Break: Return types & 4-way finalizers

Author: ashvardanian

include/simsimd/curved.h

@@ -204,23 +204,23 @@ SIMSIMD_PUBLIC void simsimd_mahalanobis_f32_accurate(simsimd_f32_t const *a, sim
                                                      simsimd_f32_t const *c, simsimd_size_t n, simsimd_f64_t *result);
 /** @copydoc simsimd_bilinear_f16 */
 SIMSIMD_PUBLIC void simsimd_bilinear_f16_accurate(simsimd_f16_t const *a, simsimd_f16_t const *b,
-                                                  simsimd_f16_t const *c, simsimd_size_t n, simsimd_f32_t *result);
+                                                  simsimd_f16_t const *c, simsimd_size_t n, simsimd_f64_t *result);
 /** @copydoc simsimd_bilinear_f16c */
 SIMSIMD_PUBLIC void simsimd_bilinear_f16c_accurate(simsimd_f16c_t const *a, simsimd_f16c_t const *b,
-                                                   simsimd_f16c_t const *c, simsimd_size_t n, simsimd_f32c_t *results);
+                                                   simsimd_f16c_t const *c, simsimd_size_t n, simsimd_f64c_t *results);
 /** @copydoc simsimd_mahalanobis_f16 */
 SIMSIMD_PUBLIC void simsimd_mahalanobis_f16_accurate(simsimd_f16_t const *a, simsimd_f16_t const *b,
-                                                     simsimd_f16_t const *c, simsimd_size_t n, simsimd_f32_t *result);
+                                                     simsimd_f16_t const *c, simsimd_size_t n, simsimd_f64_t *result);
 /** @copydoc simsimd_bilinear_bf16 */
 SIMSIMD_PUBLIC void simsimd_bilinear_bf16_accurate(simsimd_bf16_t const *a, simsimd_bf16_t const *b,
-                                                   simsimd_bf16_t const *c, simsimd_size_t n, simsimd_f32_t *result);
+                                                   simsimd_bf16_t const *c, simsimd_size_t n, simsimd_f64_t *result);
 /** @copydoc simsimd_bilinear_bf16c */
 SIMSIMD_PUBLIC void simsimd_bilinear_bf16c_accurate(simsimd_bf16c_t const *a, simsimd_bf16c_t const *b,
                                                     simsimd_bf16c_t const *c, simsimd_size_t n,
-                                                    simsimd_f32c_t *results);
+                                                    simsimd_f64c_t *results);
 /** @copydoc simsimd_mahalanobis_bf16 */
 SIMSIMD_PUBLIC void simsimd_mahalanobis_bf16_accurate(simsimd_bf16_t const *a, simsimd_bf16_t const *b,
-                                                      simsimd_bf16_t const *c, simsimd_size_t n, simsimd_f32_t *result);
+                                                      simsimd_bf16_t const *c, simsimd_size_t n, simsimd_f64_t *result);
 #if SIMSIMD_TARGET_NEON
 /** @copydoc simsimd_bilinear_f32 */
 SIMSIMD_PUBLIC void simsimd_bilinear_f32_neon(simsimd_f32_t const *a, simsimd_f32_t const *b, simsimd_f32_t const *c,

