1#include "gtest/gtest.h"
 2#include "spqlios/arithmetic/zn_arithmetic_private.h"
 3#include "testlib/zn_layouts.h"
 4
 5static void test_zn_vmp_prepare(ZN32_VMP_PREPARE_CONTIGUOUS_F prep) {
 6  MOD_Z* module = new_z_module_info(DEFAULT);
 7  for (uint64_t nrows : {1, 2, 5, 15}) {
 8    for (uint64_t ncols : {1, 2, 32, 42, 67}) {
 9      std::vector<int32_t> src(nrows * ncols);
10      zn32_pmat_layout out(nrows, ncols);
11      for (int32_t& x : src) x = uniform_i64_bits(32);
12      prep(module, out.data, src.data(), nrows, ncols);
13      for (uint64_t i = 0; i < nrows; ++i) {
14        for (uint64_t j = 0; j < ncols; ++j) {
15          int32_t in = src[i * ncols + j];
16          int32_t actual = out.get(i, j);
17          ASSERT_EQ(actual, in);
18        }
19      }
20    }
21  }
22  delete_z_module_info(module);
23}
24
25TEST(zn, zn32_vmp_prepare_contiguous) { test_zn_vmp_prepare(zn32_vmp_prepare_contiguous); }
26TEST(zn, default_zn32_vmp_prepare_contiguous_ref) { test_zn_vmp_prepare(default_zn32_vmp_prepare_contiguous_ref); }
27
28template <typename INTTYPE>
29static void test_zn_vmp_apply(void (*apply)(const MOD_Z*, int32_t*, uint64_t, const INTTYPE*, uint64_t,
30                                            const ZN32_VMP_PMAT*, uint64_t, uint64_t)) {
31  MOD_Z* module = new_z_module_info(DEFAULT);
32  for (uint64_t nrows : {1, 2, 5, 15}) {
33    for (uint64_t ncols : {1, 2, 32, 42, 67}) {
34      for (uint64_t a_size : {1, 2, 5, 15}) {
35        for (uint64_t res_size : {1, 2, 32, 42, 67}) {
36          std::vector<INTTYPE> a(a_size);
37          zn32_pmat_layout out(nrows, ncols);
38          std::vector<int32_t> res(res_size);
39          for (INTTYPE& x : a) x = uniform_i64_bits(32);
40          out.fill_random();
41          std::vector<int32_t> expect = vmp_product(a.data(), a_size, res_size, out);
42          apply(module, res.data(), res_size, a.data(), a_size, out.data, nrows, ncols);
43          for (uint64_t i = 0; i < res_size; ++i) {
44            int32_t exp = expect[i];
45            int32_t actual = res[i];
46            ASSERT_EQ(actual, exp);
47          }
48        }
49      }
50    }
51  }
52  delete_z_module_info(module);
53}
54
55TEST(zn, zn32_vmp_apply_i32) { test_zn_vmp_apply(zn32_vmp_apply_i32); }
56TEST(zn, zn32_vmp_apply_i16) { test_zn_vmp_apply(zn32_vmp_apply_i16); }
57TEST(zn, zn32_vmp_apply_i8) { test_zn_vmp_apply(zn32_vmp_apply_i8); }
58
59TEST(zn, default_zn32_vmp_apply_i32_ref) { test_zn_vmp_apply(default_zn32_vmp_apply_i32_ref); }
60TEST(zn, default_zn32_vmp_apply_i16_ref) { test_zn_vmp_apply(default_zn32_vmp_apply_i16_ref); }
61TEST(zn, default_zn32_vmp_apply_i8_ref) { test_zn_vmp_apply(default_zn32_vmp_apply_i8_ref); }
62
63#ifdef __x86_64__
64TEST(zn, default_zn32_vmp_apply_i32_avx) { test_zn_vmp_apply(default_zn32_vmp_apply_i32_avx); }
65TEST(zn, default_zn32_vmp_apply_i16_avx) { test_zn_vmp_apply(default_zn32_vmp_apply_i16_avx); }
66TEST(zn, default_zn32_vmp_apply_i8_avx) { test_zn_vmp_apply(default_zn32_vmp_apply_i8_avx); }
67#endif