bigdecimal 4.0.1 → 4.1.1

data/ext/bigdecimal/bigdecimal.h

@@ -17,24 +17,15 @@
 # include <float.h>
 #endif
 
-#if defined(HAVE_INT64_T) && !defined(BIGDECIMAL_USE_DECDIG_UINT16_T)
-# define DECDIG uint32_t
-# define DECDIG_DBL uint64_t
-# define DECDIG_DBL_SIGNED int64_t
-# define SIZEOF_DECDIG 4
-# define PRI_DECDIG_PREFIX ""
-# ifdef PRI_LL_PREFIX
-#  define PRI_DECDIG_DBL_PREFIX PRI_LL_PREFIX
-# else
-#  define PRI_DECDIG_DBL_PREFIX "l"
-# endif
+#define DECDIG uint32_t
+#define DECDIG_DBL uint64_t
+#define DECDIG_DBL_SIGNED int64_t
+#define SIZEOF_DECDIG 4
+#define PRI_DECDIG_PREFIX ""
+#ifdef PRI_LL_PREFIX
+# define PRI_DECDIG_DBL_PREFIX PRI_LL_PREFIX
 #else
-# define DECDIG uint16_t
-# define DECDIG_DBL uint32_t
-# define DECDIG_DBL_SIGNED int32_t
-# define SIZEOF_DECDIG 2
-# define PRI_DECDIG_PREFIX "h"
-# define PRI_DECDIG_DBL_PREFIX ""
+# define PRI_DECDIG_DBL_PREFIX "l"
 #endif
 
 #define PRIdDECDIG PRI_DECDIG_PREFIX"d"
@@ -51,31 +42,15 @@
 #define PRIxDECDIG_DBL PRI_DECDIG_DBL_PREFIX"x"
 #define PRIXDECDIG_DBL PRI_DECDIG_DBL_PREFIX"X"
 
-#if SIZEOF_DECDIG == 4
-# define BIGDECIMAL_BASE ((DECDIG)1000000000U)
-# define BIGDECIMAL_COMPONENT_FIGURES 9
+#define BIGDECIMAL_BASE ((DECDIG)1000000000U)
+#define BIGDECIMAL_COMPONENT_FIGURES 9
 /*
  * The number of components required for a 64-bit integer.
  *
  *   INT64_MAX:   9_223372036_854775807
  *   UINT64_MAX: 18_446744073_709551615
  */
-# define BIGDECIMAL_INT64_MAX_LENGTH 3
-
-#elif SIZEOF_DECDIG == 2
-# define BIGDECIMAL_BASE ((DECDIG)10000U)
-# define BIGDECIMAL_COMPONENT_FIGURES 4
-/*
- * The number of components required for a 64-bit integer.
- *
- *   INT64_MAX:   922_3372_0368_5477_5807
- *   UINT64_MAX: 1844_6744_0737_0955_1615
- */
-# define BIGDECIMAL_INT64_MAX_LENGTH 5
-
-#else
-# error Unknown size of DECDIG
-#endif
+#define BIGDECIMAL_INT64_MAX_LENGTH 3
 
 #define BIGDECIMAL_DOUBLE_FIGURES (1+DBL_DIG)
 
@@ -188,6 +163,16 @@ typedef struct {
     DECDIG frac[FLEXIBLE_ARRAY_SIZE]; /* Array of fraction part. */
 } Real;
 
