CRC32 implementation with gcc intrinsics instead pure asm.

.gitignore

@@ -41,3 +41,11 @@ final_fold2_test
 final_fold_constants
 final_fold_test
 slice_by_8_bench
+vec_barrett_reduction_test
+vec_final_fold_test
+vec_final_fold2_test
+vec_crc32_test
+vec_crc32_bench
+crc32_ethernet_constants.h
+crc32k_constants.h
+crc32_two_implementations

Makefile

@@ -1,12 +1,15 @@
-CFLAGS=-m64 -g -O2 -Wall
+ORIG_CFLAGS:= $(CFLAGS)
+
+CFLAGS+=-m64 -g -O2 -mcpu=power8 -mpower8-vector -Wall
 ASFLAGS=-m64 -g
 LDFLAGS=-m64 -g -static
 
+SHELL=/bin/bash
+
 # Ethernet CRC
 #CRC=0x04C11DB7
-#OPTIONS=
 
-# CRC32C
+# CRC32
 CRC=0x11EDC6F41
 OPTIONS=-r -x
 
@@ -31,10 +34,18 @@ PROGS=barrett_reduction_constants \
 
 
 PROGS_ALTIVEC=barrett_reduction_test \
-	final_fold_test \
+	final_fold_test	\
 	final_fold2_test \
-	crc32_test crc32_bench crc32_stress
+	crc32_test \
+	crc32_bench \
+	crc32_stress \
+	vec_barrett_reduction_test \
+	vec_final_fold_test \
+	vec_final_fold2_test \
+	vec_crc32_bench \
+	crc32_two_implementations
 
+CRC32_CONSTANTS_OBJS=crc32_constants.o poly_arithmetic.o crcmodel.o
 ifeq ($(call cc-option-yn,-maltivec),y)
 CFLAGS += -maltivec
 ASFLAGS += -maltivec
@@ -47,21 +58,79 @@ all: $(PROGS)
 barrett_reduction_constants: barrett_reduction_constants.o poly_arithmetic.o
 final_fold_constants: final_fold_constants.o poly_arithmetic.o
 final_fold2_constants: final_fold2_constants.o poly_arithmetic.o
-crc32_constants: crc32_constants.o poly_arithmetic.o crcmodel.o
+crc32_constants: $(CRC32_CONSTANTS_OBJS)
 
 barrett_reduction_test: barrett_reduction_test.o crcmodel.o barrett_reduction.o
 final_fold_test: final_fold_test.o crcmodel.o final_fold.o
 final_fold2_test: final_fold2_test.o crcmodel.o final_fold2.o
 
+vec_barrett_reduction_test: vec_barrett_reduction_test.o crcmodel.o \
+vec_barrett_reduction.o
+vec_final_fold_test: vec_final_fold_test.o crcmodel.o vec_final_fold.o
+vec_final_fold2_test: vec_final_fold2_test.o crcmodel.o vec_final_fold2.o
+
+
+$(CRC32_CONSTANTS_OBJS) : %.o : %.c Makefile
+	$(CC) -c $(ORIG_CFLAGS) $< -o $@
+
 crc32_constants.h: crc32_constants
 	$(EMULATOR) ./crc32_constants $(OPTIONS) $(CRC) > crc32_constants.h
+
+vec_crc32_c.o: vec_crc32.c crc32_constants.h
+	$(CC) -c $(CFLAGS) \
+		-D CRC32_FUNCTION=crc32_vpmsum_c \
+		vec_crc32.c -o $@
 crc32.o: crc32.S crc32_constants.h
 crc32_stress.o: crc32_stress.c crc32_constants.h
 crc32_test.o: crc32_test.c crc32_constants.h
 crc32_wrapper.o: crc32_wrapper.c crc32_constants.h
-crc32_test: crc32_test.o crcmodel.o crc32.o crc32_wrapper.o
-crc32_bench: crc32_bench.o crcmodel.o crc32.o crc32_wrapper.o
+
+slice_by_8_bench: crc32_bench.o
+
+crc32_test: crc32_test.o crcmodel.o crc32.o crc32_wrapper.o vec_crc32_c.o
+crc32_bench: crc32_bench.o crc32.o crc32_wrapper.o
 crc32_stress: crc32_stress.o crcmodel.o crc32.o crc32_wrapper.o
 
