1 /**************************************************************************
3 * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * 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, sub license, 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:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
28 #include <sys/errno.h>
30 #include "main/glheader.h"
31 #include "main/context.h"
32 #include "main/condrender.h"
33 #include "main/samplerobj.h"
34 #include "main/state.h"
35 #include "main/enums.h"
37 #include "vbo/vbo_context.h"
38 #include "swrast/swrast.h"
39 #include "swrast_setup/swrast_setup.h"
42 #include "brw_defines.h"
43 #include "brw_context.h"
44 #include "brw_state.h"
46 #include "intel_batchbuffer.h"
47 #include "intel_fbo.h"
48 #include "intel_mipmap_tree.h"
49 #include "intel_regions.h"
51 #define FILE_DEBUG_FLAG DEBUG_PRIMS
53 static GLuint prim_to_hw_prim
[GL_POLYGON
+1] = {
67 static const GLenum reduced_prim
[GL_POLYGON
+1] = {
81 /* When the primitive changes, set a state bit and re-validate. Not
82 * the nicest and would rather deal with this by having all the
83 * programs be immune to the active primitive (ie. cope with all
84 * possibilities). That may not be realistic however.
86 static void brw_set_prim(struct brw_context
*brw
,
87 const struct _mesa_prim
*prim
)
89 struct gl_context
*ctx
= &brw
->intel
.ctx
;
90 uint32_t hw_prim
= prim_to_hw_prim
[prim
->mode
];
92 DBG("PRIM: %s\n", _mesa_lookup_enum_by_nr(prim
->mode
));
94 /* Slight optimization to avoid the GS program when not needed:
96 if (prim
->mode
== GL_QUAD_STRIP
&&
97 ctx
->Light
.ShadeModel
!= GL_FLAT
&&
98 ctx
->Polygon
.FrontMode
== GL_FILL
&&
99 ctx
->Polygon
.BackMode
== GL_FILL
)
100 hw_prim
= _3DPRIM_TRISTRIP
;
102 if (prim
->mode
== GL_QUADS
&& prim
->count
== 4 &&
103 ctx
->Light
.ShadeModel
!= GL_FLAT
&&
104 ctx
->Polygon
.FrontMode
== GL_FILL
&&
105 ctx
->Polygon
.BackMode
== GL_FILL
) {
106 hw_prim
= _3DPRIM_TRIFAN
;
109 if (hw_prim
!= brw
->primitive
) {
110 brw
->primitive
= hw_prim
;
111 brw
->state
.dirty
.brw
|= BRW_NEW_PRIMITIVE
;
113 if (reduced_prim
[prim
->mode
] != brw
->intel
.reduced_primitive
) {
114 brw
->intel
.reduced_primitive
= reduced_prim
[prim
->mode
];
115 brw
->state
.dirty
.brw
|= BRW_NEW_REDUCED_PRIMITIVE
;
120 static void gen6_set_prim(struct brw_context
*brw
,
121 const struct _mesa_prim
*prim
)
125 DBG("PRIM: %s\n", _mesa_lookup_enum_by_nr(prim
->mode
));
128 assert(prim
->mode
== GL_TRIANGLES
);
129 hw_prim
= _3DPRIM_RECTLIST
;
131 hw_prim
= prim_to_hw_prim
[prim
->mode
];
134 if (hw_prim
!= brw
->primitive
) {
135 brw
->primitive
= hw_prim
;
136 brw
->state
.dirty
.brw
|= BRW_NEW_PRIMITIVE
;
141 static GLuint
trim(GLenum prim
, GLuint length
)
143 if (prim
== GL_QUAD_STRIP
)
144 return length
> 3 ? (length
- length
% 2) : 0;
145 else if (prim
== GL_QUADS
)
146 return length
- length
% 4;
152 static void brw_emit_prim(struct brw_context
*brw
,
153 const struct _mesa_prim
*prim
,
156 struct intel_context
*intel
= &brw
->intel
;
157 int verts_per_instance
;
158 int vertex_access_type
;
159 int start_vertex_location
;
160 int base_vertex_location
;
162 DBG("PRIM: %s %d %d\n", _mesa_lookup_enum_by_nr(prim
->mode
),
163 prim
->start
, prim
->count
);
165 start_vertex_location
= prim
->start
;
166 base_vertex_location
= prim
->basevertex
;
168 vertex_access_type
= GEN4_3DPRIM_VERTEXBUFFER_ACCESS_RANDOM
;
169 start_vertex_location
+= brw
->ib
.start_vertex_offset
;
170 base_vertex_location
+= brw
->vb
.start_vertex_bias
;
172 vertex_access_type
= GEN4_3DPRIM_VERTEXBUFFER_ACCESS_SEQUENTIAL
;
173 start_vertex_location
+= brw
->vb
.start_vertex_bias
;
176 verts_per_instance
= trim(prim
->mode
, prim
->count
);
178 /* If nothing to emit, just return. */
179 if (verts_per_instance
== 0)
182 /* If we're set to always flush, do it before and after the primitive emit.