+typedef struct {
+    VALUE bigdecimal;
+    Real *real;
+} BDVALUE;
+
+typedef struct {
+    VALUE bigdecimal_or_nil;
+    Real *real_or_null;
+} NULLABLE_BDVALUE;
+
 /*
  *  ------------------
  *   EXPORTables.
@@ -214,7 +199,7 @@ VP_EXPORT unsigned short VpSetRoundMode(unsigned short n);
 VP_EXPORT int VpException(unsigned short f,const char *str,int always);
 VP_EXPORT size_t VpNumOfChars(Real *vp,const char *pszFmt);
 VP_EXPORT size_t VpInit(DECDIG BaseVal);
-VP_EXPORT Real *VpAlloc(const char *szVal, int strict_p, int exc);
+VP_EXPORT NULLABLE_BDVALUE VpAlloc(const char *szVal, int strict_p, int exc);
 VP_EXPORT size_t VpAsgn(Real *c, Real *a, int isw);
 VP_EXPORT size_t VpAddSub(Real *c,Real *a,Real *b,int operation);
 VP_EXPORT size_t VpMult(Real *c,Real *a,Real *b);
@@ -232,10 +217,31 @@ VP_EXPORT int VpActiveRound(Real *y, Real *x, unsigned short f, ssize_t il);
 VP_EXPORT int VpMidRound(Real *y, unsigned short f, ssize_t nf);
 VP_EXPORT int VpLeftRound(Real *y, unsigned short f, ssize_t nf);
 VP_EXPORT void VpFrac(Real *y, Real *x);
+VP_EXPORT int AddExponent(Real *a, SIGNED_VALUE n);
 
 /* VP constants */
 VP_EXPORT Real *VpOne(void);
 
+/*
+ *  **** BigDecimal part ****
+ */
+VP_EXPORT VALUE BigDecimal_lt(VALUE self, VALUE r);
+VP_EXPORT VALUE BigDecimal_ge(VALUE self, VALUE r);
+VP_EXPORT VALUE BigDecimal_exponent(VALUE self);
+VP_EXPORT VALUE BigDecimal_fix(VALUE self);
+VP_EXPORT VALUE BigDecimal_frac(VALUE self);
+VP_EXPORT VALUE BigDecimal_add(VALUE self, VALUE b);
+VP_EXPORT VALUE BigDecimal_sub(VALUE self, VALUE b);
+VP_EXPORT VALUE BigDecimal_mult(VALUE self, VALUE b);
+VP_EXPORT VALUE BigDecimal_add2(VALUE self, VALUE b, VALUE n);
+VP_EXPORT VALUE BigDecimal_sub2(VALUE self, VALUE b, VALUE n);
+VP_EXPORT VALUE BigDecimal_mult2(VALUE self, VALUE b, VALUE n);
+VP_EXPORT VALUE BigDecimal_split(VALUE self);
+VP_EXPORT VALUE BigDecimal_decimal_shift(VALUE self, VALUE v);
+VP_EXPORT inline BDVALUE GetBDValueMust(VALUE v);
+VP_EXPORT inline BDVALUE rbd_allocate_struct_zero_wrap(int sign, size_t const digits);
+#define NewZeroWrap rbd_allocate_struct_zero_wrap
+
 /*
  *  ------------------
  *  MACRO definitions.

