i965: expose MESA_FORMAT_R8G8B8A8_SRGB visual
[mesa.git] / src / mesa / drivers / dri / i965 / intel_screen.c
1 /*
2 * Copyright 2003 VMware, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sublicense, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the
14 * next paragraph) shall be included in all copies or substantial portions
15 * of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
19 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
20 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
21 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
22 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
23 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
24 */
25
26 #include <drm_fourcc.h>
27 #include <errno.h>
28 #include <time.h>
29 #include <unistd.h>
30 #include "main/context.h"
31 #include "main/framebuffer.h"
32 #include "main/renderbuffer.h"
33 #include "main/texobj.h"
34 #include "main/hash.h"
35 #include "main/fbobject.h"
36 #include "main/version.h"
37 #include "swrast/s_renderbuffer.h"
38 #include "util/ralloc.h"
39 #include "util/disk_cache.h"
40 #include "brw_defines.h"
41 #include "brw_state.h"
42 #include "compiler/nir/nir.h"
43
44 #include "utils.h"
45 #include "util/disk_cache.h"
46 #include "util/xmlpool.h"
47
48 #include "common/gen_defines.h"
49
50 static const __DRIconfigOptionsExtension brw_config_options = {
51 .base = { __DRI_CONFIG_OPTIONS, 1 },
52 .xml =
53 DRI_CONF_BEGIN
54 DRI_CONF_SECTION_PERFORMANCE
55 /* Options correspond to DRI_CONF_BO_REUSE_DISABLED,
56 * DRI_CONF_BO_REUSE_ALL
57 */
58 DRI_CONF_OPT_BEGIN_V(bo_reuse, enum, 1, "0:1")
59 DRI_CONF_DESC_BEGIN(en, "Buffer object reuse")
60 DRI_CONF_ENUM(0, "Disable buffer object reuse")
61 DRI_CONF_ENUM(1, "Enable reuse of all sizes of buffer objects")
62 DRI_CONF_DESC_END
63 DRI_CONF_OPT_END
64 DRI_CONF_MESA_NO_ERROR("false")
65 DRI_CONF_SECTION_END
66
67 DRI_CONF_SECTION_QUALITY
68 DRI_CONF_PRECISE_TRIG("false")
69
70 DRI_CONF_OPT_BEGIN(clamp_max_samples, int, -1)
71 DRI_CONF_DESC(en, "Clamp the value of GL_MAX_SAMPLES to the "
72 "given integer. If negative, then do not clamp.")
73 DRI_CONF_OPT_END
74 DRI_CONF_SECTION_END
75
76 DRI_CONF_SECTION_DEBUG
77 DRI_CONF_NO_RAST("false")
78 DRI_CONF_ALWAYS_FLUSH_BATCH("false")
79 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
80 DRI_CONF_DISABLE_THROTTLING("false")
81 DRI_CONF_FORCE_GLSL_EXTENSIONS_WARN("false")
82 DRI_CONF_FORCE_GLSL_VERSION(0)
83 DRI_CONF_DISABLE_GLSL_LINE_CONTINUATIONS("false")
84 DRI_CONF_DISABLE_BLEND_FUNC_EXTENDED("false")
85 DRI_CONF_DUAL_COLOR_BLEND_BY_LOCATION("false")
86 DRI_CONF_ALLOW_GLSL_EXTENSION_DIRECTIVE_MIDSHADER("false")
87 DRI_CONF_ALLOW_GLSL_BUILTIN_VARIABLE_REDECLARATION("false")
88 DRI_CONF_ALLOW_GLSL_CROSS_STAGE_INTERPOLATION_MISMATCH("false")
89 DRI_CONF_ALLOW_HIGHER_COMPAT_VERSION("false")
90 DRI_CONF_FORCE_GLSL_ABS_SQRT("false")
91
92 DRI_CONF_OPT_BEGIN_B(shader_precompile, "true")
93 DRI_CONF_DESC(en, "Perform code generation at shader link time.")
94 DRI_CONF_OPT_END
95 DRI_CONF_SECTION_END
96
97 DRI_CONF_SECTION_MISCELLANEOUS
98 DRI_CONF_GLSL_ZERO_INIT("false")
99 DRI_CONF_ALLOW_RGB10_CONFIGS("false")
100 DRI_CONF_SECTION_END
101 DRI_CONF_END
102 };
103
104 #include "intel_batchbuffer.h"
105 #include "intel_buffers.h"
106 #include "brw_bufmgr.h"
107 #include "intel_fbo.h"
108 #include "intel_mipmap_tree.h"
109 #include "intel_screen.h"
110 #include "intel_tex.h"
111 #include "intel_image.h"
112
113 #include "brw_context.h"
114
115 #include "i915_drm.h"
116
117 /**
118 * For debugging purposes, this returns a time in seconds.
119 */
120 double
121 get_time(void)
122 {
123 struct timespec tp;
124
125 clock_gettime(CLOCK_MONOTONIC, &tp);
126
127 return tp.tv_sec + tp.tv_nsec / 1000000000.0;
128 }
129
130 static const __DRItexBufferExtension intelTexBufferExtension = {
131 .base = { __DRI_TEX_BUFFER, 3 },
132
133 .setTexBuffer = intelSetTexBuffer,
134 .setTexBuffer2 = intelSetTexBuffer2,
135 .releaseTexBuffer = intelReleaseTexBuffer,
136 };
137
138 static void
139 intel_dri2_flush_with_flags(__DRIcontext *cPriv,
140 __DRIdrawable *dPriv,
141 unsigned flags,
142 enum __DRI2throttleReason reason)
143 {
144 struct brw_context *brw = cPriv->driverPrivate;
145
146 if (!brw)
147 return;
148
149 struct gl_context *ctx = &brw->ctx;
150
151 FLUSH_VERTICES(ctx, 0);
152
153 if (flags & __DRI2_FLUSH_DRAWABLE)
154 intel_resolve_for_dri2_flush(brw, dPriv);
155
156 if (reason == __DRI2_THROTTLE_SWAPBUFFER)
157 brw->need_swap_throttle = true;
158 if (reason == __DRI2_THROTTLE_FLUSHFRONT)
159 brw->need_flush_throttle = true;
160
161 intel_batchbuffer_flush(brw);
162 }
163
164 /**
165 * Provides compatibility with loaders that only support the older (version
166 * 1-3) flush interface.
167 *
168 * That includes libGL up to Mesa 9.0, and the X Server at least up to 1.13.
169 */
170 static void
171 intel_dri2_flush(__DRIdrawable *drawable)
172 {
173 intel_dri2_flush_with_flags(drawable->driContextPriv, drawable,
174 __DRI2_FLUSH_DRAWABLE,
175 __DRI2_THROTTLE_SWAPBUFFER);
176 }
177
178 static const struct __DRI2flushExtensionRec intelFlushExtension = {
179 .base = { __DRI2_FLUSH, 4 },
180
181 .flush = intel_dri2_flush,
182 .invalidate = dri2InvalidateDrawable,
183 .flush_with_flags = intel_dri2_flush_with_flags,
184 };
185
186 static const struct intel_image_format intel_image_formats[] = {
187 { __DRI_IMAGE_FOURCC_ARGB2101010, __DRI_IMAGE_COMPONENTS_RGBA, 1,
188 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ARGB2101010, 4 } } },
189
190 { __DRI_IMAGE_FOURCC_XRGB2101010, __DRI_IMAGE_COMPONENTS_RGB, 1,
191 { { 0, 0, 0, __DRI_IMAGE_FORMAT_XRGB2101010, 4 } } },
192
193 { __DRI_IMAGE_FOURCC_ARGB8888, __DRI_IMAGE_COMPONENTS_RGBA, 1,
194 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ARGB8888, 4 } } },
195
196 { __DRI_IMAGE_FOURCC_ABGR8888, __DRI_IMAGE_COMPONENTS_RGBA, 1,
197 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ABGR8888, 4 } } },
198
199 { __DRI_IMAGE_FOURCC_SARGB8888, __DRI_IMAGE_COMPONENTS_RGBA, 1,
200 { { 0, 0, 0, __DRI_IMAGE_FORMAT_SARGB8, 4 } } },
201
202 { __DRI_IMAGE_FOURCC_XRGB8888, __DRI_IMAGE_COMPONENTS_RGB, 1,
203 { { 0, 0, 0, __DRI_IMAGE_FORMAT_XRGB8888, 4 }, } },
204
205 { __DRI_IMAGE_FOURCC_XBGR8888, __DRI_IMAGE_COMPONENTS_RGB, 1,
206 { { 0, 0, 0, __DRI_IMAGE_FORMAT_XBGR8888, 4 }, } },
207
208 { __DRI_IMAGE_FOURCC_ARGB1555, __DRI_IMAGE_COMPONENTS_RGBA, 1,
209 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ARGB1555, 2 } } },
210
211 { __DRI_IMAGE_FOURCC_RGB565, __DRI_IMAGE_COMPONENTS_RGB, 1,
212 { { 0, 0, 0, __DRI_IMAGE_FORMAT_RGB565, 2 } } },
213
214 { __DRI_IMAGE_FOURCC_R8, __DRI_IMAGE_COMPONENTS_R, 1,
215 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 }, } },
216
217 { __DRI_IMAGE_FOURCC_R16, __DRI_IMAGE_COMPONENTS_R, 1,
218 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R16, 1 }, } },
219
220 { __DRI_IMAGE_FOURCC_GR88, __DRI_IMAGE_COMPONENTS_RG, 1,
221 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR88, 2 }, } },
222
223 { __DRI_IMAGE_FOURCC_GR1616, __DRI_IMAGE_COMPONENTS_RG, 1,
224 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR1616, 2 }, } },
225
226 { __DRI_IMAGE_FOURCC_YUV410, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
227 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
228 { 1, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 },
229 { 2, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 } } },
230
231 { __DRI_IMAGE_FOURCC_YUV411, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
232 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
233 { 1, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 },
234 { 2, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
235
236 { __DRI_IMAGE_FOURCC_YUV420, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
237 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
238 { 1, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 },
239 { 2, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 } } },
240
241 { __DRI_IMAGE_FOURCC_YUV422, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
242 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
243 { 1, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 },
244 { 2, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
245
246 { __DRI_IMAGE_FOURCC_YUV444, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
247 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
248 { 1, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
249 { 2, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
250
251 { __DRI_IMAGE_FOURCC_YVU410, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
252 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
253 { 2, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 },
254 { 1, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 } } },
255
256 { __DRI_IMAGE_FOURCC_YVU411, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
257 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
258 { 2, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 },
259 { 1, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
260
261 { __DRI_IMAGE_FOURCC_YVU420, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
262 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
263 { 2, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 },
264 { 1, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 } } },
265
266 { __DRI_IMAGE_FOURCC_YVU422, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
267 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
268 { 2, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 },
269 { 1, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
270
271 { __DRI_IMAGE_FOURCC_YVU444, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
272 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
273 { 2, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
274 { 1, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
275
276 { __DRI_IMAGE_FOURCC_NV12, __DRI_IMAGE_COMPONENTS_Y_UV, 2,
277 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
278 { 1, 1, 1, __DRI_IMAGE_FORMAT_GR88, 2 } } },
279
280 { __DRI_IMAGE_FOURCC_NV16, __DRI_IMAGE_COMPONENTS_Y_UV, 2,
281 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
282 { 1, 1, 0, __DRI_IMAGE_FORMAT_GR88, 2 } } },
283
284 /* For YUYV and UYVY buffers, we set up two overlapping DRI images
285 * and treat them as planar buffers in the compositors.
