diff --git a/Makefile b/Makefile index f67791275..6c4ab9c9b 100644 --- a/Makefile +++ b/Makefile @@ -114,7 +114,7 @@ clean: ifneq (,$(filter $(shell uname),Linux Darwin GNU/kFreeBSD GNU OpenBSD FreeBSD DragonFly NetBSD MSYS_NT)) HOST_OS = POSIX -CMAKE_PARAMS = -DZSTD_BUILD_CONTRIB:BOOL=ON -DZSTD_BUILD_STATIC:BOOL=ON -DZSTD_BUILD_TESTS:BOOL=ON -DZSTD_ZLIB_SUPPORT:BOOL=ON -DZSTD_LZMA_SUPPORT:BOOL=ON +CMAKE_PARAMS = -DZSTD_BUILD_CONTRIB:BOOL=ON -DZSTD_BUILD_STATIC:BOOL=ON -DZSTD_BUILD_TESTS:BOOL=ON -DZSTD_ZLIB_SUPPORT:BOOL=ON -DZSTD_LZMA_SUPPORT:BOOL=ON -DCMAKE_BUILD_TYPE=Release .PHONY: list list: diff --git a/README.md b/README.md index 17edecb71..dc99dc0fd 100644 --- a/README.md +++ b/README.md @@ -121,6 +121,8 @@ A `cmake` project generator is provided within `build/cmake`. It can generate Makefiles or other build scripts to create `zstd` binary, and `libzstd` dynamic and static libraries. +By default, `CMAKE_BUILD_TYPE` is set to `Release`. + #### Meson A Meson project is provided within `contrib/meson`. diff --git a/build/.gitignore b/build/.gitignore index b00e709ba..1ceb70ebc 100644 --- a/build/.gitignore +++ b/build/.gitignore @@ -18,3 +18,14 @@ Studio* # CMake cmake/build/ +CMakeCache.txt +CMakeFiles +CMakeScripts +Testing +Makefile +cmake_install.cmake +install_manifest.txt +compile_commands.json +CTestTestfile.cmake +build +lib diff --git a/build/cmake/CMakeLists.txt b/build/cmake/CMakeLists.txt index fd9bc2b1e..1e2921de8 100644 --- a/build/cmake/CMakeLists.txt +++ b/build/cmake/CMakeLists.txt @@ -10,6 +10,10 @@ PROJECT(zstd) CMAKE_MINIMUM_REQUIRED(VERSION 2.8.9) SET(ZSTD_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../..") + +# Ensure Release build even if not invoked via Makefile +SET(CMAKE_BUILD_TYPE "Release") + LIST(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/CMakeModules") INCLUDE(GNUInstallDirs) diff --git a/programs/bench.c b/programs/bench.c index 76d1ff6dd..79ef42caa 100644 --- a/programs/bench.c +++ b/programs/bench.c @@ -171,6 +171,8 @@ struct BMK_timeState_t{ static void BMK_initCCtx(ZSTD_CCtx* ctx, const void* dictBuffer, size_t dictBufferSize, int cLevel, const ZSTD_compressionParameters* comprParams, const BMK_advancedParams_t* adv) { + ZSTD_CCtx_reset(ctx); + ZSTD_CCtx_resetParameters(ctx); if (adv->nbWorkers==1) { ZSTD_CCtx_setParameter(ctx, ZSTD_p_nbWorkers, 0); } else { @@ -195,6 +197,7 @@ static void BMK_initCCtx(ZSTD_CCtx* ctx, static void BMK_initDCtx(ZSTD_DCtx* dctx, const void* dictBuffer, size_t dictBufferSize) { + ZSTD_DCtx_reset(dctx); ZSTD_DCtx_loadDictionary(dctx, dictBuffer, dictBufferSize); } @@ -291,34 +294,34 @@ BMK_customReturn_t BMK_benchFunction( BMK_initFn_t initFn, void* initPayload, size_t blockCount, const void* const * const srcBlockBuffers, const size_t* srcBlockSizes, - void* const * const dstBlockBuffers, const size_t* dstBlockCapacities, + void* const * const dstBlockBuffers, const size_t* dstBlockCapacities, size_t* blockResult, unsigned nbLoops) { - size_t srcSize = 0, dstSize = 0, ind = 0; + size_t dstSize = 0; U64 totalTime; BMK_customReturn_t retval; UTIL_time_t clockStart; - { - unsigned i; - for(i = 0; i < blockCount; i++) { - memset(dstBlockBuffers[i], 0xE5, dstBlockCapacities[i]); /* warm up and erase result buffer */ - } - - UTIL_sleepMilli(5); /* give processor time to other processes */ - UTIL_waitForNextTick(); - } - if(!nbLoops) { EXM_THROW_ND(1, BMK_customReturn_t, "nbLoops must be nonzero \n"); } - for(ind = 0; ind < blockCount; ind++) { - srcSize += srcBlockSizes[ind]; + { + size_t i; + for(i = 0; i < blockCount; i++) { + memset(dstBlockBuffers[i], 0xE5, dstBlockCapacities[i]); /* warm up and erase result buffer */ + } +#if 0 + /* based on testing these seem to lower accuracy of multiple calls of 1 nbLoops vs 1 call of multiple nbLoops + * (Makes former slower) + */ + UTIL_sleepMilli(5); /* give processor time to other processes */ + UTIL_waitForNextTick(); +#endif } { - unsigned i, j, firstIter = 1; + unsigned i, j; clockStart = UTIL_getTime(); if(initFn != NULL) { initFn(initPayload); } for(i = 0; i < nbLoops; i++) { @@ -327,11 +330,13 @@ BMK_customReturn_t BMK_benchFunction( if(ZSTD_isError(res)) { EXM_THROW_ND(2, BMK_customReturn_t, "Function benchmarking failed on block %u of size %u : %s \n", j, (U32)dstBlockCapacities[j], ZSTD_getErrorName(res)); - } else if(firstIter) { + } else if(i == nbLoops - 1) { dstSize += res; - } + if(blockResult != NULL) { + blockResult[j] = res; + } + } } - firstIter = 0; } totalTime = UTIL_clockSpanNano(clockStart); } @@ -353,6 +358,9 @@ void BMK_resetTimeState(BMK_timedFnState_t* r, unsigned nbSeconds) { BMK_timedFnState_t* BMK_createTimeState(unsigned nbSeconds) { BMK_timedFnState_t* r = (BMK_timedFnState_t*)malloc(sizeof(struct BMK_timeState_t)); + if(r == NULL) { + return r; + } BMK_resetTimeState(r, nbSeconds); return r; } @@ -361,13 +369,14 @@ void BMK_freeTimeState(BMK_timedFnState_t* state) { free(state); } +/* make option for dstBlocks to be */ BMK_customTimedReturn_t BMK_benchFunctionTimed( BMK_timedFnState_t* cont, BMK_benchFn_t benchFn, void* benchPayload, BMK_initFn_t initFn, void* initPayload, size_t blockCount, const void* const* const srcBlockBuffers, const size_t* srcBlockSizes, - void* const* const dstBlockBuffers, const size_t* dstBlockCapacities) + void * const * const dstBlockBuffers, const size_t * dstBlockCapacities, size_t* blockResults) { U64 fastest = cont->fastestTime; int completed = 0; @@ -382,9 +391,9 @@ BMK_customTimedReturn_t BMK_benchFunctionTimed( UTIL_sleep(COOLPERIOD_SEC); cont->coolTime = UTIL_getTime(); } - + /* reinitialize capacity */ r.result = BMK_benchFunction(benchFn, benchPayload, initFn, initPayload, - blockCount, srcBlockBuffers, srcBlockSizes, dstBlockBuffers, dstBlockCapacities, cont->nbLoops); + blockCount, srcBlockBuffers, srcBlockSizes, dstBlockBuffers, dstBlockCapacities, blockResults, cont->nbLoops); if(r.result.error) { /* completed w/ error */ r.completed = 1; return r; @@ -393,7 +402,7 @@ BMK_customTimedReturn_t BMK_benchFunctionTimed( { U64 const loopDuration = r.result.result.nanoSecPerRun * cont->nbLoops; r.completed = (cont->timeRemaining <= loopDuration); cont->timeRemaining -= loopDuration; - if (loopDuration > 0) { + if (loopDuration > (TIMELOOP_NANOSEC / 100)) { fastest = MIN(fastest, r.result.result.nanoSecPerRun); if(loopDuration >= MINUSABLETIME) { r.result.result.nanoSecPerRun = fastest; @@ -418,9 +427,9 @@ BMK_customTimedReturn_t BMK_benchFunctionTimed( /* benchMem with no allocation */ static BMK_return_t BMK_benchMemAdvancedNoAlloc( const void ** const srcPtrs, size_t* const srcSizes, - void** const cPtrs, size_t* const cSizes, + void** const cPtrs, size_t* const cCapacities, size_t* const cSizes, void** const resPtrs, size_t* const resSizes, - void* resultBuffer, void* compressedBuffer, + void** resultBufferPtr, void* compressedBuffer, const size_t maxCompressedSize, BMK_timedFnState_t* timeStateCompress, BMK_timedFnState_t* timeStateDecompress, @@ -432,7 +441,7 @@ static BMK_return_t BMK_benchMemAdvancedNoAlloc( int displayLevel, const char* displayName, const BMK_advancedParams_t* adv) { size_t const blockSize = ((adv->blockSize>=32 && (adv->mode != BMK_decodeOnly)) ? adv->blockSize : srcSize) + (!srcSize); /* avoid div by 0 */ - BMK_return_t results; + BMK_return_t results = { { 0, 0, 0, 0 }, 0 } ; size_t const loadedCompressedSize = srcSize; size_t cSize = 0; double ratio = 0.; @@ -454,13 +463,13 @@ static BMK_return_t BMK_benchMemAdvancedNoAlloc( srcPtr += fileSizes[fileNb]; } { size_t const decodedSize = (size_t)totalDSize64; - free(resultBuffer); - resultBuffer = malloc(decodedSize); - if (!resultBuffer) { + free(*resultBufferPtr); + *resultBufferPtr = malloc(decodedSize); + if (!(*resultBufferPtr)) { EXM_THROW(33, BMK_return_t, "not enough memory"); } if (totalDSize64 > decodedSize) { - free(resultBuffer); + free(*resultBufferPtr); EXM_THROW(32, BMK_return_t, "original size is too large"); /* size_t overflow */ } cSize = srcSize; @@ -472,7 +481,7 @@ static BMK_return_t BMK_benchMemAdvancedNoAlloc( /* Init data blocks */ { const char* srcPtr = (const char*)srcBuffer; char* cPtr = (char*)compressedBuffer; - char* resPtr = (char*)resultBuffer; + char* resPtr = (char*)(*resultBufferPtr); U32 fileNb; for (nbBlocks=0, fileNb=0; fileNbmode == BMK_decodeOnly) ? thisBlockSize : ZSTD_compressBound(thisBlockSize); + cCapacities[nbBlocks] = (adv->mode == BMK_decodeOnly) ? thisBlockSize : ZSTD_compressBound(thisBlockSize); resPtrs[nbBlocks] = (void*)resPtr; resSizes[nbBlocks] = (adv->mode == BMK_decodeOnly) ? (size_t) ZSTD_findDecompressedSize(srcPtr, thisBlockSize) : thisBlockSize; srcPtr += thisBlockSize; - cPtr += cSizes[nbBlocks]; + cPtr += cCapacities[nbBlocks]; resPtr += thisBlockSize; remaining -= thisBlockSize; } @@ -538,7 +547,7 @@ static BMK_return_t BMK_benchMemAdvancedNoAlloc( while(!(intermediateResultCompress.completed && intermediateResultDecompress.completed)) { if(!intermediateResultCompress.completed) { intermediateResultCompress = BMK_benchFunctionTimed(timeStateCompress, &local_defaultCompress, (void*)ctx, &local_initCCtx, (void*)&cctxprep, - nbBlocks, srcPtrs, srcSizes, cPtrs, cSizes); + nbBlocks, srcPtrs, srcSizes, cPtrs, cCapacities, cSizes); if(intermediateResultCompress.result.error) { results.error = intermediateResultCompress.result.error; return results; @@ -546,9 +555,7 @@ static BMK_return_t BMK_benchMemAdvancedNoAlloc( ratio = (double)(srcSize / intermediateResultCompress.result.result.sumOfReturn); { int const ratioAccuracy = (ratio < 10.) ? 3 : 2; - double const compressionSpeed = ((double)srcSize / intermediateResultCompress.result.result.nanoSecPerRun) * 1000; - int const cSpeedAccuracy = (compressionSpeed < 10.) ? 2 : 1; - results.result.cSpeed = compressionSpeed * 1000000; + results.result.cSpeed = (srcSize * TIMELOOP_NANOSEC / intermediateResultCompress.result.result.nanoSecPerRun); cSize = intermediateResultCompress.result.result.sumOfReturn; results.result.cSize = cSize; ratio = (double)srcSize / results.result.cSize; @@ -556,13 +563,13 @@ static BMK_return_t BMK_benchMemAdvancedNoAlloc( DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.*f),%6.*f MB/s\r", marks[markNb], displayName, (U32)srcSize, (U32)results.result.cSize, ratioAccuracy, ratio, - cSpeedAccuracy, compressionSpeed); + results.result.cSpeed < (10 MB) ? 2 : 1, (double)results.result.cSpeed / (1 MB)); } } if(!intermediateResultDecompress.completed) { intermediateResultDecompress = BMK_benchFunctionTimed(timeStateDecompress, &local_defaultDecompress, (void*)(dctx), &local_initDCtx, (void*)&dctxprep, - nbBlocks, (const void* const*)cPtrs, cSizes, resPtrs, resSizes); + nbBlocks, (const void* const*)cPtrs, cSizes, resPtrs, resSizes, NULL); if(intermediateResultDecompress.result.error) { results.error = intermediateResultDecompress.result.error; return results; @@ -570,79 +577,79 @@ static BMK_return_t BMK_benchMemAdvancedNoAlloc( { int const ratioAccuracy = (ratio < 10.) ? 3 : 2; - double const compressionSpeed = results.result.cSpeed / 1000000; - int const cSpeedAccuracy = (compressionSpeed < 10.) ? 2 : 1; - double const decompressionSpeed = ((double)srcSize / intermediateResultDecompress.result.result.nanoSecPerRun) * 1000; - results.result.dSpeed = decompressionSpeed * 1000000; + results.result.dSpeed = (srcSize * TIMELOOP_NANOSEC/ intermediateResultDecompress.result.result.nanoSecPerRun); markNb = (markNb+1) % NB_MARKS; DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.*f),%6.*f MB/s ,%6.1f MB/s \r", marks[markNb], displayName, (U32)srcSize, (U32)results.result.cSize, ratioAccuracy, ratio, - cSpeedAccuracy, compressionSpeed, - decompressionSpeed); + results.result.cSpeed < (10 MB) ? 2 : 1, (double)results.result.cSpeed / (1 MB), + (double)results.result.dSpeed / (1 MB)); } } } - } else { + } else { //iterMode; if(adv->mode != BMK_decodeOnly) { + BMK_customReturn_t compressionResults = BMK_benchFunction(&local_defaultCompress, (void*)ctx, &local_initCCtx, (void*)&cctxprep, - nbBlocks, srcPtrs, srcSizes, cPtrs, cSizes, adv->nbSeconds); + nbBlocks, srcPtrs, srcSizes, cPtrs, cCapacities, cSizes, adv->nbSeconds); if(compressionResults.error) { results.error = compressionResults.error; return results; } + if(compressionResults.result.nanoSecPerRun == 0) { results.result.cSpeed = 0; } else { - results.result.cSpeed = (double)srcSize / compressionResults.result.nanoSecPerRun * TIMELOOP_NANOSEC; + results.result.cSpeed = srcSize * TIMELOOP_NANOSEC / compressionResults.result.nanoSecPerRun; } + results.result.cSize = compressionResults.result.sumOfReturn; { int const ratioAccuracy = (ratio < 10.) ? 3 : 2; - double const compressionSpeed = results.result.cSpeed / 1000000; - int const cSpeedAccuracy = (compressionSpeed < 10.) ? 2 : 1; + cSize = compressionResults.result.sumOfReturn; + results.result.cSize = cSize; ratio = (double)srcSize / results.result.cSize; markNb = (markNb+1) % NB_MARKS; DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.*f),%6.*f MB/s\r", marks[markNb], displayName, (U32)srcSize, (U32)results.result.cSize, ratioAccuracy, ratio, - cSpeedAccuracy, compressionSpeed); + results.result.cSpeed < (10 MB) ? 2 : 1, (double)results.result.cSpeed / (1 MB)); } } if(adv->mode != BMK_compressOnly) { BMK_customReturn_t decompressionResults = BMK_benchFunction( &local_defaultDecompress, (void*)(dctx), &local_initDCtx, (void*)&dctxprep, nbBlocks, - (const void* const*)cPtrs, cSizes, resPtrs, resSizes, + (const void* const*)cPtrs, cSizes, resPtrs, resSizes, NULL, adv->nbSeconds); if(decompressionResults.error) { results.error = decompressionResults.error; return results; } + if(decompressionResults.result.nanoSecPerRun == 0) { results.result.dSpeed = 0; } else { - results.result.dSpeed = (double)srcSize / decompressionResults.result.nanoSecPerRun * TIMELOOP_NANOSEC; + results.result.dSpeed = srcSize * TIMELOOP_NANOSEC / decompressionResults.result.nanoSecPerRun; } - { int const ratioAccuracy = (ratio < 10.) ? 3 : 2; - double const compressionSpeed = results.result.cSpeed / 1000000; - int const cSpeedAccuracy = (compressionSpeed < 10.) ? 2 : 1; - double const decompressionSpeed = ((double)srcSize / decompressionResults.result.nanoSecPerRun) * 1000; - results.result.dSpeed = decompressionSpeed * 1000000; + + { + int const ratioAccuracy = (ratio < 10.) ? 3 : 2; markNb = (markNb+1) % NB_MARKS; DISPLAYLEVEL(2, "%2s-%-17.17s :%10u ->%10u (%5.*f),%6.*f MB/s ,%6.1f MB/s \r", marks[markNb], displayName, (U32)srcSize, (U32)results.result.cSize, ratioAccuracy, ratio, - cSpeedAccuracy, compressionSpeed, - decompressionSpeed); + results.result.cSpeed < (10 MB) ? 2 : 1, (double)results.result.cSpeed / (1 MB), + (double)results.result.dSpeed / (1 MB)); } } } } /* CRC Checking */ - { U64 const crcCheck = XXH64(resultBuffer, srcSize, 0); + { void* resultBuffer = *resultBufferPtr; + U64 const crcCheck = XXH64(resultBuffer, srcSize, 0); /* adv->mode == 0 -> compress + decompress */ if ((adv->mode == BMK_both) && (crcOrig!