michael@0: /*
michael@0:  * Copyright 2012 Google Inc.
michael@0:  *
michael@0:  * Use of this source code is governed by a BSD-style license that can be
michael@0:  * found in the LICENSE file.
michael@0:  */
michael@0: 
michael@0: #ifndef SkChecksum_DEFINED
michael@0: #define SkChecksum_DEFINED
michael@0: 
michael@0: #include "SkTypes.h"
michael@0: 
michael@0: /**
michael@0:  *  Computes a 32bit checksum from a blob of 32bit aligned data. This is meant
michael@0:  *  to be very very fast, as it is used internally by the font cache, in
michael@0:  *  conjuction with the entire raw key. This algorithm does not generate
michael@0:  *  unique values as well as others (e.g. MD5) but it performs much faster.
michael@0:  *  Skia's use cases can survive non-unique values (since the entire key is
michael@0:  *  always available). Clients should only be used in circumstances where speed
michael@0:  *  over uniqueness is at a premium.
michael@0:  */
michael@0: class SkChecksum : SkNoncopyable {
michael@0: private:
michael@0:     /*
michael@0:      *  Our Rotate and Mash helpers are meant to automatically do the right
michael@0:      *  thing depending if sizeof(uintptr_t) is 4 or 8.
michael@0:      */
michael@0:     enum {
michael@0:         ROTR = 17,
michael@0:         ROTL = sizeof(uintptr_t) * 8 - ROTR,
michael@0:         HALFBITS = sizeof(uintptr_t) * 4
michael@0:     };
michael@0: 
michael@0:     static inline uintptr_t Mash(uintptr_t total, uintptr_t value) {
michael@0:         return ((total >> ROTR) | (total << ROTL)) ^ value;
michael@0:     }
michael@0: 
michael@0: public:
michael@0: 
michael@0:     /**
michael@0:      * Calculate 32-bit Murmur hash (murmur3).
michael@0:      * This should take 2-3x longer than SkChecksum::Compute, but is a considerably better hash.
michael@0:      * See en.wikipedia.org/wiki/MurmurHash.
michael@0:      *
michael@0:      *  @param data Memory address of the data block to be processed. Must be 32-bit aligned.
michael@0:      *  @param size Size of the data block in bytes. Must be a multiple of 4.
michael@0:      *  @param seed Initial hash seed. (optional)
michael@0:      *  @return hash result
michael@0:      */
michael@0:     static uint32_t Murmur3(const uint32_t* data, size_t bytes, uint32_t seed=0) {
michael@0:         SkASSERT(SkIsAlign4(bytes));
michael@0:         const size_t words = bytes/4;
michael@0: 
michael@0:         uint32_t hash = seed;
michael@0:         for (size_t i = 0; i < words; i++) {
michael@0:             uint32_t k = data[i];
michael@0:             k *= 0xcc9e2d51;
michael@0:             k = (k << 15) | (k >> 17);
michael@0:             k *= 0x1b873593;
michael@0: 
michael@0:             hash ^= k;
michael@0:             hash = (hash << 13) | (hash >> 19);
michael@0:             hash *= 5;
michael@0:             hash += 0xe6546b64;
michael@0:         }
michael@0:         hash ^= bytes;
michael@0:         hash ^= hash >> 16;
michael@0:         hash *= 0x85ebca6b;
michael@0:         hash ^= hash >> 13;
michael@0:         hash *= 0xc2b2ae35;
michael@0:         hash ^= hash >> 16;
michael@0:         return hash;
michael@0:     }
michael@0: 
michael@0:     /**
michael@0:      *  Compute a 32-bit checksum for a given data block
michael@0:      *
michael@0:      *  WARNING: this algorithm is tuned for efficiency, not backward/forward
michael@0:      *  compatibility.  It may change at any time, so a checksum generated with
michael@0:      *  one version of the Skia code may not match a checksum generated with
michael@0:      *  a different version of the Skia code.
michael@0:      *
michael@0:      *  @param data Memory address of the data block to be processed. Must be
michael@0:      *              32-bit aligned.
michael@0:      *  @param size Size of the data block in bytes. Must be a multiple of 4.
michael@0:      *  @return checksum result
michael@0:      */
michael@0:     static uint32_t Compute(const uint32_t* data, size_t size) {
michael@0:         SkASSERT(SkIsAlign4(size));
michael@0: 
michael@0:         /*
michael@0:          *  We want to let the compiler use 32bit or 64bit addressing and math
michael@0:          *  so we use uintptr_t as our magic type. This makes the code a little
michael@0:          *  more obscure (we can't hard-code 32 or 64 anywhere, but have to use
michael@0:          *  sizeof()).
michael@0:          */
michael@0:         uintptr_t result = 0;
michael@0:         const uintptr_t* ptr = reinterpret_cast<const uintptr_t*>(data);
michael@0: 
michael@0:         /*
michael@0:          *  count the number of quad element chunks. This takes into account
michael@0:          *  if we're on a 32bit or 64bit arch, since we use sizeof(uintptr_t)
michael@0:          *  to compute how much to shift-down the size.
michael@0:          */
michael@0:         size_t n4 = size / (sizeof(uintptr_t) << 2);
michael@0:         for (size_t i = 0; i < n4; ++i) {
michael@0:             result = Mash(result, *ptr++);
michael@0:             result = Mash(result, *ptr++);
michael@0:             result = Mash(result, *ptr++);
michael@0:             result = Mash(result, *ptr++);
michael@0:         }
michael@0:         size &= ((sizeof(uintptr_t) << 2) - 1);
michael@0: 
michael@0:         data = reinterpret_cast<const uint32_t*>(ptr);
michael@0:         const uint32_t* stop = data + (size >> 2);
michael@0:         while (data < stop) {
michael@0:             result = Mash(result, *data++);
michael@0:         }
michael@0: 
michael@0:         /*
michael@0:          *  smash us down to 32bits if we were 64. Note that when uintptr_t is
michael@0:          *  32bits, this code-path should go away, but I still got a warning
michael@0:          *  when I wrote
michael@0:          *      result ^= result >> 32;
michael@0:          *  since >>32 is undefined for 32bit ints, hence the wacky HALFBITS
michael@0:          *  define.
michael@0:          */
michael@0:         if (8 == sizeof(result)) {
michael@0:             result ^= result >> HALFBITS;
michael@0:         }
michael@0:         return static_cast<uint32_t>(result);
michael@0:     }
michael@0: };
michael@0: 
michael@0: #endif