2 Copyright 2003 VMware, Inc.
3 Copyright (C) Intel Corp. 2006. All Rights Reserved.
4 Intel funded Tungsten Graphics 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 <keithw@vmware.com>
34 #include "compiler/nir/nir.h"
35 #include "main/api_exec.h"
36 #include "main/context.h"
37 #include "main/fbobject.h"
38 #include "main/extensions.h"
39 #include "main/imports.h"
40 #include "main/macros.h"
41 #include "main/points.h"
42 #include "main/version.h"
43 #include "main/vtxfmt.h"
44 #include "main/texobj.h"
45 #include "main/framebuffer.h"
46 #include "main/stencil.h"
47 #include "main/state.h"
51 #include "drivers/common/driverfuncs.h"
52 #include "drivers/common/meta.h"
55 #include "brw_context.h"
56 #include "brw_defines.h"
57 #include "brw_blorp.h"
59 #include "brw_state.h"
61 #include "intel_batchbuffer.h"
62 #include "intel_buffer_objects.h"
63 #include "intel_buffers.h"
64 #include "intel_fbo.h"
65 #include "intel_mipmap_tree.h"
66 #include "intel_pixel.h"
67 #include "intel_image.h"
68 #include "intel_tex.h"
69 #include "intel_tex_obj.h"
71 #include "swrast_setup/swrast_setup.h"
73 #include "tnl/t_pipeline.h"
74 #include "util/ralloc.h"
75 #include "util/debug.h"
76 #include "util/disk_cache.h"
79 #include "common/gen_defines.h"
81 /***************************************
82 * Mesa's Driver Functions
83 ***************************************/
85 const char *const brw_vendor_string
= "Intel Open Source Technology Center";
88 get_bsw_model(const struct intel_screen
*screen
)
90 switch (screen
->eu_total
) {
101 brw_get_renderer_string(const struct intel_screen
*screen
)
104 static char buffer
[128];
107 switch (screen
->deviceID
) {
109 #define CHIPSET(id, symbol, str) case id: chipset = str; break;
110 #include "pci_ids/i965_pci_ids.h"
112 chipset
= "Unknown Intel Chipset";
116 /* Braswell branding is funny, so we have to fix it up here */
117 if (screen
->deviceID
== 0x22B1) {
118 bsw
= strdup(chipset
);
119 char *needle
= strstr(bsw
, "XXX");
121 memcpy(needle
, get_bsw_model(screen
), 3);
126 (void) driGetRendererString(buffer
, chipset
, 0);
131 static const GLubyte
*
132 intel_get_string(struct gl_context
* ctx
, GLenum name
)
134 const struct brw_context
*const brw
= brw_context(ctx
);
138 return (GLubyte
*) brw_vendor_string
;
142 (GLubyte
*) brw_get_renderer_string(brw
->screen
);
150 intel_viewport(struct gl_context
*ctx
)
152 struct brw_context
*brw
= brw_context(ctx
);
153 __DRIcontext
*driContext
= brw
->driContext
;
155 if (_mesa_is_winsys_fbo(ctx
->DrawBuffer
)) {
156 if (driContext
->driDrawablePriv
)
157 dri2InvalidateDrawable(driContext
->driDrawablePriv
);
158 if (driContext
->driReadablePriv
)
159 dri2InvalidateDrawable(driContext
->driReadablePriv
);
164 intel_update_framebuffer(struct gl_context
*ctx
,
165 struct gl_framebuffer
*fb
)
167 struct brw_context
*brw
= brw_context(ctx
);
169 /* Quantize the derived default number of samples
171 fb
->DefaultGeometry
._NumSamples
=
172 intel_quantize_num_samples(brw
->screen
,
173 fb
->DefaultGeometry
.NumSamples
);
177 intel_update_state(struct gl_context
* ctx
)
179 GLuint new_state
= ctx
->NewState
;
180 struct brw_context
*brw
= brw_context(ctx
);
182 if (ctx
->swrast_context
)
183 _swrast_InvalidateState(ctx
, new_state
);
185 brw
->NewGLState
|= new_state
;
187 if (new_state
& (_NEW_SCISSOR
| _NEW_BUFFERS
| _NEW_VIEWPORT
))
188 _mesa_update_draw_buffer_bounds(ctx
, ctx
->DrawBuffer
);
190 if (new_state
& (_NEW_STENCIL
| _NEW_BUFFERS
)) {
191 brw
->stencil_enabled
= _mesa_stencil_is_enabled(ctx
);
192 brw
->stencil_two_sided
= _mesa_stencil_is_two_sided(ctx
);
193 brw
->stencil_write_enabled
=
194 _mesa_stencil_is_write_enabled(ctx
, brw
->stencil_two_sided
);
197 if (new_state
& _NEW_POLYGON
)
198 brw
->polygon_front_bit
= _mesa_polygon_get_front_bit(ctx
);
200 if (new_state
& _NEW_BUFFERS
) {
201 intel_update_framebuffer(ctx
, ctx
->DrawBuffer
);
202 if (ctx
->DrawBuffer
!= ctx
->ReadBuffer
)
203 intel_update_framebuffer(ctx
, ctx
->ReadBuffer
);
207 #define flushFront(screen) ((screen)->image.loader ? (screen)->image.loader->flushFrontBuffer : (screen)->dri2.loader->flushFrontBuffer)
210 intel_flush_front(struct gl_context
*ctx
)
212 struct brw_context
*brw
= brw_context(ctx
);
213 __DRIcontext
*driContext
= brw
->driContext
;
214 __DRIdrawable
*driDrawable
= driContext
->driDrawablePriv
;
215 __DRIscreen
*const dri_screen
= brw
->screen
->driScrnPriv
;
217 if (brw
->front_buffer_dirty
&& _mesa_is_winsys_fbo(ctx
->DrawBuffer
)) {
218 if (flushFront(dri_screen
) && driDrawable
&&
219 driDrawable
->loaderPrivate
) {
221 /* Resolve before flushing FAKE_FRONT_LEFT to FRONT_LEFT.
223 * This potentially resolves both front and back buffer. It
224 * is unnecessary to resolve the back, but harms nothing except
225 * performance. And no one cares about front-buffer render
228 intel_resolve_for_dri2_flush(brw
, driDrawable
);
229 intel_batchbuffer_flush(brw
);
231 flushFront(dri_screen
)(driDrawable
, driDrawable
->loaderPrivate
);
233 /* We set the dirty bit in intel_prepare_render() if we're
234 * front buffer rendering once we get there.
236 brw
->front_buffer_dirty
= false;
242 intel_glFlush(struct gl_context
*ctx
)
244 struct brw_context
*brw
= brw_context(ctx
);
246 intel_batchbuffer_flush(brw
);
247 intel_flush_front(ctx
);
249 brw
->need_flush_throttle
= true;
253 intel_finish(struct gl_context
* ctx
)
255 struct brw_context
*brw
= brw_context(ctx
);
259 if (brw
->batch
.last_bo
)
260 brw_bo_wait_rendering(brw
->batch
.last_bo
);
264 brw_init_driver_functions(struct brw_context
*brw
,
265 struct dd_function_table
*functions
)
267 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
269 _mesa_init_driver_functions(functions
);
271 /* GLX uses DRI2 invalidate events to handle window resizing.
272 * Unfortunately, EGL does not - libEGL is written in XCB (not Xlib),
273 * which doesn't provide a mechanism for snooping the event queues.
275 * So EGL still relies on viewport hacks to handle window resizing.
276 * This should go away with DRI3000.
