rs[D3DRS_STENCILREF];
1284 ref.ref_value[1] = ref.ref_value[0];
1285 pipe->set_stencil_ref(pipe, &ref);
1286 }
1287 }
1288
1289 context->changed.group &=
1290 (NINE_STATE_FF | NINE_STATE_VS_CONST | NINE_STATE_PS_CONST);
1291
1292 DBG("finished\n");
1293 }
1294
1295 #define RESZ_CODE 0x7fa05000
1296
1297 static void
1298 NineDevice9_ResolveZ( struct NineDevice9 *device )
1299 {
1300 struct nine_context *context = &device->context;
1301 const struct util_format_description *desc;
1302 struct NineSurface9 *source = context->ds;
1303 struct pipe_resource *src, *dst;
1304 struct pipe_blit_info blit;
1305
1306 DBG("RESZ resolve\n");
1307
1308 if (!source || !context->texture[0].enabled ||
1309 context->texture[0].type != D3DRTYPE_TEXTURE)
1310 return;
1311
1312 src = source->base.resource;
1313 dst = context->texture[0].resource;
1314
1315 if (!src || !dst)
1316 return;
1317
1318 /* check dst is depth format. we know already for src */
1319 desc = util_format_description(dst->format);
1320 if (desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1321 return;
1322
1323 memset(&blit, 0, sizeof(blit));
1324 blit.src.resource = src;
1325 blit.src.level = 0;
1326 blit.src.format = src->format;
1327 blit.src.box.z = 0;
1328 blit.src.box.depth = 1;
1329 blit.src.box.x = 0;
1330 blit.src.box.y = 0;
1331 blit.src.box.width = src->width0;
1332 blit.src.box.height = src->height0;
1333
1334 blit.dst.resource = dst;
1335 blit.dst.level = 0;
1336 blit.dst.format = dst->format;
1337 blit.dst.box.z = 0;
1338 blit.dst.box.depth = 1;
1339 blit.dst.box.x = 0;
1340 blit.dst.box.y = 0;
1341 blit.dst.box.width = dst->width0;
1342 blit.dst.box.height = dst->height0;
1343
1344 blit.mask = PIPE_MASK_ZS;
1345 blit.filter = PIPE_TEX_FILTER_NEAREST;
1346 blit.scissor_enable = FALSE;
1347
1348 context->pipe->blit(context->pipe, &blit);
1349 }
1350
1351 #define ALPHA_TO_COVERAGE_ENABLE MAKEFOURCC('A', '2', 'M', '1')
1352 #define ALPHA_TO_COVERAGE_DISABLE MAKEFOURCC('A', '2', 'M', '0')
1353
1354 /* Nine_context functions.
1355 * Serialized through CSMT macros.
1356 */
1357
1358 static void
1359 nine_context_set_texture_apply(struct NineDevice9 *device,
1360 DWORD stage,
1361 BOOL enabled,
1362 BOOL shadow,
1363 DWORD lod,
1364 D3DRESOURCETYPE type,
1365 uint8_t pstype,
1366 struct pipe_resource *res,
1367 struct pipe_sampler_view *view0,
1368 struct pipe_sampler_view *view1);
1369 static void
1370 nine_context_set_stream_source_apply(struct NineDevice9 *device,
1371 UINT StreamNumber,
1372 struct pipe_resource *res,
1373 UINT OffsetInBytes,
1374 UINT Stride);
1375
1376 static void
1377 nine_context_set_indices_apply(struct NineDevice9 *device,
1378 struct pipe_resource *res,
1379 UINT IndexSize,
1380 UINT OffsetInBytes);
1381
1382 static void
1383 nine_context_set_pixel_shader_constant_i_transformed(struct NineDevice9 *device,
1384 UINT StartRegister,
1385 const int *pConstantData,
1386 unsigned pConstantData_size,
1387 UINT Vector4iCount);
1388
1389 CSMT_ITEM_NO_WAIT(nine_context_set_render_state,
1390 ARG_VAL(D3DRENDERSTATETYPE, State),
1391 ARG_VAL(DWORD, Value))
1392 {
1393 struct nine_context *context = &device->context;
1394
1395 /* Amd hacks (equivalent to GL extensions) */
1396 if (unlikely(State == D3DRS_POINTSIZE)) {
1397 if (Value == RESZ_CODE) {
1398 NineDevice9_ResolveZ(device);
1399 return;
1400 }
1401
1402 if (Value == ALPHA_TO_COVERAGE_ENABLE ||
1403 Value == ALPHA_TO_COVERAGE_DISABLE) {
1404 context->rs[NINED3DRS_ALPHACOVERAGE] = (Value == ALPHA_TO_COVERAGE_ENABLE);
1405 context->changed.group |= NINE_STATE_BLEND;
1406 return;
1407 }
1408 }
1409
1410 /* NV hack */
1411 if (unlikely(State == D3DRS_ADAPTIVETESS_Y)) {
1412 if (Value == D3DFMT_ATOC || (Value == D3DFMT_UNKNOWN && context->rs[NINED3DRS_ALPHACOVERAGE])) {
1413 context->rs[NINED3DRS_ALPHACOVERAGE] = (Value == D3DFMT_ATOC) ? 3 : 0;
1414 context->rs[NINED3DRS_ALPHACOVERAGE] &= context->rs[D3DRS_ALPHATESTENABLE] ? 3 : 2;
1415 context->changed.group |= NINE_STATE_BLEND;
1416 return;
1417 }
1418 }
1419 if (unlikely(State == D3DRS_ALPHATESTENABLE && (context->rs[NINED3DRS_ALPHACOVERAGE] & 2))) {
1420 DWORD alphacoverage_prev = context->rs[NINED3DRS_ALPHACOVERAGE];
1421 context->rs[NINED3DRS_ALPHACOVERAGE] = (Value ? 3 : 2);
1422 if (context->rs[NINED3DRS_ALPHACOVERAGE] != alphacoverage_prev)
1423 context->changed.group |= NINE_STATE_BLEND;
1424 }
1425
1426 context->rs[State] = nine_fix_render_state_value(State, Value);
1427 context->changed.group |= nine_render_state_group[State];
1428 }
1429
1430 CSMT_ITEM_NO_WAIT(nine_context_set_texture_apply,
1431 ARG_VAL(DWORD, stage),
1432 ARG_VAL(BOOL, enabled),
1433 ARG_VAL(BOOL, shadow),
1434 ARG_VAL(DWORD, lod),
1435 ARG_VAL(D3DRESOURCETYPE, type),
1436 ARG_VAL(uint8_t, pstype),
1437 ARG_BIND_RES(struct pipe_resource, res),
1438 ARG_BIND_VIEW(struct pipe_sampler_view, view0),
1439 ARG_BIND_VIEW(struct pipe_sampler_view, view1))
1440 {
1441 struct nine_context *context = &device->context;
1442
1443 context->texture[stage].enabled = enabled;
1444 context->samplers_shadow &= ~(1 << stage);
1445 context->samplers_shadow |= shadow << stage;
1446 context->texture[stage].shadow = shadow;
1447 context->texture[stage].lod = lod;
1448 context->texture[stage].type = type;
1449 context->texture[stage].pstype = pstype;
1450 pipe_resource_reference(&context->texture[stage].resource, res);
1451 pipe_sampler_view_reference(&context->texture[stage].view[0], view0);
1452 pipe_sampler_view_reference(&context->texture[stage].view[1], view1);
1453
1454 context->changed.group |= NINE_STATE_TEXTURE;
1455 }
1456
1457 void
1458 nine_context_set_texture(struct NineDevice9 *device,
1459 DWORD Stage,
1460 struct NineBaseTexture9 *tex)
1461 {
1462 BOOL enabled = FALSE;
1463 BOOL shadow = FALSE;
1464 DWORD lod = 0;
1465 D3DRESOURCETYPE type = D3DRTYPE_TEXTURE;
1466 uint8_t pstype = 0;
1467 struct pipe_resource *res = NULL;
1468 struct pipe_sampler_view *view0 = NULL, *view1 = NULL;
1469
1470 /* For managed pool, the data can be initially incomplete.
1471 * In that case, the texture is rebound later
1472 * (in NineBaseTexture9_Validate/NineBaseTexture9_UploadSelf). */
1473 if (tex && tex->base.resource) {
1474 enabled = TRUE;
1475 shadow = tex->shadow;
1476 lod = tex->managed.lod;
1477 type = tex->base.type;
1478 pstype = tex->pstype;
1479 res = tex->base.resource;
1480 view0 = NineBaseTexture9_GetSamplerView(tex, 0);
1481 view1 = NineBaseTexture9_GetSamplerView(tex, 1);
1482 }
1483
1484 nine_context_set_texture_apply(device, Stage, enabled,
1485 shadow, lod, type, pstype,
1486 res, view0, view1);
1487 }
1488
1489 CSMT_ITEM_NO_WAIT(nine_context_set_sampler_state,
1490 ARG_VAL(DWORD, Sampler),
1491 ARG_VAL(D3DSAMPLERSTATETYPE, Type),
1492 ARG_VAL(DWORD, Value))
1493 {
1494 struct nine_context *context = &device->context;
1495
1496 if (unlikely(!nine_check_sampler_state_value(Type, Value)))
1497 return;
1498
1499 context->samp[Sampler][Type] = Value;
1500 context->changed.group |= NINE_STATE_SAMPLER;
1501 context->changed.sampler[Sampler] |= 1 << Type;
1502 }
1503
1504 CSMT_ITEM_NO_WAIT(nine_context_set_stream_source_apply,
1505 ARG_VAL(UINT, StreamNumber),
1506 ARG_BIND_RES(struct pipe_resource, res),
1507 ARG_VAL(UINT, OffsetInBytes),
1508 ARG_VAL(UINT, Stride))
1509 {
1510 struct nine_context *context = &device->context;
1511 const unsigned i = StreamNumber;
1512
1513 context->vtxbuf[i].stride = Stride;
1514 context->vtxbuf[i].buffer_offset = OffsetInBytes;
1515 pipe_resource_reference(&context->vtxbuf[i].buffer.resource, res);
1516
1517 context->changed.vtxbuf |= 1 << StreamNumber;
1518 }
1519
1520 void
1521 nine_context_set_stream_source(struct NineDevice9 *device,
1522 UINT StreamNumber,
1523 struct NineVertexBuffer9 *pVBuf9,
1524 UINT OffsetInBytes,
1525 UINT Stride)
1526 {
1527 struct pipe_resource *res = NULL;
1528 unsigned offset = 0;
1529
1530 if (pVBuf9)
1531 res = NineVertexBuffer9_GetResource(pVBuf9, &offset);
1532 /* in the future when there is internal offset, add it
1533 * to OffsetInBytes */
1534
1535 nine_context_set_stream_source_apply(device, StreamNumber,
1536 res, offset + OffsetInBytes,
1537 Stride);
1538 }
1539
1540 CSMT_ITEM_NO_WAIT(nine_context_set_stream_source_freq,
1541 ARG_VAL(UINT, StreamNumber),
1542 ARG_VAL(UINT, Setting))
1543 {
1544 struct nine_context *context = &device->context;
1545
1546 context->stream_freq[StreamNumber] = Setting;
1547
1548 if (Setting & D3DSTREAMSOURCE_INSTANCEDATA)
1549 context->stream_instancedata_mask |= 1 << StreamNumber;
1550 else
1551 context->stream_instancedata_mask &= ~(1 << StreamNumber);
1552
1553 if (StreamNumber != 0)
1554 context->changed.group |= NINE_STATE_STREAMFREQ;
1555 }
1556
1557 CSMT_ITEM_NO_WAIT(nine_context_set_indices_apply,
1558 ARG_BIND_RES(struct pipe_resource, res),
1559 ARG_VAL(UINT, IndexSize),
1560 ARG_VAL(UINT, OffsetInBytes))
1561 {
1562 struct nine_context *context = &device->context;
1563
1564 context->index_size = IndexSize;
1565 context->index_offset = OffsetInBytes;
1566 pipe_resource_reference(&context->idxbuf, res);
1567
1568 context->changed.group |= NINE_STATE_IDXBUF;
1569 }
1570
1571 void
1572 nine_context_set_indices(struct NineDevice9 *device,
1573 struct NineIndexBuffer9 *idxbuf)
1574 {
1575 struct pipe_resource *res = NULL;
1576 UINT IndexSize = 0;
1577 unsigned OffsetInBytes = 0;
1578
1579 if (idxbuf) {
1580 res = NineIndexBuffer9_GetBuffer(idxbuf, &OffsetInBytes);
1581 IndexSize = idxbuf->index_size;
1582 }
1583
1584 nine_context_set_indices_apply(device, res, IndexSize, OffsetInBytes);
1585 }
1586
1587 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_declaration,
1588 ARG_BIND_REF(struct NineVertexDeclaration9, vdecl))
1589 {
1590 struct nine_context *context = &device->context;
1591 BOOL was_programmable_vs = context->programmable_vs;
1592
1593 nine_bind(&context->vdecl, vdecl);
1594
1595 context->programmable_vs = context->vs && !(context->vdecl && context->vdecl->position_t);
1596 if (was_programmable_vs != context->programmable_vs) {
1597 context->commit |= NINE_STATE_COMMIT_CONST_VS;
1598 context->changed.group |= NINE_STATE_VS;
1599 }
1600
1601 context->changed.group |= NINE_STATE_VDECL;
1602 }
1603
1604 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_shader,
1605 ARG_BIND_REF(struct NineVertexShader9, pShader))
1606 {
1607 struct nine_context *context = &device->context;
1608 BOOL was_programmable_vs = context->programmable_vs;
1609
1610 nine_bind(&context->vs, pShader);
1611
1612 context->programmable_vs = context->vs && !(context->vdecl && context->vdecl->position_t);
1613
1614 /* ff -> non-ff: commit back non-ff constants */
1615 if (!was_programmable_vs && context->programmable_vs)
1616 context->commit |= NINE_STATE_COMMIT_CONST_VS;
1617
1618 context->changed.group |= NINE_STATE_VS;
1619 }
1620
1621 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_shader_constant_f,
1622 ARG_VAL(UINT, StartRegister),
1623 ARG_MEM(float, pConstantData),
1624 ARG_MEM_SIZE(unsigned, pConstantData_size),
1625 ARG_VAL(UINT, Vector4fCount))
1626 {
1627 struct nine_context *context = &device->context;
1628 float *vs_const_f = device->may_swvp ? context->vs_const_f_swvp : context->vs_const_f;
1629
1630 memcpy(&vs_const_f[StartRegister * 4],
1631 pConstantData,
1632 pConstantData_size);
1633
1634 if (device->may_swvp) {
1635 Vector4fCount = MIN2(StartRegister + Vector4fCount, NINE_MAX_CONST_F) - StartRegister;
1636 if (StartRegister < NINE_MAX_CONST_F)
