From 65385d540e9cae07c5bfa92fd6847b4c8b9da190 Mon Sep 17 00:00:00 2001 From: LoRd_MuldeR Date: Sat, 20 Feb 2016 15:59:45 +0100 Subject: [PATCH] Added support for "AVX" CPU flag. --- src/Main.cpp | 2 +- src/Thread_Initialization.cpp | 22 +++++++++++++++----- src/Tools.h | 38 +++++++++++++++++++++-------------- 3 files changed, 41 insertions(+), 21 deletions(-) diff --git a/src/Main.cpp b/src/Main.cpp index c604a35c..4f8febc1 100644 --- a/src/Main.cpp +++ b/src/Main.cpp @@ -210,7 +210,7 @@ static int lamexp_main(int &argc, char **argv) qDebug(" CPU vendor id : %s (Intel=%s)", cpuFeatures.vendor, MUTILS_BOOL2STR(cpuFeatures.intel)); qDebug("CPU brand string : %s", cpuFeatures.brand); qDebug(" CPU signature : Family=%d Model=%d Stepping=%d", cpuFeatures.family, cpuFeatures.model, cpuFeatures.stepping); - qDebug("CPU capabilities : MMX=%s SSE=%s SSE2=%s SSE3=%s SSSE3=%s SSE4=%s SSE4.2=%s x64=%s", MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_MMX), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE2), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE3), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSSE3), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE4), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE42), MUTILS_BOOL2STR(cpuFeatures.x64)); + qDebug("CPU capabilities : SSE=%s SSE2=%s SSE3=%s SSSE3=%s SSE4=%s SSE4.2=%s AVX=%s x64=%s", MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE2), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE3), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSSE3), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE4), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE42), MUTILS_BOOL2STR(cpuFeatures.features & MUtils::CPUFetaures::FLAG_AVX), MUTILS_BOOL2STR(cpuFeatures.x64)); qDebug(" Number of CPU's : %d\n", cpuFeatures.count); //Initialize Qt diff --git a/src/Thread_Initialization.cpp b/src/Thread_Initialization.cpp index 0a92b72b..7a139560 100644 --- a/src/Thread_Initialization.cpp +++ b/src/Thread_Initialization.cpp @@ -353,7 +353,14 @@ double InitializationThread::doInit(const size_t threadCount) unsigned int cpuSupport = 0; if((m_cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE) && (m_cpuFeatures.features & MUtils::CPUFetaures::FLAG_SSE2) && m_cpuFeatures.intel) { - cpuSupport = m_cpuFeatures.x64 ? CPU_TYPE_X64_SSE : CPU_TYPE_X86_SSE; + if (m_cpuFeatures.features & MUtils::CPUFetaures::FLAG_AVX) + { + cpuSupport = m_cpuFeatures.x64 ? CPU_TYPE_X64_AVX : CPU_TYPE_X86_AVX; + } + else + { + cpuSupport = m_cpuFeatures.x64 ? CPU_TYPE_X64_SSE : CPU_TYPE_X86_SSE; + } } else { @@ -375,8 +382,10 @@ double InitializationThread::doInit(const size_t threadCount) { PRINT_CPU_TYPE(CPU_TYPE_X86_GEN); break; PRINT_CPU_TYPE(CPU_TYPE_X86_SSE); break; + PRINT_CPU_TYPE(CPU_TYPE_X86_AVX); break; PRINT_CPU_TYPE(CPU_TYPE_X64_GEN); break; PRINT_CPU_TYPE(CPU_TYPE_X64_SSE); break; + PRINT_CPU_TYPE(CPU_TYPE_X64_AVX); break; default: MUTILS_THROW("CPU support undefined!"); } @@ -729,20 +738,23 @@ void InitAacEncTask::initAacEncImpl(const char *const toolName, const char *cons void InitializationThread::selfTest(void) { - const unsigned int cpu[4] = {CPU_TYPE_X86_GEN, CPU_TYPE_X86_SSE, CPU_TYPE_X64_GEN, CPU_TYPE_X64_SSE}; + const unsigned int cpu[7] = {CPU_TYPE_X86_GEN, CPU_TYPE_X86_SSE, CPU_TYPE_X86_AVX, CPU_TYPE_X64_GEN, CPU_TYPE_X64_SSE, CPU_TYPE_X64_AVX, 0 }; FileHash::selfTest(); - for(size_t k = 0; k < 4; k++) + for(size_t k = 0; cpu[k]; k++) { qDebug("[TEST]"); switch(cpu[k]) { PRINT_CPU_TYPE(CPU_TYPE_X86_GEN); break; PRINT_CPU_TYPE(CPU_TYPE_X86_SSE); break; + PRINT_CPU_TYPE(CPU_TYPE_X86_AVX); break; PRINT_CPU_TYPE(CPU_TYPE_X64_GEN); break; PRINT_CPU_TYPE(CPU_TYPE_X64_SSE); break; - default: MUTILS_THROW("CPU support undefined!"); + PRINT_CPU_TYPE(CPU_TYPE_X64_AVX); break; + default: + MUTILS_THROW("CPU support undefined!"); } unsigned int n = 0; for(int i = 0; true; i++) @@ -780,7 +792,7 @@ void InitializationThread::selfTest(void) } if(n != EXPECTED_TOOL_COUNT) { - qFatal("Tool count mismatch for CPU type %u !!!", cpu[k]); + qFatal("Tool count mismatch for CPU type %u. Should be %u, but got %u !!!", cpu[k], EXPECTED_TOOL_COUNT, n); } qDebug("Done.\n"); } diff --git a/src/Tools.h b/src/Tools.h index da72a792..55da5bbc 100644 --- a/src/Tools.h +++ b/src/Tools.h @@ -33,15 +33,21 @@ /* CPU_TYPE__ */ #define CPU_TYPE_X86_GEN 0x00000001UL //x86, generic #define CPU_TYPE_X86_SSE 0x00000002UL //x86, with SSE and SSE2 support - Intel only! -#define CPU_TYPE_X64_GEN 0x00000004UL //x64, generic -#define CPU_TYPE_X64_SSE 0x00000008UL //x64, with SSE and SSE2 support - Intel only! +#define CPU_TYPE_X86_AVX 0x00000004UL //x86, with AVX support +#define CPU_TYPE_X64_GEN 0x00000008UL //x64, generic +#define CPU_TYPE_X64_SSE 0x00000010UL //x64, with SSE and SSE2 support - Intel only! +#define CPU_TYPE_X64_AVX 0x00000020UL //x64, with AVX support /* combined CPU types */ -#define CPU_TYPE_X86_ALL (CPU_TYPE_X86_GEN|CPU_TYPE_X86_SSE) //all x86 (ignore SSE/SSE2 support) -#define CPU_TYPE_X64_ALL (CPU_TYPE_X64_GEN|CPU_TYPE_X64_SSE) //all x64 (ignore SSE/SSE2 support) -#define CPU_TYPE_ALL_GEN (CPU_TYPE_X86_GEN|CPU_TYPE_X64_GEN) //all generic (ignore x86/x64) -#define CPU_TYPE_ALL_SSE (CPU_TYPE_X86_SSE|CPU_TYPE_X64_SSE) //all with SSE and SSE2 (ignore x86/x64) -#define CPU_TYPE_ALL_ALL (CPU_TYPE_X86_ALL|CPU_TYPE_X64_ALL) //use always, no exceptions +#define CPU_TYPE_X86_SSX (CPU_TYPE_X86_SSE|CPU_TYPE_X86_AVX) +#define CPU_TYPE_X86_ALL (CPU_TYPE_X86_GEN|CPU_TYPE_X86_SSX) +#define CPU_TYPE_X64_SSX (CPU_TYPE_X64_SSE|CPU_TYPE_X64_AVX) +#define CPU_TYPE_X64_ALL (CPU_TYPE_X64_GEN|CPU_TYPE_X64_SSX) +#define CPU_TYPE_ALL_GEN (CPU_TYPE_X86_GEN|CPU_TYPE_X64_GEN) +#define CPU_TYPE_ALL_SSE (CPU_TYPE_X86_SSE|CPU_TYPE_X64_SSE) +#define CPU_TYPE_ALL_SSX (CPU_TYPE_X86_SSX|CPU_TYPE_X64_SSX) +#define CPU_TYPE_ALL_AVX (CPU_TYPE_X86_AVX|CPU_TYPE_X64_AVX) +#define CPU_TYPE_ALL_ALL (CPU_TYPE_X86_ALL|CPU_TYPE_X64_ALL) //////////////////////////////////////////////////////////// // TOOLS @@ -58,7 +64,7 @@ static const struct g_lamexp_tools[] = { {"4870d7abbc3995dc5ca82147ee33f28c3ee7dc1986fbb0131912effe10c1a209d3f983c5dffd963c31707a5ce39611b4", CPU_TYPE_X86_GEN, "aften.i386.exe", 8, ""}, - {"fc4e38b11a0f52b68cca79aa2d71c02180b70eb0e592923dee4f0ccf766f1006642369b2178f6a61d1c2506446cc442d", CPU_TYPE_X86_SSE, "aften.sse2.exe", 8, ""}, + {"fc4e38b11a0f52b68cca79aa2d71c02180b70eb0e592923dee4f0ccf766f1006642369b2178f6a61d1c2506446cc442d", CPU_TYPE_X86_SSX, "aften.sse2.exe", 8, ""}, {"b44b1c0e1a8f4633c9bc2e43d0a83ea5edc890e1cccc0054c812d0e64170e5292dc1dfd7430da61b73150831aa16e9c4", CPU_TYPE_X64_ALL, "aften.x64.exe", 8, ""}, {"4558728999a05f12fe88566e8308cba0ada200814c2a1bfe2507f49faf8f3994b0d52a829148f5c6321d24faa2718439", CPU_TYPE_ALL_ALL, "avs2wav.exe", 13, ""}, {"07834b1b8ecac2f3db79ad048005eca3a284cb759e15d399eb1a560a403122e20db5f08e60319759b8463aeef2410d02", CPU_TYPE_ALL_ALL, "dcaenc.exe", 20120419, ""}, @@ -69,21 +75,23 @@ g_lamexp_tools[] = {"9d2008d06b731b0d88377e9b2ceda30826ee787fe09ae8ef98084d5ef80655199e5069672366314b67a4285b56a282f0", CPU_TYPE_ALL_ALL, "gpgv.exe", 1419, ""}, {"c66992384b6388a7b68f2659624c150b33ff1cacb0c28c0ab2a3249eca4c31e204953f325403f78c077d1735b275ac12", CPU_TYPE_ALL_ALL, "keyring.gpg", UINT_MAX, ""}, {"53cfab3896a47d48f523315f475fa07856d468ad1aefcc8cce19c18cdf509e2f92840dab92a442995df36d941cb7a6ca", CPU_TYPE_ALL_GEN, "lame.i386.exe", 3995, "Final"}, - {"9511e7ef2ad10de05386eedf7f14d637edab894a53dacd2f8f15c6f8ed582f12c25fb5bf88438e62c46b8eb92e7634b2", CPU_TYPE_ALL_SSE, "lame.sse2.exe", 3995, "Final"}, + {"9511e7ef2ad10de05386eedf7f14d637edab894a53dacd2f8f15c6f8ed582f12c25fb5bf88438e62c46b8eb92e7634b2", CPU_TYPE_ALL_SSX, "lame.sse2.exe", 3995, "Final"}, {"51af33a33bd9bde6ca6e9c612346b9c138584de59a92bc14a680729066fd849fd83fe668f79e2a06ffe117f39993d10a", CPU_TYPE_ALL_GEN, "mac.i386.exe", 416, ""}, - {"117e60cf4445cec0a777d2636091a13c50be5c6ed28afaa75ab09ab4576942f1a8584b36d3abac9db7595dfb0d90207e", CPU_TYPE_ALL_SSE, "mac.sse2.exe", 416, ""}, + {"117e60cf4445cec0a777d2636091a13c50be5c6ed28afaa75ab09ab4576942f1a8584b36d3abac9db7595dfb0d90207e", CPU_TYPE_ALL_SSX, "mac.sse2.exe", 416, ""}, {"2651dcd3c85b76bf1bbe5520e0e76b9d8db7459a1eb7fc1d49dc32c22dcf99957bcba8961b70c194ea3d4346d8c4118f", CPU_TYPE_X86_ALL, "mediainfo.i386.exe", 782, ""}, {"ee5990856760c5bd73332ddb88a126861f03ffa8070e4e84812a3f63da3387582641c61e692d9b3f432ac1c101e8704f", CPU_TYPE_X64_ALL, "mediainfo.x64.exe", 782, ""}, {"7e6346a057634ff07b2e1f427035324f7f02100cc996425990f87f71d767fce4c7b101588c7d944ba49cb2d7e51c9bdb", CPU_TYPE_ALL_ALL, "mpcdec.exe", 475, ""}, {"02f3e52fca662c97fc15ee5d5c6ded2966d716ec5e5ab958accd4761eb864cc7c62ace97198b876a78cf75ab11c555b2", CPU_TYPE_ALL_ALL, "mpg123.exe", 