Blender V5.3
metal/device_impl.h
Go to the documentation of this file.
1/* SPDX-FileCopyrightText: 2021-2022 Blender Foundation
2 *
3 * SPDX-License-Identifier: Apache-2.0 */
4
5#pragma once
6
7#ifdef WITH_METAL
8
9# include "bvh/bvh.h"
10# include "device/device.h"
11# include "device/metal/bvh.h"
12# include "device/metal/device.h"
13# include "device/metal/kernel.h"
14# include "device/metal/queue.h"
15# include "device/metal/util.h"
16
17# include <Metal/Metal.h>
18
20
21class DeviceQueue;
22
23class MetalDevice : public Device {
24 public:
25 id<MTLDevice> mtlDevice = nil;
26 id<MTLLibrary> mtlLibrary[PSO_NUM] = {nil};
27 id<MTLCommandQueue> mtlComputeCommandQueue = nil;
28 id<MTLCommandQueue> mtlGeneralCommandQueue = nil;
29 id<MTLCounterSampleBuffer> mtlCounterSampleBuffer = nil;
30 string source[PSO_NUM];
31 string kernels_md5[PSO_NUM];
32 string global_defines_md5[PSO_NUM];
33
34 bool capture_enabled = false;
35
36 /* Argument buffer for static data. */
37 id<MTLBuffer> launch_params_buffer = nil;
38 KernelParamsMetal *launch_params = nullptr;
39
40 /* MetalRT members ---------------------------------- */
41 bool use_metalrt = false;
42 bool use_metalrt_extended_limits = false;
43 bool motion_blur = false;
44 bool use_pcmi = false;
45
46 id<MTLBuffer> blas_buffer = nil;
47
48 API_AVAILABLE(macos(11.0))
49 vector<id<MTLAccelerationStructure>> unique_blas_array;
50
51 API_AVAILABLE(macos(11.0))
52 vector<id<MTLAccelerationStructure>> blas_array;
53
54 API_AVAILABLE(macos(11.0))
55 id<MTLAccelerationStructure> accel_struct = nil;
56
57 /* Residency sets -----------------------------------*/
58 void prepare_residency();
59 void metal_mem_alloc(id<MTLResource> allocation);
60 void metal_mem_free(id<MTLResource> allocation);
61
62 /* For externally-owned resources (e.g. graphics interop buffers) which need to be resident for
63 * kernels to access them, but shouldn't be included in our stats. */
64 void add_to_residency_set(id<MTLResource> allocation);
65 void remove_from_residency_set(id<MTLResource> allocation);
66
67 bool mtlResidencySet_enabled = false;
68# if defined(MAC_OS_VERSION_15_0)
69 API_AVAILABLE(macos(15.0), ios(18.0))
70 id<MTLResidencySet> mtlResidencySet = nil;
71 bool mtlResidencySet_dirty = false;
72 /* Guards mtlResidencySet mutations (may be reached from multiple threads). */
73 std::mutex mtlResidencySet_mutex;
74# endif
75
76 uint64_t kernel_features = 0;
77 bool using_nanovdb = false;
78 int max_threads_per_threadgroup;
79
80 int mtlDevId = 0;
81 bool has_error = false;
82
83 struct MetalMem {
84 device_memory *mem = nullptr;
85 int pointer_index = -1;
86 id<MTLBuffer> mtlBuffer = nil;
87 id<MTLTexture> mtlTexture = nil;
88 uint64_t offset = 0;
89 uint64_t size = 0;
90 void *hostPtr = nullptr;
91 };
92 using MetalMemMap = map<device_memory *, unique_ptr<MetalMem>>;
93 MetalMemMap metal_mem_map;
94 std::vector<id<MTLResource>> delayed_free_list;
95 std::recursive_mutex metal_mem_map_mutex;
96
97 /* Bindless Textures */
98 bool is_texture(const KernelImageInfo &info);
99 device_vector<KernelImageInfo> image_info;
100 id<MTLBuffer> image_bindings = nil;
101 std::vector<id<MTLResource>> image_info_id_map;
102
103 MetalPipelineType kernel_specialization_level = PSO_GENERIC;
104
105 int device_id = 0;
106
107 static thread_mutex existing_devices_mutex;
108 static std::map<int, MetalDevice *> active_device_ids;
109
110 static bool is_device_cancelled(const int device_id);
111
112 static MetalDevice *get_device_by_ID(const int device_idID,
113 thread_scoped_lock &existing_devices_mutex_lock);
114
115 bool is_ready(string &status) const override;
116
117 void cancel() override;
118
