This adds the UAPI for the Asahi driver targeting the GPU in the Apple M1 and M2 series systems on chip. The UAPI design is based on other modern Vulkan-capable drivers, including Xe and Panthor. Memory management is based on explicit VM management. Synchronization is exclusively explicit sync. This UAPI is validated against our open source Mesa stack, which is fully conformant to the OpenGL 4.6, OpenGL ES 3.2, OpenCL 3.0, and Vulkan 1.4 standards. The Vulkan driver supports sparse, exercising the VM_BIND mechanism. This patch adds the standalone UAPI header. It is implemented by an open source DRM driver written in Rust. We fully intend to upstream this driver when possible. However, as a production graphics driver, it depends on a significant number of Rust abstractions that will take a long time to upstream. In the mean time, our userspace is upstream in Mesa but is not allowed to probe with upstream Mesa as the UAPI is not yet reviewed and merged in the upstream kernel. Although we ship a patched Mesa in Fedora Asahi Remix, any containers shipping upstream Mesa builds are broken for our users, including upstream Flatpak and Waydroid runtimes. Additionally, it forces us to maintain forks of Mesa and virglrenderer, which complicates bisects. The intention in sending out this patch is for this UAPI to be thoroughly reviewed. Once we as the DRM community are satisfied with the UAPI, this header lands signifying that the UAPI is stable and must only be evolved in backwards-compatible ways; it will be the UAPI implemented in the DRM driver that eventually lands upstream. That promise lets us enable upstream Mesa, solving all these issues while the upstream Rust abstractions are developed. https://github.com/alyssarosenzweig/linux/commits/agx-uapi-v7 contains the DRM driver implementing this proposed UAPI. https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33984 contains the Mesa patches to implement this proposed UAPI. That Linux and Mesa branch together give a complete graphics/compute stack on top of this UAPI.
Showing the changes to include/uapi (1 of 3 files). Show all
Unified · Split
| @@ -0,0 +1,1194 @@ | |||
| 1 | /* SPDX-License-Identifier: MIT */ | ||
| 2 | /* | ||
| 3 | * Copyright (C) The Asahi Linux Contributors | ||
| 4 | * Copyright (C) 2018-2023 Collabora Ltd. | ||
| 5 | * Copyright (C) 2014-2018 Broadcom | ||
| 6 | */ | ||
| 7 | #ifndef _ASAHI_DRM_H_ | ||
| 8 | #define _ASAHI_DRM_H_ | ||
| 9 | |||
| 10 | #include "drm.h" | ||
| 11 | |||
| 12 | #if defined(__cplusplus) | ||
| 13 | extern "C" { | ||
| 14 | #endif | ||
| 15 | |||
| 16 | /** | ||
| 17 | * DOC: Introduction to the Asahi UAPI | ||
| 18 | * | ||
| 19 | * This documentation describes the Asahi IOCTLs. | ||
| 20 | * | ||
| 21 | * Just a few generic rules about the data passed to the Asahi IOCTLs (cribbed | ||
| 22 | * from Panthor): | ||
| 23 | * | ||
| 24 | * - Structures must be aligned on 64-bit/8-byte. If the object is not | ||
| 25 | * naturally aligned, a padding field must be added. | ||
| 26 | * - Fields must be explicitly aligned to their natural type alignment with | ||
| 27 | * pad[0..N] fields. | ||
| 28 | * - All padding fields will be checked by the driver to make sure they are | ||
| 29 | * zeroed. | ||
| 30 | * - Flags can be added, but not removed/replaced. | ||
| 31 | * - New fields can be added to the main structures (the structures | ||
| 32 | * directly passed to the ioctl). Those fields can be added at the end of | ||
| 33 | * the structure, or replace existing padding fields. Any new field being | ||
| 34 | * added must preserve the behavior that existed before those fields were | ||
| 35 | * added when a value of zero is passed. | ||
| 36 | * - New fields can be added to indirect objects (objects pointed by the | ||
| 37 | * main structure), iff those objects are passed a size to reflect the | ||
| 38 | * size known by the userspace driver (see | ||
| 39 | * drm_asahi_cmd_header::size). | ||
| 40 | * - If the kernel driver is too old to know some fields, those will be | ||
| 41 | * ignored if zero, and otherwise rejected (and so will be zero on output). | ||
| 42 | * - If userspace is too old to know some fields, those will be zeroed | ||
| 43 | * (input) before the structure is parsed by the kernel driver. | ||
| 44 | * - Each new flag/field addition must come with a driver version update so | ||
| 45 | * the userspace driver doesn't have to guess which flags are supported. | ||
| 46 | * - Structures should not contain unions, as this would defeat the | ||
| 47 | * extensibility of such structures. | ||
| 48 | * - IOCTLs can't be removed or replaced. New IOCTL IDs should be placed | ||
| 49 | * at the end of the drm_asahi_ioctl_id enum. | ||
| 50 | */ | ||
| 51 | |||
| 52 | /** | ||
| 53 | * enum drm_asahi_ioctl_id - IOCTL IDs | ||
| 54 | * | ||
| 55 | * Place new ioctls at the end, don't re-order, don't replace or remove entries. | ||
| 56 | * | ||
| 57 | * These IDs are not meant to be used directly. Use the DRM_IOCTL_ASAHI_xxx | ||
| 58 | * definitions instead. | ||
| 59 | */ | ||
| 60 | enum drm_asahi_ioctl_id { | ||
| 61 | /** @DRM_ASAHI_GET_PARAMS: Query device properties. */ | ||
| 62 | DRM_ASAHI_GET_PARAMS = 0, | ||
| 63 | |||
| 64 | /** @DRM_ASAHI_GET_TIME: Query device time. */ | ||
| 65 | DRM_ASAHI_GET_TIME, | ||
| 66 | |||
| 67 | /** @DRM_ASAHI_VM_CREATE: Create a GPU VM address space. */ | ||
| 68 | DRM_ASAHI_VM_CREATE, | ||
| 69 | |||
| 70 | /** @DRM_ASAHI_VM_DESTROY: Destroy a VM. */ | ||
| 71 | DRM_ASAHI_VM_DESTROY, | ||
| 72 | |||
| 73 | /** @DRM_ASAHI_VM_BIND: Bind/unbind memory to a VM. */ | ||
| 74 | DRM_ASAHI_VM_BIND, | ||
| 75 | |||
| 76 | /** @DRM_ASAHI_GEM_CREATE: Create a buffer object. */ | ||
| 77 | DRM_ASAHI_GEM_CREATE, | ||
| 78 | |||
| 79 | /** | ||
| 80 | * @DRM_ASAHI_GEM_MMAP_OFFSET: Get offset to pass to mmap() to map a | ||
| 81 | * given GEM handle. | ||
| 82 | */ | ||
| 83 | DRM_ASAHI_GEM_MMAP_OFFSET, | ||
| 84 | |||
| 85 | /** @DRM_ASAHI_GEM_BIND_OBJECT: Bind memory as a special object */ | ||
| 86 | DRM_ASAHI_GEM_BIND_OBJECT, | ||
| 87 | |||
| 88 | /** @DRM_ASAHI_QUEUE_CREATE: Create a scheduling queue. */ | ||
| 89 | DRM_ASAHI_QUEUE_CREATE, | ||
| 90 | |||
| 91 | /** @DRM_ASAHI_QUEUE_DESTROY: Destroy a scheduling queue. */ | ||
| 92 | DRM_ASAHI_QUEUE_DESTROY, | ||
| 93 | |||
| 94 | /** @DRM_ASAHI_SUBMIT: Submit commands to a queue. */ | ||
| 95 | DRM_ASAHI_SUBMIT, | ||
| 96 | }; | ||
| 97 | |||
| 98 | #define DRM_ASAHI_MAX_CLUSTERS 64 | ||
