tenflowers-core 0.1.1

Core tensor operations and execution engine for TenfloweRS
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252

// Logical operation compute shaders
// Operates on boolean tensors represented as u32 arrays

@group(0) @binding(0) var<storage, read> input_a: array<u32>; // Boolean inputs stored as u32
@group(0) @binding(1) var<storage, read> input_b: array<u32>;
@group(0) @binding(2) var<storage, read_write> output: array<u32>; // Boolean output

// Broadcasting metadata
@group(0) @binding(3) var<storage, read> shape_metadata: array<u32>;

// Helper functions for broadcasting
fn compute_flat_index_broadcast(indices: ptr<function, array<u32, 8>>, shape_offset: u32, rank: u32) -> u32 {
    var flat_idx = 0u;
    var stride = 1u;
    
    for (var i = 0u; i < rank; i++) {
        let dim_idx = rank - 1u - i;
        flat_idx += (*indices)[dim_idx] * stride;
        stride *= shape_metadata[shape_offset + dim_idx];
    }
    
    return flat_idx;
}

fn compute_multi_index_broadcast(flat_idx: u32, indices: ptr<function, array<u32, 8>>, shape_offset: u32, rank: u32) {
    var remaining = flat_idx;
    
    for (var i = 0u; i < rank; i++) {
        let dim_idx = rank - 1u - i;
        (*indices)[dim_idx] = remaining % shape_metadata[shape_offset + dim_idx];
        remaining = remaining / shape_metadata[shape_offset + dim_idx];
    }
}

fn broadcast_indices(output_indices: ptr<function, array<u32, 8>>, input_indices: ptr<function, array<u32, 8>>, 
                     input_shape_offset: u32, input_rank: u32, output_rank: u32) {
    for (var i = 0u; i < 8u; i++) {
        (*input_indices)[i] = 0u;
    }
    
    let rank_diff = output_rank - input_rank;
    for (var i = 0u; i < input_rank; i++) {
        let input_dim_idx = i;
        let output_dim_idx = i + rank_diff;
        let input_size = shape_metadata[input_shape_offset + input_dim_idx];
        
        if (input_size == 1u) {
            (*input_indices)[input_dim_idx] = 0u;
        } else {
            (*input_indices)[input_dim_idx] = (*output_indices)[output_dim_idx];
        }
    }
}

// Basic logical operations

@compute @workgroup_size(64)
fn and_op(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let index = global_id.x;
    if (index >= arrayLength(&output)) {
        return;
    }
    
    let a_idx = index % arrayLength(&input_a);
    let b_idx = index % arrayLength(&input_b);
    
    // Convert to boolean, perform AND, convert back to u32
    let a_bool = input_a[a_idx] != 0u;
    let b_bool = input_b[b_idx] != 0u;
    output[index] = select(0u, 1u, a_bool && b_bool);
}

@compute @workgroup_size(64)
fn or_op(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let index = global_id.x;
    if (index >= arrayLength(&output)) {
        return;
    }
    
    let a_idx = index % arrayLength(&input_a);
    let b_idx = index % arrayLength(&input_b);
    
    let a_bool = input_a[a_idx] != 0u;
    let b_bool = input_b[b_idx] != 0u;
    output[index] = select(0u, 1u, a_bool || b_bool);
}

@compute @workgroup_size(64)
fn xor_op(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let index = global_id.x;
    if (index >= arrayLength(&output)) {
        return;
    }
    
    let a_idx = index % arrayLength(&input_a);
    let b_idx = index % arrayLength(&input_b);
    
    let a_bool = input_a[a_idx] != 0u;
    let b_bool = input_b[b_idx] != 0u;
    output[index] = select(0u, 1u, a_bool != b_bool);
}

// Unary logical operation (NOT)
@compute @workgroup_size(64)
fn not_op(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let index = global_id.x;
    if (index >= arrayLength(&output)) {
        return;
    }
    
    let a_idx = index % arrayLength(&input_a);
    let a_bool = input_a[a_idx] != 0u;
    output[index] = select(1u, 0u, a_bool); // NOT operation
}

