numr 0.5.2

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
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

// Sparse merge count shaders - type-independent
//
// csr_merge_count:   Count output NNZ per row for CSR add/sub (union semantics)
// csr_mul_count:     Count output NNZ per row for CSR mul/div (intersection semantics)
// csc_merge_count:   Count output NNZ per col for CSC add/sub (union semantics)
// csc_mul_count:     Count output NNZ per col for CSC mul/div (intersection semantics)
// exclusive_scan_i32: Sequential exclusive prefix sum

const WORKGROUP_SIZE: u32 = 256u;

// ============================================================================
// csr_merge_count
// ============================================================================

struct CsrMergeCountParams {
    nrows: u32,
}

@group(0) @binding(0) var<storage, read> cmc_a_row_ptrs: array<i32>;
@group(0) @binding(1) var<storage, read> cmc_a_col_indices: array<i32>;
@group(0) @binding(2) var<storage, read> cmc_b_row_ptrs: array<i32>;
@group(0) @binding(3) var<storage, read> cmc_b_col_indices: array<i32>;
@group(0) @binding(4) var<storage, read_write> cmc_row_counts: array<i32>;
@group(0) @binding(5) var<uniform> cmc_params: CsrMergeCountParams;

@compute @workgroup_size(256)
fn csr_merge_count(@builtin(global_invocation_id) gid: vec3<u32>) {
    let row = gid.x;
    if (row >= cmc_params.nrows) {
        return;
    }

    let a_start = cmc_a_row_ptrs[row];
    let a_end = cmc_a_row_ptrs[row + 1u];
    let b_start = cmc_b_row_ptrs[row];
    let b_end = cmc_b_row_ptrs[row + 1u];

    var count: i32 = 0;
    var i: i32 = a_start;
    var j: i32 = b_start;

    // Merge sorted column indices, count unique columns
    while (i < a_end && j < b_end) {
        let a_col = cmc_a_col_indices[i];
        let b_col = cmc_b_col_indices[j];

        count = count + 1;
        if (a_col < b_col) {
            i = i + 1;
        } else if (a_col > b_col) {
            j = j + 1;
        } else {
            i = i + 1;
            j = j + 1;
        }
    }

    // Add remaining elements from A
    count = count + (a_end - i);
    // Add remaining elements from B
    count = count + (b_end - j);

    cmc_row_counts[row] = count;
}

// ============================================================================
// csr_mul_count
// ============================================================================

struct CsrMulCountParams {
    nrows: u32,
}

@group(0) @binding(0) var<storage, read> cmmc_a_row_ptrs: array<i32>;
@group(0) @binding(1) var<storage, read> cmmc_a_col_indices: array<i32>;
@group(0) @binding(2) var<storage, read> cmmc_b_row_ptrs: array<i32>;
@group(0) @binding(3) var<storage, read> cmmc_b_col_indices: array<i32>;
@group(0) @binding(4) var<storage, read_write> cmmc_row_counts: array<i32>;
@group(0) @binding(5) var<uniform> cmmc_params: CsrMulCountParams;

@compute @workgroup_size(256)
fn csr_mul_count(@builtin(global_invocation_id) gid: vec3<u32>) {
    let row = gid.x;
    if (row >= cmmc_params.nrows) {
        return;
    }

    let a_start = cmmc_a_row_ptrs[row];
    let a_end = cmmc_a_row_ptrs[row + 1u];
    let b_start = cmmc_b_row_ptrs[row];
    let b_end = cmmc_b_row_ptrs[row + 1u];

    var count: i32 = 0;
    var i: i32 = a_start;
    var j: i32 = b_start;