| 99 | |||
| 100 | /** | ||
| 101 | * struct drm_asahi_params_global - Global parameters. | ||
| 102 | * | ||
| 103 | * This struct may be queried by drm_asahi_get_params. | ||
| 104 | */ | ||
| 105 | struct drm_asahi_params_global { | ||
| 106 | /** @features: Feature bits from drm_asahi_feature */ | ||
| 107 | __u64 features; | ||
| 108 | |||
| 109 | /** @gpu_generation: GPU generation, e.g. 13 for G13G */ | ||
| 110 | __u32 gpu_generation; | ||
| 111 | |||
| 112 | /** @gpu_variant: GPU variant as a character, e.g. 'C' for G13C */ | ||
| 113 | __u32 gpu_variant; | ||
| 114 | |||
| 115 | /** | ||
| 116 | * @gpu_revision: GPU revision in BCD, e.g. 0x00 for 'A0' or | ||
| 117 | * 0x21 for 'C1' | ||
| 118 | */ | ||
| 119 | __u32 gpu_revision; | ||
| 120 | |||
| 121 | /** @chip_id: Chip ID in BCD, e.g. 0x8103 for T8103 */ | ||
| 122 | __u32 chip_id; | ||
| 123 | |||
| 124 | /** @num_dies: Number of dies in the SoC */ | ||
| 125 | __u32 num_dies; | ||
| 126 | |||
| 127 | /** @num_clusters_total: Number of GPU clusters (across all dies) */ | ||
| 128 | __u32 num_clusters_total; | ||
| 129 | |||
| 130 | /** | ||
| 131 | * @num_cores_per_cluster: Number of logical cores per cluster | ||
| 132 | * (including inactive/nonexistent) | ||
| 133 | */ | ||
| 134 | __u32 num_cores_per_cluster; | ||
| 135 | |||
| 136 | /** @max_frequency_khz: Maximum GPU core clock frequency */ | ||
| 137 | __u32 max_frequency_khz; | ||
| 138 | |||
| 139 | /** @core_masks: Bitmask of present/enabled cores per cluster */ | ||
| 140 | __u64 core_masks[DRM_ASAHI_MAX_CLUSTERS]; | ||
| 141 | |||
| 142 | /** | ||
| 143 | * @vm_start: VM range start VMA. Together with @vm_end, this defines | ||
| 144 | * the window of valid GPU VAs. Userspace is expected to subdivide VAs | ||
| 145 | * out of this window. | ||
| 146 | * | ||
| 147 | * This window contains all virtual addresses that userspace needs to | ||
| 148 | * know about. There may be kernel-internal GPU VAs outside this range, | ||
| 149 | * but that detail is not relevant here. | ||
| 150 | */ | ||
| 151 | __u64 vm_start; | ||
| 152 | |||
| 153 | /** @vm_end: VM range end VMA */ | ||
| 154 | __u64 vm_end; | ||
| 155 | |||
| 156 | /** | ||
| 157 | * @vm_kernel_min_size: Minimum kernel VMA window size. | ||
| 158 | * | ||
| 159 | * When creating a VM, userspace is required to carve out a section of | ||
| 160 | * virtual addresses (within the range given by @vm_start and | ||
| 161 | * @vm_end). The kernel will allocate various internal structures | ||
| 162 | * within the specified VA range. | ||
| 163 | * | ||
| 164 | * Allowing userspace to choose the VA range for the kernel, rather than | ||
| 165 | * the kernel reserving VAs and requiring userspace to cope, can assist | ||
| 166 | * in implementing SVM. | ||
| 167 | */ | ||
| 168 | __u64 vm_kernel_min_size; | ||
| 169 | |||
| 170 | /** | ||
| 171 | * @max_commands_per_submission: Maximum number of supported commands | ||
| 172 | * per submission. This mirrors firmware limits. Userspace must split up | ||
| 173 | * larger command buffers, which may require inserting additional | ||
| 174 | * synchronization. | ||
| 175 | */ | ||
| 176 | __u32 max_commands_per_submission; | ||
| 177 | |||
| 178 | /** | ||
| 179 | * @max_attachments: Maximum number of drm_asahi_attachment's per | ||
| 180 | * command | ||
| 181 | */ | ||
| 182 | __u32 max_attachments; | ||
| 183 | |||
| 184 | /** | ||
| 185 | * @command_timestamp_frequency_hz: Timebase frequency for timestamps | ||
| 186 | * written during command execution, specified via drm_asahi_timestamp | ||
| 187 | * structures. As this rate is controlled by the firmware, it is a | ||
| 188 | * queryable parameter. | ||
| 189 | * | ||
| 190 | * Userspace must divide by this frequency to convert timestamps to | ||
| 191 | * seconds, rather than hardcoding a particular firmware's rate. | ||
| 192 | */ | ||
| 193 | __u64 command_timestamp_frequency_hz; | ||
| 194 | }; | ||
| 195 | |||
| 196 | /** | ||
| 197 | * enum drm_asahi_feature - Feature bits | ||
| 198 | * | ||
| 199 | * This covers only features that userspace cannot infer from the architecture | ||
| 200 | * version. Most features don't need to be here. | ||
| 201 | */ | ||
| 202 | enum drm_asahi_feature { | ||
| 203 | /** | ||
| 204 | * @DRM_ASAHI_FEATURE_SOFT_FAULTS: GPU has "soft fault" enabled. Shader | ||
| 205 | * loads of unmapped memory will return zero. Shader stores to unmapped | ||
| 206 | * memory will be silently discarded. Note that only shader load/store | ||
| 207 | * is affected. Other hardware units are not affected, notably including | ||
| 208 | * texture sampling. | ||
| 209 | * | ||
| 210 | * Soft fault is set when initializing the GPU and cannot be runtime | ||
| 211 | * toggled. Therefore, it is exposed as a feature bit and not a | ||
| 212 | * userspace-settable flag on the VM. When soft fault is enabled, | ||
| 213 | * userspace can speculate memory accesses more aggressively. | ||
| 214 | */ | ||
| 215 | DRM_ASAHI_FEATURE_SOFT_FAULTS = (1UL) << 0, | ||
| 216 | }; | ||
| 217 | |||
| 218 | /** | ||
| 219 | * struct drm_asahi_get_params - Arguments passed to DRM_IOCTL_ASAHI_GET_PARAMS | ||
| 220 | */ | ||
| 221 | struct drm_asahi_get_params { | ||
| 222 | /** @param_group: Parameter group to fetch (MBZ) */ | ||
| 223 | __u32 param_group; | ||
| 224 | |||
| 225 | /** @pad: MBZ */ | ||
| 226 | __u32 pad; | ||
| 227 | |||
| 228 | /** @pointer: User pointer to write parameter struct */ | ||
| 229 | __u64 pointer; | ||
| 230 | |||
| 231 | /** | ||
| 232 | * @size: Size of the user buffer. In case of older userspace, this may | ||
| 233 | * be less than sizeof(struct drm_asahi_params_global). The kernel will | ||
| 234 | * not write past the length specified here, allowing extensibility. | ||
| 235 | */ | ||
| 236 | __u64 size; | ||
| 237 | }; | ||
| 238 | |||
| 239 | /** | ||
| 240 | * struct drm_asahi_vm_create - Arguments passed to DRM_IOCTL_ASAHI_VM_CREATE | ||
| 241 | */ | ||
| 242 | struct drm_asahi_vm_create { | ||
| 243 | /** | ||
| 244 | * @kernel_start: Start of the kernel-reserved address range. See | ||
| 245 | * drm_asahi_params_global::vm_kernel_min_size. | ||
| 246 | * | ||
| 247 | * Both @kernel_start and @kernel_end must be within the range of | ||
| 248 | * valid VAs given by drm_asahi_params_global::vm_start and | ||
| 249 | * drm_asahi_params_global::vm_end. The size of the kernel range | ||
| 250 | * (@kernel_end - @kernel_start) must be at least | ||
