New bitshuffle functions

.github/workflows/cmake.yml

@@ -71,6 +71,11 @@ jobs:
             compiler: clang
             cmake-args: -D DEACTIVATE_AVX2=ON
 
+          - name: Ubuntu Clang No AVX512
+            os: ubuntu-latest
+            compiler: clang
+            cmake-args: -D DEACTIVATE_AVX512=ON
+
           - name: Ubuntu Clang No ZLIB
             os: ubuntu-latest
             compiler: clang

CMakeLists.txt

@@ -29,6 +29,8 @@
 #       build a lite version (only with BloscLZ and LZ4/LZ4HC) of the blosc library
 #   DEACTIVATE_AVX2: default OFF
 #       do not attempt to build with AVX2 instructions
+#   DEACTIVATE_AVX512: default OFF
+#       do not attempt to build with AVX512 instructions
 #   DEACTIVATE_ZLIB: default OFF
 #       do not include support for the Zlib library
 #   DEACTIVATE_ZSTD: default OFF
@@ -115,6 +117,8 @@ option(BUILD_LITE
     "Build a lite version (only with BloscLZ and LZ4/LZ4HC) of the blosc library." OFF)
 option(DEACTIVATE_AVX2
     "Do not attempt to build with AVX2 instructions" OFF)
+option(DEACTIVATE_AVX512
+    "Do not attempt to build with AVX512 instructions" OFF)
 option(DEACTIVATE_ZLIB
     "Do not include support for the Zlib library." OFF)
 option(DEACTIVATE_ZSTD
@@ -281,30 +285,42 @@ if(CMAKE_SYSTEM_PROCESSOR STREQUAL i386 OR
         else()
             set(COMPILER_SUPPORT_AVX2 FALSE)
         endif()
+        if(CMAKE_C_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 4.9)
+            set(COMPILER_SUPPORT_AVX512 TRUE)
+        else()
+            set(COMPILER_SUPPORT_AVX512 FALSE)
+        endif()
     elseif(CMAKE_C_COMPILER_ID STREQUAL Clang OR CMAKE_C_COMPILER_ID STREQUAL AppleClang)
         set(COMPILER_SUPPORT_SSE2 TRUE)
         if(CMAKE_C_COMPILER_VERSION VERSION_GREATER 3.2 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 3.2)
             set(COMPILER_SUPPORT_AVX2 TRUE)
         else()
             set(COMPILER_SUPPORT_AVX2 FALSE)
         endif()
-    elseif(CMAKE_C_COMPILER_ID STREQUAL Intel)
-        set(COMPILER_SUPPORT_SSE2 TRUE)
-        if(CMAKE_C_COMPILER_VERSION VERSION_GREATER 14.0 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 14.0)
-            set(COMPILER_SUPPORT_AVX2 TRUE)
+        if(CMAKE_C_COMPILER_VERSION VERSION_GREATER 10.0 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 10.0)
+            set(COMPILER_SUPPORT_AVX512 TRUE)
         else()
-            set(COMPILER_SUPPORT_AVX2 FALSE)
+            set(COMPILER_SUPPORT_AVX512 FALSE)
         endif()
+    elseif(CMAKE_C_COMPILER_ID STREQUAL Intel)
+        # All Intel compilers since the introduction of AVX512 in 2016 should support it, so activate all SIMD flavors
+        set(COMPILER_SUPPORT_SSE2 TRUE)
+        set(COMPILER_SUPPORT_AVX2 TRUE)
+        set(COMPILER_SUPPORT_AVX512 TRUE)
     elseif(MSVC)
         set(COMPILER_SUPPORT_SSE2 TRUE)
         if(CMAKE_C_COMPILER_VERSION VERSION_GREATER 18.00.30501 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 18.00.30501)
             set(COMPILER_SUPPORT_AVX2 TRUE)
+        # AVX512 starts to be supported since Visual Studio 17 15.0
+        elseif(CMAKE_C_COMPILER_VERSION VERSION_GREATER 19.10.25017 OR CMAKE_C_COMPILER_VERSION VERSION_EQUAL 19.10.25017)
+                set(COMPILER_SUPPORT_AVX512 TRUE)
         else()
             set(COMPILER_SUPPORT_AVX2 FALSE)
         endif()
     else()
         set(COMPILER_SUPPORT_SSE2 FALSE)
         set(COMPILER_SUPPORT_AVX2 FALSE)
+        set(COMPILER_SUPPORT_AVX512 FALSE)
         # Unrecognized compiler. Emit a warning message to let the user know hardware-acceleration won't be available.
         message(WARNING "Unable to determine which ${CMAKE_SYSTEM_PROCESSOR} hardware features are supported by the C compiler (${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}).")
     endif()
@@ -345,6 +361,13 @@ endif()
 # disable AVX2 if specified
 if(DEACTIVATE_AVX2)
     set(COMPILER_SUPPORT_AVX2 FALSE)
+    # AVX512 functions in bitshuffle depend on AVX2 too
+    set(COMPILER_SUPPORT_AVX512 FALSE)
+endif()
+
+# disable AVX512 if specified
+if(DEACTIVATE_AVX512)
+    set(COMPILER_SUPPORT_AVX512 FALSE)
 endif()
 