183 * We want to catch both missed flushes that hurt instruction/state cache
184 * and missed flushes of the render cache as it heads to other parts of
185 * the besides the draw code.
187 if (intel
->always_flush_cache
) {
188 intel_batchbuffer_emit_mi_flush(intel
);
192 OUT_BATCH(CMD_3D_PRIM
<< 16 | (6 - 2) |
193 hw_prim
<< GEN4_3DPRIM_TOPOLOGY_TYPE_SHIFT
|
195 OUT_BATCH(verts_per_instance
);
196 OUT_BATCH(start_vertex_location
);
197 OUT_BATCH(1); // instance count
198 OUT_BATCH(0); // start instance location
199 OUT_BATCH(base_vertex_location
);
202 intel
->batch
.need_workaround_flush
= true;
204 if (intel
->always_flush_cache
) {
205 intel_batchbuffer_emit_mi_flush(intel
);
209 static void gen7_emit_prim(struct brw_context
*brw
,
210 const struct _mesa_prim
*prim
,
213 struct intel_context
*intel
= &brw
->intel
;
214 int verts_per_instance
;
215 int vertex_access_type
;
216 int start_vertex_location
;
217 int base_vertex_location
;
219 DBG("PRIM: %s %d %d\n", _mesa_lookup_enum_by_nr(prim
->mode
),
220 prim
->start
, prim
->count
);
222 start_vertex_location
= prim
->start
;
223 base_vertex_location
= prim
->basevertex
;
225 vertex_access_type
= GEN7_3DPRIM_VERTEXBUFFER_ACCESS_RANDOM
;
226 start_vertex_location
+= brw
->ib
.start_vertex_offset
;
227 base_vertex_location
+= brw
->vb
.start_vertex_bias
;
229 vertex_access_type
= GEN7_3DPRIM_VERTEXBUFFER_ACCESS_SEQUENTIAL
;
230 start_vertex_location
+= brw
->vb
.start_vertex_bias
;
233 verts_per_instance
= trim(prim
->mode
, prim
->count
);
235 /* If nothing to emit, just return. */
236 if (verts_per_instance
== 0)
239 /* If we're set to always flush, do it before and after the primitive emit.
240 * We want to catch both missed flushes that hurt instruction/state cache
241 * and missed flushes of the render cache as it heads to other parts of
242 * the besides the draw code.
244 if (intel
->always_flush_cache
) {
245 intel_batchbuffer_emit_mi_flush(intel
);
249 OUT_BATCH(CMD_3D_PRIM
<< 16 | (7 - 2));
250 OUT_BATCH(hw_prim
| vertex_access_type
);
251 OUT_BATCH(verts_per_instance
);
252 OUT_BATCH(start_vertex_location
);
253 OUT_BATCH(1); // instance count
254 OUT_BATCH(0); // start instance location
255 OUT_BATCH(base_vertex_location
);
258 if (intel
->always_flush_cache
) {
259 intel_batchbuffer_emit_mi_flush(intel
);
264 static void brw_merge_inputs( struct brw_context
*brw
,
265 const struct gl_client_array
*arrays
[])
267 struct brw_vertex_info old
= brw
->vb
.info
;
270 for (i
= 0; i
< brw
->vb
.nr_buffers
; i
++) {
271 drm_intel_bo_unreference(brw
->vb
.buffers
[i
].bo
);
272 brw
->vb
.buffers
[i
].bo
= NULL
;
274 brw
->vb
.nr_buffers
= 0;
276 memset(&brw
->vb
.info
, 0, sizeof(brw
->vb
.info
));
278 for (i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
279 brw
->vb
.inputs
[i
].buffer
= -1;
280 brw
->vb
.inputs
[i
].glarray
= arrays
[i
];
281 brw
->vb
.inputs
[i
].attrib
= (gl_vert_attrib
) i
;
283 if (arrays
[i
]->StrideB
!= 0)
284 brw
->vb
.info
.sizes
[i
/16] |= (brw
->vb
.inputs
[i
].glarray
->Size
- 1) <<
288 /* Raise statechanges if input sizes have changed. */
289 if (memcmp(brw
->vb
.info
.sizes
, old
.sizes
, sizeof(old
.sizes
)) != 0)
290 brw
->state
.dirty
.brw
|= BRW_NEW_INPUT_DIMENSIONS
;
294 * \brief Resolve buffers before drawing.
296 * Resolve the depth buffer's HiZ buffer and resolve the depth buffer of each
297 * enabled depth texture.
299 * (In the future, this will also perform MSAA resolves).