286 * Plane 0 is GR88 and samples YU or YV pairs and places Y into
287 * the R component, while plane 1 is ARGB/ABGR and samples YUYV/UYVY
288 * clusters and places pairs and places U into the G component and
289 * V into A. This lets the texture sampler interpolate the Y
290 * components correctly when sampling from plane 0, and interpolate
291 * U and V correctly when sampling from plane 1. */
292 { __DRI_IMAGE_FOURCC_YUYV, __DRI_IMAGE_COMPONENTS_Y_XUXV, 2,
293 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR88, 2 },
294 { 0, 1, 0, __DRI_IMAGE_FORMAT_ARGB8888, 4 } } },
295 { __DRI_IMAGE_FOURCC_UYVY, __DRI_IMAGE_COMPONENTS_Y_UXVX, 2,
296 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR88, 2 },
297 { 0, 1, 0, __DRI_IMAGE_FORMAT_ABGR8888, 4 } } }
298 };
299
300 static const struct {
301 uint64_t modifier;
302 unsigned since_gen;
303 } supported_modifiers[] = {
304 { .modifier = DRM_FORMAT_MOD_LINEAR , .since_gen = 1 },
305 { .modifier = I915_FORMAT_MOD_X_TILED , .since_gen = 1 },
306 { .modifier = I915_FORMAT_MOD_Y_TILED , .since_gen = 6 },
307 { .modifier = I915_FORMAT_MOD_Y_TILED_CCS , .since_gen = 9 },
308 };
309
310 static bool
311 modifier_is_supported(const struct gen_device_info *devinfo,
312 const struct intel_image_format *fmt, int dri_format,
313 uint64_t modifier)
314 {
315 const struct isl_drm_modifier_info *modinfo =
316 isl_drm_modifier_get_info(modifier);
317 int i;
318
319 /* ISL had better know about the modifier */
320 if (!modinfo)
321 return false;
322
323 if (modinfo->aux_usage == ISL_AUX_USAGE_CCS_E) {
324 /* If INTEL_DEBUG=norbc is set, don't support any CCS_E modifiers */
325 if (unlikely(INTEL_DEBUG & DEBUG_NO_RBC))
326 return false;
327
328 /* CCS_E is not supported for planar images */
329 if (fmt && fmt->nplanes > 1)
330 return false;
331
332 if (fmt) {
333 assert(dri_format == 0);
334 dri_format = fmt->planes[0].dri_format;
335 }
336
337 mesa_format format = driImageFormatToGLFormat(dri_format);
338 format = _mesa_get_srgb_format_linear(format);
339 if (!isl_format_supports_ccs_e(devinfo,
340 brw_isl_format_for_mesa_format(format)))
341 return false;
342 }
343
344 for (i = 0; i < ARRAY_SIZE(supported_modifiers); i++) {
345 if (supported_modifiers[i].modifier != modifier)
346 continue;
347
348 return supported_modifiers[i].since_gen <= devinfo->gen;
349 }
350
351 return false;
352 }
353
354 static uint64_t
355 tiling_to_modifier(uint32_t tiling)
356 {
357 static const uint64_t map[] = {
358 [I915_TILING_NONE] = DRM_FORMAT_MOD_LINEAR,
359 [I915_TILING_X] = I915_FORMAT_MOD_X_TILED,
360 [I915_TILING_Y] = I915_FORMAT_MOD_Y_TILED,
361 };
362
363 assert(tiling < ARRAY_SIZE(map));
364
365 return map[tiling];
366 }
367
368 static void
369 intel_image_warn_if_unaligned(__DRIimage *image, const char *func)
370 {
371 uint32_t tiling, swizzle;
372 brw_bo_get_tiling(image->bo, &tiling, &swizzle);
373
374 if (tiling != I915_TILING_NONE && (image->offset & 0xfff)) {
375 _mesa_warning(NULL, "%s: offset 0x%08x not on tile boundary",
376 func, image->offset);
377 }
378 }
379
380 static const struct intel_image_format *
381 intel_image_format_lookup(int fourcc)
382 {
383 for (unsigned i = 0; i < ARRAY_SIZE(intel_image_formats); i++) {
384 if (intel_image_formats[i].fourcc == fourcc)
385 return &intel_image_formats[i];
386 }
387
388 return NULL;
389 }
390
391 static boolean
392 intel_image_get_fourcc(__DRIimage *image, int *fourcc)
393 {
394 if (image->planar_format) {
395 *fourcc = image->planar_format->fourcc;
396 return true;
397 }
398
399 for (unsigned i = 0; i < ARRAY_SIZE(intel_image_formats); i++) {
400 if (intel_image_formats[i].planes[0].dri_format == image->dri_format) {
401 *fourcc = intel_image_formats[i].fourcc;
402 return true;
403 }
404 }
405 return false;
406 }
407
408 static __DRIimage *
409 intel_allocate_image(struct intel_screen *screen, int dri_format,
410 void *loaderPrivate)
411 {
412 __DRIimage *image;
413
414 image = calloc(1, sizeof *image);
415 if (image == NULL)
416 return NULL;
417
418 image->screen = screen;
419 image->dri_format = dri_format;
420 image->offset = 0;
421
422 image->format = driImageFormatToGLFormat(dri_format);
423 if (dri_format != __DRI_IMAGE_FORMAT_NONE &&
424 image->format == MESA_FORMAT_NONE) {
425 free(image);
426 return NULL;
427 }
428
429 image->internal_format = _mesa_get_format_base_format(image->format);
430 image->data = loaderPrivate;
431
432 return image;
433 }
434
435 /**
436 * Sets up a DRIImage structure to point to a slice out of a miptree.
437 */
438 static void
439 intel_setup_image_from_mipmap_tree(struct brw_context *brw, __DRIimage *image,
440 struct intel_mipmap_tree *mt, GLuint level,
441 GLuint zoffset)
442 {
443 intel_miptree_make_shareable(brw, mt);
444
445 intel_miptree_check_level_layer(mt, level, zoffset);
446
447 image->width = minify(mt->surf.phys_level0_sa.width,
448 level - mt->first_level);
449 image->height = minify(mt->surf.phys_level0_sa.height,
450 level - mt->first_level);
451 image->pitch = mt->surf.row_pitch;
452
453 image->offset = intel_miptree_get_tile_offsets(mt, level, zoffset,
454 &image->tile_x,
455 &image->tile_y);
456
457 brw_bo_unreference(image->bo);
458 image->bo = mt->bo;
459 brw_bo_reference(mt->bo);
460 }
461
462 static __DRIimage *
463 intel_create_image_from_name(__DRIscreen *dri_screen,
464 int width, int height, int format,
465 int name, int pitch, void *loaderPrivate)
466 {
467 struct intel_screen *screen = dri_screen->driverPrivate;
468 __DRIimage *image;
469 int cpp;
470
471 image = intel_allocate_image(screen, format, loaderPrivate);
472 if (image == NULL)
473 return NULL;
474
475 if (image->format == MESA_FORMAT_NONE)
476 cpp = 1;
477 else
478 cpp = _mesa_get_format_bytes(image->format);
479
480 image->width = width;
481 image->height = height;
482 image->pitch = pitch * cpp;
483 image->bo = brw_bo_gem_create_from_name(screen->bufmgr, "image",
484 name);
485 if (!image->bo) {
486 free(image);
487 return NULL;
488 }
489 image->modifier = tiling_to_modifier(image->bo->tiling_mode);
490
491 return image;
492 }
493
494 static __DRIimage *
495 intel_create_image_from_renderbuffer(__DRIcontext *context,
496 int renderbuffer, void *loaderPrivate)
497 {
498 __DRIimage *image;
499 struct brw_context *brw = context->driverPrivate;
500 struct gl_context *ctx = &brw->ctx;
501 struct gl_renderbuffer *rb;
502 struct intel_renderbuffer *irb;
503
504 rb = _mesa_lookup_renderbuffer(ctx, renderbuffer);
505 if (!rb) {
506 _mesa_error(ctx, GL_INVALID_OPERATION, "glRenderbufferExternalMESA");
507 return NULL;
508 }
509
510 irb = intel_renderbuffer(rb);
511 intel_miptree_make_shareable(brw, irb->mt);
512 image = calloc(1, sizeof *image);
513 if (image == NULL)
514 return NULL;
515
516 image->internal_format = rb->InternalFormat;
517 image->format = rb->Format;
518 image->modifier = tiling_to_modifier(
519 isl_tiling_to_i915_tiling(irb->mt->surf.tiling));
520 image->offset = 0;
521 image->data = loaderPrivate;
522 brw_bo_unreference(image->bo);
523 image->bo = irb->mt->bo;
524 brw_bo_reference(irb->mt->bo);
525 image->width = rb->Width;
526 image->height = rb->Height;
527 image->pitch = irb->mt->surf.row_pitch;
528 image->dri_format = driGLFormatToImageFormat(image->format);
529 image->has_depthstencil = irb->mt->stencil_mt? true : false;
530
531 rb->NeedsFinishRenderTexture = true;
532 return image;
533 }
534
535 static __DRIimage *
536 intel_create_image_from_texture(__DRIcontext *context, int target,
537 unsigned texture, int zoffset,
538 int level,
539 unsigned *error,
540 void *loaderPrivate)
541 {
542 __DRIimage *image;
543 struct brw_context *brw = context->driverPrivate;
544 struct gl_texture_object *obj;
545 struct intel_texture_object *iobj;
546 GLuint face = 0;
547
548 obj = _mesa_lookup_texture(&brw->ctx, texture);
549 if (!obj || obj->Target != target) {
550 *error = __DRI_IMAGE_ERROR_BAD_PARAMETER;
551 return NULL;
552 }
553
554 if (target == GL_TEXTURE_CUBE_MAP)
555 face = zoffset;
556
557 _mesa_test_texobj_completeness(&brw->ctx, obj);
558 iobj = intel_texture_object(obj);
559 if (!obj->_BaseComplete || (level > 0 && !obj->_MipmapComplete)) {
560 *error = __DRI_IMAGE_ERROR_BAD_PARAMETER;
561 return NULL;
562 }
563
564 if (level < obj->BaseLevel || level > obj->_MaxLevel) {
565 *error = __DRI_IMAGE_ERROR_BAD_MATCH;
566 return NULL;
567 }
568
569 if (target == GL_TEXTURE_3D && obj->Image[face][level]->Depth < zoffset) {
570 *error = __DRI_IMAGE_ERROR_BAD_MATCH;
571 return NULL;
572 }
573 image = calloc(1, sizeof *image);
574 if (image == NULL) {
575 *error = __DRI_IMAGE_ERROR_BAD_ALLOC;
576 return NULL;
577 }
578
579 image->internal_format = obj->Image[face][level]->InternalFormat;
580 image->format = obj->Image[face][level]->TexFormat;
581 image->modifier = tiling_to_modifier(
582 isl_tiling_to_i915_tiling(iobj->mt->surf.tiling));
583 image->data = loaderPrivate;
584 intel_setup_image_from_mipmap_tree(brw, image, iobj->mt, level, zoffset);
585 image->dri_format = driGLFormatToImageFormat(image->format);
586 image->has_depthstencil = iobj->mt->stencil_mt? true : false;
587 image->planar_format = iobj->planar_format;
588 if (image->dri_format == MESA_FORMAT_NONE) {
589 *error = __DRI_IMAGE_ERROR_BAD_PARAMETER;
590 free(image);
591 return NULL;
592 }
593
594 *error = __DRI_IMAGE_ERROR_SUCCESS;
595 return image;
596 }
597
598 static void
599 intel_destroy_image(__DRIimage *image)
600 {
601 brw_bo_unreference(image->bo);
602 free(image);
603 }
604
605 enum modifier_priority {
606 MODIFIER_PRIORITY_INVALID = 0,
607 MODIFIER_PRIORITY_LINEAR,
608 MODIFIER_PRIORITY_X,
609 MODIFIER_PRIORITY_Y,
610 MODIFIER_PRIORITY_Y_CCS,
611 };
612
613 const uint64_t priority_to_modifier[] = {
614 [MODIFIER_PRIORITY_INVALID] = DRM_FORMAT_MOD_INVALID,
615 [MODIFIER_PRIORITY_LINEAR] = DRM_FORMAT_MOD_LINEAR,
616 [MODIFIER_PRIORITY_X] = I915_FORMAT_MOD_X_TILED,
617 [MODIFIER_PRIORITY_Y] = I915_FORMAT_MOD_Y_TILED,
618 [MODIFIER_PRIORITY_Y_CCS] = I915_FORMAT_MOD_Y_TILED_CCS,
619 };
620
621 static uint64_t
622 select_best_modifier(struct gen_device_info *devinfo,
623 int dri_format,
624 const uint64_t *modifiers,
625 const unsigned count)
626 {
627 enum modifier_priority prio = MODIFIER_PRIORITY_INVALID;
628
629 for (int i = 0; i < count; i++) {
630 if (!modifier_is_supported(devinfo, NULL, dri_format, modifiers[i]))
631 continue;
632
633 switch (modifiers[i]) {
634 case I915_FORMAT_MOD_Y_TILED_CCS:
635 prio = MAX2(prio, MODIFIER_PRIORITY_Y_CCS);
636 break;
637 case I915_FORMAT_MOD_Y_TILED:
638 prio = MAX2(prio, MODIFIER_PRIORITY_Y);
639 break;
640 case I915_FORMAT_MOD_X_TILED:
641 prio = MAX2(prio, MODIFIER_PRIORITY_X);
642 break;
643 case DRM_FORMAT_MOD_LINEAR:
644 prio = MAX2(prio, MODIFIER_PRIORITY_LINEAR);
645 break;
646 case DRM_FORMAT_MOD_INVALID:
647 default:
648 break;
649 }
650 }
651
652 return priority_to_modifier[prio];
653 }
654
655 static __DRIimage *
656 intel_create_image_common(__DRIscreen *dri_screen,
657 int width, int height, int format,
658 unsigned int use,
659 const uint64_t *modifiers,
660 unsigned count,
661 void *loaderPrivate)
662 {
663 __DRIimage *image;
664 struct intel_screen *screen = dri_screen->driverPrivate;
665 uint64_t modifier = DRM_FORMAT_MOD_INVALID;
666 bool ok;
667
668 /* Callers of this may specify a modifier, or a dri usage, but not both. The
669 * newer modifier interface deprecates the older usage flags newer modifier
670 * interface deprecates the older usage flags.
671 */
672 assert(!(use && count));
673
674 if (use & __DRI_IMAGE_USE_CURSOR) {
675 if (width != 64 || height != 64)
676 return NULL;
677 modifier = DRM_FORMAT_MOD_LINEAR;
678 }
679
680 if (use & __DRI_IMAGE_USE_LINEAR)
681 modifier = DRM_FORMAT_MOD_LINEAR;
682
683 if (modifier == DRM_FORMAT_MOD_INVALID) {
684 if (modifiers) {
685 /* User requested specific modifiers */
686 modifier = select_best_modifier(&screen->devinfo, format,
687 modifiers, count);
688 if (modifier == DRM_FORMAT_MOD_INVALID)
689 return NULL;
690 } else {
691 /* Historically, X-tiled was the default, and so lack of modifier means
692 * X-tiled.