data/ext/bigdecimal/div.h

@@ -0,0 +1,192 @@
+// Calculate the inverse of x using the Newton-Raphson method.
+static VALUE
+newton_raphson_inverse(VALUE x, size_t prec) {
+    BDVALUE bdone = NewZeroWrap(1, 1);
+    VpSetOne(bdone.real);
+    VALUE one = bdone.bigdecimal;
+
+    // Initial approximation in 2 digits
+    BDVALUE bdx = GetBDValueMust(x);
+    BDVALUE inv0 = NewZeroWrap(1, 2 * BIGDECIMAL_COMPONENT_FIGURES);
+    VpSetOne(inv0.real);
+    DECDIG_DBL numerator = (DECDIG_DBL)BIGDECIMAL_BASE * 100;
+    DECDIG_DBL denominator = (DECDIG_DBL)bdx.real->frac[0] * 100 + (DECDIG_DBL)(bdx.real->Prec >= 2 ? bdx.real->frac[1] : 0) * 100 / BIGDECIMAL_BASE;
+    inv0.real->frac[0] = (DECDIG)(numerator / denominator);
+    inv0.real->frac[1] = (DECDIG)((numerator % denominator) * (BIGDECIMAL_BASE / 100) / denominator * 100);
+    inv0.real->Prec = 2;
+    inv0.real->exponent = 1 - bdx.real->exponent;
+    VpNmlz(inv0.real);
+    RB_GC_GUARD(bdx.bigdecimal);
+    VALUE inv = inv0.bigdecimal;
+
+    int bl = 1;
+    while (((size_t)1 << bl) < prec) bl++;
+
+    for (int i = bl; i >= 0; i--) {
+        size_t n = (prec >> i) + 2;
+        if (n > prec) n = prec;
+        // Newton-Raphson iteration: inv_next = inv + inv * (1 - x * inv)
+        VALUE one_minus_x_inv = BigDecimal_sub2(
+            one,
+            BigDecimal_mult(BigDecimal_mult2(x, one, SIZET2NUM(n + 1)), inv),
+            SIZET2NUM(SIZET2NUM(n / 2))
+        );
+        inv = BigDecimal_add2(
+            inv,
+            BigDecimal_mult(inv, one_minus_x_inv),
+            SIZET2NUM(n)
+        );
+    }
+    return inv;
+}
+
+// Calculates divmod by multiplying approximate reciprocal of y
+static void
+divmod_by_inv_mul(VALUE x, VALUE y, VALUE inv, VALUE *res_div, VALUE *res_mod) {
+    VALUE div = BigDecimal_fix(BigDecimal_mult(x, inv));
+    VALUE mod = BigDecimal_sub(x, BigDecimal_mult(div, y));
+    while (RTEST(BigDecimal_lt(mod, INT2FIX(0)))) {
+        mod = BigDecimal_add(mod, y);
+        div = BigDecimal_sub(div, INT2FIX(1));
+    }
+    while (RTEST(BigDecimal_ge(mod, y))) {
+        mod = BigDecimal_sub(mod, y);
+        div = BigDecimal_add(div, INT2FIX(1));
+    }
+    *res_div = div;
+    *res_mod = mod;
+}
+
+static void
+slice_copy(DECDIG *dest, Real *src, size_t rshift, size_t length) {
+    ssize_t start = src->exponent - rshift - length;
+    if (start >= (ssize_t)src->Prec) return;
+    if (start < 0) {
+        dest -= start;
+        length += start;
+        start = 0;
+    }
+    size_t max_length = src->Prec - start;
+    memcpy(dest, src->frac + start, Min(length, max_length) * sizeof(DECDIG));
+}
+
+/* Calculates divmod using Newton-Raphson method.
+ * x and y must be a BigDecimal representing an integer value.
+ *
+ * To calculate with low cost, we need to split x into blocks and perform divmod for each block.
+ * x_digits = remaining_digits(<= y_digits) + block_digits * num_blocks
+ *
+ * Example:
+ * xxx_xxxxx_xxxxx_xxxxx(18 digits) / yyyyy(5 digits)
+ * remaining_digits = 3, block_digits = 5, num_blocks = 3
+ * repeating xxxxx_xxxxxx.divmod(yyyyy) calculation 3 times.
+ *
+ * In each divmod step, dividend is at most (y_digits + block_digits) digits and divisor is y_digits digits.
+ * Reciprocal of y needs block_digits + 1 precision.
+ */
+static void
+divmod_newton(VALUE x, VALUE y, VALUE *div_out, VALUE *mod_out) {
+    size_t x_digits = NUM2SIZET(BigDecimal_exponent(x));
+    size_t y_digits = NUM2SIZET(BigDecimal_exponent(y));
+    if (x_digits <= y_digits) x_digits = y_digits + 1;
+
+    size_t n = x_digits / y_digits;
+    size_t block_figs = (x_digits - y_digits) / n / BIGDECIMAL_COMPONENT_FIGURES + 1;
