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"
49 #include "vbo/vbo_context.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"
78 /***************************************
79 * Mesa's Driver Functions
80 ***************************************/
82 const char *const brw_vendor_string
= "Intel Open Source Technology Center";
85 get_bsw_model(const struct intel_screen
*screen
)
87 switch (screen
->eu_total
) {
98 brw_get_renderer_string(const struct intel_screen
*screen
)
101 static char buffer
[128];
104 switch (screen
->deviceID
) {
106 #define CHIPSET(id, symbol, str) case id: chipset = str; break;
107 #include "pci_ids/i965_pci_ids.h"
109 chipset
= "Unknown Intel Chipset";
113 /* Braswell branding is funny, so we have to fix it up here */
114 if (screen
->deviceID
== 0x22B1) {
115 bsw
= strdup(chipset
);
116 char *needle
= strstr(bsw
, "XXX");
118 memcpy(needle
, get_bsw_model(screen
), 3);
123 (void) driGetRendererString(buffer
, chipset
, 0);
128 static const GLubyte
*
129 intel_get_string(struct gl_context
* ctx
, GLenum name
)
131 const struct brw_context
*const brw
= brw_context(ctx
);
135 return (GLubyte
*) brw_vendor_string
;
139 (GLubyte
*) brw_get_renderer_string(brw
->screen
);
147 intel_viewport(struct gl_context
*ctx
)
149 struct brw_context
*brw
= brw_context(ctx
);
150 __DRIcontext
*driContext
= brw
->driContext
;
152 if (_mesa_is_winsys_fbo(ctx
->DrawBuffer
)) {
153 if (driContext
->driDrawablePriv
)
154 dri2InvalidateDrawable(driContext
->driDrawablePriv
);
155 if (driContext
->driReadablePriv
)
156 dri2InvalidateDrawable(driContext
->driReadablePriv
);
161 intel_update_framebuffer(struct gl_context
*ctx
,
162 struct gl_framebuffer
*fb
)
164 struct brw_context
*brw
= brw_context(ctx
);
166 /* Quantize the derived default number of samples
168 fb
->DefaultGeometry
._NumSamples
=
169 intel_quantize_num_samples(brw
->screen
,
170 fb
->DefaultGeometry
.NumSamples
);
174 intel_update_state(struct gl_context
* ctx
)
176 GLuint new_state
= ctx
->NewState
;
177 struct brw_context
*brw
= brw_context(ctx
);
179 if (ctx
->swrast_context
)
180 _swrast_InvalidateState(ctx
, new_state
);
182 brw
->NewGLState
|= new_state
;
184 if (new_state
& (_NEW_SCISSOR
| _NEW_BUFFERS
| _NEW_VIEWPORT
))
185 _mesa_update_draw_buffer_bounds(ctx
, ctx
->DrawBuffer
);
187 if (new_state
& (_NEW_STENCIL
| _NEW_BUFFERS
)) {
188 brw
->stencil_enabled
= _mesa_stencil_is_enabled(ctx
);
189 brw
->stencil_two_sided
= _mesa_stencil_is_two_sided(ctx
);
190 brw
->stencil_write_enabled
=
191 _mesa_stencil_is_write_enabled(ctx
, brw
->stencil_two_sided
);
194 if (new_state
& _NEW_POLYGON
)
195 brw
->polygon_front_bit
= _mesa_polygon_get_front_bit(ctx
);
197 if (new_state
& _NEW_BUFFERS
) {
198 intel_update_framebuffer(ctx
, ctx
->DrawBuffer
);
199 if (ctx
->DrawBuffer
!= ctx
->ReadBuffer
)
200 intel_update_framebuffer(ctx
, ctx
->ReadBuffer
);
204 #define flushFront(screen) ((screen)->image.loader ? (screen)->image.loader->flushFrontBuffer : (screen)->dri2.loader->flushFrontBuffer)
207 intel_flush_front(struct gl_context
*ctx
)
209 struct brw_context
*brw
= brw_context(ctx
);
210 __DRIcontext
*driContext
= brw
->driContext
;
211 __DRIdrawable
*driDrawable
= driContext
->driDrawablePriv
;
212 __DRIscreen
*const dri_screen
= brw
->screen
->driScrnPriv
;
214 if (brw
->front_buffer_dirty
&& _mesa_is_winsys_fbo(ctx
->DrawBuffer
)) {
215 if (flushFront(dri_screen
) && driDrawable
&&
216 driDrawable
->loaderPrivate
) {
218 /* Resolve before flushing FAKE_FRONT_LEFT to FRONT_LEFT.
220 * This potentially resolves both front and back buffer. It
221 * is unnecessary to resolve the back, but harms nothing except
222 * performance. And no one cares about front-buffer render
225 intel_resolve_for_dri2_flush(brw
, driDrawable
);
226 intel_batchbuffer_flush(brw
);
228 flushFront(dri_screen
)(driDrawable
, driDrawable
->loaderPrivate
);
230 /* We set the dirty bit in intel_prepare_render() if we're
231 * front buffer rendering once we get there.
233 brw
->front_buffer_dirty
= false;
239 intel_glFlush(struct gl_context
*ctx
)
241 struct brw_context
*brw
= brw_context(ctx
);
243 intel_batchbuffer_flush(brw
);
244 intel_flush_front(ctx
);
246 brw
->need_flush_throttle
= true;
250 intel_finish(struct gl_context
* ctx
)
252 struct brw_context
*brw
= brw_context(ctx
);
256 if (brw
->batch
.last_bo
)
257 brw_bo_wait_rendering(brw
->batch
.last_bo
);
261 brw_init_driver_functions(struct brw_context
*brw
,
262 struct dd_function_table
*functions
)
264 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
266 _mesa_init_driver_functions(functions
);
268 /* GLX uses DRI2 invalidate events to handle window resizing.
269 * Unfortunately, EGL does not - libEGL is written in XCB (not Xlib),
270 * which doesn't provide a mechanism for snooping the event queues.
272 * So EGL still relies on viewport hacks to handle window resizing.
273 * This should go away with DRI3000.
