numkong 7.0.0 → 7.4.2

package/include/numkong/set/sve.h

@@ -8,23 +8,23 @@
  *
  *  @section set_sve_instructions ARM SVE Instructions
  *
- *      Intrinsic                   Instruction                     Latency     Throughput
- *      svld1_u8                    LD1B (Z.B, P/Z, [Xn])           4-6cy       2/cy
- *      svld1_u32                   LD1W (Z.S, P/Z, [Xn])           4-6cy       2/cy
- *      sveor_u8_m                  EOR (Z.B, P/M, Z.B, Z.B)        1cy         2/cy
- *      svand_u8_m                  AND (Z.B, P/M, Z.B, Z.B)        1cy         2/cy
- *      svorr_u8_m                  ORR (Z.B, P/M, Z.B, Z.B)        1cy         2/cy
- *      svcnt_u8_x                  CNT (Z.B, P/M, Z.B)             2cy         2/cy
- *      svadd_u8_z                  ADD (Z.B, P/M, Z.B, Z.B)        1cy         2/cy
- *      svaddv_u8                   UADDV (D, P, Z.B)               6cy         1/cy
- *      svcmpeq_u32                 CMPEQ (P.S, P/Z, Z.S, Z.S)      2cy         1/cy
- *      svcntp_b32                  CNTP (Xd, P, P.S)               2cy         1/cy
- *      svdup_n_u8                  DUP (Z.B, #imm)                 1cy         2/cy
- *      svwhilelt_b8                WHILELT (P.B, Xn, Xm)           2cy         1/cy
- *      svwhilelt_b32               WHILELT (P.S, Xn, Xm)           2cy         1/cy
- *      svptrue_b8                  PTRUE (P.B, pattern)            1cy         2/cy
- *      svcntb                      CNTB (Xd)                       1cy         2/cy
- *      svcntw                      CNTW (Xd)                       1cy         2/cy
+ *      Intrinsic      Instruction                 V1
+ *      svld1_u8       LD1B (Z.B, P/Z, [Xn])       4-6cy @ 2p
+ *      svld1_u32      LD1W (Z.S, P/Z, [Xn])       4-6cy @ 2p
+ *      sveor_u8_m     EOR (Z.B, P/M, Z.B, Z.B)    1cy @ 2p
+ *      svand_u8_m     AND (Z.B, P/M, Z.B, Z.B)    1cy @ 2p
+ *      svorr_u8_m     ORR (Z.B, P/M, Z.B, Z.B)    1cy @ 2p
+ *      svcnt_u8_x     CNT (Z.B, P/M, Z.B)         2cy @ 2p
+ *      svadd_u8_z     ADD (Z.B, P/M, Z.B, Z.B)    1cy @ 2p
+ *      svaddv_u8      UADDV (D, P, Z.B)           6cy @ 1p
+ *      svcmpeq_u32    CMPEQ (P.S, P/Z, Z.S, Z.S)  2cy @ 1p
+ *      svcntp_b32     CNTP (Xd, P, P.S)           2cy @ 1p
+ *      svdup_n_u8     DUP (Z.B, #imm)             1cy @ 2p
+ *      svwhilelt_b8   WHILELT (P.B, Xn, Xm)       2cy @ 1p
+ *      svwhilelt_b32  WHILELT (P.S, Xn, Xm)       2cy @ 1p
+ *      svptrue_b8     PTRUE (P.B, pattern)        1cy @ 2p
+ *      svcntb         CNTB (Xd)                   1cy @ 2p
+ *      svcntw         CNTW (Xd)                   1cy @ 2p
  */
 #ifndef NK_SET_SVE_H
 #define NK_SET_SVE_H
@@ -46,7 +46,7 @@ extern "C" {
 #pragma GCC target("arch=armv8.2-a+sve")
 #endif
 
