paulistrings 0.1.2

Classical simulation of quantum circuits in the Pauli basis, parallel across threads, NUMA domains, MPI ranks and GPUs
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

//! The two layer paths that bypass the fused kernel: the key-preserving rescale and the permutation scatter.

use cudarc::driver::{LaunchConfig, PushKernelArg};

use super::*;
use crate::engine::gpu::columns::{grow, DeviceColumns};
use crate::engine::gpu::error::GpuError;
use crate::engine::gpu::module::{layer_threads, thread_per, warp_per_bucket};
use crate::engine::gpu::prepared::DevicePrepared;
use crate::engine::gpu::scan::exclusive_scan;
use crate::engine::gpu::sum::GpuSum;
use crate::engine::gpu::truncation::KeepProgram;
use crate::engine::timing::timed;

/// `k_rescale`: the key-preserving fast path, into the spare columns at the input's offsets.
pub(super) fn rescale_device<const W: usize>(
    sum: &mut GpuSum<W>,
    table: &DevicePrepared<W>,
    keep: &KeepProgram,
    scratch: &mut LayerScratch<W>,
) -> Result<(), GpuError> {
    let stream = sum.stream.clone();
    let kernels = sum.kernels.clone();
    let ordinal = sum.device();
    let b = sum.hash.num_buckets();
    let extent = scratch.extent.max(sum.len());
    let mut out = match sum.spare.take() {
        Some(spare) => spare,
        None => DeviceColumns::<W>::with_capacity(&stream, ordinal, extent, b)?,
    };
    out.len = 0;
    out.buckets = 0;
    out.reserve(extent, b)?;
    grow(&stream, &mut scratch.destination_offsets, b + 1, ordinal)?;
    let amplitude = &mut scratch.table.amplitude;
    timed_transfer(&stream, &mut scratch.stats, Transfer::H2d, || {
        stream.memcpy_htod(&table.amplitude, amplitude)?;
        Ok(())
    })?;
    let b32 = b as u32;
    timed!(let t0 = scratch.event(sum)?);
    // SAFETY: arguments match `k_rescale` in rescale.cu; `out` has room for the input's extent.
    unsafe {
        stream
            .launch_builder(&kernels.rescale)
            .arg(&sum.columns.x)
            .arg(&sum.columns.z)
            .arg(&sum.columns.coeff)
            .arg(&sum.columns.fingerprint)
            .arg(&sum.columns.start)
            .arg(&sum.columns.lengths)
            .arg(&b32)
            .arg(&table.support_size)
            .arg(&table.qubit_0)
            .arg(&table.qubit_1)
            .arg(&scratch.table.amplitude)
            .arg(keep)
            .arg(&mut out.x)
            .arg(&mut out.z)
            .arg(&mut out.coeff)
            .arg(&mut out.fingerprint)
            .arg(&mut out.start)
            .arg(&mut out.lengths)
            .launch(warp_per_bucket(b))?;
    }
    exclusive_scan(
        &stream,
        &kernels,
        &out.lengths.slice(0..b),
        &mut scratch.destination_offsets.slice_mut(0..b + 1),
        b,
        &mut scratch.scan,
        &mut scratch.totals,
    )?;
    timed!(scratch.lap(sum, t0, |stats| &mut stats.rescale_ns)?);
    let totals = &scratch.totals;
    let total = timed_transfer(&stream, &mut scratch.stats, Transfer::D2h, || {
        let v = stream.clone_dtoh(totals)?;
        stream.synchronize()?;
        Ok(v[0])
    })?;
    out.len = total as usize;
    out.buckets = b;
    sum.spare = Some(std::mem::replace(&mut sum.columns, out));
    scratch.record.rescaled = true;
    sum.debug_check();
    Ok(())
}

/// `k_permute_count`'s and `k_permute_scatter`'s block width: the average source bucket rounded up to a power of two, between two warps and the fused layer's width for `w`.
pub(in crate::engine::gpu) fn permute_threads(len: usize, buckets: usize, w: usize) -> u32 {
    let avg = len.div_ceil(buckets.max(1));
    (avg.max(64).next_power_of_two() as u32).min(layer_threads(w))
}

