i965/screen: Add XRGB2101010 and ARGB2101010 support for DRI3.
[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 "brw_defines.h"
40 #include "brw_state.h"
41 #include "compiler/nir/nir.h"
42
43 #include "utils.h"
44 #include "util/xmlpool.h"
45
46 static const __DRIconfigOptionsExtension brw_config_options = {
47 .base = { __DRI_CONFIG_OPTIONS, 1 },
48 .xml =
49 DRI_CONF_BEGIN
50 DRI_CONF_SECTION_PERFORMANCE
51 /* Options correspond to DRI_CONF_BO_REUSE_DISABLED,
52 * DRI_CONF_BO_REUSE_ALL
53 */
54 DRI_CONF_OPT_BEGIN_V(bo_reuse, enum, 1, "0:1")
55 DRI_CONF_DESC_BEGIN(en, "Buffer object reuse")
56 DRI_CONF_ENUM(0, "Disable buffer object reuse")
57 DRI_CONF_ENUM(1, "Enable reuse of all sizes of buffer objects")
58 DRI_CONF_DESC_END
59 DRI_CONF_OPT_END
60 DRI_CONF_MESA_NO_ERROR("false")
61 DRI_CONF_SECTION_END
62
63 DRI_CONF_SECTION_QUALITY
64 DRI_CONF_PRECISE_TRIG("false")
65
66 DRI_CONF_OPT_BEGIN(clamp_max_samples, int, -1)
67 DRI_CONF_DESC(en, "Clamp the value of GL_MAX_SAMPLES to the "
68 "given integer. If negative, then do not clamp.")
69 DRI_CONF_OPT_END
70 DRI_CONF_SECTION_END
71
72 DRI_CONF_SECTION_DEBUG
73 DRI_CONF_NO_RAST("false")
74 DRI_CONF_ALWAYS_FLUSH_BATCH("false")
75 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
76 DRI_CONF_DISABLE_THROTTLING("false")
77 DRI_CONF_FORCE_GLSL_EXTENSIONS_WARN("false")
78 DRI_CONF_FORCE_GLSL_VERSION(0)
79 DRI_CONF_DISABLE_GLSL_LINE_CONTINUATIONS("false")
80 DRI_CONF_DISABLE_BLEND_FUNC_EXTENDED("false")
81 DRI_CONF_DUAL_COLOR_BLEND_BY_LOCATION("false")
82 DRI_CONF_ALLOW_GLSL_EXTENSION_DIRECTIVE_MIDSHADER("false")
83 DRI_CONF_ALLOW_GLSL_BUILTIN_VARIABLE_REDECLARATION("false")
84 DRI_CONF_ALLOW_GLSL_CROSS_STAGE_INTERPOLATION_MISMATCH("false")
85 DRI_CONF_ALLOW_HIGHER_COMPAT_VERSION("false")
86 DRI_CONF_FORCE_GLSL_ABS_SQRT("false")
87
88 DRI_CONF_OPT_BEGIN_B(shader_precompile, "true")
89 DRI_CONF_DESC(en, "Perform code generation at shader link time.")
90 DRI_CONF_OPT_END
91 DRI_CONF_SECTION_END
92
93 DRI_CONF_SECTION_MISCELLANEOUS
94 DRI_CONF_GLSL_ZERO_INIT("false")
95 DRI_CONF_SECTION_END
96 DRI_CONF_END
97 };
98
99 #include "intel_batchbuffer.h"
100 #include "intel_buffers.h"
101 #include "brw_bufmgr.h"
102 #include "intel_fbo.h"
103 #include "intel_mipmap_tree.h"
104 #include "intel_screen.h"
105 #include "intel_tex.h"
106 #include "intel_image.h"
107
108 #include "brw_context.h"
109
110 #include "i915_drm.h"
111
112 /**
113 * For debugging purposes, this returns a time in seconds.
114 */
115 double
116 get_time(void)
117 {
118 struct timespec tp;
119
120 clock_gettime(CLOCK_MONOTONIC, &tp);
121
122 return tp.tv_sec + tp.tv_nsec / 1000000000.0;
123 }
124
125 static const __DRItexBufferExtension intelTexBufferExtension = {
126 .base = { __DRI_TEX_BUFFER, 3 },
127
128 .setTexBuffer = intelSetTexBuffer,
129 .setTexBuffer2 = intelSetTexBuffer2,
130 .releaseTexBuffer = NULL,
131 };
132
133 static void
134 intel_dri2_flush_with_flags(__DRIcontext *cPriv,
135 __DRIdrawable *dPriv,
136 unsigned flags,
137 enum __DRI2throttleReason reason)
138 {
139 struct brw_context *brw = cPriv->driverPrivate;
140
141 if (!brw)
142 return;
143
144 struct gl_context *ctx = &brw->ctx;
145
146 FLUSH_VERTICES(ctx, 0);
147
148 if (flags & __DRI2_FLUSH_DRAWABLE)
149 intel_resolve_for_dri2_flush(brw, dPriv);
150
151 if (reason == __DRI2_THROTTLE_SWAPBUFFER)
152 brw->need_swap_throttle = true;
153 if (reason == __DRI2_THROTTLE_FLUSHFRONT)
154 brw->need_flush_throttle = true;
155
156 intel_batchbuffer_flush(brw);
157 }
158
159 /**
160 * Provides compatibility with loaders that only support the older (version
161 * 1-3) flush interface.
162 *
163 * That includes libGL up to Mesa 9.0, and the X Server at least up to 1.13.
164 */
165 static void
166 intel_dri2_flush(__DRIdrawable *drawable)
167 {
168 intel_dri2_flush_with_flags(drawable->driContextPriv, drawable,
169 __DRI2_FLUSH_DRAWABLE,
170 __DRI2_THROTTLE_SWAPBUFFER);
171 }
172
173 static const struct __DRI2flushExtensionRec intelFlushExtension = {
174 .base = { __DRI2_FLUSH, 4 },
175
176 .flush = intel_dri2_flush,
177 .invalidate = dri2InvalidateDrawable,
178 .flush_with_flags = intel_dri2_flush_with_flags,
179 };
180
181 static const struct intel_image_format intel_image_formats[] = {
182 { __DRI_IMAGE_FOURCC_ARGB2101010, __DRI_IMAGE_COMPONENTS_RGBA, 1,
183 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ARGB2101010, 4 } } },
184
185 { __DRI_IMAGE_FOURCC_XRGB2101010, __DRI_IMAGE_COMPONENTS_RGB, 1,
186 { { 0, 0, 0, __DRI_IMAGE_FORMAT_XRGB2101010, 4 } } },
187
188 { __DRI_IMAGE_FOURCC_ARGB8888, __DRI_IMAGE_COMPONENTS_RGBA, 1,
189 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ARGB8888, 4 } } },
190
191 { __DRI_IMAGE_FOURCC_ABGR8888, __DRI_IMAGE_COMPONENTS_RGBA, 1,
192 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ABGR8888, 4 } } },
193
194 { __DRI_IMAGE_FOURCC_SARGB8888, __DRI_IMAGE_COMPONENTS_RGBA, 1,
195 { { 0, 0, 0, __DRI_IMAGE_FORMAT_SARGB8, 4 } } },
196
197 { __DRI_IMAGE_FOURCC_XRGB8888, __DRI_IMAGE_COMPONENTS_RGB, 1,
198 { { 0, 0, 0, __DRI_IMAGE_FORMAT_XRGB8888, 4 }, } },
199
200 { __DRI_IMAGE_FOURCC_XBGR8888, __DRI_IMAGE_COMPONENTS_RGB, 1,
201 { { 0, 0, 0, __DRI_IMAGE_FORMAT_XBGR8888, 4 }, } },
202
203 { __DRI_IMAGE_FOURCC_ARGB1555, __DRI_IMAGE_COMPONENTS_RGBA, 1,
204 { { 0, 0, 0, __DRI_IMAGE_FORMAT_ARGB1555, 2 } } },
205
206 { __DRI_IMAGE_FOURCC_RGB565, __DRI_IMAGE_COMPONENTS_RGB, 1,
207 { { 0, 0, 0, __DRI_IMAGE_FORMAT_RGB565, 2 } } },
208
209 { __DRI_IMAGE_FOURCC_R8, __DRI_IMAGE_COMPONENTS_R, 1,
210 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 }, } },
211
212 { __DRI_IMAGE_FOURCC_R16, __DRI_IMAGE_COMPONENTS_R, 1,
213 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R16, 1 }, } },
214
215 { __DRI_IMAGE_FOURCC_GR88, __DRI_IMAGE_COMPONENTS_RG, 1,
216 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR88, 2 }, } },
217
218 { __DRI_IMAGE_FOURCC_GR1616, __DRI_IMAGE_COMPONENTS_RG, 1,
219 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR1616, 2 }, } },
220
221 { __DRI_IMAGE_FOURCC_YUV410, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
222 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
223 { 1, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 },
224 { 2, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 } } },
225
226 { __DRI_IMAGE_FOURCC_YUV411, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
227 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
228 { 1, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 },
229 { 2, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
230
231 { __DRI_IMAGE_FOURCC_YUV420, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
232 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
233 { 1, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 },
234 { 2, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 } } },
235
236 { __DRI_IMAGE_FOURCC_YUV422, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
237 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
238 { 1, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 },
239 { 2, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
240
241 { __DRI_IMAGE_FOURCC_YUV444, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
242 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
243 { 1, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
244 { 2, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
245
246 { __DRI_IMAGE_FOURCC_YVU410, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
247 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
248 { 2, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 },
249 { 1, 2, 2, __DRI_IMAGE_FORMAT_R8, 1 } } },
250
251 { __DRI_IMAGE_FOURCC_YVU411, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
252 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
253 { 2, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 },
254 { 1, 2, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
255
256 { __DRI_IMAGE_FOURCC_YVU420, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
257 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
258 { 2, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 },
259 { 1, 1, 1, __DRI_IMAGE_FORMAT_R8, 1 } } },
260
261 { __DRI_IMAGE_FOURCC_YVU422, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
262 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
263 { 2, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 },
264 { 1, 1, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
265
266 { __DRI_IMAGE_FOURCC_YVU444, __DRI_IMAGE_COMPONENTS_Y_U_V, 3,
267 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
268 { 2, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
269 { 1, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 } } },
270
271 { __DRI_IMAGE_FOURCC_NV12, __DRI_IMAGE_COMPONENTS_Y_UV, 2,
272 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
273 { 1, 1, 1, __DRI_IMAGE_FORMAT_GR88, 2 } } },
274
275 { __DRI_IMAGE_FOURCC_NV16, __DRI_IMAGE_COMPONENTS_Y_UV, 2,
276 { { 0, 0, 0, __DRI_IMAGE_FORMAT_R8, 1 },
277 { 1, 1, 0, __DRI_IMAGE_FORMAT_GR88, 2 } } },
278
279 /* For YUYV and UYVY buffers, we set up two overlapping DRI images
280 * and treat them as planar buffers in the compositors.
