thr3ads.net - Speex dev - [Speex-dev] [PATCH 3/5] resample: Add NEON optimized inner_product

If this information is useful, please help other people find it:
Share via:

Jyri Sarha

2011-Sep-01 12:30 UTC

[Speex-dev] [PATCH 0/5] ARM NEON optimization for samplerate converter

From: Jyri Sarha <jsarha at ti.com>

I optimized Speex resampler for NEON capable ARM CPUs. The first patch
should speed up resampling on any platform that can spare the
increased memory usage. It would be nice to have these merged to the
master branch. Please let me know if there is anything I can do to 
help the the merge. The patches have been rebased on top of master 
branch in http://git.xiph.org/speex.git and the relevant parts have
been tested on ARM and x86 systems.

Cheers,
Jyri

Jyri Sarha (5):
  resample: Calculate full sinc table (e.g. no sinc interpolation)
  configure.ac: Add --enable-resample-full-sinc-table conf flag
  resample: Add NEON optimized inner_product_single for fixed point
  configure.ac: Add ARM NEON support
  resample: Add NEON optimized inner_product_single for floating point

 configure.ac             |   35 ++++++++
 libspeex/arch.h          |    1 +
 libspeex/fixed_generic.h |    4 +
 libspeex/resample.c      |   14 +++-
 libspeex/resample_neon.h |  201 ++++++++++++++++++++++++++++++++++++++++++++++
 5 files changed, 253 insertions(+), 2 deletions(-)
 create mode 100644 libspeex/resample_neon.h

-- 
1.7.4.1

Jyri Sarha

2011-Sep-01 12:30 UTC

head link

[Speex-dev] [PATCH 1/5] resample: Calculate full sinc table (e.g. no sinc interpolation)

From: Jyri Sarha <jsarha at ti.com>

On many systems taking couple of kilo-bytes for full sinc table is not
a problem. With a straight filter table the number of MAC operations
needed drops to one quarter. Also the inner loop of the filter becomes
more simple and easier to optimize. So the full sinc table version
runs much faster in most cases. The trade off is some kilo-bytes
bigger memory usage and more CPU needed when initializing the
resampler.
---
 libspeex/resample.c |    4 ++++
 1 files changed, 4 insertions(+), 0 deletions(-)

diff --git a/libspeex/resample.c b/libspeex/resample.c
index 7b5a308..3829715 100644
--- a/libspeex/resample.c
+++ b/libspeex/resample.c
@@ -585,7 +585,11 @@ static void update_filter(SpeexResamplerState *st)
    }
    
    /* Choose the resampling type that requires the least amount of memory */
+#ifdef RESAMPLE_FULL_SINC_TABLE
+   if (1) 
+#else
    if (st->den_rate <= st->oversample)
+#endif
    {
       spx_uint32_t i;
       if (!st->sinc_table)
-- 
1.7.4.1

Jyri Sarha

2011-Sep-01 12:30 UTC

head link

[Speex-dev] [PATCH 2/5] configure.ac: Add --enable-resample-full-sinc-table conf flag

From: Jyri Sarha <jsarha at ti.com>

Use this flag to enable resampling with full sinc table.
---
 configure.ac |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)

diff --git a/configure.ac b/configure.ac
index 3179521..255c0b4 100644
--- a/configure.ac
+++ b/configure.ac
@@ -198,6 +198,11 @@ AC_ARG_ENABLE(fixed-point-debug, [ 
--enable-fixed-point-debug  Debug fixed-poin
   AC_DEFINE([FIXED_DEBUG], , [Debug fixed-point implementation])
 fi])
 
+AC_ARG_ENABLE(resample-full-sinc-table, [  --enable-resample-full-sinc-table
Resample full SINC table],
+[if test "$enableval" = yes; then
+  AC_DEFINE([RESAMPLE_FULL_SINC_TABLE], , [Resample with full SINC table (no
interpolation)])
+fi])
+
 AC_ARG_ENABLE(ti-c55x, [  --enable-ti-c55x        Enable support for TI C55X
DSP],
 [if test "$enableval" = yes; then
   has_char16=yes;
-- 
1.7.4.1