+    size_t block_digits = block_figs * BIGDECIMAL_COMPONENT_FIGURES;
+    size_t num_blocks = (x_digits - y_digits + block_digits - 1) / block_digits;
+    size_t y_figs = (y_digits - 1) / BIGDECIMAL_COMPONENT_FIGURES + 1;
+    VALUE yinv = newton_raphson_inverse(y, block_digits + 1);
+
+    BDVALUE divident = NewZeroWrap(1, BIGDECIMAL_COMPONENT_FIGURES * (y_figs + block_figs));
+    BDVALUE div_result = NewZeroWrap(1, BIGDECIMAL_COMPONENT_FIGURES * (num_blocks * block_figs + 1));
+    BDVALUE bdx = GetBDValueMust(x);
+
+    VALUE mod = BigDecimal_fix(BigDecimal_decimal_shift(x, SSIZET2NUM(-num_blocks * block_digits)));
+    for (ssize_t i = num_blocks - 1; i >= 0; i--) {
+        memset(divident.real->frac, 0, (y_figs + block_figs) * sizeof(DECDIG));
+
+        BDVALUE bdmod = GetBDValueMust(mod);
+        slice_copy(divident.real->frac, bdmod.real, 0, y_figs);
+        slice_copy(divident.real->frac + y_figs, bdx.real, i * block_figs, block_figs);
+        RB_GC_GUARD(bdmod.bigdecimal);
+
+        VpSetSign(divident.real, 1);
+        divident.real->exponent = y_figs + block_figs;
+        divident.real->Prec = y_figs + block_figs;
+        VpNmlz(divident.real);
+
+        VALUE div;
+        divmod_by_inv_mul(divident.bigdecimal, y, yinv, &div, &mod);
+        BDVALUE bddiv = GetBDValueMust(div);
+        slice_copy(div_result.real->frac + (num_blocks - i - 1) * block_figs, bddiv.real, 0, block_figs + 1);
+        RB_GC_GUARD(bddiv.bigdecimal);
+    }
+    VpSetSign(div_result.real, 1);
+    div_result.real->exponent = num_blocks * block_figs + 1;
+    div_result.real->Prec = num_blocks * block_figs + 1;
+    VpNmlz(div_result.real);
+    RB_GC_GUARD(bdx.bigdecimal);
+    RB_GC_GUARD(divident.bigdecimal);
+    RB_GC_GUARD(div_result.bigdecimal);
+    *div_out = div_result.bigdecimal;
+    *mod_out = mod;
+}
+
+static VALUE
+VpDivdNewtonInner(VALUE args_ptr)
+{
+    Real **args = (Real**)args_ptr;
+    Real *c = args[0], *r = args[1], *a = args[2], *b = args[3];
+    BDVALUE a2, b2, c2, r2;
+    VALUE div, mod, a2_frac = Qnil;
+    size_t div_prec = c->MaxPrec - 1;
+    size_t base_prec = b->Prec;
+
+    a2 = NewZeroWrap(1, a->Prec * BIGDECIMAL_COMPONENT_FIGURES);
+    b2 = NewZeroWrap(1, b->Prec * BIGDECIMAL_COMPONENT_FIGURES);
+    VpAsgn(a2.real, a, 1);
+    VpAsgn(b2.real, b, 1);
+    VpSetSign(a2.real, 1);
+    VpSetSign(b2.real, 1);
+    a2.real->exponent = base_prec + div_prec;
+    b2.real->exponent = base_prec;
+
+    if ((ssize_t)a2.real->Prec > a2.real->exponent) {
+        a2_frac = BigDecimal_frac(a2.bigdecimal);
+        VpMidRound(a2.real, VP_ROUND_DOWN, 0);
+    }
+    divmod_newton(a2.bigdecimal, b2.bigdecimal, &div, &mod);
+    if (a2_frac != Qnil) mod = BigDecimal_add(mod, a2_frac);
+
+    c2 = GetBDValueMust(div);
+    r2 = GetBDValueMust(mod);
+    VpAsgn(c, c2.real, VpGetSign(a) * VpGetSign(b));
+    VpAsgn(r, r2.real, VpGetSign(a));
+    AddExponent(c, a->exponent);
+    AddExponent(c, -b->exponent);
+    AddExponent(c, -div_prec);
+    AddExponent(r, a->exponent);
+    AddExponent(r, -base_prec - div_prec);
+    RB_GC_GUARD(a2.bigdecimal);
+    RB_GC_GUARD(a2.bigdecimal);
+    RB_GC_GUARD(c2.bigdecimal);
+    RB_GC_GUARD(r2.bigdecimal);
+    return Qnil;
+}
+
+static VALUE
+ensure_restore_prec_limit(VALUE limit)
+{
+    VpSetPrecLimit(NUM2SIZET(limit));
+    return Qnil;
+}
+
+static void
+VpDivdNewton(Real *c, Real *r, Real *a, Real *b)
+{
+    Real *args[4] = {c, r, a, b};
+    size_t pl = VpGetPrecLimit();
+    VpSetPrecLimit(0);
+    // Ensure restoring prec limit because some methods used in VpDivdNewtonInner may raise an exception
+    rb_ensure(VpDivdNewtonInner, (VALUE)args, ensure_restore_prec_limit, SIZET2NUM(pl));
+}