275 if (!brw
->driContext
->driScreenPriv
->dri2
.useInvalidate
)
276 functions
->Viewport
= intel_viewport
;
278 functions
->Flush
= intel_glFlush
;
279 functions
->Finish
= intel_finish
;
280 functions
->GetString
= intel_get_string
;
281 functions
->UpdateState
= intel_update_state
;
283 intelInitTextureFuncs(functions
);
284 intelInitTextureImageFuncs(functions
);
285 intelInitTextureCopyImageFuncs(functions
);
286 intelInitCopyImageFuncs(functions
);
287 intelInitClearFuncs(functions
);
288 intelInitBufferFuncs(functions
);
289 intelInitPixelFuncs(functions
);
290 intelInitBufferObjectFuncs(functions
);
291 brw_init_syncobj_functions(functions
);
292 brw_init_object_purgeable_functions(functions
);
294 brwInitFragProgFuncs( functions
);
295 brw_init_common_queryobj_functions(functions
);
296 if (devinfo
->gen
>= 8 || devinfo
->is_haswell
)
297 hsw_init_queryobj_functions(functions
);
298 else if (devinfo
->gen
>= 6)
299 gen6_init_queryobj_functions(functions
);
301 gen4_init_queryobj_functions(functions
);
302 brw_init_compute_functions(functions
);
303 brw_init_conditional_render_functions(functions
);
305 functions
->QueryInternalFormat
= brw_query_internal_format
;
307 functions
->NewTransformFeedback
= brw_new_transform_feedback
;
308 functions
->DeleteTransformFeedback
= brw_delete_transform_feedback
;
309 if (can_do_mi_math_and_lrr(brw
->screen
)) {
310 functions
->BeginTransformFeedback
= hsw_begin_transform_feedback
;
311 functions
->EndTransformFeedback
= hsw_end_transform_feedback
;
312 functions
->PauseTransformFeedback
= hsw_pause_transform_feedback
;
313 functions
->ResumeTransformFeedback
= hsw_resume_transform_feedback
;
314 } else if (devinfo
->gen
>= 7) {
315 functions
->BeginTransformFeedback
= gen7_begin_transform_feedback
;
316 functions
->EndTransformFeedback
= gen7_end_transform_feedback
;
317 functions
->PauseTransformFeedback
= gen7_pause_transform_feedback
;
318 functions
->ResumeTransformFeedback
= gen7_resume_transform_feedback
;
319 functions
->GetTransformFeedbackVertexCount
=
320 brw_get_transform_feedback_vertex_count
;
322 functions
->BeginTransformFeedback
= brw_begin_transform_feedback
;
323 functions
->EndTransformFeedback
= brw_end_transform_feedback
;
324 functions
->PauseTransformFeedback
= brw_pause_transform_feedback
;
325 functions
->ResumeTransformFeedback
= brw_resume_transform_feedback
;
326 functions
->GetTransformFeedbackVertexCount
=
327 brw_get_transform_feedback_vertex_count
;
330 if (devinfo
->gen
>= 6)
331 functions
->GetSamplePosition
= gen6_get_sample_position
;
335 brw_initialize_context_constants(struct brw_context
*brw
)
337 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
338 struct gl_context
*ctx
= &brw
->ctx
;
339 const struct brw_compiler
*compiler
= brw
->screen
->compiler
;
341 const bool stage_exists
[MESA_SHADER_STAGES
] = {
342 [MESA_SHADER_VERTEX
] = true,
343 [MESA_SHADER_TESS_CTRL
] = devinfo
->gen
>= 7,
344 [MESA_SHADER_TESS_EVAL
] = devinfo
->gen
>= 7,
345 [MESA_SHADER_GEOMETRY
] = devinfo
->gen
>= 6,
346 [MESA_SHADER_FRAGMENT
] = true,
347 [MESA_SHADER_COMPUTE
] =
348 ((ctx
->API
== API_OPENGL_COMPAT
|| ctx
->API
== API_OPENGL_CORE
) &&
349 ctx
->Const
.MaxComputeWorkGroupSize
[0] >= 1024) ||
350 (ctx
->API
== API_OPENGLES2
&&
351 ctx
->Const
.MaxComputeWorkGroupSize
[0] >= 128) ||
352 _mesa_extension_override_enables
.ARB_compute_shader
,
355 unsigned num_stages
= 0;
356 for (int i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
361 unsigned max_samplers
=
362 devinfo
->gen
>= 8 || devinfo
->is_haswell
? BRW_MAX_TEX_UNIT
: 16;
364 ctx
->Const
.MaxDualSourceDrawBuffers
= 1;
365 ctx
->Const
.MaxDrawBuffers
= BRW_MAX_DRAW_BUFFERS
;
366 ctx
->Const
.MaxCombinedShaderOutputResources
=
367 MAX_IMAGE_UNITS
+ BRW_MAX_DRAW_BUFFERS
;
369 /* The timestamp register we can read for glGetTimestamp() is
370 * sometimes only 32 bits, before scaling to nanoseconds (depending
373 * Once scaled to nanoseconds the timestamp would roll over at a
374 * non-power-of-two, so an application couldn't use
375 * GL_QUERY_COUNTER_BITS to handle rollover correctly. Instead, we
376 * report 36 bits and truncate at that (rolling over 5 times as
377 * often as the HW counter), and when the 32-bit counter rolls
378 * over, it happens to also be at a rollover in the reported value
379 * from near (1<<36) to 0.
381 * The low 32 bits rolls over in ~343 seconds. Our 36-bit result
382 * rolls over every ~69 seconds.
384 ctx
->Const
.QueryCounterBits
.Timestamp
= 36;
386 ctx
->Const
.MaxTextureCoordUnits
= 8; /* Mesa limit */
387 ctx
->Const
.MaxImageUnits
= MAX_IMAGE_UNITS
;
388 if (devinfo
->gen
>= 7) {
389 ctx
->Const
.MaxRenderbufferSize
= 16384;
390 ctx
->Const
.MaxTextureLevels
= MIN2(15 /* 16384 */, MAX_TEXTURE_LEVELS
);
391 ctx
->Const
.MaxCubeTextureLevels
= 15; /* 16384 */
393 ctx
->Const
.MaxRenderbufferSize
= 8192;
394 ctx
->Const
.MaxTextureLevels
= MIN2(14 /* 8192 */, MAX_TEXTURE_LEVELS
);
395 ctx
->Const
.MaxCubeTextureLevels
= 14; /* 8192 */
397 ctx
->Const
.Max3DTextureLevels
= 12; /* 2048 */
398 ctx
->Const
.MaxArrayTextureLayers
= devinfo
->gen
>= 7 ? 2048 : 512;
399 ctx
->Const
.MaxTextureMbytes
= 1536;
400 ctx
->Const
.MaxTextureRectSize
= devinfo
->gen
>= 7 ? 16384 : 8192;
401 ctx
->Const
.MaxTextureMaxAnisotropy
= 16.0;
402 ctx
->Const
.MaxTextureLodBias
= 15.0;
403 ctx
->Const
.StripTextureBorder
= true;
404 if (devinfo
->gen
>= 7) {
405 ctx
->Const
.MaxProgramTextureGatherComponents
= 4;
406 ctx
->Const
.MinProgramTextureGatherOffset
= -32;
407 ctx
->Const
.MaxProgramTextureGatherOffset
= 31;
408 } else if (devinfo
->gen
== 6) {
409 ctx
->Const
.MaxProgramTextureGatherComponents
= 1;
410 ctx
->Const
.MinProgramTextureGatherOffset
= -8;
411 ctx
->Const
.MaxProgramTextureGatherOffset
= 7;
414 ctx
->Const
.MaxUniformBlockSize
= 65536;
416 for (int i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
417 struct gl_program_constants
*prog
= &ctx
->Const
.Program
[i
];
419 if (!stage_exists
[i
])
422 prog
->MaxTextureImageUnits
= max_samplers
;
424 prog
->MaxUniformBlocks
= BRW_MAX_UBO
;
425 prog
->MaxCombinedUniformComponents
=
426 prog
->MaxUniformComponents
+
427 ctx
->Const
.MaxUniformBlockSize
/ 4 * prog
->MaxUniformBlocks
;
429 prog
->MaxAtomicCounters
= MAX_ATOMIC_COUNTERS
;
430 prog
->MaxAtomicBuffers
= BRW_MAX_ABO
;
431 prog
->MaxImageUniforms
= compiler
->scalar_stage
[i
] ? BRW_MAX_IMAGES
: 0;
432 prog
->MaxShaderStorageBlocks
= BRW_MAX_SSBO
;
435 ctx
->Const
.MaxTextureUnits
=
436 MIN2(ctx
->Const
.MaxTextureCoordUnits
,
437 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxTextureImageUnits
);
439 ctx
->Const
.MaxUniformBufferBindings
= num_stages
* BRW_MAX_UBO
;
440 ctx
->Const
.MaxCombinedUniformBlocks
= num_stages
* BRW_MAX_UBO
;
441 ctx
->Const
.MaxCombinedAtomicBuffers
= num_stages
* BRW_MAX_ABO
;
442 ctx
->Const
.MaxCombinedShaderStorageBlocks
= num_stages
* BRW_MAX_SSBO
;
443 ctx
->Const
.MaxShaderStorageBufferBindings
= num_stages
* BRW_MAX_SSBO
;
444 ctx
->Const
.MaxCombinedTextureImageUnits
= num_stages
* max_samplers
;
445 ctx
->Const
.MaxCombinedImageUniforms
= num_stages
* BRW_MAX_IMAGES
;
448 /* Hardware only supports a limited number of transform feedback buffers.
449 * So we need to override the Mesa default (which is based only on software
452 ctx
->Const
.MaxTransformFeedbackBuffers
= BRW_MAX_SOL_BUFFERS
;
454 /* On Gen6, in the worst case, we use up one binding table entry per
455 * transform feedback component (see comments above the definition of
456 * BRW_MAX_SOL_BINDINGS, in brw_context.h), so we need to advertise a value
457 * for MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS equal to
458 * BRW_MAX_SOL_BINDINGS.
460 * In "separate components" mode, we need to divide this value by
461 * BRW_MAX_SOL_BUFFERS, so that the total number of binding table entries
462 * used up by all buffers will not exceed BRW_MAX_SOL_BINDINGS.
