1 /**************************************************************************
3 * Copyright 2003 VMware, Inc.
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 VMWARE 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"
36 #include "main/macros.h"
37 #include "main/transformfeedback.h"
39 #include "vbo/vbo_context.h"
40 #include "swrast/swrast.h"
41 #include "swrast_setup/swrast_setup.h"
42 #include "drivers/common/meta.h"
44 #include "brw_blorp.h"
46 #include "brw_defines.h"
47 #include "brw_context.h"
48 #include "brw_state.h"
51 #include "intel_batchbuffer.h"
52 #include "intel_buffers.h"
53 #include "intel_fbo.h"
54 #include "intel_mipmap_tree.h"
55 #include "intel_buffer_objects.h"
57 #define FILE_DEBUG_FLAG DEBUG_PRIMS
59 static const GLuint prim_to_hw_prim
[GL_TRIANGLE_STRIP_ADJACENCY
+1] = {
71 _3DPRIM_LINESTRIP_ADJ
,
77 static const GLenum reduced_prim
[GL_POLYGON
+1] = {
91 get_hw_prim_for_gl_prim(int mode
)
93 if (mode
>= BRW_PRIM_OFFSET
)
94 return mode
- BRW_PRIM_OFFSET
;
96 return prim_to_hw_prim
[mode
];
100 /* When the primitive changes, set a state bit and re-validate. Not
101 * the nicest and would rather deal with this by having all the
102 * programs be immune to the active primitive (ie. cope with all
103 * possibilities). That may not be realistic however.
105 static void brw_set_prim(struct brw_context
*brw
,
106 const struct _mesa_prim
*prim
)
108 struct gl_context
*ctx
= &brw
->ctx
;
109 uint32_t hw_prim
= get_hw_prim_for_gl_prim(prim
->mode
);
111 DBG("PRIM: %s\n", _mesa_lookup_enum_by_nr(prim
->mode
));
113 /* Slight optimization to avoid the GS program when not needed:
115 if (prim
->mode
== GL_QUAD_STRIP
&&
116 ctx
->Light
.ShadeModel
!= GL_FLAT
&&
117 ctx
->Polygon
.FrontMode
== GL_FILL
&&
118 ctx
->Polygon
.BackMode
== GL_FILL
)
119 hw_prim
= _3DPRIM_TRISTRIP
;
121 if (prim
->mode
== GL_QUADS
&& prim
->count
== 4 &&
122 ctx
->Light
.ShadeModel
!= GL_FLAT
&&
123 ctx
->Polygon
.FrontMode
== GL_FILL
&&
124 ctx
->Polygon
.BackMode
== GL_FILL
) {
125 hw_prim
= _3DPRIM_TRIFAN
;
128 if (hw_prim
!= brw
->primitive
) {
129 brw
->primitive
= hw_prim
;
130 brw
->state
.dirty
.brw
|= BRW_NEW_PRIMITIVE
;
132 if (reduced_prim
[prim
->mode
] != brw
->reduced_primitive
) {
133 brw
->reduced_primitive
= reduced_prim
[prim
->mode
];
134 brw
->state
.dirty
.brw
|= BRW_NEW_REDUCED_PRIMITIVE
;
139 static void gen6_set_prim(struct brw_context
*brw
,
140 const struct _mesa_prim
*prim
)
144 DBG("PRIM: %s\n", _mesa_lookup_enum_by_nr(prim
->mode
));
146 hw_prim
= get_hw_prim_for_gl_prim(prim
->mode
);
148 if (hw_prim
!= brw
->primitive
) {
149 brw
->primitive
= hw_prim
;
150 brw
->state
.dirty
.brw
|= BRW_NEW_PRIMITIVE
;
156 * The hardware is capable of removing dangling vertices on its own; however,
157 * prior to Gen6, we sometimes convert quads into trifans (and quad strips
158 * into tristrips), since pre-Gen6 hardware requires a GS to render quads.
159 * This function manually trims dangling vertices from a draw call involving
160 * quads so that those dangling vertices won't get drawn when we convert to
163 static GLuint
trim(GLenum prim
, GLuint length
)
165 if (prim
== GL_QUAD_STRIP
)
166 return length
> 3 ? (length
- length
% 2) : 0;
167 else if (prim
== GL_QUADS
)
168 return length
- length
% 4;
174 static void brw_emit_prim(struct brw_context
*brw
,
175 const struct _mesa_prim
*prim
,
178 int verts_per_instance
;
179 int vertex_access_type
;
182 DBG("PRIM: %s %d %d\n", _mesa_lookup_enum_by_nr(prim
->mode
),
183 prim
->start
, prim
->count
);
185 int start_vertex_location
= prim
->start
;
186 int base_vertex_location
= prim
->basevertex
;
189 vertex_access_type
= brw
->gen
>= 7 ?
