2 Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
3 Copyright (C) Intel Corp. 2006. All Rights Reserved.
4 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
5 develop this 3D driver.
7 Permission is hereby granted, free of charge, to any person obtaining
8 a copy of this software and associated documentation files (the
9 "Software"), to deal in the Software without restriction, including
10 without limitation the rights to use, copy, modify, merge, publish,
11 distribute, sublicense, and/or sell copies of the Software, and to
12 permit persons to whom the Software is furnished to do so, subject to
13 the following conditions:
15 The above copyright notice and this permission notice (including the
16 next paragraph) shall be included in all copies or substantial
17 portions of the Software.
19 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
20 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
22 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
23 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
24 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
25 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **********************************************************************/
30 * Keith Whitwell <keith@tungstengraphics.com>
34 #include "main/api_exec.h"
35 #include "main/context.h"
36 #include "main/fbobject.h"
37 #include "main/imports.h"
38 #include "main/macros.h"
39 #include "main/points.h"
40 #include "main/version.h"
41 #include "main/vtxfmt.h"
43 #include "vbo/vbo_context.h"
45 #include "drivers/common/driverfuncs.h"
46 #include "drivers/common/meta.h"
49 #include "brw_context.h"
50 #include "brw_defines.h"
52 #include "brw_state.h"
54 #include "intel_batchbuffer.h"
55 #include "intel_buffer_objects.h"
56 #include "intel_buffers.h"
57 #include "intel_fbo.h"
58 #include "intel_mipmap_tree.h"
59 #include "intel_pixel.h"
60 #include "intel_regions.h"
61 #include "intel_tex.h"
62 #include "intel_tex_obj.h"
64 #include "swrast_setup/swrast_setup.h"
66 #include "tnl/t_pipeline.h"
67 #include "glsl/ralloc.h"
69 /***************************************
70 * Mesa's Driver Functions
71 ***************************************/
74 brw_query_samples_for_format(struct gl_context
*ctx
, GLenum target
,
75 GLenum internalFormat
, int samples
[16])
77 struct brw_context
*brw
= brw_context(ctx
);
97 static const GLubyte
*
98 intelGetString(struct gl_context
* ctx
, GLenum name
)
100 const struct brw_context
*const brw
= brw_context(ctx
);
102 static char buffer
[128];
106 return (GLubyte
*) "Intel Open Source Technology Center";
110 switch (brw
->intelScreen
->deviceID
) {
112 #define CHIPSET(id, family, str) case id: chipset = str; break;
113 #include "pci_ids/i965_pci_ids.h"
115 chipset
= "Unknown Intel Chipset";
119 (void) driGetRendererString(buffer
, chipset
, 0);
120 return (GLubyte
*) buffer
;
128 intel_viewport(struct gl_context
*ctx
, GLint x
, GLint y
, GLsizei w
, GLsizei h
)
130 struct brw_context
*brw
= brw_context(ctx
);
131 __DRIcontext
*driContext
= brw
->driContext
;
138 if (_mesa_is_winsys_fbo(ctx
->DrawBuffer
)) {
139 dri2InvalidateDrawable(driContext
->driDrawablePriv
);
140 dri2InvalidateDrawable(driContext
->driReadablePriv
);
145 intelInvalidateState(struct gl_context
* ctx
, GLuint new_state
)
147 struct brw_context
*brw
= brw_context(ctx
);
149 if (ctx
->swrast_context
)
150 _swrast_InvalidateState(ctx
, new_state
);
151 _vbo_InvalidateState(ctx
, new_state
);
153 brw
->NewGLState
|= new_state
;
157 intel_flush_front(struct gl_context
*ctx
)
159 struct brw_context
*brw
= brw_context(ctx
);
160 __DRIcontext
*driContext
= brw
->driContext
;
161 __DRIdrawable
*driDrawable
= driContext
->driDrawablePriv
;
162 __DRIscreen
*const screen
= brw
->intelScreen
->driScrnPriv
;
164 if (brw
->front_buffer_dirty
&& _mesa_is_winsys_fbo(ctx
->DrawBuffer
)) {
165 if (screen
->dri2
.loader
->flushFrontBuffer
!= NULL
&&
167 driDrawable
->loaderPrivate
) {
169 /* Resolve before flushing FAKE_FRONT_LEFT to FRONT_LEFT.
171 * This potentially resolves both front and back buffer. It
172 * is unnecessary to resolve the back, but harms nothing except
173 * performance. And no one cares about front-buffer render
176 intel_resolve_for_dri2_flush(brw
, driDrawable
);
177 intel_batchbuffer_flush(brw
);
179 screen
->dri2
.loader
->flushFrontBuffer(driDrawable
,
180 driDrawable
->loaderPrivate
);
182 /* We set the dirty bit in intel_prepare_render() if we're
183 * front buffer rendering once we get there.
185 brw
->front_buffer_dirty
= false;
191 intel_glFlush(struct gl_context
*ctx
)
193 struct brw_context
*brw
= brw_context(ctx
);
195 intel_batchbuffer_flush(brw
);
196 intel_flush_front(ctx
);
197 if (brw
->is_front_buffer_rendering
)
198 brw
->need_throttle
= true;
202 intelFinish(struct gl_context
* ctx
)
204 struct brw_context
*brw
= brw_context(ctx
);
208 if (brw
->batch
.last_bo
)
209 drm_intel_bo_wait_rendering(brw
->batch
.last_bo
);
213 brw_init_driver_functions(struct brw_context
*brw
,
214 struct dd_function_table
*functions
)
216 _mesa_init_driver_functions(functions
);
218 /* GLX uses DRI2 invalidate events to handle window resizing.
219 * Unfortunately, EGL does not - libEGL is written in XCB (not Xlib),
220 * which doesn't provide a mechanism for snooping the event queues.
222 * So EGL still relies on viewport hacks to handle window resizing.
223 * This should go away with DRI3000.
