1#include "fft64_layouts.h"
  2#ifdef VALGRIND_MEM_TESTS
  3#include "valgrind/memcheck.h"
  4#endif
  5
  6void* alloc64(uint64_t size) {
  7  static uint64_t _msk64 = -64;
  8  if (size == 0) return nullptr;
  9  uint64_t rsize = (size + 63) & _msk64;
 10  uint8_t* reps = (uint8_t*)spqlios_alloc(rsize);
 11  REQUIRE_DRAMATICALLY(reps != 0, "Out of memory");
 12#ifdef VALGRIND_MEM_TESTS
 13  VALGRIND_MAKE_MEM_NOACCESS(reps + size, rsize - size);
 14#endif
 15  return reps;
 16}
 17
 18fft64_vec_znx_dft_layout::fft64_vec_znx_dft_layout(uint64_t n, uint64_t size)
 19    : nn(n),                                      //
 20      size(size),                                 //
 21      data((VEC_ZNX_DFT*)alloc64(n * size * 8)),  //
 22      view(n / 2, size, (double*)data) {}
 23
 24fft64_vec_znx_dft_layout::~fft64_vec_znx_dft_layout() { spqlios_free(data); }
 25
 26double* fft64_vec_znx_dft_layout::get_addr(uint64_t idx) {
 27  REQUIRE_DRAMATICALLY(idx < size, "index overflow " << idx << " / " << size);
 28  return ((double*)data) + idx * nn;
 29}
 30const double* fft64_vec_znx_dft_layout::get_addr(uint64_t idx) const {
 31  REQUIRE_DRAMATICALLY(idx < size, "index overflow " << idx << " / " << size);
 32  return ((double*)data) + idx * nn;
 33}
 34reim_fft64vec fft64_vec_znx_dft_layout::get_copy_zext(uint64_t idx) const {
 35  if (idx < size) {
 36    return reim_fft64vec(nn, get_addr(idx));
 37  } else {
 38    return reim_fft64vec::zero(nn);
 39  }
 40}
 41void fft64_vec_znx_dft_layout::fill_dft_random_log2bound(uint64_t bits) {
 42  for (uint64_t i = 0; i < size; ++i) {
 43    set(i, simple_fft64(znx_i64::random_log2bound(nn, bits)));
 44  }
 45}
 46void fft64_vec_znx_dft_layout::set(uint64_t idx, const reim_fft64vec& value) {
 47  REQUIRE_DRAMATICALLY(value.nn() == nn, "ring dimension mismatch");
 48  value.save_as(get_addr(idx));
 49}
 50thash fft64_vec_znx_dft_layout::content_hash() const { return test_hash(data, size * nn * sizeof(double)); }
 51
 52reim4_elem fft64_vec_znx_dft_layout::get(uint64_t idx, uint64_t blk) const {
 53  REQUIRE_DRAMATICALLY(idx < size, "index overflow: " << idx << " / " << size);
 54  REQUIRE_DRAMATICALLY(blk < nn / 8, "blk overflow: " << blk << " / " << nn / 8);
 55  double* reim = ((double*)data) + idx * nn;
 56  return reim4_elem(reim + blk * 4, reim + nn / 2 + blk * 4);
 57}
 58reim4_elem fft64_vec_znx_dft_layout::get_zext(uint64_t idx, uint64_t blk) const {
 59  REQUIRE_DRAMATICALLY(blk < nn / 8, "blk overflow: " << blk << " / " << nn / 8);
 60  if (idx < size) {
 61    return get(idx, blk);
 62  } else {
 63    return reim4_elem::zero();
 64  }
 65}
 66void fft64_vec_znx_dft_layout::set(uint64_t idx, uint64_t blk, const reim4_elem& value) {
 67  REQUIRE_DRAMATICALLY(idx < size, "index overflow: " << idx << " / " << size);
 68  REQUIRE_DRAMATICALLY(blk < nn / 8, "blk overflow: " << blk << " / " << nn / 8);
 69  double* reim = ((double*)data) + idx * nn;
 70  value.save_re_im(reim + blk * 4, reim + nn / 2 + blk * 4);
 71}
 72void fft64_vec_znx_dft_layout::fill_random(double log2bound) {
 73  for (uint64_t i = 0; i < size; ++i) {
 74    set(i, reim_fft64vec::random(nn, log2bound));
 75  }
 76}
 77void fft64_vec_znx_dft_layout::fill_dft_random(uint64_t log2bound) {
 78  for (uint64_t i = 0; i < size; ++i) {