281 * Plane 0 is GR88 and samples YU or YV pairs and places Y into
282 * the R component, while plane 1 is ARGB/ABGR and samples YUYV/UYVY
283 * clusters and places pairs and places U into the G component and
284 * V into A. This lets the texture sampler interpolate the Y
285 * components correctly when sampling from plane 0, and interpolate
286 * U and V correctly when sampling from plane 1. */
287 { __DRI_IMAGE_FOURCC_YUYV, __DRI_IMAGE_COMPONENTS_Y_XUXV, 2,
288 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR88, 2 },
289 { 0, 1, 0, __DRI_IMAGE_FORMAT_ARGB8888, 4 } } },
290 { __DRI_IMAGE_FOURCC_UYVY, __DRI_IMAGE_COMPONENTS_Y_UXVX, 2,
291 { { 0, 0, 0, __DRI_IMAGE_FORMAT_GR88, 2 },
292 { 0, 1, 0, __DRI_IMAGE_FORMAT_ABGR8888, 4 } } }
293 };
294
295 static const struct {
296 uint64_t modifier;
297 unsigned since_gen;
298 } supported_modifiers[] = {
299 { .modifier = DRM_FORMAT_MOD_LINEAR , .since_gen = 1 },
300 { .modifier = I915_FORMAT_MOD_X_TILED , .since_gen = 1 },
301 { .modifier = I915_FORMAT_MOD_Y_TILED , .since_gen = 6 },
302 { .modifier = I915_FORMAT_MOD_Y_TILED_CCS , .since_gen = 9 },
303 };
304
305 static bool
306 modifier_is_supported(const struct gen_device_info *devinfo,
307 const struct intel_image_format *fmt, int dri_format,
308 uint64_t modifier)
309 {
310 const struct isl_drm_modifier_info *modinfo =
311 isl_drm_modifier_get_info(modifier);
312 int i;
313
314 /* ISL had better know about the modifier */
315 if (!modinfo)
316 return false;
317
318 if (modinfo->aux_usage == ISL_AUX_USAGE_CCS_E) {
319 /* If INTEL_DEBUG=norbc is set, don't support any CCS_E modifiers */
320 if (unlikely(INTEL_DEBUG & DEBUG_NO_RBC))
321 return false;
322
323 /* CCS_E is not supported for planar images */
324 if (fmt && fmt->nplanes > 1)
325 return false;
326
327 if (fmt) {
328 assert(dri_format == 0);
329 dri_format = fmt->planes[0].dri_format;
330 }
331
332 mesa_format format = driImageFormatToGLFormat(dri_format);
333 format = _mesa_get_srgb_format_linear(format);
334 if (!isl_format_supports_ccs_e(devinfo,
335 brw_isl_format_for_mesa_format(format)))
336 return false;
337 }
338
339 for (i = 0; i < ARRAY_SIZE(supported_modifiers); i++) {
340 if (supported_modifiers[i].modifier != modifier)
341 continue;
342
343 return supported_modifiers[i].since_gen <= devinfo->gen;
344 }
345
346 return false;
347 }
348
349 static uint64_t
350 tiling_to_modifier(uint32_t tiling)
351 {
352 static const uint64_t map[] = {
353 [I915_TILING_NONE] = DRM_FORMAT_MOD_LINEAR,
354 [I915_TILING_X] = I915_FORMAT_MOD_X_TILED,
355 [I915_TILING_Y] = I915_FORMAT_MOD_Y_TILED,
356 };
357
358 assert(tiling < ARRAY_SIZE(map));
359
360 return map[tiling];
361 }
362
363 static void
364 intel_image_warn_if_unaligned(__DRIimage *image, const char *func)
365 {
366 uint32_t tiling, swizzle;
367 brw_bo_get_tiling(image->bo, &tiling, &swizzle);
368
369 if (tiling != I915_TILING_NONE && (image->offset & 0xfff)) {
370 _mesa_warning(NULL, "%s: offset 0x%08x not on tile boundary",
371 func, image->offset);
372 }
373 }
374
375 static const struct intel_image_format *
376 intel_image_format_lookup(int fourcc)
377 {
378 for (unsigned i = 0; i < ARRAY_SIZE(intel_image_formats); i++) {
379 if (intel_image_formats[i].fourcc == fourcc)
380 return &intel_image_formats[i];
381 }
382
383 return NULL;
384 }
385
386 static boolean intel_lookup_fourcc(int dri_format, int *fourcc)
387 {
388 for (unsigned i = 0; i < ARRAY_SIZE(intel_image_formats); i++) {
389 if (intel_image_formats[i].planes[0].dri_format == dri_format) {
390 *fourcc = intel_image_formats[i].fourcc;
391 return true;
392 }
393 }
394 return false;
395 }
396
397 static __DRIimage *
398 intel_allocate_image(struct intel_screen *screen, int dri_format,
399 void *loaderPrivate)
400 {
401 __DRIimage *image;
402
403 image = calloc(1, sizeof *image);
404 if (image == NULL)
405 return NULL;
406
407 image->screen = screen;
408 image->dri_format = dri_format;
409 image->offset = 0;
410
411 image->format = driImageFormatToGLFormat(dri_format);
412 if (dri_format != __DRI_IMAGE_FORMAT_NONE &&
413 image->format == MESA_FORMAT_NONE) {
414 free(image);
415 return NULL;
416 }
417
418 image->internal_format = _mesa_get_format_base_format(image->format);
419 image->data = loaderPrivate;
420
421 return image;
422 }
423
424 /**
425 * Sets up a DRIImage structure to point to a slice out of a miptree.
426 */
427 static void
428 intel_setup_image_from_mipmap_tree(struct brw_context *brw, __DRIimage *image,
429 struct intel_mipmap_tree *mt, GLuint level,
430 GLuint zoffset)
431 {
432 intel_miptree_make_shareable(brw, mt);
433
434 intel_miptree_check_level_layer(mt, level, zoffset);
435
436 image->width = minify(mt->surf.phys_level0_sa.width,
437 level - mt->first_level);
438 image->height = minify(mt->surf.phys_level0_sa.height,
439 level - mt->first_level);
440 image->pitch = mt->surf.row_pitch;
441
442 image->offset = intel_miptree_get_tile_offsets(mt, level, zoffset,
443 &image->tile_x,
444 &image->tile_y);
445
446 brw_bo_unreference(image->bo);
447 image->bo = mt->bo;
448 brw_bo_reference(mt->bo);
449 }
450
451 static __DRIimage *
452 intel_create_image_from_name(__DRIscreen *dri_screen,
453 int width, int height, int format,
454 int name, int pitch, void *loaderPrivate)
455 {
456 struct intel_screen *screen = dri_screen->driverPrivate;
457 __DRIimage *image;
458 int cpp;
459
460 image = intel_allocate_image(screen, format, loaderPrivate);
461 if (image == NULL)
462 return NULL;
463
464 if (image->format == MESA_FORMAT_NONE)
465 cpp = 1;
466 else
467 cpp = _mesa_get_format_bytes(image->format);
468
469 image->width = width;
470 image->height = height;
471 image->pitch = pitch * cpp;
472 image->bo = brw_bo_gem_create_from_name(screen->bufmgr, "image",
473 name);
474 if (!image->bo) {
475 free(image);
476 return NULL;
477 }
478 image->modifier = tiling_to_modifier(image->bo->tiling_mode);
479
480 return image;
481 }
482
483 static __DRIimage *
484 intel_create_image_from_renderbuffer(__DRIcontext *context,
485 int renderbuffer, void *loaderPrivate)
486 {
487 __DRIimage *image;
488 struct brw_context *brw = context->driverPrivate;
489 struct gl_context *ctx = &brw->ctx;
490 struct gl_renderbuffer *rb;
491 struct intel_renderbuffer *irb;
492
493 rb = _mesa_lookup_renderbuffer(ctx, renderbuffer);
494 if (!rb) {
495 _mesa_error(ctx, GL_INVALID_OPERATION, "glRenderbufferExternalMESA");
496 return NULL;
497 }
498
499 irb = intel_renderbuffer(rb);
500 intel_miptree_make_shareable(brw, irb->mt);
501 image = calloc(1, sizeof *image);
502 if (image == NULL)
503 return NULL;
504
505 image->internal_format = rb->InternalFormat;
506 image->format = rb->Format;
507 image->modifier = tiling_to_modifier(
508 isl_tiling_to_i915_tiling(irb->mt->surf.tiling));
509 image->offset = 0;
510 image->data = loaderPrivate;
511 brw_bo_unreference(image->bo);
512 image->bo = irb->mt->bo;
513 brw_bo_reference(irb->mt->bo);
514 image->width = rb->Width;
515 image->height = rb->Height;
516 image->pitch = irb->mt->surf.row_pitch;
517 image->dri_format = driGLFormatToImageFormat(image->format);
518 image->has_depthstencil = irb->mt->stencil_mt? true : false;
519
520 rb->NeedsFinishRenderTexture = true;
521 return image;
522 }
523
524 static __DRIimage *
525 intel_create_image_from_texture(__DRIcontext *context, int target,
526 unsigned texture, int zoffset,
527 int level,
528 unsigned *error,
529 void *loaderPrivate)
530 {
531 __DRIimage *image;
532 struct brw_context *brw = context->driverPrivate;
533 struct gl_texture_object *obj;
534 struct intel_texture_object *iobj;
535 GLuint face = 0;
536
537 obj = _mesa_lookup_texture(&brw->ctx, texture);
538 if (!obj || obj->Target != target) {
539 *error = __DRI_IMAGE_ERROR_BAD_PARAMETER;
540 return NULL;
541 }
542
543 if (target == GL_TEXTURE_CUBE_MAP)
544 face = zoffset;
545
546 _mesa_test_texobj_completeness(&brw->ctx, obj);
547 iobj = intel_texture_object(obj);
548 if (!obj->_BaseComplete || (level > 0 && !obj->_MipmapComplete)) {
549 *error = __DRI_IMAGE_ERROR_BAD_PARAMETER;
550 return NULL;
551 }
552
553 if (level < obj->BaseLevel || level > obj->_MaxLevel) {
554 *error = __DRI_IMAGE_ERROR_BAD_MATCH;
555 return NULL;
556 }
557
558 if (target == GL_TEXTURE_3D && obj->Image[face][level]->Depth < zoffset) {
559 *error = __DRI_IMAGE_ERROR_BAD_MATCH;
560 return NULL;
561 }
562 image = calloc(1, sizeof *image);
563 if (image == NULL) {
564 *error = __DRI_IMAGE_ERROR_BAD_ALLOC;
565 return NULL;
566 }
567
568 image->internal_format = obj->Image[face][level]->InternalFormat;
569 image->format = obj->Image[face][level]->TexFormat;
570 image->modifier = tiling_to_modifier(
571 isl_tiling_to_i915_tiling(iobj->mt->surf.tiling));
572 image->data = loaderPrivate;
573 intel_setup_image_from_mipmap_tree(brw, image, iobj->mt, level, zoffset);
574 image->dri_format = driGLFormatToImageFormat(image->format);
575 image->has_depthstencil = iobj->mt->stencil_mt? true : false;
576 if (image->dri_format == MESA_FORMAT_NONE) {
577 *error = __DRI_IMAGE_ERROR_BAD_PARAMETER;
578 free(image);
579 return NULL;
580 }
581
582 *error = __DRI_IMAGE_ERROR_SUCCESS;
583 return image;
584 }
585
586 static void
587 intel_destroy_image(__DRIimage *image)
588 {
589 brw_bo_unreference(image->bo);
590 free(image);
591 }
592
593 enum modifier_priority {
594 MODIFIER_PRIORITY_INVALID = 0,
595 MODIFIER_PRIORITY_LINEAR,
596 MODIFIER_PRIORITY_X,
597 MODIFIER_PRIORITY_Y,
598 MODIFIER_PRIORITY_Y_CCS,
599 };
600
601 const uint64_t priority_to_modifier[] = {
602 [MODIFIER_PRIORITY_INVALID] = DRM_FORMAT_MOD_INVALID,
603 [MODIFIER_PRIORITY_LINEAR] = DRM_FORMAT_MOD_LINEAR,
604 [MODIFIER_PRIORITY_X] = I915_FORMAT_MOD_X_TILED,
605 [MODIFIER_PRIORITY_Y] = I915_FORMAT_MOD_Y_TILED,
606 [MODIFIER_PRIORITY_Y_CCS] = I915_FORMAT_MOD_Y_TILED_CCS,
607 };
608
609 static uint64_t
610 select_best_modifier(struct gen_device_info *devinfo,
611 int dri_format,
612 const uint64_t *modifiers,
613 const unsigned count)
614 {
615 enum modifier_priority prio = MODIFIER_PRIORITY_INVALID;
616
617 for (int i = 0; i < count; i++) {
618 if (!modifier_is_supported(devinfo, NULL, dri_format, modifiers[i]))
619 continue;
620
621 switch (modifiers[i]) {
622 case I915_FORMAT_MOD_Y_TILED_CCS:
623 prio = MAX2(prio, MODIFIER_PRIORITY_Y_CCS);
624 break;
625 case I915_FORMAT_MOD_Y_TILED:
626 prio = MAX2(prio, MODIFIER_PRIORITY_Y);
627 break;
628 case I915_FORMAT_MOD_X_TILED:
629 prio = MAX2(prio, MODIFIER_PRIORITY_X);
630 break;
631 case DRM_FORMAT_MOD_LINEAR:
632 prio = MAX2(prio, MODIFIER_PRIORITY_LINEAR);
633 break;
634 case DRM_FORMAT_MOD_INVALID:
635 default:
636 break;
637 }
638 }
639
640 return priority_to_modifier[prio];
641 }
642
643 static __DRIimage *
644 intel_create_image_common(__DRIscreen *dri_screen,
645 int width, int height, int format,
646 unsigned int use,
647 const uint64_t *modifiers,
648 unsigned count,
649 void *loaderPrivate)
650 {
651 __DRIimage *image;
652 struct intel_screen *screen = dri_screen->driverPrivate;
653 uint64_t modifier = DRM_FORMAT_MOD_INVALID;
654 bool ok;
655
656 /* Callers of this may specify a modifier, or a dri usage, but not both. The
657 * newer modifier interface deprecates the older usage flags newer modifier
658 * interface deprecates the older usage flags.
659 */
660 assert(!(use && count));
661
662 if (use & __DRI_IMAGE_USE_CURSOR) {
663 if (width != 64 || height != 64)
664 return NULL;
665 modifier = DRM_FORMAT_MOD_LINEAR;
666 }
667
668 if (use & __DRI_IMAGE_USE_LINEAR)
669 modifier = DRM_FORMAT_MOD_LINEAR;
670
671 if (modifier == DRM_FORMAT_MOD_INVALID) {
672 if (modifiers) {
673 /* User requested specific modifiers */
674 modifier = select_best_modifier(&screen->devinfo, format,
675 modifiers, count);
676 if (modifier == DRM_FORMAT_MOD_INVALID)
677 return NULL;
678 } else {
679 /* Historically, X-tiled was the default, and so lack of modifier means
680 * X-tiled.