Jyri Sarha

2011-Sep-01 12:30 UTC

head link

[Speex-dev] [PATCH 3/5] resample: Add NEON optimized inner_product_single for fixed point

From: Jyri Sarha <jsarha at ti.com>

Semantics of inner_product_single have also been changed to contain
the final right shift and saturation so it can also be implemented in
the optimal way for the used platform. This change affects fixed point
calculations only.

I also added a new fixed point macro SATURATE32PSHR(x, shift, a). It
does pretty much the same thing as SATURATE32(PSHR32(x, shift), a),
but it avoids over flowing in rounding up phase in the rare occasion
where x has already been saturated. It should also be slightly faster.
---
 libspeex/arch.h          |    1 +
 libspeex/fixed_generic.h |    4 ++
 libspeex/resample.c      |   10 ++++-
 libspeex/resample_neon.h |  100 ++++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 113 insertions(+), 2 deletions(-)
 create mode 100644 libspeex/resample_neon.h

diff --git a/libspeex/arch.h b/libspeex/arch.h
index 3b47ed9..daa72a7 100644
--- a/libspeex/arch.h
+++ b/libspeex/arch.h
@@ -171,6 +171,7 @@ typedef float spx_word32_t;
 #define VSHR32(a,shift) (a)
 #define SATURATE16(x,a) (x)
 #define SATURATE32(x,a) (x)
+#define SATURATE32PSHR(x,shift,a) (x)
 
 #define PSHR(a,shift)       (a)
 #define SHR(a,shift)       (a)
diff --git a/libspeex/fixed_generic.h b/libspeex/fixed_generic.h
index 3fb096e..0e012e9 100644
--- a/libspeex/fixed_generic.h
+++ b/libspeex/fixed_generic.h
@@ -52,6 +52,10 @@
 #define SATURATE16(x,a) (((x)>(a) ? (a) : (x)<-(a) ? -(a) : (x)))
 #define SATURATE32(x,a) (((x)>(a) ? (a) : (x)<-(a) ? -(a) : (x)))
 
+#define SATURATE32PSHR(x,shift,a) (((x)>=(SHL32(a,shift))) ? (a) : \
+                                   (x)<=-(SHL32(a,shift)) ? -(a) : \
+                                   (PSHR32(x, shift)))
+
 #define SHR(a,shift) ((a) >> (shift))
 #define SHL(a,shift) ((spx_word32_t)(a) << (shift))
 #define PSHR(a,shift)
(SHR((a)+((EXTEND32(1)<<((shift))>>1)),shift))
diff --git a/libspeex/resample.c b/libspeex/resample.c
index 3829715..7957c61 100644
--- a/libspeex/resample.c
+++ b/libspeex/resample.c
@@ -99,6 +99,10 @@ static void speex_free (void *ptr) {free(ptr);}
 #include "resample_sse.h"
 #endif
 
+#ifdef _USE_NEON
+#include "resample_neon.h"
+#endif
+
 /* Numer of elements to allocate on the stack */
 #ifdef VAR_ARRAYS
 #define FIXED_STACK_ALLOC 8192
@@ -360,11 +364,12 @@ static int
resampler_basic_direct_single(SpeexResamplerState *st, spx_uint32_t c
       }
       sum = accum[0] + accum[1] + accum[2] + accum[3];
 */
+      sum = SATURATE32PSHR(sum, 15, 32767);
 #else
       sum = inner_product_single(sinc, iptr, N);
 #endif
 