278 if (!brw
->driContext
->driScreenPriv
->dri2
.useInvalidate
)
279 functions
->Viewport
= intel_viewport
;
281 functions
->Flush
= intel_glFlush
;
282 functions
->Finish
= intel_finish
;
283 functions
->GetString
= intel_get_string
;
284 functions
->UpdateState
= intel_update_state
;
286 intelInitTextureFuncs(functions
);
287 intelInitTextureImageFuncs(functions
);
288 intelInitTextureCopyImageFuncs(functions
);
289 intelInitCopyImageFuncs(functions
);
290 intelInitClearFuncs(functions
);
291 intelInitBufferFuncs(functions
);
292 intelInitPixelFuncs(functions
);
293 intelInitBufferObjectFuncs(functions
);
294 brw_init_syncobj_functions(functions
);
295 brw_init_object_purgeable_functions(functions
);
297 brwInitFragProgFuncs( functions
);
298 brw_init_common_queryobj_functions(functions
);
299 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
)
300 hsw_init_queryobj_functions(functions
);
301 else if (devinfo
->gen
>= 6)
302 gen6_init_queryobj_functions(functions
);
304 gen4_init_queryobj_functions(functions
);
305 brw_init_compute_functions(functions
);
306 brw_init_conditional_render_functions(functions
);
308 functions
->GenerateMipmap
= brw_generate_mipmap
;
310 functions
->QueryInternalFormat
= brw_query_internal_format
;
312 functions
->NewTransformFeedback
= brw_new_transform_feedback
;
313 functions
->DeleteTransformFeedback
= brw_delete_transform_feedback
;
314 if (can_do_mi_math_and_lrr(brw
->screen
)) {
315 functions
->BeginTransformFeedback
= hsw_begin_transform_feedback
;
316 functions
->EndTransformFeedback
= hsw_end_transform_feedback
;
317 functions
->PauseTransformFeedback
= hsw_pause_transform_feedback
;
318 functions
->ResumeTransformFeedback
= hsw_resume_transform_feedback
;
319 } else if (devinfo
->gen
>= 7) {
320 functions
->BeginTransformFeedback
= gen7_begin_transform_feedback
;
321 functions
->EndTransformFeedback
= gen7_end_transform_feedback
;
322 functions
->PauseTransformFeedback
= gen7_pause_transform_feedback
;
323 functions
->ResumeTransformFeedback
= gen7_resume_transform_feedback
;
324 functions
->GetTransformFeedbackVertexCount
=
325 brw_get_transform_feedback_vertex_count
;
327 functions
->BeginTransformFeedback
= brw_begin_transform_feedback
;
328 functions
->EndTransformFeedback
= brw_end_transform_feedback
;
329 functions
->PauseTransformFeedback
= brw_pause_transform_feedback
;
330 functions
->ResumeTransformFeedback
= brw_resume_transform_feedback
;
331 functions
->GetTransformFeedbackVertexCount
=
332 brw_get_transform_feedback_vertex_count
;
335 if (devinfo
->gen
>= 6)
336 functions
->GetSamplePosition
= gen6_get_sample_position
;
338 /* GL_ARB_get_program_binary */
339 brw_program_binary_init(brw
->screen
->deviceID
);
340 functions
->GetProgramBinaryDriverSHA1
= brw_get_program_binary_driver_sha1
;
341 functions
->ProgramBinarySerializeDriverBlob
= brw_program_serialize_nir
;
342 functions
->ProgramBinaryDeserializeDriverBlob
=
343 brw_deserialize_program_binary
;
347 brw_initialize_context_constants(struct brw_context
*brw
)
349 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
350 struct gl_context
*ctx
= &brw
->ctx
;
351 const struct brw_compiler
*compiler
= brw
->screen
->compiler
;
353 const bool stage_exists
[MESA_SHADER_STAGES
] = {
354 [MESA_SHADER_VERTEX
] = true,
355 [MESA_SHADER_TESS_CTRL
] = devinfo
->gen
>= 7,
356 [MESA_SHADER_TESS_EVAL
] = devinfo
->gen
>= 7,
357 [MESA_SHADER_GEOMETRY
] = devinfo
->gen
>= 6,
358 [MESA_SHADER_FRAGMENT
] = true,
359 [MESA_SHADER_COMPUTE
] =
360 (_mesa_is_desktop_gl(ctx
) &&
361 ctx
->Const
.MaxComputeWorkGroupSize
[0] >= 1024) ||
362 (ctx
->API
== API_OPENGLES2
&&
363 ctx
->Const
.MaxComputeWorkGroupSize
[0] >= 128),
366 unsigned num_stages
= 0;
367 for (int i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
372 unsigned max_samplers
=
373 devinfo
->gen
>= 8 || devinfo
->is_haswell
? BRW_MAX_TEX_UNIT
: 16;
375 ctx
->Const
.MaxDualSourceDrawBuffers
= 1;
376 ctx
->Const
.MaxDrawBuffers
= BRW_MAX_DRAW_BUFFERS
;
377 ctx
->Const
.MaxCombinedShaderOutputResources
=
378 MAX_IMAGE_UNITS
+ BRW_MAX_DRAW_BUFFERS
;
380 /* The timestamp register we can read for glGetTimestamp() is
381 * sometimes only 32 bits, before scaling to nanoseconds (depending
384 * Once scaled to nanoseconds the timestamp would roll over at a
385 * non-power-of-two, so an application couldn't use
386 * GL_QUERY_COUNTER_BITS to handle rollover correctly. Instead, we
387 * report 36 bits and truncate at that (rolling over 5 times as
388 * often as the HW counter), and when the 32-bit counter rolls
389 * over, it happens to also be at a rollover in the reported value
390 * from near (1<<36) to 0.
392 * The low 32 bits rolls over in ~343 seconds. Our 36-bit result
393 * rolls over every ~69 seconds.
395 ctx
->Const
.QueryCounterBits
.Timestamp
= 36;
397 ctx
->Const
.MaxTextureCoordUnits
= 8; /* Mesa limit */
398 ctx
->Const
.MaxImageUnits
= MAX_IMAGE_UNITS
;
399 if (devinfo
->gen
>= 7) {
400 ctx
->Const
.MaxRenderbufferSize
= 16384;
401 ctx
->Const
.MaxTextureLevels
= MIN2(15 /* 16384 */, MAX_TEXTURE_LEVELS
);
402 ctx
->Const
.MaxCubeTextureLevels
= 15; /* 16384 */
404 ctx
->Const
.MaxRenderbufferSize
= 8192;
405 ctx
->Const
.MaxTextureLevels
= MIN2(14 /* 8192 */, MAX_TEXTURE_LEVELS
);
406 ctx
->Const
.MaxCubeTextureLevels
= 14; /* 8192 */
408 ctx
->Const
.Max3DTextureLevels
= 12; /* 2048 */
409 ctx
->Const
.MaxArrayTextureLayers
= devinfo
->gen
>= 7 ? 2048 : 512;
410 ctx
->Const
.MaxTextureMbytes
= 1536;
411 ctx
->Const
.MaxTextureRectSize
= devinfo
->gen
>= 7 ? 16384 : 8192;
412 ctx
->Const
.MaxTextureMaxAnisotropy
= 16.0;
413 ctx
->Const
.MaxTextureLodBias
= 15.0;
414 ctx
->Const
.StripTextureBorder
= true;
415 if (devinfo
->gen
>= 7) {
416 ctx
->Const
.MaxProgramTextureGatherComponents
= 4;
417 ctx
->Const
.MinProgramTextureGatherOffset
= -32;
418 ctx
->Const
.MaxProgramTextureGatherOffset
= 31;
419 } else if (devinfo
->gen
== 6) {
420 ctx
->Const
.MaxProgramTextureGatherComponents
= 1;
421 ctx
->Const
.MinProgramTextureGatherOffset
= -8;
422 ctx
->Const
.MaxProgramTextureGatherOffset
= 7;
425 ctx
->Const
.MaxUniformBlockSize
= 65536;
427 for (int i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
428 struct gl_program_constants
*prog
= &ctx
->Const
.Program
[i
];
430 if (!stage_exists
[i
])
433 prog
->MaxTextureImageUnits
= max_samplers
;
435 prog
->MaxUniformBlocks
= BRW_MAX_UBO
;
436 prog
->MaxCombinedUniformComponents
=
437 prog
->MaxUniformComponents
+
438 ctx
->Const
.MaxUniformBlockSize
/ 4 * prog
->MaxUniformBlocks
;
440 prog
->MaxAtomicCounters
= MAX_ATOMIC_COUNTERS
;
441 prog
->MaxAtomicBuffers
= BRW_MAX_ABO
;
442 prog
->MaxImageUniforms
= compiler
->scalar_stage
[i
] ? BRW_MAX_IMAGES
: 0;
443 prog
->MaxShaderStorageBlocks
= BRW_MAX_SSBO
;
446 ctx
->Const
.MaxTextureUnits
=
447 MIN2(ctx
->Const
.MaxTextureCoordUnits
,
448 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxTextureImageUnits
);
450 ctx
->Const
.MaxUniformBufferBindings
= num_stages
* BRW_MAX_UBO
;
451 ctx
->Const
.MaxCombinedUniformBlocks
= num_stages
* BRW_MAX_UBO
;
452 ctx
->Const
.MaxCombinedAtomicBuffers
= num_stages
* BRW_MAX_ABO
;
453 ctx
->Const
.MaxCombinedShaderStorageBlocks
= num_stages
* BRW_MAX_SSBO
;
454 ctx
->Const
.MaxShaderStorageBufferBindings
= num_stages
* BRW_MAX_SSBO
;
455 ctx
->Const
.MaxCombinedTextureImageUnits
= num_stages
* max_samplers
;
456 ctx
->Const
.MaxCombinedImageUniforms
= num_stages
* BRW_MAX_IMAGES
;
459 /* Hardware only supports a limited number of transform feedback buffers.
