Optimize nmod_poly_divrem_basecase

Author: fredrik-johansson

doc/source/nmod_vec.rst

@@ -141,6 +141,39 @@ Arithmetic operations
     `c` and all elements of ``vec`` are assumed to be less than ``mod.n``.
 
 
+Arithmetic operations without reduction
+--------------------------------------------------------------------------------
+
+The following functions set `r_i \gets r_i + v_i \cdot c`,
+`0 \le i < len`, with `r` given by ``res`` and `v` given by ``vec``,
+without performing modular reduction,
+with specific limitations on the inputs.
+
+.. function:: void _nmod_vec_nored_scalar_addmul_halflimb(nn_ptr res, nn_srcptr vec, slong len, ulong c)
+
+    Assumes that `v_i` and `c` are half-limb values and that
+    `r_i + v_i \cdot c` does not overflow a limb.
+
+.. function:: void _nmod_vec_nored_ll_scalar_addmul_halflimb(nn_ptr res, nn_srcptr vec, slong len, ulong c)
+
+    The array ``res`` contains ``2 len`` limbs, representing double-limb
+    integers contiguously.
+    Assumes that `v_i` and `c` are half-limb values and that
+    `r_i + v_i \cdot c` does not overflow two limbs.
+
+.. function:: void _nmod_vec_nored_ll_scalar_addmul(nn_ptr res, nn_srcptr vec, slong len, ulong c)
+
+    The array ``res`` contains ``2 len`` limbs, representing double-limb
+    integers contiguously.
+    Assumes that `r_i + v_i \cdot c` does not overflow two limbs.
+
+.. function:: void _nmod_vec_nored_lll_scalar_addmul(nn_ptr res, nn_srcptr vec, slong len, ulong c)
+
+    The array ``res`` contains ``3 len`` limbs, representing triple-limb
+    integers contiguously.
+    Assumes that `r_i + v_i \cdot c` does not overflow three limbs.
+
+
 Dot products
 --------------------------------------------------------------------------------
 

src/gr/nmod.c

@@ -964,9 +964,11 @@ static int
 _gr_nmod_poly_divrem(nn_ptr Q, nn_ptr R, nn_srcptr A, slong lenA,
                                   nn_srcptr B, slong lenB, gr_ctx_t ctx)
 {
-    if (lenA <= 20 || lenB <= 8 || lenA - lenB <= 6 ||
-            (NMOD_BITS(NMOD_CTX(ctx)) <= 61 && lenA <= 40) ||
-            (NMOD_BITS(NMOD_CTX(ctx)) <= 29 && lenA <= 70))
+#if FLINT_BITS == 64 && defined(__AVX2__)
+    if (lenA - lenB < 80 || lenB < ((NMOD_BITS(NMOD_CTX(ctx)) <= 27) ? 400 : 80))
+#else
+    if (lenA - lenB < 80 || lenB < ((NMOD_BITS(NMOD_CTX(ctx)) <= 27) ? 200 : 80))
+#endif
     {
         ulong invB;
         int status;

src/nmod_poly/divrem.c

@@ -21,9 +21,11 @@ _nmod_poly_divrem(nn_ptr Q, nn_ptr R, nn_srcptr A, slong lenA,
 {
     ulong invB;
 
-    if (lenA <= 20 || lenB <= 8 || lenA - lenB <= 6 ||
-            (NMOD_BITS(mod) <= 61 && lenA <= 40) ||
-            (NMOD_BITS(mod) <= 29 && lenA <= 70))
+#if FLINT_BITS == 64 && defined(__AVX2__)
+    if (lenA - lenB < 80 || lenB < ((NMOD_BITS(mod) <= 27) ? 400 : 80))
+#else
+    if (lenA - lenB < 80 || lenB < ((NMOD_BITS(mod) <= 27) ? 200 : 80))
+#endif
     {
         invB = (B[lenB - 1] == 1) ? 1 : n_invmod(B[lenB - 1], mod.n);