1224, ""}, {"f1f2ea5c9e5539620b706e7af68e543bf7a731afb06ccce3815ab34dad64d697e4d6ffcd187a396619b8b52efe7edf88", CPU_TYPE_ALL_ALL, "oggdec.exe", 1101, ""}, {"245181321625445ac42fce31d64bf03872e77e2d0dd3c19d6c17ca2771354f096a6040827dd6d00ffd7342c7dd26168e", CPU_TYPE_X86_GEN, "oggenc2.i386.exe", 288135603, "2015"}, - {"512b8efcd1003a0f67220a450d6ea4466194e8fd49fc090a69b15a858db11499acbf98f984530cd5d37b4b6abdd1c6d8", CPU_TYPE_X86_SSE, "oggenc2.sse2.exe", 288135603, "2015"}, + {"512b8efcd1003a0f67220a450d6ea4466194e8fd49fc090a69b15a858db11499acbf98f984530cd5d37b4b6abdd1c6d8", CPU_TYPE_X86_SSX, "oggenc2.sse2.exe", 288135603, "2015"}, {"a07ef67cba5a00d335d07372baf76d4d0573b425afce71a19c1e04eaabbe3f55e60bdd40af5e428224c91df1823eda08", CPU_TYPE_X64_ALL, "oggenc2.x64.exe", 288135603, "2015"}, - {"d9949ab989fa412bffb08c47f431b74a66223aa584beb9a6aeaa66885932c2ed251798748cbc489e984f0acce0e0053c", CPU_TYPE_ALL_GEN, "opusdec.i386.exe", 20150326, "v1.1"}, - {"e3fc96044491bd96734dc25c8fdcb1760ee86a8fd47ad74481999ae18593d0647f0d4a45f058950b9def0903e8a30fc3", CPU_TYPE_ALL_SSE, "opusdec.sse2.exe", 20150326, "v1.1"}, - {"ae777a525a4670df1deca0483c5087f129d8131eaf946b2cd72fa96ab65db7fb600766448d28caf2102d97b3fa26e6bc", CPU_TYPE_ALL_GEN, "opusenc.i386.exe", 20150326, "v1.1"}, - {"8eadcdfe01a6ff2d88b6cfdf203c6eae6f858bd17b894644fa1d78d293235d8dc21b0102b8ca3d48f718251b3f2e9e5a", CPU_TYPE_ALL_SSE, "opusenc.sse2.exe", 20150326, "v1.1"}, + {"563279aa9e06e45bbd15c52e012ca9bf7c70b8df5c608ba46006d59cd9a43a23bf392903c11e8f9767d330e0066b9f8b", CPU_TYPE_ALL_GEN, "opusdec.i386.exe", 20160220, "v1.1.2"}, + {"3c559169f327b98209ecf435849e68a2e1cbb7688d29cf19839dfd9499f7df64bfb6c8c91ba0b2bef45724b4fd04ef1e", CPU_TYPE_ALL_SSE, "opusdec.sse2.exe", 20160220, "v1.1.2"}, + {"42344b6838fdc232e323c4930a0640a1b4f11fd39cadd602d3b6dabd7c4ffadb201398483b6457c9c540e9b9b4a9ca41", CPU_TYPE_ALL_AVX, "opusdec.avx.exe", 20160220, "v1.1.2" }, + {"ed453ffec3a6e998a7dcbf45c29bb3b83dc4897b7845faf1bbe305b3fa75582c233ba2141027a47f69c496c9ff64b3f5", CPU_TYPE_ALL_GEN, "opusenc.i386.exe", 20160220, "v1.1.2"}, + {"3d08b877816f15991b03d967f4134b2817314a75e79be7b9ec830d3b4e14df816b356b31907b2f525ae5ac3f2131aef1", CPU_TYPE_ALL_SSE, "opusenc.sse2.exe", 20160220, "v1.1.2"}, + {"62b7ef1d2d5c87733a20d67939ce812c5a84544bc85ab30c7a7702daba41118f53a572f9b6995e6ca6d3d06f30ec8512", CPU_TYPE_ALL_AVX, "opusenc.avx.exe", 20160220, "v1.1.2" }, {"bd4c8cd7d32869f8ed332b51349147c9b987a7e552aadfd15a312e5956efc9cc0c8ee9ed8283240b049c8f5c8fc3ad9b", CPU_TYPE_X86_ALL, "refalac.i386.exe", 158, ""}, {"5821414021725248f19e71969348d04a561d81668114cabe5259e9cd65867337a8176e0f41a3cf7d2f88693e5df8850b", CPU_TYPE_X64_ALL, "refalac.x64.exe", 158, ""}, {"d041b60de6c5c6e77cbad84440db57bbeb021af59dd0f7bebd3ede047d9e2ddc2a0c14179472687ba91063743d23e337", CPU_TYPE_ALL_ALL, "shorten.exe", 361, ""},