Update of /cvsroot/pure-data/pd/src In directory sc8-pr-cvs1.sourceforge.net:/tmp/cvs-serv27025

Modified Files: Tag: devel_0_38 d_ctl.c m_simd.c m_simd_sse_gcc.c m_simd_sse_gcc.h m_simd_sse_vc.h m_simd_ve_gcc.h m_pd.h Log Message: simd code and loop unrolling for env~ (gcc / sse only at the moment)

Index: m_pd.h =================================================================== RCS file: /cvsroot/pure-data/pd/src/m_pd.h,v retrieving revision 1.4.4.5 retrieving revision 1.4.4.6 diff -C2 -d -r1.4.4.5 -r1.4.4.6 *** m_pd.h 2 Dec 2004 09:25:15 -0000 1.4.4.5 --- m_pd.h 8 Jan 2005 19:43:52 -0000 1.4.4.6 *************** *** 481,485 ****

/* tb: to be called at idle time */ ! EXTERN void set_callback(t_int (*callback) (t_int* argv), t_int* argv, t_int argc);

--- 481,485 ----

/* tb: to be called at idle time */ ! EXTERN void sys_callback(t_int (*callback) (t_int* argv), t_int* argv, t_int argc);

*************** *** 658,661 **** --- 658,662 ---- EXTERN void testcopyvec_8(t_float *dst,const t_float *src,int n); EXTERN void testaddvec_8(t_float *dst,const t_float *src,int n); + EXTERN float sumvec_8(t_float* in, t_int n);

/* vectorized, simd functions * *************** *** 667,670 **** --- 668,672 ---- EXTERN void testcopyvec_simd(t_float *dst,const t_float *src,int n); EXTERN void testaddvec_simd(t_float *dst,const t_float *src,int n); + EXTERN float sumvec_simd(t_float* in, t_int n);

EXTERN int simd_runtime_check(void);

Index: m_simd.c =================================================================== RCS file: /cvsroot/pure-data/pd/src/Attic/m_simd.c,v retrieving revision 1.1.4.1 retrieving revision 1.1.4.2 diff -C2 -d -r1.1.4.1 -r1.1.4.2 *** m_simd.c 5 Nov 2004 13:33:19 -0000 1.1.4.1 --- m_simd.c 8 Jan 2005 19:43:52 -0000 1.1.4.2 *************** *** 66,71 **** }

- - #ifdef DONTUSESIMD int simd_runtime_check() --- 66,69 ---- *************** *** 105,107 **** --- 103,112 ---- }

+ float sumvec_simd(t_float* in, t_int n) + { + return sumvec_8(in,n); + } + + #endif /* DONTUSESIMD */ +

Index: m_simd_sse_gcc.h =================================================================== RCS file: /cvsroot/pure-data/pd/src/Attic/m_simd_sse_gcc.h,v retrieving revision 1.1.4.3 retrieving revision 1.1.4.4 diff -C2 -d -r1.1.4.3 -r1.1.4.4 *** m_simd_sse_gcc.h 29 Nov 2004 18:11:38 -0000 1.1.4.3 --- m_simd_sse_gcc.h 8 Jan 2005 19:43:52 -0000 1.1.4.4 *************** *** 41,44 **** --- 41,47 ---- t_int *sigrsqrt_perf_simd(t_int *w);

+ float sumvec_simd(t_float* in, t_int n); + + //#define sum_vecsimd sumvec_8 #define sigwrap_perf_simd sigwrap_perform /* SIMD not implemented */

Index: m_simd_sse_vc.h =================================================================== RCS file: /cvsroot/pure-data/pd/src/Attic/m_simd_sse_vc.h,v retrieving revision 1.1.4.3 retrieving revision 1.1.4.4 diff -C2 -d -r1.1.4.3 -r1.1.4.4 *** m_simd_sse_vc.h 29 Nov 2004 18:11:38 -0000 1.1.4.3 --- m_simd_sse_vc.h 8 Jan 2005 19:43:52 -0000 1.1.4.4 *************** *** 41,44 **** --- 41,45 ---- t_int *sigrsqrt_perf_simd(t_int *w);

+ #define sum_vecsimd sumvec_8 #define sigwrap_perf_simd sigwrap_perform /* SIMD not implemented */