681 */
682 modifier = I915_FORMAT_MOD_X_TILED;
683 }
684 }
685
686 image = intel_allocate_image(screen, format, loaderPrivate);
687 if (image == NULL)
688 return NULL;
689
690 const struct isl_drm_modifier_info *mod_info =
691 isl_drm_modifier_get_info(modifier);
692
693 struct isl_surf surf;
694 ok = isl_surf_init(&screen->isl_dev, &surf,
695 .dim = ISL_SURF_DIM_2D,
696 .format = brw_isl_format_for_mesa_format(image->format),
697 .width = width,
698 .height = height,
699 .depth = 1,
700 .levels = 1,
701 .array_len = 1,
702 .samples = 1,
703 .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT |
704 ISL_SURF_USAGE_TEXTURE_BIT |
705 ISL_SURF_USAGE_STORAGE_BIT,
706 .tiling_flags = (1 << mod_info->tiling));
707 assert(ok);
708 if (!ok) {
709 free(image);
710 return NULL;
711 }
712
713 struct isl_surf aux_surf;
714 if (mod_info->aux_usage == ISL_AUX_USAGE_CCS_E) {
715 ok = isl_surf_get_ccs_surf(&screen->isl_dev, &surf, &aux_surf, 0);
716 if (!ok) {
717 free(image);
718 return NULL;
719 }
720 } else {
721 assert(mod_info->aux_usage == ISL_AUX_USAGE_NONE);
722 aux_surf.size = 0;
723 }
724
725 /* We request that the bufmgr zero the buffer for us for two reasons:
726 *
727 * 1) If a buffer gets re-used from the pool, we don't want to leak random
728 * garbage from our process to some other.
729 *
730 * 2) For images with CCS_E, we want to ensure that the CCS starts off in
731 * a valid state. A CCS value of 0 indicates that the given block is
732 * in the pass-through state which is what we want.
733 */
734 image->bo = brw_bo_alloc_tiled(screen->bufmgr, "image",
735 surf.size + aux_surf.size,
736 isl_tiling_to_i915_tiling(mod_info->tiling),
737 surf.row_pitch, BO_ALLOC_ZEROED);
738 if (image->bo == NULL) {
739 free(image);
740 return NULL;
741 }
742 image->width = width;
743 image->height = height;
744 image->pitch = surf.row_pitch;
745 image->modifier = modifier;
746
747 if (aux_surf.size) {
748 image->aux_offset = surf.size;
749 image->aux_pitch = aux_surf.row_pitch;
750 }
751
752 return image;
753 }
754
755 static __DRIimage *
756 intel_create_image(__DRIscreen *dri_screen,
757 int width, int height, int format,
758 unsigned int use,
759 void *loaderPrivate)
760 {
761 return intel_create_image_common(dri_screen, width, height, format, use, NULL, 0,
762 loaderPrivate);
763 }
764
765 static void *
766 intel_map_image(__DRIcontext *context, __DRIimage *image,
767 int x0, int y0, int width, int height,
768 unsigned int flags, int *stride, void **map_info)
769 {
770 struct brw_context *brw = NULL;
771 struct brw_bo *bo = NULL;
772 void *raw_data = NULL;
773 GLuint pix_w = 1;
774 GLuint pix_h = 1;
775 GLint pix_bytes = 1;
776
777 if (!context || !image || !stride || !map_info || *map_info)
778 return NULL;
779
780 if (x0 < 0 || x0 >= image->width || width > image->width - x0)
781 return NULL;
782
783 if (y0 < 0 || y0 >= image->height || height > image->height - y0)
784 return NULL;
785
786 if (flags & MAP_INTERNAL_MASK)
787 return NULL;
788
789 brw = context->driverPrivate;
790 bo = image->bo;
791
792 assert(brw);
793 assert(bo);
794
795 /* DRI flags and GL_MAP.*_BIT flags are the same, so just pass them on. */
796 raw_data = brw_bo_map(brw, bo, flags);
797 if (!raw_data)
798 return NULL;
799
800 _mesa_get_format_block_size(image->format, &pix_w, &pix_h);
801 pix_bytes = _mesa_get_format_bytes(image->format);
802
803 assert(pix_w);
804 assert(pix_h);
805 assert(pix_bytes > 0);
806
807 raw_data += (x0 / pix_w) * pix_bytes + (y0 / pix_h) * image->pitch;
808
809 brw_bo_reference(bo);
810
811 *stride = image->pitch;
812 *map_info = bo;
813
814 return raw_data;
815 }
816
817 static void
818 intel_unmap_image(__DRIcontext *context, __DRIimage *image, void *map_info)
819 {
820 struct brw_bo *bo = map_info;
821
822 brw_bo_unmap(bo);
823 brw_bo_unreference(bo);
824 }
825
826 static __DRIimage *
827 intel_create_image_with_modifiers(__DRIscreen *dri_screen,
828 int width, int height, int format,
829 const uint64_t *modifiers,
830 const unsigned count,
831 void *loaderPrivate)
832 {
833 return intel_create_image_common(dri_screen, width, height, format, 0,
834 modifiers, count, loaderPrivate);
835 }
836
837 static GLboolean
838 intel_query_image(__DRIimage *image, int attrib, int *value)
839 {
840 switch (attrib) {
841 case __DRI_IMAGE_ATTRIB_STRIDE:
842 *value = image->pitch;
843 return true;
844 case __DRI_IMAGE_ATTRIB_HANDLE:
845 *value = brw_bo_export_gem_handle(image->bo);
846 return true;
847 case __DRI_IMAGE_ATTRIB_NAME:
848 return !brw_bo_flink(image->bo, (uint32_t *) value);
849 case __DRI_IMAGE_ATTRIB_FORMAT:
850 *value = image->dri_format;
851 return true;
852 case __DRI_IMAGE_ATTRIB_WIDTH:
853 *value = image->width;
854 return true;
855 case __DRI_IMAGE_ATTRIB_HEIGHT:
856 *value = image->height;
857 return true;
858 case __DRI_IMAGE_ATTRIB_COMPONENTS:
859 if (image->planar_format == NULL)
860 return false;
861 *value = image->planar_format->components;
862 return true;
863 case __DRI_IMAGE_ATTRIB_FD:
864 return !brw_bo_gem_export_to_prime(image->bo, value);
865 case __DRI_IMAGE_ATTRIB_FOURCC:
866 return intel_lookup_fourcc(image->dri_format, value);
867 case __DRI_IMAGE_ATTRIB_NUM_PLANES:
868 if (isl_drm_modifier_has_aux(image->modifier)) {
869 assert(!image->planar_format || image->planar_format->nplanes == 1);
870 *value = 2;
871 } else if (image->planar_format) {
872 *value = image->planar_format->nplanes;
873 } else {
874 *value = 1;
875 }
876 return true;
877 case __DRI_IMAGE_ATTRIB_OFFSET:
878 *value = image->offset;
879 return true;
880 case __DRI_IMAGE_ATTRIB_MODIFIER_LOWER:
881 *value = (image->modifier & 0xffffffff);
882 return true;
883 case __DRI_IMAGE_ATTRIB_MODIFIER_UPPER:
884 *value = ((image->modifier >> 32) & 0xffffffff);
885 return true;
886
887 default:
888 return false;
889 }
890 }
891
892 static GLboolean
893 intel_query_format_modifier_attribs(__DRIscreen *dri_screen,
894 uint32_t fourcc, uint64_t modifier,
895 int attrib, uint64_t *value)
896 {
897 struct intel_screen *screen = dri_screen->driverPrivate;
898 const struct intel_image_format *f = intel_image_format_lookup(fourcc);
899
900 if (!modifier_is_supported(&screen->devinfo, f, 0, modifier))
901 return false;
902
903 switch (attrib) {
904 case __DRI_IMAGE_FORMAT_MODIFIER_ATTRIB_PLANE_COUNT:
905 *value = isl_drm_modifier_has_aux(modifier) ? 2 : f->nplanes;
906 return true;
907
908 default:
909 return false;
910 }
911 }
912
913 static __DRIimage *
914 intel_dup_image(__DRIimage *orig_image, void *loaderPrivate)
915 {
916 __DRIimage *image;
917
918 image = calloc(1, sizeof *image);
919 if (image == NULL)
920 return NULL;
921
922 brw_bo_reference(orig_image->bo);
923 image->bo = orig_image->bo;
924 image->internal_format = orig_image->internal_format;
925 image->planar_format = orig_image->planar_format;
926 image->dri_format = orig_image->dri_format;
927 image->format = orig_image->format;
928 image->modifier = orig_image->modifier;
929 image->offset = orig_image->offset;
930 image->width = orig_image->width;
931 image->height = orig_image->height;
932 image->pitch = orig_image->pitch;
933 image->tile_x = orig_image->tile_x;
934 image->tile_y = orig_image->tile_y;
935 image->has_depthstencil = orig_image->has_depthstencil;
936 image->data = loaderPrivate;
937 image->dma_buf_imported = orig_image->dma_buf_imported;
938 image->aux_offset = orig_image->aux_offset;
939 image->aux_pitch = orig_image->aux_pitch;
940
941 memcpy(image->strides, orig_image->strides, sizeof(image->strides));
942 memcpy(image->offsets, orig_image->offsets, sizeof(image->offsets));
943
944 return image;
945 }
946
947 static GLboolean
948 intel_validate_usage(__DRIimage *image, unsigned int use)
949 {
950 if (use & __DRI_IMAGE_USE_CURSOR) {
951 if (image->width != 64 || image->height != 64)
952 return GL_FALSE;
953 }
954
955 return GL_TRUE;
956 }
957
958 static __DRIimage *
959 intel_create_image_from_names(__DRIscreen *dri_screen,
960 int width, int height, int fourcc,
961 int *names, int num_names,
962 int *strides, int *offsets,
963 void *loaderPrivate)
964 {
965 const struct intel_image_format *f = NULL;
966 __DRIimage *image;
967 int i, index;
968
969 if (dri_screen == NULL || names == NULL || num_names != 1)
970 return NULL;
971
972 f = intel_image_format_lookup(fourcc);
973 if (f == NULL)
974 return NULL;
975
976 image = intel_create_image_from_name(dri_screen, width, height,
977 __DRI_IMAGE_FORMAT_NONE,
978 names[0], strides[0],
979 loaderPrivate);
980
981 if (image == NULL)
982 return NULL;
983
984 image->planar_format = f;
985 for (i = 0; i < f->nplanes; i++) {
986 index = f->planes[i].buffer_index;
987 image->offsets[index] = offsets[index];
988 image->strides[index] = strides[index];
989 }
990
991 return image;
992 }
993
994 static __DRIimage *
995 intel_create_image_from_fds_common(__DRIscreen *dri_screen,
996 int width, int height, int fourcc,
997 uint64_t modifier, int *fds, int num_fds,
998 int *strides, int *offsets,
999 void *loaderPrivate)
1000 {
1001 struct intel_screen *screen = dri_screen->driverPrivate;
1002 const struct intel_image_format *f;
1003 __DRIimage *image;
1004 int i, index;
1005 bool ok;
1006
1007 if (fds == NULL || num_fds < 1)
1008 return NULL;
1009
1010 f = intel_image_format_lookup(fourcc);
1011 if (f == NULL)
1012 return NULL;
1013
1014 if (modifier != DRM_FORMAT_MOD_INVALID &&
1015 !modifier_is_supported(&screen->devinfo, f, 0, modifier))
1016 return NULL;
1017
1018 if (f->nplanes == 1)
1019 image = intel_allocate_image(screen, f->planes[0].dri_format,
1020 loaderPrivate);
1021 else
1022 image = intel_allocate_image(screen, __DRI_IMAGE_FORMAT_NONE,
1023 loaderPrivate);
1024
1025 if (image == NULL)
1026 return NULL;
1027
1028 image->width = width;
1029 image->height = height;
1030 image->pitch = strides[0];
1031
1032 image->planar_format = f;
1033
1034 image->bo = brw_bo_gem_create_from_prime(screen->bufmgr, fds[0]);
1035 if (image->bo == NULL) {
1036 free(image);
1037 return NULL;
1038 }
1039
1040 /* We only support all planes from the same bo.
1041 * brw_bo_gem_create_from_prime() should return the same pointer for all
1042 * fds received here */
1043 for (i = 1; i < num_fds; i++) {
1044 struct brw_bo *aux = brw_bo_gem_create_from_prime(screen->bufmgr, fds[i]);
1045 brw_bo_unreference(aux);
1046 if (aux != image->bo) {
1047 brw_bo_unreference(image->bo);
1048 free(image);
1049 return NULL;
1050 }
1051 }
1052
1053 if (modifier != DRM_FORMAT_MOD_INVALID)
1054 image->modifier = modifier;
1055 else
1056 image->modifier = tiling_to_modifier(image->bo->tiling_mode);
1057
1058 const struct isl_drm_modifier_info *mod_info =
1059 isl_drm_modifier_get_info(image->modifier);
1060
1061 int size = 0;
1062 struct isl_surf surf;
1063 for (i = 0; i < f->nplanes; i++) {
1064 index = f->planes[i].buffer_index;
1065 image->offsets[index] = offsets[index];
1066 image->strides[index] = strides[index];
1067
1068 mesa_format format = driImageFormatToGLFormat(f->planes[i].dri_format);
1069
1070 ok = isl_surf_init(&screen->isl_dev, &surf,
1071 .dim = ISL_SURF_DIM_2D,
1072 .format = brw_isl_format_for_mesa_format(format),
1073 .width = image->width >> f->planes[i].width_shift,
1074 .height = image->height >> f->planes[i].height_shift,
1075 .depth = 1,
1076 .levels = 1,
1077 .array_len = 1,
1078 .samples = 1,
1079 .row_pitch = strides[index],
1080 .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT |
1081 ISL_SURF_USAGE_TEXTURE_BIT |
1082 ISL_SURF_USAGE_STORAGE_BIT,
1083 .tiling_flags = (1 << mod_info->tiling));
1084 if (!ok) {
1085 brw_bo_unreference(image->bo);
1086 free(image);
1087 return NULL;
1088 }
1089
1090 const int end = offsets[index] + surf.size;
1091 if (size < end)
1092 size = end;
1093 }
1094
1095 if (mod_info->aux_usage == ISL_AUX_USAGE_CCS_E) {
1096 /* Even though we initialize surf in the loop above, we know that
1097 * anything with CCS_E will have exactly one plane so surf is properly
1098 * initialized when we get here.
