ASCEE/lasp
ASCEE/lasp
5
1
Fork 0
Code Issues 5 Pull Requests Actions Releases 14 Activity

Bugfix, added overflow checks for debug mode. Tested AvPowerSpectra. Now works without errors, only addTimeData cannot be called twice. Improved fft implementation. Removed double wrapper for fftpack. Renamed fft.pyx to wrappers.pyx. Improved CMakeLists. Adde proper function entrance and exit tracers. Made tracers indent to function level. Added Bartlett window function

Browse Source
This commit is contained in:
Anne de Jong 2018-02-09 11:56:49 +01:00 committed by asceenl
parent 2d203647ed
commit 89e9622b1f
30 changed files with 519 additions and 417 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
.gitignore vendored
View File

@ -10,7 +10,7 @@ build
__pycache__
cython_debug
beamforming/config.pxi
beamforming/fft.c
beamforming/wrappers.c
*.so
test/test_bf
test/test_fft

10
CMakeLists.txt
View File

@ -2,8 +2,8 @@ cmake_minimum_required (VERSION 3.0)
project(beamforming)
# Whether we want to use blas yes or no
set(ASCEE_USE_BLAS 1)
add_definitions(-DASCEE_USE_BLAS=1)
set(ASCEE_USE_BLAS TRUE)
# set(ASCEE_USE_BLAS FALSE)
set(ASCEE_FLOAT double)
# set(ASCEE_FLOAT float)
@ -68,10 +68,12 @@ endif(ASCEE_DEBUG)
# The last argument here takes care of calling SIGABRT when an integer overflow
# occures.
############################## General compilation flags (independent of debug mode, windows or linux)
set(CYTHON_EXTRA_C_FLAGS "-Wno-sign-compare -Wno-cpp -Wno-implicit-fallthrough\
-Wno-incompatible-pointer-types -Wno-strict-aliasing")
# General make flags
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fPIC -std=c11 -Wall -Wextra \
-Wimplicit-function-declaration -ftrapv -Wno-type-limits")
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fPIC -std=c11 -Wall -Wextra -Wno-type-limits \
-Werror=implicit-function-declaration -Werror=incompatible-pointer-types")
# Debug make flags
set(CMAKE_C_FLAGS_DEBUG "-g" )

7
beamforming/CMakeLists.txt
View File

@ -13,7 +13,6 @@ include_directories(
)
set_source_files_properties(fft.c PROPERTIES COMPILE_FLAGS
"-Wno-cpp -Wno-sign-compare -Wno-incompatible-pointer-types -Wno-strict-aliasing -Wno-implicit-fallthrough")
cython_add_module(fft fft.pyx)
target_link_libraries(fft beamforming_lib )
set_source_files_properties(wrappers.c PROPERTIES COMPILE_FLAGS "${CMAKE_C_FLAGS} ${CYTHON_EXTRA_C_FLAGS}")
cython_add_module(wrappers wrappers.pyx)
target_link_libraries(wrappers beamforming_lib )