190 GEN7_3DPRIM_VERTEXBUFFER_ACCESS_RANDOM
:
191 GEN4_3DPRIM_VERTEXBUFFER_ACCESS_RANDOM
;
192 start_vertex_location
+= brw
->ib
.start_vertex_offset
;
193 base_vertex_location
+= brw
->vb
.start_vertex_bias
;
195 vertex_access_type
= brw
->gen
>= 7 ?
196 GEN7_3DPRIM_VERTEXBUFFER_ACCESS_SEQUENTIAL
:
197 GEN4_3DPRIM_VERTEXBUFFER_ACCESS_SEQUENTIAL
;
198 start_vertex_location
+= brw
->vb
.start_vertex_bias
;
201 /* We only need to trim the primitive count on pre-Gen6. */
203 verts_per_instance
= trim(prim
->mode
, prim
->count
);
205 verts_per_instance
= prim
->count
;
207 /* If nothing to emit, just return. */
208 if (verts_per_instance
== 0 && !prim
->is_indirect
)
211 /* If we're set to always flush, do it before and after the primitive emit.
212 * We want to catch both missed flushes that hurt instruction/state cache
213 * and missed flushes of the render cache as it heads to other parts of
214 * the besides the draw code.
216 if (brw
->always_flush_cache
) {
217 intel_batchbuffer_emit_mi_flush(brw
);
220 /* If indirect, emit a bunch of loads from the indirect BO. */
221 if (prim
->is_indirect
) {
222 struct gl_buffer_object
*indirect_buffer
= brw
->ctx
.DrawIndirectBuffer
;
223 drm_intel_bo
*bo
= intel_bufferobj_buffer(brw
,
224 intel_buffer_object(indirect_buffer
),
225 prim
->indirect_offset
, 5 * sizeof(GLuint
));
227 indirect_flag
= GEN7_3DPRIM_INDIRECT_PARAMETER_ENABLE
;
229 brw_load_register_mem(brw
, GEN7_3DPRIM_VERTEX_COUNT
, bo
,
230 I915_GEM_DOMAIN_VERTEX
, 0,
231 prim
->indirect_offset
+ 0);
232 brw_load_register_mem(brw
, GEN7_3DPRIM_INSTANCE_COUNT
, bo
,
233 I915_GEM_DOMAIN_VERTEX
, 0,
234 prim
->indirect_offset
+ 4);
236 brw_load_register_mem(brw
, GEN7_3DPRIM_START_VERTEX
, bo
,
237 I915_GEM_DOMAIN_VERTEX
, 0,
238 prim
->indirect_offset
+ 8);
240 brw_load_register_mem(brw
, GEN7_3DPRIM_BASE_VERTEX
, bo
,
241 I915_GEM_DOMAIN_VERTEX
, 0,
242 prim
->indirect_offset
+ 12);
243 brw_load_register_mem(brw
, GEN7_3DPRIM_START_INSTANCE
, bo
,
244 I915_GEM_DOMAIN_VERTEX
, 0,
245 prim
->indirect_offset
+ 16);
247 brw_load_register_mem(brw
, GEN7_3DPRIM_START_INSTANCE
, bo
,
248 I915_GEM_DOMAIN_VERTEX
, 0,
249 prim
->indirect_offset
+ 12);
251 OUT_BATCH(MI_LOAD_REGISTER_IMM
| (3 - 2));
252 OUT_BATCH(GEN7_3DPRIM_BASE_VERTEX
);
264 OUT_BATCH(CMD_3D_PRIM
<< 16 | (7 - 2) | indirect_flag
);
265 OUT_BATCH(hw_prim
| vertex_access_type
);
268 OUT_BATCH(CMD_3D_PRIM
<< 16 | (6 - 2) |
269 hw_prim
<< GEN4_3DPRIM_TOPOLOGY_TYPE_SHIFT
|
272 OUT_BATCH(verts_per_instance
);
273 OUT_BATCH(start_vertex_location
);
274 OUT_BATCH(prim
->num_instances
);
275 OUT_BATCH(prim
->base_instance
);
276 OUT_BATCH(base_vertex_location
);
279 if (brw
->always_flush_cache
) {
280 intel_batchbuffer_emit_mi_flush(brw
);
285 static void brw_merge_inputs( struct brw_context
*brw
,
286 const struct gl_client_array
*arrays
[])
288 const struct gl_context
*ctx
= &brw
->ctx
;
291 for (i
= 0; i
< brw
->vb
.nr_buffers
; i
++) {
292 drm_intel_bo_unreference(brw
->vb
.buffers
[i
].bo
);
293 brw
->vb
.buffers
[i
].bo
= NULL
;
295 brw
->vb
.nr_buffers
= 0;
297 for (i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
298 brw
->vb
.inputs
[i
].buffer
= -1;
299 brw
->vb
.inputs
[i
].glarray
= arrays
[i
];
302 if (brw
->gen
< 8 && !brw
->is_haswell
) {
303 struct gl_program
*vp
= &ctx
->VertexProgram
._Current
->Base
;
304 /* Prior to Haswell, the hardware can't natively support GL_FIXED or
305 * 2_10_10_10_REV vertex formats. Set appropriate workaround flags.