1099 */
1100 assert(f->nplanes == 1);
1101
1102 image->aux_offset = offsets[1];
1103 image->aux_pitch = strides[1];
1104
1105 /* Scanout hardware requires that the CCS be placed after the main
1106 * surface in memory. We consider any CCS that is placed any earlier in
1107 * memory to be invalid and reject it.
1108 *
1109 * At some point in the future, this restriction may be relaxed if the
1110 * hardware becomes less strict but we may need a new modifier for that.
1111 */
1112 assert(size > 0);
1113 if (image->aux_offset < size) {
1114 brw_bo_unreference(image->bo);
1115 free(image);
1116 return NULL;
1117 }
1118
1119 struct isl_surf aux_surf;
1120 ok = isl_surf_get_ccs_surf(&screen->isl_dev, &surf, &aux_surf,
1121 image->aux_pitch);
1122 if (!ok) {
1123 brw_bo_unreference(image->bo);
1124 free(image);
1125 return NULL;
1126 }
1127
1128 const int end = image->aux_offset + aux_surf.size;
1129 if (size < end)
1130 size = end;
1131 } else {
1132 assert(mod_info->aux_usage == ISL_AUX_USAGE_NONE);
1133 }
1134
1135 /* Check that the requested image actually fits within the BO. 'size'
1136 * is already relative to the offsets, so we don't need to add that. */
1137 if (image->bo->size == 0) {
1138 image->bo->size = size;
1139 } else if (size > image->bo->size) {
1140 brw_bo_unreference(image->bo);
1141 free(image);
1142 return NULL;
1143 }
1144
1145 if (f->nplanes == 1) {
1146 image->offset = image->offsets[0];
1147 intel_image_warn_if_unaligned(image, __func__);
1148 }
1149
1150 return image;
1151 }
1152
1153 static __DRIimage *
1154 intel_create_image_from_fds(__DRIscreen *dri_screen,
1155 int width, int height, int fourcc,
1156 int *fds, int num_fds, int *strides, int *offsets,
1157 void *loaderPrivate)
1158 {
1159 return intel_create_image_from_fds_common(dri_screen, width, height, fourcc,
1160 DRM_FORMAT_MOD_INVALID,
1161 fds, num_fds, strides, offsets,
1162 loaderPrivate);
1163 }
1164
1165 static __DRIimage *
1166 intel_create_image_from_dma_bufs2(__DRIscreen *dri_screen,
1167 int width, int height,
1168 int fourcc, uint64_t modifier,
1169 int *fds, int num_fds,
1170 int *strides, int *offsets,
1171 enum __DRIYUVColorSpace yuv_color_space,
1172 enum __DRISampleRange sample_range,
1173 enum __DRIChromaSiting horizontal_siting,
1174 enum __DRIChromaSiting vertical_siting,
1175 unsigned *error,
1176 void *loaderPrivate)
1177 {
1178 __DRIimage *image;
1179 const struct intel_image_format *f = intel_image_format_lookup(fourcc);
1180
1181 if (!f) {
1182 *error = __DRI_IMAGE_ERROR_BAD_MATCH;
1183 return NULL;
1184 }
1185
1186 image = intel_create_image_from_fds_common(dri_screen, width, height,
1187 fourcc, modifier,
1188 fds, num_fds, strides, offsets,
1189 loaderPrivate);
1190
1191 /*
1192 * Invalid parameters and any inconsistencies between are assumed to be
1193 * checked by the caller. Therefore besides unsupported formats one can fail
1194 * only in allocation.
1195 */
1196 if (!image) {
1197 *error = __DRI_IMAGE_ERROR_BAD_ALLOC;
1198 return NULL;
1199 }
1200
1201 image->dma_buf_imported = true;
1202 image->yuv_color_space = yuv_color_space;
1203 image->sample_range = sample_range;
1204 image->horizontal_siting = horizontal_siting;
1205 image->vertical_siting = vertical_siting;
1206
1207 *error = __DRI_IMAGE_ERROR_SUCCESS;
1208 return image;
1209 }
1210
1211 static __DRIimage *
1212 intel_create_image_from_dma_bufs(__DRIscreen *dri_screen,
1213 int width, int height, int fourcc,
1214 int *fds, int num_fds,
1215 int *strides, int *offsets,
1216 enum __DRIYUVColorSpace yuv_color_space,
1217 enum __DRISampleRange sample_range,
1218 enum __DRIChromaSiting horizontal_siting,
1219 enum __DRIChromaSiting vertical_siting,
1220 unsigned *error,
1221 void *loaderPrivate)
1222 {
1223 return intel_create_image_from_dma_bufs2(dri_screen, width, height,
1224 fourcc, DRM_FORMAT_MOD_INVALID,
1225 fds, num_fds, strides, offsets,
1226 yuv_color_space,
1227 sample_range,
1228 horizontal_siting,
1229 vertical_siting,
1230 error,
1231 loaderPrivate);
1232 }
1233
1234 static GLboolean
1235 intel_query_dma_buf_formats(__DRIscreen *screen, int max,
1236 int *formats, int *count)
1237 {
1238 int i, j = 0;
1239
1240 if (max == 0) {
1241 *count = ARRAY_SIZE(intel_image_formats) - 1; /* not SARGB */
1242 return true;
1243 }
1244
1245 for (i = 0; i < (ARRAY_SIZE(intel_image_formats)) && j < max; i++) {
1246 if (intel_image_formats[i].fourcc == __DRI_IMAGE_FOURCC_SARGB8888)
1247 continue;
1248 formats[j++] = intel_image_formats[i].fourcc;
1249 }
1250
1251 *count = j;
1252 return true;
1253 }
1254
1255 static GLboolean
1256 intel_query_dma_buf_modifiers(__DRIscreen *_screen, int fourcc, int max,
1257 uint64_t *modifiers,
1258 unsigned int *external_only,
1259 int *count)
1260 {
1261 struct intel_screen *screen = _screen->driverPrivate;
1262 const struct intel_image_format *f;
1263 int num_mods = 0, i;
1264
1265 f = intel_image_format_lookup(fourcc);
1266 if (f == NULL)
1267 return false;
1268
1269 for (i = 0; i < ARRAY_SIZE(supported_modifiers); i++) {
1270 uint64_t modifier = supported_modifiers[i].modifier;
1271 if (!modifier_is_supported(&screen->devinfo, f, 0, modifier))
1272 continue;
1273
1274 num_mods++;
1275 if (max == 0)
1276 continue;
1277
1278 modifiers[num_mods - 1] = modifier;
1279 if (num_mods >= max)
1280 break;
1281 }
1282
1283 if (external_only != NULL) {
1284 for (i = 0; i < num_mods && i < max; i++) {
1285 if (f->components == __DRI_IMAGE_COMPONENTS_Y_U_V ||
1286 f->components == __DRI_IMAGE_COMPONENTS_Y_UV ||
1287 f->components == __DRI_IMAGE_COMPONENTS_Y_XUXV) {
1288 external_only[i] = GL_TRUE;
1289 }
1290 else {
1291 external_only[i] = GL_FALSE;
1292 }
1293 }
1294 }
1295
1296 *count = num_mods;
1297 return true;
1298 }
1299
1300 static __DRIimage *
1301 intel_from_planar(__DRIimage *parent, int plane, void *loaderPrivate)
1302 {
1303 int width, height, offset, stride, dri_format, index;
1304 const struct intel_image_format *f;
1305 __DRIimage *image;
1306
1307 if (parent == NULL) {
1308 return NULL;
1309 } else if (parent->planar_format == NULL) {
1310 const bool is_aux =
1311 isl_drm_modifier_has_aux(parent->modifier) && plane == 1;
1312 if (!is_aux)
1313 return NULL;
1314
1315 width = parent->width;
1316 height = parent->height;
1317 dri_format = parent->dri_format;
1318 offset = parent->aux_offset;
1319 stride = parent->aux_pitch;
1320 } else {
1321 /* Planar formats don't support aux buffers/images */
1322 assert(!isl_drm_modifier_has_aux(parent->modifier));
1323 f = parent->planar_format;
1324
1325 if (plane >= f->nplanes)
1326 return NULL;
1327
1328 width = parent->width >> f->planes[plane].width_shift;
1329 height = parent->height >> f->planes[plane].height_shift;
1330 dri_format = f->planes[plane].dri_format;
1331 index = f->planes[plane].buffer_index;
1332 offset = parent->offsets[index];
1333 stride = parent->strides[index];
1334
1335 if (offset + height * stride > parent->bo->size) {
1336 _mesa_warning(NULL, "intel_create_sub_image: subimage out of bounds");
1337 return NULL;
1338 }
1339 }
1340
1341 image = intel_allocate_image(parent->screen, dri_format, loaderPrivate);
1342 if (image == NULL)
1343 return NULL;
1344
1345 image->bo = parent->bo;
1346 brw_bo_reference(parent->bo);
1347 image->modifier = parent->modifier;
1348
1349 image->width = width;
1350 image->height = height;
1351 image->pitch = stride;
1352 image->offset = offset;
1353
1354 intel_image_warn_if_unaligned(image, __func__);
1355
1356 return image;
1357 }
1358
1359 static const __DRIimageExtension intelImageExtension = {
1360 .base = { __DRI_IMAGE, 16 },
1361
1362 .createImageFromName = intel_create_image_from_name,
1363 .createImageFromRenderbuffer = intel_create_image_from_renderbuffer,
1364 .destroyImage = intel_destroy_image,
1365 .createImage = intel_create_image,
1366 .queryImage = intel_query_image,
1367 .dupImage = intel_dup_image,
1368 .validateUsage = intel_validate_usage,
1369 .createImageFromNames = intel_create_image_from_names,
1370 .fromPlanar = intel_from_planar,
1371 .createImageFromTexture = intel_create_image_from_texture,
1372 .createImageFromFds = intel_create_image_from_fds,
1373 .createImageFromDmaBufs = intel_create_image_from_dma_bufs,
1374 .blitImage = NULL,
1375 .getCapabilities = NULL,
1376 .mapImage = intel_map_image,
1377 .unmapImage = intel_unmap_image,
1378 .createImageWithModifiers = intel_create_image_with_modifiers,
1379 .createImageFromDmaBufs2 = intel_create_image_from_dma_bufs2,
1380 .queryDmaBufFormats = intel_query_dma_buf_formats,
1381 .queryDmaBufModifiers = intel_query_dma_buf_modifiers,
1382 .queryDmaBufFormatModifierAttribs = intel_query_format_modifier_attribs,
1383 };
1384
1385 static uint64_t
1386 get_aperture_size(int fd)
1387 {
1388 struct drm_i915_gem_get_aperture aperture;
1389
1390 if (drmIoctl(fd, DRM_IOCTL_I915_GEM_GET_APERTURE, &aperture) != 0)
1391 return 0;
1392
1393 return aperture.aper_size;
1394 }
1395
1396 static int
1397 brw_query_renderer_integer(__DRIscreen *dri_screen,
1398 int param, unsigned int *value)
1399 {
1400 const struct intel_screen *const screen =
1401 (struct intel_screen *) dri_screen->driverPrivate;
1402
1403 switch (param) {
1404 case __DRI2_RENDERER_VENDOR_ID:
1405 value[0] = 0x8086;
1406 return 0;
1407 case __DRI2_RENDERER_DEVICE_ID:
1408 value[0] = screen->deviceID;
1409 return 0;
1410 case __DRI2_RENDERER_ACCELERATED:
1411 value[0] = 1;
1412 return 0;
1413 case __DRI2_RENDERER_VIDEO_MEMORY: {
1414 /* Once a batch uses more than 75% of the maximum mappable size, we
1415 * assume that there's some fragmentation, and we start doing extra
1416 * flushing, etc. That's the big cliff apps will care about.