1
beamforming/__init__.py
View File

@ -0,0 +1 @@
from .wrappers import *

84
beamforming/c/aps.c
View File

@ -49,26 +49,31 @@ AvPowerSpectra* AvPowerSpectra_alloc(const us nfft,
/* Check nfft */
if(nfft % 2 != 0 || nfft > ASCEE_MAX_NFFT) {
WARN("nfft should be even");
feTRACE(15);
return NULL;
}
/* Check overlap percentage */
if(overlap_percentage >= 100) {
WARN("Overlap percentage >= 100!");
feTRACE(15);
return NULL;
}
if(overlap_percentage < 0) {
WARN("Overlap percentage should be positive!");
feTRACE(15);
return NULL;
}
/* Compute and check overlap offset */
us oo = nfft - (us) (overlap_percentage*nfft/100);
us oo = (us) (((d) nfft)-overlap_percentage*((d) nfft)/100);
iVARTRACE(15,oo);
if(oo == 0) {oo++;}
PowerSpectra* ps = PowerSpectra_alloc(nfft,nchannels,wt);
if(!ps) {
WARN(ALLOCFAILED "ps");
feTRACE(15);
return NULL;
}
@ -76,8 +81,10 @@ AvPowerSpectra* AvPowerSpectra_alloc(const us nfft,
if(!aps) {
WARN("Allocation of AvPowerSpectra memory failed");
PowerSpectra_free(ps);
feTRACE(15);
return NULL;
}
aps->nchannels = nchannels;
aps->nfft = nfft;
aps->ps = ps;
@ -91,7 +98,7 @@ AvPowerSpectra* AvPowerSpectra_alloc(const us nfft,
aps->ps_single = cmat_alloc(nfft/2+1,nchannels*nchannels);
cmat_set(&aps->ps_storage,0);
feTRACE(15);
return aps;
}
@ -113,26 +120,30 @@ static void AvPowerSpectra_addBlock(AvPowerSpectra* aps,
dbgassert(aps && block,NULLPTRDEREF);
dbgassert(block->n_rows == aps->nfft,"Invalid block n_rows");
dbgassert(block->n_cols == aps->nchannels,"Invalid block n_cols");
const us nchannels = aps->nchannels;
const us nfft = aps->nfft;
iVARTRACE(15,nfft);
cmat ps_single = aps->ps_single;
cmat ps_storage = aps->ps_storage;
cmat* ps_single = &aps->ps_single;
cmat* ps_storage = &aps->ps_storage;
c naverages = (++aps->naverages);
cVARTRACE(15,naverages);
/* Scale previous result */
cmat_scale(&ps_storage,
cmat_scale(ps_storage,
(naverages-1)/naverages);
uVARTRACE(15,(us) aps->ps);
PowerSpectra_compute(aps->ps,
block,
&ps_single);
ps_single);
/* Add new result, scaled properly */
cmat_add_cmat(&ps_storage,
&ps_single,1/naverages);
cmat_add_cmat(ps_storage,
ps_single,1/naverages);
feTRACE(15);
}
@ -153,60 +164,71 @@ cmat* AvPowerSpectra_addTimeData(AvPowerSpectra* aps,
const us oo = aps->oo;
us* os = &aps->os;
iVARTRACE(15,*os);
us os_timedata = 0;
dmat buffer = aps->buffer;
/* Retrieve the buffer and use it to make the first time block. */
if(*os < oo) {
TRACE(15,"Using saved data from previous run");
dbgassert(false,"not tested")
dmat tmp = dmat_alloc(nfft,nchannels);
dbgassert(0 <= *os,"BUG");
dbgassert(*os <= nfft,"BUG");
copy_dmat_rows(&tmp,
&buffer,
*os, /* Startrow_from */
0, /* Startrow to */
nfft - *os /* nrows */
);
/* copy_dmat_rows(&tmp, */
/* &buffer, */
/* *os, /\* Startrow_from *\/ */
/* 0, /\* Startrow to *\/ */
/* nfft - *os /\* nrows *\/ */
/* ); */
copy_dmat_rows(&tmp,
timedata,
0,
nfft - *os,
*os
);
/* copy_dmat_rows(&tmp, */
/* timedata, */
/* 0, */
/* nfft - *os, */
/* *os */
/* ); */
AvPowerSpectra_addBlock(aps,&tmp);
os_timedata = oo + *os - nfft;
dbgassert(os_timedata < nfft,"BUG");
dmat_free(&tmp);
}
/* Run until we cannot go any further */
while (os_timedata + nfft <= timedata->n_rows) {
dmat tmp = dmat_submat(timedata,
os_timedata,
0,nfft,nchannels);
while ((os_timedata + nfft) <= timedata->n_rows) {
dmat tmp = dmat_submat(timedata,
os_timedata, /* Startrow */
0, /* Start column */
nfft, /* Number of rows */
nchannels); /* Number of columns */
/* Process the block of time data */
AvPowerSpectra_addBlock(aps,&tmp);
iVARTRACE(15,os_timedata);
os_timedata += oo;
dmat_free(&tmp);
iVARTRACE(15,os_timedata);
}
/* We copy the last piece of samples from the timedata to the
* buffer */
copy_dmat_rows(&buffer,
timedata,
timedata->n_rows-nfft,
0,
nfft);
timedata->n_rows-nfft, /* startrow_from */
0, /* startrow_to */
nfft); /* Number of rows */
*os = os_timedata+nfft-timedata->n_rows;
dbgassert(*os < nfft,"BUG");
dbgassert(*os <= nfft,"BUG");
feTRACE(15);
return &aps->ps_storage;

