stringzillas-cuda 4.2.1__tar.gz → 4.2.2__tar.gz

{stringzillas_cuda-4.2.1/stringzillas_cuda.egg-info → stringzillas_cuda-4.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: stringzillas-cuda
-Version: 4.2.1
+Version: 4.2.2
 Summary: Search, hash, sort, fingerprint, and fuzzy-match strings faster via SWAR, SIMD, and CUDA on Nvidia GPUs
 Home-page: https://github.com/ashvardanian/StringZilla
 Author: Ash Vardanian
@@ -30,7 +30,7 @@ Classifier: Topic :: Text Processing :: Indexing
 Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: stringzilla==4.2.1
+Requires-Dist: stringzilla==4.2.2
 Dynamic: author
 Dynamic: author-email
 Dynamic: classifier

stringzillas_cuda-4.2.2/VERSION

@@ -0,0 +1 @@
+4.2.2

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/c/stringzilla.c

@@ -244,7 +244,8 @@ static void sz_dispatch_table_update_implementation_(sz_capability_t caps) {
         impl->copy = sz_copy_sve;
         impl->move = sz_move_sve;
         impl->fill = sz_fill_sve;
-
+        impl->lookup = sz_lookup_sve;
+	
         impl->find = sz_find_sve;
         // TODO: impl->rfind = sz_rfind_sve;
         impl->find_byte = sz_find_byte_sve;

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/include/stringzilla/compare.h

@@ -432,10 +432,10 @@ SZ_PUBLIC sz_bool_t sz_equal_skylake(sz_cptr_t a, sz_cptr_t b, sz_size_t length)
 #pragma region NEON Implementation
 #if SZ_USE_NEON
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+simd"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+simd"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+simd")
+#pragma GCC target("+simd")
 #endif
 
 SZ_PUBLIC sz_ordering_t sz_order_neon(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length) {
@@ -479,10 +479,10 @@ SZ_PUBLIC sz_bool_t sz_equal_neon(sz_cptr_t a, sz_cptr_t b, sz_size_t length) {
 #pragma region SVE Implementation
 #if SZ_USE_SVE
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+sve"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+sve"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+sve")
+#pragma GCC target("+sve")
 #endif
 
 SZ_PUBLIC sz_bool_t sz_equal_sve(sz_cptr_t a, sz_cptr_t b, sz_size_t length) {

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/include/stringzilla/find.h

@@ -1578,10 +1578,10 @@ SZ_PUBLIC sz_cptr_t sz_rfind_byteset_ice(sz_cptr_t text, sz_size_t length, sz_by
 #pragma region NEON Implementation
 #if SZ_USE_NEON
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+simd"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+simd"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+simd")
+#pragma GCC target("+simd")
 #endif
 
 SZ_INTERNAL sz_u64_t sz_vreinterpretq_u8_u4_(uint8x16_t vec) {
@@ -1813,10 +1813,10 @@ SZ_PUBLIC sz_cptr_t sz_rfind_byteset_neon(sz_cptr_t h, sz_size_t h_length, sz_by
 #pragma region SVE Implementation
 #if SZ_USE_SVE
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+sve"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+sve"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+sve")
+#pragma GCC target("+sve")
 #endif
 
 SZ_PUBLIC sz_cptr_t sz_find_byte_sve(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/include/stringzilla/hash.h

@@ -3132,10 +3132,10 @@ SZ_PUBLIC void sz_sha256_state_digest_ice(sz_sha256_state_t const *state_ptr, sz
 #pragma region NEON Implementation
 #if SZ_USE_NEON
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+simd"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+simd"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+simd")
+#pragma GCC target("+simd")
 #endif
 
 SZ_PUBLIC sz_u64_t sz_bytesum_neon(sz_cptr_t text, sz_size_t length) {
@@ -3167,10 +3167,10 @@ SZ_PUBLIC sz_u64_t sz_bytesum_neon(sz_cptr_t text, sz_size_t length) {
 #pragma region NEON AES Implementation
 #if SZ_USE_NEON_AES
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+simd+crypto+aes"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+simd+crypto+aes"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+simd+crypto+aes")
+#pragma GCC target("+simd+crypto+aes")
 #endif
 
