rpms/zlib
7
0
Fork 0
Code Issues Pull Requests Releases Activity
078916e735
zlib/zlib-1.2.11-s390x-vectorize-crc32.patch
Patrik Novotný 7878600f3a Add IBM CRC32 optimalization patch for s390x 
Resolves: #1959423
2021-07-13 10:18:28 +02:00

456 lines
15 KiB
Diff
Raw Blame History

From 367e79caf76bda5fdb974420b72c6ddabdcd664e Mon Sep 17 00:00:00 2001
From: Ilya Leoshkevich <iii@linux.ibm.com>
Date: Thu, 19 Mar 2020 11:52:03 +0100
Subject: [PATCH] s390x: vectorize crc32
Use vector extensions when compiling for s390x and binutils knows
about them. At runtime, check whether kernel supports vector
extensions (it has to be not just the CPU, but also the kernel) and
choose between the regular and the vectorized implementations.
---
Makefile.in | 8 ++
configure | 16 +++
contrib/s390/crc32le-vx.S | 273 ++++++++++++++++++++++++++++++++++++++
crc32.c | 66 ++++++++-
4 files changed, 361 insertions(+), 2 deletions(-)
create mode 100644 contrib/s390/crc32le-vx.S
diff --git a/Makefile.in b/Makefile.in
index 6070dcc..23e8694 100644
--- a/Makefile.in
+++ b/Makefile.in
@@ -179,6 +179,9 @@ crc32_power8.o: $(SRCDIR)contrib/power8-crc/vec_crc32.c
crc32.o: $(SRCDIR)crc32.c
$(CC) $(CFLAGS) $(ZINC) -c -o $@ $(SRCDIR)crc32.c
+crc32le-vx.o: $(SRCDIR)contrib/s390/crc32le-vx.S
+ $(CC) $(CFLAGS) -march=z13 $(ZINC) -c -o $@ $(SRCDIR)contrib/s390/crc32le-vx.S
+
deflate.o: $(SRCDIR)deflate.c
$(CC) $(CFLAGS) $(ZINC) -c -o $@ $(SRCDIR)deflate.c
@@ -234,6 +237,11 @@ crc32.lo: $(SRCDIR)crc32.c
$(CC) $(SFLAGS) $(ZINC) -DPIC -c -o objs/crc32.o $(SRCDIR)crc32.c
-@mv objs/crc32.o $@
+crc32le-vx.lo: $(SRCDIR)contrib/s390/crc32le-vx.S
+ -@mkdir objs 2>/dev/null || test -d objs
+ $(CC) $(SFLAGS) -march=z13 $(ZINC) -DPIC -c -o objs/crc32le-vx.o $(SRCDIR)contrib/s390/crc32le-vx.S
+ -@mv objs/crc32le-vx.o $@
+
deflate.lo: $(SRCDIR)deflate.c
-@mkdir objs 2>/dev/null || test -d objs
$(CC) $(SFLAGS) $(ZINC) -DPIC -c -o objs/deflate.o $(SRCDIR)deflate.c
diff --git a/configure b/configure
index 70ed86b..e658039 100755
--- a/configure
+++ b/configure
@@ -923,6 +923,22 @@ EOF
fi
fi
+# check if we are compiling for s390 and binutils support vector extensions
+cat > $test.c <<EOF
+#ifndef __s390__
+#error
+#endif
+EOF
+if try $CC -c $CFLAGS -march=z13 $test.c; then
+ CFLAGS="$CFLAGS -DHAVE_S390X_VX"
+ SFLAGS="$SFLAGS -DHAVE_S390X_VX"
+ OBJC="$OBJC crc32le-vx.o"
+ PIC_OBJC="$PIC_OBJC crc32le-vx.lo"
+ echo "Checking for s390 vector extensions... Yes." | tee -a configure.log
+else
+ echo "Checking for s390 vector extensions... No." | tee -a configure.log
+fi
+
# show the results in the log
echo >> configure.log
echo ALL = $ALL >> configure.log
diff --git a/contrib/s390/crc32le-vx.S b/contrib/s390/crc32le-vx.S
new file mode 100644
index 0000000..029cfff
--- /dev/null
+++ b/contrib/s390/crc32le-vx.S
@@ -0,0 +1,273 @@
+/*
+ * Hardware-accelerated CRC-32 variants for Linux on z Systems
+ *
+ * Use the z/Architecture Vector Extension Facility to accelerate the
+ * computing of bitreflected CRC-32 checksums.