307 for (i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
308 if (!(vp
->InputsRead
& BITFIELD64_BIT(i
)))
311 uint8_t wa_flags
= 0;
313 switch (brw
->vb
.inputs
[i
].glarray
->Type
) {
316 wa_flags
= brw
->vb
.inputs
[i
].glarray
->Size
;
319 case GL_INT_2_10_10_10_REV
:
320 wa_flags
|= BRW_ATTRIB_WA_SIGN
;
323 case GL_UNSIGNED_INT_2_10_10_10_REV
:
324 if (brw
->vb
.inputs
[i
].glarray
->Format
== GL_BGRA
)
325 wa_flags
|= BRW_ATTRIB_WA_BGRA
;
327 if (brw
->vb
.inputs
[i
].glarray
->Normalized
)
328 wa_flags
|= BRW_ATTRIB_WA_NORMALIZE
;
329 else if (!brw
->vb
.inputs
[i
].glarray
->Integer
)
330 wa_flags
|= BRW_ATTRIB_WA_SCALE
;
335 if (brw
->vb
.attrib_wa_flags
[i
] != wa_flags
) {
336 brw
->vb
.attrib_wa_flags
[i
] = wa_flags
;
337 brw
->state
.dirty
.brw
|= BRW_NEW_VS_ATTRIB_WORKAROUNDS
;
344 * \brief Call this after drawing to mark which buffers need resolving
346 * If the depth buffer was written to and if it has an accompanying HiZ
347 * buffer, then mark that it needs a depth resolve.
349 * If the color buffer is a multisample window system buffer, then
350 * mark that it needs a downsample.
352 * Also mark any render targets which will be textured as needing a render
355 static void brw_postdraw_set_buffers_need_resolve(struct brw_context
*brw
)
357 struct gl_context
*ctx
= &brw
->ctx
;
358 struct gl_framebuffer
*fb
= ctx
->DrawBuffer
;
360 struct intel_renderbuffer
*front_irb
= NULL
;
361 struct intel_renderbuffer
*back_irb
= intel_get_renderbuffer(fb
, BUFFER_BACK_LEFT
);
362 struct intel_renderbuffer
*depth_irb
= intel_get_renderbuffer(fb
, BUFFER_DEPTH
);
363 struct intel_renderbuffer
*stencil_irb
= intel_get_renderbuffer(fb
, BUFFER_STENCIL
);
364 struct gl_renderbuffer_attachment
*depth_att
= &fb
->Attachment
[BUFFER_DEPTH
];
366 if (brw_is_front_buffer_drawing(fb
))
367 front_irb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
370 front_irb
->need_downsample
= true;
372 back_irb
->need_downsample
= true;
373 if (depth_irb
&& ctx
->Depth
.Mask
) {
374 intel_renderbuffer_att_set_needs_depth_resolve(depth_att
);
375 brw_render_cache_set_add_bo(brw
, depth_irb
->mt
->bo
);
378 if (ctx
->Extensions
.ARB_stencil_texturing
&&
379 stencil_irb
&& ctx
->Stencil
._WriteEnabled
) {
380 brw_render_cache_set_add_bo(brw
, stencil_irb
->mt
->bo
);
383 for (int i
= 0; i
< fb
->_NumColorDrawBuffers
; i
++) {
384 struct intel_renderbuffer
*irb
=
385 intel_renderbuffer(fb
->_ColorDrawBuffers
[i
]);
388 brw_render_cache_set_add_bo(brw
, irb
->mt
->bo
);
392 /* May fail if out of video memory for texture or vbo upload, or on
393 * fallback conditions.