| 251 | * drm_asahi_params_global::vm_kernel_min_size. | ||
| 252 | * | ||
| 253 | * Userspace must not bind any memory on this VM into this reserved | ||
| 254 | * range, it is for kernel use only. | ||
| 255 | */ | ||
| 256 | __u64 kernel_start; | ||
| 257 | |||
| 258 | /** | ||
| 259 | * @kernel_end: End of the kernel-reserved address range. See | ||
| 260 | * @kernel_start. | ||
| 261 | */ | ||
| 262 | __u64 kernel_end; | ||
| 263 | |||
| 264 | /** @vm_id: Returned VM ID */ | ||
| 265 | __u32 vm_id; | ||
| 266 | |||
| 267 | /** @pad: MBZ */ | ||
| 268 | __u32 pad; | ||
| 269 | }; | ||
| 270 | |||
| 271 | /** | ||
| 272 | * struct drm_asahi_vm_destroy - Arguments passed to DRM_IOCTL_ASAHI_VM_DESTROY | ||
| 273 | */ | ||
| 274 | struct drm_asahi_vm_destroy { | ||
| 275 | /** @vm_id: VM ID to be destroyed */ | ||
| 276 | __u32 vm_id; | ||
| 277 | |||
| 278 | /** @pad: MBZ */ | ||
| 279 | __u32 pad; | ||
| 280 | }; | ||
| 281 | |||
| 282 | /** | ||
| 283 | * enum drm_asahi_gem_flags - Flags for GEM creation | ||
| 284 | */ | ||
| 285 | enum drm_asahi_gem_flags { | ||
| 286 | /** | ||
| 287 | * @DRM_ASAHI_GEM_WRITEBACK: BO should be CPU-mapped as writeback. | ||
| 288 | * | ||
| 289 | * Map as writeback instead of write-combine. This optimizes for CPU | ||
| 290 | * reads. | ||
| 291 | */ | ||
| 292 | DRM_ASAHI_GEM_WRITEBACK = (1L << 0), | ||
| 293 | |||
| 294 | /** | ||
| 295 | * @DRM_ASAHI_GEM_VM_PRIVATE: BO is private to this GPU VM (no exports). | ||
| 296 | */ | ||
| 297 | DRM_ASAHI_GEM_VM_PRIVATE = (1L << 1), | ||
| 298 | }; | ||
| 299 | |||
| 300 | /** | ||
| 301 | * struct drm_asahi_gem_create - Arguments passed to DRM_IOCTL_ASAHI_GEM_CREATE | ||
| 302 | */ | ||
| 303 | struct drm_asahi_gem_create { | ||
| 304 | /** @size: Size of the BO */ | ||
| 305 | __u64 size; | ||
| 306 | |||
| 307 | /** @flags: Combination of drm_asahi_gem_flags flags. */ | ||
| 308 | __u32 flags; | ||
| 309 | |||
| 310 | /** | ||
| 311 | * @vm_id: VM ID to assign to the BO, if DRM_ASAHI_GEM_VM_PRIVATE is set | ||
| 312 | */ | ||
| 313 | __u32 vm_id; | ||
| 314 | |||
| 315 | /** @handle: Returned GEM handle for the BO */ | ||
| 316 | __u32 handle; | ||
| 317 | |||
| 318 | /** @pad: MBZ */ | ||
| 319 | __u32 pad; | ||
| 320 | }; | ||
| 321 | |||
| 322 | /** | ||
| 323 | * struct drm_asahi_gem_mmap_offset - Arguments passed to | ||
| 324 | * DRM_IOCTL_ASAHI_GEM_MMAP_OFFSET | ||
| 325 | */ | ||
| 326 | struct drm_asahi_gem_mmap_offset { | ||
| 327 | /** @handle: Handle for the object being mapped. */ | ||
| 328 | __u32 handle; | ||
| 329 | |||
| 330 | /** @flags: Must be zero */ | ||
| 331 | __u32 flags; | ||
| 332 | |||
| 333 | /** @offset: The fake offset to use for subsequent mmap call */ | ||
| 334 | __u64 offset; | ||
| 335 | }; | ||
| 336 | |||
| 337 | /** | ||
| 338 | * enum drm_asahi_bind_flags - Flags for GEM binding | ||
| 339 | */ | ||
| 340 | enum drm_asahi_bind_flags { | ||
| 341 | /** | ||
| 342 | * @DRM_ASAHI_BIND_UNBIND: Instead of binding a GEM object to the range, | ||
| 343 | * simply unbind the GPU VMA range. | ||
| 344 | */ | ||
| 345 | DRM_ASAHI_BIND_UNBIND = (1L << 0), | ||
| 346 | |||
| 347 | /** @DRM_ASAHI_BIND_READ: Map BO with GPU read permission */ | ||
| 348 | DRM_ASAHI_BIND_READ = (1L << 1), | ||
| 349 | |||
| 350 | /** @DRM_ASAHI_BIND_WRITE: Map BO with GPU write permission */ | ||
| 351 | DRM_ASAHI_BIND_WRITE = (1L << 2), | ||
| 352 | |||
| 353 | /** | ||
| 354 | * @DRM_ASAHI_BIND_SINGLE_PAGE: Map a single page of the BO repeatedly | ||
| 355 | * across the VA range. | ||
| 356 | * | ||
| 357 | * This is useful to fill a VA range with scratch pages or zero pages. | ||
| 358 | * It is intended as a mechanism to accelerate sparse. | ||
| 359 | */ | ||
| 360 | DRM_ASAHI_BIND_SINGLE_PAGE = (1L << 3), | ||
| 361 | }; | ||
| 362 | |||
| 363 | /** | ||
| 364 | * struct drm_asahi_gem_bind_op - Description of a single GEM bind operation. | ||
| 365 | */ | ||
| 366 | struct drm_asahi_gem_bind_op { | ||
| 367 | /** @flags: Combination of drm_asahi_bind_flags flags. */ | ||
| 368 | __u32 flags; | ||
| 369 | |||
| 370 | /** @handle: GEM object to bind (except for UNBIND) */ | ||
| 371 | __u32 handle; | ||
| 372 | |||
| 373 | /** | ||
| 374 | * @offset: Offset into the object (except for UNBIND). | ||
| 375 | * | ||
| 376 | * For a regular bind, this is the beginning of the region of the GEM | ||
| 377 | * object to bind. | ||
| 378 | * | ||
| 379 | * For a single-page bind, this is the offset to the single page that | ||
| 380 | * will be repeatedly bound. | ||
| 381 | * | ||
| 382 | * Must be page-size aligned. | ||
| 383 | */ | ||
| 384 | __u64 offset; | ||
| 385 | |||
| 386 | /** | ||
| 387 | * @range: Number of bytes to bind/unbind to @addr. | ||
| 388 | * | ||
| 389 | * Must be page-size aligned. | ||
| 390 | */ | ||
| 391 | __u64 range; | ||
| 392 | |||
| 393 | /** | ||
| 394 | * @addr: Address to bind to. | ||
| 395 | * | ||
| 396 | * Must be page-size aligned. | ||
| 397 | */ | ||
| 398 | __u64 addr; | ||
| 399 | }; | ||
| 400 | |||
| 401 | /** | ||
| 402 | * struct drm_asahi_vm_bind - Arguments passed to | ||
| 403 | * DRM_IOCTL_ASAHI_VM_BIND | ||
| 404 | */ | ||
| 405 | struct drm_asahi_vm_bind { | ||
| 406 | /** @vm_id: The ID of the VM to bind to */ | ||
| 407 | __u32 vm_id; | ||
| 408 | |||
| 409 | /** @num_binds: number of binds in this IOCTL. */ | ||
| 410 | __u32 num_binds; | ||
| 411 | |||
| 412 | /** | ||
| 413 | * @stride: Stride in bytes between consecutive binds. This allows | ||
| 414 | * extensibility of drm_asahi_gem_bind_op. | ||
| 415 | */ | ||
| 416 | __u32 stride; | ||
| 417 | |||
| 418 | /** @pad: MBZ */ | ||
| 419 | __u32 pad; | ||
| 420 | |||
| 421 | /** | ||
| 422 | * @userptr: User pointer to an array of @num_binds structures of type | ||
| 423 | * @drm_asahi_gem_bind_op and size @stride bytes. | ||
| 424 | */ | ||
| 425 | __u64 userptr; | ||
| 426 | }; | ||
| 427 | |||
| 428 | /** | ||
| 429 | * enum drm_asahi_bind_object_op - Special object bind operation | ||
| 430 | */ | ||
| 431 | enum drm_asahi_bind_object_op { | ||
| 432 | /** @DRM_ASAHI_BIND_OBJECT_OP_BIND: Bind a BO as a special GPU object */ | ||
| 433 | DRM_ASAHI_BIND_OBJECT_OP_BIND = 0, | ||
| 434 | |||
| 435 | /** @DRM_ASAHI_BIND_OBJECT_OP_UNBIND: Unbind a special GPU object */ | ||
| 436 | DRM_ASAHI_BIND_OBJECT_OP_UNBIND = 1, | ||
| 437 | }; | ||
| 438 | |||
| 439 | /** | ||
| 440 | * enum drm_asahi_bind_object_flags - Special object bind flags | ||
| 441 | */ | ||