302 brw_predraw_resolve_buffers(struct brw_context
*brw
)
304 struct gl_context
*ctx
= &brw
->intel
.ctx
;
305 struct intel_context
*intel
= &brw
->intel
;
306 struct intel_renderbuffer
*depth_irb
;
307 struct intel_texture_object
*tex_obj
;
308 bool did_resolve
= false;
310 /* Avoid recursive HiZ op. */
315 /* Resolve the depth buffer's HiZ buffer. */
316 depth_irb
= intel_get_renderbuffer(ctx
->DrawBuffer
, BUFFER_DEPTH
);
317 if (depth_irb
&& depth_irb
->mt
) {
318 did_resolve
|= intel_renderbuffer_resolve_hiz(intel
, depth_irb
);
321 /* Resolve depth buffer of each enabled depth texture. */
322 for (int i
= 0; i
< BRW_MAX_TEX_UNIT
; i
++) {
323 if (!ctx
->Texture
.Unit
[i
]._ReallyEnabled
)
325 tex_obj
= intel_texture_object(ctx
->Texture
.Unit
[i
]._Current
);
326 if (!tex_obj
|| !tex_obj
->mt
)
328 did_resolve
|= intel_miptree_all_slices_resolve_depth(intel
, tex_obj
->mt
);
332 /* Call vbo_bind_array() to synchronize the vbo module's vertex
333 * attributes to the gl_context's.
337 * The vbo module tracks vertex attributes separately from the
338 * gl_context. Specifically, the vbo module maintins vertex attributes
339 * in vbo_exec_context::array::inputs, which is synchronized with
340 * gl_context::Array::ArrayObj::VertexAttrib by vbo_bind_array().
341 * vbo_draw_arrays() calls vbo_bind_array() to perform the
342 * synchronization before calling the real draw call,
343 * vbo_context::draw_arrays.
345 * At this point (after performing a resolve meta-op but before calling
346 * vbo_bind_array), the gl_context's vertex attributes have been
347 * restored to their original state (that is, their state before the
348 * meta-op began), but the vbo module's vertex attribute are those used
349 * in the last meta-op. Therefore we must manually synchronize the two with
350 * vbo_bind_array() before continuing with the original draw command.
352 _mesa_update_state(ctx
);
353 vbo_bind_arrays(ctx
);
354 _mesa_update_state(ctx
);
359 * \brief Call this after drawing to mark which buffers need resolving
361 * If the depth buffer was written to and if it has an accompanying HiZ
362 * buffer, then mark that it needs a depth resolve.
364 * (In the future, this will also mark needed MSAA resolves).
366 static void brw_postdraw_set_buffers_need_resolve(struct brw_context
*brw
)
368 struct gl_context
*ctx
= &brw
->intel
.ctx
;
369 struct gl_framebuffer
*fb
= ctx
->DrawBuffer
;
370 struct intel_renderbuffer
*depth_irb
=
371 intel_get_renderbuffer(fb
, BUFFER_DEPTH
);
376 intel_renderbuffer_set_needs_depth_resolve(depth_irb
);
380 /* May fail if out of video memory for texture or vbo upload, or on
381 * fallback conditions.
383 static bool brw_try_draw_prims( struct gl_context
*ctx
,
384 const struct gl_client_array
*arrays
[],
385 const struct _mesa_prim
*prim
,
387 const struct _mesa_index_buffer
*ib
,
391 struct intel_context
*intel
= intel_context(ctx
);
392 struct brw_context
*brw
= brw_context(ctx
);
395 bool fail_next
= false;
398 _mesa_update_state( ctx
);
400 /* We have to validate the textures *before* checking for fallbacks;
401 * otherwise, the software fallback won't be able to rely on the
402 * texture state, the firstLevel and lastLevel fields won't be
403 * set in the intel texture object (they'll both be 0), and the
404 * software fallback will segfault if it attempts to access any
405 * texture level other than level 0.
407 brw_validate_textures( brw
);
409 /* Resolves must occur after updating state and finalizing textures but
410 * before setting up any hardware state for this draw call.
412 brw_predraw_resolve_buffers(brw
);
414 /* Bind all inputs, derive varying and size information:
416 brw_merge_inputs( brw
, arrays
);
419 brw
->state
.dirty
.brw
|= BRW_NEW_INDICES
;
421 brw
->vb
.min_index
= min_index
;
422 brw
->vb
.max_index
= max_index
;
423 brw
->state
.dirty
.brw
|= BRW_NEW_VERTICES
;
425 /* Have to validate state quite late. Will rebuild tnl_program,
426 * which depends on varying information.
428 * Note this is where brw->vs->prog_data.inputs_read is calculated,
429 * so can't access it earlier.