464 ctx
->Const
.MaxTransformFeedbackInterleavedComponents
= BRW_MAX_SOL_BINDINGS
;
465 ctx
->Const
.MaxTransformFeedbackSeparateComponents
=
466 BRW_MAX_SOL_BINDINGS
/ BRW_MAX_SOL_BUFFERS
;
468 ctx
->Const
.AlwaysUseGetTransformFeedbackVertexCount
=
469 !can_do_mi_math_and_lrr(brw
->screen
);
472 const int *msaa_modes
= intel_supported_msaa_modes(brw
->screen
);
473 const int clamp_max_samples
=
474 driQueryOptioni(&brw
->optionCache
, "clamp_max_samples");
476 if (clamp_max_samples
< 0) {
477 max_samples
= msaa_modes
[0];
479 /* Select the largest supported MSAA mode that does not exceed
483 for (int i
= 0; msaa_modes
[i
] != 0; ++i
) {
484 if (msaa_modes
[i
] <= clamp_max_samples
) {
485 max_samples
= msaa_modes
[i
];
491 ctx
->Const
.MaxSamples
= max_samples
;
492 ctx
->Const
.MaxColorTextureSamples
= max_samples
;
493 ctx
->Const
.MaxDepthTextureSamples
= max_samples
;
494 ctx
->Const
.MaxIntegerSamples
= max_samples
;
495 ctx
->Const
.MaxImageSamples
= 0;
497 /* gen6_set_sample_maps() sets SampleMap{2,4,8}x variables which are used
498 * to map indices of rectangular grid to sample numbers within a pixel.
499 * These variables are used by GL_EXT_framebuffer_multisample_blit_scaled
500 * extension implementation. For more details see the comment above
501 * gen6_set_sample_maps() definition.
503 gen6_set_sample_maps(ctx
);
505 ctx
->Const
.MinLineWidth
= 1.0;
506 ctx
->Const
.MinLineWidthAA
= 1.0;
507 if (devinfo
->gen
>= 6) {
508 ctx
->Const
.MaxLineWidth
= 7.375;
509 ctx
->Const
.MaxLineWidthAA
= 7.375;
510 ctx
->Const
.LineWidthGranularity
= 0.125;
512 ctx
->Const
.MaxLineWidth
= 7.0;
513 ctx
->Const
.MaxLineWidthAA
= 7.0;
514 ctx
->Const
.LineWidthGranularity
= 0.5;
517 /* For non-antialiased lines, we have to round the line width to the
518 * nearest whole number. Make sure that we don't advertise a line
519 * width that, when rounded, will be beyond the actual hardware
522 assert(roundf(ctx
->Const
.MaxLineWidth
) <= ctx
->Const
.MaxLineWidth
);
524 ctx
->Const
.MinPointSize
= 1.0;
525 ctx
->Const
.MinPointSizeAA
= 1.0;
526 ctx
->Const
.MaxPointSize
= 255.0;
527 ctx
->Const
.MaxPointSizeAA
= 255.0;
528 ctx
->Const
.PointSizeGranularity
= 1.0;
530 if (devinfo
->gen
>= 5 || devinfo
->is_g4x
)
531 ctx
->Const
.MaxClipPlanes
= 8;
533 ctx
->Const
.GLSLTessLevelsAsInputs
= true;
534 ctx
->Const
.LowerTCSPatchVerticesIn
= devinfo
->gen
>= 8;
535 ctx
->Const
.LowerTESPatchVerticesIn
= true;
536 ctx
->Const
.PrimitiveRestartForPatches
= true;
538 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeInstructions
= 16 * 1024;
539 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxAluInstructions
= 0;
540 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxTexInstructions
= 0;
541 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxTexIndirections
= 0;
542 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeAluInstructions
= 0;
543 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeTexInstructions
= 0;
544 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeTexIndirections
= 0;
545 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeAttribs
= 16;
546 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeTemps
= 256;
547 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeAddressRegs
= 1;
548 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeParameters
= 1024;
549 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxEnvParams
=
550 MIN2(ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxNativeParameters
,
551 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxEnvParams
);
553 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeInstructions
= 1024;
554 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeAluInstructions
= 1024;
555 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeTexInstructions
= 1024;
556 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeTexIndirections
= 1024;
557 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeAttribs
= 12;
558 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeTemps
= 256;
559 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeAddressRegs
= 0;
560 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeParameters
= 1024;
561 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxEnvParams
=
562 MIN2(ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxNativeParameters
,
563 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxEnvParams
);
565 /* Fragment shaders use real, 32-bit twos-complement integers for all
568 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
.RangeMin
= 31;
569 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
.RangeMax
= 30;
570 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
.Precision
= 0;
571 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].HighInt
= ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
;
572 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MediumInt
= ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].LowInt
;
574 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
.RangeMin
= 31;
575 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
.RangeMax
= 30;
576 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
.Precision
= 0;
577 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].HighInt
= ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
;
578 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MediumInt
= ctx
->Const
.Program
[MESA_SHADER_VERTEX
].LowInt
;
580 /* Gen6 converts quads to polygon in beginning of 3D pipeline,
581 * but we're not sure how it's actually done for vertex order,
582 * that affect provoking vertex decision. Always use last vertex
583 * convention for quad primitive which works as expected for now.
585 if (devinfo
->gen
>= 6)
586 ctx
->Const
.QuadsFollowProvokingVertexConvention
= false;
588 ctx
->Const
.NativeIntegers
= true;
589 ctx
->Const
.VertexID_is_zero_based
= true;
591 /* Regarding the CMP instruction, the Ivybridge PRM says:
593 * "For each enabled channel 0b or 1b is assigned to the appropriate flag
594 * bit and 0/all zeros or all ones (e.g, byte 0xFF, word 0xFFFF, DWord
595 * 0xFFFFFFFF) is assigned to dst."
597 * but PRMs for earlier generations say
599 * "In dword format, one GRF may store up to 8 results. When the register
600 * is used later as a vector of Booleans, as only LSB at each channel
601 * contains meaning [sic] data, software should make sure all higher bits
602 * are masked out (e.g. by 'and-ing' an [sic] 0x01 constant)."
604 * We select the representation of a true boolean uniform to be ~0, and fix
605 * the results of Gen <= 5 CMP instruction's with -(result & 1).
607 ctx
->Const
.UniformBooleanTrue
= ~0;
609 /* From the gen4 PRM, volume 4 page 127:
611 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field specifies
612 * the base address of the first element of the surface, computed in
613 * software by adding the surface base address to the byte offset of
614 * the element in the buffer."
616 * However, unaligned accesses are slower, so enforce buffer alignment.
618 * In order to push UBO data, 3DSTATE_CONSTANT_XS imposes an additional
619 * restriction: the start of the buffer needs to be 32B aligned.
621 ctx
->Const
.UniformBufferOffsetAlignment
= 32;
623 /* ShaderStorageBufferOffsetAlignment should be a cacheline (64 bytes) so
624 * that we can safely have the CPU and GPU writing the same SSBO on
625 * non-cachecoherent systems (our Atom CPUs). With UBOs, the GPU never
626 * writes, so there's no problem. For an SSBO, the GPU and the CPU can
627 * be updating disjoint regions of the buffer simultaneously and that will
628 * break if the regions overlap the same cacheline.