 /**
@@ -3614,10 +3614,10 @@ SZ_PUBLIC void sz_fill_random_neon(sz_ptr_t text, sz_size_t length, sz_u64_t non
 #pragma region NEON SHA Implementation
 #if SZ_USE_NEON_SHA
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+simd+crypto+sha2"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+simd+crypto+sha2"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+simd+crypto+sha2")
+#pragma GCC target("+simd+crypto+sha2")
 #endif
 
 /**
@@ -3917,10 +3917,10 @@ SZ_PUBLIC void sz_sha256_state_digest_neon(sz_sha256_state_t const *state_ptr, s
 #pragma region SVE Implementation
 #if SZ_USE_SVE
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+sve"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+sve"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+sve")
+#pragma GCC target("+sve")
 #endif
 
 SZ_PUBLIC sz_u64_t sz_bytesum_sve(sz_cptr_t text, sz_size_t length) {
@@ -3955,10 +3955,10 @@ SZ_PUBLIC sz_u64_t sz_bytesum_sve(sz_cptr_t text, sz_size_t length) {
 #pragma region SVE2 Implementation
 #if SZ_USE_SVE2
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+sve+sve2"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+sve+sve2"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+sve+sve2")
+#pragma GCC target("+sve+sve2")
 #endif
 
 SZ_PUBLIC sz_u64_t sz_bytesum_sve2(sz_cptr_t text, sz_size_t length) {
@@ -3997,10 +3997,10 @@ SZ_PUBLIC sz_u64_t sz_bytesum_sve2(sz_cptr_t text, sz_size_t length) {
 #pragma region SVE2 AES Implementation
 #if SZ_USE_SVE2_AES
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+sve+sve2+sve2-aes"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+sve+sve2+sve2-aes"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+sve+sve2+sve2-aes")
+#pragma GCC target("+sve+sve2+sve2-aes")
 #endif
 
 /**

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/include/stringzilla/intersect.h

@@ -736,10 +736,10 @@ SZ_PUBLIC sz_status_t sz_sequence_intersect_ice(
 #pragma region SVE Implementation
 #if SZ_USE_SVE
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+sve"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+sve"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+sve")
+#pragma GCC target("+sve")
 #endif
 
 SZ_PUBLIC sz_status_t sz_sequence_intersect_sve(sz_sequence_t const *first_sequence,

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/include/stringzilla/memory.h

@@ -1120,10 +1120,10 @@ SZ_PUBLIC void sz_lookup_ice(sz_ptr_t target, sz_size_t length, sz_cptr_t source
 #pragma region NEON Implementation
 #if SZ_USE_NEON
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+simd"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+simd"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+simd")
+#pragma GCC target("+simd")
 #endif
 
 SZ_PUBLIC void sz_copy_neon(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
@@ -1248,10 +1248,10 @@ SZ_PUBLIC void sz_lookup_neon(sz_ptr_t target, sz_size_t length, sz_cptr_t sourc
 #pragma region SVE Implementation
 #if SZ_USE_SVE
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+sve"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+sve"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+sve")
+#pragma GCC target("+sve")
 #endif
 
 SZ_PUBLIC void sz_fill_sve(sz_ptr_t target, sz_size_t length, sz_u8_t value) {
@@ -1384,6 +1384,71 @@ SZ_PUBLIC void sz_move_sve(sz_ptr_t target, sz_cptr_t source, sz_size_t length)
     }
 }
 