1417 */
1418 const unsigned gpu_mappable_megabytes =
1419 screen->aperture_threshold / (1024 * 1024);
1420
1421 const long system_memory_pages = sysconf(_SC_PHYS_PAGES);
1422 const long system_page_size = sysconf(_SC_PAGE_SIZE);
1423
1424 if (system_memory_pages <= 0 || system_page_size <= 0)
1425 return -1;
1426
1427 const uint64_t system_memory_bytes = (uint64_t) system_memory_pages
1428 * (uint64_t) system_page_size;
1429
1430 const unsigned system_memory_megabytes =
1431 (unsigned) (system_memory_bytes / (1024 * 1024));
1432
1433 value[0] = MIN2(system_memory_megabytes, gpu_mappable_megabytes);
1434 return 0;
1435 }
1436 case __DRI2_RENDERER_UNIFIED_MEMORY_ARCHITECTURE:
1437 value[0] = 1;
1438 return 0;
1439 case __DRI2_RENDERER_HAS_TEXTURE_3D:
1440 value[0] = 1;
1441 return 0;
1442 case __DRI2_RENDERER_HAS_CONTEXT_PRIORITY:
1443 value[0] = 0;
1444 if (brw_hw_context_set_priority(screen->bufmgr,
1445 0, BRW_CONTEXT_HIGH_PRIORITY) == 0)
1446 value[0] |= __DRI2_RENDERER_HAS_CONTEXT_PRIORITY_HIGH;
1447 if (brw_hw_context_set_priority(screen->bufmgr,
1448 0, BRW_CONTEXT_LOW_PRIORITY) == 0)
1449 value[0] |= __DRI2_RENDERER_HAS_CONTEXT_PRIORITY_LOW;
1450 /* reset to default last, just in case */
1451 if (brw_hw_context_set_priority(screen->bufmgr,
1452 0, BRW_CONTEXT_MEDIUM_PRIORITY) == 0)
1453 value[0] |= __DRI2_RENDERER_HAS_CONTEXT_PRIORITY_MEDIUM;
1454 return 0;
1455 case __DRI2_RENDERER_HAS_FRAMEBUFFER_SRGB:
1456 value[0] = 1;
1457 return 0;
1458 default:
1459 return driQueryRendererIntegerCommon(dri_screen, param, value);
1460 }
1461
1462 return -1;
1463 }
1464
1465 static int
1466 brw_query_renderer_string(__DRIscreen *dri_screen,
1467 int param, const char **value)
1468 {
1469 const struct intel_screen *screen =
1470 (struct intel_screen *) dri_screen->driverPrivate;
1471
1472 switch (param) {
1473 case __DRI2_RENDERER_VENDOR_ID:
1474 value[0] = brw_vendor_string;
1475 return 0;
1476 case __DRI2_RENDERER_DEVICE_ID:
1477 value[0] = brw_get_renderer_string(screen);
1478 return 0;
1479 default:
1480 break;
1481 }
1482
1483 return -1;
1484 }
1485
1486 static const __DRI2rendererQueryExtension intelRendererQueryExtension = {
1487 .base = { __DRI2_RENDERER_QUERY, 1 },
1488
1489 .queryInteger = brw_query_renderer_integer,
1490 .queryString = brw_query_renderer_string
1491 };
1492
1493 static const __DRIrobustnessExtension dri2Robustness = {
1494 .base = { __DRI2_ROBUSTNESS, 1 }
1495 };
1496
1497 static const __DRIextension *screenExtensions[] = {
1498 &intelTexBufferExtension.base,
1499 &intelFenceExtension.base,
1500 &intelFlushExtension.base,
1501 &intelImageExtension.base,
1502 &intelRendererQueryExtension.base,
1503 &dri2ConfigQueryExtension.base,
1504 &dri2NoErrorExtension.base,
1505 &dri2FlushControlExtension.base,
1506 NULL
1507 };
1508
1509 static const __DRIextension *intelRobustScreenExtensions[] = {
1510 &intelTexBufferExtension.base,
1511 &intelFenceExtension.base,
1512 &intelFlushExtension.base,
1513 &intelImageExtension.base,
1514 &intelRendererQueryExtension.base,
1515 &dri2ConfigQueryExtension.base,
1516 &dri2FlushControlExtension.base,
1517 &dri2Robustness.base,
1518 &dri2NoErrorExtension.base,
1519 NULL
1520 };
1521
1522 static int
1523 intel_get_param(struct intel_screen *screen, int param, int *value)
1524 {
1525 int ret = 0;
1526 struct drm_i915_getparam gp;
1527
1528 memset(&gp, 0, sizeof(gp));
1529 gp.param = param;
1530 gp.value = value;
1531
1532 if (drmIoctl(screen->driScrnPriv->fd, DRM_IOCTL_I915_GETPARAM, &gp) == -1) {
1533 ret = -errno;
1534 if (ret != -EINVAL)
1535 _mesa_warning(NULL, "drm_i915_getparam: %d", ret);
1536 }
1537
1538 return ret;
1539 }
1540
1541 static bool
1542 intel_get_boolean(struct intel_screen *screen, int param)
1543 {
1544 int value = 0;
1545 return (intel_get_param(screen, param, &value) == 0) && value;
1546 }
1547
1548 static int
1549 intel_get_integer(struct intel_screen *screen, int param)
1550 {
1551 int value = -1;
1552
1553 if (intel_get_param(screen, param, &value) == 0)
1554 return value;
1555
1556 return -1;
1557 }
1558
1559 static void
1560 intelDestroyScreen(__DRIscreen * sPriv)
1561 {
1562 struct intel_screen *screen = sPriv->driverPrivate;
1563
1564 brw_bufmgr_destroy(screen->bufmgr);
1565 driDestroyOptionInfo(&screen->optionCache);
1566
1567 ralloc_free(screen);
1568 sPriv->driverPrivate = NULL;
1569 }
1570
1571
1572 /**
1573 * Create a gl_framebuffer and attach it to __DRIdrawable::driverPrivate.
1574 *
1575 *_This implements driDriverAPI::createNewDrawable, which the DRI layer calls
1576 * when creating a EGLSurface, GLXDrawable, or GLXPixmap. Despite the name,
1577 * this does not allocate GPU memory.
1578 */
1579 static GLboolean
1580 intelCreateBuffer(__DRIscreen *dri_screen,
1581 __DRIdrawable * driDrawPriv,
1582 const struct gl_config * mesaVis, GLboolean isPixmap)
1583 {
1584 struct intel_renderbuffer *rb;
1585 struct intel_screen *screen = (struct intel_screen *)
1586 dri_screen->driverPrivate;
1587 mesa_format rgbFormat;
1588 unsigned num_samples =
1589 intel_quantize_num_samples(screen, mesaVis->samples);
1590
1591 if (isPixmap)
1592 return false;
1593
1594 struct gl_framebuffer *fb = CALLOC_STRUCT(gl_framebuffer);
1595 if (!fb)
1596 return false;
1597
1598 _mesa_initialize_window_framebuffer(fb, mesaVis);
1599
1600 if (screen->winsys_msaa_samples_override != -1) {
1601 num_samples = screen->winsys_msaa_samples_override;
1602 fb->Visual.samples = num_samples;
1603 }
1604
1605 if (mesaVis->redBits == 5) {
1606 rgbFormat = mesaVis->redMask == 0x1f ? MESA_FORMAT_R5G6B5_UNORM
1607 : MESA_FORMAT_B5G6R5_UNORM;
1608 } else if (mesaVis->sRGBCapable) {
1609 rgbFormat = mesaVis->redMask == 0xff ? MESA_FORMAT_R8G8B8A8_SRGB
1610 : MESA_FORMAT_B8G8R8A8_SRGB;
1611 } else if (mesaVis->alphaBits == 0) {
1612 rgbFormat = mesaVis->redMask == 0xff ? MESA_FORMAT_R8G8B8X8_UNORM
1613 : MESA_FORMAT_B8G8R8X8_UNORM;
1614 } else {
1615 rgbFormat = mesaVis->redMask == 0xff ? MESA_FORMAT_R8G8B8A8_SRGB
1616 : MESA_FORMAT_B8G8R8A8_SRGB;
1617 fb->Visual.sRGBCapable = true;
1618 }
1619
1620 /* mesaVis->sRGBCapable was set, user is asking for sRGB */
1621 bool srgb_cap_set = mesaVis->redBits >= 8 && mesaVis->sRGBCapable;
1622
1623 /* setup the hardware-based renderbuffers */
1624 rb = intel_create_winsys_renderbuffer(screen, rgbFormat, num_samples);
1625 _mesa_attach_and_own_rb(fb, BUFFER_FRONT_LEFT, &rb->Base.Base);
1626 rb->need_srgb = srgb_cap_set;
1627
1628 if (mesaVis->doubleBufferMode) {
1629 rb = intel_create_winsys_renderbuffer(screen, rgbFormat, num_samples);
1630 _mesa_attach_and_own_rb(fb, BUFFER_BACK_LEFT, &rb->Base.Base);
1631 rb->need_srgb = srgb_cap_set;
1632 }
1633
1634 /*
1635 * Assert here that the gl_config has an expected depth/stencil bit
1636 * combination: one of d24/s8, d16/s0, d0/s0. (See intelInitScreen2(),
1637 * which constructs the advertised configs.)
1638 */
1639 if (mesaVis->depthBits == 24) {
1640 assert(mesaVis->stencilBits == 8);
1641
1642 if (screen->devinfo.has_hiz_and_separate_stencil) {
1643 rb = intel_create_private_renderbuffer(screen,
1644 MESA_FORMAT_Z24_UNORM_X8_UINT,
1645 num_samples);
1646 _mesa_attach_and_own_rb(fb, BUFFER_DEPTH, &rb->Base.Base);
1647 rb = intel_create_private_renderbuffer(screen, MESA_FORMAT_S_UINT8,
1648 num_samples);
1649 _mesa_attach_and_own_rb(fb, BUFFER_STENCIL, &rb->Base.Base);
1650 } else {
1651 /*
1652 * Use combined depth/stencil. Note that the renderbuffer is
1653 * attached to two attachment points.
1654 */
1655 rb = intel_create_private_renderbuffer(screen,
1656 MESA_FORMAT_Z24_UNORM_S8_UINT,
1657 num_samples);
1658 _mesa_attach_and_own_rb(fb, BUFFER_DEPTH, &rb->Base.Base);
1659 _mesa_attach_and_reference_rb(fb, BUFFER_STENCIL, &rb->Base.Base);
1660 }
1661 }
1662 else if (mesaVis->depthBits == 16) {
1663 assert(mesaVis->stencilBits == 0);
1664 rb = intel_create_private_renderbuffer(screen, MESA_FORMAT_Z_UNORM16,
1665 num_samples);
1666 _mesa_attach_and_own_rb(fb, BUFFER_DEPTH, &rb->Base.Base);
1667 }
1668 else {
1669 assert(mesaVis->depthBits == 0);
1670 assert(mesaVis->stencilBits == 0);
1671 }
1672
1673 /* now add any/all software-based renderbuffers we may need */
1674 _swrast_add_soft_renderbuffers(fb,
1675 false, /* never sw color */
1676 false, /* never sw depth */
1677 false, /* never sw stencil */
1678 mesaVis->accumRedBits > 0,
1679 false, /* never sw alpha */
1680 false /* never sw aux */ );
1681 driDrawPriv->driverPrivate = fb;
1682
1683 return true;
1684 }
1685
1686 static void
1687 intelDestroyBuffer(__DRIdrawable * driDrawPriv)
1688 {
1689 struct gl_framebuffer *fb = driDrawPriv->driverPrivate;
1690
1691 _mesa_reference_framebuffer(&fb, NULL);
1692 }
1693
1694 static void
1695 intel_cs_timestamp_frequency(struct intel_screen *screen)
1696 {
1697 /* We shouldn't need to update gen_device_info.timestamp_frequency prior to
1698 * gen10, PCI-id is enough to figure it out.
1699 */
1700 assert(screen->devinfo.gen >= 10);
1701
1702 int ret, freq;
1703
1704 ret = intel_get_param(screen, I915_PARAM_CS_TIMESTAMP_FREQUENCY,
1705 &freq);
1706 if (ret < 0) {
1707 _mesa_warning(NULL,
1708 "Kernel 4.15 required to read the CS timestamp frequency.\n");
1709 return;
1710 }
1711
1712 screen->devinfo.timestamp_frequency = freq;
1713 }
1714
1715 static void
1716 intel_detect_sseu(struct intel_screen *screen)
1717 {
1718 assert(screen->devinfo.gen >= 8);
1719 int ret;
1720
1721 screen->subslice_total = -1;
1722 screen->eu_total = -1;
1723
1724 ret = intel_get_param(screen, I915_PARAM_SUBSLICE_TOTAL,
1725 &screen->subslice_total);
1726 if (ret < 0 && ret != -EINVAL)
1727 goto err_out;
1728
1729 ret = intel_get_param(screen,
1730 I915_PARAM_EU_TOTAL, &screen->eu_total);
1731 if (ret < 0 && ret != -EINVAL)
1732 goto err_out;
1733
1734 /* Without this information, we cannot get the right Braswell brandstrings,
1735 * and we have to use conservative numbers for GPGPU on many platforms, but
1736 * otherwise, things will just work.
