diff --git a/contrib/libdivide/libdivide.h b/contrib/libdivide/libdivide.h index eaeaec7db6b..a153e7f9c5e 100644 --- a/contrib/libdivide/libdivide.h +++ b/contrib/libdivide/libdivide.h @@ -1,117 +1,106 @@ -/* libdivide.h - Copyright 2010 ridiculous_fish -*/ -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wold-style-cast" +// libdivide.h - Optimized integer division +// https://libdivide.com +// +// Copyright (C) 2010 - 2019 ridiculous_fish, +// Copyright (C) 2016 - 2019 Kim Walisch, +// +// libdivide is dual-licensed under the Boost or zlib licenses. +// You may use libdivide under the terms of either of these. +// See LICENSE.txt for more details. -#if defined(_WIN32) || defined(WIN32) -#define LIBDIVIDE_WINDOWS 1 -#endif +#ifndef LIBDIVIDE_H +#define LIBDIVIDE_H -#if defined(_MSC_VER) -#define LIBDIVIDE_VC 1 -#endif +#define LIBDIVIDE_VERSION "3.0" +#define LIBDIVIDE_VERSION_MAJOR 3 +#define LIBDIVIDE_VERSION_MINOR 0 -#ifdef __cplusplus -#include -#include -#include -#else -#include -#include -#include -#endif - -#if ! LIBDIVIDE_HAS_STDINT_TYPES && (! LIBDIVIDE_VC || _MSC_VER >= 1600) -/* Only Visual C++ 2010 and later include stdint.h */ #include -#define LIBDIVIDE_HAS_STDINT_TYPES 1 + +#if defined(__cplusplus) + #include + #include + #include +#else + #include + #include #endif -#if ! LIBDIVIDE_HAS_STDINT_TYPES -typedef __int32 int32_t; -typedef unsigned __int32 uint32_t; -typedef __int64 int64_t; -typedef unsigned __int64 uint64_t; -typedef __int8 int8_t; -typedef unsigned __int8 uint8_t; -#endif - -#if LIBDIVIDE_USE_SSE2 +#if defined(LIBDIVIDE_AVX512) + #include +#elif defined(LIBDIVIDE_AVX2) + #include +#elif defined(LIBDIVIDE_SSE2) #include #endif -#if LIBDIVIDE_VC +#if defined(_MSC_VER) #include + // disable warning C4146: unary minus operator applied + // to unsigned type, result still unsigned + #pragma warning(disable: 4146) + #define LIBDIVIDE_VC #endif -#ifndef __has_builtin -#define __has_builtin(x) 0 // Compatibility with non-clang compilers. +#if !defined(__has_builtin) + #define __has_builtin(x) 0 #endif -#ifdef __ICC -#define HAS_INT128_T 0 -#else -#define HAS_INT128_T __LP64__ +#if defined(__SIZEOF_INT128__) + #define HAS_INT128_T + // clang-cl on Windows does not yet support 128-bit division + #if !(defined(__clang__) && defined(LIBDIVIDE_VC)) + #define HAS_INT128_DIV + #endif #endif -#if defined(__x86_64__) || defined(_WIN64) || defined(_M_64) -#define LIBDIVIDE_IS_X86_64 1 +#if defined(__x86_64__) || defined(_M_X64) + #define LIBDIVIDE_X86_64 #endif #if defined(__i386__) -#define LIBDIVIDE_IS_i386 1 + #define LIBDIVIDE_i386 #endif -#if __GNUC__ || __clang__ -#define LIBDIVIDE_GCC_STYLE_ASM 1 +#if defined(__GNUC__) || defined(__clang__) + #define LIBDIVIDE_GCC_STYLE_ASM #endif +#if defined(__cplusplus) || defined(LIBDIVIDE_VC) + #define LIBDIVIDE_FUNCTION __FUNCTION__ +#else + #define LIBDIVIDE_FUNCTION __func__ +#endif -/* libdivide may use the pmuldq (vector signed 32x32->64 mult instruction) which is in SSE 4.1. However, signed multiplication can be emulated efficiently with unsigned multiplication, and SSE 4.1 is currently rare, so it is OK to not turn this on */ -#ifdef LIBDIVIDE_USE_SSE4_1 -#include +#define LIBDIVIDE_ERROR(msg) \ + do { \ + fprintf(stderr, "libdivide.h:%d: %s(): Error: %s\n", \ + __LINE__, LIBDIVIDE_FUNCTION, msg); \ + exit(-1); \ + } while (0) + +#if defined(LIBDIVIDE_ASSERTIONS_ON) + #define LIBDIVIDE_ASSERT(x) \ + do { \ + if (!(x)) { \ + fprintf(stderr, "libdivide.h:%d: %s(): Assertion failed: %s\n", \ + __LINE__, LIBDIVIDE_FUNCTION, #x); \ + exit(-1); \ + } \ + } while (0) +#else + #define LIBDIVIDE_ASSERT(x) #endif #ifdef __cplusplus -/* We place libdivide within the libdivide namespace, and that goes in an anonymous namespace so that the functions are only visible to files that #include this header and don't get external linkage. At least that's the theory. */ -namespace { namespace libdivide { #endif -/* Explanation of "more" field: bit 6 is whether to use shift path. If we are using the shift path, bit 7 is whether the divisor is negative in the signed case; in the unsigned case it is 0. Bits 0-4 is shift value (for shift path or mult path). In 32 bit case, bit 5 is always 0. We use bit 7 as the "negative divisor indicator" so that we can use sign extension to efficiently go to a full-width -1. - - -u32: [0-4] shift value - [5] ignored - [6] add indicator - [7] shift path - -s32: [0-4] shift value - [5] shift path - [6] add indicator - [7] indicates negative divisor - -u64: [0-5] shift value - [6] add indicator - [7] shift path - -s64: [0-5] shift value - [6] add indicator - [7] indicates negative divisor - magic number of 0 indicates shift path (we ran out of bits!) -*/ - -enum { - LIBDIVIDE_32_SHIFT_MASK = 0x1F, - LIBDIVIDE_64_SHIFT_MASK = 0x3F, - LIBDIVIDE_ADD_MARKER = 0x40, - LIBDIVIDE_U32_SHIFT_PATH = 0x80, - LIBDIVIDE_U64_SHIFT_PATH = 0x80, - LIBDIVIDE_S32_SHIFT_PATH = 0x20, - LIBDIVIDE_NEGATIVE_DIVISOR = 0x80 -}; - +// pack divider structs to prevent compilers from padding. +// This reduces memory usage by up to 43% when using a large +// array of libdivide dividers and improves performance +// by up to 10% because of reduced memory bandwidth. +#pragma pack(push, 1) struct libdivide_u32_t { uint32_t magic; @@ -133,497 +122,446 @@ struct libdivide_s64_t { uint8_t more; }; +struct libdivide_u32_branchfree_t { + uint32_t magic; + uint8_t more; +}; +struct libdivide_s32_branchfree_t { + int32_t magic; + uint8_t more; +}; -#ifndef LIBDIVIDE_API - #ifdef __cplusplus - /* In C++, we don't want our public functions to be static, because they are arguments to templates and static functions can't do that. They get internal linkage through virtue of the anonymous namespace. In C, they should be static. */ - #define LIBDIVIDE_API - #else - #define LIBDIVIDE_API static - #endif -#endif +struct libdivide_u64_branchfree_t { + uint64_t magic; + uint8_t more; +}; -#ifdef __APPLE__ -typedef signed long Int64; -typedef unsigned long UInt64; -#endif +struct libdivide_s64_branchfree_t { + int64_t magic; + uint8_t more; +}; -LIBDIVIDE_API struct libdivide_s32_t libdivide_s32_gen(int32_t y); -LIBDIVIDE_API struct libdivide_u32_t libdivide_u32_gen(uint32_t y); -LIBDIVIDE_API struct libdivide_s64_t libdivide_s64_gen(int64_t y); -LIBDIVIDE_API struct libdivide_u64_t libdivide_u64_gen(uint64_t y); -#if defined(__APPLE__) && defined(__cplusplus) -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-function" -LIBDIVIDE_API struct libdivide_s64_t libdivide_s64_gen(Int64 y) { return libdivide_s64_gen(int64_t(y)); }; -LIBDIVIDE_API struct libdivide_u64_t libdivide_u64_gen(UInt64 y) { return libdivide_u64_gen(uint64_t(y)); }; -#pragma GCC diagnostic pop -#endif +#pragma pack(pop) -LIBDIVIDE_API int32_t libdivide_s32_do(int32_t numer, const struct libdivide_s32_t *denom); -LIBDIVIDE_API uint32_t libdivide_u32_do(uint32_t numer, const struct libdivide_u32_t *denom); -LIBDIVIDE_API int64_t libdivide_s64_do(int64_t numer, const struct libdivide_s64_t *denom); -LIBDIVIDE_API uint64_t libdivide_u64_do(uint64_t y, const struct libdivide_u64_t *denom); -#if defined(__APPLE__) && defined(__cplusplus) -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-function" -LIBDIVIDE_API Int64 libdivide_s64_do(Int64 numer, const struct libdivide_s64_t *denom) { return Int64(libdivide_s64_do(int64_t(numer), denom)); }; -LIBDIVIDE_API UInt64 libdivide_u64_do(UInt64 y, const struct libdivide_u64_t *denom) { return UInt64(libdivide_u64_do(uint64_t(y), denom)); }; -#pragma GCC diagnostic pop -#endif +// Explanation of the "more" field: +// +// * Bits 0-5 is the shift value (for shift path or mult path). +// * Bit 6 is the add indicator for mult path. +// * Bit 7 is set if the divisor is negative. We use bit 7 as the negative +// divisor indicator so that we can efficiently use sign extension to +// create a bitmask with all bits set to 1 (if the divisor is negative) +// or 0 (if the divisor is positive). +// +// u32: [0-4] shift value +// [5] ignored +// [6] add indicator +// magic number of 0 indicates shift path +// +// s32: [0-4] shift value +// [5] ignored +// [6] add indicator +// [7] indicates negative divisor +// magic number of 0 indicates shift path +// +// u64: [0-5] shift value +// [6] add indicator +// magic number of 0 indicates shift path +// +// s64: [0-5] shift value +// [6] add indicator +// [7] indicates negative divisor +// magic number of 0 indicates shift path +// +// In s32 and s64 branchfree modes, the magic number is negated according to +// whether the divisor is negated. In branchfree strategy, it is not negated. -LIBDIVIDE_API int libdivide_u32_get_algorithm(const struct libdivide_u32_t *denom); -LIBDIVIDE_API uint32_t libdivide_u32_do_alg0(uint32_t numer, const struct libdivide_u32_t *denom); -LIBDIVIDE_API uint32_t libdivide_u32_do_alg1(uint32_t numer, const struct libdivide_u32_t *denom); -LIBDIVIDE_API uint32_t libdivide_u32_do_alg2(uint32_t numer, const struct libdivide_u32_t *denom); +enum { + LIBDIVIDE_32_SHIFT_MASK = 0x1F, + LIBDIVIDE_64_SHIFT_MASK = 0x3F, + LIBDIVIDE_ADD_MARKER = 0x40, + LIBDIVIDE_NEGATIVE_DIVISOR = 0x80 +}; -LIBDIVIDE_API int libdivide_u64_get_algorithm(const struct libdivide_u64_t *denom); -LIBDIVIDE_API uint64_t libdivide_u64_do_alg0(uint64_t numer, const struct libdivide_u64_t *denom); -LIBDIVIDE_API uint64_t libdivide_u64_do_alg1(uint64_t numer, const struct libdivide_u64_t *denom); -LIBDIVIDE_API uint64_t libdivide_u64_do_alg2(uint64_t numer, const struct libdivide_u64_t *denom); -#if defined(__APPLE__) && defined(__cplusplus) -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-function" -LIBDIVIDE_API UInt64 libdivide_u64_do_alg0(UInt64 numer, const struct libdivide_u64_t *denom) { return UInt64(libdivide_u64_do_alg0(uint64_t(numer), denom)); } -LIBDIVIDE_API UInt64 libdivide_u64_do_alg1(UInt64 numer, const struct libdivide_u64_t *denom) { return UInt64(libdivide_u64_do_alg1(uint64_t(numer), denom)); } -LIBDIVIDE_API UInt64 libdivide_u64_do_alg2(UInt64 numer, const struct libdivide_u64_t *denom) { return UInt64(libdivide_u64_do_alg2(uint64_t(numer), denom)); } -#pragma GCC diagnostic pop -#endif +static inline struct libdivide_s32_t libdivide_s32_gen(int32_t d); +static inline struct libdivide_u32_t libdivide_u32_gen(uint32_t d); +static inline struct libdivide_s64_t libdivide_s64_gen(int64_t d); +static inline struct libdivide_u64_t libdivide_u64_gen(uint64_t d); -LIBDIVIDE_API int libdivide_s32_get_algorithm(const struct libdivide_s32_t *denom); -LIBDIVIDE_API int32_t libdivide_s32_do_alg0(int32_t numer, const struct libdivide_s32_t *denom); -LIBDIVIDE_API int32_t libdivide_s32_do_alg1(int32_t numer, const struct libdivide_s32_t *denom); -LIBDIVIDE_API int32_t libdivide_s32_do_alg2(int32_t numer, const struct libdivide_s32_t *denom); -LIBDIVIDE_API int32_t libdivide_s32_do_alg3(int32_t numer, const struct libdivide_s32_t *denom); -LIBDIVIDE_API int32_t libdivide_s32_do_alg4(int32_t numer, const struct libdivide_s32_t *denom); +static inline struct libdivide_s32_branchfree_t libdivide_s32_branchfree_gen(int32_t d); +static inline struct libdivide_u32_branchfree_t libdivide_u32_branchfree_gen(uint32_t d); +static inline struct libdivide_s64_branchfree_t libdivide_s64_branchfree_gen(int64_t d); +static inline struct libdivide_u64_branchfree_t libdivide_u64_branchfree_gen(uint64_t d); -LIBDIVIDE_API int libdivide_s64_get_algorithm(const struct libdivide_s64_t *denom); -LIBDIVIDE_API int64_t libdivide_s64_do_alg0(int64_t numer, const struct libdivide_s64_t *denom); -LIBDIVIDE_API int64_t libdivide_s64_do_alg1(int64_t numer, const struct libdivide_s64_t *denom); -LIBDIVIDE_API int64_t libdivide_s64_do_alg2(int64_t numer, const struct libdivide_s64_t *denom); -LIBDIVIDE_API int64_t libdivide_s64_do_alg3(int64_t numer, const struct libdivide_s64_t *denom); -LIBDIVIDE_API int64_t libdivide_s64_do_alg4(int64_t numer, const struct libdivide_s64_t *denom); -#if defined(__APPLE__) && defined(__cplusplus) -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-function" -LIBDIVIDE_API Int64 libdivide_s64_do_alg0(Int64 numer, const struct libdivide_s64_t *denom) { return Int64(libdivide_s64_do_alg0(int64_t(numer), denom)); } -LIBDIVIDE_API Int64 libdivide_s64_do_alg1(Int64 numer, const struct libdivide_s64_t *denom) { return Int64(libdivide_s64_do_alg1(int64_t(numer), denom)); } -LIBDIVIDE_API Int64 libdivide_s64_do_alg2(Int64 numer, const struct libdivide_s64_t *denom) { return Int64(libdivide_s64_do_alg2(int64_t(numer), denom)); } -LIBDIVIDE_API Int64 libdivide_s64_do_alg3(Int64 numer, const struct libdivide_s64_t *denom) { return Int64(libdivide_s64_do_alg3(int64_t(numer), denom)); } -LIBDIVIDE_API Int64 libdivide_s64_do_alg4(Int64 numer, const struct libdivide_s64_t *denom) { return Int64(libdivide_s64_do_alg4(int64_t(numer), denom)); } -#pragma GCC diagnostic pop -#endif +static inline int32_t libdivide_s32_do(int32_t numer, const struct libdivide_s32_t *denom); +static inline uint32_t libdivide_u32_do(uint32_t numer, const struct libdivide_u32_t *denom); +static inline int64_t libdivide_s64_do(int64_t numer, const struct libdivide_s64_t *denom); +static inline uint64_t libdivide_u64_do(uint64_t numer, const struct libdivide_u64_t *denom); +static inline int32_t libdivide_s32_branchfree_do(int32_t numer, const struct libdivide_s32_branchfree_t *denom); +static inline uint32_t libdivide_u32_branchfree_do(uint32_t numer, const struct libdivide_u32_branchfree_t *denom); +static inline int64_t libdivide_s64_branchfree_do(int64_t numer, const struct libdivide_s64_branchfree_t *denom); +static inline uint64_t libdivide_u64_branchfree_do(uint64_t numer, const struct libdivide_u64_branchfree_t *denom); -#if LIBDIVIDE_USE_SSE2 -LIBDIVIDE_API __m128i libdivide_u32_do_vector(__m128i numers, const struct libdivide_u32_t * denom); -LIBDIVIDE_API __m128i libdivide_s32_do_vector(__m128i numers, const struct libdivide_s32_t * denom); -LIBDIVIDE_API __m128i libdivide_u64_do_vector(__m128i numers, const struct libdivide_u64_t * denom); -LIBDIVIDE_API __m128i libdivide_s64_do_vector(__m128i numers, const struct libdivide_s64_t * denom); - -LIBDIVIDE_API __m128i libdivide_u32_do_vector_alg0(__m128i numers, const struct libdivide_u32_t * denom); -LIBDIVIDE_API __m128i libdivide_u32_do_vector_alg1(__m128i numers, const struct libdivide_u32_t * denom); -LIBDIVIDE_API __m128i libdivide_u32_do_vector_alg2(__m128i numers, const struct libdivide_u32_t * denom); - -LIBDIVIDE_API __m128i libdivide_s32_do_vector_alg0(__m128i numers, const struct libdivide_s32_t * denom); -LIBDIVIDE_API __m128i libdivide_s32_do_vector_alg1(__m128i numers, const struct libdivide_s32_t * denom); -LIBDIVIDE_API __m128i libdivide_s32_do_vector_alg2(__m128i numers, const struct libdivide_s32_t * denom); -LIBDIVIDE_API __m128i libdivide_s32_do_vector_alg3(__m128i numers, const struct libdivide_s32_t * denom); -LIBDIVIDE_API __m128i libdivide_s32_do_vector_alg4(__m128i numers, const struct libdivide_s32_t * denom); - -LIBDIVIDE_API __m128i libdivide_u64_do_vector_alg0(__m128i numers, const struct libdivide_u64_t * denom); -LIBDIVIDE_API __m128i libdivide_u64_do_vector_alg1(__m128i numers, const struct libdivide_u64_t * denom); -LIBDIVIDE_API __m128i libdivide_u64_do_vector_alg2(__m128i numers, const struct libdivide_u64_t * denom); - -LIBDIVIDE_API __m128i libdivide_s64_do_vector_alg0(__m128i numers, const struct libdivide_s64_t * denom); -LIBDIVIDE_API __m128i libdivide_s64_do_vector_alg1(__m128i numers, const struct libdivide_s64_t * denom); -LIBDIVIDE_API __m128i libdivide_s64_do_vector_alg2(__m128i numers, const struct libdivide_s64_t * denom); -LIBDIVIDE_API __m128i libdivide_s64_do_vector_alg3(__m128i numers, const struct