| 442 | enum drm_asahi_bind_object_flags { | ||
| 443 | /** | ||
| 444 | * @DRM_ASAHI_BIND_OBJECT_USAGE_TIMESTAMPS: Map a BO as a timestamp | ||
| 445 | * buffer. | ||
| 446 | */ | ||
| 447 | DRM_ASAHI_BIND_OBJECT_USAGE_TIMESTAMPS = (1L << 0), | ||
| 448 | }; | ||
| 449 | |||
| 450 | /** | ||
| 451 | * struct drm_asahi_gem_bind_object - Arguments passed to | ||
| 452 | * DRM_IOCTL_ASAHI_GEM_BIND_OBJECT | ||
| 453 | */ | ||
| 454 | struct drm_asahi_gem_bind_object { | ||
| 455 | /** @op: Bind operation (enum drm_asahi_bind_object_op) */ | ||
| 456 | __u32 op; | ||
| 457 | |||
| 458 | /** @flags: Combination of drm_asahi_bind_object_flags flags. */ | ||
| 459 | __u32 flags; | ||
| 460 | |||
| 461 | /** @handle: GEM object to bind/unbind (BIND) */ | ||
| 462 | __u32 handle; | ||
| 463 | |||
| 464 | /** @vm_id: The ID of the VM to operate on (MBZ currently) */ | ||
| 465 | __u32 vm_id; | ||
| 466 | |||
| 467 | /** @offset: Offset into the object (BIND only) */ | ||
| 468 | __u64 offset; | ||
| 469 | |||
| 470 | /** @range: Number of bytes to bind/unbind (BIND only) */ | ||
| 471 | __u64 range; | ||
| 472 | |||
| 473 | /** @object_handle: Object handle (out for BIND, in for UNBIND) */ | ||
| 474 | __u32 object_handle; | ||
| 475 | |||
| 476 | /** @pad: MBZ */ | ||
| 477 | __u32 pad; | ||
| 478 | }; | ||
| 479 | |||
| 480 | /** | ||
| 481 | * enum drm_asahi_cmd_type - Command type | ||
| 482 | */ | ||
| 483 | enum drm_asahi_cmd_type { | ||
| 484 | /** | ||
| 485 | * @DRM_ASAHI_CMD_RENDER: Render command, executing on the render | ||
| 486 | * subqueue. Combined vertex and fragment operation. | ||
| 487 | * | ||
| 488 | * Followed by a @drm_asahi_cmd_render payload. | ||
| 489 | */ | ||
| 490 | DRM_ASAHI_CMD_RENDER = 0, | ||
| 491 | |||
| 492 | /** | ||
| 493 | * @DRM_ASAHI_CMD_COMPUTE: Compute command on the compute subqueue. | ||
| 494 | * | ||
| 495 | * Followed by a @drm_asahi_cmd_compute payload. | ||
| 496 | */ | ||
| 497 | DRM_ASAHI_CMD_COMPUTE = 1, | ||
| 498 | |||
| 499 | /** | ||
| 500 | * @DRM_ASAHI_SET_VERTEX_ATTACHMENTS: Software command to set | ||
| 501 | * attachments for subsequent vertex shaders in the same submit. | ||
| 502 | * | ||
| 503 | * Followed by (possibly multiple) @drm_asahi_attachment payloads. | ||
| 504 | */ | ||
| 505 | DRM_ASAHI_SET_VERTEX_ATTACHMENTS = 2, | ||
| 506 | |||
| 507 | /** | ||
| 508 | * @DRM_ASAHI_SET_FRAGMENT_ATTACHMENTS: Software command to set | ||
| 509 | * attachments for subsequent fragment shaders in the same submit. | ||
| 510 | * | ||
| 511 | * Followed by (possibly multiple) @drm_asahi_attachment payloads. | ||
| 512 | */ | ||
| 513 | DRM_ASAHI_SET_FRAGMENT_ATTACHMENTS = 3, | ||
| 514 | |||
| 515 | /** | ||
| 516 | * @DRM_ASAHI_SET_COMPUTE_ATTACHMENTS: Software command to set | ||
| 517 | * attachments for subsequent compute shaders in the same submit. | ||
| 518 | * | ||
| 519 | * Followed by (possibly multiple) @drm_asahi_attachment payloads. | ||
| 520 | */ | ||
| 521 | DRM_ASAHI_SET_COMPUTE_ATTACHMENTS = 4, | ||
| 522 | }; | ||
| 523 | |||
| 524 | /** | ||
| 525 | * enum drm_asahi_priority - Scheduling queue priority. | ||
| 526 | * | ||
| 527 | * These priorities are forwarded to the firmware to influence firmware | ||
| 528 | * scheduling. The exact policy is ultimately decided by firmware, but | ||
| 529 | * these enums allow userspace to communicate the intentions. | ||
| 530 | */ | ||
| 531 | enum drm_asahi_priority { | ||
| 532 | /** @DRM_ASAHI_PRIORITY_LOW: Low priority queue. */ | ||
| 533 | DRM_ASAHI_PRIORITY_LOW = 0, | ||
| 534 | |||
| 535 | /** @DRM_ASAHI_PRIORITY_MEDIUM: Medium priority queue. */ | ||
| 536 | DRM_ASAHI_PRIORITY_MEDIUM = 1, | ||
| 537 | |||
| 538 | /** | ||
| 539 | * @DRM_ASAHI_PRIORITY_HIGH: High priority queue. | ||
| 540 | * | ||
| 541 | * Reserved for future extension. | ||
| 542 | */ | ||
| 543 | DRM_ASAHI_PRIORITY_HIGH = 2, | ||
| 544 | |||
| 545 | /** | ||
| 546 | * @DRM_ASAHI_PRIORITY_REALTIME: Real-time priority queue. | ||
| 547 | * | ||
| 548 | * Reserved for future extension. | ||
| 549 | */ | ||
| 550 | DRM_ASAHI_PRIORITY_REALTIME = 3, | ||
| 551 | }; | ||
| 552 | |||
| 553 | /** | ||
| 554 | * struct drm_asahi_queue_create - Arguments passed to | ||
| 555 | * DRM_IOCTL_ASAHI_QUEUE_CREATE | ||
| 556 | */ | ||
| 557 | struct drm_asahi_queue_create { | ||
| 558 | /** @flags: MBZ */ | ||
| 559 | __u32 flags; | ||
| 560 | |||
| 561 | /** @vm_id: The ID of the VM this queue is bound to */ | ||
| 562 | __u32 vm_id; | ||
| 563 | |||
| 564 | /** @priority: One of drm_asahi_priority */ | ||
| 565 | __u32 priority; | ||
| 566 | |||
| 567 | /** @queue_id: The returned queue ID */ | ||
| 568 | __u32 queue_id; | ||
| 569 | |||
| 570 | /** | ||
| 571 | * @usc_exec_base: GPU base address for all USC binaries (shaders) on | ||
| 572 | * this queue. USC addresses are 32-bit relative to this 64-bit base. | ||
| 573 | * | ||
| 574 | * This sets the following registers on all queue commands: | ||
| 575 | * | ||
| 576 | * USC_EXEC_BASE_TA (vertex) | ||
| 577 | * USC_EXEC_BASE_ISP (fragment) | ||
| 578 | * USC_EXEC_BASE_CP (compute) | ||
| 579 | * | ||
| 580 | * While the hardware lets us configure these independently per command, | ||
| 581 | * we do not have a use case for this. Instead, we expect userspace to | ||
| 582 | * fix a 4GiB VA carveout for USC memory and pass its base address here. | ||
| 583 | */ | ||
| 584 | __u64 usc_exec_base; | ||
| 585 | }; | ||
| 586 | |||
| 587 | /** | ||
| 588 | * struct drm_asahi_queue_destroy - Arguments passed to | ||
| 589 | * DRM_IOCTL_ASAHI_QUEUE_DESTROY | ||
| 590 | */ | ||
| 591 | struct drm_asahi_queue_destroy { | ||
| 592 | /** @queue_id: The queue ID to be destroyed */ | ||
| 593 | __u32 queue_id; | ||
| 594 | |||
| 595 | /** @pad: MBZ */ | ||
| 596 | __u32 pad; | ||
| 597 | }; | ||
| 598 | |||
| 599 | /** | ||
| 600 | * enum drm_asahi_sync_type - Sync item type | ||
| 601 | */ | ||
| 602 | enum drm_asahi_sync_type { | ||
| 603 | /** @DRM_ASAHI_SYNC_SYNCOBJ: Binary sync object */ | ||
| 604 | DRM_ASAHI_SYNC_SYNCOBJ = 0, | ||
| 605 | |||
| 606 | /** @DRM_ASAHI_SYNC_TIMELINE_SYNCOBJ: Timeline sync object */ | ||
| 607 | DRM_ASAHI_SYNC_TIMELINE_SYNCOBJ = 1, | ||
| 608 | }; | ||
| 609 | |||
| 610 | /** | ||
| 611 | * struct drm_asahi_sync - Sync item | ||
| 612 | */ | ||
| 613 | struct drm_asahi_sync { | ||
| 614 | /** @sync_type: One of drm_asahi_sync_type */ | ||
| 615 | __u32 sync_type; | ||
| 616 | |||
| 617 | /** @handle: The sync object handle */ | ||
