MayaFlux 1.0.0
Digital-First Multimedia Processing Framework
Loading...
Searching...
No Matches
Layouts.hpp
Go to the documentation of this file.
1#pragma once
2
3/**
4 * @file Layouts.hpp
5 * @brief The push constant layouts of the tensor kernels, their flag bits and limits.
6 *
7 * This is the interface of the kernels: a layout does not change once
8 * released. A kernel that needs different constants is a new kernel with its
9 * own struct, and fields are added only by taking a reserved word.
10 */
11
13
14/** @brief Invocations per workgroup of the element-wise kernels. */
15inline constexpr uint32_t element_workgroup = 64;
16/** @brief Invocations of the one workgroup nn_lstm_sequence runs in. */
17inline constexpr uint32_t lstm_workgroup = 512;
18/** @brief Widest hidden state nn_lstm_sequence supports. */
19inline constexpr uint32_t lstm_max_hidden = 256;
20/** @brief Most workgroups the stages put on one axis; a larger dispatch continues along y. */
21inline constexpr uint32_t max_groups_per_axis = 32768;
22
23inline constexpr const char* frame_audio_shader = "nn_frame_audio.comp.spv";
24inline constexpr const char* conv1d_shader = "nn_conv1d.comp.spv";
25inline constexpr const char* magnitude_shader = "nn_magnitude.comp.spv";
26inline constexpr const char* lstm_sequence_shader = "nn_lstm_sequence.comp.spv";
27inline constexpr const char* grouped_conv1d_shader = "nn_conv1d_grouped.comp.spv";
28inline constexpr const char* normalize_shader = "nn_normalize.comp.spv";
29inline constexpr const char* add_shader = "nn_add.comp.spv";
30inline constexpr const char* attention_scores_shader = "nn_attention_scores.comp.spv";
31inline constexpr const char* softmax_shader = "nn_softmax.comp.spv";
32inline constexpr const char* attention_mix_shader = "nn_attention_mix.comp.spv";
33
34/** @brief nn_conv1d flag bit: add the bias. */
35inline constexpr uint32_t conv1d_flag_bias = 1;
36/** @brief nn_conv1d flag bit: apply ReLU to the input. */
37inline constexpr uint32_t conv1d_flag_relu_input = 2;
38/** @brief nn_conv1d and nn_conv1d_grouped flag bits 2 and 3: the output activation, 0 none, 1 ReLU, 2 sigmoid, 3 GELU. */
39inline constexpr uint32_t conv1d_activation_shift = 2;
40
41/** @brief nn_normalize flag bit: write the normalised values instead of measuring the rows. */
42inline constexpr uint32_t normalize_flag_apply = 1;
43/** @brief nn_normalize flag bit: multiply by weight and add bias. */
44inline constexpr uint32_t normalize_flag_affine = 2;
45/** @brief nn_normalize flag bit: weight and bias are indexed by the outer index instead of the position along extent. */
46inline constexpr uint32_t normalize_flag_affine_outer = 4;
47/** @brief nn_normalize flag bits 3 and 4: the output activation, 0 none, 3 GELU. */
48inline constexpr uint32_t normalize_activation_shift = 3;
49
50/** @brief Push constants of nn_conv1d. */
51struct Conv1dPC {
52 uint64_t src;
53 uint64_t weights;
54 uint64_t bias;
55 uint64_t dst;
56 uint32_t batch;
57 uint32_t c_in;
58 uint32_t t_in;
59 uint32_t c_out;
60 uint32_t t_out;
61 uint32_t kernel_size;
62 uint32_t stride;
63 uint32_t pad;
64 uint32_t flags;
65 uint32_t dilation;
66};
67
68/** @brief Push constants of nn_frame_audio. */
70 uint64_t samples;
71 uint64_t frames;
72 uint32_t chunks;
73 uint32_t chunk_size;
74 uint32_t context;
75 uint32_t pad;
76};
77
78/** @brief Push constants of nn_magnitude. */
80 uint64_t src;
81 uint64_t dst;
82 uint32_t batch;
83 uint32_t channels;
84 uint32_t time;
85 uint32_t reserved;
86};
87
88/** @brief Push constants of nn_lstm_sequence. */
89struct LstmPC {
90 uint64_t x;
91 uint64_t w_ih;
92 uint64_t w_hh;
93 uint64_t b_ih;
94 uint64_t b_hh;
95 uint64_t state;
96 uint64_t hs;
97 uint32_t steps;
98 uint32_t hidden;
99};
100
101/** @brief Push constants of nn_conv1d_grouped. */
103 uint64_t src;
104 uint64_t weights;
105 uint64_t bias;
106 uint64_t dst;
107 uint32_t batch;
108 uint32_t c_in;
109 uint32_t t_in;
110 uint32_t c_out;
111 uint32_t t_out;
112 uint32_t kernel_size;
113 uint32_t stride;
114 uint32_t pad;
115 uint32_t flags;
116 uint32_t groups;
117};
118
119/** @brief Push constants of nn_normalize. */
121 uint64_t src;