13
beamforming/c/ascee_alg.c
View File

@ -5,7 +5,6 @@
// Description:
// (Linear) algebra routine implementations
//////////////////////////////////////////////////////////////////////
#include "ascee_alg.h"
@ -13,11 +12,13 @@
void cmv_dot(const cmat* A,const vc* restrict x,vc* restrict b){
assert(A->n_rows == b->size);
assert(A->n_cols == x->size);
dbgassert(A->n_rows == b->size,SIZEINEQUAL);
dbgassert(A->n_cols == x->size,SIZEINEQUAL);
#if ASCEE_USE_BLAS == 1
dbgassert(A->_data,"Matrix-vector product only works for allocated data matrices");
/* typedef enum CBLAS_ORDER {CblasRowMajor=101, CblasColMajor=102} CBLAS_ORDER; */
/* typedef enum CBLAS_TRANSPOSE {CblasNoTrans=111, CblasTrans=112, CblasConjTrans=113, CblasConjNoTrans=114} CBLAS_TRANSPOSE; */
/*
@ -41,12 +42,12 @@ void cmv_dot(const cmat* A,const vc* restrict x,vc* restrict b){
A->n_rows,
A->n_cols,
(d*) &alpha, /* alpha */
(d*) A->data, /* A */
(d*) A->_data, /* A */
A->n_rows, /* lda */
(d*) x->data, /* */
(d*) x->ptr, /* */
1,
(d*) &beta, /* beta */
(d*) b->data,
(d*) b->ptr,
1);

40
beamforming/c/ascee_alg.h
View File

@ -86,17 +86,18 @@ static inline void dmat_add_dmat(dmat* x,dmat* y,d fac) {
* @param[in] fac
*/
static inline void cmat_add_cmat(cmat* x,cmat* y,c fac) {
dbgassert(x && y,NULLPTRDEREF);
dbgassert(x->n_cols == y->n_cols,SIZEINEQUAL);
dbgassert(x->n_rows == y->n_rows,SIZEINEQUAL);
if(likely(x->data && y->data)) {
c_add_to(x->data,y->data,fac,x->n_cols*x->n_rows);
}
else {
// dbgassert(x && y,NULLPTRDEREF);
// dbgassert(x->n_cols == y->n_cols,SIZEINEQUAL);
// dbgassert(x->n_rows == y->n_rows,SIZEINEQUAL);
// if(likely(x->data && y->data)) {
// TRACE(15,"Scale whole");
// c_add_to(x->data,y->data,fac,x->n_cols*x->n_rows);
// }
// else {
for(us col=0;col<y->n_cols;col++) {
TRACE(15,"Scale columns");
c_add_to(x->col_ptrs[col],y->col_ptrs[col],fac,x->n_rows);
}
}
}
/**
@ -107,11 +108,30 @@ static inline void cmat_add_cmat(cmat* x,cmat* y,c fac) {
* @param[in] a
* @param[in] b
*/
static inline void vc_elem_prod(vc* result,vc* a,vc* b) {
static inline void vd_elem_prod(vd* result,const vd* a,const vd* b) {
dbgassert(result && a && b,NULLPTRDEREF);
dbgassert(result->size==a->size,SIZEINEQUAL);
dbgassert(b->size==a->size,SIZEINEQUAL);
c_elem_prod_c(result->ptr,a->ptr,b->ptr,a->size);
d_elem_prod_d(result->ptr,a->ptr,b->ptr,a->size);
}
/**
* Compute the element-wise (Hadamard) product of a and b, and store
* in result
*
* @param[out] result
* @param[in] a
* @param[in] b
*/
static inline void vc_hadamard(vc* result,const vc* a,const vc* b) {
fsTRACE(15);
dbgassert(result && a && b,NULLPTRDEREF);
dbgassert(result->size==a->size,SIZEINEQUAL);
dbgassert(b->size==a->size,SIZEINEQUAL);
c_hadamard(result->ptr,a->ptr,b->ptr,a->size);
check_overflow_vx(*result);
check_overflow_vx(*a);
check_overflow_vx(*b);
feTRACE(15);
}
/**
* Compute the matrix vector product for complex-valued types: b = A*x.