1637 memcpy(&context->vs_const_f[StartRegister * 4],
1638 pConstantData,
1639 Vector4fCount * 4 * sizeof(context->vs_const_f[0]));
1640 }
1641
1642 context->changed.vs_const_f = TRUE;
1643 context->changed.group |= NINE_STATE_VS_CONST;
1644 }
1645
1646 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_shader_constant_i,
1647 ARG_VAL(UINT, StartRegister),
1648 ARG_MEM(int, pConstantData),
1649 ARG_MEM_SIZE(unsigned, pConstantData_size),
1650 ARG_VAL(UINT, Vector4iCount))
1651 {
1652 struct nine_context *context = &device->context;
1653 int i;
1654
1655 if (device->driver_caps.vs_integer) {
1656 memcpy(&context->vs_const_i[4 * StartRegister],
1657 pConstantData,
1658 pConstantData_size);
1659 } else {
1660 for (i = 0; i < Vector4iCount; i++) {
1661 context->vs_const_i[4 * (StartRegister + i)] = fui((float)(pConstantData[4 * i]));
1662 context->vs_const_i[4 * (StartRegister + i) + 1] = fui((float)(pConstantData[4 * i + 1]));
1663 context->vs_const_i[4 * (StartRegister + i) + 2] = fui((float)(pConstantData[4 * i + 2]));
1664 context->vs_const_i[4 * (StartRegister + i) + 3] = fui((float)(pConstantData[4 * i + 3]));
1665 }
1666 }
1667
1668 context->changed.vs_const_i = TRUE;
1669 context->changed.group |= NINE_STATE_VS_CONST | NINE_STATE_VS_PARAMS_MISC;
1670 }
1671
1672 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_shader_constant_b,
1673 ARG_VAL(UINT, StartRegister),
1674 ARG_MEM(BOOL, pConstantData),
1675 ARG_MEM_SIZE(unsigned, pConstantData_size),
1676 ARG_VAL(UINT, BoolCount))
1677 {
1678 struct nine_context *context = &device->context;
1679 int i;
1680 uint32_t bool_true = device->driver_caps.vs_integer ? 0xFFFFFFFF : fui(1.0f);
1681
1682 (void) pConstantData_size;
1683
1684 for (i = 0; i < BoolCount; i++)
1685 context->vs_const_b[StartRegister + i] = pConstantData[i] ? bool_true : 0;
1686
1687 context->changed.vs_const_b = TRUE;
1688 context->changed.group |= NINE_STATE_VS_CONST | NINE_STATE_VS_PARAMS_MISC;
1689 }
1690
1691 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader,
1692 ARG_BIND_REF(struct NinePixelShader9, ps))
1693 {
1694 struct nine_context *context = &device->context;
1695 unsigned old_mask = context->ps ? context->ps->rt_mask : 1;
1696 unsigned mask;
1697
1698 /* ff -> non-ff: commit back non-ff constants */
1699 if (!context->ps && ps)
1700 context->commit |= NINE_STATE_COMMIT_CONST_PS;
1701
1702 nine_bind(&context->ps, ps);
1703
1704 context->changed.group |= NINE_STATE_PS;
1705
1706 mask = context->ps ? context->ps->rt_mask : 1;
1707 /* We need to update cbufs if the pixel shader would
1708 * write to different render targets */
1709 if (mask != old_mask)
1710 context->changed.group |= NINE_STATE_FB;
1711 }
1712
1713 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader_constant_f,
1714 ARG_VAL(UINT, StartRegister),
1715 ARG_MEM(float, pConstantData),
1716 ARG_MEM_SIZE(unsigned, pConstantData_size),
1717 ARG_VAL(UINT, Vector4fCount))
1718 {
1719 struct nine_context *context = &device->context;
1720
1721 memcpy(&context->ps_const_f[StartRegister * 4],
1722 pConstantData,
1723 pConstantData_size);
1724
1725 context->changed.ps_const_f = TRUE;
1726 context->changed.group |= NINE_STATE_PS_CONST;
1727 }
1728
1729 /* For stateblocks */
1730 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader_constant_i_transformed,
1731 ARG_VAL(UINT, StartRegister),
1732 ARG_MEM(int, pConstantData),
1733 ARG_MEM_SIZE(unsigned, pConstantData_size),
1734 ARG_VAL(UINT, Vector4iCount))
1735 {
1736 struct nine_context *context = &device->context;
1737
1738 memcpy(&context->ps_const_i[StartRegister][0],
1739 pConstantData,
1740 Vector4iCount * sizeof(context->ps_const_i[0]));
1741
1742 context->changed.ps_const_i = TRUE;
1743 context->changed.group |= NINE_STATE_PS_CONST | NINE_STATE_PS_PARAMS_MISC;
1744 }
1745
1746 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader_constant_i,
1747 ARG_VAL(UINT, StartRegister),
1748 ARG_MEM(int, pConstantData),
1749 ARG_MEM_SIZE(unsigned, pConstantData_size),
1750 ARG_VAL(UINT, Vector4iCount))
1751 {
1752 struct nine_context *context = &device->context;
1753 int i;
1754
1755 if (device->driver_caps.ps_integer) {
1756 memcpy(&context->ps_const_i[StartRegister][0],
1757 pConstantData,
1758 pConstantData_size);
1759 } else {
1760 for (i = 0; i < Vector4iCount; i++) {
1761 context->ps_const_i[StartRegister+i][0] = fui((float)(pConstantData[4*i]));
1762 context->ps_const_i[StartRegister+i][1] = fui((float)(pConstantData[4*i+1]));
1763 context->ps_const_i[StartRegister+i][2] = fui((float)(pConstantData[4*i+2]));
1764 context->ps_const_i[StartRegister+i][3] = fui((float)(pConstantData[4*i+3]));
1765 }
1766 }
1767 context->changed.ps_const_i = TRUE;
1768 context->changed.group |= NINE_STATE_PS_CONST | NINE_STATE_PS_PARAMS_MISC;
1769 }
1770
1771 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader_constant_b,
1772 ARG_VAL(UINT, StartRegister),
1773 ARG_MEM(BOOL, pConstantData),
1774 ARG_MEM_SIZE(unsigned, pConstantData_size),
1775 ARG_VAL(UINT, BoolCount))
1776 {
1777 struct nine_context *context = &device->context;
1778 int i;
1779 uint32_t bool_true = device->driver_caps.ps_integer ? 0xFFFFFFFF : fui(1.0f);
1780
1781 (void) pConstantData_size;
1782
1783 for (i = 0; i < BoolCount; i++)
1784 context->ps_const_b[StartRegister + i] = pConstantData[i] ? bool_true : 0;
1785
1786 context->changed.ps_const_b = TRUE;
1787 context->changed.group |= NINE_STATE_PS_CONST | NINE_STATE_PS_PARAMS_MISC;
1788 }
1789
1790 /* XXX: use resource, as resource might change */
1791 CSMT_ITEM_NO_WAIT(nine_context_set_render_target,
1792 ARG_VAL(DWORD, RenderTargetIndex),
1793 ARG_BIND_REF(struct NineSurface9, rt))
1794 {
1795 struct nine_context *context = &device->context;
1796 const unsigned i = RenderTargetIndex;
1797
1798 if (i == 0) {
1799 context->viewport.X = 0;
1800 context->viewport.Y = 0;
1801 context->viewport.Width = rt->desc.Width;
1802 context->viewport.Height = rt->desc.Height;
1803 context->viewport.MinZ = 0.0f;
1804 context->viewport.MaxZ = 1.0f;
1805
1806 context->scissor.minx = 0;
1807 context->scissor.miny = 0;
1808 context->scissor.maxx = rt->desc.Width;
1809 context->scissor.maxy = rt->desc.Height;
1810
1811 context->changed.group |= NINE_STATE_VIEWPORT | NINE_STATE_SCISSOR | NINE_STATE_MULTISAMPLE;
1812
1813 if (context->rt[0] &&
1814 (context->rt[0]->desc.MultiSampleType <= D3DMULTISAMPLE_NONMASKABLE) !=
1815 (rt->desc.MultiSampleType <= D3DMULTISAMPLE_NONMASKABLE))
1816 context->changed.group |= NINE_STATE_SAMPLE_MASK;
1817 }
1818
1819 if (context->rt[i] != rt) {
1820 nine_bind(&context->rt[i], rt);
1821 context->changed.group |= NINE_STATE_FB;
1822 }
1823 }
1824
1825 /* XXX: use resource instead of ds, as resource might change */
1826 CSMT_ITEM_NO_WAIT(nine_context_set_depth_stencil,
1827 ARG_BIND_REF(struct NineSurface9, ds))
1828 {
1829 struct nine_context *context = &device->context;
1830
1831 nine_bind(&context->ds, ds);
1832 context->changed.group |= NINE_STATE_FB;
1833 }
1834
1835 CSMT_ITEM_NO_WAIT(nine_context_set_viewport,
1836 ARG_COPY_REF(D3DVIEWPORT9, viewport))
1837 {
1838 struct nine_context *context = &device->context;
1839
1840 context->viewport = *viewport;
1841 context->changed.group |= NINE_STATE_VIEWPORT;
1842 }
1843
1844 CSMT_ITEM_NO_WAIT(nine_context_set_scissor,
1845 ARG_COPY_REF(struct pipe_scissor_state, scissor))
1846 {
1847 struct nine_context *context = &device->context;
1848
1849 context->scissor = *scissor;
1850 context->changed.group |= NINE_STATE_SCISSOR;
1851 }
1852
1853 CSMT_ITEM_NO_WAIT(nine_context_set_transform,
1854 ARG_VAL(D3DTRANSFORMSTATETYPE, State),
1855 ARG_COPY_REF(D3DMATRIX, pMatrix))
1856 {
1857 struct nine_context *context = &device->context;
1858 D3DMATRIX *M = nine_state_access_transform(&context->ff, State, TRUE);
1859
1860 *M = *pMatrix;
1861 context->ff.changed.transform[State / 32] |= 1 << (State % 32);
1862 context->changed.group |= NINE_STATE_FF;
1863 }
1864
1865 CSMT_ITEM_NO_WAIT(nine_context_set_material,
1866 ARG_COPY_REF(D3DMATERIAL9, pMaterial))
1867 {
1868 struct nine_context *context = &device->context;
1869
1870 context->ff.material = *pMaterial;
1871 context->changed.group |= NINE_STATE_FF_MATERIAL;
1872 }
1873
1874 CSMT_ITEM_NO_WAIT(nine_context_set_light,
1875 ARG_VAL(DWORD, Index),
1876 ARG_COPY_REF(D3DLIGHT9, pLight))
1877 {
1878 struct nine_context *context = &device->context;
1879
1880 (void)nine_state_set_light(&context->ff, Index, pLight);
1881 context->changed.group |= NINE_STATE_FF_LIGHTING;
1882 }
1883
1884
1885 /* For stateblocks */
1886 static void
1887 nine_context_light_enable_stateblock(struct NineDevice9 *device,
1888 const uint16_t active_light[NINE_MAX_LIGHTS_ACTIVE], /* TODO: use pointer that convey size for csmt */
1889 unsigned int num_lights_active)
1890 {
1891 struct nine_context *context = &device->context;
1892
1893 /* TODO: Use CSMT_* to avoid calling nine_csmt_process */
1894 nine_csmt_process(device);
1895 memcpy(context->ff.active_light, active_light, NINE_MAX_LIGHTS_ACTIVE * sizeof(context->ff.active_light[0]));
1896 context->ff.num_lights_active = num_lights_active;
1897 context->changed.group |= NINE_STATE_FF_LIGHTING;
1898 }
1899
1900 CSMT_ITEM_NO_WAIT(nine_context_light_enable,
1901 ARG_VAL(DWORD, Index),
1902 ARG_VAL(BOOL, Enable))
1903 {
1904 struct nine_context *context = &device->context;
1905
1906 nine_state_light_enable(&context->ff, Index, Enable);
1907 context->changed.group |= NINE_STATE_FF_LIGHTING;
1908 }
1909
1910 CSMT_ITEM_NO_WAIT(nine_context_set_texture_stage_state,
1911 ARG_VAL(DWORD, Stage),
1912 ARG_VAL(D3DTEXTURESTAGESTATETYPE, Type),
1913 ARG_VAL(DWORD, Value))
1914 {
1915 struct nine_context *context = &device->context;
1916 int bumpmap_index = -1;
1917
1918 context->ff.tex_stage[Stage][Type] = Value;
1919 switch (Type) {
1920 case D3DTSS_BUMPENVMAT00:
1921 bumpmap_index = 4 * Stage;
1922 break;
1923 case D3DTSS_BUMPENVMAT01:
1924 bumpmap_index = 4 * Stage + 1;
1925 break;
1926 case D3DTSS_BUMPENVMAT10:
1927 bumpmap_index = 4 * Stage + 2;
1928 break;
1929 case D3DTSS_BUMPENVMAT11:
1930 bumpmap_index = 4 * Stage + 3;
1931 break;
1932 case D3DTSS_BUMPENVLSCALE:
1933 bumpmap_index = 4 * 8 + 2 * Stage;
1934 break;
1935 case D3DTSS_BUMPENVLOFFSET:
1936 bumpmap_index = 4 * 8 + 2 * Stage + 1;
1937 break;
1938 case D3DTSS_TEXTURETRANSFORMFLAGS:
1939 context->changed.group |= NINE_STATE_PS_PARAMS_MISC;
1940 break;
1941 default:
1942 break;
1943 }
1944
1945 if (bumpmap_index >= 0) {
1946 context->bumpmap_vars[bumpmap_index] = Value;
1947 context->changed.group |= NINE_STATE_PS_CONST;
1948 }
1949
1950 context->changed.group |= NINE_STATE_FF_PS_CONSTS;
1951 context->ff.changed.tex_stage[Stage][Type / 32] |= 1 << (Type % 32);
1952 }
1953
1954 CSMT_ITEM_NO_WAIT(nine_context_set_clip_plane,
1955 ARG_VAL(DWORD, Index),
1956 ARG_COPY_REF(struct nine_clipplane, pPlane))
1957 {
1958 struct nine_context *context = &device->context;
1959
1960 memcpy(&context->clip.ucp[Index][0], pPlane, sizeof(context->clip.ucp[0]));
1961 context->changed.ucp = TRUE;
1962 }
1963
1964 CSMT_ITEM_NO_WAIT(nine_context_set_swvp,
1965 ARG_VAL(boolean, swvp))
1966 {
1967 struct nine_context *context = &device->context;
1968
1969 context->swvp = swvp;
1970 context->changed.group |= NINE_STATE_SWVP;
1971 }
1972
1973 /* Do not write to nine_context directly. Slower,
1974 * but works with csmt. TODO: write a special csmt version that
1975 * would record the list of commands as much as possible,
1976 * and use the version above else.
