Author: Amaury Forgeot d'Arc <amaur...@gmail.com>
Branch:
Changeset: r75844:62a65dd49942
Date: 2015-02-13 00:08 +0100
http://bitbucket.org/pypy/pypy/changeset/62a65dd49942/
Log: Issue #1975: Port some audioop functions to cffi, this considerably
improve the performance of the "pydub" example.
pypy-no-jit is now ~30% slower than CPython.
diff --git a/lib_pypy/audioop.py b/lib_pypy/audioop.py
--- a/lib_pypy/audioop.py
+++ b/lib_pypy/audioop.py
@@ -364,19 +364,8 @@
sample_count = _sample_count(cp, size)
rv = ffi.new("unsigned char[]", len(cp) * 2)
- result = ffi.buffer(rv)
- clip = _get_clipfn(size)
-
- for i in range(sample_count):
- sample = _get_sample(cp, size, i)
-
- l_sample = clip(sample * fac1)
- r_sample = clip(sample * fac2)
-
- _put_sample(result, size, i * 2, l_sample)
- _put_sample(result, size, i * 2 + 1, r_sample)
-
- return result[:]
+ lib.tostereo(rv, cp, len(cp), size, fac1, fac2)
+ return ffi.buffer(rv)[:]
def add(cp1, cp2, size):
@@ -385,20 +374,9 @@
if len(cp1) != len(cp2):
raise error("Lengths should be the same")
- clip = _get_clipfn(size)
- sample_count = _sample_count(cp1, size)
rv = ffi.new("unsigned char[]", len(cp1))
- result = ffi.buffer(rv)
-
- for i in range(sample_count):
- sample1 = getsample(cp1, size, i)
- sample2 = getsample(cp2, size, i)
-
- sample = clip(sample1 + sample2)
-
- _put_sample(result, size, i, sample)
-
- return result[:]
+ lib.add(rv, cp1, cp2, len(cp1), size)
+ return ffi.buffer(rv)[:]
def bias(cp, size, bias):
@@ -477,11 +455,10 @@
inrate //= d
outrate //= d
- prev_i = [0] * nchannels
- cur_i = [0] * nchannels
-
if state is None:
d = -outrate
+ prev_i = ffi.new('int[]', nchannels)
+ cur_i = ffi.new('int[]', nchannels)
else:
d, samps = state
@@ -489,56 +466,37 @@
raise error("illegal state argument")
prev_i, cur_i = zip(*samps)
- prev_i, cur_i = list(prev_i), list(cur_i)
+ prev_i = ffi.new('int[]', prev_i)
+ cur_i = ffi.new('int[]', cur_i)
+ state_d = ffi.new('int[]', (d,))
q = frame_count // inrate
ceiling = (q + 1) * outrate
nbytes = ceiling * bytes_per_frame
rv = ffi.new("unsigned char[]", nbytes)
- result = ffi.buffer(rv)
-
- samples = _get_samples(cp, size)
- out_i = 0
- while True:
- while d < 0:
- if frame_count == 0:
- samps = zip(prev_i, cur_i)
- retval = result[:]
-
- # slice off extra bytes
- trim_index = (out_i * bytes_per_frame) - len(retval)
- retval = retval[:trim_index]
-
- return (retval, (d, tuple(samps)))
-
- for chan in range(nchannels):
- prev_i[chan] = cur_i[chan]
- cur_i[chan] = next(samples)
-
- cur_i[chan] = (
- (weightA * cur_i[chan] + weightB * prev_i[chan])
- // (weightA + weightB)
- )
-
- frame_count -= 1
- d += outrate
-
- while d >= 0:
- for chan in range(nchannels):
- cur_o = (
- (prev_i[chan] * d + cur_i[chan] * (outrate - d))
- // outrate
- )
- _put_sample(result, size, out_i, _overflow(cur_o, size))
- out_i += 1
- d -= inrate
+ trim_index = lib.ratecv(rv, cp, frame_count, size,
+ nchannels, inrate, outrate,
+ state_d, prev_i, cur_i,
+ weightA, weightB)
+ result = ffi.buffer(rv)[:trim_index]
+ samps = zip(prev_i, cur_i)
+ return (result, (d, tuple(samps)))
ffi = FFI()
ffi.cdef("""
typedef short PyInt16;
+int ratecv(char* rv, char* cp, size_t len, int size,
+ int nchannels, int inrate, int outrate,
+ int* state_d, int* prev_i, int* cur_i,
+ int weightA, int weightB);
+
+void tostereo(char* rv, char* cp, size_t len, int size,
+ double fac1, double fac2);
+void add(char* rv, char* cp1, char* cp2, size_t len1, int size);
+
/* 2's complement (14-bit range) */
unsigned char
st_14linear2ulaw(PyInt16 pcm_val);
@@ -829,7 +787,166 @@
#define LONGP(cp, i) ((Py_Int32 *)(cp+i))
"""
-lib = ffi.verify(_AUDIOOP_C_MODULE + """
+lib = ffi.verify(_AUDIOOP_C_MODULE + r"""
+#include <math.h>
+
+static const int maxvals[] = {0, 0x7F, 0x7FFF, 0x7FFFFF, 0x7FFFFFFF};
+/* -1 trick is needed on Windows to support -0x80000000 without a warning */
+static const int minvals[] = {0, -0x80, -0x8000, -0x800000, -0x7FFFFFFF-1};
+
+static int