395 static void brw_try_draw_prims( struct gl_context
*ctx
,
396 const struct gl_client_array
*arrays
[],
397 const struct _mesa_prim
*prims
,
399 const struct _mesa_index_buffer
*ib
,
402 struct gl_buffer_object
*indirect
)
404 struct brw_context
*brw
= brw_context(ctx
);
406 bool fail_next
= false;
409 _mesa_update_state( ctx
);
411 /* Find the highest sampler unit used by each shader program. A bit-count
412 * won't work since ARB programs use the texture unit number as the sampler
415 brw
->wm
.base
.sampler_count
=
416 _mesa_fls(ctx
->FragmentProgram
._Current
->Base
.SamplersUsed
);
417 brw
->gs
.base
.sampler_count
= ctx
->GeometryProgram
._Current
?
418 _mesa_fls(ctx
->GeometryProgram
._Current
->Base
.SamplersUsed
) : 0;
419 brw
->vs
.base
.sampler_count
=
420 _mesa_fls(ctx
->VertexProgram
._Current
->Base
.SamplersUsed
);
422 /* We have to validate the textures *before* checking for fallbacks;
423 * otherwise, the software fallback won't be able to rely on the
424 * texture state, the firstLevel and lastLevel fields won't be
425 * set in the intel texture object (they'll both be 0), and the
426 * software fallback will segfault if it attempts to access any
427 * texture level other than level 0.
429 brw_validate_textures( brw
);
431 intel_prepare_render(brw
);
433 /* This workaround has to happen outside of brw_upload_render_state()
434 * because it may flush the batchbuffer for a blit, affecting the state
437 brw_workaround_depthstencil_alignment(brw
, 0);
439 /* Bind all inputs, derive varying and size information:
441 brw_merge_inputs( brw
, arrays
);
444 brw
->state
.dirty
.brw
|= BRW_NEW_INDICES
;
446 brw
->vb
.min_index
= min_index
;
447 brw
->vb
.max_index
= max_index
;
448 brw
->state
.dirty
.brw
|= BRW_NEW_VERTICES
;
450 for (i
= 0; i
< nr_prims
; i
++) {
451 int estimated_max_prim_size
;
452 const int sampler_state_size
= 16;
454 estimated_max_prim_size
= 512; /* batchbuffer commands */
455 estimated_max_prim_size
+= BRW_MAX_TEX_UNIT
*
456 (sampler_state_size
+ sizeof(struct gen5_sampler_default_color
));
457 estimated_max_prim_size
+= 1024; /* gen6 VS push constants */
458 estimated_max_prim_size
+= 1024; /* gen6 WM push constants */
459 estimated_max_prim_size
+= 512; /* misc. pad */
461 /* Flush the batch if it's approaching full, so that we don't wrap while
462 * we've got validated state that needs to be in the same batch as the
465 intel_batchbuffer_require_space(brw
, estimated_max_prim_size
, RENDER_RING
);
466 intel_batchbuffer_save_state(brw
);
468 if (brw
->num_instances
!= prims
[i
].num_instances
||
469 brw
->basevertex
!= prims
[i
].basevertex
) {
470 brw
->num_instances
= prims
[i
].num_instances
;
471 brw
->basevertex
= prims
[i
].basevertex
;
472 if (i
> 0) { /* For i == 0 we just did this before the loop */
473 brw
->state
.dirty
.brw
|= BRW_NEW_VERTICES
;
474 brw_merge_inputs(brw
, arrays
);
478 brw
->draw
.gl_basevertex
=
479 prims
[i
].indexed
? prims
[i
].basevertex
: prims
[i
].start
;
481 drm_intel_bo_unreference(brw
->draw
.draw_params_bo
);
483 if (prims
[i
].is_indirect
) {
484 /* Point draw_params_bo at the indirect buffer. */
485 brw
->draw
.draw_params_bo
=
486 intel_buffer_object(ctx
->DrawIndirectBuffer
)->buffer
;
487 drm_intel_bo_reference(brw
->draw
.draw_params_bo
);
488 brw
->draw
.draw_params_offset
=
489 prims
[i
].indirect_offset
+ (prims
[i
].indexed
? 12 : 8);
491 /* Set draw_params_bo to NULL so brw_prepare_vertices knows it
492 * has to upload gl_BaseVertex and such if they're needed.
494 brw
->draw
.draw_params_bo
= NULL
;
495 brw
->draw
.draw_params_offset
= 0;
499 brw_set_prim(brw
, &prims
[i
]);
501 gen6_set_prim(brw
, &prims
[i
]);
505 /* Note that before the loop, brw->state.dirty.brw was set to != 0, and
506 * that the state updated in the loop outside of this block is that in
507 * *_set_prim or intel_batchbuffer_flush(), which only impacts
508 * brw->state.dirty.brw.