460 * So we need to override the Mesa default (which is based only on software
463 ctx
->Const
.MaxTransformFeedbackBuffers
= BRW_MAX_SOL_BUFFERS
;
465 /* On Gen6, in the worst case, we use up one binding table entry per
466 * transform feedback component (see comments above the definition of
467 * BRW_MAX_SOL_BINDINGS, in brw_context.h), so we need to advertise a value
468 * for MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS equal to
469 * BRW_MAX_SOL_BINDINGS.
471 * In "separate components" mode, we need to divide this value by
472 * BRW_MAX_SOL_BUFFERS, so that the total number of binding table entries
473 * used up by all buffers will not exceed BRW_MAX_SOL_BINDINGS.
475 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
= BRW_MAX_SOL_BINDINGS
;
476 ctx
->Const
.MaxTransformFeedbackSeparateComponents
=
477 BRW_MAX_SOL_BINDINGS
/ BRW_MAX_SOL_BUFFERS
;
479 ctx
->Const
.AlwaysUseGetTransformFeedbackVertexCount
=
480 !can_do_mi_math_and_lrr(brw
->screen
);
483 const int *msaa_modes
= intel_supported_msaa_modes(brw
->screen
);
484 const int clamp_max_samples
=
485 driQueryOptioni(&brw
->optionCache
, "clamp_max_samples");
487 if (clamp_max_samples
< 0) {
488 max_samples
= msaa_modes
[0];
490 /* Select the largest supported MSAA mode that does not exceed
494 for (int i
= 0; msaa_modes
[i
] != 0; ++i
) {
495 if (msaa_modes
[i
] <= clamp_max_samples
) {
496 max_samples
= msaa_modes
[i
];
502 ctx
->Const
.MaxSamples
= max_samples
;
503 ctx
->Const
.MaxColorTextureSamples
= max_samples
;
504 ctx
->Const
.MaxDepthTextureSamples
= max_samples
;
505 ctx
->Const
.MaxIntegerSamples
= max_samples
;
506 ctx
->Const
.MaxImageSamples
= 0;
508 /* gen6_set_sample_maps() sets SampleMap{2,4,8}x variables which are used
509 * to map indices of rectangular grid to sample numbers within a pixel.
510 * These variables are used by GL_EXT_framebuffer_multisample_blit_scaled
511 * extension implementation. For more details see the comment above
512 * gen6_set_sample_maps() definition.
514 gen6_set_sample_maps(ctx
);
516 ctx
->Const
.MinLineWidth
= 1.0;
517 ctx
->Const
.MinLineWidthAA
= 1.0;
518 if (devinfo
->gen
>= 6) {
519 ctx
->Const
.MaxLineWidth
= 7.375;
520 ctx
->Const
.MaxLineWidthAA
= 7.375;
521 ctx
->Const
.LineWidthGranularity
= 0.125;
523 ctx
->Const
.MaxLineWidth
= 7.0;
524 ctx
->Const
.MaxLineWidthAA
= 7.0;
525 ctx
->Const
.LineWidthGranularity
= 0.5;
528 /* For non-antialiased lines, we have to round the line width to the
529 * nearest whole number. Make sure that we don't advertise a line
530 * width that, when rounded, will be beyond the actual hardware
533 assert(roundf(ctx
->Const
.MaxLineWidth
) <= ctx
->Const
.MaxLineWidth
);
535 ctx
->Const
.MinPointSize
= 1.0;
536 ctx
->Const
.MinPointSizeAA
= 1.0;
537 ctx
->Const
.MaxPointSize
= 255.0;
538 ctx
->Const
.MaxPointSizeAA
= 255.0;
539 ctx
->Const
.PointSizeGranularity
= 1.0;
541 if (devinfo
->gen
>= 5 || devinfo
->is_g4x
)
542 ctx
->Const
.MaxClipPlanes
= 8;
544 ctx
->Const
.GLSLTessLevelsAsInputs
= true;
545 ctx
->Const
.PrimitiveRestartForPatches
= true;
547 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeInstructions
= 16 * 1024;
548 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxAluInstructions
= 0;
549 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxTexInstructions
= 0;
550 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxTexIndirections
= 0;
551 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeAluInstructions
= 0;
552 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeTexInstructions
= 0;
553 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeTexIndirections
= 0;
554 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeAttribs
= 16;
555 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeTemps
= 256;
556 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeAddressRegs
= 1;
557 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeParameters
= 1024;
558 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxEnvParams
=
559 MIN2(ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeParameters
,
560 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxEnvParams
);
562 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeInstructions
= 1024;
563 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeAluInstructions
= 1024;
564 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeTexInstructions
= 1024;
565 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeTexIndirections
= 1024;
566 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeAttribs
= 12;
567 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeTemps
= 256;
568 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeAddressRegs
= 0;
569 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeParameters
= 1024;
570 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxEnvParams
=
571 MIN2(ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeParameters
,
572 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxEnvParams
);
574 /* Fragment shaders use real, 32-bit twos-complement integers for all
577 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
.RangeMin
= 31;
578 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
.RangeMax
= 30;
579 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
.Precision
= 0;
580 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].HighInt
= ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
;
581 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MediumInt
= ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
;
583 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
.RangeMin
= 31;
584 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
.RangeMax
= 30;
585 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
.Precision
= 0;
586 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].HighInt
= ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
;
587 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MediumInt
= ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
;
589 /* Gen6 converts quads to polygon in beginning of 3D pipeline,
590 * but we're not sure how it's actually done for vertex order,
591 * that affect provoking vertex decision. Always use last vertex
592 * convention for quad primitive which works as expected for now.
594 if (devinfo
->gen
>= 6)
595 ctx
->Const
.QuadsFollowProvokingVertexConvention
= false;
597 ctx
->Const
.NativeIntegers
= true;
598 ctx
->Const
.VertexID_is_zero_based
= true;
600 /* Regarding the CMP instruction, the Ivybridge PRM says:
602 * "For each enabled channel 0b or 1b is assigned to the appropriate flag
603 * bit and 0/all zeros or all ones (e.g, byte 0xFF, word 0xFFFF, DWord
604 * 0xFFFFFFFF) is assigned to dst."
606 * but PRMs for earlier generations say
608 * "In dword format, one GRF may store up to 8 results. When the register
609 * is used later as a vector of Booleans, as only LSB at each channel
610 * contains meaning [sic] data, software should make sure all higher bits
611 * are masked out (e.g. by 'and-ing' an [sic] 0x01 constant)."
613 * We select the representation of a true boolean uniform to be ~0, and fix
614 * the results of Gen <= 5 CMP instruction's with -(result & 1).
616 ctx
->Const
.UniformBooleanTrue
= ~0;
618 /* From the gen4 PRM, volume 4 page 127:
620 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field specifies
621 * the base address of the first element of the surface, computed in
622 * software by adding the surface base address to the byte offset of
623 * the element in the buffer."
625 * However, unaligned accesses are slower, so enforce buffer alignment.
627 * In order to push UBO data, 3DSTATE_CONSTANT_XS imposes an additional
628 * restriction: the start of the buffer needs to be 32B aligned.
630 ctx
->Const
.UniformBufferOffsetAlignment
= 32;
632 /* ShaderStorageBufferOffsetAlignment should be a cacheline (64 bytes) so
633 * that we can safely have the CPU and GPU writing the same SSBO on
634 * non-cachecoherent systems (our Atom CPUs). With UBOs, the GPU never
635 * writes, so there's no problem. For an SSBO, the GPU and the CPU can
636 * be updating disjoint regions of the buffer simultaneously and that will
637 * break if the regions overlap the same cacheline.
