Skip to content

Commit 388fb9e

Browse files
Neiko2002Neiko2002
committed
fix(random): ensure cross-platform precision in DeterministicNormalDistribution
1 parent 4c58c21 commit 388fb9e

5 files changed

Lines changed: 40 additions & 18 deletions

File tree

‎cpp/deglib/include/random.h‎

Lines changed: 26 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -27,6 +27,7 @@ namespace deglib::random {
2727
namespace detail {
2828

2929
// Portable float natural logarithm (log(x)) without platform libm dependencies.
30+
// Accurate to 1e-7 (full float precision) across [1e-38, 1e38].
3031
inline float portable_log(float x) {
3132
uint32_t ix;
3233
std::memcpy(&ix, &x, sizeof(float));
@@ -42,18 +43,39 @@ inline float portable_log(float x) {
4243
float s = f / (2.0f + f);
4344
float z = s * s;
4445
float w = z * z;
45-
float R = z * (0.6666666666666666f + w * (0.4f + w * (0.2857142857142857f + w * 0.2222222222222222f)));
46+
float R = z * (0.6666666666666666f + w * (0.4f + w * (0.2857142857142857f + w * (0.2222222222222222f + w * 0.1818181818181818f))));
4647
return static_cast<float>(exp) * 0.6931471805599453f + s * (2.0f + R);
4748
}
4849

4950
// Portable float cosine (cos(x)) without platform libm dependencies.
51+
// Uses symmetry reduction to [0, PI/2] and a 10th-degree polynomial for full float precision (1e-7).
5052
inline float portable_cos(float x) {
5153
constexpr float TWO_PI = 6.28318530717958647692f;
54+
constexpr float PI = 3.14159265358979323846f;
55+
constexpr float HALF_PI = 1.57079632679489661923f;
5256
constexpr float INV_TWO_PI = 0.15915494309189533576f;
53-
float k = std::floor(x * INV_TWO_PI + 0.5f);
54-
x -= k * TWO_PI;
57+
58+
// Reduce x to [0, 2*PI)
59+
x -= std::floor(x * INV_TWO_PI) * TWO_PI;
60+
if (x < 0.0f) x += TWO_PI;
61+
62+
// Symmetry reduction to [0, PI]
63+
if (x > PI) {
64+
x = TWO_PI - x;
65+
}
66+
67+
// Symmetry reduction to [0, PI/2]
68+
bool sign = false;
69+
if (x > HALF_PI) {
70+
x = PI - x;
71+
sign = true;
72+
}
73+
74+
// Polynomial for cos(x) on [0, PI/2] accurate to 1e-7
5575
float x2 = x * x;
56-
return 1.0f - x2 * (0.5f - x2 * (0.041666666666666664f - x2 * (0.0013888888888888889f - x2 * 0.0000248015873015873f)));
76+
float res = 1.0f - x2 * (0.5f - x2 * (0.041666666666666664f - x2 * (0.0013888888888888889f - x2 * (0.0000248015873015873f - x2 * 0.000000275573192239859f))));
77+
78+
return sign ? -res : res;
5779
}
5880

5981
} // namespace detail

‎cpp/test/src/regression/builder/test_l2_builder_regression.cpp‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -36,7 +36,7 @@ TEST(DeglibBuilderRegressionL2, Benchmark_HighLID)
3636

3737
auto gt_data = compute_groundtruth_l2(base_data, base_count, query_data, query_count, dim, 10);
3838

39-
run_regression_test("HighLID", deglib::Metric::L2, 16000.0, 10.0, 0.849,
39+
run_regression_test("HighLID", deglib::Metric::L2, 16000.0, 10.0, 0.867,
4040
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
4141
deglib::distances::fp32_l2::L2Float{}, 100,
4242
deglib::builder::OptimizationTarget::HighLID);
@@ -55,7 +55,7 @@ TEST(DeglibBuilderRegressionL2, Benchmark_StreamingData)
5555

5656
auto gt_data = compute_groundtruth_l2(base_data, base_count, query_data, query_count, dim, 10);
5757