1977 */
1978 void
1979 nine_context_apply_stateblock(struct NineDevice9 *device,
1980 const struct nine_state *src)
1981 {
1982 int i;
1983
1984 /* No need to apply src->changed.group, since all calls do
1985 * set context->changed.group */
1986
1987 for (i = 0; i < ARRAY_SIZE(src->changed.rs); ++i) {
1988 uint32_t m = src->changed.rs[i];
1989 while (m) {
1990 const int r = ffs(m) - 1;
1991 m &= ~(1 << r);
1992 nine_context_set_render_state(device, i * 32 + r, src->rs_advertised[i * 32 + r]);
1993 }
1994 }
1995
1996 /* Textures */
1997 if (src->changed.texture) {
1998 uint32_t m = src->changed.texture;
1999 unsigned s;
2000
2001 for (s = 0; m; ++s, m >>= 1) {
2002 struct NineBaseTexture9 *tex = src->texture[s];
2003 if (!(m & 1))
2004 continue;
2005 nine_context_set_texture(device, s, tex);
2006 }
2007 }
2008
2009 /* Sampler state */
2010 if (src->changed.group & NINE_STATE_SAMPLER) {
2011 unsigned s;
2012
2013 for (s = 0; s < NINE_MAX_SAMPLERS; ++s) {
2014 uint32_t m = src->changed.sampler[s];
2015 while (m) {
2016 const int i = ffs(m) - 1;
2017 m &= ~(1 << i);
2018 nine_context_set_sampler_state(device, s, i, src->samp_advertised[s][i]);
2019 }
2020 }
2021 }
2022
2023 /* Vertex buffers */
2024 if (src->changed.vtxbuf | src->changed.stream_freq) {
2025 uint32_t m = src->changed.vtxbuf | src->changed.stream_freq;
2026 for (i = 0; m; ++i, m >>= 1) {
2027 if (src->changed.vtxbuf & (1 << i))
2028 nine_context_set_stream_source(device, i, src->stream[i], src->vtxbuf[i].buffer_offset, src->vtxbuf[i].stride);
2029 if (src->changed.stream_freq & (1 << i))
2030 nine_context_set_stream_source_freq(device, i, src->stream_freq[i]);
2031 }
2032 }
2033
2034 /* Index buffer */
2035 if (src->changed.group & NINE_STATE_IDXBUF)
2036 nine_context_set_indices(device, src->idxbuf);
2037
2038 /* Vertex declaration */
2039 if ((src->changed.group & NINE_STATE_VDECL) && src->vdecl)
2040 nine_context_set_vertex_declaration(device, src->vdecl);
2041
2042 /* Vertex shader */
2043 if (src->changed.group & NINE_STATE_VS)
2044 nine_context_set_vertex_shader(device, src->vs);
2045
2046 /* Pixel shader */
2047 if (src->changed.group & NINE_STATE_PS)
2048 nine_context_set_pixel_shader(device, src->ps);
2049
2050 /* Vertex constants */
2051 if (src->changed.group & NINE_STATE_VS_CONST) {
2052 struct nine_range *r;
2053 for (r = src->changed.vs_const_f; r; r = r->next)
2054 nine_context_set_vertex_shader_constant_f(device, r->bgn,
2055 &src->vs_const_f[r->bgn * 4],
2056 sizeof(float[4]) * (r->end - r->bgn),
2057 r->end - r->bgn);
2058 for (r = src->changed.vs_const_i; r; r = r->next)
2059 nine_context_set_vertex_shader_constant_i(device, r->bgn,
2060 &src->vs_const_i[r->bgn * 4],
2061 sizeof(int[4]) * (r->end - r->bgn),
2062 r->end - r->bgn);
2063 for (r = src->changed.vs_const_b; r; r = r->next)
2064 nine_context_set_vertex_shader_constant_b(device, r->bgn,
2065 &src->vs_const_b[r->bgn * 4],
2066 sizeof(BOOL) * (r->end - r->bgn),
2067 r->end - r->bgn);
2068 }
2069
2070 /* Pixel constants */
2071 if (src->changed.group & NINE_STATE_PS_CONST) {
2072 struct nine_range *r;
2073 for (r = src->changed.ps_const_f; r; r = r->next)
2074 nine_context_set_pixel_shader_constant_f(device, r->bgn,
2075 &src->ps_const_f[r->bgn * 4],
2076 sizeof(float[4]) * (r->end - r->bgn),
2077 r->end - r->bgn);
2078 if (src->changed.ps_const_i) {
2079 uint16_t m = src->changed.ps_const_i;
2080 for (i = ffs(m) - 1, m >>= i; m; ++i, m >>= 1)
2081 if (m & 1)
2082 nine_context_set_pixel_shader_constant_i_transformed(device, i,
2083 src->ps_const_i[i], sizeof(int[4]), 1);
2084 }
2085 if (src->changed.ps_const_b) {
2086 uint16_t m = src->changed.ps_const_b;
2087 for (i = ffs(m) - 1, m >>= i; m; ++i, m >>= 1)
2088 if (m & 1)
2089 nine_context_set_pixel_shader_constant_b(device, i,
2090 &src->ps_const_b[i], sizeof(BOOL), 1);
2091 }
2092 }
2093
2094 /* Viewport */
2095 if (src->changed.group & NINE_STATE_VIEWPORT)
2096 nine_context_set_viewport(device, &src->viewport);
2097
2098 /* Scissor */
2099 if (src->changed.group & NINE_STATE_SCISSOR)
2100 nine_context_set_scissor(device, &src->scissor);
2101
2102 /* User Clip Planes */
2103 if (src->changed.ucp)
2104 for (i = 0; i < PIPE_MAX_CLIP_PLANES; ++i)
2105 if (src->changed.ucp & (1 << i))
2106 nine_context_set_clip_plane(device, i, (struct nine_clipplane*)&src->clip.ucp[i][0]);
2107
2108 if (!(src->changed.group & NINE_STATE_FF))
2109 return;
2110
2111 /* Fixed function state. */
2112
2113 if (src->changed.group & NINE_STATE_FF_MATERIAL)
2114 nine_context_set_material(device, &src->ff.material);
2115
2116 if (src->changed.group & NINE_STATE_FF_PS_CONSTS) {
2117 unsigned s;
2118 for (s = 0; s < NINE_MAX_TEXTURE_STAGES; ++s) {
2119 for (i = 0; i < NINED3DTSS_COUNT; ++i)
2120 if (src->ff.changed.tex_stage[s][i / 32] & (1 << (i % 32)))
2121 nine_context_set_texture_stage_state(device, s, i, src->ff.tex_stage[s][i]);
2122 }
2123 }
2124 if (src->changed.group & NINE_STATE_FF_LIGHTING) {
2125 for (i = 0; i < src->ff.num_lights; ++i)
2126 if (src->ff.light[i].Type != NINED3DLIGHT_INVALID)
2127 nine_context_set_light(device, i, &src->ff.light[i]);
2128
2129 nine_context_light_enable_stateblock(device, src->ff.active_light, src->ff.num_lights_active);
2130 }
2131 if (src->changed.group & NINE_STATE_FF_VSTRANSF) {
2132 for (i = 0; i < ARRAY_SIZE(src->ff.changed.transform); ++i) {
2133 unsigned s;
2134 if (!src->ff.changed.transform[i])
2135 continue;
2136 for (s = i * 32; s < (i * 32 + 32); ++s) {
2137 if (!(src->ff.changed.transform[i] & (1 << (s % 32))))
2138 continue;
2139 /* MaxVertexBlendMatrixIndex is 8, which means
2140 * we don't read past index D3DTS_WORLDMATRIX(8).
2141 * swvp is supposed to allow all 256, but we don't
2142 * implement it for now. */
2143 if (s > D3DTS_WORLDMATRIX(8))
2144 break;
2145 nine_context_set_transform(device, s,
2146 nine_state_access_transform(
2147 (struct nine_ff_state *)&src->ff,
2148 s, FALSE));
2149 }
2150 }
2151 }
2152 }
2153
2154 static void
2155 nine_update_state_framebuffer_clear(struct NineDevice9 *device)
2156 {
2157 struct nine_context *context = &device->context;
2158
2159 if (context->changed.group & NINE_STATE_FB)
2160 update_framebuffer(device, TRUE);
2161 }
2162
2163 CSMT_ITEM_NO_WAIT(nine_context_clear_fb,
2164 ARG_VAL(DWORD, Count),
2165 ARG_COPY_REF(D3DRECT, pRects),
2166 ARG_VAL(DWORD, Flags),
2167 ARG_VAL(D3DCOLOR, Color),
2168 ARG_VAL(float, Z),
2169 ARG_VAL(DWORD, Stencil))
2170 {
2171 struct nine_context *context = &device->context;
2172 const int sRGB = context->rs[D3DRS_SRGBWRITEENABLE] ? 1 : 0;
2173 struct pipe_surface *cbuf, *zsbuf;
2174 struct pipe_context *pipe = context->pipe;
2175 struct NineSurface9 *zsbuf_surf = context->ds;
2176 struct NineSurface9 *rt;
2177 unsigned bufs = 0;
2178 unsigned r, i;
2179 union pipe_color_union rgba;
2180 unsigned rt_mask = 0;
2181 D3DRECT rect;
2182
2183 nine_update_state_framebuffer_clear(device);
2184
2185 if (Flags & D3DCLEAR_TARGET) bufs |= PIPE_CLEAR_COLOR;
2186 /* Ignore Z buffer if not bound */
2187 if (context->pipe_data.fb.zsbuf != NULL) {
2188 if (Flags & D3DCLEAR_ZBUFFER) bufs |= PIPE_CLEAR_DEPTH;
2189 if (Flags & D3DCLEAR_STENCIL) bufs |= PIPE_CLEAR_STENCIL;
2190 }
2191 if (!bufs)
2192 return;
2193 d3dcolor_to_pipe_color_union(&rgba, Color);
2194
2195 rect.x1 = context->viewport.X;
2196 rect.y1 = context->viewport.Y;
2197 rect.x2 = context->viewport.Width + rect.x1;
2198 rect.y2 = context->viewport.Height + rect.y1;
2199
2200 /* Both rectangles apply, which is weird, but that's D3D9. */
2201 if (context->rs[D3DRS_SCISSORTESTENABLE]) {
2202 rect.x1 = MAX2(rect.x1, context->scissor.minx);
2203 rect.y1 = MAX2(rect.y1, context->scissor.miny);
2204 rect.x2 = MIN2(rect.x2, context->scissor.maxx);
2205 rect.y2 = MIN2(rect.y2, context->scissor.maxy);
2206 }
2207
2208 if (Count) {
2209 /* Maybe apps like to specify a large rect ? */
2210 if (pRects[0].x1 <= rect.x1 && pRects[0].x2 >= rect.x2 &&
2211 pRects[0].y1 <= rect.y1 && pRects[0].y2 >= rect.y2) {
2212 DBG("First rect covers viewport.\n");
2213 Count = 0;
2214 pRects = NULL;
2215 }
2216 }
2217
2218 if (rect.x1 >= context->pipe_data.fb.width || rect.y1 >= context->pipe_data.fb.height)
2219 return;
2220
2221 for (i = 0; i < device->caps.NumSimultaneousRTs; ++i) {
2222 if (context->rt[i] && context->rt[i]->desc.Format != D3DFMT_NULL)
2223 rt_mask |= 1 << i;
2224 }
2225
2226 /* fast path, clears everything at once */
2227 if (!Count &&
2228 (!(bufs & PIPE_CLEAR_COLOR) || (rt_mask == context->rt_mask)) &&
2229 rect.x1 == 0 && rect.y1 == 0 &&
2230 /* Case we clear only render target. Check clear region vs rt. */
2231 ((!(bufs & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) &&
2232 rect.x2 >= context->pipe_data.fb.width &&
2233 rect.y2 >= context->pipe_data.fb.height) ||
2234 /* Case we clear depth buffer (and eventually rt too).
