Skip to contentSkip to conversation

Audio quality improvement for 16-bit fixed point SBC encoder

PATCH In review Siarhei Siamashka 4 participants 7 messages Jan 21, 2009 – Jan 23, 2009 Bluetooth drivers

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Audio quality improvement for 16-bit fixed point SBC encoder

Hello all,

The attached patch quite noticeably minimizes rounding errors and improves audio quality.

I decided to drop non-SIMD variant because it would require quite a bit of work to update for better precision. Most of the CPU cores which are relevant nowadays have support for some kind of SIMD extension anyway. I will also do ARMv6 SIMD version of the analysis filter after all the high level SBC optimizations are in place.

Audio quality estimation done with tiny_psnr (lower stddev value is better):

=== before patch (4 subbands) ===

./sbc_encode_test.rb BigBuckBunny-stereo.flac
[2, 48000]
["-j -s4 -B16 -b128", "-p -j -l16 -n4 -r1584000"]
--- comparing original / sbcenc + sbcdec ---
stddev: 3.58 PSNR: 85.23 bytes:114519660/114520000

--- comparing original / sbcenc + sbc_decoder.exe ---
stddev: 1.70 PSNR: 91.71 bytes:114519660/114520000

--- comparing original / sbc_encoder.exe + sbc_decoder.exe ---
stddev: 1.44 PSNR: 93.09 bytes:114519660/114520000

--- comparing sbcenc + sbc_decoder.exe / sbc_encoder.exe + sbc_decoder.exe
stddev: 0.99 PSNR: 96.36 bytes:114519808/114519808

=== after patch (4 subbands) ===

./sbc_encode_test.rb BigBuckBunny-stereo.flac
[2, 48000]
["-j -s4 -B16 -b128", "-p -j -l16 -n4 -r1584000"]
--- comparing original / sbcenc + sbcdec ---
stddev: 3.55 PSNR: 85.31 bytes:114519660/114520000

--- comparing original / sbcenc + sbc_decoder.exe ---
stddev: 1.62 PSNR: 92.09 bytes:114519660/114520000

--- comparing original / sbc_encoder.exe + sbc_decoder.exe ---
stddev: 1.44 PSNR: 93.09 bytes:114519660/114520000

--- comparing sbcenc + sbc_decoder.exe / sbc_encoder.exe + sbc_decoder.exe
stddev: 0.77 PSNR: 98.57 bytes:114519808/114519808

=== before patch (8 subbands) ===

./sbc_encode_test.rb BigBuckBunny-stereo.flac
[2, 48000]
["-j -s8 -B16 -b255", "-p -j -l16 -n8 -r1569000"]
--- comparing original / sbcenc + sbcdec ---
stddev: 4.85 PSNR: 82.60 bytes:114519260/114520000

--- comparing original / sbcenc + sbc_decoder.exe ---
stddev: 2.07 PSNR: 89.98 bytes:114519260/114520000

--- comparing original / sbc_encoder.exe + sbc_decoder.exe ---
stddev: 1.09 PSNR: 95.56 bytes:114519260/114520000

--- comparing sbcenc + sbc_decoder.exe / sbc_encoder.exe + sbc_decoder.exe
stddev: 1.77 PSNR: 91.34 bytes:114519552/114519552

=== after patch (8 subbands) ===

./sbc_encode_test.rb BigBuckBunny-stereo.flac
[2, 48000]
["-j -s8 -B16 -b255", "-p -j -l16 -n8 -r1569000"]
--- comparing original / sbcenc + sbcdec ---
stddev: 4.55 PSNR: 83.16 bytes:114519260/114520000

--- comparing original / sbcenc + sbc_decoder.exe ---
stddev: 1.28 PSNR: 94.11 bytes:114519260/114520000

--- comparing original / sbc_encoder.exe + sbc_decoder.exe ---
stddev: 1.09 PSNR: 95.56 bytes:114519260/114520000

--- comparing sbcenc + sbc_decoder.exe / sbc_encoder.exe + sbc_decoder.exe
stddev: 0.73 PSNR: 98.96 bytes:114519552/114519552

===

So for 4 subbands encode, stddev is down from 1.70 to 1.62 (1.44 for the reference encoder). For 8 subbands encode stddev is down from 2.07 to 1.28 (1.09 for the reference encoder).

