projects / mesa.git / blob
? search:


227e953959206906a0ddec8dbe54e6ce41af0c0c
[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 cso_velems_state ve;
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.velems[n] = vdecl->elems[index];
860 b = ve.velems[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.velems[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.velems[n].vertex_buffer_index = dummy_vbo_stream;
871 ve.velems[n].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
872 ve.velems[n].src_offset = 0;
873 ve.velems[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 ve.count = vs->num_inputs;
888 cso_set_vertex_elements(context->cso, &ve);
889 }
890
891 static void
892 update_vertex_buffers(struct NineDevice9 *device)
893 {
894 struct nine_context *context = &device->context;
895 struct pipe_context *pipe = context->pipe;
896 struct pipe_vertex_buffer dummy_vtxbuf;
897 uint32_t mask = context->changed.vtxbuf;
898 unsigned i;
899
900 DBG("mask=%x\n", mask);
901
902 if (context->dummy_vbo_bound_at >= 0) {
903 if (!context->vbo_bound_done) {
904 dummy_vtxbuf.buffer.resource = device->dummy_vbo;
905 dummy_vtxbuf.stride = 0;
906 dummy_vtxbuf.is_user_buffer = false;
907 dummy_vtxbuf.buffer_offset = 0;
908 pipe->set_vertex_buffers(pipe, context->dummy_vbo_bound_at,
909 1, &dummy_vtxbuf);
910 context->vbo_bound_done = TRUE;
911 }
912 mask &= ~(1 << context->dummy_vbo_bound_at);
913 }
914
915 for (i = 0; mask; mask >>= 1, ++i) {
916 if (mask & 1) {
917 if (context->vtxbuf[i].buffer.resource)
918 pipe->set_vertex_buffers(pipe, i, 1, &context->vtxbuf[i]);
919 else
920 pipe->set_vertex_buffers(pipe, i, 1, NULL);
921 }
922 }
923
924 context->changed.vtxbuf = 0;
925 }
926
927 static inline boolean
928 update_sampler_derived(struct nine_context *context, unsigned s)
929 {
930 boolean changed = FALSE;
931
932 if (context->samp[s][NINED3DSAMP_SHADOW] != context->texture[s].shadow) {
933 changed = TRUE;
934 context->samp[s][NINED3DSAMP_SHADOW] = context->texture[s].shadow;
935 }
936
937 if (context->samp[s][NINED3DSAMP_CUBETEX] !=
938 (context->texture[s].type == D3DRTYPE_CUBETEXTURE)) {
939 changed = TRUE;
940 context->samp[s][NINED3DSAMP_CUBETEX] =
941 context->texture[s].type == D3DRTYPE_CUBETEXTURE;
942 }
943
944 if (context->samp[s][D3DSAMP_MIPFILTER] != D3DTEXF_NONE) {
945 int lod = context->samp[s][D3DSAMP_MAXMIPLEVEL] - context->texture[s].lod;
946 if (lod < 0)
947 lod = 0;
948 if (context->samp[s][NINED3DSAMP_MINLOD] != lod) {
949 changed = TRUE;
950 context->samp[s][NINED3DSAMP_MINLOD] = lod;
951 }
952 } else {
953 context->changed.sampler[s] &= ~0x300; /* lod changes irrelevant */
954 }
955
956 return changed;
957 }
958
959 /* TODO: add sRGB override to pipe_sampler_state ? */
960 static void
961 update_textures_and_samplers(struct NineDevice9 *device)
962 {
963 struct nine_context *context = &device->context;
964 struct pipe_sampler_view *view[NINE_MAX_SAMPLERS];
965 unsigned num_textures;
966 unsigned i;
967 boolean commit_samplers;
968 uint16_t sampler_mask = context->ps ? context->ps->sampler_mask :
969 device->ff.ps->sampler_mask;
970
971 /* TODO: Can we reduce iterations here ? */
972
973 commit_samplers = FALSE;
974 context->bound_samplers_mask_ps = 0;
975 for (num_textures = 0, i = 0; i < NINE_MAX_SAMPLERS_PS; ++i) {
976 const unsigned s = NINE_SAMPLER_PS(i);
977 int sRGB;
978
979 if (!context->texture[s].enabled && !(sampler_mask & (1 << i))) {
980 view[i] = NULL;
981 continue;
982 }
983
984 if (context->texture[s].enabled) {
985 sRGB = context->samp[s][D3DSAMP_SRGBTEXTURE] ? 1 : 0;
986
987 view[i] = context->texture[s].view[sRGB];
988 num_textures = i + 1;
989
990 if (update_sampler_derived(context, s) || (context->changed.sampler[s] & 0x05fe)) {
991 context->changed.sampler[s] = 0;
992 commit_samplers = TRUE;
993 nine_convert_sampler_state(context->cso, s, context->samp[s]);
994 }
995 } else {
996 /* Bind dummy sampler. We do not bind dummy sampler when
997 * it is not needed because it could add overhead. The
998 * dummy sampler should have r=g=b=0 and a=1. We do not
999 * unbind dummy sampler directly when they are not needed
1000 * anymore, but they're going to be removed as long as texture
1001 * or sampler states are changed. */
1002 view[i] = device->dummy_sampler_view;
1003 num_textures = i + 1;
1004
1005 cso_single_sampler(context->cso, PIPE_SHADER_FRAGMENT,
1006 s - NINE_SAMPLER_PS(0), &device->dummy_sampler_state);
1007
1008 commit_samplers = TRUE;
1009 context->changed.sampler[s] = ~0;
1010 }
1011
1012 context->bound_samplers_mask_ps |= (1 << s);
1013 }
1014
1015 cso_set_sampler_views(context->cso, PIPE_SHADER_FRAGMENT, num_textures, view);
1016
1017 if (commit_samplers)
1018 cso_single_sampler_done(context->cso, PIPE_SHADER_FRAGMENT);
1019
1020 commit_samplers = FALSE;
1021 sampler_mask = context->programmable_vs ? context->vs->sampler_mask : 0;
1022 context->bound_samplers_mask_vs = 0;
1023 for (num_textures = 0, i = 0; i < NINE_MAX_SAMPLERS_VS; ++i) {
1024 const unsigned s = NINE_SAMPLER_VS(i);
1025 int sRGB;
1026
1027 if (!context->texture[s].enabled && !(sampler_mask & (1 << i))) {
1028 view[i] = NULL;
1029 continue;
1030 }
1031
1032 if (context->texture[s].enabled) {
1033 sRGB = context->samp[s][D3DSAMP_SRGBTEXTURE] ? 1 : 0;
1034
1035 view[i] = context->texture[s].view[sRGB];
1036 num_textures = i + 1;
1037
1038 if (update_sampler_derived(context, s) || (context->changed.sampler[s] & 0x05fe)) {
1039 context->changed.sampler[s] = 0;
1040 commit_samplers = TRUE;
1041 nine_convert_sampler_state(context->cso, s, context->samp[s]);
1042 }
1043 } else {
1044 /* Bind dummy sampler. We do not bind dummy sampler when
1045 * it is not needed because it could add overhead. The
1046 * dummy sampler should have r=g=b=0 and a=1. We do not
1047 * unbind dummy sampler directly when they are not needed
1048 * anymore, but they're going to be removed as long as texture
1049 * or sampler states are changed. */
1050 view[i] = device->dummy_sampler_view;
1051 num_textures = i + 1;
1052
1053 cso_single_sampler(context->cso, PIPE_SHADER_VERTEX,
1054 s - NINE_SAMPLER_VS(0), &device->dummy_sampler_state);
1055
1056 commit_samplers = TRUE;
1057 context->changed.sampler[s] = ~0;
1058 }
1059
1060 context->bound_samplers_mask_vs |= (1 << i);
1061 }
1062
1063 cso_set_sampler_views(context->cso, PIPE_SHADER_VERTEX, num_textures, view);
1064
1065 if (commit_samplers)
1066 cso_single_sampler_done(context->cso, PIPE_SHADER_VERTEX);
1067 }
1068
1069 /* State commit only */
1070
1071 static inline void
1072 commit_blend(struct NineDevice9 *device)
1073 {
1074 struct nine_context *context = &device->context;
1075
1076 cso_set_blend(context->cso, &context->pipe_data.blend);
1077 }
1078
1079 static inline void
1080 commit_dsa(struct NineDevice9 *device)
1081 {
1082 struct nine_context *context = &device->context;
1083
1084 cso_set_depth_stencil_alpha(context->cso, &context->pipe_data.dsa);
1085 }
1086
1087 static inline void
1088 commit_scissor(struct NineDevice9 *device)
1089 {
1090 struct nine_context *context = &device->context;
1091 struct pipe_context *pipe = context->pipe;
1092
1093 pipe->set_scissor_states(pipe, 0, 1, &context->scissor);
1094 }
1095
1096 static inline void
1097 commit_rasterizer(struct NineDevice9 *device)
1098 {
1099 struct nine_context *context = &device->context;
1100
1101 cso_set_rasterizer(context->cso, &context->pipe_data.rast);
1102 }
1103
1104 static inline void
1105 commit_vs_constants(struct NineDevice9 *device)
1106 {
1107 struct nine_context *context = &device->context;
1108 struct pipe_context *pipe = context->pipe;
1109
1110 if (unlikely(!context->programmable_vs))
1111 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &context->pipe_data.cb_vs_ff);
1112 else {
1113 if (context->swvp) {
1114 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &context->pipe_data.cb0_swvp);
1115 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 1, &context->pipe_data.cb1_swvp);
1116 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 2, &context->pipe_data.cb2_swvp);
1117 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 3, &context->pipe_data.cb3_swvp);
1118 } else {
1119 pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &context->pipe_data.cb_vs);
1120 }
1121 }
1122 }
1123
1124 static inline void
1125 commit_ps_constants(struct NineDevice9 *device)
1126 {
1127 struct nine_context *context = &device->context;
1128 struct pipe_context *pipe = context->pipe;
1129
1130 if (unlikely(!context->ps))
1131 pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, 0, &context->pipe_data.cb_ps_ff);
1132 else
1133 pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, 0, &context->pipe_data.cb_ps);
1134 }
1135
1136 static inline void
1137 commit_vs(struct NineDevice9 *device)
1138 {
1139 struct nine_context *context = &device->context;
1140
1141 context->pipe->bind_vs_state(context->pipe, context->cso_shader.vs);
1142 }
1143
1144
1145 static inline void
1146 commit_ps(struct NineDevice9 *device)
1147 {
1148 struct nine_context *context = &device->context;
1149
1150 context->pipe->bind_fs_state(context->pipe, context->cso_shader.ps);
1151 }
1152 /* State Update */
1153
1154 #define NINE_STATE_SHADER_CHANGE_VS \
1155 (NINE_STATE_VS | \
1156 NINE_STATE_TEXTURE | \
1157 NINE_STATE_VS_PARAMS_MISC | \
1158 NINE_STATE_SWVP)
1159
1160 #define NINE_STATE_SHADER_CHANGE_PS \
1161 (NINE_STATE_PS | \
1162 NINE_STATE_TEXTURE | \
1163 NINE_STATE_PS_PARAMS_MISC)
1164
1165 #define NINE_STATE_FREQUENT \
1166 (NINE_STATE_RASTERIZER | \
1167 NINE_STATE_TEXTURE | \
1168 NINE_STATE_SAMPLER | \
1169 NINE_STATE_VS_CONST | \
1170 NINE_STATE_PS_CONST | \
1171 NINE_STATE_MULTISAMPLE)
1172
1173 #define NINE_STATE_COMMON \
1174 (NINE_STATE_FB | \
1175 NINE_STATE_BLEND | \
1176 NINE_STATE_DSA | \
1177 NINE_STATE_VIEWPORT | \
1178 NINE_STATE_VDECL | \
1179 NINE_STATE_IDXBUF | \
1180 NINE_STATE_STREAMFREQ)
1181
1182 #define NINE_STATE_RARE \
1183 (NINE_STATE_SCISSOR | \
1184 NINE_STATE_BLEND_COLOR | \
1185 NINE_STATE_STENCIL_REF | \
1186 NINE_STATE_SAMPLE_MASK)
1187
1188 static void
1189 nine_update_state(struct NineDevice9 *device)
1190 {
1191 struct nine_context *context = &device->context;
1192 struct pipe_context *pipe = context->pipe;
1193 uint32_t group;
1194
1195 DBG("changed state groups: %x\n", context->changed.group);
1196
1197 /* NOTE: We may want to use the cso cache for everything, or let
1198 * NineDevice9.RestoreNonCSOState actually set the states, then we wouldn't
1199 * have to care about state being clobbered here and could merge this back
1200 * into update_textures. Except, we also need to re-validate textures that
1201 * may be dirty anyway, even if no texture bindings changed.
1202 */
1203
1204 /* ff_update may change VS/PS dirty bits */
1205 if (unlikely(!context->programmable_vs || !context->ps))
1206 nine_ff_update(device);
1207 group = context->changed.group;
1208
1209 if (group & (NINE_STATE_SHADER_CHANGE_VS | NINE_STATE_SHADER_CHANGE_PS)) {
1210 if (group & NINE_STATE_SHADER_CHANGE_VS)
1211 group |= prepare_vs(device, (group & NINE_STATE_VS) != 0); /* may set NINE_STATE_RASTERIZER and NINE_STATE_SAMPLER*/
1212 if (group & NINE_STATE_SHADER_CHANGE_PS)
1213 group |= prepare_ps(device, (group & NINE_STATE_PS) != 0);
1214 }
1215
1216 if (group & (NINE_STATE_COMMON | NINE_STATE_VS)) {
1217 if (group & NINE_STATE_FB)
1218 update_framebuffer(device, FALSE);
1219 if (group & NINE_STATE_BLEND)
1220 prepare_blend(device);
1221 if (group & NINE_STATE_DSA)
1222 prepare_dsa(device);
1223 if (group & NINE_STATE_VIEWPORT)
1224 update_viewport(device);
1225 if (group & (NINE_STATE_VDECL | NINE_STATE_VS | NINE_STATE_STREAMFREQ))
1226 update_vertex_elements(device);
1227 }
1228
1229 if (likely(group & (NINE_STATE_FREQUENT | NINE_STATE_VS | NINE_STATE_PS | NINE_STATE_SWVP))) {
1230 if (group & NINE_STATE_MULTISAMPLE)
1231 group |= check_multisample(device);
1232 if (group & NINE_STATE_RASTERIZER)
1233 prepare_rasterizer(device);
1234 if (group & (NINE_STATE_TEXTURE | NINE_STATE_SAMPLER))
1235 update_textures_and_samplers(device);
1236 if ((group & (NINE_STATE_VS_CONST | NINE_STATE_VS | NINE_STATE_SWVP)) && context->programmable_vs)
1237 prepare_vs_constants_userbuf(device);
1238 if ((group & (NINE_STATE_PS_CONST | NINE_STATE_PS)) && context->ps)
1239 prepare_ps_constants_userbuf(device);
1240 }
1241
1242 if (context->changed.vtxbuf)
1243 update_vertex_buffers(device);
1244
1245 if (context->commit & NINE_STATE_COMMIT_BLEND)
1246 commit_blend(device);
1247 if (context->commit & NINE_STATE_COMMIT_DSA)
1248 commit_dsa(device);
1249 if (context->commit & NINE_STATE_COMMIT_RASTERIZER)
1250 commit_rasterizer(device);
1251 if (context->commit & NINE_STATE_COMMIT_CONST_VS)
1252 commit_vs_constants(device);
1253 if (context->commit & NINE_STATE_COMMIT_CONST_PS)
1254 commit_ps_constants(device);
1255 if (context->commit & NINE_STATE_COMMIT_VS)
1256 commit_vs(device);
1257 if (context->commit & NINE_STATE_COMMIT_PS)
1258 commit_ps(device);
1259
1260 context->commit = 0;
1261
1262 if (unlikely(context->changed.ucp)) {
1263 pipe->set_clip_state(pipe, &context->clip);
1264 context->changed.ucp = FALSE;
1265 }
1266
1267 if (unlikely(group & NINE_STATE_RARE)) {
1268 if (group & NINE_STATE_SCISSOR)
1269 commit_scissor(device);
1270 if (group & NINE_STATE_BLEND_COLOR) {
1271 struct pipe_blend_color color;
1272 d3dcolor_to_rgba(&color.color[0], context->rs[D3DRS_BLENDFACTOR]);
1273 pipe->set_blend_color(pipe, &color);
1274 }
1275 if (group & NINE_STATE_SAMPLE_MASK) {
1276 if (context->rt[0]->desc.MultiSampleType <= D3DMULTISAMPLE_NONMASKABLE) {
1277 pipe->set_sample_mask(pipe, ~0);
1278 } else {
1279 pipe->set_sample_mask(pipe, context->rs[D3DRS_MULTISAMPLEMASK]);
1280 }
1281 }
1282 if (group & NINE_STATE_STENCIL_REF) {
1283 struct pipe_stencil_ref ref;
1284 ref.ref_value[0] = context->rs[D3DRS_STENCILREF];
1285 ref.ref_value[1] = ref.ref_value[0];
1286 pipe->set_stencil_ref(pipe, &ref);
1287 }
1288 }
1289
1290 context->changed.group &=
1291 (NINE_STATE_FF | NINE_STATE_VS_CONST | NINE_STATE_PS_CONST);
1292
1293 DBG("finished\n");
1294 }
1295
1296 #define RESZ_CODE 0x7fa05000
1297
1298 static void
1299 NineDevice9_ResolveZ( struct NineDevice9 *device )
1300 {
1301 struct nine_context *context = &device->context;
1302 const struct util_format_description *desc;
1303 struct NineSurface9 *source = context->ds;
1304 struct pipe_resource *src, *dst;
1305 struct pipe_blit_info blit;
1306
1307 DBG("RESZ resolve\n");
1308
1309 if (!source || !context->texture[0].enabled ||
1310 context->texture[0].type != D3DRTYPE_TEXTURE)
1311 return;
1312
1313 src = source->base.resource;
1314 dst = context->texture[0].resource;
1315
1316 if (!src || !dst)
1317 return;
1318
1319 /* check dst is depth format. we know already for src */
1320 desc = util_format_description(dst->format);
1321 if (desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1322 return;
1323
1324 memset(&blit, 0, sizeof(blit));
1325 blit.src.resource = src;
1326 blit.src.level = 0;
1327 blit.src.format = src->format;
1328 blit.src.box.z = 0;
1329 blit.src.box.depth = 1;
1330 blit.src.box.x = 0;
1331 blit.src.box.y = 0;
1332 blit.src.box.width = src->width0;
1333 blit.src.box.height = src->height0;
1334
1335 blit.dst.resource = dst;
1336 blit.dst.level = 0;
1337 blit.dst.format = dst->format;
1338 blit.dst.box.z = 0;
1339 blit.dst.box.depth = 1;
1340 blit.dst.box.x = 0;
1341 blit.dst.box.y = 0;
1342 blit.dst.box.width = dst->width0;
1343 blit.dst.box.height = dst->height0;
1344
1345 blit.mask = PIPE_MASK_ZS;
1346 blit.filter = PIPE_TEX_FILTER_NEAREST;
1347 blit.scissor_enable = FALSE;
1348
1349 context->pipe->blit(context->pipe, &blit);
1350 }
1351
1352 #define ALPHA_TO_COVERAGE_ENABLE MAKEFOURCC('A', '2', 'M', '1')
1353 #define ALPHA_TO_COVERAGE_DISABLE MAKEFOURCC('A', '2', 'M', '0')
1354
1355 /* Nine_context functions.