2
beamforming/c/ascee_assert.h
View File

@ -34,6 +34,8 @@ void DBG_AssertFailedExtImplementation(const char* file,
#define dbgassert(assertion, assert_string)
#define assertvalidptr(ptr) dbgassert(ptr,NULLPTRDEREF)
#endif // ASCEE_DEBUG
#endif // ASCEE_ASSERT_H

24
beamforming/c/ascee_math.c
View File

@ -12,13 +12,14 @@
#include "ascee_math.h"
#include "ascee_tracer.h"
#include <math.h>
#ifdef ASCEE_DEBUG
void print_dmat(const dmat* m) {
feTRACE(50);
size_t row,col;
for(row=0;row<m->n_rows;row++){
indent_trace();
for(col=0;col<m->n_cols;col++){
d val = *getdmatval(m,row,col);
printf("%c%2.2e ", val<0?'-':' ' ,d_abs(val));
@ -27,10 +28,13 @@ void print_dmat(const dmat* m) {
printf("\n");
}
fsTRACE(50);
}
void print_cmat(const cmat* m) {
feTRACE(50);
size_t row,col;
for(row=0;row<m->n_rows;row++){
indent_trace();
for(col=0;col<m->n_cols;col++){
c val = *getcmatval(m,row,col);
@ -44,32 +48,36 @@ void print_cmat(const cmat* m) {
printf("\n");
}
feTRACE(50);
}
void print_vc(const vc* m) {
TRACE(20,"print_vc");
fsTRACE(50);
size_t row;
for(row=0;row<m->size;row++){
c val = *getvcval(m,row);
d rval = creal(m->data[row]);
d ival = cimag(m->data[row]);
d rval = creal(val);
d ival = cimag(val);
indent_trace();
printf("%c%2.2e%c%2.2ei ",rval< 0 ?'-': ' ', d_abs(rval),ival<0 ? '-' : '+',d_abs(ival) ) ;
printf("\n");
}
feTRACE(50);
}
void print_vd(const vd* m) {
TRACE(20,"print_vd");
fsTRACE(50);
size_t row;
iVARTRACE(20,m->size);
for(row=0;row<m->size;row++){
d rval = m->data[row];
d rval = m->ptr[row];
indent_trace();
printf("%c%2.2e ",rval< 0 ? '\r': ' ',rval);
printf("\n");
}
feTRACE(50);
}
#endif

