Sweetbread/kolibrios
1
0
Fork 0
forked from KolibriOS/kolibrios
Code Issues Pull Requests Actions Activity

Fast two-stream mixer

Browse Source 
git-svn-id: svn://kolibrios.org@697 a494cfbc-eb01-0410-851d-a64ba20cac60
This commit is contained in:
Sergey Semyonov (Serge) 2008-01-17 08:27:36 +00:00
parent 48ecbd9104
commit df2ca4bbd5
1 changed files with 80 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

81
kernel/trunk/drivers/mixer.asm
View File

@ -36,6 +36,8 @@ proc new_mix stdcall, output:dword
test eax, eax test eax, eax
je .done je .done
if USE_SSE2_MIXER
cmp eax, 1 cmp eax, 1
ja @F ja @F
;use fast path ;use fast path
@ -44,6 +46,17 @@ proc new_mix stdcall, output:dword
call mix_fast call mix_fast
jmp .next jmp .next
@@: @@:
cmp eax, 2
ja @F
mov edi, [output]
lea edx, [mix_list]
call mix_fast_2_stream
jmp .next
@@:
end if
lea ebx, [mix_list] lea ebx, [mix_list]
stdcall mix_all, [output], ebx, eax stdcall mix_all, [output], ebx, eax
.next: .next:
@ -346,7 +359,7 @@ mix_fast:
shufps xmm2, xmm2, 0x88 ; volume level Rf Lf Rf Lf shufps xmm2, xmm2, 0x88 ; volume level Rf Lf Rf Lf
.mix: .mix:
movdqa xmm1, [eax] ; R3w L3w R2w L2w R1w L1w R0w L0w movdqa xmm1, [eax] ; R3w L3w R2w L2w R1w L1w R0w L0w
add eax, 16 add eax, 16
punpcklwd xmm0, xmm1 ; R1w R1w L1w L1W R0w R0w L0w L0w punpcklwd xmm0, xmm1 ; R1w R1w L1w L1W R0w R0w L0w L0w
punpckhwd xmm1, xmm1 ; R3w R3w L3w L3w R2w R2w L2w L2w punpckhwd xmm1, xmm1 ; R3w R3w L3w L3w R2w R2w L2w L2w
@ -372,6 +385,72 @@ mix_fast:
ret ret
align 4
mix_fast_2_stream:
mov ebx, 32
mov eax, [edx]
movss xmm4, [edx+4] ; vol Lf
movss xmm5, [edx+8] ; vol Rf
mov ecx, [edx+12]
movss xmm6, [edx+16] ; vol Lf
movss xmm7, [edx+20] ; vol Rf
shufps xmm4, xmm5, 0 ; Rf Rf Lf Lf
shufps xmm4, xmm4, 0x88 ; volume level Rf Lf Rf Lf
shufps xmm6, xmm7, 0 ; Rf Rf Lf Lf
shufps xmm6, xmm6, 0x88 ; volume level Rf Lf Rf Lf
.mix:
movdqa xmm1, [eax] ; R3w L3w R2w L2w R1w L1w R0w L0w
movdqa xmm3, [ecx] ; R3w L3w R2w L2w R1w L1w R0w L0w
add eax, 16
add ecx, 16
punpcklwd xmm0, xmm1 ; R1w R1w L1w L1W R0w R0w L0w L0w
punpckhwd xmm1, xmm1 ; R3w R3w L3w L3w R2w R2w L2w L2w
psrad xmm0, 16 ; R1d L1d R0d L0d
psrad xmm1, 16 ; R3d L3d R2d L2d
cvtdq2ps xmm0, xmm0 ; time to use all power
cvtdq2ps xmm1, xmm1 ; of the dark side
mulps xmm0, xmm4 ; R1f' L1f' R0f' L0f'
mulps xmm1, xmm4 ; R3f' L3f' R2f' L2f'
punpcklwd xmm2, xmm3 ; R1w R1w L1w L1W R0w R0w L0w L0w
punpckhwd xmm3, xmm3 ; R3w R3w L3w L3w R2w R2w L2w L2w
psrad xmm2, 16 ; R1d L1d R0d L0d
psrad xmm3, 16 ; R3d L3d R2d L2d
cvtdq2ps xmm2, xmm2 ; time to use all power
cvtdq2ps xmm3, xmm3 ; of the dark side
mulps xmm2, xmm6 ; R1f' L1f' R0f' L0f'
mulps xmm3, xmm6 ; R3f' L3f' R2f' L2f'
addps xmm0, xmm2
addps xmm1, xmm3
cvtps2dq xmm0, xmm0 ; R1d' L1d' R0d' L0d'
cvtps2dq xmm1, xmm1 ; R3d' L3d' R2d' L2d'
packssdw xmm0, xmm0 ; R1w' L1w' R0w' L0w' R1w' L1w' R0w' L0w'
packssdw xmm1, xmm1 ; R3w' L3w' R2w' L2w' R3w' L3w' R2w' L2w'
punpcklqdq xmm0, xmm1 ; R3w' L3w' R2w' L2w' R1w' L1w' R0w' L0w'
movntdq [edi], xmm0
add edi, 16
dec ebx
jnz .mix
ret
else ; fixed point mmx version else ; fixed point mmx version
align 4 align 4