1356 * Serialized through CSMT macros.
1357 */
1358
1359 static void
1360 nine_context_set_texture_apply(struct NineDevice9 *device,
1361 DWORD stage,
1362 BOOL enabled,
1363 BOOL shadow,
1364 DWORD lod,
1365 D3DRESOURCETYPE type,
1366 uint8_t pstype,
1367 struct pipe_resource *res,
1368 struct pipe_sampler_view *view0,
1369 struct pipe_sampler_view *view1);
1370 static void
1371 nine_context_set_stream_source_apply(struct NineDevice9 *device,
1372 UINT StreamNumber,
1373 struct pipe_resource *res,
1374 UINT OffsetInBytes,
1375 UINT Stride);
1376
1377 static void
1378 nine_context_set_indices_apply(struct NineDevice9 *device,
1379 struct pipe_resource *res,
1380 UINT IndexSize,
1381 UINT OffsetInBytes);
1382
1383 static void
1384 nine_context_set_pixel_shader_constant_i_transformed(struct NineDevice9 *device,
1385 UINT StartRegister,
1386 const int *pConstantData,
1387 unsigned pConstantData_size,
1388 UINT Vector4iCount);
1389
1390 CSMT_ITEM_NO_WAIT(nine_context_set_render_state,
1391 ARG_VAL(D3DRENDERSTATETYPE, State),
1392 ARG_VAL(DWORD, Value))
1393 {
1394 struct nine_context *context = &device->context;
1395
1396 /* Amd hacks (equivalent to GL extensions) */
1397 if (unlikely(State == D3DRS_POINTSIZE)) {
1398 if (Value == RESZ_CODE) {
1399 NineDevice9_ResolveZ(device);
1400 return;
1401 }
1402
1403 if (Value == ALPHA_TO_COVERAGE_ENABLE ||
1404 Value == ALPHA_TO_COVERAGE_DISABLE) {
1405 context->rs[NINED3DRS_ALPHACOVERAGE] = (Value == ALPHA_TO_COVERAGE_ENABLE);
1406 context->changed.group |= NINE_STATE_BLEND;
1407 return;
1408 }
1409 }
1410
1411 /* NV hack */
1412 if (unlikely(State == D3DRS_ADAPTIVETESS_Y)) {
1413 if (Value == D3DFMT_ATOC || (Value == D3DFMT_UNKNOWN && context->rs[NINED3DRS_ALPHACOVERAGE])) {
1414 context->rs[NINED3DRS_ALPHACOVERAGE] = (Value == D3DFMT_ATOC) ? 3 : 0;
1415 context->rs[NINED3DRS_ALPHACOVERAGE] &= context->rs[D3DRS_ALPHATESTENABLE] ? 3 : 2;
1416 context->changed.group |= NINE_STATE_BLEND;
1417 return;
1418 }
1419 }
1420 if (unlikely(State == D3DRS_ALPHATESTENABLE && (context->rs[NINED3DRS_ALPHACOVERAGE] & 2))) {
1421 DWORD alphacoverage_prev = context->rs[NINED3DRS_ALPHACOVERAGE];
1422 context->rs[NINED3DRS_ALPHACOVERAGE] = (Value ? 3 : 2);
1423 if (context->rs[NINED3DRS_ALPHACOVERAGE] != alphacoverage_prev)
1424 context->changed.group |= NINE_STATE_BLEND;
1425 }
1426
1427 context->rs[State] = nine_fix_render_state_value(State, Value);
1428 context->changed.group |= nine_render_state_group[State];
1429 }
1430
1431 CSMT_ITEM_NO_WAIT(nine_context_set_texture_apply,
1432 ARG_VAL(DWORD, stage),
1433 ARG_VAL(BOOL, enabled),
1434 ARG_VAL(BOOL, shadow),
1435 ARG_VAL(DWORD, lod),
1436 ARG_VAL(D3DRESOURCETYPE, type),
1437 ARG_VAL(uint8_t, pstype),
1438 ARG_BIND_RES(struct pipe_resource, res),
1439 ARG_BIND_VIEW(struct pipe_sampler_view, view0),
1440 ARG_BIND_VIEW(struct pipe_sampler_view, view1))
1441 {
1442 struct nine_context *context = &device->context;
1443
1444 context->texture[stage].enabled = enabled;
1445 context->samplers_shadow &= ~(1 << stage);
1446 context->samplers_shadow |= shadow << stage;
1447 context->texture[stage].shadow = shadow;
1448 context->texture[stage].lod = lod;
1449 context->texture[stage].type = type;
1450 context->texture[stage].pstype = pstype;
1451 pipe_resource_reference(&context->texture[stage].resource, res);
1452 pipe_sampler_view_reference(&context->texture[stage].view[0], view0);
1453 pipe_sampler_view_reference(&context->texture[stage].view[1], view1);
1454
1455 context->changed.group |= NINE_STATE_TEXTURE;
1456 }
1457
1458 void
1459 nine_context_set_texture(struct NineDevice9 *device,
1460 DWORD Stage,
1461 struct NineBaseTexture9 *tex)
1462 {
1463 BOOL enabled = FALSE;
1464 BOOL shadow = FALSE;
1465 DWORD lod = 0;
1466 D3DRESOURCETYPE type = D3DRTYPE_TEXTURE;
1467 uint8_t pstype = 0;
1468 struct pipe_resource *res = NULL;
1469 struct pipe_sampler_view *view0 = NULL, *view1 = NULL;
1470
1471 /* For managed pool, the data can be initially incomplete.
1472 * In that case, the texture is rebound later
1473 * (in NineBaseTexture9_Validate/NineBaseTexture9_UploadSelf). */
1474 if (tex && tex->base.resource) {
1475 enabled = TRUE;
1476 shadow = tex->shadow;
1477 lod = tex->managed.lod;
1478 type = tex->base.type;
1479 pstype = tex->pstype;
1480 res = tex->base.resource;
1481 view0 = NineBaseTexture9_GetSamplerView(tex, 0);
1482 view1 = NineBaseTexture9_GetSamplerView(tex, 1);
1483 }
1484
1485 nine_context_set_texture_apply(device, Stage, enabled,
1486 shadow, lod, type, pstype,
1487 res, view0, view1);
1488 }
1489
1490 CSMT_ITEM_NO_WAIT(nine_context_set_sampler_state,
1491 ARG_VAL(DWORD, Sampler),
1492 ARG_VAL(D3DSAMPLERSTATETYPE, Type),
1493 ARG_VAL(DWORD, Value))
1494 {
1495 struct nine_context *context = &device->context;
1496
1497 if (unlikely(!nine_check_sampler_state_value(Type, Value)))
1498 return;
1499
1500 context->samp[Sampler][Type] = Value;
1501 context->changed.group |= NINE_STATE_SAMPLER;
1502 context->changed.sampler[Sampler] |= 1 << Type;
1503 }
1504
1505 CSMT_ITEM_NO_WAIT(nine_context_set_stream_source_apply,
1506 ARG_VAL(UINT, StreamNumber),
1507 ARG_BIND_RES(struct pipe_resource, res),
1508 ARG_VAL(UINT, OffsetInBytes),
1509 ARG_VAL(UINT, Stride))
1510 {
1511 struct nine_context *context = &device->context;
1512 const unsigned i = StreamNumber;
1513
1514 context->vtxbuf[i].stride = Stride;
1515 context->vtxbuf[i].buffer_offset = OffsetInBytes;
1516 pipe_resource_reference(&context->vtxbuf[i].buffer.resource, res);
1517
1518 context->changed.vtxbuf |= 1 << StreamNumber;
1519 }
1520
1521 void
1522 nine_context_set_stream_source(struct NineDevice9 *device,
1523 UINT StreamNumber,
1524 struct NineVertexBuffer9 *pVBuf9,
1525 UINT OffsetInBytes,
1526 UINT Stride)
1527 {
1528 struct pipe_resource *res = NULL;
1529 unsigned offset = 0;
1530
1531 if (pVBuf9)
1532 res = NineVertexBuffer9_GetResource(pVBuf9, &offset);
1533 /* in the future when there is internal offset, add it
1534 * to OffsetInBytes */
1535
1536 nine_context_set_stream_source_apply(device, StreamNumber,
1537 res, offset + OffsetInBytes,
1538 Stride);
1539 }
1540
1541 CSMT_ITEM_NO_WAIT(nine_context_set_stream_source_freq,
1542 ARG_VAL(UINT, StreamNumber),
1543 ARG_VAL(UINT, Setting))
1544 {
1545 struct nine_context *context = &device->context;
1546
1547 context->stream_freq[StreamNumber] = Setting;
1548
1549 if (Setting & D3DSTREAMSOURCE_INSTANCEDATA)
1550 context->stream_instancedata_mask |= 1 << StreamNumber;
1551 else
1552 context->stream_instancedata_mask &= ~(1 << StreamNumber);
1553
1554 if (StreamNumber != 0)
1555 context->changed.group |= NINE_STATE_STREAMFREQ;
1556 }
1557
1558 CSMT_ITEM_NO_WAIT(nine_context_set_indices_apply,
1559 ARG_BIND_RES(struct pipe_resource, res),
1560 ARG_VAL(UINT, IndexSize),
1561 ARG_VAL(UINT, OffsetInBytes))
1562 {
1563 struct nine_context *context = &device->context;
1564
1565 context->index_size = IndexSize;
1566 context->index_offset = OffsetInBytes;
1567 pipe_resource_reference(&context->idxbuf, res);
1568
1569 context->changed.group |= NINE_STATE_IDXBUF;
1570 }
1571
1572 void
1573 nine_context_set_indices(struct NineDevice9 *device,
1574 struct NineIndexBuffer9 *idxbuf)
1575 {
1576 struct pipe_resource *res = NULL;
1577 UINT IndexSize = 0;
1578 unsigned OffsetInBytes = 0;
1579
1580 if (idxbuf) {
1581 res = NineIndexBuffer9_GetBuffer(idxbuf, &OffsetInBytes);
1582 IndexSize = idxbuf->index_size;
1583 }
1584
1585 nine_context_set_indices_apply(device, res, IndexSize, OffsetInBytes);
1586 }
1587
1588 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_declaration,
1589 ARG_BIND_REF(struct NineVertexDeclaration9, vdecl))
1590 {
1591 struct nine_context *context = &device->context;
1592 BOOL was_programmable_vs = context->programmable_vs;
1593
1594 nine_bind(&context->vdecl, vdecl);
1595
1596 context->programmable_vs = context->vs && !(context->vdecl && context->vdecl->position_t);
1597 if (was_programmable_vs != context->programmable_vs) {
1598 context->commit |= NINE_STATE_COMMIT_CONST_VS;
1599 context->changed.group |= NINE_STATE_VS;
1600 }
1601
1602 context->changed.group |= NINE_STATE_VDECL;
1603 }
1604
1605 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_shader,
1606 ARG_BIND_REF(struct NineVertexShader9, pShader))
1607 {
1608 struct nine_context *context = &device->context;
1609 BOOL was_programmable_vs = context->programmable_vs;
1610
1611 nine_bind(&context->vs, pShader);
1612
1613 context->programmable_vs = context->vs && !(context->vdecl && context->vdecl->position_t);
1614
1615 /* ff -> non-ff: commit back non-ff constants */
1616 if (!was_programmable_vs && context->programmable_vs)
1617 context->commit |= NINE_STATE_COMMIT_CONST_VS;
1618
1619 context->changed.group |= NINE_STATE_VS;
1620 }
1621
1622 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_shader_constant_f,
1623 ARG_VAL(UINT, StartRegister),
1624 ARG_MEM(float, pConstantData),
1625 ARG_MEM_SIZE(unsigned, pConstantData_size),
1626 ARG_VAL(UINT, Vector4fCount))
1627 {
1628 struct nine_context *context = &device->context;
1629 float *vs_const_f = device->may_swvp ? context->vs_const_f_swvp : context->vs_const_f;
1630
1631 memcpy(&vs_const_f[StartRegister * 4],
1632 pConstantData,
1633 pConstantData_size);
1634
1635 if (device->may_swvp) {
1636 Vector4fCount = MIN2(StartRegister + Vector4fCount, NINE_MAX_CONST_F) - StartRegister;
1637 if (StartRegister < NINE_MAX_CONST_F)
1638 memcpy(&context->vs_const_f[StartRegister * 4],
1639 pConstantData,
1640 Vector4fCount * 4 * sizeof(context->vs_const_f[0]));
1641 }
1642
1643 context->changed.vs_const_f = TRUE;
1644 context->changed.group |= NINE_STATE_VS_CONST;
1645 }
1646
1647 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_shader_constant_i,
1648 ARG_VAL(UINT, StartRegister),
1649 ARG_MEM(int, pConstantData),
1650 ARG_MEM_SIZE(unsigned, pConstantData_size),
1651 ARG_VAL(UINT, Vector4iCount))
1652 {
1653 struct nine_context *context = &device->context;
1654 int i;
1655
1656 if (device->driver_caps.vs_integer) {
1657 memcpy(&context->vs_const_i[4 * StartRegister],
1658 pConstantData,
1659 pConstantData_size);
1660 } else {
1661 for (i = 0; i < Vector4iCount; i++) {
1662 context->vs_const_i[4 * (StartRegister + i)] = fui((float)(pConstantData[4 * i]));
1663 context->vs_const_i[4 * (StartRegister + i) + 1] = fui((float)(pConstantData[4 * i + 1]));
1664 context->vs_const_i[4 * (StartRegister + i) + 2] = fui((float)(pConstantData[4 * i + 2]));
1665 context->vs_const_i[4 * (StartRegister + i) + 3] = fui((float)(pConstantData[4 * i + 3]));
1666 }
1667 }
1668
1669 context->changed.vs_const_i = TRUE;
1670 context->changed.group |= NINE_STATE_VS_CONST | NINE_STATE_VS_PARAMS_MISC;
1671 }
1672
1673 CSMT_ITEM_NO_WAIT(nine_context_set_vertex_shader_constant_b,
1674 ARG_VAL(UINT, StartRegister),
1675 ARG_MEM(BOOL, pConstantData),
1676 ARG_MEM_SIZE(unsigned, pConstantData_size),
1677 ARG_VAL(UINT, BoolCount))
1678 {
1679 struct nine_context *context = &device->context;
1680 int i;
1681 uint32_t bool_true = device->driver_caps.vs_integer ? 0xFFFFFFFF : fui(1.0f);
1682
1683 (void) pConstantData_size;
1684
1685 for (i = 0; i < BoolCount; i++)
1686 context->vs_const_b[StartRegister + i] = pConstantData[i] ? bool_true : 0;
1687
1688 context->changed.vs_const_b = TRUE;
1689 context->changed.group |= NINE_STATE_VS_CONST | NINE_STATE_VS_PARAMS_MISC;
1690 }
1691
1692 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader,
1693 ARG_BIND_REF(struct NinePixelShader9, ps))
1694 {
1695 struct nine_context *context = &device->context;
1696 unsigned old_mask = context->ps ? context->ps->rt_mask : 1;
1697 unsigned mask;
1698
1699 /* ff -> non-ff: commit back non-ff constants */
1700 if (!context->ps && ps)
1701 context->commit |= NINE_STATE_COMMIT_CONST_PS;
1702
1703 nine_bind(&context->ps, ps);
1704
1705 context->changed.group |= NINE_STATE_PS;
1706
1707 mask = context->ps ? context->ps->rt_mask : 1;
1708 /* We need to update cbufs if the pixel shader would
1709 * write to different render targets */
1710 if (mask != old_mask)
1711 context->changed.group |= NINE_STATE_FB;
1712 }
1713
1714 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader_constant_f,
1715 ARG_VAL(UINT, StartRegister),
1716 ARG_MEM(float, pConstantData),
1717 ARG_MEM_SIZE(unsigned, pConstantData_size),
1718 ARG_VAL(UINT, Vector4fCount))
1719 {
1720 struct nine_context *context = &device->context;
1721
1722 memcpy(&context->ps_const_f[StartRegister * 4],
1723 pConstantData,
1724 pConstantData_size);
1725
1726 context->changed.ps_const_f = TRUE;
1727 context->changed.group |= NINE_STATE_PS_CONST;
1728 }
1729
1730 /* For stateblocks */
1731 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader_constant_i_transformed,
1732 ARG_VAL(UINT, StartRegister),
1733 ARG_MEM(int, pConstantData),
1734 ARG_MEM_SIZE(unsigned, pConstantData_size),
1735 ARG_VAL(UINT, Vector4iCount))
1736 {
1737 struct nine_context *context = &device->context;
1738
1739 memcpy(&context->ps_const_i[StartRegister][0],
1740 pConstantData,
1741 Vector4iCount * sizeof(context->ps_const_i[0]));
1742
1743 context->changed.ps_const_i = TRUE;
1744 context->changed.group |= NINE_STATE_PS_CONST | NINE_STATE_PS_PARAMS_MISC;
1745 }
1746
1747 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader_constant_i,