-#pragma region - Binary Sets
+#pragma region Binary Sets
 
 NK_PUBLIC void nk_hamming_u1_sve(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n, nk_u32_t *result) {
     nk_size_t n_bytes = nk_size_divide_round_up_(n, NK_BITS_PER_BYTE);
@@ -61,20 +61,20 @@ NK_PUBLIC void nk_hamming_u1_sve(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size
     // On larger register sizes, SVE is faster.
     nk_size_t i = 0, cycle = 0;
     nk_u32_t differences = 0;
-    svuint8_t popcount_u8 = svdup_n_u8(0);
-    svbool_t const all_predicate = svptrue_b8();
+    svuint8_t popcount_u8x = svdup_n_u8(0);
+    svbool_t const all_predicate_b8x = svptrue_b8();
     while (i < n_bytes) {
         do {
-            svbool_t active_predicate = svwhilelt_b8_u64(i, n_bytes);
-            svuint8_t a_u8 = svld1_u8(active_predicate, a + i);
-            svuint8_t b_u8 = svld1_u8(active_predicate, b + i);
-            popcount_u8 = svadd_u8_z(all_predicate, popcount_u8,
-                                     svcnt_u8_x(all_predicate, sveor_u8_m(all_predicate, a_u8, b_u8)));
+            svbool_t active_predicate_b8x = svwhilelt_b8_u64(i, n_bytes);
+            svuint8_t a_u8x = svld1_u8(active_predicate_b8x, a + i);
+            svuint8_t b_u8x = svld1_u8(active_predicate_b8x, b + i);
+            popcount_u8x = svadd_u8_z(all_predicate_b8x, popcount_u8x,
+                                      svcnt_u8_x(all_predicate_b8x, sveor_u8_m(all_predicate_b8x, a_u8x, b_u8x)));
             i += words_per_register;
             ++cycle;
         } while (i < n_bytes && cycle < 31);
-        differences += svaddv_u8(all_predicate, popcount_u8);
-        popcount_u8 = svdup_n_u8(0);
+        differences += svaddv_u8(all_predicate_b8x, popcount_u8x);
+        popcount_u8x = svdup_n_u8(0);
         cycle = 0; // Reset the cycle counter.
     }
 
@@ -94,45 +94,46 @@ NK_PUBLIC void nk_jaccard_u1_sve(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size
     // On larger register sizes, SVE is faster.
     nk_size_t i = 0, cycle = 0;
     nk_u32_t intersection_count = 0, union_count = 0;
-    svuint8_t intersection_popcount_u8 = svdup_n_u8(0);
-    svuint8_t union_popcount_u8 = svdup_n_u8(0);
-    svbool_t const all_predicate = svptrue_b8();
+    svuint8_t intersection_popcount_u8x = svdup_n_u8(0);
+    svuint8_t union_popcount_u8x = svdup_n_u8(0);
+    svbool_t const all_predicate_b8x = svptrue_b8();
     while (i < n_bytes) {
         do {
-            svbool_t active_predicate = svwhilelt_b8_u64(i, n_bytes);
-            svuint8_t a_u8 = svld1_u8(active_predicate, a + i);
-            svuint8_t b_u8 = svld1_u8(active_predicate, b + i);
-            intersection_popcount_u8 = svadd_u8_z(all_predicate, intersection_popcount_u8,
-                                                  svcnt_u8_x(all_predicate, svand_u8_m(all_predicate, a_u8, b_u8)));
-            union_popcount_u8 = svadd_u8_z(all_predicate, union_popcount_u8,
-                                           svcnt_u8_x(all_predicate, svorr_u8_m(all_predicate, a_u8, b_u8)));
+            svbool_t active_predicate_b8x = svwhilelt_b8_u64(i, n_bytes);
+            svuint8_t a_u8x = svld1_u8(active_predicate_b8x, a + i);
+            svuint8_t b_u8x = svld1_u8(active_predicate_b8x, b + i);
+            intersection_popcount_u8x = svadd_u8_z(
+                all_predicate_b8x, intersection_popcount_u8x,
+                svcnt_u8_x(all_predicate_b8x, svand_u8_m(all_predicate_b8x, a_u8x, b_u8x)));
+            union_popcount_u8x = svadd_u8_z(all_predicate_b8x, union_popcount_u8x,
+                                            svcnt_u8_x(all_predicate_b8x, svorr_u8_m(all_predicate_b8x, a_u8x, b_u8x)));
             i += words_per_register;
             ++cycle;
         } while (i < n_bytes && cycle < 31);
-        intersection_count += svaddv_u8(all_predicate, intersection_popcount_u8);
-        intersection_popcount_u8 = svdup_n_u8(0);
-        union_count += svaddv_u8(all_predicate, union_popcount_u8);
-        union_popcount_u8 = svdup_n_u8(0);
+        intersection_count += svaddv_u8(all_predicate_b8x, intersection_popcount_u8x);
+        intersection_popcount_u8x = svdup_n_u8(0);
+        union_count += svaddv_u8(all_predicate_b8x, union_popcount_u8x);
+        union_popcount_u8x = svdup_n_u8(0);
         cycle = 0; // Reset the cycle counter.
     }
 