| 618 | __u32 handle; | ||
| 619 | |||
| 620 | /** @timeline_value: Timeline value for timeline sync objects */ | ||
| 621 | __u64 timeline_value; | ||
| 622 | }; | ||
| 623 | |||
| 624 | /** | ||
| 625 | * define DRM_ASAHI_BARRIER_NONE - Command index for no barrier | ||
| 626 | * | ||
| 627 | * This special value may be passed in to drm_asahi_command::vdm_barrier or | ||
| 628 | * drm_asahi_command::cdm_barrier to indicate that the respective subqueue | ||
| 629 | * should not wait on any previous work. | ||
| 630 | */ | ||
| 631 | #define DRM_ASAHI_BARRIER_NONE (0xFFFFu) | ||
| 632 | |||
| 633 | /** | ||
| 634 | * struct drm_asahi_cmd_header - Top level command structure | ||
| 635 | * | ||
| 636 | * This struct is core to the command buffer definition and therefore is not | ||
| 637 | * extensible. | ||
| 638 | */ | ||
| 639 | struct drm_asahi_cmd_header { | ||
| 640 | /** @cmd_type: One of drm_asahi_cmd_type */ | ||
| 641 | __u16 cmd_type; | ||
| 642 | |||
| 643 | /** | ||
| 644 | * @size: Size of this command, not including this header. | ||
| 645 | * | ||
| 646 | * For hardware commands, this enables extensibility of commands without | ||
| 647 | * requiring extra command types. Passing a command that is shorter | ||
| 648 | * than expected is explicitly allowed for backwards-compatibility. | ||
| 649 | * Truncated fields will be zeroed. | ||
| 650 | * | ||
| 651 | * For the synthetic attachment setting commands, this implicitly | ||
| 652 | * encodes the number of attachments. These commands take multiple | ||
| 653 | * fixed-size @drm_asahi_attachment structures as their payload, so size | ||
| 654 | * equals number of attachments * sizeof(struct drm_asahi_attachment). | ||
| 655 | */ | ||
| 656 | __u16 size; | ||
| 657 | |||
| 658 | /** | ||
| 659 | * @vdm_barrier: VDM (render) command index to wait on. | ||
| 660 | * | ||
| 661 | * Barriers are indices relative to the beginning of a given submit. A | ||
| 662 | * barrier of 0 waits on commands submitted to the respective subqueue | ||
| 663 | * in previous submit ioctls. A barrier of N waits on N previous | ||
| 664 | * commands on the subqueue within the current submit ioctl. As a | ||
| 665 | * special case, passing @DRM_ASAHI_BARRIER_NONE avoids waiting on any | ||
| 666 | * commands in the subqueue. | ||
| 667 | * | ||
| 668 | * Examples: | ||
| 669 | * | ||
| 670 | * 0: This waits on all previous work. | ||
| 671 | * | ||
| 672 | * NONE: This does not wait for anything on this subqueue. | ||
| 673 | * | ||
| 674 | * 1: This waits on the first render command in the submit. | ||
| 675 | * This is valid only if there are multiple render commands in the | ||
| 676 | * same submit. | ||
| 677 | * | ||
| 678 | * Barriers are valid only for hardware commands. Synthetic software | ||
| 679 | * commands to set attachments must pass NONE here. | ||
| 680 | */ | ||
| 681 | __u16 vdm_barrier; | ||
| 682 | |||
| 683 | /** | ||
| 684 | * @cdm_barrier: CDM (compute) command index to wait on. | ||
| 685 | * | ||
| 686 | * See @vdm_barrier, and replace VDM/render with CDM/compute. | ||
| 687 | */ | ||
| 688 | __u16 cdm_barrier; | ||
| 689 | }; | ||
| 690 | |||
| 691 | /** | ||
| 692 | * struct drm_asahi_submit - Arguments passed to DRM_IOCTL_ASAHI_SUBMIT | ||
| 693 | */ | ||
| 694 | struct drm_asahi_submit { | ||
| 695 | /** | ||
| 696 | * @syncs: An optional pointer to an array of drm_asahi_sync. The first | ||
| 697 | * @in_sync_count elements are in-syncs, then the remaining | ||
| 698 | * @out_sync_count elements are out-syncs. Using a single array with | ||
| 699 | * explicit partitioning simplifies handling. | ||
| 700 | */ | ||
| 701 | __u64 syncs; | ||
| 702 | |||
| 703 | /** | ||
| 704 | * @cmdbuf: Pointer to the command buffer to submit. | ||
| 705 | * | ||
| 706 | * This is a flat command buffer. By design, it contains no CPU | ||
| 707 | * pointers, which makes it suitable for a virtgpu wire protocol without | ||
| 708 | * requiring any serializing/deserializing step. | ||
| 709 | * | ||
| 710 | * It consists of a series of commands. Each command begins with a | ||
| 711 | * fixed-size @drm_asahi_cmd_header header and is followed by a | ||
| 712 | * variable-length payload according to the type and size in the header. | ||
| 713 | * | ||
| 714 | * The combined count of "real" hardware commands must be nonzero and at | ||
| 715 | * most drm_asahi_params_global::max_commands_per_submission. | ||
| 716 | */ | ||
| 717 | __u64 cmdbuf; | ||
| 718 | |||
| 719 | /** @flags: Flags for command submission (MBZ) */ | ||
| 720 | __u32 flags; | ||
| 721 | |||
| 722 | /** @queue_id: The queue ID to be submitted to */ | ||
| 723 | __u32 queue_id; | ||
| 724 | |||
| 725 | /** | ||
| 726 | * @in_sync_count: Number of sync objects to wait on before starting | ||
| 727 | * this job. | ||
| 728 | */ | ||
| 729 | __u32 in_sync_count; | ||
| 730 | |||
| 731 | /** | ||
| 732 | * @out_sync_count: Number of sync objects to signal upon completion of | ||
| 733 | * this job. | ||
| 734 | */ | ||
| 735 | __u32 out_sync_count; | ||
| 736 | |||
| 737 | /** @cmdbuf_size: Command buffer size in bytes */ | ||
| 738 | __u32 cmdbuf_size; | ||
| 739 | |||
| 740 | /** @pad: MBZ */ | ||
| 741 | __u32 pad; | ||
| 742 | }; | ||
| 743 | |||
| 744 | /** | ||
| 745 | * struct drm_asahi_attachment - Describe an "attachment". | ||
| 746 | * | ||
| 747 | * Attachments are any memory written by shaders, notably including render | ||
| 748 | * target attachments written by the end-of-tile program. This is purely a hint | ||
| 749 | * about the accessed memory regions. It is optional to specify, which is | ||
| 750 | * fortunate as it cannot be specified precisely with bindless access anyway. | ||
| 751 | * But where possible, it's probably a good idea for userspace to include these | ||
| 752 | * hints, forwarded to the firmware. | ||
| 753 | * | ||
| 754 | * This struct is implicitly sized and therefore is not extensible. | ||
| 755 | */ | ||
| 756 | struct drm_asahi_attachment { | ||
| 757 | /** @pointer: Base address of the attachment */ | ||
| 758 | __u64 pointer; | ||
| 759 | |||
| 760 | /** @size: Size of the attachment in bytes */ | ||
| 761 | __u64 size; | ||
| 762 | |||
| 763 | /** @pad: MBZ */ | ||
| 764 | __u32 pad; | ||
| 765 | |||
| 766 | /** @flags: MBZ */ | ||
| 767 | __u32 flags; | ||
| 768 | }; | ||
| 769 | |||
| 770 | enum drm_asahi_render_flags { | ||
| 771 | /** | ||
| 772 | * @DRM_ASAHI_RENDER_VERTEX_SCRATCH: A vertex stage shader uses scratch | ||