225 if (!brw
->driContext
->driScreenPriv
->dri2
.useInvalidate
)
226 functions
->Viewport
= intel_viewport
;
228 functions
->Flush
= intel_glFlush
;
229 functions
->Finish
= intelFinish
;
230 functions
->GetString
= intelGetString
;
231 functions
->UpdateState
= intelInvalidateState
;
233 intelInitTextureFuncs(functions
);
234 intelInitTextureImageFuncs(functions
);
235 intelInitTextureSubImageFuncs(functions
);
236 intelInitTextureCopyImageFuncs(functions
);
237 intelInitClearFuncs(functions
);
238 intelInitBufferFuncs(functions
);
239 intelInitPixelFuncs(functions
);
240 intelInitBufferObjectFuncs(functions
);
241 intel_init_syncobj_functions(functions
);
242 brw_init_object_purgeable_functions(functions
);
244 brwInitFragProgFuncs( functions
);
245 brw_init_common_queryobj_functions(functions
);
247 gen6_init_queryobj_functions(functions
);
249 gen4_init_queryobj_functions(functions
);
251 functions
->QuerySamplesForFormat
= brw_query_samples_for_format
;
253 functions
->NewTransformFeedback
= brw_new_transform_feedback
;
254 functions
->DeleteTransformFeedback
= brw_delete_transform_feedback
;
255 functions
->GetTransformFeedbackVertexCount
=
256 brw_get_transform_feedback_vertex_count
;
258 functions
->BeginTransformFeedback
= gen7_begin_transform_feedback
;
259 functions
->EndTransformFeedback
= gen7_end_transform_feedback
;
260 functions
->PauseTransformFeedback
= gen7_pause_transform_feedback
;
261 functions
->ResumeTransformFeedback
= gen7_resume_transform_feedback
;
263 functions
->BeginTransformFeedback
= brw_begin_transform_feedback
;
264 functions
->EndTransformFeedback
= brw_end_transform_feedback
;
268 functions
->GetSamplePosition
= gen6_get_sample_position
;
272 * Return array of MSAA modes supported by the hardware. The array is
273 * zero-terminated and sorted in decreasing order.
276 brw_supported_msaa_modes(const struct brw_context
*brw
)
278 static const int gen7_samples
[] = {8, 4, 0};
279 static const int gen6_samples
[] = {4, 0};
280 static const int gen4_samples
[] = {0};
283 } else if (brw
->gen
== 6) {
291 * Override GL_MAX_SAMPLES and related constants according to value of driconf
292 * option 'clamp_max_samples'.
295 brw_override_max_samples(struct brw_context
*brw
)
297 const int clamp_max_samples
= driQueryOptioni(&brw
->optionCache
,
298 "clamp_max_samples");
299 if (clamp_max_samples
< 0)
302 const int *supported_msaa_modes
= brw_supported_msaa_modes(brw
);
305 /* Select the largest supported MSAA mode that does not exceed
308 for (int i
= 0; supported_msaa_modes
[i
] != 0; ++i
) {
309 if (supported_msaa_modes
[i
] <= clamp_max_samples
) {
310 max_samples
= supported_msaa_modes
[i
];
315 brw
->ctx
.Const
.MaxSamples
= max_samples
;
316 brw
->ctx
.Const
.MaxColorTextureSamples
= max_samples
;
317 brw
->ctx
.Const
.MaxDepthTextureSamples
= max_samples
;
318 brw
->ctx
.Const
.MaxIntegerSamples
= max_samples
;
322 brw_initialize_context_constants(struct brw_context
*brw
)
324 struct gl_context
*ctx
= &brw
->ctx
;
326 ctx
->Const
.QueryCounterBits
.Timestamp
= 36;
328 ctx
->Const
.StripTextureBorder
= true;
330 ctx
->Const
.MaxDualSourceDrawBuffers
= 1;
331 ctx
->Const
.MaxDrawBuffers
= BRW_MAX_DRAW_BUFFERS
;
332 ctx
->Const
.FragmentProgram
.MaxTextureImageUnits
= BRW_MAX_TEX_UNIT
;
333 ctx
->Const
.MaxTextureCoordUnits
= 8; /* Mesa limit */
334 ctx
->Const
.MaxTextureUnits
=
335 MIN2(ctx
->Const
.MaxTextureCoordUnits
,
336 ctx
->Const
.FragmentProgram
.MaxTextureImageUnits
);
337 ctx
->Const
.VertexProgram
.MaxTextureImageUnits
= BRW_MAX_TEX_UNIT
;
339 ctx
->Const
.GeometryProgram
.MaxTextureImageUnits
= BRW_MAX_TEX_UNIT
;
341 ctx
->Const
.GeometryProgram
.MaxTextureImageUnits
= 0;
342 ctx
->Const
.MaxCombinedTextureImageUnits
=
343 ctx
->Const
.VertexProgram
.MaxTextureImageUnits
+
344 ctx
->Const
.FragmentProgram
.MaxTextureImageUnits
+
345 ctx
->Const
.GeometryProgram
.MaxTextureImageUnits
;
347 ctx
->Const
.MaxTextureLevels
= 14; /* 8192 */
348 if (ctx
->Const
.MaxTextureLevels
> MAX_TEXTURE_LEVELS
)
349 ctx
->Const
.MaxTextureLevels
= MAX_TEXTURE_LEVELS
;
350 ctx
->Const
.Max3DTextureLevels
= 9;
351 ctx
->Const
.MaxCubeTextureLevels
= 12;
354 ctx
->Const
.MaxArrayTextureLayers
= 2048;
356 ctx
->Const
.MaxArrayTextureLayers
= 512;
358 ctx
->Const
.MaxTextureRectSize
= 1 << 12;
360 ctx
->Const
.MaxTextureMaxAnisotropy
= 16.0;
362 ctx
->Const
.MaxRenderbufferSize
= 8192;
364 /* Hardware only supports a limited number of transform feedback buffers.
