Standardize fill_window implementations and abstract out slide_hash_n…

…eon for ARM.
zlib-ng · Apr 30, 2020 · e09d131 · e09d131
1 parent 343596f
commit e09d131
Show file tree

Hide file tree

Showing 23 changed files with 150 additions and 424 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -624,16 +624,19 @@ endif()
 
 if(WITH_OPTIM)
     if(BASEARCH_ARM_FOUND)
-        add_definitions(-DARM_GETAUXVAL)
-        list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/armfeature.c ${ARCHDIR}/fill_window_arm.c)
+        add_definitions(-DARM_CPUID)
+        list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/arm.h)
+        list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/armfeature.c)
         if(WITH_NEON)
-            list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/adler32_neon.c)
-            add_definitions(-DARM_NEON_ADLER32)
+            list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/adler32_neon.h)
+            list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/adler32_neon.c ${ARCHDIR}/slide_neon.c)
+            add_definitions(-DARM_NEON_ADLER32 -DARM_NEON_SLIDEHASH)
             add_intrinsics_option("${NEONFLAG}")
             if(MSVC)
                 add_definitions(-D__ARM_NEON__)
             endif()
-            add_feature_info(NEON_FILLWINDOW 1 "Support NEON instructions in fill_window_arm, using \"${NEONFLAG}\"")
+            add_feature_info(NEON_ALDER32 1 "Support NEON instructions in adler32, using \"${NEONFLAG}\"")
+            add_feature_info(NEON_SLIDEHASH 1 "Support NEON instructions in slide_hash, using \"${NEONFLAG}\"")
         endif()
         if(WITH_ACLE)
             list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/crc32_acle.c ${ARCHDIR}/insert_string_acle.c)
@@ -659,6 +662,7 @@ if(WITH_OPTIM)
         endif()
     elseif(BASEARCH_X86_FOUND)
         add_definitions(-DX86_CPUID)
+        list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/x86.h)
         list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/x86.c)
         if(MSVC)
             list(APPEND ZLIB_ARCH_HDRS fallback_builtins.h)
@@ -685,7 +689,7 @@ if(WITH_OPTIM)
         endif()
         if(WITH_SSE2 AND HAVE_SSE2_INTRIN)
             add_definitions(-DX86_SSE2)
-            list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/fill_window_sse.c ${ARCHDIR}/slide_sse.c)
+            list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/slide_sse.c)
             if(NOT ${ARCH} MATCHES "x86_64")
                 add_intrinsics_option("${SSE2FLAG}")
                 add_feature_info(FORCE_SSE2 FORCE_SSE2 "Assume CPU is SSE2 capable")

diff --git a/arch/arm/Makefile.in b/arch/arm/Makefile.in
@@ -12,7 +12,7 @@ SRCDIR=.
 SRCTOP=../..
 TOPDIR=$(SRCTOP)
 
-all: adler32_neon.o adler32_neon.lo armfeature.o armfeature.lo crc32_acle.o crc32_acle.lo fill_window_arm.o fill_window_arm.lo insert_string_acle.o insert_string_acle.lo
+all: adler32_neon.o adler32_neon.lo armfeature.o armfeature.lo crc32_acle.o crc32_acle.lo slide_neon.o slide_neon.lo insert_string_acle.o insert_string_acle.lo
 
 adler32_neon.o:
 	$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/adler32_neon.c
@@ -32,11 +32,11 @@ crc32_acle.o:
 crc32_acle.lo:
 	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/crc32_acle.c
 
-fill_window_arm.o:
-	$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/fill_window_arm.c
+slide_neon.o:
+	$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/slide_neon.c
 
-fill_window_arm.lo:
-	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/fill_window_arm.c
+slide_neon.lo:
+	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/slide_neon.c
 
 insert_string_acle.o:
 	$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/insert_string_acle.c

