// SPDX-FileCopyrightText: 2026 Filip Leonarski, Paul Scherrer Institute // SPDX-License-Identifier: GPL-3.0-only #include #ifdef __APPLE__ #include #include #include #include #include "gemmi/it92.hpp" #include "gemmi/symmetry.hpp" #include "../image_analysis/structure_refinement/SideChainField.h" #include "../image_analysis/structure_refinement/SideChainFieldMetal.h" // The side-chain field's direct sums on the Apple GPU (SideChainFieldMetal.h) are the CPU's, bit for bit, so the // contributions are compared as bit patterns and the energies exactly. A Mac without a Metal device fails rather than // skips, as in ModelDensityMetalTest.cpp. TEST_CASE("SideChainFieldMetal_BitIdenticalToCPU", "[portable][Metal]") { std::mt19937 rng(23); std::uniform_real_distribution unit(0.0f, 1.0f); std::uniform_int_distribution index(-25, 25); // reflections of a 50 A cell to about 2 A, in no particular order, a tenth of them free FieldReflections fr; for (int i = 0; i < 20000; i++) { const float h = static_cast(index(rng)), k = static_cast(index(rng)), l = static_cast(index(rng)); fr.hkl.push_back({h, k, l}); fr.stol2.push_back((h * h + k * k + l * l) / (4.0f * 50 * 50)); fr.fo.push_back(100 * unit(rng)); fr.d.push_back(0.9f); fr.var.push_back(50 + 100 * unit(rng)); fr.centric.push_back(i % 7 == 0); fr.work.push_back(i % 10 != 0); fr.k_total.push_back(1.0f); fr.fmask.push_back({10 * unit(rng), 10 * unit(rng)}); } const std::vector elements = {gemmi::IT92::get(gemmi::El::C, 0).coefs, gemmi::IT92::get(gemmi::El::N, 0).coefs, gemmi::IT92::get(gemmi::El::O, 0).coefs}; const auto ops = FieldOperators(*gemmi::find_spacegroup_by_name("P 21 21 21")); // 70 states of six atoms each - more than one batch - and combinations of one and of two states FieldStates states; for (int s = 0; s < 70; s++) { std::vector atoms; for (int a = 0; a < 6; a++) atoms.push_back({unit(rng), unit(rng), unit(rng), 10 + 50 * unit(rng), 0.5f + 0.5f * unit(rng), a % 3}); states.Add(atoms); } states.current = {0}; for (int c = 0; c < 69; c++) states.AddCombo(c % 2 == 0 ? std::vector{c} : std::vector{c, c + 1}); auto cpu = MakeSideChainFieldEvaluatorCPU(fr, ops, elements, 4, false); auto gpu = MakeSideChainFieldEvaluatorCPU(fr, ops, elements, 4, true); const size_t before = SideChainFieldMetal::States(); size_t different = 0; for (int s : {0, 33, 69}) { const std::vector> c = cpu->Contribution(states, s), g = gpu->Contribution(states, s); REQUIRE(c.size() == g.size()); for (size_t m = 0; m < c.size(); m++) different += std::memcmp(&c[m], &g[m], sizeof(c[m])) != 0; } REQUIRE(SideChainFieldMetal::States() == before + 3); // the GPU summed them CHECK(different == 0); const std::vector> total = cpu->Contribution(states, 0); cpu->SetTotal(total); gpu->SetTotal(total); const std::vector ec = cpu->Evaluate(states), eg = gpu->Evaluate(states); REQUIRE(SideChainFieldMetal::States() == before + 3 + 70); CHECK(ec == eg); } #endif