 79    set(i, reim_fft64vec::dft_random(nn, log2bound));
 80  }
 81}
 82
 83fft64_vec_znx_big_layout::fft64_vec_znx_big_layout(uint64_t n, uint64_t size)
 84    : nn(n),       //
 85      size(size),  //
 86      data((VEC_ZNX_BIG*)alloc64(n * size * 8)) {}
 87
 88znx_i64 fft64_vec_znx_big_layout::get_copy(uint64_t index) const {
 89  REQUIRE_DRAMATICALLY(index < size, "index overflow: " << index << " / " << size);
 90  return znx_i64(nn, ((int64_t*)data) + index * nn);
 91}
 92znx_i64 fft64_vec_znx_big_layout::get_copy_zext(uint64_t index) const {
 93  if (index < size) {
 94    return znx_i64(nn, ((int64_t*)data) + index * nn);
 95  } else {
 96    return znx_i64::zero(nn);
 97  }
 98}
 99void fft64_vec_znx_big_layout::set(uint64_t index, const znx_i64& value) {
100  REQUIRE_DRAMATICALLY(index < size, "index overflow: " << index << " / " << size);
101  value.save_as(((int64_t*)data) + index * nn);
102}
103void fft64_vec_znx_big_layout::fill_random() {
104  for (uint64_t i = 0; i < size; ++i) {
105    set(i, znx_i64::random_log2bound(nn, 1));
106  }
107}
108fft64_vec_znx_big_layout::~fft64_vec_znx_big_layout() { spqlios_free(data); }
109
110fft64_vmp_pmat_layout::fft64_vmp_pmat_layout(uint64_t n, uint64_t nrows, uint64_t ncols)
111    : nn(n),
112      nrows(nrows),
113      ncols(ncols),  //
114      data((VMP_PMAT*)alloc64(nrows * ncols * nn * 8)) {}
115
116double* fft64_vmp_pmat_layout::get_addr(uint64_t row, uint64_t col, uint64_t blk) const {
117  REQUIRE_DRAMATICALLY(row < nrows, "row overflow: " << row << " / " << nrows);
118  REQUIRE_DRAMATICALLY(col < ncols, "col overflow: " << col << " / " << ncols);
119  REQUIRE_DRAMATICALLY(blk < nn / 8, "block overflow: " << blk << " / " << (nn / 8));
120  double* d = (double*)data;
121  if (col == (ncols - 1) && (ncols % 2 == 1)) {
122    // special case: last column out of an odd column number
123    return d + blk * nrows * ncols * 8  // major: blk
124           + col * nrows * 8            // col == ncols-1
125           + row * 8;
126  } else {
127    // general case: columns go by pair
128    return d + blk * nrows * ncols * 8    // major: blk
129           + (col / 2) * (2 * nrows) * 8  // second: col pair index
130           + row * 2 * 8                  // third: row index
131           + (col % 2) * 8;               // minor: col in colpair
132  }
133}
134
135reim4_elem fft64_vmp_pmat_layout::get(uint64_t row, uint64_t col, uint64_t blk) const {
136  return reim4_elem(get_addr(row, col, blk));
137}
138reim4_elem fft64_vmp_pmat_layout::get_zext(uint64_t row, uint64_t col, uint64_t blk) const {
139  REQUIRE_DRAMATICALLY(blk < nn / 8, "block overflow: " << blk << " / " << (nn / 8));
140  if (row < nrows && col < ncols) {
141    return reim4_elem(get_addr(row, col, blk));
142  } else {
143    return reim4_elem::zero();
144  }
145}
146void fft64_vmp_pmat_layout::set(uint64_t row, uint64_t col, uint64_t blk, const reim4_elem& value) const {
147  value.save_as(get_addr(row, col, blk));
148}
149
150fft64_vmp_pmat_layout::~fft64_vmp_pmat_layout() { spqlios_free(data); }
151
152reim_fft64vec fft64_vmp_pmat_layout::get_zext(uint64_t row, uint64_t col) const {
153  if (row >= nrows || col >= ncols) {
154    return reim_fft64vec::zero(nn);
155  }
156  if (nn < 8) {
157    // the pmat is just col major
