ubergames/lilium-voyager
0
0
Fork 0
mirror of https://github.com/UberGames/lilium-voyager.git synced 2024-12-13 21:51:09 +00:00
Code Issues Projects Releases Packages Wiki Activity

Implement Mathias Benthrup's suggestion for x86 ASM snapvector implementation which reduces cache misses.

Browse source
This commit is contained in:
Thilo Schulz 2011-09-19 18:30:24 +00:00
parent 98af5f4bb0
commit c927fab58f
2 changed files with 22 additions and 18 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

25
code/asm/snapvector.asm
View file

@ -48,14 +48,15 @@ IFDEF idx64
stmxcsr [rsp] ; save SSE control word stmxcsr [rsp] ; save SSE control word
ldmxcsr ssecw ; set to round nearest ldmxcsr ssecw ; set to round nearest
push rdi movaps xmm1, ssemask ; initialize the mask register
mov rdi, rcx ; maskmovdqu uses rdi as implicit memory operand movups xmm0, [rcx] ; here is stored our vector. Read 4 values in one go
movaps xmm1, ssemask ; initialize the mask register for maskmovdqu movaps xmm2, xmm0 ; keep a copy of the original data
movups xmm0, [rdi] ; here is stored our vector. Read 4 values in one go andps xmm0, xmm1 ; set the fourth value to zero in xmm0
andnps xmm1, xmm2 ; copy fourth value to xmm1 and set rest to zero
cvtps2dq xmm0, xmm0 ; convert 4 single fp to int cvtps2dq xmm0, xmm0 ; convert 4 single fp to int
cvtdq2ps xmm0, xmm0 ; convert 4 int to single fp cvtdq2ps xmm0, xmm0 ; convert 4 int to single fp
maskmovdqu xmm0, xmm1 ; write 3 values back to memory orps xmm0, xmm1 ; combine all 4 values again
pop rdi movups [rcx], xmm0 ; write 3 rounded and 1 unchanged values back to memory
ldmxcsr [rsp] ; restore sse control word to old value ldmxcsr [rsp] ; restore sse control word to old value
add rsp, 8 add rsp, 8
@ -69,14 +70,16 @@ ELSE
stmxcsr [esp] ; save SSE control word stmxcsr [esp] ; save SSE control word
ldmxcsr ssecw ; set to round nearest ldmxcsr ssecw ; set to round nearest
push edi mov eax, dword ptr 16[esp] ; store address of vector in eax
mov edi, dword ptr 16[esp] ; maskmovdqu uses edi as implicit memory operand
movaps xmm1, ssemask ; initialize the mask register for maskmovdqu movaps xmm1, ssemask ; initialize the mask register for maskmovdqu
movups xmm0, [edi] ; here is stored our vector. Read 4 values in one go movups xmm0, [eax] ; here is stored our vector. Read 4 values in one go
movaps xmm2, xmm0 ; keep a copy of the original data
andps xmm0, xmm1 ; set the fourth value to zero in xmm0
andnps xmm1, xmm2 ; copy fourth value to xmm1 and set rest to zero
cvtps2dq xmm0, xmm0 ; convert 4 single fp to int cvtps2dq xmm0, xmm0 ; convert 4 single fp to int
cvtdq2ps xmm0, xmm0 ; convert 4 int to single fp cvtdq2ps xmm0, xmm0 ; convert 4 int to single fp
maskmovdqu xmm0, xmm1 ; write 3 values back to memory orps xmm0, xmm1 ; combine all 4 values again
pop edi movups [eax], xmm0 ; write 3 rounded and 1 unchanged values back to memory
ldmxcsr [esp] ; restore sse control word to old value ldmxcsr [esp] ; restore sse control word to old value
add esp, 8 add esp, 8

13
code/asm/snapvector.c
View file

@ -47,17 +47,18 @@ void qsnapvectorsse(vec3_t vec)
"movaps (%0), %%xmm1\n" "movaps (%0), %%xmm1\n"
"movups (%2), %%xmm0\n" "movups (%2), %%xmm0\n"
"movaps %%xmm0, %%xmm2\n"
"andps %%xmm1, %%xmm0\n"
"andnps %%xmm2, %%xmm1\n"
"cvtps2dq %%xmm0, %%xmm0\n" "cvtps2dq %%xmm0, %%xmm0\n"
"cvtdq2ps %%xmm0, %%xmm0\n" "cvtdq2ps %%xmm0, %%xmm0\n"
// vec MUST reside in register rdi as maskmovdqu uses "orps %%xmm1, %%xmm0\n"
// it as an implicit operand. The "D" constraint makes "movups %%xmm0, (%2)\n"
// sure of that.
"maskmovdqu %%xmm1, %%xmm0\n"
"ldmxcsr %3\n" "ldmxcsr %3\n"
: :
: "r" (ssemask), "m" (ssecw), "D" (vec), "m" (oldcw) : "r" (ssemask), "m" (ssecw), "r" (vec), "m" (oldcw)
: "memory", "%xmm0", "%xmm1" : "memory", "%xmm0", "%xmm1", "%xmm2"
); );
} }