1 /*
2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
3  Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
4  develop this 3D driver.
5 
6  Permission is hereby granted, free of charge, to any person obtaining
7  a copy of this software and associated documentation files (the
8  "Software"), to deal in the Software without restriction, including
9  without limitation the rights to use, copy, modify, merge, publish,
10  distribute, sublicense, and/or sell copies of the Software, and to
11  permit persons to whom the Software is furnished to do so, subject to
12  the following conditions:
13 
14  The above copyright notice and this permission notice (including the
15  next paragraph) shall be included in all copies or substantial
16  portions of the Software.
17 
18  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21  IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 
26  **********************************************************************/
27  /*
28   * Authors:
29   *   Keith Whitwell <keith@tungstengraphics.com>
30   */
31 
32 
33 
34 #include "intel_batchbuffer.h"
35 #include "intel_fbo.h"
36 #include "intel_mipmap_tree.h"
37 #include "intel_regions.h"
38 
39 #include "brw_context.h"
40 #include "brw_state.h"
41 #include "brw_defines.h"
42 
43 #include "main/fbobject.h"
44 
45 /* Constant single cliprect for framebuffer object or DRI2 drawing */
upload_drawing_rect(struct brw_context * brw)46 static void upload_drawing_rect(struct brw_context *brw)
47 {
48    struct intel_context *intel = &brw->intel;
49    struct gl_context *ctx = &intel->ctx;
50 
51    BEGIN_BATCH(4);
52    OUT_BATCH(_3DSTATE_DRAWING_RECTANGLE << 16 | (4 - 2));
53    OUT_BATCH(0); /* xmin, ymin */
54    OUT_BATCH(((ctx->DrawBuffer->Width - 1) & 0xffff) |
55 	    ((ctx->DrawBuffer->Height - 1) << 16));
56    OUT_BATCH(0);
57    ADVANCE_BATCH();
58 }
59 
60 const struct brw_tracked_state brw_drawing_rect = {
61    .dirty = {
62       .mesa = _NEW_BUFFERS,
63       .brw = BRW_NEW_CONTEXT,
64       .cache = 0
65    },
66    .emit = upload_drawing_rect
67 };
68 
69 /**
70  * Upload the binding table pointers, which point each stage's array of surface
71  * state pointers.
72  *
73  * The binding table pointers are relative to the surface state base address,
74  * which points at the batchbuffer containing the streamed batch state.
75  */
upload_binding_table_pointers(struct brw_context * brw)76 static void upload_binding_table_pointers(struct brw_context *brw)
77 {
78    struct intel_context *intel = &brw->intel;
79 
80    BEGIN_BATCH(6);
81    OUT_BATCH(_3DSTATE_BINDING_TABLE_POINTERS << 16 | (6 - 2));
82    OUT_BATCH(brw->vs.bind_bo_offset);
83    OUT_BATCH(0); /* gs */
84    OUT_BATCH(0); /* clip */
85    OUT_BATCH(0); /* sf */
86    OUT_BATCH(brw->wm.bind_bo_offset);
87    ADVANCE_BATCH();
88 }
89 
90 const struct brw_tracked_state brw_binding_table_pointers = {
91    .dirty = {
92       .mesa = 0,
93       .brw = (BRW_NEW_BATCH |
94 	      BRW_NEW_STATE_BASE_ADDRESS |
95 	      BRW_NEW_VS_BINDING_TABLE |
96 	      BRW_NEW_GS_BINDING_TABLE |
97 	      BRW_NEW_PS_BINDING_TABLE),
98       .cache = 0,
99    },
100    .emit = upload_binding_table_pointers,
101 };
102 
103 /**
104  * Upload the binding table pointers, which point each stage's array of surface
105  * state pointers.
106  *
107  * The binding table pointers are relative to the surface state base address,
108  * which points at the batchbuffer containing the streamed batch state.
109  */
upload_gen6_binding_table_pointers(struct brw_context * brw)110 static void upload_gen6_binding_table_pointers(struct brw_context *brw)
111 {
112    struct intel_context *intel = &brw->intel;
113 
114    BEGIN_BATCH(4);
115    OUT_BATCH(_3DSTATE_BINDING_TABLE_POINTERS << 16 |
116 	     GEN6_BINDING_TABLE_MODIFY_VS |
117 	     GEN6_BINDING_TABLE_MODIFY_GS |
118 	     GEN6_BINDING_TABLE_MODIFY_PS |
119 	     (4 - 2));
120    OUT_BATCH(brw->vs.bind_bo_offset); /* vs */
121    OUT_BATCH(brw->gs.bind_bo_offset); /* gs */
122    OUT_BATCH(brw->wm.bind_bo_offset); /* wm/ps */
123    ADVANCE_BATCH();
124 }
125 
126 const struct brw_tracked_state gen6_binding_table_pointers = {
127    .dirty = {
128       .mesa = 0,
129       .brw = (BRW_NEW_BATCH |
130 	      BRW_NEW_STATE_BASE_ADDRESS |
131 	      BRW_NEW_VS_BINDING_TABLE |
132 	      BRW_NEW_GS_BINDING_TABLE |
133 	      BRW_NEW_PS_BINDING_TABLE),
134       .cache = 0,
135    },
136    .emit = upload_gen6_binding_table_pointers,
137 };
138 
139 /**
140  * Upload pointers to the per-stage state.