639 ctx
->Const
.ShaderStorageBufferOffsetAlignment
= 64;
640 ctx
->Const
.TextureBufferOffsetAlignment
= 16;
641 ctx
->Const
.MaxTextureBufferSize
= 128 * 1024 * 1024;
643 if (devinfo
->gen
>= 6) {
644 ctx
->Const
.MaxVarying
= 32;
645 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxOutputComponents
= 128;
646 ctx
->Const
.Program
[MESA_SHADER_GEOMETRY
].MaxInputComponents
=
647 compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
] ? 128 : 64;
648 ctx
->Const
.Program
[MESA_SHADER_GEOMETRY
].MaxOutputComponents
= 128;
649 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxInputComponents
= 128;
650 ctx
->Const
.Program
[MESA_SHADER_TESS_CTRL
].MaxInputComponents
= 128;
651 ctx
->Const
.Program
[MESA_SHADER_TESS_CTRL
].MaxOutputComponents
= 128;
652 ctx
->Const
.Program
[MESA_SHADER_TESS_EVAL
].MaxInputComponents
= 128;
653 ctx
->Const
.Program
[MESA_SHADER_TESS_EVAL
].MaxOutputComponents
= 128;
656 /* We want the GLSL compiler to emit code that uses condition codes */
657 for (int i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
658 ctx
->Const
.ShaderCompilerOptions
[i
] =
659 brw
->screen
->compiler
->glsl_compiler_options
[i
];
662 if (devinfo
->gen
>= 7) {
663 ctx
->Const
.MaxViewportWidth
= 32768;
664 ctx
->Const
.MaxViewportHeight
= 32768;
667 /* ARB_viewport_array, OES_viewport_array */
668 if (devinfo
->gen
>= 6) {
669 ctx
->Const
.MaxViewports
= GEN6_NUM_VIEWPORTS
;
670 ctx
->Const
.ViewportSubpixelBits
= 0;
672 /* Cast to float before negating because MaxViewportWidth is unsigned.
674 ctx
->Const
.ViewportBounds
.Min
= -(float)ctx
->Const
.MaxViewportWidth
;
675 ctx
->Const
.ViewportBounds
.Max
= ctx
->Const
.MaxViewportWidth
;
678 /* ARB_gpu_shader5 */
679 if (devinfo
->gen
>= 7)
680 ctx
->Const
.MaxVertexStreams
= MIN2(4, MAX_VERTEX_STREAMS
);
682 /* ARB_framebuffer_no_attachments */
683 ctx
->Const
.MaxFramebufferWidth
= 16384;
684 ctx
->Const
.MaxFramebufferHeight
= 16384;
685 ctx
->Const
.MaxFramebufferLayers
= ctx
->Const
.MaxArrayTextureLayers
;
686 ctx
->Const
.MaxFramebufferSamples
= max_samples
;
688 /* OES_primitive_bounding_box */
689 ctx
->Const
.NoPrimitiveBoundingBoxOutput
= true;
691 /* TODO: We should be able to use STD430 packing by default on all hardware
692 * but some piglit tests [1] currently fail on SNB when this is enabled.
693 * The problem is the messages we're using for doing uniform pulls
694 * in the vec4 back-end on SNB is the OWORD block load instruction, which
695 * takes its offset in units of OWORDS (16 bytes). On IVB+, we use the
696 * sampler which doesn't have these restrictions.
698 * In the scalar back-end, we use the sampler for dynamic uniform loads and
699 * pull an entire cache line at a time for constant offset loads both of
700 * which support almost any alignment.
702 * [1] glsl-1.40/uniform_buffer/vs-float-array-variable-index.shader_test
704 if (devinfo
->gen
>= 7)
705 ctx
->Const
.UseSTD430AsDefaultPacking
= true;
707 if (!(ctx
->Const
.ContextFlags
& GL_CONTEXT_FLAG_DEBUG_BIT
))
708 ctx
->Const
.AllowMappedBuffersDuringExecution
= true;
710 /* GL_ARB_get_program_binary */
711 ctx
->Const
.NumProgramBinaryFormats
= 1;
715 brw_initialize_cs_context_constants(struct brw_context
*brw
)
717 struct gl_context
*ctx
= &brw
->ctx
;
718 const struct intel_screen
*screen
= brw
->screen
;
719 struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
721 /* FINISHME: Do this for all platforms that the kernel supports */
722 if (devinfo
->is_cherryview
&&
723 screen
->subslice_total
> 0 && screen
->eu_total
> 0) {
724 /* Logical CS threads = EUs per subslice * 7 threads per EU */
725 uint32_t max_cs_threads
= screen
->eu_total
/ screen
->subslice_total
* 7;
727 /* Fuse configurations may give more threads than expected, never less. */
728 if (max_cs_threads
> devinfo
->max_cs_threads
)
729 devinfo
->max_cs_threads
= max_cs_threads
;
732 /* Maximum number of scalar compute shader invocations that can be run in
733 * parallel in the same subslice assuming SIMD32 dispatch.
735 * We don't advertise more than 64 threads, because we are limited to 64 by
736 * our usage of thread_width_max in the gpgpu walker command. This only
737 * currently impacts Haswell, which otherwise might be able to advertise 70
738 * threads. With SIMD32 and 64 threads, Haswell still provides twice the
739 * required the number of invocation needed for ARB_compute_shader.
741 const unsigned max_threads
= MIN2(64, devinfo
->max_cs_threads
);
742 const uint32_t max_invocations
= 32 * max_threads
;
743 ctx
->Const
.MaxComputeWorkGroupSize
[0] = max_invocations
;
744 ctx
->Const
.MaxComputeWorkGroupSize
[1] = max_invocations
;
745 ctx
->Const
.MaxComputeWorkGroupSize
[2] = max_invocations
;
746 ctx
->Const
.MaxComputeWorkGroupInvocations
= max_invocations
;
747 ctx
->Const
.MaxComputeSharedMemorySize
= 64 * 1024;
751 * Process driconf (drirc) options, setting appropriate context flags.
753 * intelInitExtensions still pokes at optionCache directly, in order to
754 * avoid advertising various extensions. No flags are set, so it makes
755 * sense to continue doing that there.
758 brw_process_driconf_options(struct brw_context
*brw
)
760 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
761 struct gl_context
*ctx
= &brw
->ctx
;
763 driOptionCache
*options
= &brw
->optionCache
;
764 driParseConfigFiles(options
, &brw
->screen
->optionCache
,
765 brw
->driContext
->driScreenPriv
->myNum
, "i965");
767 int bo_reuse_mode
= driQueryOptioni(options
, "bo_reuse");
768 switch (bo_reuse_mode
) {
769 case DRI_CONF_BO_REUSE_DISABLED
:
771 case DRI_CONF_BO_REUSE_ALL
:
772 brw_bufmgr_enable_reuse(brw
->bufmgr
);
776 if (INTEL_DEBUG
& DEBUG_NO_HIZ
) {
777 brw
->has_hiz
= false;
778 /* On gen6, you can only do separate stencil with HIZ. */
779 if (devinfo
->gen
== 6)
780 brw
->has_separate_stencil
= false;
783 if (driQueryOptionb(options
, "mesa_no_error"))
784 ctx
->Const
.ContextFlags
|= GL_CONTEXT_FLAG_NO_ERROR_BIT_KHR
;
786 if (driQueryOptionb(options
, "always_flush_batch")) {
787 fprintf(stderr
, "flushing batchbuffer before/after each draw call\n");
788 brw
->always_flush_batch
= true;
791 if (driQueryOptionb(options
, "always_flush_cache")) {
792 fprintf(stderr
, "flushing GPU caches before/after each draw call\n");
793 brw
->always_flush_cache
= true;
796 if (driQueryOptionb(options
, "disable_throttling")) {
797 fprintf(stderr
, "disabling flush throttling\n");
798 brw
->disable_throttling
= true;
801 brw
->precompile
= driQueryOptionb(&brw
->optionCache
, "shader_precompile");
803 if (driQueryOptionb(&brw
->optionCache
, "precise_trig"))
804 brw
->screen
->compiler
->precise_trig
= true;
806 ctx
->Const
.ForceGLSLExtensionsWarn
=
807 driQueryOptionb(options
, "force_glsl_extensions_warn");
809 ctx
->Const
.ForceGLSLVersion
=
810 driQueryOptioni(options
, "force_glsl_version");
812 ctx
->Const
.DisableGLSLLineContinuations
=
813 driQueryOptionb(options
, "disable_glsl_line_continuations");
815 ctx
->Const
.AllowGLSLExtensionDirectiveMidShader
=
816 driQueryOptionb(options
, "allow_glsl_extension_directive_midshader");
818 ctx
->Const
.AllowGLSLBuiltinVariableRedeclaration
=
819 driQueryOptionb(options
, "allow_glsl_builtin_variable_redeclaration");
821 ctx
->Const
.AllowHigherCompatVersion
=
822 driQueryOptionb(options
, "allow_higher_compat_version");
824 ctx
->Const
.ForceGLSLAbsSqrt
=
825 driQueryOptionb(options
, "force_glsl_abs_sqrt");
827 ctx
->Const
.GLSLZeroInit
= driQueryOptionb(options
, "glsl_zero_init");
829 brw
->dual_color_blend_by_location
=
830 driQueryOptionb(options
, "dual_color_blend_by_location");
832 ctx
->Const
.AllowGLSLCrossStageInterpolationMismatch
=
833 driQueryOptionb(options
, "allow_glsl_cross_stage_interpolation_mismatch");
835 ctx
->Const
.dri_config_options_sha1
= ralloc_array(brw
, unsigned char, 20);
836 driComputeOptionsSha1(&brw
->screen
->optionCache
,
837 ctx
->Const
.dri_config_options_sha1
);
841 brwCreateContext(gl_api api
,
842 const struct gl_config
*mesaVis
,
843 __DRIcontext
*driContextPriv
,
844 const struct __DriverContextConfig
*ctx_config
,
845 unsigned *dri_ctx_error
,
846 void *sharedContextPrivate
)
848 struct gl_context
*shareCtx
= (struct gl_context
*) sharedContextPrivate
;
849 struct intel_screen
*screen
= driContextPriv
->driScreenPriv
->driverPrivate
;
850 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
851 struct dd_function_table functions
;
853 /* Only allow the __DRI_CTX_FLAG_ROBUST_BUFFER_ACCESS flag if the kernel
854 * provides us with context reset notifications.