1748 ARG_VAL(UINT, StartRegister),
1749 ARG_MEM(int, pConstantData),
1750 ARG_MEM_SIZE(unsigned, pConstantData_size),
1751 ARG_VAL(UINT, Vector4iCount))
1752 {
1753 struct nine_context *context = &device->context;
1754 int i;
1755
1756 if (device->driver_caps.ps_integer) {
1757 memcpy(&context->ps_const_i[StartRegister][0],
1758 pConstantData,
1759 pConstantData_size);
1760 } else {
1761 for (i = 0; i < Vector4iCount; i++) {
1762 context->ps_const_i[StartRegister+i][0] = fui((float)(pConstantData[4*i]));
1763 context->ps_const_i[StartRegister+i][1] = fui((float)(pConstantData[4*i+1]));
1764 context->ps_const_i[StartRegister+i][2] = fui((float)(pConstantData[4*i+2]));
1765 context->ps_const_i[StartRegister+i][3] = fui((float)(pConstantData[4*i+3]));
1766 }
1767 }
1768 context->changed.ps_const_i = TRUE;
1769 context->changed.group |= NINE_STATE_PS_CONST | NINE_STATE_PS_PARAMS_MISC;
1770 }
1771
1772 CSMT_ITEM_NO_WAIT(nine_context_set_pixel_shader_constant_b,
1773 ARG_VAL(UINT, StartRegister),
1774 ARG_MEM(BOOL, pConstantData),
1775 ARG_MEM_SIZE(unsigned, pConstantData_size),
1776 ARG_VAL(UINT, BoolCount))
1777 {
1778 struct nine_context *context = &device->context;
1779 int i;
1780 uint32_t bool_true = device->driver_caps.ps_integer ? 0xFFFFFFFF : fui(1.0f);
1781
1782 (void) pConstantData_size;
1783
1784 for (i = 0; i < BoolCount; i++)
1785 context->ps_const_b[StartRegister + i] = pConstantData[i] ? bool_true : 0;
1786
1787 context->changed.ps_const_b = TRUE;
1788 context->changed.group |= NINE_STATE_PS_CONST | NINE_STATE_PS_PARAMS_MISC;
1789 }
1790
1791 /* XXX: use resource, as resource might change */
1792 CSMT_ITEM_NO_WAIT(nine_context_set_render_target,
1793 ARG_VAL(DWORD, RenderTargetIndex),
1794 ARG_BIND_REF(struct NineSurface9, rt))
1795 {
1796 struct nine_context *context = &device->context;
1797 const unsigned i = RenderTargetIndex;
1798
1799 if (i == 0) {
1800 context->viewport.X = 0;
1801 context->viewport.Y = 0;
1802 context->viewport.Width = rt->desc.Width;
1803 context->viewport.Height = rt->desc.Height;
1804 context->viewport.MinZ = 0.0f;
1805 context->viewport.MaxZ = 1.0f;
1806
1807 context->scissor.minx = 0;
1808 context->scissor.miny = 0;
1809 context->scissor.maxx = rt->desc.Width;
1810 context->scissor.maxy = rt->desc.Height;
1811
1812 context->changed.group |= NINE_STATE_VIEWPORT | NINE_STATE_SCISSOR | NINE_STATE_MULTISAMPLE;
1813
1814 if (context->rt[0] &&
1815 (context->rt[0]->desc.MultiSampleType <= D3DMULTISAMPLE_NONMASKABLE) !=
1816 (rt->desc.MultiSampleType <= D3DMULTISAMPLE_NONMASKABLE))
1817 context->changed.group |= NINE_STATE_SAMPLE_MASK;
1818 }
1819
1820 if (context->rt[i] != rt) {
1821 nine_bind(&context->rt[i], rt);
1822 context->changed.group |= NINE_STATE_FB;
1823 }
1824 }
1825
1826 /* XXX: use resource instead of ds, as resource might change */
1827 CSMT_ITEM_NO_WAIT(nine_context_set_depth_stencil,
1828 ARG_BIND_REF(struct NineSurface9, ds))
1829 {
1830 struct nine_context *context = &device->context;
1831
1832 nine_bind(&context->ds, ds);
1833 context->changed.group |= NINE_STATE_FB;
1834 }
1835
1836 CSMT_ITEM_NO_WAIT(nine_context_set_viewport,
1837 ARG_COPY_REF(D3DVIEWPORT9, viewport))
1838 {
1839 struct nine_context *context = &device->context;
1840
1841 context->viewport = *viewport;
1842 context->changed.group |= NINE_STATE_VIEWPORT;
1843 }
1844
1845 CSMT_ITEM_NO_WAIT(nine_context_set_scissor,
1846 ARG_COPY_REF(struct pipe_scissor_state, scissor))
1847 {
1848 struct nine_context *context = &device->context;
1849
1850 context->scissor = *scissor;
1851 context->changed.group |= NINE_STATE_SCISSOR;
1852 }
1853
1854 CSMT_ITEM_NO_WAIT(nine_context_set_transform,
1855 ARG_VAL(D3DTRANSFORMSTATETYPE, State),
1856 ARG_COPY_REF(D3DMATRIX, pMatrix))
1857 {
1858 struct nine_context *context = &device->context;
1859 D3DMATRIX *M = nine_state_access_transform(&context->ff, State, TRUE);
1860
1861 *M = *pMatrix;
1862 context->ff.changed.transform[State / 32] |= 1 << (State % 32);
1863 context->changed.group |= NINE_STATE_FF;
1864 }
1865
1866 CSMT_ITEM_NO_WAIT(nine_context_set_material,
1867 ARG_COPY_REF(D3DMATERIAL9, pMaterial))
1868 {
1869 struct nine_context *context = &device->context;
1870
1871 context->ff.material = *pMaterial;
1872 context->changed.group |= NINE_STATE_FF_MATERIAL;
1873 }
1874
1875 CSMT_ITEM_NO_WAIT(nine_context_set_light,
1876 ARG_VAL(DWORD, Index),
1877 ARG_COPY_REF(D3DLIGHT9, pLight))
1878 {
1879 struct nine_context *context = &device->context;
1880
1881 (void)nine_state_set_light(&context->ff, Index, pLight);
1882 context->changed.group |= NINE_STATE_FF_LIGHTING;
1883 }
1884
1885
1886 /* For stateblocks */
1887 static void
1888 nine_context_light_enable_stateblock(struct NineDevice9 *device,
1889 const uint16_t active_light[NINE_MAX_LIGHTS_ACTIVE], /* TODO: use pointer that convey size for csmt */
1890 unsigned int num_lights_active)
1891 {
1892 struct nine_context *context = &device->context;
1893
1894 /* TODO: Use CSMT_* to avoid calling nine_csmt_process */
1895 nine_csmt_process(device);
1896 memcpy(context->ff.active_light, active_light, NINE_MAX_LIGHTS_ACTIVE * sizeof(context->ff.active_light[0]));
1897 context->ff.num_lights_active = num_lights_active;
1898 context->changed.group |= NINE_STATE_FF_LIGHTING;
1899 }
1900
1901 CSMT_ITEM_NO_WAIT(nine_context_light_enable,
1902 ARG_VAL(DWORD, Index),
1903 ARG_VAL(BOOL, Enable))
1904 {
1905 struct nine_context *context = &device->context;
1906
1907 nine_state_light_enable(&context->ff, Index, Enable);
1908 context->changed.group |= NINE_STATE_FF_LIGHTING;
1909 }
1910
1911 CSMT_ITEM_NO_WAIT(nine_context_set_texture_stage_state,
1912 ARG_VAL(DWORD, Stage),
1913 ARG_VAL(D3DTEXTURESTAGESTATETYPE, Type),
1914 ARG_VAL(DWORD, Value))
1915 {
1916 struct nine_context *context = &device->context;
1917 int bumpmap_index = -1;
1918
1919 context->ff.tex_stage[Stage][Type] = Value;
1920 switch (Type) {
1921 case D3DTSS_BUMPENVMAT00:
1922 bumpmap_index = 4 * Stage;
1923 break;
1924 case D3DTSS_BUMPENVMAT01:
1925 bumpmap_index = 4 * Stage + 1;
1926 break;
1927 case D3DTSS_BUMPENVMAT10:
1928 bumpmap_index = 4 * Stage + 2;
1929 break;
1930 case D3DTSS_BUMPENVMAT11:
1931 bumpmap_index = 4 * Stage + 3;
1932 break;
1933 case D3DTSS_BUMPENVLSCALE:
1934 bumpmap_index = 4 * 8 + 2 * Stage;
1935 break;
1936 case D3DTSS_BUMPENVLOFFSET:
1937 bumpmap_index = 4 * 8 + 2 * Stage + 1;
1938 break;
1939 case D3DTSS_TEXTURETRANSFORMFLAGS:
1940 context->changed.group |= NINE_STATE_PS_PARAMS_MISC;
1941 break;
1942 default:
1943 break;
1944 }
1945
1946 if (bumpmap_index >= 0) {
1947 context->bumpmap_vars[bumpmap_index] = Value;
1948 context->changed.group |= NINE_STATE_PS_CONST;
1949 }
1950
1951 context->changed.group |= NINE_STATE_FF_PS_CONSTS;
1952 context->ff.changed.tex_stage[Stage][Type / 32] |= 1 << (Type % 32);
1953 }
1954
1955 CSMT_ITEM_NO_WAIT(nine_context_set_clip_plane,
1956 ARG_VAL(DWORD, Index),
1957 ARG_COPY_REF(struct nine_clipplane, pPlane))
1958 {
1959 struct nine_context *context = &device->context;
1960
1961 memcpy(&context->clip.ucp[Index][0], pPlane, sizeof(context->clip.ucp[0]));
1962 context->changed.ucp = TRUE;
1963 }
1964
1965 CSMT_ITEM_NO_WAIT(nine_context_set_swvp,
1966 ARG_VAL(boolean, swvp))
1967 {
1968 struct nine_context *context = &device->context;
1969
1970 context->swvp = swvp;
1971 context->changed.group |= NINE_STATE_SWVP;
1972 }
1973
1974 /* Do not write to nine_context directly. Slower,
1975 * but works with csmt. TODO: write a special csmt version that
1976 * would record the list of commands as much as possible,
1977 * and use the version above else.
1978 */
1979 void
1980 nine_context_apply_stateblock(struct NineDevice9 *device,
1981 const struct nine_state *src)
1982 {
1983 int i;
1984
1985 /* No need to apply src->changed.group, since all calls do
1986 * set context->changed.group */
1987
1988 for (i = 0; i < ARRAY_SIZE(src->changed.rs); ++i) {
1989 uint32_t m = src->changed.rs[i];
1990 while (m) {
1991 const int r = ffs(m) - 1;
1992 m &= ~(1 << r);
1993 nine_context_set_render_state(device, i * 32 + r, src->rs_advertised[i * 32 + r]);
1994 }
1995 }
1996
1997 /* Textures */
1998 if (src->changed.texture) {
1999 uint32_t m = src->changed.texture;
2000 unsigned s;
2001
2002 for (s = 0; m; ++s, m >>= 1) {
2003 struct NineBaseTexture9 *tex = src->texture[s];
2004 if (!(m & 1))
2005 continue;
2006 nine_context_set_texture(device, s, tex);
2007 }
2008 }
2009
2010 /* Sampler state */
2011 if (src->changed.group & NINE_STATE_SAMPLER) {
2012 unsigned s;
2013
2014 for (s = 0; s < NINE_MAX_SAMPLERS; ++s) {
2015 uint32_t m = src->changed.sampler[s];
2016 while (m) {
2017 const int i = ffs(m) - 1;
2018 m &= ~(1 << i);
2019 nine_context_set_sampler_state(device, s, i, src->samp_advertised[s][i]);
2020 }
2021 }
2022 }
2023
2024 /* Vertex buffers */
2025 if (src->changed.vtxbuf | src->changed.stream_freq) {
2026 uint32_t m = src->changed.vtxbuf | src->changed.stream_freq;
2027 for (i = 0; m; ++i, m >>= 1) {
2028 if (src->changed.vtxbuf & (1 << i))
2029 nine_context_set_stream_source(device, i, src->stream[i], src->vtxbuf[i].buffer_offset, src->vtxbuf[i].stride);
2030 if (src->changed.stream_freq & (1 << i))
2031 nine_context_set_stream_source_freq(device, i, src->stream_freq[i]);
2032 }
2033 }
2034
2035 /* Index buffer */
2036 if (src->changed.group & NINE_STATE_IDXBUF)
2037 nine_context_set_indices(device, src->idxbuf);
2038
2039 /* Vertex declaration */
2040 if ((src->changed.group & NINE_STATE_VDECL) && src->vdecl)
2041 nine_context_set_vertex_declaration(device, src->vdecl);
2042
2043 /* Vertex shader */
2044 if (src->changed.group & NINE_STATE_VS)
2045 nine_context_set_vertex_shader(device, src->vs);
2046
2047 /* Pixel shader */
2048 if (src->changed.group & NINE_STATE_PS)
2049 nine_context_set_pixel_shader(device, src->ps);
2050
2051 /* Vertex constants */
2052 if (src->changed.group & NINE_STATE_VS_CONST) {
2053 struct nine_range *r;
2054 for (r = src->changed.vs_const_f; r; r = r->next)
2055 nine_context_set_vertex_shader_constant_f(device, r->bgn,
2056 &src->vs_const_f[r->bgn * 4],
2057 sizeof(float[4]) * (r->end - r->bgn),
2058 r->end - r->bgn);
2059 for (r = src->changed.vs_const_i; r; r = r->next)
2060 nine_context_set_vertex_shader_constant_i(device, r->bgn,
2061 &src->vs_const_i[r->bgn * 4],
2062 sizeof(int[4]) * (r->end - r->bgn),
2063 r->end - r->bgn);
2064 for (r = src->changed.vs_const_b; r; r = r->next)
2065 nine_context_set_vertex_shader_constant_b(device, r->bgn,
2066 &src->vs_const_b[r->bgn * 4],
2067 sizeof(BOOL) * (r->end - r->bgn),
2068 r->end - r->bgn);
2069 }
2070
2071 /* Pixel constants */
2072 if (src->changed.group & NINE_STATE_PS_CONST) {
2073 struct nine_range *r;
2074 for (r = src->changed.ps_const_f; r; r = r->next)
2075 nine_context_set_pixel_shader_constant_f(device, r->bgn,
2076 &src->ps_const_f[r->bgn * 4],
2077 sizeof(float[4]) * (r->end - r->bgn),
2078 r->end - r->bgn);
2079 if (src->changed.ps_const_i) {
2080 uint16_t m = src->changed.ps_const_i;
2081 for (i = ffs(m) - 1, m >>= i; m; ++i, m >>= 1)
2082 if (m & 1)
2083 nine_context_set_pixel_shader_constant_i_transformed(device, i,
2084 src->ps_const_i[i], sizeof(int[4]), 1);
2085 }
2086 if (src->changed.ps_const_b) {
2087 uint16_t m = src->changed.ps_const_b;
2088 for (i = ffs(m) - 1, m >>= i; m; ++i, m >>= 1)
2089 if (m & 1)
2090 nine_context_set_pixel_shader_constant_b(device, i,
2091 &src->ps_const_b[i], sizeof(BOOL), 1);
2092 }
2093 }
2094
2095 /* Viewport */
2096 if (src->changed.group & NINE_STATE_VIEWPORT)
2097 nine_context_set_viewport(device, &src->viewport);
2098
2099 /* Scissor */
2100 if (src->changed.group & NINE_STATE_SCISSOR)
2101 nine_context_set_scissor(device, &src->scissor);
2102
2103 /* User Clip Planes */
2104 if (src->changed.ucp)
2105 for (i = 0; i < PIPE_MAX_CLIP_PLANES; ++i)
2106 if (src->changed.ucp & (1 << i))
2107 nine_context_set_clip_plane(device, i, (struct nine_clipplane*)&src->clip.ucp[i][0]);
2108
2109 if (!(src->changed.group & NINE_STATE_FF))
2110 return;
2111
2112 /* Fixed function state. */
2113
2114 if (src->changed.group & NINE_STATE_FF_MATERIAL)
2115 nine_context_set_material(device, &src->ff.material);
2116
2117 if (src->changed.group & NINE_STATE_FF_PS_CONSTS) {
2118 unsigned s;
2119 for (s = 0; s < NINE_MAX_TEXTURE_STAGES; ++s) {
2120 for (i = 0; i < NINED3DTSS_COUNT; ++i)
2121 if (src->ff.changed.tex_stage[s][i / 32] & (1 << (i % 32)))
2122 nine_context_set_texture_stage_state(device, s, i, src->ff.tex_stage[s][i]);
2123 }
2124 }
2125 if (src->changed.group & NINE_STATE_FF_LIGHTING) {
2126 for (i = 0; i < src->ff.num_lights; ++i)
2127 if (src->ff.light[i].Type != NINED3DLIGHT_INVALID)
2128 nine_context_set_light(device, i, &src->ff.light[i]);
2129
2130 nine_context_light_enable_stateblock(device, src->ff.active_light, src->ff.num_lights_active);
2131 }
2132 if (src->changed.group & NINE_STATE_FF_VSTRANSF) {
2133 for (i = 0; i < ARRAY_SIZE(src->ff.changed.transform); ++i) {
2134 unsigned s;
2135 if (!src->ff.changed.transform[i])
2136 continue;
2137 for (s = i * 32; s < (i * 32 + 32); ++s) {
2138 if (!(src->ff.changed.transform[i] & (1 << (s % 32))))
2139 continue;
2140 /* MaxVertexBlendMatrixIndex is 8, which means
2141 * we don't read past index D3DTS_WORLDMATRIX(8).
