From 0a9a9abaf7afceeb923292630b93b0fe4830efea Mon Sep 17 00:00:00 2001
From: psychocrypt <psychocryptHPC@gmail.com>
Date: Sun, 16 Sep 2018 22:24:56 +0200
Subject: [PATCH] infrastructure to load asm code
- add new option to `cpu.txt` named `asm` to select the asm code version
- extent function selection method to choose assembler code for `cryptonight_v8`
- update auto adjustment to add default value for option `asm`
---
CMakeLists.txt | 11 +++++-
xmrstak/backend/cpu/autoAdjust.hpp | 2 +-
xmrstak/backend/cpu/autoAdjustHwloc.hpp | 2 +-
xmrstak/backend/cpu/config.tpl | 13 ++++---
.../backend/cpu/crypto/cryptonight_aesni.h | 21 +++++++++++
xmrstak/backend/cpu/jconf.cpp | 7 +++-
xmrstak/backend/cpu/jconf.hpp | 1 +
xmrstak/backend/cpu/minethd.cpp | 35 +++++++++++++++----
xmrstak/backend/cpu/minethd.hpp | 5 +--
9 files changed, 81 insertions(+), 16 deletions(-)
diff --git a/CMakeLists.txt b/CMakeLists.txt
index a642b38..067bbd0 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -458,6 +458,15 @@ if(MICROHTTPD_ENABLE)
endif()
target_link_libraries(xmr-stak-c ${LIBS})
+enable_language(ASM)
+# asm optimized monero v8 code
+add_library(xmr-stak-asm
+ STATIC
+ "xmrstak/backend/cpu/crypto/asm/cryptonigh_v8_main_loop.S"
+)
+set_property(TARGET xmr-stak-asm PROPERTY LINKER_LANGUAGE C)
+
+
# compile generic backend files
file(GLOB BACKEND_CPP
"xmrstak/*.cpp"
@@ -472,7 +481,7 @@ add_library(xmr-stak-backend
STATIC
${BACKEND_CPP}
)
-target_link_libraries(xmr-stak-backend xmr-stak-c ${CMAKE_DL_LIBS})
+target_link_libraries(xmr-stak-backend xmr-stak-c ${CMAKE_DL_LIBS} xmr-stak-asm)
# compile CUDA backend
if(CUDA_FOUND)
diff --git a/xmrstak/backend/cpu/autoAdjust.hpp b/xmrstak/backend/cpu/autoAdjust.hpp
index 57dbef0..8588fea 100644
--- a/xmrstak/backend/cpu/autoAdjust.hpp
+++ b/xmrstak/backend/cpu/autoAdjust.hpp
@@ -82,7 +82,7 @@ public:
conf += std::string(" { \"low_power_mode\" : ");
conf += std::string(double_mode ? "true" : "false");
- conf += std::string(", \"no_prefetch\" : true, \"affine_to_cpu\" : ");
+ conf += std::string(", \"no_prefetch\" : true, \"asm\" : \"auto\", \"affine_to_cpu\" : ");
conf += std::to_string(aff_id);
conf += std::string(" },\n");
diff --git a/xmrstak/backend/cpu/autoAdjustHwloc.hpp b/xmrstak/backend/cpu/autoAdjustHwloc.hpp
index 01d2280..a73de86 100644
--- a/xmrstak/backend/cpu/autoAdjustHwloc.hpp
+++ b/xmrstak/backend/cpu/autoAdjustHwloc.hpp
@@ -70,7 +70,7 @@ public:
{
conf += std::string(" { \"low_power_mode\" : ");
conf += std::string((id & 0x8000000) != 0 ? "true" : "false");
- conf += std::string(", \"no_prefetch\" : true, \"affine_to_cpu\" : ");
+ conf += std::string(", \"no_prefetch\" : true, \"asm\" : \"auto\", \"affine_to_cpu\" : ");
conf += std::to_string(id & 0x7FFFFFF);
conf += std::string(" },\n");
}
diff --git a/xmrstak/backend/cpu/config.tpl b/xmrstak/backend/cpu/config.tpl
index 2fc9a47..bfffc85 100644
--- a/xmrstak/backend/cpu/config.tpl
+++ b/xmrstak/backend/cpu/config.tpl
@@ -7,10 +7,15 @@ R"===(
* the maximum performance. When set to a number N greater than 1, this mode will increase the
* cache usage and single thread performance by N times.
