Refactor Sollya integration and enhance trigonometric function precision

- Completely rewrote Sollya integration with a new Python CLI tool (tools/sollya/cli.py)
- Replaced the old generate.py script with a more robust implementation
- Added comprehensive Sollya core library (tools/sollya/core.sol) with utility functions
- Created new SIMD-optimized lookup table implementation (npsr/lut-inl.h)
- Significantly enhanced the Precise class for floating-point precision control
- Refactored trigonometric data generation scripts with better documentation
- Updated constants generation with improved Cody-Waite and Payne-Hanek reduction
- Added kpi16-inl.h.sol for high-precision π/16 table generation
- Removed deprecated utils-inl.h and old trigonometric data files
- Improved project configuration in pyproject.toml and .spin/cmds.py
- Enhanced code documentation throughout the codebase

Author: seiko2plus

.spin/cmds.py

@@ -1,18 +1,16 @@
-import os
 import pathlib
 import sys
 import click
-import spin
-from spin.cmds import meson
 
 curdir = pathlib.Path(__file__).parent
 rootdir = curdir.parent
+toolsdir = rootdir / "tools"
+sys.path.insert(0, str(toolsdir))
 
 
-@click.command(help="Generate sollya python based files")
+@click.command(help="Generate sollya c++/python based files")
 @click.option("-f", "--force", is_flag=True, help="Force regenerate all files")
-def generate(*, force):
-    spin.util.run(
-        ["python", str(rootdir / "tools" / "sollya" / "generate.py")]
-        + (["--force"] if force else []),
-    )
+def sollya(*, force):
+    import sollya  # type: ignore[import]
+
+    sollya.main(force=force)

npsr/common.h

@@ -0,0 +1,31 @@
+#ifndef NPSR_HWY_H_
+#define NPSR_HWY_H_
+
+#include <hwy/highway.h>
+
+#endif  // NPSR_HWY_H_
+
+#if defined(NPSR_HWY_FOREACH_H_) == defined(HWY_TARGET_TOGGLE)  // NOLINT
+#ifdef NPSR_HWY_FOREACH_H_
+#undef NPSR_HWY_FOREACH_H_
+#else
+#define NPSR_HWY_FOREACH_H_
+#endif
+
+HWY_BEFORE_NAMESPACE();
+namespace npsr::HWY_NAMESPACE {
+namespace hn = hwy::HWY_NAMESPACE;
+using hn::DFromV;
+using hn::MFromD;
+using hn::Rebind;
+using hn::RebindToUnsigned;
+using hn::TFromD;
+using hn::TFromV;
+using hn::VFromD;
+constexpr bool kNativeFMA = HWY_NATIVE_FMA != 0;
+
+HWY_ATTR void DummyToSuppressUnusedWarning() {}
+}  // namespace npsr::HWY_NAMESPACE
+HWY_AFTER_NAMESPACE();
+
+#endif  // NPSR_HWY_FOREACH_H_