     *result = (union_count != 0) ? 1.0f - (nk_f32_t)intersection_count / (nk_f32_t)union_count : 0.0f;
 }
 
-#pragma endregion - Binary Sets
+#pragma endregion Binary Sets
 
-#pragma region - Integer Sets
+#pragma region Integer Sets
 
 NK_PUBLIC void nk_jaccard_u32_sve(nk_u32_t const *a, nk_u32_t const *b, nk_size_t n, nk_f32_t *result) {
     nk_size_t const words_per_register = svcntw();
     nk_size_t i = 0;
     nk_u32_t intersection_count = 0;
     while (i < n) {
-        svbool_t active_predicate = svwhilelt_b32_u64(i, n);
-        svuint32_t a_u32 = svld1_u32(active_predicate, a + i);
-        svuint32_t b_u32 = svld1_u32(active_predicate, b + i);
-        svbool_t equality_predicate = svcmpeq_u32(active_predicate, a_u32, b_u32);
-        intersection_count += svcntp_b32(active_predicate, equality_predicate);
+        svbool_t active_predicate_b32x = svwhilelt_b32_u64(i, n);
+        svuint32_t a_u32x = svld1_u32(active_predicate_b32x, a + i);
+        svuint32_t b_u32x = svld1_u32(active_predicate_b32x, b + i);
+        svbool_t equality_predicate_b32x = svcmpeq_u32(active_predicate_b32x, a_u32x, b_u32x);
+        intersection_count += svcntp_b32(active_predicate_b32x, equality_predicate_b32x);
         i += words_per_register;
     }
     *result = (n != 0) ? 1.0f - (nk_f32_t)intersection_count / (nk_f32_t)n : 0.0f;
@@ -143,11 +144,11 @@ NK_PUBLIC void nk_hamming_u8_sve(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n
     nk_size_t i = 0;
     nk_u32_t differences = 0;
     while (i < n) {
-        svbool_t active_predicate = svwhilelt_b8_u64(i, n);
-        svuint8_t a_u8 = svld1_u8(active_predicate, a + i);
-        svuint8_t b_u8 = svld1_u8(active_predicate, b + i);
-        svbool_t neq_predicate = svcmpne_u8(active_predicate, a_u8, b_u8);
-        differences += svcntp_b8(active_predicate, neq_predicate);
+        svbool_t active_predicate_b8x = svwhilelt_b8_u64(i, n);
+        svuint8_t a_u8x = svld1_u8(active_predicate_b8x, a + i);
+        svuint8_t b_u8x = svld1_u8(active_predicate_b8x, b + i);
+        svbool_t neq_predicate_b8x = svcmpne_u8(active_predicate_b8x, a_u8x, b_u8x);
+        differences += svcntp_b8(active_predicate_b8x, neq_predicate_b8x);
         i += bytes_per_register;
     }
     *result = differences;
@@ -158,17 +159,17 @@ NK_PUBLIC void nk_jaccard_u16_sve(nk_u16_t const *a, nk_u16_t const *b, nk_size_
     nk_size_t i = 0;
     nk_u32_t intersection_count = 0;
     while (i < n) {
-        svbool_t active_predicate = svwhilelt_b16_u64(i, n);
-        svuint16_t a_u16 = svld1_u16(active_predicate, a + i);
-        svuint16_t b_u16 = svld1_u16(active_predicate, b + i);
-        svbool_t equality_predicate = svcmpeq_u16(active_predicate, a_u16, b_u16);
-        intersection_count += svcntp_b16(active_predicate, equality_predicate);
+        svbool_t active_predicate_b16x = svwhilelt_b16_u64(i, n);
+        svuint16_t a_u16x = svld1_u16(active_predicate_b16x, a + i);
+        svuint16_t b_u16x = svld1_u16(active_predicate_b16x, b + i);
+        svbool_t equality_predicate_b16x = svcmpeq_u16(active_predicate_b16x, a_u16x, b_u16x);
+        intersection_count += svcntp_b16(active_predicate_b16x, equality_predicate_b16x);
         i += halfwords_per_register;
     }
     *result = (n != 0) ? 1.0f - (nk_f32_t)intersection_count / (nk_f32_t)n : 0.0f;
 }
 
