diff --git a/CMakeLists.txt b/CMakeLists.txt index 99a19bfdf..3d0042afe 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -5,6 +5,7 @@ option(WITH_LIBCPUID "Use Libcpuid" ON) option(WITH_AEON "CryptoNight-Lite support" ON) option(WITH_SUMO "CryptoNight-Heavy support" ON) option(WITH_CN_PICO "CryptoNight-Pico support" ON) +option(WITH_CN_GPU "CryptoNight-GPU support" ON) option(WITH_HTTPD "HTTP REST API" ON) option(WITH_DEBUG_LOG "Enable debug log output" OFF) option(WITH_TLS "Enable OpenSSL support" ON) @@ -200,6 +201,7 @@ endif() include(cmake/OpenSSL.cmake) include(cmake/asm.cmake) +include(cmake/cn-gpu.cmake) CHECK_INCLUDE_FILE (syslog.h HAVE_SYSLOG_H) if (HAVE_SYSLOG_H) @@ -262,5 +264,5 @@ if (WITH_DEBUG_LOG) add_definitions(/DAPP_DEBUG) endif() -add_executable(${CMAKE_PROJECT_NAME} ${HEADERS} ${SOURCES} ${SOURCES_OS} ${SOURCES_CPUID} ${HEADERS_CRYPTO} ${SOURCES_CRYPTO} ${SOURCES_SYSLOG} ${HTTPD_SOURCES} ${TLS_SOURCES} ${XMRIG_ASM_SOURCES}) +add_executable(${CMAKE_PROJECT_NAME} ${HEADERS} ${SOURCES} ${SOURCES_OS} ${SOURCES_CPUID} ${HEADERS_CRYPTO} ${SOURCES_CRYPTO} ${SOURCES_SYSLOG} ${HTTPD_SOURCES} ${TLS_SOURCES} ${XMRIG_ASM_SOURCES} ${CN_GPU_SOURCES}) target_link_libraries(${CMAKE_PROJECT_NAME} ${XMRIG_ASM_LIBRARY} ${OPENSSL_LIBRARIES} ${UV_LIBRARIES} ${MHD_LIBRARY} ${EXTRA_LIBS} ${CPUID_LIB}) diff --git a/cmake/cn-gpu.cmake b/cmake/cn-gpu.cmake new file mode 100644 index 000000000..81092519b --- /dev/null +++ b/cmake/cn-gpu.cmake @@ -0,0 +1,25 @@ +if (WITH_CN_GPU AND CMAKE_SIZEOF_VOID_P EQUAL 8) + + if (XMRIG_ARM) + set(CN_GPU_SOURCES src/crypto/cn_gpu_arm.cpp) + + if (CMAKE_CXX_COMPILER_ID MATCHES GNU OR CMAKE_CXX_COMPILER_ID MATCHES Clang) + set_source_files_properties(src/crypto/cn_gpu_arm.cpp PROPERTIES COMPILE_FLAGS "-O3") + endif() + else() + set(CN_GPU_SOURCES src/crypto/cn_gpu_avx.cpp src/crypto/cn_gpu_ssse3.cpp) + + if (CMAKE_CXX_COMPILER_ID MATCHES GNU) + set_source_files_properties(src/crypto/cn_gpu_avx.cpp PROPERTIES COMPILE_FLAGS "-O3 -mavx2") + set_source_files_properties(src/crypto/cn_gpu_ssse3.cpp PROPERTIES COMPILE_FLAGS "-O3") + elseif (CMAKE_CXX_COMPILER_ID MATCHES Clang) + set_source_files_properties(src/crypto/cn_gpu_avx.cpp PROPERTIES COMPILE_FLAGS "-mavx2") + elseif (CMAKE_CXX_COMPILER_ID MATCHES MSVC) + set_source_files_properties(src/crypto/cn_gpu_avx.cpp PROPERTIES COMPILE_FLAGS "/arch:AVX") + endif() + endif() +else() + set(CN_GPU_SOURCES "") + + add_definitions(/DXMRIG_NO_CN_GPU) +endif() diff --git a/src/Summary.cpp b/src/Summary.cpp index f010d70ba..dc671438e 100644 --- a/src/Summary.cpp +++ b/src/Summary.cpp @@ -4,9 +4,9 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2017-2018 XMR-Stak , - * Copyright 2018 SChernykh - * Copyright 2016-2018 XMRig , + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig * * This program is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -74,18 +74,25 @@ static void print_cpu(xmrig::Config *config) using namespace xmrig; if (config->isColors()) { - Log::i()->text(GREEN_BOLD(" * ") WHITE_BOLD("%-13s%s (%d)") " %sx64 %sAES", + Log::i()->text(GREEN_BOLD(" * ") WHITE_BOLD("%-13s%s (%d)") " %sx64 %sAES %sAVX2", "CPU", Cpu::info()->brand(), Cpu::info()->sockets(), - Cpu::info()->isX64() ? "\x1B[1;32m" : "\x1B[1;31m-", - Cpu::info()->hasAES() ? "\x1B[1;32m" : "\x1B[1;31m-"); + Cpu::info()->isX64() ? "\x1B[1;32m" : "\x1B[1;31m-", + Cpu::info()->hasAES() ? "\x1B[1;32m" : "\x1B[1;31m-", + Cpu::info()->hasAVX2() ? "\x1B[1;32m" : "\x1B[1;31m-"); # ifndef XMRIG_NO_LIBCPUID Log::i()->text(GREEN_BOLD(" * ") WHITE_BOLD("%-13s%.1f MB/%.1f MB"), "CPU L2/L3", Cpu::info()->L2() / 1024.0, Cpu::info()->L3() / 1024.0); # endif } else { - Log::i()->text(" * %-13s%s (%d) %sx64 %sAES", "CPU", Cpu::info()->brand(), Cpu::info()->sockets(), Cpu::info()->isX64() ? "" : "-", Cpu::info()->hasAES() ? "" : "-"); + Log::i()->text(" * %-13s%s (%d) %sx64 %sAES %sAVX2", + "CPU", + Cpu::info()->brand(), + Cpu::info()->sockets(), + Cpu::info()->isX64() ? "" : "-", + Cpu::info()->hasAES() ? "" : "-", + Cpu::info()->hasAVX2() ? "" : "-"); # ifndef XMRIG_NO_LIBCPUID Log::i()->text(" * %-13s%.1f MB/%.1f MB", "CPU L2/L3", Cpu::info()->L2() / 1024.0, Cpu::info()->L3() / 1024.0); # endif diff --git a/src/common/cpu/BasicCpuInfo.cpp b/src/common/cpu/BasicCpuInfo.cpp index cb1e6d1de..e9018c98a 100644 --- a/src/common/cpu/BasicCpuInfo.cpp +++ b/src/common/cpu/BasicCpuInfo.cpp @@ -4,8 +4,9 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2017-2018 XMR-Stak , - * Copyright 2016-2018 XMRig , + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig * * This program is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -35,6 +36,10 @@ # define bit_AES (1 << 25) #endif +#ifndef bit_AVX2 +# define bit_AVX2 (1 << 5) +#endif + #include "common/cpu/BasicCpuInfo.h" @@ -93,9 +98,19 @@ static inline bool has_aes_ni() } +static inline bool has_avx2() +{ + int32_t cpu_info[4] = { 0 }; + cpuid(EXTENDED_FEATURES, cpu_info); + + return (cpu_info[EBX_Reg] & bit_AVX2) != 0; +} + + xmrig::BasicCpuInfo::BasicCpuInfo() : m_assembly(ASM_NONE), m_aes(has_aes_ni()), + m_avx2(has_avx2()), m_brand(), m_threads(std::thread::hardware_concurrency()) { diff --git a/src/common/cpu/BasicCpuInfo.h b/src/common/cpu/BasicCpuInfo.h index 911674ea1..95857ed27 100644 --- a/src/common/cpu/BasicCpuInfo.h +++ b/src/common/cpu/BasicCpuInfo.h @@ -4,8 +4,9 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2017-2018 XMR-Stak , - * Copyright 2016-2018 XMRig , + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig * * This program