npsr/hwy.h

@@ -0,0 +1,220 @@
+#if defined(NPSR_LUT_INL_H_) == defined(HWY_TARGET_TOGGLE)  // NOLINT
+#ifdef NPSR_LUT_INL_H_
+#undef NPSR_LUT_INL_H_
+#else
+#define NPSR_LUT_INL_H_
+#endif
+
+#include "npsr/hwy.h"
+
+HWY_BEFORE_NAMESPACE();
+
+namespace npsr::HWY_NAMESPACE {
+
+/**
+ * @brief SIMD-optimized lookup table implementation
+ *
+ * This class provides an efficient lookup table.
+ * It stores data in both row-major and column-major
+ * formats to optimize different access patterns.
+ *
+ * @tparam T Element type (must match SIMD vector element type)
+ * @tparam kRows Number of rows in the lookup table
+ * @tparam kCols Number of columns in the lookup table
+ *
+ * Example usage:
+ * @code
+ *   // Create a 2x4 lookup table
+ *   constexpr Lut lut{{1.0f, 2.0f, 3.0f, 4.0f}, {5.0f, 6.0f, 7.0f, 8.0f}};
+ *   // Load values using SIMD indices
+ *   auto indices = Set(d, 2);  // SIMD vector of indices
+ *   Vec<D> out0, out1;
+ *   lut.Load(indices, out0, out1);
+ * @endcode
+ */
+template <typename T, size_t kRows, size_t kCols>
+class Lut {
+ public:
+  static constexpr size_t kLength = kRows * kCols;
+
+  /**
+   * @brief Construct a lookup table from row arrays
+   *
+   * @tparam ColSizes Size of each row array (deduced)
+   * @param rows Variable number of arrays, each representing a row
+   *
+   * @note All rows must have exactly kCols elements
+   * @note The constructor is constexpr for compile-time initialization
+   */
+  template <size_t... ColSizes>
+  constexpr Lut(const T (&...rows)[ColSizes]) : row_{} {
+    // Check that we have the right number of rows
+    static_assert(sizeof...(rows) == kRows,
+                  "Number of rows doesn't match template parameter");
+    // Check that all rows have the same number of columns
+    static_assert(((ColSizes == kCols) && ...),
+                  "All rows must have the same number of columns");
+
+    // Copy data using recursive template approach
+    ToRowMajor_<0>(rows...);
+  }
+
+  /**
+   * @brief Load values from the LUT using SIMD indices
+   *
+   * This method performs efficient SIMD lookups by selecting the optimal
+   * implementation based on the vector size and LUT dimensions.
+   *
+   * @tparam VU SIMD vector type for indices
+   * @tparam OutV Output SIMD vector types (must match number of rows)
+   * @param idx SIMD vector of column indices
+   * @param out Output vectors (one per row)
+   *
+   * @note The number of output vectors must exactly match kRows
+   * @note Index values must be in range [0, kCols)
+   */
+  template <typename VU, typename... OutV>
+  HWY_ATTR void Load(VU idx, OutV &...out) const {
+    static_assert(sizeof...(OutV) == kRows,
+                  "Number of output vectors must match number of rows in LUT");
+    using namespace hn;
+    using TU = TFromV<VU>;
+    static_assert(sizeof(TU) == sizeof(T),
+                  "Index type must match LUT element type");
+    // Row-major based optimization
+    LoadRow_(idx, out...);
+  }
+
+ private:
+  /// Convert input rows to row-major storage format
+  template <size_t RowIDX, size_t... ColSizes>
+  constexpr void ToRowMajor_(const T (&...rows)[ColSizes]) {
+    if constexpr (RowIDX < kRows) {
+      auto row_array = std::get<RowIDX>(std::make_tuple(rows...));
+      for (size_t col = 0; col < kCols; ++col) {
+        row_[RowIDX * kCols + col] = row_array[col];
+      }
+      ToRowMajor_<RowIDX + 1>(rows...);
+    }
+  }
+
+  /// Dispatch to optimal row-load implementation based on vector/LUT size
+  template <size_t Off = 0, typename VU, typename... OutV>
+  HWY_ATTR void LoadRow_(VU idx, OutV &...out) const {
+    using namespace hn;
+    using DU = DFromV<VU>;
+    const DU du;
+    using D = Rebind<T, DU>;
+    const D d;
+
+#if !HWY_HAVE_SCALABLE
+    constexpr size_t kLanes = Lanes(du);
+    if constexpr (kLanes == kCols) {
+      // Vector size matches table width - use single table lookup