630 ctx
->Const
.ShaderStorageBufferOffsetAlignment
= 64;
631 ctx
->Const
.TextureBufferOffsetAlignment
= 16;
632 ctx
->Const
.MaxTextureBufferSize
= 128 * 1024 * 1024;
634 if (devinfo
->gen
>= 6) {
635 ctx
->Const
.MaxVarying
= 32;
636 ctx
->Const
.Program
[MESA_SHADER_VERTEX
].MaxOutputComponents
= 128;
637 ctx
->Const
.Program
[MESA_SHADER_GEOMETRY
].MaxInputComponents
=
638 compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
] ? 128 : 64;
639 ctx
->Const
.Program
[MESA_SHADER_GEOMETRY
].MaxOutputComponents
= 128;
640 ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxInputComponents
= 128;
641 ctx
->Const
.Program
[MESA_SHADER_TESS_CTRL
].MaxInputComponents
= 128;
642 ctx
->Const
.Program
[MESA_SHADER_TESS_CTRL
].MaxOutputComponents
= 128;
643 ctx
->Const
.Program
[MESA_SHADER_TESS_EVAL
].MaxInputComponents
= 128;
644 ctx
->Const
.Program
[MESA_SHADER_TESS_EVAL
].MaxOutputComponents
= 128;
647 /* We want the GLSL compiler to emit code that uses condition codes */
648 for (int i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
649 ctx
->Const
.ShaderCompilerOptions
[i
] =
650 brw
->screen
->compiler
->glsl_compiler_options
[i
];
653 if (devinfo
->gen
>= 7) {
654 ctx
->Const
.MaxViewportWidth
= 32768;
655 ctx
->Const
.MaxViewportHeight
= 32768;
658 /* ARB_viewport_array, OES_viewport_array */
659 if (devinfo
->gen
>= 6) {
660 ctx
->Const
.MaxViewports
= GEN6_NUM_VIEWPORTS
;
661 ctx
->Const
.ViewportSubpixelBits
= 0;
663 /* Cast to float before negating because MaxViewportWidth is unsigned.
665 ctx
->Const
.ViewportBounds
.Min
= -(float)ctx
->Const
.MaxViewportWidth
;
666 ctx
->Const
.ViewportBounds
.Max
= ctx
->Const
.MaxViewportWidth
;
669 /* ARB_gpu_shader5 */
670 if (devinfo
->gen
>= 7)
671 ctx
->Const
.MaxVertexStreams
= MIN2(4, MAX_VERTEX_STREAMS
);
673 /* ARB_framebuffer_no_attachments */
674 ctx
->Const
.MaxFramebufferWidth
= 16384;
675 ctx
->Const
.MaxFramebufferHeight
= 16384;
676 ctx
->Const
.MaxFramebufferLayers
= ctx
->Const
.MaxArrayTextureLayers
;
677 ctx
->Const
.MaxFramebufferSamples
= max_samples
;
679 /* OES_primitive_bounding_box */
680 ctx
->Const
.NoPrimitiveBoundingBoxOutput
= true;
682 /* TODO: We should be able to use STD430 packing by default on all hardware
683 * but some piglit tests [1] currently fail on SNB when this is enabled.
684 * The problem is the messages we're using for doing uniform pulls
685 * in the vec4 back-end on SNB is the OWORD block load instruction, which
686 * takes its offset in units of OWORDS (16 bytes). On IVB+, we use the
687 * sampler which doesn't have these restrictions.
689 * In the scalar back-end, we use the sampler for dynamic uniform loads and
690 * pull an entire cache line at a time for constant offset loads both of
691 * which support almost any alignment.
693 * [1] glsl-1.40/uniform_buffer/vs-float-array-variable-index.shader_test
695 if (devinfo
->gen
>= 7)
696 ctx
->Const
.UseSTD430AsDefaultPacking
= true;
698 if (!(ctx
->Const
.ContextFlags
& GL_CONTEXT_FLAG_DEBUG_BIT
))
699 ctx
->Const
.AllowMappedBuffersDuringExecution
= true;
703 brw_initialize_cs_context_constants(struct brw_context
*brw
)
705 struct gl_context
*ctx
= &brw
->ctx
;
706 const struct intel_screen
*screen
= brw
->screen
;
707 struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
709 /* FINISHME: Do this for all platforms that the kernel supports */
710 if (devinfo
->is_cherryview
&&
711 screen
->subslice_total
> 0 && screen
->eu_total
> 0) {
712 /* Logical CS threads = EUs per subslice * 7 threads per EU */
713 uint32_t max_cs_threads
= screen
->eu_total
/ screen
->subslice_total
* 7;
715 /* Fuse configurations may give more threads than expected, never less. */
716 if (max_cs_threads
> devinfo
->max_cs_threads
)
717 devinfo
->max_cs_threads
= max_cs_threads
;
720 /* Maximum number of scalar compute shader invocations that can be run in
721 * parallel in the same subslice assuming SIMD32 dispatch.
723 * We don't advertise more than 64 threads, because we are limited to 64 by
724 * our usage of thread_width_max in the gpgpu walker command. This only
725 * currently impacts Haswell, which otherwise might be able to advertise 70
726 * threads. With SIMD32 and 64 threads, Haswell still provides twice the
727 * required the number of invocation needed for ARB_compute_shader.
729 const unsigned max_threads
= MIN2(64, devinfo
->max_cs_threads
);
730 const uint32_t max_invocations
= 32 * max_threads
;
731 ctx
->Const
.MaxComputeWorkGroupSize
[0] = max_invocations
;
732 ctx
->Const
.MaxComputeWorkGroupSize
[1] = max_invocations
;
733 ctx
->Const
.MaxComputeWorkGroupSize
[2] = max_invocations
;
734 ctx
->Const
.MaxComputeWorkGroupInvocations
= max_invocations
;
735 ctx
->Const
.MaxComputeSharedMemorySize
= 64 * 1024;
739 * Process driconf (drirc) options, setting appropriate context flags.
741 * intelInitExtensions still pokes at optionCache directly, in order to
742 * avoid advertising various extensions. No flags are set, so it makes
743 * sense to continue doing that there.
746 brw_process_driconf_options(struct brw_context
*brw
)
748 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
749 struct gl_context
*ctx
= &brw
->ctx
;
751 driOptionCache
*options
= &brw
->optionCache
;
752 driParseConfigFiles(options
, &brw
->screen
->optionCache
,
753 brw
->driContext
->driScreenPriv
->myNum
, "i965");
755 int bo_reuse_mode
= driQueryOptioni(options
, "bo_reuse");
756 switch (bo_reuse_mode
) {
757 case DRI_CONF_BO_REUSE_DISABLED
:
759 case DRI_CONF_BO_REUSE_ALL
:
760 brw_bufmgr_enable_reuse(brw
->bufmgr
);
764 if (INTEL_DEBUG
& DEBUG_NO_HIZ
) {
765 brw
->has_hiz
= false;
766 /* On gen6, you can only do separate stencil with HIZ. */
767 if (devinfo
->gen
== 6)
768 brw
->has_separate_stencil
= false;
771 if (driQueryOptionb(options
, "mesa_no_error"))
772 ctx
->Const
.ContextFlags
|= GL_CONTEXT_FLAG_NO_ERROR_BIT_KHR
;
774 if (driQueryOptionb(options
, "always_flush_batch")) {
775 fprintf(stderr
, "flushing batchbuffer before/after each draw call\n");
776 brw
->always_flush_batch
= true;
779 if (driQueryOptionb(options
, "always_flush_cache")) {
780 fprintf(stderr
, "flushing GPU caches before/after each draw call\n");
781 brw
->always_flush_cache
= true;
784 if (driQueryOptionb(options
, "disable_throttling")) {
785 fprintf(stderr
, "disabling flush throttling\n");
786 brw
->disable_throttling
= true;
789 brw
->precompile
= driQueryOptionb(&brw
->optionCache
, "shader_precompile");
791 if (driQueryOptionb(&brw
->optionCache
, "precise_trig"))
792 brw
->screen
->compiler
->precise_trig
= true;
794 ctx
->Const
.ForceGLSLExtensionsWarn
=
795 driQueryOptionb(options
, "force_glsl_extensions_warn");
797 ctx
->Const
.ForceGLSLVersion
=
798 driQueryOptioni(options
, "force_glsl_version");
800 ctx
->Const
.DisableGLSLLineContinuations
=
801 driQueryOptionb(options
, "disable_glsl_line_continuations");
803 ctx
->Const
.AllowGLSLExtensionDirectiveMidShader
=
804 driQueryOptionb(options
, "allow_glsl_extension_directive_midshader");
806 ctx
->Const
.AllowGLSLBuiltinVariableRedeclaration
=
807 driQueryOptionb(options
, "allow_glsl_builtin_variable_redeclaration");
809 ctx
->Const
.AllowHigherCompatVersion
=
810 driQueryOptionb(options
, "allow_higher_compat_version");
812 ctx
->Const
.ForceGLSLAbsSqrt
=
813 driQueryOptionb(options
, "force_glsl_abs_sqrt");
815 ctx
->Const
.GLSLZeroInit
= driQueryOptionb(options
, "glsl_zero_init");
817 brw
->dual_color_blend_by_location
=
818 driQueryOptionb(options
, "dual_color_blend_by_location");
822 brwCreateContext(gl_api api
,
823 const struct gl_config
*mesaVis
,
824 __DRIcontext
*driContextPriv
,
825 unsigned major_version
,
826 unsigned minor_version
,
830 unsigned *dri_ctx_error
,
831 void *sharedContextPrivate
)
833 struct gl_context
*shareCtx
= (struct gl_context
*) sharedContextPrivate
;
834 struct intel_screen
*screen
= driContextPriv
->driScreenPriv
->driverPrivate
;
835 const struct gen_device_info
*devinfo
= &screen
->devinfo
;
836 struct dd_function_table functions
;
838 /* Only allow the __DRI_CTX_FLAG_ROBUST_BUFFER_ACCESS flag if the kernel
839 * provides us with context reset notifications.