2235 * depth buffer size is always >= rt size. Compare to clear region */
2236 ((bufs & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) &&
2237 rect.x2 >= zsbuf_surf->desc.Width &&
2238 rect.y2 >= zsbuf_surf->desc.Height))) {
2239 DBG("Clear fast path\n");
2240 pipe->clear(pipe, bufs, &rgba, Z, Stencil);
2241 return;
2242 }
2243
2244 if (!Count) {
2245 Count = 1;
2246 pRects = &rect;
2247 }
2248
2249 for (i = 0; i < device->caps.NumSimultaneousRTs; ++i) {
2250 rt = context->rt[i];
2251 if (!rt || rt->desc.Format == D3DFMT_NULL ||
2252 !(bufs & PIPE_CLEAR_COLOR))
2253 continue; /* save space, compiler should hoist this */
2254 cbuf = NineSurface9_GetSurface(rt, sRGB);
2255 for (r = 0; r < Count; ++r) {
2256 /* Don't trust users to pass these in the right order. */
2257 unsigned x1 = MIN2(pRects[r].x1, pRects[r].x2);
2258 unsigned y1 = MIN2(pRects[r].y1, pRects[r].y2);
2259 unsigned x2 = MAX2(pRects[r].x1, pRects[r].x2);
2260 unsigned y2 = MAX2(pRects[r].y1, pRects[r].y2);
2261 #ifndef NINE_LAX
2262 /* Drop negative rectangles (like wine expects). */
2263 if (pRects[r].x1 > pRects[r].x2) continue;
2264 if (pRects[r].y1 > pRects[r].y2) continue;
2265 #endif
2266
2267 x1 = MAX2(x1, rect.x1);
2268 y1 = MAX2(y1, rect.y1);
2269 x2 = MIN3(x2, rect.x2, rt->desc.Width);
2270 y2 = MIN3(y2, rect.y2, rt->desc.Height);
2271
2272 DBG("Clearing (%u..%u)x(%u..%u)\n", x1, x2, y1, y2);
2273 pipe->clear_render_target(pipe, cbuf, &rgba,
2274 x1, y1, x2 - x1, y2 - y1, false);
2275 }
2276 }
2277 if (!(bufs & PIPE_CLEAR_DEPTHSTENCIL))
2278 return;
2279
2280 bufs &= PIPE_CLEAR_DEPTHSTENCIL;
2281
2282 for (r = 0; r < Count; ++r) {
2283 unsigned x1 = MIN2(pRects[r].x1, pRects[r].x2);
2284 unsigned y1 = MIN2(pRects[r].y1, pRects[r].y2);
2285 unsigned x2 = MAX2(pRects[r].x1, pRects[r].x2);
2286 unsigned y2 = MAX2(pRects[r].y1, pRects[r].y2);
2287 #ifndef NINE_LAX
2288 /* Drop negative rectangles. */
2289 if (pRects[r].x1 > pRects[r].x2) continue;
2290 if (pRects[r].y1 > pRects[r].y2) continue;
2291 #endif
2292
2293 x1 = MIN2(x1, rect.x1);
2294 y1 = MIN2(y1, rect.y1);
2295 x2 = MIN3(x2, rect.x2, zsbuf_surf->desc.Width);
2296 y2 = MIN3(y2, rect.y2, zsbuf_surf->desc.Height);
2297
2298 zsbuf = NineSurface9_GetSurface(zsbuf_surf, 0);
2299 assert(zsbuf);
2300 pipe->clear_depth_stencil(pipe, zsbuf, bufs, Z, Stencil,
2301 x1, y1, x2 - x1, y2 - y1, false);
2302 }
2303 return;
2304 }
2305
2306
2307 static inline void
2308 init_draw_info(struct pipe_draw_info *info,
2309 struct NineDevice9 *dev, D3DPRIMITIVETYPE type, UINT count)
2310 {
2311 info->mode = d3dprimitivetype_to_pipe_prim(type);
2312 info->count = prim_count_to_vertex_count(type, count);
2313 info->start_instance = 0;
2314 info->instance_count = 1;
2315 if (dev->context.stream_instancedata_mask & dev->context.stream_usage_mask)
2316 info->instance_count = MAX2(dev->context.stream_freq[0] & 0x7FFFFF, 1);
2317 info->primitive_restart = FALSE;
2318 info->has_user_indices = FALSE;
2319 info->restart_index = 0;
2320 info->count_from_stream_output = NULL;
2321 info->indirect = NULL;
2322 }
2323
2324 CSMT_ITEM_NO_WAIT(nine_context_draw_primitive,
2325 ARG_VAL(D3DPRIMITIVETYPE, PrimitiveType),
2326 ARG_VAL(UINT, StartVertex),
2327 ARG_VAL(UINT, PrimitiveCount))
2328 {
2329 struct nine_context *context = &device->context;
2330 struct pipe_draw_info info;
2331
2332 nine_update_state(device);
2333
2334 init_draw_info(&info, device, PrimitiveType, PrimitiveCount);
2335 info.index_size = 0;
2336 info.start = StartVertex;
2337 info.index_bias = 0;
2338 info.min_index = info.start;
2339 info.max_index = info.count - 1;
2340 info.index.resource = NULL;
2341
2342 context->pipe->draw_vbo(context->pipe, &info);
2343 }
2344
2345 CSMT_ITEM_NO_WAIT(nine_context_draw_indexed_primitive,
2346 ARG_VAL(D3DPRIMITIVETYPE, PrimitiveType),
2347 ARG_VAL(INT, BaseVertexIndex),
2348 ARG_VAL(UINT, MinVertexIndex),
2349 ARG_VAL(UINT, NumVertices),
2350 ARG_VAL(UINT, StartIndex),
2351 ARG_VAL(UINT, PrimitiveCount))
2352 {
2353 struct nine_context *context = &device->context;
2354 struct pipe_draw_info info;
2355
2356 nine_update_state(device);
2357
2358 init_draw_info(&info, device, PrimitiveType, PrimitiveCount);
2359 info.index_size = context->index_size;
2360 info.start = context->index_offset / context->index_size + StartIndex;
2361 info.index_bias = BaseVertexIndex;
2362 /* These don't include index bias: */
2363 info.min_index = MinVertexIndex;
2364 info.max_index = MinVertexIndex + NumVertices - 1;
2365 info.index.resource = context->idxbuf;
2366
2367 context->pipe->draw_vbo(context->pipe, &info);
2368 }
2369
2370 CSMT_ITEM_NO_WAIT(nine_context_draw_primitive_from_vtxbuf,
2371 ARG_VAL(D3DPRIMITIVETYPE, PrimitiveType),
2372 ARG_VAL(UINT, PrimitiveCount),
2373 ARG_BIND_VBUF(struct pipe_vertex_buffer, vtxbuf))
2374 {
2375 struct nine_context *context = &device->context;
2376 struct pipe_draw_info info;
2377
2378 nine_update_state(device);
2379
2380 init_draw_info(&info, device, PrimitiveType, PrimitiveCount);
2381 info.index_size = 0;
2382 info.start = 0;
2383 info.index_bias = 0;
2384 info.min_index = 0;
2385 info.max_index = info.count - 1;
2386 info.index.resource = NULL;
2387
2388 context->pipe->set_vertex_buffers(context->pipe, 0, 1, vtxbuf);
2389
2390 context->pipe->draw_vbo(context->pipe, &info);
2391 }
2392
2393 CSMT_ITEM_NO_WAIT(nine_context_draw_indexed_primitive_from_vtxbuf_idxbuf,
2394 ARG_VAL(D3DPRIMITIVETYPE, PrimitiveType),
2395 ARG_VAL(UINT, MinVertexIndex),
2396 ARG_VAL(UINT, NumVertices),
2397 ARG_VAL(UINT, PrimitiveCount),
2398 ARG_BIND_VBUF(struct pipe_vertex_buffer, vbuf),
2399 ARG_BIND_RES(struct pipe_resource, ibuf),
2400 ARG_VAL(void *, user_ibuf),
2401 ARG_VAL(UINT, index_offset),
2402 ARG_VAL(UINT, index_size))
2403 {
2404 struct nine_context *context = &device->context;
2405 struct pipe_draw_info info;
2406
2407 nine_update_state(device);
2408
2409 init_draw_info(&info, device, PrimitiveType, PrimitiveCount);
2410 info.index_size = index_size;
2411 info.start = index_offset / info.index_size;
2412 info.index_bias = 0;
2413 info.min_index = MinVertexIndex;
2414 info.max_index = MinVertexIndex + NumVertices - 1;
2415 info.has_user_indices = ibuf == NULL;
2416 if (ibuf)
2417 info.index.resource = ibuf;
2418 else
2419 info.index.user = user_ibuf;
2420
2421 context->pipe->set_vertex_buffers(context->pipe, 0, 1, vbuf);
2422
2423 context->pipe->draw_vbo(context->pipe, &info);
2424 }
2425
2426 CSMT_ITEM_NO_WAIT(nine_context_resource_copy_region,
2427 ARG_BIND_REF(struct NineUnknown, dst),
2428 ARG_BIND_REF(struct NineUnknown, src),
2429 ARG_BIND_RES(struct pipe_resource, dst_res),
2430 ARG_VAL(unsigned, dst_level),
2431 ARG_COPY_REF(struct pipe_box, dst_box),
2432 ARG_BIND_RES(struct pipe_resource, src_res),
2433 ARG_VAL(unsigned, src_level),
2434 ARG_COPY_REF(struct pipe_box, src_box))
2435 {
2436 struct nine_context *context = &device->context;
2437
2438 (void) dst;
2439 (void) src;
2440
2441 context->pipe->resource_copy_region(context->pipe,
2442 dst_res, dst_level,
2443 dst_box->x, dst_box->y, dst_box->z,
2444 src_res, src_level,
2445 src_box);
2446 }
2447
2448 CSMT_ITEM_NO_WAIT(nine_context_blit,
2449 ARG_BIND_REF(struct NineUnknown, dst),
2450 ARG_BIND_REF(struct NineUnknown, src),
2451 ARG_BIND_BLIT(struct pipe_blit_info, blit))
2452 {
2453 struct nine_context *context = &device->context;
2454
2455 (void) dst;
2456 (void) src;
2457
2458 context->pipe->blit(context->pipe, blit);
2459 }
2460
2461 CSMT_ITEM_NO_WAIT(nine_context_clear_render_target,
2462 ARG_BIND_REF(struct NineSurface9, surface),
2463 ARG_VAL(D3DCOLOR, color),
2464 ARG_VAL(UINT, x),
2465 ARG_VAL(UINT, y),
2466 ARG_VAL(UINT, width),
2467 ARG_VAL(UINT, height))
2468 {
2469 struct nine_context *context = &device->context;
2470 struct pipe_surface *surf;
2471 union pipe_color_union rgba;
2472
2473 d3dcolor_to_pipe_color_union(&rgba, color);
2474 surf = NineSurface9_GetSurface(surface, 0);
2475 context->pipe->clear_render_target(context->pipe, surf, &rgba, x, y, width, height, false);
2476 }
2477
2478 CSMT_ITEM_NO_WAIT(nine_context_gen_mipmap,
2479 ARG_BIND_REF(struct NineUnknown, dst),
2480 ARG_BIND_RES(struct pipe_resource, res),
2481 ARG_VAL(UINT, base_level),
2482 ARG_VAL(UINT, last_level),
2483 ARG_VAL(UINT, first_layer),
2484 ARG_VAL(UINT, last_layer),
2485 ARG_VAL(UINT, filter))
2486 {
2487 struct nine_context *context = &device->context;
2488
2489 /* We just bind dst for the bind count */
2490 (void)dst;
2491
2492 util_gen_mipmap(context->pipe, res, res->format, base_level,
2493 last_level, first_layer, last_layer, filter);
2494 }
2495
2496 CSMT_ITEM_NO_WAIT_WITH_COUNTER(nine_context_range_upload,
2497 ARG_BIND_REF(struct NineUnknown, src_ref),
2498 ARG_BIND_RES(struct pipe_resource, res),
2499 ARG_VAL(unsigned, offset),
2500 ARG_VAL(unsigned, size),
2501 ARG_VAL(const void *, data))
2502 {
2503 struct nine_context *context = &device->context;
2504
2505 /* Binding src_ref avoids release before upload */
2506 (void)src_ref;
2507
2508 context->pipe->buffer_subdata(context->pipe, res, 0, offset, size, data);
2509 }
2510
2511 CSMT_ITEM_NO_WAIT_WITH_COUNTER(nine_context_box_upload,
2512 ARG_BIND_REF(struct NineUnknown, src_ref),
2513 ARG_BIND_RES(struct pipe_resource, res),
2514 ARG_VAL(unsigned, level),
2515 ARG_COPY_REF(struct pipe_box, dst_box),
2516 ARG_VAL(enum pipe_format, src_format),
2517 ARG_VAL(const void *, src),
2518 ARG_VAL(unsigned, src_stride),
2519 ARG_VAL(unsigned, src_layer_stride),
2520 ARG_COPY_REF(struct pipe_box, src_box))
2521 {
2522 struct nine_context *context = &device->context;
2523 struct pipe_context *pipe = context->pipe;
2524 struct pipe_transfer *transfer = NULL;
2525 uint8_t *map;
2526
2527 /* Binding src_ref avoids release before upload */
2528 (void)src_ref;
2529
2530 map = pipe->transfer_map(pipe,
2531 res,
2532 level,
2533 PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE,
2534 dst_box, &transfer);
2535 if (!map)
2536 return;
2537
2538 /* Note: if formats are the sames, it will revert
2539 * to normal memcpy */
2540 (void) util_format_translate_3d(res->format,
2541 map, transfer->stride,
2542 transfer->layer_stride,
2543 0, 0, 0,
2544 src_format,
2545 src, src_stride,
2546 src_layer_stride,
2547 src_box->x, src_box->y, src_box->z,
2548 dst_box->width, dst_box->height,
2549 dst_box->depth);
2550
2551 pipe_transfer_unmap(pipe, transfer);
2552 }
2553
2554 struct pipe_query *
2555 nine_context_create_query(struct NineDevice9 *device, unsigned query_type)
2556 {
2557 struct pipe_context *pipe;
2558 struct pipe_query *res;
2559
2560 pipe = nine_context_get_pipe_acquire(device);
2561 res = pipe->create_query(pipe, query_type, 0);
2562 nine_context_get_pipe_release(device);
2563 return res;
2564 }
2565
2566 CSMT_ITEM_DO_WAIT(nine_context_destroy_query,
2567 ARG_REF(struct pipe_query, query))
2568 {
2569 struct nine_context *context = &device->context;
2570
2571 context->pipe->destroy_query(context->pipe, query);
2572 }
2573
2574 CSMT_ITEM_NO_WAIT_WITH_COUNTER(nine_context_begin_query,
2575 ARG_REF(struct pipe_query, query))
2576 {
2577 struct nine_context *context = &device->context;
2578
2579 (void) context->pipe->begin_query(context->pipe, query);
2580 }
2581
2582 CSMT_ITEM_NO_WAIT_WITH_COUNTER(nine_context_end_query,
2583 ARG_REF(struct pipe_query, query))
2584 {
2585 struct nine_context *context = &device->context;
2586
2587 (void) context->pipe->end_query(context->pipe, query);