-      out[out_stride * out_sample++] = SATURATE32(PSHR32(sum, 15), 32767);
+      out[out_stride * out_sample++] = sum;
       last_sample += int_advance;
       samp_frac_num += frac_advance;
       if (samp_frac_num >= den_rate)
@@ -470,12 +475,13 @@ static int
resampler_basic_interpolate_single(SpeexResamplerState *st, spx_uint3
 
       cubic_coef(frac, interp);
       sum = MULT16_32_Q15(interp[0],SHR32(accum[0], 1)) +
MULT16_32_Q15(interp[1],SHR32(accum[1], 1)) +
MULT16_32_Q15(interp[2],SHR32(accum[2], 1)) +
MULT16_32_Q15(interp[3],SHR32(accum[3], 1));
+      sum = SATURATE32PSHR(sum, 15, 32767);
 #else
       cubic_coef(frac, interp);
       sum = interpolate_product_single(iptr, st->sinc_table +
st->oversample + 4 - offset - 2, N, st->oversample, interp);
 #endif
       
-      out[out_stride * out_sample++] = SATURATE32(PSHR32(sum, 14), 32767);
+      out[out_stride * out_sample++] = sum;
       last_sample += int_advance;
       samp_frac_num += frac_advance;
       if (samp_frac_num >= den_rate)
diff --git a/libspeex/resample_neon.h b/libspeex/resample_neon.h
new file mode 100644
index 0000000..ba93e41
--- /dev/null
+++ b/libspeex/resample_neon.h
@@ -0,0 +1,100 @@
+/* Copyright (C) 2007-2008 Jean-Marc Valin
+ * Copyright (C) 2008 Thorvald Natvig
+ * Copyright (C) 2011 Texas Instruments
+ *               author Jyri Sarha
+ */
+/**
+   @file resample_neon.h
+   @brief Resampler functions (NEON version)
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   - Neither the name of the Xiph.org Foundation nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
+   CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#include <arm_neon.h>
+
+#ifdef FIXED_POINT
+#define OVERRIDE_INNER_PRODUCT_SINGLE
+/* Only works when len % 4 == 0 */
+static inline int32_t inner_product_single(const int16_t *a, const int16_t *b,
unsigned int len)
+{
+    int32_t ret;
+    uint32_t remainder = len % 16;
+    len = len - remainder;
+
+    asm volatile ("	 cmp %[len], #0\n"
+		  "	 bne 1f\n"
+		  "	 vld1.16 {d16}, [%[b]]!\n"
+		  "	 vld1.16 {d20}, [%[a]]!\n"
+		  "	 subs %[remainder], %[remainder], #4\n"
+		  "	 vmull.s16 q0, d16, d20\n"
+		  "      beq 5f\n" 
+		  "	 b 4f\n"
+		  "1:"
+		  "	 vld1.16 {d16, d17, d18, d19}, [%[b]]!\n"
+		  "	 vld1.16 {d20, d21, d22, d23}, [%[a]]!\n"
+		  "	 subs %[len], %[len], #16\n"
+		  "	 vmull.s16 q0, d16, d20\n"
+		  "	 vmlal.s16 q0, d17, d21\n"
+		  "	 vmlal.s16 q0, d18, d22\n"
+		  "	 vmlal.s16 q0, d19, d23\n"
+		  "	 beq 3f\n"
+		  "2:"
+		  "	 vld1.16 {d16, d17, d18, d19}, [%[b]]!\n"
+		  "	 vld1.16 {d20, d21, d22, d23}, [%[a]]!\n"
+		  "	 subs %[len], %[len], #16\n"
+		  "	 vmlal.s16 q0, d16, d20\n"
+		  "	 vmlal.s16 q0, d17, d21\n"
+		  "	 vmlal.s16 q0, d18, d22\n"
+		  "	 vmlal.s16 q0, d19, d23\n"
+		  "	 bne 2b\n"
+		  "3:"
+		  "	 cmp %[remainder], #0\n"
+		  "	 beq 5f\n"
+		  "4:"
+		  "	 vld1.16 {d16}, [%[b]]!\n"
+		  "	 vld1.16 {d20}, [%[a]]!\n"
+		  "	 subs %[remainder], %[remainder], #4\n"
+		  "	 vmlal.s16 q0, d16, d20\n"
+		  "	 bne 4b\n"
+		  "5:"
+		  "	 vaddl.s32 q0, d0, d1\n"
+		  "	 vadd.s64 d0, d0, d1\n"
+		  "	 vqmovn.s64 d0, q0\n"
+		  "	 vqrshrn.s32 d0, q0, #15\n"
+		  "	 vmov.s16 %[ret], d0[0]\n"
+		  : [ret] "=&r" (ret), [a] "+r" (a), [b]
"+r" (b),
+		    [len] "+r" (len), [remainder] "+r" (remainder)
+		  :
+		  : "cc", "q0",
+		    "d16", "d17", "d18", "d19",
+		    "d20", "d21", "d22", "d23");
+
+    return ret;
+}
+#endif
-- 
1.7.4.1