150
beamforming/c/ascee_math.h
View File

@ -18,7 +18,7 @@ typedef struct {
us size;
d* ptr; /**< This pointer points to the data
of this vector */
d* data; /**< Pointer set if data storage is
d* _data; /**< Pointer set if data storage is
intern. If this is set to zero, the
vector is a sub-vector. */
} vd;
@ -27,7 +27,7 @@ typedef struct {
us size;
c* ptr; /**< This pointer points to the data
of this vector */
c* data; /**< Pointer set if data storage is
c* _data; /**< Pointer set if data storage is
intern. If this is set to zero, the
vector is a sub-vector. */
} vc;
@ -36,16 +36,39 @@ typedef struct {
us n_rows;
us n_cols;
d** col_ptrs;
d* data;
d* _data;
} dmat;
/// Dense matrix of complex floating point values
typedef struct {
us n_rows;
us n_cols;
c** col_ptrs;
c* data;
c* _data;
} cmat;
#ifdef ASCEE_DEBUG
#define OVERFLOW_MAGIC_NUMBER (-10e-45)
#define check_overflow_vx(vx) \
TRACE(15,"Checking overflow " #vx); \
if((vx)._data) { \
dbgassert((vx)._data[(vx).size] == OVERFLOW_MAGIC_NUMBER, \
"Buffer overflow detected on" #vx ); \
}
#define check_overflow_xmat(xmat) \
TRACE(15,"Checking overflow " #xmat); \
if((xmat)._data) { \
for(us _overflow=0;_overflow<(xmat).n_cols;_overflow++) \
dbgassert((xmat)._data[(xmat).n_rows*(xmat).n_cols+ \
_overflow] == OVERFLOW_MAGIC_NUMBER, \
"Buffer overflow detected on" #xmat ); \
}
#else
#define check_overflow_vx(vx)
#define check_overflow_xmat(xmat)
#endif
/**
* Sets all values in a vector to the value
*
@ -74,15 +97,9 @@ static inline void vc_set(vc* vec,const c value){
*/
static inline void dmat_set(dmat* mat,const d value){
dbgassert(mat,NULLPTRDEREF);
us size = mat->n_cols*mat->n_rows;
if(likely(mat->data)){
d_set(mat->data,value,size);
}
else {
for(us col=0;col<mat->n_cols;col++) {
d_set(mat->col_ptrs[col],value,mat->n_rows);
}
}
}
@ -94,15 +111,9 @@ static inline void dmat_set(dmat* mat,const d value){
*/
static inline void cmat_set(cmat* mat,const c value){
dbgassert(mat,NULLPTRDEREF);
us size = mat->n_cols*mat->n_rows;
if(likely(mat->data)){
c_set(mat->data,value,size);
}
else {
for(us col=0;col<mat->n_cols;col++) {
c_set(mat->col_ptrs[col],value,mat->n_rows);
}
}
}
/**
@ -114,9 +125,15 @@ static inline void cmat_set(cmat* mat,const c value){
*/
static inline vd vd_alloc(us size) {
vd result = { size, NULL,NULL};
result.data = (d*) a_malloc(size*sizeof(d));
result.ptr = result.data;
dbgassert(result.data,ALLOCFAILED);
#ifdef ASCEE_DEBUG
result._data = (d*) a_malloc((size+1)*sizeof(d));
result._data[size] = OVERFLOW_MAGIC_NUMBER;
#else
result._data = (d*) a_malloc(size*sizeof(d));
#endif // ASCEE_DEBUG
result.ptr = result._data;
dbgassert(result._data,ALLOCFAILED);
#ifdef ASCEE_DEBUG
vd_set(&result,NAN);
#endif // ASCEE_DEBUG
@ -131,8 +148,14 @@ static inline vd vd_alloc(us size) {
*/
static inline vc vc_alloc(us size) {
vc result = { size, NULL, NULL};
result.data = (c*) a_malloc(size*sizeof(c));
result.ptr = result.data;
#ifdef ASCEE_DEBUG
result._data = (c*) a_malloc((size+1)*sizeof(c));
result._data[size] = OVERFLOW_MAGIC_NUMBER;
#else
result._data = (c*) a_malloc(size*sizeof(c));
#endif // ASCEE_DEBUG
dbgassert(result._data,ALLOCFAILED);
result.ptr = result._data;
#ifdef ASCEE_DEBUG
vc_set(&result,NAN+I*NAN);
#endif // ASCEE_DEBUG
@ -153,16 +176,21 @@ static inline dmat dmat_alloc(us n_rows,
/**
* Here storage is allocated for both the data, as well as the
* column pointers. The column pointer data is stored at the end
* of the allocated block.
* column pointers. In debug mode, extra memory is allocated to
* check for possible buffer overflows.
*/
result.data = (d*) a_malloc(n_rows*n_cols*sizeof(d)
+sizeof(d*)*n_cols);
#ifdef ASCEE_DEBUG
result._data = (d*) a_malloc((n_rows*n_cols+n_cols)*sizeof(d));
for(us i=0;i<n_cols;i++)
result._data[n_rows*n_cols+i] = OVERFLOW_MAGIC_NUMBER;
#else
result._data = (d*) a_malloc((n_rows*n_cols+1)*sizeof(d));
#endif // ASCEE_DEBUG
dbgassert(result.data,ALLOCFAILED);
result.col_ptrs = (d**) &result.data[n_rows*n_cols];
dbgassert(result._data,ALLOCFAILED);
result.col_ptrs = a_malloc(n_cols*sizeof(d*));
for(us col=0;col<n_cols;col++) {
result.col_ptrs[col] = &result.data[n_rows*col];
result.col_ptrs[col] = &result._data[n_rows*col];
}
#ifdef ASCEE_DEBUG
dmat_set(&result,NAN);
@ -186,16 +214,19 @@ static inline cmat cmat_alloc(const us n_rows,
cmat result = { n_rows, n_cols, NULL, NULL};
/**