libdivide_s64_t * denom); -LIBDIVIDE_API __m128i libdivide_s64_do_vector_alg4(__m128i numers, const struct libdivide_s64_t * denom); -#endif - +static inline int32_t libdivide_s32_recover(const struct libdivide_s32_t *denom); +static inline uint32_t libdivide_u32_recover(const struct libdivide_u32_t *denom); +static inline int64_t libdivide_s64_recover(const struct libdivide_s64_t *denom); +static inline uint64_t libdivide_u64_recover(const struct libdivide_u64_t *denom); +static inline int32_t libdivide_s32_branchfree_recover(const struct libdivide_s32_branchfree_t *denom); +static inline uint32_t libdivide_u32_branchfree_recover(const struct libdivide_u32_branchfree_t *denom); +static inline int64_t libdivide_s64_branchfree_recover(const struct libdivide_s64_branchfree_t *denom); +static inline uint64_t libdivide_u64_branchfree_recover(const struct libdivide_u64_branchfree_t *denom); //////// Internal Utility Functions -static inline uint32_t libdivide__mullhi_u32(uint32_t x, uint32_t y) { +static inline uint32_t libdivide_mullhi_u32(uint32_t x, uint32_t y) { uint64_t xl = x, yl = y; uint64_t rl = xl * yl; return (uint32_t)(rl >> 32); } -static uint64_t libdivide__mullhi_u64(uint64_t x, uint64_t y) { -#if HAS_INT128_T +static inline int32_t libdivide_mullhi_s32(int32_t x, int32_t y) { + int64_t xl = x, yl = y; + int64_t rl = xl * yl; + // needs to be arithmetic shift + return (int32_t)(rl >> 32); +} + +static inline uint64_t libdivide_mullhi_u64(uint64_t x, uint64_t y) { +#if defined(LIBDIVIDE_VC) && \ + defined(LIBDIVIDE_X86_64) + return __umulh(x, y); +#elif defined(HAS_INT128_T) __uint128_t xl = x, yl = y; __uint128_t rl = xl * yl; return (uint64_t)(rl >> 64); #else - //full 128 bits are x0 * y0 + (x0 * y1 << 32) + (x1 * y0 << 32) + (x1 * y1 << 64) - const uint32_t mask = 0xFFFFFFFF; - const uint32_t x0 = (uint32_t)(x & mask), x1 = (uint32_t)(x >> 32); - const uint32_t y0 = (uint32_t)(y & mask), y1 = (uint32_t)(y >> 32); - const uint32_t x0y0_hi = libdivide__mullhi_u32(x0, y0); - const uint64_t x0y1 = x0 * (uint64_t)y1; - const uint64_t x1y0 = x1 * (uint64_t)y0; - const uint64_t x1y1 = x1 * (uint64_t)y1; - + // full 128 bits are x0 * y0 + (x0 * y1 << 32) + (x1 * y0 << 32) + (x1 * y1 << 64) + uint32_t mask = 0xFFFFFFFF; + uint32_t x0 = (uint32_t)(x & mask); + uint32_t x1 = (uint32_t)(x >> 32); + uint32_t y0 = (uint32_t)(y & mask); + uint32_t y1 = (uint32_t)(y >> 32); + uint32_t x0y0_hi = libdivide_mullhi_u32(x0, y0); + uint64_t x0y1 = x0 * (uint64_t)y1; + uint64_t x1y0 = x1 * (uint64_t)y0; + uint64_t x1y1 = x1 * (uint64_t)y1; uint64_t temp = x1y0 + x0y0_hi; - uint64_t temp_lo = temp & mask, temp_hi = temp >> 32; + uint64_t temp_lo = temp & mask; + uint64_t temp_hi = temp >> 32; + return x1y1 + temp_hi + ((temp_lo + x0y1) >> 32); #endif } -static inline int64_t libdivide__mullhi_s64(int64_t x, int64_t y) { -#if HAS_INT128_T +static inline int64_t libdivide_mullhi_s64(int64_t x, int64_t y) { +#if defined(LIBDIVIDE_VC) && \ + defined(LIBDIVIDE_X86_64) + return __mulh(x, y); +#elif defined(HAS_INT128_T) __int128_t xl = x, yl = y; __int128_t rl = xl * yl; return (int64_t)(rl >> 64); #else - //full 128 bits are x0 * y0 + (x0 * y1 << 32) + (x1 * y0 << 32) + (x1 * y1 << 64) - const uint32_t mask = 0xFFFFFFFF; - const uint32_t x0 = (uint32_t)(x & mask), y0 = (uint32_t)(y & mask); - const int32_t x1 = (int32_t)(x >> 32), y1 = (int32_t)(y >> 32); - const uint32_t x0y0_hi = libdivide__mullhi_u32(x0, y0); - const int64_t t = x1*(int64_t)y0 + x0y0_hi; - const int64_t w1 = x0*(int64_t)y1 + (t & mask); - return x1*(int64_t)y1 + (t >> 32) + (w1 >> 32); + // full 128 bits are x0 * y0 + (x0 * y1 << 32) + (x1 * y0 << 32) + (x1 * y1 << 64) + uint32_t mask = 0xFFFFFFFF; + uint32_t x0 = (uint32_t)(x & mask); + uint32_t y0 = (uint32_t)(y & mask); + int32_t x1 = (int32_t)(x >> 32); + int32_t y1 = (int32_t)(y >> 32); + uint32_t x0y0_hi = libdivide_mullhi_u32(x0, y0); + int64_t t = x1 * (int64_t)y0 + x0y0_hi; + int64_t w1 = x0 * (int64_t)y1 + (t & mask); + + return x1 * (int64_t)y1 + (t >> 32) + (w1 >> 32); #endif } -#if LIBDIVIDE_USE_SSE2 - -static inline __m128i libdivide__u64_to_m128(uint64_t x) { -#if LIBDIVIDE_VC && ! _WIN64 - //64 bit windows doesn't seem to have an implementation of any of these load intrinsics, and 32 bit Visual C++ crashes - _declspec(align(16)) uint64_t temp[2] = {x, x}; - return _mm_load_si128((const __m128i*)temp); -#elif defined(__ICC) - uint64_t __attribute__((aligned(16))) temp[2] = {x,x}; - return _mm_load_si128((const __m128i*)temp); -#elif __clang__ -#pragma clang diagnostic push -#pragma clang diagnostic ignored "-Wc++11-narrowing" // narrowing from uint64_t (aka 'unsigned long') to 'long long' - // clang does not provide this intrinsic either - return (__m128i){x, x}; -#pragma clang diagnostic pop -#else - // everyone else gets it right - return _mm_set1_epi64x(x); -#endif -} - -static inline __m128i libdivide_get_FFFFFFFF00000000(void) { - //returns the same as _mm_set1_epi64(0xFFFFFFFF00000000ULL) without touching memory - __m128i result = _mm_set1_epi8(-1); //optimizes to pcmpeqd on OS X - return _mm_slli_epi64(result, 32); -} - -static inline __m128i libdivide_get_00000000FFFFFFFF(void) { - //returns the same as _mm_set1_epi64(0x00000000FFFFFFFFULL) without touching memory - __m128i result = _mm_set1_epi8(-1); //optimizes to pcmpeqd on OS X - result = _mm_srli_epi64(result, 32); - return result; -} - -#if __clang__ -#pragma clang diagnostic push -#pragma clang diagnostic ignored "-Wuninitialized" -#endif -static inline __m128i libdivide_get_0000FFFF(void) { - //returns the same as _mm_set1_epi32(0x0000FFFFULL) without touching memory - __m128i result; //we don't care what its contents are - result = _mm_cmpeq_epi8(result, result); //all 1s - result = _mm_srli_epi32(result, 16); - return result; -} -#if __clang__ -#pragma clang diagnostic pop -#endif - -/// This is a bug in gcc-8, _MM_SHUFFLE was forgotten, though in trunk it is ok https://github.com/gcc-mirror/gcc/blob/master/gcc/config/rs6000/xmmintrin.h#L61 -#if defined(__PPC__) -#ifndef _MM_SHUFFLE -#define _MM_SHUFFLE(w,x,y,z) (((w) << 6) | ((x) << 4) | ((y) << 2) | (z)) -#endif -#endif - -static inline __m128i libdivide_s64_signbits(__m128i v) { - //we want to compute v >> 63, that is, _mm_srai_epi64(v, 63). But there is no 64 bit shift right arithmetic instruction in SSE2. So we have to fake it by first duplicating the high 32 bit values, and then using a 32 bit shift. Another option would be to use _mm_srli_epi64(v, 63) and then subtract that from 0, but that approach appears to be substantially slower for unknown reasons - __m128i hiBitsDuped = _mm_shuffle_epi32(v, _MM_SHUFFLE(3, 3, 1, 1)); - __m128i signBits = _mm_srai_epi32(hiBitsDuped, 31); - return signBits; -} - -/* Returns an __m128i whose low 32 bits are equal to amt and has zero elsewhere. */ -static inline __m128i libdivide_u32_to_m128i(uint32_t amt) { - return _mm_set_epi32(0, 0, 0, amt); -} - -static inline __m128i libdivide_s64_shift_right_vector(__m128i v, int amt) { - //implementation of _mm_sra_epi64. Here we have two 64 bit values which are shifted right to logically become (64 - amt) values, and are then sign extended from a (64 - amt) bit number. - const int b = 64 - amt; - __m128i m = libdivide__u64_to_m128(1ULL << (b - 1)); - __m128i x = _mm_srl_epi64(v, libdivide_u32_to_m128i(amt)); - __m128i result = _mm_sub_epi64(_mm_xor_si128(x, m), m); //result = x^m - m - return result; -} - -/* Here, b is assumed to contain one 32 bit value repeated four times. If it did not, the function would not work. */ -static inline __m128i libdivide__mullhi_u32_flat_vector(__m128i a, __m128i b) { - __m128i hi_product_0Z2Z = _mm_srli_epi64(_mm_mul_epu32(a, b), 32); - __m128i a1X3X = _mm_srli_epi64(a, 32); - __m128i hi_product_Z1Z3 = _mm_and_si128(_mm_mul_epu32(a1X3X, b), libdivide_get_FFFFFFFF00000000()); - return _mm_or_si128(hi_product_0Z2Z, hi_product_Z1Z3); // = hi_product_0123 -} - - -/* Here, y is assumed to contain one 64 bit value repeated twice. */ -static inline __m128i libdivide_mullhi_u64_flat_vector(__m128i x, __m128i y) { - //full 128 bits are x0 * y0 + (x0 * y1 << 32) + (x1 * y0 << 32) + (x1 * y1 << 64) - const __m128i mask = libdivide_get_00000000FFFFFFFF(); - const __m128i x0 = _mm_and_si128(x, mask), x1 = _mm_srli_epi64(x, 32); //x0 is low half of 2 64 bit values, x1 is high half in low slots - const __m128i y0 = _mm_and_si128(y, mask), y1 = _mm_srli_epi64(y, 32); - const __m128i x0y0_hi = _mm_srli_epi64(_mm_mul_epu32(x0, y0), 32); //x0 happens to have the low half of the two 64 bit values in 32 bit slots 0 and 2, so _mm_mul_epu32 computes their full product, and then we shift right by 32 to get just the high values - const __m128i x0y1 = _mm_mul_epu32(x0, y1); - const __m128i x1y0 = _mm_mul_epu32(x1, y0); - const __m128i x1y1 = _mm_mul_epu32(x1, y1); - - const __m128i temp = _mm_add_epi64(x1y0, x0y0_hi); - __m128i temp_lo = _mm_and_si128(temp, mask), temp_hi = _mm_srli_epi64(temp, 32); - temp_lo = _mm_srli_epi64(_mm_add_epi64(temp_lo, x0y1), 32); - temp_hi = _mm_add_epi64(x1y1, temp_hi); - - return _mm_add_epi64(temp_lo, temp_hi); -} - -/* y is one 64 bit value repeated twice */ -static inline __m128i libdivide_mullhi_s64_flat_vector(__m128i x, __m128i y) { - __m128i p = libdivide_mullhi_u64_flat_vector(x, y); - __m128i t1 = _mm_and_si128(libdivide_s64_signbits(x), y); - p = _mm_sub_epi64(p, t1); - __m128i t2 = _mm_and_si128(libdivide_s64_signbits(y), x); - p = _mm_sub_epi64(p, t2); - return p; -} - -#ifdef LIBDIVIDE_USE_SSE4_1 - -/* b is one 32 bit value repeated four times. */ -static inline __m128i libdivide_mullhi_s32_flat_vector(__m128i a, __m128i b) { - __m128i hi_product_0Z2Z = _mm_srli_epi64(_mm_mul_epi32(a, b), 32); - __m128i a1X3X = _mm_srli_epi64(a, 32); - __m128i hi_product_Z1Z3 = _mm_and_si128(_mm_mul_epi32(a1X3X, b), libdivide_get_FFFFFFFF00000000()); - return _mm_or_si128(hi_product_0Z2Z, hi_product_Z1Z3); // = hi_product_0123 -} - -#else - -/* SSE2 does not have a signed multiplication instruction, but we can convert unsigned to signed pretty efficiently. Again, b is just a 32 bit value repeated four times. */ -static inline __m128i libdivide_mullhi_s32_flat_vector(__m128i a, __m128i b) { - __m128i p = libdivide__mullhi_u32_flat_vector(a, b); - __m128i t1 = _mm_and_si128(_mm_srai_epi32(a, 31), b); //t1 = (a >> 31) & y, arithmetic shift - __m128i t2 = _mm_and_si128(_mm_srai_epi32(b, 31), a); - p = _mm_sub_epi32(p, t1); - p = _mm_sub_epi32(p, t2); - return p; -} -#endif -#endif - -static inline int32_t libdivide__count_trailing_zeros32(uint32_t val) { -#if __GNUC__ || __has_builtin(__builtin_ctz) - /* Fast way to count trailing zeros */ - return __builtin_ctz(val); -#elif LIBDIVIDE_VC - unsigned long result; - if (_BitScanForward(&result, val)) { - return result; - } - return 0; -#else - /* Dorky way to count trailing zeros. Note that this hangs for val = 0! */ - int32_t result = 0; - val = (val ^ (val - 1)) >> 1; // Set v's trailing 0s to 1s and zero rest - while (val) { - val >>= 1; - result++; - } - return result; -#endif -} - -static inline int32_t libdivide__count_trailing_zeros64(uint64_t val) { -#if __LP64__ && (__GNUC__ || __has_builtin(__builtin_ctzll)) - /* Fast way to count trailing zeros. Note that we disable this in 32 bit because gcc does something horrible - it calls through to a dynamically bound function. */ - return __builtin_ctzll(val); -#elif LIBDIVIDE_VC && _WIN64 - unsigned long result; - if (_BitScanForward64(&result, val)) { - return result; - } - return 0; -#else - /* Pretty good way to count trailing zeros. Note that this hangs for val = 0! */ - uint32_t lo = val & 0xFFFFFFFF; - if (lo != 0) return libdivide__count_trailing_zeros32(lo); - return 32 + libdivide__count_trailing_zeros32(val >> 32); -#endif -} - -static inline int32_t libdivide__count_leading_zeros32(uint32_t val) { -#if __GNUC__ || __has_builtin(__builtin_clzll) - /* Fast way to count leading zeros */ +static inline int32_t libdivide_count_leading_zeros32(uint32_t val) { +#if defined(__GNUC__) || \ + __has_builtin(__builtin_clz) + // Fast way to count leading zeros return __builtin_clz(val); -#elif LIBDIVIDE_VC +#elif defined(LIBDIVIDE_VC) unsigned long result; if (_BitScanReverse(&result, val)) { return 31 - result; } return 0; #else - /* Dorky way to count leading zeros. Note that this hangs for val = 0! */ int32_t result = 0; - while (! (val & (1U << 31))) { - val <<= 1; + uint32_t hi = 1U << 31; + for (; ~val & hi; hi >>= 1) { result++; } return result; #endif } -static inline int32_t libdivide__count_leading_zeros64(uint64_t val) { -#if __GNUC__ || __has_builtin(__builtin_clzll) - /* Fast way to count leading zeros */ +static inline int32_t libdivide_count_leading_zeros64(uint64_t val) { +#if defined(__GNUC__) || \ + __has_builtin(__builtin_clzll) + // Fast way to count leading zeros return __builtin_clzll(val); -#elif LIBDIVIDE_VC && _WIN64 +#elif defined(LIBDIVIDE_VC) && defined(_WIN64) unsigned long result; if (_BitScanReverse64(&result, val)) { return 63 - result; } return 0; #else - /* Dorky way to count leading zeros. Note that this hangs for val = 0! */ - int32_t result = 0; - while (! (val & (1ULL << 63))) { - val <<= 1; - result++; - } - return result; + uint32_t hi = val >> 32; + uint32_t lo = val & 0xFFFFFFFF; + if (hi != 0) return libdivide_count_leading_zeros32(hi); + return 32 + libdivide_count_leading_zeros32(lo); #endif } -//libdivide_64_div_32_to_32: divides a 64 bit uint {u1, u0} by a 32 bit uint {v}. The result must fit in 32 bits. Returns the quotient directly and the remainder in *r -#if (LIBDIVIDE_IS_i386 || LIBDIVIDE_IS_X86_64) && LIBDIVIDE_GCC_STYLE_ASM -static uint32_t libdivide_64_div_32_to_32(uint32_t u1, uint32_t u0, uint32_t v, uint32_t *r) { +// libdivide_64_div_32_to_32: divides a 64-bit uint {u1, u0} by a 32-bit +// uint {v}. The result must fit in 32 bits. +// Returns the quotient directly and the remainder in *r +static inline uint32_t libdivide_64_div_32_to_32(uint32_t u1, uint32_t u0, uint32_t v, uint32_t *r) { +#if (defined(LIBDIVIDE_i386) || defined(LIBDIVIDE_X86_64)) && \ + defined(LIBDIVIDE_GCC_STYLE_ASM) uint32_t result; __asm__("divl %[v]" : "=a"(result), "=d"(*r) : [v] "r"(v), "a"(u0), "d"(u1) ); return result; -} #else -static uint32_t libdivide_64_div_32_to_32(uint32_t u1, uint32_t u0, uint32_t v, uint32_t *r) { - uint64_t n = (((uint64_t)u1) << 32) | u0; + uint64_t n = ((uint64_t)u1 << 32) | u0; uint32_t result = (uint32_t)(n / v); *r = (uint32_t)(n - result * (uint64_t)v); return result; -} #endif +} -#if LIBDIVIDE_IS_X86_64 && LIBDIVIDE_GCC_STYLE_ASM +// libdivide_128_div_64_to_64: divides a 128-bit uint {u1, u0} by a 64-bit +// uint {v}. The result must fit in 64 bits. +// Returns the quotient directly and the remainder in *r static uint64_t libdivide_128_div_64_to_64(uint64_t u1, uint64_t u0, uint64_t v, uint64_t *r) { - //u0 -> rax - //u1 -> rdx - //divq +#if defined(LIBDIVIDE_X86_64) && \ + defined(LIBDIVIDE_GCC_STYLE_ASM) uint64_t result; __asm__("divq %[v]" : "=a"(result), "=d"(*r) : [v] "r"(v), "a"(u0), "d"(u1) ); return result; - -} +#elif defined(HAS_INT128_T) && \ + defined(HAS_INT128_DIV) + __uint128_t n = ((__uint128_t)u1 << 64) | u0; + uint64_t result = (uint64_t)(n / v); + *r = (uint64_t)(n - result * (__uint128_t)v); + return result; #else + // Code taken from Hacker's Delight: + // http://www.hackersdelight.org/HDcode/divlu.c. + // License permits inclusion here per: + // http://www.hackersdelight.org/permissions.htm -/* Code taken from Hacker's Delight, http://www.hackersdelight.org/HDcode/divlu.c . License permits inclusion here per http://www.hackersdelight.org/permissions.htm - */ -static uint64_t libdivide_128_div_64_to_64(uint64_t