It is very interesting to see what a more advanced PEAQ test will show.

Best regards,
Siarhei Siamashka

Quoted 1 line

From 7b6b60b25fbd20b9eab0c3f50d4ab121e4aee058 Mon Sep 17 00:00:00 2001

From: Siarhei Siamashka <[email protected]>
Date: Thu, 22 Jan 2009 00:12:40 +0200
Subject: [PATCH] Audio quality improvement for 16-bit fixed point SBC encoder

Multiplying the first part of the analysis filter constant tables by some coefficients and dividing the second part by the same coefficients is a transformation which should produce the same results if rounding errors are not taken into account. These additional C0/C1/... coefficients can be varied in a certain range (the requirement is that we still do not get overflows). The 'magic' values for these coefficients are selected in such a way that the rounding errors are minimized (rounding errors are unavoidable when putting all the floating constants into 16-bit tables and losing some of the fractional part).

Also non-SIMD variant of the analysis filter is dropped because keeping it would require applying a similar change to its tables, which is a bit tricky and just increases maintenance overhead.

---
 sbc/sbc_primitives.c |  157 ++----------------
 sbc/sbc_tables.h     |  460 ++++++++++++++++++++++++++++----------------------
 2 files changed, 270 insertions(+), 347 deletions(-)
sbc/sbc_primitives.c +12 -145
index e3a7764..602b473 100644
@@ -34,155 +34,22 @@
 #include "sbc_primitives_neon.h"
 