* Here storage is allocated for both the data, as well as the
* column pointers. The column pointer data is stored at the end
* of the allocated block.
* column pointers. In debug mode, extra memory is allocated to
* check for possible buffer overflows.
*/
result.data = (c*) a_malloc(n_rows*n_cols*sizeof(c)
+sizeof(c*)*n_cols);
dbgassert(result.data,ALLOCFAILED);
result.col_ptrs = (c**) &result.data[n_rows*n_cols];
#ifdef ASCEE_DEBUG
result._data = (c*) a_malloc((n_rows*n_cols+n_cols)*sizeof(c));
for(us i=0;i<n_cols;i++)
result._data[n_rows*n_cols+i] = OVERFLOW_MAGIC_NUMBER;
#else
result._data = (c*) a_malloc((n_rows*n_cols+1)*sizeof(c));
#endif // ASCEE_DEBUG
result.col_ptrs = a_malloc(n_cols*sizeof(c*));
for(us col=0;col<n_cols;col++) {
result.col_ptrs[col] = &result.data[n_rows*col];
result.col_ptrs[col] = &result._data[n_rows*col];
}
#ifdef ASCEE_DEBUG
@ -222,6 +253,7 @@ static inline dmat dmat_foreign(const us n_rows,
dbgassert(data,NULLPTRDEREF);
dmat result = {n_rows,n_cols,NULL,NULL};
d** colptrs = malloc(sizeof(d*)*n_cols);
dbgassert(colptrs,ALLOCFAILED);
result.col_ptrs = colptrs;
for(us i=0;i<n_cols;i++) {
@ -263,7 +295,7 @@ static inline cmat cmat_foreign(const us n_rows,
*/
static inline void vd_free(vd* f) {
dbgassert(f,NULLPTRDEREF);
if(likely(f->data)) a_free(f->data);
if(likely(f->_data)) a_free(f->_data);
}
/**
* Free's data of a vector. Is safe to run on sub-vecs as well, to
@ -273,7 +305,7 @@ static inline void vd_free(vd* f) {
*/
static inline void vc_free(vc* f) {
dbgassert(f,NULLPTRDEREF);
if(likely(f->data)) a_free(f->data);
if(likely(f->_data)) a_free(f->_data);
}
/**
* Free's data of dmat. Safe to run on sub-matrices as well.
@ -281,14 +313,9 @@ static inline void vc_free(vc* f) {
* @param m Matrix to free
*/
static inline void dmat_free(dmat* m) {
if(likely(m->data)) {
a_free(m->data);
}
else {
// Only column pointers allocated. This was a submat
if(m->_data) a_free(m->_data);
dbgassert(m->col_ptrs,NULLPTRDEREF);
a_free(m->col_ptrs);
}
}
/**
* Free's data of dmat. Safe to run on sub-matrices as well.
@ -296,19 +323,14 @@ static inline void dmat_free(dmat* m) {
* @param m Matrix to free
*/
static inline void cmat_free(cmat* m) {
if(likely(m->data)) {
a_free(m->data);
}
else {
// Only column pointers allocated. This was a submat
if(m->_data) a_free(m->_data);
dbgassert(m->col_ptrs,NULLPTRDEREF);
a_free(m->col_ptrs);
}
}
#define setvecval(vec,index,val) \
dbgassert((((us) index) <= (vec)->size),OUTOFBOUNDSVEC); \
(vec)->data[index] = val;
(vec)->ptr[index] = val;
#define setmatval(mat,row,col,val) \
@ -418,7 +440,8 @@ static inline dmat dmat_submat(const dmat* parent,
startcol+col);
}
dmat result = { n_rows,n_cols,col_ptrs,NULL};
dmat result = { n_rows,n_cols,NULL,NULL};
result.col_ptrs = col_ptrs;
return result;
}
/**
@ -437,7 +460,7 @@ static inline cmat cmat_submat(cmat* parent,
const us startcol,
const us n_rows,
const us n_cols) {
dbgassert(false,"untested");
dbgassert(parent,NULLPTRDEREF);
dbgassert(n_rows+startrow <= parent->n_rows,OUTOFBOUNDSMATR);
dbgassert(n_cols+startcol <= parent->n_cols,OUTOFBOUNDSMATC);
@ -515,7 +538,10 @@ static inline void cmat_copy(cmat* to,const cmat* from) {
* @return vector with reference to column
*/
static inline vd dmat_column(dmat* x,us col) {
vd res = { x->n_rows, getdmatval(x,0,col),NULL};
vd res;
res.size = x->n_rows;
res.ptr = getdmatval(x,0,col);
res._data = NULL;
return res;
}
@ -528,7 +554,10 @@ static inline vd dmat_column(dmat* x,us col) {
* @return vector with reference to column
*/
static inline vc cmat_column(cmat* x,us col) {
vc res = { x->n_rows, getcmatval(x,0,col),NULL};
vc res;
res.size = x->n_rows;
res.ptr = getcmatval(x,0,col);
res._data = NULL;
return res;
}
@ -538,10 +567,12 @@ static inline vc cmat_column(cmat* x,us col) {
* @param a
* @param b
*/
static inline void vc_conj_vc(vc* a,const vc* b) {
static inline void vc_conj(vc* a,const vc* b) {
fsTRACE(15);
dbgassert(a && b,NULLPTRDEREF);
dbgassert(a->size == b->size,SIZEINEQUAL);
c_conj_c(a->ptr,b->ptr,a->size);
carray_conj(a->ptr,b->ptr,a->size);
feTRACE(15);
}
/**
@ -551,14 +582,9 @@ static inline void vc_conj_vc(vc* a,const vc* b) {
*/
static inline void cmat_conj(cmat* x) {
dbgassert(x,NULLPTRDEREF);
if(likely(x->data)) {
c_conj_inplace(x->data,x->n_cols*x->n_rows);
}
else {
for(us col=0;col<x->n_cols;col++) {
c_conj_inplace(x->col_ptrs[col],x->n_rows);
}
}
}