*
- * no_prefetch - Some systems can gain up to extra 5% here, but sometimes it will have no difference or make
+ * no_prefetch - Some systems can gain up to extra 5% here, but sometimes it will have no difference or make
* things slower.
*
- * affine_to_cpu - This can be either false (no affinity), or the CPU core number. Note that on hyperthreading
+ * asm - Allow to switch to a assembler version of cryptonight_v8; allowed value [auto, intel, ryzen]
+ * - auto: used the default implementation (no assembler version)
+ * - intel: supports Intel Ivy Bridge (Xeon v2, Core i7/i5/i3 3xxx, Pentium G2xxx, Celeron G1xxx)
+ * - ryzen: AMD Ryzen (1xxx and 2xxx series)
+ *
+ * affine_to_cpu - This can be either false (no affinity), or the CPU core number. Note that on hyperthreading
* systems it is better to assign threads to physical cores. On Windows this usually means selecting
* even or odd numbered cpu numbers. For Linux it will be usually the lower CPU numbers, so for a 4
* physical core CPU you should select cpu numbers 0-3.
@@ -21,8 +26,8 @@ R"===(
* A filled out configuration should look like this:
* "cpu_threads_conf" :
* [
- * { "low_power_mode" : false, "no_prefetch" : true, "affine_to_cpu" : 0 },
- * { "low_power_mode" : false, "no_prefetch" : true, "affine_to_cpu" : 1 },
+ * { "low_power_mode" : false, "no_prefetch" : true, "asm" : "auto", "affine_to_cpu" : 0 },
+ * { "low_power_mode" : false, "no_prefetch" : true, "asm" : "auto", "affine_to_cpu" : 1 },
* ],
* If you do not wish to mine with your CPU(s) then use:
* "cpu_threads_conf" :
diff --git a/xmrstak/backend/cpu/crypto/cryptonight_aesni.h b/xmrstak/backend/cpu/crypto/cryptonight_aesni.h
index 2734760..0ab47e3 100644
--- a/xmrstak/backend/cpu/crypto/cryptonight_aesni.h
+++ b/xmrstak/backend/cpu/crypto/cryptonight_aesni.h
@@ -876,3 +876,24 @@ struct Cryptonight_hash<5>
REPEAT_5(0, CN_FINALIZE);
}
};
+
+extern "C" void cryptonigh_v8_mainloop_ivybridge_asm(cryptonight_ctx* ctx0);
+extern "C" void cryptonigh_v8_mainloop_ryzen_asm(cryptonight_ctx* ctx0);
+
+template<xmrstak_algo ALGO, int asm_version>
+void cryptonight_hash_v2_asm(const void* input, size_t len, void* output, cryptonight_ctx** ctx)
+{
+ constexpr size_t MEM = cn_select_memory<ALGO>();
+
+ keccak((const uint8_t *)input, len, ctx[0]->hash_state, 200);
+ cn_explode_scratchpad<MEM, false, false, ALGO>((__m128i*)ctx[0]->hash_state, (__m128i*)ctx[0]->long_state);
+
+ if (asm_version == 1)
+ cryptonigh_v8_mainloop_ivybridge_asm(ctx[0]);
+ else
+ cryptonigh_v8_mainloop_ryzen_asm(ctx[0]);
+
+ cn_implode_scratchpad<MEM, false, false, ALGO>((__m128i*)ctx[0]->long_state, (__m128i*)ctx[0]->hash_state);
+ keccakf((uint64_t*)ctx[0]->hash_state, 24);
+ extra_hashes[ctx[0]->hash_state[0] & 3](ctx[0]->hash_state, 200, (char*)output);
+}
diff --git a/xmrstak/backend/cpu/jconf.cpp b/xmrstak/backend/cpu/jconf.cpp