+      const auto ind = IndicesFromVec(d, idx);
+      LoadX1_<Off>(ind, out...);
+    } else if constexpr (kLanes * 2 == kCols) {
+      // Vector size is half table width - use two table lookup
+      const auto ind = IndicesFromVec(d, idx);
+      LoadX2_<Off>(ind, out...);
+    }
+#else
+    if constexpr (0) {
+    }
+#endif
+    else {
+      // Fallback to gather for other configurations
+      LoadGather_<Off>(idx, out...);
+    }
+  }
+
+  // Load using single table lookup (vector size == table width)
+  template <size_t Off = 0, typename VInd, typename OutV0, typename... OutV>
+  HWY_ATTR void LoadX1_(const VInd &ind, OutV0 &out0, OutV &...out) const {
+    using namespace hn;
+    using D = DFromV<OutV0>;
+    const D d;
+
+    const OutV0 lut0 = Load(d, row_ + Off);
+    out0 = TableLookupLanes(d, lut0, ind);
+
+    if constexpr (sizeof...(OutV) > 0) {
+      LoadX1_<Off + kCols>(ind, out...);
+    }
+  }
+
+  // Load using two table lookups (vector size == table width / 2)
+  template <size_t Off = 0, typename VInd, typename OutV0, typename... OutV>
+  HWY_ATTR void LoadX2_(const VInd &ind, OutV0 &out0, OutV &...out) const {
+    using namespace hn;
+    using D = DFromV<OutV0>;
+    const D d;
+
+    constexpr size_t kLanes = kCols / 2;
+    const OutV0 lut0 = LoadU(d, row_ + Off);
+    const OutV0 lut1 = LoadU(d, row_ + Off + kLanes);
+    out0 = TwoTablesLookupLanes(d, lut0, lut1, ind);
+
+    if constexpr (sizeof...(OutV) > 0) {
+      LoadX2_<Off + kCols>(ind, out...);
+    }
+  }
+
+  //  General fallback using gather instructions
+  template <size_t Off = 0, typename VU, typename OutV0, typename... OutV>
+  HWY_ATTR void LoadGather_(const VU &idx, OutV0 &out0, OutV &...out) const {
+    using namespace hn;
+    using D = DFromV<OutV0>;
+    const D d;
+    out0 = GatherIndex(d, row_ + Off, BitCast(RebindToSigned<D>(), idx));
+    if constexpr (sizeof...(OutV) > 0) {
+      LoadGather_<Off + kCols>(idx, out...);
+    }
+  }
+
+  // Row-major
+  HWY_ALIGN T row_[kLength];
+};
+
+/**
+ * @brief Deduction guide for automatic dimension detection
+ *
+ * Allows constructing a Lut without explicitly specifying dimensions:
+ * @code
+ *   Lut lut{row0, row1, row2};  // Dimensions deduced from arrays
+ * @endcode
+ */
+template <typename T, size_t First, size_t... Rest>
+Lut(const T (&first)[First], const T (&...rest)[Rest])
+    -> Lut<T, 1 + sizeof...(Rest), First>;
+
+/**
+ * @brief Factory function that requires explicit type specification
+ *
+ * This approach forces users to specify the type T explicitly while
+ * automatically deducing the dimensions from the array arguments.
+ *
+ * Note: We use MakeLut since partial deduction guides (e.g., Lut<float>{...})
+ * require C++20, but this codebase targets C++17.
+ *
+ * @tparam T Element type (must be explicitly specified)
+ * @param first First row array
+ * @param rest Additional row arrays
+ * @return Lut with deduced dimensions
+ *
+ * Usage:
+ * @code
+ *   auto lut = MakeLut<float>(row0, row1, row2);  // T explicit, dimensions
+ * deduced
+ * @endcode
+ */
+template <typename T, size_t First, size_t... Rest>
+constexpr auto MakeLut(const T (&first)[First], const T (&...rest)[Rest]) {
+  return Lut<T, 1 + sizeof...(Rest), First>{first, rest...};
+}
+
+}  // namespace npsr::HWY_NAMESPACE
+
+HWY_AFTER_NAMESPACE();
+
+#endif  // NPSR_LUT_INL_H_

npsr/lut-inl.h

@@ -1,12 +1,11 @@
 // To include them once per target, which is ensured by the toggle check.
-// clang-format off
-#if defined(_NPSR_NPSR_H_) == defined(HWY_TARGET_TOGGLE) // NOLINT
-#ifdef _NPSR_NPSR_H_
-#undef _NPSR_NPSR_H_
+#if defined(NPSR_NPSR_H_) == defined(HWY_TARGET_TOGGLE)  // NOLINT
+#ifdef NPSR_NPSR_H_
+#undef NPSR_NPSR_H_
 #else
-#define _NPSR_NPSR_H_
+#define NPSR_NPSR_H_
 #endif
 
 #include "npsr/trig/inl.h"
 
-#endif // _NPSR_NPSR_H_
+#endif  // NPSR_NPSR_H_