+ *
+ * This CRC-32 implementation algorithm is bitreflected and processes
+ * the least-significant bit first (Little-Endian).
+ *
+ * This code has been originally written by Hendrik Brueckner
+ * <brueckner@linux.vnet.ibm.com> and included in the Linux kernel:
+ *
+ * https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/s390/crypto/crc32le-vx.S?h=v5.5
+ *
+ * Hendrik Brueckner has allowed reusing it under zlib license.
+ *
+ * The following adjustments were made:
+ *
+ * - Reformatted in order to match the zlib code style.
+ * - Changed the vector register numbers in order to avoid clobbering the call-saved %v8-%v16.
+ * - Fixed clang compatibility.
+ * - Added 31-bit compatibility.
+ */
+
+#ifndef __clang__
+.machinemode zarch
+#endif
+
+#define PART1 %v16
+#define PART2 %v17
+#define PART3 %v18
+#define PART4 %v19
+#define SHIFTS %v20
+
+/* Vector register range containing CRC-32 constants */
+#define CONST_PERM_LE2BE %v21
+#define CONST_R2R1 %v22
+#define CONST_R4R3 %v23
+#define CONST_R5 %v24
+#define CONST_RU_POLY %v25
+#define CONST_CRC_POLY %v26
+
+#if defined(__s390x__)
+#define AGHI aghi
+#define CGHI cghi
+#else
+#define AGHI ahi
+#define CGHI chi
+#endif
+
+.data
+.align 8
+
+/*
+ * The CRC-32 constant block contains reduction constants to fold and
+ * process particular chunks of the input data stream in parallel.
+ *
+ * For the CRC-32 variants, the constants are precomputed according to
+ * these definitions:
+ *
+ * R1 = [(x4*128+32 mod P'(x) << 32)]' << 1
+ * R2 = [(x4*128-32 mod P'(x) << 32)]' << 1
+ * R3 = [(x128+32 mod P'(x) << 32)]' << 1
+ * R4 = [(x128-32 mod P'(x) << 32)]' << 1
+ * R5 = [(x64 mod P'(x) << 32)]' << 1
+ * R6 = [(x32 mod P'(x) << 32)]' << 1
+ *
+ * The bitreflected Barret reduction constant, u', is defined as
+ * the bit reversal of floor(x**64 / P(x)).
+ *
+ * where P(x) is the polynomial in the normal domain and the P'(x) is the
+ * polynomial in the reversed (bitreflected) domain.
+ *
+ * CRC-32 (IEEE 802.3 Ethernet, ...) polynomials:
+ *
+ * P(x) = 0x04C11DB7
+ * P'(x) = 0xEDB88320
+ */
+
+.Lconstants_CRC_32_LE:
+ .octa 0x0F0E0D0C0B0A09080706050403020100 # BE->LE mask
+ .quad 0x1c6e41596, 0x154442bd4 # R2, R1
+ .quad 0x0ccaa009e, 0x1751997d0 # R4, R3
+ .octa 0x163cd6124 # R5
+ .octa 0x1F7011641 # u'
+ .octa 0x1DB710641 # P'(x) << 1
+
+.text
+
+/*
+ * The CRC-32 functions use these calling conventions:
+ *
+ * Parameters:
+ *
+ * %r2: Initial CRC value, typically ~0; and final CRC (return) value.
+ * %r3: Input buffer pointer, performance might be improved if the
+ * buffer is on a doubleword boundary.