141  *
142  * The state pointers in this packet are all relative to the general state
143  * base address set by CMD_STATE_BASE_ADDRESS, which is 0.
144  */
upload_pipelined_state_pointers(struct brw_context * brw)145 static void upload_pipelined_state_pointers(struct brw_context *brw )
146 {
147    struct intel_context *intel = &brw->intel;
148 
149    if (intel->gen == 5) {
150       /* Need to flush before changing clip max threads for errata. */
151       BEGIN_BATCH(1);
152       OUT_BATCH(MI_FLUSH);
153       ADVANCE_BATCH();
154    }
155 
156    BEGIN_BATCH(7);
157    OUT_BATCH(_3DSTATE_PIPELINED_POINTERS << 16 | (7 - 2));
158    OUT_RELOC(intel->batch.bo, I915_GEM_DOMAIN_INSTRUCTION, 0,
159 	     brw->vs.state_offset);
160    if (brw->gs.prog_active)
161       OUT_RELOC(brw->intel.batch.bo, I915_GEM_DOMAIN_INSTRUCTION, 0,
162 		brw->gs.state_offset | 1);
163    else
164       OUT_BATCH(0);
165    OUT_RELOC(brw->intel.batch.bo, I915_GEM_DOMAIN_INSTRUCTION, 0,
166 	     brw->clip.state_offset | 1);
167    OUT_RELOC(brw->intel.batch.bo, I915_GEM_DOMAIN_INSTRUCTION, 0,
168 	     brw->sf.state_offset);
169    OUT_RELOC(brw->intel.batch.bo, I915_GEM_DOMAIN_INSTRUCTION, 0,
170 	     brw->wm.state_offset);
171    OUT_RELOC(brw->intel.batch.bo, I915_GEM_DOMAIN_INSTRUCTION, 0,
172 	     brw->cc.state_offset);
173    ADVANCE_BATCH();
174 
175    brw->state.dirty.brw |= BRW_NEW_PSP;
176 }
177 
upload_psp_urb_cbs(struct brw_context * brw)178 static void upload_psp_urb_cbs(struct brw_context *brw )
179 {
180    upload_pipelined_state_pointers(brw);
181    brw_upload_urb_fence(brw);
182    brw_upload_cs_urb_state(brw);
183 }
184 
185 const struct brw_tracked_state brw_psp_urb_cbs = {
186    .dirty = {
187       .mesa = 0,
188       .brw = (BRW_NEW_URB_FENCE |
189 	      BRW_NEW_BATCH |
190 	      BRW_NEW_STATE_BASE_ADDRESS),
191       .cache = (CACHE_NEW_VS_UNIT |
192 		CACHE_NEW_GS_UNIT |
193 		CACHE_NEW_GS_PROG |
194 		CACHE_NEW_CLIP_UNIT |
195 		CACHE_NEW_SF_UNIT |
196 		CACHE_NEW_WM_UNIT |
197 		CACHE_NEW_CC_UNIT)
198    },
199    .emit = upload_psp_urb_cbs,
200 };
201 
202 uint32_t
brw_depthbuffer_format(struct brw_context * brw)203 brw_depthbuffer_format(struct brw_context *brw)
204 {
205    struct intel_context *intel = &brw->intel;
206    struct gl_context *ctx = &intel->ctx;
207    struct gl_framebuffer *fb = ctx->DrawBuffer;
208    struct intel_renderbuffer *drb = intel_get_renderbuffer(fb, BUFFER_DEPTH);
209    struct intel_renderbuffer *srb;
210 
211    if (!drb &&
212        (srb = intel_get_renderbuffer(fb, BUFFER_STENCIL)) &&
213        !srb->mt->stencil_mt &&
214        (intel_rb_format(srb) == MESA_FORMAT_S8_Z24 ||
215 	intel_rb_format(srb) == MESA_FORMAT_Z32_FLOAT_X24S8)) {
216       drb = srb;
217    }
218 
219    if (!drb)
220       return BRW_DEPTHFORMAT_D32_FLOAT;
221 
222    switch (drb->mt->format) {
223    case MESA_FORMAT_Z16:
224       return BRW_DEPTHFORMAT_D16_UNORM;
225    case MESA_FORMAT_Z32_FLOAT:
226       return BRW_DEPTHFORMAT_D32_FLOAT;
227    case MESA_FORMAT_X8_Z24:
228       if (intel->gen >= 6) {
229 	 return BRW_DEPTHFORMAT_D24_UNORM_X8_UINT;
230       } else {
231 	 /* Use D24_UNORM_S8, not D24_UNORM_X8.
232 	  *
233 	  * D24_UNORM_X8 was not introduced until Gen5. (See the Ironlake PRM,
234 	  * Volume 2, Part 1, Section 8.4.6 "Depth/Stencil Buffer State", Bits
235 	  * 3DSTATE_DEPTH_BUFFER.Surface_Format).