 # flags

blosc/CMakeLists.txt

@@ -262,7 +262,7 @@ list(APPEND SOURCES
     blosc/directories.c
     blosc/blosc2-stdio.c
     blosc/b2nd.c
-    blosc/b2nd_utils.c    
+    blosc/b2nd_utils.c
 )
 if(NOT CMAKE_SYSTEM_PROCESSOR STREQUAL arm64)
     if(COMPILER_SUPPORT_SSE2)
@@ -273,6 +273,10 @@ if(NOT CMAKE_SYSTEM_PROCESSOR STREQUAL arm64)
         message(STATUS "Adding run-time support for AVX2")
         list(APPEND SOURCES blosc/shuffle-avx2.c blosc/bitshuffle-avx2.c)
     endif()
+    if(COMPILER_SUPPORT_AVX512)
+        message(STATUS "Adding run-time support for AVX512")
+        list(APPEND SOURCES blosc/bitshuffle-avx512.c)
+    endif()
 endif()
 if(COMPILER_SUPPORT_NEON)
     message(STATUS "Adding run-time support for NEON")
@@ -343,6 +347,30 @@ if(COMPILER_SUPPORT_AVX2)
             SOURCE shuffle.c
             APPEND PROPERTY COMPILE_DEFINITIONS SHUFFLE_AVX2_ENABLED)
 endif()
+if(COMPILER_SUPPORT_AVX512)
+    if(MSVC)
+        set_source_files_properties(
+                bitshuffle-avx512.c
+		PROPERTIES COMPILE_OPTIONS "/arch:AVX512")
+        set_property(
+                SOURCE shuffle.c
+		APPEND PROPERTY COMPILE_OPTIONS "/arch:AVX512")
+    else()
+        set_source_files_properties(
+                bitshuffle-avx512.c
+                PROPERTIES COMPILE_OPTIONS "-mavx512f;-mavx512bw")
+        set_property(
+                SOURCE shuffle.c
+                APPEND PROPERTY COMPILE_OPTIONS "-mavx512f;-mavx512bw")
+    endif()
+
+    # Define a symbol for the shuffle-dispatch implementation
+    # so it knows AVX512 is supported even though that file is
+    # compiled without AVX512 support (for portability).
+    set_property(
+            SOURCE shuffle.c
+            APPEND PROPERTY COMPILE_DEFINITIONS SHUFFLE_AVX512_ENABLED)
+endif()
 if(COMPILER_SUPPORT_NEON)
     set_source_files_properties(
             shuffle-neon.c bitshuffle-neon.c

blosc/bitshuffle-altivec.c

@@ -258,18 +258,6 @@ int64_t bshuf_trans_byte_elem_128(void* in, void* out, const size_t size) {
 }
 
 
-/* Memory copy with bshuf call signature. */
-int64_t bshuf_copy(void* in, void* out, const size_t size,
-                   const size_t elem_size) {
-
-  char* in_b = (char*)in;
-  char* out_b = (char*)out;
-
-  memcpy(out_b, in_b, size * elem_size);
-  return size * elem_size;
-}
-
-
 /* Transpose bytes within elements using best SSE algorithm available. */
 int64_t bshuf_trans_byte_elem_altivec(void* in, void* out, const size_t size,
                                       const size_t elem_size, void* tmp_buf) {
@@ -373,18 +361,24 @@ int64_t bshuf_trans_bit_byte_altivec(void* in, void* out, const size_t size,
 
 /* Transpose bits within elements. */
 int64_t bshuf_trans_bit_elem_altivec(void* in, void* out, const size_t size,
-                                     const size_t elem_size, void* tmp_buf) {
+                                     const size_t elem_size) {
 
   int64_t count;
 
   CHECK_MULT_EIGHT(size);
 
+  void* tmp_buf = malloc(size * elem_size);
+  if (tmp_buf == NULL) return -1;
+
   count = bshuf_trans_byte_elem_altivec(in, out, size, elem_size, tmp_buf);
   CHECK_ERR(count);
   // bshuf_trans_bit_byte_altivec / bitshuffle1_altivec
   count = bshuf_trans_bit_byte_altivec(out, tmp_buf, size, elem_size);
   CHECK_ERR(count);
   count = bshuf_trans_bitrow_eight(tmp_buf, out, size, elem_size);
+
+  free(tmp_buf);
+
   return count;
 }
 
@@ -580,16 +574,20 @@ int64_t bshuf_shuffle_bit_eightelem_altivec(void* in, void* out, const size_t si
 
 /* Untranspose bits within elements. */
 int64_t bshuf_untrans_bit_elem_altivec(void* in, void* out, const size_t size,
-                                       const size_t elem_size, void* tmp_buf) {
+                                       const size_t elem_size) {
 
   int64_t count;
 
   CHECK_MULT_EIGHT(size);
 
+  void* tmp_buf = malloc(size * elem_size);
+  if (tmp_buf == NULL) return -1;
+
   count = bshuf_trans_byte_bitrow_altivec(in, tmp_buf, size, elem_size);
   CHECK_ERR(count);
   count = bshuf_shuffle_bit_eightelem_altivec(tmp_buf, out, size, elem_size);
 
+  free(tmp_buf);
   return count;
 }