Index: m_simd_ve_gcc.h =================================================================== RCS file: /cvsroot/pure-data/pd/src/Attic/m_simd_ve_gcc.h,v retrieving revision 1.1.4.1 retrieving revision 1.1.4.2 diff -C2 -d -r1.1.4.1 -r1.1.4.2 *** m_simd_ve_gcc.h 5 Nov 2004 13:33:20 -0000 1.1.4.1 --- m_simd_ve_gcc.h 8 Jan 2005 19:43:52 -0000 1.1.4.2 *************** *** 41,43 **** --- 41,45 ---- t_int *sigrsqrt_perf_simd(t_int *w);

+ #define sum_vecsimd sumvec_8 /* SIMD not implemented */ + #endif /* __M_SIMD_VE_GCC_H */

Index: m_simd_sse_gcc.c =================================================================== RCS file: /cvsroot/pure-data/pd/src/Attic/m_simd_sse_gcc.c,v retrieving revision 1.1.4.7 retrieving revision 1.1.4.8 diff -C2 -d -r1.1.4.7 -r1.1.4.8 *** m_simd_sse_gcc.c 29 Nov 2004 18:11:34 -0000 1.1.4.7 --- m_simd_sse_gcc.c 8 Jan 2005 19:43:52 -0000 1.1.4.8 *************** *** 813,816 **** --- 813,861 ---- }

+ float sumvec_simd(t_float* in, t_int n) + { + float ret = 0; + asm( + ".set T_FLOAT,4 \n" + + "shrl $4, %2 \n" /* divide by 16 */ + "xorps %%xmm4, %%xmm4 \n" /* zero values */ + "xorps %%xmm5, %%xmm5 \n" + "xorps %%xmm6, %%xmm6 \n" + "xorps %0, %0 \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" + + "addps %%xmm0,%%xmm4 \n" + "addps %%xmm1,%%xmm4 \n" + "addps %%xmm2,%%xmm4 \n" + "addps %%xmm3,%%xmm4 \n" + + "addl $16*T_FLOAT,%1 \n" + "loop 1b \n" + + "movhlps %%xmm4, %%xmm5 \n" + "movups %%xmm4, %%xmm6 \n" + "movups %%xmm5, %0 \n" + "shufps $81, %%xmm6, %%xmm6 \n" + "shufps $81, %0, %0 \n" + + "addss %%xmm4, %%xmm5 \n" + "addss %%xmm5, %%xmm6 \n" + "addss %%xmm6, %0 \n" + + + :"=x"(ret) + :"r"(in),"c"(n) + :"%xmm0","%xmm1","%xmm2","%xmm3", "%xmm4","%xmm5","%xmm6"); + return ret; + } + +

#endif

Index: d_ctl.c =================================================================== RCS file: /cvsroot/pure-data/pd/src/d_ctl.c,v retrieving revision 1.3.4.3 retrieving revision 1.3.4.4 diff -C2 -d -r1.3.4.3 -r1.3.4.4 *** d_ctl.c 8 Jan 2005 09:41:02 -0000 1.3.4.3 --- d_ctl.c 8 Jan 2005 19:43:51 -0000 1.3.4.4 *************** *** 653,656 **** --- 653,658 ---- float x_sumbuf[MAXOVERLAP]; /* summing buffer */ float x_f; + float *x_tmpbuf; /* tb: temporary buffer for simd */ + int x_blocksize; /* tb: only for freealignedbytes */ } t_sigenv;

*************** *** 670,674 **** if (period < npoints / MAXOVERLAP + 1) period = npoints / MAXOVERLAP + 1; ! if (!(buf = getbytes(sizeof(float) * (npoints + MAXVSTAKEN)))) { error("env: couldn't allocate buffer"); --- 672,676 ---- if (period < npoints / MAXOVERLAP + 1) period = npoints / MAXOVERLAP + 1; ! if (!(buf = getalignedbytes(sizeof(float) * (npoints + MAXVSTAKEN)))) { error("env: couldn't allocate buffer"); *************** *** 687,690 **** --- 689,694 ---- x->x_outlet = outlet_new(&x->x_obj, gensym("float")); x->x_f = 0; + x->x_blocksize = 64; + x->x_tmpbuf = getalignedbytes(x->x_blocksize * sizeof(float)); return (x); } *************** *** 728,731 **** --- 732,851 ---- }