u1, uint64_t u0, uint64_t v, uint64_t *r) { - const uint64_t b = (1ULL << 32); // Number base (16 bits). - uint64_t un1, un0, // Norm. dividend LSD's. - vn1, vn0, // Norm. divisor digits. - q1, q0, // Quotient digits. - un64, un21, un10,// Dividend digit pairs. - rhat; // A remainder. - int s; // Shift amount for norm. + const uint64_t b = (1ULL << 32); // Number base (32 bits) + uint64_t un1, un0; // Norm. dividend LSD's + uint64_t vn1, vn0; // Norm. divisor digits + uint64_t q1, q0; // Quotient digits + uint64_t un64, un21, un10; // Dividend digit pairs + uint64_t rhat; // A remainder + int32_t s; // Shift amount for norm - if (u1 >= v) { // If overflow, set rem. - if (r != NULL) // to an impossible value, - *r = (uint64_t)(-1); // and return the largest - return (uint64_t)(-1);} // possible quotient. + // If overflow, set rem. to an impossible value, + // and return the largest possible quotient + if (u1 >= v) { + *r = (uint64_t) -1; + return (uint64_t) -1; + } - /* count leading zeros */ - s = libdivide__count_leading_zeros64(v); // 0 <= s <= 63. + // count leading zeros + s = libdivide_count_leading_zeros64(v); if (s > 0) { - v = v << s; // Normalize divisor. - un64 = (u1 << s) | ((u0 >> (64 - s)) & (-s >> 31)); - un10 = u0 << s; // Shift dividend left. + // Normalize divisor + v = v << s; + un64 = (u1 << s) | (u0 >> (64 - s)); + un10 = u0 << s; // Shift dividend left } else { - // Avoid undefined behavior. - un64 = u1 | u0; + // Avoid undefined behavior of (u0 >> 64). + // The behavior is undefined if the right operand is + // negative, or greater than or equal to the length + // in bits of the promoted left operand. + un64 = u1; un10 = u0; } - vn1 = v >> 32; // Break divisor up into - vn0 = v & 0xFFFFFFFF; // two 32-bit digits. + // Break divisor up into two 32-bit digits + vn1 = v >> 32; + vn0 = v & 0xFFFFFFFF; - un1 = un10 >> 32; // Break right half of - un0 = un10 & 0xFFFFFFFF; // dividend into two digits. + // Break right half of dividend into two digits + un1 = un10 >> 32; + un0 = un10 & 0xFFFFFFFF; - q1 = un64/vn1; // Compute the first - rhat = un64 - q1*vn1; // quotient digit, q1. -again1: - if (q1 >= b || q1*vn0 > b*rhat + un1) { + // Compute the first quotient digit, q1 + q1 = un64 / vn1; + rhat = un64 - q1 * vn1; + + while (q1 >= b || q1 * vn0 > b * rhat + un1) { q1 = q1 - 1; rhat = rhat + vn1; - if (rhat < b) goto again1;} + if (rhat >= b) + break; + } - un21 = un64*b + un1 - q1*v; // Multiply and subtract. + // Multiply and subtract + un21 = un64 * b + un1 - q1 * v; - q0 = un21/vn1; // Compute the second - rhat = un21 - q0*vn1; // quotient digit, q0. -again2: - if (q0 >= b || q0*vn0 > b*rhat + un0) { + // Compute the second quotient digit + q0 = un21 / vn1; + rhat = un21 - q0 * vn1; + + while (q0 >= b || q0 * vn0 > b * rhat + un0) { q0 = q0 - 1; rhat = rhat + vn1; - if (rhat < b) goto again2;} + if (rhat >= b) + break; + } - if (r != NULL) // If remainder is wanted, - *r = (un21*b + un0 - q0*v) >> s; // return it. - return q1*b + q0; + *r = (un21 * b + un0 - q0 * v) >> s; + return q1 * b + q0; +#endif } -#endif -#if LIBDIVIDE_ASSERTIONS_ON -#define LIBDIVIDE_ASSERT(x) do { if (! (x)) { fprintf(stderr, "Assertion failure on line %ld: %s\n", (long)__LINE__, #x); exit(-1); } } while (0) +// Bitshift a u128 in place, left (signed_shift > 0) or right (signed_shift < 0) +static inline void libdivide_u128_shift(uint64_t *u1, uint64_t *u0, int32_t signed_shift) { + if (signed_shift > 0) { + uint32_t shift = signed_shift; + *u1 <<= shift; + *u1 |= *u0 >> (64 - shift); + *u0 <<= shift; + } + else if (signed_shift < 0) { + uint32_t shift = -signed_shift; + *u0 >>= shift; + *u0 |= *u1 << (64 - shift); + *u1 >>= shift; + } +} + +// Computes a 128 / 128 -> 64 bit division, with a 128 bit remainder. +static uint64_t libdivide_128_div_128_to_64(uint64_t u_hi, uint64_t u_lo, uint64_t v_hi, uint64_t v_lo, uint64_t *r_hi, uint64_t *r_lo) { +#if defined(HAS_INT128_T) && \ + defined(HAS_INT128_DIV) + __uint128_t ufull = u_hi; + __uint128_t vfull = v_hi; + ufull = (ufull << 64) | u_lo; + vfull = (vfull << 64) | v_lo; + uint64_t res = (uint64_t)(ufull / vfull); + __uint128_t remainder = ufull - (vfull * res); + *r_lo = (uint64_t)remainder; + *r_hi = (uint64_t)(remainder >> 64); + return res; #else -#define LIBDIVIDE_ASSERT(x) -#endif + // Adapted from "Unsigned Doubleword Division" in Hacker's Delight + // We want to compute u / v + typedef struct { uint64_t hi; uint64_t lo; } u128_t; + u128_t u = {u_hi, u_lo}; + u128_t v = {v_hi, v_lo}; -#ifndef LIBDIVIDE_HEADER_ONLY + if (v.hi == 0) { + // divisor v is a 64 bit value, so we just need one 128/64 division + // Note that we are simpler than Hacker's Delight here, because we know + // the quotient fits in 64 bits whereas Hacker's Delight demands a full + // 128 bit quotient + *r_hi = 0; + return libdivide_128_div_64_to_64(u.hi, u.lo, v.lo, r_lo); + } + // Here v >= 2**64 + // We know that v.hi != 0, so count leading zeros is OK + // We have 0 <= n <= 63 + uint32_t n = libdivide_count_leading_zeros64(v.hi); + + // Normalize the divisor so its MSB is 1 + u128_t v1t = v; + libdivide_u128_shift(&v1t.hi, &v1t.lo, n); + uint64_t v1 = v1t.hi; // i.e. v1 = v1t >> 64 + + // To ensure no overflow + u128_t u1 = u; + libdivide_u128_shift(&u1.hi, &u1.lo, -1); + + // Get quotient from divide unsigned insn. + uint64_t rem_ignored; + uint64_t q1 = libdivide_128_div_64_to_64(u1.hi, u1.lo, v1, &rem_ignored); + + // Undo normalization and division of u by 2. + u128_t q0 = {0, q1}; + libdivide_u128_shift(&q0.hi, &q0.lo, n); + libdivide_u128_shift(&q0.hi, &q0.lo, -63); + + // Make q0 correct or too small by 1 + // Equivalent to `if (q0 != 0) q0 = q0 - 1;` + if (q0.hi != 0 || q0.lo != 0) { + q0.hi -= (q0.lo == 0); // borrow + q0.lo -= 1; + } + + // Now q0 is correct. + // Compute q0 * v as q0v + // = (q0.hi << 64 + q0.lo) * (v.hi << 64 + v.lo) + // = (q0.hi * v.hi << 128) + (q0.hi * v.lo << 64) + + // (q0.lo * v.hi << 64) + q0.lo * v.lo) + // Each term is 128 bit + // High half of full product (upper 128 bits!) are dropped + u128_t q0v = {0, 0}; + q0v.hi = q0.hi*v.lo + q0.lo*v.hi + libdivide_mullhi_u64(q0.lo, v.lo); + q0v.lo = q0.lo*v.lo; + + // Compute u - q0v as u_q0v + // This is the remainder + u128_t u_q0v = u; + u_q0v.hi -= q0v.hi + (u.lo < q0v.lo); // second term is borrow + u_q0v.lo -= q0v.lo; + + // Check if u_q0v >= v + // This checks if our remainder is larger than the divisor + if ((u_q0v.hi > v.hi) || + (u_q0v.hi == v.hi && u_q0v.lo >= v.lo)) { + // Increment q0 + q0.lo += 1; + q0.hi += (q0.lo == 0); // carry + + // Subtract v from remainder + u_q0v.hi -= v.hi + (u_q0v.lo < v.lo); + u_q0v.lo -= v.lo; + } + + *r_hi = u_q0v.hi; + *r_lo = u_q0v.lo; + + LIBDIVIDE_ASSERT(q0.hi == 0); + return q0.lo; +#endif +} ////////// UINT32 -struct libdivide_u32_t libdivide_u32_gen(uint32_t d) { - struct libdivide_u32_t result; - if ((d & (d - 1)) == 0) { - result.magic = 0; - result.more = libdivide__count_trailing_zeros32(d) | LIBDIVIDE_U32_SHIFT_PATH; +static inline struct libdivide_u32_t libdivide_internal_u32_gen(uint32_t d, int branchfree) { + if (d == 0) { + LIBDIVIDE_ERROR("divider must be != 0"); } - else { - const uint32_t floor_log_2_d = 31 - libdivide__count_leading_zeros32(d); + struct libdivide_u32_t result; + uint32_t floor_log_2_d = 31 - libdivide_count_leading_zeros32(d); + + // Power of 2 + if ((d & (d - 1)) == 0) { + // We need to subtract 1 from the shift value in case of an unsigned + // branchfree divider because there is a hardcoded right shift by 1 + // in its division algorithm. Because of this we also need to add back + // 1 in its recovery algorithm. + result.magic = 0; + result.more = (uint8_t)(floor_log_2_d - (branchfree != 0)); + } else { uint8_t more; uint32_t rem, proposed_m; proposed_m = libdivide_64_div_32_to_32(1U << floor_log_2_d, 0, d, &rem); @@ -631,570 +569,1358 @@ struct libdivide_u32_t libdivide_u32_gen(uint32_t d) { LIBDIVIDE_ASSERT(rem > 0 && rem < d); const uint32_t e = d - rem; - /* This power works if e < 2**floor_log_2_d. */ - if (e < (1U << floor_log_2_d)) { - /* This power works */ + // This power works if e < 2**floor_log_2_d. + if (!branchfree && (e < (1U << floor_log_2_d))) { + // This power works more = floor_log_2_d; - } - else { - /* We have to use the general 33-bit algorithm. We need to compute (2**power) / d. However, we already have (2**(power-1))/d and its remainder. By doubling both, and then correcting the remainder, we can compute the larger division. */ - proposed_m += proposed_m; //don't care about overflow here - in fact, we expect it + } else { + // We have to use the general 33-bit algorithm. We need to compute + // (2**power) / d. However, we already have (2**(power-1))/d and + // its remainder. By doubling both, and then correcting the + // remainder, we can compute the larger division. + // don't care about overflow here - in fact, we expect it + proposed_m += proposed_m; const uint32_t twice_rem = rem + rem; if (twice_rem >= d || twice_rem < rem) proposed_m += 1; more = floor_log_2_d | LIBDIVIDE_ADD_MARKER; } result.magic = 1 + proposed_m; result.more = more; - //result.more's shift should in general be ceil_log_2_d. But if we used the smaller power, we subtract one from the shift because we're using the smaller power. If we're using the larger power, we subtract one from the shift because it's taken care of by the add indicator. So floor_log_2_d happens to be correct in both cases. - + // result.more's shift should in general be ceil_log_2_d. But if we + // used the smaller power, we subtract one from the shift because we're + // using the smaller power. If we're using the larger power, we + // subtract one from the shift because it's taken care of by the add + // indicator. So floor_log_2_d happens to be correct in both cases. } return result; } +struct libdivide_u32_t libdivide_u32_gen(uint32_t d) { + return libdivide_internal_u32_gen(d, 0); +} + +struct libdivide_u32_branchfree_t libdivide_u32_branchfree_gen(uint32_t d) { + if (d == 1) { + LIBDIVIDE_ERROR("branchfree divider must be != 1"); + } + struct libdivide_u32_t tmp = libdivide_internal_u32_gen(d, 1); + struct libdivide_u32_branchfree_t ret = {tmp.magic, (uint8_t)(tmp.more & LIBDIVIDE_32_SHIFT_MASK)}; + return ret; +} + uint32_t libdivide_u32_do(uint32_t numer, const struct libdivide_u32_t *denom) { uint8_t more = denom->more; - if (more & LIBDIVIDE_U32_SHIFT_PATH) { - return numer >> (more & LIBDIVIDE_32_SHIFT_MASK); + if (!denom->magic) { + return numer >> more; } else { - uint32_t q = libdivide__mullhi_u32(denom->magic, numer); + uint32_t q = libdivide_mullhi_u32(denom->magic, numer); if (more & LIBDIVIDE_ADD_MARKER) { uint32_t t = ((numer - q) >> 1) + q; return t >> (more & LIBDIVIDE_32_SHIFT_MASK); } else { - return q >> more; //all upper bits are 0 - don't need to mask them off + // All upper bits are 0, + // don't need to mask them off. + return q >> more; } } } - -int libdivide_u32_get_algorithm(const struct libdivide_u32_t *denom) { - uint8_t more = denom->more; - if (more & LIBDIVIDE_U32_SHIFT_PATH) return 0; - else if (! (more & LIBDIVIDE_ADD_MARKER)) return 1; - else return 2; -} - -uint32_t libdivide_u32_do_alg0(uint32_t numer, const struct libdivide_u32_t *denom) { - return numer >> (denom->more & LIBDIVIDE_32_SHIFT_MASK); -} - -uint32_t libdivide_u32_do_alg1(uint32_t numer, const struct libdivide_u32_t *denom) { - uint32_t q = libdivide__mullhi_u32(denom->magic, numer); - return q >> denom->more; -} - -uint32_t libdivide_u32_do_alg2(uint32_t numer, const struct libdivide_u32_t *denom) { - // denom->add != 0 - uint32_t q = libdivide__mullhi_u32(denom->magic, numer); +uint32_t libdivide_u32_branchfree_do(uint32_t numer, const struct libdivide_u32_branchfree_t *denom) { + uint32_t q = libdivide_mullhi_u32(denom->magic, numer); uint32_t t = ((numer - q) >> 1) + q; - return t >> (denom->more & LIBDIVIDE_32_SHIFT_MASK); + return t >> denom->more; } - - - -#if LIBDIVIDE_USE_SSE2 -__m128i libdivide_u32_do_vector(__m128i numers, const struct libdivide_u32_t *denom) { +uint32_t libdivide_u32_recover(const struct libdivide_u32_t *denom) { uint8_t more = denom->more; - if (more & LIBDIVIDE_U32_SHIFT_PATH) { - return _mm_srl_epi32(numers, libdivide_u32_to_m128i(more & LIBDIVIDE_32_SHIFT_MASK)); - } - else { - __m128i q = libdivide__mullhi_u32_flat_vector(numers, _mm_set1_epi32(denom->magic)); - if (more & LIBDIVIDE_ADD_MARKER) { - //uint32_t t = ((numer - q) >> 1) + q; - //return t >> denom->shift; - __m128i t = _mm_add_epi32(_mm_srli_epi32(_mm_sub_epi32(numers, q), 1), q); - return _mm_srl_epi32(t, libdivide_u32_to_m128i(more & LIBDIVIDE_32_SHIFT_MASK)); + uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK; - } - else { - //q >> denom->shift - return _mm_srl_epi32(q, libdivide_u32_to_m128i(more)); - } + if (!denom->magic) { + return 1U << shift; + } else if (!(more & LIBDIVIDE_ADD_MARKER)) { + // We compute q = n/d = n*m / 2^(32 + shift) + // Therefore we have d = 2^(32 + shift) / m + // We need to ceil it. + // We know d is not a power of 2, so m is not a power of 2, + // so we can just add 1 to the floor + uint32_t hi_dividend = 1U << shift; + uint32_t rem_ignored; + return 1 + libdivide_64_div_32_to_32(hi_dividend, 0, denom->magic, &rem_ignored); + } else { + // Here we wish to compute d = 2^(32+shift+1)/(m+2^32). + // Notice (m + 2^32) is a 33 bit number. Use 64 bit division for now + // Also note that shift may be as high as 31, so shift + 1 will + // overflow. So we have to compute it as 2^(32+shift)/(m+2^32), and + // then double the quotient and remainder. + uint64_t half_n = 1ULL << (32 + shift); + uint64_t d = (1ULL << 32) | denom->magic; + // Note that the quotient is guaranteed <= 32 bits, but the remainder + // may need 33! + uint32_t half_q = (uint32_t)(half_n / d); + uint64_t rem = half_n % d; + // We computed 2^(32+shift)/(m+2^32) + // Need to double it, and then add 1 to the quotient if doubling th + // remainder would increase the quotient. + // Note that rem<<1 cannot overflow, since rem < d and d is 33 bits + uint32_t full_q = half_q + half_q + ((rem<<1) >= d); + + // We rounded down in gen (hence +1) + return full_q + 1; } } -__m128i libdivide_u32_do_vector_alg0(__m128i numers, const struct libdivide_u32_t *denom) { - return _mm_srl_epi32(numers, libdivide_u32_to_m128i(denom->more & LIBDIVIDE_32_SHIFT_MASK)); -} +uint32_t libdivide_u32_branchfree_recover(const struct libdivide_u32_branchfree_t *denom) { + uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK; -__m128i libdivide_u32_do_vector_alg1(__m128i numers, const struct libdivide_u32_t *denom) { - __m128i q = libdivide__mullhi_u32_flat_vector(numers, _mm_set1_epi32(denom->magic)); - return _mm_srl_epi32(q, libdivide_u32_to_m128i(denom->more)); -} + if (!denom->magic) { + return 1U << (shift + 1); + } else { + // Here we wish to compute d = 2^(32+shift+1)/(m+2^32). + // Notice (m + 2^32) is a 33 bit number. Use 64 bit division for now + // Also note that shift may be as high as 31, so shift + 1 will + // overflow. So we have to compute it as 2^(32+shift)/(m+2^32), and + // then double the quotient and remainder. + uint64_t half_n = 1ULL << (32 + shift); + uint64_t d = (1ULL << 32) | denom->magic; + // Note that the quotient is guaranteed <= 32 bits, but the remainder + // may need 33! + uint32_t half_q = (uint32_t)(half_n / d); + uint64_t rem = half_n % d; + // We computed 2^(32+shift)/(m+2^32) + // Need to double it, and then add 1 to the quotient if doubling th + // remainder would increase the quotient. + // Note that rem<<1 cannot overflow, since rem < d and d is 33 bits + uint32_t full_q = half_q + half_q + ((rem<<1) >= d); -__m128i libdivide_u32_do_vector_alg2(__m128i numers, const struct