236 	  *
237 	  * However, on Gen5, D24_UNORM_X8 may be used only if separate
238 	  * stencil is enabled, and we never enable it. From the Ironlake PRM,
239 	  * same section as above, Bit 3DSTATE_DEPTH_BUFFER.Separate_Stencil_Buffer_Enable:
240 	  *     If this field is disabled, the Surface Format of the depth
241 	  *     buffer cannot be D24_UNORM_X8_UINT.
242 	  */
243 	 return BRW_DEPTHFORMAT_D24_UNORM_S8_UINT;
244       }
245    case MESA_FORMAT_S8_Z24:
246       return BRW_DEPTHFORMAT_D24_UNORM_S8_UINT;
247    case MESA_FORMAT_Z32_FLOAT_X24S8:
248       return BRW_DEPTHFORMAT_D32_FLOAT_S8X24_UINT;
249    default:
250       _mesa_problem(ctx, "Unexpected depth format %s\n",
251 		    _mesa_get_format_name(intel_rb_format(drb)));
252       return BRW_DEPTHFORMAT_D16_UNORM;
253    }
254 }
255 
emit_depthbuffer(struct brw_context * brw)256 static void emit_depthbuffer(struct brw_context *brw)
257 {
258    struct intel_context *intel = &brw->intel;
259    struct gl_context *ctx = &intel->ctx;
260    struct gl_framebuffer *fb = ctx->DrawBuffer;
261    /* _NEW_BUFFERS */
262    struct intel_renderbuffer *depth_irb = intel_get_renderbuffer(fb, BUFFER_DEPTH);
263    struct intel_renderbuffer *stencil_irb = intel_get_renderbuffer(fb, BUFFER_STENCIL);
264    struct intel_mipmap_tree *stencil_mt = NULL;
265    struct intel_region *hiz_region = NULL;
266    unsigned int len;
267    bool separate_stencil = false;
268 
269    /* Amount by which drawing should be offset in order to draw to the
270     * appropriate miplevel/zoffset/cubeface.  We will extract these values
271     * from depth_irb or stencil_irb once we determine which is present.
272     */
273    uint32_t draw_x = 0, draw_y = 0;
274 
275    /* Masks used to determine how much of the draw_x and draw_y offsets should
276     * be performed using the fine adjustment of "depth coordinate offset X/Y"
277     * (dw5 of 3DSTATE_DEPTH_BUFFER).  Any remaining coarse adjustment will be
278     * performed by changing the base addresses of the buffers.
279     *
280     * Since the HiZ, depth, and stencil buffers all use the same "depth
281     * coordinate offset X/Y" values, we need to make sure that the coarse
282     * adjustment will be possible to apply to all three buffers.  Since coarse
283     * adjustment can only be applied in multiples of the tile size, we will OR
284     * together the tile masks of all the buffers to determine which offsets to
285     * perform as fine adjustments.
286     */
287    uint32_t tile_mask_x = 0, tile_mask_y = 0;
288 
289    if (depth_irb) {
290       intel_region_get_tile_masks(depth_irb->mt->region,
291                                   &tile_mask_x, &tile_mask_y, false);
292    }
293 
294    if (depth_irb &&
295        depth_irb->mt &&
296        depth_irb->mt->hiz_mt) {
297       hiz_region = depth_irb->mt->hiz_mt->region;
298 
299       uint32_t hiz_tile_mask_x, hiz_tile_mask_y;
300       intel_region_get_tile_masks(hiz_region,
301                                   &hiz_tile_mask_x, &hiz_tile_mask_y, false);
302 
303       /* Each HiZ row represents 2 rows of pixels */
304       hiz_tile_mask_y = hiz_tile_mask_y << 1 | 1;
305 
306       tile_mask_x |= hiz_tile_mask_x;
307       tile_mask_y |= hiz_tile_mask_y;
308    }
309 
310    /* 3DSTATE_DEPTH_BUFFER, 3DSTATE_STENCIL_BUFFER are both
311     * non-pipelined state that will need the PIPE_CONTROL workaround.
312     */
313    if (intel->gen == 6) {
314       intel_emit_post_sync_nonzero_flush(intel);
315       intel_emit_depth_stall_flushes(intel);
316    }
317 
318    /* Find the real separate stencil mt if present. */
319    if (stencil_irb) {
320       stencil_mt = stencil_irb->mt;
321       if (stencil_mt->stencil_mt)
322 	 stencil_mt = stencil_mt->stencil_mt;
323 
324       if (stencil_mt->format == MESA_FORMAT_S8) {
325 	 separate_stencil = true;
326 
327          /* Separate stencil buffer uses 64x64 tiles. */
328          tile_mask_x |= 63;
329          tile_mask_y |= 63;
330       } else {
331          uint32_t stencil_tile_mask_x, stencil_tile_mask_y;
332          intel_region_get_tile_masks(stencil_mt->region,
333                                      &stencil_tile_mask_x,
334                                      &stencil_tile_mask_y, false);
335 
336          tile_mask_x |= stencil_tile_mask_x;
337          tile_mask_y |= stencil_tile_mask_y;
338       }
339    }
340 
341    /* If there's a packed depth/stencil bound to stencil only, we need to
342     * emit the packed depth/stencil buffer packet.