index 49da7ae..1f9501c 100644
--- a/xmrstak/backend/cpu/jconf.cpp
+++ b/xmrstak/backend/cpu/jconf.cpp
@@ -108,10 +108,11 @@ bool jconf::GetThreadConfig(size_t id, thd_cfg &cfg)
if(!oThdConf.IsObject())
return false;
- const Value *mode, *no_prefetch, *aff;
+ const Value *mode, *no_prefetch, *aff, *asm_version;
mode = GetObjectMember(oThdConf, "low_power_mode");
no_prefetch = GetObjectMember(oThdConf, "no_prefetch");
aff = GetObjectMember(oThdConf, "affine_to_cpu");
+ asm_version = GetObjectMember(oThdConf, "asm");
if(mode == nullptr || no_prefetch == nullptr || aff == nullptr)
return false;
@@ -140,6 +141,10 @@ bool jconf::GetThreadConfig(size_t id, thd_cfg &cfg)
else
cfg.iCpuAff = -1;
+ if(!asm_version->IsString())
+ return false;
+ cfg.asm_version_str = asm_version->GetString();
+
return true;
}
diff --git a/xmrstak/backend/cpu/jconf.hpp b/xmrstak/backend/cpu/jconf.hpp
index be85503..4ec9165 100644
--- a/xmrstak/backend/cpu/jconf.hpp
+++ b/xmrstak/backend/cpu/jconf.hpp
@@ -24,6 +24,7 @@ public:
struct thd_cfg {
int iMultiway;
bool bNoPrefetch;
+ std::string asm_version_str;
long long iCpuAff;
};
diff --git a/xmrstak/backend/cpu/minethd.cpp b/xmrstak/backend/cpu/minethd.cpp
index 87f4d32..f07c714 100644
--- a/xmrstak/backend/cpu/minethd.cpp
+++ b/xmrstak/backend/cpu/minethd.cpp
@@ -104,7 +104,7 @@ bool minethd::thd_setaffinity(std::thread::native_handle_type h, uint64_t cpu_id
#endif
}
-minethd::minethd(miner_work& pWork, size_t iNo, int iMultiway, bool no_prefetch, int64_t affinity)
+minethd::minethd(miner_work& pWork, size_t iNo, int iMultiway, bool no_prefetch, int64_t affinity, const std::string& asm_version)
{
this->backendType = iBackend::CPU;
oWork = pWork;
@@ -113,6 +113,7 @@ minethd::minethd(miner_work& pWork, size_t iNo, int iMultiway, bool no_prefetch,
iJobNo = 0;
bNoPrefetch = no_prefetch;
this->affinity = affinity;
+ asm_version_str = asm_version;
std::unique_lock<std::mutex> lck(thd_aff_set);
std::future<void> order_guard = order_fix.get_future();
@@ -441,7 +442,7 @@ std::vector<iBackend*> minethd::thread_starter(uint32_t threadOffset, miner_work
else
printer::inst()->print_msg(L1, "Starting %dx thread, no affinity.", cfg.iMultiway);
- minethd* thd = new minethd(pWork, i + threadOffset, cfg.iMultiway, cfg.bNoPrefetch, cfg.iCpuAff);
+ minethd* thd = new minethd(pWork, i + threadOffset, cfg.iMultiway, cfg.bNoPrefetch, cfg.iCpuAff, cfg.asm_version_str);
pvThreads.push_back(thd);
}
@@ -449,9 +450,31 @@ std::vector<iBackend*> minethd::thread_starter(uint32_t threadOffset, miner_work
}
template<size_t N>
-minethd::cn_hash_fun minethd::func_multi_selector(bool bHaveAes, bool bNoPrefetch, xmrstak_algo algo)
+minethd::cn_hash_fun minethd::func_multi_selector(bool bHaveAes, bool bNoPrefetch, xmrstak_algo algo, const std::string& asm_version_str)
{
static_assert(N >= 1, "number of threads must be >= 1" );
+