365 * So we need to override the Mesa default (which is based only on software
368 ctx
->Const
.MaxTransformFeedbackBuffers
= BRW_MAX_SOL_BUFFERS
;
370 /* On Gen6, in the worst case, we use up one binding table entry per
371 * transform feedback component (see comments above the definition of
372 * BRW_MAX_SOL_BINDINGS, in brw_context.h), so we need to advertise a value
373 * for MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS equal to
374 * BRW_MAX_SOL_BINDINGS.
376 * In "separate components" mode, we need to divide this value by
377 * BRW_MAX_SOL_BUFFERS, so that the total number of binding table entries
378 * used up by all buffers will not exceed BRW_MAX_SOL_BINDINGS.
380 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
= BRW_MAX_SOL_BINDINGS
;
381 ctx
->Const
.MaxTransformFeedbackSeparateComponents
=
382 BRW_MAX_SOL_BINDINGS
/ BRW_MAX_SOL_BUFFERS
;
384 ctx
->Const
.AlwaysUseGetTransformFeedbackVertexCount
= true;
386 const int max_samples
= brw_supported_msaa_modes(brw
)[0];
387 ctx
->Const
.MaxSamples
= max_samples
;
388 ctx
->Const
.MaxColorTextureSamples
= max_samples
;
389 ctx
->Const
.MaxDepthTextureSamples
= max_samples
;
390 ctx
->Const
.MaxIntegerSamples
= max_samples
;
393 ctx
->Const
.MaxProgramTextureGatherComponents
= 4;
395 ctx
->Const
.MinLineWidth
= 1.0;
396 ctx
->Const
.MinLineWidthAA
= 1.0;
397 ctx
->Const
.MaxLineWidth
= 5.0;
398 ctx
->Const
.MaxLineWidthAA
= 5.0;
399 ctx
->Const
.LineWidthGranularity
= 0.5;
401 ctx
->Const
.MinPointSize
= 1.0;
402 ctx
->Const
.MinPointSizeAA
= 1.0;
403 ctx
->Const
.MaxPointSize
= 255.0;
404 ctx
->Const
.MaxPointSizeAA
= 255.0;
405 ctx
->Const
.PointSizeGranularity
= 1.0;
407 if (brw
->gen
>= 5 || brw
->is_g4x
)
408 ctx
->Const
.MaxClipPlanes
= 8;
410 ctx
->Const
.VertexProgram
.MaxNativeInstructions
= 16 * 1024;
411 ctx
->Const
.VertexProgram
.MaxAluInstructions
= 0;
412 ctx
->Const
.VertexProgram
.MaxTexInstructions
= 0;
413 ctx
->Const
.VertexProgram
.MaxTexIndirections
= 0;
414 ctx
->Const
.VertexProgram
.MaxNativeAluInstructions
= 0;
415 ctx
->Const
.VertexProgram
.MaxNativeTexInstructions
= 0;
416 ctx
->Const
.VertexProgram
.MaxNativeTexIndirections
= 0;
417 ctx
->Const
.VertexProgram
.MaxNativeAttribs
= 16;
418 ctx
->Const
.VertexProgram
.MaxNativeTemps
= 256;
419 ctx
->Const
.VertexProgram
.MaxNativeAddressRegs
= 1;
420 ctx
->Const
.VertexProgram
.MaxNativeParameters
= 1024;
421 ctx
->Const
.VertexProgram
.MaxEnvParams
=
422 MIN2(ctx
->Const
.VertexProgram
.MaxNativeParameters
,
423 ctx
->Const
.VertexProgram
.MaxEnvParams
);
425 ctx
->Const
.FragmentProgram
.MaxNativeInstructions
= 1024;
426 ctx
->Const
.FragmentProgram
.MaxNativeAluInstructions
= 1024;
427 ctx
->Const
.FragmentProgram
.MaxNativeTexInstructions
= 1024;
428 ctx
->Const
.FragmentProgram
.MaxNativeTexIndirections
= 1024;
429 ctx
->Const
.FragmentProgram
.MaxNativeAttribs
= 12;
430 ctx
->Const
.FragmentProgram
.MaxNativeTemps
= 256;
431 ctx
->Const
.FragmentProgram
.MaxNativeAddressRegs
= 0;
432 ctx
->Const
.FragmentProgram
.MaxNativeParameters
= 1024;
433 ctx
->Const
.FragmentProgram
.MaxEnvParams
=
434 MIN2(ctx
->Const
.FragmentProgram
.MaxNativeParameters
,
435 ctx
->Const
.FragmentProgram
.MaxEnvParams
);
437 /* Fragment shaders use real, 32-bit twos-complement integers for all
440 ctx
->Const
.FragmentProgram
.LowInt
.RangeMin
= 31;
441 ctx
->Const
.FragmentProgram
.LowInt
.RangeMax
= 30;
442 ctx
->Const
.FragmentProgram
.LowInt
.Precision
= 0;
443 ctx
->Const
.FragmentProgram
.HighInt
= ctx
->Const
.FragmentProgram
.LowInt
;
444 ctx
->Const
.FragmentProgram
.MediumInt
= ctx
->Const
.FragmentProgram
.LowInt
;
447 ctx
->Const
.FragmentProgram
.MaxAtomicCounters
= MAX_ATOMIC_COUNTERS
;
448 ctx
->Const
.VertexProgram
.MaxAtomicCounters
= MAX_ATOMIC_COUNTERS
;
449 ctx
->Const
.GeometryProgram
.MaxAtomicCounters
= MAX_ATOMIC_COUNTERS
;
450 ctx
->Const
.FragmentProgram
.MaxAtomicBuffers
= BRW_MAX_ABO
;
451 ctx
->Const
.VertexProgram
.MaxAtomicBuffers
= BRW_MAX_ABO
;
452 ctx
->Const
.GeometryProgram
.MaxAtomicBuffers
= BRW_MAX_ABO
;
453 ctx
->Const
.MaxCombinedAtomicBuffers
= 3 * BRW_MAX_ABO
;
456 /* Gen6 converts quads to polygon in beginning of 3D pipeline,
457 * but we're not sure how it's actually done for vertex order,
458 * that affect provoking vertex decision. Always use last vertex
459 * convention for quad primitive which works as expected for now.