841 uint32_t allowed_flags
= __DRI_CTX_FLAG_DEBUG
|
842 __DRI_CTX_FLAG_FORWARD_COMPATIBLE
|
843 __DRI_CTX_FLAG_NO_ERROR
;
845 if (screen
->has_context_reset_notification
)
846 allowed_flags
|= __DRI_CTX_FLAG_ROBUST_BUFFER_ACCESS
;
848 if (flags
& ~allowed_flags
) {
849 *dri_ctx_error
= __DRI_CTX_ERROR_UNKNOWN_FLAG
;
853 struct brw_context
*brw
= rzalloc(NULL
, struct brw_context
);
855 fprintf(stderr
, "%s: failed to alloc context\n", __func__
);
856 *dri_ctx_error
= __DRI_CTX_ERROR_NO_MEMORY
;
860 driContextPriv
->driverPrivate
= brw
;
861 brw
->driContext
= driContextPriv
;
862 brw
->screen
= screen
;
863 brw
->bufmgr
= screen
->bufmgr
;
865 brw
->has_hiz
= devinfo
->has_hiz_and_separate_stencil
;
866 brw
->has_separate_stencil
= devinfo
->has_hiz_and_separate_stencil
;
868 brw
->has_swizzling
= screen
->hw_has_swizzling
;
870 brw
->isl_dev
= screen
->isl_dev
;
872 brw
->vs
.base
.stage
= MESA_SHADER_VERTEX
;
873 brw
->tcs
.base
.stage
= MESA_SHADER_TESS_CTRL
;
874 brw
->tes
.base
.stage
= MESA_SHADER_TESS_EVAL
;
875 brw
->gs
.base
.stage
= MESA_SHADER_GEOMETRY
;
876 brw
->wm
.base
.stage
= MESA_SHADER_FRAGMENT
;
877 if (devinfo
->gen
>= 8) {
878 brw
->vtbl
.emit_depth_stencil_hiz
= gen8_emit_depth_stencil_hiz
;
879 } else if (devinfo
->gen
>= 7) {
880 brw
->vtbl
.emit_depth_stencil_hiz
= gen7_emit_depth_stencil_hiz
;
881 } else if (devinfo
->gen
>= 6) {
882 brw
->vtbl
.emit_depth_stencil_hiz
= gen6_emit_depth_stencil_hiz
;
884 brw
->vtbl
.emit_depth_stencil_hiz
= brw_emit_depth_stencil_hiz
;
887 brw_init_driver_functions(brw
, &functions
);
890 functions
.GetGraphicsResetStatus
= brw_get_graphics_reset_status
;
892 struct gl_context
*ctx
= &brw
->ctx
;
894 if (!_mesa_initialize_context(ctx
, api
, mesaVis
, shareCtx
, &functions
)) {
895 *dri_ctx_error
= __DRI_CTX_ERROR_NO_MEMORY
;
896 fprintf(stderr
, "%s: failed to init mesa context\n", __func__
);
897 intelDestroyContext(driContextPriv
);
901 driContextSetFlags(ctx
, flags
);
903 /* Initialize the software rasterizer and helper modules.
905 * As of GL 3.1 core, the gen4+ driver doesn't need the swrast context for
906 * software fallbacks (which we have to support on legacy GL to do weird
907 * glDrawPixels(), glBitmap(), and other functions).
909 if (api
!= API_OPENGL_CORE
&& api
!= API_OPENGLES2
) {
910 _swrast_CreateContext(ctx
);
913 _vbo_CreateContext(ctx
);
914 if (ctx
->swrast_context
) {
915 _tnl_CreateContext(ctx
);
916 TNL_CONTEXT(ctx
)->Driver
.RunPipeline
= _tnl_run_pipeline
;
917 _swsetup_CreateContext(ctx
);
919 /* Configure swrast to match hardware characteristics: */
920 _swrast_allow_pixel_fog(ctx
, false);
921 _swrast_allow_vertex_fog(ctx
, true);
924 _mesa_meta_init(ctx
);
926 brw_process_driconf_options(brw
);
928 if (INTEL_DEBUG
& DEBUG_PERF
)
929 brw
->perf_debug
= true;
931 brw_initialize_cs_context_constants(brw
);
932 brw_initialize_context_constants(brw
);
934 ctx
->Const
.ResetStrategy
= notify_reset
935 ? GL_LOSE_CONTEXT_ON_RESET_ARB
: GL_NO_RESET_NOTIFICATION_ARB
;
937 /* Reinitialize the context point state. It depends on ctx->Const values. */
938 _mesa_init_point(ctx
);
942 intel_batchbuffer_init(brw
);
944 if (devinfo
->gen
>= 6) {
945 /* Create a new hardware context. Using a hardware context means that
946 * our GPU state will be saved/restored on context switch, allowing us
947 * to assume that the GPU is in the same state we left it in.
949 * This is required for transform feedback buffer offsets, query objects,
950 * and also allows us to reduce how much state we have to emit.