is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -41,6 +42,7 @@ protected: inline Assembly assembly() const override { return m_assembly; } inline bool hasAES() const override { return m_aes; } + inline bool hasAVX2() const override { return m_avx2; } inline bool isSupported() const override { return true; } inline const char *brand() const override { return m_brand; } inline int32_t cores() const override { return -1; } @@ -59,6 +61,7 @@ protected: private: Assembly m_assembly; bool m_aes; + bool m_avx2; char m_brand[64]; int32_t m_threads; }; diff --git a/src/common/cpu/BasicCpuInfo_arm.cpp b/src/common/cpu/BasicCpuInfo_arm.cpp index c1c127db6..e1df86df4 100644 --- a/src/common/cpu/BasicCpuInfo_arm.cpp +++ b/src/common/cpu/BasicCpuInfo_arm.cpp @@ -4,8 +4,9 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2017-2018 XMR-Stak , - * Copyright 2016-2018 XMRig , + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig * * This program is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -30,6 +31,7 @@ xmrig::BasicCpuInfo::BasicCpuInfo() : m_aes(false), + m_avx2(false), m_brand(), m_threads(std::thread::hardware_concurrency()) { diff --git a/src/common/crypto/Algorithm.cpp b/src/common/crypto/Algorithm.cpp index 6d408cd2f..9d424c575 100644 --- a/src/common/crypto/Algorithm.cpp +++ b/src/common/crypto/Algorithm.cpp @@ -86,6 +86,10 @@ static AlgoData const algorithms[] = { { "cryptonight-ultralite", "cn-ultralite", xmrig::CRYPTONIGHT_PICO, xmrig::VARIANT_TRTL }, { "cryptonight_turtle", "cn_turtle", xmrig::CRYPTONIGHT_PICO, xmrig::VARIANT_TRTL }, # endif + +# ifndef XMRIG_NO_CN_GPU + { "cryptonight/gpu", "cn/gpu", xmrig::CRYPTONIGHT, xmrig::VARIANT_GPU }, +# endif }; @@ -120,7 +124,8 @@ static const char *variants[] = { "rto", "2", "half", - "trtl" + "trtl", + "gpu" }; diff --git a/src/common/interfaces/ICpuInfo.h b/src/common/interfaces/ICpuInfo.h index 267616d0a..dd4034b33 100644 --- a/src/common/interfaces/ICpuInfo.h +++ b/src/common/interfaces/ICpuInfo.h @@ -4,7 +4,9 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2016-2018 XMRig + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig * * This program is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -40,6 +42,7 @@ public: virtual ~ICpuInfo() {} virtual bool hasAES() const = 0; + virtual bool hasAVX2() const = 0; virtual bool isSupported() const = 0; virtual bool isX64() const = 0; virtual const char *brand() const = 0; diff --git a/src/common/net/Pool.cpp b/src/common/net/Pool.cpp index 7e8fa1dc1..2010aade6 100644 --- a/src/common/net/Pool.cpp +++ b/src/common/net/Pool.cpp @@ -4,7 +4,7 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2017-2018 XMR-Stak , + * Copyright 2017-2019 XMR-Stak , * Copyright 2018-2019 SChernykh * Copyright 2016-2019 XMRig , * @@ -419,6 +419,7 @@ void Pool::rebuild() addVariant(xmrig::VARIANT_XHV); addVariant(xmrig::VARIANT_XAO); addVariant(xmrig::VARIANT_RTO); + addVariant(xmrig::VARIANT_GPU); addVariant(xmrig::VARIANT_AUTO); # endif } diff --git a/src/common/xmrig.h b/src/common/xmrig.h index b20513318..d738028f0 100644 --- a/src/common/xmrig.h +++ b/src/common/xmrig.h @@ -73,6 +73,7 @@ enum Variant { VARIANT_2 = 8, // CryptoNight variant 2 VARIANT_HALF = 9, // CryptoNight variant 2 with half iterations (Masari/Stellite) VARIANT_TRTL = 10, // CryptoNight Turtle (TRTL) + VARIANT_GPU = 11, // CryptoNight-GPU (Ryo) VARIANT_MAX }; diff --git a/src/core/cpu/AdvancedCpuInfo.cpp b/src/core/cpu/AdvancedCpuInfo.cpp index a91521361..c41378e53 100644 --- a/src/core/cpu/AdvancedCpuInfo.cpp +++ b/src/core/cpu/AdvancedCpuInfo.cpp @@ -4,9 +4,9 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2017-2018 XMR-Stak , - * Copyright 2018 SChernykh - * Copyright 2016-2018 XMRig , + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig * * This program is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -33,6 +33,7 @@ xmrig::AdvancedCpuInfo::AdvancedCpuInfo() : m_assembly(ASM_NONE), m_aes(false), + m_avx2(false), m_L2_exclusive(false), m_brand(), m_cores(0), @@ -83,6 +84,8 @@ xmrig::AdvancedCpuInfo::AdvancedCpuInfo() : m_assembly = ASM_INTEL; } } + + m_avx2 = data.flags[CPU_FEATURE_AVX2]; } diff --git a/src/core/cpu/AdvancedCpuInfo.h b/src/core/cpu/AdvancedCpuInfo.h index 5e8967ad2..0765da339 100644 --- a/src/core/cpu/AdvancedCpuInfo.h +++ b/src/core/cpu/AdvancedCpuInfo.h @@ -4,8 +4,9 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2017-2018 XMR-Stak , - * Copyright 2016-2018 XMRig , + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig * * This program is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -41,6 +42,7 @@ protected: inline Assembly assembly() const override { return m_assembly; } inline bool hasAES() const override { return m_aes; } + inline bool hasAVX2() const override { return m_avx2; } inline bool isSupported() const override { return true; } inline const char *brand() const override { return m_brand; } inline int32_t cores() const override { return m_cores; } @@ -59,6 +61,7 @@ protected: private: Assembly m_assembly; bool m_aes; + bool m_avx2; bool m_L2_exclusive; char m_brand[64]; int32_t m_cores; diff --git a/src/crypto/CryptoNight_arm.h b/src/crypto/CryptoNight_arm.h index b8c5092f1..456b5d744 100644 --- a/src/crypto/CryptoNight_arm.h +++ b/src/crypto/CryptoNight_arm.h @@ -5,9 +5,9 @@ * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee * Copyright 2016 Imran Yusuff - * Copyright 2017-2018 XMR-Stak , + * Copyright 2017-2019 XMR-Stak , * Copyright 2018 Lee Clagett - * Copyright 2018 SChernykh + * Copyright 2018-2019 SChernykh * Copyright 