693 */
694 modifier = I915_FORMAT_MOD_X_TILED;
695 }
696 }
697
698 image = intel_allocate_image(screen, format, loaderPrivate);
699 if (image == NULL)
700 return NULL;
701
702 const struct isl_drm_modifier_info *mod_info =
703 isl_drm_modifier_get_info(modifier);
704
705 struct isl_surf surf;
706 ok = isl_surf_init(&screen->isl_dev, &surf,
707 .dim = ISL_SURF_DIM_2D,
708 .format = brw_isl_format_for_mesa_format(image->format),
709 .width = width,
710 .height = height,
711 .depth = 1,
712 .levels = 1,
713 .array_len = 1,
714 .samples = 1,
715 .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT |
716 ISL_SURF_USAGE_TEXTURE_BIT |
717 ISL_SURF_USAGE_STORAGE_BIT,
718 .tiling_flags = (1 << mod_info->tiling));
719 assert(ok);
720 if (!ok) {
721 free(image);
722 return NULL;
723 }
724
725 struct isl_surf aux_surf;
726 if (mod_info->aux_usage == ISL_AUX_USAGE_CCS_E) {
727 ok = isl_surf_get_ccs_surf(&screen->isl_dev, &surf, &aux_surf, 0);
728 if (!ok) {
729 free(image);
730 return NULL;
731 }
732 } else {
733 assert(mod_info->aux_usage == ISL_AUX_USAGE_NONE);
734 aux_surf.size = 0;
735 }
736
737 /* We request that the bufmgr zero the buffer for us for two reasons:
738 *
739 * 1) If a buffer gets re-used from the pool, we don't want to leak random
740 * garbage from our process to some other.
741 *
742 * 2) For images with CCS_E, we want to ensure that the CCS starts off in
743 * a valid state. A CCS value of 0 indicates that the given block is
744 * in the pass-through state which is what we want.
745 */
746 image->bo = brw_bo_alloc_tiled(screen->bufmgr, "image",
747 surf.size + aux_surf.size,
748 isl_tiling_to_i915_tiling(mod_info->tiling),
749 surf.row_pitch, BO_ALLOC_ZEROED);
750 if (image->bo == NULL) {
751 free(image);
752 return NULL;
753 }
754 image->width = width;
755 image->height = height;
756 image->pitch = surf.row_pitch;
757 image->modifier = modifier;
758
759 if (aux_surf.size) {
760 image->aux_offset = surf.size;
761 image->aux_pitch = aux_surf.row_pitch;
762 image->aux_size = aux_surf.size;
763 }
764
765 return image;
766 }
767
768 static __DRIimage *
769 intel_create_image(__DRIscreen *dri_screen,
770 int width, int height, int format,
771 unsigned int use,
772 void *loaderPrivate)
773 {
774 return intel_create_image_common(dri_screen, width, height, format, use, NULL, 0,
775 loaderPrivate);
776 }
777
778 static void *
779 intel_map_image(__DRIcontext *context, __DRIimage *image,
780 int x0, int y0, int width, int height,
781 unsigned int flags, int *stride, void **map_info)
782 {
783 struct brw_context *brw = NULL;
784 struct brw_bo *bo = NULL;
785 void *raw_data = NULL;
786 GLuint pix_w = 1;
787 GLuint pix_h = 1;
788 GLint pix_bytes = 1;
789
790 if (!context || !image || !stride || !map_info || *map_info)
791 return NULL;
792
793 if (x0 < 0 || x0 >= image->width || width > image->width - x0)
794 return NULL;
795
796 if (y0 < 0 || y0 >= image->height || height > image->height - y0)
797 return NULL;
798
799 if (flags & MAP_INTERNAL_MASK)
800 return NULL;
801
802 brw = context->driverPrivate;
803 bo = image->bo;
804
805 assert(brw);
806 assert(bo);
807
808 /* DRI flags and GL_MAP.*_BIT flags are the same, so just pass them on. */
809 raw_data = brw_bo_map(brw, bo, flags);
810 if (!raw_data)
811 return NULL;
812
813 _mesa_get_format_block_size(image->format, &pix_w, &pix_h);
814 pix_bytes = _mesa_get_format_bytes(image->format);
815
816 assert(pix_w);
817 assert(pix_h);
818 assert(pix_bytes > 0);
819
820 raw_data += (x0 / pix_w) * pix_bytes + (y0 / pix_h) * image->pitch;
821
822 brw_bo_reference(bo);
823
824 *stride = image->pitch;
825 *map_info = bo;
826
827 return raw_data;
828 }
829
830 static void
831 intel_unmap_image(__DRIcontext *context, __DRIimage *image, void *map_info)
832 {
833 struct brw_bo *bo = map_info;
834
835 brw_bo_unmap(bo);
836 brw_bo_unreference(bo);
837 }
838
839 static __DRIimage *
840 intel_create_image_with_modifiers(__DRIscreen *dri_screen,
841 int width, int height, int format,
842 const uint64_t *modifiers,
843 const unsigned count,
844 void *loaderPrivate)
845 {
846 return intel_create_image_common(dri_screen, width, height, format, 0,
847 modifiers, count, loaderPrivate);
848 }
849
850 static GLboolean
851 intel_query_image(__DRIimage *image, int attrib, int *value)
852 {
853 switch (attrib) {
854 case __DRI_IMAGE_ATTRIB_STRIDE:
855 *value = image->pitch;
856 return true;
857 case __DRI_IMAGE_ATTRIB_HANDLE:
858 *value = brw_bo_export_gem_handle(image->bo);
859 return true;
860 case __DRI_IMAGE_ATTRIB_NAME:
861 return !brw_bo_flink(image->bo, (uint32_t *) value);
862 case __DRI_IMAGE_ATTRIB_FORMAT:
863 *value = image->dri_format;
864 return true;
865 case __DRI_IMAGE_ATTRIB_WIDTH:
866 *value = image->width;
867 return true;
868 case __DRI_IMAGE_ATTRIB_HEIGHT:
869 *value = image->height;
870 return true;
871 case __DRI_IMAGE_ATTRIB_COMPONENTS:
872 if (image->planar_format == NULL)
873 return false;
874 *value = image->planar_format->components;
875 return true;
876 case __DRI_IMAGE_ATTRIB_FD:
877 return !brw_bo_gem_export_to_prime(image->bo, value);
878 case __DRI_IMAGE_ATTRIB_FOURCC:
879 return intel_image_get_fourcc(image, value);
880 case __DRI_IMAGE_ATTRIB_NUM_PLANES:
881 if (isl_drm_modifier_has_aux(image->modifier)) {
882 assert(!image->planar_format || image->planar_format->nplanes == 1);
883 *value = 2;
884 } else if (image->planar_format) {
885 *value = image->planar_format->nplanes;
886 } else {
887 *value = 1;
888 }
889 return true;
890 case __DRI_IMAGE_ATTRIB_OFFSET:
891 *value = image->offset;
892 return true;
893 case __DRI_IMAGE_ATTRIB_MODIFIER_LOWER:
894 *value = (image->modifier & 0xffffffff);
895 return true;
896 case __DRI_IMAGE_ATTRIB_MODIFIER_UPPER:
897 *value = ((image->modifier >> 32) & 0xffffffff);
898 return true;
899
900 default:
901 return false;
902 }
903 }
904
905 static GLboolean
906 intel_query_format_modifier_attribs(__DRIscreen *dri_screen,
907 uint32_t fourcc, uint64_t modifier,
908 int attrib, uint64_t *value)
909 {
910 struct intel_screen *screen = dri_screen->driverPrivate;
911 const struct intel_image_format *f = intel_image_format_lookup(fourcc);
912
913 if (!modifier_is_supported(&screen->devinfo, f, 0, modifier))
914 return false;
915
916 switch (attrib) {
917 case __DRI_IMAGE_FORMAT_MODIFIER_ATTRIB_PLANE_COUNT:
918 *value = isl_drm_modifier_has_aux(modifier) ? 2 : f->nplanes;
919 return true;
920
921 default:
922 return false;
923 }
924 }
925
926 static __DRIimage *
927 intel_dup_image(__DRIimage *orig_image, void *loaderPrivate)
928 {
929 __DRIimage *image;
930
931 image = calloc(1, sizeof *image);
932 if (image == NULL)
933 return NULL;
934
935 brw_bo_reference(orig_image->bo);
936 image->bo = orig_image->bo;
937 image->internal_format = orig_image->internal_format;
938 image->planar_format = orig_image->planar_format;
939 image->dri_format = orig_image->dri_format;
940 image->format = orig_image->format;
941 image->modifier = orig_image->modifier;
942 image->offset = orig_image->offset;
943 image->width = orig_image->width;
944 image->height = orig_image->height;
945 image->pitch = orig_image->pitch;
946 image->tile_x = orig_image->tile_x;
947 image->tile_y = orig_image->tile_y;
948 image->has_depthstencil = orig_image->has_depthstencil;
949 image->data = loaderPrivate;
950 image->dma_buf_imported = orig_image->dma_buf_imported;
951 image->aux_offset = orig_image->aux_offset;
952 image->aux_pitch = orig_image->aux_pitch;
953
954 memcpy(image->strides, orig_image->strides, sizeof(image->strides));
955 memcpy(image->offsets, orig_image->offsets, sizeof(image->offsets));
956
957 return image;
958 }
959
960 static GLboolean
961 intel_validate_usage(__DRIimage *image, unsigned int use)
962 {
963 if (use & __DRI_IMAGE_USE_CURSOR) {
964 if (image->width != 64 || image->height != 64)
965 return GL_FALSE;
966 }
967
968 return GL_TRUE;
969 }
970
971 static __DRIimage *
972 intel_create_image_from_names(__DRIscreen *dri_screen,
973 int width, int height, int fourcc,
974 int *names, int num_names,
975 int *strides, int *offsets,
976 void *loaderPrivate)
977 {
978 const struct intel_image_format *f = NULL;
979 __DRIimage *image;
980 int i, index;
981
982 if (dri_screen == NULL || names == NULL || num_names != 1)
983 return NULL;
984
985 f = intel_image_format_lookup(fourcc);
986 if (f == NULL)
987 return NULL;
988
989 image = intel_create_image_from_name(dri_screen, width, height,
990 __DRI_IMAGE_FORMAT_NONE,
991 names[0], strides[0],
992 loaderPrivate);
993
994 if (image == NULL)
995 return NULL;
996
997 image->planar_format = f;
998 for (i = 0; i < f->nplanes; i++) {
999 index = f->planes[i].buffer_index;
1000 image->offsets[index] = offsets[index];
1001 image->strides[index] = strides[index];
1002 }
1003
1004 return image;
1005 }
1006
1007 static __DRIimage *
1008 intel_create_image_from_fds_common(__DRIscreen *dri_screen,
1009 int width, int height, int fourcc,
1010 uint64_t modifier, int *fds, int num_fds,
1011 int *strides, int *offsets,
1012 void *loaderPrivate)
1013 {
1014 struct intel_screen *screen = dri_screen->driverPrivate;
1015 const struct intel_image_format *f;
1016 __DRIimage *image;
1017 int i, index;
1018 bool ok;
1019
1020 if (fds == NULL || num_fds < 1)
1021 return NULL;
1022
1023 f = intel_image_format_lookup(fourcc);
1024 if (f == NULL)
1025 return NULL;
1026
1027 if (modifier != DRM_FORMAT_MOD_INVALID &&
1028 !modifier_is_supported(&screen->devinfo, f, 0, modifier))
1029 return NULL;
1030
1031 if (f->nplanes == 1)
1032 image = intel_allocate_image(screen, f->planes[0].dri_format,
1033 loaderPrivate);
1034 else
1035 image = intel_allocate_image(screen, __DRI_IMAGE_FORMAT_NONE,
1036 loaderPrivate);
1037
1038 if (image == NULL)
1039 return NULL;
1040
1041 image->width = width;
1042 image->height = height;
1043 image->pitch = strides[0];
1044
1045 image->planar_format = f;
1046
1047 if (modifier != DRM_FORMAT_MOD_INVALID) {
1048 const struct isl_drm_modifier_info *mod_info =
1049 isl_drm_modifier_get_info(modifier);
1050 uint32_t tiling = isl_tiling_to_i915_tiling(mod_info->tiling);
1051 image->bo = brw_bo_gem_create_from_prime_tiled(screen->bufmgr, fds[0],
1052 tiling, strides[0]);
1053 } else {
1054 image->bo = brw_bo_gem_create_from_prime(screen->bufmgr, fds[0]);
1055 }
1056
1057 if (image->bo == NULL) {
1058 free(image);
1059 return NULL;
1060 }
1061
1062 /* We only support all planes from the same bo.
1063 * brw_bo_gem_create_from_prime() should return the same pointer for all
1064 * fds received here */
1065 for (i = 1; i < num_fds; i++) {
1066 struct brw_bo *aux = brw_bo_gem_create_from_prime(screen->bufmgr, fds[i]);
1067 brw_bo_unreference(aux);
1068 if (aux != image->bo) {
1069 brw_bo_unreference(image->bo);
1070 free(image);
1071 return NULL;
1072 }
1073 }
1074
1075 if (modifier != DRM_FORMAT_MOD_INVALID)
1076 image->modifier = modifier;
1077 else
1078 image->modifier = tiling_to_modifier(image->bo->tiling_mode);
1079
1080 const struct isl_drm_modifier_info *mod_info =
1081 isl_drm_modifier_get_info(image->modifier);
1082
1083 int size = 0;
1084 struct isl_surf surf;
1085 for (i = 0; i < f->nplanes; i++) {
1086 index = f->planes[i].buffer_index;
1087 image->offsets[index] = offsets[index];
1088 image->strides[index] = strides[index];
1089
1090 mesa_format format = driImageFormatToGLFormat(f->planes[i].dri_format);
1091
1092 ok = isl_surf_init(&screen->isl_dev, &surf,
1093 .dim = ISL_SURF_DIM_2D,
1094 .format = brw_isl_format_for_mesa_format(format),
1095 .width = image->width >> f->planes[i].width_shift,
1096 .height = image->height >> f->planes[i].height_shift,
1097 .depth = 1,
1098 .levels = 1,
1099 .array_len = 1,
1100 .samples = 1,
1101 .row_pitch = strides[index],
1102 .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT |
1103 ISL_SURF_USAGE_TEXTURE_BIT |
1104 ISL_SURF_USAGE_STORAGE_BIT,
1105 .tiling_flags = (1 << mod_info->tiling));
1106 if (!ok) {
1107 brw_bo_unreference(image->bo);
1108 free(image);
1109 return NULL;
1110 }
1111
1112 const int end = offsets[index] + surf.size;
1113 if (size < end)
1114 size = end;
1115 }
1116
1117 if (mod_info->aux_usage == ISL_AUX_USAGE_CCS_E) {
1118 /* Even though we initialize surf in the loop above, we know that
1119 * anything with CCS_E will have exactly one plane so surf is properly
1120 * initialized when we get here.