| 773 | * memory. | ||
| 774 | */ | ||
| 775 | DRM_ASAHI_RENDER_VERTEX_SCRATCH = (1U << 0), | ||
| 776 | |||
| 777 | /** | ||
| 778 | * @DRM_ASAHI_RENDER_PROCESS_EMPTY_TILES: Process even empty tiles. | ||
| 779 | * This must be set when clearing render targets. | ||
| 780 | */ | ||
| 781 | DRM_ASAHI_RENDER_PROCESS_EMPTY_TILES = (1U << 1), | ||
| 782 | |||
| 783 | /** | ||
| 784 | * @DRM_ASAHI_RENDER_NO_VERTEX_CLUSTERING: Run vertex stage on a single | ||
| 785 | * cluster (on multi-cluster GPUs) | ||
| 786 | * | ||
| 787 | * This harms performance but can workaround certain sync/coherency | ||
| 788 | * bugs, and therefore is useful for debugging. | ||
| 789 | */ | ||
| 790 | DRM_ASAHI_RENDER_NO_VERTEX_CLUSTERING = (1U << 2), | ||
| 791 | |||
| 792 | /** | ||
| 793 | * @DRM_ASAHI_RENDER_DBIAS_IS_INT: Use integer depth bias formula. | ||
| 794 | * | ||
| 795 | * Graphics specifications contain two alternate formulas for depth | ||
| 796 | * bias, a float formula used with floating-point depth buffers and an | ||
| 797 | * integer formula using with unorm depth buffers. This flag specifies | ||
| 798 | * that the integer formula should be used. If omitted, the float | ||
| 799 | * formula is used instead. | ||
| 800 | * | ||
| 801 | * This corresponds to bit 18 of the relevant hardware control register, | ||
| 802 | * so we match that here for efficiency. | ||
| 803 | */ | ||
| 804 | DRM_ASAHI_RENDER_DBIAS_IS_INT = (1U << 18), | ||
| 805 | }; | ||
| 806 | |||
| 807 | /** | ||
| 808 | * struct drm_asahi_zls_buffer - Describe a depth or stencil buffer. | ||
| 809 | * | ||
| 810 | * These fields correspond to hardware registers in the ZLS (Z Load/Store) unit. | ||
| 811 | * There are three hardware registers for each field respectively for loads, | ||
| 812 | * stores, and partial renders. In practice, it makes sense to set all to the | ||
| 813 | * same values, except in exceptional cases not yet implemented in userspace, so | ||
| 814 | * we do not duplicate here for simplicity/efficiency. | ||
| 815 | * | ||
| 816 | * This struct is embedded in other structs and therefore is not extensible. | ||
| 817 | */ | ||
| 818 | struct drm_asahi_zls_buffer { | ||
| 819 | /** @base: Base address of the buffer */ | ||
| 820 | __u64 base; | ||
| 821 | |||
| 822 | /** | ||
| 823 | * @comp_base: If the load buffer is compressed, address of the | ||
| 824 | * compression metadata section. | ||
| 825 | */ | ||
| 826 | __u64 comp_base; | ||
| 827 | |||
| 828 | /** | ||
| 829 | * @stride: If layered rendering is enabled, the number of bytes | ||
| 830 | * between each layer of the buffer. | ||
| 831 | */ | ||
| 832 | __u32 stride; | ||
| 833 | |||
| 834 | /** | ||
| 835 | * @comp_stride: If layered rendering is enabled, the number of bytes | ||
| 836 | * between each layer of the compression metadata. | ||
| 837 | */ | ||
| 838 | __u32 comp_stride; | ||
| 839 | }; | ||
| 840 | |||
| 841 | /** | ||
| 842 | * struct drm_asahi_timestamp - Describe a timestamp write. | ||
| 843 | * | ||
| 844 | * The firmware can optionally write the GPU timestamp at render pass | ||
| 845 | * granularities, but it needs to be mapped specially via | ||
| 846 | * DRM_IOCTL_ASAHI_GEM_BIND_OBJECT. This structure therefore describes where to | ||
| 847 | * write as a handle-offset pair, rather than a GPU address like normal. | ||
| 848 | * | ||
| 849 | * This struct is embedded in other structs and therefore is not extensible. | ||
| 850 | */ | ||
| 851 | struct drm_asahi_timestamp { | ||
| 852 | /** | ||
| 853 | * @handle: Handle of the timestamp buffer, or 0 to skip this | ||
| 854 | * timestamp. If nonzero, this must equal the value returned in | ||
| 855 | * drm_asahi_gem_bind_object::object_handle. | ||
| 856 | */ | ||
| 857 | __u32 handle; | ||
| 858 | |||
| 859 | /** @offset: Offset to write into the timestamp buffer */ | ||
| 860 | __u32 offset; | ||
| 861 | }; | ||
| 862 | |||
| 863 | /** | ||
| 864 | * struct drm_asahi_timestamps - Describe timestamp writes. | ||
| 865 | * | ||
| 866 | * Each operation that can be timestamped, can be timestamped at the start and | ||
| 867 | * end. Therefore, drm_asahi_timestamp structs always come in pairs, bundled | ||
| 868 | * together into drm_asahi_timestamps. | ||
| 869 | * | ||
| 870 | * This struct is embedded in other structs and therefore is not extensible. | ||
| 871 | */ | ||
| 872 | struct drm_asahi_timestamps { | ||
| 873 | /** @start: Timestamp recorded at the start of the operation */ | ||
| 874 | struct drm_asahi_timestamp start; | ||
| 875 | |||
| 876 | /** @end: Timestamp recorded at the end of the operation */ | ||
| 877 | struct drm_asahi_timestamp end; | ||
| 878 | }; | ||
| 879 | |||
| 880 | /** | ||
| 881 | * struct drm_asahi_helper_program - Describe helper program configuration. | ||
| 882 | * | ||
| 883 | * The helper program is a compute-like kernel required for various hardware | ||
| 884 | * functionality. Its most important role is dynamically allocating | ||
| 885 | * scratch/stack memory for individual subgroups, by partitioning a static | ||
| 886 | * allocation shared for the whole device. It is supplied by userspace via | ||
| 887 | * drm_asahi_helper_program and internally dispatched by the hardware as needed. | ||
| 888 | * | ||
| 889 | * This struct is embedded in other structs and therefore is not extensible. | ||
| 890 | */ | ||
| 891 | struct drm_asahi_helper_program { | ||
| 892 | /** | ||
| 893 | * @binary: USC address to the helper program binary. This is a tagged | ||
| 894 | * pointer with configuration in the bottom bits. | ||
| 895 | */ | ||
| 896 | __u32 binary; | ||
| 897 | |||
| 898 | /** @cfg: Additional configuration bits for the helper program. */ | ||
| 899 | __u32 cfg; | ||
| 900 | |||
| 901 | /** | ||
| 902 | * @data: Data passed to the helper program. This value is not | ||
| 903 | * interpreted by the kernel, firmware, or hardware in any way. It is | ||
| 904 | * simply a sideband for userspace, set with the submit ioctl and read | ||
| 905 | * via special registers inside the helper program. | ||
| 906 | * | ||
| 907 | * In practice, userspace will pass a 64-bit GPU VA here pointing to the | ||
| 908 | * actual arguments, which presumably don't fit in 64-bits. | ||
| 909 | */ | ||
| 910 | __u64 data; | ||
| 911 | }; | ||
| 912 | |||
| 913 | /** | ||
| 914 | * struct drm_asahi_bg_eot - Describe a background or end-of-tile program. | ||
| 915 | * | ||