-#pragma endregion - Integer Sets
+#pragma endregion Integer Sets
 
 #if defined(__clang__)
 #pragma clang attribute pop

package/include/numkong/set/v128relaxed.h

@@ -30,7 +30,7 @@ extern "C" {
 #pragma clang attribute push(__attribute__((target("relaxed-simd"))), apply_to = function)
 #endif
 
-#pragma region - Binary Sets
+#pragma region Binary Sets
 
 NK_PUBLIC void nk_hamming_u1_v128relaxed(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n, nk_u32_t *result) {
     nk_u8_t const *a_bytes = (nk_u8_t const *)a;
@@ -121,9 +121,9 @@ NK_PUBLIC void nk_jaccard_u1_v128relaxed(nk_u1x8_t const *a, nk_u1x8_t const *b,
     *result = union_count > 0 ? 1.0f - ((nk_f32_t)intersection / (nk_f32_t)union_count) : 0.0f;
 }
 
-#pragma endregion - Binary Sets
+#pragma endregion Binary Sets
 
-#pragma region - Integer Sets
+#pragma region Integer Sets
 
 NK_PUBLIC void nk_hamming_u8_v128relaxed(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk_u32_t *result) {
     nk_u32_t sum_total = 0;
@@ -197,9 +197,9 @@ NK_PUBLIC void nk_jaccard_u16_v128relaxed(nk_u16_t const *a, nk_u16_t const *b,
     *result = (n != 0) ? 1.0f - (nk_f32_t)matches / (nk_f32_t)n : 0.0f;
 }
 
-#pragma endregion - Integer Sets
+#pragma endregion Integer Sets
 
-#pragma region - Binary Sets from Dot
+#pragma region Binary Sets from Dot
 
 NK_INTERNAL void nk_hamming_u32x4_from_dot_v128relaxed_( //
     nk_b128_vec_t dots, nk_u32_t query_pop, nk_b128_vec_t target_pops, nk_b128_vec_t *results) {
@@ -226,7 +226,7 @@ NK_INTERNAL void nk_jaccard_f32x4_from_dot_v128relaxed_( //
     results->v128 = wasm_i32x4_relaxed_laneselect(zero_f32x4, jaccard_f32x4, zero_mask_u32x4);
 }
 
-#pragma endregion - Binary Sets from Dot
+#pragma endregion Binary Sets from Dot
 