1121 */
1122 assert(f->nplanes == 1);
1123
1124 image->aux_offset = offsets[1];
1125 image->aux_pitch = strides[1];
1126
1127 /* Scanout hardware requires that the CCS be placed after the main
1128 * surface in memory. We consider any CCS that is placed any earlier in
1129 * memory to be invalid and reject it.
1130 *
1131 * At some point in the future, this restriction may be relaxed if the
1132 * hardware becomes less strict but we may need a new modifier for that.
1133 */
1134 assert(size > 0);
1135 if (image->aux_offset < size) {
1136 brw_bo_unreference(image->bo);
1137 free(image);
1138 return NULL;
1139 }
1140
1141 struct isl_surf aux_surf;
1142 ok = isl_surf_get_ccs_surf(&screen->isl_dev, &surf, &aux_surf,
1143 image->aux_pitch);
1144 if (!ok) {
1145 brw_bo_unreference(image->bo);
1146 free(image);
1147 return NULL;
1148 }
1149
1150 image->aux_size = aux_surf.size;
1151
1152 const int end = image->aux_offset + aux_surf.size;
1153 if (size < end)
1154 size = end;
1155 } else {
1156 assert(mod_info->aux_usage == ISL_AUX_USAGE_NONE);
1157 }
1158
1159 /* Check that the requested image actually fits within the BO. 'size'
1160 * is already relative to the offsets, so we don't need to add that. */
1161 if (image->bo->size == 0) {
1162 image->bo->size = size;
1163 } else if (size > image->bo->size) {
1164 brw_bo_unreference(image->bo);
1165 free(image);
1166 return NULL;
1167 }
1168
1169 if (f->nplanes == 1) {
1170 image->offset = image->offsets[0];
1171 intel_image_warn_if_unaligned(image, __func__);
1172 }
1173
1174 return image;
1175 }
1176
1177 static __DRIimage *
1178 intel_create_image_from_fds(__DRIscreen *dri_screen,
1179 int width, int height, int fourcc,
1180 int *fds, int num_fds, int *strides, int *offsets,
1181 void *loaderPrivate)
1182 {
1183 return intel_create_image_from_fds_common(dri_screen, width, height, fourcc,
1184 DRM_FORMAT_MOD_INVALID,
1185 fds, num_fds, strides, offsets,
1186 loaderPrivate);
1187 }
1188
1189 static __DRIimage *
1190 intel_create_image_from_dma_bufs2(__DRIscreen *dri_screen,
1191 int width, int height,
1192 int fourcc, uint64_t modifier,
1193 int *fds, int num_fds,
1194 int *strides, int *offsets,
1195 enum __DRIYUVColorSpace yuv_color_space,
1196 enum __DRISampleRange sample_range,
1197 enum __DRIChromaSiting horizontal_siting,
1198 enum __DRIChromaSiting vertical_siting,
1199 unsigned *error,
1200 void *loaderPrivate)
1201 {
1202 __DRIimage *image;
1203 const struct intel_image_format *f = intel_image_format_lookup(fourcc);
1204
1205 if (!f) {
1206 *error = __DRI_IMAGE_ERROR_BAD_MATCH;
1207 return NULL;
1208 }
1209
1210 image = intel_create_image_from_fds_common(dri_screen, width, height,
1211 fourcc, modifier,
1212 fds, num_fds, strides, offsets,
1213 loaderPrivate);
1214
1215 /*
1216 * Invalid parameters and any inconsistencies between are assumed to be
1217 * checked by the caller. Therefore besides unsupported formats one can fail
1218 * only in allocation.
1219 */
1220 if (!image) {
1221 *error = __DRI_IMAGE_ERROR_BAD_ALLOC;
1222 return NULL;
1223 }
1224
1225 image->dma_buf_imported = true;
1226 image->yuv_color_space = yuv_color_space;
1227 image->sample_range = sample_range;
1228 image->horizontal_siting = horizontal_siting;
1229 image->vertical_siting = vertical_siting;
1230
1231 *error = __DRI_IMAGE_ERROR_SUCCESS;
1232 return image;
1233 }
1234
1235 static __DRIimage *
1236 intel_create_image_from_dma_bufs(__DRIscreen *dri_screen,
1237 int width, int height, int fourcc,
1238 int *fds, int num_fds,
1239 int *strides, int *offsets,
1240 enum __DRIYUVColorSpace yuv_color_space,
1241 enum __DRISampleRange sample_range,
1242 enum __DRIChromaSiting horizontal_siting,
1243 enum __DRIChromaSiting vertical_siting,
1244 unsigned *error,
1245 void *loaderPrivate)
1246 {
1247 return intel_create_image_from_dma_bufs2(dri_screen, width, height,
1248 fourcc, DRM_FORMAT_MOD_INVALID,
1249 fds, num_fds, strides, offsets,
1250 yuv_color_space,
1251 sample_range,
1252 horizontal_siting,
1253 vertical_siting,
1254 error,
1255 loaderPrivate);
1256 }
1257
1258 static GLboolean
1259 intel_query_dma_buf_formats(__DRIscreen *screen, int max,
1260 int *formats, int *count)
1261 {
1262 int i, j = 0;
1263
1264 if (max == 0) {
1265 /* Note, sRGB formats not included. */
1266 *count = ARRAY_SIZE(intel_image_formats) - 2;
1267 return true;
1268 }
1269
1270 for (i = 0; i < (ARRAY_SIZE(intel_image_formats)) && j < max; i++) {
1271 if (intel_image_formats[i].fourcc == __DRI_IMAGE_FOURCC_SARGB8888 ||
1272 intel_image_formats[i].fourcc == __DRI_IMAGE_FOURCC_SABGR8888)
1273 continue;
1274 formats[j++] = intel_image_formats[i].fourcc;
1275 }
1276
1277 *count = j;
1278 return true;
1279 }
1280
1281 static GLboolean
1282 intel_query_dma_buf_modifiers(__DRIscreen *_screen, int fourcc, int max,
1283 uint64_t *modifiers,
1284 unsigned int *external_only,
1285 int *count)
1286 {
1287 struct intel_screen *screen = _screen->driverPrivate;
1288 const struct intel_image_format *f;
1289 int num_mods = 0, i;
1290
1291 f = intel_image_format_lookup(fourcc);
1292 if (f == NULL)
1293 return false;
1294
1295 for (i = 0; i < ARRAY_SIZE(supported_modifiers); i++) {
1296 uint64_t modifier = supported_modifiers[i].modifier;
1297 if (!modifier_is_supported(&screen->devinfo, f, 0, modifier))
1298 continue;
1299
1300 num_mods++;
1301 if (max == 0)
1302 continue;
1303
1304 modifiers[num_mods - 1] = modifier;
1305 if (num_mods >= max)
1306 break;
1307 }
1308
1309 if (external_only != NULL) {
1310 for (i = 0; i < num_mods && i < max; i++) {
1311 if (f->components == __DRI_IMAGE_COMPONENTS_Y_U_V ||
1312 f->components == __DRI_IMAGE_COMPONENTS_Y_UV ||
1313 f->components == __DRI_IMAGE_COMPONENTS_Y_XUXV) {
1314 external_only[i] = GL_TRUE;
1315 }
1316 else {
1317 external_only[i] = GL_FALSE;
1318 }
1319 }
1320 }
1321
1322 *count = num_mods;
1323 return true;
1324 }
1325
1326 static __DRIimage *
1327 intel_from_planar(__DRIimage *parent, int plane, void *loaderPrivate)
1328 {
1329 int width, height, offset, stride, size, dri_format;
1330 __DRIimage *image;
1331
1332 if (parent == NULL)
1333 return NULL;
1334
1335 width = parent->width;
1336 height = parent->height;
1337
1338 const struct intel_image_format *f = parent->planar_format;
1339
1340 if (f && plane < f->nplanes) {
1341 /* Use the planar format definition. */
1342 width >>= f->planes[plane].width_shift;
1343 height >>= f->planes[plane].height_shift;
1344 dri_format = f->planes[plane].dri_format;
1345 int index = f->planes[plane].buffer_index;
1346 offset = parent->offsets[index];
1347 stride = parent->strides[index];
1348 size = height * stride;
1349 } else if (plane == 0) {
1350 /* The only plane of a non-planar image: copy the parent definition
1351 * directly. */
1352 dri_format = parent->dri_format;
1353 offset = parent->offset;
1354 stride = parent->pitch;
1355 size = height * stride;
1356 } else if (plane == 1 && parent->modifier != DRM_FORMAT_MOD_INVALID &&
1357 isl_drm_modifier_has_aux(parent->modifier)) {
1358 /* Auxiliary plane */
1359 dri_format = parent->dri_format;
1360 offset = parent->aux_offset;
1361 stride = parent->aux_pitch;
1362 size = parent->aux_size;
1363 } else {
1364 return NULL;
1365 }
1366
1367 if (offset + size > parent->bo->size) {
1368 _mesa_warning(NULL, "intel_from_planar: subimage out of bounds");
1369 return NULL;
1370 }
1371
1372 image = intel_allocate_image(parent->screen, dri_format, loaderPrivate);
1373 if (image == NULL)
1374 return NULL;
1375
1376 image->bo = parent->bo;
1377 brw_bo_reference(parent->bo);
1378 image->modifier = parent->modifier;
1379
1380 image->width = width;
1381 image->height = height;
1382 image->pitch = stride;
1383 image->offset = offset;
1384
1385 intel_image_warn_if_unaligned(image, __func__);
1386
1387 return image;
1388 }
1389
1390 static const __DRIimageExtension intelImageExtension = {
1391 .base = { __DRI_IMAGE, 16 },
1392
1393 .createImageFromName = intel_create_image_from_name,
1394 .createImageFromRenderbuffer = intel_create_image_from_renderbuffer,
1395 .destroyImage = intel_destroy_image,
1396 .createImage = intel_create_image,
1397 .queryImage = intel_query_image,
1398 .dupImage = intel_dup_image,
1399 .validateUsage = intel_validate_usage,
1400 .createImageFromNames = intel_create_image_from_names,
1401 .fromPlanar = intel_from_planar,
1402 .createImageFromTexture = intel_create_image_from_texture,
1403 .createImageFromFds = intel_create_image_from_fds,
1404 .createImageFromDmaBufs = intel_create_image_from_dma_bufs,
1405 .blitImage = NULL,
1406 .getCapabilities = NULL,
1407 .mapImage = intel_map_image,
1408 .unmapImage = intel_unmap_image,
1409 .createImageWithModifiers = intel_create_image_with_modifiers,
1410 .createImageFromDmaBufs2 = intel_create_image_from_dma_bufs2,
1411 .queryDmaBufFormats = intel_query_dma_buf_formats,
1412 .queryDmaBufModifiers = intel_query_dma_buf_modifiers,
1413 .queryDmaBufFormatModifierAttribs = intel_query_format_modifier_attribs,
1414 };
1415
1416 static uint64_t
1417 get_aperture_size(int fd)
1418 {
1419 struct drm_i915_gem_get_aperture aperture;
1420
1421 if (drmIoctl(fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture) != 0)
1422 return 0;
1423
1424 return aperture.aper_size;
1425 }
1426
1427 static int
1428 brw_query_renderer_integer(__DRIscreen *dri_screen,
1429 int param, unsigned int *value)
1430 {
1431 const struct intel_screen *const screen =
1432 (struct intel_screen *) dri_screen->driverPrivate;
1433
1434 switch (param) {
1435 case __DRI2_RENDERER_VENDOR_ID:
1436 value[0] = 0x8086;
1437 return 0;
1438 case __DRI2_RENDERER_DEVICE_ID:
1439 value[0] = screen->deviceID;
1440 return 0;
1441 case __DRI2_RENDERER_ACCELERATED:
1442 value[0] = 1;
1443 return 0;
1444 case __DRI2_RENDERER_VIDEO_MEMORY: {
1445 /* Once a batch uses more than 75% of the maximum mappable size, we
1446 * assume that there's some fragmentation, and we start doing extra
1447 * flushing, etc. That's the big cliff apps will care about.