856 uint32_t allowed_flags
= __DRI_CTX_FLAG_DEBUG
|
857 __DRI_CTX_FLAG_FORWARD_COMPATIBLE
|
858 __DRI_CTX_FLAG_NO_ERROR
;
860 if (screen
->has_context_reset_notification
)
861 allowed_flags
|= __DRI_CTX_FLAG_ROBUST_BUFFER_ACCESS
;
863 if (ctx_config
->flags
& ~allowed_flags
) {
864 *dri_ctx_error
= __DRI_CTX_ERROR_UNKNOWN_FLAG
;
868 if (ctx_config
->attribute_mask
&
869 ~(__DRIVER_CONTEXT_ATTRIB_RESET_STRATEGY
|
870 __DRIVER_CONTEXT_ATTRIB_PRIORITY
)) {
871 *dri_ctx_error
= __DRI_CTX_ERROR_UNKNOWN_ATTRIBUTE
;
876 ((ctx_config
->attribute_mask
& __DRIVER_CONTEXT_ATTRIB_RESET_STRATEGY
) &&
877 ctx_config
->reset_strategy
!= __DRI_CTX_RESET_NO_NOTIFICATION
);
879 struct brw_context
*brw
= rzalloc(NULL
, struct brw_context
);
881 fprintf(stderr
, "%s: failed to alloc context\n", __func__
);
882 *dri_ctx_error
= __DRI_CTX_ERROR_NO_MEMORY
;
886 driContextPriv
->driverPrivate
= brw
;
887 brw
->driContext
= driContextPriv
;
888 brw
->screen
= screen
;
889 brw
->bufmgr
= screen
->bufmgr
;
891 brw
->has_hiz
= devinfo
->has_hiz_and_separate_stencil
;
892 brw
->has_separate_stencil
= devinfo
->has_hiz_and_separate_stencil
;
894 brw
->has_swizzling
= screen
->hw_has_swizzling
;
896 brw
->isl_dev
= screen
->isl_dev
;
898 brw
->vs
.base
.stage
= MESA_SHADER_VERTEX
;
899 brw
->tcs
.base
.stage
= MESA_SHADER_TESS_CTRL
;
900 brw
->tes
.base
.stage
= MESA_SHADER_TESS_EVAL
;
901 brw
->gs
.base
.stage
= MESA_SHADER_GEOMETRY
;
902 brw
->wm
.base
.stage
= MESA_SHADER_FRAGMENT
;
903 brw
->cs
.base
.stage
= MESA_SHADER_COMPUTE
;
904 if (devinfo
->gen
>= 8) {
905 brw
->vtbl
.emit_depth_stencil_hiz
= gen8_emit_depth_stencil_hiz
;
906 } else if (devinfo
->gen
>= 7) {
907 brw
->vtbl
.emit_depth_stencil_hiz
= gen7_emit_depth_stencil_hiz
;
908 } else if (devinfo
->gen
>= 6) {
909 brw
->vtbl
.emit_depth_stencil_hiz
= gen6_emit_depth_stencil_hiz
;
911 brw
->vtbl
.emit_depth_stencil_hiz
= brw_emit_depth_stencil_hiz
;
914 brw_init_driver_functions(brw
, &functions
);
917 functions
.GetGraphicsResetStatus
= brw_get_graphics_reset_status
;
919 struct gl_context
*ctx
= &brw
->ctx
;
921 if (!_mesa_initialize_context(ctx
, api
, mesaVis
, shareCtx
, &functions
)) {
922 *dri_ctx_error
= __DRI_CTX_ERROR_NO_MEMORY
;
923 fprintf(stderr
, "%s: failed to init mesa context\n", __func__
);
924 intelDestroyContext(driContextPriv
);
928 driContextSetFlags(ctx
, ctx_config
->flags
);
930 /* Initialize the software rasterizer and helper modules.
932 * As of GL 3.1 core, the gen4+ driver doesn't need the swrast context for
933 * software fallbacks (which we have to support on legacy GL to do weird
934 * glDrawPixels(), glBitmap(), and other functions).
936 if (api
!= API_OPENGL_CORE
&& api
!= API_OPENGLES2
) {
937 _swrast_CreateContext(ctx
);
940 _vbo_CreateContext(ctx
);
941 if (ctx
->swrast_context
) {
942 _tnl_CreateContext(ctx
);
943 TNL_CONTEXT(ctx
)->Driver
.RunPipeline
= _tnl_run_pipeline
;
944 _swsetup_CreateContext(ctx
);
946 /* Configure swrast to match hardware characteristics: */
947 _swrast_allow_pixel_fog(ctx
, false);
948 _swrast_allow_vertex_fog(ctx
, true);
951 _mesa_meta_init(ctx
);
953 brw_process_driconf_options(brw
);
955 if (INTEL_DEBUG
& DEBUG_PERF
)
956 brw
->perf_debug
= true;
958 brw_initialize_cs_context_constants(brw
);
959 brw_initialize_context_constants(brw
);
961 ctx
->Const
.ResetStrategy
= notify_reset
962 ? GL_LOSE_CONTEXT_ON_RESET_ARB
: GL_NO_RESET_NOTIFICATION_ARB
;
964 /* Reinitialize the context point state. It depends on ctx->Const values. */
965 _mesa_init_point(ctx
);
969 intel_batchbuffer_init(brw
);
971 if (devinfo
->gen
>= 6) {
972 /* Create a new hardware context. Using a hardware context means that
973 * our GPU state will be saved/restored on context switch, allowing us
974 * to assume that the GPU is in the same state we left it in.
976 * This is required for transform feedback buffer offsets, query objects,
977 * and also allows us to reduce how much state we have to emit.