343     */
344    if (!depth_irb && stencil_irb && !separate_stencil)
345       depth_irb = stencil_irb;
346 
347    if (intel->gen >= 6)
348       len = 7;
349    else if (intel->is_g4x || intel->gen == 5)
350       len = 6;
351    else
352       len = 5;
353 
354    if (!depth_irb && !separate_stencil) {
355       BEGIN_BATCH(len);
356       OUT_BATCH(_3DSTATE_DEPTH_BUFFER << 16 | (len - 2));
357       OUT_BATCH((BRW_DEPTHFORMAT_D32_FLOAT << 18) |
358 		(BRW_SURFACE_NULL << 29));
359       OUT_BATCH(0);
360       OUT_BATCH(0);
361       OUT_BATCH(0);
362 
363       if (intel->is_g4x || intel->gen >= 5)
364          OUT_BATCH(0);
365 
366       if (intel->gen >= 6)
367 	 OUT_BATCH(0);
368 
369       ADVANCE_BATCH();
370 
371    } else if (!depth_irb && separate_stencil) {
372       uint32_t tile_x, tile_y;
373 
374       /*
375        * There exists a separate stencil buffer but no depth buffer.
376        *
377        * The stencil buffer inherits most of its fields from
378        * 3DSTATE_DEPTH_BUFFER: namely the tile walk, surface type, width, and
379        * height.
380        *
381        * Enable the hiz bit because it and the separate stencil bit must have
382        * the same value. From Section 2.11.5.6.1.1 3DSTATE_DEPTH_BUFFER, Bit
383        * 1.21 "Separate Stencil Enable":
384        *     [DevIL]: If this field is enabled, Hierarchical Depth Buffer
385        *     Enable must also be enabled.
386        *
387        *     [DevGT]: This field must be set to the same value (enabled or
388        *     disabled) as Hierarchical Depth Buffer Enable
389        *
390        * The tiled bit must be set. From the Sandybridge PRM, Volume 2, Part 1,
391        * Section 7.5.5.1.1 3DSTATE_DEPTH_BUFFER, Bit 1.27 Tiled Surface:
392        *     [DevGT+]: This field must be set to TRUE.
393        */
394       assert(intel->has_separate_stencil);
395 
396       draw_x = stencil_irb->draw_x;
397       draw_y = stencil_irb->draw_y;
398       tile_x = draw_x & tile_mask_x;
399       tile_y = draw_y & tile_mask_y;
400 
401       /* According to the Sandy Bridge PRM, volume 2 part 1, pp326-327
402        * (3DSTATE_DEPTH_BUFFER dw5), in the documentation for "Depth
403        * Coordinate Offset X/Y":
404        *
405        *   "The 3 LSBs of both offsets must be zero to ensure correct
406        *   alignment"
407        *
408        * We have no guarantee that tile_x and tile_y are correctly aligned,
409        * since they are determined by the mipmap layout, which is only aligned
410        * to multiples of 4.
411        *
412        * So, to avoid hanging the GPU, just smash the low order 3 bits of
413        * tile_x and tile_y to 0.  This is a temporary workaround until we come
414        * up with a better solution.
415        */
416       if (intel->gen >= 6) {
417          tile_x &= ~7;
418          tile_y &= ~7;
419       }
420 
421       BEGIN_BATCH(len);
422       OUT_BATCH(_3DSTATE_DEPTH_BUFFER << 16 | (len - 2));
423       OUT_BATCH((BRW_DEPTHFORMAT_D32_FLOAT << 18) |
424 	        (1 << 21) | /* separate stencil enable */
425 	        (1 << 22) | /* hiz enable */
426 	        (BRW_TILEWALK_YMAJOR << 26) |
427 	        (1 << 27) | /* tiled surface */
428 	        (BRW_SURFACE_2D << 29));
429       OUT_BATCH(0);
430       OUT_BATCH(((stencil_irb->Base.Base.Width + tile_x - 1) << 6) |
431 	         (stencil_irb->Base.Base.Height + tile_y - 1) << 19);
432       OUT_BATCH(0);
433 
434       if (intel->is_g4x || intel->gen >= 5)
435          OUT_BATCH(tile_x | (tile_y << 16));
436       else
437 	 assert(tile_x == 0 && tile_y == 0);
438 
439       if (intel->gen >= 6)
440 	 OUT_BATCH(0);
441 
442       ADVANCE_BATCH();
443 
444    } else {
445       struct intel_region *region = depth_irb->mt->region;
446       uint32_t tile_x, tile_y, offset;
447 
448       /* If using separate stencil, hiz must be enabled. */
449       assert(!separate_stencil || hiz_region);
450 
451       assert(intel->gen < 6 || region->tiling == I915_TILING_Y);
452       assert(!hiz_region || region->tiling == I915_TILING_Y);
453 
454       draw_x = depth_irb->draw_x;
455       draw_y = depth_irb->draw_y;
456       tile_x = draw_x & tile_mask_x;
457       tile_y = draw_y & tile_mask_y;
458 
459       /* According to the Sandy Bridge PRM, volume 2 part 1, pp326-327
460        * (3DSTATE_DEPTH_BUFFER dw5), in the documentation for "Depth
461        * Coordinate Offset X/Y":
462        *
463        *   "The 3 LSBs of both offsets must be zero to ensure correct
464        *   alignment"
465        *
466        * We have no guarantee that tile_x and tile_y are correctly aligned,
467        * since they are determined by the mipmap layout, which is only aligned
468        * to multiples of 4.