+SZ_PUBLIC void sz_lookup_sve(sz_ptr_t target, sz_size_t length, sz_cptr_t source, sz_cptr_t lut) {
+
+    if (length <= 128) {
+        sz_lookup_serial(target, length, source, lut);
+        return;
+    }
+
+    // SVE vector length in bytes
+    sz_size_t vl = svcntb();
+
+    // Load the 256-byte lookup table into 4 SVE vectors
+    svuint8_t lut_0_to_63_vec = svld1_u8(svptrue_b8(), (sz_u8_t const *)(lut + 0));
+    svuint8_t lut_64_to_127_vec = svld1_u8(svptrue_b8(), (sz_u8_t const *)(lut + 64));
+    svuint8_t lut_128_to_191_vec = svld1_u8(svptrue_b8(), (sz_u8_t const *)(lut + 128));
+    svuint8_t lut_192_to_255_vec = svld1_u8(svptrue_b8(), (sz_u8_t const *)(lut + 192));
+
+    svuint8_t mask_0x3f = svdup_u8(0x3f);
+
+    sz_size_t i = 0;
+
+    // Main loop: process full vectors
+    while (i + vl <= length) {
+        svuint8_t source_vec = svld1_u8(svptrue_b8(), (sz_u8_t const *)(source + i));
+
+        // Create predicates based on top 2 bits (which 64-byte range)
+        svbool_t pred_0_63 = svcmplt_n_u8(svptrue_b8(), source_vec, 64);
+        svbool_t pred_64_127 = svcmpge_n_u8(svcmplt_n_u8(svptrue_b8(), source_vec, 128), source_vec, 64);
+        svbool_t pred_128_191 = svcmpge_n_u8(svcmplt_n_u8(svptrue_b8(), source_vec, 192), source_vec, 128);
+        svbool_t pred_192_255 = svcmpge_n_u8(svptrue_b8(), source_vec, 192);
+
+        // Mask indices to bottom 6 bits for indexing within each 64-byte table
+        svuint8_t idx = svand_u8_x(svptrue_b8(), source_vec, mask_0x3f);
+
+        // Perform lookups and blend results based on predicates
+        svuint8_t result = svsel_u8(pred_0_63, svtbl_u8(lut_0_to_63_vec, idx), svdup_u8(0));
+        result = svsel_u8(pred_64_127, svtbl_u8(lut_64_to_127_vec, idx), result);
+        result = svsel_u8(pred_128_191, svtbl_u8(lut_128_to_191_vec, idx), result);
+        result = svsel_u8(pred_192_255, svtbl_u8(lut_192_to_255_vec, idx), result);
+
+        svst1_u8(svptrue_b8(), (sz_u8_t *)target + i, result);
+        i += vl;
+    }
+
+    // Handle tail: process remaining elements with predicated operations
+    if (i < length) {
+        svbool_t pred = svwhilelt_b8_u64(i, length);
+        svuint8_t source_vec = svld1_u8(pred, (sz_u8_t const *)(source + i));
+
+        // Create predicates for each range (comparison already uses pred as governing predicate)
+        svbool_t pred_0_63 = svcmplt_n_u8(pred, source_vec, 64);
+        svbool_t pred_64_127 = svcmpge_n_u8(svcmplt_n_u8(pred, source_vec, 128), source_vec, 64);
+        svbool_t pred_128_191 = svcmpge_n_u8(svcmplt_n_u8(pred, source_vec, 192), source_vec, 128);
+        svbool_t pred_192_255 = svcmpge_n_u8(pred, source_vec, 192);
+
+        // Mask indices to bottom 6 bits
+        svuint8_t idx = svand_u8_x(pred, source_vec, mask_0x3f);
+
+        svuint8_t result = svsel_u8(pred_0_63, svtbl_u8(lut_0_to_63_vec, idx), svdup_u8(0));
+        result = svsel_u8(pred_64_127, svtbl_u8(lut_64_to_127_vec, idx), result);
+        result = svsel_u8(pred_128_191, svtbl_u8(lut_128_to_191_vec, idx), result);
+        result = svsel_u8(pred_192_255, svtbl_u8(lut_192_to_255_vec, idx), result);
+
+        svst1_u8(pred, (sz_u8_t *)target + i, result);
+    }
+}
 #if defined(__clang__)
 #pragma clang attribute pop
 #elif defined(__GNUC__)

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/include/stringzilla/sort.h

@@ -931,10 +931,10 @@ SZ_PUBLIC sz_status_t sz_sequence_argsort_skylake(sz_sequence_t const *sequence,
 #pragma region SVE Implementation
 #if SZ_USE_SVE
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.2-a+sve"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+sve"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.2-a+sve")
+#pragma GCC target("+sve")
 #endif
 
