@@ -134,27 +134,27 @@ struct LZSSMatch
134134};
135135
136136/* *
137- @return The number of bytes written into 'dst' or a negative value if this function fails.
138- @param capacity ... Byte size of a destination buffer 'dst'. This is at least more than the size which is reterned by the compressBound.
139- @param dst ... A destination buffer.
140- @param size ... Byte size of a source data. The maximum supported value for this is MAX_BLOCK_SIZE.
141- @param src ... A source data.
142- */
137+ @return The number of bytes written into 'dst' or a negative value if this function fails.
138+ @param capacity ... Byte size of a destination buffer 'dst'. This is at least more than the size which is reterned by the compressBound.
139+ @param dst ... A destination buffer.
140+ @param size ... Byte size of a source data. The maximum supported value for this is MAX_BLOCK_SIZE.
141+ @param src ... A source data.
142+ */
143143s32 compress (SLZ4Context& context, u32 capacity, u8* dst, u32 size, const u8* src);
144144
145145/* *
146- @return The maximum number of bytes that compression may write into a destination buffer.
147- @param size ... Byte size of a source data. The maximum supported value for this is MAX_BLOCK_SIZE.
148- */
146+ @return The maximum number of bytes that compression may write into a destination buffer.
147+ @param size ... Byte size of a source data. The maximum supported value for this is MAX_BLOCK_SIZE.
148+ */
149149s32 compressBound (u32 size);
150150
151151/* *
152- @return The number of bytes written into 'dst' or a negative value if this function fails.
153- @param capacity ... Byte size of a destination buffer 'dst'. The maximum supported value for this is MAX_BLOCK_SIZE.
154- @param dst ... A destination buffer.
155- @param size ... Byte size of a compressed data 'src'.
156- @param src ... A source compressed data.
157- */
152+ @return The number of bytes written into 'dst' or a negative value if this function fails.
153+ @param capacity ... Byte size of a destination buffer 'dst'. The maximum supported value for this is MAX_BLOCK_SIZE.
154+ @param dst ... A destination buffer.
155+ @param size ... Byte size of a compressed data 'src'.
156+ @param src ... A source compressed data.
157+ */
158158s32 decompress (u32 capacity, u8* dst, u32 size, const u8* src);
159159
160160} // namespace slz4
@@ -165,8 +165,8 @@ s32 decompress(u32 capacity, u8* dst, u32 size, const u8* src);
165165#include < immintrin.h>
166166
167167#define XXH_INLINE_ALL
168- #define XXH_STATIC_LINKING_ONLY /* access advanced declarations */
169- #define XXH_IMPLEMENTATION /* access definitions */
168+ #define XXH_STATIC_LINKING_ONLY /* access advanced declarations */
169+ #define XXH_IMPLEMENTATION /* access definitions */
170170#include < xxhash.h>
171171
172172namespace slz4
@@ -187,7 +187,7 @@ namespace
187187
188188 u32 push (u32 code, u32 u)
189189 {
190- return (code >> 8 ) | (u<< 24 );
190+ return (code >> 8 ) | (u << 24 );
191191 }
192192
193193 u32 hash (u32 code)
@@ -365,27 +365,42 @@ namespace
365365 }
366366
367367 // -------------------------------------------------------------------
368- void copy (u8* dst, const u8* src , u32 size)
368+ void set (u8* dst, s32 value , u32 size)
369369 {
370- u32 s = size>>6 ;
370+ u32 s = size>>7 ;
371+ SLZ4_ASSERT (size == (s<<7 ));
372+ __m256i x = _mm256_set1_epi32 (value);
371373 for (u32 i=0 ; i<s; ++i){
374+ _mm256_store_si256 (reinterpret_cast <__m256i*>(dst), x);
375+ _mm256_store_si256 (reinterpret_cast <__m256i*>(dst+32 ), x);
376+ _mm256_store_si256 (reinterpret_cast <__m256i*>(dst+64 ), x);
377+ _mm256_store_si256 (reinterpret_cast <__m256i*>(dst+96 ), x);
378+ dst += 128 ;
379+ }
380+ }
381+
382+ // -------------------------------------------------------------------
383+ void copy (u8* dst, const u8* src, u32 size)
384+ {
385+ u32 s = size >> 6 ;
386+ for (u32 i = 0 ; i < s; ++i) {
372387 _mm_storeu_ps (reinterpret_cast <f32*>(dst), _mm_loadu_ps (reinterpret_cast <const f32*>(src)));
373- _mm_storeu_ps (reinterpret_cast <f32*>(dst+ 16 ), _mm_loadu_ps (reinterpret_cast <const f32*>(src+ 16 )));