diff --git a/arch/arm/fill_window_arm.c b/arch/arm/fill_window_arm.c
diff --git a/arch/arm/slide_neon.c b/arch/arm/slide_neon.c
@@ -0,0 +1,48 @@
+/* slide_neon.c -- Optimized hash table shifting for ARM with support for NEON instructions
+ * Copyright (C) 2017 Mika T. Lindqvist
+ *
+ * Authors:
+ * Mika T. Lindqvist <[email protected]>
+ * Jun He <[email protected]>
+ *
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#if defined(ARM_NEON_SLIDEHASH)
+#include <arm_neon.h>
+#include "../../zbuild.h"
+#include "../../deflate.h"
+
+/* SIMD version of hash_chain rebase */
+static inline void slide_hash_chain(Pos *table, unsigned int entries, uint16_t window_size) {
+    register uint16x8_t v, *p;
+    register size_t n;
+
+    size_t size = entries*sizeof(table[0]);
+    Assert((size % sizeof(uint16x8_t) * 8 == 0), "hash table size err");
+
+    Assert(sizeof(Pos) == 2, "Wrong Pos size");
+    v = vdupq_n_u16(window_size);
+
+    p = (uint16x8_t *)table;
+    n = size / (sizeof(uint16x8_t) * 8);
+    do {
+        p[0] = vqsubq_u16(p[0], v);
+        p[1] = vqsubq_u16(p[1], v);
+        p[2] = vqsubq_u16(p[2], v);
+        p[3] = vqsubq_u16(p[3], v);
+        p[4] = vqsubq_u16(p[4], v);
+        p[5] = vqsubq_u16(p[5], v);
+        p[6] = vqsubq_u16(p[6], v);
+        p[7] = vqsubq_u16(p[7], v);
+        p += 8;
+    } while (--n);
+}
+
+ZLIB_INTERNAL void slide_hash_neon(deflate_state *s) {
+    unsigned int wsize = s->w_size;
+
+    slide_hash_chain(s->head, s->hash_size, wsize);
+    slide_hash_chain(s->prev, wsize, wsize);
+}
+#endif
diff --git a/arch/x86/Makefile.in b/arch/x86/Makefile.in
@@ -17,20 +17,14 @@ SRCDIR=.
 SRCTOP=../..
 TOPDIR=$(SRCTOP)
 
-all: x86.o x86.lo fill_window_sse.o fill_window_sse.lo deflate_quick.o deflate_quick.lo insert_string_sse.o insert_string_sse.lo crc_folding.o crc_folding.lo slide_avx.o slide_avx.lo slide_sse.o slide_sse.lo
+all: x86.o x86.lo deflate_quick.o deflate_quick.lo insert_string_sse.o insert_string_sse.lo crc_folding.o crc_folding.lo slide_avx.o slide_avx.lo slide_sse.o slide_sse.lo
 
 x86.o:
 	$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/x86.c
 
 x86.lo:
 	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/x86.c
 
-fill_window_sse.o:
-	$(CC) $(CFLAGS) $(SSE2FLAG) $(INCLUDES) -c -o $@ $(SRCDIR)/fill_window_sse.c
-
-fill_window_sse.lo:
-	$(CC) $(SFLAGS) $(SSE2FLAG) -DPIC $(INCLUDES) -c -o $@ $(SRCDIR)/fill_window_sse.c
-
 deflate_quick.o:
 	$(CC) $(CFLAGS) $(SSE4FLAG) $(INCLUDES) -c -o $@ $(SRCDIR)/deflate_quick.c
 

diff --git a/arch/x86/README.md b/arch/x86/README.md
@@ -3,6 +3,6 @@ Contents
 
 |Name|Description|
 |:-|:-|
-|fill_window_sse.c|SSE2 optimized fill_window|
 |deflate_quick.c|SSE4 optimized deflate strategy for use as level 1|
 |crc_folding.c|SSE4 + PCLMULQDQ optimized CRC folding implementation|
+|slide_sse2.c|SSE2 optimized slide_hash|
diff --git a/arch/x86/deflate_quick.c b/arch/x86/deflate_quick.c
@@ -30,7 +30,6 @@
 #  include <ctype.h>
 #endif
 
-extern void fill_window_sse(deflate_state *s);
 extern void flush_pending(PREFIX3(stream) *strm);
 
 static inline long compare258(const unsigned char *const src0, const unsigned char *const src1) {
@@ -209,7 +208,7 @@ ZLIB_INTERNAL block_state deflate_quick(deflate_state *s, int flush) {
         }
 
         if (s->lookahead < MIN_LOOKAHEAD) {
-            fill_window_sse(s);
+            fill_window(s);
             if (s->lookahead < MIN_LOOKAHEAD && flush == Z_NO_FLUSH) {
                 static_emit_end_block(s, 0);
                 return need_more;