2016-2019 XMRig , * * This program is free software: you can redistribute it and/or modify @@ -284,6 +284,34 @@ static inline void cn_explode_scratchpad(const __m128i *input, __m128i *output) } +#ifndef XMRIG_NO_CN_GPU +template +void cn_explode_scratchpad_gpu(const uint8_t *input, uint8_t *output) +{ + constexpr size_t hash_size = 200; // 25x8 bytes + alignas(16) uint64_t hash[25]; + + for (uint64_t i = 0; i < MEM / 512; i++) + { + memcpy(hash, input, hash_size); + hash[0] ^= i; + + xmrig::keccakf(hash, 24); + memcpy(output, hash, 160); + output += 160; + + xmrig::keccakf(hash, 24); + memcpy(output, hash, 176); + output += 176; + + xmrig::keccakf(hash, 24); + memcpy(output, hash, 176); + output += 176; + } +} +#endif + + template static inline void cn_implode_scratchpad(const __m128i *input, __m128i *output) { @@ -541,6 +569,33 @@ inline void cryptonight_single_hash(const uint8_t *__restrict__ input, size_t si } +#ifndef XMRIG_NO_CN_GPU +template +void cn_gpu_inner_arm(const uint8_t *spad, uint8_t *lpad); + + +template +inline void cryptonight_single_hash_gpu(const uint8_t *__restrict__ input, size_t size, uint8_t *__restrict__ output, cryptonight_ctx **__restrict__ ctx) +{ + constexpr size_t MASK = xmrig::CRYPTONIGHT_GPU_MASK; + constexpr size_t ITERATIONS = xmrig::cn_select_iter(); + constexpr size_t MEM = xmrig::cn_select_memory(); + + static_assert(MASK > 0 && ITERATIONS > 0 && MEM > 0, "unsupported algorithm/variant"); + + xmrig::keccak(input, size, ctx[0]->state); + cn_explode_scratchpad_gpu(ctx[0]->state, ctx[0]->memory); + + cn_gpu_inner_arm(ctx[0]->state, ctx[0]->memory); + + cn_implode_scratchpad((__m128i*) ctx[0]->memory, (__m128i*) ctx[0]->state); + + xmrig::keccakf((uint64_t*) ctx[0]->state, 24); + memcpy(output, ctx[0]->state, 32); +} +#endif + + template inline void cryptonight_double_hash(const uint8_t *__restrict__ input, size_t size, uint8_t *__restrict__ output, struct cryptonight_ctx **__restrict__ ctx) { diff --git a/src/crypto/CryptoNight_constants.h b/src/crypto/CryptoNight_constants.h index 654e37409..f7a9542bf 100644 --- a/src/crypto/CryptoNight_constants.h +++ b/src/crypto/CryptoNight_constants.h @@ -4,7 +4,7 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2017-2018 XMR-Stak , + * Copyright 2017-2019 XMR-Stak , * Copyright 2018 Lee Clagett * Copyright 2018-2019 SChernykh * Copyright 2016-2019 XMRig , @@ -27,6 +27,7 @@ #define XMRIG_CRYPTONIGHT_CONSTANTS_H +#include #include @@ -42,6 +43,9 @@ constexpr const uint32_t CRYPTONIGHT_ITER = 0x80000; constexpr const uint32_t CRYPTONIGHT_HALF_ITER = 0x40000; constexpr const uint32_t CRYPTONIGHT_XAO_ITER = 0x100000; +constexpr const uint32_t CRYPTONIGHT_GPU_ITER = 0xC000; +constexpr const uint32_t CRYPTONIGHT_GPU_MASK = 0x1FFFC0; + constexpr const size_t CRYPTONIGHT_LITE_MEMORY = 1 * 1024 * 1024; constexpr const uint32_t CRYPTONIGHT_LITE_MASK = 0xFFFF0; constexpr const uint32_t CRYPTONIGHT_LITE_ITER = 0x40000; @@ -127,6 +131,7 @@ template<> inline constexpr uint32_t cn_select_iter() template<> inline constexpr uint32_t cn_select_iter() { return CRYPTONIGHT_HALF_ITER; } template<> inline constexpr uint32_t cn_select_iter() { return CRYPTONIGHT_XAO_ITER; } template<> inline constexpr uint32_t cn_select_iter() { return CRYPTONIGHT_ITER; } +template<> inline constexpr uint32_t cn_select_iter() { return CRYPTONIGHT_GPU_ITER; } template<> inline constexpr uint32_t cn_select_iter() { return CRYPTONIGHT_LITE_ITER; } template<> inline constexpr uint32_t cn_select_iter() { return CRYPTONIGHT_LITE_ITER; } template<> inline constexpr uint32_t cn_select_iter() { return CRYPTONIGHT_HEAVY_ITER; } @@ -142,6 +147,9 @@ inline uint32_t cn_select_iter(Algo algorithm, Variant variant) case VARIANT_HALF: return CRYPTONIGHT_HALF_ITER; + case VARIANT_GPU: + return CRYPTONIGHT_GPU_ITER; + case VARIANT_RTO: return CRYPTONIGHT_XAO_ITER; @@ -183,6 +191,7 @@ template<> inline constexpr Variant cn_base_variant() { return VA template<> inline constexpr Variant cn_base_variant() { return VARIANT_2; } template<> inline constexpr Variant cn_base_variant() { return VARIANT_2; } template<> inline constexpr Variant cn_base_variant() { return VARIANT_2; } +template<> inline constexpr Variant cn_base_variant() { return VARIANT_GPU; } } /* namespace xmrig */ diff --git a/src/crypto/CryptoNight_test.h b/src/crypto/CryptoNight_test.h index ed2c9e0ad..aadddc2eb 100644 --- a/src/crypto/CryptoNight_test.h +++ b/src/crypto/CryptoNight_test.h @@ -272,4 +272,21 @@ const static uint8_t test_output_pico_trtl[160] = { #endif +#ifndef XMRIG_NO_CN_GPU +// "cn/gpu" +const static uint8_t test_output_gpu[160] = { + 0xE5, 0x5C, 0xB2, 0x3E, 0x51, 0x64, 0x9A, 0x59, 0xB1, 0x27, 0xB9, 0x6B, 0x51, 0x5F, 0x2B, 0xF7, + 0xBF, 0xEA, 0x19, 0x97, 0x41, 0xA0, 0x21, 0x6C, 0xF8, 0x38, 0xDE, 0xD0, 0x6E, 0xFF, 0x82, 0xDF, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, +}; +#endif + + #endif /* XMRIG_CRYPTONIGHT_TEST_H */ diff --git a/src/crypto/CryptoNight_x86.h b/src/crypto/CryptoNight_x86.h index a8c2c309c..2d1aef0d6 100644 --- a/src/crypto/CryptoNight_x86.h +++ b/src/crypto/CryptoNight_x86.h @@ -4,7 +4,7 @@ * Copyright 2014 Lucas Jones * Copyright 2014-2016 Wolf9466 * Copyright 2016 Jay D Dee - * Copyright 2017-2018 XMR-Stak , + * Copyright 2017-2019 XMR-Stak , * Copyright 2018 Lee Clagett * Copyright 2018-2019 SChernykh * Copyright 2016-2019 XMRig , @@ -35,6 +35,7 @@ #endif +#include "common/cpu/Cpu.h" #include "common/crypto/keccak.h" #include "crypto/CryptoNight.h" #include "crypto/CryptoNight_constants.h" @@ -289,6 +290,34 @@ static