1448 */
1449 const unsigned gpu_mappable_megabytes =
1450 screen->aperture_threshold / (1024 * 1024);
1451
1452 const long system_memory_pages = sysconf(_SC_PHYS_PAGES);
1453 const long system_page_size = sysconf(_SC_PAGE_SIZE);
1454
1455 if (system_memory_pages <= 0 || system_page_size <= 0)
1456 return -1;
1457
1458 const uint64_t system_memory_bytes = (uint64_t) system_memory_pages
1459 * (uint64_t) system_page_size;
1460
1461 const unsigned system_memory_megabytes =
1462 (unsigned) (system_memory_bytes / (1024 * 1024));
1463
1464 value[0] = MIN2(system_memory_megabytes, gpu_mappable_megabytes);
1465 return 0;
1466 }
1467 case __DRI2_RENDERER_UNIFIED_MEMORY_ARCHITECTURE:
1468 value[0] = 1;
1469 return 0;
1470 case __DRI2_RENDERER_HAS_TEXTURE_3D:
1471 value[0] = 1;
1472 return 0;
1473 case __DRI2_RENDERER_HAS_CONTEXT_PRIORITY:
1474 value[0] = 0;
1475 if (brw_hw_context_set_priority(screen->bufmgr,
1476 0, GEN_CONTEXT_HIGH_PRIORITY) == 0)
1477 value[0] |= __DRI2_RENDERER_HAS_CONTEXT_PRIORITY_HIGH;
1478 if (brw_hw_context_set_priority(screen->bufmgr,
1479 0, GEN_CONTEXT_LOW_PRIORITY) == 0)
1480 value[0] |= __DRI2_RENDERER_HAS_CONTEXT_PRIORITY_LOW;
1481 /* reset to default last, just in case */
1482 if (brw_hw_context_set_priority(screen->bufmgr,
1483 0, GEN_CONTEXT_MEDIUM_PRIORITY) == 0)
1484 value[0] |= __DRI2_RENDERER_HAS_CONTEXT_PRIORITY_MEDIUM;
1485 return 0;
1486 case __DRI2_RENDERER_HAS_FRAMEBUFFER_SRGB:
1487 value[0] = 1;
1488 return 0;
1489 default:
1490 return driQueryRendererIntegerCommon(dri_screen, param, value);
1491 }
1492
1493 return -1;
1494 }
1495
1496 static int
1497 brw_query_renderer_string(__DRIscreen *dri_screen,
1498 int param, const char **value)
1499 {
1500 const struct intel_screen *screen =
1501 (struct intel_screen *) dri_screen->driverPrivate;
1502
1503 switch (param) {
1504 case __DRI2_RENDERER_VENDOR_ID:
1505 value[0] = brw_vendor_string;
1506 return 0;
1507 case __DRI2_RENDERER_DEVICE_ID:
1508 value[0] = brw_get_renderer_string(screen);
1509 return 0;
1510 default:
1511 break;
1512 }
1513
1514 return -1;
1515 }
1516
1517 static void
1518 brw_set_cache_funcs(__DRIscreen *dri_screen,
1519 __DRIblobCacheSet set, __DRIblobCacheGet get)
1520 {
1521 const struct intel_screen *const screen =
1522 (struct intel_screen *) dri_screen->driverPrivate;
1523
1524 if (!screen->disk_cache)
1525 return;
1526
1527 disk_cache_set_callbacks(screen->disk_cache, set, get);
1528 }
1529
1530 static const __DRI2rendererQueryExtension intelRendererQueryExtension = {
1531 .base = { __DRI2_RENDERER_QUERY, 1 },
1532
1533 .queryInteger = brw_query_renderer_integer,
1534 .queryString = brw_query_renderer_string
1535 };
1536
1537 static const __DRIrobustnessExtension dri2Robustness = {
1538 .base = { __DRI2_ROBUSTNESS, 1 }
1539 };
1540
1541 static const __DRI2blobExtension intelBlobExtension = {
1542 .base = { __DRI2_BLOB, 1 },
1543 .set_cache_funcs = brw_set_cache_funcs
1544 };
1545
1546 static const __DRIextension *screenExtensions[] = {
1547 &intelTexBufferExtension.base,
1548 &intelFenceExtension.base,
1549 &intelFlushExtension.base,
1550 &intelImageExtension.base,
1551 &intelRendererQueryExtension.base,
1552 &dri2ConfigQueryExtension.base,
1553 &dri2NoErrorExtension.base,
1554 &intelBlobExtension.base,
1555 NULL
1556 };
1557
1558 static const __DRIextension *intelRobustScreenExtensions[] = {
1559 &intelTexBufferExtension.base,
1560 &intelFenceExtension.base,
1561 &intelFlushExtension.base,
1562 &intelImageExtension.base,
1563 &intelRendererQueryExtension.base,
1564 &dri2ConfigQueryExtension.base,
1565 &dri2Robustness.base,
1566 &dri2NoErrorExtension.base,
1567 &intelBlobExtension.base,
1568 NULL
1569 };
1570
1571 static int
1572 intel_get_param(struct intel_screen *screen, int param, int *value)
1573 {
1574 int ret = 0;
1575 struct drm_i915_getparam gp;
1576
1577 memset(&gp, 0, sizeof(gp));
1578 gp.param = param;
1579 gp.value = value;
1580
1581 if (drmIoctl(screen->driScrnPriv->fd, DRM_IOCTL_I915_GETPARAM, &gp) == -1) {
1582 ret = -errno;
1583 if (ret != -EINVAL)
1584 _mesa_warning(NULL, "drm_i915_getparam: %d", ret);
1585 }
1586
1587 return ret;
1588 }
1589
1590 static bool
1591 intel_get_boolean(struct intel_screen *screen, int param)
1592 {
1593 int value = 0;
1594 return (intel_get_param(screen, param, &value) == 0) && value;
1595 }
1596
1597 static int
1598 intel_get_integer(struct intel_screen *screen, int param)
1599 {
1600 int value = -1;
1601
1602 if (intel_get_param(screen, param, &value) == 0)
1603 return value;
1604
1605 return -1;
1606 }
1607
1608 static void
1609 intelDestroyScreen(__DRIscreen * sPriv)
1610 {
1611 struct intel_screen *screen = sPriv->driverPrivate;
1612
1613 brw_bufmgr_destroy(screen->bufmgr);
1614 driDestroyOptionInfo(&screen->optionCache);
1615
1616 disk_cache_destroy(screen->disk_cache);
1617
1618 ralloc_free(screen);
1619 sPriv->driverPrivate = NULL;
1620 }
1621
1622
1623 /**
1624 * Create a gl_framebuffer and attach it to __DRIdrawable::driverPrivate.
1625 *
1626 *_This implements driDriverAPI::createNewDrawable, which the DRI layer calls
1627 * when creating a EGLSurface, GLXDrawable, or GLXPixmap. Despite the name,
1628 * this does not allocate GPU memory.
1629 */
1630 static GLboolean
1631 intelCreateBuffer(__DRIscreen *dri_screen,
1632 __DRIdrawable * driDrawPriv,
1633 const struct gl_config * mesaVis, GLboolean isPixmap)
1634 {
1635 struct intel_renderbuffer *rb;
1636 struct intel_screen *screen = (struct intel_screen *)
1637 dri_screen->driverPrivate;
1638 mesa_format rgbFormat;
1639 unsigned num_samples =
1640 intel_quantize_num_samples(screen, mesaVis->samples);
1641
1642 if (isPixmap)
1643 return false;
1644
1645 struct gl_framebuffer *fb = CALLOC_STRUCT(gl_framebuffer);
1646 if (!fb)
1647 return false;
1648
1649 _mesa_initialize_window_framebuffer(fb, mesaVis);
1650
1651 if (screen->winsys_msaa_samples_override != -1) {
1652 num_samples = screen->winsys_msaa_samples_override;
1653 fb->Visual.samples = num_samples;
1654 }
1655
1656 if (mesaVis->redBits == 10 && mesaVis->alphaBits > 0) {
1657 rgbFormat = mesaVis->redMask == 0x3ff00000 ? MESA_FORMAT_B10G10R10A2_UNORM
1658 : MESA_FORMAT_R10G10B10A2_UNORM;
1659 } else if (mesaVis->redBits == 10) {
1660 rgbFormat = mesaVis->redMask == 0x3ff00000 ? MESA_FORMAT_B10G10R10X2_UNORM
1661 : MESA_FORMAT_R10G10B10X2_UNORM;
1662 } else if (mesaVis->redBits == 5) {
1663 rgbFormat = mesaVis->redMask == 0x1f ? MESA_FORMAT_R5G6B5_UNORM
1664 : MESA_FORMAT_B5G6R5_UNORM;
1665 } else if (mesaVis->sRGBCapable) {
1666 rgbFormat = mesaVis->redMask == 0xff ? MESA_FORMAT_R8G8B8A8_SRGB
1667 : MESA_FORMAT_B8G8R8A8_SRGB;
1668 } else if (mesaVis->alphaBits == 0) {
1669 rgbFormat = mesaVis->redMask == 0xff ? MESA_FORMAT_R8G8B8X8_UNORM
1670 : MESA_FORMAT_B8G8R8X8_UNORM;
1671 } else {
1672 rgbFormat = mesaVis->redMask == 0xff ? MESA_FORMAT_R8G8B8A8_SRGB
1673 : MESA_FORMAT_B8G8R8A8_SRGB;
1674 fb->Visual.sRGBCapable = true;
1675 }
1676
1677 /* mesaVis->sRGBCapable was set, user is asking for sRGB */
1678 bool srgb_cap_set = mesaVis->redBits >= 8 && mesaVis->sRGBCapable;
1679
1680 /* setup the hardware-based renderbuffers */
1681 rb = intel_create_winsys_renderbuffer(screen, rgbFormat, num_samples);
1682 _mesa_attach_and_own_rb(fb, BUFFER_FRONT_LEFT, &rb->Base.Base);
1683 rb->need_srgb = srgb_cap_set;
1684
1685 if (mesaVis->doubleBufferMode) {
1686 rb = intel_create_winsys_renderbuffer(screen, rgbFormat, num_samples);
1687 _mesa_attach_and_own_rb(fb, BUFFER_BACK_LEFT, &rb->Base.Base);
1688 rb->need_srgb = srgb_cap_set;
1689 }
1690
1691 /*
1692 * Assert here that the gl_config has an expected depth/stencil bit
1693 * combination: one of d24/s8, d16/s0, d0/s0. (See intelInitScreen2(),
1694 * which constructs the advertised configs.)
1695 */
1696 if (mesaVis->depthBits == 24) {
1697 assert(mesaVis->stencilBits == 8);
1698
1699 if (screen->devinfo.has_hiz_and_separate_stencil) {
1700 rb = intel_create_private_renderbuffer(screen,
1701 MESA_FORMAT_Z24_UNORM_X8_UINT,
1702 num_samples);
1703 _mesa_attach_and_own_rb(fb, BUFFER_DEPTH, &rb->Base.Base);
1704 rb = intel_create_private_renderbuffer(screen, MESA_FORMAT_S_UINT8,
1705 num_samples);
1706 _mesa_attach_and_own_rb(fb, BUFFER_STENCIL, &rb->Base.Base);
1707 } else {
1708 /*
1709 * Use combined depth/stencil. Note that the renderbuffer is
1710 * attached to two attachment points.
1711 */
1712 rb = intel_create_private_renderbuffer(screen,
1713 MESA_FORMAT_Z24_UNORM_S8_UINT,
1714 num_samples);
1715 _mesa_attach_and_own_rb(fb, BUFFER_DEPTH, &rb->Base.Base);
1716 _mesa_attach_and_reference_rb(fb, BUFFER_STENCIL, &rb->Base.Base);
1717 }
1718 }
1719 else if (mesaVis->depthBits == 16) {
1720 assert(mesaVis->stencilBits == 0);
1721 rb = intel_create_private_renderbuffer(screen, MESA_FORMAT_Z_UNORM16,
1722 num_samples);
1723 _mesa_attach_and_own_rb(fb, BUFFER_DEPTH, &rb->Base.Base);
1724 }
1725 else {
1726 assert(mesaVis->depthBits == 0);
1727 assert(mesaVis->stencilBits == 0);
1728 }
1729
1730 /* now add any/all software-based renderbuffers we may need */
1731 _swrast_add_soft_renderbuffers(fb,
1732 false, /* never sw color */
1733 false, /* never sw depth */
1734 false, /* never sw stencil */
1735 mesaVis->accumRedBits > 0,
1736 false, /* never sw alpha */
1737 false /* never sw aux */ );
1738 driDrawPriv->driverPrivate = fb;
1739
1740 return true;
1741 }
1742
1743 static void
1744 intelDestroyBuffer(__DRIdrawable * driDrawPriv)
1745 {
1746 struct gl_framebuffer *fb = driDrawPriv->driverPrivate;
1747
1748 _mesa_reference_framebuffer(&fb, NULL);
1749 }
1750
1751 static void
1752 intel_cs_timestamp_frequency(struct intel_screen *screen)
1753 {
1754 /* We shouldn't need to update gen_device_info.timestamp_frequency prior to
1755 * gen10, PCI-id is enough to figure it out.
1756 */
1757 assert(screen->devinfo.gen >= 10);
1758
1759 int ret, freq;
1760
1761 ret = intel_get_param(screen, I915_PARAM_CS_TIMESTAMP_FREQUENCY,
1762 &freq);
1763 if (ret < 0) {
1764 _mesa_warning(NULL,
1765 "Kernel 4.15 required to read the CS timestamp frequency.\n");
1766 return;
1767 }
1768
1769 screen->devinfo.timestamp_frequency = freq;
1770 }
1771
1772 static void
1773 intel_detect_sseu(struct intel_screen *screen)
1774 {
1775 assert(screen->devinfo.gen >= 8);
1776 int ret;
1777
1778 screen->subslice_total = -1;
1779 screen->eu_total = -1;
1780
1781 ret = intel_get_param(screen, I915_PARAM_SUBSLICE_TOTAL,
1782 &screen->subslice_total);
1783 if (ret < 0 && ret != -EINVAL)
1784 goto err_out;
1785
1786 ret = intel_get_param(screen,
1787 I915_PARAM_EU_TOTAL, &screen->eu_total);
1788 if (ret < 0 && ret != -EINVAL)
1789 goto err_out;
1790
1791 /* Without this information, we cannot get the right Braswell brandstrings,
1792 * and we have to use conservative numbers for GPGPU on many platforms, but
1793 * otherwise, things will just work.