952 brw
->hw_ctx
= brw_create_hw_context(brw
->bufmgr
);
955 fprintf(stderr
, "Failed to create hardware context.\n");
956 intelDestroyContext(driContextPriv
);
960 int hw_priority
= BRW_CONTEXT_MEDIUM_PRIORITY
;
962 case __DRI_CTX_PRIORITY_LOW
:
963 hw_priority
= BRW_CONTEXT_LOW_PRIORITY
;
965 case __DRI_CTX_PRIORITY_HIGH
:
966 hw_priority
= BRW_CONTEXT_HIGH_PRIORITY
;
969 if (hw_priority
!= I915_CONTEXT_DEFAULT_PRIORITY
&&
970 brw_hw_context_set_priority(brw
->bufmgr
, brw
->hw_ctx
, hw_priority
)) {
972 "Failed to set priority [%d:%d] for hardware context.\n",
973 priority
, hw_priority
);
974 intelDestroyContext(driContextPriv
);
979 if (brw_init_pipe_control(brw
, devinfo
)) {
980 *dri_ctx_error
= __DRI_CTX_ERROR_NO_MEMORY
;
981 intelDestroyContext(driContextPriv
);
987 intelInitExtensions(ctx
);
989 brw_init_surface_formats(brw
);
993 brw
->urb
.size
= devinfo
->urb
.size
;
995 if (devinfo
->gen
== 6)
996 brw
->urb
.gs_present
= false;
998 brw
->prim_restart
.in_progress
= false;
999 brw
->prim_restart
.enable_cut_index
= false;
1000 brw
->gs
.enabled
= false;
1001 brw
->clip
.viewport_count
= 1;
1003 brw
->predicate
.state
= BRW_PREDICATE_STATE_RENDER
;
1005 brw
->max_gtt_map_object_size
= screen
->max_gtt_map_object_size
;
1007 ctx
->VertexProgram
._MaintainTnlProgram
= true;
1008 ctx
->FragmentProgram
._MaintainTexEnvProgram
= true;
1010 brw_draw_init( brw
);
1012 if ((flags
& __DRI_CTX_FLAG_DEBUG
) != 0) {
1013 /* Turn on some extra GL_ARB_debug_output generation. */
1014 brw
->perf_debug
= true;
1017 if ((flags
& __DRI_CTX_FLAG_ROBUST_BUFFER_ACCESS
) != 0) {
1018 ctx
->Const
.ContextFlags
|= GL_CONTEXT_FLAG_ROBUST_ACCESS_BIT_ARB
;
1019 ctx
->Const
.RobustAccess
= GL_TRUE
;
1022 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
1023 brw_init_shader_time(brw
);
1025 _mesa_compute_version(ctx
);
1027 _mesa_initialize_dispatch_tables(ctx
);
1028 _mesa_initialize_vbo_vtxfmt(ctx
);
1030 if (ctx
->Extensions
.INTEL_performance_query
)
1031 brw_init_performance_queries(brw
);
1033 vbo_use_buffer_objects(ctx
);
1034 vbo_always_unmap_buffers(ctx
);
1040 intelDestroyContext(__DRIcontext
* driContextPriv
)
1042 struct brw_context
*brw
=
1043 (struct brw_context
*) driContextPriv
->driverPrivate
;
1044 struct gl_context
*ctx
= &brw
->ctx
;
1045 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1047 _mesa_meta_free(&brw
->ctx
);
1049 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
) {
1050 /* Force a report. */
1051 brw
->shader_time
.report_time
= 0;
1053 brw_collect_and_report_shader_time(brw
);
1054 brw_destroy_shader_time(brw
);
1057 if (devinfo
->gen
>= 6)
1058 blorp_finish(&brw
->blorp
);
1060 brw_destroy_state(brw
);
1061 brw_draw_destroy(brw
);
1063 brw_bo_unreference(brw
->curbe
.curbe_bo
);
1064 if (brw
->vs
.base
.scratch_bo
)
1065 brw_bo_unreference(brw
->vs
.base
.scratch_bo
);
1066 if (brw
->tcs
.base
.scratch_bo
)
1067 brw_bo_unreference(brw
->tcs
.base
.scratch_bo
);
1068 if (brw
->tes
.base
.scratch_bo
)
1069 brw_bo_unreference(brw
->tes
.base
.scratch_bo
);
1070 if (brw
->gs
.base
.scratch_bo
)
1071 brw_bo_unreference(brw
->gs
.base
.scratch_bo
);
1072 if (brw
->wm
.base
.scratch_bo
)
1073 brw_bo_unreference(brw
->wm
.base
.scratch_bo
);
1075 brw_destroy_hw_context(brw
->bufmgr
, brw
->hw_ctx
);
1077 if (ctx
->swrast_context
) {
1078 _swsetup_DestroyContext(&brw
->ctx
);
1079 _tnl_DestroyContext(&brw
->ctx
);
1081 _vbo_DestroyContext(&brw
->ctx
);
1083 if (ctx
->swrast_context
)
1084 _swrast_DestroyContext(&brw
->ctx
);
1086 brw_fini_pipe_control(brw
);
1087 intel_batchbuffer_free(&brw
->batch
);
1089 brw_bo_unreference(brw
->throttle_batch
[1]);
1090 brw_bo_unreference(brw
->throttle_batch
[0]);
1091 brw
->throttle_batch
[1] = NULL
;
1092 brw
->throttle_batch
[0] = NULL
;
1094 driDestroyOptionCache(&brw
->optionCache
);
1096 /* free the Mesa context */
1097 _mesa_free_context_data(&brw
->ctx
);
1100 driContextPriv
->driverPrivate
= NULL
;
1104 intelUnbindContext(__DRIcontext
* driContextPriv
)
1106 /* Unset current context and dispath table */
1107 _mesa_make_current(NULL
, NULL
, NULL
);
1113 * Fixes up the context for GLES23 with our default-to-sRGB-capable behavior
1114 * on window system framebuffers.
1116 * Desktop GL is fairly reasonable in its handling of sRGB: You can ask if
1117 * your renderbuffer can do sRGB encode, and you can flip a switch that does
1118 * sRGB encode if the renderbuffer can handle it. You can ask specifically
1119 * for a visual where you're guaranteed to be capable, but it turns out that
1120 * everyone just makes all their ARGB8888 visuals capable and doesn't offer
1121 * incapable ones, because there's no difference between the two in resources
1122 * used. Applications thus get built that accidentally rely on the default
1123 * visual choice being sRGB, so we make ours sRGB capable. Everything sounds
1126 * But for GLES2/3, they decided that it was silly to not turn on sRGB encode
1127 * for sRGB renderbuffers you made with the GL_EXT_texture_sRGB equivalent.
1128 * So they removed the enable knob and made it "if the renderbuffer is sRGB
1129 * capable, do sRGB encode". Then, for your window system renderbuffers, you
1130 * can ask for sRGB visuals and get sRGB encode, or not ask for sRGB visuals
1131 * and get no sRGB encode (assuming that both kinds of visual are available).
1132 * Thus our choice to support sRGB by default on our visuals for desktop would
1133 * result in broken rendering of GLES apps that aren't expecting sRGB encode.
1135 * Unfortunately, renderbuffer setup happens before a context is created. So
1136 * in intel_screen.c we always set up sRGB, and here, if you're a GLES2/3
1137 * context (without an sRGB visual, though we don't have sRGB visuals exposed
1138 * yet), we go turn that back off before anyone finds out.
1141 intel_gles3_srgb_workaround(struct brw_context
*brw
,
1142 struct gl_framebuffer
*fb
)
1144 struct gl_context
*ctx
= &brw
->ctx
;
1146 if (_mesa_is_desktop_gl(ctx
) || !fb
->Visual
.sRGBCapable
)
1149 /* Some day when we support the sRGB capable bit on visuals available for
1150 * GLES, we'll need to respect that and not disable things here.
1152 fb
->Visual
.sRGBCapable
= false;
1153 for (int i
= 0; i
< BUFFER_COUNT
; i
++) {
1154 struct gl_renderbuffer
*rb
= fb
->Attachment
[i
].Renderbuffer
;
1156 rb
->Format
= _mesa_get_srgb_format_linear(rb
->Format
);
1161 intelMakeCurrent(__DRIcontext
* driContextPriv
,
1162 __DRIdrawable
* driDrawPriv
,
1163 __DRIdrawable
* driReadPriv
)
1165 struct brw_context
*brw
;
1166 GET_CURRENT_CONTEXT(curCtx
);
1169 brw
= (struct brw_context
*) driContextPriv
->driverPrivate
;
1173 /* According to the glXMakeCurrent() man page: "Pending commands to
1174 * the previous context, if any, are flushed before it is released."
1175 * But only flush if we're actually changing contexts.
1177 if (brw_context(curCtx
) && brw_context(curCtx
) != brw
) {
1178 _mesa_flush(curCtx
);
1181 if (driContextPriv
) {
1182 struct gl_context
*ctx
= &brw
->ctx
;
1183 struct gl_framebuffer
*fb
, *readFb
;
1185 if (driDrawPriv
== NULL
) {
1186 fb
= _mesa_get_incomplete_framebuffer();
1188 fb
= driDrawPriv
->driverPrivate
;
1189 driContextPriv
->dri2
.draw_stamp
= driDrawPriv
->dri2
.stamp
- 1;
1192 if (driReadPriv
== NULL
) {
1193 readFb
= _mesa_get_incomplete_framebuffer();
1195 readFb
= driReadPriv
->driverPrivate
;
1196 driContextPriv
->dri2
.read_stamp
= driReadPriv
->dri2
.stamp
- 1;
1199 /* The sRGB workaround changes the renderbuffer's format. We must change
1200 * the format before the renderbuffer's miptree get's allocated, otherwise
1201 * the formats of the renderbuffer and its miptree will differ.