data/ext/bigdecimal/extconf.rb

@@ -1,4 +1,4 @@
-# frozen_string_literal: false
+# frozen_string_literal: true
 require 'mkmf'
 
 def have_builtin_func(name, check_expr, opt = "", &b)
@@ -46,13 +46,16 @@ have_func("rb_opts_exception_p", "ruby.h")
 have_func("rb_category_warn", "ruby.h")
 have_const("RB_WARN_CATEGORY_DEPRECATED", "ruby.h")
 
+if RUBY_ENGINE == "ruby"
+  have_const("RUBY_TYPED_EMBEDDABLE", "ruby.h") # RUBY_VERSION >= 3.3
+end
+
 if File.file?(File.expand_path('../lib/bigdecimal.rb', __FILE__))
   bigdecimal_rb = "$(srcdir)/lib/bigdecimal.rb"
 else
   bigdecimal_rb = "$(srcdir)/../../lib/bigdecimal.rb"
 end
 
-$defs.push '-DBIGDECIMAL_USE_DECDIG_UINT16_T' if ENV['BIGDECIMAL_USE_DECDIG_UINT16_T'] == 'true'
 $defs.push '-DBIGDECIMAL_USE_VP_TEST_METHODS' if ENV['BIGDECIMAL_USE_VP_TEST_METHODS'] == 'true'
 
 create_makefile('bigdecimal') {|mf|

data/ext/bigdecimal/missing.h

@@ -58,7 +58,7 @@ char *BigDecimal_dtoa(double d_, int mode, int ndigits, int *decpt, int *sign, c
 
 #ifndef HAVE_RB_COMPLEX_REAL
 static inline VALUE
-rb_complex_real(VALUE cmp)
+rb_complex_real_fallback(VALUE cmp)
 {
 #ifdef RCOMPLEX
   return RCOMPLEX(cmp)->real;
@@ -66,11 +66,12 @@ rb_complex_real(VALUE cmp)
   return rb_funcall(cmp, rb_intern("real"), 0);
 #endif
 }
+#define rb_complex_real rb_complex_real_fallback
 #endif
 
 #ifndef HAVE_RB_COMPLEX_IMAG
 static inline VALUE
-rb_complex_imag(VALUE cmp)
+rb_complex_imag_fallback(VALUE cmp)
 {
 # ifdef RCOMPLEX
   return RCOMPLEX(cmp)->imag;
@@ -78,6 +79,7 @@ rb_complex_imag(VALUE cmp)
   return rb_funcall(cmp, rb_intern("imag"), 0);
 # endif
 }
+#define rb_complex_imag rb_complex_imag_fallback
 #endif
 