+ * %r4: Length of the buffer, must be 64 bytes or greater.
+ *
+ * Register usage:
+ *
+ * %r5: CRC-32 constant pool base pointer.
+ * V0: Initial CRC value and intermediate constants and results.
+ * V1..V4: Data for CRC computation.
+ * V16..V19: Next data chunks that are fetched from the input buffer.
+ * V20: Constant for BE->LE conversion and shift operations
+ *
+ * V21..V26: CRC-32 constants.
+ */
+
+ .globl crc32_le_vgfm_16
+ .align 4, 0x07
+crc32_le_vgfm_16:
+ /* Load CRC-32 constants */
+ larl %r5,.Lconstants_CRC_32_LE
+ VLM CONST_PERM_LE2BE,CONST_CRC_POLY,0(%r5)
+
+ /*
+ * Load the initial CRC value.
+ *
+ * The CRC value is loaded into the rightmost word of the
+ * vector register and is later XORed with the LSB portion
+ * of the loaded input data.
+ */
+ VZERO %v0 /* Clear V0 */
+ VLVGF %v0,%r2,3 /* Load CRC into rightmost word */
+
+ /* Load a 64-byte data chunk and XOR with CRC */
+ VLM %v1,%v4,0(%r3) /* 64-bytes into V1..V4 */
+ VPERM %v1,%v1,%v1,CONST_PERM_LE2BE
+ VPERM %v2,%v2,%v2,CONST_PERM_LE2BE
+ VPERM %v3,%v3,%v3,CONST_PERM_LE2BE
+ VPERM %v4,%v4,%v4,CONST_PERM_LE2BE
+
+ VX %v1,%v0,%v1 /* V1 ^= CRC */
+ AGHI %r3,64 /* BUF = BUF + 64 */
+ AGHI %r4,-64 /* LEN = LEN - 64 */
+
+ CGHI %r4,64
+ jl .Lless_than_64bytes
+
+.Lfold_64bytes_loop:
+ /* Load the next 64-byte data chunk into PART1 to PART4 */
+ VLM PART1,PART4,0(%r3)
+ VPERM PART1,PART1,PART1,CONST_PERM_LE2BE
+ VPERM PART2,PART2,PART2,CONST_PERM_LE2BE
+ VPERM PART3,PART3,PART3,CONST_PERM_LE2BE
+ VPERM PART4,PART4,PART4,CONST_PERM_LE2BE
+
+ /*
+ * Perform a GF(2) multiplication of the doublewords in V1 with
+ * the R1 and R2 reduction constants in V0. The intermediate result
+ * is then folded (accumulated) with the next data chunk in PART1 and
+ * stored in V1. Repeat this step for the register contents
+ * in V2, V3, and V4 respectively.
+ */
+ VGFMAG %v1,CONST_R2R1,%v1,PART1
+ VGFMAG %v2,CONST_R2R1,%v2,PART2
+ VGFMAG %v3,CONST_R2R1,%v3,PART3
+ VGFMAG %v4,CONST_R2R1,%v4,PART4
+
+ AGHI %r3,64 /* BUF = BUF + 64 */
+ AGHI %r4,-64 /* LEN = LEN - 64 */
+
+ CGHI %r4,64
+ jnl .Lfold_64bytes_loop
+
+.Lless_than_64bytes:
+ /*
+ * Fold V1 to V4 into a single 128-bit value in V1. Multiply V1 with R3
+ * and R4 and accumulating the next 128-bit chunk until a single 128-bit
+ * value remains.