| 916 | * The background and end-of-tile programs are dispatched by the hardware at the | ||
| 917 | * beginning and end of rendering. As the hardware "tilebuffer" is simply local | ||
| 918 | * memory, these programs are necessary to implement API-level render targets. | ||
| 919 | * The fragment-like background program is responsible for loading either the | ||
| 920 | * clear colour or the existing render target contents, while the compute-like | ||
| 921 | * end-of-tile program stores the tilebuffer contents to memory. | ||
| 922 | * | ||
| 923 | * This struct is embedded in other structs and therefore is not extensible. | ||
| 924 | */ | ||
| 925 | struct drm_asahi_bg_eot { | ||
| 926 | /** | ||
| 927 | * @usc: USC address of the hardware USC words binding resources | ||
| 928 | * (including images and uniforms) and the program itself. Note this is | ||
| 929 | * an additional layer of indirection compared to the helper program, | ||
| 930 | * avoiding the need for a sideband for data. This is a tagged pointer | ||
| 931 | * with additional configuration in the bottom bits. | ||
| 932 | */ | ||
| 933 | __u32 usc; | ||
| 934 | |||
| 935 | /** | ||
| 936 | * @rsrc_spec: Resource specifier for the program. This is a packed | ||
| 937 | * hardware data structure describing the required number of registers, | ||
| 938 | * uniforms, bound textures, and bound samplers. | ||
| 939 | */ | ||
| 940 | __u32 rsrc_spec; | ||
| 941 | }; | ||
| 942 | |||
| 943 | /** | ||
| 944 | * struct drm_asahi_cmd_render - Command to submit 3D | ||
| 945 | * | ||
| 946 | * This command submits a single render pass. The hardware control stream may | ||
| 947 | * include many draws and subpasses, but within the command, the framebuffer | ||
| 948 | * dimensions and attachments are fixed. | ||
| 949 | * | ||
| 950 | * The hardware requires the firmware to set a large number of Control Registers | ||
| 951 | * setting up state at render pass granularity before each command rendering 3D. | ||
| 952 | * The firmware bundles this state into data structures. Unfortunately, we | ||
| 953 | * cannot expose either any of that directly to userspace, because the | ||
| 954 | * kernel-firmware ABI is not stable. Although we can guarantee the firmware | ||
| 955 | * updates in tandem with the kernel, we cannot break old userspace when | ||
| 956 | * upgrading the firmware and kernel. Therefore, we need to abstract well the | ||
| 957 | * data structures to avoid tying our hands with future firmwares. | ||
| 958 | * | ||
| 959 | * The bulk of drm_asahi_cmd_render therefore consists of values of hardware | ||
| 960 | * control registers, marshalled via the firmware interface. | ||
| 961 | * | ||
| 962 | * The framebuffer/tilebuffer dimensions are also specified here. In addition to | ||
| 963 | * being passed to the firmware/hardware, the kernel requires these dimensions | ||
| 964 | * to calculate various essential tiling-related data structures. It is | ||
| 965 | * unfortunate that our submits are heavier than on vendors with saner | ||
| 966 | * hardware-software interfaces. The upshot is all of this information is | ||
| 967 | * readily available to userspace with all current APIs. | ||
| 968 | * | ||
| 969 | * It looks odd - but it's not overly burdensome and it ensures we can remain | ||
| 970 | * compatible with old userspace. | ||
| 971 | */ | ||
| 972 | struct drm_asahi_cmd_render { | ||
| 973 | /** @flags: Combination of drm_asahi_render_flags flags. */ | ||
| 974 | __u32 flags; | ||
| 975 | |||
| 976 | /** | ||
| 977 | * @isp_zls_pixels: ISP_ZLS_PIXELS register value. This contains the | ||
| 978 | * depth/stencil width/height, which may differ from the framebuffer | ||
| 979 | * width/height. | ||
| 980 | */ | ||
| 981 | __u32 isp_zls_pixels; | ||
| 982 | |||
| 983 | /** | ||
| 984 | * @vdm_ctrl_stream_base: VDM_CTRL_STREAM_BASE register value. GPU | ||
| 985 | * address to the beginning of the VDM control stream. | ||
| 986 | */ | ||
| 987 | __u64 vdm_ctrl_stream_base; | ||
| 988 | |||
| 989 | /** @vertex_helper: Helper program used for the vertex shader */ | ||
| 990 | struct drm_asahi_helper_program vertex_helper; | ||
| 991 | |||
| 992 | /** @fragment_helper: Helper program used for the fragment shader */ | ||
| 993 | struct drm_asahi_helper_program fragment_helper; | ||
| 994 | |||
| 995 | /** | ||
| 996 | * @isp_scissor_base: ISP_SCISSOR_BASE register value. GPU address of an | ||
| 997 | * array of scissor descriptors indexed in the render pass. | ||
| 998 | */ | ||
| 999 | __u64 isp_scissor_base; | ||
| 1000 | |||
| 1001 | /** | ||
| 1002 | * @isp_dbias_base: ISP_DBIAS_BASE register value. GPU address of an | ||
| 1003 | * array of depth bias values indexed in the render pass. | ||
| 1004 | */ | ||
| 1005 | __u64 isp_dbias_base; | ||
| 1006 | |||
| 1007 | /** | ||
| 1008 | * @isp_oclqry_base: ISP_OCLQRY_BASE register value. GPU address of an | ||
| 1009 | * array of occlusion query results written by the render pass. | ||
| 1010 | */ | ||
| 1011 | __u64 isp_oclqry_base; | ||
| 1012 | |||
| 1013 | /** @depth: Depth buffer */ | ||
| 1014 | struct drm_asahi_zls_buffer depth; | ||
| 1015 | |||
| 1016 | /** @stencil: Stencil buffer */ | ||
| 1017 | struct drm_asahi_zls_buffer stencil; | ||
| 1018 | |||
| 1019 | /** @zls_ctrl: ZLS_CTRL register value */ | ||
| 1020 | __u64 zls_ctrl; | ||
| 1021 | |||
| 1022 | /** @ppp_multisamplectl: PPP_MULTISAMPLECTL register value */ | ||
| 1023 | __u64 ppp_multisamplectl; | ||
| 1024 | |||
| 1025 | /** | ||
| 1026 | * @sampler_heap: Base address of the sampler heap. This heap is used | ||
| 1027 | * for both vertex shaders and fragment shaders. The registers are | ||
| 1028 | * per-stage, but there is no known use case for separate heaps. | ||
| 1029 | */ | ||
| 1030 | __u64 sampler_heap; | ||
| 1031 | |||
| 1032 | /** @ppp_ctrl: PPP_CTRL register value */ | ||
| 1033 | __u32 ppp_ctrl; | ||
| 1034 | |||
| 1035 | /** @width_px: Framebuffer width in pixels */ | ||
| 1036 | __u16 width_px; | ||
| 1037 | |||
| 1038 | /** @height_px: Framebuffer height in pixels */ | ||
| 1039 | __u16 height_px; | ||
| 1040 | |||
| 1041 | /** @layers: Number of layers in the framebuffer */ | ||
| 1042 | __u16 layers; | ||
| 1043 | |||
| 1044 | /** @sampler_count: Number of samplers in the sampler heap. */ | ||
| 1045 | __u16 sampler_count; | ||
| 1046 | |||
| 1047 | /** @utile_width_px: Width of a logical tilebuffer tile in pixels */ | ||
| 1048 | __u8 utile_width_px; | ||
| 1049 | |||