1203 intel_gles3_srgb_workaround(brw
, fb
);
1204 intel_gles3_srgb_workaround(brw
, readFb
);
1206 /* If the context viewport hasn't been initialized, force a call out to
1207 * the loader to get buffers so we have a drawable size for the initial
1209 if (!brw
->ctx
.ViewportInitialized
)
1210 intel_prepare_render(brw
);
1212 _mesa_make_current(ctx
, fb
, readFb
);
1214 _mesa_make_current(NULL
, NULL
, NULL
);
1221 intel_resolve_for_dri2_flush(struct brw_context
*brw
,
1222 __DRIdrawable
*drawable
)
1224 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1226 if (devinfo
->gen
< 6) {
1227 /* MSAA and fast color clear are not supported, so don't waste time
1228 * checking whether a resolve is needed.
1233 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1234 struct intel_renderbuffer
*rb
;
1236 /* Usually, only the back buffer will need to be downsampled. However,
1237 * the front buffer will also need it if the user has rendered into it.
1239 static const gl_buffer_index buffers
[2] = {
1244 for (int i
= 0; i
< 2; ++i
) {
1245 rb
= intel_get_renderbuffer(fb
, buffers
[i
]);
1246 if (rb
== NULL
|| rb
->mt
== NULL
)
1248 if (rb
->mt
->surf
.samples
== 1) {
1249 assert(rb
->mt_layer
== 0 && rb
->mt_level
== 0 &&
1250 rb
->layer_count
== 1);
1251 intel_miptree_prepare_external(brw
, rb
->mt
);
1253 intel_renderbuffer_downsample(brw
, rb
);
1259 intel_bits_per_pixel(const struct intel_renderbuffer
*rb
)
1261 return _mesa_get_format_bytes(intel_rb_format(rb
)) * 8;
1265 intel_query_dri2_buffers(struct brw_context
*brw
,
1266 __DRIdrawable
*drawable
,
1267 __DRIbuffer
**buffers
,
1271 intel_process_dri2_buffer(struct brw_context
*brw
,
1272 __DRIdrawable
*drawable
,
1273 __DRIbuffer
*buffer
,
1274 struct intel_renderbuffer
*rb
,
1275 const char *buffer_name
);
1278 intel_update_image_buffers(struct brw_context
*brw
, __DRIdrawable
*drawable
);
1281 intel_update_dri2_buffers(struct brw_context
*brw
, __DRIdrawable
*drawable
)
1283 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1284 struct intel_renderbuffer
*rb
;
1285 __DRIbuffer
*buffers
= NULL
;
1287 const char *region_name
;
1289 /* Set this up front, so that in case our buffers get invalidated
1290 * while we're getting new buffers, we don't clobber the stamp and
1291 * thus ignore the invalidate. */
1292 drawable
->lastStamp
= drawable
->dri2
.stamp
;
1294 if (unlikely(INTEL_DEBUG
& DEBUG_DRI
))
1295 fprintf(stderr
, "enter %s, drawable %p\n", __func__
, drawable
);
1297 intel_query_dri2_buffers(brw
, drawable
, &buffers
, &count
);
1299 if (buffers
== NULL
)
1302 for (int i
= 0; i
< count
; i
++) {
1303 switch (buffers
[i
].attachment
) {
1304 case __DRI_BUFFER_FRONT_LEFT
:
1305 rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1306 region_name
= "dri2 front buffer";
1309 case __DRI_BUFFER_FAKE_FRONT_LEFT
:
1310 rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1311 region_name
= "dri2 fake front buffer";
1314 case __DRI_BUFFER_BACK_LEFT
:
1315 rb
= intel_get_renderbuffer(fb
, BUFFER_BACK_LEFT
);
1316 region_name
= "dri2 back buffer";
1319 case __DRI_BUFFER_DEPTH
:
1320 case __DRI_BUFFER_HIZ
:
1321 case __DRI_BUFFER_DEPTH_STENCIL
:
1322 case __DRI_BUFFER_STENCIL
:
1323 case __DRI_BUFFER_ACCUM
:
1326 "unhandled buffer attach event, attachment type %d\n",
1327 buffers
[i
].attachment
);
1331 intel_process_dri2_buffer(brw
, drawable
, &buffers
[i
], rb
, region_name
);
1337 intel_update_renderbuffers(__DRIcontext
*context
, __DRIdrawable
*drawable
)
1339 struct brw_context
*brw
= context
->driverPrivate
;
1340 __DRIscreen
*dri_screen
= brw
->screen
->driScrnPriv
;
1342 /* Set this up front, so that in case our buffers get invalidated
1343 * while we're getting new buffers, we don't clobber the stamp and
1344 * thus ignore the invalidate. */
1345 drawable
->lastStamp
= drawable
->dri2
.stamp
;
1347 if (unlikely(INTEL_DEBUG
& DEBUG_DRI
))
1348 fprintf(stderr
, "enter %s, drawable %p\n", __func__
, drawable
);
1350 if (dri_screen
->image
.loader
)
1351 intel_update_image_buffers(brw
, drawable
);
1353 intel_update_dri2_buffers(brw
, drawable
);
1355 driUpdateFramebufferSize(&brw
->ctx
, drawable
);
1359 * intel_prepare_render should be called anywhere that curent read/drawbuffer
1360 * state is required.
1363 intel_prepare_render(struct brw_context
*brw
)
1365 struct gl_context
*ctx
= &brw
->ctx
;
1366 __DRIcontext
*driContext
= brw
->driContext
;
1367 __DRIdrawable
*drawable
;
1369 drawable
= driContext
->driDrawablePriv
;
1370 if (drawable
&& drawable
->dri2
.stamp
!= driContext
->dri2
.draw_stamp
) {
1371 if (drawable
->lastStamp
!= drawable
->dri2
.stamp
)
1372 intel_update_renderbuffers(driContext
, drawable
);
1373 driContext
->dri2
.draw_stamp
= drawable
->dri2
.stamp
;
1376 drawable
= driContext
->driReadablePriv
;
1377 if (drawable
&& drawable
->dri2
.stamp
!= driContext
->dri2
.read_stamp
) {
1378 if (drawable
->lastStamp
!= drawable
->dri2
.stamp
)
1379 intel_update_renderbuffers(driContext
, drawable
);
1380 driContext
->dri2
.read_stamp
= drawable
->dri2
.stamp
;
1383 /* If we're currently rendering to the front buffer, the rendering
1384 * that will happen next will probably dirty the front buffer. So
1385 * mark it as dirty here.
1387 if (_mesa_is_front_buffer_drawing(ctx
->DrawBuffer
))
1388 brw
->front_buffer_dirty
= true;
1392 * \brief Query DRI2 to obtain a DRIdrawable's buffers.
1394 * To determine which DRI buffers to request, examine the renderbuffers
1395 * attached to the drawable's framebuffer. Then request the buffers with
1396 * DRI2GetBuffers() or DRI2GetBuffersWithFormat().
1398 * This is called from intel_update_renderbuffers().
1400 * \param drawable Drawable whose buffers are queried.
1401 * \param buffers [out] List of buffers returned by DRI2 query.
1402 * \param buffer_count [out] Number of buffers returned.
1404 * \see intel_update_renderbuffers()
1405 * \see DRI2GetBuffers()
1406 * \see DRI2GetBuffersWithFormat()
1409 intel_query_dri2_buffers(struct brw_context
*brw
,
1410 __DRIdrawable
*drawable
,
1411 __DRIbuffer
**buffers
,
1414 __DRIscreen
*dri_screen
= brw
->screen
->driScrnPriv
;
1415 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1417 unsigned attachments
[8];
1419 struct intel_renderbuffer
*front_rb
;
1420 struct intel_renderbuffer
*back_rb
;
1422 front_rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1423 back_rb
= intel_get_renderbuffer(fb
, BUFFER_BACK_LEFT
);
1425 memset(attachments
, 0, sizeof(attachments
));
1426 if ((_mesa_is_front_buffer_drawing(fb
) ||
1427 _mesa_is_front_buffer_reading(fb
) ||
1428 !back_rb
) && front_rb
) {
1429 /* If a fake front buffer is in use, then querying for
1430 * __DRI_BUFFER_FRONT_LEFT will cause the server to copy the image from
1431 * the real front buffer to the fake front buffer. So before doing the
1432 * query, we need to make sure all the pending drawing has landed in the
1433 * real front buffer.
