fzorb/REDACTED-rig
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

RandomX: optimized loading from scratchpad

Browse source 
Prefetches scratchpad data as soon as possible to calculate data address for the next load.

Up to ~1.4% speedup on Ryzen 7 3700X @ 4.1 GHz, RAM 3200 MHz 14-14-14-28 with optimized sub-timings:
Variant|Before H/S|After H/S
-------|----------|---------
rx/0|8663|8777
rx/wow|9867|10009
rx/loki|8652|8731
This commit is contained in:
SChernykh 2019-09-11 19:10:01 +02:00
parent 01b2c952ea
commit 2322e3bcf7
9 changed files with 75 additions and 21 deletions
Download patch file Download diff file Expand all files Collapse all files

27
src/crypto/randomx/jit_compiler_x86.cpp
View file

@ -76,6 +76,8 @@ namespace randomx {
*/
const uint8_t* codePrefetchScratchpad = (uint8_t*)&randomx_prefetch_scratchpad;
const uint8_t* codePrefetchScratchpadEnd = (uint8_t*)&randomx_prefetch_scratchpad_end;
const uint8_t* codePrologue = (uint8_t*)&randomx_program_prologue;
const uint8_t* codeLoopBegin = (uint8_t*)&randomx_program_loop_begin;
const uint8_t* codeLoopLoad = (uint8_t*)&randomx_program_loop_load;
@ -93,6 +95,7 @@ namespace randomx {
const uint8_t* codeShhEnd = (uint8_t*)&randomx_sshash_end;
const uint8_t* codeShhInit = (uint8_t*)&randomx_sshash_init;
const int32_t prefetchScratchpadSize = codePrefetchScratchpadEnd - codePrefetchScratchpad;
const int32_t prologueSize = codeLoopBegin - codePrologue;
const int32_t loopLoadSize = codeProgamStart - codeLoopLoad;
const int32_t readDatasetSize = codeReadDatasetLightSshInit - codeReadDataset;
@ -214,7 +217,7 @@ namespace randomx {
generateProgramPrologue(prog, pcfg);
memcpy(code + codePos, RandomX_CurrentConfig.codeReadDatasetTweaked, readDatasetSize);
codePos += readDatasetSize;
generateProgramEpilogue(prog);
generateProgramEpilogue(prog, pcfg);
}
void JitCompilerX86::generateProgramLight(Program& prog, ProgramConfiguration& pcfg, uint32_t datasetOffset) {
@ -225,7 +228,7 @@ namespace randomx {
emitByte(CALL, code, codePos);
emit32(superScalarHashOffset - (codePos + 4), code, codePos);
emit(codeReadDatasetLightSshFin, readDatasetLightFinSize, code, codePos);
generateProgramEpilogue(prog);
generateProgramEpilogue(prog, pcfg);
}
template<size_t N>
@ -266,13 +269,16 @@ namespace randomx {
void JitCompilerX86::generateProgramPrologue(Program& prog, ProgramConfiguration& pcfg) {
memset(registerUsage, -1, sizeof(registerUsage));
codePos = ((uint8_t*)randomx_program_prologue_first_load) - ((uint8_t*)randomx_program_prologue);
code[codePos + 2] = 0xc0 + pcfg.readReg0;
code[codePos + 5] = 0xc0 + pcfg.readReg1;
*(uint32_t*)(code + codePos + 10) = RandomX_CurrentConfig.ScratchpadL3Mask64_Calculated;
*(uint32_t*)(code + codePos + 20) = RandomX_CurrentConfig.ScratchpadL3Mask64_Calculated;
codePos = prologueSize;
memcpy(code + codePos - 48, &pcfg.eMask, sizeof(pcfg.eMask));
emit(REX_XOR_RAX_R64, code, codePos);
emitByte(0xc0 + pcfg.readReg0, code, codePos);
emit(REX_XOR_RAX_R64, code, codePos);
emitByte(0xc0 + pcfg.readReg1, code, codePos);
memcpy(code + codePos, RandomX_CurrentConfig.codeLoopLoadTweaked, loopLoadSize);
memcpy(code + codePos, codeLoopLoad, loopLoadSize);
codePos += loopLoadSize;
for (unsigned i = 0; i < prog.getSize(); ++i) {
Instruction& instr = prog(i);
@ -287,7 +293,12 @@ namespace randomx {
emitByte(0xc0 + pcfg.readReg3, code, codePos);
}
void JitCompilerX86::generateProgramEpilogue(Program& prog) {
void JitCompilerX86::generateProgramEpilogue(Program& prog, ProgramConfiguration& pcfg) {
emit(REX_MOV_RR64, code, codePos);
emitByte(0xc0 + pcfg.readReg0, code, codePos);
emit(REX_XOR_RAX_R64, code, codePos);
emitByte(0xc0 + pcfg.readReg1, code, codePos);
emit(RandomX_CurrentConfig.codePrefetchScratchpadTweaked, prefetchScratchpadSize, code, codePos);
memcpy(code + codePos, codeLoopStore, loopStoreSize);
codePos += loopStoreSize;
emit(SUB_EBX, code, codePos);