libdivide_u32_t *denom) { - __m128i q = libdivide__mullhi_u32_flat_vector(numers, _mm_set1_epi32(denom->magic)); - __m128i t = _mm_add_epi32(_mm_srli_epi32(_mm_sub_epi32(numers, q), 1), q); - return _mm_srl_epi32(t, libdivide_u32_to_m128i(denom->more & LIBDIVIDE_32_SHIFT_MASK)); + // We rounded down in gen (hence +1) + return full_q + 1; + } } -#endif - /////////// UINT64 -struct libdivide_u64_t libdivide_u64_gen(uint64_t d) { - struct libdivide_u64_t result; - if ((d & (d - 1)) == 0) { - result.more = libdivide__count_trailing_zeros64(d) | LIBDIVIDE_U64_SHIFT_PATH; - result.magic = 0; +static inline struct libdivide_u64_t libdivide_internal_u64_gen(uint64_t d, int branchfree) { + if (d == 0) { + LIBDIVIDE_ERROR("divider must be != 0"); } - else { - const uint32_t floor_log_2_d = 63 - libdivide__count_leading_zeros64(d); + struct libdivide_u64_t result; + uint32_t floor_log_2_d = 63 - libdivide_count_leading_zeros64(d); + + // Power of 2 + if ((d & (d - 1)) == 0) { + // We need to subtract 1 from the shift value in case of an unsigned + // branchfree divider because there is a hardcoded right shift by 1 + // in its division algorithm. Because of this we also need to add back + // 1 in its recovery algorithm. + result.magic = 0; + result.more = (uint8_t)(floor_log_2_d - (branchfree != 0)); + } else { uint64_t proposed_m, rem; uint8_t more; - proposed_m = libdivide_128_div_64_to_64(1ULL << floor_log_2_d, 0, d, &rem); //== (1 << (64 + floor_log_2_d)) / d + // (1 << (64 + floor_log_2_d)) / d + proposed_m = libdivide_128_div_64_to_64(1ULL << floor_log_2_d, 0, d, &rem); LIBDIVIDE_ASSERT(rem > 0 && rem < d); const uint64_t e = d - rem; - /* This power works if e < 2**floor_log_2_d. */ - if (e < (1ULL << floor_log_2_d)) { - /* This power works */ + // This power works if e < 2**floor_log_2_d. + if (!branchfree && e < (1ULL << floor_log_2_d)) { + // This power works more = floor_log_2_d; - } - else { - /* We have to use the general 65-bit algorithm. We need to compute (2**power) / d. However, we already have (2**(power-1))/d and its remainder. By doubling both, and then correcting the remainder, we can compute the larger division. */ - proposed_m += proposed_m; //don't care about overflow here - in fact, we expect it + } else { + // We have to use the general 65-bit algorithm. We need to compute + // (2**power) / d. However, we already have (2**(power-1))/d and + // its remainder. By doubling both, and then correcting the + // remainder, we can compute the larger division. + // don't care about overflow here - in fact, we expect it + proposed_m += proposed_m; const uint64_t twice_rem = rem + rem; if (twice_rem >= d || twice_rem < rem) proposed_m += 1; - more = floor_log_2_d | LIBDIVIDE_ADD_MARKER; + more = floor_log_2_d | LIBDIVIDE_ADD_MARKER; } result.magic = 1 + proposed_m; result.more = more; - //result.more's shift should in general be ceil_log_2_d. But if we used the smaller power, we subtract one from the shift because we're using the smaller power. If we're using the larger power, we subtract one from the shift because it's taken care of by the add indicator. So floor_log_2_d happens to be correct in both cases, which is why we do it outside of the if statement. + // result.more's shift should in general be ceil_log_2_d. But if we + // used the smaller power, we subtract one from the shift because we're + // using the smaller power. If we're using the larger power, we + // subtract one from the shift because it's taken care of by the add + // indicator. So floor_log_2_d happens to be correct in both cases, + // which is why we do it outside of the if statement. } return result; } +struct libdivide_u64_t libdivide_u64_gen(uint64_t d) { + return libdivide_internal_u64_gen(d, 0); +} + +struct libdivide_u64_branchfree_t libdivide_u64_branchfree_gen(uint64_t d) { + if (d == 1) { + LIBDIVIDE_ERROR("branchfree divider must be != 1"); + } + struct libdivide_u64_t tmp = libdivide_internal_u64_gen(d, 1); + struct libdivide_u64_branchfree_t ret = {tmp.magic, (uint8_t)(tmp.more & LIBDIVIDE_64_SHIFT_MASK)}; + return ret; +} + uint64_t libdivide_u64_do(uint64_t numer, const struct libdivide_u64_t *denom) { uint8_t more = denom->more; - if (more & LIBDIVIDE_U64_SHIFT_PATH) { - return numer >> (more & LIBDIVIDE_64_SHIFT_MASK); + if (!denom->magic) { + return numer >> more; } else { - uint64_t q = libdivide__mullhi_u64(denom->magic, numer); + uint64_t q = libdivide_mullhi_u64(denom->magic, numer); if (more & LIBDIVIDE_ADD_MARKER) { uint64_t t = ((numer - q) >> 1) + q; return t >> (more & LIBDIVIDE_64_SHIFT_MASK); } else { - return q >> more; //all upper bits are 0 - don't need to mask them off + // All upper bits are 0, + // don't need to mask them off. + return q >> more; } } } - -int libdivide_u64_get_algorithm(const struct libdivide_u64_t *denom) { - uint8_t more = denom->more; - if (more & LIBDIVIDE_U64_SHIFT_PATH) return 0; - else if (! (more & LIBDIVIDE_ADD_MARKER)) return 1; - else return 2; -} - -uint64_t libdivide_u64_do_alg0(uint64_t numer, const struct libdivide_u64_t *denom) { - return numer >> (denom->more & LIBDIVIDE_64_SHIFT_MASK); -} - -uint64_t libdivide_u64_do_alg1(uint64_t numer, const struct libdivide_u64_t *denom) { - uint64_t q = libdivide__mullhi_u64(denom->magic, numer); - return q >> denom->more; -} - -uint64_t libdivide_u64_do_alg2(uint64_t numer, const struct libdivide_u64_t *denom) { - uint64_t q = libdivide__mullhi_u64(denom->magic, numer); +uint64_t libdivide_u64_branchfree_do(uint64_t numer, const struct libdivide_u64_branchfree_t *denom) { + uint64_t q = libdivide_mullhi_u64(denom->magic, numer); uint64_t t = ((numer - q) >> 1) + q; - return t >> (denom->more & LIBDIVIDE_64_SHIFT_MASK); + return t >> denom->more; } -#if LIBDIVIDE_USE_SSE2 -__m128i libdivide_u64_do_vector(__m128i numers, const struct libdivide_u64_t * denom) { +uint64_t libdivide_u64_recover(const struct libdivide_u64_t *denom) { uint8_t more = denom->more; - if (more & LIBDIVIDE_U64_SHIFT_PATH) { - return _mm_srl_epi64(numers, libdivide_u32_to_m128i(more & LIBDIVIDE_64_SHIFT_MASK)); - } - else { - __m128i q = libdivide_mullhi_u64_flat_vector(numers, libdivide__u64_to_m128(denom->magic)); - if (more & LIBDIVIDE_ADD_MARKER) { - //uint32_t t = ((numer - q) >> 1) + q; - //return t >> denom->shift; - __m128i t = _mm_add_epi64(_mm_srli_epi64(_mm_sub_epi64(numers, q), 1), q); - return _mm_srl_epi64(t, libdivide_u32_to_m128i(more & LIBDIVIDE_64_SHIFT_MASK)); - } - else { - //q >> denom->shift - return _mm_srl_epi64(q, libdivide_u32_to_m128i(more)); - } + uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + + if (!denom->magic) { + return 1ULL << shift; + } else if (!(more & LIBDIVIDE_ADD_MARKER)) { + // We compute q = n/d = n*m / 2^(64 + shift) + // Therefore we have d = 2^(64 + shift) / m + // We need to ceil it. + // We know d is not a power of 2, so m is not a power of 2, + // so we can just add 1 to the floor + uint64_t hi_dividend = 1ULL << shift; + uint64_t rem_ignored; + return 1 + libdivide_128_div_64_to_64(hi_dividend, 0, denom->magic, &rem_ignored); + } else { + // Here we wish to compute d = 2^(64+shift+1)/(m+2^64). + // Notice (m + 2^64) is a 65 bit number. This gets hairy. See + // libdivide_u32_recover for more on what we do here. + // TODO: do something better than 128 bit math + + // Full n is a (potentially) 129 bit value + // half_n is a 128 bit value + // Compute the hi half of half_n. Low half is 0. + uint64_t half_n_hi = 1ULL << shift, half_n_lo = 0; + // d is a 65 bit value. The high bit is always set to 1. + const uint64_t d_hi = 1, d_lo = denom->magic; + // Note that the quotient is guaranteed <= 64 bits, + // but the remainder may need 65! + uint64_t r_hi, r_lo; + uint64_t half_q = libdivide_128_div_128_to_64(half_n_hi, half_n_lo, d_hi, d_lo, &r_hi, &r_lo); + // We computed 2^(64+shift)/(m+2^64) + // Double the remainder ('dr') and check if that is larger than d + // Note that d is a 65 bit value, so r1 is small and so r1 + r1 + // cannot overflow + uint64_t dr_lo = r_lo + r_lo; + uint64_t dr_hi = r_hi + r_hi + (dr_lo < r_lo); // last term is carry + int dr_exceeds_d = (dr_hi > d_hi) || (dr_hi == d_hi && dr_lo >= d_lo); + uint64_t full_q = half_q + half_q + (dr_exceeds_d ? 1 : 0); + return full_q + 1; } } -__m128i libdivide_u64_do_vector_alg0(__m128i numers, const struct libdivide_u64_t *denom) { - return _mm_srl_epi64(numers, libdivide_u32_to_m128i(denom->more & LIBDIVIDE_64_SHIFT_MASK)); +uint64_t libdivide_u64_branchfree_recover(const struct libdivide_u64_branchfree_t *denom) { + uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + + if (!denom->magic) { + return 1ULL << (shift + 1); + } else { + // Here we wish to compute d = 2^(64+shift+1)/(m+2^64). + // Notice (m + 2^64) is a 65 bit number. This gets hairy. See + // libdivide_u32_recover for more on what we do here. + // TODO: do something better than 128 bit math + + // Full n is a (potentially) 129 bit value + // half_n is a 128 bit value + // Compute the hi half of half_n. Low half is 0. + uint64_t half_n_hi = 1ULL << shift, half_n_lo = 0; + // d is a 65 bit value. The high bit is always set to 1. + const uint64_t d_hi = 1, d_lo = denom->magic; + // Note that the quotient is guaranteed <= 64 bits, + // but the remainder may need 65! + uint64_t r_hi, r_lo; + uint64_t half_q = libdivide_128_div_128_to_64(half_n_hi, half_n_lo, d_hi, d_lo, &r_hi, &r_lo); + // We computed 2^(64+shift)/(m+2^64) + // Double the remainder ('dr') and check if that is larger than d + // Note that d is a 65 bit value, so r1 is small and so r1 + r1 + // cannot overflow + uint64_t dr_lo = r_lo + r_lo; + uint64_t dr_hi = r_hi + r_hi + (dr_lo < r_lo); // last term is carry + int dr_exceeds_d = (dr_hi > d_hi) || (dr_hi == d_hi && dr_lo >= d_lo); + uint64_t full_q = half_q + half_q + (dr_exceeds_d ? 1 : 0); + return full_q + 1; + } } -__m128i libdivide_u64_do_vector_alg1(__m128i numers, const struct libdivide_u64_t *denom) { - __m128i q = libdivide_mullhi_u64_flat_vector(numers, libdivide__u64_to_m128(denom->magic)); - return _mm_srl_epi64(q, libdivide_u32_to_m128i(denom->more)); -} - -__m128i libdivide_u64_do_vector_alg2(__m128i numers, const struct libdivide_u64_t *denom) { - __m128i q = libdivide_mullhi_u64_flat_vector(numers, libdivide__u64_to_m128(denom->magic)); - __m128i t = _mm_add_epi64(_mm_srli_epi64(_mm_sub_epi64(numers, q), 1), q); - return _mm_srl_epi64(t, libdivide_u32_to_m128i(denom->more & LIBDIVIDE_64_SHIFT_MASK)); -} - - -#endif - /////////// SINT32 +static inline struct libdivide_s32_t libdivide_internal_s32_gen(int32_t d, int branchfree) { + if (d == 0) { + LIBDIVIDE_ERROR("divider must be != 0"); + } -static inline int32_t libdivide__mullhi_s32(int32_t x, int32_t y) { - int64_t xl = x, yl = y; - int64_t rl = xl * yl; - return (int32_t)(rl >> 32); //needs to be arithmetic shift -} - -struct libdivide_s32_t libdivide_s32_gen(int32_t d) { struct libdivide_s32_t result; - /* If d is a power of 2, or negative a power of 2, we have to use a shift. This is especially important because the magic algorithm fails for -1. To check if d is a power of 2 or its inverse, it suffices to check whether its absolute value has exactly one bit set. This works even for INT_MIN, because abs(INT_MIN) == INT_MIN, and INT_MIN has one bit set and is a power of 2. */ - uint32_t absD = (uint32_t)(d < 0 ? -d : d); //gcc optimizes this to the fast abs trick - if ((absD & (absD - 1)) == 0) { //check if exactly one bit is set, don't care if absD is 0 since that's divide by zero + // If d is a power of 2, or negative a power of 2, we have to use a shift. + // This is especially important because the magic algorithm fails for -1. + // To check if d is a power of 2 or its inverse, it suffices to check + // whether its absolute value has exactly one bit set. This works even for + // INT_MIN, because abs(INT_MIN) == INT_MIN, and INT_MIN has one bit set + // and is a power of 2. + uint32_t ud = (uint32_t)d; + uint32_t absD = (d < 0) ? -ud : ud; + uint32_t floor_log_2_d = 31 - libdivide_count_leading_zeros32(absD); + // check if exactly one bit is set, + // don't care if absD is 0 since that's divide by zero + if ((absD & (absD - 1)) == 0) { + // Branchfree and normal paths are exactly the same result.magic = 0; - result.more = libdivide__count_trailing_zeros32(absD) | (d < 0 ? LIBDIVIDE_NEGATIVE_DIVISOR : 0) | LIBDIVIDE_S32_SHIFT_PATH; - } - else { - const uint32_t floor_log_2_d = 31 - libdivide__count_leading_zeros32(absD); + result.more = floor_log_2_d | (d < 0 ? LIBDIVIDE_NEGATIVE_DIVISOR : 0); + } else { LIBDIVIDE_ASSERT(floor_log_2_d >= 1); uint8_t more; - //the dividend here is 2**(floor_log_2_d + 31), so the low 32 bit word is 0 and the high word is floor_log_2_d - 1 + // the dividend here is 2**(floor_log_2_d + 31), so the low 32 bit word + // is 0 and the high word is floor_log_2_d - 1 uint32_t rem, proposed_m; proposed_m = libdivide_64_div_32_to_32(1U << (floor_log_2_d - 1), 0, absD, &rem); const uint32_t e = absD - rem; - /* We are going to start with a power of floor_log_2_d - 1. This works if works if e < 2**floor_log_2_d. */ - if (e < (1U << floor_log_2_d)) { - /* This power works */ + // We are going to start with a power of floor_log_2_d - 1. + // This works if works if e < 2**floor_log_2_d. + if (!branchfree && e < (1U << floor_log_2_d)) { + // This power works more = floor_log_2_d - 1; - } - else { - /* We need to go one higher. This should not make proposed_m overflow, but it will make it negative when interpreted as an int32_t. */ + } else { + // We need to go one higher. This should not make proposed_m + // overflow, but it will make it negative when interpreted as an + // int32_t. proposed_m += proposed_m; const uint32_t twice_rem = rem + rem; if (twice_rem >= absD || twice_rem < rem) proposed_m += 1; - more = floor_log_2_d | LIBDIVIDE_ADD_MARKER | (d < 0 ? LIBDIVIDE_NEGATIVE_DIVISOR : 0); //use the general algorithm + more = floor_log_2_d | LIBDIVIDE_ADD_MARKER; } - proposed_m += 1; - result.magic = (d < 0 ? -(int32_t)proposed_m : (int32_t)proposed_m); - result.more = more; + proposed_m += 1; + int32_t magic = (int32_t)proposed_m; + + // Mark if we are negative. Note we only negate the magic number in the + // branchfull case. + if (d < 0) { + more |= LIBDIVIDE_NEGATIVE_DIVISOR; + if (!branchfree) { + magic = -magic; + } + } + + result.more = more; + result.magic = magic; } return result; } +struct libdivide_s32_t libdivide_s32_gen(int32_t d) { + return libdivide_internal_s32_gen(d, 0); +} + +struct libdivide_s32_branchfree_t libdivide_s32_branchfree_gen(int32_t d) { + struct libdivide_s32_t tmp = libdivide_internal_s32_gen(d, 1); + struct libdivide_s32_branchfree_t result = {tmp.magic, tmp.more}; + return result; +} + int32_t libdivide_s32_do(int32_t numer, const struct libdivide_s32_t *denom) { uint8_t more = denom->more; - if (more & LIBDIVIDE_S32_SHIFT_PATH) { - uint8_t shifter = more & LIBDIVIDE_32_SHIFT_MASK; - int32_t q = numer + ((numer >> 31) & ((1 << shifter) - 1)); - q = q >> shifter; - int32_t shiftMask = (int8_t)more >> 7; //must be arithmetic shift and then sign-extend - q = (q ^ shiftMask) - shiftMask; + uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + + if (!denom->magic) { + uint32_t sign = (int8_t)more >> 7; + uint32_t mask = (1U << shift) - 1; + uint32_t uq = numer + ((numer >> 31) & mask); + int32_t q = (int32_t)uq; + q >>= shift; + q = (q ^ sign) - sign; return q; - } - else { - int32_t q = libdivide__mullhi_s32(denom->magic, numer); + } else { + uint32_t uq = (uint32_t)libdivide_mullhi_s32(denom->magic, numer); if (more & LIBDIVIDE_ADD_MARKER) { - int32_t sign = (int8_t)more >> 7; //must be arithmetic shift and then sign extend - q += ((numer ^ sign) - sign); + // must be arithmetic shift and then sign extend + int32_t sign = (int8_t)more >> 7; + // q += (more < 0 ? -numer : numer) + // cast required to avoid UB + uq += ((uint32_t)numer ^ sign) - sign; } - q >>= more & LIBDIVIDE_32_SHIFT_MASK; + int32_t q = (int32_t)uq; + q >>= shift; q += (q < 0); return q; } } -int libdivide_s32_get_algorithm(const struct libdivide_s32_t *denom) { +int32_t libdivide_s32_branchfree_do(int32_t numer, const struct libdivide_s32_branchfree_t *denom) { uint8_t more = denom->more; - int positiveDivisor = ! (more & LIBDIVIDE_NEGATIVE_DIVISOR); - if (more & LIBDIVIDE_S32_SHIFT_PATH) return (positiveDivisor ? 