Jyri Sarha

2011-Sep-01 12:30 UTC

head link

[Speex-dev] [PATCH 4/5] configure.ac: Add ARM NEON support

From: Jyri Sarha <jsarha at ti.com>

Use --enable-neon to force NEON optimization on. The auto detection
should also work if your CFLAGS supports NEON.
---
 configure.ac |   32 +++++++++++++++++++++++++++++++-
 1 files changed, 31 insertions(+), 1 deletions(-)

diff --git a/configure.ac b/configure.ac
index 255c0b4..08d3d5f 100644
--- a/configure.ac
+++ b/configure.ac
@@ -89,6 +89,23 @@ has_sse=no
 )
 AC_MSG_RESULT($has_sse)  
 
+AC_MSG_CHECKING(for NEON in current arch/CFLAGS)
+AC_LINK_IFELSE([
+AC_LANG_PROGRAM([[
+#include <arm_neon.h>
+int32x4_t testfunc(int16_t *a, int16_t *b) {
+      return vmull_s16(vld1_s16(a), vld1_s16(b));
+}
+]])],
+[
+has_neon=yes
+],
+[
+has_neon=no
+]
+)
+AC_MSG_RESULT($has_neon)  
+
 SAVE_CFLAGS="$CFLAGS"
 CFLAGS="$CFLAGS -fvisibility=hidden"
 AC_MSG_CHECKING(for ELF visibility)
@@ -148,6 +165,15 @@ has_sse=no
 fi
 ])
 
+AC_ARG_ENABLE(neon, [  --enable-neon           Enable NEON support], [
+if test "x$enableval" != xno; then
+has_neon=yes
+CFLAGS="$CFLAGS -O3 -march=armv7-a -mfpu=neon"
+else
+has_neon=no
+fi
+])
+
 
 FFT=smallft
 
@@ -165,6 +191,10 @@ if test "$has_sse" = yes; then
   AC_DEFINE([_USE_SSE], , [Enable SSE support])
 fi
 
+if test "$has_neon" = yes; then
+  AC_DEFINE([_USE_NEON], , [Enable NEON support])
+fi
+
 AC_ARG_ENABLE(float-api, [  --disable-float-api     Disable the floating-point
API],
 [if test "$enableval" = no; then
   AC_DEFINE([DISABLE_FLOAT_API], , [Disable all parts of the API that are using
floats])
@@ -198,7 +228,7 @@ AC_ARG_ENABLE(fixed-point-debug, [ 
--enable-fixed-point-debug  Debug fixed-poin
   AC_DEFINE([FIXED_DEBUG], , [Debug fixed-point implementation])
 fi])
 
-AC_ARG_ENABLE(resample-full-sinc-table, [  --enable-resample-full-sinc-table
Resample full SINC table],
+AC_ARG_ENABLE(resample-full-sinc-table, [  --enable-resample-full-sinc-table
Resample full SINC table (no interpolation)],
 [if test "$enableval" = yes; then
   AC_DEFINE([RESAMPLE_FULL_SINC_TABLE], , [Resample with full SINC table (no
interpolation)])
 fi])
-- 
1.7.4.1