979 brw
->hw_ctx
= brw_create_hw_context(brw
->bufmgr
);
982 fprintf(stderr
, "Failed to create hardware context.\n");
983 intelDestroyContext(driContextPriv
);
987 int hw_priority
= GEN_CONTEXT_MEDIUM_PRIORITY
;
988 if (ctx_config
->attribute_mask
& __DRIVER_CONTEXT_ATTRIB_PRIORITY
) {
989 switch (ctx_config
->priority
) {
990 case __DRI_CTX_PRIORITY_LOW
:
991 hw_priority
= GEN_CONTEXT_LOW_PRIORITY
;
993 case __DRI_CTX_PRIORITY_HIGH
:
994 hw_priority
= GEN_CONTEXT_HIGH_PRIORITY
;
998 if (hw_priority
!= I915_CONTEXT_DEFAULT_PRIORITY
&&
999 brw_hw_context_set_priority(brw
->bufmgr
, brw
->hw_ctx
, hw_priority
)) {
1001 "Failed to set priority [%d:%d] for hardware context.\n",
1002 ctx_config
->priority
, hw_priority
);
1003 intelDestroyContext(driContextPriv
);
1008 if (brw_init_pipe_control(brw
, devinfo
)) {
1009 *dri_ctx_error
= __DRI_CTX_ERROR_NO_MEMORY
;
1010 intelDestroyContext(driContextPriv
);
1014 brw_init_state(brw
);
1016 intelInitExtensions(ctx
);
1018 brw_init_surface_formats(brw
);
1020 brw_blorp_init(brw
);
1022 brw
->urb
.size
= devinfo
->urb
.size
;
1024 if (devinfo
->gen
== 6)
1025 brw
->urb
.gs_present
= false;
1027 brw
->prim_restart
.in_progress
= false;
1028 brw
->prim_restart
.enable_cut_index
= false;
1029 brw
->gs
.enabled
= false;
1030 brw
->clip
.viewport_count
= 1;
1032 brw
->predicate
.state
= BRW_PREDICATE_STATE_RENDER
;
1034 brw
->max_gtt_map_object_size
= screen
->max_gtt_map_object_size
;
1036 ctx
->VertexProgram
._MaintainTnlProgram
= true;
1037 ctx
->FragmentProgram
._MaintainTexEnvProgram
= true;
1039 brw_draw_init( brw
);
1041 if ((ctx_config
->flags
& __DRI_CTX_FLAG_DEBUG
) != 0) {
1042 /* Turn on some extra GL_ARB_debug_output generation. */
1043 brw
->perf_debug
= true;
1046 if ((ctx_config
->flags
& __DRI_CTX_FLAG_ROBUST_BUFFER_ACCESS
) != 0) {
1047 ctx
->Const
.ContextFlags
|= GL_CONTEXT_FLAG_ROBUST_ACCESS_BIT_ARB
;
1048 ctx
->Const
.RobustAccess
= GL_TRUE
;
1051 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
1052 brw_init_shader_time(brw
);
1054 _mesa_override_extensions(ctx
);
1055 _mesa_compute_version(ctx
);
1057 _mesa_initialize_dispatch_tables(ctx
);
1058 _mesa_initialize_vbo_vtxfmt(ctx
);
1060 if (ctx
->Extensions
.INTEL_performance_query
)
1061 brw_init_performance_queries(brw
);
1063 vbo_use_buffer_objects(ctx
);
1064 vbo_always_unmap_buffers(ctx
);
1066 brw
->ctx
.Cache
= brw
->screen
->disk_cache
;
1072 intelDestroyContext(__DRIcontext
* driContextPriv
)
1074 struct brw_context
*brw
=
1075 (struct brw_context
*) driContextPriv
->driverPrivate
;
1076 struct gl_context
*ctx
= &brw
->ctx
;
1077 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1079 _mesa_meta_free(&brw
->ctx
);
1081 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
) {
1082 /* Force a report. */
1083 brw
->shader_time
.report_time
= 0;
1085 brw_collect_and_report_shader_time(brw
);
1086 brw_destroy_shader_time(brw
);
1089 if (devinfo
->gen
>= 6)
1090 blorp_finish(&brw
->blorp
);
1092 brw_destroy_state(brw
);
1093 brw_draw_destroy(brw
);
1095 brw_bo_unreference(brw
->curbe
.curbe_bo
);
1097 brw_bo_unreference(brw
->vs
.base
.scratch_bo
);
1098 brw_bo_unreference(brw
->tcs
.base
.scratch_bo
);
1099 brw_bo_unreference(brw
->tes
.base
.scratch_bo
);
1100 brw_bo_unreference(brw
->gs
.base
.scratch_bo
);
1101 brw_bo_unreference(brw
->wm
.base
.scratch_bo
);
1103 brw_bo_unreference(brw
->vs
.base
.push_const_bo
);
1104 brw_bo_unreference(brw
->tcs
.base
.push_const_bo
);
1105 brw_bo_unreference(brw
->tes
.base
.push_const_bo
);
1106 brw_bo_unreference(brw
->gs
.base
.push_const_bo
);
1107 brw_bo_unreference(brw
->wm
.base
.push_const_bo
);
1109 brw_destroy_hw_context(brw
->bufmgr
, brw
->hw_ctx
);
1111 if (ctx
->swrast_context
) {
1112 _swsetup_DestroyContext(&brw
->ctx
);
1113 _tnl_DestroyContext(&brw
->ctx
);
1115 _vbo_DestroyContext(&brw
->ctx
);
1117 if (ctx
->swrast_context
)
1118 _swrast_DestroyContext(&brw
->ctx
);
1120 brw_fini_pipe_control(brw
);
1121 intel_batchbuffer_free(&brw
->batch
);
1123 brw_bo_unreference(brw
->throttle_batch
[1]);
1124 brw_bo_unreference(brw
->throttle_batch
[0]);
1125 brw
->throttle_batch
[1] = NULL
;
1126 brw
->throttle_batch
[0] = NULL
;
1128 driDestroyOptionCache(&brw
->optionCache
);
1130 /* free the Mesa context */
1131 _mesa_free_context_data(&brw
->ctx
);
1134 driContextPriv
->driverPrivate
= NULL
;
1138 intelUnbindContext(__DRIcontext
* driContextPriv
)
1140 /* Unset current context and dispath table */
1141 _mesa_make_current(NULL
, NULL
, NULL
);
1147 * Fixes up the context for GLES23 with our default-to-sRGB-capable behavior
1148 * on window system framebuffers.
1150 * Desktop GL is fairly reasonable in its handling of sRGB: You can ask if
1151 * your renderbuffer can do sRGB encode, and you can flip a switch that does
1152 * sRGB encode if the renderbuffer can handle it. You can ask specifically
1153 * for a visual where you're guaranteed to be capable, but it turns out that
1154 * everyone just makes all their ARGB8888 visuals capable and doesn't offer
1155 * incapable ones, because there's no difference between the two in resources
1156 * used. Applications thus get built that accidentally rely on the default
1157 * visual choice being sRGB, so we make ours sRGB capable. Everything sounds
1160 * But for GLES2/3, they decided that it was silly to not turn on sRGB encode
1161 * for sRGB renderbuffers you made with the GL_EXT_texture_sRGB equivalent.
1162 * So they removed the enable knob and made it "if the renderbuffer is sRGB
1163 * capable, do sRGB encode". Then, for your window system renderbuffers, you
1164 * can ask for sRGB visuals and get sRGB encode, or not ask for sRGB visuals
1165 * and get no sRGB encode (assuming that both kinds of visual are available).
1166 * Thus our choice to support sRGB by default on our visuals for desktop would
1167 * result in broken rendering of GLES apps that aren't expecting sRGB encode.
1169 * Unfortunately, renderbuffer setup happens before a context is created. So
1170 * in intel_screen.c we always set up sRGB, and here, if you're a GLES2/3
1171 * context (without an sRGB visual), we go turn that back off before anyone
1175 intel_gles3_srgb_workaround(struct brw_context
*brw
,
1176 struct gl_framebuffer
*fb
)
1178 struct gl_context
*ctx
= &brw
->ctx
;
1180 if (_mesa_is_desktop_gl(ctx
) || !fb
->Visual
.sRGBCapable
)
1183 for (int i
= 0; i
< BUFFER_COUNT
; i
++) {
1184 struct gl_renderbuffer
*rb
= fb
->Attachment
[i
].Renderbuffer
;
1186 /* Check if sRGB was specifically asked for. */
1187 struct intel_renderbuffer
*irb
= intel_get_renderbuffer(fb
, i
);
1188 if (irb
&& irb
->need_srgb
)
1192 rb
->Format
= _mesa_get_srgb_format_linear(rb
->Format
);
1194 /* Disable sRGB from framebuffers that are not compatible. */
1195 fb
->Visual
.sRGBCapable
= false;
1199 intelMakeCurrent(__DRIcontext
* driContextPriv
,
1200 __DRIdrawable
* driDrawPriv
,
1201 __DRIdrawable
* driReadPriv
)
1203 struct brw_context
*brw
;
1206 brw
= (struct brw_context
*) driContextPriv
->driverPrivate
;
1210 if (driContextPriv
) {
1211 struct gl_context
*ctx
= &brw
->ctx
;
1212 struct gl_framebuffer
*fb
, *readFb
;
1214 if (driDrawPriv
== NULL
) {
1215 fb
= _mesa_get_incomplete_framebuffer();
1217 fb
= driDrawPriv
->driverPrivate
;
1218 driContextPriv
->dri2
.draw_stamp
= driDrawPriv
->dri2
.stamp
- 1;
1221 if (driReadPriv
== NULL
) {
1222 readFb
= _mesa_get_incomplete_framebuffer();
1224 readFb
= driReadPriv
->driverPrivate
;
1225 driContextPriv
->dri2
.read_stamp
= driReadPriv
->dri2
.stamp
- 1;
1228 /* The sRGB workaround changes the renderbuffer's format. We must change
1229 * the format before the renderbuffer's miptree get's allocated, otherwise
1230 * the formats of the renderbuffer and its miptree will differ.