469        *
470        * So, to avoid hanging the GPU, just smash the low order 3 bits of
471        * tile_x and tile_y to 0.  This is a temporary workaround until we come
472        * up with a better solution.
473        */
474       tile_x &= ~7;
475       tile_y &= ~7;
476 
477       offset = intel_region_get_aligned_offset(region,
478                                                draw_x & ~tile_mask_x,
479                                                draw_y & ~tile_mask_y, false);
480 
481       BEGIN_BATCH(len);
482       OUT_BATCH(_3DSTATE_DEPTH_BUFFER << 16 | (len - 2));
483       OUT_BATCH(((region->pitch * region->cpp) - 1) |
484 		(brw_depthbuffer_format(brw) << 18) |
485 		((hiz_region ? 1 : 0) << 21) | /* separate stencil enable */
486 		((hiz_region ? 1 : 0) << 22) | /* hiz enable */
487 		(BRW_TILEWALK_YMAJOR << 26) |
488 		((region->tiling != I915_TILING_NONE) << 27) |
489 		(BRW_SURFACE_2D << 29));
490       OUT_RELOC(region->bo,
491 		I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
492 		offset);
493       OUT_BATCH((BRW_SURFACE_MIPMAPLAYOUT_BELOW << 1) |
494 		(((depth_irb->Base.Base.Width + tile_x) - 1) << 6) |
495 		(((depth_irb->Base.Base.Height + tile_y) - 1) << 19));
496       OUT_BATCH(0);
497 
498       if (intel->is_g4x || intel->gen >= 5)
499          OUT_BATCH(tile_x | (tile_y << 16));
500       else
501 	 assert(tile_x == 0 && tile_y == 0);
502 
503       if (intel->gen >= 6)
504 	 OUT_BATCH(0);
505 
506       ADVANCE_BATCH();
507    }
508 
509    if (hiz_region || separate_stencil) {
510       /*
511        * In the 3DSTATE_DEPTH_BUFFER batch emitted above, the 'separate
512        * stencil enable' and 'hiz enable' bits were set. Therefore we must
513        * emit 3DSTATE_HIER_DEPTH_BUFFER and 3DSTATE_STENCIL_BUFFER. Even if
514        * there is no stencil buffer, 3DSTATE_STENCIL_BUFFER must be emitted;
515        * failure to do so causes hangs on gen5 and a stall on gen6.
516        */
517 
518       /* Emit hiz buffer. */
519       if (hiz_region) {
520          uint32_t hiz_offset =
521             intel_region_get_aligned_offset(hiz_region,
522                                             draw_x & ~tile_mask_x,
523                                             (draw_y & ~tile_mask_y) / 2,
524                                             false);
525 
526 	 BEGIN_BATCH(3);
527 	 OUT_BATCH((_3DSTATE_HIER_DEPTH_BUFFER << 16) | (3 - 2));
528 	 OUT_BATCH(hiz_region->pitch * hiz_region->cpp - 1);
529 	 OUT_RELOC(hiz_region->bo,
530 		   I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
531 		   hiz_offset);
532 	 ADVANCE_BATCH();
533       } else {
534 	 BEGIN_BATCH(3);
535 	 OUT_BATCH((_3DSTATE_HIER_DEPTH_BUFFER << 16) | (3 - 2));
536 	 OUT_BATCH(0);
537 	 OUT_BATCH(0);
538 	 ADVANCE_BATCH();
539       }
540 
541       /* Emit stencil buffer. */
542       if (separate_stencil) {
543 	 struct intel_region *region = stencil_mt->region;
544 
545          /* Note: we can't compute the stencil offset using
546           * intel_region_get_aligned_offset(), because stencil_region claims
547           * that the region is untiled; in fact it's W tiled.
548           */
549          uint32_t stencil_offset =
550             (draw_y & ~tile_mask_y) * region->pitch +
551             (draw_x & ~tile_mask_x) * 64;
552 
553 	 BEGIN_BATCH(3);
554 	 OUT_BATCH((_3DSTATE_STENCIL_BUFFER << 16) | (3 - 2));
555          /* The stencil buffer has quirky pitch requirements.  From Vol 2a,
556           * 11.5.6.2.1 3DSTATE_STENCIL_BUFFER, field "Surface Pitch":
557           *    The pitch must be set to 2x the value computed based on width, as
558           *    the stencil buffer is stored with two rows interleaved.
559           */
560 	 OUT_BATCH(2 * region->pitch * region->cpp - 1);
561 	 OUT_RELOC(region->bo,
562 		   I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
563 		   stencil_offset);
564 	 ADVANCE_BATCH();
565       } else {
566 	 BEGIN_BATCH(3);
567 	 OUT_BATCH((_3DSTATE_STENCIL_BUFFER << 16) | (3 - 2));
568 	 OUT_BATCH(0);
569 	 OUT_BATCH(0);
570 	 ADVANCE_BATCH();
571       }
572    }
573 
574    /*
575     * On Gen >= 6, emit clear params for safety. If using hiz, then clear
576     * params must be emitted.