 /**

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/include/stringzilla/stringzilla.h

@@ -67,7 +67,7 @@
 
 #define STRINGZILLA_H_VERSION_MAJOR 4
 #define STRINGZILLA_H_VERSION_MINOR 2
-#define STRINGZILLA_H_VERSION_PATCH 1
+#define STRINGZILLA_H_VERSION_PATCH 2
 
 #include "types.h"        // `sz_size_t`, `sz_bool_t`, `sz_ordering_t`
 #include "compare.h"      // `sz_equal`, `sz_order`
@@ -105,7 +105,8 @@
 
 /* On Windows ARM, we use IsProcessorFeaturePresent API for capability detection */
 #if defined(SZ_IS_WINDOWS_) && SZ_IS_64BIT_ARM_
-#include <processthreadsapi.h>
+#define NOMINMAX
+#include <windows.h>
 #endif
 
 #ifdef __cplusplus
@@ -261,10 +262,10 @@ SZ_PUBLIC sz_capability_t sz_capabilities_comptime_implementation_(void) {
  *  function targeting newer ones.
  */
 #if defined(__clang__)
-#pragma clang attribute push(__attribute__((target("arch=armv8.5-a+sve"))), apply_to = function)
+#pragma clang attribute push(__attribute__((target("+sve"))), apply_to = function)
 #elif defined(__GNUC__)
 #pragma GCC push_options
-#pragma GCC target("arch=armv8.5-a+sve")
+#pragma GCC target("+sve")
 #endif
 
 #if SZ_HAS_POSIX_EXTENSIONS_

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/include/stringzilla/types.h

@@ -267,7 +267,7 @@
 #if !defined(SZ_USE_NEON)
 #if SZ_IS_64BIT_ARM_ && defined(__ARM_NEON)
 #define SZ_USE_NEON (1)
-#elif SZ_IS_64BIT_X86_ && defined(_MSC_VER) && defined(__ARM_ARCH) && __ARM_ARCH >= 800
+#elif SZ_IS_64BIT_ARM_ && defined(_MSC_VER) && defined(_M_ARM64)
 #define SZ_USE_NEON (1) // ! MSVC doesn't expose `__ARM_NEON` macros
 #else
 #define SZ_USE_NEON (0)

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2}/setup.py

@@ -217,6 +217,7 @@ def is_big_endian() -> bool:
 def linux_settings(use_cpp: bool = False) -> Tuple[List[str], List[str], List[Tuple[str]]]:
     compile_args = [
         "-std=c++17" if use_cpp else "-std=c99",  # use C++17 for StringZillas, C99 for StringZilla
+        "-D_GNU_SOURCE",  # enable POSIX extensions (sigaction, sigjmp_buf, etc.) when using -std=c99
         "-O2",  # optimization level
         "-fdiagnostics-color=always",  # color console output
         "-Wno-unknown-pragmas",  # like: `pragma region` and some unrolls

{stringzillas_cuda-4.2.1 → stringzillas_cuda-4.2.2/stringzillas_cuda.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: stringzillas-cuda
-Version: 4.2.1
+Version: 4.2.2
 Summary: Search, hash, sort, fingerprint, and fuzzy-match strings faster via SWAR, SIMD, and CUDA on Nvidia GPUs
 Home-page: https://github.com/ashvardanian/StringZilla
 Author: Ash Vardanian
@@ -30,7 +30,7 @@ Classifier: Topic :: Text Processing :: Indexing
 Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: stringzilla==4.2.1
+Requires-Dist: stringzilla==4.2.2
 Dynamic: author
 Dynamic: author-email
 Dynamic: classifier

stringzillas_cuda-4.2.2/stringzillas_cuda.egg-info/requires.txt

@@ -0,0 +1 @@
+stringzilla==4.2.2

stringzillas_cuda-4.2.1/VERSION

@@ -1 +0,0 @@
-4.2.1

stringzillas_cuda-4.2.1/stringzillas_cuda.egg-info/requires.txt

@@ -1 +0,0 @@
-stringzilla==4.2.1