2142 * swvp is supposed to allow all 256, but we don't
2143 * implement it for now. */
2144 if (s > D3DTS_WORLDMATRIX(8))
2145 break;
2146 nine_context_set_transform(device, s,
2147 nine_state_access_transform(
2148 (struct nine_ff_state *)&src->ff,
2149 s, FALSE));
2150 }
2151 }
2152 }
2153 }
2154
2155 static void
2156 nine_update_state_framebuffer_clear(struct NineDevice9 *device)
2157 {
2158 struct nine_context *context = &device->context;
2159
2160 if (context->changed.group & NINE_STATE_FB)
2161 update_framebuffer(device, TRUE);
2162 }
2163
2164 CSMT_ITEM_NO_WAIT(nine_context_clear_fb,
2165 ARG_VAL(DWORD, Count),
2166 ARG_COPY_REF(D3DRECT, pRects),
2167 ARG_VAL(DWORD, Flags),
2168 ARG_VAL(D3DCOLOR, Color),
2169 ARG_VAL(float, Z),
2170 ARG_VAL(DWORD, Stencil))
2171 {
2172 struct nine_context *context = &device->context;
2173 const int sRGB = context->rs[D3DRS_SRGBWRITEENABLE] ? 1 : 0;
2174 struct pipe_surface *cbuf, *zsbuf;
2175 struct pipe_context *pipe = context->pipe;
2176 struct NineSurface9 *zsbuf_surf = context->ds;
2177 struct NineSurface9 *rt;
2178 unsigned bufs = 0;
2179 unsigned r, i;
2180 union pipe_color_union rgba;
2181 unsigned rt_mask = 0;
2182 D3DRECT rect;
2183
2184 nine_update_state_framebuffer_clear(device);
2185
2186 if (Flags & D3DCLEAR_TARGET) bufs |= PIPE_CLEAR_COLOR;
2187 /* Ignore Z buffer if not bound */
2188 if (context->pipe_data.fb.zsbuf != NULL) {
2189 if (Flags & D3DCLEAR_ZBUFFER) bufs |= PIPE_CLEAR_DEPTH;
2190 if (Flags & D3DCLEAR_STENCIL) bufs |= PIPE_CLEAR_STENCIL;
2191 }
2192 if (!bufs)
2193 return;
2194 d3dcolor_to_pipe_color_union(&rgba, Color);
2195
2196 rect.x1 = context->viewport.X;
2197 rect.y1 = context->viewport.Y;
2198 rect.x2 = context->viewport.Width + rect.x1;
2199 rect.y2 = context->viewport.Height + rect.y1;
2200
2201 /* Both rectangles apply, which is weird, but that's D3D9. */
2202 if (context->rs[D3DRS_SCISSORTESTENABLE]) {
2203 rect.x1 = MAX2(rect.x1, context->scissor.minx);
2204 rect.y1 = MAX2(rect.y1, context->scissor.miny);
2205 rect.x2 = MIN2(rect.x2, context->scissor.maxx);
2206 rect.y2 = MIN2(rect.y2, context->scissor.maxy);
2207 }
2208
2209 if (Count) {
2210 /* Maybe apps like to specify a large rect ? */
2211 if (pRects[0].x1 <= rect.x1 && pRects[0].x2 >= rect.x2 &&
2212 pRects[0].y1 <= rect.y1 && pRects[0].y2 >= rect.y2) {
2213 DBG("First rect covers viewport.\n");
2214 Count = 0;
2215 pRects = NULL;
2216 }
2217 }
2218
2219 if (rect.x1 >= context->pipe_data.fb.width || rect.y1 >= context->pipe_data.fb.height)
2220 return;
2221
2222 for (i = 0; i < device->caps.NumSimultaneousRTs; ++i) {
2223 if (context->rt[i] && context->rt[i]->desc.Format != D3DFMT_NULL)
2224 rt_mask |= 1 << i;
2225 }
2226
2227 /* fast path, clears everything at once */
2228 if (!Count &&
2229 (!(bufs & PIPE_CLEAR_COLOR) || (rt_mask == context->rt_mask)) &&
2230 rect.x1 == 0 && rect.y1 == 0 &&
2231 /* Case we clear only render target. Check clear region vs rt. */
2232 ((!(bufs & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) &&
2233 rect.x2 >= context->pipe_data.fb.width &&
2234 rect.y2 >= context->pipe_data.fb.height) ||
2235 /* Case we clear depth buffer (and eventually rt too).
2236 * depth buffer size is always >= rt size. Compare to clear region */
2237 ((bufs & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) &&
2238 rect.x2 >= zsbuf_surf->desc.Width &&
2239 rect.y2 >= zsbuf_surf->desc.Height))) {
2240 DBG("Clear fast path\n");
2241 pipe->clear(pipe, bufs, &rgba, Z, Stencil);
2242 return;
2243 }
2244
2245 if (!Count) {
2246 Count = 1;
2247 pRects = &rect;
2248 }
2249
2250 for (i = 0; i < device->caps.NumSimultaneousRTs; ++i) {
2251 rt = context->rt[i];
2252 if (!rt || rt->desc.Format == D3DFMT_NULL ||
2253 !(bufs & PIPE_CLEAR_COLOR))
2254 continue; /* save space, compiler should hoist this */
2255 cbuf = NineSurface9_GetSurface(rt, sRGB);
2256 for (r = 0; r < Count; ++r) {
2257 /* Don't trust users to pass these in the right order. */
2258 unsigned x1 = MIN2(pRects[r].x1, pRects[r].x2);
2259 unsigned y1 = MIN2(pRects[r].y1, pRects[r].y2);
2260 unsigned x2 = MAX2(pRects[r].x1, pRects[r].x2);
2261 unsigned y2 = MAX2(pRects[r].y1, pRects[r].y2);
2262 #ifndef NINE_LAX
2263 /* Drop negative rectangles (like wine expects). */
2264 if (pRects[r].x1 > pRects[r].x2) continue;
2265 if (pRects[r].y1 > pRects[r].y2) continue;
2266 #endif
2267
2268 x1 = MAX2(x1, rect.x1);
2269 y1 = MAX2(y1, rect.y1);
2270 x2 = MIN3(x2, rect.x2, rt->desc.Width);
2271 y2 = MIN3(y2, rect.y2, rt->desc.Height);
2272
2273 DBG("Clearing (%u..%u)x(%u..%u)\n", x1, x2, y1, y2);
2274 pipe->clear_render_target(pipe, cbuf, &rgba,
2275 x1, y1, x2 - x1, y2 - y1, false);
2276 }
2277 }
2278 if (!(bufs & PIPE_CLEAR_DEPTHSTENCIL))
2279 return;
2280
2281 bufs &= PIPE_CLEAR_DEPTHSTENCIL;
2282
2283 for (r = 0; r < Count; ++r) {
2284 unsigned x1 = MIN2(pRects[r].x1, pRects[r].x2);
2285 unsigned y1 = MIN2(pRects[r].y1, pRects[r].y2);
2286 unsigned x2 = MAX2(pRects[r].x1, pRects[r].x2);
2287 unsigned y2 = MAX2(pRects[r].y1, pRects[r].y2);
2288 #ifndef NINE_LAX
2289 /* Drop negative rectangles. */
2290 if (pRects[r].x1 > pRects[r].x2) continue;
2291 if (pRects[r].y1 > pRects[r].y2) continue;
2292 #endif
2293
2294 x1 = MIN2(x1, rect.x1);
2295 y1 = MIN2(y1, rect.y1);
2296 x2 = MIN3(x2, rect.x2, zsbuf_surf->desc.Width);
2297 y2 = MIN3(y2, rect.y2, zsbuf_surf->desc.Height);
2298
2299 zsbuf = NineSurface9_GetSurface(zsbuf_surf, 0);
2300 assert(zsbuf);
2301 pipe->clear_depth_stencil(pipe, zsbuf, bufs, Z, Stencil,
2302 x1, y1, x2 - x1, y2 - y1, false);
2303 }
2304 return;
2305 }
2306
2307
2308 static inline void
2309 init_draw_info(struct pipe_draw_info *info,
2310 struct NineDevice9 *dev, D3DPRIMITIVETYPE type, UINT count)
2311 {
2312 info->mode = d3dprimitivetype_to_pipe_prim(type);
2313 info->count = prim_count_to_vertex_count(type, count);
2314 info->start_instance = 0;
2315 info->instance_count = 1;
2316 if (dev->context.stream_instancedata_mask & dev->context.stream_usage_mask)
2317 info->instance_count = MAX2(dev->context.stream_freq[0] & 0x7FFFFF, 1);
2318 info->primitive_restart = FALSE;
2319 info->has_user_indices = FALSE;
2320 info->restart_index = 0;
2321 info->count_from_stream_output = NULL;
2322 info->indirect = NULL;
2323 }
2324
2325 CSMT_ITEM_NO_WAIT(nine_context_draw_primitive,
2326 ARG_VAL(D3DPRIMITIVETYPE, PrimitiveType),
2327 ARG_VAL(UINT, StartVertex),
2328 ARG_VAL(UINT, PrimitiveCount))
2329 {
2330 struct nine_context *context = &device->context;
2331 struct pipe_draw_info info;
2332
2333 nine_update_state(device);
2334
2335 init_draw_info(&info, device, PrimitiveType, PrimitiveCount);
2336 info.index_size = 0;
2337 info.start = StartVertex;
2338 info.index_bias = 0;
2339 info.min_index = info.start;
2340 info.max_index = info.count - 1;
2341 info.index.resource = NULL;
2342
2343 context->pipe->draw_vbo(context->pipe, &info);
2344 }
2345
2346 CSMT_ITEM_NO_WAIT(nine_context_draw_indexed_primitive,
2347 ARG_VAL(D3DPRIMITIVETYPE, PrimitiveType),
2348 ARG_VAL(INT, BaseVertexIndex),
2349 ARG_VAL(UINT, MinVertexIndex),
2350 ARG_VAL(UINT, NumVertices),
2351 ARG_VAL(UINT, StartIndex),
2352 ARG_VAL(UINT, PrimitiveCount))
2353 {
2354 struct nine_context *context = &device->context;
2355 struct pipe_draw_info info;
2356
2357 nine_update_state(device);
2358
2359 init_draw_info(&info, device, PrimitiveType, PrimitiveCount);
2360 info.index_size = context->index_size;
2361 info.start = context->index_offset / context->index_size + StartIndex;
2362 info.index_bias = BaseVertexIndex;
2363 /* These don't include index bias: */
2364 info.min_index = MinVertexIndex;
2365 info.max_index = MinVertexIndex + NumVertices - 1;
2366 info.index.resource = context->idxbuf;
2367
2368 context->pipe->draw_vbo(context->pipe, &info);
2369 }
2370
2371 CSMT_ITEM_NO_WAIT(nine_context_draw_primitive_from_vtxbuf,
2372 ARG_VAL(D3DPRIMITIVETYPE, PrimitiveType),
2373 ARG_VAL(UINT, PrimitiveCount),
2374 ARG_BIND_VBUF(struct pipe_vertex_buffer, vtxbuf))
2375 {
2376 struct nine_context *context = &device->context;
2377 struct pipe_draw_info info;
2378
2379 nine_update_state(device);
2380
2381 init_draw_info(&info, device, PrimitiveType, PrimitiveCount);
2382 info.index_size = 0;
2383 info.start = 0;
2384 info.index_bias = 0;
2385 info.min_index = 0;
2386 info.max_index = info.count - 1;
2387 info.index.resource = NULL;
2388
2389 context->pipe->set_vertex_buffers(context->pipe, 0, 1, vtxbuf);
2390
2391 context->pipe->draw_vbo(context->pipe, &info);
2392 }
2393
2394 CSMT_ITEM_NO_WAIT(nine_context_draw_indexed_primitive_from_vtxbuf_idxbuf,
2395 ARG_VAL(D3DPRIMITIVETYPE, PrimitiveType),
2396 ARG_VAL(UINT, MinVertexIndex),
2397 ARG_VAL(UINT, NumVertices),
2398 ARG_VAL(UINT, PrimitiveCount),
2399 ARG_BIND_VBUF(struct pipe_vertex_buffer, vbuf),
2400 ARG_BIND_RES(struct pipe_resource, ibuf),
2401 ARG_VAL(void *, user_ibuf),
2402 ARG_VAL(UINT, index_offset),
2403 ARG_VAL(UINT, index_size))
2404 {
2405 struct nine_context *context = &device->context;
2406 struct pipe_draw_info info;
2407
2408 nine_update_state(device);
2409
2410 init_draw_info(&info, device, PrimitiveType, PrimitiveCount);
2411 info.index_size = index_size;
2412 info.start = index_offset / info.index_size;
2413 info.index_bias = 0;
2414 info.min_index = MinVertexIndex;
2415 info.max_index = MinVertexIndex + NumVertices - 1;
2416 info.has_user_indices = ibuf == NULL;
2417 if (ibuf)
2418 info.index.resource = ibuf;
2419 else
2420 info.index.user = user_ibuf;
2421
2422 context->pipe->set_vertex_buffers(context->pipe, 0, 1, vbuf);
2423
2424 context->pipe->draw_vbo(context->pipe, &info);
2425 }
2426
2427 CSMT_ITEM_NO_WAIT(nine_context_resource_copy_region,
2428 ARG_BIND_REF(struct NineUnknown, dst),
2429 ARG_BIND_REF(struct NineUnknown, src),
2430 ARG_BIND_RES(struct pipe_resource, dst_res),
2431 ARG_VAL(unsigned, dst_level),
2432 ARG_COPY_REF(struct pipe_box, dst_box),
2433 ARG_BIND_RES(struct pipe_resource, src_res),
2434 ARG_VAL(unsigned, src_level),
2435 ARG_COPY_REF(struct pipe_box, src_box))
2436 {
2437 struct nine_context *context = &device->context;
2438
2439 (void) dst;
2440 (void) src;
2441
2442 context->pipe->resource_copy_region(context->pipe,
2443 dst_res, dst_level,
2444 dst_box->x, dst_box->y, dst_box->z,
2445 src_res, src_level,
2446 src_box);
2447 }
2448
2449 CSMT_ITEM_NO_WAIT(nine_context_blit,
2450 ARG_BIND_REF(struct NineUnknown, dst),
2451 ARG_BIND_REF(struct NineUnknown, src),
2452 ARG_BIND_BLIT(struct pipe_blit_info, blit))
2453 {
2454 struct nine_context *context = &device->context;
2455
2456 (void) dst;
2457 (void) src;
2458
2459 context->pipe->blit(context->pipe, blit);
2460 }
2461
2462 CSMT_ITEM_NO_WAIT(nine_context_clear_render_target,
2463 ARG_BIND_REF(struct NineSurface9, surface),
2464 ARG_VAL(D3DCOLOR, color),
2465 ARG_VAL(UINT, x),
2466 ARG_VAL(UINT, y),
2467 ARG_VAL(UINT, width),
2468 ARG_VAL(UINT, height))
2469 {
2470 struct nine_context *context = &device->context;
2471 struct pipe_surface *surf;
2472 union pipe_color_union rgba;
2473