1794 */
1795 if (screen->subslice_total < 1 || screen->eu_total < 1)
1796 _mesa_warning(NULL,
1797 "Kernel 4.1 required to properly query GPU properties.\n");
1798
1799 return;
1800
1801 err_out:
1802 screen->subslice_total = -1;
1803 screen->eu_total = -1;
1804 _mesa_warning(NULL, "Failed to query GPU properties (%s).\n", strerror(-ret));
1805 }
1806
1807 static bool
1808 intel_init_bufmgr(struct intel_screen *screen)
1809 {
1810 __DRIscreen *dri_screen = screen->driScrnPriv;
1811
1812 if (getenv("INTEL_NO_HW") != NULL)
1813 screen->no_hw = true;
1814
1815 screen->bufmgr = brw_bufmgr_init(&screen->devinfo, dri_screen->fd);
1816 if (screen->bufmgr == NULL) {
1817 fprintf(stderr, "[%s:%u] Error initializing buffer manager.\n",
1818 __func__, __LINE__);
1819 return false;
1820 }
1821
1822 if (!intel_get_boolean(screen, I915_PARAM_HAS_EXEC_NO_RELOC)) {
1823 fprintf(stderr, "[%s: %u] Kernel 3.9 required.\n", __func__, __LINE__);
1824 return false;
1825 }
1826
1827 return true;
1828 }
1829
1830 static bool
1831 intel_detect_swizzling(struct intel_screen *screen)
1832 {
1833 uint32_t tiling = I915_TILING_X;
1834 uint32_t swizzle_mode = 0;
1835 struct brw_bo *buffer =
1836 brw_bo_alloc_tiled(screen->bufmgr, "swizzle test", 32768,
1837 tiling, 512, 0);
1838 if (buffer == NULL)
1839 return false;
1840
1841 brw_bo_get_tiling(buffer, &tiling, &swizzle_mode);
1842 brw_bo_unreference(buffer);
1843
1844 return swizzle_mode != I915_BIT_6_SWIZZLE_NONE;
1845 }
1846
1847 static int
1848 intel_detect_timestamp(struct intel_screen *screen)
1849 {
1850 uint64_t dummy = 0, last = 0;
1851 int upper, lower, loops;
1852
1853 /* On 64bit systems, some old kernels trigger a hw bug resulting in the
1854 * TIMESTAMP register being shifted and the low 32bits always zero.
1855 *
1856 * More recent kernels offer an interface to read the full 36bits
1857 * everywhere.
1858 */
1859 if (brw_reg_read(screen->bufmgr, TIMESTAMP | 1, &dummy) == 0)
1860 return 3;
1861
1862 /* Determine if we have a 32bit or 64bit kernel by inspecting the
1863 * upper 32bits for a rapidly changing timestamp.
1864 */
1865 if (brw_reg_read(screen->bufmgr, TIMESTAMP, &last))
1866 return 0;
1867
1868 upper = lower = 0;
1869 for (loops = 0; loops < 10; loops++) {
1870 /* The TIMESTAMP should change every 80ns, so several round trips
1871 * through the kernel should be enough to advance it.
1872 */
1873 if (brw_reg_read(screen->bufmgr, TIMESTAMP, &dummy))
1874 return 0;
1875
1876 upper += (dummy >> 32) != (last >> 32);
1877 if (upper > 1) /* beware 32bit counter overflow */
1878 return 2; /* upper dword holds the low 32bits of the timestamp */
1879
1880 lower += (dummy & 0xffffffff) != (last & 0xffffffff);
1881 if (lower > 1)
1882 return 1; /* timestamp is unshifted */
1883
1884 last = dummy;
1885 }
1886
1887 /* No advancement? No timestamp! */
1888 return 0;
1889 }
1890
1891 /**
1892 * Test if we can use MI_LOAD_REGISTER_MEM from an untrusted batchbuffer.
1893 *
1894 * Some combinations of hardware and kernel versions allow this feature,
1895 * while others don't. Instead of trying to enumerate every case, just
1896 * try and write a register and see if works.
1897 */
1898 static bool
1899 intel_detect_pipelined_register(struct intel_screen *screen,
1900 int reg, uint32_t expected_value, bool reset)
1901 {
1902 if (screen->no_hw)
1903 return false;
1904
1905 struct brw_bo *results, *bo;
1906 uint32_t *batch;
1907 uint32_t offset = 0;
1908 void *map;
1909 bool success = false;
1910
1911 /* Create a zero'ed temporary buffer for reading our results */
1912 results = brw_bo_alloc(screen->bufmgr, "registers", 4096);
1913 if (results == NULL)
1914 goto err;
1915
1916 bo = brw_bo_alloc(screen->bufmgr, "batchbuffer", 4096);
1917 if (bo == NULL)
1918 goto err_results;
1919
1920 map = brw_bo_map(NULL, bo, MAP_WRITE);
1921 if (!map)
1922 goto err_batch;
1923
1924 batch = map;
1925
1926 /* Write the register. */
1927 *batch++ = MI_LOAD_REGISTER_IMM | (3 - 2);
1928 *batch++ = reg;
1929 *batch++ = expected_value;
1930
1931 /* Save the register's value back to the buffer. */
1932 *batch++ = MI_STORE_REGISTER_MEM | (3 - 2);
1933 *batch++ = reg;
1934 struct drm_i915_gem_relocation_entry reloc = {
1935 .offset = (char *) batch - (char *) map,
1936 .delta = offset * sizeof(uint32_t),
1937 .target_handle = results->gem_handle,
1938 .read_domains = I915_GEM_DOMAIN_INSTRUCTION,
1939 .write_domain = I915_GEM_DOMAIN_INSTRUCTION,
1940 };
1941 *batch++ = reloc.presumed_offset + reloc.delta;
1942
1943 /* And afterwards clear the register */
1944 if (reset) {
1945 *batch++ = MI_LOAD_REGISTER_IMM | (3 - 2);
1946 *batch++ = reg;
1947 *batch++ = 0;
1948 }
1949
1950 *batch++ = MI_BATCH_BUFFER_END;
1951
1952 struct drm_i915_gem_exec_object2 exec_objects[2] = {
1953 {
1954 .handle = results->gem_handle,
1955 },
1956 {
1957 .handle = bo->gem_handle,
1958 .relocation_count = 1,
1959 .relocs_ptr = (uintptr_t) &reloc,
1960 }
1961 };
1962
1963 struct drm_i915_gem_execbuffer2 execbuf = {
1964 .buffers_ptr = (uintptr_t) exec_objects,
1965 .buffer_count = 2,
1966 .batch_len = ALIGN((char *) batch - (char *) map, 8),
1967 .flags = I915_EXEC_RENDER,
1968 };
1969
1970 /* Don't bother with error checking - if the execbuf fails, the
1971 * value won't be written and we'll just report that there's no access.
1972 */
1973 __DRIscreen *dri_screen = screen->driScrnPriv;
1974 drmIoctl(dri_screen->fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
1975
1976 /* Check whether the value got written. */
1977 void *results_map = brw_bo_map(NULL, results, MAP_READ);
1978 if (results_map) {
1979 success = *((uint32_t *)results_map + offset) == expected_value;
1980 brw_bo_unmap(results);
1981 }
1982
1983 err_batch:
1984 brw_bo_unreference(bo);
1985 err_results:
1986 brw_bo_unreference(results);
1987 err:
1988 return success;
1989 }
1990
1991 static bool
1992 intel_detect_pipelined_so(struct intel_screen *screen)
1993 {
1994 const struct gen_device_info *devinfo = &screen->devinfo;
1995
1996 /* Supposedly, Broadwell just works. */
1997 if (devinfo->gen >= 8)
1998 return true;
1999
2000 if (devinfo->gen <= 6)
2001 return false;
2002
2003 /* See the big explanation about command parser versions below */
2004 if (screen->cmd_parser_version >= (devinfo->is_haswell ? 7 : 2))
2005 return true;
2006
2007 /* We use SO_WRITE_OFFSET0 since you're supposed to write it (unlike the
2008 * statistics registers), and we already reset it to zero before using it.
2009 */
2010 return intel_detect_pipelined_register(screen,
2011 GEN7_SO_WRITE_OFFSET(0),
2012 0x1337d0d0,
2013 false);
2014 }
2015
2016 /**
2017 * Return array of MSAA modes supported by the hardware. The array is
2018 * zero-terminated and sorted in decreasing order.
2019 */
2020 const int*
2021 intel_supported_msaa_modes(const struct intel_screen *screen)
2022 {
2023 static const int gen9_modes[] = {16, 8, 4, 2, 0, -1};
2024 static const int gen8_modes[] = {8, 4, 2, 0, -1};
2025 static const int gen7_modes[] = {8, 4, 0, -1};
2026 static const int gen6_modes[] = {4, 0, -1};
2027 static const int gen4_modes[] = {0, -1};
2028
2029 if (screen->devinfo.gen >= 9) {
2030 return gen9_modes;
2031 } else if (screen->devinfo.gen >= 8) {
2032 return gen8_modes;
2033 } else if (screen->devinfo.gen >= 7) {
2034 return gen7_modes;
2035 } else if (screen->devinfo.gen == 6) {
2036 return gen6_modes;
2037 } else {
2038 return gen4_modes;
2039 }
2040 }
2041
2042 static unsigned
2043 intel_loader_get_cap(const __DRIscreen *dri_screen, enum dri_loader_cap cap)
2044 {
2045 if (dri_screen->dri2.loader && dri_screen->dri2.loader->base.version >= 4 &&
2046 dri_screen->dri2.loader->getCapability)
2047 return dri_screen->dri2.loader->getCapability(dri_screen->loaderPrivate, cap);
2048
2049 if (dri_screen->image.loader && dri_screen->image.loader->base.version >= 2 &&
2050 dri_screen->image.loader->getCapability)
2051 return dri_screen->image.loader->getCapability(dri_screen->loaderPrivate, cap);
2052
2053 return 0;
2054 }
2055
2056 static __DRIconfig**
2057 intel_screen_make_configs(__DRIscreen *dri_screen)
2058 {
2059 static const mesa_format formats[] = {
2060 MESA_FORMAT_B5G6R5_UNORM,
2061 MESA_FORMAT_B8G8R8A8_UNORM,
2062 MESA_FORMAT_B8G8R8X8_UNORM,
2063
2064 MESA_FORMAT_B8G8R8A8_SRGB,
2065
2066 /* For 10 bpc, 30 bit depth framebuffers. */
2067 MESA_FORMAT_B10G10R10A2_UNORM,
2068 MESA_FORMAT_B10G10R10X2_UNORM,
2069
2070 /* The 32-bit RGBA format must not precede the 32-bit BGRA format.
2071 * Likewise for RGBX and BGRX. Otherwise, the GLX client and the GLX
2072 * server may disagree on which format the GLXFBConfig represents,
2073 * resulting in swapped color channels.
2074 *
2075 * The problem, as of 2017-05-30:
2076 * When matching a GLXFBConfig to a __DRIconfig, GLX ignores the channel
2077 * order and chooses the first __DRIconfig with the expected channel
2078 * sizes. Specifically, GLX compares the GLXFBConfig's and __DRIconfig's
2079 * __DRI_ATTRIB_{CHANNEL}_SIZE but ignores __DRI_ATTRIB_{CHANNEL}_MASK.
2080 *
2081 * EGL does not suffer from this problem. It correctly compares the
2082 * channel masks when matching EGLConfig to __DRIconfig.
2083 */
2084
2085 /* Required by Android, for HAL_PIXEL_FORMAT_RGBA_8888. */
2086 MESA_FORMAT_R8G8B8A8_UNORM,
2087
2088 /* Required by Android, for HAL_PIXEL_FORMAT_RGBX_8888. */
2089 MESA_FORMAT_R8G8B8X8_UNORM,
2090
2091 MESA_FORMAT_R8G8B8A8_SRGB,
2092 };
2093
2094 /* GLX_SWAP_COPY_OML is not supported due to page flipping. */
2095 static const GLenum back_buffer_modes[] = {
2096 __DRI_ATTRIB_SWAP_UNDEFINED, __DRI_ATTRIB_SWAP_NONE
2097 };
2098
2099 static const uint8_t singlesample_samples[1] = {0};
2100
2101 struct intel_screen *screen = dri_screen->driverPrivate;
2102 const struct gen_device_info *devinfo = &screen->devinfo;
2103 uint8_t depth_bits[4], stencil_bits[4];
2104 __DRIconfig **configs = NULL;
2105
2106 /* Expose only BGRA ordering if the loader doesn't support RGBA ordering. */
2107 unsigned num_formats;
2108 if (intel_loader_get_cap(dri_screen, DRI_LOADER_CAP_RGBA_ORDERING))
2109 num_formats = ARRAY_SIZE(formats);
2110 else
2111 num_formats = ARRAY_SIZE(formats) - 3; /* all - RGBA_ORDERING formats */
2112
2113 /* Shall we expose 10 bpc formats? */
2114 bool allow_rgb10_configs = driQueryOptionb(&screen->optionCache,
2115 "allow_rgb10_configs");
2116
2117 /* Generate singlesample configs without accumulation buffer. */
2118 for (unsigned i = 0; i < num_formats; i++) {
2119 __DRIconfig **new_configs;
2120 int num_depth_stencil_bits = 2;
2121
2122 if (!allow_rgb10_configs &&
2123 (formats[i] == MESA_FORMAT_B10G10R10A2_UNORM ||
2124 formats[i] == MESA_FORMAT_B10G10R10X2_UNORM))
2125 continue;
2126
2127 /* Starting with DRI2 protocol version 1.1 we can request a depth/stencil
2128 * buffer that has a different number of bits per pixel than the color
2129 * buffer, gen >= 6 supports this.