Jyri Sarha

2011-Sep-01 12:30 UTC

head link

[Speex-dev] [PATCH 5/5] resample: Add NEON optimized inner_product_single for floating point

From: Jyri Sarha <jsarha at ti.com>

Also adds inline asm implementations of WORD2INT(x) macro for fixed
and floating point.
---
 libspeex/resample_neon.h |  101 ++++++++++++++++++++++++++++++++++++++++++++++
 1 files changed, 101 insertions(+), 0 deletions(-)

diff --git a/libspeex/resample_neon.h b/libspeex/resample_neon.h
index ba93e41..e7e981e 100644
--- a/libspeex/resample_neon.h
+++ b/libspeex/resample_neon.h
@@ -39,6 +39,30 @@
 #include <arm_neon.h>
 
 #ifdef FIXED_POINT
+#ifdef __thumb2__ 
+static inline int32_t saturate_32bit_to_16bit(int32_t a) {
+    int32_t ret;
+    asm ("ssat %[ret], #16, %[a]"
+         : [ret] "=&r" (ret)
+         : [a] "r" (a)
+         : );
+    return ret;
+}
+#else
+static inline int32_t saturate_32bit_to_16bit(int32_t a) {
+    int32_t ret;
+    asm ("vmov.s32 d0[0], %[a]\n"
+         "vqmovn.s32 d0, q0\n"
+         "vmov.s16 %[ret], d0[0]\n"
+         : [ret] "=&r" (ret)
+         : [a] "r" (a)
+         : "q0");
+    return ret;
+}
+#endif
+#undef WORD2INT
+#define WORD2INT(x) (saturate_32bit_to_16bit(x))
+
 #define OVERRIDE_INNER_PRODUCT_SINGLE
 /* Only works when len % 4 == 0 */
 static inline int32_t inner_product_single(const int16_t *a, const int16_t *b,
unsigned int len)
@@ -97,4 +121,81 @@ static inline int32_t inner_product_single(const int16_t *a,
const int16_t *b, u
 