 /* st */

data/ext/bigdecimal/ntt.h

@@ -0,0 +1,191 @@
+// NTT (Number Theoretic Transform) implementation for BigDecimal multiplication
+
+#define NTT_PRIMITIVE_ROOT 17
+#define NTT_PRIME_BASE1 24
+#define NTT_PRIME_BASE2 26
+#define NTT_PRIME_BASE3 29
+#define NTT_PRIME_SHIFT 27
+#define NTT_PRIME1 (((uint32_t)NTT_PRIME_BASE1 << NTT_PRIME_SHIFT) | 1)
+#define NTT_PRIME2 (((uint32_t)NTT_PRIME_BASE2 << NTT_PRIME_SHIFT) | 1)
+#define NTT_PRIME3 (((uint32_t)NTT_PRIME_BASE3 << NTT_PRIME_SHIFT) | 1)
+#define MAX_NTT32_BITS 27
+#define NTT_DECDIG_BASE 1000000000
+
+// Calculates base**ex % mod
+static uint32_t
+mod_pow(uint32_t base, uint32_t ex, uint32_t mod) {
+    uint32_t res = 1;
+    uint32_t bit = 1;
+    while (true) {
+        if (ex & bit) {
+            ex ^= bit;
+            res = ((uint64_t)res * base) % mod;
+        }
+        if (!ex) break;
+        base = ((uint64_t)base * base) % mod;
+        bit <<= 1;
+    }
+    return res;
+}
+
+// Recursively performs butterfly operations of NTT
+static void
+ntt_recursive(int size_bits, uint32_t *input, uint32_t *output, uint32_t *tmp, int depth, uint32_t r, uint32_t prime) {
+    if (depth > 0) {
+        ntt_recursive(size_bits, input, tmp, output, depth - 1, ((uint64_t)r * r) % prime, prime);
+    } else {
+        tmp = input;
+    }
+    uint32_t size_half = (uint32_t)1 << (size_bits - 1);
+    uint32_t stride = (uint32_t)1 << (size_bits - depth - 1);
+    uint32_t n = size_half / stride;
+    uint32_t rn = 1, rm = prime - 1;
+    for (uint32_t i = 0; i < n; i++) {
+        uint32_t *aptr = tmp + i * 2 * stride;
+        uint32_t *bptr = aptr + stride;
+        uint32_t *out1 = output + stride * i;
+        uint32_t *out2 = out1 + size_half;
+        for (uint32_t k = 0; k < stride; k++) {
+            uint32_t a = aptr[k], b = bptr[k];
+            out1[k] = (a + (uint64_t)rn * b) % prime;
+            out2[k] = (a + (uint64_t)rm * b) % prime;
+        }
+        rn = ((uint64_t)rn * r) % prime;
+        rm = ((uint64_t)rm * r) % prime;
+    }
+}
+
+/* Perform NTT on input array.
+ * base, shift: Represent the prime number as (base << shift | 1)
+ * r_base: Primitive root of unity modulo prime
+ * size_bits: log2 of the size of the input array. Should be less or equal to shift
+ * input: input array of size (1 << size_bits)
+ */
+static void
+ntt(int size_bits, uint32_t *input, uint32_t *output, uint32_t *tmp, int r_base, int base, int shift, int dir) {
+    uint32_t size = (uint32_t)1 << size_bits;
+    uint32_t prime = ((uint32_t)base << shift) | 1;
+
+    // rmax**(1 << shift) % prime == 1
+    // r**size % prime == 1
+    uint32_t rmax = mod_pow(r_base, base, prime);
+    uint32_t r = mod_pow(rmax, (uint32_t)1 << (shift - size_bits), prime);
+
+    if (dir < 0) r = mod_pow(r, prime - 2, prime);
+    ntt_recursive(size_bits, input, output, tmp, size_bits - 1, r, prime);
+    if (dir < 0) {
+        uint32_t n_inv = mod_pow((uint32_t)size, prime - 2, prime);
