Files
leonarski_fandClaude Opus 5.5 55972f4b67 Refinement CPU engine: the gradient's coefficients scattered from a table, its buffers kept
What Evaluate() did on one thread between the parallel parts - on a 1.2 A case a quarter of an evaluation on 16
threads (eval 44.6 ms, 105 ms on one):
- the target index of each composed term in the half-l coefficient grid worked out with three integer modulos per
  term per evaluation: now once, at the engine's construction (Scatter), the terms added in the same order;
- the coefficient grid and the map allocated (and their pages faulted in) every evaluation, the map copied from
  the template grid and scaled on one thread: now kept by the engine, zeroed and scaled on the threads, the inverse
  transform written into the map itself;
- the density calculator's grid made anew per Compose(): now the engine's own, kept;
- MapToFPhi()'s conjugate-and-scale pass over the coefficients on one thread: now on the call's threads.
The same values in the same order: the refined models of cytC, thau and 9zlo are bit for bit the ones before.
Engine benchmark (16 threads): 1.22 A thau 44.6 -> 30.5 ms per evaluation, cytC 32.6 -> 29.7 ms. --mode refine
--no-gpu -N 16 (on the CPU protocol): thau 115 -> 84 s, 9zlo 115 -> 108 s, cytC 154 -> 150 s.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UqXYdJDrM65gxsCn2PrbfK
2026-11-16 04:58:33 +01:00

120 lines
5.4 KiB
C++

// SPDX-FileCopyrightText: 2026 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only
#include "ModelFFT.h"
#include <algorithm>
#include <array>
#include <cmath>
#include <complex>
#include <vector>
#include "gemmi/fail.hpp"
#include "../../common/CpuFFT.h"
#include "../../common/ParallelFor.h"
// An (nu, nv, nw) map is stored u fastest - row-major (nw, nv, nu) - and its half-l coefficients
// likewise as (nw/2 + 1, nv, nu): the layout gemmi's own transforms read and write. The real transform
// halves w, so w is the last axis transformed.
static const cpu_fft::Shape kUVW{2, 1, 0};
static cpu_fft::Shape MapShape(int nu, int nv, int nw) {
return {static_cast<size_t>(nw), static_cast<size_t>(nv), static_cast<size_t>(nu)};
}
gemmi::FPhiGrid<float> MapToFPhi(const gemmi::Grid<float> &map, size_t nthreads) {
if (map.axis_order == gemmi::AxisOrder::ZYX)
gemmi::fail("MapToFPhi(): ZYX order is not supported");
gemmi::FPhiGrid<float> hkl;
hkl.unit_cell = map.unit_cell;
hkl.spacegroup = map.spacegroup;
hkl.axis_order = map.axis_order;
hkl.half_l = true;
hkl.set_size_without_checking(map.nu, map.nv, map.nw / 2 + 1);
const float norm = static_cast<float>(map.unit_cell.volume / map.point_count());
cpu_fft::R2C(MapShape(map.nu, map.nv, map.nw), kUVW, map.data.data(), hkl.data.data(), nthreads);
// Conjugated, as gemmi does: the forward transform carries exp(-2 pi i h.x), and a structure
// factor is the sum over exp(+2 pi i h.x).
const int n = static_cast<int>(hkl.data.size());
ParallelChunks(n, nthreads, [&](int lo, int hi) {
for (int i = lo; i < hi; i++)
hkl.data[i] = std::conj(hkl.data[i]) * norm;
});
return hkl;
}
void MapFromFPhiExecute(int nu, int nv, int nw, std::complex<float> *in, float *out, size_t nthreads) {
cpu_fft::C2R(MapShape(nu, nv, nw), kUVW, in, out, nthreads);
}
gemmi::FPhiGrid<float> MapToFPhiInParallel(const gemmi::Grid<float> &map, int reach, size_t nthreads) {
if (map.axis_order == gemmi::AxisOrder::ZYX)
gemmi::fail("MapToFPhiInParallel(): ZYX order is not supported");
const int nu = map.nu, nv = map.nv, nw = map.nw, nwh = nw / 2 + 1, nk = 2 * reach + 1;
const size_t plane = static_cast<size_t>(nu) * nv;
std::vector<std::complex<float>> half(plane * nwh);
// The r2c along w of every (u, v) column, then the 2D transform of the planes l <= reach only - the
// rest are never read. Each is one call split over the threads by lines.