=crcCheck)) { size_t u; @@ -682,8 +689,8 @@ static BMK_return_t BMK_benchMemAdvancedNoAlloc( } /* CRC Checking */ if (displayLevel == 1) { /* hidden display mode -q, used by python speed benchmark */ - double const cSpeed = results.result.cSpeed / 1000000; - double const dSpeed = results.result.dSpeed / 1000000; + double const cSpeed = (double)results.result.cSpeed / (1 MB); + double const dSpeed = (double)results.result.dSpeed / (1 MB); if (adv->additionalParam) { DISPLAY("-%-3i%11i (%5.3f) %6.2f MB/s %6.1f MB/s %s (param=%d)\n", cLevel, (int)cSize, ratio, cSpeed, dSpeed, displayName, adv->additionalParam); } else { @@ -692,14 +699,16 @@ static BMK_return_t BMK_benchMemAdvancedNoAlloc( } DISPLAYLEVEL(2, "%2i#\n", cLevel); } /* Bench */ + results.result.cMem = (1ULL << (comprParams->windowLog)) + ZSTD_sizeof_CCtx(ctx); + results.error = 0; return results; } BMK_return_t BMK_benchMemAdvanced(const void* srcBuffer, size_t srcSize, + void* dstBuffer, size_t dstCapacity, const size_t* fileSizes, unsigned nbFiles, const int cLevel, const ZSTD_compressionParameters* comprParams, const void* dictBuffer, size_t dictBufferSize, - ZSTD_CCtx* ctx, ZSTD_DCtx* dctx, int displayLevel, const char* displayName, const BMK_advancedParams_t* adv) { @@ -707,47 +716,72 @@ BMK_return_t BMK_benchMemAdvanced(const void* srcBuffer, size_t srcSize, U32 const maxNbBlocks = (U32) ((srcSize + (blockSize-1)) / blockSize) + nbFiles; /* these are the blockTable parameters, just split up */ - const void ** const srcPtrs = (const void** const)malloc(maxNbBlocks * sizeof(void*)); - size_t* const srcSizes = (size_t* const)malloc(maxNbBlocks * sizeof(size_t)); + const void ** const srcPtrs = (const void**)malloc(maxNbBlocks * sizeof(void*)); + size_t* const srcSizes = (size_t*)malloc(maxNbBlocks * sizeof(size_t)); - void ** const cPtrs = (void** const)malloc(maxNbBlocks * sizeof(void*)); - size_t* const cSizes = (size_t* const)malloc(maxNbBlocks * sizeof(size_t)); - void ** const resPtrs = (void** const)malloc(maxNbBlocks * sizeof(void*)); - size_t* const resSizes = (size_t* const)malloc(maxNbBlocks * sizeof(size_t)); + void ** const cPtrs = (void**)malloc(maxNbBlocks * sizeof(void*)); + size_t* const cSizes = (size_t*)malloc(maxNbBlocks * sizeof(size_t)); + size_t* const cCapacities = (size_t*)malloc(maxNbBlocks * sizeof(size_t)); + + void ** const resPtrs = (void**)malloc(maxNbBlocks * sizeof(void*)); + size_t* const resSizes = (size_t*)malloc(maxNbBlocks * sizeof(size_t)); - const size_t maxCompressedSize = ZSTD_compressBound(srcSize) + (maxNbBlocks * 1024); /* add some room for safety */ - void* compressedBuffer = malloc(maxCompressedSize); - void* resultBuffer = malloc(srcSize); BMK_timedFnState_t* timeStateCompress = BMK_createTimeState(adv->nbSeconds); BMK_timedFnState_t* timeStateDecompress = BMK_createTimeState(adv->nbSeconds); - BMK_return_t results; - int allocationincomplete = !compressedBuffer || !resultBuffer || - !srcPtrs || !srcSizes || !cPtrs || !cSizes || !resPtrs || !resSizes; - if (!allocationincomplete) { - results = BMK_benchMemAdvancedNoAlloc(srcPtrs, srcSizes, cPtrs, cSizes, - resPtrs, resSizes, resultBuffer, compressedBuffer, maxCompressedSize, timeStateCompress, timeStateDecompress, + ZSTD_CCtx* ctx = ZSTD_createCCtx(); + ZSTD_DCtx* dctx = ZSTD_createDCtx(); + + const size_t maxCompressedSize = dstCapacity ? dstCapacity : ZSTD_compressBound(srcSize) + (maxNbBlocks * 1024); + + void* const internalDstBuffer = dstBuffer ? NULL : malloc(maxCompressedSize); + void* const compressedBuffer = dstBuffer ? dstBuffer : internalDstBuffer; + + BMK_return_t results = { { 0, 0, 0, 0 }, 0 }; + + int parametersConflict = !dstBuffer ^ !dstCapacity; + + void* resultBuffer = srcSize ? malloc(srcSize) : NULL; + + int allocationincomplete = !srcPtrs || !srcSizes || !cPtrs || + !cSizes || !cCapacities || !resPtrs || !resSizes || + !timeStateCompress || !timeStateDecompress || !compressedBuffer || !resultBuffer; + + + + if (!allocationincomplete && !parametersConflict) { + results = BMK_benchMemAdvancedNoAlloc(srcPtrs, srcSizes, cPtrs, cCapacities, cSizes, + resPtrs, resSizes, &resultBuffer, compressedBuffer, maxCompressedSize, timeStateCompress, timeStateDecompress, srcBuffer, srcSize, fileSizes, nbFiles, cLevel, comprParams, dictBuffer, dictBufferSize, ctx, dctx, displayLevel, displayName, adv); } + /* clean up */ BMK_freeTimeState(timeStateCompress); BMK_freeTimeState(timeStateDecompress); - free(compressedBuffer); + + ZSTD_freeCCtx(ctx); + ZSTD_freeDCtx(dctx); + + free(internalDstBuffer); free(resultBuffer); free((void*)srcPtrs); free(srcSizes); free(cPtrs); free(cSizes); + free(cCapacities); free(resPtrs); free(resSizes); if(allocationincomplete) { EXM_THROW(31, BMK_return_t, "allocation error : not enough memory"); } - results.error = 0; + + if(parametersConflict) { + EXM_THROW(32, BMK_return_t, "Conflicting input results"); + } return results; } @@ -755,42 +789,17 @@ BMK_return_t BMK_benchMem(const void* srcBuffer, size_t srcSize, const size_t* fileSizes, unsigned nbFiles, const int cLevel, const ZSTD_compressionParameters* comprParams, const void* dictBuffer, size_t dictBufferSize, - ZSTD_CCtx* ctx, ZSTD_DCtx* dctx, int displayLevel, const char* displayName) { const BMK_advancedParams_t adv = BMK_initAdvancedParams(); return BMK_benchMemAdvanced(srcBuffer, srcSize, + NULL, 0, fileSizes, nbFiles, cLevel, comprParams, dictBuffer, dictBufferSize, - ctx, dctx, displayLevel, displayName, &adv); } -static BMK_return_t BMK_benchMemCtxless(const void* srcBuffer, size_t srcSize, - const size_t* fileSizes, unsigned nbFiles, - int cLevel, const ZSTD_compressionParameters* const comprParams, - const void* dictBuffer, size_t dictBufferSize, - int displayLevel, const char* displayName, - const BMK_advancedParams_t* const adv) -{ - BMK_return_t res; - ZSTD_CCtx* ctx = ZSTD_createCCtx(); - ZSTD_DCtx* dctx = ZSTD_createDCtx(); - if(ctx == NULL || dctx == NULL) { - EXM_THROW(12, BMK_return_t, "not enough memory for contexts"); - } - res = BMK_benchMemAdvanced(srcBuffer, srcSize, - fileSizes, nbFiles, - cLevel, comprParams, - dictBuffer, dictBufferSize, - ctx, dctx, - displayLevel, displayName, adv); - ZSTD_freeCCtx(ctx); - ZSTD_freeDCtx(dctx); - return res; -} - static size_t BMK_findMaxMem(U64 requiredMem) { size_t const step = 64 MB; @@ -803,7 +812,7 @@ static size_t BMK_findMaxMem(U64 requiredMem) do { testmem = (BYTE*)malloc((size_t)requiredMem); requiredMem -= step; - } while (!testmem); + } while (!testmem && requiredMem > 0); free(testmem); return (size_t)(requiredMem); @@ -831,12 +840,12 @@ static BMK_return_t BMK_benchCLevel(const void* srcBuffer, size_t benchedSize, if (displayLevel == 1 && !adv->additionalParam) DISPLAY("bench %s %s: input %u bytes, %u seconds, %u KB blocks\n", ZSTD_VERSION_STRING, ZSTD_GIT_COMMIT_STRING, (U32)benchedSize, adv->nbSeconds, (U32)(adv->blockSize>>10)); - res = BMK_benchMemCtxless(srcBuffer, benchedSize, - fileSizes, nbFiles, - cLevel, comprParams, - dictBuffer, dictBufferSize, - displayLevel, displayName, - adv); + res = BMK_benchMemAdvanced(srcBuffer, benchedSize, + NULL, 0, + fileSizes, nbFiles, + cLevel, comprParams, + dictBuffer, dictBufferSize, + displayLevel, displayName, adv); return res; } @@ -931,7 +940,8 @@ BMK_return_t BMK_benchFilesAdvanced(const char* const * const fileNamesTable, un if ((U64)benchedSize > totalSizeToLoad) benchedSize = (size_t)totalSizeToLoad; if (benchedSize < totalSizeToLoad) DISPLAY("Not enough memory; testing %u MB only...\n", (U32)(benchedSize >> 20)); - srcBuffer = malloc(benchedSize); + + srcBuffer = benchedSize ? malloc(benchedSize) : NULL; if (!srcBuffer) { free(dictBuffer); free(fileSizes); diff --git a/programs/bench.h b/programs/bench.h index 87cf56380..ad4ede1f0 100644 --- a/programs/bench.h +++ b/programs/bench.h @@ -32,8 +32,9 @@ extern "C" { typedef struct { size_t cSize; - double cSpeed; /* bytes / sec */ - double dSpeed; + U64 cSpeed; /* bytes / sec */ + U64 dSpeed; + size_t cMem; } BMK_result_t; ERROR_STRUCT(BMK_result_t, BMK_return_t); @@ -125,8 +126,6 @@ BMK_return_t BMK_syntheticTest(int cLevel, double compressibility, * comprParams - basic compression parameters * dictBuffer - a dictionary if used, null otherwise * dictBufferSize - size of dictBuffer, 0 otherwise - * ctx - Compression Context (must be provided) - * dctx - Decompression Context (must be provided) * diplayLevel - see BMK_benchFiles * displayName - name used by display * return @@ -138,15 +137,17 @@ BMK_return_t BMK_benchMem(const void* srcBuffer, size_t srcSize, const size_t* fileSizes, unsigned nbFiles, const int cLevel, const ZSTD_compressionParameters* comprParams, const void* dictBuffer, size_t dictBufferSize, - ZSTD_CCtx* ctx, ZSTD_DCtx* dctx, int displayLevel, const char* displayName); -/* See benchMem for normal parameter uses and return, see advancedParams_t for adv */ +/* See benchMem for normal parameter uses and return, see advancedParams_t for adv + * dstBuffer - destination buffer to write compressed output in, NULL if none provided. + * dstCapacity - capacity of destination buffer, give 0 if dstBuffer = NULL + */ BMK_return_t BMK_benchMemAdvanced(const void* srcBuffer, size_t srcSize, + void* dstBuffer, size_t dstCapacity, const size_t* fileSizes, unsigned nbFiles, const int cLevel, const ZSTD_compressionParameters* comprParams, const void* dictBuffer, size_t dictBufferSize, - ZSTD_CCtx* ctx, ZSTD_DCtx* dctx, int displayLevel, const char* displayName, const BMK_advancedParams_t* adv); @@ -170,8 +171,10 @@ typedef size_t (*BMK_initFn_t)(void*); * srcBuffers - an array of buffers to be operated on by benchFn * srcSizes - an array of the sizes of above buffers * dstBuffers - an array of buffers to be written into by benchFn - * dstCapacities - an array of the capacities of above buffers. + * dstCapacities - an array of the capacities of above buffers + * blockResults - the return value of benchFn called on each block. * nbLoops - defines number of times benchFn is run. + * assumed array of size blockCount, will have compressed size of each block written to it. * return * .error will give a nonzero value if ZSTD_isError() is nonzero for any of the return * of the calls to initFn and benchFn, or if benchFunction errors internally @@ -182,12 +185,11 @@ typedef size_t (*BMK_initFn_t)(void*); * into dstBuffer, hence this value will be the total amount of bytes written to * dstBuffer. */ -BMK_customReturn_t BMK_benchFunction( - BMK_benchFn_t benchFn, void* benchPayload, +BMK_customReturn_t BMK_benchFunction(BMK_benchFn_t benchFn, void* benchPayload, BMK_initFn_t initFn, void* initPayload, size_t blockCount, const void* const * const srcBuffers, const size_t* srcSizes, - void* const * const dstBuffers, const size_t* dstCapacities, + void * const * const dstBuffers, const size_t* dstCapacities, size_t* blockResults, unsigned nbLoops); @@ -216,7 +218,7 @@ BMK_customTimedReturn_t BMK_benchFunctionTimed(BMK_timedFnState_t* cont, BMK_initFn_t initFn, void* initPayload, size_t blockCount, const void* const * const srcBlockBuffers, const size_t* srcBlockSizes, - void* const * const dstBlockBuffers, const size_t* dstBlockCapacities); + void* const * const dstBlockBuffers, const size_t* dstBlockCapacities, size_t* blockResults); #endif /* BENCH_H_121279284357 */ diff --git a/tests/Makefile b/tests/Makefile index 81e685780..88a5d763c 100644 --- a/tests/Makefile +++ b/tests/Makefile @@ -129,7 +129,7 @@ zstdmt_d_%.o : $(ZSTDDIR)/decompress/%.c fullbench32: CPPFLAGS += -m32 fullbench fullbench32 : CPPFLAGS += $(MULTITHREAD_CPP) fullbench fullbench32 : LDFLAGS += $(MULTITHREAD_LD) -fullbench fullbench32 : DEBUGFLAGS = # turn off assert() for speed measurements +fullbench fullbench32 : DEBUGFLAGS = -DNDEBUG # turn off assert() for speed measurements fullbench fullbench32 : $(ZSTD_FILES) fullbench fullbench32 : $(PRGDIR)/datagen.c $(PRGDIR)/bench.c fullbench.c $(CC) $(FLAGS) $^ -o $@$(EXT) @@ -200,7 +200,7 @@ zstreamtest-dll : $(ZSTDDIR)/common/xxhash.c # xxh symbols not exposed from dll zstreamtest-dll : $(ZSTREAM_LOCAL_FILES) $(CC) $(CPPFLAGS) $(CFLAGS) $(filter %.c,$^) $(LDFLAGS) -o $@$(EXT) -paramgrill : DEBUGFLAGS = # turn off assert() for speed measurements +paramgrill : DEBUGFLAGS = -DNDEBUG # turn off assert() for speed measurements paramgrill : $(ZSTD_FILES) $(PRGDIR)/bench.c $(PRGDIR)/datagen.c paramgrill.c $(CC) $(FLAGS) $^ -lm -o $@$(EXT) diff --git a/tests/README.md b/tests/README.md index 24a28ab7b..f28766bd1 100644 --- a/tests/README.md +++ b/tests/README.md @@ -88,3 +88,56 @@ as well as the 10,000 original files for more detailed comparison of decompressi will choose a random seed, and for 1 minute, generate random test frames and ensure that the zstd library correctly decompresses them in both simple and streaming modes. + +#### `paramgrill` - tool for generating compression table parameters and optimizing parameters on file given constraints + +Full list of arguments +``` + -T# : set level 1 speed objective + -B# : cut input into blocks of size # (default : single block) + -S : benchmarks a single run (example command: -Sl3w10h12) + w# - windowLog + h# - hashLog + c# - chainLog + s# - searchLog + l# - searchLength + t# - targetLength + S# - strategy + L# - level + --zstd= : Single run, parameter selection syntax same as zstdcli with more parameters + (Added forceAttachDictionary / fadt) + When invoked with --optimize, this represents the sample to exceed. + --optimize= : find parameters to maximize compression ratio given parameters + Can use all --zstd= commands to constrain the type of solution found in addition to the following constraints + cSpeed= : Minimum compression speed + dSpeed= : Minimum decompression speed + cMem= : Maximum compression memory + lvl= : Searches for solutions which are strictly better than that compression lvl in ratio and cSpeed, + stc= : When invoked with lvl=, represents percentage slack in ratio/cSpeed allowed for a solution to be considered (Default 100%) + : In normal operation, represents percentage slack in choosing viable starting strategy selection in choosing the default parameters + (Lower value will begin with stronger strategies) (Default 90%) + speedRatio= (accepts decimals) + : determines value of gains in speed vs gains in ratio + when determining overall winner (default 5 (1% ratio = 5% speed)). + tries= : Maximum number of random restarts on a single strategy before switching (Default 5) + Higher values will make optimizer run longer, more chances to find better solution. + memLog : Limits the log of the size of each memotable (1 per strategy). Will use hash tables when state space is larger than max size. + Setting memLog = 0 turns off memoization + --display= : specifiy which parameters are included in the output + can use all --zstd parameter names and 'cParams' as a shorthand for all parameters used in ZSTD_compressionParameters + (Default: display all params available) + -P# : generated sample compressibility (when no file is provided) + -t# : Caps runtime of operation in seconds (default : 99999 seconds (about 27 hours )) + -v : Prints Benchmarking output + -D : Next argument dictionary file + -s : Benchmark all files separately + -q : Quiet, repeat for more quiet + -q Prints parameters + results whenever a new best is found + -qq Only prints parameters whenever a new best is found, prints final parameters + results + -qqq Only print final parameters + results + -qqqq Only prints final parameter set in the form --zstd= + -v : Verbose, cancels quiet, repeat for more volume + -v Prints all candidate parameters and results + +``` + Any inputs afterwards are treated as files to benchmark. diff --git a/tests/fullbench.c