+ + /* tb: loop unrolling and simd */ + float sumvec_8(t_float* in, t_int n) + { + int i; + float result = 0; + + n>>=3; + for (i = 0; i != n; ++i) + { + result += *in++; + result += *in++; + result += *in++; + result += *in++; + result += *in++; + result += *in++; + result += *in++; + result += *in++; + } + return result; + } + + + static t_int *env_tilde_perf8(t_int *w) + { + t_sigenv *x = (t_sigenv *)(w[1]); + t_float *in = (t_float *)(w[2]); + int n = (int)(w[3]); + int count; + float *sump; + for (count = x->x_phase, sump = x->x_sumbuf; + count < x->x_npoints; count += x->x_realperiod, sump++) + { + float *hp = x->x_buf + count; + float *fp = in; + float sum = *sump; + float *tmp = x->x_tmpbuf; + int i; + t_int sqrargs[4]; + t_int timesargs[5]; + + sqrargs[1] = (t_int)in; + sqrargs[2] = (t_int)tmp; + sqrargs[3] = (t_int) n; + timesargs[1] = (t_int)tmp; + timesargs[2] = (t_int)hp; + timesargs[3] = (t_int)tmp; + timesargs[4] = (t_int)n; + + sqr_perf8(sqrargs); + times_perf8(timesargs); + + *sump += sumvec_8(tmp,n); + } + sump[0] = 0; + x->x_phase -= n; + if (x->x_phase < 0) + { + x->x_result = x->x_sumbuf[0]; + for (count = x->x_realperiod, sump = x->x_sumbuf; + count < x->x_npoints; count += x->x_realperiod, sump++) + sump[0] = sump[1]; + sump[0] = 0; + x->x_phase = x->x_realperiod - n; + clock_delay(x->x_clock, 0L); + } + return (w+4); + } + + static t_int *env_tilde_perf_simd(t_int *w) + { + t_sigenv *x = (t_sigenv *)(w[1]); + t_float *in = (t_float *)(w[2]); + int n = (int)(w[3]); + int count; + float *sump; + for (count = x->x_phase, sump = x->x_sumbuf; + count < x->x_npoints; count += x->x_realperiod, sump++) + { + float *hp = x->x_buf + count; + float *fp = in; + float sum = *sump; + float *tmp = x->x_tmpbuf; + int i; + t_int sqrargs[4]; + t_int timesargs[5]; + + sqrargs[1] = (t_int)in; + sqrargs[2] = (t_int)tmp; + sqrargs[3] = (t_int) n; + timesargs[1] = (t_int)tmp; + timesargs[2] = (t_int)hp; + timesargs[3] = (t_int)tmp; + timesargs[4] = (t_int)n; + + sqr_perf_simd(sqrargs); + times_perf_simd(timesargs); + + *sump += sumvec_simd(tmp,n); + } + sump[0] = 0; + x->x_phase -= n; + if (x->x_phase < 0) + { + x->x_result = x->x_sumbuf[0]; + for (count = x->x_realperiod, sump = x->x_sumbuf; + count < x->x_npoints; count += x->x_realperiod, sump++) + sump[0] = sump[1]; + sump[0] = 0; + x->x_phase = x->x_realperiod - n; + clock_delay(x->x_clock, 0L); + } + return (w+4); + } + + static void env_tilde_dsp(t_sigenv *x, t_signal **sp) { *************** *** 733,738 **** x->x_period + sp[0]->s_n - (x->x_period % sp[0]->s_n); else x->x_realperiod = x->x_period; ! dsp_add(env_tilde_perform, 3, x, sp[0]->s_vec, sp[0]->s_n); ! if (sp[0]->s_n > MAXVSTAKEN) bug("env_tilde_dsp"); }

--- 853,868 ---- x->x_period + sp[0]->s_n - (x->x_period % sp[0]->s_n); else x->x_realperiod = x->x_period; ! ! if (sp[0]->s_n & 7) ! dsp_add(env_tilde_perform, 3, x, sp[0]->s_vec, sp[0]->s_n); ! else ! if (SIMD_CHECK1(sp[0]->s_n, sp[0]->s_vec)) ! dsp_add(env_tilde_perf_simd, 3, x, sp[0]->s_vec, sp[0]->s_n); ! else ! dsp_add(env_tilde_perf8, 3, x, sp[0]->s_vec, sp[0]->s_n); ! ! if (sp[0]->s_n > MAXVSTAKEN) bug("env_tilde_dsp"); ! ! x->x_blocksize = sp[0]->s_n; }

*************** *** 745,749 **** { clock_free(x->x_clock); ! freebytes(x->x_buf, (x->x_npoints + MAXVSTAKEN) * sizeof(float)); }

--- 875,879 ---- { clock_free(x->x_clock); ! freealignedbytes(x->x_buf, (x->x_npoints + MAXVSTAKEN) * sizeof(float)); }