1737 */
1738 if (screen->subslice_total < 1 || screen->eu_total < 1)
1739 _mesa_warning(NULL,
1740 "Kernel 4.1 required to properly query GPU properties.\n");
1741
1742 return;
1743
1744 err_out:
1745 screen->subslice_total = -1;
1746 screen->eu_total = -1;
1747 _mesa_warning(NULL, "Failed to query GPU properties (%s).\n", strerror(-ret));
1748 }
1749
1750 static bool
1751 intel_init_bufmgr(struct intel_screen *screen)
1752 {
1753 __DRIscreen *dri_screen = screen->driScrnPriv;
1754
1755 if (getenv("INTEL_NO_HW") != NULL)
1756 screen->no_hw = true;
1757
1758 screen->bufmgr = brw_bufmgr_init(&screen->devinfo, dri_screen->fd);
1759 if (screen->bufmgr == NULL) {
1760 fprintf(stderr, "[%s:%u] Error initializing buffer manager.\n",
1761 __func__, __LINE__);
1762 return false;
1763 }
1764
1765 if (!intel_get_boolean(screen, I915_PARAM_HAS_WAIT_TIMEOUT)) {
1766 fprintf(stderr, "[%s: %u] Kernel 3.6 required.\n", __func__, __LINE__);
1767 return false;
1768 }
1769
1770 return true;
1771 }
1772
1773 static bool
1774 intel_detect_swizzling(struct intel_screen *screen)
1775 {
1776 struct brw_bo *buffer;
1777 unsigned flags = 0;
1778 uint32_t aligned_pitch;
1779 uint32_t tiling = I915_TILING_X;
1780 uint32_t swizzle_mode = 0;
1781
1782 buffer = brw_bo_alloc_tiled_2d(screen->bufmgr, "swizzle test",
1783 64, 64, 4, tiling, &aligned_pitch, flags);
1784 if (buffer == NULL)
1785 return false;
1786
1787 brw_bo_get_tiling(buffer, &tiling, &swizzle_mode);
1788 brw_bo_unreference(buffer);
1789
1790 if (swizzle_mode == I915_BIT_6_SWIZZLE_NONE)
1791 return false;
1792 else
1793 return true;
1794 }
1795
1796 static int
1797 intel_detect_timestamp(struct intel_screen *screen)
1798 {
1799 uint64_t dummy = 0, last = 0;
1800 int upper, lower, loops;
1801
1802 /* On 64bit systems, some old kernels trigger a hw bug resulting in the
1803 * TIMESTAMP register being shifted and the low 32bits always zero.
1804 *
1805 * More recent kernels offer an interface to read the full 36bits
1806 * everywhere.
1807 */
1808 if (brw_reg_read(screen->bufmgr, TIMESTAMP | 1, &dummy) == 0)
1809 return 3;
1810
1811 /* Determine if we have a 32bit or 64bit kernel by inspecting the
1812 * upper 32bits for a rapidly changing timestamp.
1813 */
1814 if (brw_reg_read(screen->bufmgr, TIMESTAMP, &last))
1815 return 0;
1816
1817 upper = lower = 0;
1818 for (loops = 0; loops < 10; loops++) {
1819 /* The TIMESTAMP should change every 80ns, so several round trips
1820 * through the kernel should be enough to advance it.
1821 */
1822 if (brw_reg_read(screen->bufmgr, TIMESTAMP, &dummy))
1823 return 0;
1824
1825 upper += (dummy >> 32) != (last >> 32);
1826 if (upper > 1) /* beware 32bit counter overflow */
1827 return 2; /* upper dword holds the low 32bits of the timestamp */
1828
1829 lower += (dummy & 0xffffffff) != (last & 0xffffffff);
1830 if (lower > 1)
1831 return 1; /* timestamp is unshifted */
1832
1833 last = dummy;
1834 }
1835
1836 /* No advancement? No timestamp! */
1837 return 0;
1838 }
1839
1840 /**
1841 * Test if we can use MI_LOAD_REGISTER_MEM from an untrusted batchbuffer.
1842 *
1843 * Some combinations of hardware and kernel versions allow this feature,
1844 * while others don't. Instead of trying to enumerate every case, just
1845 * try and write a register and see if works.
1846 */
1847 static bool
1848 intel_detect_pipelined_register(struct intel_screen *screen,
1849 int reg, uint32_t expected_value, bool reset)
1850 {
1851 if (screen->no_hw)
1852 return false;
1853
1854 struct brw_bo *results, *bo;
1855 uint32_t *batch;
1856 uint32_t offset = 0;
1857 void *map;
1858 bool success = false;
1859
1860 /* Create a zero'ed temporary buffer for reading our results */
1861 results = brw_bo_alloc(screen->bufmgr, "registers", 4096, 0);
1862 if (results == NULL)
1863 goto err;
1864
1865 bo = brw_bo_alloc(screen->bufmgr, "batchbuffer", 4096, 0);
1866 if (bo == NULL)
1867 goto err_results;
1868
1869 map = brw_bo_map(NULL, bo, MAP_WRITE);
1870 if (!map)
1871 goto err_batch;
1872
1873 batch = map;
1874
1875 /* Write the register. */
1876 *batch++ = MI_LOAD_REGISTER_IMM | (3 - 2);
1877 *batch++ = reg;
1878 *batch++ = expected_value;
1879
1880 /* Save the register's value back to the buffer. */
1881 *batch++ = MI_STORE_REGISTER_MEM | (3 - 2);
1882 *batch++ = reg;
1883 struct drm_i915_gem_relocation_entry reloc = {
1884 .offset = (char *) batch - (char *) map,
1885 .delta = offset * sizeof(uint32_t),
1886 .target_handle = results->gem_handle,
1887 .read_domains = I915_GEM_DOMAIN_INSTRUCTION,
1888 .write_domain = I915_GEM_DOMAIN_INSTRUCTION,
1889 };
1890 *batch++ = reloc.presumed_offset + reloc.delta;
1891
1892 /* And afterwards clear the register */
1893 if (reset) {
1894 *batch++ = MI_LOAD_REGISTER_IMM | (3 - 2);
1895 *batch++ = reg;
1896 *batch++ = 0;
1897 }
1898
1899 *batch++ = MI_BATCH_BUFFER_END;
1900
1901 struct drm_i915_gem_exec_object2 exec_objects[2] = {
1902 {
1903 .handle = results->gem_handle,
1904 },
1905 {
1906 .handle = bo->gem_handle,
1907 .relocation_count = 1,
1908 .relocs_ptr = (uintptr_t) &reloc,
1909 }
1910 };
1911
1912 struct drm_i915_gem_execbuffer2 execbuf = {
1913 .buffers_ptr = (uintptr_t) exec_objects,
1914 .buffer_count = 2,
1915 .batch_len = ALIGN((char *) batch - (char *) map, 8),
1916 .flags = I915_EXEC_RENDER,
1917 };
1918
1919 /* Don't bother with error checking - if the execbuf fails, the
1920 * value won't be written and we'll just report that there's no access.
1921 */
1922 __DRIscreen *dri_screen = screen->driScrnPriv;
1923 drmIoctl(dri_screen->fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf);
1924
1925 /* Check whether the value got written. */
1926 void *results_map = brw_bo_map(NULL, results, MAP_READ);
1927 if (results_map) {
1928 success = *((uint32_t *)results_map + offset) == expected_value;
1929 brw_bo_unmap(results);
1930 }
1931
1932 err_batch:
1933 brw_bo_unreference(bo);
1934 err_results:
1935 brw_bo_unreference(results);
1936 err:
1937 return success;
1938 }
1939
1940 static bool
1941 intel_detect_pipelined_so(struct intel_screen *screen)
1942 {
1943 const struct gen_device_info *devinfo = &screen->devinfo;
1944
1945 /* Supposedly, Broadwell just works. */
1946 if (devinfo->gen >= 8)
1947 return true;
1948
1949 if (devinfo->gen <= 6)
1950 return false;
1951
1952 /* See the big explanation about command parser versions below */
1953 if (screen->cmd_parser_version >= (devinfo->is_haswell ? 7 : 2))
1954 return true;
1955
1956 /* We use SO_WRITE_OFFSET0 since you're supposed to write it (unlike the
1957 * statistics registers), and we already reset it to zero before using it.
1958 */
1959 return intel_detect_pipelined_register(screen,
1960 GEN7_SO_WRITE_OFFSET(0),
1961 0x1337d0d0,
1962 false);
1963 }
1964
1965 /**
1966 * Return array of MSAA modes supported by the hardware. The array is
1967 * zero-terminated and sorted in decreasing order.
1968 */
1969 const int*
1970 intel_supported_msaa_modes(const struct intel_screen *screen)
1971 {
1972 static const int gen9_modes[] = {16, 8, 4, 2, 0, -1};
1973 static const int gen8_modes[] = {8, 4, 2, 0, -1};
1974 static const int gen7_modes[] = {8, 4, 0, -1};
1975 static const int gen6_modes[] = {4, 0, -1};
1976 static const int gen4_modes[] = {0, -1};
1977
1978 if (screen->devinfo.gen >= 9) {
1979 return gen9_modes;
1980 } else if (screen->devinfo.gen >= 8) {
1981 return gen8_modes;
1982 } else if (screen->devinfo.gen >= 7) {
1983 return gen7_modes;
1984 } else if (screen->devinfo.gen == 6) {
1985 return gen6_modes;
1986 } else {
1987 return gen4_modes;
1988 }
1989 }
1990
1991 static unsigned
1992 intel_loader_get_cap(const __DRIscreen *dri_screen, enum dri_loader_cap cap)
1993 {
1994 if (dri_screen->dri2.loader && dri_screen->dri2.loader->base.version >= 4 &&
1995 dri_screen->dri2.loader->getCapability)
1996 return dri_screen->dri2.loader->getCapability(dri_screen->loaderPrivate, cap);
1997
1998 if (dri_screen->image.loader && dri_screen->image.loader->base.version >= 2 &&
1999 dri_screen->image.loader->getCapability)
2000 return dri_screen->image.loader->getCapability(dri_screen->loaderPrivate, cap);
2001
2002 return 0;
2003 }
2004
2005 static __DRIconfig**
2006 intel_screen_make_configs(__DRIscreen *dri_screen)
2007 {
2008 static const mesa_format formats[] = {
2009 MESA_FORMAT_B5G6R5_UNORM,
2010 MESA_FORMAT_B8G8R8A8_UNORM,
2011 MESA_FORMAT_B8G8R8X8_UNORM,
2012
2013 MESA_FORMAT_B8G8R8A8_SRGB,
2014
2015 /* The 32-bit RGBA format must not precede the 32-bit BGRA format.
2016 * Likewise for RGBX and BGRX. Otherwise, the GLX client and the GLX
2017 * server may disagree on which format the GLXFBConfig represents,
2018 * resulting in swapped color channels.
2019 *
2020 * The problem, as of 2017-05-30:
2021 * When matching a GLXFBConfig to a __DRIconfig, GLX ignores the channel
2022 * order and chooses the first __DRIconfig with the expected channel
2023 * sizes. Specifically, GLX compares the GLXFBConfig's and __DRIconfig's
2024 * __DRI_ATTRIB_{CHANNEL}_SIZE but ignores __DRI_ATTRIB_{CHANNEL}_MASK.
2025 *
2026 * EGL does not suffer from this problem. It correctly compares the
2027 * channel masks when matching EGLConfig to __DRIconfig.
2028 */
2029
2030 /* Required by Android, for HAL_PIXEL_FORMAT_RGBA_8888. */
2031 MESA_FORMAT_R8G8B8A8_UNORM,
2032
2033 /* Required by Android, for HAL_PIXEL_FORMAT_RGBX_8888. */
2034 MESA_FORMAT_R8G8B8X8_UNORM,
2035 };
2036
2037 /* GLX_SWAP_COPY_OML is not supported due to page flipping. */
2038 static const GLenum back_buffer_modes[] = {
2039 __DRI_ATTRIB_SWAP_UNDEFINED, __DRI_ATTRIB_SWAP_NONE
2040 };
2041
2042 static const uint8_t singlesample_samples[1] = {0};
2043
2044 struct intel_screen *screen = dri_screen->driverPrivate;
2045 const struct gen_device_info *devinfo = &screen->devinfo;
2046 uint8_t depth_bits[4], stencil_bits[4];
2047 __DRIconfig **configs = NULL;
2048
2049 /* Expose only BGRA ordering if the loader doesn't support RGBA ordering. */
2050 unsigned num_formats;
2051 if (intel_loader_get_cap(dri_screen, DRI_LOADER_CAP_RGBA_ORDERING))
2052 num_formats = ARRAY_SIZE(formats);
2053 else
2054 num_formats = ARRAY_SIZE(formats) - 2; /* all - RGBA_ORDERING formats */
2055
2056 /* Generate singlesample configs without accumulation buffer. */
2057 for (unsigned i = 0; i < num_formats; i++) {
2058 __DRIconfig **new_configs;
2059 int num_depth_stencil_bits = 2;
2060
2061 /* Starting with DRI2 protocol version 1.1 we can request a depth/stencil
2062 * buffer that has a different number of bits per pixel than the color
2063 * buffer, gen >= 6 supports this.
2064 */
2065 depth_bits[0] = 0;
2066 stencil_bits[0] = 0;
2067
2068 if (formats[i] == MESA_FORMAT_B5G6R5_UNORM) {
2069 depth_bits[1] = 16;
2070 stencil_bits[1] = 0;
2071 if (devinfo->gen >= 6) {
2072 depth_bits[2] = 24;
2073 stencil_bits[2] = 8;
2074 num_depth_stencil_bits = 3;
2075 }
2076 } else {
2077 depth_bits[1] = 24;
2078 stencil_bits[1] = 8;
2079 }
2080
2081 new_configs = driCreateConfigs(formats[i],
2082 depth_bits,
2083 stencil_bits,
2084 num_depth_stencil_bits,
2085 back_buffer_modes, 2,
2086 singlesample_samples, 1,
2087 false, false);
2088 configs = driConcatConfigs(configs, new_configs);
2089 }
2090
2091 /* Generate the minimum possible set of configs that include an
2092 * accumulation buffer.
2093 */
2094 for (unsigned i = 0; i < num_formats; i++) {
2095 __DRIconfig **new_configs;
2096
2097 if (formats[i] == MESA_FORMAT_B5G6R5_UNORM) {
2098 depth_bits[0] = 16;
2099 stencil_bits[0] = 0;
2100 } else {
2101 depth_bits[0] = 24;
2102 stencil_bits[0] = 8;
2103 }
2104
2105 new_configs = driCreateConfigs(formats[i],
2106 depth_bits, stencil_bits, 1,
2107 back_buffer_modes, 1,
2108 singlesample_samples, 1,
2109 true, false);
2110 configs = driConcatConfigs(configs, new_configs);
2111 }
2112
2113 /* Generate multisample configs.