2130 */
2131 depth_bits[0] = 0;
2132 stencil_bits[0] = 0;
2133
2134 if (formats[i] == MESA_FORMAT_B5G6R5_UNORM) {
2135 depth_bits[1] = 16;
2136 stencil_bits[1] = 0;
2137 if (devinfo->gen >= 6) {
2138 depth_bits[2] = 24;
2139 stencil_bits[2] = 8;
2140 num_depth_stencil_bits = 3;
2141 }
2142 } else {
2143 depth_bits[1] = 24;
2144 stencil_bits[1] = 8;
2145 }
2146
2147 new_configs = driCreateConfigs(formats[i],
2148 depth_bits,
2149 stencil_bits,
2150 num_depth_stencil_bits,
2151 back_buffer_modes, 2,
2152 singlesample_samples, 1,
2153 false, false);
2154 configs = driConcatConfigs(configs, new_configs);
2155 }
2156
2157 /* Generate the minimum possible set of configs that include an
2158 * accumulation buffer.
2159 */
2160 for (unsigned i = 0; i < num_formats; i++) {
2161 __DRIconfig **new_configs;
2162
2163 if (!allow_rgb10_configs &&
2164 (formats[i] == MESA_FORMAT_B10G10R10A2_UNORM ||
2165 formats[i] == MESA_FORMAT_B10G10R10X2_UNORM))
2166 continue;
2167
2168 if (formats[i] == MESA_FORMAT_B5G6R5_UNORM) {
2169 depth_bits[0] = 16;
2170 stencil_bits[0] = 0;
2171 } else {
2172 depth_bits[0] = 24;
2173 stencil_bits[0] = 8;
2174 }
2175
2176 new_configs = driCreateConfigs(formats[i],
2177 depth_bits, stencil_bits, 1,
2178 back_buffer_modes, 1,
2179 singlesample_samples, 1,
2180 true, false);
2181 configs = driConcatConfigs(configs, new_configs);
2182 }
2183
2184 /* Generate multisample configs.
2185 *
2186 * This loop breaks early, and hence is a no-op, on gen < 6.
2187 *
2188 * Multisample configs must follow the singlesample configs in order to
2189 * work around an X server bug present in 1.12. The X server chooses to
2190 * associate the first listed RGBA888-Z24S8 config, regardless of its
2191 * sample count, with the 32-bit depth visual used for compositing.
2192 *
2193 * Only doublebuffer configs with GLX_SWAP_UNDEFINED_OML behavior are
2194 * supported. Singlebuffer configs are not supported because no one wants
2195 * them.
2196 */
2197 for (unsigned i = 0; i < num_formats; i++) {
2198 if (devinfo->gen < 6)
2199 break;
2200
2201 if (!allow_rgb10_configs &&
2202 (formats[i] == MESA_FORMAT_B10G10R10A2_UNORM ||
2203 formats[i] == MESA_FORMAT_B10G10R10X2_UNORM))
2204 continue;
2205
2206 __DRIconfig **new_configs;
2207 const int num_depth_stencil_bits = 2;
2208 int num_msaa_modes = 0;
2209 const uint8_t *multisample_samples = NULL;
2210
2211 depth_bits[0] = 0;
2212 stencil_bits[0] = 0;
2213
2214 if (formats[i] == MESA_FORMAT_B5G6R5_UNORM) {
2215 depth_bits[1] = 16;
2216 stencil_bits[1] = 0;
2217 } else {
2218 depth_bits[1] = 24;
2219 stencil_bits[1] = 8;
2220 }
2221
2222 if (devinfo->gen >= 9) {
2223 static const uint8_t multisample_samples_gen9[] = {2, 4, 8, 16};
2224 multisample_samples = multisample_samples_gen9;
2225 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen9);
2226 } else if (devinfo->gen == 8) {
2227 static const uint8_t multisample_samples_gen8[] = {2, 4, 8};
2228 multisample_samples = multisample_samples_gen8;
2229 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen8);
2230 } else if (devinfo->gen == 7) {
2231 static const uint8_t multisample_samples_gen7[] = {4, 8};
2232 multisample_samples = multisample_samples_gen7;
2233 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen7);
2234 } else if (devinfo->gen == 6) {
2235 static const uint8_t multisample_samples_gen6[] = {4};
2236 multisample_samples = multisample_samples_gen6;
2237 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen6);
2238 }
2239
2240 new_configs = driCreateConfigs(formats[i],
2241 depth_bits,
2242 stencil_bits,
2243 num_depth_stencil_bits,
2244 back_buffer_modes, 1,
2245 multisample_samples,
2246 num_msaa_modes,
2247 false, false);
2248 configs = driConcatConfigs(configs, new_configs);
2249 }
2250
2251 if (configs == NULL) {
2252 fprintf(stderr, "[%s:%u] Error creating FBConfig!\n", __func__,
2253 __LINE__);
2254 return NULL;
2255 }
2256
2257 return configs;
2258 }
2259
2260 static void
2261 set_max_gl_versions(struct intel_screen *screen)
2262 {
2263 __DRIscreen *dri_screen = screen->driScrnPriv;
2264 const bool has_astc = screen->devinfo.gen >= 9;
2265
2266 switch (screen->devinfo.gen) {
2267 case 11:
2268 case 10:
2269 case 9:
2270 case 8:
2271 dri_screen->max_gl_core_version = 45;
2272 dri_screen->max_gl_compat_version = 30;
2273 dri_screen->max_gl_es1_version = 11;
2274 dri_screen->max_gl_es2_version = has_astc ? 32 : 31;
2275 break;
2276 case 7:
2277 dri_screen->max_gl_core_version = 33;
2278 if (can_do_pipelined_register_writes(screen)) {
2279 dri_screen->max_gl_core_version = 42;
2280 if (screen->devinfo.is_haswell && can_do_compute_dispatch(screen))
2281 dri_screen->max_gl_core_version = 43;
2282 if (screen->devinfo.is_haswell && can_do_mi_math_and_lrr(screen))
2283 dri_screen->max_gl_core_version = 45;
2284 }
2285 dri_screen->max_gl_compat_version = 30;
2286 dri_screen->max_gl_es1_version = 11;
2287 dri_screen->max_gl_es2_version = screen->devinfo.is_haswell ? 31 : 30;
2288 break;
2289 case 6:
2290 dri_screen->max_gl_core_version = 33;
2291 dri_screen->max_gl_compat_version = 30;
2292 dri_screen->max_gl_es1_version = 11;
2293 dri_screen->max_gl_es2_version = 30;
2294 break;
2295 case 5:
2296 case 4:
2297 dri_screen->max_gl_core_version = 0;
2298 dri_screen->max_gl_compat_version = 21;
2299 dri_screen->max_gl_es1_version = 11;
2300 dri_screen->max_gl_es2_version = 20;
2301 break;
2302 default:
2303 unreachable("unrecognized intel_screen::gen");
2304 }
2305 }
2306
2307 /**
2308 * Return the revision (generally the revid field of the PCI header) of the
2309 * graphics device.
2310 */
2311 int
2312 intel_device_get_revision(int fd)
2313 {
2314 struct drm_i915_getparam gp;
2315 int revision;
2316 int ret;
2317
2318 memset(&gp, 0, sizeof(gp));
2319 gp.param = I915_PARAM_REVISION;
2320 gp.value = &revision;
2321
2322 ret = drmCommandWriteRead(fd, DRM_I915_GETPARAM, &gp, sizeof(gp));
2323 if (ret)
2324 revision = -1;
2325
2326 return revision;
2327 }
2328
2329 static void
2330 shader_debug_log_mesa(void *data, const char *fmt, ...)
2331 {
2332 struct brw_context *brw = (struct brw_context *)data;
2333 va_list args;
2334
2335 va_start(args, fmt);
2336 GLuint msg_id = 0;
2337 _mesa_gl_vdebug(&brw->ctx, &msg_id,
2338 MESA_DEBUG_SOURCE_SHADER_COMPILER,
2339 MESA_DEBUG_TYPE_OTHER,
2340 MESA_DEBUG_SEVERITY_NOTIFICATION, fmt, args);
2341 va_end(args);
2342 }
2343
2344 static void
2345 shader_perf_log_mesa(void *data, const char *fmt, ...)
2346 {
2347 struct brw_context *brw = (struct brw_context *)data;
2348
2349 va_list args;
2350 va_start(args, fmt);
2351
2352 if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
2353 va_list args_copy;
2354 va_copy(args_copy, args);
2355 vfprintf(stderr, fmt, args_copy);
2356 va_end(args_copy);
2357 }
2358
2359 if (brw->perf_debug) {
2360 GLuint msg_id = 0;
2361 _mesa_gl_vdebug(&brw->ctx, &msg_id,
2362 MESA_DEBUG_SOURCE_SHADER_COMPILER,
2363 MESA_DEBUG_TYPE_PERFORMANCE,
2364 MESA_DEBUG_SEVERITY_MEDIUM, fmt, args);
2365 }
2366 va_end(args);
2367 }
2368
2369 /**
2370 * This is the driver specific part of the createNewScreen entry point.
2371 * Called when using DRI2.
2372 *
2373 * \return the struct gl_config supported by this driver
2374 */
2375 static const
2376 __DRIconfig **intelInitScreen2(__DRIscreen *dri_screen)
2377 {
2378 struct intel_screen *screen;
2379
2380 if (dri_screen->image.loader) {
2381 } else if (dri_screen->dri2.loader->base.version <= 2 ||
2382 dri_screen->dri2.loader->getBuffersWithFormat == NULL) {
2383 fprintf(stderr,
2384 "\nERROR! DRI2 loader with getBuffersWithFormat() "
2385 "support required\n");
2386 return NULL;
2387 }
2388
2389 /* Allocate the private area */
2390 screen = rzalloc(NULL, struct intel_screen);
2391 if (!screen) {
2392 fprintf(stderr, "\nERROR! Allocating private area failed\n");
2393 return NULL;
2394 }
2395 /* parse information in __driConfigOptions */
2396 driOptionCache options;
2397 memset(&options, 0, sizeof(options));
2398
2399 driParseOptionInfo(&options, brw_config_options.xml);
2400 driParseConfigFiles(&screen->optionCache, &options, dri_screen->myNum, "i965");
2401 driDestroyOptionCache(&options);
2402
2403 screen->driScrnPriv = dri_screen;
2404 dri_screen->driverPrivate = (void *) screen;
2405
2406 screen->deviceID = gen_get_pci_device_id_override();
2407 if (screen->deviceID < 0)
2408 screen->deviceID = intel_get_integer(screen, I915_PARAM_CHIPSET_ID);
2409 else
2410 screen->no_hw = true;
2411
2412 if (!gen_get_device_info(screen->deviceID, &screen->devinfo))
2413 return NULL;
2414
2415 if (!intel_init_bufmgr(screen))
2416 return NULL;
2417
2418 const struct gen_device_info *devinfo = &screen->devinfo;
2419
2420 brw_process_intel_debug_variable();
2421
2422 if ((INTEL_DEBUG & DEBUG_SHADER_TIME) && devinfo->gen < 7) {
2423 fprintf(stderr,
2424 "shader_time debugging requires gen7 (Ivybridge) or better.\n");
2425 INTEL_DEBUG &= ~DEBUG_SHADER_TIME;
2426 }
2427
2428 if (intel_get_integer(screen, I915_PARAM_MMAP_GTT_VERSION) >= 1) {
2429 /* Theorectically unlimited! At least for individual objects...
2430 *
2431 * Currently the entire (global) address space for all GTT maps is
2432 * limited to 64bits. That is all objects on the system that are
2433 * setup for GTT mmapping must fit within 64bits. An attempt to use
2434 * one that exceeds the limit with fail in brw_bo_map_gtt().
2435 *
2436 * Long before we hit that limit, we will be practically limited by
2437 * that any single object must fit in physical memory (RAM). The upper
2438 * limit on the CPU's address space is currently 48bits (Skylake), of
2439 * which only 39bits can be physical memory. (The GPU itself also has
2440 * a 48bit addressable virtual space.) We can fit over 32 million
2441 * objects of the current maximum allocable size before running out
2442 * of mmap space.
2443 */
2444 screen->max_gtt_map_object_size = UINT64_MAX;
2445 } else {
2446 /* Estimate the size of the mappable aperture into the GTT. There's an
2447 * ioctl to get the whole GTT size, but not one to get the mappable subset.
2448 * It turns out it's basically always 256MB, though some ancient hardware
2449 * was smaller.
2450 */
2451 uint32_t gtt_size = 256 * 1024 * 1024;
2452
2453 /* We don't want to map two objects such that a memcpy between them would
2454 * just fault one mapping in and then the other over and over forever. So
2455 * we would need to divide the GTT size by 2. Additionally, some GTT is
2456 * taken up by things like the framebuffer and the ringbuffer and such, so
2457 * be more conservative.
2458 */
2459 screen->max_gtt_map_object_size = gtt_size / 4;
2460 }
2461
2462 screen->aperture_threshold = get_aperture_size(dri_screen->fd) * 3 / 4;
2463
2464 screen->hw_has_swizzling = intel_detect_swizzling(screen);
2465 screen->hw_has_timestamp = intel_detect_timestamp(screen);
2466
2467 isl_device_init(&screen->isl_dev, &screen->devinfo,
2468 screen->hw_has_swizzling);
2469
2470 if (devinfo->gen >= 10)
2471 intel_cs_timestamp_frequency(screen);
2472
2473 /* GENs prior to 8 do not support EU/Subslice info */
2474 if (devinfo->gen >= 8) {
2475 intel_detect_sseu(screen);
2476 } else if (devinfo->gen == 7) {
2477 screen->subslice_total = 1 << (devinfo->gt - 1);
2478 }
2479
2480 /* Gen7-7.5 kernel requirements / command parser saga:
2481 *
2482 * - pre-v3.16:
2483 * Haswell and Baytrail cannot use any privileged batchbuffer features.
2484 *
2485 * Ivybridge has aliasing PPGTT on by default, which accidentally marks
2486 * all batches secure, allowing them to use any feature with no checking.