 #if defined(__clang__)
 #pragma clang attribute pop

package/include/numkong/set.h

@@ -41,12 +41,12 @@
  *  On binary vectors, when computing Jaccard distance, the CPU often struggles to compute the
  *  large number of required population counts. There are several instructions we should keep in mind:
  *
- *      Intrinsic                   Instruction                     Ice         Genoa
- *      _mm512_popcnt_epi64         VPOPCNTQ (ZMM, K, ZMM)          3cy @ p5     2cy @ p01
- *      _mm512_shuffle_epi8         VPSHUFB (ZMM, ZMM, ZMM)         1cy @ p5     2cy @ p12
- *      _mm512_sad_epu8             VPSADBW (ZMM, ZMM, ZMM)         3cy @ p5     3cy @ p01
- *      _mm512_ternarylogic_epi64   VPTERNLOGQ (ZMM, ZMM, ZMM, I8)  1cy @ p05    1cy @ p0123
- *      _mm512_gf2p8mul_epi8        VGF2P8MULB (ZMM, ZMM, ZMM)      5cy @ p0     3cy @ p01
+ *      Intrinsic                  Instruction                     Icelake    Genoa
+ *      _mm512_popcnt_epi64        VPOPCNTQ (ZMM, K, ZMM)          3cy @ p5   2cy @ p01
+ *      _mm512_shuffle_epi8        VPSHUFB (ZMM, ZMM, ZMM)         1cy @ p5   2cy @ p12
+ *      _mm512_sad_epu8            VPSADBW (ZMM, ZMM, ZMM)         3cy @ p5   3cy @ p01
+ *      _mm512_ternarylogic_epi64  VPTERNLOGQ (ZMM, ZMM, ZMM, I8)  1cy @ p05  1cy @ p0123
+ *      _mm512_gf2p8mul_epi8       VGF2P8MULB (ZMM, ZMM, ZMM)      5cy @ p0   3cy @ p01
  *
  *  On Ice Lake, VPOPCNTQ bottlenecks on port 5. On AMD Genoa/Turin, it dual-issues
  *  on ports 0-1, making native popcount significantly faster without CSA tricks.
@@ -123,7 +123,7 @@
  *
  *  @section references References
  *
- *  - Intel Intrinsics Guide: https://www.intel.com/content/www/us/en/docs/intrinsics-guide/
+ *  - Intel Intrinsics Guide: https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html
  *  - Arm Intrinsics Reference: https://developer.arm.com/architectures/instruction-sets/intrinsics/
  *  - Muła et al. "Faster Population Counts": https://arxiv.org/pdf/1611.07612
  *  - Muła SSE POPCOUNT experiments: https://github.com/WojciechMula/sse-popcount
@@ -344,9 +344,11 @@ NK_INTERNAL nk_dtype_t nk_jaccard_output_dtype(nk_dtype_t dtype) {
 #include "numkong/set/sve.h"
 #include "numkong/set/icelake.h"
 #include "numkong/set/haswell.h"
+#include "numkong/set/powervsx.h"
 #include "numkong/set/v128relaxed.h"
 #include "numkong/set/rvv.h"
 #include "numkong/set/rvvbb.h"
+#include "numkong/set/loongsonasx.h"
 
 #if defined(__cplusplus)
 extern "C" {
@@ -365,6 +367,10 @@ NK_PUBLIC void nk_hamming_u1(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n
     nk_hamming_u1_icelake(a, b, n, result);
 #elif NK_TARGET_HASWELL
     nk_hamming_u1_haswell(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_hamming_u1_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_hamming_u1_loongsonasx(a, b, n, result);
 #elif NK_TARGET_RVVBB
     nk_hamming_u1_rvvbb(a, b, n, result);
 #elif NK_TARGET_RVV
@@ -385,6 +391,10 @@ NK_PUBLIC void nk_jaccard_u1(nk_u1x8_t const *a, nk_u1x8_t const *b, nk_size_t n
     nk_jaccard_u1_icelake(a, b, n, result);
 #elif NK_TARGET_HASWELL
     nk_jaccard_u1_haswell(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_jaccard_u1_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_jaccard_u1_loongsonasx(a, b, n, result);
 #elif NK_TARGET_RVVBB
     nk_jaccard_u1_rvvbb(a, b, n, result);
 #elif NK_TARGET_RVV
@@ -423,6 +433,10 @@ NK_PUBLIC void nk_hamming_u8(nk_u8_t const *a, nk_u8_t const *b, nk_size_t n, nk
     nk_hamming_u8_icelake(a, b, n, result);
 #elif NK_TARGET_HASWELL
     nk_hamming_u8_haswell(a, b, n, result);
+#elif NK_TARGET_POWERVSX
+    nk_hamming_u8_powervsx(a, b, n, result);
+#elif NK_TARGET_LOONGSONASX
+    nk_hamming_u8_loongsonasx(a, b, n, result);
 #elif NK_TARGET_RVV
     nk_hamming_u8_rvv(a, b, n, result);
 #else

package/include/numkong/sets/README.md

@@ -4,17 +4,17 @@ NumKong implements batched M×N Hamming and Jaccard distance matrices for binary
 
 Hamming distance from batched dot products:
 
-```math
+$$
 D_{ij} = \|A_i\|_1 + \|B_j\|_1 - 2 \cdot \text{dot}(A_i, B_j)
-```
+$$
 
 Where dot = popcount(AND), measuring intersection size.
 