2474 d3dcolor_to_pipe_color_union(&rgba, color);
2475 surf = NineSurface9_GetSurface(surface, 0);
2476 context->pipe->clear_render_target(context->pipe, surf, &rgba, x, y, width, height, false);
2477 }
2478
2479 CSMT_ITEM_NO_WAIT(nine_context_gen_mipmap,
2480 ARG_BIND_REF(struct NineUnknown, dst),
2481 ARG_BIND_RES(struct pipe_resource, res),
2482 ARG_VAL(UINT, base_level),
2483 ARG_VAL(UINT, last_level),
2484 ARG_VAL(UINT, first_layer),
2485 ARG_VAL(UINT, last_layer),
2486 ARG_VAL(UINT, filter))
2487 {
2488 struct nine_context *context = &device->context;
2489
2490 /* We just bind dst for the bind count */
2491 (void)dst;
2492
2493 util_gen_mipmap(context->pipe, res, res->format, base_level,
2494 last_level, first_layer, last_layer, filter);
2495 }
2496
2497 CSMT_ITEM_NO_WAIT_WITH_COUNTER(nine_context_range_upload,
2498 ARG_BIND_REF(struct NineUnknown, src_ref),
2499 ARG_BIND_RES(struct pipe_resource, res),
2500 ARG_VAL(unsigned, offset),
2501 ARG_VAL(unsigned, size),
2502 ARG_VAL(const void *, data))
2503 {
2504 struct nine_context *context = &device->context;
2505
2506 /* Binding src_ref avoids release before upload */
2507 (void)src_ref;
2508
2509 context->pipe->buffer_subdata(context->pipe, res, 0, offset, size, data);
2510 }
2511
2512 CSMT_ITEM_NO_WAIT_WITH_COUNTER(nine_context_box_upload,
2513 ARG_BIND_REF(struct NineUnknown, src_ref),
2514 ARG_BIND_RES(struct pipe_resource, res),
2515 ARG_VAL(unsigned, level),
2516 ARG_COPY_REF(struct pipe_box, dst_box),
2517 ARG_VAL(enum pipe_format, src_format),
2518 ARG_VAL(const void *, src),
2519 ARG_VAL(unsigned, src_stride),
2520 ARG_VAL(unsigned, src_layer_stride),
2521 ARG_COPY_REF(struct pipe_box, src_box))
2522 {
2523 struct nine_context *context = &device->context;
2524 struct pipe_context *pipe = context->pipe;
2525 struct pipe_transfer *transfer = NULL;
2526 uint8_t *map;
2527
2528 /* Binding src_ref avoids release before upload */
2529 (void)src_ref;
2530
2531 map = pipe->transfer_map(pipe,
2532 res,
2533 level,
2534 PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE,
2535 dst_box, &transfer);
2536 if (!map)
2537 return;
2538
2539 /* Note: if formats are the sames, it will revert
2540 * to normal memcpy */
2541 (void) util_format_translate_3d(res->format,
2542 map, transfer->stride,
2543 transfer->layer_stride,
2544 0, 0, 0,
2545 src_format,
2546 src, src_stride,
2547 src_layer_stride,
2548 src_box->x, src_box->y, src_box->z,
2549 dst_box->width, dst_box->height,
2550 dst_box->depth);
2551
2552 pipe_transfer_unmap(pipe, transfer);
2553 }
2554
2555 struct pipe_query *
2556 nine_context_create_query(struct NineDevice9 *device, unsigned query_type)
2557 {
2558 struct pipe_context *pipe;
2559 struct pipe_query *res;
2560
2561 pipe = nine_context_get_pipe_acquire(device);
2562 res = pipe->create_query(pipe, query_type, 0);
2563 nine_context_get_pipe_release(device);
2564 return res;
2565 }
2566
2567 CSMT_ITEM_DO_WAIT(nine_context_destroy_query,
2568 ARG_REF(struct pipe_query, query))
2569 {
2570 struct nine_context *context = &device->context;
2571
2572 context->pipe->destroy_query(context->pipe, query);
2573 }
2574
2575 CSMT_ITEM_NO_WAIT_WITH_COUNTER(nine_context_begin_query,
2576 ARG_REF(struct pipe_query, query))
2577 {
2578 struct nine_context *context = &device->context;
2579
2580 (void) context->pipe->begin_query(context->pipe, query);
2581 }
2582
2583 CSMT_ITEM_NO_WAIT_WITH_COUNTER(nine_context_end_query,
2584 ARG_REF(struct pipe_query, query))
2585 {
2586 struct nine_context *context = &device->context;
2587
2588 (void) context->pipe->end_query(context->pipe, query);
2589 }
2590
2591 boolean
2592 nine_context_get_query_result(struct NineDevice9 *device, struct pipe_query *query,
2593 unsigned *counter, boolean flush, boolean wait,
2594 union pipe_query_result *result)
2595 {
2596 struct pipe_context *pipe;
2597 boolean ret;
2598
2599 if (wait)
2600 nine_csmt_process(device);
2601 else if (p_atomic_read(counter) > 0) {
2602 if (flush && device->csmt_active)
2603 nine_queue_flush(device->csmt_ctx->pool);
2604 DBG("Pending begin/end. Returning\n");
2605 return false;
2606 }
2607
2608 pipe = nine_context_get_pipe_acquire(device);
2609 ret = pipe->get_query_result(pipe, query, wait, result);
2610 nine_context_get_pipe_release(device);
2611
2612 DBG("Query result %s\n", ret ? "found" : "not yet available");
2613 return ret;
2614 }
2615
2616 /* State defaults */
2617
2618 static const DWORD nine_render_state_defaults[NINED3DRS_LAST + 1] =
2619 {
2620 /* [D3DRS_ZENABLE] = D3DZB_TRUE; wine: auto_depth_stencil */
2621 [D3DRS_ZENABLE] = D3DZB_FALSE,
2622 [D3DRS_FILLMODE] = D3DFILL_SOLID,
2623 [D3DRS_SHADEMODE] = D3DSHADE_GOURAUD,
2624 /* [D3DRS_LINEPATTERN] = 0x00000000, */
2625 [D3DRS_ZWRITEENABLE] = TRUE,
2626 [D3DRS_ALPHATESTENABLE] = FALSE,
2627 [D3DRS_LASTPIXEL] = TRUE,
2628 [D3DRS_SRCBLEND] = D3DBLEND_ONE,
2629 [D3DRS_DESTBLEND] = D3DBLEND_ZERO,
2630 [D3DRS_CULLMODE] = D3DCULL_CCW,
2631 [D3DRS_ZFUNC] = D3DCMP_LESSEQUAL,
2632 [D3DRS_ALPHAFUNC] = D3DCMP_ALWAYS,
2633 [D3DRS_ALPHAREF] = 0,
2634 [D3DRS_DITHERENABLE] = FALSE,
2635 [D3DRS_ALPHABLENDENABLE] = FALSE,
2636 [D3DRS_FOGENABLE] = FALSE,
2637 [D3DRS_SPECULARENABLE] = FALSE,
2638 /* [D3DRS_ZVISIBLE] = 0, */
2639 [D3DRS_FOGCOLOR] = 0,
2640 [D3DRS_FOGTABLEMODE] = D3DFOG_NONE,
2641 [D3DRS_FOGSTART] = 0x00000000,
2642 [D3DRS_FOGEND] = 0x3F800000,
2643 [D3DRS_FOGDENSITY] = 0x3F800000,
2644 /* [D3DRS_EDGEANTIALIAS] = FALSE, */
2645 [D3DRS_RANGEFOGENABLE] = FALSE,
2646 [D3DRS_STENCILENABLE] = FALSE,
2647 [D3DRS_STENCILFAIL] = D3DSTENCILOP_KEEP,
2648 [D3DRS_STENCILZFAIL] = D3DSTENCILOP_KEEP,
2649 [D3DRS_STENCILPASS] = D3DSTENCILOP_KEEP,
2650 [D3DRS_STENCILREF] = 0,
2651 [D3DRS_STENCILMASK] = 0xFFFFFFFF,
2652 [D3DRS_STENCILFUNC] = D3DCMP_ALWAYS,
2653 [D3DRS_STENCILWRITEMASK] = 0xFFFFFFFF,
2654 [D3DRS_TEXTUREFACTOR] = 0xFFFFFFFF,
2655 [D3DRS_WRAP0] = 0,
2656 [D3DRS_WRAP1] = 0,
2657 [D3DRS_WRAP2] = 0,
2658 [D3DRS_WRAP3] = 0,
2659 [D3DRS_WRAP4] = 0,
2660 [D3DRS_WRAP5] = 0,
2661 [D3DRS_WRAP6] = 0,
2662 [D3DRS_WRAP7] = 0,
2663 [D3DRS_CLIPPING] = TRUE,
2664 [D3DRS_LIGHTING] = TRUE,
2665 [D3DRS_AMBIENT] = 0,
2666 [D3DRS_FOGVERTEXMODE] = D3DFOG_NONE,
2667 [D3DRS_COLORVERTEX] = TRUE,
2668 [D3DRS_LOCALVIEWER] = TRUE,
2669 [D3DRS_NORMALIZENORMALS] = FALSE,
2670 [D3DRS_DIFFUSEMATERIALSOURCE] = D3DMCS_COLOR1,
2671 [D3DRS_SPECULARMATERIALSOURCE] = D3DMCS_COLOR2,
2672 [D3DRS_AMBIENTMATERIALSOURCE] = D3DMCS_MATERIAL,
2673 [D3DRS_EMISSIVEMATERIALSOURCE] = D3DMCS_MATERIAL,
2674 [D3DRS_VERTEXBLEND] = D3DVBF_DISABLE,
2675 [D3DRS_CLIPPLANEENABLE] = 0,
2676 /* [D3DRS_SOFTWAREVERTEXPROCESSING] = FALSE, */
2677 [D3DRS_POINTSIZE] = 0x3F800000,
2678 [D3DRS_POINTSIZE_MIN] = 0x3F800000,
2679 [D3DRS_POINTSPRITEENABLE] = FALSE,
2680 [D3DRS_POINTSCALEENABLE] = FALSE,
2681 [D3DRS_POINTSCALE_A] = 0x3F800000,
2682 [D3DRS_POINTSCALE_B] = 0x00000000,
2683 [D3DRS_POINTSCALE_C] = 0x00000000,
2684 [D3DRS_MULTISAMPLEANTIALIAS] = TRUE,
2685 [D3DRS_MULTISAMPLEMASK] = 0xFFFFFFFF,
2686 [D3DRS_PATCHEDGESTYLE] = D3DPATCHEDGE_DISCRETE,
2687 /* [D3DRS_PATCHSEGMENTS] = 0x3F800000, */
2688 [D3DRS_DEBUGMONITORTOKEN] = 0xDEADCAFE,
2689 [D3DRS_POINTSIZE_MAX] = 0x3F800000, /* depends on cap */
2690 [D3DRS_INDEXEDVERTEXBLENDENABLE] = FALSE,
2691 [D3DRS_COLORWRITEENABLE] = 0x0000000f,
2692 [D3DRS_TWEENFACTOR] = 0x00000000,
2693 [D3DRS_BLENDOP] = D3DBLENDOP_ADD,
2694 [D3DRS_POSITIONDEGREE] = D3DDEGREE_CUBIC,
2695 [D3DRS_NORMALDEGREE] = D3DDEGREE_LINEAR,
2696 [D3DRS_SCISSORTESTENABLE] = FALSE,
2697 [D3DRS_SLOPESCALEDEPTHBIAS] = 0,
2698 [D3DRS_MINTESSELLATIONLEVEL] = 0x3F800000,
2699 [D3DRS_MAXTESSELLATIONLEVEL] = 0x3F800000,
2700 [D3DRS_ANTIALIASEDLINEENABLE] = FALSE,
2701 [D3DRS_ADAPTIVETESS_X] = 0x00000000,
2702 [D3DRS_ADAPTIVETESS_Y] = 0x00000000,
2703 [D3DRS_ADAPTIVETESS_Z] = 0x3F800000,
2704 [D3DRS_ADAPTIVETESS_W] = 0x00000000,
2705 [D3DRS_ENABLEADAPTIVETESSELLATION] = FALSE,
2706 [D3DRS_TWOSIDEDSTENCILMODE] = FALSE,
2707 [D3DRS_CCW_STENCILFAIL] = D3DSTENCILOP_KEEP,
2708 [D3DRS_CCW_STENCILZFAIL] = D3DSTENCILOP_KEEP,
2709 [D3DRS_CCW_STENCILPASS] = D3DSTENCILOP_KEEP,
2710 [D3DRS_CCW_STENCILFUNC] = D3DCMP_ALWAYS,
2711 [D3DRS_COLORWRITEENABLE1] = 0x0000000F,
2712 [D3DRS_COLORWRITEENABLE2] = 0x0000000F,
2713 [D3DRS_COLORWRITEENABLE3] = 0x0000000F,
2714 [D3DRS_BLENDFACTOR] = 0xFFFFFFFF,
2715 [D3DRS_SRGBWRITEENABLE] = 0,
2716 [D3DRS_DEPTHBIAS] = 0,
2717 [D3DRS_WRAP8] = 0,
2718 [D3DRS_WRAP9] = 0,
2719 [D3DRS_WRAP10] = 0,
2720 [D3DRS_WRAP11] = 0,
2721 [D3DRS_WRAP12] = 0,
2722 [D3DRS_WRAP13] = 0,
2723 [D3DRS_WRAP14] = 0,
2724 [D3DRS_WRAP15] = 0,
2725 [D3DRS_SEPARATEALPHABLENDENABLE] = FALSE,
2726 [D3DRS_SRCBLENDALPHA] = D3DBLEND_ONE,
2727 [D3DRS_DESTBLENDALPHA] = D3DBLEND_ZERO,
2728 [D3DRS_BLENDOPALPHA] = D3DBLENDOP_ADD,
2729 [NINED3DRS_VSPOINTSIZE] = FALSE,
2730 [NINED3DRS_RTMASK] = 0xf,
2731 [NINED3DRS_ALPHACOVERAGE] = FALSE,
2732 [NINED3DRS_MULTISAMPLE] = FALSE
2733 };
2734 static const DWORD nine_tex_stage_state_defaults[NINED3DTSS_LAST + 1] =
2735 {
2736 [D3DTSS_COLOROP] = D3DTOP_DISABLE,
2737 [D3DTSS_ALPHAOP] = D3DTOP_DISABLE,
2738 [D3DTSS_COLORARG1] = D3DTA_TEXTURE,
2739 [D3DTSS_COLORARG2] = D3DTA_CURRENT,
2740 [D3DTSS_COLORARG0] = D3DTA_CURRENT,
2741 [D3DTSS_ALPHAARG1] = D3DTA_TEXTURE,
2742 [D3DTSS_ALPHAARG2] = D3DTA_CURRENT,
2743 [D3DTSS_ALPHAARG0] = D3DTA_CURRENT,
2744 [D3DTSS_RESULTARG] = D3DTA_CURRENT,
2745 [D3DTSS_BUMPENVMAT00] = 0,
2746 [D3DTSS_BUMPENVMAT01] = 0,
2747 [D3DTSS_BUMPENVMAT10] = 0,
2748 [D3DTSS_BUMPENVMAT11] = 0,
2749 [D3DTSS_BUMPENVLSCALE] = 0,
2750 [D3DTSS_BUMPENVLOFFSET] = 0,
2751 [D3DTSS_TEXCOORDINDEX] = 0,
2752 [D3DTSS_TEXTURETRANSFORMFLAGS] = D3DTTFF_DISABLE,
2753 };
2754 static const DWORD nine_samp_state_defaults[NINED3DSAMP_LAST + 1] =
2755 {
2756 [D3DSAMP_ADDRESSU] = D3DTADDRESS_WRAP,
2757 [D3DSAMP_ADDRESSV] = D3DTADDRESS_WRAP,
2758 [D3DSAMP_ADDRESSW] = D3DTADDRESS_WRAP,
2759 [D3DSAMP_BORDERCOLOR] = 0,
2760 [D3DSAMP_MAGFILTER] = D3DTEXF_POINT,
2761 [D3DSAMP_MINFILTER] = D3DTEXF_POINT,
2762 [D3DSAMP_MIPFILTER] = D3DTEXF_NONE,
2763 [D3DSAMP_MIPMAPLODBIAS] = 0,
2764 [D3DSAMP_MAXMIPLEVEL] = 0,
2765 [D3DSAMP_MAXANISOTROPY] = 1,
2766 [D3DSAMP_SRGBTEXTURE] = 0,
2767 [D3DSAMP_ELEMENTINDEX] = 0,
2768 [D3DSAMP_DMAPOFFSET] = 0,
2769 [NINED3DSAMP_MINLOD] = 0,
2770 [NINED3DSAMP_SHADOW] = 0,
2771 [NINED3DSAMP_CUBETEX] = 0
2772 };
2773
2774 /* Note: The following 4 functions assume there is no
2775 * pending commands */
2776
2777 void nine_state_restore_non_cso(struct NineDevice9 *device)
2778 {
2779 struct nine_context *context = &device->context;
2780
2781 context->changed.group = NINE_STATE_ALL;
2782 context->changed.vtxbuf = (1ULL << device->caps.MaxStreams) - 1;
2783 context->changed.ucp = TRUE;
2784 context->commit |= NINE_STATE_COMMIT_CONST_VS | NINE_STATE_COMMIT_CONST_PS;
2785 }
2786
2787 void
2788 nine_state_set_defaults(struct NineDevice9 *device, const D3DCAPS9 *caps,
2789 boolean is_reset)
2790 {
2791 struct nine_state *state = &device->state;
2792 struct nine_context *context = &device->context;
2793 unsigned s;
2794
2795 /* Initialize defaults.