577     *
578     * From Section 2.11.5.6.4.1 3DSTATE_CLEAR_PARAMS:
579     *     3DSTATE_CLEAR_PARAMS packet must follow the DEPTH_BUFFER_STATE packet
580     *     when HiZ is enabled and the DEPTH_BUFFER_STATE changes.
581     */
582    if (intel->gen >= 6 || hiz_region) {
583       if (intel->gen == 6)
584 	 intel_emit_post_sync_nonzero_flush(intel);
585 
586       BEGIN_BATCH(2);
587       OUT_BATCH(_3DSTATE_CLEAR_PARAMS << 16 |
588 		GEN5_DEPTH_CLEAR_VALID |
589 		(2 - 2));
590       OUT_BATCH(depth_irb ? depth_irb->mt->depth_clear_value : 0);
591       ADVANCE_BATCH();
592    }
593 }
594 
595 const struct brw_tracked_state brw_depthbuffer = {
596    .dirty = {
597       .mesa = _NEW_BUFFERS,
598       .brw = BRW_NEW_BATCH,
599       .cache = 0,
600    },
601    .emit = emit_depthbuffer,
602 };
603 
604 
605 
606 /***********************************************************************
607  * Polygon stipple packet
608  */
609 
upload_polygon_stipple(struct brw_context * brw)610 static void upload_polygon_stipple(struct brw_context *brw)
611 {
612    struct intel_context *intel = &brw->intel;
613    struct gl_context *ctx = &brw->intel.ctx;
614    GLuint i;
615 
616    /* _NEW_POLYGON */
617    if (!ctx->Polygon.StippleFlag)
618       return;
619 
620    if (intel->gen == 6)
621       intel_emit_post_sync_nonzero_flush(intel);
622 
623    BEGIN_BATCH(33);
624    OUT_BATCH(_3DSTATE_POLY_STIPPLE_PATTERN << 16 | (33 - 2));
625 
626    /* Polygon stipple is provided in OpenGL order, i.e. bottom
627     * row first.  If we're rendering to a window (i.e. the
628     * default frame buffer object, 0), then we need to invert
629     * it to match our pixel layout.  But if we're rendering
630     * to a FBO (i.e. any named frame buffer object), we *don't*
631     * need to invert - we already match the layout.
632     */
633    if (_mesa_is_winsys_fbo(ctx->DrawBuffer)) {
634       for (i = 0; i < 32; i++)
635 	  OUT_BATCH(ctx->PolygonStipple[31 - i]); /* invert */
636    }
637    else {
638       for (i = 0; i < 32; i++)
639 	 OUT_BATCH(ctx->PolygonStipple[i]);
640    }
641    CACHED_BATCH();
642 }
643 
644 const struct brw_tracked_state brw_polygon_stipple = {
645    .dirty = {
646       .mesa = (_NEW_POLYGONSTIPPLE |
647 	       _NEW_POLYGON),
648       .brw = BRW_NEW_CONTEXT,
649       .cache = 0
650    },
651    .emit = upload_polygon_stipple
652 };
653 
654 
655 /***********************************************************************
656  * Polygon stipple offset packet
657  */
658 
upload_polygon_stipple_offset(struct brw_context * brw)659 static void upload_polygon_stipple_offset(struct brw_context *brw)
660 {
661    struct intel_context *intel = &brw->intel;
662    struct gl_context *ctx = &brw->intel.ctx;
663 
664    /* _NEW_POLYGON */
665    if (!ctx->Polygon.StippleFlag)
666       return;
667 
668    if (intel->gen == 6)
669       intel_emit_post_sync_nonzero_flush(intel);
670 
671    BEGIN_BATCH(2);
672    OUT_BATCH(_3DSTATE_POLY_STIPPLE_OFFSET << 16 | (2-2));
673 
674    /* _NEW_BUFFERS
675     *
676     * If we're drawing to a system window we have to invert the Y axis
677     * in order to match the OpenGL pixel coordinate system, and our
678     * offset must be matched to the window position.  If we're drawing
679     * to a user-created FBO then our native pixel coordinate system
680     * works just fine, and there's no window system to worry about.