 Jaccard distance from batched dot products:
 
-```math
+$$
 D_{ij} = 1 - \frac{\text{dot}(A_i, B_j)}{\|A_i\|_1 + \|B_j\|_1 - \text{dot}(A_i, B_j)}
-```
+$$
 
 Reformulating as Python pseudocode:
 
@@ -112,38 +112,38 @@ Measured with Wasmtime v42 (Cranelift backend).
 | `nk_jaccards_symmetric_u1_serial`      |        3.57 gso/s, 0 ulp |        13.3 gso/s, 0 ulp |        40.6 gso/s, 0 ulp |
 | `nk_jaccards_symmetric_u1_v128relaxed` |        3.65 gso/s, 0 ulp |        13.9 gso/s, 0 ulp |        42.2 gso/s, 0 ulp |
 
-### Apple M4
+### Apple M5
 
 #### Native
 
 | Kernel                             |                     256³ |                    1024³ |                    4096³ |
 | :--------------------------------- | -----------------------: | -----------------------: | -----------------------: |
 | __u1__                             | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
-| `nk_hammings_packed_u1_serial`     |                154 gso/s |                204 gso/s |                221 gso/s |
-| `nk_hammings_symmetric_u1_serial`  |                101 gso/s |                159 gso/s |                172 gso/s |
-| `nk_jaccards_packed_u1_serial`     |         116 gso/s, 0 ulp |         203 gso/s, 0 ulp |         232 gso/s, 0 ulp |
-| `nk_jaccards_symmetric_u1_serial`  |        86.3 gso/s, 0 ulp |         157 gso/s, 0 ulp |         176 gso/s, 0 ulp |
-| `nk_hammings_packed_u1_neon`       |                315 gso/s |                428 gso/s |                481 gso/s |
-| `nk_hammings_symmetric_u1_neon`    |                132 gso/s |                240 gso/s |                294 gso/s |
-| `nk_jaccards_packed_u1_neon`       |       266 gso/s, 8.6 ulp |       416 gso/s, 8.6 ulp |       488 gso/s, 8.6 ulp |
-| `nk_jaccards_symmetric_u1_neon`    |       129 gso/s, 8.5 ulp |       242 gso/s, 8.5 ulp |       294 gso/s, 8.5 ulp |
-| `nk_hammings_packed_u1_smebi32`    |              1,420 gso/s |              2,928 gso/s |              4,027 gso/s |
-| `nk_hammings_symmetric_u1_smebi32` |                629 gso/s |              1,438 gso/s |              1,111 gso/s |
-| `nk_jaccards_packed_u1_smebi32`    |         273 gso/s, 0 ulp |       1,381 gso/s, 0 ulp |       3,280 gso/s, 0 ulp |
-| `nk_jaccards_symmetric_u1_smebi32` |        45.1 gso/s, 0 ulp |         267 gso/s, 0 ulp |         618 gso/s, 0 ulp |
+| `nk_hammings_packed_u1_serial`     |                156 gso/s |                231 gso/s |                262 gso/s |
+| `nk_hammings_symmetric_u1_serial`  |                106 gso/s |                196 gso/s |                246 gso/s |
+| `nk_jaccards_packed_u1_serial`     |         136 gso/s, 0 ulp |         221 gso/s, 0 ulp |         262 gso/s, 0 ulp |
+| `nk_jaccards_symmetric_u1_serial`  |        96.5 gso/s, 0 ulp |         183 gso/s, 0 ulp |         244 gso/s, 0 ulp |
+| `nk_hammings_packed_u1_neon`       |                321 gso/s |                436 gso/s |                508 gso/s |
+| `nk_hammings_symmetric_u1_neon`    |                126 gso/s |                239 gso/s |                318 gso/s |
+| `nk_jaccards_packed_u1_neon`       |         271 gso/s, 0 ulp |         423 gso/s, 0 ulp |         503 gso/s, 0 ulp |
+| `nk_jaccards_symmetric_u1_neon`    |         120 gso/s, 0 ulp |         233 gso/s, 0 ulp |         316 gso/s, 0 ulp |
+| `nk_hammings_packed_u1_smebi32`    |              3,286 gso/s |              7,303 gso/s |             11,269 gso/s |
+| `nk_hammings_symmetric_u1_smebi32` |              1,872 gso/s |              5,332 gso/s |              4,079 gso/s |
+| `nk_jaccards_packed_u1_smebi32`    |         371 gso/s, 0 ulp |       1,735 gso/s, 0 ulp |       4,348 gso/s, 0 ulp |
+| `nk_jaccards_symmetric_u1_smebi32` |        83.1 gso/s, 0 ulp |         358 gso/s, 0 ulp |       1,005 gso/s, 0 ulp |
 