462 ctx
->Const
.QuadsFollowProvokingVertexConvention
= false;
464 ctx
->Const
.NativeIntegers
= true;
465 ctx
->Const
.UniformBooleanTrue
= 1;
467 /* From the gen4 PRM, volume 4 page 127:
469 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field specifies
470 * the base address of the first element of the surface, computed in
471 * software by adding the surface base address to the byte offset of
472 * the element in the buffer."
474 * However, unaligned accesses are slower, so enforce buffer alignment.
476 ctx
->Const
.UniformBufferOffsetAlignment
= 16;
477 ctx
->Const
.TextureBufferOffsetAlignment
= 16;
480 ctx
->Const
.MaxVarying
= 32;
481 ctx
->Const
.VertexProgram
.MaxOutputComponents
= 128;
482 ctx
->Const
.GeometryProgram
.MaxInputComponents
= 64;
483 ctx
->Const
.GeometryProgram
.MaxOutputComponents
= 128;
484 ctx
->Const
.FragmentProgram
.MaxInputComponents
= 128;
487 /* We want the GLSL compiler to emit code that uses condition codes */
488 for (int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
489 ctx
->ShaderCompilerOptions
[i
].MaxIfDepth
= brw
->gen
< 6 ? 16 : UINT_MAX
;
490 ctx
->ShaderCompilerOptions
[i
].EmitCondCodes
= true;
491 ctx
->ShaderCompilerOptions
[i
].EmitNoNoise
= true;
492 ctx
->ShaderCompilerOptions
[i
].EmitNoMainReturn
= true;
493 ctx
->ShaderCompilerOptions
[i
].EmitNoIndirectInput
= true;
494 ctx
->ShaderCompilerOptions
[i
].EmitNoIndirectOutput
= true;
496 ctx
->ShaderCompilerOptions
[i
].EmitNoIndirectUniform
=
497 (i
== MESA_SHADER_FRAGMENT
);
498 ctx
->ShaderCompilerOptions
[i
].EmitNoIndirectTemp
=
499 (i
== MESA_SHADER_FRAGMENT
);
500 ctx
->ShaderCompilerOptions
[i
].LowerClipDistance
= true;
503 ctx
->ShaderCompilerOptions
[MESA_SHADER_VERTEX
].PreferDP4
= true;
507 * Process driconf (drirc) options, setting appropriate context flags.
509 * intelInitExtensions still pokes at optionCache directly, in order to
510 * avoid advertising various extensions. No flags are set, so it makes
511 * sense to continue doing that there.
514 brw_process_driconf_options(struct brw_context
*brw
)
516 struct gl_context
*ctx
= &brw
->ctx
;
518 driOptionCache
*options
= &brw
->optionCache
;
519 driParseConfigFiles(options
, &brw
->intelScreen
->optionCache
,
520 brw
->driContext
->driScreenPriv
->myNum
, "i965");
522 int bo_reuse_mode
= driQueryOptioni(options
, "bo_reuse");
523 switch (bo_reuse_mode
) {
524 case DRI_CONF_BO_REUSE_DISABLED
:
526 case DRI_CONF_BO_REUSE_ALL
:
527 intel_bufmgr_gem_enable_reuse(brw
->bufmgr
);
531 if (!driQueryOptionb(options
, "hiz")) {
532 brw
->has_hiz
= false;
533 /* On gen6, you can only do separate stencil with HIZ. */
535 brw
->has_separate_stencil
= false;
538 if (driQueryOptionb(options
, "always_flush_batch")) {
539 fprintf(stderr
, "flushing batchbuffer before/after each draw call\n");
540 brw
->always_flush_batch
= true;
543 if (driQueryOptionb(options
, "always_flush_cache")) {
544 fprintf(stderr
, "flushing GPU caches before/after each draw call\n");
545 brw
->always_flush_cache
= true;
548 if (driQueryOptionb(options
, "disable_throttling")) {
549 fprintf(stderr
, "disabling flush throttling\n");
550 brw
->disable_throttling
= true;
553 brw
->disable_derivative_optimization
=
554 driQueryOptionb(&brw
->optionCache
, "disable_derivative_optimization");
556 brw
->precompile
= driQueryOptionb(&brw
->optionCache
, "shader_precompile");
558 ctx
->Const
.ForceGLSLExtensionsWarn
=
559 driQueryOptionb(options
, "force_glsl_extensions_warn");
561 ctx
->Const
.DisableGLSLLineContinuations
=
562 driQueryOptionb(options
, "disable_glsl_line_continuations");
566 brwCreateContext(gl_api api
,
567 const struct gl_config
*mesaVis
,
568 __DRIcontext
*driContextPriv
,
569 unsigned major_version
,
570 unsigned minor_version
,
572 unsigned *dri_ctx_error
,
573 void *sharedContextPrivate
)
575 __DRIscreen
*sPriv
= driContextPriv
->driScreenPriv
;
576 struct gl_context
*shareCtx
= (struct gl_context
*) sharedContextPrivate
;
577 struct intel_screen
*screen
= sPriv
->driverPrivate
;
578 const struct brw_device_info
*devinfo
= screen
->devinfo
;
579 struct dd_function_table functions
;
580 struct gl_config visual
;
582 struct brw_context
*brw
= rzalloc(NULL
, struct brw_context
);
584 printf("%s: failed to alloc context\n", __FUNCTION__
);
585 *dri_ctx_error
= __DRI_CTX_ERROR_NO_MEMORY
;
589 driContextPriv
->driverPrivate
= brw
;
590 brw
->driContext
= driContextPriv
;
591 brw
->intelScreen
= screen
;
592 brw
->bufmgr
= screen
->bufmgr
;
594 brw
->gen
= devinfo
->gen
;
595 brw
->gt
= devinfo
->gt
;
596 brw
->is_g4x
= devinfo
->is_g4x
;
597 brw
->is_baytrail
= devinfo
->is_baytrail
;
598 brw
->is_haswell
= devinfo
->is_haswell
;
599 brw
->has_llc
= devinfo
->has_llc
;
600 brw
->has_hiz
= devinfo
->has_hiz_and_separate_stencil
;
601 brw
->has_separate_stencil
= devinfo
->has_hiz_and_separate_stencil
;
602 brw
->has_pln
= devinfo
->has_pln
;
603 brw
->has_compr4
= devinfo