b/tests/fullbench.c index 7859745a3..12c1e1ae4 100644 --- a/tests/fullbench.c +++ b/tests/fullbench.c @@ -336,7 +336,7 @@ size_t local_ZSTD_decompressContinue(const void* src, size_t srcSize, void* dst, static size_t benchMem(const void* src, size_t srcSize, U32 benchNb, int cLevel, ZSTD_compressionParameters* cparams) { BYTE* dstBuff; - size_t const dstBuffSize = ZSTD_compressBound(srcSize); + size_t dstBuffSize = ZSTD_compressBound(srcSize); void* buff2, *buff1; const char* benchName; BMK_benchFn_t benchFunction; @@ -514,9 +514,10 @@ static size_t benchMem(const void* src, size_t srcSize, U32 benchNb, int cLevel, /* benchmark loop */ { + void* dstBuffv = (void*)dstBuff; r = BMK_benchFunction(benchFunction, buff2, NULL, NULL, 1, &src, &srcSize, - (void * const * const)&dstBuff, &dstBuffSize, g_nbIterations); + &dstBuffv, &dstBuffSize, NULL, g_nbIterations); if(r.error) { DISPLAY("ERROR %d ! ! \n", r.error); errorcode = r.error; diff --git a/tests/paramgrill.c b/tests/paramgrill.c index 025bc6aad..f42afe403 100644 --- a/tests/paramgrill.c +++ b/tests/paramgrill.c @@ -17,7 +17,6 @@ #include /* fprintf, fopen, ftello64 */ #include /* strcmp */ #include /* log */ -#include #include #include "mem.h" @@ -27,7 +26,8 @@ #include "xxhash.h" #include "util.h" #include "bench.h" - +#include "zstd_errors.h" +#include "zstd_internal.h" /*-************************************ * Constants @@ -36,13 +36,7 @@ #define AUTHOR "Yann Collet" #define WELCOME_MESSAGE "*** %s %s %i-bits, by %s ***\n", PROGRAM_DESCRIPTION, ZSTD_VERSION_STRING, (int)(sizeof(void*)*8), AUTHOR - -#define KB *(1<<10) -#define MB *(1<<20) -#define GB *(1ULL<<30) - -#define NBLOOPS 2 -#define TIMELOOP (2 * SEC_TO_MICRO) +#define TIMELOOP_NANOSEC (1*1000000000ULL) /* 1 second */ #define NB_LEVELS_TRACKED 22 /* ensured being >= ZSTD_maxCLevel() in BMK_init_level_constraints() */ static const size_t maxMemory = (sizeof(size_t)==4) ? (2 GB - 64 MB) : (size_t)(1ULL << ((sizeof(size_t)*8)-31)); @@ -56,39 +50,302 @@ static const int g_maxNbVariations = 64; * Macros **************************************/ #define DISPLAY(...) fprintf(stderr, __VA_ARGS__) +#define DISPLAYLEVEL(n, ...) if(g_displayLevel >= n) { fprintf(stderr, __VA_ARGS__); } +#define DEBUGOUTPUT(...) { if (DEBUG) DISPLAY(__VA_ARGS__); } + +#define TIMED 0 +#ifndef DEBUG +# define DEBUG 0 +#endif #undef MIN #undef MAX #define MIN(a,b) ( (a) < (b) ? (a) : (b) ) #define MAX(a,b) ( (a) > (b) ? (a) : (b) ) #define CUSTOM_LEVEL 99 +#define BASE_CLEVEL 1 + +#define FADT_MIN 0 +#define FADT_MAX ((U32)-1) + +#define ZSTD_TARGETLENGTH_MIN 0 +#define ZSTD_TARGETLENGTH_MAX 999 + +#define WLOG_RANGE (ZSTD_WINDOWLOG_MAX - ZSTD_WINDOWLOG_MIN + 1) +#define CLOG_RANGE (ZSTD_CHAINLOG_MAX - ZSTD_CHAINLOG_MIN + 1) +#define HLOG_RANGE (ZSTD_HASHLOG_MAX - ZSTD_HASHLOG_MIN + 1) +#define SLOG_RANGE (ZSTD_SEARCHLOG_MAX - ZSTD_SEARCHLOG_MIN + 1) +#define SLEN_RANGE (ZSTD_SEARCHLENGTH_MAX - ZSTD_SEARCHLENGTH_MIN + 1) +#define TLEN_RANGE 17 +#define STRT_RANGE (ZSTD_btultra - ZSTD_fast + 1) +#define FADT_RANGE 3 + +#define CHECKTIME(r) { if(BMK_timeSpan(g_time) > g_timeLimit_s) { DEBUGOUTPUT("Time Limit Reached\n"); return r; } } +#define CHECKTIMEGT(ret, val, _gototag) {if(BMK_timeSpan(g_time) > g_timeLimit_s) { DEBUGOUTPUT("Time Limit Reached\n"); ret = val; goto _gototag; } } + +#define PARAM_UNSET ((U32)-2) /* can't be -1 b/c fadt uses -1 */ + +static const char* g_stratName[ZSTD_btultra+1] = { + "(none) ", "ZSTD_fast ", "ZSTD_dfast ", + "ZSTD_greedy ", "ZSTD_lazy ", "ZSTD_lazy2 ", + "ZSTD_btlazy2 ", "ZSTD_btopt ", "ZSTD_btultra "}; + + +static const U32 tlen_table[TLEN_RANGE] = { 0, 1, 2, 4, 6, 8, 12, 16, 24, 32, 48, 64, 96, 128, 256, 512, 999 }; /*-************************************ -* Benchmark Parameters +* Setup for Adding new params **************************************/ -static double g_grillDuration_s = 99999; /* about 27 hours */ -static U32 g_nbIterations = NBLOOPS; -static double g_compressibility = COMPRESSIBILITY_DEFAULT; -static U32 g_blockSize = 0; -static U32 g_rand = 1; -static U32 g_singleRun = 0; -static U32 g_target = 0; -static U32 g_noSeed = 0; -static ZSTD_compressionParameters g_params = { 0, 0, 0, 0, 0, 0, ZSTD_greedy }; +/* indices for each of the variables */ +typedef enum { + wlog_ind = 0, + clog_ind = 1, + hlog_ind = 2, + slog_ind = 3, + slen_ind = 4, + tlen_ind = 5, + strt_ind = 6, + fadt_ind = 7, /* forceAttachDict */ + NUM_PARAMS = 8 +} varInds_t; -void BMK_SetNbIterations(int nbLoops) -{ - g_nbIterations = nbLoops; - DISPLAY("- %u iterations -\n", g_nbIterations); +typedef struct { + U32 vals[NUM_PARAMS]; +} paramValues_t; + +/* maximum value of parameters */ +static const U32 mintable[NUM_PARAMS] = + { ZSTD_WINDOWLOG_MIN, ZSTD_CHAINLOG_MIN, ZSTD_HASHLOG_MIN, ZSTD_SEARCHLOG_MIN, ZSTD_SEARCHLENGTH_MIN, ZSTD_TARGETLENGTH_MIN, ZSTD_fast, FADT_MIN }; + +/* minimum value of parameters */ +static const U32 maxtable[NUM_PARAMS] = + { ZSTD_WINDOWLOG_MAX, ZSTD_CHAINLOG_MAX, ZSTD_HASHLOG_MAX, ZSTD_SEARCHLOG_MAX, ZSTD_SEARCHLENGTH_MAX, ZSTD_TARGETLENGTH_MAX, ZSTD_btultra, FADT_MAX }; + +/* # of values parameters can take on */ +static const U32 rangetable[NUM_PARAMS] = + { WLOG_RANGE, CLOG_RANGE, HLOG_RANGE, SLOG_RANGE, SLEN_RANGE, TLEN_RANGE, STRT_RANGE, FADT_RANGE }; + +/* ZSTD_cctxSetParameter() index to set */ +static const ZSTD_cParameter cctxSetParamTable[NUM_PARAMS] = + { ZSTD_p_windowLog, ZSTD_p_chainLog, ZSTD_p_hashLog, ZSTD_p_searchLog, ZSTD_p_minMatch, ZSTD_p_targetLength, ZSTD_p_compressionStrategy, ZSTD_p_forceAttachDict }; + +/* names of parameters */ +static const char* g_paramNames[NUM_PARAMS] = + { "windowLog", "chainLog", "hashLog","searchLog", "searchLength", "targetLength", "strategy", "forceAttachDict" }; + +/* shortened names of parameters */ +static const char* g_shortParamNames[NUM_PARAMS] = + { "wlog", "clog", "hlog","slog", "slen", "tlen", "strt", "fadt" }; + +/* maps value from { 0 to rangetable[param] - 1 } to valid paramvalues */ +static U32 rangeMap(varInds_t param, int ind) { + ind = MAX(MIN(ind, (int)rangetable[param] - 1), 0); + switch(param) { + case tlen_ind: + return tlen_table[ind]; + case fadt_ind: /* 0, 1, 2 -> -1, 0, 1 */ + return ind - 1; + case wlog_ind: /* using default: triggers -Wswitch-enum */ + case clog_ind: + case hlog_ind: + case slog_ind: + case slen_ind: + case strt_ind: + return mintable[param] + ind; + case NUM_PARAMS: + DISPLAY("Error, not a valid param\n "); + return (U32)-1; + } + return 0; /* should never happen, stop compiler warnings */ } +/* inverse of rangeMap */ +static int invRangeMap(varInds_t param, U32 value) { + value = MIN(MAX(mintable[param], value), maxtable[param]); + switch(param) { + case tlen_ind: /* bin search */ + { + int lo = 0; + int hi = TLEN_RANGE; + while(lo < hi) { + int mid = (lo + hi) / 2; + if(tlen_table[mid] < value) { + lo = mid + 1; + } if(tlen_table[mid] == value) { + return mid; + } else { + hi = mid; + } + } + return lo; + } + case fadt_ind: + return (int)value + 1; + case wlog_ind: + case clog_ind: + case hlog_ind: + case slog_ind: + case slen_ind: + case strt_ind: + return value - mintable[param]; + case NUM_PARAMS: + DISPLAY("Error, not a valid param\n "); + return -2; + } + return 0; /* should never happen, stop compiler warnings */ +} + +/* display of params */ +static void displayParamVal(FILE* f, varInds_t param, U32 value, int width) { + switch(param) { + case fadt_ind: if(width) { fprintf(f, "%*d", width, (int)value); } else { fprintf(f, "%d", (int)value); } break; + case strt_ind: if(width) { fprintf(f, "%*s", width, g_stratName[value]); } else { fprintf(f, "%s", g_stratName[value]); } break; + case wlog_ind: + case clog_ind: + case hlog_ind: + case slog_ind: + case slen_ind: + case tlen_ind: if(width) { fprintf(f, "%*u", width, value); } else { fprintf(f, "%u", value); } break; + case NUM_PARAMS: + DISPLAY("Error, not a valid param\n "); break; + } +} + +/*-************************************ +* Benchmark Parameters/Global Variables +**************************************/ + +typedef BYTE U8; + +/* General Utility */ +static U32 g_timeLimit_s = 99999; /* about 27 hours */ +static UTIL_time_t g_time; /* to be used to compare solution finding speeds to compare to original */ +static U32 g_blockSize = 0; +static U32 g_rand = 1; + +/* Display */ +static int g_displayLevel = 3; +static BYTE g_silenceParams[NUM_PARAMS]; + +/* Mode Selection */ +static U32 g_singleRun = 0; +static U32 g_optimizer = 0; +static int g_optmode = 0; + +/* For cLevel Table generation */ +static U32 g_target = 0; +static U32 g_noSeed = 0; + +/* For optimizer */ +static paramValues_t g_params; /* Initialized at the beginning of main w/ emptyParams() function */ +static double g_ratioMultiplier = 5.; +static U32 g_strictness = PARAM_UNSET; /* range 1 - 100, measure of how strict */ +static BMK_result_t g_lvltarget; + +typedef enum { + directMap, + xxhashMap, + noMemo +} memoTableType_t; + +typedef struct { + memoTableType_t tableType; + BYTE* table; + size_t tableLen; + varInds_t varArray[NUM_PARAMS]; + size_t varLen; +} memoTable_t; + +typedef struct { + BMK_result_t result; + paramValues_t params; +} winnerInfo_t; + +typedef struct { + U32 cSpeed; /* bytes / sec */ + U32 dSpeed; + U32 cMem; /* bytes */ +} constraint_t; + +typedef struct winner_ll_node winner_ll_node; +struct winner_ll_node { + winnerInfo_t res; + winner_ll_node* next; +}; + +static winner_ll_node* g_winners; /* linked list sorted ascending by cSize & cSpeed */ + +/* + * Additional Global Variables (Defined Above Use) + * g_level_constraint + * g_alreadyTested + * g_maxTries + * g_clockGranularity + */ + /*-******************************************************* -* Private functions +* General Util Functions *********************************************************/ -/* accuracy in seconds only, span can be multiple years */ -static double BMK_timeSpan(time_t tStart) { return difftime(time(NULL), tStart); } +/* nullified useless params, to ensure count stats */ +/* cleans up params for memoizing / display */ +static paramValues_t sanitizeParams(paramValues_t params) +{ + if (params.vals[strt_ind] == ZSTD_fast) + params.vals[clog_ind] = 0, params.vals[slog_ind] = 0; + if (params.vals[strt_ind] == ZSTD_dfast) + params.vals[slog_ind] = 0; + if (params.vals[strt_ind] != ZSTD_btopt && params.vals[strt_ind] != ZSTD_btultra && params.vals[strt_ind] != ZSTD_fast) + params.vals[tlen_ind] = 0; + + return params; +} + +static ZSTD_compressionParameters pvalsToCParams(paramValues_t p) { + ZSTD_compressionParameters c; + memset(&c, 0, sizeof(ZSTD_compressionParameters)); + c.windowLog = p.vals[wlog_ind]; + c.chainLog = p.vals[clog_ind]; + c.hashLog = p.vals[hlog_ind]; + c.searchLog = p.vals[slog_ind]; + c.searchLength = p.vals[slen_ind]; + c.targetLength = p.vals[tlen_ind]; + c.strategy = p.vals[strt_ind]; + /* no forceAttachDict */ + return c; +} + +static paramValues_t cParamsToPVals(ZSTD_compressionParameters c) { + paramValues_t p; + varInds_t i; + p.vals[wlog_ind] = c.windowLog; + p.vals[clog_ind] = c.chainLog; + p.vals[hlog_ind] = c.hashLog; + p.vals[slog_ind] = c.searchLog; + p.vals[slen_ind] = c.searchLength; + p.vals[tlen_ind] = c.targetLength; + p.vals[strt_ind] = c.strategy; + + /* set all other params to their minimum value */ + for(i = strt_ind + 1; i < NUM_PARAMS; i++) { + p.vals[i] = mintable[i]; + } + return p; +} + +/* equivalent of ZSTD_adjustCParams for paramValues_t */ +static paramValues_t adjustParams(paramValues_t p, const size_t maxBlockSize, const size_t dictSize) { + paramValues_t ot = p; + varInds_t i; + p = cParamsToPVals(ZSTD_adjustCParams(pvalsToCParams(p), maxBlockSize, dictSize)); + if(!dictSize) { p.vals[fadt_ind] = 0; } + /* retain value of all other parameters */ + for(i = strt_ind + 1; i < NUM_PARAMS; i++) { + p.vals[i] = ot.vals[i]; + } + return p; +} static size_t BMK_findMaxMem(U64 requiredMem) { @@ -98,16 +355,18 @@ static size_t BMK_findMaxMem(U64 requiredMem) requiredMem = (((requiredMem >> 26) + 1) << 26); if (requiredMem > maxMemory) requiredMem = maxMemory; - requiredMem += 2*step; - while (!testmem) { - requiredMem -= step; + requiredMem += 2 * step; + while (!testmem && requiredMem > 0) { testmem = malloc ((size_t)requiredMem); + requiredMem -= step; } free (testmem); - return (size_t) (requiredMem - step); + return (size_t) requiredMem; } +/* accuracy in seconds only, span can be multiple years */ +static U32 BMK_timeSpan(const UTIL_time_t tStart) { return (U32)(UTIL_clockSpanMicro(tStart) / 1000000ULL); } static U32 FUZ_rotl32(U32 x, U32 r) { @@ -126,75 +385,467 @@ U32 FUZ_rand(U32* src) return rand32 >> 5; } -/** longCommandWArg() : - * check if *stringPtr is the same as longCommand. - * If yes, @return 1 and advances *stringPtr to the position which immediately follows longCommand. - * @return 0 and doesn't modify *stringPtr otherwise. - * from zstdcli.c +/* allows zeros */ +#define CLAMPCHECK(val,min,max) { \ + if (((val)<(min)) | ((val)>(max))) { \ + DISPLAY("INVALID PARAMETER CONSTRAINTS\n"); \ + return 0; \ +} } + +static int paramValid(const paramValues_t paramTarget) { + U32 i; + for(i = 0; i < NUM_PARAMS; i++) { + CLAMPCHECK(paramTarget.vals[i], mintable[i], maxtable[i]); + } + return 1; +} + +static paramValues_t cParamUnsetMin(paramValues_t paramTarget) { + varInds_t i; + for(i = 0; i < NUM_PARAMS; i++) { + if(paramTarget.vals[i] == PARAM_UNSET) { + paramTarget.vals[i] = mintable[i]; + } + } + return paramTarget; +} + +static paramValues_t emptyParams(void) { + U32 i; + paramValues_t p; + for(i = 0; i < NUM_PARAMS; i++) { + p.vals[i] = PARAM_UNSET; + } + return p; +} + +static winnerInfo_t initWinnerInfo(const paramValues_t p) { + winnerInfo_t w1; + w1.result.cSpeed = 0.; + w1.result.dSpeed = 0.; + w1.result.cMem = (size_t)-1; + w1.result.cSize = (size_t)-1; + w1.params = p; + return w1; +} + +static paramValues_t overwriteParams(paramValues_t base, const paramValues_t mask) { + U32 i; + for(i = 0; i < NUM_PARAMS; i++) { + if(mask.vals[i] != PARAM_UNSET) { + base.vals[i] = mask.vals[i]; + } + } + return base; +} + +static void paramVaryOnce(const varInds_t paramIndex, const int amt, paramValues_t* ptr) { + ptr->vals[paramIndex] = rangeMap(paramIndex, invRangeMap(paramIndex, ptr->vals[paramIndex]) + amt); +} + +/* varies ptr by nbChanges respecting varyParams*/ +static void paramVariation(paramValues_t* ptr, memoTable_t* mtAll, const U32 nbChanges) +{ + paramValues_t p; + U32 validated = 0; + while (!validated) { + U32 i; + p = *ptr; + for (i = 0 ; i < nbChanges ; i++) { + const U32 changeID = (U32)FUZ_rand(&g_rand) % (mtAll[p.vals[strt_ind]].varLen << 1); + paramVaryOnce(mtAll[p.vals[strt_ind]].varArray[changeID >> 1], ((changeID & 1) << 1) - 1, &p); + } + validated = paramValid(p); + } + *ptr = p; +} + +/* Completely random parameter selection */ +static paramValues_t randomParams(void) +{ + varInds_t v; paramValues_t p; + for(v = 0; v <= NUM_PARAMS; v++) { + p.vals[v] = rangeMap(v, FUZ_rand(&g_rand) % rangetable[v]); + } + return p; +} + +static U64 g_clockGranularity = 100000000ULL; + +static void findClockGranularity(void) { + UTIL_time_t clockStart = UTIL_getTime(); + U64 el1 = 0, el2 = 0; + int i = 0; + do { + el1 = el2; + el2 = UTIL_clockSpanNano(clockStart); + if(el1 < el2) { + U64 iv = el2 - el1; + if(g_clockGranularity > iv) { + g_clockGranularity = iv; + i = 0; + } else { + i++; + } + } + } while(i < 10); + DEBUGOUTPUT("Granularity: %llu\n", (unsigned long long)g_clockGranularity); +} + +/*-************************************ +* Optimizer Util Functions +**************************************/ + +/* checks results are feasible */ +static int feasible(const BMK_result_t results, const constraint_t target) { + return (results.cSpeed >= target.cSpeed) && (results.dSpeed >= target.dSpeed) && (results.cMem <= target.cMem) && (!g_optmode || results.cSize <= g_lvltarget.cSize); +} + +/* hill climbing value for part 1 */ +/* Scoring here is a linear reward for all set constraints normalized between 0 to 1 + * (with 0 at 0 and 1 being fully fulfilling the constraint), summed with a logarithmic + * bonus to exceeding the constraint value. We also give linear ratio for compression ratio. + * The constant factors are experimental. */ -static unsigned longCommandWArg(const char** stringPtr, const char* longCommand) -{ - size_t const comSize = strlen(longCommand); - int const result = !strncmp(*stringPtr, longCommand, comSize); - if (result) *stringPtr += comSize; - return result; +static double resultScore(const BMK_result_t res, const size_t srcSize, const constraint_t target) { + double cs = 0., ds = 0., rt, cm = 0.; + const double r1 = 1, r2 = 0.1, rtr = 0.5; + double ret; + if(target.cSpeed) { cs = res.cSpeed / (double)target.cSpeed; } + if(target.dSpeed) { ds = res.dSpeed / (double)target.dSpeed; } + if(target.cMem != (U32)-1) { cm = (double)target.cMem / res.cMem; } + rt = ((double)srcSize / res.cSize); + + ret = (MIN(1, cs) + MIN(1, ds) + MIN(1, cm))*r1 + rt * rtr + + (MAX(0, log(cs))+ MAX(0, log(ds))+ MAX(0, log(cm))) * r2; + + return ret; } -/*-******************************************************* -* Bench functions -*********************************************************/ - -typedef struct -{ - const char* srcPtr; - size_t srcSize; - char* cPtr; - size_t cRoom; - size_t cSize; - char* resPtr; - size_t resSize; -} blockParam_t; - - -const char* g_stratName[ZSTD_btultra+1] = { - "(none) ", "ZSTD_fast ", "ZSTD_dfast ", - "ZSTD_greedy ", "ZSTD_lazy ", "ZSTD_lazy2 ", - "ZSTD_btlazy2 ", "ZSTD_btopt ", "ZSTD_btultra "}; - -/* TODO: support additional parameters (more files, fileSizes) */ -static size_t -BMK_benchParam(BMK_result_t* resultPtr, - const void* srcBuffer, size_t srcSize, - ZSTD_CCtx* ctx, ZSTD_DCtx* dctx, - const ZSTD_compressionParameters cParams) { - - BMK_return_t res = BMK_benchMem(srcBuffer,srcSize, &srcSize, 1, 0, &cParams, NULL, 0, ctx, dctx, 0, "File"); - *resultPtr = res.result; - return res.error; +/* calculates normalized squared euclidean distance of result1 if it is in the first quadrant relative to lvlRes */ +static double resultDistLvl(const BMK_result_t result1, const BMK_result_t lvlRes) { + double normalizedCSpeedGain1 = (result1.cSpeed / lvlRes.cSpeed) - 1; + double normalizedRatioGain1 = ((double)lvlRes.cSize / result1.cSize) - 1; + if(normalizedRatioGain1 < 0 || normalizedCSpeedGain1 < 0) { + return 0.0; + } + return normalizedRatioGain1 * g_ratioMultiplier + normalizedCSpeedGain1; } -static void BMK_printWinner(FILE* f, U32 cLevel, BMK_result_t result, ZSTD_compressionParameters params, size_t srcSize) +/* return true if r2 strictly better than r1 */ +static int compareResultLT(const BMK_result_t result1, const BMK_result_t result2, const constraint_t target, size_t srcSize) { + if(feasible(result1, target) && feasible(result2, target)) { + if(g_optmode) { + return resultDistLvl(result1, g_lvltarget) < resultDistLvl(result2, g_lvltarget); + } else { + return (result1.cSize > result2.cSize) || (result1.cSize == result2.cSize && result2.cSpeed > result1.cSpeed) + || (result1.cSize == result2.cSize && result2.cSpeed == result1.cSpeed && result2.dSpeed > result1.dSpeed); + } + } + return feasible(result2, target) || (!feasible(result1, target) && (resultScore(result1, srcSize, target) < resultScore(result2, srcSize, target))); +} + +static constraint_t relaxTarget(constraint_t target) { + target.cMem = (U32)-1; + target.cSpeed *= ((double)g_strictness) / 100; + target.dSpeed *= ((double)g_strictness) / 100; + return target; +} + +static void optimizerAdjustInput(paramValues_t* pc, const size_t maxBlockSize) { + varInds_t v; + for(v = 0; v < NUM_PARAMS; v++) { + if(pc->vals[v] != PARAM_UNSET) { + U32 newval = MIN(MAX(pc->vals[v], mintable[v]), maxtable[v]); + if(newval != pc->vals[v]) { + pc->vals[v] = newval; + DISPLAY("Warning: parameter %s not in valid range, adjusting to ", g_paramNames[v]); displayParamVal(stderr, v, newval, 0); DISPLAY("\n"); + } + } + } + + if(pc->vals[wlog_ind] != PARAM_UNSET) { + + U32 sshb = maxBlockSize > 1 ? ZSTD_highbit32((U32)(maxBlockSize-1)) + 1 : 1; + /* edge case of highBit not working for 0 */ + + if(maxBlockSize < (1ULL << 31) && sshb + 1 < pc->vals[wlog_ind]) { + U32 adjust = MAX(mintable[wlog_ind], sshb); + if(adjust != pc->vals[wlog_ind]) { + pc->vals[wlog_ind] = adjust; + DISPLAY("Warning: windowLog larger than src/block size, adjusted to %u\n", pc->vals[wlog_ind]); + } + } + } + + if(pc->vals[wlog_ind] != PARAM_UNSET && pc->vals[clog_ind] != PARAM_UNSET) { + U32 maxclog; + if(pc->vals[strt_ind] == PARAM_UNSET || pc->vals[strt_ind] >= (U32)ZSTD_btlazy2) { + maxclog = pc->vals[wlog_ind] + 1; + } else { + maxclog = pc->vals[wlog_ind]; + } + + if(pc->vals[clog_ind] > maxclog) { + pc->vals[clog_ind] = maxclog; + DISPLAY("Warning: chainlog too much larger than windowLog size, adjusted to %u\n", pc->vals[clog_ind]); + } + } + + if(pc->vals[wlog_ind] != PARAM_UNSET && pc->vals[hlog_ind] != PARAM_UNSET) { + if(pc->vals[wlog_ind] + 1 < pc->vals[hlog_ind]) { + pc->vals[hlog_ind] = pc->vals[wlog_ind] + 1; + DISPLAY("Warning: hashlog too much larger than windowLog size, adjusted to %u\n", pc->vals[hlog_ind]); + } + } + + if(pc->vals[slog_ind] != PARAM_UNSET && pc->vals[clog_ind] != PARAM_UNSET) { + if(pc->vals[slog_ind] > pc->vals[clog_ind]) { + pc->vals[clog_ind] = pc->vals[slog_ind]; + DISPLAY("Warning: searchLog larger than chainLog, adjusted to %u\n", pc->vals[slog_ind]); + } + } +} + +static int redundantParams(const paramValues_t paramValues, const constraint_t target, const size_t maxBlockSize) { + return + (ZSTD_estimateCStreamSize_usingCParams(pvalsToCParams(paramValues)) > (size_t)target.cMem) /* Uses too much memory */ + || ((1ULL << (paramValues.vals[wlog_ind] - 1)) >= maxBlockSize && paramValues.vals[wlog_ind] != mintable[wlog_ind]) /* wlog too much bigger than src size */ + || (paramValues.vals[clog_ind] > (paramValues.vals[wlog_ind] + (paramValues.vals[strt_ind] > ZSTD_btlazy2))) /* chainLog larger than windowLog*/ + || (paramValues.vals[slog_ind] > paramValues.vals[clog_ind]) /* searchLog larger than chainLog */ + || (paramValues.vals[hlog_ind] > paramValues.vals[wlog_ind] + 1); /* hashLog larger than windowLog + 1 */ + +} + +/*-************************************ +* Display Functions +**************************************/ + +static void BMK_translateAdvancedParams(FILE* f, const paramValues_t params) { + varInds_t v; + int first = 1; + fprintf(f,"--zstd="); + for(v = 0; v < NUM_PARAMS; v++) { + if(g_silenceParams[v]) { continue; } + if(!first) { fprintf(f, ","); } + fprintf(f,"%s=", g_paramNames[v]); + + if(v == strt_ind) { fprintf(f,"%u", params.vals[v]); } + else { displayParamVal(f, v, params.vals[v], 0); } + first = 0; + } + fprintf(f, "\n"); +} + +static void BMK_displayOneResult(FILE* f, winnerInfo_t res, const size_t srcSize) { + varInds_t v; + int first = 1; + res.params = cParamUnsetMin(res.params); + fprintf(f," {"); + for(v = 0; v < NUM_PARAMS; v++) { + if(g_silenceParams[v]) { continue; } + if(!first) { fprintf(f, ","); } + displayParamVal(f, v, res.params.vals[v], 3); + first = 0; + } + + fprintf(f, " }, /* R:%5.3f at %5.1f MB/s - %5.1f MB/s */\n", + (double)srcSize / res.result.cSize, (double)res.result.cSpeed / (1 MB), (double)res.result.dSpeed / (1 MB)); +} + +/* Writes to f the results of a parameter benchmark */ +/* when used with --optimize, will only print results better than previously discovered */ +static void BMK_printWinner(FILE* f, const int cLevel, const BMK_result_t result, const paramValues_t params, const size_t srcSize) { char lvlstr[15] = "Custom Level"; - DISPLAY("\r%79s\r", ""); - fprintf(f," {%3u,%3u,%3u,%3u,%3u,%3u, %s }, ", - params.windowLog, params.chainLog, params.hashLog, params.searchLog, params.searchLength, - params.targetLength, g_stratName[(U32)(params.strategy)]); + winnerInfo_t w; + w.params = params; + w.result = result; + + fprintf(f, "\r%79s\r", ""); + if(cLevel != CUSTOM_LEVEL) { - snprintf(lvlstr, 15, " Level %2u ", cLevel); + snprintf(lvlstr, 15, " Level %2d ", cLevel); } - fprintf(f, - "/* %s */ /* R:%5.3f at %5.1f MB/s - %5.1f MB/s */\n", - lvlstr, (double)srcSize / result.cSize, result.cSpeed / 1000000., result.dSpeed / 1000000.); + + if(TIMED) { + const U64 time = UTIL_clockSpanNano(g_time); + const U64 minutes = time / (60ULL * TIMELOOP_NANOSEC); + fprintf(f, "%1lu:%2lu:%05.2f - ", (unsigned long) minutes / 60,(unsigned long) minutes % 60, (double)(time - minutes * TIMELOOP_NANOSEC * 60ULL)/TIMELOOP_NANOSEC); + } + + fprintf(f, "/* %s */ ", lvlstr); + BMK_displayOneResult(f, w, srcSize); } +/* comparison function: */ +/* strictly better, strictly worse, equal, speed-side adv, size-side adv */ +#define WORSE_RESULT 0 +#define BETTER_RESULT 1 +#define ERROR_RESULT 2 -typedef struct { - BMK_result_t result; - ZSTD_compressionParameters params; -} winnerInfo_t; +#define SPEED_RESULT 4 +#define SIZE_RESULT 5 +/* maybe have epsilon-eq to limit table size? */ +static int speedSizeCompare(const BMK_result_t r1, const BMK_result_t r2) { + if(r1.cSpeed < r2.cSpeed) { + if(r1.cSize >= r2.cSize) { + return BETTER_RESULT; + } + return SPEED_RESULT; /* r2 is smaller but not faster. */ + } else { + if(r1.cSize <= r2.cSize) { + return WORSE_RESULT; + } + return SIZE_RESULT; /* r2 is faster but not smaller */ + } +} -static void BMK_printWinners2(FILE* f, const winnerInfo_t* winners, size_t srcSize) +/* 0 for insertion, 1 for no insert */ +/* maintain invariant speedSizeCompare(n, n->next) = SPEED_RESULT */ +static int insertWinner(const winnerInfo_t w, const constraint_t targetConstraints) { + BMK_result_t r = w.result; + winner_ll_node* cur_node = g_winners; + /* first node to insert */ + if(!feasible(r, targetConstraints)) { + return 1; + } + + if(g_winners == NULL) { + winner_ll_node* first_node = malloc(sizeof(winner_ll_node)); + if(first_node == NULL) { + return 1; + } + first_node->next = NULL; + first_node->res = w; + g_winners = first_node; + return 0; + } + + while(cur_node->next != NULL) { + switch(speedSizeCompare(cur_node->res.result, r)) { + case WORSE_RESULT: + { + return 1; /* never insert if better */ + } + case BETTER_RESULT: + { + winner_ll_node* tmp; + cur_node->res = cur_node->next->res; + tmp = cur_node->next; + cur_node->next = cur_node->next->next; + free(tmp); + break; + } + case SIZE_RESULT: + { + cur_node = cur_node->next; + break; + } + case SPEED_RESULT: /* insert after first size result, then return */ + { + winner_ll_node* newnode = malloc(sizeof(winner_ll_node)); + if(newnode == NULL) { + return 1; + } + newnode->res = cur_node->res; + cur_node->res = w; + newnode->next = cur_node->next; + cur_node->next = newnode; + return 0; + } + } + + } + + assert(cur_node->next == NULL); + switch(speedSizeCompare(cur_node->res.result, r)) { + case WORSE_RESULT: + { + return 1; /* never insert if better */ + } + case BETTER_RESULT: + { + cur_node->res = w; + return 0; + } + case SIZE_RESULT: + { + winner_ll_node* newnode = malloc(sizeof(winner_ll_node)); + if(newnode == NULL) { + return 1; + } + newnode->res = w; + newnode->next = NULL; + cur_node->next = newnode; + return 0; + } + case SPEED_RESULT: /* insert before first size result, then return */ + { + winner_ll_node* newnode = malloc(sizeof(winner_ll_node)); + if(newnode == NULL) { + return 1; + } + newnode->res = cur_node->res; + cur_node->res = w; + newnode->next = cur_node->next; + cur_node->next = newnode; + return 0; + } + default: + return 1; + } +} + +static void BMK_printWinnerOpt(FILE* f, const U32 cLevel, const BMK_result_t result, const paramValues_t params, const constraint_t targetConstraints, const size_t srcSize) +{ + /* global winner used for constraints */ + /* cSize, cSpeed, dSpeed, cMem */ + static winnerInfo_t g_winner = { { (size_t)-1LL, 0, 0, (size_t)-1LL }, { { PARAM_UNSET, PARAM_UNSET, PARAM_UNSET, PARAM_UNSET, PARAM_UNSET, PARAM_UNSET, PARAM_UNSET, PARAM_UNSET } } }; + if(DEBUG || compareResultLT(g_winner.result, result, targetConstraints, srcSize) || g_displayLevel >= 4) { + if(DEBUG && compareResultLT(g_winner.result, result, targetConstraints, srcSize)) { + DISPLAY("New Winner: \n"); + } + + if(g_displayLevel >= 2) { BMK_printWinner(f, cLevel, result, params, srcSize); } + + if(compareResultLT(g_winner.result, result, targetConstraints, srcSize)) { + if(g_displayLevel >= 1) { BMK_translateAdvancedParams(f, params); } + g_winner.result = result; + g_winner.params = params; + } + } + + if(g_optmode && g_optimizer && (DEBUG || g_displayLevel == 3)) { + winnerInfo_t w; + winner_ll_node* n; + w.result = result; + w.params = params; + insertWinner(w, targetConstraints); + + if(!DEBUG) { fprintf(f, "\033c"); } + fprintf(f, "\n"); + + /* the table */ + fprintf(f, "================================\n"); + for(n = g_winners; n != NULL; n = n->next) { + BMK_displayOneResult(f, n->res, srcSize); + } + fprintf(f, "================================\n"); + fprintf(f, "Level Bounds: R: > %.3f AND C: < %.1f MB/s \n\n", + (double)srcSize / g_lvltarget.cSize, (double)g_lvltarget.cSpeed / (1 MB)); + + + fprintf(f, "Overall Winner: \n"); + BMK_displayOneResult(f, g_winner, srcSize); + BMK_translateAdvancedParams(f, g_winner.params); + + fprintf(f, "Latest BMK: \n");\ + BMK_displayOneResult(f, w, srcSize); + } +} + +static void BMK_printWinners2(FILE* f, const winnerInfo_t* winners, const size_t srcSize) { int cLevel; @@ -206,7 +857,7 @@ static void BMK_printWinners2(FILE* f, const winnerInfo_t* winners, size_t srcSi } -static void BMK_printWinners(FILE* f, const winnerInfo_t* winners, size_t srcSize) +static void BMK_printWinners(FILE* f, const winnerInfo_t* winners, const size_t srcSize) { fseek(f, 0, SEEK_SET); BMK_printWinners2(f, winners, srcSize); @@ -215,9 +866,774 @@ static void BMK_printWinners(FILE* f, const winnerInfo_t* winners, size_t srcSiz } +/*-******************************************************* +* Functions to Benchmark +*********************************************************/ + typedef struct { - double cSpeed_min; - double dSpeed_min; + ZSTD_CCtx* ctx; + const void* dictBuffer; + size_t dictBufferSize; + int cLevel; + const paramValues_t* comprParams; +} BMK_initCCtxArgs; + +static size_t local_initCCtx(void* payload) { + const BMK_initCCtxArgs* ag = (const BMK_initCCtxArgs*)payload; + varInds_t i; + ZSTD_CCtx_reset(ag->ctx); + ZSTD_CCtx_resetParameters(ag->ctx); + ZSTD_CCtx_setParameter(ag->ctx, ZSTD_p_compressionLevel, ag->cLevel); + + for(i = 0; i < NUM_PARAMS; i++) { + if(ag->comprParams->vals[i] != PARAM_UNSET) + ZSTD_CCtx_setParameter(ag->ctx, cctxSetParamTable[i], ag->comprParams->vals[i]); + } + ZSTD_CCtx_loadDictionary(ag->ctx, ag->dictBuffer, ag->dictBufferSize); + + return 