 /*
- * A standard C code of analysis filter.
- */
-static inline void sbc_analyze_four(const int16_t *in, int32_t *out)
-{
-	FIXED_A t1[4];
-	FIXED_T t2[4];
-	int i = 0, hop = 0;
-
-	/* rounding coefficient */
-	t1[0] = t1[1] = t1[2] = t1[3] =
-		(FIXED_A) 1 << (SBC_PROTO_FIXED4_SCALE - 1);
-
-	/* low pass polyphase filter */
-	for (hop = 0; hop < 40; hop += 8) {
-		t1[0] += (FIXED_A) in[hop] * _sbc_proto_fixed4[hop];
-		t1[1] += (FIXED_A) in[hop + 1] * _sbc_proto_fixed4[hop + 1];
-		t1[2] += (FIXED_A) in[hop + 2] * _sbc_proto_fixed4[hop + 2];
-		t1[1] += (FIXED_A) in[hop + 3] * _sbc_proto_fixed4[hop + 3];
-		t1[0] += (FIXED_A) in[hop + 4] * _sbc_proto_fixed4[hop + 4];
-		t1[3] += (FIXED_A) in[hop + 5] * _sbc_proto_fixed4[hop + 5];
-		t1[3] += (FIXED_A) in[hop + 7] * _sbc_proto_fixed4[hop + 7];
-	}
-
-	/* scaling */
-	t2[0] = t1[0] >> SBC_PROTO_FIXED4_SCALE;
-	t2[1] = t1[1] >> SBC_PROTO_FIXED4_SCALE;
-	t2[2] = t1[2] >> SBC_PROTO_FIXED4_SCALE;
-	t2[3] = t1[3] >> SBC_PROTO_FIXED4_SCALE;
-
-	/* do the cos transform */
-	for (i = 0, hop = 0; i < 4; hop += 8, i++) {
-		out[i] = ((FIXED_A) t2[0] * cos_table_fixed_4[0 + hop] +
-			(FIXED_A) t2[1] * cos_table_fixed_4[1 + hop] +
-			(FIXED_A) t2[2] * cos_table_fixed_4[2 + hop] +
-			(FIXED_A) t2[3] * cos_table_fixed_4[5 + hop]) >>
-			(SBC_COS_TABLE_FIXED4_SCALE - SCALE_OUT_BITS);
-	}
-}
-
-static void sbc_analyze_4b_4s(int16_t *pcm, int16_t *x,
-						int32_t *out, int out_stride)
-{
-	int i;
-
-	/* Input 4 x 4 Audio Samples */
-	for (i = 0; i < 16; i += 4) {
-		x[64 + i] = x[0 + i] = pcm[15 - i];
-		x[65 + i] = x[1 + i] = pcm[14 - i];
-		x[66 + i] = x[2 + i] = pcm[13 - i];
-		x[67 + i] = x[3 + i] = pcm[12 - i];
-	}
-
-	/* Analyze four blocks */
-	sbc_analyze_four(x + 12, out);
-	out += out_stride;
-	sbc_analyze_four(x + 8, out);
-	out += out_stride;
-	sbc_analyze_four(x + 4, out);
-	out += out_stride;
-	sbc_analyze_four(x, out);
-}
-
-static inline void sbc_analyze_eight(const int16_t *in, int32_t *out)
-{
-	FIXED_A t1[8];
-	FIXED_T t2[8];
-	int i, hop;
-
-	/* rounding coefficient */
-	t1[0] = t1[1] = t1[2] = t1[3] = t1[4] = t1[5] = t1[6] = t1[7] =
-		(FIXED_A) 1 << (SBC_PROTO_FIXED8_SCALE-1);
-
-	/* low pass polyphase filter */
-	for (hop = 0; hop < 80; hop += 16) {
-		t1[0] += (FIXED_A) in[hop] * _sbc_proto_fixed8[hop];
-		t1[1] += (FIXED_A) in[hop + 1] * _sbc_proto_fixed8[hop + 1];
-		t1[2] += (FIXED_A) in[hop + 2] * _sbc_proto_fixed8[hop + 2];
-		t1[3] += (FIXED_A) in[hop + 3] * _sbc_proto_fixed8[hop + 3];
-		t1[4] += (FIXED_A) in[hop + 4] * _sbc_proto_fixed8[hop + 4];
-		t1[3] += (FIXED_A) in[hop + 5] * _sbc_proto_fixed8[hop + 5];
-		t1[2] += (FIXED_A) in[hop + 6] * _sbc_proto_fixed8[hop + 6];
-		t1[1] += (FIXED_A) in[hop + 7] * _sbc_proto_fixed8[hop + 7];
-		t1[0] += (FIXED_A) in[hop + 8] * _sbc_proto_fixed8[hop + 8];
-		t1[5] += (FIXED_A) in[hop + 9] * _sbc_proto_fixed8[hop + 9];
-		t1[6] += (FIXED_A) in[hop + 10] * _sbc_proto_fixed8[hop + 10];
-		t1[7] += (FIXED_A) in[hop + 11] * _sbc_proto_fixed8[hop + 11];
-		t1[7] += (FIXED_A) in[hop + 13] * _sbc_proto_fixed8[hop + 13];
-		t1[6] += (FIXED_A) in[hop + 14] * _sbc_proto_fixed8[hop + 14];
-		t1[5] += (FIXED_A) in[hop + 15] * _sbc_proto_fixed8[hop + 15];
-	}
-
-	/* scaling */
-	t2[0] = t1[0] >> SBC_PROTO_FIXED8_SCALE;
-	t2[1] = t1[1] >> SBC_PROTO_FIXED8_SCALE;
-	t2[2] = t1[2] >> SBC_PROTO_FIXED8_SCALE;
-	t2[3] = t1[3] >> SBC_PROTO_FIXED8_SCALE;
-	t2[4] = t1[4] >> SBC_PROTO_FIXED8_SCALE;
-	t2[5] = t1[5] >> SBC_PROTO_FIXED8_SCALE;
-	t2[6] = t1[6] >> SBC_PROTO_FIXED8_SCALE;
-	t2[7] = t1[7] >> SBC_PROTO_FIXED8_SCALE;
-
-	/* do the cos transform */
-	for (i = 0, hop = 0; i < 8; hop += 16, i++) {
-		out[i] = ((FIXED_A) t2[0] * cos_table_fixed_8[0 + hop] +
-			(FIXED_A) t2[1] * cos_table_fixed_8[1 + hop] +
-			(FIXED_A) t2[2] * cos_table_fixed_8[2 + hop] +
-			(FIXED_A) t2[3] * cos_table_fixed_8[3 + hop] +
-			(FIXED_A) t2[4] * cos_table_fixed_8[4 + hop] +
-			(FIXED_A) t2[5] * cos_table_fixed_8[9 + hop] +
-			(FIXED_A) t2[6] * cos_table_fixed_8[10 + hop] +
-			(FIXED_A) t2[7] * cos_table_fixed_8[11 + hop]) >>
-			(SBC_COS_TABLE_FIXED8_SCALE - SCALE_OUT_BITS);
-	}
-}
-
-static void sbc_analyze_4b_8s(int16_t *pcm, int16_t *x,
-						int32_t *out, int out_stride)
-{
-	int i;
-
-	/* Input 4 x 8 Audio Samples */
-	for (i = 0; i < 32; i += 8) {
-		x[128 + i] = x[0 + i] = pcm[31 - i];
-		x[129 + i] = x[1 + i] = pcm[30 - i];
-		x[130 + i] = x[2 + i] = pcm[29 - i];
-		x[131 + i] = x[3 + i] = pcm[28 - i];
-		x[132 + i] = x[4 + i] = pcm[27 - i];
-		x[133 + i] = x[5 + i] = pcm[26 - i];
-		x[134 + i] = x[6 + i] = pcm[25 - i];
-		x[135 + i] = x[7 + i] = pcm[24 - i];
-	}
-
-	/* Analyze four blocks */
-	sbc_analyze_eight(x + 24, out);
-	out += out_stride;
-	sbc_analyze_eight(x + 16, out);
-	out += out_stride;
-	sbc_analyze_eight(x + 8, out);
-	out += out_stride;
-	sbc_analyze_eight(x, out);
-}
-
-/*
  * A reference C code of analysis filter with SIMD-friendly tables
  * reordering and code layout. This code can be used to develop platform
  * specific SIMD optimizations. Also it may be used as some kind of test
  * for compiler autovectorization capabilities (who knows, if the compiler
  * is very good at this stuff, hand optimized assembly may be not strictly
  * needed for some platform).
+ *
+ * Note: It is also possible to make a simple variant of analysis filter,
+ * which needs only a single constants table without taking care about
+ * even/odd cases. This simple variant of filter can be implemented without
+ * input data permutation. The only thing that would be lost is the
+ * possibility to use pairwise SIMD multiplications. But for some simple
+ * CPU cores without SIMD extensions it can be useful. If anybody is
+ * interested in implementing such variant of a filter, sourcecode from
+ * bluez versions 4.26/4.27 can be used as a reference and the history of
+ * the changes in git repository done around that time may be worth checking.
  */
 