0 : 1); - else if (more & LIBDIVIDE_ADD_MARKER) return (positiveDivisor ? 2 : 3); - else return 4; -} - -int32_t libdivide_s32_do_alg0(int32_t numer, const struct libdivide_s32_t *denom) { - uint8_t shifter = denom->more & LIBDIVIDE_32_SHIFT_MASK; - int32_t q = numer + ((numer >> 31) & ((1 << shifter) - 1)); - return q >> shifter; -} - -int32_t libdivide_s32_do_alg1(int32_t numer, const struct libdivide_s32_t *denom) { - uint8_t shifter = denom->more & LIBDIVIDE_32_SHIFT_MASK; - int32_t q = numer + ((numer >> 31) & ((1 << shifter) - 1)); - return - (q >> shifter); -} - -int32_t libdivide_s32_do_alg2(int32_t numer, const struct libdivide_s32_t *denom) { - int32_t q = libdivide__mullhi_s32(denom->magic, numer); + uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + // must be arithmetic shift and then sign extend + int32_t sign = (int8_t)more >> 7; + int32_t magic = denom->magic; + int32_t q = libdivide_mullhi_s32(magic, numer); q += numer; - q >>= denom->more & LIBDIVIDE_32_SHIFT_MASK; - q += (q < 0); + + // If q is non-negative, we have nothing to do + // If q is negative, we want to add either (2**shift)-1 if d is a power of + // 2, or (2**shift) if it is not a power of 2 + uint32_t is_power_of_2 = (magic == 0); + uint32_t q_sign = (uint32_t)(q >> 31); + q += q_sign & ((1U << shift) - is_power_of_2); + + // Now arithmetic right shift + q >>= shift; + // Negate if needed + q = (q ^ sign) - sign; + return q; } -int32_t libdivide_s32_do_alg3(int32_t numer, const struct libdivide_s32_t *denom) { - int32_t q = libdivide__mullhi_s32(denom->magic, numer); - q -= numer; - q >>= denom->more & LIBDIVIDE_32_SHIFT_MASK; - q += (q < 0); - return q; -} - -int32_t libdivide_s32_do_alg4(int32_t numer, const struct libdivide_s32_t *denom) { - int32_t q = libdivide__mullhi_s32(denom->magic, numer); - q >>= denom->more & LIBDIVIDE_32_SHIFT_MASK; - q += (q < 0); - return q; -} - -#if LIBDIVIDE_USE_SSE2 -__m128i libdivide_s32_do_vector(__m128i numers, const struct libdivide_s32_t * denom) { +int32_t libdivide_s32_recover(const struct libdivide_s32_t *denom) { uint8_t more = denom->more; - if (more & LIBDIVIDE_S32_SHIFT_PATH) { - uint32_t shifter = more & LIBDIVIDE_32_SHIFT_MASK; - __m128i roundToZeroTweak = _mm_set1_epi32((1 << shifter) - 1); //could use _mm_srli_epi32 with an all -1 register - __m128i q = _mm_add_epi32(numers, _mm_and_si128(_mm_srai_epi32(numers, 31), roundToZeroTweak)); //q = numer + ((numer >> 31) & roundToZeroTweak); - q = _mm_sra_epi32(q, libdivide_u32_to_m128i(shifter)); // q = q >> shifter - __m128i shiftMask = _mm_set1_epi32((int32_t)((int8_t)more >> 7)); //set all bits of shift mask = to the sign bit of more - q = _mm_sub_epi32(_mm_xor_si128(q, shiftMask), shiftMask); //q = (q ^ shiftMask) - shiftMask; - return q; - } - else { - __m128i q = libdivide_mullhi_s32_flat_vector(numers, _mm_set1_epi32(denom->magic)); - if (more & LIBDIVIDE_ADD_MARKER) { - __m128i sign = _mm_set1_epi32((int32_t)(int8_t)more >> 7); //must be arithmetic shift - q = _mm_add_epi32(q, _mm_sub_epi32(_mm_xor_si128(numers, sign), sign)); // q += ((numer ^ sign) - sign); + uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + if (!denom->magic) { + uint32_t absD = 1U << shift; + if (more & LIBDIVIDE_NEGATIVE_DIVISOR) { + absD = -absD; } - q = _mm_sra_epi32(q, libdivide_u32_to_m128i(more & LIBDIVIDE_32_SHIFT_MASK)); //q >>= shift - q = _mm_add_epi32(q, _mm_srli_epi32(q, 31)); // q += (q < 0) - return q; + return (int32_t)absD; + } else { + // Unsigned math is much easier + // We negate the magic number only in the branchfull case, and we don't + // know which case we're in. However we have enough information to + // determine the correct sign of the magic number. The divisor was + // negative if LIBDIVIDE_NEGATIVE_DIVISOR is set. If ADD_MARKER is set, + // the magic number's sign is opposite that of the divisor. + // We want to compute the positive magic number. + int negative_divisor = (more & LIBDIVIDE_NEGATIVE_DIVISOR); + int magic_was_negated = (more & LIBDIVIDE_ADD_MARKER) + ? denom->magic > 0 : denom->magic < 0; + + // Handle the power of 2 case (including branchfree) + if (denom->magic == 0) { + int32_t result = 1U << shift; + return negative_divisor ? -result : result; + } + + uint32_t d = (uint32_t)(magic_was_negated ? -denom->magic : denom->magic); + uint64_t n = 1ULL << (32 + shift); // this shift cannot exceed 30 + uint32_t q = (uint32_t)(n / d); + int32_t result = (int32_t)q; + result += 1; + return negative_divisor ? -result : result; } } -__m128i libdivide_s32_do_vector_alg0(__m128i numers, const struct libdivide_s32_t *denom) { - uint8_t shifter = denom->more & LIBDIVIDE_32_SHIFT_MASK; - __m128i roundToZeroTweak = _mm_set1_epi32((1 << shifter) - 1); - __m128i q = _mm_add_epi32(numers, _mm_and_si128(_mm_srai_epi32(numers, 31), roundToZeroTweak)); - return _mm_sra_epi32(q, libdivide_u32_to_m128i(shifter)); +int32_t libdivide_s32_branchfree_recover(const struct libdivide_s32_branchfree_t *denom) { + return libdivide_s32_recover((const struct libdivide_s32_t *)denom); } -__m128i libdivide_s32_do_vector_alg1(__m128i numers, const struct libdivide_s32_t *denom) { - uint8_t shifter = denom->more & LIBDIVIDE_32_SHIFT_MASK; - __m128i roundToZeroTweak = _mm_set1_epi32((1 << shifter) - 1); - __m128i q = _mm_add_epi32(numers, _mm_and_si128(_mm_srai_epi32(numers, 31), roundToZeroTweak)); - return _mm_sub_epi32(_mm_setzero_si128(), _mm_sra_epi32(q, libdivide_u32_to_m128i(shifter))); -} - -__m128i libdivide_s32_do_vector_alg2(__m128i numers, const struct libdivide_s32_t *denom) { - __m128i q = libdivide_mullhi_s32_flat_vector(numers, _mm_set1_epi32(denom->magic)); - q = _mm_add_epi32(q, numers); - q = _mm_sra_epi32(q, libdivide_u32_to_m128i(denom->more & LIBDIVIDE_32_SHIFT_MASK)); - q = _mm_add_epi32(q, _mm_srli_epi32(q, 31)); - return q; -} - -__m128i libdivide_s32_do_vector_alg3(__m128i numers, const struct libdivide_s32_t *denom) { - __m128i q = libdivide_mullhi_s32_flat_vector(numers, _mm_set1_epi32(denom->magic)); - q = _mm_sub_epi32(q, numers); - q = _mm_sra_epi32(q, libdivide_u32_to_m128i(denom->more & LIBDIVIDE_32_SHIFT_MASK)); - q = _mm_add_epi32(q, _mm_srli_epi32(q, 31)); - return q; -} - -__m128i libdivide_s32_do_vector_alg4(__m128i numers, const struct libdivide_s32_t *denom) { - __m128i q = libdivide_mullhi_s32_flat_vector(numers, _mm_set1_epi32(denom->magic)); - q = _mm_sra_epi32(q, libdivide_u32_to_m128i(denom->more)); //q >>= shift - q = _mm_add_epi32(q, _mm_srli_epi32(q, 31)); // q += (q < 0) - return q; -} -#endif - ///////////// SINT64 +static inline struct libdivide_s64_t libdivide_internal_s64_gen(int64_t d, int branchfree) { + if (d == 0) { + LIBDIVIDE_ERROR("divider must be != 0"); + } -struct libdivide_s64_t libdivide_s64_gen(int64_t d) { struct libdivide_s64_t result; - /* If d is a power of 2, or negative a power of 2, we have to use a shift. This is especially important because the magic algorithm fails for -1. To check if d is a power of 2 or its inverse, it suffices to check whether its absolute value has exactly one bit set. This works even for INT_MIN, because abs(INT_MIN) == INT_MIN, and INT_MIN has one bit set and is a power of 2. */ - const uint64_t absD = (uint64_t)(d < 0 ? -d : d); //gcc optimizes this to the fast abs trick - if ((absD & (absD - 1)) == 0) { //check if exactly one bit is set, don't care if absD is 0 since that's divide by zero - result.more = libdivide__count_trailing_zeros64(absD) | (d < 0 ? LIBDIVIDE_NEGATIVE_DIVISOR : 0); + // If d is a power of 2, or negative a power of 2, we have to use a shift. + // This is especially important because the magic algorithm fails for -1. + // To check if d is a power of 2 or its inverse, it suffices to check + // whether its absolute value has exactly one bit set. This works even for + // INT_MIN, because abs(INT_MIN) == INT_MIN, and INT_MIN has one bit set + // and is a power of 2. + uint64_t ud = (uint64_t)d; + uint64_t absD = (d < 0) ? -ud : ud; + uint32_t floor_log_2_d = 63 - libdivide_count_leading_zeros64(absD); + // check if exactly one bit is set, + // don't care if absD is 0 since that's divide by zero + if ((absD & (absD - 1)) == 0) { + // Branchfree and non-branchfree cases are the same result.magic = 0; - } - else { - const uint32_t floor_log_2_d = 63 - libdivide__count_leading_zeros64(absD); - - //the dividend here is 2**(floor_log_2_d + 63), so the low 64 bit word is 0 and the high word is floor_log_2_d - 1 + result.more = floor_log_2_d | (d < 0 ? LIBDIVIDE_NEGATIVE_DIVISOR : 0); + } else { + // the dividend here is 2**(floor_log_2_d + 63), so the low 64 bit word + // is 0 and the high word is floor_log_2_d - 1 uint8_t more; uint64_t rem, proposed_m; proposed_m = libdivide_128_div_64_to_64(1ULL << (floor_log_2_d - 1), 0, absD, &rem); const uint64_t e = absD - rem; - /* We are going to start with a power of floor_log_2_d - 1. This works if works if e < 2**floor_log_2_d. */ - if (e < (1ULL << floor_log_2_d)) { - /* This power works */ + // We are going to start with a power of floor_log_2_d - 1. + // This works if works if e < 2**floor_log_2_d. + if (!branchfree && e < (1ULL << floor_log_2_d)) { + // This power works more = floor_log_2_d - 1; - } - else { - /* We need to go one higher. This should not make proposed_m overflow, but it will make it negative when interpreted as an int32_t. */ + } else { + // We need to go one higher. This should not make proposed_m + // overflow, but it will make it negative when interpreted as an + // int32_t. proposed_m += proposed_m; const uint64_t twice_rem = rem + rem; if (twice_rem >= absD || twice_rem < rem) proposed_m += 1; - more = floor_log_2_d | LIBDIVIDE_ADD_MARKER | (d < 0 ? LIBDIVIDE_NEGATIVE_DIVISOR : 0); + // note that we only set the LIBDIVIDE_NEGATIVE_DIVISOR bit if we + // also set ADD_MARKER this is an annoying optimization that + // enables algorithm #4 to avoid the mask. However we always set it + // in the branchfree case + more = floor_log_2_d | LIBDIVIDE_ADD_MARKER; } proposed_m += 1; + int64_t magic = (int64_t)proposed_m; + + // Mark if we are negative + if (d < 0) { + more |= LIBDIVIDE_NEGATIVE_DIVISOR; + if (!branchfree) { + magic = -magic; + } + } + result.more = more; - result.magic = (d < 0 ? -(int64_t)proposed_m : (int64_t)proposed_m); + result.magic = magic; } return result; } +struct libdivide_s64_t libdivide_s64_gen(int64_t d) { + return libdivide_internal_s64_gen(d, 0); +} + +struct libdivide_s64_branchfree_t libdivide_s64_branchfree_gen(int64_t d) { + struct libdivide_s64_t tmp = libdivide_internal_s64_gen(d, 1); + struct libdivide_s64_branchfree_t ret = {tmp.magic, tmp.more}; + return ret; +} + int64_t libdivide_s64_do(int64_t numer, const struct libdivide_s64_t *denom) { uint8_t more = denom->more; - int64_t magic = denom->magic; - if (magic == 0) { //shift path - uint32_t shifter = more & LIBDIVIDE_64_SHIFT_MASK; - int64_t q = numer + ((numer >> 63) & ((1LL << shifter) - 1)); - q = q >> shifter; - int64_t shiftMask = (int8_t)more >> 7; //must be arithmetic shift and then sign-extend - q = (q ^ shiftMask) - shiftMask; + uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + + if (!denom->magic) { // shift path + uint64_t mask = (1ULL << shift) - 1; + uint64_t uq = numer + ((numer >> 63) & mask); + int64_t q = (int64_t)uq; + q >>= shift; + // must be arithmetic shift and then sign-extend + int64_t sign = (int8_t)more >> 7; + q = (q ^ sign) - sign; return q; - } - else { - int64_t q = libdivide__mullhi_s64(magic, numer); + } else { + uint64_t uq = (uint64_t)libdivide_mullhi_s64(denom->magic, numer); if (more & LIBDIVIDE_ADD_MARKER) { - int64_t sign = (int8_t)more >> 7; //must be arithmetic shift and then sign extend - q += ((numer ^ sign) - sign); + // must be arithmetic shift and then sign extend + int64_t sign = (int8_t)more >> 7; + // q += (more < 0 ? -numer : numer) + // cast required to avoid UB + uq += ((uint64_t)numer ^ sign) - sign; } - q >>= more & LIBDIVIDE_64_SHIFT_MASK; + int64_t q = (int64_t)uq; + q >>= shift; q += (q < 0); return q; } } - -int libdivide_s64_get_algorithm(const struct libdivide_s64_t *denom) { - uint8_t more = denom->more; - int positiveDivisor = ! (more & LIBDIVIDE_NEGATIVE_DIVISOR); - if (denom->magic == 0) return (positiveDivisor ? 0 : 1); //shift path - else if (more & LIBDIVIDE_ADD_MARKER) return (positiveDivisor ? 2 : 3); - else return 4; -} - -int64_t libdivide_s64_do_alg0(int64_t numer, const struct libdivide_s64_t *denom) { - uint32_t shifter = denom->more & LIBDIVIDE_64_SHIFT_MASK; - int64_t q = numer + ((numer >> 63) & ((1LL << shifter) - 1)); - return q >> shifter; -} - -int64_t libdivide_s64_do_alg1(int64_t numer, const struct libdivide_s64_t *denom) { - //denom->shifter != -1 && demo->shiftMask != 0 - uint32_t shifter = denom->more & LIBDIVIDE_64_SHIFT_MASK; - int64_t q = numer + ((numer >> 63) & ((1LL << shifter) - 1)); - return - (q >> shifter); -} - -int64_t libdivide_s64_do_alg2(int64_t numer, const struct libdivide_s64_t *denom) { - int64_t q = libdivide__mullhi_s64(denom->magic, numer); - q += numer; - q >>= denom->more & LIBDIVIDE_64_SHIFT_MASK; - q += (q < 0); - return q; -} - -int64_t libdivide_s64_do_alg3(int64_t numer, const struct libdivide_s64_t *denom) { - int64_t q = libdivide__mullhi_s64(denom->magic, numer); - q -= numer; - q >>= denom->more & LIBDIVIDE_64_SHIFT_MASK; - q += (q < 0); - return q; -} - -int64_t libdivide_s64_do_alg4(int64_t numer, const struct libdivide_s64_t *denom) { - int64_t q = libdivide__mullhi_s64(denom->magic, numer); - q >>= denom->more; - q += (q < 0); - return q; -} - - -#if LIBDIVIDE_USE_SSE2 -__m128i libdivide_s64_do_vector(__m128i numers, const struct libdivide_s64_t * denom) { +int64_t libdivide_s64_branchfree_do(int64_t numer, const struct libdivide_s64_branchfree_t *denom) { uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + // must be arithmetic shift and then sign extend + int64_t sign = (int8_t)more >> 7; int64_t magic = denom->magic; - if (magic == 0) { //shift path - uint32_t shifter = more & LIBDIVIDE_64_SHIFT_MASK; - __m128i roundToZeroTweak = libdivide__u64_to_m128((1LL << shifter) - 1); - __m128i q = _mm_add_epi64(numers, _mm_and_si128(libdivide_s64_signbits(numers), roundToZeroTweak)); //q = numer + ((numer >> 63) & roundToZeroTweak); - q = libdivide_s64_shift_right_vector(q, shifter); // q = q >> shifter - __m128i shiftMask = _mm_set1_epi32((int32_t)((int8_t)more >> 7)); - q = _mm_sub_epi64(_mm_xor_si128(q, shiftMask), shiftMask); //q = (q ^ shiftMask) - shiftMask; + int64_t q = libdivide_mullhi_s64(magic, numer); + q += numer; + + // If q is non-negative, we have nothing to do. + // If q is negative, we want to add either (2**shift)-1 if d is a power of + // 2, or (2**shift) if it is not a power of 2. + uint64_t is_power_of_2 = (magic == 0); + uint64_t q_sign = (uint64_t)(q >> 63); + q += q_sign & ((1ULL << shift) - is_power_of_2); + + // Arithmetic right shift + q >>= shift; + // Negate if needed + q = (q ^ sign) - sign; + + return q; +} + +int64_t libdivide_s64_recover(const struct libdivide_s64_t *denom) { + uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + if (denom->magic == 0) { // shift path + uint64_t absD = 1ULL << shift; + if (more & LIBDIVIDE_NEGATIVE_DIVISOR) { + absD = -absD; + } + return (int64_t)absD; + } else { + // Unsigned math is much easier + int negative_divisor = (more & LIBDIVIDE_NEGATIVE_DIVISOR); + int magic_was_negated = (more & LIBDIVIDE_ADD_MARKER) + ? denom->magic > 0 : denom->magic < 0; + + uint64_t d = (uint64_t)(magic_was_negated ? -denom->magic : denom->magic); + uint64_t n_hi = 1ULL << shift, n_lo = 0; + uint64_t rem_ignored; + uint64_t q = libdivide_128_div_64_to_64(n_hi, n_lo, d, &rem_ignored); + int64_t result = (int64_t)(q + 1); + if (negative_divisor) { + result = -result; + } + return result; + } +} + +int64_t libdivide_s64_branchfree_recover(const struct libdivide_s64_branchfree_t *denom) { + return libdivide_s64_recover((const struct libdivide_s64_t *)denom); +} + +#if defined(LIBDIVIDE_AVX512) + +static inline __m512i libdivide_u32_do_vector(__m512i numers, const struct libdivide_u32_t *denom); +static inline __m512i libdivide_s32_do_vector(__m512i numers, const struct libdivide_s32_t *denom); +static inline __m512i libdivide_u64_do_vector(__m512i numers, const struct libdivide_u64_t *denom); +static inline __m512i libdivide_s64_do_vector(__m512i numers, const struct libdivide_s64_t *denom); + +static inline __m512i libdivide_u32_branchfree_do_vector(__m512i numers, const struct libdivide_u32_branchfree_t *denom); +static inline __m512i libdivide_s32_branchfree_do_vector(__m512i numers, const struct libdivide_s32_branchfree_t *denom); +static inline __m512i libdivide_u64_branchfree_do_vector(__m512i numers, const struct libdivide_u64_branchfree_t *denom); +static inline __m512i libdivide_s64_branchfree_do_vector(__m512i numers, const struct libdivide_s64_branchfree_t *denom); + +//////// Internal Utility Functions + +static inline __m512i libdivide_s64_signbits(__m512i v) {; + return _mm512_srai_epi64(v, 63); +} + +static inline __m512i libdivide_s64_shift_right_vector(__m512i v, int amt) { + return _mm512_srai_epi64(v, amt); +} + +// Here, b is assumed to contain one 32-bit value repeated. +static inline __m512i libdivide_mullhi_u32_vector(__m512i a, __m512i b) { + __m512i hi_product_0Z2Z = _mm512_srli_epi64(_mm512_mul_epu32(a, b), 32); + __m512i a1X3X = _mm512_srli_epi64(a, 32); + __m512i mask = _mm512_set_epi32(-1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0); + __m512i hi_product_Z1Z3 = _mm512_and_si512(_mm512_mul_epu32(a1X3X, b), mask); + return _mm512_or_si512(hi_product_0Z2Z, hi_product_Z1Z3); +} + +// b is one 32-bit value repeated. +static inline __m512i libdivide_mullhi_s32_vector(__m512i a, __m512i b) { + __m512i hi_product_0Z2Z = _mm512_srli_epi64(_mm512_mul_epi32(a, b), 32); + __m512i a1X3X = _mm512_srli_epi64(a, 32); + __m512i mask = _mm512_set_epi32(-1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0); + __m512i hi_product_Z1Z3 = _mm512_and_si512(_mm512_mul_epi32(a1X3X, b), mask); + return _mm512_or_si512(hi_product_0Z2Z, hi_product_Z1Z3); +} + +// Here, y is assumed to contain one 64-bit value repeated. +// https://stackoverflow.com/a/28827013 +static inline __m512i libdivide_mullhi_u64_vector(__m512i x, __m512i y) { + __m512i lomask = _mm512_set1_epi64(0xffffffff); + __m512i xh = _mm512_shuffle_epi32(x, (_MM_PERM_ENUM) 0xB1); + __m512i yh = _mm512_shuffle_epi32(y, (_MM_PERM_ENUM) 0xB1); + __m512i w0 = _mm512_mul_epu32(x, y); + __m512i w1 = _mm512_mul_epu32(x, yh); + __m512i w2 = _mm512_mul_epu32(xh, y); + __m512i w3 = _mm512_mul_epu32(xh, yh); + __m512i w0h = _mm512_srli_epi64(w0, 32); + __m512i s1 = _mm512_add_epi64(w1, w0h); + __m512i s1l = _mm512_and_si512(s1, lomask); + __m512i s1h = _mm512_srli_epi64(s1, 32); + __m512i s2 = _mm512_add_epi64(w2, s1l); + __m512i s2h = _mm512_srli_epi64(s2, 32); + __m512i hi = _mm512_add_epi64(w3, s1h); + hi = _mm512_add_epi64(hi, s2h); + + return hi; +} + +// y is one 64-bit value repeated. +static inline __m512i libdivide_mullhi_s64_vector(__m512i x, __m512i y) { + __m512i p = libdivide_mullhi_u64_vector(x, y); + __m512i t1 = _mm512_and_si512(libdivide_s64_signbits(x), y); + __m512i t2 = _mm512_and_si512(libdivide_s64_signbits(y), x); + p = _mm512_sub_epi64(p, t1); + p = _mm512_sub_epi64(p, t2); + return p; +} + +////////// UINT32 + +__m512i libdivide_u32_do_vector(__m512i numers, const struct libdivide_u32_t *denom) { + uint8_t more = denom->more; + if (!denom->magic) { + return _mm512_srli_epi32(numers, more); + } + else { + __m512i q = libdivide_mullhi_u32_vector(numers, _mm512_set1_epi32(denom->magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // uint32_t t = ((numer - q) >> 1) + q; + // return t >> denom->shift; + uint32_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + __m512i t = _mm512_add_epi32(_mm512_srli_epi32(_mm512_sub_epi32(numers, q), 1), q); + return _mm512_srli_epi32(t, shift); + } + else { + return _mm512_srli_epi32(q, more); + } + } +} + +__m512i libdivide_u32_branchfree_do_vector(__m512i numers, const struct libdivide_u32_branchfree_t *denom) { + __m512i q = libdivide_mullhi_u32_vector(numers, _mm512_set1_epi32(denom->magic)); + __m512i t = _mm512_add_epi32(_mm512_srli_epi32(_mm512_sub_epi32(numers, q), 1), q); + return _mm512_srli_epi32(t, denom->more); +} + +////////// UINT64 + +__m512i libdivide_u64_do_vector(__m512i numers, const struct libdivide_u64_t *denom) { + uint8_t more = denom->more; + if (!denom->magic) { + return _mm512_srli_epi64(numers, more); + } + else { + __m512i q = libdivide_mullhi_u64_vector(numers, _mm512_set1_epi64(denom->magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // uint32_t t = ((numer - q) >> 1) + q; + // return t >> denom->shift; + uint32_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + __m512i t = _mm512_add_epi64(_mm512_srli_epi64(_mm512_sub_epi64(numers, q), 1), q); + return _mm512_srli_epi64(t, shift); + } + else { + return _mm512_srli_epi64(q, more); + } + } +} + +__m512i libdivide_u64_branchfree_do_vector(__m512i numers, const struct libdivide_u64_branchfree_t *denom) { + __m512i q = libdivide_mullhi_u64_vector(numers, _mm512_set1_epi64(denom->magic)); + __m512i t = _mm512_add_epi64(_mm512_srli_epi64(_mm512_sub_epi64(numers, q), 1), q); + return _mm512_srli_epi64(t, denom->more); +} + +////////// SINT32 + +__m512i libdivide_s32_do_vector(__m512i numers, const struct libdivide_s32_t *denom) { + uint8_t more = denom->more; + if (!denom->magic) { + uint32_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + uint32_t mask = (1U << shift) - 1; + __m512i roundToZeroTweak = _mm512_set1_epi32(mask); + // q = numer + ((numer >> 31) & roundToZeroTweak); + __m512i q = _mm512_add_epi32(numers, _mm512_and_si512(_mm512_srai_epi32(numers, 31), roundToZeroTweak)); + q = _mm512_srai_epi32(q, shift); + __m512i sign = _mm512_set1_epi32((int8_t)more >> 7); + // q = (q ^ sign) - sign; + q = _mm512_sub_epi32(_mm512_xor_si512(q, sign), sign); return q; } else { - __m128i q = libdivide_mullhi_s64_flat_vector(numers, libdivide__u64_to_m128(magic)); + __m512i q = libdivide_mullhi_s32_vector(numers, _mm512_set1_epi32(denom->magic)); if (more & LIBDIVIDE_ADD_MARKER) { - __m128i sign = _mm_set1_epi32((int32_t)((int8_t)more >> 7)); //must be arithmetic shift - q = _mm_add_epi64(q, _mm_sub_epi64(_mm_xor_si128(numers, sign), sign)); // q += ((numer ^ sign) - sign); + // must be arithmetic shift + __m512i sign = _mm512_set1_epi32((int8_t)more >> 7); + // q += ((numer ^ sign) - sign); + q = _mm512_add_epi32(q, _mm512_sub_epi32(_mm512_xor_si512(numers, sign), sign)); } - q = libdivide_s64_shift_right_vector(q, more & LIBDIVIDE_64_SHIFT_MASK); //q >>= denom->mult_path.shift + // q >>= shift + q = _mm512_srai_epi32(q, more & LIBDIVIDE_32_SHIFT_MASK); + q = _mm512_add_epi32(q, _mm512_srli_epi32(q, 31)); // q += (q < 0) + return q; + } +} + +__m512i libdivide_s32_branchfree_do_vector(__m512i numers, const struct libdivide_s32_branchfree_t *denom) { + int32_t magic = denom->magic; + uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + // must be arithmetic shift + __m512i sign = _mm512_set1_epi32((int8_t)more >> 7); + __m512i q = libdivide_mullhi_s32_vector(numers, _mm512_set1_epi32(magic)); + q = _mm512_add_epi32(q, numers); // q += numers + + // If q is non-negative, we have nothing to do + // If q is negative, we want to add either (2**shift)-1 if d is + // a power of 2, or (2**shift) if it is not a power of 2 + uint32_t is_power_of_2 = (magic == 0); + __m512i q_sign = _mm512_srai_epi32(q, 31); // q_sign = q >> 31 + __m512i mask = _mm512_set1_epi32((1U << shift) - is_power_of_2); + q = _mm512_add_epi32(q, _mm512_and_si512(q_sign, mask)); // q = q + (q_sign & mask) + q = _mm512_srai_epi32(q, shift); // q >>= shift + q = _mm512_sub_epi32(_mm512_xor_si512(q, sign), sign); // q = (q ^ sign) - sign + return q; +} + +////////// SINT64 + +__m512i libdivide_s64_do_vector(__m512i numers, const struct libdivide_s64_t *denom) { + uint8_t more = denom->more; + int64_t magic = denom->magic; + if (magic == 0) { // shift path + uint32_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + uint64_t mask = (1ULL << shift) - 1; + __m512i roundToZeroTweak = _mm512_set1_epi64(mask); + // q = numer + ((numer >> 63) & roundToZeroTweak); + __m512i q = _mm512_add_epi64(numers, _mm512_and_si512(libdivide_s64_signbits(numers), roundToZeroTweak)); + q = libdivide_s64_shift_right_vector(q, shift); + __m512i sign = _mm512_set1_epi32((int8_t)more >> 7); + // q = (q ^ sign) - sign; + q = _mm512_sub_epi64(_mm512_xor_si512(q, sign), sign); + return q; + } + else { + __m512i q = libdivide_mullhi_s64_vector(numers, _mm512_set1_epi64(magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // must be arithmetic shift + __m512i sign = _mm512_set1_epi32((int8_t)more >> 7); + // q += ((numer ^ sign) - sign); + q = _mm512_add_epi64(q, _mm512_sub_epi64(_mm512_xor_si512(numers, sign), sign)); + } + // q >>= denom->mult_path.shift + q = libdivide_s64_shift_right_vector(q, more & LIBDIVIDE_64_SHIFT_MASK); + q = _mm512_add_epi64(q, _mm512_srli_epi64(q, 63)); // q += (q < 0) + return q; + } +} + +__m512i libdivide_s64_branchfree_do_vector(__m512i numers, const struct libdivide_s64_branchfree_t *denom) { + int64_t magic = denom->magic; + uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + // must be arithmetic shift + __m512i sign = _mm512_set1_epi32((int8_t)more >> 7); + + // libdivide_mullhi_s64(numers, magic); + __m512i q = libdivide_mullhi_s64_vector(numers, _mm512_set1_epi64(magic)); + q = _mm512_add_epi64(q, numers); // q += numers + + // If q is non-negative, we have nothing to do. + // If q is negative, we want to add either (2**shift)-1 if d is + // a power of 2, or (2**shift) if it is not a power of 2. + uint32_t is_power_of_2 = (magic == 0); + __m512i q_sign = libdivide_s64_signbits(q); // q_sign = q >> 63 + __m512i mask = _mm512_set1_epi64((1ULL << shift) - is_power_of_2); + q = _mm512_add_epi64(q, _mm512_and_si512(q_sign, mask)); // q = q + (q_sign & mask) + q = libdivide_s64_shift_right_vector(q, shift); // q >>= shift + q = _mm512_sub_epi64(_mm512_xor_si512(q, sign), sign); // q = (q ^ sign) - sign + return q; +} + +#elif defined(LIBDIVIDE_AVX2) + +static inline __m256i libdivide_u32_do_vector(__m256i numers, const struct libdivide_u32_t *denom); +static inline __m256i libdivide_s32_do_vector(__m256i numers, const struct libdivide_s32_t *denom); +static inline __m256i libdivide_u64_do_vector(__m256i numers, const struct libdivide_u64_t *denom); +static inline __m256i libdivide_s64_do_vector(__m256i numers, const struct libdivide_s64_t *denom); + +static inline __m256i libdivide_u32_branchfree_do_vector(__m256i numers, const struct libdivide_u32_branchfree_t *denom); +static inline __m256i libdivide_s32_branchfree_do_vector(__m256i numers, const struct libdivide_s32_branchfree_t *denom); +static inline __m256i libdivide_u64_branchfree_do_vector(__m256i numers, const struct libdivide_u64_branchfree_t *denom); +static inline __m256i libdivide_s64_branchfree_do_vector(__m256i numers, const struct libdivide_s64_branchfree_t *denom); + +//////// Internal Utility Functions + +// Implementation of _mm256_srai_epi64(v, 63) (from AVX512). +static inline __m256i libdivide_s64_signbits(__m256i v) { + __m256i hiBitsDuped = _mm256_shuffle_epi32(v, _MM_SHUFFLE(3, 3, 1, 1)); + __m256i signBits = _mm256_srai_epi32(hiBitsDuped, 31); + return signBits; +} + +// Implementation of _mm256_srai_epi64 (from AVX512). +static inline __m256i libdivide_s64_shift_right_vector(__m256i v, int amt) { + const int b = 64 - amt; + __m256i m = _mm256_set1_epi64x(1ULL << (b - 1)); + __m256i x = _mm256_srli_epi64(v, amt); + __m256i result = _mm256_sub_epi64(_mm256_xor_si256(x, m), m); + return result; +} + +// Here, b is assumed to contain one 32-bit value repeated. +static inline __m256i libdivide_mullhi_u32_vector(__m256i a, __m256i b) { + __m256i hi_product_0Z2Z = _mm256_srli_epi64(_mm256_mul_epu32(a, b), 32); + __m256i a1X3X = _mm256_srli_epi64(a, 32); + __m256i mask = _mm256_set_epi32(-1, 0, -1, 0, -1, 0, -1, 0); + __m256i hi_product_Z1Z3 = _mm256_and_si256(_mm256_mul_epu32(a1X3X, b), mask); + return _mm256_or_si256(hi_product_0Z2Z, hi_product_Z1Z3); +} + +// b is one 32-bit value repeated. +static inline __m256i libdivide_mullhi_s32_vector(__m256i a, __m256i b) { + __m256i hi_product_0Z2Z = _mm256_srli_epi64(_mm256_mul_epi32(a, b), 32); + __m256i a1X3X = _mm256_srli_epi64(a, 32); + __m256i mask = _mm256_set_epi32(-1, 0, -1, 0, -1, 0, -1, 0); + __m256i hi_product_Z1Z3 = _mm256_and_si256(_mm256_mul_epi32(a1X3X, b), mask); + return _mm256_or_si256(hi_product_0Z2Z, hi_product_Z1Z3); +} + +// Here, y is assumed to contain one 64-bit value repeated. +// https://stackoverflow.com/a/28827013 +static inline __m256i libdivide_mullhi_u64_vector(__m256i x, __m256i y) { + __m256i lomask = _mm256_set1_epi64x(0xffffffff); + __m256i xh = _mm256_shuffle_epi32(x, 0xB1); // x0l, x0h, x1l, x1h + __m256i yh = _mm256_shuffle_epi32(y, 0xB1); // y0l, y0h, y1l, y1h + __m256i w0 = _mm256_mul_epu32(x, y); // x0l*y0l, x1l*y1l + __m256i w1 = _mm256_mul_epu32(x, yh); // x0l*y0h, x1l*y1h + __m256i w2 = _mm256_mul_epu32(xh, y); // x0h*y0l, x1h*y0l + __m256i w3 = _mm256_mul_epu32(xh, yh); // x0h*y0h, x1h*y1h + __m256i w0h = _mm256_srli_epi64(w0, 32); + __m256i s1 = _mm256_add_epi64(w1, w0h); + __m256i s1l = _mm256_and_si256(s1, lomask); + __m256i s1h = _mm256_srli_epi64(s1, 32); + __m256i s2 = _mm256_add_epi64(w2, s1l); + __m256i s2h = _mm256_srli_epi64(s2, 32); + __m256i hi = _mm256_add_epi64(w3, s1h); + hi = _mm256_add_epi64(hi, s2h); + + return hi; +} + +// y is one 64-bit value repeated. +static inline __m256i libdivide_mullhi_s64_vector(__m256i x, __m256i y) { + __m256i p = libdivide_mullhi_u64_vector(x, y); + __m256i t1 = _mm256_and_si256(libdivide_s64_signbits(x), y); + __m256i t2 = _mm256_and_si256(libdivide_s64_signbits(y), x); + p = _mm256_sub_epi64(p, t1); + p = _mm256_sub_epi64(p, t2); + return p; +} + +////////// UINT32 + +__m256i libdivide_u32_do_vector(__m256i numers, const struct libdivide_u32_t *denom) { + uint8_t more = denom->more; + if (!denom->magic) { + return _mm256_srli_epi32(numers, more); + } + else { + __m256i q = libdivide_mullhi_u32_vector(numers, _mm256_set1_epi32(denom->magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // uint32_t t = ((numer - q) >> 1) + q; + // return t >> denom->shift; + uint32_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + __m256i t = _mm256_add_epi32(_mm256_srli_epi32(_mm256_sub_epi32(numers, q), 1), q); + return _mm256_srli_epi32(t, shift); + } + else { + return _mm256_srli_epi32(q, more); + } + } +} + +__m256i libdivide_u32_branchfree_do_vector(__m256i numers, const struct libdivide_u32_branchfree_t *denom) { + __m256i q = libdivide_mullhi_u32_vector(numers, _mm256_set1_epi32(denom->magic)); + __m256i