1232 intel_gles3_srgb_workaround(brw
, fb
);
1233 intel_gles3_srgb_workaround(brw
, readFb
);
1235 /* If the context viewport hasn't been initialized, force a call out to
1236 * the loader to get buffers so we have a drawable size for the initial
1238 if (!brw
->ctx
.ViewportInitialized
)
1239 intel_prepare_render(brw
);
1241 _mesa_make_current(ctx
, fb
, readFb
);
1243 _mesa_make_current(NULL
, NULL
, NULL
);
1250 intel_resolve_for_dri2_flush(struct brw_context
*brw
,
1251 __DRIdrawable
*drawable
)
1253 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1255 if (devinfo
->gen
< 6) {
1256 /* MSAA and fast color clear are not supported, so don't waste time
1257 * checking whether a resolve is needed.
1262 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1263 struct intel_renderbuffer
*rb
;
1265 /* Usually, only the back buffer will need to be downsampled. However,
1266 * the front buffer will also need it if the user has rendered into it.
1268 static const gl_buffer_index buffers
[2] = {
1273 for (int i
= 0; i
< 2; ++i
) {
1274 rb
= intel_get_renderbuffer(fb
, buffers
[i
]);
1275 if (rb
== NULL
|| rb
->mt
== NULL
)
1277 if (rb
->mt
->surf
.samples
== 1) {
1278 assert(rb
->mt_layer
== 0 && rb
->mt_level
== 0 &&
1279 rb
->layer_count
== 1);
1280 intel_miptree_prepare_external(brw
, rb
->mt
);
1282 intel_renderbuffer_downsample(brw
, rb
);
1284 /* Call prepare_external on the single-sample miptree to do any
1285 * needed resolves prior to handing it off to the window system.
1286 * This is needed in the case that rb->singlesample_mt is Y-tiled
1287 * with CCS_E enabled but without I915_FORMAT_MOD_Y_TILED_CCS_E. In
1288 * this case, the MSAA resolve above will write compressed data into
1289 * rb->singlesample_mt.
1291 * TODO: Some day, if we decide to care about the tiny performance
1292 * hit we're taking by doing the MSAA resolve and then a CCS resolve,
1293 * we could detect this case and just allocate the single-sampled
1294 * miptree without aux. However, that would be a lot of plumbing and
1295 * this is a rather exotic case so it's not really worth it.
1297 intel_miptree_prepare_external(brw
, rb
->singlesample_mt
);
1303 intel_bits_per_pixel(const struct intel_renderbuffer
*rb
)
1305 return _mesa_get_format_bytes(intel_rb_format(rb
)) * 8;
1309 intel_query_dri2_buffers(struct brw_context
*brw
,
1310 __DRIdrawable
*drawable
,
1311 __DRIbuffer
**buffers
,
1315 intel_process_dri2_buffer(struct brw_context
*brw
,
1316 __DRIdrawable
*drawable
,
1317 __DRIbuffer
*buffer
,
1318 struct intel_renderbuffer
*rb
,
1319 const char *buffer_name
);
1322 intel_update_image_buffers(struct brw_context
*brw
, __DRIdrawable
*drawable
);
1325 intel_update_dri2_buffers(struct brw_context
*brw
, __DRIdrawable
*drawable
)
1327 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1328 struct intel_renderbuffer
*rb
;
1329 __DRIbuffer
*buffers
= NULL
;
1331 const char *region_name
;
1333 /* Set this up front, so that in case our buffers get invalidated
1334 * while we're getting new buffers, we don't clobber the stamp and
1335 * thus ignore the invalidate. */
1336 drawable
->lastStamp
= drawable
->dri2
.stamp
;
1338 if (unlikely(INTEL_DEBUG
& DEBUG_DRI
))
1339 fprintf(stderr
, "enter %s, drawable %p\n", __func__
, drawable
);
1341 intel_query_dri2_buffers(brw
, drawable
, &buffers
, &count
);
1343 if (buffers
== NULL
)
1346 for (int i
= 0; i
< count
; i
++) {
1347 switch (buffers
[i
].attachment
) {
1348 case __DRI_BUFFER_FRONT_LEFT
:
1349 rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1350 region_name
= "dri2 front buffer";
1353 case __DRI_BUFFER_FAKE_FRONT_LEFT
:
1354 rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1355 region_name
= "dri2 fake front buffer";
1358 case __DRI_BUFFER_BACK_LEFT
:
1359 rb
= intel_get_renderbuffer(fb
, BUFFER_BACK_LEFT
);
1360 region_name
= "dri2 back buffer";
1363 case __DRI_BUFFER_DEPTH
:
1364 case __DRI_BUFFER_HIZ
:
1365 case __DRI_BUFFER_DEPTH_STENCIL
:
1366 case __DRI_BUFFER_STENCIL
:
1367 case __DRI_BUFFER_ACCUM
:
1370 "unhandled buffer attach event, attachment type %d\n",
1371 buffers
[i
].attachment
);
1375 intel_process_dri2_buffer(brw
, drawable
, &buffers
[i
], rb
, region_name
);
1381 intel_update_renderbuffers(__DRIcontext
*context
, __DRIdrawable
*drawable
)
1383 struct brw_context
*brw
= context
->driverPrivate
;
1384 __DRIscreen
*dri_screen
= brw
->screen
->driScrnPriv
;
1386 /* Set this up front, so that in case our buffers get invalidated
1387 * while we're getting new buffers, we don't clobber the stamp and
1388 * thus ignore the invalidate. */
1389 drawable
->lastStamp
= drawable
->dri2
.stamp
;
1391 if (unlikely(INTEL_DEBUG
& DEBUG_DRI
))
1392 fprintf(stderr
, "enter %s, drawable %p\n", __func__
, drawable
);
1394 if (dri_screen
->image
.loader
)
1395 intel_update_image_buffers(brw
, drawable
);
1397 intel_update_dri2_buffers(brw
, drawable
);
1399 driUpdateFramebufferSize(&brw
->ctx
, drawable
);
1403 * intel_prepare_render should be called anywhere that curent read/drawbuffer
1404 * state is required.
1407 intel_prepare_render(struct brw_context
*brw
)
1409 struct gl_context
*ctx
= &brw
->ctx
;
1410 __DRIcontext
*driContext
= brw
->driContext
;
1411 __DRIdrawable
*drawable
;
1413 drawable
= driContext
->driDrawablePriv
;
1414 if (drawable
&& drawable
->dri2
.stamp
!= driContext
->dri2
.draw_stamp
) {
1415 if (drawable
->lastStamp
!= drawable
->dri2
.stamp
)
1416 intel_update_renderbuffers(driContext
, drawable
);
1417 driContext
->dri2
.draw_stamp
= drawable
->dri2
.stamp
;
1420 drawable
= driContext
->driReadablePriv
;
1421 if (drawable
&& drawable
->dri2
.stamp
!= driContext
->dri2
.read_stamp
) {
1422 if (drawable
->lastStamp
!= drawable
->dri2
.stamp
)
1423 intel_update_renderbuffers(driContext
, drawable
);
1424 driContext
->dri2
.read_stamp
= drawable
->dri2
.stamp
;
1427 /* If we're currently rendering to the front buffer, the rendering
1428 * that will happen next will probably dirty the front buffer. So
1429 * mark it as dirty here.
1431 if (_mesa_is_front_buffer_drawing(ctx
->DrawBuffer
))
1432 brw
->front_buffer_dirty
= true;
1436 * \brief Query DRI2 to obtain a DRIdrawable's buffers.
1438 * To determine which DRI buffers to request, examine the renderbuffers
1439 * attached to the drawable's framebuffer. Then request the buffers with
1440 * DRI2GetBuffers() or DRI2GetBuffersWithFormat().
1442 * This is called from intel_update_renderbuffers().
1444 * \param drawable Drawable whose buffers are queried.
1445 * \param buffers [out] List of buffers returned by DRI2 query.
1446 * \param buffer_count [out] Number of buffers returned.
1448 * \see intel_update_renderbuffers()
1449 * \see DRI2GetBuffers()
1450 * \see DRI2GetBuffersWithFormat()
1453 intel_query_dri2_buffers(struct brw_context
*brw
,
1454 __DRIdrawable
*drawable
,
1455 __DRIbuffer
**buffers
,
1458 __DRIscreen
*dri_screen
= brw
->screen
->driScrnPriv
;
1459 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1461 unsigned attachments
[8];
1463 struct intel_renderbuffer
*front_rb
;
1464 struct intel_renderbuffer
*back_rb
;
1466 front_rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1467 back_rb
= intel_get_renderbuffer(fb
, BUFFER_BACK_LEFT
);
1469 memset(attachments
, 0, sizeof(attachments
));
1470 if ((_mesa_is_front_buffer_drawing(fb
) ||
1471 _mesa_is_front_buffer_reading(fb
) ||
1472 !back_rb
) && front_rb
) {
1473 /* If a fake front buffer is in use, then querying for
1474 * __DRI_BUFFER_FRONT_LEFT will cause the server to copy the image from
1475 * the real front buffer to the fake front buffer. So before doing the
1476 * query, we need to make sure all the pending drawing has landed in the
1477 * real front buffer.