2588 }
2589
2590 boolean
2591 nine_context_get_query_result(struct NineDevice9 *device, struct pipe_query *query,
2592 unsigned *counter, boolean flush, boolean wait,
2593 union pipe_query_result *result)
2594 {
2595 struct pipe_context *pipe;
2596 boolean ret;
2597
2598 if (wait)
2599 nine_csmt_process(device);
2600 else if (p_atomic_read(counter) > 0) {
2601 if (flush && device->csmt_active)
2602 nine_queue_flush(device->csmt_ctx->pool);
2603 DBG("Pending begin/end. Returning\n");
2604 return false;
2605 }
2606
2607 pipe = nine_context_get_pipe_acquire(device);
2608 ret = pipe->get_query_result(pipe, query, wait, result);
2609 nine_context_get_pipe_release(device);
2610
2611 DBG("Query result %s\n", ret ? "found" : "not yet available");
2612 return ret;
2613 }
2614
2615 /* State defaults */
2616
2617 static const DWORD nine_render_state_defaults[NINED3DRS_LAST + 1] =
2618 {
2619 /* [D3DRS_ZENABLE] = D3DZB_TRUE; wine: auto_depth_stencil */
2620 [D3DRS_ZENABLE] = D3DZB_FALSE,
2621 [D3DRS_FILLMODE] = D3DFILL_SOLID,
2622 [D3DRS_SHADEMODE] = D3DSHADE_GOURAUD,
2623 /* [D3DRS_LINEPATTERN] = 0x00000000, */
2624 [D3DRS_ZWRITEENABLE] = TRUE,
2625 [D3DRS_ALPHATESTENABLE] = FALSE,
2626 [D3DRS_LASTPIXEL] = TRUE,
2627 [D3DRS_SRCBLEND] = D3DBLEND_ONE,
2628 [D3DRS_DESTBLEND] = D3DBLEND_ZERO,
2629 [D3DRS_CULLMODE] = D3DCULL_CCW,
2630 [D3DRS_ZFUNC] = D3DCMP_LESSEQUAL,
2631 [D3DRS_ALPHAFUNC] = D3DCMP_ALWAYS,
2632 [D3DRS_ALPHAREF] = 0,
2633 [D3DRS_DITHERENABLE] = FALSE,
2634 [D3DRS_ALPHABLENDENABLE] = FALSE,
2635 [D3DRS_FOGENABLE] = FALSE,
2636 [D3DRS_SPECULARENABLE] = FALSE,
2637 /* [D3DRS_ZVISIBLE] = 0, */
2638 [D3DRS_FOGCOLOR] = 0,
2639 [D3DRS_FOGTABLEMODE] = D3DFOG_NONE,
2640 [D3DRS_FOGSTART] = 0x00000000,
2641 [D3DRS_FOGEND] = 0x3F800000,
2642 [D3DRS_FOGDENSITY] = 0x3F800000,
2643 /* [D3DRS_EDGEANTIALIAS] = FALSE, */
2644 [D3DRS_RANGEFOGENABLE] = FALSE,
2645 [D3DRS_STENCILENABLE] = FALSE,
2646 [D3DRS_STENCILFAIL] = D3DSTENCILOP_KEEP,
2647 [D3DRS_STENCILZFAIL] = D3DSTENCILOP_KEEP,
2648 [D3DRS_STENCILPASS] = D3DSTENCILOP_KEEP,
2649 [D3DRS_STENCILREF] = 0,
2650 [D3DRS_STENCILMASK] = 0xFFFFFFFF,
2651 [D3DRS_STENCILFUNC] = D3DCMP_ALWAYS,
2652 [D3DRS_STENCILWRITEMASK] = 0xFFFFFFFF,
2653 [D3DRS_TEXTUREFACTOR] = 0xFFFFFFFF,
2654 [D3DRS_WRAP0] = 0,
2655 [D3DRS_WRAP1] = 0,
2656 [D3DRS_WRAP2] = 0,
2657 [D3DRS_WRAP3] = 0,
2658 [D3DRS_WRAP4] = 0,
2659 [D3DRS_WRAP5] = 0,
2660 [D3DRS_WRAP6] = 0,
2661 [D3DRS_WRAP7] = 0,
2662 [D3DRS_CLIPPING] = TRUE,
2663 [D3DRS_LIGHTING] = TRUE,
2664 [D3DRS_AMBIENT] = 0,
2665 [D3DRS_FOGVERTEXMODE] = D3DFOG_NONE,
2666 [D3DRS_COLORVERTEX] = TRUE,
2667 [D3DRS_LOCALVIEWER] = TRUE,
2668 [D3DRS_NORMALIZENORMALS] = FALSE,
2669 [D3DRS_DIFFUSEMATERIALSOURCE] = D3DMCS_COLOR1,
2670 [D3DRS_SPECULARMATERIALSOURCE] = D3DMCS_COLOR2,
2671 [D3DRS_AMBIENTMATERIALSOURCE] = D3DMCS_MATERIAL,
2672 [D3DRS_EMISSIVEMATERIALSOURCE] = D3DMCS_MATERIAL,
2673 [D3DRS_VERTEXBLEND] = D3DVBF_DISABLE,
2674 [D3DRS_CLIPPLANEENABLE] = 0,
2675 /* [D3DRS_SOFTWAREVERTEXPROCESSING] = FALSE, */
2676 [D3DRS_POINTSIZE] = 0x3F800000,
2677 [D3DRS_POINTSIZE_MIN] = 0x3F800000,
2678 [D3DRS_POINTSPRITEENABLE] = FALSE,
2679 [D3DRS_POINTSCALEENABLE] = FALSE,
2680 [D3DRS_POINTSCALE_A] = 0x3F800000,
2681 [D3DRS_POINTSCALE_B] = 0x00000000,
2682 [D3DRS_POINTSCALE_C] = 0x00000000,
2683 [D3DRS_MULTISAMPLEANTIALIAS] = TRUE,
2684 [D3DRS_MULTISAMPLEMASK] = 0xFFFFFFFF,
2685 [D3DRS_PATCHEDGESTYLE] = D3DPATCHEDGE_DISCRETE,
2686 /* [D3DRS_PATCHSEGMENTS] = 0x3F800000, */
2687 [D3DRS_DEBUGMONITORTOKEN] = 0xDEADCAFE,
2688 [D3DRS_POINTSIZE_MAX] = 0x3F800000, /* depends on cap */
2689 [D3DRS_INDEXEDVERTEXBLENDENABLE] = FALSE,
2690 [D3DRS_COLORWRITEENABLE] = 0x0000000f,
2691 [D3DRS_TWEENFACTOR] = 0x00000000,
2692 [D3DRS_BLENDOP] = D3DBLENDOP_ADD,
2693 [D3DRS_POSITIONDEGREE] = D3DDEGREE_CUBIC,
2694 [D3DRS_NORMALDEGREE] = D3DDEGREE_LINEAR,
2695 [D3DRS_SCISSORTESTENABLE] = FALSE,
2696 [D3DRS_SLOPESCALEDEPTHBIAS] = 0,
2697 [D3DRS_MINTESSELLATIONLEVEL] = 0x3F800000,
2698 [D3DRS_MAXTESSELLATIONLEVEL] = 0x3F800000,
2699 [D3DRS_ANTIALIASEDLINEENABLE] = FALSE,
2700 [D3DRS_ADAPTIVETESS_X] = 0x00000000,
2701 [D3DRS_ADAPTIVETESS_Y] = 0x00000000,
2702 [D3DRS_ADAPTIVETESS_Z] = 0x3F800000,
2703 [D3DRS_ADAPTIVETESS_W] = 0x00000000,
2704 [D3DRS_ENABLEADAPTIVETESSELLATION] = FALSE,
2705 [D3DRS_TWOSIDEDSTENCILMODE] = FALSE,
2706 [D3DRS_CCW_STENCILFAIL] = D3DSTENCILOP_KEEP,
2707 [D3DRS_CCW_STENCILZFAIL] = D3DSTENCILOP_KEEP,
2708 [D3DRS_CCW_STENCILPASS] = D3DSTENCILOP_KEEP,
2709 [D3DRS_CCW_STENCILFUNC] = D3DCMP_ALWAYS,
2710 [D3DRS_COLORWRITEENABLE1] = 0x0000000F,
2711 [D3DRS_COLORWRITEENABLE2] = 0x0000000F,
2712 [D3DRS_COLORWRITEENABLE3] = 0x0000000F,
2713 [D3DRS_BLENDFACTOR] = 0xFFFFFFFF,
2714 [D3DRS_SRGBWRITEENABLE] = 0,
2715 [D3DRS_DEPTHBIAS] = 0,
2716 [D3DRS_WRAP8] = 0,
2717 [D3DRS_WRAP9] = 0,
2718 [D3DRS_WRAP10] = 0,
2719 [D3DRS_WRAP11] = 0,
2720 [D3DRS_WRAP12] = 0,
2721 [D3DRS_WRAP13] = 0,
2722 [D3DRS_WRAP14] = 0,
2723 [D3DRS_WRAP15] = 0,
2724 [D3DRS_SEPARATEALPHABLENDENABLE] = FALSE,
2725 [D3DRS_SRCBLENDALPHA] = D3DBLEND_ONE,
2726 [D3DRS_DESTBLENDALPHA] = D3DBLEND_ZERO,
2727 [D3DRS_BLENDOPALPHA] = D3DBLENDOP_ADD,
2728 [NINED3DRS_VSPOINTSIZE] = FALSE,
2729 [NINED3DRS_RTMASK] = 0xf,
2730 [NINED3DRS_ALPHACOVERAGE] = FALSE,
2731 [NINED3DRS_MULTISAMPLE] = FALSE
2732 };
2733 static const DWORD nine_tex_stage_state_defaults[NINED3DTSS_LAST + 1] =
2734 {
2735 [D3DTSS_COLOROP] = D3DTOP_DISABLE,
2736 [D3DTSS_ALPHAOP] = D3DTOP_DISABLE,
2737 [D3DTSS_COLORARG1] = D3DTA_TEXTURE,
2738 [D3DTSS_COLORARG2] = D3DTA_CURRENT,
2739 [D3DTSS_COLORARG0] = D3DTA_CURRENT,
2740 [D3DTSS_ALPHAARG1] = D3DTA_TEXTURE,
2741 [D3DTSS_ALPHAARG2] = D3DTA_CURRENT,
2742 [D3DTSS_ALPHAARG0] = D3DTA_CURRENT,
2743 [D3DTSS_RESULTARG] = D3DTA_CURRENT,
2744 [D3DTSS_BUMPENVMAT00] = 0,
2745 [D3DTSS_BUMPENVMAT01] = 0,
2746 [D3DTSS_BUMPENVMAT10] = 0,
2747 [D3DTSS_BUMPENVMAT11] = 0,
2748 [D3DTSS_BUMPENVLSCALE] = 0,
2749 [D3DTSS_BUMPENVLOFFSET] = 0,
2750 [D3DTSS_TEXCOORDINDEX] = 0,
2751 [D3DTSS_TEXTURETRANSFORMFLAGS] = D3DTTFF_DISABLE,
2752 };
2753 static const DWORD nine_samp_state_defaults[NINED3DSAMP_LAST + 1] =
2754 {
2755 [D3DSAMP_ADDRESSU] = D3DTADDRESS_WRAP,
2756 [D3DSAMP_ADDRESSV] = D3DTADDRESS_WRAP,
2757 [D3DSAMP_ADDRESSW] = D3DTADDRESS_WRAP,
2758 [D3DSAMP_BORDERCOLOR] = 0,
2759 [D3DSAMP_MAGFILTER] = D3DTEXF_POINT,
2760 [D3DSAMP_MINFILTER] = D3DTEXF_POINT,
2761 [D3DSAMP_MIPFILTER] = D3DTEXF_NONE,
2762 [D3DSAMP_MIPMAPLODBIAS] = 0,
2763 [D3DSAMP_MAXMIPLEVEL] = 0,
2764 [D3DSAMP_MAXANISOTROPY] = 1,
2765 [D3DSAMP_SRGBTEXTURE] = 0,
2766 [D3DSAMP_ELEMENTINDEX] = 0,
2767 [D3DSAMP_DMAPOFFSET] = 0,
2768 [NINED3DSAMP_MINLOD] = 0,
2769 [NINED3DSAMP_SHADOW] = 0,
2770 [NINED3DSAMP_CUBETEX] = 0
2771 };
2772
2773 /* Note: The following 4 functions assume there is no
2774 * pending commands */
2775
2776 void nine_state_restore_non_cso(struct NineDevice9 *device)
2777 {
2778 struct nine_context *context = &device->context;
2779
2780 context->changed.group = NINE_STATE_ALL;
2781 context->changed.vtxbuf = (1ULL << device->caps.MaxStreams) - 1;
2782 context->changed.ucp = TRUE;
2783 context->commit |= NINE_STATE_COMMIT_CONST_VS | NINE_STATE_COMMIT_CONST_PS;
2784 }
2785
2786 void
2787 nine_state_set_defaults(struct NineDevice9 *device, const D3DCAPS9 *caps,
2788 boolean is_reset)
2789 {
2790 struct nine_state *state = &device->state;
2791 struct nine_context *context = &device->context;
2792 unsigned s;
2793
2794 /* Initialize defaults.
2795 */
2796 memcpy(context->rs, nine_render_state_defaults, sizeof(context->rs));
2797
2798 for (s = 0; s < ARRAY_SIZE(state->ff.tex_stage); ++s) {
2799 memcpy(&state->ff.tex_stage[s], nine_tex_stage_state_defaults,
2800 sizeof(state->ff.tex_stage[s]));
2801 state->ff.tex_stage[s][D3DTSS_TEXCOORDINDEX] = s;
2802 }
2803 state->ff.tex_stage[0][D3DTSS_COLOROP] = D3DTOP_MODULATE;
2804 state->ff.tex_stage[0][D3DTSS_ALPHAOP] = D3DTOP_SELECTARG1;
2805
2806 for (s = 0; s < ARRAY_SIZE(state->ff.tex_stage); ++s)
2807 memcpy(&context->ff.tex_stage[s], state->ff.tex_stage[s],
2808 sizeof(state->ff.tex_stage[s]));
2809
2810 memset(&context->bumpmap_vars, 0, sizeof(context->bumpmap_vars));
2811
2812 for (s = 0; s < NINE_MAX_SAMPLERS; ++s) {
2813 memcpy(&context->samp[s], nine_samp_state_defaults,
2814 sizeof(context->samp[s]));
2815 memcpy(&state->samp_advertised[s], nine_samp_state_defaults,
2816 sizeof(state->samp_advertised[s]));
2817 }
2818
2819 memset(state->vs_const_f, 0, VS_CONST_F_SIZE(device));
2820 memset(context->vs_const_f, 0, device->vs_const_size);
2821 if (context->vs_const_f_swvp)
2822 memset(context->vs_const_f_swvp, 0, NINE_MAX_CONST_F_SWVP * sizeof(float[4]));
2823 memset(state->vs_const_i, 0, VS_CONST_I_SIZE(device));
2824 memset(context->vs_const_i, 0, VS_CONST_I_SIZE(device));
2825 memset(state->vs_const_b, 0, VS_CONST_B_SIZE(device));
2826 memset(context->vs_const_b, 0, VS_CONST_B_SIZE(device));
2827 memset(state->ps_const_f, 0, device->ps_const_size);
2828 memset(context->ps_const_f, 0, device->ps_const_size);
2829 memset(state->ps_const_i, 0, sizeof(state->ps_const_i));
2830 memset(context->ps_const_i, 0, sizeof(context->ps_const_i));
2831 memset(state->ps_const_b, 0, sizeof(state->ps_const_b));
2832 memset(context->ps_const_b, 0, sizeof(context->ps_const_b));
2833
2834 /* Cap dependent initial state:
2835 */
2836 context->rs[D3DRS_POINTSIZE_MAX] = fui(caps->MaxPointSize);
2837
2838 memcpy(state->rs_advertised, context->rs, sizeof(context->rs));
2839
2840 /* Set changed flags to initialize driver.
2841 */
2842 context->changed.group = NINE_STATE_ALL;
2843 context->changed.vtxbuf = (1ULL << device->caps.MaxStreams) - 1;
2844 context->changed.ucp = TRUE;
2845
2846 context->ff.changed.transform[0] = ~0;
2847 context->ff.changed.transform[D3DTS_WORLD / 32] |= 1 << (D3DTS_WORLD % 32);
2848
2849 if (!is_reset) {
2850 state->viewport.MinZ = context->viewport.MinZ = 0.0f;
2851 state->viewport.MaxZ = context->viewport.MaxZ = 1.0f;
2852 }
2853
2854 for (s = 0; s < NINE_MAX_SAMPLERS; ++s)
2855 context->changed.sampler[s] = ~0;
2856
2857 if (!is_reset) {
2858 context->dummy_vbo_bound_at = -1;
2859 context->vbo_bound_done = FALSE;
2860 }
2861 }
2862
2863 void
2864 nine_device_state_clear(struct NineDevice9 *device)
2865 {
2866 struct nine_state *state = &device->state;
2867 unsigned i;
2868
2869 for (i = 0; i < ARRAY_SIZE(state->rt); ++i)
2870 nine_bind(&state->rt[i], NULL);
2871 nine_bind(&state->ds, NULL);
2872 nine_bind(&state->vs, NULL);
2873 nine_bind(&state->ps, NULL);
2874 nine_bind(&state->vdecl, NULL);
2875 for (i = 0; i < PIPE_MAX_ATTRIBS; ++i)
2876 NineBindBufferToDevice(device,
2877 (struct NineBuffer9 **)&state->stream[i],
2878 NULL);
2879 NineBindBufferToDevice(device,
2880 (struct NineBuffer9 **)&state->idxbuf,
2881 NULL);
2882
2883 for (i = 0; i < NINE_MAX_SAMPLERS; ++i)
2884 NineBindTextureToDevice(device, &state->texture[i], NULL);
2885 }
2886
2887 void
2888 nine_context_clear(struct NineDevice9 *device)
2889 {
2890 struct nine_context *context = &device->context;
2891 struct pipe_context *pipe = context->pipe;
2892 struct cso_context *cso = context->cso;
2893 unsigned i;
2894
2895 /* Early device ctor failure. Nothing to do */
2896 if (!pipe || !cso)
2897 return;
2898
2899 pipe->bind_vs_state(pipe, NULL);
2900 pipe->bind_fs_state(pipe, NULL);
2901
2902 /* Don't unbind constant buffers, they're device-private and
2903 * do not change on Reset.