 static inline void sbc_analyze_four_simd(const int16_t *in, int32_t *out,
@@ -398,8 +265,8 @@ static inline void sbc_analyze_4b_8s_simd(int16_t *pcm, int16_t *x,
 void sbc_init_primitives(struct sbc_encoder_state *state)
 {
 	/* Default implementation for analyze functions */
-	state->sbc_analyze_4b_4s = sbc_analyze_4b_4s;
-	state->sbc_analyze_4b_8s = sbc_analyze_4b_8s;
+	state->sbc_analyze_4b_4s = sbc_analyze_4b_4s_simd;
+	state->sbc_analyze_4b_8s = sbc_analyze_4b_8s_simd;
 
 	/* X86/AMD64 optimizations */
 #ifdef SBC_BUILD_WITH_MMX_SUPPORT
sbc/sbc_tables.h +258 -202

495 lines not on this page · raw message

Signature · 1 line
1.5.6.5
#1 lore raw

Hello all,

Quoted 8 lines

Hello all,

The attached patch quite noticeably minimizes rounding errors and improves audio quality.

It is very interesting to see what a more advanced PEAQ test will show.

The PEAQ results for latest version and the latest plus your latest patch can be found
http://net.cs.uni-tuebingen.de/html/nexgenvoip/ in latest and latest+patch.

I need to write a diff webpage...

Greetings

Christian

#2 lore raw
In reply to Siarhei Siamashka

Hello Siarhei,

Quoted 8 lines

Hello all,

The attached patch quite noticeably minimizes rounding errors and improves audio quality.

It is very interesting to see what a more advanced PEAQ test will show.