158    double* addr = (double*)data + (row + col * nrows) * nn;
159    return reim_fft64vec(nn, addr);
160  }
161  // otherwise, reconstruct it block by block
162  reim_fft64vec res(nn);
163  for (uint64_t blk = 0; blk < nn / 8; ++blk) {
164    reim4_elem v = get(row, col, blk);
165    res.set_blk(blk, v);
166  }
167  return res;
168}
169void fft64_vmp_pmat_layout::set(uint64_t row, uint64_t col, const reim_fft64vec& value) {
170  REQUIRE_DRAMATICALLY(row < nrows, "row overflow: " << row << " / " << nrows);
171  REQUIRE_DRAMATICALLY(col < ncols, "row overflow: " << col << " / " << ncols);
172  if (nn < 8) {
173    // the pmat is just col major
174    double* addr = (double*)data + (row + col * nrows) * nn;
175    value.save_as(addr);
176    return;
177  }
178  // otherwise, reconstruct it block by block
179  for (uint64_t blk = 0; blk < nn / 8; ++blk) {
180    reim4_elem v = value.get_blk(blk);
181    set(row, col, blk, v);
182  }
183}
184void fft64_vmp_pmat_layout::fill_random(double log2bound) {
185  for (uint64_t row = 0; row < nrows; ++row) {
186    for (uint64_t col = 0; col < ncols; ++col) {
187      set(row, col, reim_fft64vec::random(nn, log2bound));
188    }
189  }
190}
191void fft64_vmp_pmat_layout::fill_dft_random(uint64_t log2bound) {
192  for (uint64_t row = 0; row < nrows; ++row) {
193    for (uint64_t col = 0; col < ncols; ++col) {
194      set(row, col, reim_fft64vec::dft_random(nn, log2bound));
195    }
196  }
197}
198
199fft64_svp_ppol_layout::fft64_svp_ppol_layout(uint64_t n)
200    : nn(n),  //
201      data((SVP_PPOL*)alloc64(nn * 8)) {}
202
203reim_fft64vec fft64_svp_ppol_layout::get_copy() const { return reim_fft64vec(nn, (double*)data); }
204
205void fft64_svp_ppol_layout::set(const reim_fft64vec& value) { value.save_as((double*)data); }
206
207void fft64_svp_ppol_layout::fill_dft_random(uint64_t log2bound) { set(reim_fft64vec::dft_random(nn, log2bound)); }
208
209void fft64_svp_ppol_layout::fill_random(double log2bound) { set(reim_fft64vec::random(nn, log2bound)); }
210
211fft64_svp_ppol_layout::~fft64_svp_ppol_layout() { spqlios_free(data); }
212thash fft64_svp_ppol_layout::content_hash() const { return test_hash(data, nn * sizeof(double)); }
213
214fft64_cnv_left_layout::fft64_cnv_left_layout(uint64_t n, uint64_t size)
215    : nn(n),  //
216      size(size),
217      data((CNV_PVEC_L*)alloc64(size * nn * 8)) {}
218
219reim4_elem fft64_cnv_left_layout::get(uint64_t idx, uint64_t blk) {
220  REQUIRE_DRAMATICALLY(idx < size, "idx overflow: " << idx << " / " << size);
221  REQUIRE_DRAMATICALLY(blk < nn / 8, "block overflow: " << blk << " / " << (nn / 8));
222  return reim4_elem(((double*)data) + blk * size + idx);
223}
224
225fft64_cnv_left_layout::~fft64_cnv_left_layout() { spqlios_free(data); }
226
227fft64_cnv_right_layout::fft64_cnv_right_layout(uint64_t n, uint64_t size)
228    : nn(n),  //
229      size(size),
230      data((CNV_PVEC_R*)alloc64(size * nn * 8)) {}
231
232reim4_elem fft64_cnv_right_layout::get(uint64_t idx, uint64_t blk) {
233  REQUIRE_DRAMATICALLY(idx < size, "idx overflow: " << idx << " / " << size);
234  REQUIRE_DRAMATICALLY(blk < nn / 8, "block overflow: " << blk << " / " << (nn / 8));
235  return reim4_elem(((double*)data) + blk * size + idx);
236}
237
238fft64_cnv_right_layout::~fft64_cnv_right_layout() { spqlios_free(data); }