122 uint64_t stats;
123 uint64_t weight;
124 uint64_t bias;
125 uint64_t dst;
126 uint32_t outer;
127 uint32_t extent;
128 uint32_t inner;
129 float epsilon;
130 uint32_t flags;
131 uint32_t reserved;
132};
133
134/** @brief Push constants of nn_add. */
135struct AddPC {
136 uint64_t a;
137 uint64_t b;
138 uint64_t dst;
139 uint32_t count;
140 uint32_t reserved;
141};
142
143/** @brief Push constants of nn_attention_scores. */
145 uint64_t q;
146 uint64_t k;
147 uint64_t scores;
148 uint32_t heads;
149 uint32_t dim;
150 uint32_t time;
151 float scale;
152};
153
154/** @brief Push constants of nn_softmax. */
155struct SoftmaxPC {
156 uint64_t data;
157 uint32_t outer;
158 uint32_t extent;
159 uint32_t inner;
160 uint32_t reserved;
161};
162
163/** @brief Push constants of nn_attention_mix. */
165 uint64_t weights;
166 uint64_t v;
167 uint64_t dst;
168 uint32_t heads;
169 uint32_t dim;
170 uint32_t time;
171 uint32_t reserved;
172};
173
174static_assert(sizeof(Conv1dPC) == 72);
175static_assert(sizeof(GroupedConv1dPC) == 72);
176static_assert(sizeof(NormalizePC) == 64);
177static_assert(sizeof(AddPC) == 32);
178static_assert(sizeof(AttentionScoresPC) == 40);
179static_assert(sizeof(SoftmaxPC) == 24);
180static_assert(sizeof(AttentionMixPC) == 40);
181static_assert(sizeof(FrameAudioPC) == 32);
182static_assert(sizeof(MagnitudePC) == 32);
183static_assert(sizeof(LstmPC) == 64);
184
185} // namespace MayaFlux::Yantra::TensorKernels
constexpr uint32_t normalize_flag_affine_outer
nn_normalize flag bit: weight and bias are indexed by the outer index instead of the position along e...
Definition Layouts.hpp:46
constexpr uint32_t normalize_flag_affine
nn_normalize flag bit: multiply by weight and add bias.
Definition Layouts.hpp:44
constexpr uint32_t conv1d_activation_shift
nn_conv1d and nn_conv1d_grouped flag bits 2 and 3: the output activation, 0 none, 1 ReLU,...
Definition Layouts.hpp:39
constexpr uint32_t max_groups_per_axis
Most workgroups the stages put on one axis; a larger dispatch continues along y.
Definition Layouts.hpp:21
constexpr const char * attention_mix_shader
Definition Layouts.hpp:32
constexpr const char * conv1d_shader
Definition Layouts.hpp:24
constexpr const char * softmax_shader
Definition Layouts.hpp:31
constexpr const char * frame_audio_shader
Definition Layouts.hpp:23
constexpr const char * normalize_shader
Definition Layouts.hpp:28
constexpr const char * grouped_conv1d_shader
Definition Layouts.hpp:27
constexpr uint32_t element_workgroup
Invocations per workgroup of the element-wise kernels.
Definition Layouts.hpp:15
constexpr uint32_t normalize_flag_apply
nn_normalize flag bit: write the normalised values instead of measuring the rows.
Definition Layouts.hpp:42
constexpr uint32_t conv1d_flag_bias
nn_conv1d flag bit: add the bias.
Definition Layouts.hpp:35
constexpr uint32_t normalize_activation_shift
nn_normalize flag bits 3 and 4: the output activation, 0 none, 3 GELU.
Definition Layouts.hpp:48
constexpr const char * magnitude_shader
Definition Layouts.hpp:25
constexpr const char * add_shader
Definition Layouts.hpp:29
constexpr const char * attention_scores_shader
Definition Layouts.hpp:30
constexpr uint32_t lstm_max_hidden
Widest hidden state nn_lstm_sequence supports.
Definition Layouts.hpp:19
constexpr uint32_t lstm_workgroup
Invocations of the one workgroup nn_lstm_sequence runs in.
Definition Layouts.hpp:17
constexpr const char * lstm_sequence_shader
Definition Layouts.hpp:26
constexpr uint32_t conv1d_flag_relu_input
nn_conv1d flag bit: apply ReLU to the input.
Definition Layouts.hpp:37
Push constants of nn_add.
Definition Layouts.hpp:135
Push constants of nn_attention_mix.
Definition Layouts.hpp:164
Push constants of nn_attention_scores.
Definition Layouts.hpp:144
Push constants of nn_conv1d.
Definition Layouts.hpp:51
Push constants of nn_frame_audio.
Definition Layouts.hpp:69
Push constants of nn_conv1d_grouped.
Definition Layouts.hpp:102
Push constants of nn_lstm_sequence.
Definition Layouts.hpp:89
Push constants of nn_magnitude.
Definition Layouts.hpp:79
Push constants of nn_normalize.
Definition Layouts.hpp:120
Push constants of nn_softmax.
Definition Layouts.hpp:155