+ // check for asm optimized version for cryptonight_v8
+ if(N == 1 && algo == cryptonight_monero_v8 && bHaveAes)
+ {
+ if(asm_version_str != "auto")
+ {
+ if(asm_version_str == "intel")
+ {
+ // Intel Ivy Bridge (Xeon v2, Core i7/i5/i3 3xxx, Pentium G2xxx, Celeron G1xxx)
+ return cryptonight_hash_v2_asm<cryptonight_monero_v8, 1>;
+ }
+ if(asm_version_str == "ryzen")
+ {
+ // AMD Ryzen (1xxx and 2xxx series)
+ return cryptonight_hash_v2_asm<cryptonight_monero_v8, 2>;
+ }
+ else
+ {
+ printer::inst()->print_msg(L1, "Assembler %s unknown, fallback to non asm version of cryptonight_v8", asm_version_str.c_str());
+ }
+ }
+ }
// We have two independent flag bits in the functions
// therefore we will build a binary digit and select the
// function as a two digit binary
@@ -636,7 +659,7 @@ void minethd::multiway_work_main()
// start with root algorithm and switch later if fork version is reached
auto miner_algo = ::jconf::inst()->GetCurrentCoinSelection().GetDescription(1).GetMiningAlgoRoot();
- cn_hash_fun hash_fun_multi = func_multi_selector<N>(::jconf::inst()->HaveHardwareAes(), bNoPrefetch, miner_algo);
+ cn_hash_fun hash_fun_multi = func_multi_selector<N>(::jconf::inst()->HaveHardwareAes(), bNoPrefetch, miner_algo, asm_version_str);
uint8_t version = 0;
size_t lastPoolId = 0;
@@ -671,12 +694,12 @@ void minethd::multiway_work_main()
if(new_version >= coinDesc.GetMiningForkVersion())
{
miner_algo = coinDesc.GetMiningAlgo();
- hash_fun_multi = func_multi_selector<N>(::jconf::inst()->HaveHardwareAes(), bNoPrefetch, miner_algo);
+ hash_fun_multi = func_multi_selector<N>(::jconf::inst()->HaveHardwareAes(), bNoPrefetch, miner_algo, asm_version_str);
}
else
{
miner_algo = coinDesc.GetMiningAlgoRoot();
- hash_fun_multi = func_multi_selector<N>(::jconf::inst()->HaveHardwareAes(), bNoPrefetch, miner_algo);
+ hash_fun_multi = func_multi_selector<N>(::jconf::inst()->HaveHardwareAes(), bNoPrefetch, miner_algo, asm_version_str);
}
lastPoolId = oWork.iPoolId;
version = new_version;
diff --git a/xmrstak/backend/cpu/minethd.hpp b/xmrstak/backend/cpu/minethd.hpp
index 2647854..53ff93c 100644
--- a/xmrstak/backend/cpu/minethd.hpp
+++ b/xmrstak/backend/cpu/minethd.hpp
@@ -32,9 +32,9 @@ public:
private:
template<size_t N>
- static cn_hash_fun func_multi_selector(bool bHaveAes, bool bNoPrefetch, xmrstak_algo algo);
+ static cn_hash_fun func_multi_selector(bool bHaveAes, bool bNoPrefetch, xmrstak_algo algo, const std::string& asm_version_str = "auto");
- minethd(miner_work& pWork, size_t iNo, int iMultiway, bool no_prefetch, int64_t affinity);
+ minethd(miner_work& pWork, size_t iNo, int iMultiway, bool no_prefetch, int64_t affinity, const std::string& asm_version);
template<uint32_t N>
void multiway_work_main();
@@ -60,6 +60,7 @@ private:
bool bQuit;
bool bNoPrefetch;
+ std::string asm_version_str = "auto";
};
} // namespace cpu
--
GitLab