27
beamforming/c/ascee_math_raw.c
View File

@ -15,7 +15,7 @@ void d_elem_prod_d(d res[],
const d arr2[],
const us size) {
#if ASCEE_USE_BLAS
#if ASCEE_USE_BLAS == 1
#if ASCEE_DEBUG
@ -77,15 +77,16 @@ void d_elem_prod_d(d res[],
#endif
}
void c_elem_prod_c(c res[],
void c_hadamard(c res[],
const c arr1[],
const c arr2[],
const us size) {
TRACE(15,"c_elem_prod_c");
fsTRACE(15);
uVARTRACE(15,size);
dbgassert(arr1 && arr2 && res,NULLPTRDEREF);
#if ASCEE_USE_BLAS
#if ASCEE_USE_BLAS == 1
#if ASCEE_DEBUG
@ -105,16 +106,13 @@ void c_elem_prod_c(c res[],
#define elem_prod_fun cblas_cgbmv
#endif
/* These parameters do not matter for this specific case */
const CBLAS_ORDER mat_order= CblasColMajor;
const CBLAS_TRANSPOSE tr = CblasNoTrans;
c alpha = 1.0;
c beta = 0.0;
const c alpha = 1.0;
const c beta = 0.0;
TRACE(15,"Calling " annestr(elem_prod_fun));
elem_prod_fun(mat_order,
tr,
uVARTRACE(15,size);
elem_prod_fun(CblasColMajor,
CblasNoTrans,
(blasint) size, /* M: Number of rows */
(blasint) size, /* B: Number of columns */
0, /* KL: Number of sub-diagonals */
@ -126,17 +124,18 @@ void c_elem_prod_c(c res[],
1, /* incX = 1 */
(d*) &beta,
(d*) res, /* The Y matrix to write to */
1); /* incY */
1); /* incY (increment in res) */
#undef elem_prod_fun
#else /* No blas routines, routine is very simple, but here we
* go! */
DBGWARN("Performing slow non-blas vector-vector multiplication");
DBGWARN("Performing slower non-blas vector-vector multiplication");
for(us i=0;i<size;i++) {
res[i] = arr1[i]*arr2[i];
}
#endif
feTRACE(15);
}