2796 */
2797 memcpy(context->rs, nine_render_state_defaults, sizeof(context->rs));
2798
2799 for (s = 0; s < ARRAY_SIZE(state->ff.tex_stage); ++s) {
2800 memcpy(&state->ff.tex_stage[s], nine_tex_stage_state_defaults,
2801 sizeof(state->ff.tex_stage[s]));
2802 state->ff.tex_stage[s][D3DTSS_TEXCOORDINDEX] = s;
2803 }
2804 state->ff.tex_stage[0][D3DTSS_COLOROP] = D3DTOP_MODULATE;
2805 state->ff.tex_stage[0][D3DTSS_ALPHAOP] = D3DTOP_SELECTARG1;
2806
2807 for (s = 0; s < ARRAY_SIZE(state->ff.tex_stage); ++s)
2808 memcpy(&context->ff.tex_stage[s], state->ff.tex_stage[s],
2809 sizeof(state->ff.tex_stage[s]));
2810
2811 memset(&context->bumpmap_vars, 0, sizeof(context->bumpmap_vars));
2812
2813 for (s = 0; s < NINE_MAX_SAMPLERS; ++s) {
2814 memcpy(&context->samp[s], nine_samp_state_defaults,
2815 sizeof(context->samp[s]));
2816 memcpy(&state->samp_advertised[s], nine_samp_state_defaults,
2817 sizeof(state->samp_advertised[s]));
2818 }
2819
2820 memset(state->vs_const_f, 0, VS_CONST_F_SIZE(device));
2821 memset(context->vs_const_f, 0, device->vs_const_size);
2822 if (context->vs_const_f_swvp)
2823 memset(context->vs_const_f_swvp, 0, NINE_MAX_CONST_F_SWVP * sizeof(float[4]));
2824 memset(state->vs_const_i, 0, VS_CONST_I_SIZE(device));
2825 memset(context->vs_const_i, 0, VS_CONST_I_SIZE(device));
2826 memset(state->vs_const_b, 0, VS_CONST_B_SIZE(device));
2827 memset(context->vs_const_b, 0, VS_CONST_B_SIZE(device));
2828 memset(state->ps_const_f, 0, device->ps_const_size);
2829 memset(context->ps_const_f, 0, device->ps_const_size);
2830 memset(state->ps_const_i, 0, sizeof(state->ps_const_i));
2831 memset(context->ps_const_i, 0, sizeof(context->ps_const_i));
2832 memset(state->ps_const_b, 0, sizeof(state->ps_const_b));
2833 memset(context->ps_const_b, 0, sizeof(context->ps_const_b));
2834
2835 /* Cap dependent initial state:
2836 */
2837 context->rs[D3DRS_POINTSIZE_MAX] = fui(caps->MaxPointSize);
2838
2839 memcpy(state->rs_advertised, context->rs, sizeof(context->rs));
2840
2841 /* Set changed flags to initialize driver.
2842 */
2843 context->changed.group = NINE_STATE_ALL;
2844 context->changed.vtxbuf = (1ULL << device->caps.MaxStreams) - 1;
2845 context->changed.ucp = TRUE;
2846
2847 context->ff.changed.transform[0] = ~0;
2848 context->ff.changed.transform[D3DTS_WORLD / 32] |= 1 << (D3DTS_WORLD % 32);
2849
2850 if (!is_reset) {
2851 state->viewport.MinZ = context->viewport.MinZ = 0.0f;
2852 state->viewport.MaxZ = context->viewport.MaxZ = 1.0f;
2853 }
2854
2855 for (s = 0; s < NINE_MAX_SAMPLERS; ++s)
2856 context->changed.sampler[s] = ~0;
2857
2858 if (!is_reset) {
2859 context->dummy_vbo_bound_at = -1;
2860 context->vbo_bound_done = FALSE;
2861 }
2862 }
2863
2864 void
2865 nine_device_state_clear(struct NineDevice9 *device)
2866 {
2867 struct nine_state *state = &device->state;
2868 unsigned i;
2869
2870 for (i = 0; i < ARRAY_SIZE(state->rt); ++i)
2871 nine_bind(&state->rt[i], NULL);
2872 nine_bind(&state->ds, NULL);
2873 nine_bind(&state->vs, NULL);
2874 nine_bind(&state->ps, NULL);
2875 nine_bind(&state->vdecl, NULL);
2876 for (i = 0; i < PIPE_MAX_ATTRIBS; ++i)
2877 NineBindBufferToDevice(device,
2878 (struct NineBuffer9 **)&state->stream[i],
2879 NULL);
2880 NineBindBufferToDevice(device,
2881 (struct NineBuffer9 **)&state->idxbuf,
2882 NULL);
2883
2884 for (i = 0; i < NINE_MAX_SAMPLERS; ++i)
2885 NineBindTextureToDevice(device, &state->texture[i], NULL);
2886 }
2887
2888 void
2889 nine_context_clear(struct NineDevice9 *device)
2890 {
2891 struct nine_context *context = &device->context;
2892 struct pipe_context *pipe = context->pipe;
2893 struct cso_context *cso = context->cso;
2894 unsigned i;
2895
2896 /* Early device ctor failure. Nothing to do */
2897 if (!pipe || !cso)
2898 return;
2899
2900 pipe->bind_vs_state(pipe, NULL);
2901 pipe->bind_fs_state(pipe, NULL);
2902
2903 /* Don't unbind constant buffers, they're device-private and
2904 * do not change on Reset.
2905 */
2906
2907 cso_set_samplers(cso, PIPE_SHADER_VERTEX, 0, NULL);
2908 cso_set_samplers(cso, PIPE_SHADER_FRAGMENT, 0, NULL);
2909
2910 cso_set_sampler_views(cso, PIPE_SHADER_VERTEX, 0, NULL);
2911 cso_set_sampler_views(cso, PIPE_SHADER_FRAGMENT, 0, NULL);
2912
2913 pipe->set_vertex_buffers(pipe, 0, device->caps.MaxStreams, NULL);
2914
2915 for (i = 0; i < ARRAY_SIZE(context->rt); ++i)
2916 nine_bind(&context->rt[i], NULL);
2917 nine_bind(&context->ds, NULL);
2918 nine_bind(&context->vs, NULL);
2919 nine_bind(&context->ps, NULL);
2920 nine_bind(&context->vdecl, NULL);
2921 for (i = 0; i < PIPE_MAX_ATTRIBS; ++i)
2922 pipe_vertex_buffer_unreference(&context->vtxbuf[i]);
2923 pipe_resource_reference(&context->idxbuf, NULL);
2924 pipe_resource_reference(&context->pipe_data.cb_vs.buffer, NULL);
2925 pipe_resource_reference(&context->pipe_data.cb_ps.buffer, NULL);
2926
2927 for (i = 0; i < NINE_MAX_SAMPLERS; ++i) {
2928 context->texture[i].enabled = FALSE;
2929 pipe_resource_reference(&context->texture[i].resource,
2930 NULL);
2931 pipe_sampler_view_reference(&context->texture[i].view[0],
2932 NULL);
2933 pipe_sampler_view_reference(&context->texture[i].view[1],
2934 NULL);
2935 }
2936 }
2937
2938 void
2939 nine_state_init_sw(struct NineDevice9 *device)
2940 {
2941 struct pipe_context *pipe_sw = device->pipe_sw;
2942 struct pipe_rasterizer_state rast;
2943 struct pipe_blend_state blend;
2944 struct pipe_depth_stencil_alpha_state dsa;
2945 struct pipe_framebuffer_state fb;
2946
2947 /* Only used with Streamout */
2948 memset(&rast, 0, sizeof(rast));
2949 rast.rasterizer_discard = true;
2950 rast.point_quad_rasterization = 1; /* to make llvmpipe happy */
2951 cso_set_rasterizer(device->cso_sw, &rast);
2952
2953 /* dummy settings */
2954 memset(&blend, 0, sizeof(blend));
2955 memset(&dsa, 0, sizeof(dsa));
2956 memset(&fb, 0, sizeof(fb));
2957 cso_set_blend(device->cso_sw, &blend);
2958 cso_set_depth_stencil_alpha(device->cso_sw, &dsa);
2959 cso_set_framebuffer(device->cso_sw, &fb);
2960 cso_set_viewport_dims(device->cso_sw, 1.0, 1.0, false);
2961 cso_set_fragment_shader_handle(device->cso_sw, util_make_empty_fragment_shader(pipe_sw));
2962 }
2963
2964 /* There is duplication with update_vertex_elements.
2965 * TODO: Share the code */
2966
2967 static void
2968 update_vertex_elements_sw(struct NineDevice9 *device)
2969 {
2970 struct nine_state *state = &device->state;
2971 const struct NineVertexDeclaration9 *vdecl = device->state.vdecl;
2972 const struct NineVertexShader9 *vs;
2973 unsigned n, b, i;
2974 int index;
2975 char vdecl_index_map[16]; /* vs->num_inputs <= 16 */
2976 char used_streams[device->caps.MaxStreams];
2977 int dummy_vbo_stream = -1;
2978 BOOL need_dummy_vbo = FALSE;
2979 struct cso_velems_state ve;
2980 bool programmable_vs = state->vs && !(state->vdecl && state->vdecl->position_t);
2981
2982 memset(vdecl_index_map, -1, 16);
2983 memset(used_streams, 0, device->caps.MaxStreams);
2984 vs = programmable_vs ? device->state.vs : device->ff.vs;
2985
2986 if (vdecl) {
2987 for (n = 0; n < vs->num_inputs; ++n) {
2988 DBG("looking up input %u (usage %u) from vdecl(%p)\n",
2989 n, vs->input_map[n].ndecl, vdecl);
2990
2991 for (i = 0; i < vdecl->nelems; i++) {
2992 if (vdecl->usage_map[i] == vs->input_map[n].ndecl) {
2993 vdecl_index_map[n] = i;
2994 used_streams[vdecl->elems[i].vertex_buffer_index] = 1;
2995 break;
2996 }
2997 }
2998 if (vdecl_index_map[n] < 0)
2999 need_dummy_vbo = TRUE;
3000 }
3001 } else {
3002 /* No vertex declaration. Likely will never happen in practice,
3003 * but we need not crash on this */
3004 need_dummy_vbo = TRUE;
3005 }
3006
3007 if (need_dummy_vbo) {
3008 for (i = 0; i < device->caps.MaxStreams; i++ ) {
3009 if (!used_streams[i]) {
3010 dummy_vbo_stream = i;
3011 break;
3012 }
3013 }
3014 }
3015 /* TODO handle dummy_vbo */
3016 assert (!need_dummy_vbo);
3017
3018 for (n = 0; n < vs->num_inputs; ++n) {
3019 index = vdecl_index_map[n];
3020 if (index >= 0) {
3021 ve.velems[n] = vdecl->elems[index];
3022 b = ve.velems[n].vertex_buffer_index;
3023 /* XXX wine just uses 1 here: */
3024 if (state->stream_freq[b] & D3DSTREAMSOURCE_INSTANCEDATA)
3025 ve.velems[n].instance_divisor = state->stream_freq[b] & 0x7FFFFF;
3026 } else {
3027 /* if the vertex declaration is incomplete compared to what the
3028 * vertex shader needs, we bind a dummy vbo with 0 0 0 0.