1435 intel_batchbuffer_flush(brw
);
1436 intel_flush_front(&brw
->ctx
);
1438 attachments
[i
++] = __DRI_BUFFER_FRONT_LEFT
;
1439 attachments
[i
++] = intel_bits_per_pixel(front_rb
);
1440 } else if (front_rb
&& brw
->front_buffer_dirty
) {
1441 /* We have pending front buffer rendering, but we aren't querying for a
1442 * front buffer. If the front buffer we have is a fake front buffer,
1443 * the X server is going to throw it away when it processes the query.
1444 * So before doing the query, make sure all the pending drawing has
1445 * landed in the real front buffer.
1447 intel_batchbuffer_flush(brw
);
1448 intel_flush_front(&brw
->ctx
);
1452 attachments
[i
++] = __DRI_BUFFER_BACK_LEFT
;
1453 attachments
[i
++] = intel_bits_per_pixel(back_rb
);
1456 assert(i
<= ARRAY_SIZE(attachments
));
1459 dri_screen
->dri2
.loader
->getBuffersWithFormat(drawable
,
1464 drawable
->loaderPrivate
);
1468 * \brief Assign a DRI buffer's DRM region to a renderbuffer.
1470 * This is called from intel_update_renderbuffers().
1473 * DRI buffers whose attachment point is DRI2BufferStencil or
1474 * DRI2BufferDepthStencil are handled as special cases.
1476 * \param buffer_name is a human readable name, such as "dri2 front buffer",
1477 * that is passed to brw_bo_gem_create_from_name().
1479 * \see intel_update_renderbuffers()
1482 intel_process_dri2_buffer(struct brw_context
*brw
,
1483 __DRIdrawable
*drawable
,
1484 __DRIbuffer
*buffer
,
1485 struct intel_renderbuffer
*rb
,
1486 const char *buffer_name
)
1488 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1494 unsigned num_samples
= rb
->Base
.Base
.NumSamples
;
1496 /* We try to avoid closing and reopening the same BO name, because the first
1497 * use of a mapping of the buffer involves a bunch of page faulting which is
1498 * moderately expensive.
1500 struct intel_mipmap_tree
*last_mt
;
1501 if (num_samples
== 0)
1504 last_mt
= rb
->singlesample_mt
;
1506 uint32_t old_name
= 0;
1508 /* The bo already has a name because the miptree was created by a
1509 * previous call to intel_process_dri2_buffer(). If a bo already has a
1510 * name, then brw_bo_flink() is a low-cost getter. It does not
1511 * create a new name.
1513 brw_bo_flink(last_mt
->bo
, &old_name
);
1516 if (old_name
== buffer
->name
)
1519 if (unlikely(INTEL_DEBUG
& DEBUG_DRI
)) {
1521 "attaching buffer %d, at %d, cpp %d, pitch %d\n",
1522 buffer
->name
, buffer
->attachment
,
1523 buffer
->cpp
, buffer
->pitch
);
1526 bo
= brw_bo_gem_create_from_name(brw
->bufmgr
, buffer_name
,
1530 "Failed to open BO for returned DRI2 buffer "
1531 "(%dx%d, %s, named %d).\n"
1532 "This is likely a bug in the X Server that will lead to a "
1534 drawable
->w
, drawable
->h
, buffer_name
, buffer
->name
);
1538 struct intel_mipmap_tree
*mt
=
1539 intel_miptree_create_for_bo(brw
,
1541 intel_rb_format(rb
),
1547 MIPTREE_CREATE_DEFAULT
);
1549 brw_bo_unreference(bo
);
1553 /* We got this BO from X11. We cana't assume that we have coherent texture
1554 * access because X may suddenly decide to use it for scan-out which would
1555 * destroy coherency.
1557 bo
->cache_coherent
= false;
1559 if (!intel_update_winsys_renderbuffer_miptree(brw
, rb
, mt
,
1560 drawable
->w
, drawable
->h
,
1562 brw_bo_unreference(bo
);
1563 intel_miptree_release(&mt
);
1567 if (_mesa_is_front_buffer_drawing(fb
) &&
1568 (buffer
->attachment
== __DRI_BUFFER_FRONT_LEFT
||
1569 buffer
->attachment
== __DRI_BUFFER_FAKE_FRONT_LEFT
) &&
1570 rb
->Base
.Base
.NumSamples
> 1) {
1571 intel_renderbuffer_upsample(brw
, rb
);
1576 brw_bo_unreference(bo
);
1580 * \brief Query DRI image loader to obtain a DRIdrawable's buffers.
1582 * To determine which DRI buffers to request, examine the renderbuffers
1583 * attached to the drawable's framebuffer. Then request the buffers from
1586 * This is called from intel_update_renderbuffers().
1588 * \param drawable Drawable whose buffers are queried.
1589 * \param buffers [out] List of buffers returned by DRI2 query.
1590 * \param buffer_count [out] Number of buffers returned.
1592 * \see intel_update_renderbuffers()
1596 intel_update_image_buffer(struct brw_context
*intel
,
1597 __DRIdrawable
*drawable
,
1598 struct intel_renderbuffer
*rb
,
1600 enum __DRIimageBufferMask buffer_type
)
1602 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1604 if (!rb
|| !buffer
->bo
)
1607 unsigned num_samples
= rb
->Base
.Base
.NumSamples
;
1609 /* Check and see if we're already bound to the right
1612 struct intel_mipmap_tree
*last_mt
;
1613 if (num_samples
== 0)
1616 last_mt
= rb
->singlesample_mt
;
1618 if (last_mt
&& last_mt
->bo
== buffer
->bo
)
1621 struct intel_mipmap_tree
*mt
=
1622 intel_miptree_create_for_dri_image(intel
, buffer
, GL_TEXTURE_2D
,
1623 intel_rb_format(rb
), true);
1627 if (!intel_update_winsys_renderbuffer_miptree(intel
, rb
, mt
,
1628 buffer
->width
, buffer
->height
,
1630 intel_miptree_release(&mt
);
1634 if (_mesa_is_front_buffer_drawing(fb
) &&
1635 buffer_type
== __DRI_IMAGE_BUFFER_FRONT
&&
1636 rb
->Base
.Base
.NumSamples
> 1) {
1637 intel_renderbuffer_upsample(intel
, rb
);
1642 intel_update_image_buffers(struct brw_context
*brw
, __DRIdrawable
*drawable
)
1644 struct gl_framebuffer
*fb
= drawable
->driverPrivate
;
1645 __DRIscreen
*dri_screen
= brw
->screen
->driScrnPriv
;
1646 struct intel_renderbuffer
*front_rb
;
1647 struct intel_renderbuffer
*back_rb
;
1648 struct __DRIimageList images
;
1650 uint32_t buffer_mask
= 0;
1653 front_rb
= intel_get_renderbuffer(fb
, BUFFER_FRONT_LEFT
);
1654 back_rb
= intel_get_renderbuffer(fb
, BUFFER_BACK_LEFT
);
1657 format
= intel_rb_format(back_rb
);
1659 format
= intel_rb_format(front_rb
);
1663 if (front_rb
&& (_mesa_is_front_buffer_drawing(fb
) ||
1664 _mesa_is_front_buffer_reading(fb
) || !back_rb
)) {
1665 buffer_mask
|= __DRI_IMAGE_BUFFER_FRONT
;
1669 buffer_mask
|= __DRI_IMAGE_BUFFER_BACK
;
1671 ret
= dri_screen
->image
.loader
->getBuffers(drawable
,
1672 driGLFormatToImageFormat(format
),
1673 &drawable
->dri2
.stamp
,
1674 drawable
->loaderPrivate
,
1680 if (images
.image_mask
& __DRI_IMAGE_BUFFER_FRONT
) {
1681 drawable
->w
= images
.front
->width
;
1682 drawable
->h
= images
.front
->height
;
1683 intel_update_image_buffer(brw
,
1687 __DRI_IMAGE_BUFFER_FRONT
);
1690 if (images
.image_mask
& __DRI_IMAGE_BUFFER_BACK
) {
1691 drawable
->w
= images
.back
->width
;
1692 drawable
->h
= images
.back
->height
;
1693 intel_update_image_buffer(brw
,
1697 __DRI_IMAGE_BUFFER_BACK
);