374- _mm_storeu_ps (reinterpret_cast <f32*>(dst+ 32 ), _mm_loadu_ps (reinterpret_cast <const f32*>(src+ 32 )));
375- _mm_storeu_ps (reinterpret_cast <f32*>(dst+ 48 ), _mm_loadu_ps (reinterpret_cast <const f32*>(src+ 48 )));
388+ _mm_storeu_ps (reinterpret_cast <f32*>(dst + 16 ), _mm_loadu_ps (reinterpret_cast <const f32*>(src + 16 )));
389+ _mm_storeu_ps (reinterpret_cast <f32*>(dst + 32 ), _mm_loadu_ps (reinterpret_cast <const f32*>(src + 32 )));
390+ _mm_storeu_ps (reinterpret_cast <f32*>(dst + 48 ), _mm_loadu_ps (reinterpret_cast <const f32*>(src + 48 )));
376391 dst += 64 ;
377392 src += 64 ;
378393 }
379- size -= (s<< 6 );
380- s = size>> 4 ;
381- for (u32 i= 0 ; i< s; ++i){
394+ size -= (s << 6 );
395+ s = size >> 4 ;
396+ for (u32 i = 0 ; i < s; ++i) {
382397 _mm_storeu_ps (reinterpret_cast <f32*>(dst), _mm_loadu_ps (reinterpret_cast <const f32*>(src)));
383398 dst += 16 ;
384399 src += 16 ;
385400 }
386401
387- s = size - (s<< 4 );
388- for (u32 i= 0 ; i< s; ++i){
402+ s = size - (s << 4 );
403+ for (u32 i = 0 ; i < s; ++i) {
389404 dst[i] = src[i];
390405 }
391406 }
@@ -406,7 +421,8 @@ s32 compress(SLZ4Context& context, u32 capacity, u8* dst, u32 size, const u8* sr
406421 : -1 ;
407422 }
408423
409- memset (context.entries_ , -1 , sizeof (s32) * DICTIONARY_SIZE);
424+ // memset(context.entries_, -1, sizeof(s32) * DICTIONARY_SIZE);
425+ set (reinterpret_cast <u8*>(context.entries_ ), -1 , sizeof (s32)*DICTIONARY_SIZE);
410426 { // Add the first code to our dictionary
411427 u32 code = pack (src);
412428 u32 index = hash (code) & (DICTIONARY_SIZE - 1 );
@@ -418,7 +434,7 @@ s32 compress(SLZ4Context& context, u32 capacity, u8* dst, u32 size, const u8* sr
418434 u32 pending = 0 ;
419435 u32 position = 4 ;
420436 while (position < endMatch) {
421- u32 code = pack (src+ position);
437+ u32 code = pack (src + position);
422438 LZSSMatch match = findLongestMatch (context, code, src, position, end);
423439
424440 if (MIN_MATCH_LENGTH <= match.length_ ) {
@@ -454,7 +470,7 @@ s32 decompress(u32 capacity, u8* dst, u32 size, const u8* src)
454470{
455471 SLZ4_ASSERT (0 <= size);
456472 SLZ4_ASSERT (0 <= capacity && capacity <= MAX_BLOCK_SIZE);
457- if (MAX_BLOCK_SIZE< capacity){
473+ if (MAX_BLOCK_SIZE < capacity) {
458474 return -1 ;
459475 }
460476 const u8* current = src;
@@ -464,10 +480,11 @@ s32 decompress(u32 capacity, u8* dst, u32 size, const u8* src)
464480 u8* d = dst;
465481
466482 for (;;) {
467- // Decode token
468483 if (end0 <= current) {
469- return - 1 ;
484+ break ;
470485 }
486+
487+ // Decode token
471488 u32 literalLength = (current[0 ] >> 4 );
472489 u32 matchLength = (current[0 ]) & 0xFU ;
473490 ++current;
@@ -478,13 +495,12 @@ s32 decompress(u32 capacity, u8* dst, u32 size, const u8* src)
478495 }
479496
480497 // Read literals
481- if (end0< (current + literalLength)){
498+ if (end0 < (current + literalLength)) {
482499 return -1 ;
483500 }
484- if (dend< (d + literalLength)){
501+ if (dend < (d + literalLength)) {
485502 return -1 ;
486503 }
487- // memcpy(d, current, literalLength);
488504 copy (d, current, literalLength);
489505 d += literalLength;
490506 current += literalLength;
@@ -503,15 +519,15 @@ s32 decompress(u32 capacity, u8* dst, u32 size, const u8* src)
503519 }
504520
505521 // Copy match
506- if (static_cast <s32>(d- dst) < offset) {
522+ if (static_cast <s32>(d - dst) < offset) {
507523 return -1 ;
508524 }
509525 matchLength += MIN_MATCH_LENGTH;
510- if (dend< (d + matchLength)){
526+ if (dend < (d + matchLength)) {
511527 return -1 ;
512528 }
513529 if (16 <= offset) {
514- copy (d, d- offset, matchLength);
530+ copy (d, d - offset, matchLength);
515531 d += matchLength;
516532 } else {
517533 while (0 < matchLength) {
@@ -521,7 +537,8 @@ s32 decompress(u32 capacity, u8* dst, u32 size, const u8* src)
521537 }
522538 }
523539 }
524- return static_cast <s32>(d- dst);
540+ return static_cast <s32>(d - dst);
525541}
542+
526543} // namespace slz4
527544#endif // SLZ4_IMPLEMENTATION
0 commit comments