+        for (uint32_t i = 0; i < size; i++) {
+            output[i] = ((uint64_t)output[i] * n_inv) % prime;
+        }
+    }
+}
+
+/* Calculate c that satisfies: c % PRIME1 == mod1 && c % PRIME2 == mod2 && c % PRIME3 == mod3
+ * c = (mod1 * 35002755423056150739595925972 + mod2 * 14584479687667766215746868453 + mod3 * 37919651490985126265126719818) % (PRIME1 * PRIME2 * PRIME3)
+ * Assume c <= 999999999**2*(1<<27)
+ */
+static inline void
+mod_restore_prime_24_26_29_shift_27(uint32_t mod1, uint32_t mod2, uint32_t mod3, uint32_t *digits) {
+    // Use mixed radix notation to eliminate modulo by PRIME1 * PRIME2 * PRIME3
+    // [DIG0, DIG1, DIG2] = DIG0 + DIG1 * PRIME1 + DIG2 * PRIME1 * PRIME2
+    // DIG0: 0...PRIME1, DIG1: 0...PRIME2, DIG2: 0...PRIME3
+    // 35002755423056150739595925972 = [1, 3489660916, 3113851359]
+    // 14584479687667766215746868453 = [0, 13, 1297437912]
+    // 37919651490985126265126719818 = [0, 0, 3373338954]
+    uint64_t c0 = mod1;
+    uint64_t c1 = (uint64_t)mod2 * 13 + (uint64_t)mod1 * 3489660916;
+    uint64_t c2 = (uint64_t)mod3 * 3373338954 % NTT_PRIME3 + (uint64_t)mod2 * 1297437912 % NTT_PRIME3 + (uint64_t)mod1 * 3113851359 % NTT_PRIME3;
+    c2 += c1 / NTT_PRIME2;
+    c1 %= NTT_PRIME2;
+    c2 %= NTT_PRIME3;
+    // Base conversion. c fits in 3 digits.
+    c1 += c2 % NTT_DECDIG_BASE * NTT_PRIME2;
+    c0 += c1 % NTT_DECDIG_BASE * NTT_PRIME1;
+    c1 /= NTT_DECDIG_BASE;
+    digits[0] = c0 % NTT_DECDIG_BASE;
+    c0 /= NTT_DECDIG_BASE;
+    c1 += c2 / NTT_DECDIG_BASE % NTT_DECDIG_BASE * NTT_PRIME2;
+    c0 += c1 % NTT_DECDIG_BASE * NTT_PRIME1;
+    c1 /= NTT_DECDIG_BASE;
+    digits[1] = c0 % NTT_DECDIG_BASE;
+    digits[2] = (uint32_t)(c0 / NTT_DECDIG_BASE + c1 % NTT_DECDIG_BASE * NTT_PRIME1);
+}
+
+/*
+ * NTT multiplication
+ * Uses three NTTs with mod (24 << 27 | 1), (26 << 27 | 1), and (29 << 27 | 1)
+ */
+static void
+ntt_multiply(size_t a_size, size_t b_size, uint32_t *a, uint32_t *b, uint32_t *c) {
+    if (a_size < b_size) {
+      ntt_multiply(b_size, a_size, b, a, c);
+      return;
+    }
+
+    int ntt_size_bits = bit_length(b_size - 1) + 1;
+    if (ntt_size_bits > MAX_NTT32_BITS) {
+      rb_raise(rb_eArgError, "Multiply size too large");
+    }
+
+    // To calculate large_a * small_b faster, split into several batches.
+    uint32_t ntt_size = (uint32_t)1 << ntt_size_bits;
+    uint32_t batch_size = ntt_size - (uint32_t)b_size;
+    uint32_t batch_count = (uint32_t)((a_size + batch_size - 1) / batch_size);
+
+    uint32_t *mem = ruby_xcalloc(sizeof(uint32_t), ntt_size * 9);
+    uint32_t *ntt1 = mem;
+    uint32_t *ntt2 = mem + ntt_size;
+    uint32_t *ntt3 = mem + ntt_size * 2;
+    uint32_t *tmp1 = mem + ntt_size * 3;
+    uint32_t *tmp2 = mem + ntt_size * 4;
+    uint32_t *tmp3 = mem + ntt_size * 5;
+    uint32_t *conv1 = mem + ntt_size * 6;
+    uint32_t *conv2 = mem + ntt_size * 7;
+    uint32_t *conv3 = mem + ntt_size * 8;