2114 *
2115 * This loop breaks early, and hence is a no-op, on gen < 6.
2116 *
2117 * Multisample configs must follow the singlesample configs in order to
2118 * work around an X server bug present in 1.12. The X server chooses to
2119 * associate the first listed RGBA888-Z24S8 config, regardless of its
2120 * sample count, with the 32-bit depth visual used for compositing.
2121 *
2122 * Only doublebuffer configs with GLX_SWAP_UNDEFINED_OML behavior are
2123 * supported. Singlebuffer configs are not supported because no one wants
2124 * them.
2125 */
2126 for (unsigned i = 0; i < num_formats; i++) {
2127 if (devinfo->gen < 6)
2128 break;
2129
2130 __DRIconfig **new_configs;
2131 const int num_depth_stencil_bits = 2;
2132 int num_msaa_modes = 0;
2133 const uint8_t *multisample_samples = NULL;
2134
2135 depth_bits[0] = 0;
2136 stencil_bits[0] = 0;
2137
2138 if (formats[i] == MESA_FORMAT_B5G6R5_UNORM) {
2139 depth_bits[1] = 16;
2140 stencil_bits[1] = 0;
2141 } else {
2142 depth_bits[1] = 24;
2143 stencil_bits[1] = 8;
2144 }
2145
2146 if (devinfo->gen >= 9) {
2147 static const uint8_t multisample_samples_gen9[] = {2, 4, 8, 16};
2148 multisample_samples = multisample_samples_gen9;
2149 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen9);
2150 } else if (devinfo->gen == 8) {
2151 static const uint8_t multisample_samples_gen8[] = {2, 4, 8};
2152 multisample_samples = multisample_samples_gen8;
2153 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen8);
2154 } else if (devinfo->gen == 7) {
2155 static const uint8_t multisample_samples_gen7[] = {4, 8};
2156 multisample_samples = multisample_samples_gen7;
2157 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen7);
2158 } else if (devinfo->gen == 6) {
2159 static const uint8_t multisample_samples_gen6[] = {4};
2160 multisample_samples = multisample_samples_gen6;
2161 num_msaa_modes = ARRAY_SIZE(multisample_samples_gen6);
2162 }
2163
2164 new_configs = driCreateConfigs(formats[i],
2165 depth_bits,
2166 stencil_bits,
2167 num_depth_stencil_bits,
2168 back_buffer_modes, 1,
2169 multisample_samples,
2170 num_msaa_modes,
2171 false, false);
2172 configs = driConcatConfigs(configs, new_configs);
2173 }
2174
2175 if (configs == NULL) {
2176 fprintf(stderr, "[%s:%u] Error creating FBConfig!\n", __func__,
2177 __LINE__);
2178 return NULL;
2179 }
2180
2181 return configs;
2182 }
2183
2184 static void
2185 set_max_gl_versions(struct intel_screen *screen)
2186 {
2187 __DRIscreen *dri_screen = screen->driScrnPriv;
2188 const bool has_astc = screen->devinfo.gen >= 9;
2189
2190 switch (screen->devinfo.gen) {
2191 case 10:
2192 case 9:
2193 case 8:
2194 dri_screen->max_gl_core_version = 45;
2195 dri_screen->max_gl_compat_version = 30;
2196 dri_screen->max_gl_es1_version = 11;
2197 dri_screen->max_gl_es2_version = has_astc ? 32 : 31;
2198 break;
2199 case 7:
2200 dri_screen->max_gl_core_version = 33;
2201 if (can_do_pipelined_register_writes(screen)) {
2202 dri_screen->max_gl_core_version = 42;
2203 if (screen->devinfo.is_haswell && can_do_compute_dispatch(screen))
2204 dri_screen->max_gl_core_version = 43;
2205 if (screen->devinfo.is_haswell && can_do_mi_math_and_lrr(screen))
2206 dri_screen->max_gl_core_version = 45;
2207 }
2208 dri_screen->max_gl_compat_version = 30;
2209 dri_screen->max_gl_es1_version = 11;
2210 dri_screen->max_gl_es2_version = screen->devinfo.is_haswell ? 31 : 30;
2211 break;
2212 case 6:
2213 dri_screen->max_gl_core_version = 33;
2214 dri_screen->max_gl_compat_version = 30;
2215 dri_screen->max_gl_es1_version = 11;
2216 dri_screen->max_gl_es2_version = 30;
2217 break;
2218 case 5:
2219 case 4:
2220 dri_screen->max_gl_core_version = 0;
2221 dri_screen->max_gl_compat_version = 21;
2222 dri_screen->max_gl_es1_version = 11;
2223 dri_screen->max_gl_es2_version = 20;
2224 break;
2225 default:
2226 unreachable("unrecognized intel_screen::gen");
2227 }
2228 }
2229
2230 /**
2231 * Return the revision (generally the revid field of the PCI header) of the
2232 * graphics device.
2233 */
2234 int
2235 intel_device_get_revision(int fd)
2236 {
2237 struct drm_i915_getparam gp;
2238 int revision;
2239 int ret;
2240
2241 memset(&gp, 0, sizeof(gp));
2242 gp.param = I915_PARAM_REVISION;
2243 gp.value = &revision;
2244
2245 ret = drmCommandWriteRead(fd, DRM_I915_GETPARAM, &gp, sizeof(gp));
2246 if (ret)
2247 revision = -1;
2248
2249 return revision;
2250 }
2251
2252 static void
2253 shader_debug_log_mesa(void *data, const char *fmt, ...)
2254 {
2255 struct brw_context *brw = (struct brw_context *)data;
2256 va_list args;
2257
2258 va_start(args, fmt);
2259 GLuint msg_id = 0;
2260 _mesa_gl_vdebug(&brw->ctx, &msg_id,
2261 MESA_DEBUG_SOURCE_SHADER_COMPILER,
2262 MESA_DEBUG_TYPE_OTHER,
2263 MESA_DEBUG_SEVERITY_NOTIFICATION, fmt, args);
2264 va_end(args);
2265 }
2266
2267 static void
2268 shader_perf_log_mesa(void *data, const char *fmt, ...)
2269 {
2270 struct brw_context *brw = (struct brw_context *)data;
2271
2272 va_list args;
2273 va_start(args, fmt);
2274
2275 if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
2276 va_list args_copy;
2277 va_copy(args_copy, args);
2278 vfprintf(stderr, fmt, args_copy);
2279 va_end(args_copy);
2280 }
2281
2282 if (brw->perf_debug) {
2283 GLuint msg_id = 0;
2284 _mesa_gl_vdebug(&brw->ctx, &msg_id,
2285 MESA_DEBUG_SOURCE_SHADER_COMPILER,
2286 MESA_DEBUG_TYPE_PERFORMANCE,
2287 MESA_DEBUG_SEVERITY_MEDIUM, fmt, args);
2288 }
2289 va_end(args);
2290 }
2291
2292 static int
2293 parse_devid_override(const char *devid_override)
2294 {
2295 static const struct {
2296 const char *name;
2297 int pci_id;
2298 } name_map[] = {
2299 { "brw", 0x2a02 },
2300 { "g4x", 0x2a42 },
2301 { "ilk", 0x0042 },
2302 { "snb", 0x0126 },
2303 { "ivb", 0x016a },
2304 { "hsw", 0x0d2e },
2305 { "byt", 0x0f33 },
2306 { "bdw", 0x162e },
2307 { "chv", 0x22B3 },
2308 { "skl", 0x1912 },
2309 { "bxt", 0x5A85 },
2310 { "kbl", 0x5912 },
2311 { "glk", 0x3185 },
2312 { "cnl", 0x5a52 },
2313 };
2314
2315 for (unsigned i = 0; i < ARRAY_SIZE(name_map); i++) {
2316 if (!strcmp(name_map[i].name, devid_override))
2317 return name_map[i].pci_id;
2318 }
2319
2320 return strtol(devid_override, NULL, 0);
2321 }
2322
2323 /**
2324 * Get the PCI ID for the device. This can be overridden by setting the
2325 * INTEL_DEVID_OVERRIDE environment variable to the desired ID.
2326 *
2327 * Returns -1 on ioctl failure.
2328 */
2329 static int
2330 get_pci_device_id(struct intel_screen *screen)
2331 {
2332 if (geteuid() == getuid()) {
2333 char *devid_override = getenv("INTEL_DEVID_OVERRIDE");
2334 if (devid_override) {
2335 screen->no_hw = true;
2336 return parse_devid_override(devid_override);
2337 }
2338 }
2339
2340 return intel_get_integer(screen, I915_PARAM_CHIPSET_ID);
2341 }
2342
2343 /**
2344 * This is the driver specific part of the createNewScreen entry point.
2345 * Called when using DRI2.
2346 *
2347 * \return the struct gl_config supported by this driver
2348 */
2349 static const
2350 __DRIconfig **intelInitScreen2(__DRIscreen *dri_screen)
2351 {
2352 struct intel_screen *screen;
2353
2354 if (dri_screen->image.loader) {
2355 } else if (dri_screen->dri2.loader->base.version <= 2 ||
2356 dri_screen->dri2.loader->getBuffersWithFormat == NULL) {
2357 fprintf(stderr,
2358 "\nERROR! DRI2 loader with getBuffersWithFormat() "
2359 "support required\n");
2360 return NULL;
2361 }
2362
2363 /* Allocate the private area */
2364 screen = rzalloc(NULL, struct intel_screen);
2365 if (!screen) {
2366 fprintf(stderr, "\nERROR! Allocating private area failed\n");
2367 return NULL;
2368 }
2369 /* parse information in __driConfigOptions */
2370 driParseOptionInfo(&screen->optionCache, brw_config_options.xml);
2371
2372 screen->driScrnPriv = dri_screen;
2373 dri_screen->driverPrivate = (void *) screen;
2374
2375 screen->deviceID = get_pci_device_id(screen);
2376
2377 if (!gen_get_device_info(screen->deviceID, &screen->devinfo))
2378 return NULL;
2379
2380 if (!intel_init_bufmgr(screen))
2381 return NULL;
2382
2383 const struct gen_device_info *devinfo = &screen->devinfo;
2384
2385 brw_process_intel_debug_variable();
2386
2387 if ((INTEL_DEBUG & DEBUG_SHADER_TIME) && devinfo->gen < 7) {
2388 fprintf(stderr,
2389 "shader_time debugging requires gen7 (Ivybridge) or better.\n");
2390 INTEL_DEBUG &= ~DEBUG_SHADER_TIME;
2391 }
2392
2393 if (intel_get_integer(screen, I915_PARAM_MMAP_GTT_VERSION) >= 1) {
2394 /* Theorectically unlimited! At least for individual objects...
2395 *
2396 * Currently the entire (global) address space for all GTT maps is
2397 * limited to 64bits. That is all objects on the system that are
2398 * setup for GTT mmapping must fit within 64bits. An attempt to use
2399 * one that exceeds the limit with fail in brw_bo_map_gtt().
2400 *
2401 * Long before we hit that limit, we will be practically limited by
2402 * that any single object must fit in physical memory (RAM). The upper
2403 * limit on the CPU's address space is currently 48bits (Skylake), of
2404 * which only 39bits can be physical memory. (The GPU itself also has
2405 * a 48bit addressable virtual space.) We can fit over 32 million
2406 * objects of the current maximum allocable size before running out
2407 * of mmap space.
2408 */
2409 screen->max_gtt_map_object_size = UINT64_MAX;
2410 } else {
2411 /* Estimate the size of the mappable aperture into the GTT. There's an
2412 * ioctl to get the whole GTT size, but not one to get the mappable subset.
2413 * It turns out it's basically always 256MB, though some ancient hardware
2414 * was smaller.
2415 */
2416 uint32_t gtt_size = 256 * 1024 * 1024;
2417
2418 /* We don't want to map two objects such that a memcpy between them would
2419 * just fault one mapping in and then the other over and over forever. So
2420 * we would need to divide the GTT size by 2. Additionally, some GTT is
2421 * taken up by things like the framebuffer and the ringbuffer and such, so
2422 * be more conservative.
2423 */
2424 screen->max_gtt_map_object_size = gtt_size / 4;
2425 }
2426
2427 screen->aperture_threshold = get_aperture_size(dri_screen->fd) * 3 / 4;
2428
2429 screen->hw_has_swizzling = intel_detect_swizzling(screen);
2430 screen->hw_has_timestamp = intel_detect_timestamp(screen);
2431
2432 isl_device_init(&screen->isl_dev, &screen->devinfo,
2433 screen->hw_has_swizzling);
2434
2435 if (devinfo->gen >= 10)
2436 intel_cs_timestamp_frequency(screen);
2437
2438 /* GENs prior to 8 do not support EU/Subslice info */
2439 if (devinfo->gen >= 8) {
2440 intel_detect_sseu(screen);
2441 } else if (devinfo->gen == 7) {
2442 screen->subslice_total = 1 << (devinfo->gt - 1);
2443 }
2444
2445 /* Gen7-7.5 kernel requirements / command parser saga:
2446 *
2447 * - pre-v3.16:
2448 * Haswell and Baytrail cannot use any privileged batchbuffer features.