 #### WASM
 
-Measured with Wasmtime v42 (Cranelift backend).
+Measured with Wasmtime v43 (Cranelift backend).
 
 | Kernel                                 |                     256³ |                    1024³ |                    4096³ |
 | :------------------------------------- | -----------------------: | -----------------------: | -----------------------: |
 | __u1__                                 | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
-| `nk_hammings_packed_u1_serial`         |               35.2 gso/s |               47.6 gso/s |               52.8 gso/s |
-| `nk_hammings_symmetric_u1_serial`      |               25.4 gso/s |               51.5 gso/s |                129 gso/s |
-| `nk_jaccards_packed_u1_serial`         |        30.9 gso/s, 0 ulp |        46.0 gso/s, 0 ulp |        52.7 gso/s, 0 ulp |
-| `nk_jaccards_symmetric_u1_serial`      |        22.8 gso/s, 0 ulp |        48.9 gso/s, 0 ulp |         123 gso/s, 0 ulp |
-| `nk_hammings_packed_u1_v128relaxed`    |                102 gso/s |                144 gso/s |                160 gso/s |
-| `nk_hammings_symmetric_u1_v128relaxed` |               28.2 gso/s |               61.7 gso/s |                175 gso/s |
-| `nk_jaccards_packed_u1_v128relaxed`    |        91.2 gso/s, 0 ulp |         140 gso/s, 0 ulp |         172 gso/s, 0 ulp |
-| `nk_jaccards_symmetric_u1_v128relaxed` |        26.9 gso/s, 0 ulp |        60.3 gso/s, 0 ulp |         177 gso/s, 0 ulp |
+| `nk_hammings_packed_u1_serial`         |               99.3 gso/s |                127 gso/s |                154 gso/s |
+| `nk_hammings_symmetric_u1_serial`      |               63.7 gso/s |                142 gso/s |                210 gso/s |
+| `nk_jaccards_packed_u1_serial`         |        92.2 gso/s, 0 ulp |         123 gso/s, 0 ulp |         153 gso/s, 0 ulp |
+| `nk_jaccards_symmetric_u1_serial`      |        59.3 gso/s, 0 ulp |         142 gso/s, 0 ulp |         207 gso/s, 0 ulp |
+| `nk_hammings_packed_u1_v128relaxed`    |                266 gso/s |                378 gso/s |                426 gso/s |
+| `nk_hammings_symmetric_u1_v128relaxed` |               72.2 gso/s |                185 gso/s |                259 gso/s |
+| `nk_jaccards_packed_u1_v128relaxed`    |         243 gso/s, 0 ulp |         370 gso/s, 0 ulp |         424 gso/s, 0 ulp |
+| `nk_jaccards_symmetric_u1_v128relaxed` |        72.9 gso/s, 0 ulp |         183 gso/s, 0 ulp |         257 gso/s, 0 ulp |