2904 */
2905
2906 cso_set_samplers(cso, PIPE_SHADER_VERTEX, 0, NULL);
2907 cso_set_samplers(cso, PIPE_SHADER_FRAGMENT, 0, NULL);
2908
2909 cso_set_sampler_views(cso, PIPE_SHADER_VERTEX, 0, NULL);
2910 cso_set_sampler_views(cso, PIPE_SHADER_FRAGMENT, 0, NULL);
2911
2912 pipe->set_vertex_buffers(pipe, 0, device->caps.MaxStreams, NULL);
2913
2914 for (i = 0; i < ARRAY_SIZE(context->rt); ++i)
2915 nine_bind(&context->rt[i], NULL);
2916 nine_bind(&context->ds, NULL);
2917 nine_bind(&context->vs, NULL);
2918 nine_bind(&context->ps, NULL);
2919 nine_bind(&context->vdecl, NULL);
2920 for (i = 0; i < PIPE_MAX_ATTRIBS; ++i)
2921 pipe_vertex_buffer_unreference(&context->vtxbuf[i]);
2922 pipe_resource_reference(&context->idxbuf, NULL);
2923 pipe_resource_reference(&context->pipe_data.cb_vs.buffer, NULL);
2924 pipe_resource_reference(&context->pipe_data.cb_ps.buffer, NULL);
2925
2926 for (i = 0; i < NINE_MAX_SAMPLERS; ++i) {
2927 context->texture[i].enabled = FALSE;
2928 pipe_resource_reference(&context->texture[i].resource,
2929 NULL);
2930 pipe_sampler_view_reference(&context->texture[i].view[0],
2931 NULL);
2932 pipe_sampler_view_reference(&context->texture[i].view[1],
2933 NULL);
2934 }
2935 }
2936
2937 void
2938 nine_state_init_sw(struct NineDevice9 *device)
2939 {
2940 struct pipe_context *pipe_sw = device->pipe_sw;
2941 struct pipe_rasterizer_state rast;
2942 struct pipe_blend_state blend;
2943 struct pipe_depth_stencil_alpha_state dsa;
2944 struct pipe_framebuffer_state fb;
2945
2946 /* Only used with Streamout */
2947 memset(&rast, 0, sizeof(rast));
2948 rast.rasterizer_discard = true;
2949 rast.point_quad_rasterization = 1; /* to make llvmpipe happy */
2950 cso_set_rasterizer(device->cso_sw, &rast);
2951
2952 /* dummy settings */
2953 memset(&blend, 0, sizeof(blend));
2954 memset(&dsa, 0, sizeof(dsa));
2955 memset(&fb, 0, sizeof(fb));
2956 cso_set_blend(device->cso_sw, &blend);
2957 cso_set_depth_stencil_alpha(device->cso_sw, &dsa);
2958 cso_set_framebuffer(device->cso_sw, &fb);
2959 cso_set_viewport_dims(device->cso_sw, 1.0, 1.0, false);
2960 cso_set_fragment_shader_handle(device->cso_sw, util_make_empty_fragment_shader(pipe_sw));
2961 }
2962
2963 /* There is duplication with update_vertex_elements.
2964 * TODO: Share the code */
2965
2966 static void
2967 update_vertex_elements_sw(struct NineDevice9 *device)
2968 {
2969 struct nine_state *state = &device->state;
2970 const struct NineVertexDeclaration9 *vdecl = device->state.vdecl;
2971 const struct NineVertexShader9 *vs;
2972 unsigned n, b, i;
2973 int index;
2974 char vdecl_index_map[16]; /* vs->num_inputs <= 16 */
2975 char used_streams[device->caps.MaxStreams];
2976 int dummy_vbo_stream = -1;
2977 BOOL need_dummy_vbo = FALSE;
2978 struct pipe_vertex_element ve[PIPE_MAX_ATTRIBS];
2979 bool programmable_vs = state->vs && !(state->vdecl && state->vdecl->position_t);
2980
2981 memset(vdecl_index_map, -1, 16);
2982 memset(used_streams, 0, device->caps.MaxStreams);
2983 vs = programmable_vs ? device->state.vs : device->ff.vs;
2984
2985 if (vdecl) {
2986 for (n = 0; n < vs->num_inputs; ++n) {
2987 DBG("looking up input %u (usage %u) from vdecl(%p)\n",
2988 n, vs->input_map[n].ndecl, vdecl);
2989
2990 for (i = 0; i < vdecl->nelems; i++) {
2991 if (vdecl->usage_map[i] == vs->input_map[n].ndecl) {
2992 vdecl_index_map[n] = i;
2993 used_streams[vdecl->elems[i].vertex_buffer_index] = 1;
2994 break;
2995 }
2996 }
2997 if (vdecl_index_map[n] < 0)
2998 need_dummy_vbo = TRUE;
2999 }
3000 } else {
3001 /* No vertex declaration. Likely will never happen in practice,
3002 * but we need not crash on this */
3003 need_dummy_vbo = TRUE;
3004 }
3005
3006 if (need_dummy_vbo) {
3007 for (i = 0; i < device->caps.MaxStreams; i++ ) {
3008 if (!used_streams[i]) {
3009 dummy_vbo_stream = i;
3010 break;
3011 }
3012 }
3013 }
3014 /* TODO handle dummy_vbo */
3015 assert (!need_dummy_vbo);
3016
3017 for (n = 0; n < vs->num_inputs; ++n) {
3018 index = vdecl_index_map[n];
3019 if (index >= 0) {
3020 ve[n] = vdecl->elems[index];
3021 b = ve[n].vertex_buffer_index;
3022 /* XXX wine just uses 1 here: */
3023 if (state->stream_freq[b] & D3DSTREAMSOURCE_INSTANCEDATA)
3024 ve[n].instance_divisor = state->stream_freq[b] & 0x7FFFFF;
3025 } else {
3026 /* if the vertex declaration is incomplete compared to what the
3027 * vertex shader needs, we bind a dummy vbo with 0 0 0 0.
3028 * This is not precised by the spec, but is the behaviour
3029 * tested on win */
3030 ve[n].vertex_buffer_index = dummy_vbo_stream;
3031 ve[n].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
3032 ve[n].src_offset = 0;
3033 ve[n].instance_divisor = 0;
3034 }
3035 }
3036
3037 cso_set_vertex_elements(device->cso_sw, vs->num_inputs, ve);
3038 }
3039
3040 static void
3041 update_vertex_buffers_sw(struct NineDevice9 *device, int start_vertice, int num_vertices)
3042 {
3043 struct pipe_context *pipe = nine_context_get_pipe_acquire(device);
3044 struct pipe_context *pipe_sw = device->pipe_sw;
3045 struct nine_state *state = &device->state;
3046 struct nine_state_sw_internal *sw_internal = &device->state_sw_internal;
3047 struct pipe_vertex_buffer vtxbuf;
3048 uint32_t mask = 0xf;
3049 unsigned i;
3050
3051 DBG("mask=%x\n", mask);
3052
3053 /* TODO: handle dummy_vbo_bound_at */
3054
3055 for (i = 0; mask; mask >>= 1, ++i) {
3056 if (mask & 1) {
3057 if (state->stream[i]) {
3058 unsigned offset;
3059 struct pipe_resource *buf;
3060 struct pipe_box box;
3061 void *userbuf;
3062
3063 vtxbuf = state->vtxbuf[i];
3064 buf = NineVertexBuffer9_GetResource(state->stream[i], &offset);
3065
3066 DBG("Locking %p (offset %d, length %d)\n", buf,
3067 vtxbuf.buffer_offset, num_vertices * vtxbuf.stride);
3068
3069 u_box_1d(vtxbuf.buffer_offset + offset + start_vertice * vtxbuf.stride,
3070 num_vertices * vtxbuf.stride, &box);
3071
3072 userbuf = pipe->transfer_map(pipe, buf, 0, PIPE_TRANSFER_READ, &box,
3073 &(sw_internal->transfers_so[i]));
3074 vtxbuf.is_user_buffer = true;
3075 vtxbuf.buffer.user = userbuf;
3076
3077 if (!device->driver_caps.user_sw_vbufs) {
3078 vtxbuf.buffer.resource = NULL;
3079 vtxbuf.is_user_buffer = false;
3080 u_upload_data(device->pipe_sw->stream_uploader,
3081 0,
3082 box.width,
3083 16,
3084 userbuf,
3085 &(vtxbuf.buffer_offset),
3086 &(vtxbuf.buffer.resource));
3087 u_upload_unmap(device->pipe_sw->stream_uploader);
3088 }
3089 pipe_sw->set_vertex_buffers(pipe_sw, i, 1, &vtxbuf);
3090 pipe_vertex_buffer_unreference(&vtxbuf);
3091 } else
3092 pipe_sw->set_vertex_buffers(pipe_sw, i, 1, NULL);
3093 }
3094 }
3095 nine_context_get_pipe_release(device);
3096 }
3097
3098 static void
3099 update_vs_constants_sw(struct NineDevice9 *device)
3100 {
3101 struct nine_state *state = &device->state;
3102 struct pipe_context *pipe_sw = device->pipe_sw;
3103
3104 DBG("updating\n");
3105
3106 {
3107 struct pipe_constant_buffer cb;
3108 const void *buf;
3109
3110 cb.buffer = NULL;
3111 cb.buffer_offset = 0;
3112 cb.buffer_size = 4096 * sizeof(float[4]);
3113 cb.user_buffer = state->vs_const_f;
3114
3115 if (state->vs->lconstf.ranges) {
3116 const struct nine_lconstf *lconstf = &device->state.vs->lconstf;
3117 const struct nine_range *r = lconstf->ranges;
3118 unsigned n = 0;
3119 float *dst = device->state.vs_lconstf_temp;
3120 float *src = (float *)cb.user_buffer;
3121 memcpy(dst, src, 8192 * sizeof(float[4]));
3122 while (r) {
3123 unsigned p = r->bgn;
3124 unsigned c = r->end - r->bgn;
3125 memcpy(&dst[p * 4], &lconstf->data[n * 4], c * 4 * sizeof(float));
3126 n += c;
3127 r = r->next;
3128 }
3129 cb.user_buffer = dst;
3130 }
3131
3132 buf = cb.user_buffer;
3133
3134 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 0, &cb);
3135 if (cb.buffer)
3136 pipe_resource_reference(&cb.buffer, NULL);
3137
3138 cb.user_buffer = (char *)buf + 4096 * sizeof(float[4]);
3139
3140 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 1, &cb);
3141 if (cb.buffer)
3142 pipe_resource_reference(&cb.buffer, NULL);
3143 }
3144
3145 {
3146 struct pipe_constant_buffer cb;
3147
3148 cb.buffer = NULL;
3149 cb.buffer_offset = 0;
3150 cb.buffer_size = 2048 * sizeof(float[4]);
3151 cb.user_buffer = state->vs_const_i;
3152
3153 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 2, &cb);
3154 if (cb.buffer)
3155 pipe_resource_reference(&cb.buffer, NULL);
3156 }
3157
3158 {
3159 struct pipe_constant_buffer cb;
3160
3161 cb.buffer = NULL;
3162 cb.buffer_offset = 0;
3163 cb.buffer_size = 512 * sizeof(float[4]);
3164 cb.user_buffer = state->vs_const_b;
3165
3166 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 3, &cb);
3167 if (cb.buffer)
3168 pipe_resource_reference(&cb.buffer, NULL);
3169 }
3170
3171 {
3172 struct pipe_constant_buffer cb;
3173 const D3DVIEWPORT9 *vport = &device->state.viewport;
3174 float viewport_data[8] = {(float)vport->Width * 0.5f,
3175 (float)vport->Height * -0.5f, vport->MaxZ - vport->MinZ, 0.f,
3176 (float)vport->Width * 0.5f + (float)vport->X,
3177 (float)vport->Height * 0.5f + (float)vport->Y,
3178 vport->MinZ, 0.f};
3179
3180 cb.buffer = NULL;
3181 cb.buffer_offset = 0;
3182 cb.buffer_size = 2 * sizeof(float[4]);
3183 cb.user_buffer = viewport_data;
3184
3185 {
3186 u_upload_data(device->pipe_sw->const_uploader,
3187 0,
3188 cb.buffer_size,
3189 16,
3190 cb.user_buffer,
3191 &(cb.buffer_offset),
3192 &(cb.buffer));