+
+    // Calculate NTT for b in three primes. Result is reused for each batch of a.
+    memcpy(tmp1, b, b_size * sizeof(uint32_t));
+    memset(tmp1 + b_size, 0, (ntt_size - b_size) * sizeof(uint32_t));
+    ntt(ntt_size_bits, tmp1, ntt1, tmp2, NTT_PRIMITIVE_ROOT, NTT_PRIME_BASE1, NTT_PRIME_SHIFT, +1);
+    ntt(ntt_size_bits, tmp1, ntt2, tmp2, NTT_PRIMITIVE_ROOT, NTT_PRIME_BASE2, NTT_PRIME_SHIFT, +1);
+    ntt(ntt_size_bits, tmp1, ntt3, tmp2, NTT_PRIMITIVE_ROOT, NTT_PRIME_BASE3, NTT_PRIME_SHIFT, +1);
+
+    memset(c, 0, (a_size + b_size) * sizeof(uint32_t));
+    for (uint32_t idx = 0; idx < batch_count; idx++) {
+        uint32_t len = idx == batch_count - 1 ? (uint32_t)a_size - idx * batch_size : batch_size;
+        memcpy(tmp1, a + idx * batch_size, len * sizeof(uint32_t));
+        memset(tmp1 + len, 0, (ntt_size - len) * sizeof(uint32_t));
+        // Calculate convolution for this batch in three primes
+        ntt(ntt_size_bits, tmp1, tmp2, tmp3, NTT_PRIMITIVE_ROOT, NTT_PRIME_BASE1, NTT_PRIME_SHIFT, +1);
+        for (uint32_t i = 0; i < ntt_size; i++) tmp2[i] = ((uint64_t)tmp2[i] * ntt1[i]) % NTT_PRIME1;
+        ntt(ntt_size_bits, tmp2, conv1, tmp3, NTT_PRIMITIVE_ROOT, NTT_PRIME_BASE1, NTT_PRIME_SHIFT, -1);
+        ntt(ntt_size_bits, tmp1, tmp2, tmp3, NTT_PRIMITIVE_ROOT, NTT_PRIME_BASE2, NTT_PRIME_SHIFT, +1);
+        for (uint32_t i = 0; i < ntt_size; i++) tmp2[i] = ((uint64_t)tmp2[i] * ntt2[i]) % NTT_PRIME2;
+        ntt(ntt_size_bits, tmp2, conv2, tmp3, NTT_PRIMITIVE_ROOT, NTT_PRIME_BASE2, NTT_PRIME_SHIFT, -1);
+        ntt(ntt_size_bits, tmp1, tmp2, tmp3, NTT_PRIMITIVE_ROOT, NTT_PRIME_BASE3, NTT_PRIME_SHIFT, +1);
+        for (uint32_t i = 0; i < ntt_size; i++) tmp2[i] = ((uint64_t)tmp2[i] * ntt3[i]) % NTT_PRIME3;
+        ntt(ntt_size_bits, tmp2, conv3, tmp3, NTT_PRIMITIVE_ROOT, NTT_PRIME_BASE3, NTT_PRIME_SHIFT, -1);
+
+        // Restore the original convolution value from three convolutions calculated in three primes.
+        // Each convolution value is maximum 999999999**2*(1<<27)/2
+        for (uint32_t i = 0; i < ntt_size; i++) {
+            uint32_t dig[3];
+            mod_restore_prime_24_26_29_shift_27(conv1[i], conv2[i], conv3[i], dig);
+            // Maximum values of dig[0], dig[1], and dig[2] are 999999999, 999999999 and 67108863 respectively
+            // Maximum overlapped sum (considering overlaps between 2 batches) is less than 4134217722
+            // so this sum doesn't overflow uint32_t.
+            for (int j = 0; j < 3; j++) {
+                // Index check: if dig[j] is non-zero, assign index is within valid range.
+                if (dig[j]) c[idx * batch_size + i + 1 - j] += dig[j];
+            }
+        }
+    }
+    uint32_t carry = 0;
+    for (int32_t i = (uint32_t)(a_size + b_size - 1); i >= 0; i--) {
+        uint32_t v = c[i] + carry;
+        c[i] = v % NTT_DECDIG_BASE;
+        carry = v / NTT_DECDIG_BASE;
+    }
+    ruby_xfree(mem);
+}