3029 * This is not precised by the spec, but is the behaviour
3030 * tested on win */
3031 ve.velems[n].vertex_buffer_index = dummy_vbo_stream;
3032 ve.velems[n].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
3033 ve.velems[n].src_offset = 0;
3034 ve.velems[n].instance_divisor = 0;
3035 }
3036 }
3037
3038 ve.count = vs->num_inputs;
3039 cso_set_vertex_elements(device->cso_sw, &ve);
3040 }
3041
3042 static void
3043 update_vertex_buffers_sw(struct NineDevice9 *device, int start_vertice, int num_vertices)
3044 {
3045 struct pipe_context *pipe = nine_context_get_pipe_acquire(device);
3046 struct pipe_context *pipe_sw = device->pipe_sw;
3047 struct nine_state *state = &device->state;
3048 struct nine_state_sw_internal *sw_internal = &device->state_sw_internal;
3049 struct pipe_vertex_buffer vtxbuf;
3050 uint32_t mask = 0xf;
3051 unsigned i;
3052
3053 DBG("mask=%x\n", mask);
3054
3055 /* TODO: handle dummy_vbo_bound_at */
3056
3057 for (i = 0; mask; mask >>= 1, ++i) {
3058 if (mask & 1) {
3059 if (state->stream[i]) {
3060 unsigned offset;
3061 struct pipe_resource *buf;
3062 struct pipe_box box;
3063 void *userbuf;
3064
3065 vtxbuf = state->vtxbuf[i];
3066 buf = NineVertexBuffer9_GetResource(state->stream[i], &offset);
3067
3068 DBG("Locking %p (offset %d, length %d)\n", buf,
3069 vtxbuf.buffer_offset, num_vertices * vtxbuf.stride);
3070
3071 u_box_1d(vtxbuf.buffer_offset + offset + start_vertice * vtxbuf.stride,
3072 num_vertices * vtxbuf.stride, &box);
3073
3074 userbuf = pipe->transfer_map(pipe, buf, 0, PIPE_TRANSFER_READ, &box,
3075 &(sw_internal->transfers_so[i]));
3076 vtxbuf.is_user_buffer = true;
3077 vtxbuf.buffer.user = userbuf;
3078
3079 if (!device->driver_caps.user_sw_vbufs) {
3080 vtxbuf.buffer.resource = NULL;
3081 vtxbuf.is_user_buffer = false;
3082 u_upload_data(device->pipe_sw->stream_uploader,
3083 0,
3084 box.width,
3085 16,
3086 userbuf,
3087 &(vtxbuf.buffer_offset),
3088 &(vtxbuf.buffer.resource));
3089 u_upload_unmap(device->pipe_sw->stream_uploader);
3090 }
3091 pipe_sw->set_vertex_buffers(pipe_sw, i, 1, &vtxbuf);
3092 pipe_vertex_buffer_unreference(&vtxbuf);
3093 } else
3094 pipe_sw->set_vertex_buffers(pipe_sw, i, 1, NULL);
3095 }
3096 }
3097 nine_context_get_pipe_release(device);
3098 }
3099
3100 static void
3101 update_vs_constants_sw(struct NineDevice9 *device)
3102 {
3103 struct nine_state *state = &device->state;
3104 struct pipe_context *pipe_sw = device->pipe_sw;
3105
3106 DBG("updating\n");
3107
3108 {
3109 struct pipe_constant_buffer cb;
3110 const void *buf;
3111
3112 cb.buffer = NULL;
3113 cb.buffer_offset = 0;
3114 cb.buffer_size = 4096 * sizeof(float[4]);
3115 cb.user_buffer = state->vs_const_f;
3116
3117 if (state->vs->lconstf.ranges) {
3118 const struct nine_lconstf *lconstf = &device->state.vs->lconstf;
3119 const struct nine_range *r = lconstf->ranges;
3120 unsigned n = 0;
3121 float *dst = device->state.vs_lconstf_temp;
3122 float *src = (float *)cb.user_buffer;
3123 memcpy(dst, src, 8192 * sizeof(float[4]));
3124 while (r) {
3125 unsigned p = r->bgn;
3126 unsigned c = r->end - r->bgn;
3127 memcpy(&dst[p * 4], &lconstf->data[n * 4], c * 4 * sizeof(float));
3128 n += c;
3129 r = r->next;
3130 }
3131 cb.user_buffer = dst;
3132 }
3133
3134 buf = cb.user_buffer;
3135
3136 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 0, &cb);
3137 if (cb.buffer)
3138 pipe_resource_reference(&cb.buffer, NULL);
3139
3140 cb.user_buffer = (char *)buf + 4096 * sizeof(float[4]);
3141
3142 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 1, &cb);
3143 if (cb.buffer)
3144 pipe_resource_reference(&cb.buffer, NULL);
3145 }
3146
3147 {
3148 struct pipe_constant_buffer cb;
3149
3150 cb.buffer = NULL;
3151 cb.buffer_offset = 0;
3152 cb.buffer_size = 2048 * sizeof(float[4]);
3153 cb.user_buffer = state->vs_const_i;
3154
3155 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 2, &cb);
3156 if (cb.buffer)
3157 pipe_resource_reference(&cb.buffer, NULL);
3158 }
3159
3160 {
3161 struct pipe_constant_buffer cb;
3162
3163 cb.buffer = NULL;
3164 cb.buffer_offset = 0;
3165 cb.buffer_size = 512 * sizeof(float[4]);
3166 cb.user_buffer = state->vs_const_b;
3167
3168 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 3, &cb);
3169 if (cb.buffer)
3170 pipe_resource_reference(&cb.buffer, NULL);
3171 }
3172
3173 {
3174 struct pipe_constant_buffer cb;
3175 const D3DVIEWPORT9 *vport = &device->state.viewport;
3176 float viewport_data[8] = {(float)vport->Width * 0.5f,
3177 (float)vport->Height * -0.5f, vport->MaxZ - vport->MinZ, 0.f,
3178 (float)vport->Width * 0.5f + (float)vport->X,
3179 (float)vport->Height * 0.5f + (float)vport->Y,
3180 vport->MinZ, 0.f};
3181
3182 cb.buffer = NULL;
3183 cb.buffer_offset = 0;
3184 cb.buffer_size = 2 * sizeof(float[4]);
3185 cb.user_buffer = viewport_data;
3186
3187 {
3188 u_upload_data(device->pipe_sw->const_uploader,
3189 0,
3190 cb.buffer_size,
3191 16,
3192 cb.user_buffer,
3193 &(cb.buffer_offset),
3194 &(cb.buffer));
3195 u_upload_unmap(device->pipe_sw->const_uploader);
3196 cb.user_buffer = NULL;
3197 }
3198
3199 pipe_sw->set_constant_buffer(pipe_sw, PIPE_SHADER_VERTEX, 4, &cb);
3200 if (cb.buffer)
3201 pipe_resource_reference(&cb.buffer, NULL);
3202 }
3203
3204 }
3205
3206 void
3207 nine_state_prepare_draw_sw(struct NineDevice9 *device, struct NineVertexDeclaration9 *vdecl_out,
3208 int start_vertice, int num_vertices, struct pipe_stream_output_info *so)
3209 {
3210 struct nine_state *state = &device->state;
3211 bool programmable_vs = state->vs && !(state->vdecl && state->vdecl->position_t);
3212 struct NineVertexShader9 *vs = programmable_vs ? device->state.vs : device->ff.vs;
3213
3214 assert(programmable_vs);
3215
3216 DBG("Preparing draw\n");
3217 cso_set_vertex_shader_handle(device->cso_sw,
3218 NineVertexShader9_GetVariantProcessVertices(vs, vdecl_out, so));
3219 update_vertex_elements_sw(device);
3220 update_vertex_buffers_sw(device, start_vertice, num_vertices);
3221 update_vs_constants_sw(device);
3222 DBG("Preparation succeeded\n");
3223 }
3224
3225 void
3226 nine_state_after_draw_sw(struct NineDevice9 *device)
3227 {
3228 struct nine_state_sw_internal *sw_internal = &device->state_sw_internal;
3229 struct pipe_context *pipe = nine_context_get_pipe_acquire(device);
3230 struct pipe_context *pipe_sw = device->pipe_sw;
3231 int i;
3232
3233 for (i = 0; i < 4; i++) {
3234 pipe_sw->set_vertex_buffers(pipe_sw, i, 1, NULL);
3235 if (sw_internal->transfers_so[i])
3236 pipe->transfer_unmap(pipe, sw_internal->transfers_so[i]);
3237 sw_internal->transfers_so[i] = NULL;
3238 }
3239 nine_context_get_pipe_release(device);
3240 }
3241
3242 void
3243 nine_state_destroy_sw(struct NineDevice9 *device)
3244 {
3245 (void) device;
3246 /* Everything destroyed with cso */
3247 }
3248
3249 /*
3250 static const DWORD nine_render_states_pixel[] =
3251 {
3252 D3DRS_ALPHABLENDENABLE,
3253 D3DRS_ALPHAFUNC,
3254 D3DRS_ALPHAREF,
3255 D3DRS_ALPHATESTENABLE,
3256 D3DRS_ANTIALIASEDLINEENABLE,
3257 D3DRS_BLENDFACTOR,
3258 D3DRS_BLENDOP,
3259 D3DRS_BLENDOPALPHA,
3260 D3DRS_CCW_STENCILFAIL,
3261 D3DRS_CCW_STENCILPASS,
3262 D3DRS_CCW_STENCILZFAIL,
3263 D3DRS_COLORWRITEENABLE,
3264 D3DRS_COLORWRITEENABLE1,
3265 D3DRS_COLORWRITEENABLE2,
3266 D3DRS_COLORWRITEENABLE3,
3267 D3DRS_DEPTHBIAS,
3268 D3DRS_DESTBLEND,
3269 D3DRS_DESTBLENDALPHA,
3270 D3DRS_DITHERENABLE,
3271 D3DRS_FILLMODE,
3272 D3DRS_FOGDENSITY,
3273 D3DRS_FOGEND,
3274 D3DRS_FOGSTART,
3275 D3DRS_LASTPIXEL,
3276 D3DRS_SCISSORTESTENABLE,
3277 D3DRS_SEPARATEALPHABLENDENABLE,
3278 D3DRS_SHADEMODE,
3279 D3DRS_SLOPESCALEDEPTHBIAS,
3280 D3DRS_SRCBLEND,
3281 D3DRS_SRCBLENDALPHA,
3282 D3DRS_SRGBWRITEENABLE,
3283 D3DRS_STENCILENABLE,
3284 D3DRS_STENCILFAIL,
3285 D3DRS_STENCILFUNC,
3286 D3DRS_STENCILMASK,
3287 D3DRS_STENCILPASS,
3288 D3DRS_STENCILREF,
3289 D3DRS_STENCILWRITEMASK,
3290 D3DRS_STENCILZFAIL,
3291 D3DRS_TEXTUREFACTOR,
3292 D3DRS_TWOSIDEDSTENCILMODE,
3293 D3DRS_WRAP0,
3294 D3DRS_WRAP1,
3295 D3DRS_WRAP10,
3296 D3DRS_WRAP11,
3297 D3DRS_WRAP12,
3298 D3DRS_WRAP13,
3299 D3DRS_WRAP14,
3300 D3DRS_WRAP15,
3301 D3DRS_WRAP2,
3302 D3DRS_WRAP3,
3303 D3DRS_WRAP4,
3304 D3DRS_WRAP5,
3305 D3DRS_WRAP6,
3306 D3DRS_WRAP7,
3307 D3DRS_WRAP8,
3308 D3DRS_WRAP9,
3309 D3DRS_ZENABLE,
3310 D3DRS_ZFUNC,
3311 D3DRS_ZWRITEENABLE
3312 };
3313 */
3314 const uint32_t nine_render_states_pixel[(NINED3DRS_LAST + 31) / 32] =
3315 {
3316 0x0f99c380, 0x1ff00070, 0x00000000, 0x00000000,
3317 0x000000ff, 0xde01c900, 0x0003ffcf
3318 };
3319
3320 /*
3321 static const DWORD nine_render_states_vertex[] =
3322 {
3323 D3DRS_ADAPTIVETESS_W,
3324 D3DRS_ADAPTIVETESS_X,
3325 D3DRS_ADAPTIVETESS_Y,
3326 D3DRS_ADAPTIVETESS_Z,
3327 D3DRS_AMBIENT,
3328 D3DRS_AMBIENTMATERIALSOURCE,
3329 D3DRS_CLIPPING,
3330 D3DRS_CLIPPLANEENABLE,
3331 D3DRS_COLORVERTEX,
3332 D3DRS_CULLMODE,
3333 D3DRS_DIFFUSEMATERIALSOURCE,
3334 D3DRS_EMISSIVEMATERIALSOURCE,
3335 D3DRS_ENABLEADAPTIVETESSELLATION,
3336 D3DRS_FOGCOLOR,
3337 D3DRS_FOGDENSITY,
3338 D3DRS_FOGENABLE,
3339 D3DRS_FOGEND,
3340 D3DRS_FOGSTART,
3341 D3DRS_FOGTABLEMODE,
3342 D3DRS_FOGVERTEXMODE,
3343 D3DRS_INDEXEDVERTEXBLENDENABLE,
3344 D3DRS_LIGHTING,
3345 D3DRS_LOCALVIEWER,
3346 D3DRS_MAXTESSELLATIONLEVEL,
3347 D3DRS_MINTESSELLATIONLEVEL,
3348 D3DRS_MULTISAMPLEANTIALIAS,
3349 D3DRS_MULTISAMPLEMASK,
3350 D3DRS_NORMALDEGREE,
3351 D3DRS_NORMALIZENORMALS,
3352 D3DRS_PATCHEDGESTYLE,
3353 D3DRS_POINTSCALE_A,
3354 D3DRS_POINTSCALE_B,
3355 D3DRS_POINTSCALE_C,
3356 D3DRS_POINTSCALEENABLE,
3357 D3DRS_POINTSIZE,
3358 D3DRS_POINTSIZE_MAX,
3359 D3DRS_POINTSIZE_MIN,
3360 D3DRS_POINTSPRITEENABLE,
3361 D3DRS_POSITIONDEGREE,
3362 D3DRS_RANGEFOGENABLE,
3363 D3DRS_SHADEMODE,
3364 D3DRS_SPECULARENABLE,
3365 D3DRS_SPECULARMATERIALSOURCE,
3366 D3DRS_TWEENFACTOR,
3367 D3DRS_VERTEXBLEND
3368 };
3369 */
3370 const uint32_t nine_render_states_vertex[(NINED3DRS_LAST + 31) / 32] =
3371 {
3372 0x30400200, 0x0001007c, 0x00000000, 0x00000000,
3373 0xfd9efb00, 0x01fc34cf, 0x00000000
3374 };
3375
3376 /* TODO: put in the right values */
3377 const uint32_t nine_render_state_group[NINED3DRS_LAST + 1] =
3378 {
3379 [D3DRS_ZENABLE] = NINE_STATE_DSA | NINE_STATE_MULTISAMPLE,
3380 [D3DRS_FILLMODE] = NINE_STATE_RASTERIZER,
3381 [D3DRS_SHADEMODE] = NINE_STATE_RASTERIZER,
3382 [D3DRS_ZWRITEENABLE] = NINE_STATE_DSA,
3383 [D3DRS_ALPHATESTENABLE] = NINE_STATE_DSA,
3384 [D3DRS_LASTPIXEL] = NINE_STATE_RASTERIZER,
3385 [D3DRS_SRCBLEND] = NINE_STATE_BLEND,
3386 [D3DRS_DESTBLEND] = NINE_STATE_BLEND,
3387 [D3DRS_CULLMODE] = NINE_STATE_RASTERIZER,
3388 [D3DRS_ZFUNC] = NINE_STATE_DSA,
3389 [D3DRS_ALPHAREF] = NINE_STATE_DSA,
3390 [D3DRS_ALPHAFUNC] = NINE_STATE_DSA,
3391 [D3DRS_DITHERENABLE] = NINE_STATE_BLEND,
3392 [D3DRS_ALPHABLENDENABLE] = NINE_STATE_BLEND,
3393 [D3DRS_FOGENABLE] = NINE_STATE_FF_SHADER | NINE_STATE_VS_PARAMS_MISC | NINE_STATE_PS_PARAMS_MISC | NINE_STATE_PS_CONST,
3394 [D3DRS_SPECULARENABLE] = NINE_STATE_FF_LIGHTING,
3395 [D3DRS_FOGCOLOR] = NINE_STATE_FF_PS_CONSTS | NINE_STATE_PS_CONST,
3396 [D3DRS_FOGTABLEMODE] = NINE_STATE_FF_SHADER | NINE_STATE_PS_PARAMS_MISC | NINE_STATE_PS_CONST,
3397 [D3DRS_FOGSTART] = NINE_STATE_FF_VS_OTHER | NINE_STATE_FF_PS_CONSTS | NINE_STATE_PS_CONST,
3398 [D3DRS_FOGEND] = NINE_STATE_FF_VS_OTHER | NINE_STATE_FF_PS_CONSTS | NINE_STATE_PS_CONST,
3399 [D3DRS_FOGDENSITY] = NINE_STATE_FF_VS_OTHER | NINE_STATE_FF_PS_CONSTS | NINE_STATE_PS_CONST,
3400 [D3DRS_RANGEFOGENABLE] = NINE_STATE_FF_SHADER,
3401 [D3DRS_STENCILENABLE] = NINE_STATE_DSA | NINE_STATE_MULTISAMPLE,
3402 [D3DRS_STENCILFAIL] = NINE_STATE_DSA,
3403 [D3DRS_STENCILZFAIL] = NINE_STATE_DSA,
3404 [D3DRS_STENCILPASS] = NINE_STATE_DSA,
3405 [D3DRS_STENCILFUNC] = NINE_STATE_DSA,
3406 [D3DRS_STENCILREF] = NINE_STATE_STENCIL_REF,
3407 [D3DRS_STENCILMASK] = NINE_STATE_DSA,
3408 [D3DRS_STENCILWRITEMASK] = NINE_STATE_DSA,