3193 u_upload_unmap(device->pipe_sw->const_uploader);
3194 cb.user_buffer = NULL;
3195 }
3196
3197 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 4, &cb);
3198 if (cb.buffer)
3199 pipe_resource_reference(&cb.buffer, NULL);
3200 }
3201
3202 }
3203
3204 void
3205 nine_state_prepare_draw_sw(struct NineDevice9 *device, struct NineVertexDeclaration9 *vdecl_out,
3206 int start_vertice, int num_vertices, struct pipe_stream_output_info *so)
3207 {
3208 struct nine_state *state = &device->state;
3209 bool programmable_vs = state->vs && !(state->vdecl && state->vdecl->position_t);
3210 struct NineVertexShader9 *vs = programmable_vs ? device->state.vs : device->ff.vs;
3211
3212 assert(programmable_vs);
3213
3214 DBG("Preparing draw\n");
3215 cso_set_vertex_shader_handle(device->cso_sw,
3216 NineVertexShader9_GetVariantProcessVertices(vs, vdecl_out, so));
3217 update_vertex_elements_sw(device);
3218 update_vertex_buffers_sw(device, start_vertice, num_vertices);
3219 update_vs_constants_sw(device);
3220 DBG("Preparation succeeded\n");
3221 }
3222
3223 void
3224 nine_state_after_draw_sw(struct NineDevice9 *device)
3225 {
3226 struct nine_state_sw_internal *sw_internal = &device->state_sw_internal;
3227 struct pipe_context *pipe = nine_context_get_pipe_acquire(device);
3228 struct pipe_context *pipe_sw = device->pipe_sw;
3229 int i;
3230
3231 for (i = 0; i < 4; i++) {
3232 pipe_sw->set_vertex_buffers(pipe_sw, i, 1, NULL);
3233 if (sw_internal->transfers_so[i])
3234 pipe->transfer_unmap(pipe, sw_internal->transfers_so[i]);
3235 sw_internal->transfers_so[i] = NULL;
3236 }
3237 nine_context_get_pipe_release(device);
3238 }
3239
3240 void
3241 nine_state_destroy_sw(struct NineDevice9 *device)
3242 {
3243 (void) device;
3244 /* Everything destroyed with cso */
3245 }
3246
3247 /*
3248 static const DWORD nine_render_states_pixel[] =
3249 {
3250 D3DRS_ALPHABLENDENABLE,
3251 D3DRS_ALPHAFUNC,
3252 D3DRS_ALPHAREF,
3253 D3DRS_ALPHATESTENABLE,
3254 D3DRS_ANTIALIASEDLINEENABLE,
3255 D3DRS_BLENDFACTOR,
3256 D3DRS_BLENDOP,
3257 D3DRS_BLENDOPALPHA,
3258 D3DRS_CCW_STENCILFAIL,
3259 D3DRS_CCW_STENCILPASS,
3260 D3DRS_CCW_STENCILZFAIL,
3261 D3DRS_COLORWRITEENABLE,
3262 D3DRS_COLORWRITEENABLE1,
3263 D3DRS_COLORWRITEENABLE2,
3264 D3DRS_COLORWRITEENABLE3,
3265 D3DRS_DEPTHBIAS,
3266 D3DRS_DESTBLEND,
3267 D3DRS_DESTBLENDALPHA,
3268 D3DRS_DITHERENABLE,
3269 D3DRS_FILLMODE,
3270 D3DRS_FOGDENSITY,
3271 D3DRS_FOGEND,
3272 D3DRS_FOGSTART,
3273 D3DRS_LASTPIXEL,
3274 D3DRS_SCISSORTESTENABLE,
3275 D3DRS_SEPARATEALPHABLENDENABLE,
3276 D3DRS_SHADEMODE,
3277 D3DRS_SLOPESCALEDEPTHBIAS,
3278 D3DRS_SRCBLEND,
3279 D3DRS_SRCBLENDALPHA,
3280 D3DRS_SRGBWRITEENABLE,
3281 D3DRS_STENCILENABLE,
3282 D3DRS_STENCILFAIL,
3283 D3DRS_STENCILFUNC,
3284 D3DRS_STENCILMASK,
3285 D3DRS_STENCILPASS,
3286 D3DRS_STENCILREF,
3287 D3DRS_STENCILWRITEMASK,
3288 D3DRS_STENCILZFAIL,
3289 D3DRS_TEXTUREFACTOR,
3290 D3DRS_TWOSIDEDSTENCILMODE,
3291 D3DRS_WRAP0,
3292 D3DRS_WRAP1,
3293 D3DRS_WRAP10,
3294 D3DRS_WRAP11,
3295 D3DRS_WRAP12,
3296 D3DRS_WRAP13,
3297 D3DRS_WRAP14,
3298 D3DRS_WRAP15,
3299 D3DRS_WRAP2,
3300 D3DRS_WRAP3,
3301 D3DRS_WRAP4,
3302 D3DRS_WRAP5,
3303 D3DRS_WRAP6,
3304 D3DRS_WRAP7,
3305 D3DRS_WRAP8,
3306 D3DRS_WRAP9,
3307 D3DRS_ZENABLE,
3308 D3DRS_ZFUNC,
3309 D3DRS_ZWRITEENABLE
3310 };
3311 */
3312 const uint32_t nine_render_states_pixel[(NINED3DRS_LAST + 31) / 32] =
3313 {
3314 0x0f99c380, 0x1ff00070, 0x00000000, 0x00000000,
3315 0x000000ff, 0xde01c900, 0x0003ffcf
3316 };
3317
3318 /*
3319 static const DWORD nine_render_states_vertex[] =
3320 {
3321 D3DRS_ADAPTIVETESS_W,
3322 D3DRS_ADAPTIVETESS_X,
3323 D3DRS_ADAPTIVETESS_Y,
3324 D3DRS_ADAPTIVETESS_Z,
3325 D3DRS_AMBIENT,
3326 D3DRS_AMBIENTMATERIALSOURCE,
3327 D3DRS_CLIPPING,
3328 D3DRS_CLIPPLANEENABLE,
3329 D3DRS_COLORVERTEX,
3330 D3DRS_CULLMODE,
3331 D3DRS_DIFFUSEMATERIALSOURCE,
3332 D3DRS_EMISSIVEMATERIALSOURCE,
3333 D3DRS_ENABLEADAPTIVETESSELLATION,
3334 D3DRS_FOGCOLOR,
3335 D3DRS_FOGDENSITY,
3336 D3DRS_FOGENABLE,
3337 D3DRS_FOGEND,
3338 D3DRS_FOGSTART,
3339 D3DRS_FOGTABLEMODE,
3340 D3DRS_FOGVERTEXMODE,
3341 D3DRS_INDEXEDVERTEXBLENDENABLE,
3342 D3DRS_LIGHTING,
3343 D3DRS_LOCALVIEWER,
3344 D3DRS_MAXTESSELLATIONLEVEL,
3345 D3DRS_MINTESSELLATIONLEVEL,
3346 D3DRS_MULTISAMPLEANTIALIAS,
3347 D3DRS_MULTISAMPLEMASK,
3348 D3DRS_NORMALDEGREE,
3349 D3DRS_NORMALIZENORMALS,
3350 D3DRS_PATCHEDGESTYLE,
3351 D3DRS_POINTSCALE_A,
3352 D3DRS_POINTSCALE_B,
3353 D3DRS_POINTSCALE_C,
3354 D3DRS_POINTSCALEENABLE,
3355 D3DRS_POINTSIZE,
3356 D3DRS_POINTSIZE_MAX,
3357 D3DRS_POINTSIZE_MIN,
3358 D3DRS_POINTSPRITEENABLE,
3359 D3DRS_POSITIONDEGREE,
3360 D3DRS_RANGEFOGENABLE,
3361 D3DRS_SHADEMODE,
3362 D3DRS_SPECULARENABLE,
3363 D3DRS_SPECULARMATERIALSOURCE,
3364 D3DRS_TWEENFACTOR,
3365 D3DRS_VERTEXBLEND
3366 };
3367 */
3368 const uint32_t nine_render_states_vertex[(NINED3DRS_LAST + 31) / 32] =
3369 {
3370 0x30400200, 0x0001007c, 0x00000000, 0x00000000,
3371 0xfd9efb00, 0x01fc34cf, 0x00000000
3372 };
3373
3374 /* TODO: put in the right values */
3375 const uint32_t nine_render_state_group[NINED3DRS_LAST + 1] =
3376 {
3377 [D3DRS_ZENABLE] = NINE_STATE_DSA | NINE_STATE_MULTISAMPLE,
3378 [D3DRS_FILLMODE] = NINE_STATE_RASTERIZER,
3379 [D3DRS_SHADEMODE] = NINE_STATE_RASTERIZER,
3380 [D3DRS_ZWRITEENABLE] = NINE_STATE_DSA,
3381 [D3DRS_ALPHATESTENABLE] = NINE_STATE_DSA,
3382 [D3DRS_LASTPIXEL] = NINE_STATE_RASTERIZER,
3383 [D3DRS_SRCBLEND] = NINE_STATE_BLEND,
3384 [D3DRS_DESTBLEND] = NINE_STATE_BLEND,
3385 [D3DRS_CULLMODE] = NINE_STATE_RASTERIZER,
3386 [D3DRS_ZFUNC] = NINE_STATE_DSA,
3387 [D3DRS_ALPHAREF] = NINE_STATE_DSA,
3388 [D3DRS_ALPHAFUNC] = NINE_STATE_DSA,
3389 [D3DRS_DITHERENABLE] = NINE_STATE_BLEND,
3390 [D3DRS_ALPHABLENDENABLE] = NINE_STATE_BLEND,
3391 [D3DRS_FOGENABLE] = NINE_STATE_FF_SHADER | NINE_STATE_VS_PARAMS_MISC | NINE_STATE_PS_PARAMS_MISC | NINE_STATE_PS_CONST,
3392 [D3DRS_SPECULARENABLE] = NINE_STATE_FF_LIGHTING,
3393 [D3DRS_FOGCOLOR] = NINE_STATE_FF_PS_CONSTS | NINE_STATE_PS_CONST,
3394 [D3DRS_FOGTABLEMODE] = NINE_STATE_FF_SHADER | NINE_STATE_PS_PARAMS_MISC | NINE_STATE_PS_CONST,
3395 [D3DRS_FOGSTART] = NINE_STATE_FF_VS_OTHER | NINE_STATE_FF_PS_CONSTS | NINE_STATE_PS_CONST,
3396 [D3DRS_FOGEND] = NINE_STATE_FF_VS_OTHER | NINE_STATE_FF_PS_CONSTS | NINE_STATE_PS_CONST,
3397 [D3DRS_FOGDENSITY] = NINE_STATE_FF_VS_OTHER | NINE_STATE_FF_PS_CONSTS | NINE_STATE_PS_CONST,
3398 [D3DRS_RANGEFOGENABLE] = NINE_STATE_FF_SHADER,
3399 [D3DRS_STENCILENABLE] = NINE_STATE_DSA | NINE_STATE_MULTISAMPLE,
3400 [D3DRS_STENCILFAIL] = NINE_STATE_DSA,
3401 [D3DRS_STENCILZFAIL] = NINE_STATE_DSA,
3402 [D3DRS_STENCILPASS] = NINE_STATE_DSA,
3403 [D3DRS_STENCILFUNC] = NINE_STATE_DSA,
3404 [D3DRS_STENCILREF] = NINE_STATE_STENCIL_REF,
3405 [D3DRS_STENCILMASK] = NINE_STATE_DSA,
3406 [D3DRS_STENCILWRITEMASK] = NINE_STATE_DSA,
3407 [D3DRS_TEXTUREFACTOR] = NINE_STATE_FF_PS_CONSTS,
3408 [D3DRS_WRAP0] = NINE_STATE_UNHANDLED, /* cylindrical wrap is crazy */
3409 [D3DRS_WRAP1] = NINE_STATE_UNHANDLED,
3410 [D3DRS_WRAP2] = NINE_STATE_UNHANDLED,
3411 [D3DRS_WRAP3] = NINE_STATE_UNHANDLED,
3412 [D3DRS_WRAP4] = NINE_STATE_UNHANDLED,
3413 [D3DRS_WRAP5] = NINE_STATE_UNHANDLED,
3414 [D3DRS_WRAP6] = NINE_STATE_UNHANDLED,
3415 [D3DRS_WRAP7] = NINE_STATE_UNHANDLED,
3416 [D3DRS_CLIPPING] = 0, /* software vertex processing only */
3417 [D3DRS_LIGHTING] = NINE_STATE_FF_LIGHTING,
3418 [D3DRS_AMBIENT] = NINE_STATE_FF_LIGHTING | NINE_STATE_FF_MATERIAL,
3419 [D3DRS_FOGVERTEXMODE] = NINE_STATE_FF_SHADER,
3420 [D3DRS_COLORVERTEX] = NINE_STATE_FF_LIGHTING,
3421 [D3DRS_LOCALVIEWER] = NINE_STATE_FF_LIGHTING,
3422 [D3DRS_NORMALIZENORMALS] = NINE_STATE_FF_SHADER,
3423 [D3DRS_DIFFUSEMATERIALSOURCE] = NINE_STATE_FF_LIGHTING,
3424 [D3DRS_SPECULARMATERIALSOURCE] = NINE_STATE_FF_LIGHTING,
3425 [D3DRS_AMBIENTMATERIALSOURCE] = NINE_STATE_FF_LIGHTING,
3426 [D3DRS_EMISSIVEMATERIALSOURCE] = NINE_STATE_FF_LIGHTING,
3427 [D3DRS_VERTEXBLEND] = NINE_STATE_FF_SHADER,
3428 [D3DRS_CLIPPLANEENABLE] = NINE_STATE_RASTERIZER,
3429 [D3DRS_POINTSIZE] = NINE_STATE_RASTERIZER | NINE_STATE_FF_VS_OTHER,
3430 [D3DRS_POINTSIZE_MIN] = NINE_STATE_RASTERIZER | NINE_STATE_FF_VS_OTHER | NINE_STATE_VS_PARAMS_MISC,
3431 [D3DRS_POINTSPRITEENABLE] = NINE_STATE_RASTERIZER,
3432 [D3DRS_POINTSCALEENABLE] = NINE_STATE_FF_SHADER,
3433 [D3DRS_POINTSCALE_A] = NINE_STATE_FF_VS_OTHER,
3434 [D3DRS_POINTSCALE_B] = NINE_STATE_FF_VS_OTHER,
3435 [D3DRS_POINTSCALE_C] = NINE_STATE_FF_VS_OTHER,
3436 [D3DRS_MULTISAMPLEANTIALIAS] = NINE_STATE_MULTISAMPLE,
3437 [D3DRS_MULTISAMPLEMASK] = NINE_STATE_SAMPLE_MASK,
3438 [D3DRS_PATCHEDGESTYLE] = NINE_STATE_UNHANDLED,
3439 [D3DRS_DEBUGMONITORTOKEN] = NINE_STATE_UNHANDLED,
3440 [D3DRS_POINTSIZE_MAX] = NINE_STATE_RASTERIZER | NINE_STATE_FF_VS_OTHER | NINE_STATE_VS_PARAMS_MISC,
3441 [D3DRS_INDEXEDVERTEXBLENDENABLE] = NINE_STATE_FF_SHADER,
3442 [D3DRS_COLORWRITEENABLE] = NINE_STATE_BLEND,
3443 [D3DRS_TWEENFACTOR] = NINE_STATE_FF_VS_OTHER,
3444 [D3DRS_BLENDOP] = NINE_STATE_BLEND,
3445 [D3DRS_POSITIONDEGREE] = NINE_STATE_UNHANDLED,
3446 [D3DRS_NORMALDEGREE] = NINE_STATE_UNHANDLED,
3447 [D3DRS_SCISSORTESTENABLE] = NINE_STATE_RASTERIZER,
3448 [D3DRS_SLOPESCALEDEPTHBIAS] = NINE_STATE_RASTERIZER,
3449 [D3DRS_ANTIALIASEDLINEENABLE] = NINE_STATE_RASTERIZER,
3450 [D3DRS_MINTESSELLATIONLEVEL] = NINE_STATE_UNHANDLED,