t = _mm256_add_epi32(_mm256_srli_epi32(_mm256_sub_epi32(numers, q), 1), q); + return _mm256_srli_epi32(t, denom->more); +} + +////////// UINT64 + +__m256i libdivide_u64_do_vector(__m256i numers, const struct libdivide_u64_t *denom) { + uint8_t more = denom->more; + if (!denom->magic) { + return _mm256_srli_epi64(numers, more); + } + else { + __m256i q = libdivide_mullhi_u64_vector(numers, _mm256_set1_epi64x(denom->magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // uint32_t t = ((numer - q) >> 1) + q; + // return t >> denom->shift; + uint32_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + __m256i t = _mm256_add_epi64(_mm256_srli_epi64(_mm256_sub_epi64(numers, q), 1), q); + return _mm256_srli_epi64(t, shift); + } + else { + return _mm256_srli_epi64(q, more); + } + } +} + +__m256i libdivide_u64_branchfree_do_vector(__m256i numers, const struct libdivide_u64_branchfree_t *denom) { + __m256i q = libdivide_mullhi_u64_vector(numers, _mm256_set1_epi64x(denom->magic)); + __m256i t = _mm256_add_epi64(_mm256_srli_epi64(_mm256_sub_epi64(numers, q), 1), q); + return _mm256_srli_epi64(t, denom->more); +} + +////////// SINT32 + +__m256i libdivide_s32_do_vector(__m256i numers, const struct libdivide_s32_t *denom) { + uint8_t more = denom->more; + if (!denom->magic) { + uint32_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + uint32_t mask = (1U << shift) - 1; + __m256i roundToZeroTweak = _mm256_set1_epi32(mask); + // q = numer + ((numer >> 31) & roundToZeroTweak); + __m256i q = _mm256_add_epi32(numers, _mm256_and_si256(_mm256_srai_epi32(numers, 31), roundToZeroTweak)); + q = _mm256_srai_epi32(q, shift); + __m256i sign = _mm256_set1_epi32((int8_t)more >> 7); + // q = (q ^ sign) - sign; + q = _mm256_sub_epi32(_mm256_xor_si256(q, sign), sign); + return q; + } + else { + __m256i q = libdivide_mullhi_s32_vector(numers, _mm256_set1_epi32(denom->magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // must be arithmetic shift + __m256i sign = _mm256_set1_epi32((int8_t)more >> 7); + // q += ((numer ^ sign) - sign); + q = _mm256_add_epi32(q, _mm256_sub_epi32(_mm256_xor_si256(numers, sign), sign)); + } + // q >>= shift + q = _mm256_srai_epi32(q, more & LIBDIVIDE_32_SHIFT_MASK); + q = _mm256_add_epi32(q, _mm256_srli_epi32(q, 31)); // q += (q < 0) + return q; + } +} + +__m256i libdivide_s32_branchfree_do_vector(__m256i numers, const struct libdivide_s32_branchfree_t *denom) { + int32_t magic = denom->magic; + uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + // must be arithmetic shift + __m256i sign = _mm256_set1_epi32((int8_t)more >> 7); + __m256i q = libdivide_mullhi_s32_vector(numers, _mm256_set1_epi32(magic)); + q = _mm256_add_epi32(q, numers); // q += numers + + // If q is non-negative, we have nothing to do + // If q is negative, we want to add either (2**shift)-1 if d is + // a power of 2, or (2**shift) if it is not a power of 2 + uint32_t is_power_of_2 = (magic == 0); + __m256i q_sign = _mm256_srai_epi32(q, 31); // q_sign = q >> 31 + __m256i mask = _mm256_set1_epi32((1U << shift) - is_power_of_2); + q = _mm256_add_epi32(q, _mm256_and_si256(q_sign, mask)); // q = q + (q_sign & mask) + q = _mm256_srai_epi32(q, shift); // q >>= shift + q = _mm256_sub_epi32(_mm256_xor_si256(q, sign), sign); // q = (q ^ sign) - sign + return q; +} + +////////// SINT64 + +__m256i libdivide_s64_do_vector(__m256i numers, const struct libdivide_s64_t *denom) { + uint8_t more = denom->more; + int64_t magic = denom->magic; + if (magic == 0) { // shift path + uint32_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + uint64_t mask = (1ULL << shift) - 1; + __m256i roundToZeroTweak = _mm256_set1_epi64x(mask); + // q = numer + ((numer >> 63) & roundToZeroTweak); + __m256i q = _mm256_add_epi64(numers, _mm256_and_si256(libdivide_s64_signbits(numers), roundToZeroTweak)); + q = libdivide_s64_shift_right_vector(q, shift); + __m256i sign = _mm256_set1_epi32((int8_t)more >> 7); + // q = (q ^ sign) - sign; + q = _mm256_sub_epi64(_mm256_xor_si256(q, sign), sign); + return q; + } + else { + __m256i q = libdivide_mullhi_s64_vector(numers, _mm256_set1_epi64x(magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // must be arithmetic shift + __m256i sign = _mm256_set1_epi32((int8_t)more >> 7); + // q += ((numer ^ sign) - sign); + q = _mm256_add_epi64(q, _mm256_sub_epi64(_mm256_xor_si256(numers, sign), sign)); + } + // q >>= denom->mult_path.shift + q = libdivide_s64_shift_right_vector(q, more & LIBDIVIDE_64_SHIFT_MASK); + q = _mm256_add_epi64(q, _mm256_srli_epi64(q, 63)); // q += (q < 0) + return q; + } +} + +__m256i libdivide_s64_branchfree_do_vector(__m256i numers, const struct libdivide_s64_branchfree_t *denom) { + int64_t magic = denom->magic; + uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + // must be arithmetic shift + __m256i sign = _mm256_set1_epi32((int8_t)more >> 7); + + // libdivide_mullhi_s64(numers, magic); + __m256i q = libdivide_mullhi_s64_vector(numers, _mm256_set1_epi64x(magic)); + q = _mm256_add_epi64(q, numers); // q += numers + + // If q is non-negative, we have nothing to do. + // If q is negative, we want to add either (2**shift)-1 if d is + // a power of 2, or (2**shift) if it is not a power of 2. + uint32_t is_power_of_2 = (magic == 0); + __m256i q_sign = libdivide_s64_signbits(q); // q_sign = q >> 63 + __m256i mask = _mm256_set1_epi64x((1ULL << shift) - is_power_of_2); + q = _mm256_add_epi64(q, _mm256_and_si256(q_sign, mask)); // q = q + (q_sign & mask) + q = libdivide_s64_shift_right_vector(q, shift); // q >>= shift + q = _mm256_sub_epi64(_mm256_xor_si256(q, sign), sign); // q = (q ^ sign) - sign + return q; +} + +#elif defined(LIBDIVIDE_SSE2) + +static inline __m128i libdivide_u32_do_vector(__m128i numers, const struct libdivide_u32_t *denom); +static inline __m128i libdivide_s32_do_vector(__m128i numers, const struct libdivide_s32_t *denom); +static inline __m128i libdivide_u64_do_vector(__m128i numers, const struct libdivide_u64_t *denom); +static inline __m128i libdivide_s64_do_vector(__m128i numers, const struct libdivide_s64_t *denom); + +static inline __m128i libdivide_u32_branchfree_do_vector(__m128i numers, const struct libdivide_u32_branchfree_t *denom); +static inline __m128i libdivide_s32_branchfree_do_vector(__m128i numers, const struct libdivide_s32_branchfree_t *denom); +static inline __m128i libdivide_u64_branchfree_do_vector(__m128i numers, const struct libdivide_u64_branchfree_t *denom); +static inline __m128i libdivide_s64_branchfree_do_vector(__m128i numers, const struct libdivide_s64_branchfree_t *denom); + +//////// Internal Utility Functions + +// Implementation of _mm_srai_epi64(v, 63) (from AVX512). +static inline __m128i libdivide_s64_signbits(__m128i v) { + __m128i hiBitsDuped = _mm_shuffle_epi32(v, _MM_SHUFFLE(3, 3, 1, 1)); + __m128i signBits = _mm_srai_epi32(hiBitsDuped, 31); + return signBits; +} + +// Implementation of _mm_srai_epi64 (from AVX512). +static inline __m128i libdivide_s64_shift_right_vector(__m128i v, int amt) { + const int b = 64 - amt; + __m128i m = _mm_set1_epi64x(1ULL << (b - 1)); + __m128i x = _mm_srli_epi64(v, amt); + __m128i result = _mm_sub_epi64(_mm_xor_si128(x, m), m); + return result; +} + +// Here, b is assumed to contain one 32-bit value repeated. +static inline __m128i libdivide_mullhi_u32_vector(__m128i a, __m128i b) { + __m128i hi_product_0Z2Z = _mm_srli_epi64(_mm_mul_epu32(a, b), 32); + __m128i a1X3X = _mm_srli_epi64(a, 32); + __m128i mask = _mm_set_epi32(-1, 0, -1, 0); + __m128i hi_product_Z1Z3 = _mm_and_si128(_mm_mul_epu32(a1X3X, b), mask); + return _mm_or_si128(hi_product_0Z2Z, hi_product_Z1Z3); +} + +// SSE2 does not have a signed multiplication instruction, but we can convert +// unsigned to signed pretty efficiently. Again, b is just a 32 bit value +// repeated four times. +static inline __m128i libdivide_mullhi_s32_vector(__m128i a, __m128i b) { + __m128i p = libdivide_mullhi_u32_vector(a, b); + // t1 = (a >> 31) & y, arithmetic shift + __m128i t1 = _mm_and_si128(_mm_srai_epi32(a, 31), b); + __m128i t2 = _mm_and_si128(_mm_srai_epi32(b, 31), a); + p = _mm_sub_epi32(p, t1); + p = _mm_sub_epi32(p, t2); + return p; +} + +// Here, y is assumed to contain one 64-bit value repeated. +// https://stackoverflow.com/a/28827013 +static inline __m128i libdivide_mullhi_u64_vector(__m128i x, __m128i y) { + __m128i lomask = _mm_set1_epi64x(0xffffffff); + __m128i xh = _mm_shuffle_epi32(x, 0xB1); // x0l, x0h, x1l, x1h + __m128i yh = _mm_shuffle_epi32(y, 0xB1); // y0l, y0h, y1l, y1h + __m128i w0 = _mm_mul_epu32(x, y); // x0l*y0l, x1l*y1l + __m128i w1 = _mm_mul_epu32(x, yh); // x0l*y0h, x1l*y1h + __m128i w2 = _mm_mul_epu32(xh, y); // x0h*y0l, x1h*y0l + __m128i w3 = _mm_mul_epu32(xh, yh); // x0h*y0h, x1h*y1h + __m128i w0h = _mm_srli_epi64(w0, 32); + __m128i s1 = _mm_add_epi64(w1, w0h); + __m128i s1l = _mm_and_si128(s1, lomask); + __m128i s1h = _mm_srli_epi64(s1, 32); + __m128i s2 = _mm_add_epi64(w2, s1l); + __m128i s2h = _mm_srli_epi64(s2, 32); + __m128i hi = _mm_add_epi64(w3, s1h); + hi = _mm_add_epi64(hi, s2h); + + return hi; +} + +// y is one 64-bit value repeated. +static inline __m128i libdivide_mullhi_s64_vector(__m128i x, __m128i y) { + __m128i p = libdivide_mullhi_u64_vector(x, y); + __m128i t1 = _mm_and_si128(libdivide_s64_signbits(x), y); + __m128i t2 = _mm_and_si128(libdivide_s64_signbits(y), x); + p = _mm_sub_epi64(p, t1); + p = _mm_sub_epi64(p, t2); + return p; +} + +////////// UINT32 + +__m128i libdivide_u32_do_vector(__m128i numers, const struct libdivide_u32_t *denom) { + uint8_t more = denom->more; + if (!denom->magic) { + return _mm_srli_epi32(numers, more); + } + else { + __m128i q = libdivide_mullhi_u32_vector(numers, _mm_set1_epi32(denom->magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // uint32_t t = ((numer - q) >> 1) + q; + // return t >> denom->shift; + uint32_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + __m128i t = _mm_add_epi32(_mm_srli_epi32(_mm_sub_epi32(numers, q), 1), q); + return _mm_srli_epi32(t, shift); + } + else { + return _mm_srli_epi32(q, more); + } + } +} + +__m128i libdivide_u32_branchfree_do_vector(__m128i numers, const struct libdivide_u32_branchfree_t *denom) { + __m128i q = libdivide_mullhi_u32_vector(numers, _mm_set1_epi32(denom->magic)); + __m128i t = _mm_add_epi32(_mm_srli_epi32(_mm_sub_epi32(numers, q), 1), q); + return _mm_srli_epi32(t, denom->more); +} + +////////// UINT64 + +__m128i libdivide_u64_do_vector(__m128i numers, const struct libdivide_u64_t *denom) { + uint8_t more = denom->more; + if (!denom->magic) { + return _mm_srli_epi64(numers, more); + } + else { + __m128i q = libdivide_mullhi_u64_vector(numers, _mm_set1_epi64x(denom->magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // uint32_t t = ((numer - q) >> 1) + q; + // return t >> denom->shift; + uint32_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + __m128i t = _mm_add_epi64(_mm_srli_epi64(_mm_sub_epi64(numers, q), 1), q); + return _mm_srli_epi64(t, shift); + } + else { + return _mm_srli_epi64(q, more); + } + } +} + +__m128i libdivide_u64_branchfree_do_vector(__m128i numers, const struct libdivide_u64_branchfree_t *denom) { + __m128i q = libdivide_mullhi_u64_vector(numers, _mm_set1_epi64x(denom->magic)); + __m128i t = _mm_add_epi64(_mm_srli_epi64(_mm_sub_epi64(numers, q), 1), q); + return _mm_srli_epi64(t, denom->more); +} + +////////// SINT32 + +__m128i libdivide_s32_do_vector(__m128i numers, const struct libdivide_s32_t *denom) { + uint8_t more = denom->more; + if (!denom->magic) { + uint32_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + uint32_t mask = (1U << shift) - 1; + __m128i roundToZeroTweak = _mm_set1_epi32(mask); + // q = numer + ((numer >> 31) & roundToZeroTweak); + __m128i q = _mm_add_epi32(numers, _mm_and_si128(_mm_srai_epi32(numers, 31), roundToZeroTweak)); + q = _mm_srai_epi32(q, shift); + __m128i sign = _mm_set1_epi32((int8_t)more >> 7); + // q = (q ^ sign) - sign; + q = _mm_sub_epi32(_mm_xor_si128(q, sign), sign); + return q; + } + else { + __m128i q = libdivide_mullhi_s32_vector(numers, _mm_set1_epi32(denom->magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // must be arithmetic shift + __m128i sign = _mm_set1_epi32((int8_t)more >> 7); + // q += ((numer ^ sign) - sign); + q = _mm_add_epi32(q, _mm_sub_epi32(_mm_xor_si128(numers, sign), sign)); + } + // q >>= shift + q = _mm_srai_epi32(q, more & LIBDIVIDE_32_SHIFT_MASK); + q = _mm_add_epi32(q, _mm_srli_epi32(q, 31)); // q += (q < 0) + return q; + } +} + +__m128i libdivide_s32_branchfree_do_vector(__m128i numers, const struct libdivide_s32_branchfree_t *denom) { + int32_t magic = denom->magic; + uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK; + // must be arithmetic shift + __m128i sign = _mm_set1_epi32((int8_t)more >> 7); + __m128i q = libdivide_mullhi_s32_vector(numers, _mm_set1_epi32(magic)); + q = _mm_add_epi32(q, numers); // q += numers + + // If q is non-negative, we have nothing to do + // If q is negative, we want to add either (2**shift)-1 if d is + // a power of 2, or (2**shift) if it is not a power of 2 + uint32_t is_power_of_2 = (magic == 0); + __m128i q_sign = _mm_srai_epi32(q, 31); // q_sign = q >> 31 + __m128i mask = _mm_set1_epi32((1U << shift) - is_power_of_2); + q = _mm_add_epi32(q, _mm_and_si128(q_sign, mask)); // q = q + (q_sign & mask) + q = _mm_srai_epi32(q, shift); // q >>= shift + q = _mm_sub_epi32(_mm_xor_si128(q, sign), sign); // q = (q ^ sign) - sign + return q; +} + +////////// SINT64 + +__m128i libdivide_s64_do_vector(__m128i numers, const struct libdivide_s64_t *denom) { + uint8_t more = denom->more; + int64_t magic = denom->magic; + if (magic == 0) { // shift path + uint32_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + uint64_t mask = (1ULL << shift) - 1; + __m128i roundToZeroTweak = _mm_set1_epi64x(mask); + // q = numer + ((numer >> 63) & roundToZeroTweak); + __m128i q = _mm_add_epi64(numers, _mm_and_si128(libdivide_s64_signbits(numers), roundToZeroTweak)); + q = libdivide_s64_shift_right_vector(q, shift); + __m128i sign = _mm_set1_epi32((int8_t)more >> 7); + // q = (q ^ sign) - sign; + q = _mm_sub_epi64(_mm_xor_si128(q, sign), sign); + return q; + } + else { + __m128i q = libdivide_mullhi_s64_vector(numers, _mm_set1_epi64x(magic)); + if (more & LIBDIVIDE_ADD_MARKER) { + // must be arithmetic shift + __m128i sign = _mm_set1_epi32((int8_t)more >> 7); + // q += ((numer ^ sign) - sign); + q = _mm_add_epi64(q, _mm_sub_epi64(_mm_xor_si128(numers, sign), sign)); + } + // q >>= denom->mult_path.shift + q = libdivide_s64_shift_right_vector(q, more & LIBDIVIDE_64_SHIFT_MASK); q = _mm_add_epi64(q, _mm_srli_epi64(q, 63)); // q += (q < 0) return q; } } -__m128i libdivide_s64_do_vector_alg0(__m128i numers, const struct libdivide_s64_t *denom) { - uint32_t shifter = denom->more & LIBDIVIDE_64_SHIFT_MASK; - __m128i roundToZeroTweak = libdivide__u64_to_m128((1LL << shifter) - 1); - __m128i q = _mm_add_epi64(numers, _mm_and_si128(libdivide_s64_signbits(numers), roundToZeroTweak)); - q = libdivide_s64_shift_right_vector(q, shifter); - return q; -} +__m128i libdivide_s64_branchfree_do_vector(__m128i numers, const struct libdivide_s64_branchfree_t *denom) { + int64_t magic = denom->magic; + uint8_t more = denom->more; + uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK; + // must be arithmetic shift + __m128i sign = _mm_set1_epi32((int8_t)more >> 7); -__m128i libdivide_s64_do_vector_alg1(__m128i numers, const struct libdivide_s64_t *denom) { - uint32_t shifter = denom->more & LIBDIVIDE_64_SHIFT_MASK; - __m128i roundToZeroTweak = libdivide__u64_to_m128((1LL << shifter) - 1); - __m128i q = _mm_add_epi64(numers, _mm_and_si128(libdivide_s64_signbits(numers), roundToZeroTweak)); - q = libdivide_s64_shift_right_vector(q, shifter); - return _mm_sub_epi64(_mm_setzero_si128(), q); -} + // libdivide_mullhi_s64(numers, magic); + __m128i q = libdivide_mullhi_s64_vector(numers, _mm_set1_epi64x(magic)); + q = _mm_add_epi64(q, numers); // q += numers -__m128i libdivide_s64_do_vector_alg2(__m128i numers, const struct libdivide_s64_t *denom) { - __m128i q = libdivide_mullhi_s64_flat_vector(numers, libdivide__u64_to_m128(denom->magic)); - q = _mm_add_epi64(q, numers); - q = libdivide_s64_shift_right_vector(q, denom->more & LIBDIVIDE_64_SHIFT_MASK); - q = _mm_add_epi64(q, _mm_srli_epi64(q, 63)); // q += (q < 0) - return q; -} - -__m128i libdivide_s64_do_vector_alg3(__m128i numers, const struct libdivide_s64_t *denom) { - __m128i q = libdivide_mullhi_s64_flat_vector(numers, libdivide__u64_to_m128(denom->magic)); - q = _mm_sub_epi64(q, numers); - q = libdivide_s64_shift_right_vector(q, denom->more & LIBDIVIDE_64_SHIFT_MASK); - q = _mm_add_epi64(q, _mm_srli_epi64(q, 63)); // q += (q < 0) - return q; -} - -__m128i libdivide_s64_do_vector_alg4(__m128i numers, const struct libdivide_s64_t *denom) { - __m128i q = libdivide_mullhi_s64_flat_vector(numers, libdivide__u64_to_m128(denom->magic)); - q = libdivide_s64_shift_right_vector(q, denom->more); - q = _mm_add_epi64(q, _mm_srli_epi64(q, 63)); + // If q is non-negative, we have nothing to do. + // If q is negative, we want to add either (2**shift)-1 if d is + // a power of 2, or (2**shift) if it is not a power of 2. + uint32_t is_power_of_2 = (magic == 0); + __m128i q_sign = libdivide_s64_signbits(q); // q_sign = q >> 63 + __m128i mask = _mm_set1_epi64x((1ULL << shift) - is_power_of_2); + q = _mm_add_epi64(q, _mm_and_si128(q_sign, mask)); // q = q + (q_sign & mask) + q = libdivide_s64_shift_right_vector(q, shift); // q >>= shift + q = _mm_sub_epi64(_mm_xor_si128(q, sign), sign); // q = (q ^ sign) - sign return q; } @@ -1204,228 +1930,143 @@ __m128i libdivide_s64_do_vector_alg4(__m128i numers, const struct libdivide_s64_ #ifdef __cplusplus -/* The C++ template design here is a total mess. This needs to be fixed by someone better at templates than I. The current design is: - -- The base is a template divider_base that takes the integer type, the libdivide struct, a generating function, a get algorithm function, a do function, and either a do vector function or a dummy int. -- The base has storage for the libdivide struct. This is the only storage (so the C++ class should be no larger than the libdivide struct). - -- Above that, there's divider_mid. This is an empty struct by default, but it is specialized against our four int types. divider_mid contains a template struct algo, that contains a typedef for a specialization of divider_base. struct algo is specialized to take an "algorithm number," where -1 means to use the general algorithm. - -- Publicly we have class divider, which inherits from divider_mid::algo. This also take an algorithm number, which defaults to -1 (the general algorithm). -- divider has a operator / which allows you to use a divider as the divisor in a quotient expression. - -*/ - -namespace libdivide_internal { - -#if LIBDIVIDE_USE_SSE2 -#define MAYBE_VECTOR(x) x -#define MAYBE_VECTOR_PARAM __m128i vector_func(__m128i, const DenomType *) -#else -#define MAYBE_VECTOR(x) 0 -#define MAYBE_VECTOR_PARAM int vector_func -#endif - - /* Some bogus unswitch functions for unsigned types so the same (presumably templated) code can work for both signed and unsigned. */ - uint32_t crash_u32(uint32_t, const libdivide_u32_t *) { abort(); } - uint64_t crash_u64(uint64_t, const libdivide_u64_t *) { abort(); } -#ifdef __APPLE__ - UInt64 crash_u64(UInt64, const libdivide_u64_t *) { abort(); } -#endif -#if LIBDIVIDE_USE_SSE2 - __m128i crash_u32_vector(__m128i, const libdivide_u32_t *) { abort(); } - __m128i crash_u64_vector(__m128i, const libdivide_u64_t *) { abort(); } -#endif - - template - class divider_base { - public: - DenomType denom; - divider_base(IntType d) : denom(gen_func(d)) { } - divider_base(const DenomType & d) : denom(d) { } - - IntType perform_divide(IntType val) const { return do_func(val, &denom); } -#if LIBDIVIDE_USE_SSE2 - __m128i perform_divide_vector(__m128i val) const { return vector_func(val, &denom); } -#endif - - int get_algorithm() const { return get_algo(&denom); } - }; - - - template struct divider_mid { }; - - template<> struct divider_mid { - typedef uint32_t IntType; - typedef struct libdivide_u32_t DenomType; - template struct denom { - typedef divider_base divider; - }; - - template struct algo { }; - template struct algo<-1, J> { typedef denom::divider divider; }; - template struct algo<0, J> { typedef denom::divider divider; }; - template struct algo<1, J> { typedef denom::divider divider; }; - template struct algo<2, J> { typedef denom::divider divider; }; - - /* Define two more bogus ones so that the same (templated, presumably) code can handle both signed and unsigned */ - template struct algo<3, J> { typedef denom::divider divider; }; - template struct algo<4, J> { typedef denom::divider divider; }; - - }; - - template<> struct divider_mid { - typedef int32_t IntType; - typedef struct libdivide_s32_t DenomType; - template struct denom { - typedef divider_base divider; - }; - - - template struct algo { }; - template struct algo<-1, J> { typedef denom::divider divider; }; - template struct algo<0, J> { typedef denom::divider divider; }; - template struct algo<1, J> { typedef denom::divider divider; }; - template struct algo<2, J> { typedef denom::divider divider; }; - template struct algo<3, J> { typedef denom::divider divider; }; - template struct algo<4, J> { typedef denom::divider divider; }; - - }; - -#ifdef __APPLE__ - template<> struct divider_mid { - typedef Int64 IntType; - typedef struct libdivide_s64_t DenomType; - template struct denom { - typedef divider_base divider; - }; - - template struct algo { }; - template struct algo<-1, J> { typedef denom::divider divider; }; - template struct algo<0, J> { typedef denom::divider divider; }; - template struct algo<1, J> { typedef denom::divider divider; }; - template struct algo<2, J> { typedef denom::divider divider; }; - template struct algo<3, J> { typedef denom::divider divider; }; - template struct algo<4, J> { typedef denom::divider divider; }; - }; - - template<> struct divider_mid { - typedef UInt64 IntType; - typedef struct libdivide_u64_t DenomType; - template struct denom { - typedef divider_base divider; - }; - - template struct algo { }; - template struct algo<-1, J> { typedef denom::divider divider; }; - template struct algo<0, J> { typedef denom::divider divider; }; - template struct algo<1, J> { typedef denom::divider divider; }; - template struct algo<2, J> { typedef denom::divider divider; }; - - /* Define two more bogus ones so that the same (templated, presumably) code can handle both signed and unsigned */ - template struct algo<3, J> { typedef denom::divider divider; }; - template struct algo<4, J> { typedef denom::divider divider; }; - - - }; -#endif - - template<> struct divider_mid { - typedef uint64_t IntType; - typedef struct libdivide_u64_t DenomType; - template struct denom { - typedef divider_base divider; - }; - - template struct algo { }; - template struct algo<-1, J> { typedef denom::divider divider; }; - template struct algo<0, J> { typedef denom::divider divider; }; - template struct algo<1, J> { typedef denom::divider divider; }; - template struct algo<2, J> { typedef denom::divider divider; }; - - /* Define two more bogus ones so that the same (templated, presumably) code can handle both signed and unsigned */ - template struct algo<3, J> { typedef denom::divider divider; }; - template struct algo<4, J> { typedef denom::divider divider; }; - - - }; - - template<> struct divider_mid { - typedef int64_t IntType; - typedef struct libdivide_s64_t DenomType; - template struct denom { - typedef divider_base divider; - }; - - template struct algo { }; - template struct algo<-1, J> { typedef denom::divider divider; }; - template struct algo<0, J> { typedef denom::divider divider; }; - template struct algo<1, J> { typedef denom::divider divider; }; - template struct algo<2, J> { typedef denom::divider divider; }; - template struct algo<3, J> { typedef denom::divider divider; }; - template struct algo<4, J> { typedef denom::divider divider; }; - }; - -} - -template -class divider -{ - private: - typename libdivide_internal::divider_mid::template algo::divider sub; - template friend divider unswitch(const divider & d); - divider(const typename libdivide_internal::divider_mid::DenomType & denom) : sub(denom) { } - - public: - - /* Ordinary constructor, that takes the divisor as a parameter. */ - divider(T n) : sub(n) { } - - /* Default constructor, that divides by 1 */ - divider() : sub(1) { } - - /* Divides the parameter by the divisor, returning the quotient */ - T perform_divide(T val) const { return sub.perform_divide(val); } - -#if LIBDIVIDE_USE_SSE2 - /* Treats the vector as either two or four packed values (depending on the size), and divides each of them by the divisor, returning the packed quotients. */ - __m128i perform_divide_vector(__m128i val) const { return sub.perform_divide_vector(val); } -#endif - - /* Returns the index of algorithm, for use in the unswitch function */ - int get_algorithm() const { return sub.get_algorithm(); } // returns the algorithm for unswitching - - /* operator== */ - bool operator==(const divider & him) const { return sub.denom.magic == him.sub.denom.magic && sub.denom.more == him.sub.denom.more; } - - bool operator!=(const divider & him) const { return ! (*this == him); } +// The C++ divider class is templated on both an integer type +// (like uint64_t) and an algorithm type. +// * BRANCHFULL is the default algorithm type. +// * BRANCHFREE is the branchfree algorithm type. +enum { + BRANCHFULL, + BRANCHFREE }; -/* Returns a divider specialized for the given algorithm. */ -template -divider unswitch(const divider & d) { return divider(d.sub.denom); } - -/* Overload of the / operator for scalar division. */ -template -int_type operator/(int_type numer, const divider & denom) { - return denom.perform_divide(numer); -} - -#if LIBDIVIDE_USE_SSE2 -/* Overload of the / operator for vector division. */ -template -__m128i operator/(__m128i numer, const divider & denom) { - return denom.perform_divide_vector(numer); -} +#if defined(LIBDIVIDE_AVX512) + #define LIBDIVIDE_VECTOR_TYPE __m512i +#elif defined(LIBDIVIDE_AVX2) + #define LIBDIVIDE_VECTOR_TYPE __m256i +#elif defined(LIBDIVIDE_SSE2) + #define LIBDIVIDE_VECTOR_TYPE __m128i #endif - -#endif //__cplusplus - -#endif //LIBDIVIDE_HEADER_ONLY -#ifdef __cplusplus -} //close namespace libdivide -} //close anonymous namespace +#if !defined(LIBDIVIDE_VECTOR_TYPE) + #define LIBDIVIDE_DIVIDE_VECTOR(ALGO) +#else + #define LIBDIVIDE_DIVIDE_VECTOR(ALGO) \ + LIBDIVIDE_VECTOR_TYPE divide(LIBDIVIDE_VECTOR_TYPE n) const { \ + return libdivide_##ALGO##_do_vector(n, &denom); \ + } #endif -#pragma GCC diagnostic pop +// The DISPATCHER_GEN() macro generates C++ methods (for the given integer +// and algorithm types) that redirect to libdivide's C API. +#define DISPATCHER_GEN(T, ALGO) \ + libdivide_##ALGO##_t denom; \ + dispatcher() { } \ + dispatcher(T d) \ + : denom(libdivide_##ALGO##_gen(d)) \ + { } \ + T divide(T n) const { \ + return libdivide_##ALGO##_do(n, &denom); \ + } \ + LIBDIVIDE_DIVIDE_VECTOR(ALGO) \ + T recover() const { \ + return libdivide_##ALGO##_recover(&denom); \ + } + +// The dispatcher selects a specific division algorithm for a given +// type and ALGO using partial template specialization. +template struct dispatcher { }; + +template<> struct dispatcher { DISPATCHER_GEN(int32_t, s32) }; +template<> struct dispatcher { DISPATCHER_GEN(int32_t, s32_branchfree) }; +template<> struct dispatcher { DISPATCHER_GEN(uint32_t, u32) }; +template<> struct dispatcher { DISPATCHER_GEN(uint32_t, u32_branchfree) }; +template<> struct dispatcher { DISPATCHER_GEN(int64_t, s64) }; +template<> struct dispatcher { DISPATCHER_GEN(int64_t, s64_branchfree) }; +template<> struct dispatcher { DISPATCHER_GEN(uint64_t, u64) }; +template<> struct dispatcher { DISPATCHER_GEN(uint64_t, u64_branchfree) }; + +// This is the main divider class for use by the user (C++ API). +// The actual division algorithm is selected using the dispatcher struct +// based on the integer and algorithm template parameters. +template +class divider { +public: + // We leave the default constructor empty so that creating + // an array of dividers and then initializing them + // later doesn't slow us down. + divider() { } + + // Constructor that takes the divisor as a parameter + divider(T d) : div(d) { } + + // Divides n by the divisor + T divide(T n) const { + return div.divide(n); + } + + // Recovers the divisor, returns the value that was + // used to initialize this divider object. + T recover() const { + return div.recover(); + } + + bool operator==(const divider& other) const { + return div.denom.magic == other.denom.magic && + div.denom.more == other.denom.more; + } + + bool operator!=(const divider& other) const { + return !(*this == other); + } + +#if defined(LIBDIVIDE_VECTOR_TYPE) + // Treats the vector as packed integer values with the same type as + // the divider (e.g. s32, u32, s64, u64) and divides each of + // them by the divider, returning the packed quotients. + LIBDIVIDE_VECTOR_TYPE divide(LIBDIVIDE_VECTOR_TYPE n) const { + return div.divide(n); + } +#endif + +private: + // Storage for the actual divisor + dispatcher::value, + std::is_signed::value, sizeof(T), ALGO> div; +}; + +// Overload of operator / for scalar division +template +T operator/(T n, const divider& div) { + return div.divide(n); +} + +// Overload of operator /= for scalar division +template +T& operator/=(T& n, const divider& div) { + n = div.divide(n); + return n; +} + +#if defined(LIBDIVIDE_VECTOR_TYPE) + // Overload of operator / for vector division + template + LIBDIVIDE_VECTOR_TYPE operator/(LIBDIVIDE_VECTOR_TYPE n, const divider& div) { + return div.divide(n); + } + // Overload of operator /= for vector division + template + LIBDIVIDE_VECTOR_TYPE& operator/=(LIBDIVIDE_VECTOR_TYPE& n, const divider& div) { + n = div.divide(n); + return n; + } +#endif + +// libdivdie::branchfree_divider +template +using branchfree_divider = divider; + +} // namespace libdivide + +#endif // __cplusplus + +#endif // LIBDIVIDE_H diff --git a/src/Functions/intDiv.cpp b/src/Functions/intDiv.cpp index 0b6734c0136..062a374c00f 100644 --- a/src/Functions/intDiv.cpp +++ b/src/Functions/intDiv.cpp @@ -1,8 +1,9 @@ #include #include -#ifdef __SSE2__ - #define LIBDIVIDE_USE_SSE2 1 +#if defined(__SSE2__) +# define LIBDIVIDE_SSE2 1 +# define LIBDIVIDE_VECTOR_TYPE #endif #include @@ -45,7 +46,7 @@ struct DivideIntegralByConstantImpl const A * a_end = a_pos + size; -#ifdef __SSE2__ +#if defined(__SSE2__) static constexpr size_t values_per_sse_register = 16 / sizeof(A); const A * a_end_sse = a_pos + size / values_per_sse_register * values_per_sse_register; diff --git a/src/Functions/modulo.cpp b/src/Functions/modulo.cpp index 9e4409ca91b..631b7d12263 100644 --- a/src/Functions/modulo.cpp +++ b/src/Functions/modulo.cpp @@ -1,8 +1,8 @@ #include #include -#ifdef __SSE2__ - #define LIBDIVIDE_USE_SSE2 1 +#if defined(__SSE2__) +# define LIBDIVIDE_SSE2 1 #endif #include diff --git a/src/Interpreters/createBlockSelector.cpp b/src/Interpreters/createBlockSelector.cpp index 2b08ca0845c..0759b9d9601 100644 --- a/src/Interpreters/createBlockSelector.cpp +++ b/src/Interpreters/createBlockSelector.cpp @@ -5,8 +5,8 @@ #include -#ifdef __SSE2__ - #define LIBDIVIDE_USE_SSE2 1 +#if defined(__SSE2__) +# define LIBDIVIDE_SSE2 1 #endif #include