/// `k_permute_*`: the permutation path, into the spare columns as a tight CSR in bucket order.
pub(super) fn permute_device<const W: usize>(
    sum: &mut GpuSum<W>,
    table: &DevicePrepared<W>,
    keep: &KeepProgram,
    scratch: &mut LayerScratch<W>,
) -> Result<(), GpuError> {
    let stream = sum.stream.clone();
    let kernels = sum.kernels.clone();
    let ordinal = sum.device();
    let b = sum.hash.num_buckets();
    let n_in = sum.len();
    let (b32, e32) = (b as u32, table.entries as u32);
    grow(&stream, &mut scratch.counts, b * table.entries, ordinal)?;
    let mut out = match sum.spare.take() {
        Some(spare) => spare,
        None => DeviceColumns::<W>::with_capacity(&stream, ordinal, n_in, b)?,
    };
    out.len = 0;
    out.buckets = 0;
    out.reserve(n_in, b)?;
    let (buffers, entry_of) = (&mut scratch.table, &mut scratch.entry_of);
    timed_transfer(&stream, &mut scratch.stats, Transfer::H2d, || {
        stream.memcpy_htod(&table.amplitude, &mut buffers.amplitude)?;
        stream.memcpy_htod(&table.mask, &mut buffers.mask)?;
        stream.memcpy_htod(&table.bucket_delta, &mut buffers.bucket_delta)?;
        stream.memcpy_htod(&table.fingerprint_delta, &mut buffers.fingerprint_delta)?;
        stream.memcpy_htod(&table.entry_of, entry_of)?;
        Ok(())
    })?;
    let block_per_bucket = LaunchConfig {
        grid_dim: (b32, 1, 1),
        block_dim: (permute_threads(n_in, b, W), 1, 1),
        shared_mem_bytes: 0,
    };
    timed!(let t0 = scratch.event(sum)?);
    // SAFETY: arguments match `k_permute_count` in permute.cu; `counts` holds `b * e` entries.
    unsafe {
        stream
            .launch_builder(&kernels.permute_count)
            .arg(&sum.columns.x)
            .arg(&sum.columns.z)
            .arg(&sum.columns.coeff)
            .arg(&sum.columns.start)
            .arg(&sum.columns.lengths)
            .arg(&table.mode)
            .arg(&e32)
            .arg(&table.support_size)
            .arg(&table.qubit_0)
            .arg(&table.qubit_1)
            .arg(&table.rotation_cos)
            .arg(&table.rotation_sin)
            .arg(&scratch.table.amplitude)
            .arg(&scratch.table.mask)
            .arg(&scratch.table.nonzero_patterns)
            .arg(&scratch.entry_of)
            .arg(keep)
            .arg(&mut scratch.counts)
            .launch(block_per_bucket)?;
    }
    timed!(scratch.lap(sum, t0, |stats| &mut stats.gather_ns)?);
    timed!(let t1 = scratch.event(sum)?);
    // SAFETY: arguments match `k_permute_lengths`; `out.lens` holds `b` entries.
    unsafe {
        stream
            .launch_builder(&kernels.permute_lengths)
            .arg(&scratch.counts)
            .arg(&scratch.table.bucket_delta)
            .arg(&b32)
            .arg(&e32)
            .arg(&mut out.lengths)
            .launch(thread_per(b, 256))?;
    }
    exclusive_scan(
        &stream,
        &kernels,
        &out.lengths.slice(0..b),
        &mut out.start.slice_mut(0..b + 1),
        b,
        &mut scratch.scan,
        &mut scratch.totals,
    )?;
    timed!(scratch.lap(sum, t1, |stats| &mut stats.gather_ns)?);
    timed!(let t2 = scratch.event(sum)?);
    // SAFETY: arguments match `k_permute_scatter`; `out` has room for every input row, the scan's total at most.
    unsafe {
        stream
            .launch_builder(&kernels.permute_scatter)
            .arg(&sum.columns.x)
            .arg(&sum.columns.z)
            .arg(&sum.columns.coeff)
            .arg(&sum.columns.fingerprint)
            .arg(&sum.columns.start)
            .arg(&sum.columns.lengths)
            .arg(&table.mode)
            .arg(&e32)
            .arg(&table.support_size)
            .arg(&table.qubit_0)
            .arg(&table.qubit_1)
            .arg(&table.rotation_cos)
            .arg(&table.rotation_sin)
            .arg(&scratch.table.amplitude)
            .arg(&scratch.table.mask)
            .arg(&scratch.table.nonzero_patterns)
            .arg(&scratch.entry_of)
            .arg(&scratch.table.bucket_delta)
            .arg(&scratch.table.fingerprint_delta)
            .arg(&scratch.counts)
            .arg(&out.start)
            .arg(keep)
            .arg(&mut out.x)
            .arg(&mut out.z)
            .arg(&mut out.coeff)
            .arg(&mut out.fingerprint)
            .launch(block_per_bucket)?;
    }
    timed!(scratch.lap(sum, t2, |stats| &mut stats.merge_ns)?);
    let totals = &scratch.totals;
    let total = timed_transfer(&stream, &mut scratch.stats, Transfer::D2h, || {
        let v = stream.clone_dtoh(totals)?;
        stream.synchronize()?;
        Ok(v[0])
    })?;
    out.len = total as usize;
    out.buckets = b;
    sum.spare = Some(std::mem::replace(&mut sum.columns, out));
    scratch.extent = sum.len();
    scratch.record.records = n_in as u64;
    scratch.record.permuted = true;
    sum.debug_check();
    Ok(())
}