// Broadcasting versions

@compute @workgroup_size(64)
fn and_broadcast(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let output_idx = global_id.x;
    if (output_idx >= arrayLength(&output)) {
        return;
    }
    
    let a_rank = shape_metadata[0];
    let b_rank = shape_metadata[1];
    let output_rank = shape_metadata[2];
    
    let a_shape_offset = 3u;
    let b_shape_offset = 3u + a_rank;
    let output_shape_offset = 3u + a_rank + b_rank;
    
    var output_indices: array<u32, 8>;
    compute_multi_index_broadcast(output_idx, &output_indices, output_shape_offset, output_rank);
    
    var a_indices: array<u32, 8>;
    var b_indices: array<u32, 8>;
    broadcast_indices(&output_indices, &a_indices, a_shape_offset, a_rank, output_rank);
    broadcast_indices(&output_indices, &b_indices, b_shape_offset, b_rank, output_rank);
    
    let a_flat_idx = compute_flat_index_broadcast(&a_indices, a_shape_offset, a_rank);
    let b_flat_idx = compute_flat_index_broadcast(&b_indices, b_shape_offset, b_rank);
    
    let a_bool = input_a[a_flat_idx] != 0u;
    let b_bool = input_b[b_flat_idx] != 0u;
    output[output_idx] = select(0u, 1u, a_bool && b_bool);
}

@compute @workgroup_size(64)
fn or_broadcast(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let output_idx = global_id.x;
    if (output_idx >= arrayLength(&output)) {
        return;
    }
    
    let a_rank = shape_metadata[0];
    let b_rank = shape_metadata[1];
    let output_rank = shape_metadata[2];
    
    let a_shape_offset = 3u;
    let b_shape_offset = 3u + a_rank;
    let output_shape_offset = 3u + a_rank + b_rank;
    
    var output_indices: array<u32, 8>;
    compute_multi_index_broadcast(output_idx, &output_indices, output_shape_offset, output_rank);
    
    var a_indices: array<u32, 8>;
    var b_indices: array<u32, 8>;
    broadcast_indices(&output_indices, &a_indices, a_shape_offset, a_rank, output_rank);
    broadcast_indices(&output_indices, &b_indices, b_shape_offset, b_rank, output_rank);
    
    let a_flat_idx = compute_flat_index_broadcast(&a_indices, a_shape_offset, a_rank);
    let b_flat_idx = compute_flat_index_broadcast(&b_indices, b_shape_offset, b_rank);
    
    let a_bool = input_a[a_flat_idx] != 0u;
    let b_bool = input_b[b_flat_idx] != 0u;
    output[output_idx] = select(0u, 1u, a_bool || b_bool);
}

@compute @workgroup_size(64)
fn xor_broadcast(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let output_idx = global_id.x;
    if (output_idx >= arrayLength(&output)) {
        return;
    }
    
    let a_rank = shape_metadata[0];
    let b_rank = shape_metadata[1];
    let output_rank = shape_metadata[2];
    
    let a_shape_offset = 3u;
    let b_shape_offset = 3u + a_rank;
    let output_shape_offset = 3u + a_rank + b_rank;
    
    var output_indices: array<u32, 8>;
    compute_multi_index_broadcast(output_idx, &output_indices, output_shape_offset, output_rank);
    
    var a_indices: array<u32, 8>;
    var b_indices: array<u32, 8>;
    broadcast_indices(&output_indices, &a_indices, a_shape_offset, a_rank, output_rank);
    broadcast_indices(&output_indices, &b_indices, b_shape_offset, b_rank, output_rank);
    
    let a_flat_idx = compute_flat_index_broadcast(&a_indices, a_shape_offset, a_rank);
    let b_flat_idx = compute_flat_index_broadcast(&b_indices, b_shape_offset, b_rank);
    
    let a_bool = input_a[a_flat_idx] != 0u;
    let b_bool = input_b[b_flat_idx] != 0u;
    output[output_idx] = select(0u, 1u, a_bool != b_bool);
}

// Short-circuit optimized operations
// These operations can potentially skip computation when one operand guarantees the result

@compute @workgroup_size(64)
fn and_short_circuit(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let index = global_id.x;
    if (index >= arrayLength(&output)) {
        return;
    }
    
    let a_idx = index % arrayLength(&input_a);
    let b_idx = index % arrayLength(&input_b);
    
    // Early exit if first operand is false
    if (input_a[a_idx] == 0u) {
        output[index] = 0u;
        return;
    }
    
    // Check second operand
    output[index] = select(0u, 1u, input_b[b_idx] != 0u);
}

@compute @workgroup_size(64)
fn or_short_circuit(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let index = global_id.x;
    if (index >= arrayLength(&output)) {
        return;
    }
    
    let a_idx = index % arrayLength(&input_a);
    let b_idx = index % arrayLength(&input_b);
    
    // Early exit if first operand is true
    if (input_a[a_idx] != 0u) {
        output[index] = 1u;
        return;
    }
    
    // Check second operand
    output[index] = select(0u, 1u, input_b[b_idx] != 0u);
}