Merge pull request #181 from ashvardanian/main-dev

MSVC Builds & Python Argument Parsing

CMakeLists.txt

@@ -327,6 +327,16 @@ if(${STRINGZILLA_BUILD_SHARED})
         "SZ_USE_ARM_NEON=1"
         "SZ_USE_ARM_SVE=1")
     endif()
+
+    if (MSVC)
+      # Add dependencies for necessary runtime libraries in case of static linking
+      # This ensures that basic runtime functions are available:
+      # msvcrt.lib: Microsoft Visual C Runtime, required for basic C runtime functions on Windows.
+      # vcruntime.lib: Microsoft Visual C++ Runtime library for basic runtime functions.
+      # ucrt.lib: Universal C Runtime, necessary for linking basic C functions like I/O.
+      target_link_libraries(${target} PRIVATE msvcrt.lib vcruntime.lib ucrt.lib)
+    endif()
+
   endfunction()
 
   define_shared(stringzilla_shared)
@@ -344,4 +354,6 @@ if(${STRINGZILLA_BUILD_SHARED})
     "$<$<CXX_COMPILER_ID:MSVC>:/Oi-;/GS->")
   target_link_options(stringzillite PRIVATE "$<$<CXX_COMPILER_ID:GNU,Clang>:-nostdlib>")
   target_link_options(stringzillite PRIVATE "$<$<CXX_COMPILER_ID:MSVC>:/NODEFAULTLIB>")
+
+
 endif()

README.md

@@ -171,7 +171,7 @@ __Who is this for?__
       <span style="color:#ABABAB;">arm:</span> <b>9.4</b> MB/s
     </td>
     <td align="center">
-      <code>uniform_int_distribution</code><br/>
+      <code>std::uniform_int_distribution</code><br/>
       <span style="color:#ABABAB;">x86:</span> <b>47.2</b> &centerdot;
       <span style="color:#ABABAB;">arm:</span> <b>20.4</b> MB/s
     </td>
@@ -193,7 +193,7 @@ __Who is this for?__
   <tr>
     <td align="center">⚪</td>
     <td align="center">
-      <code>transform</code><br/>
+      <code>std::transform</code><br/>
       <span style="color:#ABABAB;">x86:</span> <b>3.81</b> &centerdot;
       <span style="color:#ABABAB;">arm:</span> <b>2.65</b> GB/s
     </td>

c/lib.c

@@ -232,21 +232,20 @@ static void sz_dispatch_table_init(void) {
 }
 
 #if defined(_MSC_VER)
+#pragma section(".CRT$XCU", read)
+__declspec(allocate(".CRT$XCU")) void (*_sz_dispatch_table_init)() = sz_dispatch_table_init;
+
 BOOL WINAPI DllMain(HINSTANCE hints, DWORD forward_reason, LPVOID lp) {
     switch (forward_reason) {
-    case DLL_PROCESS_ATTACH: sz_dispatch_table_init(); return TRUE;
+    case DLL_PROCESS_ATTACH:
+        sz_dispatch_table_init(); // Ensure initialization
+        return TRUE;
     case DLL_THREAD_ATTACH: return TRUE;
     case DLL_THREAD_DETACH: return TRUE;
     case DLL_PROCESS_DETACH: return TRUE;
     }
-}
-
-#if SZ_AVOID_LIBC
-BOOL WINAPI _DllMainCRTStartup(HINSTANCE hints, DWORD forward_reason, LPVOID lp) {
-    DllMain(hints, forward_reason, lp);
     return TRUE;
 }
-#endif
 
 #else
 __attribute__((constructor)) static void sz_dispatch_table_init_on_gcc_or_clang(void) { sz_dispatch_table_init(); }

include/stringzilla/stringzilla.h

@@ -5323,8 +5323,16 @@ SZ_PUBLIC void sz_look_up_transform_avx512(sz_cptr_t source, sz_size_t length, s
     // operate on 4 registers, it might be cleaner to use 2x separate `_mm512_permutexvar_epi8` calls.
     // Combining the results with 2x `_mm512_test_epi8_mask` and 3x blends afterwards.
     //
-    //  - `_mm512_mask_blend_epi8` - 1 cycle latency, and generally 2x can run in parallel.
-    //  - `_mm512_test_epi8_mask` - 3 cycles latency, same as most comparison functions in AVX-512.
+    //  - 4x `_mm512_permutexvar_epi8` maps to "VPERMB (ZMM, ZMM, ZMM)":
+    //      - On Ice Lake: 3 cycles latency, ports: 1*p5
+    //      - On Genoa: 6 cycles latency, ports: 1*FP12
+    //  - 3x `_mm512_mask_blend_epi8` maps to "VPBLENDMB_Z (ZMM, K, ZMM, ZMM)":
+    //      - On Ice Lake: 3 cycles latency, ports: 1*p05
+    //      - On Genoa: 1 cycle latency, ports: 1*FP0123
+    //  - 2x `_mm512_test_epi8_mask` maps to "VPTESTMB (K, ZMM, ZMM)":
+    //      - On Ice Lake: 3 cycles latency, ports: 1*p5
+    //      - On Genoa: 4 cycles latency, ports: 1*FP01
+    //
     sz_u512_vec_t lut_0_to_63_vec, lut_64_to_127_vec, lut_128_to_191_vec, lut_192_to_255_vec;
     lut_0_to_63_vec.zmm = _mm512_loadu_si512((lut));
     lut_64_to_127_vec.zmm = _mm512_loadu_si512((lut + 64));