The PEAQ results for latest version and the latest plus your latest patch can be found
http://net.cs.uni-tuebingen.de/html/nexgenvoip/ in latest and latest+patch.

Congratulations, the encoder is perfect now. Sometimes even better than the reference!

Greetings

Christian

#3 lore raw
Quoted 8 lines · On Thursday 22 January 2009 13:58:57 ext Christian Hoene wrote:

Hello Siarhei,

Quoted 6 lines

Hello all,

The attached patch quite noticeably minimizes rounding errors and improves audio quality.

Quoted 8 lines
Quoted 1 line

It is very interesting to see what a more advanced PEAQ test will show.

The PEAQ results for latest version and the latest plus your latest patch can be found
http://net.cs.uni-tuebingen.de/html/nexgenvoip/ in latest and latest+patch.

Congratulations, the encoder is perfect now. Sometimes even better than the reference!

Thanks. The results have really exceeded my expectations. Looks like the precision loss on rounding is now really insignificant so that the rounding errors are now smaller than the sensitivity of PEAQ method. My guess is that very minor differences in results in both directions are just some kind of random deviation and can't be clearly interpreted as an advantage of either implementation.

So appears that the perceived quality should be really good now (PSNR rating is a bit worse than reference, but it is not an objective way to measure audio quality). Looks like there is even no need to introduce a high precision configuration option for enabling 32-bit fixed point implementation in practice. It makes everything a bit easier :)

Signature · 2 lines
Best regards,
Siarhei Siamashka
#4 lore raw
In reply to Siarhei Siamashka

Hi Siarhei,

Quoted 5 lines

I decided to drop non-SIMD variant because it would require quite a bit of work to update for better precision. Most of the CPU cores which are relevant nowadays have support for some kind of SIMD extension anyway. I will also do ARMv6 SIMD version of the analysis filter after all the high level SBC optimizations are in place.

Perhaps we can just disable it, since it is probably useful to maintain a version in C as a reference code just in case someone want to do its own optimizations in the future.

Signature · 2 lines
Luiz Augusto von Dentz
Engenheiro de Computação
#5 lore raw
Quoted 11 lines · On Thursday 22 January 2009 15:36:36 ext Luiz Augusto von Dentz wrote:

Hi Siarhei,

Quoted 5 lines

I decided to drop non-SIMD variant because it would require quite a bit of work to update for better precision. Most of the CPU cores which are relevant nowadays have support for some kind of SIMD extension anyway. I will also do ARMv6 SIMD version of the analysis filter after all the high level SBC optimizations are in place.

Perhaps we can just disable it, since it is probably useful to maintain a version in C as a reference code just in case someone want to do its own optimizations in the future.

Right now there are two reference C versions:
1. "simple" one which uses smaller constant tables and may be modified not to require any input data reordering (actually it reverses the order of audio samples, but this can be avoided).
2. "simd-friendly" one with larger data tables and it also has to reorder input data in all cases.

Extra size for constant tables is not an issue because a good optimizing compiler should be able to optimize the constants pool. Let's consider the following simplified example:

/*************************/
const short table1[4] = { 0x1234, 0x4321, 0x0000, 0x1234 }; const short table2[4] = { 0x4321, 0x1234, 0x1234, 0x0000 };

static inline int dotproduct(const short *x, const short *y)

{
    return x[0] * y[0] + x[1] * y[1] + x[2] * y[2] + x[3] * y[3];
}

int f(const short *in, int *out)

{
    out[0] = dotproduct(in + 0, table1);
    out[1] = dotproduct(in + 4, table2);
}

/*************************/

It compiles into the following code for x86 (gcc 4.3.2):

00000000 <f>:

   0:   53                      push   %ebx
   1:   8b 4c 24 08             mov    0x8(%esp),%ecx
   5:   8b 5c 24 0c             mov    0xc(%esp),%ebx
   9:   0f bf 51 02             movswl 0x2(%ecx),%edx
   d:   0f bf 41 06             movswl 0x6(%ecx),%eax
  11:   69 d2 21 43 00 00       imul   $0x4321,%edx,%edx
  17:   69 c0 34 12 00 00       imul   $0x1234,%eax,%eax
  1d:   01 c2                   add    %eax,%edx
  1f:   0f bf 01                movswl (%ecx),%eax
  22:   69 c0 34 12 00 00       imul   $0x1234,%eax,%eax
  28:   01 c2                   add    %eax,%edx
  2a:   8d 41 08                lea    0x8(%ecx),%eax
  2d:   89 13                   mov    %edx,(%ebx)
  2f:   0f bf 50 02             movswl 0x2(%eax),%edx
  33:   0f bf 40 04             movswl 0x4(%eax),%eax
  37:   01 d0                   add    %edx,%eax
  39:   0f bf 51 08             movswl 0x8(%ecx),%edx
  3d:   69 c0 34 12 00 00       imul   $0x1234,%eax,%eax
  43:   69 d2 21 43 00 00       imul   $0x4321,%edx,%edx
  49:   01 d0                   add    %edx,%eax
  4b:   89 43 04                mov    %eax,0x4(%ebx)
  4e:   5b                      pop    %ebx
  4f:   c3                      ret

The compiler did not use any tables at all, but emitted all the constants as immediate operands for instructions. Also it eliminated all the multiplications with zero constants (so we have only 6 IMUL instructions in the code). So gcc seems to be clever enough to optimize this code well.

On ARM the generated code is the following (gcc 4.2.1, -mcpu=arm926ej-s):

00000000 <f>:

   0:   e92d41f0        push    {r4, r5, r6, r7, r8, lr}
   4:   e59fc040        ldr     ip, [pc, #64]   ; 4c <table2+0x44>
   8:   e2808008        add     r8, r0, #8      ; 0x8
   c:   e59f703c        ldr     r7, [pc, #60]   ; 50 <table2+0x48>
  10:   e1d030b2        ldrh    r3, [r0, #2]
  14:   e1d820b2        ldrh    r2, [r8, #2]
  18:   e1d0e0f0        ldrsh   lr, [r0]
  1c:   e1d050f8        ldrsh   r5, [r0, #8]
  20:   e1630783        smulbb  r3, r3, r7
  24:   e1620c82        smulbb  r2, r2, ip
  28:   e0263c9e        mla     r6, lr, ip, r3
  2c:   e0242795        mla     r4, r5, r7, r2
  30:   e1d830f4        ldrsh   r3, [r8, #4]
  34:   e1d020f6        ldrsh   r2, [r0, #6]
  38:   e0204c93        mla     r0, r3, ip, r4
  3c:   e02e6c92        mla     lr, r2, ip, r6
  40:   e5810004        str     r0, [r1, #4]
  44:   e581e000        str     lr, [r1]
  48:   e8bd81f0        pop     {r4, r5, r6, r7, r8, pc}
  4c:   00001234        .word   0x00001234
  50:   00004321        .word   0x00004321

Here the compiler reduced the tables to only 2 constants. It was also able to eliminate multiplications by zero. Regarding 16-bit constants, it could use only 2 fast 16-bit SMULBB instructions, performing the rest of multiplications with a slower 32-bit MLA. So the compiler is not very clever about generating optimal code, but it at least could perform some basic optimizations.

Of course, when handling a more complex code, the compiler may screw up something and miss some optimization opportunities. But if it happens, bugreport should be submitted to gcc. In any case, handwritten assembly is still much better for such type of code at the moment, at least on ARM.

So the only reason to have "simple" C reference version are the potential savings on input samples reordering. But it is probably not worth the efforts. In addition, when having non-native byte order for input data, "simple" version will gain nothing because processing and copying data will be still unavoidable.

The more I think about it, the more I'm getting inclined to the idea that only SIMD-style version of C reference code should be kept in order to have better maintainability.

Signature · 2 lines
Best regards,
Siarhei Siamashka
#6 lore raw
In reply to Siarhei Siamashka

Hi,

Quoted 6 lines · Siarhei Siamashka

The attached patch quite noticeably minimizes rounding errors and improves audio quality.

I decided to drop non-SIMD variant because it would require quite a bit of work to update for better precision. Most of the CPU cores which are relevant nowadays have support for some kind of SIMD extension anyway.

This patch is now also pushed upstream. Thanks.

Johan