From f2b6ecd5fef3487daf9493b25f880b52af1bf3e0 Mon Sep 17 00:00:00 2001 From: Michael Sevakis Date: Sat, 3 Mar 2007 01:19:35 +0000 Subject: SPC codec on Coldfire: Move movclrs into the light and out of the long dark shadow cast my emac latency as much as possible. Put in a faster interpolation routine (emac saves the day...again). Add comments about what's going on. git-svn-id: svn://svn.rockbox.org/rockbox/trunk@12558 a1c6a512-1295-4272-9138-f99709370657 --- apps/codecs/spc/Spc_Dsp.h | 93 +++++++++++++++++++++++++++++++---------------- 1 file changed, 62 insertions(+), 31 deletions(-) (limited to 'apps/codecs/spc/Spc_Dsp.h') diff --git a/apps/codecs/spc/Spc_Dsp.h b/apps/codecs/spc/Spc_Dsp.h index 6b530a7a62..fdcd37f226 100644 --- a/apps/codecs/spc/Spc_Dsp.h +++ b/apps/codecs/spc/Spc_Dsp.h @@ -974,23 +974,35 @@ static void DSP_run_( struct Spc_Dsp* this, long count, int32_t* out_buf ) if ( (this->r.g.noise_enables & vbit) == 0 ) { uint32_t f = voice->position; - int32_t y1; + int32_t y0; + /** + * Formula (fastest found so far of MANY): + * output = y0 + f*y1 - f*y0 + */ asm volatile ( - "move.l %[f], %[y0] \r\n" /* separate fraction */ - "and.l #0xfff, %[f] \r\n" /* and whole parts */ - "lsr.l %[sh], %[y0] \r\n" - "move.l 2(%[s], %[y0].l*2), %[y1] \r\n" /* load two samples */ - "move.l %[y1], %[y0] \r\n" /* separate samples */ - "ext.l %[y1] \r\n" /* y0=s[1], y1=s[2] */ - "swap %[y0] \r\n" - "ext.l %[y0] \r\n" - "sub.l %[y0], %[y1] \r\n" /* diff = y1 - y0 */ - "muls.l %[f], %[y1] \r\n" /* y0 += f*diff */ - "asr.l %[sh], %[y1] \r\n" - "add.l %[y1], %[y0] \r\n" - : [f]"+&d"(f), [y0]"=&d"(output), [y1]"=&d"(y1) - : [s]"a"(voice->samples), [sh]"r"(12) + /* separate fractional and whole parts */ + "move.l %[f], %[y1] \r\n" + "and.l #0xfff, %[f] \r\n" + "lsr.l %[sh], %[y1] \r\n" + /* load samples y0 (upper) & y1 (lower) */ + "move.l 2(%[s], %[y1].l*2), %[y1] \r\n" + /* %acc0 = f*y1 */ + "mac.w %[f]l, %[y1]l, %%acc0 \r\n" + /* msac.w is 2% boostier so add negative */ + "neg.l %[f] \r\n" + /* %acc0 -= f*y0 */ + "mac.w %[f]l, %[y1]u, %%acc0 \r\n" + /* separate out y0 and sign extend */ + "swap %[y1] \r\n" + "movea.w %[y1], %[y0] \r\n" + /* fetch result, scale down and add y0 */ + "movclr.l %%acc0, %[y1] \r\n" + /* output = y0 + (result >> 12) */ + "asr.l %[sh], %[y1] \r\n" + "add.l %[y0], %[y1] \r\n" + : [f]"+&d"(f), [y0]"=&a"(y0), [y1]"=&d"(output) + : [s]"a"(voice->samples), [sh]"d"(12) ); } @@ -1093,9 +1105,13 @@ static void DSP_run_( struct Spc_Dsp* this, long count, int32_t* out_buf ) *this->last_fir_ptr = fb; this->last_fir_ptr = this->fir_ptr; - /* Apply echo FIR filter to output - circular buffer is hardware - incremented and masked; FIR coefficients and buffer history are - loaded in parallel with multiply accumulate operations. */ + /* Apply echo FIR filter to output samples read from echo buffer - + circular buffer is hardware incremented and masked; FIR + coefficients and buffer history are loaded in parallel with + multiply accumulate