432 intel_prepare_render(intel
);
434 for (i
= 0; i
< nr_prims
; i
++) {
435 int estimated_max_prim_size
;
437 estimated_max_prim_size
= 512; /* batchbuffer commands */
438 estimated_max_prim_size
+= (BRW_MAX_TEX_UNIT
*
439 (sizeof(struct brw_sampler_state
) +
440 sizeof(struct gen5_sampler_default_color
)));
441 estimated_max_prim_size
+= 1024; /* gen6 VS push constants */
442 estimated_max_prim_size
+= 1024; /* gen6 WM push constants */
443 estimated_max_prim_size
+= 512; /* misc. pad */
445 /* Flush the batch if it's approaching full, so that we don't wrap while
446 * we've got validated state that needs to be in the same batch as the
449 intel_batchbuffer_require_space(intel
, estimated_max_prim_size
, false);
450 intel_batchbuffer_save_state(intel
);
453 brw_set_prim(brw
, &prim
[i
]);
455 gen6_set_prim(brw
, &prim
[i
]);
458 /* Note that before the loop, brw->state.dirty.brw was set to != 0, and
459 * that the state updated in the loop outside of this block is that in
460 * *_set_prim or intel_batchbuffer_flush(), which only impacts
461 * brw->state.dirty.brw.
463 if (brw
->state
.dirty
.brw
) {
464 intel
->no_batch_wrap
= true;
465 brw_upload_state(brw
);
467 if (unlikely(brw
->intel
.Fallback
)) {
468 intel
->no_batch_wrap
= false;
475 gen7_emit_prim(brw
, &prim
[i
], brw
->primitive
);
477 brw_emit_prim(brw
, &prim
[i
], brw
->primitive
);
479 intel
->no_batch_wrap
= false;
481 if (dri_bufmgr_check_aperture_space(&intel
->batch
.bo
, 1)) {
483 intel_batchbuffer_reset_to_saved(intel
);
484 intel_batchbuffer_flush(intel
);
488 if (intel_batchbuffer_flush(intel
) == -ENOSPC
) {
489 static bool warned
= false;
492 fprintf(stderr
, "i965: Single primitive emit exceeded"
493 "available aperture space\n");
503 if (intel
->always_flush_batch
)
504 intel_batchbuffer_flush(intel
);
507 brw_state_cache_check_size(brw
);
508 brw_postdraw_set_buffers_need_resolve(brw
);
513 void brw_draw_prims( struct gl_context
*ctx
,
514 const struct gl_client_array
*arrays
[],
515 const struct _mesa_prim
*prim
,
517 const struct _mesa_index_buffer
*ib
,
518 GLboolean index_bounds_valid
,
521 struct gl_transform_feedback_object
*tfb_vertcount
)
525 if (!_mesa_check_conditional_render(ctx
))
528 if (!vbo_all_varyings_in_vbos(arrays
)) {
529 if (!index_bounds_valid
)
530 vbo_get_minmax_index(ctx
, prim
, ib
, &min_index
, &max_index
);
532 /* Decide if we want to rebase. If so we end up recursing once
533 * only into this function.
535 if (min_index
!= 0 && !vbo_any_varyings_in_vbos(arrays
)) {
536 vbo_rebase_prims(ctx
, arrays
,
538 ib
, min_index
, max_index
,
544 /* Make a first attempt at drawing:
546 retval
= brw_try_draw_prims(ctx
, arrays
, prim
, nr_prims
, ib
, min_index
, max_index
);
548 /* Otherwise, we really are out of memory. Pass the drawing
549 * command to the software tnl module and which will in turn call
550 * swrast to do the drawing.
553 _swsetup_Wakeup(ctx
);
555 _tnl_draw_prims(ctx
, arrays
, prim
, nr_prims
, ib
, min_index
, max_index
);
560 void brw_draw_init( struct brw_context
*brw
)
562 struct gl_context
*ctx
= &brw
->intel
.ctx
;
563 struct vbo_context
*vbo
= vbo_context(ctx
);
566 /* Register our drawing function:
568 vbo
->draw_prims
= brw_draw_prims
;
570 for (i
= 0; i
< VERT_ATTRIB_MAX
; i
++)
571 brw
->vb
.inputs
[i
].buffer
= -1;
572 brw
->vb
.nr_buffers
= 0;
573 brw
->vb
.nr_enabled
= 0;
576 void brw_draw_destroy( struct brw_context
*brw
)
580 for (i
= 0; i
< brw
->vb
.nr_buffers
; i
++) {
581 drm_intel_bo_unreference(brw
->vb
.buffers
[i
].bo
);
582 brw
->vb
.buffers
[i
].bo
= NULL
;
584 brw
->vb
.nr_buffers
= 0;
586 for (i
= 0; i
< brw
->vb
.nr_enabled
; i
++) {
587 brw
->vb
.enabled
[i
]->buffer
= -1;
589 brw
->vb
.nr_enabled
= 0;
591 drm_intel_bo_unreference(brw
->ib
.bo
);