inline void cn_explode_scratchpad(const __m128i *input, __m128i *output) } +#ifndef XMRIG_NO_CN_GPU +template +void cn_explode_scratchpad_gpu(const uint8_t *input, uint8_t *output) +{ + constexpr size_t hash_size = 200; // 25x8 bytes + alignas(16) uint64_t hash[25]; + + for (uint64_t i = 0; i < MEM / 512; i++) + { + memcpy(hash, input, hash_size); + hash[0] ^= i; + + xmrig::keccakf(hash, 24); + memcpy(output, hash, 160); + output += 160; + + xmrig::keccakf(hash, 24); + memcpy(output, hash, 176); + output += 176; + + xmrig::keccakf(hash, 24); + memcpy(output, hash, 176); + output += 176; + } +} +#endif + + template static inline void cn_implode_scratchpad(const __m128i *input, __m128i *output) { @@ -566,6 +595,41 @@ inline void cryptonight_single_hash(const uint8_t *__restrict__ input, size_t si } +#ifndef XMRIG_NO_CN_GPU +template +void cn_gpu_inner_avx(const uint8_t *spad, uint8_t *lpad); + + +template +void cn_gpu_inner_ssse3(const uint8_t *spad, uint8_t *lpad); + + +template +inline void cryptonight_single_hash_gpu(const uint8_t *__restrict__ input, size_t size, uint8_t *__restrict__ output, cryptonight_ctx **__restrict__ ctx) +{ + constexpr size_t MASK = xmrig::CRYPTONIGHT_GPU_MASK; + constexpr size_t ITERATIONS = xmrig::cn_select_iter(); + constexpr size_t MEM = xmrig::cn_select_memory(); + + static_assert(MASK > 0 && ITERATIONS > 0 && MEM > 0, "unsupported algorithm/variant"); + + xmrig::keccak(input, size, ctx[0]->state); + cn_explode_scratchpad_gpu(ctx[0]->state, ctx[0]->memory); + + if (xmrig::Cpu::info()->hasAVX2()) { + cn_gpu_inner_avx(ctx[0]->state, ctx[0]->memory); + } else { + cn_gpu_inner_ssse3(ctx[0]->state, ctx[0]->memory); + } + + cn_implode_scratchpad((__m128i*) ctx[0]->memory, (__m128i*) ctx[0]->state); + + xmrig::keccakf((uint64_t*) ctx[0]->state, 24); + memcpy(output, ctx[0]->state, 32); +} +#endif + + #ifndef XMRIG_NO_ASM extern "C" void cnv2_mainloop_ivybridge_asm(cryptonight_ctx *ctx); extern "C" void cnv2_mainloop_ryzen_asm(cryptonight_ctx *ctx); diff --git a/src/crypto/cn_gpu_arm.cpp b/src/crypto/cn_gpu_arm.cpp new file mode 100644 index 000000000..b463dd2ec --- /dev/null +++ b/src/crypto/cn_gpu_arm.cpp @@ -0,0 +1,240 @@ +/* XMRig + * Copyright 2010 Jeff Garzik + * Copyright 2012-2014 pooler + * Copyright 2014 Lucas Jones + * Copyright 2014-2016 Wolf9466 + * Copyright 2016 Jay D Dee + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig + * + * This program is free software: you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation, either version 3 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program. If not, see . + */ + + +#include + + +#include "crypto/CryptoNight_constants.h" + + +inline void vandq_f32(float32x4_t &v, uint32_t v2) +{ + uint32x4_t vc = vdupq_n_u32(v2); + v = (float32x4_t)vandq_u32((uint32x4_t)v, vc); +} + + +inline void vorq_f32(float32x4_t &v, uint32_t v2) +{ + uint32x4_t vc = vdupq_n_u32(v2); + v = (float32x4_t)vorrq_u32((uint32x4_t)v, vc); +} + + +template +inline void vrot_si32(int32x4_t &r) +{ + r = (int32x4_t)vextq_s8((int8x16_t)r, (int8x16_t)r, v); +} + +template <> +inline void vrot_si32<0>(int32x4_t &r) +{ +} + + +inline uint32_t vheor_s32(const int32x4_t &v) +{ + int32x4_t v0 = veorq_s32(v, vrev64q_s32(v)); + int32x2_t vf = veor_s32(vget_high_s32(v0), vget_low_s32(v0)); + return (uint32_t)vget_lane_s32(vf, 0); +} + + +inline void prep_dv(int32_t *idx, int32x4_t &v, float32x4_t &n) +{ + v = vld1q_s32(idx); + n = vcvtq_f32_s32(v); +} + + +inline void sub_round(const float32x4_t &n0, const float32x4_t &n1, const float32x4_t &n2, const float32x4_t &n3, const float32x4_t &rnd_c, float32x4_t &n, float32x4_t &d, float32x4_t &c) +{ + float32x4_t ln1 = vaddq_f32(n1, c); + float32x4_t nn = vmulq_f32(n0, c); + nn = vmulq_f32(ln1, vmulq_f32(nn, nn)); + vandq_f32(nn, 0xFEFFFFFF); + vorq_f32(nn, 0x00800000); + n = vaddq_f32(n, nn); + + float32x4_t ln3 = vsubq_f32(n3, c); + float32x4_t dd = vmulq_f32(n2, c); + dd = vmulq_f32(ln3, vmulq_f32(dd, dd)); + vandq_f32(dd, 0xFEFFFFFF); + vorq_f32(dd, 0x00800000); + d = vaddq_f32(d, dd); + + //Constant feedback + c = vaddq_f32(c, rnd_c); + c = vaddq_f32(c, vdupq_n_f32(0.734375f)); + float32x4_t r = vaddq_f32(nn, dd); + vandq_f32(r, 0x807FFFFF); + vorq_f32(r, 0x40000000); + c = vaddq_f32(c, r); +} + + +inline void round_compute(const float32x4_t &n0, const float32x4_t &n1, const float32x4_t &n2, const float32x4_t &n3, const float32x4_t &rnd_c, float32x4_t &c, float32x4_t &r) +{ + float32x4_t n = vdupq_n_f32(0.0f), d = vdupq_n_f32(0.0f); + + sub_round(n0, n1, n2, n3, rnd_c, n, d, c); + sub_round(n1, n2, n3, n0, rnd_c, n, d, c); + sub_round(n2, n3, n0, n1, rnd_c, n, d, c); + sub_round(n3, n0, n1, n2, rnd_c, n, d, c); + sub_round(n3, n2, n1, n0, rnd_c, n, d, c); + sub_round(n2, n1, n0, n3, rnd_c, n, d, c); + sub_round(n1, n0, n3, n2, rnd_c, n, d, c); + sub_round(n0, n3, n2, n1, rnd_c, n, d, c); + + // Make sure abs(d) > 2.0 - this prevents division by zero and accidental overflows by division by < 1.0 + vandq_f32(d, 0xFF7FFFFF); + vorq_f32(d, 0x40000000); + r = vaddq_f32(r, vdivq_f32(n, d)); +} + + +// 112×4 = 448 +template +inline int32x4_t single_compute(const float32x4_t &n0, const float32x4_t &n1, const float32x4_t &n2, const float32x4_t &n3, float cnt, const float32x4_t &rnd_c, float32x4_t &sum) +{ + float32x4_t c = vdupq_n_f32(cnt); + float32x4_t r = vdupq_n_f32(0.0f); + + round_compute(n0, n1, n2, n3, rnd_c, c, r); + round_compute(n0, n1, n2, n3, rnd_c, c, r); + round_compute(n0, n1, n2, n3, rnd_c, c, r); + round_compute(n0, n1, n2, n3, rnd_c, c, r); + + // do