+vec_crc32.o: vec_crc32.c crc32_constants.h
+vec_crc32_bench: crc32_bench.o vec_crc32.o
+
+vec_crc32_bench:
+	$(CC) $(LDFLAGS) $^ -o $@
+
+# This is an example of multiple crc32 polynomials being used
+# in a single linked file.
+crc32_ethernet_constants.h: crc32_constants
+	$(EMULATOR) ./crc32_constants -c -r -x 0x4c11db7 > $@
+
+vec_crc32_ethernet.o: vec_crc32.c crc32_ethernet_constants.h
+	$(CC) -c $(CFLAGS) \
+		-D CRC32_FUNCTION=crc32ethernet \
+		-D CRC32_CONSTANTS_HEADER=\"crc32_ethernet_constants.h\" \
+		vec_crc32.c -o $@
+
+crc32k_constants.h: crc32_constants
+	$(EMULATOR) ./crc32_constants -a -r -x 0x741B8CD7 > $@
+
+crc32k_wrapper.o: crc32_wrapper.c crc32k_constants.h
+crc32k.o: crc32.S crc32k_constants.h
+
+crc32k_wrapper.o crc32k.o:
+	$(CC) -c $(CFLAGS) \
+		-D CRC32_FUNCTION=crc32k \
+		-D CRC32_FUNCTION_ASM=crc32k_asm \
+		-D CRC32_CONSTANTS_HEADER=\"crc32k_constants.h\" \
+		$< -o $@
+
+crc32_two_implementations: crc32k_wrapper.o crc32k.o vec_crc32_ethernet.o
+
+# implementation has boundaries on datasizes 16 and 256, 32768 so ensure coverage and correctness
+test: crc32_test
+	set -e ; \
+	for len in `seq 0 257`  32767 32768 32769 65535 65536 65537 ; do \
+		./crc32_test  $${RANDOM} $$len $${RANDOM} ; \
+	done ; \
+
 clean:
-	rm -f crc32_constants.h *.o $(PROGS) $(PROGS_ALTIVEC)
+	rm -f crc32_constants.h crc32k_constants.h crc32_ethernet_constants.h *.o $(PROGS) $(PROGS_ALTIVEC)
+
+.PHONY: clean test all

README.md

@@ -34,18 +34,53 @@ to mitigate any I/O induced variability.
 Quick start
 -----------
 
+There's two different versions of crc32. They are, basically, the same 
+algorithm. The only difference is that one is implemented in pure assembly 
+(crc32.S) and the other in C using gcc (power8) vector intrinsics and 
+builtins (vec_crc32.c) to make the compiler generate the asm instructions 
+instead.  
+
 - Modify CRC and OPTIONS in the Makefile. There are examples for the two most
   common crc32s.
 
 - Type make to create the constants (crc32_constants.h)
 
+**If you will use the pure asm version**
+
 - Import the code into your application (crc32.S crc32_wrapper.c
-  crc32_constants.h ppc-opcode.h) and call the CRC:
+  crc32_constants.h ppc-opcode.h)
+
+**If you will use the C version**
+
+- Import the code into your application (vec_crc32.c crc32_constants.h)
+
+- Call the CRC:
+
 
 ```
 unsigned int crc32_vpmsum(unsigned int crc, unsigned char *p, unsigned long len);
 ```
 
+Advanced Usage
+--------------
+
+Occasionally you may have a number of CRC32 polynomial implementations.
+
+To do this you'll need to compile the C or assembler implementation with a
+different constants header file and change the function names to avoid linker
+conflicts.
+
+To facilitate this optional defines can be introduced:
+
+- CRC32_CONSTANTS_HEADER to be set to the *quoted* header filename.
+
+- CRC32_FUNCTION to be set to the crc32 function name (instead of crc32_vpmsum)
+
+- CRC32_FUNCTION_ASM (asm version only) to be set to the assember function name used
+by crc32_wrapper.c (defaults to __crc32_vpmsum).
+
+An example of this is with crc32_two_implementations as found in the Makefile.
+
 CRC background
 --------------
 
@@ -233,3 +268,7 @@ Acknowledgements
 
 Thanks to Michael Gschwind, Jeff Derby, Lorena Pesantez and Stewart Smith
 for their ideas and assistance.
+
+Thanks Rogerio Alves for writing the C implementation.
+
+Thanks Daniel Black for cleanup and testing.

crc32.S

@@ -46,8 +46,11 @@
 	/* byte reverse permute constant */
 	.octa 0x0F0E0D0C0B0A09080706050403020100
 
-#define __ASSEMBLY__
+#ifdef CRC32_CONSTANTS_HEADER
+#include CRC32_CONSTANTS_HEADER
+#else
 #include "crc32_constants.h"
+#endif
 
 	.text
 
@@ -81,8 +84,12 @@
 #define VPERM(A, B, C, D)
 #endif
 
+#ifndef CRC32_FUNCTION_ASM
+#define CRC32_FUNCTION_ASM __crc32_vpmsum
+#endif
+
 /* unsigned int __crc32_vpmsum(unsigned int crc, void *p, unsigned long len) */
-FUNC_START(__crc32_vpmsum)
+FUNC_START(CRC32_FUNCTION_ASM)
 	std	r31,-8(r1)
 	std	r30,-16(r1)
 	std	r29,-24(r1)
@@ -769,4 +776,4 @@ FUNC_START(__crc32_vpmsum)
 	mr	r3,r10
 	b	.Lout
 
-FUNC_END(__crc32_vpmsum)
+FUNC_END(CRC32_FUNCTION_ASM)

crc32_bench.c

@@ -1,6 +1,7 @@
 #include <malloc.h>
 #include <unistd.h>
 #include <stdlib.h>
+#include <stdio.h>
 
 unsigned int crc32_vpmsum(unsigned int crc, unsigned char *p, unsigned long len);
 
@@ -12,7 +13,8 @@ int main(int argc, char *argv[])
 	unsigned int crc = 0;
 
 	if (argc != 3) {
-		fprintf(stderr, "Usage: crc32_bench length iterations\n");
+		fprintf(stderr, "Usage: %s length iterations\n", argv[0]);
+		fprintf(stderr, "Performs crc32 checksum an [iterations] number of times on a buffer filled with junk data of [length] bytes\n");
 		exit(1);
 	}