operations. Shift left by one here and once + again when calculating feedback to have sample values justified + to bit 31 in the output to ease endian swap, interleaving and + clamping before placing result in the program's echo buffer. */ int _0, _1, _2; asm volatile ( "move.l (%[fir_c]) , %[_2] \r\n" @@ -1115,53 +1131,68 @@ static void DSP_run_( struct Spc_Dsp* this, long count, int32_t* out_buf ) "mac.w %[_1]l, %[_2]u, << , %%acc1 \r\n" "mac.w %[_0]u, %[_2]l, << , %%acc0 \r\n" "mac.w %[_0]l, %[_2]l, << , %%acc1 \r\n" - "movclr.l %%acc0, %[out_0] \r\n" - "movclr.l %%acc1, %[out_1] \r\n" : [_0]"=&r"(_0), [_1]"=&r"(_1), [_2]"=&r"(_2), - [fir_p]"+a"(this->fir_ptr), - [out_0]"=r"(out_0), [out_1]"=r"(out_1) + [fir_p]"+a"(this->fir_ptr) : [fir_c]"a"(this->fir_coeff), [fb]"r"(fb) ); /* Generate output */ asm volatile ( + /* fetch filter results to eliminate stalls */ + "movclr.l %%acc0, %[out_0] \r\n" + "movclr.l %%acc1, %[out_1] \r\n" + /* apply global volume */ "mac.l %[chans_0], %[gv_0] , %%acc2 \r\n" "mac.l %[chans_1], %[gv_1] , %%acc3 \r\n" + /* apply echo volume and add to final output */ "mac.l %[ev_0], %[out_0], >>, %%acc2 \r\n" "mac.l %[ev_1], %[out_1], >>, %%acc3 \r\n" - : + : [out_0]"=&r"(out_0), [out_1]"=&r"(out_1) : [chans_0]"r"(chans_0), [gv_0]"r"(global_vol_0), [ev_0]"r"((int)this->r.g.echo_volume_0), [chans_1]"r"(chans_1), [gv_1]"r"(global_vol_1), - [ev_1]"r"((int)this->r.g.echo_volume_1), - [out_0]"r"(out_0), [out_1]"r"(out_1) + [ev_1]"r"((int)this->r.g.echo_volume_1) ); /* Feedback into echo buffer */ if ( !(this->r.g.flags & 0x20) ) { asm volatile ( - "mac.l %[sh], %[e0] , %%acc0 \r\n" - "mac.l %[out_0], %[ef], <<, %%acc0 \r\n" + /* scale echo voices; saturate if overflow */ "mac.l %[sh], %[e1] , %%acc1 \r\n" + "mac.l %[sh], %[e0] , %%acc0 \r\n" + /* add scaled output from FIR filter */ "mac.l %[out_1], %[ef], <<, %%acc1 \r\n" - "movclr.l %%acc0, %[e0] \r\n" + "mac.l %[out_0], %[ef], <<, %%acc0 \r\n" + /* swap and fetch feedback results - simply + swap_odd_even32 mixed in between macs and + movclrs to mitigate stall issues */ + "move.l #0x00ff00ff, %[sh] \r\n" "movclr.l %%acc1, %[e1] \r\n" "swap %[e1] \r\n" + "movclr.l %%acc0, %[e0] \r\n" "move.w %[e1], %[e0] \r\n" + "and.l %[e0], %[sh] \r\n" + "eor.l %[sh], %[e0] \r\n" + "lsl.l #8, %[sh] \r\n" + "lsr.l #8, %[e0] \r\n" + "or.l %[sh], %[e0] \r\n" + /* save final feedback into echo buffer */ + "move.l %[e0], (%[echo_ptr]) \r\n" : [e0]"+&d"(echo_0), [e1]"+&d"(echo_1) : [out_0]"r"(out_0), [out_1]"r"(out_1), [ef]"r"((int)this->r.g.echo_feedback), - [sh]"r"(1 << 9) + [echo_ptr]"a"((int32_t *)echo_ptr), + [sh]"d"(1 << 9) ); - - *(int32_t *)echo_ptr = swap_odd_even32(echo_0); } /* Output final samples */ asm volatile ( + /* fetch output saved in %acc2 and %acc3 */ "movclr.l %%acc2, %[out_0] \r\n" "movclr.l %%acc3, %[out_1] \r\n" + /* scale right by global_muting shift */ "asr.l %[gm], %[out_0] \r\n" "asr.l %[gm], %[out_1] \r\n" : [out_0]"=&d"(out_0), [out_1]"=&d"(out_1) -- cgit v1.2.3