3409 [D3DRS_TEXTUREFACTOR] = NINE_STATE_FF_PS_CONSTS,
3410 [D3DRS_WRAP0] = NINE_STATE_UNHANDLED, /* cylindrical wrap is crazy */
3411 [D3DRS_WRAP1] = NINE_STATE_UNHANDLED,
3412 [D3DRS_WRAP2] = NINE_STATE_UNHANDLED,
3413 [D3DRS_WRAP3] = NINE_STATE_UNHANDLED,
3414 [D3DRS_WRAP4] = NINE_STATE_UNHANDLED,
3415 [D3DRS_WRAP5] = NINE_STATE_UNHANDLED,
3416 [D3DRS_WRAP6] = NINE_STATE_UNHANDLED,
3417 [D3DRS_WRAP7] = NINE_STATE_UNHANDLED,
3418 [D3DRS_CLIPPING] = 0, /* software vertex processing only */
3419 [D3DRS_LIGHTING] = NINE_STATE_FF_LIGHTING,
3420 [D3DRS_AMBIENT] = NINE_STATE_FF_LIGHTING | NINE_STATE_FF_MATERIAL,
3421 [D3DRS_FOGVERTEXMODE] = NINE_STATE_FF_SHADER,
3422 [D3DRS_COLORVERTEX] = NINE_STATE_FF_LIGHTING,
3423 [D3DRS_LOCALVIEWER] = NINE_STATE_FF_LIGHTING,
3424 [D3DRS_NORMALIZENORMALS] = NINE_STATE_FF_SHADER,
3425 [D3DRS_DIFFUSEMATERIALSOURCE] = NINE_STATE_FF_LIGHTING,
3426 [D3DRS_SPECULARMATERIALSOURCE] = NINE_STATE_FF_LIGHTING,
3427 [D3DRS_AMBIENTMATERIALSOURCE] = NINE_STATE_FF_LIGHTING,
3428 [D3DRS_EMISSIVEMATERIALSOURCE] = NINE_STATE_FF_LIGHTING,
3429 [D3DRS_VERTEXBLEND] = NINE_STATE_FF_SHADER,
3430 [D3DRS_CLIPPLANEENABLE] = NINE_STATE_RASTERIZER,
3431 [D3DRS_POINTSIZE] = NINE_STATE_RASTERIZER | NINE_STATE_FF_VS_OTHER,
3432 [D3DRS_POINTSIZE_MIN] = NINE_STATE_RASTERIZER | NINE_STATE_FF_VS_OTHER | NINE_STATE_VS_PARAMS_MISC,
3433 [D3DRS_POINTSPRITEENABLE] = NINE_STATE_RASTERIZER,
3434 [D3DRS_POINTSCALEENABLE] = NINE_STATE_FF_SHADER,
3435 [D3DRS_POINTSCALE_A] = NINE_STATE_FF_VS_OTHER,
3436 [D3DRS_POINTSCALE_B] = NINE_STATE_FF_VS_OTHER,
3437 [D3DRS_POINTSCALE_C] = NINE_STATE_FF_VS_OTHER,
3438 [D3DRS_MULTISAMPLEANTIALIAS] = NINE_STATE_MULTISAMPLE,
3439 [D3DRS_MULTISAMPLEMASK] = NINE_STATE_SAMPLE_MASK,
3440 [D3DRS_PATCHEDGESTYLE] = NINE_STATE_UNHANDLED,
3441 [D3DRS_DEBUGMONITORTOKEN] = NINE_STATE_UNHANDLED,
3442 [D3DRS_POINTSIZE_MAX] = NINE_STATE_RASTERIZER | NINE_STATE_FF_VS_OTHER | NINE_STATE_VS_PARAMS_MISC,
3443 [D3DRS_INDEXEDVERTEXBLENDENABLE] = NINE_STATE_FF_SHADER,
3444 [D3DRS_COLORWRITEENABLE] = NINE_STATE_BLEND,
3445 [D3DRS_TWEENFACTOR] = NINE_STATE_FF_VS_OTHER,
3446 [D3DRS_BLENDOP] = NINE_STATE_BLEND,
3447 [D3DRS_POSITIONDEGREE] = NINE_STATE_UNHANDLED,
3448 [D3DRS_NORMALDEGREE] = NINE_STATE_UNHANDLED,
3449 [D3DRS_SCISSORTESTENABLE] = NINE_STATE_RASTERIZER,
3450 [D3DRS_SLOPESCALEDEPTHBIAS] = NINE_STATE_RASTERIZER,
3451 [D3DRS_ANTIALIASEDLINEENABLE] = NINE_STATE_RASTERIZER,
3452 [D3DRS_MINTESSELLATIONLEVEL] = NINE_STATE_UNHANDLED,
3453 [D3DRS_MAXTESSELLATIONLEVEL] = NINE_STATE_UNHANDLED,
3454 [D3DRS_ADAPTIVETESS_X] = NINE_STATE_UNHANDLED,
3455 [D3DRS_ADAPTIVETESS_Y] = NINE_STATE_UNHANDLED,
3456 [D3DRS_ADAPTIVETESS_Z] = NINE_STATE_UNHANDLED,
3457 [D3DRS_ADAPTIVETESS_W] = NINE_STATE_UNHANDLED,
3458 [D3DRS_ENABLEADAPTIVETESSELLATION] = NINE_STATE_UNHANDLED,
3459 [D3DRS_TWOSIDEDSTENCILMODE] = NINE_STATE_DSA,
3460 [D3DRS_CCW_STENCILFAIL] = NINE_STATE_DSA,
3461 [D3DRS_CCW_STENCILZFAIL] = NINE_STATE_DSA,
3462 [D3DRS_CCW_STENCILPASS] = NINE_STATE_DSA,
3463 [D3DRS_CCW_STENCILFUNC] = NINE_STATE_DSA,
3464 [D3DRS_COLORWRITEENABLE1] = NINE_STATE_BLEND,
3465 [D3DRS_COLORWRITEENABLE2] = NINE_STATE_BLEND,
3466 [D3DRS_COLORWRITEENABLE3] = NINE_STATE_BLEND,
3467 [D3DRS_BLENDFACTOR] = NINE_STATE_BLEND_COLOR,
3468 [D3DRS_SRGBWRITEENABLE] = NINE_STATE_FB,
3469 [D3DRS_DEPTHBIAS] = NINE_STATE_RASTERIZER,
3470 [D3DRS_WRAP8] = NINE_STATE_UNHANDLED, /* cylwrap has to be done via GP */
3471 [D3DRS_WRAP9] = NINE_STATE_UNHANDLED,
3472 [D3DRS_WRAP10] = NINE_STATE_UNHANDLED,
3473 [D3DRS_WRAP11] = NINE_STATE_UNHANDLED,
3474 [D3DRS_WRAP12] = NINE_STATE_UNHANDLED,
3475 [D3DRS_WRAP13] = NINE_STATE_UNHANDLED,
3476 [D3DRS_WRAP14] = NINE_STATE_UNHANDLED,
3477 [D3DRS_WRAP15] = NINE_STATE_UNHANDLED,
3478 [D3DRS_SEPARATEALPHABLENDENABLE] = NINE_STATE_BLEND,
3479 [D3DRS_SRCBLENDALPHA] = NINE_STATE_BLEND,
3480 [D3DRS_DESTBLENDALPHA] = NINE_STATE_BLEND,
3481 [D3DRS_BLENDOPALPHA] = NINE_STATE_BLEND
3482 };
3483
3484 /* Misc */
3485
3486 static D3DMATRIX nine_state_identity = { .m[0] = { 1, 0, 0, 0 },
3487 .m[1] = { 0, 1, 0, 0 },
3488 .m[2] = { 0, 0, 1, 0 },
3489 .m[3] = { 0, 0, 0, 1 } };
3490
3491 void
3492 nine_state_resize_transform(struct nine_ff_state *ff_state, unsigned N)
3493 {
3494 unsigned n = ff_state->num_transforms;
3495
3496 if (N <= n)
3497 return;
3498
3499 ff_state->transform = REALLOC(ff_state->transform,
3500 n * sizeof(D3DMATRIX),
3501 N * sizeof(D3DMATRIX));
3502 for (; n < N; ++n)
3503 ff_state->transform[n] = nine_state_identity;
3504 ff_state->num_transforms = N;
3505 }
3506
3507 D3DMATRIX *
3508 nine_state_access_transform(struct nine_ff_state *ff_state, D3DTRANSFORMSTATETYPE t,
3509 boolean alloc)
3510 {
3511 unsigned index;
3512
3513 switch (t) {
3514 case D3DTS_VIEW: index = 0; break;
3515 case D3DTS_PROJECTION: index = 1; break;
3516 case D3DTS_TEXTURE0: index = 2; break;
3517 case D3DTS_TEXTURE1: index = 3; break;
3518 case D3DTS_TEXTURE2: index = 4; break;
3519 case D3DTS_TEXTURE3: index = 5; break;
3520 case D3DTS_TEXTURE4: index = 6; break;
3521 case D3DTS_TEXTURE5: index = 7; break;
3522 case D3DTS_TEXTURE6: index = 8; break;
3523 case D3DTS_TEXTURE7: index = 9; break;
3524 default:
3525 if (!(t >= D3DTS_WORLDMATRIX(0) && t <= D3DTS_WORLDMATRIX(255)))
3526 return NULL;
3527 index = 10 + (t - D3DTS_WORLDMATRIX(0));
3528 break;
3529 }
3530
3531 if (index >= ff_state->num_transforms) {
3532 if (!alloc)
3533 return &nine_state_identity;
3534 nine_state_resize_transform(ff_state, index + 1);
3535 }
3536 return &ff_state->transform[index];
3537 }
3538
3539 HRESULT
3540 nine_state_set_light(struct nine_ff_state *ff_state, DWORD Index,
3541 const D3DLIGHT9 *pLight)
3542 {
3543 if (Index >= ff_state->num_lights) {
3544 unsigned n = ff_state->num_lights;
3545 unsigned N = Index + 1;
3546
3547 ff_state->light = REALLOC(ff_state->light, n * sizeof(D3DLIGHT9),
3548 N * sizeof(D3DLIGHT9));
3549 if (!ff_state->light)
3550 return E_OUTOFMEMORY;
3551 ff_state->num_lights = N;
3552
3553 for (; n < Index; ++n) {
3554 memset(&ff_state->light[n], 0, sizeof(D3DLIGHT9));
3555 ff_state->light[n].Type = (D3DLIGHTTYPE)NINED3DLIGHT_INVALID;
3556 }
3557 }
3558 ff_state->light[Index] = *pLight;
3559
3560 if (pLight->Type == D3DLIGHT_SPOT && pLight->Theta >= pLight->Phi) {
3561 DBG("Warning: clamping D3DLIGHT9.Theta\n");
3562 ff_state->light[Index].Theta = ff_state->light[Index].Phi;
3563 }
3564 return D3D_OK;
3565 }
3566
3567 HRESULT
3568 nine_state_light_enable(struct nine_ff_state *ff_state,
3569 DWORD Index, BOOL Enable)
3570 {
3571 unsigned i;
3572
3573 user_assert(Index < ff_state->num_lights, D3DERR_INVALIDCALL);
3574
3575 for (i = 0; i < ff_state->num_lights_active; ++i) {
3576 if (ff_state->active_light[i] == Index)
3577 break;
3578 }
3579
3580 if (Enable) {
3581 if (i < ff_state->num_lights_active)
3582 return D3D_OK;
3583 /* XXX wine thinks this should still succeed:
3584 */
3585 user_assert(i < NINE_MAX_LIGHTS_ACTIVE, D3DERR_INVALIDCALL);
3586
3587 ff_state->active_light[i] = Index;
3588 ff_state->num_lights_active++;
3589 } else {
3590 if (i == ff_state->num_lights_active)
3591 return D3D_OK;
3592 --ff_state->num_lights_active;
3593 for (; i < ff_state->num_lights_active; ++i)
3594 ff_state->active_light[i] = ff_state->active_light[i + 1];
3595 }
3596
3597 return D3D_OK;
3598 }
3599
3600 #define D3DRS_TO_STRING_CASE(n) case D3DRS_##n: return "D3DRS_"#n
3601 const char *nine_d3drs_to_string(DWORD State)
3602 {
3603 switch (State) {
3604 D3DRS_TO_STRING_CASE(ZENABLE);
3605 D3DRS_TO_STRING_CASE(FILLMODE);
3606 D3DRS_TO_STRING_CASE(SHADEMODE);
3607 D3DRS_TO_STRING_CASE(ZWRITEENABLE);
3608 D3DRS_TO_STRING_CASE(ALPHATESTENABLE);
3609 D3DRS_TO_STRING_CASE(LASTPIXEL);
3610 D3DRS_TO_STRING_CASE(SRCBLEND);
3611 D3DRS_TO_STRING_CASE(DESTBLEND);
3612 D3DRS_TO_STRING_CASE(CULLMODE);
3613 D3DRS_TO_STRING_CASE(ZFUNC);
3614 D3DRS_TO_STRING_CASE(ALPHAREF);
3615 D3DRS_TO_STRING_CASE(ALPHAFUNC);
3616 D3DRS_TO_STRING_CASE(DITHERENABLE);
3617 D3DRS_TO_STRING_CASE(ALPHABLENDENABLE);
3618 D3DRS_TO_STRING_CASE(FOGENABLE);
3619 D3DRS_TO_STRING_CASE(SPECULARENABLE);
3620 D3DRS_TO_STRING_CASE(FOGCOLOR);
3621 D3DRS_TO_STRING_CASE(FOGTABLEMODE);
3622 D3DRS_TO_STRING_CASE(FOGSTART);
3623 D3DRS_TO_STRING_CASE(FOGEND);
3624 D3DRS_TO_STRING_CASE(FOGDENSITY);
3625 D3DRS_TO_STRING_CASE(RANGEFOGENABLE);
3626 D3DRS_TO_STRING_CASE(STENCILENABLE);
3627 D3DRS_TO_STRING_CASE(STENCILFAIL);
3628 D3DRS_TO_STRING_CASE(STENCILZFAIL);
3629 D3DRS_TO_STRING_CASE(STENCILPASS);
3630 D3DRS_TO_STRING_CASE(STENCILFUNC);
3631 D3DRS_TO_STRING_CASE(STENCILREF);
3632 D3DRS_TO_STRING_CASE(STENCILMASK);
3633 D3DRS_TO_STRING_CASE(STENCILWRITEMASK);
3634 D3DRS_TO_STRING_CASE(TEXTUREFACTOR);
3635 D3DRS_TO_STRING_CASE(WRAP0);
3636 D3DRS_TO_STRING_CASE(WRAP1);
3637 D3DRS_TO_STRING_CASE(WRAP2);
3638 D3DRS_TO_STRING_CASE(WRAP3);
3639 D3DRS_TO_STRING_CASE(WRAP4);
3640 D3DRS_TO_STRING_CASE(WRAP5);
3641 D3DRS_TO_STRING_CASE(WRAP6);
3642 D3DRS_TO_STRING_CASE(WRAP7);
3643 D3DRS_TO_STRING_CASE(CLIPPING);
3644 D3DRS_TO_STRING_CASE(LIGHTING);
3645 D3DRS_TO_STRING_CASE(AMBIENT);
3646 D3DRS_TO_STRING_CASE(FOGVERTEXMODE);
3647 D3DRS_TO_STRING_CASE(COLORVERTEX);
3648 D3DRS_TO_STRING_CASE(LOCALVIEWER);
3649 D3DRS_TO_STRING_CASE(NORMALIZENORMALS);
3650 D3DRS_TO_STRING_CASE(DIFFUSEMATERIALSOURCE);
3651 D3DRS_TO_STRING_CASE(SPECULARMATERIALSOURCE);
3652 D3DRS_TO_STRING_CASE(AMBIENTMATERIALSOURCE);
3653 D3DRS_TO_STRING_CASE(EMISSIVEMATERIALSOURCE);
3654 D3DRS_TO_STRING_CASE(VERTEXBLEND);
3655 D3DRS_TO_STRING_CASE(CLIPPLANEENABLE);
3656 D3DRS_TO_STRING_CASE(POINTSIZE);
3657 D3DRS_TO_STRING_CASE(POINTSIZE_MIN);
3658 D3DRS_TO_STRING_CASE(POINTSPRITEENABLE);
3659 D3DRS_TO_STRING_CASE(POINTSCALEENABLE);
3660 D3DRS_TO_STRING_CASE(POINTSCALE_A);
3661 D3DRS_TO_STRING_CASE(POINTSCALE_B);
3662 D3DRS_TO_STRING_CASE(POINTSCALE_C);
3663 D3DRS_TO_STRING_CASE(MULTISAMPLEANTIALIAS);
3664 D3DRS_TO_STRING_CASE(MULTISAMPLEMASK);
3665 D3DRS_TO_STRING_CASE(PATCHEDGESTYLE);
3666 D3DRS_TO_STRING_CASE(DEBUGMONITORTOKEN);
3667 D3DRS_TO_STRING_CASE(POINTSIZE_MAX);
3668 D3DRS_TO_STRING_CASE(INDEXEDVERTEXBLENDENABLE);
3669 D3DRS_TO_STRING_CASE(COLORWRITEENABLE);
3670 D3DRS_TO_STRING_CASE(TWEENFACTOR);
3671 D3DRS_TO_STRING_CASE(BLENDOP);
3672 D3DRS_TO_STRING_CASE(POSITIONDEGREE);
3673 D3DRS_TO_STRING_CASE(NORMALDEGREE);
3674 D3DRS_TO_STRING_CASE(SCISSORTESTENABLE);
3675 D3DRS_TO_STRING_CASE(SLOPESCALEDEPTHBIAS);
3676 D3DRS_TO_STRING_CASE(ANTIALIASEDLINEENABLE);
3677 D3DRS_TO_STRING_CASE(MINTESSELLATIONLEVEL);
3678 D3DRS_TO_STRING_CASE(MAXTESSELLATIONLEVEL);
3679 D3DRS_TO_STRING_CASE(ADAPTIVETESS_X);
3680 D3DRS_TO_STRING_CASE(ADAPTIVETESS_Y);
3681 D3DRS_TO_STRING_CASE(ADAPTIVETESS_Z);
3682 D3DRS_TO_STRING_CASE(ADAPTIVETESS_W);
3683 D3DRS_TO_STRING_CASE(ENABLEADAPTIVETESSELLATION);
3684 D3DRS_TO_STRING_CASE(TWOSIDEDSTENCILMODE);
3685 D3DRS_TO_STRING_CASE(CCW_STENCILFAIL);
3686 D3DRS_TO_STRING_CASE(CCW_STENCILZFAIL);
3687 D3DRS_TO_STRING_CASE(CCW_STENCILPASS);
3688 D3DRS_TO_STRING_CASE(CCW_STENCILFUNC);
3689 D3DRS_TO_STRING_CASE(COLORWRITEENABLE1);
3690 D3DRS_TO_STRING_CASE(COLORWRITEENABLE2);
3691 D3DRS_TO_STRING_CASE(COLORWRITEENABLE3);
3692 D3DRS_TO_STRING_CASE(BLENDFACTOR);
3693 D3DRS_TO_STRING_CASE(SRGBWRITEENABLE);
3694 D3DRS_TO_STRING_CASE(DEPTHBIAS);
3695 D3DRS_TO_STRING_CASE(WRAP8);
3696 D3DRS_TO_STRING_CASE(WRAP9);
3697 D3DRS_TO_STRING_CASE(WRAP10);
3698 D3DRS_TO_STRING_CASE(WRAP11);
3699 D3DRS_TO_STRING_CASE(WRAP12);
3700 D3DRS_TO_STRING_CASE(WRAP13);
3701 D3DRS_TO_STRING_CASE(WRAP14);
3702 D3DRS_TO_STRING_CASE(WRAP15);
3703 D3DRS_TO_STRING_CASE(SEPARATEALPHABLENDENABLE);
3704 D3DRS_TO_STRING_CASE(SRCBLENDALPHA);
3705 D3DRS_TO_STRING_CASE(DESTBLENDALPHA);
3706 D3DRS_TO_STRING_CASE(BLENDOPALPHA);
3707 default:
3708 return "(invalid)";
3709 }
3710 }