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

21
src/crypto/randomx/jit_compiler_x86_static.S
View file

@ -37,7 +37,10 @@
#define WINABI
#endif
.global DECL(randomx_prefetch_scratchpad)
.global DECL(randomx_prefetch_scratchpad_end)
.global DECL(randomx_program_prologue)
.global DECL(randomx_program_prologue_first_load)
.global DECL(randomx_program_loop_begin)
.global DECL(randomx_program_loop_load)
.global DECL(randomx_program_start)
@ -61,6 +64,16 @@
#define db .byte
DECL(randomx_prefetch_scratchpad):
mov rdx, rax
and eax, RANDOMX_SCRATCHPAD_MASK
prefetcht0 [rsi+rax]
ror rdx, 32
and edx, RANDOMX_SCRATCHPAD_MASK
prefetcht0 [rsi+rdx]
DECL(randomx_prefetch_scratchpad_end):
.balign 64
DECL(randomx_program_prologue):
#if defined(WINABI)
@ -71,6 +84,14 @@ DECL(randomx_program_prologue):
movapd xmm13, xmmword ptr [mantissaMask+rip]
movapd xmm14, xmmword ptr [exp240+rip]
movapd xmm15, xmmword ptr [scaleMask+rip]
DECL(randomx_program_prologue_first_load):
xor rax, r8
xor rax, r8
mov rdx, rax
and eax, RANDOMX_SCRATCHPAD_MASK
ror rdx, 32
and edx, RANDOMX_SCRATCHPAD_MASK
jmp DECL(randomx_program_loop_begin)
.balign 64