diff options
| author | Tim Blechmann <timblech@users.sourceforge.net> | 2005-01-24 21:22:31 +0000 |
|---|---|---|
| committer | IOhannes m zmölnig <zmoelnig@iem.at> | 2015-10-14 15:11:58 +0200 |
| commit | bee409f85b7facc33e9e2b9fb7920edb7c3520bd (patch) | |
| tree | a9601e5a65e2405e13731db6e094ee689029d88f /volctl~ | |
| parent | 29520a2b958bcefe3dd774bd71a7b9e0fd2fa18e (diff) | |
improvement for factor 1
svn path=/trunk/externals/tb/; revision=2539
Diffstat (limited to 'volctl~')
| -rw-r--r-- | volctl~/volctl~.c | 71 |
1 files changed, 50 insertions, 21 deletions
diff --git a/volctl~/volctl~.c b/volctl~/volctl~.c index 1a9e451..e49c267 100644 --- a/volctl~/volctl~.c +++ b/volctl~/volctl~.c @@ -53,7 +53,7 @@ typedef struct _volctl void *volctl_new(t_symbol *s, int argc, t_atom *argv) { - if (argc > 2) post("volctl~: extra arguments ignored"); + if (argc > 3) post("volctl~: extra arguments ignored"); t_volctl *x = (t_volctl *)pd_new(volctl_class); inlet_new(&x->x_obj, &x->x_obj.ob_pd, &s_float, gensym("f1")); @@ -171,6 +171,7 @@ static t_int *volctl_perf8(t_int *w) return (w+5); } + static t_int *volctl_perf_simd(t_int *w) { t_volctl * x = (t_volctl *)(w[1]); @@ -229,7 +230,53 @@ static t_int *volctl_perf_simd(t_int *w) } else { - if(x->x_target) + switch(x->x_target) + { + case 0: + asm( + ".set T_FLOAT,4 \n" + + "xorps %%xmm0, %%xmm0 \n" + "shrl $4, %1 \n" + + "1: \n" + "movaps %%xmm0, (%0) \n" + "movaps %%xmm0, 4*T_FLOAT(%0) \n" + "movaps %%xmm0, 8*T_FLOAT(%0) \n" + "movaps %%xmm0, 12*T_FLOAT(%0) \n" + "addl $16*T_FLOAT, %0 \n" + "loop 1b \n" + : + :"r"(out), + "c"(w[4]) + : "%xmm0"); + break; + + case 1: + asm( + ".set T_FLOAT,4 \n" + + "shrl $4, %1 \n" + + "1: \n" + "movaps (%1), %%xmm0 \n" + "movaps 4*T_FLOAT(%1), %%xmm1 \n" + "movaps 8*T_FLOAT(%1), %%xmm2 \n" + "movaps 12*T_FLOAT(%1), %%xmm3 \n" + "movaps %%xmm0, (%2) \n" + "movaps %%xmm1, 4*T_FLOAT(%2) \n" + "movaps %%xmm2, 8*T_FLOAT(%2) \n" + "movaps %%xmm3, 12*T_FLOAT(%2) \n" + + "addl $16*T_FLOAT,%1 \n" + "addl $16*T_FLOAT,%2 \n" + "loop 1b \n" + : + :"c"(w[4]),"r"(in),"r"(out) + :"%xmm0","%xmm1","%xmm2","%xmm3"); + break; + + default: asm( ".set T_FLOAT,4 \n" @@ -257,25 +304,7 @@ static t_int *volctl_perf_simd(t_int *w) : "r"(in), "r"(out), "c"(w[4]),"r"(&(t_float)(x->x_target)) : "%xmm0", "%xmm1","%xmm2","%xmm3","%xmm4"); - else - asm( - ".set T_FLOAT,4 \n" - - "xorps %%xmm0, %%xmm0 \n" - "shrl $4, %1 \n" - - "1: \n" - "movaps %%xmm0, (%0) \n" - "movaps %%xmm0, 4*T_FLOAT(%0) \n" - "movaps %%xmm0, 8*T_FLOAT(%0) \n" - "movaps %%xmm0, 12*T_FLOAT(%0) \n" - "addl $16*T_FLOAT, %0 \n" - "loop 1b \n" - : - :"r"(out), - "c"(w[4]) - : "%xmm0"); - + } } return (w+5); } |