package/include/numkong/sets/loongsonasx.h

@@ -0,0 +1,52 @@
+/**
+ *  @brief Batched Set Operations for LoongArch LASX (256-bit).
+ *  @file include/numkong/sets/loongsonasx.h
+ *  @author Ash Vardanian
+ *  @date March 25, 2026
+ *
+ *  @sa include/numkong/sets.h
+ */
+#ifndef NK_SETS_LOONGSONASX_H
+#define NK_SETS_LOONGSONASX_H
+
+#if NK_TARGET_LOONGARCH_
+#if NK_TARGET_LOONGSONASX
+
+#include "numkong/set/loongsonasx.h"
+#include "numkong/dots/loongsonasx.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+nk_define_cross_normalized_packed_(hamming, u1, loongsonasx, u1x8, u1x8, u32, /*norm_value_type=*/u32, u32,
+                                   nk_b128_vec_t, nk_dots_packed_u1_loongsonasx, nk_hamming_u32x4_from_dot_loongsonasx_,
+                                   nk_dots_reduce_sum_u1_, nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_,
+                                   /*dimensions_per_value=*/8)
+
+nk_define_cross_normalized_packed_(jaccard, u1, loongsonasx, u1x8, u1x8, u32, /*norm_value_type=*/u32, f32,
+                                   nk_b128_vec_t, nk_dots_packed_u1_loongsonasx, nk_jaccard_f32x4_from_dot_loongsonasx_,
+                                   nk_dots_reduce_sum_u1_, nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                   nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_,
+                                   /*dimensions_per_value=*/8)
+
+nk_define_cross_normalized_symmetric_(hamming, u1, loongsonasx, u1x8, u32, /*norm_value_type=*/u32, u32, nk_b128_vec_t,
+                                      nk_dots_symmetric_u1_loongsonasx, nk_hamming_u32x4_from_dot_loongsonasx_,
+                                      nk_dots_reduce_sum_u1_, nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_,
+                                      /*dimensions_per_value=*/8)
+
+nk_define_cross_normalized_symmetric_(jaccard, u1, loongsonasx, u1x8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_u1_loongsonasx, nk_jaccard_f32x4_from_dot_loongsonasx_,
+                                      nk_dots_reduce_sum_u1_, nk_load_b128_loongsonasx_, nk_partial_load_b32x4_serial_,
+                                      nk_store_b128_loongsonasx_, nk_partial_store_b32x4_serial_,
+                                      /*dimensions_per_value=*/8)
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#endif // NK_TARGET_LOONGSONASX
+#endif // NK_TARGET_LOONGARCH_
+#endif // NK_SETS_LOONGSONASX_H

package/include/numkong/sets/powervsx.h

@@ -0,0 +1,65 @@
+/**
+ *  @brief Batched Set Operations for Power VSX.
+ *  @file include/numkong/sets/powervsx.h
+ *  @author Ash Vardanian
+ *  @date March 23, 2026
+ *
+ *  @sa include/numkong/sets.h
+ */
+#ifndef NK_SETS_POWERVSX_H
+#define NK_SETS_POWERVSX_H
+
+#if NK_TARGET_POWER_
+#if NK_TARGET_POWERVSX
+
+#include "numkong/set/powervsx.h"
+#include "numkong/dots/powervsx.h"
+
+#if defined(__clang__)
+#pragma clang attribute push(__attribute__((target("power9-vector"))), apply_to = function)
+#elif defined(__GNUC__)
+#pragma GCC push_options
+#pragma GCC target("power9-vector")
+#endif
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+nk_define_cross_normalized_packed_(hamming, u1, powervsx, u1x8, u1x8, u32, /*norm_value_type=*/u32, u32, nk_b128_vec_t,
+                                   nk_dots_packed_u1_powervsx, nk_hamming_u32x4_from_dot_powervsx_,
+                                   nk_dots_reduce_sum_u1_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                   nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_,
+                                   /*dimensions_per_value=*/8)
+
+nk_define_cross_normalized_packed_(jaccard, u1, powervsx, u1x8, u1x8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                   nk_dots_packed_u1_powervsx, nk_jaccard_f32x4_from_dot_powervsx_,
+                                   nk_dots_reduce_sum_u1_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                   nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_,
+                                   /*dimensions_per_value=*/8)
+
+nk_define_cross_normalized_symmetric_(hamming, u1, powervsx, u1x8, u32, /*norm_value_type=*/u32, u32, nk_b128_vec_t,
+                                      nk_dots_symmetric_u1_powervsx, nk_hamming_u32x4_from_dot_powervsx_,
+                                      nk_dots_reduce_sum_u1_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                      nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_,
+                                      /*dimensions_per_value=*/8)
+
+nk_define_cross_normalized_symmetric_(jaccard, u1, powervsx, u1x8, u32, /*norm_value_type=*/u32, f32, nk_b128_vec_t,
+                                      nk_dots_symmetric_u1_powervsx, nk_jaccard_f32x4_from_dot_powervsx_,
+                                      nk_dots_reduce_sum_u1_, nk_load_b128_powervsx_, nk_partial_load_b32x4_powervsx_,
+                                      nk_store_b128_powervsx_, nk_partial_store_b32x4_powervsx_,
+                                      /*dimensions_per_value=*/8)
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
+
+#if defined(__clang__)
+#pragma clang attribute pop
+#elif defined(__GNUC__)
+#pragma GCC pop_options
+#endif
+
+#endif // NK_TARGET_POWERVSX
+#endif // NK_TARGET_POWER_
+#endif // NK_SETS_POWERVSX_H