0; +} + +typedef struct { + ZSTD_DCtx* dctx; + const void* dictBuffer; + size_t dictBufferSize; +} BMK_initDCtxArgs; + +static size_t local_initDCtx(void* payload) { + const BMK_initDCtxArgs* ag = (const BMK_initDCtxArgs*)payload; + ZSTD_DCtx_reset(ag->dctx); + ZSTD_DCtx_loadDictionary(ag->dctx, ag->dictBuffer, ag->dictBufferSize); + return 0; +} + +/* additional argument is just the context */ +static size_t local_defaultCompress( + const void* srcBuffer, size_t srcSize, + void* dstBuffer, size_t dstSize, + void* addArgs) { + size_t moreToFlush = 1; + ZSTD_CCtx* ctx = (ZSTD_CCtx*)addArgs; + ZSTD_inBuffer in; + ZSTD_outBuffer out; + in.src = srcBuffer; + in.size = srcSize; + in.pos = 0; + out.dst = dstBuffer; + out.size = dstSize; + out.pos = 0; + assert(dstSize == ZSTD_compressBound(srcSize)); /* specific to this version, which is only used in paramgrill */ + while (moreToFlush) { + if(out.pos == out.size) { + return (size_t)-ZSTD_error_dstSize_tooSmall; + } + moreToFlush = ZSTD_compress_generic(ctx, &out, &in, ZSTD_e_end); + if (ZSTD_isError(moreToFlush)) { + return moreToFlush; + } + } + return out.pos; +} + +/* additional argument is just the context */ +static size_t local_defaultDecompress( + const void* srcBuffer, size_t srcSize, + void* dstBuffer, size_t dstSize, + void* addArgs) { + size_t moreToFlush = 1; + ZSTD_DCtx* dctx = (ZSTD_DCtx*)addArgs; + ZSTD_inBuffer in; + ZSTD_outBuffer out; + in.src = srcBuffer; + in.size = srcSize; + in.pos = 0; + out.dst = dstBuffer; + out.size = dstSize; + out.pos = 0; + while (moreToFlush) { + if(out.pos == out.size) { + return (size_t)-ZSTD_error_dstSize_tooSmall; + } + moreToFlush = ZSTD_decompress_generic(dctx, + &out, &in); + if (ZSTD_isError(moreToFlush)) { + return moreToFlush; + } + } + return out.pos; + +} + +/*-************************************ +* Data Initialization Functions +**************************************/ + +typedef struct { + void* srcBuffer; + size_t srcSize; + const void** srcPtrs; + size_t* srcSizes; + void** dstPtrs; + size_t* dstCapacities; + size_t* dstSizes; + void** resPtrs; + size_t* resSizes; + size_t nbBlocks; + size_t maxBlockSize; +} buffers_t; + +typedef struct { + size_t dictSize; + void* dictBuffer; + ZSTD_CCtx* cctx; + ZSTD_DCtx* dctx; +} contexts_t; + +static void freeNonSrcBuffers(const buffers_t b) { + free(b.srcPtrs); + free(b.srcSizes); + + if(b.dstPtrs != NULL) { + free(b.dstPtrs[0]); + } + free(b.dstPtrs); + free(b.dstCapacities); + free(b.dstSizes); + + if(b.resPtrs != NULL) { + free(b.resPtrs[0]); + } + free(b.resPtrs); + free(b.resSizes); +} + +static void freeBuffers(const buffers_t b) { + if(b.srcPtrs != NULL) { + free(b.srcBuffer); + } + freeNonSrcBuffers(b); +} + +/* srcBuffer will be freed by freeBuffers now */ +static int createBuffersFromMemory(buffers_t* buff, void * srcBuffer, const size_t nbFiles, + const size_t* fileSizes) +{ + size_t pos = 0, n, blockSize; + U32 maxNbBlocks, blockNb = 0; + buff->srcSize = 0; + for(n = 0; n < nbFiles; n++) { + buff->srcSize += fileSizes[n]; + } + + if(buff->srcSize == 0) { + DISPLAY("No data to bench\n"); + return 1; + } + + blockSize = g_blockSize ? g_blockSize : buff->srcSize; + maxNbBlocks = (U32) ((buff->srcSize + (blockSize-1)) / blockSize) + (U32)nbFiles; + + buff->srcPtrs = (const void**)calloc(maxNbBlocks, sizeof(void*)); + buff->srcSizes = (size_t*)malloc(maxNbBlocks * sizeof(size_t)); + + buff->dstPtrs = (void**)calloc(maxNbBlocks, sizeof(void*)); + buff->dstCapacities = (size_t*)malloc(maxNbBlocks * sizeof(size_t)); + buff->dstSizes = (size_t*)malloc(maxNbBlocks * sizeof(size_t)); + + buff->resPtrs = (void**)calloc(maxNbBlocks, sizeof(void*)); + buff->resSizes = (size_t*)malloc(maxNbBlocks * sizeof(size_t)); + + if(!buff->srcPtrs || !buff->srcSizes || !buff->dstPtrs || !buff->dstCapacities || !buff->dstSizes || !buff->resPtrs || !buff->resSizes) { + DISPLAY("alloc error\n"); + freeNonSrcBuffers(*buff); + return 1; + } + + buff->srcBuffer = srcBuffer; + buff->srcPtrs[0] = (const void*)buff->srcBuffer; + buff->dstPtrs[0] = malloc(ZSTD_compressBound(buff->srcSize) + (maxNbBlocks * 1024)); + buff->resPtrs[0] = malloc(buff->srcSize); + + if(!buff->dstPtrs[0] || !buff->resPtrs[0]) { + DISPLAY("alloc error\n"); + freeNonSrcBuffers(*buff); + return 1; + } + + for(n = 0; n < nbFiles; n++) { + size_t pos_end = pos + fileSizes[n]; + for(; pos < pos_end; blockNb++) { + buff->srcPtrs[blockNb] = (const void*)((char*)srcBuffer + pos); + buff->srcSizes[blockNb] = blockSize; + pos += blockSize; + } + + if(fileSizes[n] > 0) { buff->srcSizes[blockNb - 1] = ((fileSizes[n] - 1) % blockSize) + 1; } + pos = pos_end; + } + + buff->dstCapacities[0] = ZSTD_compressBound(buff->srcSizes[0]); + buff->dstSizes[0] = buff->dstCapacities[0]; + buff->resSizes[0] = buff->srcSizes[0]; + buff->maxBlockSize = buff->srcSizes[0]; + + for(n = 1; n < blockNb; n++) { + buff->dstPtrs[n] = ((char*)buff->dstPtrs[n-1]) + buff->dstCapacities[n-1]; + buff->resPtrs[n] = ((char*)buff->resPtrs[n-1]) + buff->resSizes[n-1]; + buff->dstCapacities[n] = ZSTD_compressBound(buff->srcSizes[n]); + buff->dstSizes[n] = buff->dstCapacities[n]; + buff->resSizes[n] = buff->srcSizes[n]; + + buff->maxBlockSize = MAX(buff->maxBlockSize, buff->srcSizes[n]); + } + + buff->nbBlocks = blockNb; + + return 0; +} + +/* allocates buffer's arguments. returns success / failuere */ +static int createBuffers(buffers_t* buff, const char* const * const fileNamesTable, + size_t nbFiles) { + size_t pos = 0; + size_t n; + size_t totalSizeToLoad = UTIL_getTotalFileSize(fileNamesTable, (U32)nbFiles); + size_t benchedSize = MIN(BMK_findMaxMem(totalSizeToLoad * 3) / 3, totalSizeToLoad); + size_t* fileSizes = calloc(sizeof(size_t), nbFiles); + void* srcBuffer = NULL; + int ret = 0; + + if(!totalSizeToLoad || !benchedSize) { + ret = 1; + DISPLAY("Nothing to Bench\n"); + goto _cleanUp; + } + + srcBuffer = malloc(benchedSize); + + if(!fileSizes || !srcBuffer) { + ret = 1; + goto _cleanUp; + } + + for(n = 0; n < nbFiles; n++) { + FILE* f; + U64 fileSize = UTIL_getFileSize(fileNamesTable[n]); + if (UTIL_isDirectory(fileNamesTable[n])) { + DISPLAY("Ignoring %s directory... \n", fileNamesTable[n]); + continue; + } + if (fileSize == UTIL_FILESIZE_UNKNOWN) { + DISPLAY("Cannot evaluate size of %s, ignoring ... \n", fileNamesTable[n]); + continue; + } + f = fopen(fileNamesTable[n], "rb"); + if (f==NULL) { + DISPLAY("impossible to open file %s\n", fileNamesTable[n]); + fclose(f); + ret = 10; + goto _cleanUp; + } + + DISPLAYLEVEL(2, "Loading %s... \r", fileNamesTable[n]); + + if (fileSize + pos > benchedSize) fileSize = benchedSize - pos, nbFiles=n; /* buffer too small - stop after this file */ + { + char* buffer = (char*)(srcBuffer); + size_t const readSize = fread((buffer)+pos, 1, (size_t)fileSize, f); + fclose(f); + if (readSize != (size_t)fileSize) { + DISPLAY("could not read %s", fileNamesTable[n]); + ret = 1; + goto _cleanUp; + } + + fileSizes[n] = readSize; + pos += readSize; + } + } + + ret = createBuffersFromMemory(buff, srcBuffer, nbFiles, fileSizes); + +_cleanUp: + if(ret) { free(srcBuffer); } + free(fileSizes); + return ret; +} + +static void freeContexts(const contexts_t ctx) { + free(ctx.dictBuffer); + ZSTD_freeCCtx(ctx.cctx); + ZSTD_freeDCtx(ctx.dctx); +} + +static int createContexts(contexts_t* ctx, const char* dictFileName) { + FILE* f; + size_t readSize; + ctx->cctx = ZSTD_createCCtx(); + ctx->dctx = ZSTD_createDCtx(); + if(dictFileName == NULL) { + ctx->dictSize = 0; + ctx->dictBuffer = NULL; + return 0; + } + ctx->dictSize = UTIL_getFileSize(dictFileName); + ctx->dictBuffer = malloc(ctx->dictSize); + + f = fopen(dictFileName, "rb"); + + if(!f) { + DISPLAY("unable to open file\n"); + fclose(f); + freeContexts(*ctx); + return 1; + } + + if(ctx->dictSize > 64 MB || !(ctx->dictBuffer)) { + DISPLAY("dictionary too large\n"); + fclose(f); + freeContexts(*ctx); + return 1; + } + readSize = fread(ctx->dictBuffer, 1, ctx->dictSize, f); + if(readSize != ctx->dictSize) { + DISPLAY("unable to read file\n"); + fclose(f); + freeContexts(*ctx); + return 1; + } + fclose(f); + return 0; +} + +/*-************************************ +* Optimizer Memoization Functions +**************************************/ + +/* return: new length */ +/* keep old array, will need if iter over strategy. */ +/* prunes useless params */ +static size_t sanitizeVarArray(varInds_t* varNew, const size_t varLength, const varInds_t* varArray, const ZSTD_strategy strat) { + size_t i, j = 0; + for(i = 0; i < varLength; i++) { + if( !((varArray[i] == clog_ind && strat == ZSTD_fast) + || (varArray[i] == slog_ind && strat == ZSTD_fast) + || (varArray[i] == slog_ind && strat == ZSTD_dfast) + || (varArray[i] == tlen_ind && strat != ZSTD_btopt && strat != ZSTD_btultra && strat != ZSTD_fast))) { + varNew[j] = varArray[i]; + j++; + } + } + return j; +} + +/* res should be NUM_PARAMS size */ +/* constructs varArray from paramValues_t style parameter */ +/* pass in using dict. */ +static size_t variableParams(const paramValues_t paramConstraints, varInds_t* res, const int usingDictionary) { + varInds_t i; + size_t j = 0; + for(i = 0; i < NUM_PARAMS; i++) { + if(paramConstraints.vals[i] == PARAM_UNSET) { + if(i == fadt_ind && !usingDictionary) continue; /* don't use fadt if no dictionary */ + res[j] = i; j++; + } + } + return j; +} + +/* length of memo table given free variables */ +static size_t memoTableLen(const varInds_t* varyParams, const size_t varyLen) { + size_t arrayLen = 1; + size_t i; + for(i = 0; i < varyLen; i++) { + if(varyParams[i] == strt_ind) continue; /* strategy separated by table */ + arrayLen *= rangetable[varyParams[i]]; + } + return arrayLen; +} + +/* returns unique index in memotable of compression parameters */ +static unsigned memoTableIndDirect(const paramValues_t* ptr, const varInds_t* varyParams, const size_t varyLen) { + size_t i; + unsigned ind = 0; + for(i = 0; i < varyLen; i++) { + varInds_t v = varyParams[i]; + if(v == strt_ind) continue; /* exclude strategy from memotable */ + ind *= rangetable[v]; ind += (unsigned)invRangeMap(v, ptr->vals[v]); + } + return ind; +} + +static size_t memoTableGet(const memoTable_t* memoTableArray, const paramValues_t p) { + const memoTable_t mt = memoTableArray[p.vals[strt_ind]]; + switch(mt.tableType) { + case directMap: + return mt.table[memoTableIndDirect(&p, mt.varArray, mt.varLen)]; + case xxhashMap: + return mt.table[(XXH64(&p.vals, sizeof(U32) * NUM_PARAMS, 0) >> 3) % mt.tableLen]; + case noMemo: + return 0; + } + return 0; /* should never happen, stop compiler warnings */ +} + +static void memoTableSet(const memoTable_t* memoTableArray, const paramValues_t p, const BYTE value) { + const memoTable_t mt = memoTableArray[p.vals[strt_ind]]; + switch(mt.tableType) { + case directMap: + mt.table[memoTableIndDirect(&p, mt.varArray, mt.varLen)] = value; break; + case xxhashMap: + mt.table[(XXH64(&p.vals, sizeof(U32) * NUM_PARAMS, 0) >> 3) % mt.tableLen] = value; break; + case noMemo: + break; + } +} + +/* frees all allocated memotables */ +static void freeMemoTableArray(memoTable_t* const mtAll) { + int i; + if(mtAll == NULL) { return; } + for(i = 1; i <= (int)ZSTD_btultra; i++) { + free(mtAll[i].table); + } + free(mtAll); +} + +/* inits memotables for all (including mallocs), all strategies */ +/* takes unsanitized varyParams */ +static memoTable_t* createMemoTableArray(const paramValues_t p, const varInds_t* const varyParams, const size_t varyLen, const U32 memoTableLog) { + memoTable_t* mtAll = (memoTable_t*)calloc(sizeof(memoTable_t),(ZSTD_btultra + 1)); + ZSTD_strategy i, stratMin = ZSTD_fast, stratMax = ZSTD_btultra; + + if(mtAll == NULL) { + return NULL; + } + + for(i = 1; i <= (int)ZSTD_btultra; i++) { + mtAll[i].varLen = sanitizeVarArray(mtAll[i].varArray, varyLen, varyParams, i); + } + + /* no memoization */ + if(memoTableLog == 0) { + for(i = 1; i <= (int)ZSTD_btultra; i++) { + mtAll[i].tableType = noMemo; + mtAll[i].table = NULL; + mtAll[i].tableLen = 0; + } + return mtAll; + } + + + if(p.vals[strt_ind] != PARAM_UNSET) { + stratMin = p.vals[strt_ind]; + stratMax = p.vals[strt_ind]; + } + + + for(i = stratMin; i <= stratMax; i++) { + size_t mtl = memoTableLen(mtAll[i].varArray, mtAll[i].varLen); + mtAll[i].tableType = directMap; + + if(memoTableLog != PARAM_UNSET && mtl > (1ULL << memoTableLog)) { /* use hash table */ /* provide some option to only use hash tables? */ + mtAll[i].tableType = xxhashMap; + mtl = (1ULL << memoTableLog); + } + + mtAll[i].table = (BYTE*)calloc(sizeof(BYTE), mtl); + mtAll[i].tableLen = mtl; + + if(mtAll[i].table == NULL) { + freeMemoTableArray(mtAll); + return NULL; + } + } + + return mtAll; +} + +/* Sets pc to random unmeasured set of parameters */ +/* specifiy strategy */ +static void randomConstrainedParams(paramValues_t* pc, const memoTable_t* memoTableArray, const ZSTD_strategy st) +{ + size_t j; + const memoTable_t mt = memoTableArray[st]; + pc->vals[strt_ind] = st; + for(j = 0; j < mt.tableLen; j++) { + int i; + for(i = 0; i < NUM_PARAMS; i++) { + varInds_t v = mt.varArray[i]; + if(v == strt_ind) continue; + pc->vals[v] = rangeMap(v, FUZ_rand(&g_rand) % rangetable[v]); + } + + if(!