681     */
682    if (_mesa_is_winsys_fbo(brw->intel.ctx.DrawBuffer))
683       OUT_BATCH((32 - (ctx->DrawBuffer->Height & 31)) & 31);
684    else
685       OUT_BATCH(0);
686    CACHED_BATCH();
687 }
688 
689 const struct brw_tracked_state brw_polygon_stipple_offset = {
690    .dirty = {
691       .mesa = (_NEW_BUFFERS |
692 	       _NEW_POLYGON),
693       .brw = BRW_NEW_CONTEXT,
694       .cache = 0
695    },
696    .emit = upload_polygon_stipple_offset
697 };
698 
699 /**********************************************************************
700  * AA Line parameters
701  */
upload_aa_line_parameters(struct brw_context * brw)702 static void upload_aa_line_parameters(struct brw_context *brw)
703 {
704    struct intel_context *intel = &brw->intel;
705    struct gl_context *ctx = &brw->intel.ctx;
706 
707    if (!ctx->Line.SmoothFlag || !brw->has_aa_line_parameters)
708       return;
709 
710    if (intel->gen == 6)
711       intel_emit_post_sync_nonzero_flush(intel);
712 
713    OUT_BATCH(_3DSTATE_AA_LINE_PARAMETERS << 16 | (3 - 2));
714    /* use legacy aa line coverage computation */
715    OUT_BATCH(0);
716    OUT_BATCH(0);
717    CACHED_BATCH();
718 }
719 
720 const struct brw_tracked_state brw_aa_line_parameters = {
721    .dirty = {
722       .mesa = _NEW_LINE,
723       .brw = BRW_NEW_CONTEXT,
724       .cache = 0
725    },
726    .emit = upload_aa_line_parameters
727 };
728 
729 /***********************************************************************
730  * Line stipple packet
731  */
732 
upload_line_stipple(struct brw_context * brw)733 static void upload_line_stipple(struct brw_context *brw)
734 {
735    struct intel_context *intel = &brw->intel;
736    struct gl_context *ctx = &brw->intel.ctx;
737    GLfloat tmp;
738    GLint tmpi;
739 
740    if (!ctx->Line.StippleFlag)
741       return;
742 
743    if (intel->gen == 6)
744       intel_emit_post_sync_nonzero_flush(intel);
745 
746    BEGIN_BATCH(3);
747    OUT_BATCH(_3DSTATE_LINE_STIPPLE_PATTERN << 16 | (3 - 2));
748    OUT_BATCH(ctx->Line.StipplePattern);
749    tmp = 1.0 / (GLfloat) ctx->Line.StippleFactor;
750    tmpi = tmp * (1<<13);
751    OUT_BATCH(tmpi << 16 | ctx->Line.StippleFactor);
752    CACHED_BATCH();
753 }
754 
755 const struct brw_tracked_state brw_line_stipple = {
756    .dirty = {
757       .mesa = _NEW_LINE,
758       .brw = BRW_NEW_CONTEXT,
759       .cache = 0
760    },
761    .emit = upload_line_stipple
762 };
763 
764 
765 /***********************************************************************
766  * Misc invariant state packets
767  */
768 
upload_invariant_state(struct brw_context * brw)769 static void upload_invariant_state( struct brw_context *brw )
770 {
771    struct intel_context *intel = &brw->intel;
772 
773    /* 3DSTATE_SIP, 3DSTATE_MULTISAMPLE, etc. are nonpipelined. */
774    if (intel->gen == 6)
775       intel_emit_post_sync_nonzero_flush(intel);
776 
777    /* Select the 3D pipeline (as opposed to media) */
778    BEGIN_BATCH(1);
779    OUT_BATCH(brw->CMD_PIPELINE_SELECT << 16 | 0);
780    ADVANCE_BATCH();
781 
782    if (intel->gen < 6) {
783       /* Disable depth offset clamping. */
784       BEGIN_BATCH(2);
785       OUT_BATCH(_3DSTATE_GLOBAL_DEPTH_OFFSET_CLAMP << 16 | (2 - 2));
786       OUT_BATCH_F(0.0);
787       ADVANCE_BATCH();
788    }
789 
790    if (intel->gen == 6) {
791       int i;
792 
793       for (i = 0; i < 4; i++) {
794          BEGIN_BATCH(4);
795          OUT_BATCH(_3DSTATE_GS_SVB_INDEX << 16 | (4 - 2));
796          OUT_BATCH(i << SVB_INDEX_SHIFT);
797          OUT_BATCH(0);
798          OUT_BATCH(0xffffffff);
799          ADVANCE_BATCH();
800       }
801    }
802 
803    BEGIN_BATCH(2);
804    OUT_BATCH(CMD_STATE_SIP << 16 | (2 - 2));
805    OUT_BATCH(0);
806    ADVANCE_BATCH();
807 
808    BEGIN_BATCH(1);
809    OUT_BATCH(brw->CMD_VF_STATISTICS << 16 |
810 	     (unlikely(INTEL_DEBUG & DEBUG_STATS) ? 1 : 0));
811    ADVANCE_BATCH();
812 }
813 
814 const struct brw_tracked_state brw_invariant_state = {
815    .dirty = {
816       .mesa = 0,
817       .brw = BRW_NEW_CONTEXT,
818       .cache = 0
819    },
820    .emit = upload_invariant_state
821 };
822 
823 /**
824  * Define the base addresses which some state is referenced from.
825  *
826  * This allows us to avoid having to emit relocations for the objects,
827  * and is actually required for binding table pointers on gen6.
828  *
829  * Surface state base address covers binding table pointers and
830  * surface state objects, but not the surfaces that the surface state
831  * objects point to.
832  */
upload_state_base_address(struct brw_context * brw)833 static void upload_state_base_address( struct brw_context *brw )
834 {
835    struct intel_context *intel = &brw->intel;
836 
837    /* FINISHME: According to section 3.6.1 "STATE_BASE_ADDRESS" of
838     * vol1a of the G45 PRM, MI_FLUSH with the ISC invalidate should be
839     * programmed prior to STATE_BASE_ADDRESS.