510 if (brw
->state
.dirty
.brw
) {
511 brw
->no_batch_wrap
= true;
512 brw_upload_render_state(brw
);
515 brw_emit_prim(brw
, &prims
[i
], brw
->primitive
);
517 brw
->no_batch_wrap
= false;
519 if (dri_bufmgr_check_aperture_space(&brw
->batch
.bo
, 1)) {
521 intel_batchbuffer_reset_to_saved(brw
);
522 intel_batchbuffer_flush(brw
);
526 int ret
= intel_batchbuffer_flush(brw
);
527 WARN_ONCE(ret
== -ENOSPC
,
528 "i965: Single primitive emit exceeded "
529 "available aperture space\n");
533 /* Now that we know we haven't run out of aperture space, we can safely
534 * reset the dirty bits.
536 if (brw
->state
.dirty
.brw
)
537 brw_clear_dirty_bits(brw
);
540 if (brw
->always_flush_batch
)
541 intel_batchbuffer_flush(brw
);
543 brw_state_cache_check_size(brw
);
544 brw_postdraw_set_buffers_need_resolve(brw
);
549 void brw_draw_prims( struct gl_context
*ctx
,
550 const struct _mesa_prim
*prims
,
552 const struct _mesa_index_buffer
*ib
,
553 GLboolean index_bounds_valid
,
556 struct gl_transform_feedback_object
*unused_tfb_object
,
557 struct gl_buffer_object
*indirect
)
559 struct brw_context
*brw
= brw_context(ctx
);
560 const struct gl_client_array
**arrays
= ctx
->Array
._DrawArrays
;
562 assert(unused_tfb_object
== NULL
);
564 if (ctx
->Query
.CondRenderQuery
) {
565 perf_debug("Conditional rendering is implemented in software and may "
566 "stall. This should be fixed in the driver.\n");
569 if (!_mesa_check_conditional_render(ctx
))
572 /* Handle primitive restart if needed */
573 if (brw_handle_primitive_restart(ctx
, prims
, nr_prims
, ib
, indirect
)) {
574 /* The draw was handled, so we can exit now */
578 /* Do GL_SELECT and GL_FEEDBACK rendering using swrast, even though it
579 * won't support all the extensions we support.
581 if (ctx
->RenderMode
!= GL_RENDER
) {
582 perf_debug("%s render mode not supported in hardware\n",
583 _mesa_lookup_enum_by_nr(ctx
->RenderMode
));
584 _swsetup_Wakeup(ctx
);
586 _tnl_draw_prims(ctx
, prims
, nr_prims
, ib
,
587 index_bounds_valid
, min_index
, max_index
, NULL
, NULL
);
591 /* If we're going to have to upload any of the user's vertex arrays, then
592 * get the minimum and maximum of their index buffer so we know what range
595 if (!index_bounds_valid
&& !vbo_all_varyings_in_vbos(arrays
)) {
596 perf_debug("Scanning index buffer to compute index buffer bounds. "
597 "Use glDrawRangeElements() to avoid this.\n");
598 vbo_get_minmax_indices(ctx
, prims
, ib
, &min_index
, &max_index
, nr_prims
);
601 /* Try drawing with the hardware, but don't do anything else if we can't
602 * manage it. swrast doesn't support our featureset, so we can't fall back
605 brw_try_draw_prims(ctx
, arrays
, prims
, nr_prims
, ib
, min_index
, max_index
, indirect
);
608 void brw_draw_init( struct brw_context
*brw
)
610 struct gl_context
*ctx
= &brw
->ctx
;
611 struct vbo_context
*vbo
= vbo_context(ctx
);
614 /* Register our drawing function:
616 vbo
->draw_prims
= brw_draw_prims
;
618 for (i
= 0; i
< VERT_ATTRIB_MAX
; i
++)
619 brw
->vb
.inputs
[i
].buffer
= -1;
620 brw
->vb
.nr_buffers
= 0;
621 brw
->vb
.nr_enabled
= 0;
624 void brw_draw_destroy( struct brw_context
*brw
)
628 for (i
= 0; i
< brw
->vb
.nr_buffers
; i
++) {
629 drm_intel_bo_unreference(brw
->vb
.buffers
[i
].bo
);
630 brw
->vb
.buffers
[i
].bo
= NULL
;
632 brw
->vb
.nr_buffers
= 0;
634 for (i
= 0; i
< brw
->vb
.nr_enabled
; i
++) {
635 brw
->vb
.enabled
[i
]->buffer
= -1;
637 brw
->vb
.nr_enabled
= 0;
639 drm_intel_bo_unreference(brw
->ib
.bo
);