a quick fmod by setting exp to 2 + vandq_f32(r, 0x807FFFFF); + vorq_f32(r, 0x40000000); + + if (add) { + sum = vaddq_f32(sum, r); + } else { + sum = r; + } + + const float32x4_t cc2 = vdupq_n_f32(536870880.0f); + r = vmulq_f32(r, cc2); // 35 + return vcvtq_s32_f32(r); +} + + +template +inline void single_compute_wrap(const float32x4_t &n0, const float32x4_t &n1, const float32x4_t &n2, const float32x4_t &n3, float cnt, const float32x4_t &rnd_c, float32x4_t &sum, int32x4_t &out) +{ + int32x4_t r = single_compute(n0, n1, n2, n3, cnt, rnd_c, sum); + vrot_si32(r); + out = veorq_s32(out, r); +} + + +template +inline int32_t *scratchpad_ptr(uint8_t* lpad, uint32_t idx, size_t n) { return reinterpret_cast(lpad + (idx & MASK) + n * 16); } + + +template +void cn_gpu_inner_arm(const uint8_t *spad, uint8_t *lpad) +{ + uint32_t s = reinterpret_cast(spad)[0] >> 8; + int32_t *idx0 = scratchpad_ptr(lpad, s, 0); + int32_t *idx1 = scratchpad_ptr(lpad, s, 1); + int32_t *idx2 = scratchpad_ptr(lpad, s, 2); + int32_t *idx3 = scratchpad_ptr(lpad, s, 3); + float32x4_t sum0 = vdupq_n_f32(0.0f); + + for (size_t i = 0; i < ITER; i++) { + float32x4_t n0, n1, n2, n3; + int32x4_t v0, v1, v2, v3; + float32x4_t suma, sumb, sum1, sum2, sum3; + + prep_dv(idx0, v0, n0); + prep_dv(idx1, v1, n1); + prep_dv(idx2, v2, n2); + prep_dv(idx3, v3, n3); + float32x4_t rc = sum0; + + int32x4_t out, out2; + out = vdupq_n_s32(0); + single_compute_wrap<0>(n0, n1, n2, n3, 1.3437500f, rc, suma, out); + single_compute_wrap<1>(n0, n2, n3, n1, 1.2812500f, rc, suma, out); + single_compute_wrap<2>(n0, n3, n1, n2, 1.3593750f, rc, sumb, out); + single_compute_wrap<3>(n0, n3, n2, n1, 1.3671875f, rc, sumb, out); + sum0 = vaddq_f32(suma, sumb); + vst1q_s32(idx0, veorq_s32(v0, out)); + out2 = out; + + out = vdupq_n_s32(0); + single_compute_wrap<0>(n1, n0, n2, n3, 1.4296875f, rc, suma, out); + single_compute_wrap<1>(n1, n2, n3, n0, 1.3984375f, rc, suma, out); + single_compute_wrap<2>(n1, n3, n0, n2, 1.3828125f, rc, sumb, out); + single_compute_wrap<3>(n1, n3, n2, n0, 1.3046875f, rc, sumb, out); + sum1 = vaddq_f32(suma, sumb); + vst1q_s32(idx1, veorq_s32(v1, out)); + out2 = veorq_s32(out2, out); + + out = vdupq_n_s32(0); + single_compute_wrap<0>(n2, n1, n0, n3, 1.4140625f, rc, suma, out); + single_compute_wrap<1>(n2, n0, n3, n1, 1.2734375f, rc, suma, out); + single_compute_wrap<2>(n2, n3, n1, n0, 1.2578125f, rc, sumb, out); + single_compute_wrap<3>(n2, n3, n0, n1, 1.2890625f, rc, sumb, out); + sum2 = vaddq_f32(suma, sumb); + vst1q_s32(idx2, veorq_s32(v2, out)); + out2 = veorq_s32(out2, out); + + out = vdupq_n_s32(0); + single_compute_wrap<0>(n3, n1, n2, n0, 1.3203125f, rc, suma, out); + single_compute_wrap<1>(n3, n2, n0, n1, 1.3515625f, rc, suma, out); + single_compute_wrap<2>(n3, n0, n1, n2, 1.3359375f, rc, sumb, out); + single_compute_wrap<3>(n3, n0, n2, n1, 1.4609375f, rc, sumb, out); + sum3 = vaddq_f32(suma, sumb); + vst1q_s32(idx3, veorq_s32(v3, out)); + out2 = veorq_s32(out2, out); + + sum0 = vaddq_f32(sum0, sum1); + sum2 = vaddq_f32(sum2, sum3); + sum0 = vaddq_f32(sum0, sum2); + + const float32x4_t cc1 = vdupq_n_f32(16777216.0f); + const float32x4_t cc2 = vdupq_n_f32(64.0f); + vandq_f32(sum0, 0x7fffffff); // take abs(va) by masking the float sign bit + // vs range 0 - 64 + n0 = vmulq_f32(sum0, cc1); + v0 = vcvtq_s32_f32(n0); + v0 = veorq_s32(v0, out2); + uint32_t n = vheor_s32(v0); + + // vs is now between 0 and 1 + sum0 = vdivq_f32(sum0, cc2); + idx0 = scratchpad_ptr(lpad, n, 0); + idx1 = scratchpad_ptr(lpad, n, 1); + idx2 = scratchpad_ptr(lpad, n, 2); + idx3 = scratchpad_ptr(lpad, n, 3); + } +} + +template void cn_gpu_inner_arm(const uint8_t* spad, uint8_t* lpad); diff --git a/src/crypto/cn_gpu_avx.cpp b/src/crypto/cn_gpu_avx.cpp new file mode 100644 index 000000000..3dc7cacb7 --- /dev/null +++ b/src/crypto/cn_gpu_avx.cpp @@ -0,0 +1,203 @@ +/* XMRig + * Copyright 2010 Jeff Garzik + * Copyright 2012-2014 pooler + * Copyright 2014 Lucas Jones + * Copyright 2014-2016 Wolf9466 + * Copyright 2016 Jay D Dee + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig + * + * This program is free software: you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation, either version 3 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program. If not, see . + */ + +#include "crypto/CryptoNight_constants.h" + +#ifdef __GNUC__ +# include +#else +# include +# define __restrict__ __restrict +#endif + +inline void prep_dv_avx(__m256i* idx, __m256i& v, __m256& n01) +{ + v = _mm256_load_si256(idx); + n01 = _mm256_cvtepi32_ps(v); +} + +inline __m256 fma_break(const __m256& x) +{ + // Break the dependency chain by setitng the exp to ?????01 + __m256 xx = _mm256_and_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0xFEFFFFFF)), x); + return _mm256_or_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0x00800000)), xx); +} + +// 14 +inline void sub_round(const __m256& n0, const __m256& n1, const __m256& n2, const __m256& n3, const __m256& rnd_c, __m256& n, __m256& d, __m256& c) +{ + __m256 nn = _mm256_mul_ps(n0, c); + nn = _mm256_mul_ps(_mm256_add_ps(n1, c), _mm256_mul_ps(nn, nn)); + nn = fma_break(nn); + n = _mm256_add_ps(n, nn); + + __m256 dd = _mm256_mul_ps(n2, c); + dd = _mm256_mul_ps(_mm256_sub_ps(n3, c), _mm256_mul_ps(dd, dd)); + dd = fma_break(dd); + d = _mm256_add_ps(d, dd); + + //Constant feedback + c = _mm256_add_ps(c, rnd_c); + c = _mm256_add_ps(c, _mm256_set1_ps(0.734375f)); + __m256 r = _mm256_add_ps(nn, dd); + r = _mm256_and_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0x807FFFFF)), r); + r = _mm256_or_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0x40000000)), r); + c = _mm256_add_ps(c, r); +} + +// 14*8 + 2 = 112 +inline void round_compute(const __m256& n0, const __m256& n1, const __m256& n2, const __m256& n3, const __m256& rnd_c, __m256& c, __m256& r) +{ + __m256 n = _mm256_setzero_ps(), d = _mm256_setzero_ps(); + + sub_round(n0, n1, n2, n3, rnd_c, n, d, c); + sub_round(n1, n2, n3, n0, rnd_c, n, d, c); + sub_round(n2, n3, n0, n1, rnd_c, n, d, c); + sub_round(n3, n0, n1, n2, rnd_c, n, d, c); + sub_round(n3, n2, n1, n0, rnd_c, n, d, c); + sub_round(n2, n1, n0, n3, rnd_c, n, d, c); + sub_round(n1, n0, n3, n2, rnd_c, n, d, c); + sub_round(n0, n3, n2, n1, rnd_c, n, d, c); + + // Make sure abs(d) > 2.0 - this prevents division by zero and accidental overflows by division by < 1.0 + d = _mm256_and_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0xFF7FFFFF)), d); + d = _mm256_or_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0x40000000)), d); + r = _mm256_add_ps(r, _mm256_div_ps(n, d)); +} + +// 112×4 = 448 +template +inline __m256i double_compute(const __m256& n0, const __m256& n1, const __m256& n2, const __m256& n3, + float lcnt, float hcnt, const __m256& rnd_c, __m256& sum) +{ + __m256 c = _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_set1_ps(lcnt)), _mm_set1_ps(hcnt), 1); + __m256 r = _mm256_setzero_ps(); + + round_compute(n0, n1, n2, n3, rnd_c, c, r); + round_compute(n0, n1, n2, n3, rnd_c, c, r); + round_compute(n0, n1, n2, n3, rnd_c, c, r); + round_compute(n0, n1, n2, n3, rnd_c, c, r); + + // do a quick fmod by setting exp to 2 + r = _mm256_and_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0x807FFFFF)), r); + r = _mm256_or_ps(_mm256_castsi256_ps(_mm256_set1_epi32(0x40000000)), r); + + if(add) + sum = _mm256_add_ps(sum, r); + else + sum = r; + + r = _mm256_mul_ps(r, _mm256_set1_ps(536870880.0f)); // 35 + return _mm256_cvttps_epi32(r); +} + +template +inline void double_compute_wrap(const __m256& n0, const __m256& n1, const __m256& n2, const __m256& n3, + float lcnt, float hcnt, const __m256& rnd_c, __m256& sum, __m256i& out) +{ + __m256i r = double_compute(n0, n1, n2, n3, lcnt, hcnt, rnd_c, sum); + if(rot != 0) + r = _mm256_or_si256(_mm256_bslli_epi128(r, 16 - rot), _mm256_bsrli_epi128(r, rot)); + + out = _mm256_xor_si256(out, r); +} + +template +inline __m256i* scratchpad_ptr(uint8_t* lpad, uint32_t idx, size_t n) { return reinterpret_cast<__m256i*>(lpad + (idx & MASK) + n*16); } + +template +void cn_gpu_inner_avx(const uint8_t* spad, uint8_t* lpad) +{ + uint32_t s = reinterpret_cast(spad)[0] >> 8; + __m256i* idx0 = scratchpad_ptr(lpad, s, 0); + __m256i* idx2 = scratchpad_ptr(lpad, s, 2); + __m256 sum0 = _mm256_setzero_ps(); + + for(size_t i = 0; i < ITER; i++) + { + __m256i v01, v23; + __m256 suma, sumb, sum1; + __m256 rc = sum0; + + __m256 n01, n23; + prep_dv_avx(idx0, v01, n01); + prep_dv_avx(idx2, v23, n23); + + __m256i out, out2; + __m256 n10, n22, n33; + n10 = _mm256_permute2f128_ps(n01, n01, 0x01); + n22 = _mm256_permute2f128_ps(n23, n23, 0x00); + n33 = _mm256_permute2f128_ps(n23, n23, 0x11); + + out = _mm256_setzero_si256(); + double_compute_wrap<0>(n01, n10, n22, n33, 1.3437500f, 1.4296875f, rc, suma, out); + double_compute_wrap<1>(n01, n22, n33, n10, 1.2812500f, 1.3984375f, rc, suma, out); + double_compute_wrap<2>(n01, n33, n10, n22, 1.3593750f, 1.3828125f, rc, sumb, out); + double_compute_wrap<3>(n01, n33, n22, n10, 1.3671875f, 1.3046875f, rc, sumb, out); + _mm256_store_si256(idx0, _mm256_xor_si256(v01, out)); + sum0 = _mm256_add_ps(suma, sumb); + out2 = out; + + __m256 n11, n02, n30; + n11 = _mm256_permute2f128_ps(n01, n01, 0x11); + n02 = _mm256_permute2f128_ps(n01, n23, 0x20); + n30 = _mm256_permute2f128_ps(n01, n23, 0x03); + + out = _mm256_setzero_si256(); + double_compute_wrap<0>(n23, n11, n02, n30, 1.4140625f, 1.3203125f, rc, suma, out); + double_compute_wrap<1>(n23, n02, n30, n11, 1.2734375f, 1.3515625f, rc, suma, out); + double_compute_wrap<2>(n23, n30, n11, n02, 1.2578125f, 1.3359375f, rc, sumb, out); + double_compute_wrap<3>(n23, n30, n02, n11, 1.2890625f, 1.4609375f, rc, sumb, out); + _mm256_store_si256(idx2, _mm256_xor_si256(v23, out)); + sum1 = _mm256_add_ps(suma, sumb); + + out2 = _mm256_xor_si256(out2, out); + out2 = _mm256_xor_si256(_mm256_permute2x128_si256(out2,out2,0x41), out2); + suma = _mm256_permute2f128_ps(sum0, sum1, 0x30); + sumb = _mm256_permute2f128_ps(sum0, sum1, 0x21); + sum0 = _mm256_add_ps(suma, sumb); + sum0 = _mm256_add_ps(sum0, _mm256_permute2f128_ps(sum0, sum0, 0x41)); + + // Clear the high 128 bits + __m128 sum = _mm256_castps256_ps128(sum0); + + sum = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0x7fffffff)), sum); // take abs(va) by masking the float sign bit + // vs range 0 - 64 + __m128i v0 = _mm_cvttps_epi32(_mm_mul_ps(sum, _mm_set1_ps(16777216.0f))); + v0 = _mm_xor_si128(v0, _mm256_castsi256_si128(out2)); + __m128i v1 = _mm_shuffle_epi32(v0, _MM_SHUFFLE(0, 1, 2, 3)); + v0 = _mm_xor_si128(v0, v1); + v1 = _mm_shuffle_epi32(v0, _MM_SHUFFLE(0, 1, 0, 1)); + v0 = _mm_xor_si128(v0, v1); + + // vs is now between 0 and 1 + sum = _mm_div_ps(sum, _mm_set1_ps(64.0f)); + sum0 = _mm256_insertf128_ps(_mm256_castps128_ps256(sum), sum, 1); + uint32_t n = _mm_cvtsi128_si32(v0); + idx0 = scratchpad_ptr(lpad, n, 0); + idx2 = scratchpad_ptr(lpad, n, 2); + } +} + +template void cn_gpu_inner_avx(const uint8_t* spad, uint8_t* lpad); diff --git a/src/crypto/cn_gpu_ssse3.cpp b/src/crypto/cn_gpu_ssse3.cpp new file mode 100644 index 000000000..ce3d19add --- /dev/null +++ b/src/crypto/cn_gpu_ssse3.cpp @@ -0,0 +1,210 @@ +/* XMRig + * Copyright 2010 Jeff Garzik + * Copyright 2012-2014 pooler + * Copyright 2014 Lucas Jones + * Copyright 2014-2016 Wolf9466 + * Copyright 2016 Jay D Dee + * Copyright 2017-2019 XMR-Stak , + * Copyright 2018-2019 SChernykh + * Copyright 2016-2019 XMRig + * + * This program is free software: you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation, either version 3 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program. If not, see . + */ + +#include "crypto/CryptoNight_constants.h" + +#ifdef __GNUC__ +# include +#else +# include +# define __restrict__ __restrict +#endif + +inline void prep_dv(__m128i* idx, __m128i& v, __m128& n) +{ + v = _mm_load_si128(idx); + n = _mm_cvtepi32_ps(v); +} + +inline __m128 fma_break(__m128 x) +{ + // Break the dependency chain by setitng the exp to ?????01 + x = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0xFEFFFFFF)), x); + return _mm_or_ps(_mm_castsi128_ps(_mm_set1_epi32(0x00800000)), x); +} + +// 14 +inline void sub_round(__m128 n0, __m128 n1, __m128 n2, __m128 n3, __m128 rnd_c, __m128& n, __m128& d, __m128& c) +{ + n1 = _mm_add_ps(n1, c); + __m128 nn = _mm_mul_ps(n0, c); + nn = _mm_mul_ps(n1, _mm_mul_ps(nn,nn)); + nn = fma_break(nn); + n = _mm_add_ps(n, nn); + + n3 = _mm_sub_ps(n3, c); + __m128 dd = _mm_mul_ps(n2, c); + dd = _mm_mul_ps(n3, _mm_mul_ps(dd,dd)); + dd = fma_break(dd); + d = _mm_add_ps(d, dd); + + //Constant feedback + c = _mm_add_ps(c, rnd_c); + c = _mm_add_ps(c, _mm_set1_ps(0.734375f)); + __m128 r = _mm_add_ps(nn, dd); + r = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0x807FFFFF)), r); + r = _mm_or_ps(_mm_castsi128_ps(_mm_set1_epi32(0x40000000)), r); + c = _mm_add_ps(c, r); +} + +// 14*8 + 2 = 112 +inline void round_compute(__m128 n0, __m128 n1, __m128 n2, __m128 n3, __m128 rnd_c, __m128& c, __m128& r) +{ + __m128 n = _mm_setzero_ps(), d = _mm_setzero_ps(); + + sub_round(n0, n1, n2, n3, rnd_c, n, d, c); + sub_round(n1, n2, n3, n0, rnd_c, n, d, c); + sub_round(n2, n3, n0, n1, rnd_c, n, d, c); + sub_round(n3, n0, n1, n2, rnd_c, n, d, c); + sub_round(n3, n2, n1, n0, rnd_c, n, d, c); + sub_round(n2, n1, n0, n3, rnd_c, n, d, c); + sub_round(n1, n0, n3, n2, rnd_c, n, d, c); + sub_round(n0, n3, n2, n1, rnd_c, n, d, c); + + // Make sure abs(d) > 2.0 - this prevents division by zero and accidental overflows by division by < 1.0 + d = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0xFF7FFFFF)), d); + d = _mm_or_ps(_mm_castsi128_ps(_mm_set1_epi32(0x40000000)), d); + r =_mm_add_ps(r, _mm_div_ps(n,d)); +} + +// 112×4 = 448 +template +inline __m128i single_compute(__m128 n0, __m128 n1, __m128 n2, __m128 n3, float cnt, __m128 rnd_c, __m128& sum) +{ + __m128 c = _mm_set1_ps(cnt); + __m128 r = _mm_setzero_ps(); + + round_compute(n0, n1, n2, n3, rnd_c, c, r); + round_compute(n0, n1, n2, n3, rnd_c, c, r); + round_compute(n0, n1, n2, n3, rnd_c, c, r); + round_compute(n0, n1, n2, n3, rnd_c, c, r); + + // do a quick fmod by setting exp to 2 + r = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0x807FFFFF)), r); + r = _mm_or_ps(_mm_castsi128_ps(_mm_set1_epi32(0x40000000)), r); + + if(add) + sum = _mm_add_ps(sum, r); + else + sum = r; + + r = _mm_mul_ps(r, _mm_set1_ps(536870880.0f)); // 35 + return _mm_cvttps_epi32(r); +} + +template +inline void single_compute_wrap(__m128 n0, __m128 n1, __m128 n2, __m128 n3, float cnt, __m128 rnd_c, __m128& sum, __m128i& out) +{ + __m128i r = single_compute(n0, n1, n2, n3, cnt, rnd_c, sum); + if(rot != 0) + r = _mm_or_si128(_mm_slli_si128(r, 16 - rot), _mm_srli_si128(r, rot)); + out = _mm_xor_si128(out, r); +} + +template +inline __m128i* scratchpad_ptr(uint8_t* lpad, uint32_t idx, size_t n) { return reinterpret_cast<__m128i*>(lpad + (idx & MASK) + n*16); } + +template +void cn_gpu_inner_ssse3(const uint8_t* spad, uint8_t* lpad) +{ + uint32_t s = reinterpret_cast(spad)[0] >> 8; + __m128i* idx0 = scratchpad_ptr(lpad, s, 0); + __m128i* idx1 = scratchpad_ptr(lpad, s, 1); + __m128i* idx2 = scratchpad_ptr(lpad, s, 2); + __m128i* idx3 = scratchpad_ptr(lpad, s, 3); + __m128 sum0 = _mm_setzero_ps(); + + for(size_t i = 0; i < ITER; i++) + { + __m128 n0, n1, n2, n3; + __m128i v0, v1, v2, v3; + __m128 suma, sumb, sum1, sum2, sum3; + + prep_dv(idx0, v0, n0); + prep_dv(idx1, v1, n1); + prep_dv(idx2, v2, n2); + prep_dv(idx3, v3, n3); + __m128 rc = sum0; + + __m128i out, out2; + out = _mm_setzero_si128(); + single_compute_wrap<0>(n0, n1, n2, n3, 1.3437500f, rc, suma, out); + single_compute_wrap<1>(n0, n2, n3, n1, 1.2812500f, rc, suma, out); + single_compute_wrap<2>(n0, n3, n1, n2, 1.3593750f, rc, sumb, out); + single_compute_wrap<3>(n0, n3, n2, n1, 1.3671875f, rc, sumb, out); + sum0 = _mm_add_ps(suma, sumb); + _mm_store_si128(idx0, _mm_xor_si128(v0, out)); + out2 = out; + + out = _mm_setzero_si128(); + single_compute_wrap<0>(n1, n0, n2, n3, 1.4296875f, rc, suma, out); + single_compute_wrap<1>(n1, n2, n3, n0, 1.3984375f, rc, suma, out); + single_compute_wrap<2>(n1, n3, n0, n2, 1.3828125f, rc, sumb, out); + single_compute_wrap<3>(n1, n3, n2, n0, 1.3046875f, rc, sumb, out); + sum1 = _mm_add_ps(suma, sumb); + _mm_store_si128(idx1, _mm_xor_si128(v1, out)); + out2 = _mm_xor_si128(out2, out); + + out = _mm_setzero_si128(); + single_compute_wrap<0>(n2, n1, n0, n3, 1.4140625f, rc, suma, out); + single_compute_wrap<1>(n2, n0, n3, n1, 1.2734375f, rc, suma, out); + single_compute_wrap<2>(n2, n3, n1, n0, 1.2578125f, rc, sumb, out); + single_compute_wrap<3>(n2, n3, n0, n1, 1.2890625f, rc, sumb, out); + sum2 = _mm_add_ps(suma, sumb); + _mm_store_si128(idx2, _mm_xor_si128(v2, out)); + out2 = _mm_xor_si128(out2, out); + + out = _mm_setzero_si128(); + single_compute_wrap<0>(n3, n1, n2, n0, 1.3203125f, rc, suma, out); + single_compute_wrap<1>(n3, n2, n0, n1, 1.3515625f, rc, suma, out); + single_compute_wrap<2>(n3, n0, n1, n2, 1.3359375f, rc, sumb, out); + single_compute_wrap<3>(n3, n0, n2, n1, 1.4609375f, rc, sumb, out); + sum3 = _mm_add_ps(suma, sumb); + _mm_store_si128(idx3, _mm_xor_si128(v3, out)); + out2 = _mm_xor_si128(out2, out); + sum0 = _mm_add_ps(sum0, sum1); + sum2 = _mm_add_ps(sum2, sum3); + sum0 = _mm_add_ps(sum0, sum2); + + sum0 = _mm_and_ps(_mm_castsi128_ps(_mm_set1_epi32(0x7fffffff)), sum0); // take abs(va) by masking the float sign bit + // vs range 0 - 64 + n0 = _mm_mul_ps(sum0, _mm_set1_ps(16777216.0f)); + v0 = _mm_cvttps_epi32(n0); + v0 = _mm_xor_si128(v0, out2); + v1 = _mm_shuffle_epi32(v0, _MM_SHUFFLE(0, 1, 2, 3)); + v0 = _mm_xor_si128(v0, v1); + v1 = _mm_shuffle_epi32(v0, _MM_SHUFFLE(0, 1, 0, 1)); + v0 = _mm_xor_si128(v0, v1); + + // vs is now between 0 and 1 + sum0 = _mm_div_ps(sum0, _mm_set1_ps(64.0f)); + uint32_t n = _mm_cvtsi128_si32(v0); + idx0 = scratchpad_ptr(lpad, n, 0); + idx1 = scratchpad_ptr(lpad, n, 1); + idx2 = scratchpad_ptr(lpad, n, 2); + idx3 = scratchpad_ptr(lpad, n, 3); + } +} + +template void cn_gpu_inner_ssse3(const uint8_t* spad, uint8_t* lpad); diff --git a/src/workers/CpuThread.cpp b/src/workers/CpuThread.cpp index af55b0dac..d709fc6ef 100644 --- a/src/workers/CpuThread.cpp +++ b/src/workers/CpuThread.cpp @@ -250,6 +250,21 @@ xmrig::CpuThread::cn_hash_fun xmrig::CpuThread::fn(Algo algorithm, AlgoVariant a nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_TRTL +# ifndef XMRIG_NO_CN_GPU + cryptonight_single_hash_gpu, + nullptr, + cryptonight_single_hash_gpu, + nullptr, + nullptr, + nullptr, + nullptr, + nullptr, + nullptr, + nullptr, +# else + nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_GPU +# endif + # ifndef XMRIG_NO_AEON cryptonight_single_hash, cryptonight_double_hash, @@ -282,6 +297,7 @@ xmrig::CpuThread::cn_hash_fun xmrig::CpuThread::fn(Algo algorithm, AlgoVariant a nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_2 nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_HALF nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_TRTL + nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_GPU # else nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_0 nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_1 @@ -294,6 +310,7 @@ xmrig::CpuThread::cn_hash_fun xmrig::CpuThread::fn(Algo algorithm, AlgoVariant a nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_2 nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_HALF nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_TRTL + nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_GPU # endif # ifndef XMRIG_NO_SUMO @@ -340,6 +357,7 @@ xmrig::CpuThread::cn_hash_fun xmrig::CpuThread::fn(Algo algorithm, AlgoVariant a nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_2 nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_HALF nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_TRTL + nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_GPU # else nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_0 nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_1 @@ -352,6 +370,7 @@ xmrig::CpuThread::cn_hash_fun xmrig::CpuThread::fn(Algo algorithm, AlgoVariant a nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_2 nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_HALF nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_TRTL + nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_GPU # endif # ifndef XMRIG_NO_CN_PICO nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_0 @@ -375,6 +394,8 @@ xmrig::CpuThread::cn_hash_fun xmrig::CpuThread::fn(Algo algorithm, AlgoVariant a cryptonight_triple_hash, cryptonight_quad_hash, cryptonight_penta_hash, + + nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_GPU # else nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_0 nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_1 @@ -387,6 +408,7 @@ xmrig::CpuThread::cn_hash_fun xmrig::CpuThread::fn(Algo algorithm, AlgoVariant a nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_2 nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_HALF nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_TRTL + nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, // VARIANT_GPU # endif # ifndef XMRIG_NO_ASM cryptonight_single_hash_asm, diff --git a/src/workers/MultiWorker.cpp b/src/workers/MultiWorker.cpp index eff330624..a25665df6 100644 --- a/src/workers/MultiWorker.cpp +++ b/src/workers/MultiWorker.cpp @@ -54,13 +54,16 @@ bool MultiWorker::selfTest() using namespace xmrig; if (m_thread->algorithm() == CRYPTONIGHT) { - return verify(VARIANT_0, test_output_v0) && - verify(VARIANT_1, test_output_v1) && - verify(VARIANT_2, test_output_v2) && - verify(VARIANT_XTL, test_output_xtl) && - verify(VARIANT_MSR, test_output_msr) && - verify(VARIANT_XAO, test_output_xao) && - verify(VARIANT_RTO, test_output_rto) && + return verify(VARIANT_0, test_output_v0) && + verify(VARIANT_1, test_output_v1) && + verify(VARIANT_2, test_output_v2) && + verify(VARIANT_XTL, test_output_xtl) && + verify(VARIANT_MSR, test_output_msr) && + verify(VARIANT_XAO, test_output_xao) && + verify(VARIANT_RTO, test_output_rto) && +# ifndef XMRIG_NO_CN_GPU + verify(VARIANT_GPU, test_output_gpu) && +# endif verify(VARIANT_HALF, test_output_half); }