include/simsimd/dots.h

@@ -968,7 +968,7 @@ SIMSIMD_MAKE_DOTS_INNER(i8i8i32_neon, i8, i32, simsimd_b128_vec_t, simsimd_dot_i
 // U8 GEMM: k_tile=16 (16 u8s = 16 bytes = NEON register width)
 SIMSIMD_MAKE_DOTS_SERIAL_PACKED_SIZE(neon, u8, i32, 16)
 SIMSIMD_MAKE_DOTS_SERIAL_PACK(neon, u8, i32, 16)
-SIMSIMD_MAKE_DOTS_INNER(u8u8i32_neon, u8, i32, simsimd_b128_vec_t, simsimd_dot_u8x16_state_neon_t,
+SIMSIMD_MAKE_DOTS_INNER(u8u8i32_neon, u8, u32, simsimd_b128_vec_t, simsimd_dot_u8x16_state_neon_t,
                         simsimd_dot_u8x16_init_neon, _simsimd_load_b128_neon, _simsimd_partial_load_b8x16_neon,
                         simsimd_dot_u8x16_update_neon, simsimd_dot_u8x16_finalize_neon,
                         /*k_tile=*/16, /*MR=*/4, /*MC=*/128, /*NC=*/2048, /*KC=*/256)
@@ -1090,7 +1090,7 @@ SIMSIMD_MAKE_DOTS_INNER(i8i8i32_haswell, i8, i32, simsimd_b256_vec_t, simsimd_do
 // U8 GEMM: k_tile=32 (32 u8s = 32 bytes = AVX2 register width)
 SIMSIMD_MAKE_DOTS_SERIAL_PACKED_SIZE(haswell, u8, i32, 32)
 SIMSIMD_MAKE_DOTS_SERIAL_PACK(haswell, u8, i32, 32)
-SIMSIMD_MAKE_DOTS_INNER(u8u8i32_haswell, u8, i32, simsimd_b256_vec_t, simsimd_dot_u8x32_state_haswell_t,
+SIMSIMD_MAKE_DOTS_INNER(u8u8i32_haswell, u8, u32, simsimd_b256_vec_t, simsimd_dot_u8x32_state_haswell_t,
                         simsimd_dot_u8x32_init_haswell, _simsimd_load_b256_haswell, _simsimd_partial_load_b8x32_haswell,
                         simsimd_dot_u8x32_update_haswell, simsimd_dot_u8x32_finalize_haswell,
                         /*k_tile=*/32, /*MR=*/4, /*MC=*/128, /*NC=*/2048, /*KC=*/256)
@@ -1199,7 +1199,7 @@ SIMSIMD_MAKE_DOTS_INNER(i8i8i32_ice, i8, i32, simsimd_b512_vec_t, simsimd_dot_i8
 // U8 GEMM: k_tile=64 (64 u8s = 64 bytes = 1 cache line)
 SIMSIMD_MAKE_DOTS_SERIAL_PACKED_SIZE(ice, u8, i32, SIMSIMD_DOTS_SERIAL_TILE_K_U8)
 SIMSIMD_MAKE_DOTS_SERIAL_PACK(ice, u8, i32, SIMSIMD_DOTS_SERIAL_TILE_K_U8)
-SIMSIMD_MAKE_DOTS_INNER(u8u8i32_ice, u8, i32, simsimd_b512_vec_t, simsimd_dot_u8x64_state_ice_t,
+SIMSIMD_MAKE_DOTS_INNER(u8u8i32_ice, u8, u32, simsimd_b512_vec_t, simsimd_dot_u8x64_state_ice_t,
                         simsimd_dot_u8x64_init_ice, _simsimd_load_b512_skylake, _simsimd_partial_load_b8x64_skylake,
                         simsimd_dot_u8x64_update_ice, simsimd_dot_u8x64_finalize_ice,
                         /*k_tile=*/64, /*MR=*/4, /*MC=*/128, /*NC=*/2048, /*KC=*/256)

include/simsimd/sparse.h

@@ -341,15 +341,12 @@ SIMSIMD_MAKE_SPARSE_DOT(accurate, u16, bf16, f64, simsimd_bf16_to_f64) // simsim
         *count = (simsimd_u32_t)intersection_size;                                                                   \
     }
 
-SIMSIMD_MAKE_INTERSECT_GALLOPING(serial, u16, size)                  // simsimd_intersect_u16_serial
-SIMSIMD_MAKE_INTERSECT_GALLOPING(serial, u32, size)                  // simsimd_intersect_u32_serial
-SIMSIMD_MAKE_SPARSE_DOT(serial, u16, bf16, f32, simsimd_bf16_to_f32) // simsimd_sparse_dot_u16bf16_serial
-
-SIMSIMD_INTERNAL void _simsimd_f32_to_f32(simsimd_f32_t const *x, simsimd_f32_t *y) { *y = *x; }
-SIMSIMD_INTERNAL void _simsimd_f32_to_f64(simsimd_f32_t const *x, simsimd_f64_t *y) { *y = (simsimd_f64_t)*x; }
-
-SIMSIMD_MAKE_SPARSE_DOT(serial, u32, f32, f32, _simsimd_f32_to_f32)   // simsimd_sparse_dot_u32f32_serial
-SIMSIMD_MAKE_SPARSE_DOT(accurate, u32, f32, f64, _simsimd_f32_to_f64) // simsimd_sparse_dot_u32f32_accurate
+SIMSIMD_MAKE_INTERSECT_GALLOPING(serial, u16, size)                     // simsimd_intersect_u16_serial
+SIMSIMD_MAKE_INTERSECT_GALLOPING(serial, u32, size)                     // simsimd_intersect_u32_serial
+SIMSIMD_MAKE_SPARSE_DOT(serial, u16, bf16, f32, simsimd_bf16_to_f32)    // simsimd_sparse_dot_u16bf16_serial
+SIMSIMD_MAKE_SPARSE_DOT(serial, u32, f32, f32, SIMSIMD_ASSIGN_FROM_TO)  // simsimd_sparse_dot_u32f32_serial
+SIMSIMD_MAKE_SPARSE_DOT(accurate, u16, bf16, f64, _simsimd_bf16_to_f64) // simsimd_sparse_dot_u16bf16_accurate
+SIMSIMD_MAKE_SPARSE_DOT(accurate, u32, f32, f64, _simsimd_f32_to_f64)   // simsimd_sparse_dot_u32f32_accurate
 
 /*  The AVX-512 implementations are inspired by the "Faster-Than-Native Alternatives
  *  for x86 VP2INTERSECT Instructions" paper by Guille Diez-Canas, 2022.