->has_compr4
;
604 brw
->has_surface_tile_offset
= devinfo
->has_surface_tile_offset
;
605 brw
->has_negative_rhw_bug
= devinfo
->has_negative_rhw_bug
;
606 brw
->needs_unlit_centroid_workaround
=
607 devinfo
->needs_unlit_centroid_workaround
;
609 brw
->must_use_separate_stencil
= screen
->hw_must_use_separate_stencil
;
610 brw
->has_swizzling
= screen
->hw_has_swizzling
;
613 gen7_init_vtable_surface_functions(brw
);
614 gen7_init_vtable_sampler_functions(brw
);
615 brw
->vtbl
.emit_depth_stencil_hiz
= gen7_emit_depth_stencil_hiz
;
617 gen4_init_vtable_surface_functions(brw
);
618 gen4_init_vtable_sampler_functions(brw
);
619 brw
->vtbl
.emit_depth_stencil_hiz
= brw_emit_depth_stencil_hiz
;
622 brw_init_driver_functions(brw
, &functions
);
624 struct gl_context
*ctx
= &brw
->ctx
;
626 if (mesaVis
== NULL
) {
627 memset(&visual
, 0, sizeof visual
);
631 if (!_mesa_initialize_context(ctx
, api
, mesaVis
, shareCtx
, &functions
)) {
632 *dri_ctx_error
= __DRI_CTX_ERROR_NO_MEMORY
;
633 printf("%s: failed to init mesa context\n", __FUNCTION__
);
634 intelDestroyContext(driContextPriv
);
638 /* Initialize the software rasterizer and helper modules.
640 * As of GL 3.1 core, the gen4+ driver doesn't need the swrast context for
641 * software fallbacks (which we have to support on legacy GL to do weird
642 * glDrawPixels(), glBitmap(), and other functions).
644 if (api
!= API_OPENGL_CORE
&& api
!= API_OPENGLES2
) {
645 _swrast_CreateContext(ctx
);
648 _vbo_CreateContext(ctx
);
649 if (ctx
->swrast_context
) {
650 _tnl_CreateContext(ctx
);
651 TNL_CONTEXT(ctx
)->Driver
.RunPipeline
= _tnl_run_pipeline
;
652 _swsetup_CreateContext(ctx
);
654 /* Configure swrast to match hardware characteristics: */
655 _swrast_allow_pixel_fog(ctx
, false);
656 _swrast_allow_vertex_fog(ctx
, true);
659 _mesa_meta_init(ctx
);
661 brw_process_driconf_options(brw
);
662 brw_process_intel_debug_variable(brw
);
663 brw_initialize_context_constants(brw
);
665 /* Reinitialize the context point state. It depends on ctx->Const values. */
666 _mesa_init_point(ctx
);
668 intel_batchbuffer_init(brw
);
672 intelInitExtensions(ctx
);
677 /* Create a new hardware context. Using a hardware context means that
678 * our GPU state will be saved/restored on context switch, allowing us
679 * to assume that the GPU is in the same state we left it in.
681 * This is required for transform feedback buffer offsets, query objects,
682 * and also allows us to reduce how much state we have to emit.
684 brw
->hw_ctx
= drm_intel_gem_context_create(brw
->bufmgr
);
687 fprintf(stderr
, "Gen6+ requires Kernel 3.6 or later.\n");
688 intelDestroyContext(driContextPriv
);
693 brw_init_surface_formats(brw
);
695 if (brw
->is_g4x
|| brw
->gen
>= 5) {
696 brw
->CMD_VF_STATISTICS
= GM45_3DSTATE_VF_STATISTICS
;
697 brw
->CMD_PIPELINE_SELECT
= CMD_PIPELINE_SELECT_GM45
;
699 brw
->CMD_VF_STATISTICS
= GEN4_3DSTATE_VF_STATISTICS
;
700 brw
->CMD_PIPELINE_SELECT
= CMD_PIPELINE_SELECT_965
;
703 brw
->max_vs_threads
= devinfo
->max_vs_threads
;
704 brw
->max_gs_threads
= devinfo
->max_gs_threads
;
705 brw
->max_wm_threads
= devinfo
->max_wm_threads
;
706 brw
->urb
.size
= devinfo
->urb
.size
;
707 brw
->urb
.min_vs_entries
= devinfo
->urb
.min_vs_entries
;
708 brw
->urb
.max_vs_entries
= devinfo
->urb
.max_vs_entries
;
709 brw
->urb
.max_gs_entries
= devinfo
->urb
.max_gs_entries
;
711 /* Estimate the size of the mappable aperture into the GTT. There's an
712 * ioctl to get the whole GTT size, but not one to get the mappable subset.
713 * It turns out it's basically always 256MB, though some ancient hardware
716 uint32_t gtt_size
= 256 * 1024 * 1024;
718 /* We don't want to map two objects such that a memcpy between them would
719 * just fault one mapping in and then the other over and over forever. So
720 * we would need to divide the GTT size by 2. Additionally, some GTT is
721 * taken up by things like the framebuffer and the ringbuffer and such, so
722 * be more conservative.
724 brw
->max_gtt_map_object_size
= gtt_size
/ 4;
727 brw
->urb
.gen6_gs_previously_active
= false;
729 brw
->prim_restart
.in_progress
= false;
730 brw
->prim_restart
.enable_cut_index
= false;
733 brw
->curbe
.last_buf
= calloc(1, 4096);
734 brw
->curbe
.next_buf
= calloc(1, 4096);
737 ctx
->VertexProgram
._MaintainTnlProgram
= true;
738 ctx
->FragmentProgram
._MaintainTexEnvProgram
= true;
740 brw_draw_init( brw
);
742 if ((flags
& __DRI_CTX_FLAG_DEBUG
) != 0) {
743 /* Turn on some extra GL_ARB_debug_output generation. */
744 brw
->perf_debug
= true;
747 brw_fs_alloc_reg_sets(brw
);
748 brw_vec4_alloc_reg_set(brw
);
750 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
751 brw_init_shader_time(brw
);
753 _mesa_compute_version(ctx
);
755 /* Here we override context constants. We apply the overrides after
756 * calculation of the context version because we do not want the overridden
757 * constants to change the version.