+ */
+ VGFMAG %v1,CONST_R4R3,%v1,%v2
+ VGFMAG %v1,CONST_R4R3,%v1,%v3
+ VGFMAG %v1,CONST_R4R3,%v1,%v4
+
+ CGHI %r4,16
+ jl .Lfinal_fold
+
+.Lfold_16bytes_loop:
+
+ VL %v2,0(%r3) /* Load next data chunk */
+ VPERM %v2,%v2,%v2,CONST_PERM_LE2BE
+ VGFMAG %v1,CONST_R4R3,%v1,%v2 /* Fold next data chunk */
+
+ AGHI %r3,16
+ AGHI %r4,-16
+
+ CGHI %r4,16
+ jnl .Lfold_16bytes_loop
+
+.Lfinal_fold:
+ /*
+ * Set up a vector register for byte shifts. The shift value must
+ * be loaded in bits 1-4 in byte element 7 of a vector register.
+ * Shift by 8 bytes: 0x40
+ * Shift by 4 bytes: 0x20
+ */
+ VLEIB SHIFTS,0x40,7
+
+ /*
+ * Prepare V0 for the next GF(2) multiplication: shift V0 by 8 bytes
+ * to move R4 into the rightmost doubleword and set the leftmost
+ * doubleword to 0x1.
+ */
+ VSRLB %v0,CONST_R4R3,SHIFTS
+ VLEIG %v0,1,0
+
+ /*
+ * Compute GF(2) product of V1 and V0. The rightmost doubleword
+ * of V1 is multiplied with R4. The leftmost doubleword of V1 is
+ * multiplied by 0x1 and is then XORed with rightmost product.
+ * Implicitly, the intermediate leftmost product becomes padded
+ */
+ VGFMG %v1,%v0,%v1
+
+ /*
+ * Now do the final 32-bit fold by multiplying the rightmost word
+ * in V1 with R5 and XOR the result with the remaining bits in V1.
+ *
+ * To achieve this by a single VGFMAG, right shift V1 by a word
+ * and store the result in V2 which is then accumulated. Use the
+ * vector unpack instruction to load the rightmost half of the
+ * doubleword into the rightmost doubleword element of V1; the other
+ * half is loaded in the leftmost doubleword.
+ * The vector register with CONST_R5 contains the R5 constant in the
+ * rightmost doubleword and the leftmost doubleword is zero to ignore
+ * the leftmost product of V1.
+ */
+ VLEIB SHIFTS,0x20,7 /* Shift by words */
+ VSRLB %v2,%v1,SHIFTS /* Store remaining bits in V2 */
+ VUPLLF %v1,%v1 /* Split rightmost doubleword */
+ VGFMAG %v1,CONST_R5,%v1,%v2 /* V1 = (V1 * R5) XOR V2 */
+
+ /*
+ * Apply a Barret reduction to compute the final 32-bit CRC value.
+ *
+ * The input values to the Barret reduction are the degree-63 polynomial
+ * in V1 (R(x)), degree-32 generator polynomial, and the reduction
+ * constant u. The Barret reduction result is the CRC value of R(x) mod
+ * P(x).
+ *
+ * The Barret reduction algorithm is defined as:
+ *
+ * 1. T1(x) = floor( R(x) / x^32 ) GF2MUL u
+ * 2. T2(x) = floor( T1(x) / x^32 ) GF2MUL P(x)
+ * 3. C(x) = R(x) XOR T2(x) mod x^32
+ *
+ * Note: The leftmost doubleword of vector register containing
+ * CONST_RU_POLY is zero and, thus, the intermediate GF(2) product
+ * is zero and does not contribute to the final result.
+ */
+
+ /* T1(x) = floor( R(x) / x^32 ) GF2MUL u */
+ VUPLLF %v2,%v1
+ VGFMG %v2,CONST_RU_POLY,%v2
+
+ /*
+ * Compute the GF(2) product of the CRC polynomial with T1(x) in
+ * V2 and XOR the intermediate result, T2(x), with the value in V1.
+ * The final result is stored in word element 2 of V2.