     return ret;
 }
+#elif defined(FLOATING_POINT)
+
+static inline int32_t saturate_float_to_16bit(float a) {
+    int32_t ret;
+    asm ("vmov.f32 d0[0], %[a]\n"
+         "vcvt.s32.f32 d0, d0, #15\n"
+         "vqrshrn.s32 d0, q0, #15\n"
+         "vmov.s16 %[ret], d0[0]\n"
+         : [ret] "=&r" (ret)
+         : [a] "r" (a)
+         : "q0");
+    return ret;
+}
+#undef WORD2INT
+#define WORD2INT(x) (saturate_float_to_16bit(x))
+
+#define OVERRIDE_INNER_PRODUCT_SINGLE
+/* Only works when len % 4 == 0 */
+static inline float inner_product_single(const float *a, const float *b,
unsigned int len)
+{
+    float ret;
+    uint32_t remainder = len % 16;
+    len = len - remainder;
+
+    asm volatile ("	 cmp %[len], #0\n"
+		  "	 bne 1f\n"
+		  "	 vld1.32 {q4}, [%[b]]!\n"
+		  "	 vld1.32 {q8}, [%[a]]!\n"
+		  "	 subs %[remainder], %[remainder], #4\n"
+		  "	 vmul.f32 q0, q4, q8\n"
+		  "      bne 4f\n" 
+		  "	 b 5f\n"
+		  "1:"
+		  "	 vld1.32 {q4, q5}, [%[b]]!\n"
+		  "	 vld1.32 {q8, q9}, [%[a]]!\n"
+		  "	 vld1.32 {q6, q7}, [%[b]]!\n"
+		  "	 vld1.32 {q10, q11}, [%[a]]!\n"
+		  "	 subs %[len], %[len], #16\n"
+		  "	 vmul.f32 q0, q4, q8\n"
+		  "	 vmul.f32 q1, q5, q9\n"
+		  "	 vmul.f32 q2, q6, q10\n"
+		  "	 vmul.f32 q3, q7, q11\n"
+		  "	 beq 3f\n"
+		  "2:"
+		  "	 vld1.32 {q4, q5}, [%[b]]!\n"
+		  "	 vld1.32 {q8, q9}, [%[a]]!\n"
+		  "	 vld1.32 {q6, q7}, [%[b]]!\n"
+		  "	 vld1.32 {q10, q11}, [%[a]]!\n"
+		  "	 subs %[len], %[len], #16\n"
+		  "	 vmla.f32 q0, q4, q8\n"
+		  "	 vmla.f32 q1, q5, q9\n"
+		  "	 vmla.f32 q2, q6, q10\n"
+		  "	 vmla.f32 q3, q7, q11\n"
+		  "	 bne 2b\n"
+		  "3:"
+		  "	 vadd.f32 q4, q0, q1\n"
+		  "	 vadd.f32 q5, q2, q3\n"
+		  "	 cmp %[remainder], #0\n"
+		  "	 vadd.f32 q0, q4, q5\n"
+		  "	 beq 5f\n"
+		  "4:"
+		  "	 vld1.32 {q6}, [%[b]]!\n"
+		  "	 vld1.32 {q10}, [%[a]]!\n"
+		  "	 subs %[remainder], %[remainder], #4\n"
+		  "	 vmla.f32 q0, q6, q10\n"
+		  "	 bne 4b\n"
+		  "5:"
+		  "	 vadd.f32 d0, d0, d1\n"
+		  "	 vpadd.f32 d0, d0, d0\n"
+		  "	 vmov.f32 %[ret], d0[0]\n"
+		  : [ret] "=&r" (ret), [a] "+r" (a), [b]
"+r" (b),
+		    [len] "+l" (len), [remainder] "+l" (remainder)
+		  :
+		  : "cc", "q0", "q1", "q2",
"q3", "q4", "q5", "q6", "q7",
"q8",
+                    "q9", "q10", "q11");
+    return ret;
+}
 #endif
-- 
1.7.4.1

Arun Raghavan

2012-Nov-26 10:10 UTC

head link

[Speex-dev] [PATCH 0/5] ARM NEON optimization for samplerate converter

On Thu, 2011-09-01 at 15:30 +0300, Jyri Sarha wrote:> From: Jyri Sarha <jsarha at ti.com>
> 
> I optimized Speex resampler for NEON capable ARM CPUs. The first patch
> should speed up resampling on any platform that can spare the
> increased memory usage. It would be nice to have these merged to the
> master branch. Please let me know if there is anything I can do to 
> help the the merge. The patches have been rebased on top of master 
> branch in http://git.xiph.org/speex.git and the relevant parts have
> been tested on ARM and x86 systems.
Ping? Would be quite nice to have this upstream.

Cheers,
Arun

Reasonably Related Threads

Search for more reasonably related threads

Speex dev - Sep 2011 - [PATCH 3/5] resample: Add NEON optimized inner_product_single for fixed point

[Speex-dev] [PATCH 0/5] ARM NEON optimization for samplerate converter

[Speex-dev] [PATCH 1/5] resample: Calculate full sinc table (e.g. no sinc interpolation)

[Speex-dev] [PATCH 2/5] configure.ac: Add --enable-resample-full-sinc-table conf flag

[Speex-dev] [PATCH 3/5] resample: Add NEON optimized inner_product_single for fixed point

[Speex-dev] [PATCH 4/5] configure.ac: Add ARM NEON support

[Speex-dev] [PATCH 5/5] resample: Add NEON optimized inner_product_single for floating point

[Speex-dev] [PATCH 0/5] ARM NEON optimization for samplerate converter

Reasonably Related Threads