759 brw_override_max_samples(brw
);
761 _mesa_initialize_dispatch_tables(ctx
);
762 _mesa_initialize_vbo_vtxfmt(ctx
);
768 intelDestroyContext(__DRIcontext
* driContextPriv
)
770 struct brw_context
*brw
=
771 (struct brw_context
*) driContextPriv
->driverPrivate
;
772 struct gl_context
*ctx
= &brw
->ctx
;
774 assert(brw
); /* should never be null */
778 /* Dump a final BMP in case the application doesn't call SwapBuffers */
779 if (INTEL_DEBUG
& DEBUG_AUB
) {
780 intel_batchbuffer_flush(brw
);
781 aub_dump_bmp(&brw
->ctx
);
784 _mesa_meta_free(&brw
->ctx
);
786 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
) {
787 /* Force a report. */
788 brw
->shader_time
.report_time
= 0;
790 brw_collect_and_report_shader_time(brw
);
791 brw_destroy_shader_time(brw
);
794 brw_destroy_state(brw
);
795 brw_draw_destroy(brw
);
797 drm_intel_bo_unreference(brw
->curbe
.curbe_bo
);
798 drm_intel_bo_unreference(brw
->vs
.base
.const_bo
);
799 drm_intel_bo_unreference(brw
->wm
.base
.const_bo
);
801 free(brw
->curbe
.last_buf
);
802 free(brw
->curbe
.next_buf
);
804 drm_intel_gem_context_destroy(brw
->hw_ctx
);
806 if (ctx
->swrast_context
) {
807 _swsetup_DestroyContext(&brw
->ctx
);
808 _tnl_DestroyContext(&brw
->ctx
);
810 _vbo_DestroyContext(&brw
->ctx
);
812 if (ctx
->swrast_context
)
813 _swrast_DestroyContext(&brw
->ctx
);
815 intel_batchbuffer_free(brw
);
817 drm_intel_bo_unreference(brw
->first_post_swapbuffers_batch
);
818 brw
->first_post_swapbuffers_batch
= NULL
;
820 driDestroyOptionCache(&brw
->optionCache
);
822 /* free the Mesa context */
823 _mesa_free_context_data(&brw
->ctx
);
826 driContextPriv
->driverPrivate
= NULL
;
830 intelUnbindContext(__DRIcontext
* driContextPriv
)
832 /* Unset current context and dispath table */
833 _mesa_make_current(NULL
, NULL
, NULL
);
839 * Fixes up the context for GLES23 with our default-to-sRGB-capable behavior
840 * on window system framebuffers.
842 * Desktop GL is fairly reasonable in its handling of sRGB: You can ask if
843 * your renderbuffer can do sRGB encode, and you can flip a switch that does
844 * sRGB encode if the renderbuffer can handle it. You can ask specifically
845 * for a visual where you're guaranteed to be capable, but it turns out that
846 * everyone just makes all their ARGB8888 visuals capable and doesn't offer
847 * incapable ones, becuase there's no difference between the two in resources
848 * used. Applications thus get built that accidentally rely on the default
849 * visual choice being sRGB, so we make ours sRGB capable. Everything sounds
852 * But for GLES2/3, they decided that it was silly to not turn on sRGB encode
853 * for sRGB renderbuffers you made with the GL_EXT_texture_sRGB equivalent.
854 * So they removed the enable knob and made it "if the renderbuffer is sRGB
855 * capable, do sRGB encode". Then, for your window system renderbuffers, you
856 * can ask for sRGB visuals and get sRGB encode, or not ask for sRGB visuals
857 * and get no sRGB encode (assuming that both kinds of visual are available).
858 * Thus our choice to support sRGB by default on our visuals for desktop would
859 * result in broken rendering of GLES apps that aren't expecting sRGB encode.
861 * Unfortunately, renderbuffer setup happens before a context is created. So
862 * in intel_screen.c we always set up sRGB, and here, if you're a GLES2/3
863 * context (without an sRGB visual, though we don't have sRGB visuals exposed
864 * yet), we go turn that back off before anyone finds out.
867 intel_gles3_srgb_workaround(struct brw_context
*brw
,
868 struct gl_framebuffer
*fb
)
870 struct gl_context
*ctx
= &brw
->ctx
;
872 if (_mesa_is_desktop_gl(ctx
) || !fb
->Visual
.sRGBCapable
)
875 /* Some day when we support the sRGB capable bit on visuals available for
876 * GLES, we'll need to respect that and not disable things here.
878 fb
->Visual
.sRGBCapable
= false;
879 for (int i
= 0; i
< BUFFER_COUNT
; i
++) {
880 if (fb
->Attachment
[i
].Renderbuffer
&&
881 fb
->Attachment
[i
].Renderbuffer
->Format
== MESA_FORMAT_SARGB8
) {
882 fb
->Attachment
[i
].Renderbuffer
->Format
= MESA_FORMAT_ARGB8888
;
888 intelMakeCurrent(__DRIcontext
* driContextPriv
,
889 __DRIdrawable
* driDrawPriv
,
890 __DRIdrawable
* driReadPriv
)
892 struct brw_context
*brw
;
893 GET_CURRENT_CONTEXT(curCtx
);
896 brw
= (struct brw_context
*) driContextPriv
->driverPrivate
;
900 /* According to the glXMakeCurrent() man page: "Pending commands to
901 * the previous context, if any, are flushed before it is released."
902 * But only flush if we're actually changing contexts.