119 BVHLayoutMask get_bvh_layout_mask(uint64_t kernel_features) const override;
120
121 void set_error(const string &error) override;
122
123 MetalDevice(const DeviceInfo &info, Stats &stats, Profiler &profiler, bool headless);
124
125 ~MetalDevice() override;
126
127 bool support_device(uint64_t kernel_features);
128
129 bool check_peer_access(Device *peer_device) override;
130
131 bool use_adaptive_compilation();
132
133 bool use_local_atomic_sort() const;
134
135 string preprocess_source(MetalPipelineType pso_type,
136 uint64_t kernel_features,
137 string *source = nullptr);
138
139 void refresh_source_and_kernels_md5(MetalPipelineType pso_type);
140
141 void make_source(MetalPipelineType pso_type, uint64_t kernel_features);
142
143 bool load_kernels(uint64_t kernel_features) override;
144
145 void load_image_info();
146
147 void erase_allocation(device_memory &mem);
148
149 bool should_use_graphics_interop(const GraphicsInteropDevice &interop_device,
150 const bool log) override;
151
152 void *get_native_buffer(device_ptr ptr) override;
153
154 unique_ptr<DeviceQueue> gpu_queue_create() override;
155
156 void build_bvh(BVH *bvh, Progress &progress, bool refit) override;
157
158 bool set_bvh_limits(size_t instance_count, size_t max_prim_count) override;
159
160 void optimize_for_scene(Scene *scene) override;
161
162 static void compile_and_load(const int device_id, MetalPipelineType pso_type);
163
164 /* ------------------------------------------------------------------ */
165 /* low-level memory management */
166
167 bool max_working_set_exceeded(const size_t safety_margin = 8 * 1024 * 1024) const;
168
169 MetalMem *generic_alloc(device_memory &mem);
170
171 void generic_copy_to(device_memory &mem);
172
173 void generic_free(device_memory &mem);
174
175 void mem_alloc(device_memory &mem) override;
176
177 void mem_copy_to(device_memory &mem) override;
178
179 void mem_move_to_host(device_memory &mem) override;
180
181 void mem_copy_from(device_memory &mem)
182 {
183 mem_copy_from(mem, -1, -1, -1, -1);
184 }
185 void mem_copy_from(
186 device_memory &mem, const size_t y, size_t w, const size_t h, size_t elem) override;
187
188 void mem_or_from_device(device_memory &mem) override;
189
190 void mem_zero(device_memory &mem) override;
191
192 void mem_free(device_memory &mem) override;
193
194 device_ptr mem_alloc_sub_ptr(device_memory &mem, const size_t offset, size_t /*size*/) override;
195
196 void const_copy_to(const char *name, void *host, const size_t size) override;
197
198 void global_alloc(device_memory &mem);
199 void global_free(device_memory &mem);
200
201 void image_alloc(device_image &mem);
202 void image_alloc_as_buffer(device_image &mem);
203 void image_copy_to(device_image &mem);
204 void image_free(device_image &mem);
205
206 void flush_delayed_free_list();
207
208 void free_bvh();
209
210 void update_bvh(BVHMetal *bvh_metal);
211};
212
214
215#endif
unsigned long long int uint64_t
void refit(btStridingMeshInterface *triangles, const btVector3 &aabbMin, const btVector3 &aabbMax)
SIMD_FORCE_INLINE const btScalar & w() const
Return the w value.
Definition btQuadWord.h:119
#define CCL_NAMESPACE_END
#define log
ccl_device bool BVH_FUNCTION_FULL_NAME BVH(KernelGlobals kg, const ccl_private Ray *ray, ccl_private LocalIntersection *local_isect, const int local_object, ccl_private uint *lcg_state, const int max_hits)
Definition local.h:27
int size(const AssetLibraryReference *library_reference)
int BVHLayoutMask
Definition params.h:50
std::mutex thread_mutex
Definition thread.h:27
std::unique_lock< std::mutex > thread_scoped_lock
Definition thread.h:28
uint64_t device_ptr
Definition types_base.h:44