mpv/mp3lib/decode_sse.s

202 lines
4.9 KiB
ArmAsm

///
/// Replacement of synth_1to1() with Intel's SSE SIMD operations support
///
/// This code based 'decode_k7.s' by Nick Kurshev
/// <squash@mb.kcom.ne.jp>,only some types of changes have been made:
///
/// - SSE optimization
/// - change function name for support SSE automatic detect
///
/// Modified by Nick Kurshev <nickols_k@mail.ru>
///
/ synth_1to1_3dnow works the same way as the c version of
/ synth_1to1. this assembler code based 'decode-i586.s'
/ (by Stefan Bieschewski <stb@acm.org>), two types of changes
/ have been made:
/ - use {MMX,3DNow!} instruction for reduce cpu
/ - remove unused(?) local symbols
/
/ useful sources of information on optimizing 3DNow! code include:
/ AMD 3DNow! Technology Manual (Publication #21928)
/ English: http://www.amd.com/K6/k6docs/pdf/21928d.pdf
/ (Japanese: http://www.amd.com/japan/K6/k6docs/j21928c.pdf)
/ AMD-K6-2 Processor Code Optimization Application Note (Publication #21924)
/ English: http://www.amd.com/K6/k6docs/pdf/21924b.pdf
/
/ This code was tested only AMD-K6-2 processor Linux systems,
/ please tell me:
/ - whether this code works on other 3DNow! capable processors
/ (ex.IDT-C6-2) or not
/ - whether this code works on other OSes or not
/
/ by KIMURA Takuhiro <kim@hannah.ipc.miyakyo-u.ac.jp> - until 31.Mar.1998
/ <kim@comtec.co.jp> - after 1.Apr.1998
/ Enhancments for q-word operation by Michael Hipp
.bss
.comm buffs,4352,4
.data
.align 4
bo:
.long 1
.text
/* int synth_1to1(real *bandPtr,int channel,unsigned char *out) */
.globl synth_1to1_sse
synth_1to1_sse:
subl $12,%esp
pushl %ebp
pushl %edi
pushl %esi
pushl %ebx
movl 32(%esp),%eax
movl 40(%esp),%esi
movl $0,%edi
movl bo,%ebp
cmpl %edi,36(%esp)
jne .L48
decl %ebp
andl $15,%ebp
movl %ebp,bo
movl $buffs,%ecx
jmp .L49
.L48:
addl $2,%esi
movl $buffs+2176,%ecx
.L49:
testl $1,%ebp
je .L50
movl %ecx,%ebx
movl %ebp,16(%esp)
pushl %eax
movl 20(%esp),%edx
leal (%ebx,%edx,4),%eax
pushl %eax
movl 24(%esp),%eax
incl %eax
andl $15,%eax
leal 1088(,%eax,4),%eax
addl %ebx,%eax
jmp .L74
.L50:
leal 1088(%ecx),%ebx
leal 1(%ebp),%edx
movl %edx,16(%esp)
pushl %eax
leal 1092(%ecx,%ebp,4),%eax
pushl %eax
leal (%ecx,%ebp,4),%eax
.L74:
pushl %eax
call dct64
addl $12, %esp
movl 16(%esp),%edx
leal 0(,%edx,4),%edx
movl $decwin+64,%eax
movl %eax,%ecx
subl %edx,%ecx
movl $16,%ebp
.L55:
movups (%ecx), %xmm0
mulps (%ebx), %xmm0
movups 16(%ecx), %xmm1
mulps 16(%ebx), %xmm1
addps %xmm1, %xmm0
movups 32(%ecx), %xmm1
mulps 32(%ebx), %xmm1
addps %xmm1, %xmm0
movups 48(%ecx), %xmm1
mulps 48(%ebx), %xmm1
addps %xmm1, %xmm0
/* pfnacc -> PFNACC mmreg1, mmreg2 performs the following operations: */
/* temp = mmreg2 */
/* mmreg1[31:0] = mmreg1[31:0] - mmreg1[63:32] */
/* mmreg1[63:32]= temp [31:0] - temp[63:32] */
/* save difference of mmreg1's low-word and high-word into mmreg1's low-word */
/* save difference of mmreg2's low-word and high-word into mmreg1's high-word */
movhlps %xmm0, %xmm1
addps %xmm1, %xmm0
movaps %xmm0, %xmm1
shufps $0x55, %xmm1, %xmm1 /* fake of pfnacc. 1|1|1|1 */
subss %xmm1, %xmm0
cvtss2si %xmm0, %eax
/ sar $16,%eax
movw %ax,(%esi)
addl $64,%ebx
subl $-128,%ecx
addl $4,%esi
decl %ebp
jnz .L55
/ --- end of loop 1 ---
movups (%ecx), %xmm0
mulps (%ebx), %xmm0
movups 16(%ecx), %xmm1
mulps 16(%ebx), %xmm1
addps %xmm1, %xmm0
movups 32(%ecx), %xmm1
mulps 32(%ebx), %xmm1
addps %xmm1, %xmm0
movups 48(%ecx), %xmm1
mulps 48(%ebx), %xmm1
addps %xmm1, %xmm0
movhlps %xmm0, %xmm1
addss %xmm1, %xmm0
cvtss2si %xmm0, %eax
/ sar $16,%eax
movw %ax,(%esi)
addl $-64,%ebx
addl $4,%esi
addl $256,%ecx
movl $15,%ebp
.L68:
xorps %xmm0, %xmm0
movups (%ecx), %xmm1
mulps (%ebx), %xmm1
subps %xmm1, %xmm0
movups 16(%ecx), %xmm1
mulps 16(%ebx), %xmm1
subps %xmm1, %xmm0
movups 32(%ecx), %xmm1
mulps 32(%ebx), %xmm1
subps %xmm1, %xmm0
movups 48(%ecx), %xmm1
mulps 48(%ebx), %xmm1
subps %xmm1, %xmm0
movhlps %xmm0, %xmm1
subps %xmm1, %xmm0
movaps %xmm0, %xmm1
shufps $0x55, %xmm1, %xmm1 /* fake of pfacc 1|1|1|1 */
addss %xmm1, %xmm0
cvtss2si %xmm0, %eax
/ sar $16,%eax
movw %ax,(%esi)
addl $-64,%ebx
subl $-128,%ecx
addl $4,%esi
decl %ebp
jnz .L68
/ --- end of loop 2
movl %edi,%eax
popl %ebx
popl %esi
popl %edi
popl %ebp
addl $12,%esp
ret