904 if (brw_context(curCtx
) && brw_context(curCtx
) != brw
) {
908 if (driContextPriv
) {
909 struct gl_context
*ctx
= &brw
->ctx
;
910 struct gl_framebuffer
*fb
, *readFb
;
912 if (driDrawPriv
== NULL
&& driReadPriv
== NULL
) {
913 fb
= _mesa_get_incomplete_framebuffer();
914 readFb
= _mesa_get_incomplete_framebuffer();
916 fb
= driDrawPriv
->driverPrivate
;
917 readFb
= driReadPriv
->driverPrivate
;
918 driContextPriv
->dri2
.draw_stamp
= driDrawPriv
->dri2
.stamp
- 1;
919 driContextPriv
->dri2
.read_stamp
= driReadPriv
->dri2
.stamp
- 1;
922 /* The sRGB workaround changes the renderbuffer's format. We must change
923 * the format before the renderbuffer's miptree get's allocated, otherwise
924 * the formats of the renderbuffer and its miptree will differ.
926 intel_gles3_srgb_workaround(brw
, fb
);
927 intel_gles3_srgb_workaround(brw
, readFb
);
929 intel_prepare_render(brw
);
930 _mesa_make_current(ctx
, fb
, readFb
);
932 _mesa_make_current(NULL
, NULL
, NULL
);
939 intel_resolve_for_dri2_flush(struct brw_context
*brw
,
940 __DRIdrawable
*drawable
)
943 /* MSAA and fast color clear are not supported, so don't waste time
944 * checking whether a resolve is needed.
949 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
950 struct intel_renderbuffer
*rb
;
952 /* Usually, only the back buffer will need to be downsampled. However,
953 * the front buffer will also need it if the user has rendered into it.
955 static const gl_buffer_index buffers
[2] = {
960 for (int i
= 0; i
< 2; ++i
) {
961 rb
= intel_get_renderbuffer(fb
, buffers
[i
]);
962 if (rb
== NULL
|| rb
->mt
== NULL
)
964 if (rb
->mt
->num_samples
<= 1)
965 intel_miptree_resolve_color(brw
, rb
->mt
);
967 intel_miptree_downsample(brw
, rb
->mt
);
972 intel_bits_per_pixel(const struct intel_renderbuffer
*rb
)
974 return _mesa_get_format_bytes(intel_rb_format(rb
)) * 8;
978 intel_query_dri2_buffers(struct brw_context
*brw
,
979 __DRIdrawable
*drawable
,
980 __DRIbuffer
**buffers
,
984 intel_process_dri2_buffer(struct brw_context
*brw
,
985 __DRIdrawable
*drawable
,
987 struct intel_renderbuffer
*rb
,
988 const char *buffer_name
);
991 intel_update_renderbuffers(__DRIcontext
*context
, __DRIdrawable
*drawable
)
993 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
994 struct intel_renderbuffer
*rb
;
995 struct brw_context
*brw
= context
->driverPrivate
;
996 __DRIbuffer
*buffers
= NULL
;
998 const char *region_name
;
1000 /* Set this up front, so that in case our buffers get invalidated
1001 * while we're getting new buffers, we don't clobber the stamp and
1002 * thus ignore the invalidate. */
1003 drawable
->lastStamp
= drawable
->dri2
.stamp
;
1005 if (unlikely(INTEL_DEBUG
& DEBUG_DRI
))
1006 fprintf(stderr
, "enter %s, drawable %p\n", __func__
, drawable
);
1008 intel_query_dri2_buffers(brw
, drawable
, &buffers
, &count
);
1010 if (buffers
== NULL
)
1013 for (i
= 0; i
< count
; i
++) {
1014 switch (buffers
[i
].attachment
) {
1015 case __DRI_BUFFER_FRONT_LEFT
:
1016 rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1017 region_name
= "dri2 front buffer";
1020 case __DRI_BUFFER_FAKE_FRONT_LEFT
:
1021 rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1022 region_name
= "dri2 fake front buffer";
1025 case __DRI_BUFFER_BACK_LEFT
:
1026 rb
= intel_get_renderbuffer(fb
, BUFFER_BACK_LEFT
);
1027 region_name
= "dri2 back buffer";
1030 case __DRI_BUFFER_DEPTH
:
1031 case __DRI_BUFFER_HIZ
:
1032 case __DRI_BUFFER_DEPTH_STENCIL
:
1033 case __DRI_BUFFER_STENCIL
:
1034 case __DRI_BUFFER_ACCUM
:
1037 "unhandled buffer attach event, attachment type %d\n",
1038 buffers
[i
].attachment
);
1042 intel_process_dri2_buffer(brw
, drawable
, &buffers
[i
], rb
, region_name
);
1045 driUpdateFramebufferSize(&brw
->ctx
, drawable
);
1049 * intel_prepare_render should be called anywhere that curent read/drawbuffer
1050 * state is required.
1053 intel_prepare_render(struct brw_context
*brw
)
1055 __DRIcontext
*driContext
= brw
->driContext
;
1056 __DRIdrawable
*drawable
;
1058 drawable
= driContext
->driDrawablePriv
;
1059 if (drawable
&& drawable
->dri2
.stamp
!= driContext
->dri2
.draw_stamp
) {
1060 if (drawable
->lastStamp
!= drawable
->dri2
.stamp
)
1061 intel_update_renderbuffers(driContext
, drawable
);
1062 driContext
->dri2
.draw_stamp
= drawable
->dri2
.stamp
;
1065 drawable
= driContext
->driReadablePriv
;
1066 if (drawable
&& drawable
->dri2
.stamp
!= driContext
->dri2
.read_stamp
) {
1067 if (drawable
->lastStamp
!= drawable
->dri2
.stamp
)
1068 intel_update_renderbuffers(driContext
, drawable
);
1069 driContext
->dri2
.read_stamp
= drawable
->dri2
.stamp
;
1072 /* If we're currently rendering to the front buffer, the rendering
1073 * that will happen next will probably dirty the front buffer. So
1074 * mark it as dirty here.