+ */
+ VUPLLF %v2,%v2
+ VGFMAG %v2,CONST_CRC_POLY,%v2,%v1
+
+.Ldone:
+ VLGVF %r2,%v2,2
+ BR %r14
+ .type crc32_le_vgfm_16, @function
+ .size crc32_le_vgfm_16, .-crc32_le_vgfm_16
+
+.previous
diff --git a/crc32.c b/crc32.c
index 34132ea..af5d3cd 100644
--- a/crc32.c
+++ b/crc32.c
@@ -252,12 +252,26 @@ unsigned long crc32_vpmsum(unsigned long, const unsigned char FAR *, z_size_t);
#endif
#endif
+#ifdef HAVE_S390X_VX
+#include <sys/auxv.h>
+
+local unsigned long crc32_s390_vx(unsigned long crc,
+ const unsigned char FAR *buf,
+ z_size_t len);
+#endif
+
/* due to a quirk of gnu_indirect_function - "local" (aka static) is applied to
* crc32_z which is not desired. crc32_z_ifunc is implictly "local" */
#ifndef Z_IFUNC_ASM
local
#endif
-unsigned long (*(crc32_z_ifunc(void)))(unsigned long, const unsigned char FAR *, z_size_t)
+unsigned long (*(crc32_z_ifunc(
+#ifdef __s390__
+unsigned long hwcap
+#else
+void
+#endif
+)))(unsigned long, const unsigned char FAR *, z_size_t)
{
#if _ARCH_PWR8==1
#if defined(__BUILTIN_CPU_SUPPORTS__)
@@ -269,6 +283,11 @@ unsigned long (*(crc32_z_ifunc(void)))(unsigned long, const unsigned char FAR *,
#endif
#endif /* _ARCH_PWR8 */
+#ifdef HAVE_S390X_VX
+ if (hwcap & HWCAP_S390_VX)
+ return crc32_s390_vx;
+#endif
+
/* return a function pointer for optimized arches here */
#ifdef DYNAMIC_CRC_TABLE
@@ -301,7 +320,11 @@ unsigned long ZEXPORT crc32_z(crc, buf, len)
static unsigned long ZEXPORT (*crc32_func)(unsigned long, const unsigned char FAR *, z_size_t) = NULL;
if (!crc32_func)
- crc32_func = crc32_z_ifunc();
+ crc32_func = crc32_z_ifunc(
+#ifdef __s390__
+ getauxval(AT_HWCAP)
+#endif
+ );
return (*crc32_func)(crc, buf, len);
}
@@ -500,6 +523,45 @@ local uLong crc32_combine_(crc1, crc2, len2)
return crc1;
}
+#ifdef HAVE_S390X_VX
+#define VX_MIN_LEN 64
+#define VX_ALIGNMENT 16L
+#define VX_ALIGN_MASK (VX_ALIGNMENT - 1)
+
+unsigned int crc32_le_vgfm_16(unsigned int crc,
+ unsigned char const *buf,
+ size_t size);
+
+local unsigned long crc32_s390_vx(crc, buf, len)
+ unsigned long crc;
+ const unsigned char FAR *buf;
+ z_size_t len;
+{
+ unsigned long prealign, aligned, remaining;
+
+ if (buf == Z_NULL) return 0UL;
+
+ if (len < VX_MIN_LEN + VX_ALIGN_MASK)
+ return crc32_big(crc, buf, len);
+
+ if ((unsigned long)buf & VX_ALIGN_MASK) {
+ prealign = VX_ALIGNMENT - ((unsigned long)buf & VX_ALIGN_MASK);
+ len -= prealign;
+ crc = crc32_big(crc, buf, prealign);
+ buf = (void *)((unsigned long)buf + prealign);
+ }
+ aligned = len & ~VX_ALIGN_MASK;
+ remaining = len & VX_ALIGN_MASK;
+
+ crc = crc32_le_vgfm_16(crc ^ 0xffffffff, buf, aligned) ^ 0xffffffff;
+
+ if (remaining)
+ crc = crc32_big(crc, buf + aligned, remaining);
+
+ return crc;
+}
+#endif
+
/* ========================================================================= */
uLong ZEXPORT crc32_combine(crc1, crc2, len2)
uLong crc1;
--
2.25.1
Reference in New Issue View Git Blame Copy Permalink