(memoTableGet(memoTableArray, *pc))) break; /* only pick unpicked params. */ + } +} + +/*-************************************ +* Benchmarking Functions +**************************************/ + +/* Replicate functionality of benchMemAdvanced, but with pre-split src / dst buffers */ +/* The purpose is so that sufficient information is returned so that a decompression call to benchMemInvertible is possible */ +/* BMK_benchMemAdvanced(srcBuffer,srcSize, dstBuffer, dstSize, fileSizes, nbFiles, 0, &cParams, dictBuffer, dictSize, ctx, dctx, 0, "File", &adv); */ +/* nbSeconds used in same way as in BMK_advancedParams_t, as nbIters when in iterMode */ + +/* if in decodeOnly, then srcPtr's will be compressed blocks, and uncompressedBlocks will be written to dstPtrs */ +/* dictionary nullable, nothing else though. */ +static BMK_return_t BMK_benchMemInvertible(const buffers_t buf, const contexts_t ctx, + const int cLevel, const paramValues_t* comprParams, + const BMK_mode_t mode, const BMK_loopMode_t loopMode, const unsigned nbSeconds) { + + U32 i; + BMK_return_t results = { { 0, 0., 0., 0 }, 0 } ; + const void *const *const srcPtrs = (const void *const *const)buf.srcPtrs; + size_t const *const srcSizes = buf.srcSizes; + void** const dstPtrs = buf.dstPtrs; + size_t const *const dstCapacities = buf.dstCapacities; + size_t* const dstSizes = buf.dstSizes; + void** const resPtrs = buf.resPtrs; + size_t const *const resSizes = buf.resSizes; + const void* dictBuffer = ctx.dictBuffer; + const size_t dictBufferSize = ctx.dictSize; + const size_t nbBlocks = buf.nbBlocks; + const size_t srcSize = buf.srcSize; + ZSTD_CCtx* cctx = ctx.cctx; + ZSTD_DCtx* dctx = ctx.dctx; + + /* warmimg up memory */ + for(i = 0; i < buf.nbBlocks; i++) { + if(mode != BMK_decodeOnly) { + RDG_genBuffer(dstPtrs[i], dstCapacities[i], 0.10, 0.50, 1); + } else { + RDG_genBuffer(resPtrs[i], resSizes[i], 0.10, 0.50, 1); + } + } + + /* Bench */ + { + /* init args */ + BMK_initCCtxArgs cctxprep; + BMK_initDCtxArgs dctxprep; + cctxprep.ctx = cctx; + cctxprep.dictBuffer = dictBuffer; + cctxprep.dictBufferSize = dictBufferSize; + cctxprep.cLevel = cLevel; + cctxprep.comprParams = comprParams; + dctxprep.dctx = dctx; + dctxprep.dictBuffer = dictBuffer; + dctxprep.dictBufferSize = dictBufferSize; + + if(loopMode == BMK_timeMode) { + BMK_customTimedReturn_t intermediateResultCompress; + BMK_customTimedReturn_t intermediateResultDecompress; + BMK_timedFnState_t* timeStateCompress = BMK_createTimeState(nbSeconds); + BMK_timedFnState_t* timeStateDecompress = BMK_createTimeState(nbSeconds); + if(mode == BMK_compressOnly) { + intermediateResultCompress.completed = 0; + intermediateResultDecompress.completed = 1; + } else if (mode == BMK_decodeOnly) { + intermediateResultCompress.completed = 1; + intermediateResultDecompress.completed = 0; + } else { /* both */ + intermediateResultCompress.completed = 0; + intermediateResultDecompress.completed = 0; + } + + while(!intermediateResultCompress.completed) { + intermediateResultCompress = BMK_benchFunctionTimed(timeStateCompress, &local_defaultCompress, (void*)cctx, &local_initCCtx, (void*)&cctxprep, + nbBlocks, srcPtrs, srcSizes, dstPtrs, dstCapacities, dstSizes); + + if(intermediateResultCompress.result.error) { + results.error = intermediateResultCompress.result.error; + BMK_freeTimeState(timeStateCompress); + BMK_freeTimeState(timeStateDecompress); + return results; + } + results.result.cSpeed = (srcSize * TIMELOOP_NANOSEC) / intermediateResultCompress.result.result.nanoSecPerRun; + results.result.cSize = intermediateResultCompress.result.result.sumOfReturn; + } + + while(!intermediateResultDecompress.completed) { + intermediateResultDecompress = BMK_benchFunctionTimed(timeStateDecompress, &local_defaultDecompress, (void*)(dctx), &local_initDCtx, (void*)&dctxprep, + nbBlocks, (const void* const*)dstPtrs, dstSizes, resPtrs, resSizes, NULL); + + if(intermediateResultDecompress.result.error) { + results.error = intermediateResultDecompress.result.error; + BMK_freeTimeState(timeStateCompress); + BMK_freeTimeState(timeStateDecompress); + return results; + } + results.result.dSpeed = (srcSize * TIMELOOP_NANOSEC) / intermediateResultDecompress.result.result.nanoSecPerRun; + } + + BMK_freeTimeState(timeStateCompress); + BMK_freeTimeState(timeStateDecompress); + + } else { /* iterMode; */ + if(mode != BMK_decodeOnly) { + + BMK_customReturn_t compressionResults = BMK_benchFunction(&local_defaultCompress, (void*)cctx, &local_initCCtx, (void*)&cctxprep, + nbBlocks, srcPtrs, srcSizes, dstPtrs, dstCapacities, dstSizes, nbSeconds); + if(compressionResults.error) { + results.error = compressionResults.error; + return results; + } + if(compressionResults.result.nanoSecPerRun == 0) { + results.result.cSpeed = 0; + } else { + results.result.cSpeed = srcSize * TIMELOOP_NANOSEC / compressionResults.result.nanoSecPerRun; + } + results.result.cSize = compressionResults.result.sumOfReturn; + } + + if(mode != BMK_compressOnly) { + BMK_customReturn_t decompressionResults; + decompressionResults = BMK_benchFunction( + &local_defaultDecompress, (void*)(dctx), + &local_initDCtx, (void*)&dctxprep, nbBlocks, + (const void* const*)dstPtrs, dstSizes, resPtrs, resSizes, NULL, + nbSeconds); + + if(decompressionResults.error) { + results.error = decompressionResults.error; + return results; + } + + if(decompressionResults.result.nanoSecPerRun == 0) { + results.result.dSpeed = 0; + } else { + results.result.dSpeed = srcSize * TIMELOOP_NANOSEC / decompressionResults.result.nanoSecPerRun; + } + } + } + } + /* Bench */ + results.result.cMem = (1 << (comprParams->vals[wlog_ind])) + ZSTD_sizeof_CCtx(cctx); + return results; +} + +static int BMK_benchParam(BMK_result_t* resultPtr, + const buffers_t buf, const contexts_t ctx, + const paramValues_t cParams) { + BMK_return_t res = BMK_benchMemInvertible(buf, ctx, BASE_CLEVEL, &cParams, BMK_both, BMK_timeMode, 3); + *resultPtr = res.result; + return res.error; +} + + +#define CBENCHMARK(conditional, resultvar, tmpret, mode, loopmode, sec) { \ + if(conditional) { \ + BMK_return_t tmpret = BMK_benchMemInvertible(buf, ctx, BASE_CLEVEL, &cParams, mode, loopmode, sec); \ + if(tmpret.error) { DEBUGOUTPUT("Benchmarking failed\n"); return ERROR_RESULT; } \ + if(mode != BMK_decodeOnly) { \ + resultvar.cSpeed = tmpret.result.cSpeed; \ + resultvar.cSize = tmpret.result.cSize; \ + resultvar.cMem = tmpret.result.cMem; \ + } \ + if(mode != BMK_compressOnly) { resultvar.dSpeed = tmpret.result.dSpeed; } \ + } \ +} + +/* Benchmarking which stops when we are sufficiently sure the solution is infeasible / worse than the winner */ +#define VARIANCE 1.2 +static int allBench(BMK_result_t* resultPtr, + const buffers_t buf, const contexts_t ctx, + const paramValues_t cParams, + const constraint_t target, + BMK_result_t* winnerResult, int feas) { + BMK_result_t resultMax, benchres; + U64 loopDurationC = 0, loopDurationD = 0; + double uncertaintyConstantC = 3., uncertaintyConstantD = 3.; + double winnerRS; + /* initial benchmarking, gives exact ratio and memory, warms up future runs */ + CBENCHMARK(1, benchres, tmp, BMK_both, BMK_iterMode, 1); + + winnerRS = resultScore(*winnerResult, buf.srcSize, target); + DEBUGOUTPUT("WinnerScore: %f\n ", winnerRS); + + *resultPtr = benchres; + + /* calculate uncertainty in compression / decompression runs */ + if(benchres.cSpeed) { + loopDurationC = ((buf.srcSize * TIMELOOP_NANOSEC) / benchres.cSpeed); + uncertaintyConstantC = ((loopDurationC + (double)(2 * g_clockGranularity))/loopDurationC); + } + + if(benchres.dSpeed) { + loopDurationD = ((buf.srcSize * TIMELOOP_NANOSEC) / benchres.dSpeed); + uncertaintyConstantD = ((loopDurationD + (double)(2 * g_clockGranularity))/loopDurationD); + } + + /* anything with worse ratio in feas is definitely worse, discard */ + if(feas && benchres.cSize < winnerResult->cSize && !g_optmode) { + return WORSE_RESULT; + } + + /* second run, if first run is too short, gives approximate cSpeed + dSpeed */ + CBENCHMARK(loopDurationC < TIMELOOP_NANOSEC / 10, benchres, tmp, BMK_compressOnly, BMK_iterMode, 1); + CBENCHMARK(loopDurationD < TIMELOOP_NANOSEC / 10, benchres, tmp, BMK_decodeOnly, BMK_iterMode, 1); + + *resultPtr = benchres; + + /* optimistic assumption of benchres */ + resultMax = benchres; + resultMax.cSpeed *= uncertaintyConstantC * VARIANCE; + resultMax.dSpeed *= uncertaintyConstantD * VARIANCE; + + /* disregard infeasible results in feas mode */ + /* disregard if resultMax < winner in infeas mode */ + if((feas && !feasible(resultMax, target)) || + (!feas && (winnerRS > resultScore(resultMax, buf.srcSize, target)))) { + return WORSE_RESULT; + } + + CBENCHMARK(loopDurationC < TIMELOOP_NANOSEC, benchres, tmp, BMK_compressOnly, BMK_timeMode, 1); + CBENCHMARK(loopDurationD < TIMELOOP_NANOSEC, benchres, tmp, BMK_decodeOnly, BMK_timeMode, 1); + + *resultPtr = benchres; + + /* compare by resultScore when in infeas */ + /* compare by compareResultLT when in feas */ + if((!feas && (resultScore(benchres, buf.srcSize, target) > resultScore(*winnerResult, buf.srcSize, target))) || + (feas && (compareResultLT(*winnerResult, benchres, target, buf.srcSize))) ) { + return BETTER_RESULT; + } else { + return WORSE_RESULT; + } +} + +#define INFEASIBLE_THRESHOLD 200 +/* Memoized benchmarking, won't benchmark anything which has already been benchmarked before. */ +static int benchMemo(BMK_result_t* resultPtr, + const buffers_t buf, const contexts_t ctx, + const paramValues_t cParams, + const constraint_t target, + BMK_result_t* winnerResult, memoTable_t* const memoTableArray, + const int feas) { + static int bmcount = 0; + int res; + + if(memoTableGet(memoTableArray, cParams) >= INFEASIBLE_THRESHOLD || redundantParams(cParams, target, buf.maxBlockSize)) { return WORSE_RESULT; } + + res = allBench(resultPtr, buf, ctx, cParams, target, winnerResult, feas); + + if(DEBUG && !(bmcount % 250)) { + DISPLAY("Count: %d\n", bmcount); + bmcount++; + } + BMK_printWinnerOpt(stdout, CUSTOM_LEVEL, *resultPtr, cParams, target, buf.srcSize); + + if(res == BETTER_RESULT || feas) { + memoTableSet(memoTableArray, cParams, 255); /* what happens if collisions are frequent */ + } + return res; +} + +typedef struct { + U64 cSpeed_min; + U64 dSpeed_min; U32 windowLog_max; ZSTD_strategy strategy_max; } level_constraints_t; @@ -243,15 +1659,14 @@ static void BMK_init_level_constraints(int bytePerSec_level1) } } } -static int BMK_seed(winnerInfo_t* winners, const ZSTD_compressionParameters params, - const void* srcBuffer, size_t srcSize, - ZSTD_CCtx* ctx, ZSTD_DCtx* dctx) +static int BMK_seed(winnerInfo_t* winners, const paramValues_t params, + const buffers_t buf, const contexts_t ctx) { BMK_result_t testResult; int better = 0; int cLevel; - BMK_benchParam(&testResult, srcBuffer, srcSize, ctx, dctx, params); + BMK_benchParam(&testResult, buf, ctx, params); for (cLevel = 1; cLevel <= NB_LEVELS_TRACKED; cLevel++) { @@ -259,32 +1674,32 @@ static int BMK_seed(winnerInfo_t* winners, const ZSTD_compressionParameters para continue; /* not fast enough for this level */ if (testResult.dSpeed < g_level_constraint[cLevel].dSpeed_min) continue; /* not fast enough for this level */ - if (params.windowLog > g_level_constraint[cLevel].windowLog_max) + if (params.vals[wlog_ind] > g_level_constraint[cLevel].windowLog_max) continue; /* too much memory for this level */ - if (params.strategy > g_level_constraint[cLevel].strategy_max) + if (params.vals[strt_ind] > g_level_constraint[cLevel].strategy_max) continue; /* forbidden strategy for this level */ if (winners[cLevel].result.cSize==0) { /* first solution for this cLevel */ winners[cLevel].result = testResult; winners[cLevel].params = params; - BMK_printWinner(stdout, cLevel, testResult, params, srcSize); + BMK_printWinner(stdout, cLevel, testResult, params, buf.srcSize); better = 1; continue; } if ((double)testResult.cSize <= ((double)winners[cLevel].result.cSize * (1. + (0.02 / cLevel))) ) { /* Validate solution is "good enough" */ - double W_ratio = (double)srcSize / testResult.cSize; - double O_ratio = (double)srcSize / winners[cLevel].result.cSize; + double W_ratio = (double)buf.srcSize / testResult.cSize; + double O_ratio = (double)buf.srcSize / winners[cLevel].result.cSize; double W_ratioNote = log (W_ratio); double O_ratioNote = log (O_ratio); - size_t W_DMemUsed = (1 << params.windowLog) + (16 KB); - size_t O_DMemUsed = (1 << winners[cLevel].params.windowLog) + (16 KB); + size_t W_DMemUsed = (1 << params.vals[wlog_ind]) + (16 KB); + size_t O_DMemUsed = (1 << winners[cLevel].params.vals[wlog_ind]) + (16 KB); double W_DMemUsed_note = W_ratioNote * ( 40 + 9*cLevel) - log((double)W_DMemUsed); double O_DMemUsed_note = O_ratioNote * ( 40 + 9*cLevel) - log((double)O_DMemUsed); - size_t W_CMemUsed = (1 << params.windowLog) + ZSTD_estimateCCtxSize_usingCParams(params); - size_t O_CMemUsed = (1 << winners[cLevel].params.windowLog) + ZSTD_estimateCCtxSize_usingCParams(winners[cLevel].params); + size_t W_CMemUsed = (1 << params.vals[wlog_ind]) + ZSTD_estimateCCtxSize_usingCParams(pvalsToCParams(params)); + size_t O_CMemUsed = (1 << winners[cLevel].params.vals[wlog_ind]) + ZSTD_estimateCCtxSize_usingCParams(pvalsToCParams(winners[cLevel].params)); double W_CMemUsed_note = W_ratioNote * ( 50 + 13*cLevel) - log((double)W_CMemUsed); double O_CMemUsed_note = O_ratioNote * ( 50 + 13*cLevel) - log((double)O_CMemUsed); @@ -314,16 +1729,16 @@ static int BMK_seed(winnerInfo_t* winners, const ZSTD_compressionParameters para /* too large compression speed difference for the compression benefit */ if (W_ratio > O_ratio) DISPLAY ("Compression Speed : %5.3f @ %4.1f MB/s vs %5.3f @ %4.1f MB/s : not enough for level %i\n", - W_ratio, testResult.cSpeed / 1000000, - O_ratio, winners[cLevel].result.cSpeed / 1000000., cLevel); + W_ratio, (double)testResult.cSpeed / (1 MB), + O_ratio, (double)winners[cLevel].result.cSpeed / (1 MB), cLevel); continue; } if (W_DSpeed_note < O_DSpeed_note ) { /* too large decompression speed difference for the compression benefit */ if (W_ratio > O_ratio) DISPLAY ("Decompression Speed : %5.3f @ %4.1f MB/s vs %5.3f @ %4.1f MB/s : not enough for level %i\n", - W_ratio, testResult.dSpeed / 1000000., - O_ratio, winners[cLevel].result.dSpeed / 1000000., cLevel); + W_ratio, (double)testResult.dSpeed / (1 MB), + O_ratio, (double)winners[cLevel].result.dSpeed / (1 MB), cLevel); continue; } @@ -332,7 +1747,7 @@ static int BMK_seed(winnerInfo_t* winners, const ZSTD_compressionParameters para winners[cLevel].result = testResult; winners[cLevel].params = params; - BMK_printWinner(stdout, cLevel, testResult, params, srcSize); + BMK_printWinner(stdout, cLevel, testResult, params, buf.srcSize); better = 1; } } @@ -340,158 +1755,72 @@ static int BMK_seed(winnerInfo_t* winners, const ZSTD_compressionParameters para return better; } - -/* nullified useless params, to ensure count stats */ -static ZSTD_compressionParameters* sanitizeParams(ZSTD_compressionParameters params) -{ - g_params = params; - if (params.strategy == ZSTD_fast) - g_params.chainLog = 0, g_params.searchLog = 0; - if (params.strategy == ZSTD_dfast) - g_params.searchLog = 0; - if (params.strategy != ZSTD_btopt && params.strategy != ZSTD_btultra) - g_params.targetLength = 0; - return &g_params; -} - - -static void paramVariation(ZSTD_compressionParameters* ptr) -{ - ZSTD_compressionParameters p; - U32 validated = 0; - while (!validated) { - U32 nbChanges = (FUZ_rand(&g_rand) & 3) + 1; - p = *ptr; - for ( ; nbChanges ; nbChanges--) { - const U32 changeID = FUZ_rand(&g_rand) % 14; - switch(changeID) - { - case 0: - p.chainLog++; break; - case 1: - p.chainLog--; break; - case 2: - p.hashLog++; break; - case 3: - p.hashLog--; break; - case 4: - p.searchLog++; break; - case 5: - p.searchLog--; break; - case 6: - p.windowLog++; break; - case 7: - p.windowLog--; break; - case 8: - p.searchLength++; break; - case 9: - p.searchLength--; break; - case 10: - p.strategy = (ZSTD_strategy)(((U32)p.strategy)+1); break; - case 11: - p.strategy = (ZSTD_strategy)(((U32)p.strategy)-1); break; - case 12: - p.targetLength *= 1 + ((double)(FUZ_rand(&g_rand)&255)) / 256.; break; - case 13: - p.targetLength /= 1 + ((double)(FUZ_rand(&g_rand)&255)) / 256.; break; - } - } - validated = !ZSTD_isError(ZSTD_checkCParams(p)); - } - *ptr = p; -} - +/*-************************************ +* Compression Level Table Generation Functions +**************************************/ #define PARAMTABLELOG 25 #define PARAMTABLESIZE (1<> 3) & PARAMTABLEMASK] - +static BYTE* NB_TESTS_PLAYED(paramValues_t p) { + ZSTD_compressionParameters p2 = pvalsToCParams(sanitizeParams(p)); + return &g_alreadyTested[(XXH64((void*)&p2, sizeof(p2), 0) >> 3) & PARAMTABLEMASK]; +} static void playAround(FILE* f, winnerInfo_t* winners, - ZSTD_compressionParameters params, - const void* srcBuffer, size_t srcSize, - ZSTD_CCtx* ctx, ZSTD_DCtx* dctx) + paramValues_t p, + const buffers_t buf, const contexts_t ctx) { - int nbVariations = 0; + int nbVariations = 0, i; UTIL_time_t const clockStart = UTIL_getTime(); while (UTIL_clockSpanMicro(clockStart) < g_maxVariationTime) { - ZSTD_compressionParameters p = params; + BYTE* b; if (nbVariations++ > g_maxNbVariations) break; - paramVariation(&p); + + do { for(i = 0; i < 4; i++) { paramVaryOnce(FUZ_rand(&g_rand) % (strt_ind + 1), ((FUZ_rand(&g_rand) & 1) << 1) - 1, &p); } } + while(!paramValid(p)); /* exclude faster if already played params */ - if (FUZ_rand(&g_rand) & ((1 << NB_TESTS_PLAYED(p))-1)) + if (FUZ_rand(&g_rand) & ((1 << *NB_TESTS_PLAYED(p))-1)) continue; /* test */ - NB_TESTS_PLAYED(p)++; - if (!BMK_seed(winners, p, srcBuffer, srcSize, ctx, dctx)) continue; + b = NB_TESTS_PLAYED(p); + (*b)++; + if (!BMK_seed(winners, p, buf, ctx)) continue; /* improvement found => search more */ - BMK_printWinners(f, winners, srcSize); - playAround(f, winners, p, srcBuffer, srcSize, ctx, dctx); + BMK_printWinners(f, winners, buf.srcSize); + playAround(f, winners, p, buf, ctx); } } - -static ZSTD_compressionParameters randomParams(void) -{ - ZSTD_compressionParameters p; - U32 validated = 0; - while (!validated) { - /* totally random entry */ - p.chainLog = (FUZ_rand(&g_rand) % (ZSTD_CHAINLOG_MAX+1 - ZSTD_CHAINLOG_MIN)) + ZSTD_CHAINLOG_MIN; - p.hashLog = (FUZ_rand(&g_rand) % (ZSTD_HASHLOG_MAX+1 - ZSTD_HASHLOG_MIN)) + ZSTD_HASHLOG_MIN; - p.searchLog = (FUZ_rand(&g_rand) % (ZSTD_SEARCHLOG_MAX+1 - ZSTD_SEARCHLOG_MIN)) + ZSTD_SEARCHLOG_MIN; - p.windowLog = (FUZ_rand(&g_rand) % (ZSTD_WINDOWLOG_MAX+1 - ZSTD_WINDOWLOG_MIN)) + ZSTD_WINDOWLOG_MIN; - p.searchLength=(FUZ_rand(&g_rand) % (ZSTD_SEARCHLENGTH_MAX+1 - ZSTD_SEARCHLENGTH_MIN)) + ZSTD_SEARCHLENGTH_MIN; - p.targetLength=(FUZ_rand(&g_rand) % (512)); - p.strategy = (ZSTD_strategy) (FUZ_rand(&g_rand) % (ZSTD_btultra +1)); - validated = !ZSTD_isError(ZSTD_checkCParams(p)); - } - return p; -} - static void BMK_selectRandomStart( FILE* f, winnerInfo_t* winners, - const void* srcBuffer, size_t srcSize, - ZSTD_CCtx* ctx, ZSTD_DCtx* dctx) + const buffers_t buf, const contexts_t ctx) { U32 const id = FUZ_rand(&g_rand) % (NB_LEVELS_TRACKED+1); - if ((id==0) || (winners[id].params.windowLog==0)) { + if ((id==0) || (winners[id].params.vals[wlog_ind]==0)) { /* use some random entry */ - ZSTD_compressionParameters const p = ZSTD_adjustCParams(randomParams(), srcSize, 0); - playAround(f, winners, p, srcBuffer, srcSize, ctx, dctx); + paramValues_t const p = adjustParams(cParamsToPVals(pvalsToCParams(randomParams())), /* defaults nonCompression parameters */ + buf.srcSize, 0); + playAround(f, winners, p, buf, ctx); } else { - playAround(f, winners, winners[id].params, srcBuffer, srcSize, ctx, dctx); + playAround(f, winners, winners[id].params, buf, ctx); } } - -static void BMK_benchOnce(ZSTD_CCtx* cctx, ZSTD_DCtx* dctx, const void* srcBuffer, size_t srcSize) +static void BMK_benchFullTable(const buffers_t buf, const contexts_t ctx) { - BMK_result_t testResult; - g_params = ZSTD_adjustCParams(g_params, srcSize, 0); - BMK_benchParam(&testResult, srcBuffer, srcSize, cctx, dctx, g_params); - DISPLAY("Compression Ratio: %.3f Compress Speed: %.1f MB/s Decompress Speed: %.1f MB/s\n", (double)srcSize / testResult.cSize, - testResult.cSpeed / 1000000, testResult.dSpeed / 1000000); - return; -} - -static void BMK_benchFullTable(ZSTD_CCtx* cctx, ZSTD_DCtx* dctx, const void* srcBuffer, size_t srcSize) -{ - ZSTD_compressionParameters params; + paramValues_t params; winnerInfo_t winners[NB_LEVELS_TRACKED+1]; const char* const rfName = "grillResults.txt"; FILE* const f = fopen(rfName, "w"); - const size_t blockSize = g_blockSize ? g_blockSize : srcSize; /* cut by block or not ? */ /* init */ assert(g_singleRun==0); @@ -499,12 +1828,12 @@ static void BMK_benchFullTable(ZSTD_CCtx* cctx, ZSTD_DCtx* dctx, const void* src if (f==NULL) { DISPLAY("error opening %s \n", rfName); exit(1); } if (g_target) { - BMK_init_level_constraints(g_target*1000000); + BMK_init_level_constraints(g_target * (1 MB)); } else { /* baseline config for level 1 */ - ZSTD_compressionParameters const l1params = ZSTD_getCParams(1, blockSize, 0); + paramValues_t const l1params = cParamsToPVals(ZSTD_getCParams(1, buf.maxBlockSize, ctx.dictSize)); BMK_result_t testResult; - BMK_benchParam(&testResult, srcBuffer, srcSize, cctx, dctx, l1params); + BMK_benchParam(&testResult, buf, ctx, l1params); BMK_init_level_constraints((int)((testResult.cSpeed * 31) / 32)); } @@ -512,251 +1841,543 @@ static void BMK_benchFullTable(ZSTD_CCtx* cctx, ZSTD_DCtx* dctx, const void* src { const int maxSeeds = g_noSeed ? 1 : ZSTD_maxCLevel(); int i; for (i=0; i<=maxSeeds; i++) { - params = ZSTD_getCParams(i, blockSize, 0); - BMK_seed(winners, params, srcBuffer, srcSize, cctx, dctx); + params = cParamsToPVals(ZSTD_getCParams(i, buf.maxBlockSize, 0)); + BMK_seed(winners, params, buf, ctx); } } - BMK_printWinners(f, winners, srcSize); + BMK_printWinners(f, winners, buf.srcSize); /* start tests */ - { const time_t grillStart = time(NULL); + { const UTIL_time_t grillStart = UTIL_getTime(); do { - BMK_selectRandomStart(f, winners, srcBuffer, srcSize, cctx, dctx); - } while (BMK_timeSpan(grillStart) < g_grillDuration_s); + BMK_selectRandomStart(f, winners, buf, ctx); + } while (BMK_timeSpan(grillStart) < g_timeLimit_s); } /* end summary */ - BMK_printWinners(f, winners, srcSize); + BMK_printWinners(f, winners, buf.srcSize); DISPLAY("grillParams operations completed \n"); /* clean up*/ fclose(f); } -static void BMK_benchMem_usingCCtx(ZSTD_CCtx* const cctx, ZSTD_DCtx* const dctx, const void* srcBuffer, size_t srcSize) -{ - if (g_singleRun) - return BMK_benchOnce(cctx, dctx, srcBuffer, srcSize); - else - return BMK_benchFullTable(cctx, dctx, srcBuffer, srcSize); -} +/*-************************************ +* Single Benchmark Functions +**************************************/ -static void BMK_benchMemCCtxInit(const void* srcBuffer, size_t srcSize) -{ - ZSTD_CCtx* const cctx = ZSTD_createCCtx(); - ZSTD_DCtx* const dctx = ZSTD_createDCtx(); - if (cctx==NULL || dctx==NULL) { DISPLAY("Context Creation failed \n"); exit(1); } - BMK_benchMem_usingCCtx(cctx, dctx, srcBuffer, srcSize); - ZSTD_freeCCtx(cctx); -} +static int benchOnce(const buffers_t buf, const contexts_t ctx, const int cLevel) { + BMK_result_t testResult; + g_params = adjustParams(overwriteParams(cParamsToPVals(ZSTD_getCParams(cLevel, buf.maxBlockSize, ctx.dictSize)), g_params), buf.maxBlockSize, ctx.dictSize); + if(BMK_benchParam(&testResult, buf, ctx, g_params)) { + DISPLAY("Error during benchmarking\n"); + return 1; + } -static int benchSample(void) -{ - const char* const name = "Sample 10MB"; - size_t const benchedSize = 10000000; + BMK_printWinner(stdout, CUSTOM_LEVEL, testResult, g_params, buf.srcSize); - void* origBuff = malloc(benchedSize); - if (!origBuff) { perror("not enough memory"); return 12; } - - /* Fill buffer */ - RDG_genBuffer(origBuff, benchedSize, g_compressibility, 0.0, 0); - - /* bench */ - DISPLAY("\r%79s\r", ""); - DISPLAY("using %s %i%%: \n", name, (int)(g_compressibility*100)); - BMK_benchMemCCtxInit(origBuff, benchedSize); - - free(origBuff); return 0; } +static int benchSample(double compressibility, int cLevel) +{ + const char* const name = "Sample 10MB"; + size_t const benchedSize = 10 MB; + void* srcBuffer = malloc(benchedSize); + int ret = 0; + + buffers_t buf; + contexts_t ctx; + + if(srcBuffer == NULL) { + DISPLAY("Out of Memory\n"); + return 2; + } + + RDG_genBuffer(srcBuffer, benchedSize, compressibility, 0.0, 0); + + if(createBuffersFromMemory(&buf, srcBuffer, 1, &benchedSize)) { + DISPLAY("Buffer Creation Error\n"); + free(srcBuffer); + return 3; + } + + if(createContexts(&ctx, NULL)) { + DISPLAY("Context Creation Error\n"); + freeBuffers(buf); + return 1; + } + + /* bench */ + DISPLAY("\r%79s\r", ""); + DISPLAY("using %s %i%%: \n", name, (int)(compressibility*100)); + + if(g_singleRun) { + ret = benchOnce(buf, ctx, cLevel); + } else { + BMK_benchFullTable(buf, ctx); + } + + freeBuffers(buf); + freeContexts(ctx); + + return ret; +} /* benchFiles() : * note: while this function takes a table of filenames, * in practice, only the first filename will be used */ -int benchFiles(const char** fileNamesTable, int nbFiles) +int benchFiles(const char** fileNamesTable, int nbFiles, const char* dictFileName, const int cLevel) { - int fileIdx=0; + buffers_t buf; + contexts_t ctx; + int ret = 0; - /* Loop for each file */ - while (fileIdx inFileSize) benchedSize = (size_t)inFileSize; - if (benchedSize < inFileSize) - DISPLAY("Not enough memory for '%s' full size; testing %i MB only...\n", inFileName, (int)(benchedSize>>20)); - origBuff = malloc(benchedSize); - if (origBuff==NULL) { - DISPLAY("\nError: not enough memory!\n"); - fclose(inFile); - return 12; - } - - /* Fill input buffer */ - DISPLAY("Loading %s... \r", inFileName); - { size_t const readSize = fread(origBuff, 1, benchedSize, inFile); - fclose(inFile); - if(readSize != benchedSize) { - DISPLAY("\nError: problem reading file '%s' !! \n", inFileName); - free(origBuff); - return 13; - } } - - /* bench */ - DISPLAY("\r%79s\r", ""); - DISPLAY("using %s : \n", inFileName); - BMK_benchMemCCtxInit(origBuff, benchedSize); - - /* clean */ - free(origBuff); + if(createBuffers(&buf, fileNamesTable, nbFiles)) { + DISPLAY("unable to load files\n"); + return 1; } - return 0; -} - - -static void BMK_translateAdvancedParams(ZSTD_compressionParameters params) -{ - DISPLAY("--zstd=windowLog=%u,chainLog=%u,hashLog=%u,searchLog=%u,searchLength=%u,targetLength=%u,strategy=%u \n", - params.windowLog, params.chainLog, params.hashLog, params.searchLog, params.searchLength, params.targetLength, (U32)(params.strategy)); -} - -/* optimizeForSize(): - * targetSpeed : expressed in MB/s */ -int optimizeForSize(const char* inFileName, U32 targetSpeed) -{ - FILE* const inFile = fopen( inFileName, "rb" ); - U64 const inFileSize = UTIL_getFileSize(inFileName); - size_t benchedSize = BMK_findMaxMem(inFileSize*3) / 3; - void* origBuff; - /* Init */ - if (inFile==NULL) { DISPLAY( "Pb opening %s\n", inFileName); return 11; } - if (inFileSize == UTIL_FILESIZE_UNKNOWN) { - DISPLAY("Pb evaluatin size of %s \n", inFileName); - fclose(inFile); - return 11; + if(createContexts(&ctx, dictFileName)) { + DISPLAY("unable to load dictionary\n"); + freeBuffers(buf); + return 2; } - /* Memory allocation & restrictions */ - if ((U64)benchedSize > inFileSize) benchedSize = (size_t)inFileSize; - if (benchedSize < inFileSize) { - DISPLAY("Not enough memory for '%s' \n", inFileName); - fclose(inFile); - return 11; - } - - /* Alloc */ - origBuff = malloc(benchedSize); - if(!origBuff) { - DISPLAY("\nError: not enough memory!\n"); - fclose(inFile); - return 12; - } - - /* Fill input buffer */ - DISPLAY("Loading %s... \r", inFileName); - { size_t const readSize = fread(origBuff, 1, benchedSize, inFile); - fclose(inFile); - if(readSize != benchedSize) { - DISPLAY("\nError: problem reading file '%s' !! \n", inFileName); - free(origBuff); - return 13; - } } - - /* bench */ DISPLAY("\r%79s\r", ""); - DISPLAY("optimizing for %s - limit speed %u MB/s \n", inFileName, targetSpeed); - targetSpeed *= 1000000; - { ZSTD_CCtx* const ctx = ZSTD_createCCtx(); - ZSTD_DCtx* const dctx = ZSTD_createDCtx(); - winnerInfo_t winner; - BMK_result_t candidate; - const size_t blockSize = g_blockSize ? g_blockSize : benchedSize; + if(nbFiles == 1) { + DISPLAY("using %s : \n", fileNamesTable[0]); + } else { + DISPLAY("using %d Files : \n", nbFiles); + } - /* init */ - if (ctx==NULL) { DISPLAY("\n ZSTD_createCCtx error \n"); free(origBuff); return 14;} - memset(&winner, 0, sizeof(winner)); - winner.result.cSize = (size_t)(-1); + if(g_singleRun) { + ret = benchOnce(buf, ctx, cLevel); + } else { + BMK_benchFullTable(buf, ctx); + } - /* find best solution from default params */ - { const int maxSeeds = g_noSeed ? 1 : ZSTD_maxCLevel(); - int i; - for (i=1; i<=maxSeeds; i++) { - ZSTD_compressionParameters const CParams = ZSTD_getCParams(i, blockSize, 0); - BMK_benchParam(&candidate, origBuff, benchedSize, ctx, dctx, CParams); - if (candidate.cSpeed < (double)targetSpeed) { + freeBuffers(buf); + freeContexts(ctx); + return ret; +} + + +/*-************************************ +* Local Optimization Functions +**************************************/ + +/* One iteration of hill climbing. Specifically, it first tries all + * valid parameter configurations w/ manhattan distance 1 and picks the best one + * failing that, it progressively tries candidates further and further away (up to #dim + 2) + * if it finds a candidate exceeding winnerInfo, it will repeat. Otherwise, it will stop the + * current stage of hill climbing. + * Each iteration of hill climbing proceeds in 2 'phases'. Phase 1 climbs according to + * the resultScore function, which is effectively a linear increase in reward until it reaches + * the constraint-satisfying value, it which point any excess results in only logarithmic reward. + * This aims to find some constraint-satisfying point. + * Phase 2 optimizes in accordance with what the original function sets out to maximize, with + * all feasible solutions valued over all infeasible solutions. + */ + +/* sanitize all params here. + * all generation after random should be sanitized. (maybe sanitize random) + */ +static winnerInfo_t climbOnce(const constraint_t target, + memoTable_t* mtAll, + const buffers_t buf, const contexts_t ctx, + const paramValues_t init) { + /* + * cparam - currently considered 'center' + * candidate - params to benchmark/results + * winner - best option found so far. + */ + paramValues_t cparam = init; + winnerInfo_t candidateInfo, winnerInfo; + int better = 1; + int feas = 0; + + winnerInfo = initWinnerInfo(init); + candidateInfo = winnerInfo; + + { + winnerInfo_t bestFeasible1 = initWinnerInfo(cparam); + DEBUGOUTPUT("Climb Part 1\n"); + while(better) { + + int offset; + size_t i, dist; + const size_t varLen = mtAll[cparam.vals[strt_ind]].varLen; + better = 0; + DEBUGOUTPUT("Start\n"); + cparam = winnerInfo.params; + candidateInfo.params = cparam; + /* all dist-1 candidates */ + for(i = 0; i < varLen; i++) { + for(offset = -1; offset <= 1; offset += 2) { + CHECKTIME(winnerInfo); + candidateInfo.params = cparam; + paramVaryOnce(mtAll[cparam.vals[strt_ind]].varArray[i], offset, &candidateInfo.params); + + if(paramValid(candidateInfo.params)) { + int res; + res = benchMemo(&candidateInfo.result, buf, ctx, + sanitizeParams(candidateInfo.params), target, &winnerInfo.result, mtAll, feas); + DEBUGOUTPUT("Res: %d\n", res); + if(res == BETTER_RESULT) { /* synonymous with better when called w/ infeasibleBM */ + winnerInfo = candidateInfo; + better = 1; + if(compareResultLT(bestFeasible1.result, winnerInfo.result, target, buf.srcSize)) { + bestFeasible1 = winnerInfo; + } + } + } + } + } + + if(better) { + continue; + } + + for(dist = 2; dist < varLen + 2; dist++) { /* varLen is # dimensions */ + for(i = 0; i < (1 << varLen) / varLen + 2; i++) { + int res; + CHECKTIME(winnerInfo); + candidateInfo.params = cparam; + /* param error checking already done here */ + paramVariation(&candidateInfo.params, mtAll, (U32)dist); + + res = benchMemo(&candidateInfo.result, buf, ctx, + sanitizeParams(candidateInfo.params), target, &winnerInfo.result, mtAll, feas); + DEBUGOUTPUT("Res: %d\n", res); + if(res == BETTER_RESULT) { /* synonymous with better in this case*/ + winnerInfo = candidateInfo; + better = 1; + if(compareResultLT(bestFeasible1.result, winnerInfo.result, target, buf.srcSize)) { + bestFeasible1 = winnerInfo; + } + break; + } + + } + if(better) { break; } - if ( (candidate.cSize < winner.result.cSize) - | ((candidate.cSize == winner.result.cSize) & (candidate.cSpeed > winner.result.cSpeed)) ) - { - winner.params = CParams; - winner.result = candidate; - BMK_printWinner(stdout, i, winner.result, winner.params, benchedSize); - } } + } + + if(!better) { /* infeas -> feas -> stop */ + if(feas) { return winnerInfo; } + + feas = 1; + better = 1; + winnerInfo = bestFeasible1; /* note with change, bestFeasible may not necessarily be feasible, but if one has been benchmarked, it will be. */ + DEBUGOUTPUT("Climb Part 2\n"); + } } - - BMK_printWinner(stdout, CUSTOM_LEVEL, winner.result, winner.params, benchedSize); - - BMK_translateAdvancedParams(winner.params); - - /* start tests */ - { time_t const grillStart = time(NULL); - do { - ZSTD_compressionParameters params = winner.params; - paramVariation(¶ms); - if ((FUZ_rand(&g_rand) & 31) == 3) params = randomParams(); /* totally random config to improve search space */ - params = ZSTD_adjustCParams(params, blockSize, 0); - - /* exclude faster if already played set of params */ - if (FUZ_rand(&g_rand) & ((1 << NB_TESTS_PLAYED(params))-1)) continue; - - /* test */ - NB_TESTS_PLAYED(params)++; - BMK_benchParam(&candidate, origBuff, benchedSize, ctx, dctx, params); - - /* improvement found => new winner */ - if ( (candidate.cSpeed > targetSpeed) - & ( (candidate.cSize < winner.result.cSize) - | ((candidate.cSize == winner.result.cSize) & (candidate.cSpeed > winner.result.cSpeed)) ) ) - { - winner.params = params; - winner.result = candidate; - BMK_printWinner(stdout, CUSTOM_LEVEL, winner.result, winner.params, benchedSize); - BMK_translateAdvancedParams(winner.params); - } - } while (BMK_timeSpan(grillStart) < g_grillDuration_s); - } - /* end summary */ - - BMK_printWinner(stdout, CUSTOM_LEVEL, winner.result, winner.params, benchedSize); - BMK_translateAdvancedParams(winner.params); - DISPLAY("grillParams size - optimizer completed \n"); - - /* clean up*/ - ZSTD_freeCCtx(ctx); - ZSTD_freeDCtx(dctx); + winnerInfo = bestFeasible1; } - free(origBuff); - return 0; + return winnerInfo; +} + +/* Optimizes for a fixed strategy */ + +/* flexible parameters: iterations of failed climbing (or if we do non-random, maybe this is when everything is close to visitied) + weight more on visit for bad results, less on good results/more on later results / ones with more failures. + allocate memoTable here. + */ +static winnerInfo_t optimizeFixedStrategy( + const buffers_t buf, const contexts_t ctx, + const constraint_t target, paramValues_t paramTarget, + const ZSTD_strategy strat, + memoTable_t* memoTableArray, const int tries) { + int i = 0; + + paramValues_t init; + winnerInfo_t winnerInfo, candidateInfo; + winnerInfo = initWinnerInfo(emptyParams()); + /* so climb is given the right fixed strategy */ + paramTarget.vals[strt_ind] = strat; + /* to pass ZSTD_checkCParams */ + paramTarget = cParamUnsetMin(paramTarget); + + init = paramTarget; + + for(i = 0; i < tries; i++) { + DEBUGOUTPUT("Restart\n"); + do { randomConstrainedParams(&init, memoTableArray, strat); } while(redundantParams(init, target, buf.maxBlockSize)); + candidateInfo = climbOnce(target, memoTableArray, buf, ctx, init); + if(compareResultLT(winnerInfo.result, candidateInfo.result, target, buf.srcSize)) { + winnerInfo = candidateInfo; + BMK_printWinnerOpt(stdout, CUSTOM_LEVEL, winnerInfo.result, winnerInfo.params, target, buf.srcSize); + i = 0; + continue; + } + CHECKTIME(winnerInfo); + i++; + } + return winnerInfo; +} + +/* goes best, best-1, best+1, best-2, ... */ +/* return 0 if nothing remaining */ +static int nextStrategy(const int currentStrategy, const int bestStrategy) { + if(bestStrategy <= currentStrategy) { + int candidate = 2 * bestStrategy - currentStrategy - 1; + if(candidate < 1) { + candidate = currentStrategy + 1; + if(candidate > (int)ZSTD_btultra) { + return 0; + } else { + return candidate; + } + } else { + return candidate; + } + } else { /* bestStrategy >= currentStrategy */ + int candidate = 2 * bestStrategy - currentStrategy; + if(candidate > (int)ZSTD_btultra) { + candidate = currentStrategy - 1; + if(candidate < 1) { + return 0; + } else { + return candidate; + } + } else { + return candidate; + } + } +} + +/* experiment with playing with this and decay value */ + +/* main fn called when using --optimize */ +/* Does strategy selection by benchmarking default compression levels + * then optimizes by strategy, starting with the best one and moving + * progressively moving further away by number + * args: + * fileNamesTable - list of files to benchmark + * nbFiles - length of fileNamesTable + * dictFileName - name of dictionary file if one, else NULL + * target - performance constraints (cSpeed, dSpeed, cMem) + * paramTarget - parameter constraints (i.e. restriction search space to where strategy = ZSTD_fast) + * cLevel - compression level to exceed (all solutions must be > lvl in cSpeed + ratio) + */ + +static int g_maxTries = 5; +#define TRY_DECAY 1 + +static int optimizeForSize(const char* const * const fileNamesTable, const size_t nbFiles, const char* dictFileName, constraint_t target, paramValues_t paramTarget, + const int cLevelOpt, const int cLevelRun, const U32 memoTableLog) +{ + varInds_t varArray [NUM_PARAMS]; + int ret = 0; + const size_t varLen = variableParams(paramTarget, varArray, dictFileName != NULL); + winnerInfo_t winner = initWinnerInfo(emptyParams()); + memoTable_t* allMT = NULL; + paramValues_t paramBase; + contexts_t ctx; + buffers_t buf; + g_time = UTIL_getTime(); + + if(createBuffers(&buf, fileNamesTable, nbFiles)) { + DISPLAY("unable to load files\n"); + return 1; + } + + if(createContexts(&ctx, dictFileName)) { + DISPLAY("unable to load dictionary\n"); + freeBuffers(buf); + return 2; + } + + if(nbFiles == 1) { + DISPLAYLEVEL(2, "Loading %s... \r", fileNamesTable[0]); + } else { + DISPLAYLEVEL(2, "Loading %lu Files... \r", (unsigned long)nbFiles); + } + + /* sanitize paramTarget */ + optimizerAdjustInput(¶mTarget, buf.maxBlockSize); + paramBase = cParamUnsetMin(paramTarget); + + allMT = createMemoTableArray(paramTarget, varArray, varLen, memoTableLog); + + if(!allMT) { + DISPLAY("MemoTable Init Error\n"); + ret = 2; + goto _cleanUp; + } + + /* default strictnesses */ + if(g_strictness == PARAM_UNSET) { + if(g_optmode) { + g_strictness = 100; + } else { + g_strictness = 90; + } + } else { + if(0 >= g_strictness || g_strictness > 100) { + DISPLAY("Strictness Outside of Bounds\n"); + ret = 4; + goto _cleanUp; + } + } + + /* use level'ing mode instead of normal target mode */ + if(g_optmode) { + winner.params = cParamsToPVals(ZSTD_getCParams(cLevelOpt, buf.maxBlockSize, ctx.dictSize)); + if(BMK_benchParam(&winner.result, buf, ctx, winner.params)) { + ret = 3; + goto _cleanUp; + } + + g_lvltarget = winner.result; + g_lvltarget.cSpeed *= ((double)g_strictness) / 100; + g_lvltarget.dSpeed *= ((double)g_strictness) / 100; + g_lvltarget.cSize /= ((double)g_strictness) / 100; + + target.cSpeed = (U32)g_lvltarget.cSpeed; + target.dSpeed = (U32)g_lvltarget.dSpeed; + + BMK_printWinnerOpt(stdout, cLevelOpt, winner.result, winner.params, target, buf.srcSize); + } + + /* Don't want it to return anything worse than the best known result */ + if(g_singleRun) { + BMK_result_t res; + g_params = adjustParams(overwriteParams(cParamsToPVals(ZSTD_getCParams(cLevelRun, buf.maxBlockSize, ctx.dictSize)), g_params), buf.maxBlockSize, ctx.dictSize); + if(BMK_benchParam(&res, buf, ctx, g_params)) { + ret = 45; + goto _cleanUp; + } + if(compareResultLT(winner.result, res, relaxTarget(target), buf.srcSize)) { + winner.result = res; + winner.params = g_params; + } + } + + /* bench */ + DISPLAYLEVEL(2, "\r%79s\r", ""); + if(nbFiles == 1) { + DISPLAYLEVEL(2, "optimizing for %s", fileNamesTable[0]); + } else { + DISPLAYLEVEL(2, "optimizing for %lu Files", (unsigned long)nbFiles); + } + + if(target.cSpeed != 0) { DISPLAYLEVEL(2," - limit compression speed %u MB/s", target.cSpeed >> 20); } + if(target.dSpeed != 0) { DISPLAYLEVEL(2, " - limit decompression speed %u MB/s", target.dSpeed >> 20); } + if(target.cMem != (U32)-1) { DISPLAYLEVEL(2, " - limit memory %u MB", target.cMem >> 20); } + + DISPLAYLEVEL(2, "\n"); + findClockGranularity(); + + { + paramValues_t CParams; + + /* find best solution from default params */ + { + /* strategy selection */ + const int maxSeeds = g_noSeed ? 1 : ZSTD_maxCLevel(); + DEBUGOUTPUT("Strategy Selection\n"); + if(paramTarget.vals[strt_ind] == PARAM_UNSET) { + BMK_result_t candidate; + int i; + for (i=1; i<=maxSeeds; i++) { + int ec; + CParams = overwriteParams(cParamsToPVals(ZSTD_getCParams(i, buf.maxBlockSize, ctx.dictSize)), paramTarget); + ec = BMK_benchParam(&candidate, buf, ctx, CParams); + BMK_printWinnerOpt(stdout, i, candidate, CParams, target, buf.srcSize); + + if(!ec && compareResultLT(winner.result, candidate, relaxTarget(target), buf.srcSize)) { + winner.result = candidate; + winner.params = CParams; + } + + CHECKTIMEGT(ret, 0, _displayCleanUp); /* if pass time limit, stop */ + /* if the current params are too slow, just stop. */ + if(target.cSpeed > candidate.cSpeed * 3 / 2) { break; } + } + + BMK_printWinnerOpt(stdout, CUSTOM_LEVEL, winner.result, winner.params, target, buf.srcSize); + } + } + + DEBUGOUTPUT("Real Opt\n"); + /* start 'real' optimization */ + { + int bestStrategy = (int)winner.params.vals[strt_ind]; + if(paramTarget.vals[strt_ind] == PARAM_UNSET) { + int st = bestStrategy; + int tries = g_maxTries; + + { + /* one iterations of hill climbing with the level-defined parameters. */ + winnerInfo_t w1 = climbOnce(target, allMT, buf, ctx, winner.params); + if(compareResultLT(winner.result, w1.result, target, buf.srcSize)) { + winner = w1; + } + CHECKTIMEGT(ret, 0, _displayCleanUp); + } + + while(st && tries > 0) { + winnerInfo_t wc; + DEBUGOUTPUT("StrategySwitch: %s\n", g_stratName[st]); + + wc = optimizeFixedStrategy(buf, ctx, target, paramBase, st, allMT, tries); + + if(compareResultLT(winner.result, wc.result, target, buf.srcSize)) { + winner = wc; + tries = g_maxTries; + bestStrategy = st; + } else { + st = nextStrategy(st, bestStrategy); + tries -= TRY_DECAY; + } + CHECKTIMEGT(ret, 0, _displayCleanUp); + } + } else { + winner = optimizeFixedStrategy(buf, ctx, target, paramBase, paramTarget.vals[strt_ind], allMT, g_maxTries); + } + + } + + /* no solution found */ + if(winner.result.cSize == (size_t)-1) { + ret = 1; + DISPLAY("No feasible solution found\n"); + goto _cleanUp; + } + /* end summary */ +_displayCleanUp: + if(g_displayLevel >= 0) { BMK_displayOneResult(stdout, winner, buf.srcSize); } + BMK_translateAdvancedParams(stdout, winner.params); + DISPLAYLEVEL(1, "grillParams size - optimizer completed \n"); + + } +_cleanUp: + freeContexts(ctx); + freeBuffers(buf); + freeMemoTableArray(allMT); + return ret; +} + +/*-************************************ +* CLI parsing functions +**************************************/ + +/** longCommandWArg() : + * check if *stringPtr is the same as longCommand. + * If yes, @return 1 and advances *stringPtr to the position which immediately follows longCommand. + * @return 0 and doesn't modify *stringPtr otherwise. + * from zstdcli.c + */ +static unsigned longCommandWArg(const char** stringPtr, const char* longCommand) +{ + size_t const comSize = strlen(longCommand); + int const result = !strncmp(*stringPtr, longCommand, comSize); + if (result) *stringPtr += comSize; + return result; } static void errorOut(const char* msg) @@ -772,7 +2393,9 @@ static void errorOut(const char* msg) static unsigned readU32FromChar(const char** stringPtr) { const char errorMsg[] = "error: numeric value too large"; + unsigned sign = 1; unsigned result = 0; + if(**stringPtr == '-') { sign = (unsigned)-1; (*stringPtr)++; } while ((**stringPtr >='0') && (**stringPtr <='9')) { unsigned const max = (((unsigned)(-1)) / 10) - 1; if (result > max) errorOut(errorMsg); @@ -790,6 +2413,22 @@ static unsigned readU32FromChar(const char** stringPtr) if (**stringPtr=='i') (*stringPtr)++; if (**stringPtr=='B') (*stringPtr)++; } + return result * sign; +} + +static double readDoubleFromChar(const char** stringPtr) +{ + double result = 0, divide = 10; + while ((**stringPtr >='0') && (**stringPtr <='9')) { + result *= 10, result += **stringPtr - '0', (*stringPtr)++ ; + } + if(**stringPtr!='.') { + return result; + } + (*stringPtr)++; + while ((**stringPtr >='0') && (**stringPtr <='9')) { + result += (double)(**stringPtr - '0') / divide, divide *= 10, (*stringPtr)++ ; + } return result; } @@ -806,15 +2445,16 @@ static int usage(const char* exename) static int usage_advanced(void) { DISPLAY( "\nAdvanced options :\n"); - DISPLAY( " -T# : set level 1 speed objective \n"); - DISPLAY( " -B# : cut input into blocks of size # (default : single block) \n"); - DISPLAY( " -i# : iteration loops [1-9](default : %i) \n", NBLOOPS); - DISPLAY( " -O# : find Optimized parameters for # MB/s compression speed (default : 0) \n"); - DISPLAY( " -S : Single run \n"); - DISPLAY( " --zstd : Single run, parameter selection same as zstdcli \n"); - DISPLAY( " -P# : generated sample compressibility (default : %.1f%%) \n", COMPRESSIBILITY_DEFAULT * 100); - DISPLAY( " -t# : Caps runtime of operation in seconds (default : %u seconds (%.1f hours)) \n", (U32)g_grillDuration_s, g_grillDuration_s / 3600); - DISPLAY( " -v : Prints Benchmarking output\n"); + DISPLAY( " -T# : set level 1 speed objective \n"); + DISPLAY( " -B# : cut input into blocks of size # (default : single block) \n"); + DISPLAY( " --optimize= : same as -O with more verbose syntax (see README.md)\n"); + DISPLAY( " -S : Single run \n"); + DISPLAY( " --zstd : Single run, parameter selection same as zstdcli \n"); + DISPLAY( " -P# : generated sample compressibility (default : %.1f%%) \n", COMPRESSIBILITY_DEFAULT * 100); + DISPLAY( " -t# : Caps runtime of operation in seconds (default : %u seconds (%.1f hours)) \n", g_timeLimit_s, (double)g_timeLimit_s / 3600); + DISPLAY( " -v : Prints Benchmarking output\n"); + DISPLAY( " -D : Next argument dictionary file\n"); + DISPLAY( " -s : Seperate Files\n"); return 0; } @@ -825,41 +2465,87 @@ static int badusage(const char* exename) return 1; } +#define PARSE_SUB_ARGS(stringLong, stringShort, variable) { if (longCommandWArg(&argument, stringLong) || longCommandWArg(&argument, stringShort)) { variable = readU32FromChar(&argument); if (argument[0]==',') { argument++; continue; } else break; } } +/* 1 if successful parse, 0 otherwise */ +static int parse_params(const char** argptr, paramValues_t* pv) { + int matched = 0; + const char* argOrig = *argptr; + varInds_t v; + for(v = 0; v < NUM_PARAMS; v++) { + if(longCommandWArg(argptr,g_shortParamNames[v]) || longCommandWArg(argptr, g_paramNames[v])) { + if(**argptr == '=') { + (*argptr)++; + pv->vals[v] = readU32FromChar(argptr); + matched = 1; + break; + } + } + /* reset and try again */ + *argptr = argOrig; + } + return matched; +} + +/*-************************************ +* Main +**************************************/ + int main(int argc, const char** argv) { int i, filenamesStart=0, result; const char* exename=argv[0]; - const char* input_filename=0; - U32 optimizer = 0; + const char* input_filename = NULL; + const char* dictFileName = NULL; U32 main_pause = 0; - U32 targetSpeed = 0; + int cLevelOpt = 0, cLevelRun = 0; + int seperateFiles = 0; + double compressibility = COMPRESSIBILITY_DEFAULT; + U32 memoTableLog = PARAM_UNSET; + constraint_t target = { 0, 0, (U32)-1 }; + + paramValues_t paramTarget = emptyParams(); + g_params = emptyParams(); assert(argc>=1); /* for exename */ - /* Welcome message */ - DISPLAY(WELCOME_MESSAGE); - for(i=1; i