3451 [D3DRS_MAXTESSELLATIONLEVEL] = NINE_STATE_UNHANDLED,
3452 [D3DRS_ADAPTIVETESS_X] = NINE_STATE_UNHANDLED,
3453 [D3DRS_ADAPTIVETESS_Y] = NINE_STATE_UNHANDLED,
3454 [D3DRS_ADAPTIVETESS_Z] = NINE_STATE_UNHANDLED,
3455 [D3DRS_ADAPTIVETESS_W] = NINE_STATE_UNHANDLED,
3456 [D3DRS_ENABLEADAPTIVETESSELLATION] = NINE_STATE_UNHANDLED,
3457 [D3DRS_TWOSIDEDSTENCILMODE] = NINE_STATE_DSA,
3458 [D3DRS_CCW_STENCILFAIL] = NINE_STATE_DSA,
3459 [D3DRS_CCW_STENCILZFAIL] = NINE_STATE_DSA,
3460 [D3DRS_CCW_STENCILPASS] = NINE_STATE_DSA,
3461 [D3DRS_CCW_STENCILFUNC] = NINE_STATE_DSA,
3462 [D3DRS_COLORWRITEENABLE1] = NINE_STATE_BLEND,
3463 [D3DRS_COLORWRITEENABLE2] = NINE_STATE_BLEND,
3464 [D3DRS_COLORWRITEENABLE3] = NINE_STATE_BLEND,
3465 [D3DRS_BLENDFACTOR] = NINE_STATE_BLEND_COLOR,
3466 [D3DRS_SRGBWRITEENABLE] = NINE_STATE_FB,
3467 [D3DRS_DEPTHBIAS] = NINE_STATE_RASTERIZER,
3468 [D3DRS_WRAP8] = NINE_STATE_UNHANDLED, /* cylwrap has to be done via GP */
3469 [D3DRS_WRAP9] = NINE_STATE_UNHANDLED,
3470 [D3DRS_WRAP10] = NINE_STATE_UNHANDLED,
3471 [D3DRS_WRAP11] = NINE_STATE_UNHANDLED,
3472 [D3DRS_WRAP12] = NINE_STATE_UNHANDLED,
3473 [D3DRS_WRAP13] = NINE_STATE_UNHANDLED,
3474 [D3DRS_WRAP14] = NINE_STATE_UNHANDLED,
3475 [D3DRS_WRAP15] = NINE_STATE_UNHANDLED,
3476 [D3DRS_SEPARATEALPHABLENDENABLE] = NINE_STATE_BLEND,
3477 [D3DRS_SRCBLENDALPHA] = NINE_STATE_BLEND,
3478 [D3DRS_DESTBLENDALPHA] = NINE_STATE_BLEND,
3479 [D3DRS_BLENDOPALPHA] = NINE_STATE_BLEND
3480 };
3481
3482 /* Misc */
3483
3484 static D3DMATRIX nine_state_identity = { .m[0] = { 1, 0, 0, 0 },
3485 .m[1] = { 0, 1, 0, 0 },
3486 .m[2] = { 0, 0, 1, 0 },
3487 .m[3] = { 0, 0, 0, 1 } };
3488
3489 void
3490 nine_state_resize_transform(struct nine_ff_state *ff_state, unsigned N)
3491 {
3492 unsigned n = ff_state->num_transforms;
3493
3494 if (N <= n)
3495 return;
3496
3497 ff_state->transform = REALLOC(ff_state->transform,
3498 n * sizeof(D3DMATRIX),
3499 N * sizeof(D3DMATRIX));
3500 for (; n < N; ++n)
3501 ff_state->transform[n] = nine_state_identity;
3502 ff_state->num_transforms = N;
3503 }
3504
3505 D3DMATRIX *
3506 nine_state_access_transform(struct nine_ff_state *ff_state, D3DTRANSFORMSTATETYPE t,
3507 boolean alloc)
3508 {
3509 unsigned index;
3510
3511 switch (t) {
3512 case D3DTS_VIEW: index = 0; break;
3513 case D3DTS_PROJECTION: index = 1; break;
3514 case D3DTS_TEXTURE0: index = 2; break;
3515 case D3DTS_TEXTURE1: index = 3; break;
3516 case D3DTS_TEXTURE2: index = 4; break;
3517 case D3DTS_TEXTURE3: index = 5; break;
3518 case D3DTS_TEXTURE4: index = 6; break;
3519 case D3DTS_TEXTURE5: index = 7; break;
3520 case D3DTS_TEXTURE6: index = 8; break;
3521 case D3DTS_TEXTURE7: index = 9; break;
3522 default:
3523 if (!(t >= D3DTS_WORLDMATRIX(0) && t <= D3DTS_WORLDMATRIX(255)))
3524 return NULL;
3525 index = 10 + (t - D3DTS_WORLDMATRIX(0));
3526 break;
3527 }
3528
3529 if (index >= ff_state->num_transforms) {
3530 if (!alloc)
3531 return &nine_state_identity;
3532 nine_state_resize_transform(ff_state, index + 1);
3533 }
3534 return &ff_state->transform[index];
3535 }
3536
3537 HRESULT
3538 nine_state_set_light(struct nine_ff_state *ff_state, DWORD Index,
3539 const D3DLIGHT9 *pLight)
3540 {
3541 if (Index >= ff_state->num_lights) {
3542 unsigned n = ff_state->num_lights;
3543 unsigned N = Index + 1;
3544
3545 ff_state->light = REALLOC(ff_state->light, n * sizeof(D3DLIGHT9),
3546 N * sizeof(D3DLIGHT9));
3547 if (!ff_state->light)
3548 return E_OUTOFMEMORY;
3549 ff_state->num_lights = N;
3550
3551 for (; n < Index; ++n) {
3552 memset(&ff_state->light[n], 0, sizeof(D3DLIGHT9));
3553 ff_state->light[n].Type = (D3DLIGHTTYPE)NINED3DLIGHT_INVALID;
3554 }
3555 }
3556 ff_state->light[Index] = *pLight;
3557
3558 if (pLight->Type == D3DLIGHT_SPOT && pLight->Theta >= pLight->Phi) {
3559 DBG("Warning: clamping D3DLIGHT9.Theta\n");
3560 ff_state->light[Index].Theta = ff_state->light[Index].Phi;
3561 }
3562 return D3D_OK;
3563 }
3564
3565 HRESULT
3566 nine_state_light_enable(struct nine_ff_state *ff_state,
3567 DWORD Index, BOOL Enable)
3568 {
3569 unsigned i;
3570
3571 user_assert(Index < ff_state->num_lights, D3DERR_INVALIDCALL);
3572
3573 for (i = 0; i < ff_state->num_lights_active; ++i) {
3574 if (ff_state->active_light[i] == Index)
3575 break;
3576 }
3577
3578 if (Enable) {
3579 if (i < ff_state->num_lights_active)
3580 return D3D_OK;
3581 /* XXX wine thinks this should still succeed:
3582 */
3583 user_assert(i < NINE_MAX_LIGHTS_ACTIVE, D3DERR_INVALIDCALL);
3584
3585 ff_state->active_light[i] = Index;
3586 ff_state->num_lights_active++;
3587 } else {
3588 if (i == ff_state->num_lights_active)
3589 return D3D_OK;
3590 --ff_state->num_lights_active;
3591 for (; i < ff_state->num_lights_active; ++i)
3592 ff_state->active_light[i] = ff_state->active_light[i + 1];
3593 }
3594
3595 return D3D_OK;
3596 }
3597
3598 #define D3DRS_TO_STRING_CASE(n) case D3DRS_##n: return "D3DRS_"#n
3599 const char *nine_d3drs_to_string(DWORD State)
3600 {
3601 switch (State) {
3602 D3DRS_TO_STRING_CASE(ZENABLE);
3603 D3DRS_TO_STRING_CASE(FILLMODE);
3604 D3DRS_TO_STRING_CASE(SHADEMODE);
3605 D3DRS_TO_STRING_CASE(ZWRITEENABLE);
3606 D3DRS_TO_STRING_CASE(ALPHATESTENABLE);
3607 D3DRS_TO_STRING_CASE(LASTPIXEL);
3608 D3DRS_TO_STRING_CASE(SRCBLEND);
3609 D3DRS_TO_STRING_CASE(DESTBLEND);
3610 D3DRS_TO_STRING_CASE(CULLMODE);
3611 D3DRS_TO_STRING_CASE(ZFUNC);
3612 D3DRS_TO_STRING_CASE(ALPHAREF);
3613 D3DRS_TO_STRING_CASE(ALPHAFUNC);
3614 D3DRS_TO_STRING_CASE(DITHERENABLE);
3615 D3DRS_TO_STRING_CASE(ALPHABLENDENABLE);
3616 D3DRS_TO_STRING_CASE(FOGENABLE);
3617 D3DRS_TO_STRING_CASE(SPECULARENABLE);
3618 D3DRS_TO_STRING_CASE(FOGCOLOR);
3619 D3DRS_TO_STRING_CASE(FOGTABLEMODE);
3620 D3DRS_TO_STRING_CASE(FOGSTART);
3621 D3DRS_TO_STRING_CASE(FOGEND);
3622 D3DRS_TO_STRING_CASE(FOGDENSITY);
3623 D3DRS_TO_STRING_CASE(RANGEFOGENABLE);
3624 D3DRS_TO_STRING_CASE(STENCILENABLE);
3625 D3DRS_TO_STRING_CASE(STENCILFAIL);
3626 D3DRS_TO_STRING_CASE(STENCILZFAIL);
3627 D3DRS_TO_STRING_CASE(STENCILPASS);
3628 D3DRS_TO_STRING_CASE(STENCILFUNC);
3629 D3DRS_TO_STRING_CASE(STENCILREF);
3630 D3DRS_TO_STRING_CASE(STENCILMASK);
3631 D3DRS_TO_STRING_CASE(STENCILWRITEMASK);
3632 D3DRS_TO_STRING_CASE(TEXTUREFACTOR);
3633 D3DRS_TO_STRING_CASE(WRAP0);
3634 D3DRS_TO_STRING_CASE(WRAP1);
3635 D3DRS_TO_STRING_CASE(WRAP2);
3636 D3DRS_TO_STRING_CASE(WRAP3);
3637 D3DRS_TO_STRING_CASE(WRAP4);
3638 D3DRS_TO_STRING_CASE(WRAP5);
3639 D3DRS_TO_STRING_CASE(WRAP6);
3640 D3DRS_TO_STRING_CASE(WRAP7);
3641 D3DRS_TO_STRING_CASE(CLIPPING);
3642 D3DRS_TO_STRING_CASE(LIGHTING);
3643 D3DRS_TO_STRING_CASE(AMBIENT);
3644 D3DRS_TO_STRING_CASE(FOGVERTEXMODE);
3645 D3DRS_TO_STRING_CASE(COLORVERTEX);
3646 D3DRS_TO_STRING_CASE(LOCALVIEWER);
3647 D3DRS_TO_STRING_CASE(NORMALIZENORMALS);
3648 D3DRS_TO_STRING_CASE(DIFFUSEMATERIALSOURCE);
3649 D3DRS_TO_STRING_CASE(SPECULARMATERIALSOURCE);
3650 D3DRS_TO_STRING_CASE(AMBIENTMATERIALSOURCE);
3651 D3DRS_TO_STRING_CASE(EMISSIVEMATERIALSOURCE);
3652 D3DRS_TO_STRING_CASE(VERTEXBLEND);
3653 D3DRS_TO_STRING_CASE(CLIPPLANEENABLE);
3654 D3DRS_TO_STRING_CASE(POINTSIZE);
3655 D3DRS_TO_STRING_CASE(POINTSIZE_MIN);
3656 D3DRS_TO_STRING_CASE(POINTSPRITEENABLE);
3657 D3DRS_TO_STRING_CASE(POINTSCALEENABLE);
3658 D3DRS_TO_STRING_CASE(POINTSCALE_A);
3659 D3DRS_TO_STRING_CASE(POINTSCALE_B);
3660 D3DRS_TO_STRING_CASE(POINTSCALE_C);
3661 D3DRS_TO_STRING_CASE(MULTISAMPLEANTIALIAS);
3662 D3DRS_TO_STRING_CASE(MULTISAMPLEMASK);
3663 D3DRS_TO_STRING_CASE(PATCHEDGESTYLE);
3664 D3DRS_TO_STRING_CASE(DEBUGMONITORTOKEN);
3665 D3DRS_TO_STRING_CASE(POINTSIZE_MAX);
3666 D3DRS_TO_STRING_CASE(INDEXEDVERTEXBLENDENABLE);
3667 D3DRS_TO_STRING_CASE(COLORWRITEENABLE);
3668 D3DRS_TO_STRING_CASE(TWEENFACTOR);
3669 D3DRS_TO_STRING_CASE(BLENDOP);
3670 D3DRS_TO_STRING_CASE(POSITIONDEGREE);
3671 D3DRS_TO_STRING_CASE(NORMALDEGREE);
3672 D3DRS_TO_STRING_CASE(SCISSORTESTENABLE);
3673 D3DRS_TO_STRING_CASE(SLOPESCALEDEPTHBIAS);
3674 D3DRS_TO_STRING_CASE(ANTIALIASEDLINEENABLE);
3675 D3DRS_TO_STRING_CASE(MINTESSELLATIONLEVEL);
3676 D3DRS_TO_STRING_CASE(MAXTESSELLATIONLEVEL);
3677 D3DRS_TO_STRING_CASE(ADAPTIVETESS_X);
3678 D3DRS_TO_STRING_CASE(ADAPTIVETESS_Y);
3679 D3DRS_TO_STRING_CASE(ADAPTIVETESS_Z);
3680 D3DRS_TO_STRING_CASE(ADAPTIVETESS_W);
3681 D3DRS_TO_STRING_CASE(ENABLEADAPTIVETESSELLATION);
3682 D3DRS_TO_STRING_CASE(TWOSIDEDSTENCILMODE);
3683 D3DRS_TO_STRING_CASE(CCW_STENCILFAIL);
3684 D3DRS_TO_STRING_CASE(CCW_STENCILZFAIL);
3685 D3DRS_TO_STRING_CASE(CCW_STENCILPASS);
3686 D3DRS_TO_STRING_CASE(CCW_STENCILFUNC);
3687 D3DRS_TO_STRING_CASE(COLORWRITEENABLE1);
3688 D3DRS_TO_STRING_CASE(COLORWRITEENABLE2);
3689 D3DRS_TO_STRING_CASE(COLORWRITEENABLE3);
3690 D3DRS_TO_STRING_CASE(BLENDFACTOR);
3691 D3DRS_TO_STRING_CASE(SRGBWRITEENABLE);
3692 D3DRS_TO_STRING_CASE(DEPTHBIAS);
3693 D3DRS_TO_STRING_CASE(WRAP8);
3694 D3DRS_TO_STRING_CASE(WRAP9);
3695 D3DRS_TO_STRING_CASE(WRAP10);
3696 D3DRS_TO_STRING_CASE(WRAP11);
3697 D3DRS_TO_STRING_CASE(WRAP12);
3698 D3DRS_TO_STRING_CASE(WRAP13);
3699 D3DRS_TO_STRING_CASE(WRAP14);
3700 D3DRS_TO_STRING_CASE(WRAP15);
3701 D3DRS_TO_STRING_CASE(SEPARATEALPHABLENDENABLE);
3702 D3DRS_TO_STRING_CASE(SRCBLENDALPHA);
3703 D3DRS_TO_STRING_CASE(DESTBLENDALPHA);
3704 D3DRS_TO_STRING_CASE(BLENDOPALPHA);
3705 default:
3706 return "(invalid)";
3707 }
3708 }