58-
run_regression_test("StreamingData", deglib::Metric::L2, 23000.0, 23.8, 0.89,
58+
run_regression_test("StreamingData", deglib::Metric::L2, 23000.0, 23.8, 0.93,
5959
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
6060
deglib::distances::fp32_l2::L2Float{}, 100,
6161
deglib::builder::OptimizationTarget::StreamingData);

‎cpp/test/src/regression/metric/test_innerproduct_regression.cpp‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -17,7 +17,7 @@ TEST(DeglibRegressionIP, Benchmark_AVX512_16Ext)
1717

1818
auto gt_data = compute_groundtruth_innerproduct(base_data, base_count, query_data, query_count, dim, 10);
1919

20-
run_regression_test("AVX512_16Ext", deglib::Metric::InnerProduct, 18000.0, 10.3, 0.829,
20+
run_regression_test("AVX512_16Ext", deglib::Metric::InnerProduct, 18000.0, 10.3, 0.79,
2121
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
2222
deglib::distances::fp32_ip::InnerProductFloat16Ext_AVX512{}, 50);
2323
#else
@@ -42,7 +42,7 @@ TEST(DeglibRegressionIP, Benchmark_AVX2_16Ext)
4242

4343
auto gt_data = compute_groundtruth_innerproduct(base_data, base_count, query_data, query_count, dim, 10);
4444

45-
run_regression_test("AVX2_16Ext", deglib::Metric::InnerProduct, 18000.0, 10.3, 0.829,
45+
run_regression_test("AVX2_16Ext", deglib::Metric::InnerProduct, 18000.0, 10.3, 0.79,
4646
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
4747
deglib::distances::fp32_ip::InnerProductFloat16Ext_AVX2{}, 50);
4848
#else
@@ -67,7 +67,7 @@ TEST(DeglibRegressionIP, Benchmark_SSE_16Ext)
6767

6868
auto gt_data = compute_groundtruth_innerproduct(base_data, base_count, query_data, query_count, dim, 10);
6969

70-
run_regression_test("SSE_16Ext", deglib::Metric::InnerProduct, 16000.0, 11.8, 0.82,
70+
run_regression_test("SSE_16Ext", deglib::Metric::InnerProduct, 16000.0, 11.8, 0.78,
7171
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
7272
deglib::distances::fp32_ip::InnerProductFloat16Ext_SSE{}, 50);
7373
#else
@@ -88,7 +88,7 @@ TEST(DeglibRegressionIP, Benchmark_Scalar)
8888

8989
auto gt_data = compute_groundtruth_innerproduct(base_data, base_count, query_data, query_count, dim, 10);
9090

91-
run_regression_test("Scalar", deglib::Metric::InnerProduct, 13500.0, 15.8, 0.789,
91+
run_regression_test("Scalar", deglib::Metric::InnerProduct, 13500.0, 15.8, 0.82,
9292
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
9393
deglib::distances::fp32_ip::InnerProductFloat{}, 50);
9494
}

‎cpp/test/src/regression/metric/test_l2_regression.cpp‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -17,7 +17,7 @@ TEST(DeglibRegressionL2, Benchmark_AVX512_16Ext)
1717

1818
auto gt_data = compute_groundtruth_l2(base_data, base_count, query_data, query_count, dim, 10);
1919

20-
run_regression_test("AVX512_16Ext", deglib::Metric::L2, 42000.0, 6.0, 0.951,
20+
run_regression_test("AVX512_16Ext", deglib::Metric::L2, 42000.0, 6.0, 0.949,
2121
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
2222
deglib::distances::fp32_l2::L2Float16Ext_AVX512{}, 100);
2323
#else
@@ -42,7 +42,7 @@ TEST(DeglibRegressionL2, Benchmark_AVX2_16Ext)
4242

4343
auto gt_data = compute_groundtruth_l2(base_data, base_count, query_data, query_count, dim, 10);
4444