| 1050 | /** @utile_height_px: Height of a logical tilebuffer tile in pixels */ | ||
| 1051 | __u8 utile_height_px; | ||
| 1052 | |||
| 1053 | /** @samples: # of samples in the framebuffer. Must be 1, 2, or 4. */ | ||
| 1054 | __u8 samples; | ||
| 1055 | |||
| 1056 | /** @sample_size_B: # of bytes in the tilebuffer required per sample. */ | ||
| 1057 | __u8 sample_size_B; | ||
| 1058 | |||
| 1059 | /** | ||
| 1060 | * @isp_merge_upper_x: 32-bit float used in the hardware triangle | ||
| 1061 | * merging. Calculate as: tan(60 deg) * width. | ||
| 1062 | * | ||
| 1063 | * Making these values UAPI avoids requiring floating-point calculations | ||
| 1064 | * in the kernel in the hot path. | ||
| 1065 | */ | ||
| 1066 | __u32 isp_merge_upper_x; | ||
| 1067 | |||
| 1068 | /** | ||
| 1069 | * @isp_merge_upper_y: 32-bit float. Calculate as: tan(60 deg) * height. | ||
| 1070 | * See @isp_merge_upper_x. | ||
| 1071 | */ | ||
| 1072 | __u32 isp_merge_upper_y; | ||
| 1073 | |||
| 1074 | /** @bg: Background program run for each tile at the start */ | ||
| 1075 | struct drm_asahi_bg_eot bg; | ||
| 1076 | |||
| 1077 | /** @eot: End-of-tile program ran for each tile at the end */ | ||
| 1078 | struct drm_asahi_bg_eot eot; | ||
| 1079 | |||
| 1080 | /** | ||
| 1081 | * @partial_bg: Background program ran at the start of each tile when | ||
| 1082 | * resuming the render pass during a partial render. | ||
| 1083 | */ | ||
| 1084 | struct drm_asahi_bg_eot partial_bg; | ||
| 1085 | |||
| 1086 | /** | ||
| 1087 | * @partial_eot: End-of-tile program ran at the end of each tile when | ||
| 1088 | * pausing the render pass during a partial render. | ||
| 1089 | */ | ||
| 1090 | struct drm_asahi_bg_eot partial_eot; | ||
| 1091 | |||
| 1092 | /** | ||
| 1093 | * @isp_bgobjdepth: ISP_BGOBJDEPTH register value. This is the depth | ||
| 1094 | * buffer clear value, encoded in the depth buffer's format: either a | ||
| 1095 | * 32-bit float or a 16-bit unorm (with upper bits zeroed). | ||
| 1096 | */ | ||
| 1097 | __u32 isp_bgobjdepth; | ||
| 1098 | |||
| 1099 | /** | ||
| 1100 | * @isp_bgobjvals: ISP_BGOBJVALS register value. The bottom 8-bits | ||
| 1101 | * contain the stencil buffer clear value. | ||
| 1102 | */ | ||
| 1103 | __u32 isp_bgobjvals; | ||
| 1104 | |||
| 1105 | /** @ts_vtx: Timestamps for the vertex portion of the render */ | ||
| 1106 | struct drm_asahi_timestamps ts_vtx; | ||
| 1107 | |||
| 1108 | /** @ts_frag: Timestamps for the fragment portion of the render */ | ||
| 1109 | struct drm_asahi_timestamps ts_frag; | ||
| 1110 | }; | ||
| 1111 | |||
| 1112 | /** | ||
| 1113 | * struct drm_asahi_cmd_compute - Command to submit compute | ||
| 1114 | * | ||
| 1115 | * This command submits a control stream consisting of compute dispatches. There | ||
| 1116 | * is essentially no limit on how many compute dispatches may be included in a | ||
| 1117 | * single compute command, although timestamps are at command granularity. | ||
| 1118 | */ | ||
| 1119 | struct drm_asahi_cmd_compute { | ||
| 1120 | /** @flags: MBZ */ | ||
| 1121 | __u32 flags; | ||
| 1122 | |||
| 1123 | /** @sampler_count: Number of samplers in the sampler heap. */ | ||
| 1124 | __u32 sampler_count; | ||
| 1125 | |||
| 1126 | /** | ||
| 1127 | * @cdm_ctrl_stream_base: CDM_CTRL_STREAM_BASE register value. GPU | ||
| 1128 | * address to the beginning of the CDM control stream. | ||
| 1129 | */ | ||
| 1130 | __u64 cdm_ctrl_stream_base; | ||
| 1131 | |||
| 1132 | /** | ||
| 1133 | * @cdm_ctrl_stream_end: GPU base address to the end of the hardware | ||
| 1134 | * control stream. Note this only considers the first contiguous segment | ||
| 1135 | * of the control stream, as the stream might jump elsewhere. | ||
| 1136 | */ | ||
| 1137 | __u64 cdm_ctrl_stream_end; | ||
| 1138 | |||
| 1139 | /** @sampler_heap: Base address of the sampler heap. */ | ||
| 1140 | __u64 sampler_heap; | ||
| 1141 | |||
| 1142 | /** @helper: Helper program used for this compute command */ | ||
| 1143 | struct drm_asahi_helper_program helper; | ||
| 1144 | |||
| 1145 | /** @ts: Timestamps for the compute command */ | ||
| 1146 | struct drm_asahi_timestamps ts; | ||
| 1147 | }; | ||
| 1148 | |||
| 1149 | /** | ||
| 1150 | * struct drm_asahi_get_time - Arguments passed to DRM_IOCTL_ASAHI_GET_TIME | ||
| 1151 | */ | ||
| 1152 | struct drm_asahi_get_time { | ||
| 1153 | /** @flags: MBZ. */ | ||
| 1154 | __u64 flags; | ||
| 1155 | |||
| 1156 | /** @gpu_timestamp: On return, the GPU timestamp in nanoseconds. */ | ||
| 1157 | __u64 gpu_timestamp; | ||
| 1158 | }; | ||
| 1159 | |||
| 1160 | /** | ||
| 1161 | * DRM_IOCTL_ASAHI() - Build an Asahi IOCTL number | ||
| 1162 | * @__access: Access type. Must be R, W or RW. | ||
| 1163 | * @__id: One of the DRM_ASAHI_xxx id. | ||
| 1164 | * @__type: Suffix of the type being passed to the IOCTL. | ||
| 1165 | * | ||
| 1166 | * Don't use this macro directly, use the DRM_IOCTL_ASAHI_xxx | ||
| 1167 | * values instead. | ||
| 1168 | * | ||
| 1169 | * Return: An IOCTL number to be passed to ioctl() from userspace. | ||
| 1170 | */ | ||
| 1171 | #define DRM_IOCTL_ASAHI(__access, __id, __type) \ | ||
| 1172 | DRM_IO ## __access(DRM_COMMAND_BASE + DRM_ASAHI_ ## __id, \ | ||
| 1173 | struct drm_asahi_ ## __type) | ||
| 1174 | |||
| 1175 | /* Note: this is an enum so that it can be resolved by Rust bindgen. */ | ||
| 1176 | enum { | ||
| 1177 | DRM_IOCTL_ASAHI_GET_PARAMS = DRM_IOCTL_ASAHI(W, GET_PARAMS, get_params), | ||
| 1178 | DRM_IOCTL_ASAHI_GET_TIME = DRM_IOCTL_ASAHI(WR, GET_TIME, get_time), | ||
| 1179 | DRM_IOCTL_ASAHI_VM_CREATE = DRM_IOCTL_ASAHI(WR, VM_CREATE, vm_create), | ||
| 1180 | DRM_IOCTL_ASAHI_VM_DESTROY = DRM_IOCTL_ASAHI(W, VM_DESTROY, vm_destroy), | ||
| 1181 | DRM_IOCTL_ASAHI_VM_BIND = DRM_IOCTL_ASAHI(W, VM_BIND, vm_bind), | ||
| 1182 | DRM_IOCTL_ASAHI_GEM_CREATE = DRM_IOCTL_ASAHI(WR, GEM_CREATE, gem_create), | ||
| 1183 | DRM_IOCTL_ASAHI_GEM_MMAP_OFFSET = DRM_IOCTL_ASAHI(WR, GEM_MMAP_OFFSET, gem_mmap_offset), | ||
| 1184 | DRM_IOCTL_ASAHI_GEM_BIND_OBJECT = DRM_IOCTL_ASAHI(WR, GEM_BIND_OBJECT, gem_bind_object), | ||
| 1185 | DRM_IOCTL_ASAHI_QUEUE_CREATE = DRM_IOCTL_ASAHI(WR, QUEUE_CREATE, queue_create), | ||
| 1186 | DRM_IOCTL_ASAHI_QUEUE_DESTROY = DRM_IOCTL_ASAHI(W, QUEUE_DESTROY, queue_destroy), | ||
| 1187 | DRM_IOCTL_ASAHI_SUBMIT = DRM_IOCTL_ASAHI(W, SUBMIT, submit), | ||
| 1188 | }; | ||
| 1189 | |||
| 1190 | #if defined(__cplusplus) | ||
| 1191 | } | ||
| 1192 | #endif | ||
| 1193 | |||
| 1194 | #endif /* _ASAHI_DRM_H_ */ | ||