2487 * This is effectively equivalent to a command parser version of
2488 * \infinity - everything is possible.
2489 *
2490 * The command parser does not exist, and querying the version will
2491 * return -EINVAL.
2492 *
2493 * - v3.16:
2494 * The kernel enables the command parser by default, for systems with
2495 * aliasing PPGTT enabled (Ivybridge and Haswell). However, the
2496 * hardware checker is still enabled, so Haswell and Baytrail cannot
2497 * do anything.
2498 *
2499 * Ivybridge goes from "everything is possible" to "only what the
2500 * command parser allows" (if the user boots with i915.cmd_parser=0,
2501 * then everything is possible again). We can only safely use features
2502 * allowed by the supported command parser version.
2503 *
2504 * Annoyingly, I915_PARAM_CMD_PARSER_VERSION reports the static version
2505 * implemented by the kernel, even if it's turned off. So, checking
2506 * for version > 0 does not mean that you can write registers. We have
2507 * to try it and see. The version does, however, indicate the age of
2508 * the kernel.
2509 *
2510 * Instead of matching the hardware checker's behavior of converting
2511 * privileged commands to MI_NOOP, it makes execbuf2 start returning
2512 * -EINVAL, making it dangerous to try and use privileged features.
2513 *
2514 * Effective command parser versions:
2515 * - Haswell: 0 (reporting 1, writes don't work)
2516 * - Baytrail: 0 (reporting 1, writes don't work)
2517 * - Ivybridge: 1 (enabled) or infinite (disabled)
2518 *
2519 * - v3.17:
2520 * Baytrail aliasing PPGTT is enabled, making it like Ivybridge:
2521 * effectively version 1 (enabled) or infinite (disabled).
2522 *
2523 * - v3.19: f1f55cc0556031c8ee3fe99dae7251e78b9b653b
2524 * Command parser v2 supports predicate writes.
2525 *
2526 * - Haswell: 0 (reporting 1, writes don't work)
2527 * - Baytrail: 2 (enabled) or infinite (disabled)
2528 * - Ivybridge: 2 (enabled) or infinite (disabled)
2529 *
2530 * So version >= 2 is enough to know that Ivybridge and Baytrail
2531 * will work. Haswell still can't do anything.
2532 *
2533 * - v4.0: Version 3 happened. Largely not relevant.
2534 *
2535 * - v4.1: 6702cf16e0ba8b0129f5aa1b6609d4e9c70bc13b
2536 * L3 config registers are properly saved and restored as part
2537 * of the hardware context. We can approximately detect this point
2538 * in time by checking if I915_PARAM_REVISION is recognized - it
2539 * landed in a later commit, but in the same release cycle.
2540 *
2541 * - v4.2: 245054a1fe33c06ad233e0d58a27ec7b64db9284
2542 * Command parser finally gains secure batch promotion. On Haswell,
2543 * the hardware checker gets disabled, which finally allows it to do
2544 * privileged commands.
2545 *
2546 * I915_PARAM_CMD_PARSER_VERSION reports 3. Effective versions:
2547 * - Haswell: 3 (enabled) or 0 (disabled)
2548 * - Baytrail: 3 (enabled) or infinite (disabled)
2549 * - Ivybridge: 3 (enabled) or infinite (disabled)
2550 *
2551 * Unfortunately, detecting this point in time is tricky, because
2552 * no version bump happened when this important change occurred.
2553 * On Haswell, if we can write any register, then the kernel is at
2554 * least this new, and we can start trusting the version number.
2555 *
2556 * - v4.4: 2bbe6bbb0dc94fd4ce287bdac9e1bd184e23057b and
2557 * Command parser reaches version 4, allowing access to Haswell
2558 * atomic scratch and chicken3 registers. If version >= 4, we know
2559 * the kernel is new enough to support privileged features on all
2560 * hardware. However, the user might have disabled it...and the
2561 * kernel will still report version 4. So we still have to guess
2562 * and check.
2563 *
2564 * - v4.4: 7b9748cb513a6bef4af87b79f0da3ff7e8b56cd8
2565 * Command parser v5 whitelists indirect compute shader dispatch
2566 * registers, needed for OpenGL 4.3 and later.
2567 *
2568 * - v4.8:
2569 * Command parser v7 lets us use MI_MATH on Haswell.
2570 *
2571 * Additionally, the kernel begins reporting version 0 when
2572 * the command parser is disabled, allowing us to skip the
2573 * guess-and-check step on Haswell. Unfortunately, this also
2574 * means that we can no longer use it as an indicator of the
2575 * age of the kernel.
2576 */
2577 if (intel_get_param(screen, I915_PARAM_CMD_PARSER_VERSION,
2578 &screen->cmd_parser_version) < 0) {
2579 /* Command parser does not exist - getparam is unrecognized */
2580 screen->cmd_parser_version = 0;
2581 }
2582
2583 /* Kernel 4.13 retuired for exec object capture */
2584 if (intel_get_boolean(screen, I915_PARAM_HAS_EXEC_CAPTURE)) {
2585 screen->kernel_features |= KERNEL_ALLOWS_EXEC_CAPTURE;
2586 }
2587
2588 if (intel_get_boolean(screen, I915_PARAM_HAS_EXEC_BATCH_FIRST)) {
2589 screen->kernel_features |= KERNEL_ALLOWS_EXEC_BATCH_FIRST;
2590 }
2591
2592 if (!intel_detect_pipelined_so(screen)) {
2593 /* We can't do anything, so the effective version is 0. */
2594 screen->cmd_parser_version = 0;
2595 } else {
2596 screen->kernel_features |= KERNEL_ALLOWS_SOL_OFFSET_WRITES;
2597 }
2598
2599 if (devinfo->gen >= 8 || screen->cmd_parser_version >= 2)
2600 screen->kernel_features |= KERNEL_ALLOWS_PREDICATE_WRITES;
2601
2602 /* Haswell requires command parser version 4 in order to have L3
2603 * atomic scratch1 and chicken3 bits
2604 */
2605 if (devinfo->is_haswell && screen->cmd_parser_version >= 4) {
2606 screen->kernel_features |=
2607 KERNEL_ALLOWS_HSW_SCRATCH1_AND_ROW_CHICKEN3;
2608 }
2609
2610 /* Haswell requires command parser version 6 in order to write to the
2611 * MI_MATH GPR registers, and version 7 in order to use
2612 * MI_LOAD_REGISTER_REG (which all users of MI_MATH use).
2613 */
2614 if (devinfo->gen >= 8 ||
2615 (devinfo->is_haswell && screen->cmd_parser_version >= 7)) {
2616 screen->kernel_features |= KERNEL_ALLOWS_MI_MATH_AND_LRR;
2617 }
2618
2619 /* Gen7 needs at least command parser version 5 to support compute */
2620 if (devinfo->gen >= 8 || screen->cmd_parser_version >= 5)
2621 screen->kernel_features |= KERNEL_ALLOWS_COMPUTE_DISPATCH;
2622
2623 if (intel_get_boolean(screen, I915_PARAM_HAS_CONTEXT_ISOLATION))
2624 screen->kernel_features |= KERNEL_ALLOWS_CONTEXT_ISOLATION;
2625
2626 const char *force_msaa = getenv("INTEL_FORCE_MSAA");
2627 if (force_msaa) {
2628 screen->winsys_msaa_samples_override =
2629 intel_quantize_num_samples(screen, atoi(force_msaa));
2630 printf("Forcing winsys sample count to %d\n",
2631 screen->winsys_msaa_samples_override);
2632 } else {
2633 screen->winsys_msaa_samples_override = -1;
2634 }
2635
2636 set_max_gl_versions(screen);
2637
2638 /* Notification of GPU resets requires hardware contexts and a kernel new
2639 * enough to support DRM_IOCTL_I915_GET_RESET_STATS. If the ioctl is
2640 * supported, calling it with a context of 0 will either generate EPERM or
2641 * no error. If the ioctl is not supported, it always generate EINVAL.
2642 * Use this to determine whether to advertise the __DRI2_ROBUSTNESS
2643 * extension to the loader.
2644 *
2645 * Don't even try on pre-Gen6, since we don't attempt to use contexts there.
2646 */
2647 if (devinfo->gen >= 6) {
2648 struct drm_i915_reset_stats stats;
2649 memset(&stats, 0, sizeof(stats));
2650
2651 const int ret = drmIoctl(dri_screen->fd, DRM_IOCTL_I915_GET_RESET_STATS, &stats);
2652
2653 screen->has_context_reset_notification =
2654 (ret != -1 || errno != EINVAL);
2655 }
2656
2657 dri_screen->extensions = !screen->has_context_reset_notification
2658 ? screenExtensions : intelRobustScreenExtensions;
2659
2660 screen->compiler = brw_compiler_create(screen, devinfo);
2661 screen->compiler->shader_debug_log = shader_debug_log_mesa;
2662 screen->compiler->shader_perf_log = shader_perf_log_mesa;
2663
2664 /* Changing the meaning of constant buffer pointers from a dynamic state
2665 * offset to an absolute address is only safe if the kernel isolates other
2666 * contexts from our changes.
2667 */
2668 screen->compiler->constant_buffer_0_is_relative = devinfo->gen < 8 ||
2669 !(screen->kernel_features & KERNEL_ALLOWS_CONTEXT_ISOLATION);
2670
2671 screen->compiler->supports_pull_constants = true;
2672
2673 screen->has_exec_fence =
2674 intel_get_boolean(screen, I915_PARAM_HAS_EXEC_FENCE);
2675
2676 intel_screen_init_surface_formats(screen);
2677
2678 if (INTEL_DEBUG & (DEBUG_BATCH | DEBUG_SUBMIT)) {
2679 unsigned int caps = intel_get_integer(screen, I915_PARAM_HAS_SCHEDULER);
2680 if (caps) {
2681 fprintf(stderr, "Kernel scheduler detected: %08x\n", caps);
2682 if (caps & I915_SCHEDULER_CAP_PRIORITY)
2683 fprintf(stderr, " - User priority sorting enabled\n");
2684 if (caps & I915_SCHEDULER_CAP_PREEMPTION)
2685 fprintf(stderr, " - Preemption enabled\n");
2686 }
2687 }
2688
2689 brw_disk_cache_init(screen);
2690
2691 return (const __DRIconfig**) intel_screen_make_configs(dri_screen);
2692 }
2693
2694 struct intel_buffer {
2695 __DRIbuffer base;
2696 struct brw_bo *bo;
2697 };
2698
2699 static __DRIbuffer *
2700 intelAllocateBuffer(__DRIscreen *dri_screen,
2701 unsigned attachment, unsigned format,
2702 int width, int height)
2703 {
2704 struct intel_buffer *intelBuffer;
2705 struct intel_screen *screen = dri_screen->driverPrivate;
2706
2707 assert(attachment == __DRI_BUFFER_FRONT_LEFT ||
2708 attachment == __DRI_BUFFER_BACK_LEFT);
2709
2710 intelBuffer = calloc(1, sizeof *intelBuffer);
2711 if (intelBuffer == NULL)
2712 return NULL;
2713
2714 /* The front and back buffers are color buffers, which are X tiled. GEN9+
2715 * supports Y tiled and compressed buffers, but there is no way to plumb that
2716 * through to here. */
2717 uint32_t pitch;
2718 int cpp = format / 8;
2719 intelBuffer->bo = brw_bo_alloc_tiled_2d(screen->bufmgr,
2720 "intelAllocateBuffer",
2721 width,
2722 height,
2723 cpp,
2724 I915_TILING_X, &pitch,
2725 BO_ALLOC_BUSY);
2726
2727 if (intelBuffer->bo == NULL) {
2728 free(intelBuffer);
2729 return NULL;
2730 }
2731
2732 brw_bo_flink(intelBuffer->bo, &intelBuffer->base.name);
2733
2734 intelBuffer->base.attachment = attachment;
2735 intelBuffer->base.cpp = cpp;
2736 intelBuffer->base.pitch = pitch;
2737
2738 return &intelBuffer->base;
2739 }
2740
2741 static void
2742 intelReleaseBuffer(__DRIscreen *dri_screen, __DRIbuffer *buffer)
2743 {
2744 struct intel_buffer *intelBuffer = (struct intel_buffer *) buffer;
2745
2746 brw_bo_unreference(intelBuffer->bo);
2747 free(intelBuffer);
2748 }
2749
2750 static const struct __DriverAPIRec brw_driver_api = {
2751 .InitScreen = intelInitScreen2,
2752 .DestroyScreen = intelDestroyScreen,
2753 .CreateContext = brwCreateContext,
2754 .DestroyContext = intelDestroyContext,
2755 .CreateBuffer = intelCreateBuffer,
2756 .DestroyBuffer = intelDestroyBuffer,
2757 .MakeCurrent = intelMakeCurrent,
2758 .UnbindContext = intelUnbindContext,
2759 .AllocateBuffer = intelAllocateBuffer,
2760 .ReleaseBuffer = intelReleaseBuffer
2761 };
2762
2763 static const struct __DRIDriverVtableExtensionRec brw_vtable = {
2764 .base = { __DRI_DRIVER_VTABLE, 1 },
2765 .vtable = &brw_driver_api,
2766 };
2767
2768 static const __DRIextension *brw_driver_extensions[] = {
2769 &driCoreExtension.base,
2770 &driImageDriverExtension.base,
2771 &driDRI2Extension.base,
2772 &brw_vtable.base,
2773 &brw_config_options.base,
2774 NULL
2775 };
2776
2777 PUBLIC const __DRIextension **__driDriverGetExtensions_i965(void)
2778 {
2779 globalDriverAPI = &brw_driver_api;
2780
2781 return brw_driver_extensions;
2782 }