1479 intel_batchbuffer_flush(brw
);
1480 intel_flush_front(&brw
->ctx
);
1482 attachments
[i
++] = __DRI_BUFFER_FRONT_LEFT
;
1483 attachments
[i
++] = intel_bits_per_pixel(front_rb
);
1484 } else if (front_rb
&& brw
->front_buffer_dirty
) {
1485 /* We have pending front buffer rendering, but we aren't querying for a
1486 * front buffer. If the front buffer we have is a fake front buffer,
1487 * the X server is going to throw it away when it processes the query.
1488 * So before doing the query, make sure all the pending drawing has
1489 * landed in the real front buffer.
1491 intel_batchbuffer_flush(brw
);
1492 intel_flush_front(&brw
->ctx
);
1496 attachments
[i
++] = __DRI_BUFFER_BACK_LEFT
;
1497 attachments
[i
++] = intel_bits_per_pixel(back_rb
);
1500 assert(i
<= ARRAY_SIZE(attachments
));
1503 dri_screen
->dri2
.loader
->getBuffersWithFormat(drawable
,
1508 drawable
->loaderPrivate
);
1512 * \brief Assign a DRI buffer's DRM region to a renderbuffer.
1514 * This is called from intel_update_renderbuffers().
1517 * DRI buffers whose attachment point is DRI2BufferStencil or
1518 * DRI2BufferDepthStencil are handled as special cases.
1520 * \param buffer_name is a human readable name, such as "dri2 front buffer",
1521 * that is passed to brw_bo_gem_create_from_name().
1523 * \see intel_update_renderbuffers()
1526 intel_process_dri2_buffer(struct brw_context
*brw
,
1527 __DRIdrawable
*drawable
,
1528 __DRIbuffer
*buffer
,
1529 struct intel_renderbuffer
*rb
,
1530 const char *buffer_name
)
1532 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1538 unsigned num_samples
= rb
->Base
.Base
.NumSamples
;
1540 /* We try to avoid closing and reopening the same BO name, because the first
1541 * use of a mapping of the buffer involves a bunch of page faulting which is
1542 * moderately expensive.
1544 struct intel_mipmap_tree
*last_mt
;
1545 if (num_samples
== 0)
1548 last_mt
= rb
->singlesample_mt
;
1550 uint32_t old_name
= 0;
1552 /* The bo already has a name because the miptree was created by a
1553 * previous call to intel_process_dri2_buffer(). If a bo already has a
1554 * name, then brw_bo_flink() is a low-cost getter. It does not
1555 * create a new name.
1557 brw_bo_flink(last_mt
->bo
, &old_name
);
1560 if (old_name
== buffer
->name
)
1563 if (unlikely(INTEL_DEBUG
& DEBUG_DRI
)) {
1565 "attaching buffer %d, at %d, cpp %d, pitch %d\n",
1566 buffer
->name
, buffer
->attachment
,
1567 buffer
->cpp
, buffer
->pitch
);
1570 bo
= brw_bo_gem_create_from_name(brw
->bufmgr
, buffer_name
,
1574 "Failed to open BO for returned DRI2 buffer "
1575 "(%dx%d, %s, named %d).\n"
1576 "This is likely a bug in the X Server that will lead to a "
1578 drawable
->w
, drawable
->h
, buffer_name
, buffer
->name
);
1582 uint32_t tiling
, swizzle
;
1583 brw_bo_get_tiling(bo
, &tiling
, &swizzle
);
1585 struct intel_mipmap_tree
*mt
=
1586 intel_miptree_create_for_bo(brw
,
1588 intel_rb_format(rb
),
1594 isl_tiling_from_i915_tiling(tiling
),
1595 MIPTREE_CREATE_DEFAULT
);
1597 brw_bo_unreference(bo
);
1601 /* We got this BO from X11. We cana't assume that we have coherent texture
1602 * access because X may suddenly decide to use it for scan-out which would
1603 * destroy coherency.
1605 bo
->cache_coherent
= false;
1607 if (!intel_update_winsys_renderbuffer_miptree(brw
, rb
, mt
,
1608 drawable
->w
, drawable
->h
,
1610 brw_bo_unreference(bo
);
1611 intel_miptree_release(&mt
);
1615 if (_mesa_is_front_buffer_drawing(fb
) &&
1616 (buffer
->attachment
== __DRI_BUFFER_FRONT_LEFT
||
1617 buffer
->attachment
== __DRI_BUFFER_FAKE_FRONT_LEFT
) &&
1618 rb
->Base
.Base
.NumSamples
> 1) {
1619 intel_renderbuffer_upsample(brw
, rb
);
1624 brw_bo_unreference(bo
);
1628 * \brief Query DRI image loader to obtain a DRIdrawable's buffers.
1630 * To determine which DRI buffers to request, examine the renderbuffers
1631 * attached to the drawable's framebuffer. Then request the buffers from
1634 * This is called from intel_update_renderbuffers().
1636 * \param drawable Drawable whose buffers are queried.
1637 * \param buffers [out] List of buffers returned by DRI2 query.
1638 * \param buffer_count [out] Number of buffers returned.
1640 * \see intel_update_renderbuffers()
1644 intel_update_image_buffer(struct brw_context
*intel
,
1645 __DRIdrawable
*drawable
,
1646 struct intel_renderbuffer
*rb
,
1648 enum __DRIimageBufferMask buffer_type
)
1650 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1652 if (!rb
|| !buffer
->bo
)
1655 unsigned num_samples
= rb
->Base
.Base
.NumSamples
;
1657 /* Check and see if we're already bound to the right
1660 struct intel_mipmap_tree
*last_mt
;
1661 if (num_samples
== 0)
1664 last_mt
= rb
->singlesample_mt
;
1666 if (last_mt
&& last_mt
->bo
== buffer
->bo
)
1669 struct intel_mipmap_tree
*mt
=
1670 intel_miptree_create_for_dri_image(intel
, buffer
, GL_TEXTURE_2D
,
1671 intel_rb_format(rb
), true);
1675 if (!intel_update_winsys_renderbuffer_miptree(intel
, rb
, mt
,
1676 buffer
->width
, buffer
->height
,
1678 intel_miptree_release(&mt
);
1682 if (_mesa_is_front_buffer_drawing(fb
) &&
1683 buffer_type
== __DRI_IMAGE_BUFFER_FRONT
&&
1684 rb
->Base
.Base
.NumSamples
> 1) {
1685 intel_renderbuffer_upsample(intel
, rb
);
1690 intel_update_image_buffers(struct brw_context
*brw
, __DRIdrawable
*drawable
)
1692 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1693 __DRIscreen
*dri_screen
= brw
->screen
->driScrnPriv
;
1694 struct intel_renderbuffer
*front_rb
;
1695 struct intel_renderbuffer
*back_rb
;
1696 struct __DRIimageList images
;
1698 uint32_t buffer_mask
= 0;
1701 front_rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1702 back_rb
= intel_get_renderbuffer(fb
, BUFFER_BACK_LEFT
);
1705 format
= intel_rb_format(back_rb
);
1707 format
= intel_rb_format(front_rb
);
1711 if (front_rb
&& (_mesa_is_front_buffer_drawing(fb
) ||
1712 _mesa_is_front_buffer_reading(fb
) || !back_rb
)) {
1713 buffer_mask
|= __DRI_IMAGE_BUFFER_FRONT
;
1717 buffer_mask
|= __DRI_IMAGE_BUFFER_BACK
;
1719 ret
= dri_screen
->image
.loader
->getBuffers(drawable
,
1720 driGLFormatToImageFormat(format
),
1721 &drawable
->dri2
.stamp
,
1722 drawable
->loaderPrivate
,
1728 if (images
.image_mask
& __DRI_IMAGE_BUFFER_FRONT
) {
1729 drawable
->w
= images
.front
->width
;
1730 drawable
->h
= images
.front
->height
;
1731 intel_update_image_buffer(brw
,
1735 __DRI_IMAGE_BUFFER_FRONT
);
1738 if (images
.image_mask
& __DRI_IMAGE_BUFFER_BACK
) {
1739 drawable
->w
= images
.back
->width
;
1740 drawable
->h
= images
.back
->height
;
1741 intel_update_image_buffer(brw
,
1745 __DRI_IMAGE_BUFFER_BACK
);