+fbound(double val, double minval, double maxval)
+{
+ if (val > maxval)
+ val = maxval;
+ else if (val < minval + 1)
+ val = minval;
+ return val;
+}
+
+static int
+gcd(int a, int b)
+{
+ while (b > 0) {
+ int tmp = a % b;
+ a = b;
+ b = tmp;
+ }
+ return a;
+}
+
+int ratecv(char* rv, char* cp, size_t len, int size,
+ int nchannels, int inrate, int outrate,
+ int* state_d, int* prev_i, int* cur_i,
+ int weightA, int weightB)
+{
+ char *ncp = rv;
+ int d, chan;
+
+ /* divide inrate and outrate by their greatest common divisor */
+ d = gcd(inrate, outrate);
+ inrate /= d;
+ outrate /= d;
+ /* divide weightA and weightB by their greatest common divisor */
+ d = gcd(weightA, weightB);
+ weightA /= d;
+ weightA /= d;
+
+ d = *state_d;
+
+ for (;;) {
+ while (d < 0) {
+ if (len == 0) {
+ *state_d = d;
+ return ncp - rv;
+ }
+ for (chan = 0; chan < nchannels; chan++) {
+ prev_i[chan] = cur_i[chan];
+ if (size == 1)
+ cur_i[chan] = ((int)*CHARP(cp, 0)) << 24;
+ else if (size == 2)
+ cur_i[chan] = ((int)*SHORTP(cp, 0)) << 16;
+ else if (size == 4)
+ cur_i[chan] = (int)*LONGP(cp, 0);
+ cp += size;
+ /* implements a simple digital filter */
+ cur_i[chan] = (int)(
+ ((double)weightA * (double)cur_i[chan] +
+ (double)weightB * (double)prev_i[chan]) /
+ ((double)weightA + (double)weightB));
+ }
+ len--;
+ d += outrate;
+ }
+ while (d >= 0) {
+ for (chan = 0; chan < nchannels; chan++) {
+ int cur_o;
+ cur_o = (int)(((double)prev_i[chan] * (double)d +
+ (double)cur_i[chan] * (double)(outrate - d)) /
+ (double)outrate);
+ if (size == 1)
+ *CHARP(ncp, 0) = (signed char)(cur_o >> 24);
+ else if (size == 2)
+ *SHORTP(ncp, 0) = (short)(cur_o >> 16);
+ else if (size == 4)
+ *LONGP(ncp, 0) = (Py_Int32)(cur_o);
+ ncp += size;
+ }
+ d -= inrate;
+ }
+ }
+}
+
+void tostereo(char* rv, char* cp, size_t len, int size,
+ double fac1, double fac2)
+{
+ int val1, val2, val = 0;
+ double fval, maxval, minval;
+ char *ncp = rv;
+ int i;
+
+ maxval = (double) maxvals[size];
+ minval = (double) minvals[size];
+
+ for ( i=0; i < len; i += size ) {
+ if ( size == 1 ) val = (int)*CHARP(cp, i);
+ else if ( size == 2 ) val = (int)*SHORTP(cp, i);
+ else if ( size == 4 ) val = (int)*LONGP(cp, i);
+
+ fval = (double)val*fac1;
+ val1 = (int)floor(fbound(fval, minval, maxval));
+
+ fval = (double)val*fac2;
+ val2 = (int)floor(fbound(fval, minval, maxval));
+
+ if ( size == 1 ) *CHARP(ncp, i*2) = (signed char)val1;
+ else if ( size == 2 ) *SHORTP(ncp, i*2) = (short)val1;
+ else if ( size == 4 ) *LONGP(ncp, i*2) = (Py_Int32)val1;
+
+ if ( size == 1 ) *CHARP(ncp, i*2+1) = (signed char)val2;
+ else if ( size == 2 ) *SHORTP(ncp, i*2+2) = (short)val2;
+ else if ( size == 4 ) *LONGP(ncp, i*2+4) = (Py_Int32)val2;
+ }
+}
+
+void add(char* rv, char* cp1, char* cp2, size_t len1, int size)
+{
+ int i;
+ int val1 = 0, val2 = 0, minval, maxval, newval;
+ char* ncp = rv;
+
+ maxval = maxvals[size];
+ minval = minvals[size];
+
+ for ( i=0; i < len1; i += size ) {
+ if ( size == 1 ) val1 = (int)*CHARP(cp1, i);
+ else if ( size == 2 ) val1 = (int)*SHORTP(cp1, i);
+ else if ( size == 4 ) val1 = (int)*LONGP(cp1, i);
+
+ if ( size == 1 ) val2 = (int)*CHARP(cp2, i);
+ else if ( size == 2 ) val2 = (int)*SHORTP(cp2, i);
+ else if ( size == 4 ) val2 = (int)*LONGP(cp2, i);
+
+ if (size < 4) {
+ newval = val1 + val2;
+ /* truncate in case of overflow */
+ if (newval > maxval)
+ newval = maxval;
+ else if (newval < minval)
+ newval = minval;
+ }
+ else {
+ double fval = (double)val1 + (double)val2;
+ /* truncate in case of overflow */
+ newval = (int)floor(fbound(fval, minval, maxval));
+ }
+
+ if ( size == 1 ) *CHARP(ncp, i) = (signed char)newval;
+ else if ( size == 2 ) *SHORTP(ncp, i) = (short)newval;
+ else if ( size == 4 ) *LONGP(ncp, i) = (Py_Int32)newval;
+ }
+}
+
void lin2adcpm(unsigned char* ncp, unsigned char* cp, size_t len,
size_t size, int* state)
{
_______________________________________________
pypy-commit mailing list
pypy-commit@python.org
https://mail.python.org/mailman/listinfo/pypy-commit
