MCPcopy Create free account
hub / github.com/DeepRec-AI/DeepRec / CompilationOptions

Class CompilationOptions

tensorflow/lite/delegates/gpu/gl/compiler_options.h:26–65  ·  view source on GitHub ↗

Default constructor for options turns on all optimizations.

Source from the content-addressed store, hash-verified

24
25// Default constructor for options turns on all optimizations.
26struct CompilationOptions {
27 // Allows to quantify tensors, downcast values, process in float16 etc.
28 bool allow_precision_loss = false;
29
30 // When set few operations are fused into a single shader. Therefore, there
31 // will be less shaders, but each shader will become larger.
32 bool fuse_operations = true;
33
34 // Parameters will be inlined into a shader. This in turn will generated more
35 // unique shaders where each will need to be compiled.
36 bool inline_parameters = false;
37
38 // If true, shaders, that have auto-input and auto-output, will use a single
39 // object for reading and writing.
40 bool inline_objects = true; // TODO(akulik): unsupported
41
42 // Can be only Textures or Buffers
43 ObjectType preferred_obj_type = ObjectType::UNKNOWN;
44 // User has an option to choose between textures and buffers. Textures work
45 // better on Adreno and buffers are better for Mali.
46
47 // Chooses object type to represent intermediate tensors. Buffers have more
48 // efficient memory usage because they represent opaque memory blob, but
49 // textures work better on Adreno.
50 // TODO(akulik): may be better name?
51 ObjectType ref_obj_type = ObjectType::UNKNOWN;
52
53 // If true, a user may change BATCH dimension at runtime. Otherwise, static
54 // batch size will be fixed during compile time.
55 // Dynamic mode uses less memory, while static mode may yield better
56 // performance for small models.
57 bool dynamic_batch = false;
58
59 // Fuses consequent nodes which have auto output and auto input.
60 bool auto_input_fusion = true;
61
62 // If true sampler2D and texelFetch will be used to access read only textures.
63 // This feature is not supported yet by the OpenGL runtime.
64 bool sampler_textures = false;
65};
66
67} // namespace gl
68} // namespace gpu

Callers 1

InvokeMethod · 0.85

Calls

no outgoing calls

Tested by 1

InvokeMethod · 0.68