    // Count matching column indices only (intersection)
    while (i < a_end && j < b_end) {
        let a_col = cmmc_a_col_indices[i];
        let b_col = cmmc_b_col_indices[j];

        if (a_col < b_col) {
            i = i + 1;
        } else if (a_col > b_col) {
            j = j + 1;
        } else {
            count = count + 1;
            i = i + 1;
            j = j + 1;
        }
    }

    cmmc_row_counts[row] = count;
}

// ============================================================================
// csc_merge_count
// ============================================================================

struct CscMergeCountParams {
    ncols: u32,
}

@group(0) @binding(0) var<storage, read> csmc_a_col_ptrs: array<i32>;
@group(0) @binding(1) var<storage, read> csmc_a_row_indices: array<i32>;
@group(0) @binding(2) var<storage, read> csmc_b_col_ptrs: array<i32>;
@group(0) @binding(3) var<storage, read> csmc_b_row_indices: array<i32>;
@group(0) @binding(4) var<storage, read_write> csmc_col_counts: array<i32>;
@group(0) @binding(5) var<uniform> csmc_params: CscMergeCountParams;

@compute @workgroup_size(256)
fn csc_merge_count(@builtin(global_invocation_id) gid: vec3<u32>) {
    let col = gid.x;
    if (col >= csmc_params.ncols) {
        return;
    }

    let a_start = csmc_a_col_ptrs[col];
    let a_end = csmc_a_col_ptrs[col + 1u];
    let b_start = csmc_b_col_ptrs[col];
    let b_end = csmc_b_col_ptrs[col + 1u];

    var count: i32 = 0;
    var i: i32 = a_start;
    var j: i32 = b_start;

    while (i < a_end && j < b_end) {
        let a_row = csmc_a_row_indices[i];
        let b_row = csmc_b_row_indices[j];

        count = count + 1;
        if (a_row < b_row) {
            i = i + 1;
        } else if (a_row > b_row) {
            j = j + 1;
        } else {
            i = i + 1;
            j = j + 1;
        }
    }

    count = count + (a_end - i);
    count = count + (b_end - j);

    csmc_col_counts[col] = count;
}

// ============================================================================
// csc_mul_count
// ============================================================================

struct CscMulCountParams {
    ncols: u32,
}

@group(0) @binding(0) var<storage, read> csmmc_a_col_ptrs: array<i32>;
@group(0) @binding(1) var<storage, read> csmmc_a_row_indices: array<i32>;
@group(0) @binding(2) var<storage, read> csmmc_b_col_ptrs: array<i32>;
@group(0) @binding(3) var<storage, read> csmmc_b_row_indices: array<i32>;
@group(0) @binding(4) var<storage, read_write> csmmc_col_counts: array<i32>;
@group(0) @binding(5) var<uniform> csmmc_params: CscMulCountParams;

@compute @workgroup_size(256)
fn csc_mul_count(@builtin(global_invocation_id) gid: vec3<u32>) {
    let col = gid.x;
    if (col >= csmmc_params.ncols) {
        return;
    }

    let a_start = csmmc_a_col_ptrs[col];
    let a_end = csmmc_a_col_ptrs[col + 1u];
    let b_start = csmmc_b_col_ptrs[col];
    let b_end = csmmc_b_col_ptrs[col + 1u];

    var count: i32 = 0;
    var i: i32 = a_start;
    var j: i32 = b_start;

    while (i < a_end && j < b_end) {
        let a_row = csmmc_a_row_indices[i];
        let b_row = csmmc_b_row_indices[j];

        if (a_row < b_row) {
            i = i + 1;
        } else if (a_row > b_row) {
            j = j + 1;
        } else {
            count = count + 1;
            i = i + 1;
            j = j + 1;
        }
    }

    csmmc_col_counts[col] = count;
}

// ============================================================================
// exclusive_scan_i32
// ============================================================================

struct ScanParams {
    n: u32,
}

@group(0) @binding(0) var<storage, read> scan_input: array<i32>;
@group(0) @binding(1) var<storage, read_write> scan_output: array<i32>;
@group(0) @binding(2) var<uniform> scan_params: ScanParams;

// Sequential exclusive scan - only first thread does work
@compute @workgroup_size(1)
fn exclusive_scan_i32(@builtin(global_invocation_id) gid: vec3<u32>) {
    if (gid.x != 0u) {
        return;
    }

    var sum: i32 = 0;
    for (var i: u32 = 0u; i < scan_params.n; i = i + 1u) {
        let val = scan_input[i];
        scan_output[i] = sum;
        sum = sum + val;
    }
    // Final element is total sum
    scan_output[scan_params.n] = sum;
}