1076 if (brw
->is_front_buffer_rendering
)
1077 brw
->front_buffer_dirty
= true;
1079 /* Wait for the swapbuffers before the one we just emitted, so we
1080 * don't get too many swaps outstanding for apps that are GPU-heavy
1081 * but not CPU-heavy.
1083 * We're using intelDRI2Flush (called from the loader before
1084 * swapbuffer) and glFlush (for front buffer rendering) as the
1085 * indicator that a frame is done and then throttle when we get
1086 * here as we prepare to render the next frame. At this point for
1087 * round trips for swap/copy and getting new buffers are done and
1088 * we'll spend less time waiting on the GPU.
1090 * Unfortunately, we don't have a handle to the batch containing
1091 * the swap, and getting our hands on that doesn't seem worth it,
1092 * so we just us the first batch we emitted after the last swap.
1094 if (brw
->need_throttle
&& brw
->first_post_swapbuffers_batch
) {
1095 if (!brw
->disable_throttling
)
1096 drm_intel_bo_wait_rendering(brw
->first_post_swapbuffers_batch
);
1097 drm_intel_bo_unreference(brw
->first_post_swapbuffers_batch
);
1098 brw
->first_post_swapbuffers_batch
= NULL
;
1099 brw
->need_throttle
= false;
1104 * \brief Query DRI2 to obtain a DRIdrawable's buffers.
1106 * To determine which DRI buffers to request, examine the renderbuffers
1107 * attached to the drawable's framebuffer. Then request the buffers with
1108 * DRI2GetBuffers() or DRI2GetBuffersWithFormat().
1110 * This is called from intel_update_renderbuffers().
1112 * \param drawable Drawable whose buffers are queried.
1113 * \param buffers [out] List of buffers returned by DRI2 query.
1114 * \param buffer_count [out] Number of buffers returned.
1116 * \see intel_update_renderbuffers()
1117 * \see DRI2GetBuffers()
1118 * \see DRI2GetBuffersWithFormat()
1121 intel_query_dri2_buffers(struct brw_context
*brw
,
1122 __DRIdrawable
*drawable
,
1123 __DRIbuffer
**buffers
,
1126 __DRIscreen
*screen
= brw
->intelScreen
->driScrnPriv
;
1127 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1129 unsigned attachments
[8];
1131 struct intel_renderbuffer
*front_rb
;
1132 struct intel_renderbuffer
*back_rb
;
1134 front_rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1135 back_rb
= intel_get_renderbuffer(fb
, BUFFER_BACK_LEFT
);
1137 memset(attachments
, 0, sizeof(attachments
));
1138 if ((brw
->is_front_buffer_rendering
||
1139 brw
->is_front_buffer_reading
||
1140 !back_rb
) && front_rb
) {
1141 /* If a fake front buffer is in use, then querying for
1142 * __DRI_BUFFER_FRONT_LEFT will cause the server to copy the image from
1143 * the real front buffer to the fake front buffer. So before doing the
1144 * query, we need to make sure all the pending drawing has landed in the
1145 * real front buffer.
1147 intel_batchbuffer_flush(brw
);
1148 intel_flush_front(&brw
->ctx
);
1150 attachments
[i
++] = __DRI_BUFFER_FRONT_LEFT
;
1151 attachments
[i
++] = intel_bits_per_pixel(front_rb
);
1152 } else if (front_rb
&& brw
->front_buffer_dirty
) {
1153 /* We have pending front buffer rendering, but we aren't querying for a
1154 * front buffer. If the front buffer we have is a fake front buffer,
1155 * the X server is going to throw it away when it processes the query.
1156 * So before doing the query, make sure all the pending drawing has
1157 * landed in the real front buffer.
1159 intel_batchbuffer_flush(brw
);
1160 intel_flush_front(&brw
->ctx
);
1164 attachments
[i
++] = __DRI_BUFFER_BACK_LEFT
;
1165 attachments
[i
++] = intel_bits_per_pixel(back_rb
);
1168 assert(i
<= ARRAY_SIZE(attachments
));
1170 *buffers
= screen
->dri2
.loader
->getBuffersWithFormat(drawable
,
1175 drawable
->loaderPrivate
);
1179 * \brief Assign a DRI buffer's DRM region to a renderbuffer.
1181 * This is called from intel_update_renderbuffers().
1184 * DRI buffers whose attachment point is DRI2BufferStencil or
1185 * DRI2BufferDepthStencil are handled as special cases.
1187 * \param buffer_name is a human readable name, such as "dri2 front buffer",
1188 * that is passed to intel_region_alloc_for_handle().
1190 * \see intel_update_renderbuffers()
1191 * \see intel_region_alloc_for_handle()
1194 intel_process_dri2_buffer(struct brw_context
*brw
,
1195 __DRIdrawable
*drawable
,
1196 __DRIbuffer
*buffer
,
1197 struct intel_renderbuffer
*rb
,
1198 const char *buffer_name
)
1200 struct intel_region
*region
= NULL
;
1205 unsigned num_samples
= rb
->Base
.Base
.NumSamples
;
1207 /* We try to avoid closing and reopening the same BO name, because the first
1208 * use of a mapping of the buffer involves a bunch of page faulting which is
1209 * moderately expensive.
1211 if (num_samples
== 0) {
1214 rb
->mt
->region
->name
== buffer
->name
)
1218 rb
->mt
->singlesample_mt
&&
1219 rb
->mt
->singlesample_mt
->region
&&
1220 rb
->mt
->singlesample_mt
->region
->name
== buffer
->name
)
1224 if (unlikely(INTEL_DEBUG
& DEBUG_DRI
)) {
1226 "attaching buffer %d, at %d, cpp %d, pitch %d\n",
1227 buffer
->name
, buffer
->attachment
,
1228 buffer
->cpp
, buffer
->pitch
);
1231 intel_miptree_release(&rb
->mt
);
1232 region
= intel_region_alloc_for_handle(brw
->intelScreen
,
1242 rb
->mt
= intel_miptree_create_for_dri2_buffer(brw
,
1244 intel_rb_format(rb
),
1247 intel_region_release(®ion
);