840     *
841     * However, given that the instruction SBA (general state base
842     * address) on this chipset is always set to 0 across X and GL,
843     * maybe this isn't required for us in particular.
844     */
845 
846    if (intel->gen >= 6) {
847       if (intel->gen == 6)
848 	 intel_emit_post_sync_nonzero_flush(intel);
849 
850        BEGIN_BATCH(10);
851        OUT_BATCH(CMD_STATE_BASE_ADDRESS << 16 | (10 - 2));
852        /* General state base address: stateless DP read/write requests */
853        OUT_BATCH(1);
854        /* Surface state base address:
855 	* BINDING_TABLE_STATE
856 	* SURFACE_STATE
857 	*/
858        OUT_RELOC(intel->batch.bo, I915_GEM_DOMAIN_SAMPLER, 0, 1);
859         /* Dynamic state base address:
860 	 * SAMPLER_STATE
861 	 * SAMPLER_BORDER_COLOR_STATE
862 	 * CLIP, SF, WM/CC viewport state
863 	 * COLOR_CALC_STATE
864 	 * DEPTH_STENCIL_STATE
865 	 * BLEND_STATE
866 	 * Push constants (when INSTPM: CONSTANT_BUFFER Address Offset
867 	 * Disable is clear, which we rely on)
868 	 */
869        OUT_RELOC(intel->batch.bo, (I915_GEM_DOMAIN_RENDER |
870 				   I915_GEM_DOMAIN_INSTRUCTION), 0, 1);
871 
872        OUT_BATCH(1); /* Indirect object base address: MEDIA_OBJECT data */
873        OUT_RELOC(brw->cache.bo, I915_GEM_DOMAIN_INSTRUCTION, 0,
874 		 1); /* Instruction base address: shader kernels (incl. SIP) */
875 
876        OUT_BATCH(1); /* General state upper bound */
877        /* Dynamic state upper bound.  Although the documentation says that
878 	* programming it to zero will cause it to be ignored, that is a lie.
879 	* If this isn't programmed to a real bound, the sampler border color
880 	* pointer is rejected, causing border color to mysteriously fail.
881 	*/
882        OUT_BATCH(0xfffff001);
883        OUT_BATCH(1); /* Indirect object upper bound */
884        OUT_BATCH(1); /* Instruction access upper bound */
885        ADVANCE_BATCH();
886    } else if (intel->gen == 5) {
887        BEGIN_BATCH(8);
888        OUT_BATCH(CMD_STATE_BASE_ADDRESS << 16 | (8 - 2));
889        OUT_BATCH(1); /* General state base address */
890        OUT_RELOC(intel->batch.bo, I915_GEM_DOMAIN_SAMPLER, 0,
891 		 1); /* Surface state base address */
892        OUT_BATCH(1); /* Indirect object base address */
893        OUT_RELOC(brw->cache.bo, I915_GEM_DOMAIN_INSTRUCTION, 0,
894 		 1); /* Instruction base address */
895        OUT_BATCH(0xfffff001); /* General state upper bound */
896        OUT_BATCH(1); /* Indirect object upper bound */
897        OUT_BATCH(1); /* Instruction access upper bound */
898        ADVANCE_BATCH();
899    } else {
900        BEGIN_BATCH(6);
901        OUT_BATCH(CMD_STATE_BASE_ADDRESS << 16 | (6 - 2));
902        OUT_BATCH(1); /* General state base address */
903        OUT_RELOC(intel->batch.bo, I915_GEM_DOMAIN_SAMPLER, 0,
904 		 1); /* Surface state base address */
905        OUT_BATCH(1); /* Indirect object base address */
906        OUT_BATCH(1); /* General state upper bound */
907        OUT_BATCH(1); /* Indirect object upper bound */
908        ADVANCE_BATCH();
909    }
910 
911    /* According to section 3.6.1 of VOL1 of the 965 PRM,
912     * STATE_BASE_ADDRESS updates require a reissue of:
913     *
914     * 3DSTATE_PIPELINE_POINTERS
915     * 3DSTATE_BINDING_TABLE_POINTERS
916     * MEDIA_STATE_POINTERS
917     *
918     * and this continues through Ironlake.  The Sandy Bridge PRM, vol
919     * 1 part 1 says that the folowing packets must be reissued:
920     *
921     * 3DSTATE_CC_POINTERS
922     * 3DSTATE_BINDING_TABLE_POINTERS
923     * 3DSTATE_SAMPLER_STATE_POINTERS
924     * 3DSTATE_VIEWPORT_STATE_POINTERS
925     * MEDIA_STATE_POINTERS
926     *
927     * Those are always reissued following SBA updates anyway (new
928     * batch time), except in the case of the program cache BO
929     * changing.  Having a separate state flag makes the sequence more
930     * obvious.
931     */
932 
933    brw->state.dirty.brw |= BRW_NEW_STATE_BASE_ADDRESS;
934 }
935 
936 const struct brw_tracked_state brw_state_base_address = {
937    .dirty = {
938       .mesa = 0,
939       .brw = (BRW_NEW_BATCH |
940 	      BRW_NEW_PROGRAM_CACHE),
941       .cache = 0,
942    },
943    .emit = upload_state_base_address
944 };
945