45-
run_regression_test("AVX2_16Ext", deglib::Metric::L2, 38000.0, 6.2, 0.95,
45+
run_regression_test("AVX2_16Ext", deglib::Metric::L2, 38000.0, 6.2, 0.92,
4646
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
4747
deglib::distances::fp32_l2::L2Float16Ext_AVX2{}, 100);
4848
#else
@@ -67,7 +67,7 @@ TEST(DeglibRegressionL2, Benchmark_SSE_16Ext)
6767

6868
auto gt_data = compute_groundtruth_l2(base_data, base_count, query_data, query_count, dim, 10);
6969

70-
run_regression_test("SSE_16Ext", deglib::Metric::L2, 33000.0, 7.0, 0.97,
70+
run_regression_test("SSE_16Ext", deglib::Metric::L2, 33000.0, 7.0, 0.979,
7171
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
7272
deglib::distances::fp32_l2::L2Float16Ext_SSE{}, 100);
7373
#else
@@ -88,7 +88,7 @@ TEST(DeglibRegressionL2, Benchmark_Scalar)
8888

8989
auto gt_data = compute_groundtruth_l2(base_data, base_count, query_data, query_count, dim, 10);
9090

91-
run_regression_test("Scalar", deglib::Metric::L2, 29000.0, 9.3, 0.969,
91+
run_regression_test("Scalar", deglib::Metric::L2, 29000.0, 9.3, 0.94,
9292
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
9393
deglib::distances::fp32_l2::L2Float{}, 100);
9494
}

‎cpp/test/src/regression/metric/test_l2_uint8_regression.cpp‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -17,7 +17,7 @@ TEST(DeglibRegressionL2Uint8, Benchmark_AVX512_Ext32)
1717

1818
auto gt_data = compute_groundtruth_l2_uint8(base_data, base_count, query_data, query_count, dim, 10);
1919

20-
run_regression_test("AVX512_Ext32", deglib::Metric::L2_Uint8, 77000.0, 4.6, 0.97,
20+
run_regression_test("AVX512_Ext32", deglib::Metric::L2_Uint8, 77000.0, 4.6, 0.98,
2121
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
2222
deglib::distances::uint8_l2::L2Uint8Ext32_AVX512{}, 500);
2323
#else
@@ -42,7 +42,7 @@ TEST(DeglibRegressionL2Uint8, Benchmark_AVX2_Ext32)
4242

4343
auto gt_data = compute_groundtruth_l2_uint8(base_data, base_count, query_data, query_count, dim, 10);
4444

45-
run_regression_test("AVX2_Ext32", deglib::Metric::L2_Uint8, 78000.0, 4.6, 0.97,
45+
run_regression_test("AVX2_Ext32", deglib::Metric::L2_Uint8, 78000.0, 4.6, 0.98,
4646
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
4747
deglib::distances::uint8_l2::L2Uint8Ext32_AVX2{}, 500);
4848
#else
@@ -67,7 +67,7 @@ TEST(DeglibRegressionL2Uint8, Benchmark_SSE_Ext32)
6767

6868
auto gt_data = compute_groundtruth_l2_uint8(base_data, base_count, query_data, query_count, dim, 10);
6969

70-
run_regression_test("SSE_Ext32", deglib::Metric::L2_Uint8, 73000.0, 5.1, 0.97,
70+
run_regression_test("SSE_Ext32", deglib::Metric::L2_Uint8, 73000.0, 5.1, 0.98,
7171
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
7272
deglib::distances::uint8_l2::L2Uint8Ext32_SSE{}, 500);
7373
#else
@@ -88,7 +88,7 @@ TEST(DeglibRegressionL2Uint8, Benchmark_Scalar)
8888

8989
auto gt_data = compute_groundtruth_l2_uint8(base_data, base_count, query_data, query_count, dim, 10);
9090

91-
run_regression_test("Scalar", deglib::Metric::L2_Uint8, 66000.0, 5.5, 0.97,
91+
run_regression_test("Scalar", deglib::Metric::L2_Uint8, 66000.0, 5.5, 0.98,
9292
base_data.data(), query_data.data(), base_count, query_count, dim, gt_data,
9393
deglib::distances::uint8_l2::L2Uint8{}, 500);
9494
}

0 commit comments

Comments
 (0)