cpu_fft::R2C({static_cast<size_t>(nw), plane}, {0}, map.data.data(), half.data(), nthreads);
cpu_fft::C2C({static_cast<size_t>(reach) + 1, static_cast<size_t>(nv), static_cast<size_t>(nu)}, {1, 2}, true,
half.data(), nthreads);
gemmi::FPhiGrid<float> hkl;
hkl.unit_cell = map.unit_cell;
hkl.spacegroup = map.spacegroup;
hkl.axis_order = map.axis_order;
hkl.half_l = true;
hkl.set_size_without_checking(nu, nk, reach + 1);
const float norm = static_cast<float>(map.unit_cell.volume / map.point_count());
// Conjugated and scaled as MapToFPhi does it, and cropped as CropHalfL crops it.
ParallelFor(reach + 1, nthreads, [&](int l) {
for (int k = -reach; k <= reach; k++) {
const std::complex<float> *row = &half[l * plane + static_cast<size_t>(nu) * gemmi::modulo(k, nv)];
std::complex<float> *out = &hkl.data[static_cast<size_t>(nu) * (gemmi::modulo(k, nk) + static_cast<size_t>(nk) * l)];
for (int u = 0; u < nu; u++)
out[u] = std::conj(row[u]) * norm;
}
});
return hkl;
}
gemmi::Grid<float> MapFromFPhi(const gemmi::FPhiGrid<float> &hkl) {
if (hkl.axis_order == gemmi::AxisOrder::ZYX || !hkl.half_l)
gemmi::fail("MapFromFPhi(): only an XYZ half-l grid is supported");
gemmi::Grid<float> map;
map.spacegroup = hkl.spacegroup;
map.unit_cell = hkl.unit_cell;
map.set_size(hkl.nu, hkl.nv, 2 * (hkl.nw - 1));
map.axis_order = hkl.axis_order;
const float norm = static_cast<float>(1.0 / hkl.unit_cell.volume);
// Conjugated, as gemmi does: rho(x) = 1/V sum F exp(-2 pi i h.x), and the backward transform
// carries exp(+2 pi i h.x). A missing coefficient (NaN) counts as zero, also as gemmi does.
std::vector<std::complex<float>> in(hkl.data.size());
for (size_t i = 0; i < in.size(); ++i)
in[i] = std::isnan(hkl.data[i].imag()) ? 0.0f : std::conj(hkl.data[i]);
cpu_fft::C2R(MapShape(map.nu, map.nv, map.nw), kUVW, in.data(), map.data.data());
for (auto &x : map.data)
x *= norm;
return map;
}
// The rows keep their length: cut to 2 reach + 1 as well, the rotation function read them 15% slower on the
// GPU (measured on a large cell).
gemmi::FPhiGrid<float> CropHalfL(const gemmi::FPhiGrid<float> &full, int reach) {
gemmi::FPhiGrid<float> out;
out.unit_cell = full.unit_cell;
out.spacegroup = full.spacegroup;
out.axis_order = full.axis_order;
out.half_l = true;
const int nu = full.nu, n = 2 * reach + 1;
out.set_size_without_checking(nu, n, reach + 1);
for (int l = 0; l <= reach; l++)
for (int k = -reach; k <= reach; k++)
std::copy_n(&full.data[static_cast<size_t>(nu) * (gemmi::modulo(k, full.nv) + static_cast<size_t>(full.nv) * l)],
nu, &out.data[static_cast<size_t>(nu) * (gemmi::modulo(k, n) + static_cast<size_t>(n) * l)]);
return out;
}