2449 *
2450 * Ivybridge has aliasing PPGTT on by default, which accidentally marks
2451 * all batches secure, allowing them to use any feature with no checking.
2452 * This is effectively equivalent to a command parser version of
2453 * \infinity - everything is possible.
2454 *
2455 * The command parser does not exist, and querying the version will
2456 * return -EINVAL.
2457 *
2458 * - v3.16:
2459 * The kernel enables the command parser by default, for systems with
2460 * aliasing PPGTT enabled (Ivybridge and Haswell). However, the
2461 * hardware checker is still enabled, so Haswell and Baytrail cannot
2462 * do anything.
2463 *
2464 * Ivybridge goes from "everything is possible" to "only what the
2465 * command parser allows" (if the user boots with i915.cmd_parser=0,
2466 * then everything is possible again). We can only safely use features
2467 * allowed by the supported command parser version.
2468 *
2469 * Annoyingly, I915_PARAM_CMD_PARSER_VERSION reports the static version
2470 * implemented by the kernel, even if it's turned off. So, checking
2471 * for version > 0 does not mean that you can write registers. We have
2472 * to try it and see. The version does, however, indicate the age of
2473 * the kernel.
2474 *
2475 * Instead of matching the hardware checker's behavior of converting
2476 * privileged commands to MI_NOOP, it makes execbuf2 start returning
2477 * -EINVAL, making it dangerous to try and use privileged features.
2478 *
2479 * Effective command parser versions:
2480 * - Haswell: 0 (reporting 1, writes don't work)
2481 * - Baytrail: 0 (reporting 1, writes don't work)
2482 * - Ivybridge: 1 (enabled) or infinite (disabled)
2483 *
2484 * - v3.17:
2485 * Baytrail aliasing PPGTT is enabled, making it like Ivybridge:
2486 * effectively version 1 (enabled) or infinite (disabled).
2487 *
2488 * - v3.19: f1f55cc0556031c8ee3fe99dae7251e78b9b653b
2489 * Command parser v2 supports predicate writes.
2490 *
2491 * - Haswell: 0 (reporting 1, writes don't work)
2492 * - Baytrail: 2 (enabled) or infinite (disabled)
2493 * - Ivybridge: 2 (enabled) or infinite (disabled)
2494 *
2495 * So version >= 2 is enough to know that Ivybridge and Baytrail
2496 * will work. Haswell still can't do anything.
2497 *
2498 * - v4.0: Version 3 happened. Largely not relevant.
2499 *
2500 * - v4.1: 6702cf16e0ba8b0129f5aa1b6609d4e9c70bc13b
2501 * L3 config registers are properly saved and restored as part
2502 * of the hardware context. We can approximately detect this point
2503 * in time by checking if I915_PARAM_REVISION is recognized - it
2504 * landed in a later commit, but in the same release cycle.
2505 *
2506 * - v4.2: 245054a1fe33c06ad233e0d58a27ec7b64db9284
2507 * Command parser finally gains secure batch promotion. On Haswell,
2508 * the hardware checker gets disabled, which finally allows it to do
2509 * privileged commands.
2510 *
2511 * I915_PARAM_CMD_PARSER_VERSION reports 3. Effective versions:
2512 * - Haswell: 3 (enabled) or 0 (disabled)
2513 * - Baytrail: 3 (enabled) or infinite (disabled)
2514 * - Ivybridge: 3 (enabled) or infinite (disabled)
2515 *
2516 * Unfortunately, detecting this point in time is tricky, because
2517 * no version bump happened when this important change occurred.
2518 * On Haswell, if we can write any register, then the kernel is at
2519 * least this new, and we can start trusting the version number.
2520 *
2521 * - v4.4: 2bbe6bbb0dc94fd4ce287bdac9e1bd184e23057b and
2522 * Command parser reaches version 4, allowing access to Haswell
2523 * atomic scratch and chicken3 registers. If version >= 4, we know
2524 * the kernel is new enough to support privileged features on all
2525 * hardware. However, the user might have disabled it...and the
2526 * kernel will still report version 4. So we still have to guess
2527 * and check.
2528 *
2529 * - v4.4: 7b9748cb513a6bef4af87b79f0da3ff7e8b56cd8
2530 * Command parser v5 whitelists indirect compute shader dispatch
2531 * registers, needed for OpenGL 4.3 and later.
2532 *
2533 * - v4.8:
2534 * Command parser v7 lets us use MI_MATH on Haswell.
2535 *
2536 * Additionally, the kernel begins reporting version 0 when
2537 * the command parser is disabled, allowing us to skip the
2538 * guess-and-check step on Haswell. Unfortunately, this also
2539 * means that we can no longer use it as an indicator of the
2540 * age of the kernel.
2541 */
2542 if (intel_get_param(screen, I915_PARAM_CMD_PARSER_VERSION,
2543 &screen->cmd_parser_version) < 0) {
2544 /* Command parser does not exist - getparam is unrecognized */
2545 screen->cmd_parser_version = 0;
2546 }
2547
2548 /* Kernel 4.13 retuired for exec object capture */
2549 if (intel_get_boolean(screen, I915_PARAM_HAS_EXEC_CAPTURE)) {
2550 screen->kernel_features |= KERNEL_ALLOWS_EXEC_CAPTURE;
2551 }
2552
2553 if (intel_get_boolean(screen, I915_PARAM_HAS_EXEC_BATCH_FIRST)) {
2554 screen->kernel_features |= KERNEL_ALLOWS_EXEC_BATCH_FIRST;
2555 }
2556
2557 if (!intel_detect_pipelined_so(screen)) {
2558 /* We can't do anything, so the effective version is 0. */
2559 screen->cmd_parser_version = 0;
2560 } else {
2561 screen->kernel_features |= KERNEL_ALLOWS_SOL_OFFSET_WRITES;
2562 }
2563
2564 if (devinfo->gen >= 8 || screen->cmd_parser_version >= 2)
2565 screen->kernel_features |= KERNEL_ALLOWS_PREDICATE_WRITES;
2566
2567 /* Haswell requires command parser version 4 in order to have L3
2568 * atomic scratch1 and chicken3 bits
2569 */
2570 if (devinfo->is_haswell && screen->cmd_parser_version >= 4) {
2571 screen->kernel_features |=
2572 KERNEL_ALLOWS_HSW_SCRATCH1_AND_ROW_CHICKEN3;
2573 }
2574
2575 /* Haswell requires command parser version 6 in order to write to the
2576 * MI_MATH GPR registers, and version 7 in order to use
2577 * MI_LOAD_REGISTER_REG (which all users of MI_MATH use).
2578 */
2579 if (devinfo->gen >= 8 ||
2580 (devinfo->is_haswell && screen->cmd_parser_version >= 7)) {
2581 screen->kernel_features |= KERNEL_ALLOWS_MI_MATH_AND_LRR;
2582 }
2583
2584 /* Gen7 needs at least command parser version 5 to support compute */
2585 if (devinfo->gen >= 8 || screen->cmd_parser_version >= 5)
2586 screen->kernel_features |= KERNEL_ALLOWS_COMPUTE_DISPATCH;
2587
2588 const char *force_msaa = getenv("INTEL_FORCE_MSAA");
2589 if (force_msaa) {
2590 screen->winsys_msaa_samples_override =
2591 intel_quantize_num_samples(screen, atoi(force_msaa));
2592 printf("Forcing winsys sample count to %d\n",
2593 screen->winsys_msaa_samples_override);
2594 } else {
2595 screen->winsys_msaa_samples_override = -1;
2596 }
2597
2598 set_max_gl_versions(screen);
2599
2600 /* Notification of GPU resets requires hardware contexts and a kernel new
2601 * enough to support DRM_IOCTL_I915_GET_RESET_STATS. If the ioctl is
2602 * supported, calling it with a context of 0 will either generate EPERM or
2603 * no error. If the ioctl is not supported, it always generate EINVAL.
2604 * Use this to determine whether to advertise the __DRI2_ROBUSTNESS
2605 * extension to the loader.
2606 *
2607 * Don't even try on pre-Gen6, since we don't attempt to use contexts there.
2608 */
2609 if (devinfo->gen >= 6) {
2610 struct drm_i915_reset_stats stats;
2611 memset(&stats, 0, sizeof(stats));
2612
2613 const int ret = drmIoctl(dri_screen->fd, DRM_IOCTL_I915_GET_RESET_STATS, &stats);
2614
2615 screen->has_context_reset_notification =
2616 (ret != -1 || errno != EINVAL);
2617 }
2618
2619 dri_screen->extensions = !screen->has_context_reset_notification
2620 ? screenExtensions : intelRobustScreenExtensions;
2621
2622 screen->compiler = brw_compiler_create(screen, devinfo);
2623 screen->compiler->shader_debug_log = shader_debug_log_mesa;
2624 screen->compiler->shader_perf_log = shader_perf_log_mesa;
2625 screen->compiler->constant_buffer_0_is_relative = true;
2626 screen->compiler->supports_pull_constants = true;
2627
2628 screen->has_exec_fence =
2629 intel_get_boolean(screen, I915_PARAM_HAS_EXEC_FENCE);
2630
2631 intel_screen_init_surface_formats(screen);
2632
2633 if (INTEL_DEBUG & (DEBUG_BATCH | DEBUG_SUBMIT)) {
2634 unsigned int caps = intel_get_integer(screen, I915_PARAM_HAS_SCHEDULER);
2635 if (caps) {
2636 fprintf(stderr, "Kernel scheduler detected: %08x\n", caps);
2637 if (caps & I915_SCHEDULER_CAP_PRIORITY)
2638 fprintf(stderr, " - User priority sorting enabled\n");
2639 if (caps & I915_SCHEDULER_CAP_PREEMPTION)
2640 fprintf(stderr, " - Preemption enabled\n");
2641 }
2642 }
2643
2644 return (const __DRIconfig**) intel_screen_make_configs(dri_screen);
2645 }
2646
2647 struct intel_buffer {
2648 __DRIbuffer base;
2649 struct brw_bo *bo;
2650 };
2651
2652 static __DRIbuffer *
2653 intelAllocateBuffer(__DRIscreen *dri_screen,
2654 unsigned attachment, unsigned format,
2655 int width, int height)
2656 {
2657 struct intel_buffer *intelBuffer;
2658 struct intel_screen *screen = dri_screen->driverPrivate;
2659
2660 assert(attachment == __DRI_BUFFER_FRONT_LEFT ||
2661 attachment == __DRI_BUFFER_BACK_LEFT);
2662
2663 intelBuffer = calloc(1, sizeof *intelBuffer);
2664 if (intelBuffer == NULL)
2665 return NULL;
2666
2667 /* The front and back buffers are color buffers, which are X tiled. GEN9+
2668 * supports Y tiled and compressed buffers, but there is no way to plumb that
2669 * through to here. */
2670 uint32_t pitch;
2671 int cpp = format / 8;
2672 intelBuffer->bo = brw_bo_alloc_tiled_2d(screen->bufmgr,
2673 "intelAllocateBuffer",
2674 width,
2675 height,
2676 cpp,
2677 I915_TILING_X, &pitch,
2678 BO_ALLOC_BUSY);
2679
2680 if (intelBuffer->bo == NULL) {
2681 free(intelBuffer);
2682 return NULL;
2683 }
2684
2685 brw_bo_flink(intelBuffer->bo, &intelBuffer->base.name);
2686
2687 intelBuffer->base.attachment = attachment;
2688 intelBuffer->base.cpp = cpp;
2689 intelBuffer->base.pitch = pitch;
2690
2691 return &intelBuffer->base;
2692 }
2693
2694 static void
2695 intelReleaseBuffer(__DRIscreen *dri_screen, __DRIbuffer *buffer)
2696 {
2697 struct intel_buffer *intelBuffer = (struct intel_buffer *) buffer;
2698
2699 brw_bo_unreference(intelBuffer->bo);
2700 free(intelBuffer);
2701 }
2702
2703 static const struct __DriverAPIRec brw_driver_api = {
2704 .InitScreen = intelInitScreen2,
2705 .DestroyScreen = intelDestroyScreen,
2706 .CreateContext = brwCreateContext,
2707 .DestroyContext = intelDestroyContext,
2708 .CreateBuffer = intelCreateBuffer,
2709 .DestroyBuffer = intelDestroyBuffer,
2710 .MakeCurrent = intelMakeCurrent,
2711 .UnbindContext = intelUnbindContext,
2712 .AllocateBuffer = intelAllocateBuffer,
2713 .ReleaseBuffer = intelReleaseBuffer
2714 };
2715
2716 static const struct __DRIDriverVtableExtensionRec brw_vtable = {
2717 .base = { __DRI_DRIVER_VTABLE, 1 },
2718 .vtable = &brw_driver_api,
2719 };
2720
2721 static const __DRIextension *brw_driver_extensions[] = {
2722 &driCoreExtension.base,
2723 &driImageDriverExtension.base,
2724 &driDRI2Extension.base,
2725 &brw_vtable.base,
2726 &brw_config_options.base,
2727 NULL
2728 };
2729
2730 PUBLIC const __DRIextension **__driDriverGetExtensions_i965(void)
2731 {
2732 globalDriverAPI = &brw_driver_api;
2733
2734 return brw_driver_extensions;
2735 }