24
beamforming/c/ascee_math_raw.h
View File

@ -15,6 +15,9 @@
#if ASCEE_USE_BLAS == 1
#include <cblas.h>
#elif ASCEE_USE_BLAS == 0
#else
#error "ASCEE_USE_BLAS should be set to either 0 or 1"
#endif
#ifdef ASCEE_DOUBLE_PRECISION
@ -249,6 +252,7 @@ static inline void d_add_to(d x[],const d y[],d fac,us size){
* @param size Size of the arrays
*/
static inline void c_add_to(c x[],const c y[],c fac,us size){
fsTRACE(15);
#if ASCEE_USE_BLAS == 1
#if ASCEE_DOUBLE_PRECISION
cblas_zaxpy(size,(d*) &fac,(d*) y,1,(d*) x,1);
@ -260,6 +264,7 @@ static inline void c_add_to(c x[],const c y[],c fac,us size){
for(i=0;i<size;i++)
x[i]+=fac*y[i];
#endif
feTRACE(15);
}
/**
@ -298,12 +303,11 @@ static inline void c_scale(c a[],const c scale_fac,us size){
// part. Fortunately the (c) type stores the two values in this
// order. To be portable and absolutely sure anything goes well,
// we convert it explicitly here.
d scale_fac_d [] = {creal(scale_fac),cimag(scale_fac)};
#if ASCEE_DOUBLE_PRECISION
cblas_zscal(size,scale_fac_d,(d*) a,1);
cblas_zscal(size,(d*) &scale_fac,(d*) a,1);
#else
cblas_cscal(size,scale_fac_d,(d*) a,1);
cblas_cscal(size,(d*) &scale_fac,(d*) a,1);
#endif
#else
us i;
@ -411,7 +415,7 @@ void d_elem_prod_d(d res[],
* @param vec2 Array 2
* @param size: Size of the arrays
*/
void c_elem_prod_c(c res[],
void c_hadamard(c res[],
const c arr1[],
const c arr2[],
const us size);
@ -424,15 +428,16 @@ void c_elem_prod_c(c res[],
* @param in Input vector
* @param size Size of the vector
*/
static inline void c_conj_c(c res[],const c in[],const us size) {
static inline void carray_conj(c res[],const c in[],const us size) {
// First set the result vector to zero
fsTRACE(15);
c_set(res,0,size);
#ifdef ASCEE_USE_BLAS
#if ASCEE_USE_BLAS == 1
#if ASCEE_DOUBLE_PRECISION
// Cast as a float, scale all odd elements with minus one to find
// the complex conjugate.
cblas_daxpy(size ,1,(d*) in,2,(d*) res,2);
cblas_daxpy(size,-1,&((d*) in)[1],2,&((d*) res)[1],2);
cblas_daxpy(size ,1.0,(d*) in,2,(d*) res,2);
cblas_daxpy(size,-1.0,&((d*) in)[1],2,&((d*) res)[1],2);
#else
cblas_faxpy(size ,1,(d*) in,2,(d*) res,2);
cblas_faxpy(size,-1,&((d*) in)[1],2,&((d*) res)[1],2);
@ -442,6 +447,7 @@ static inline void c_conj_c(c res[],const c in[],const us size) {
res[i] = c_conj(in[i]);
}
#endif // ASCEE_USE_BLAS
feTRACE(15);
}
/**
* In place complex conjugation
@ -450,7 +456,7 @@ static inline void c_conj_c(c res[],const c in[],const us size) {
* @param size Size of the vector
*/
static inline void c_conj_inplace(c res[],us size) {
#ifdef ASCEE_USE_BLAS
#if ASCEE_USE_BLAS
#if ASCEE_DOUBLE_PRECISION
// Cast as a float, scale all odd elements with minus one to find
// the complex conjugate.

23
beamforming/c/ascee_tracer.c
View File

@ -12,8 +12,10 @@
#ifdef _REENTRANT
#include <stdatomic.h>
static _Thread_local us ASCEE_FN_LEVEL = 0;
static _Atomic(int) TRACERNAME = ATOMIC_VAR_INIT(DEFAULTTRACERLEVEL);
_Atomic(int) TRACERNAME = ATOMIC_VAR_INIT(DEFAULTTRACERLEVEL);
void setTracerLevel(int level) {
atomic_store(&TRACERNAME,level);
@ -21,35 +23,52 @@ void setTracerLevel(int level) {
int getTracerLevel() {
return atomic_load(&TRACERNAME);
}
#else
int TRACERNAME;
static us ASCEE_FN_LEVEL = 0;
/* setTracerLevel and getTracerLevel are defined as macros in
* tracer.h */
#endif
void indent_trace() {
for(us i=0;i<ASCEE_FN_LEVEL;i++) {
printf("--");
}
printf("* ");
}
void trace_impl(const char* file,int pos, const char * string){
indent_trace();
printf(annestr(TRACERNAME) ":%s:%i: %s\n",file,pos,string);
}
void fstrace_impl(const char* file,int pos,const char* fn){
ASCEE_FN_LEVEL++;
indent_trace();
printf(annestr(TRACERNAME) ":%s:%i: start function: %s()\n",file,pos,fn);
}
void fetrace_impl(const char* file,int pos,const char* fn){
indent_trace();
printf(annestr(TRACERNAME) ":%s:%i: end function: %s()\n",file,pos,fn);
ASCEE_FN_LEVEL--;
}
void ivartrace_impl(const char* pos,int line,const char* varname, int var){
indent_trace();
printf(annestr(TRACERNAME) ":%s:%i: %s = %i\n",pos,line,varname,var);
}
void uvartrace_impl(const char* pos,int line,const char* varname,size_t var){
indent_trace();
printf(annestr(TRACERNAME) ":%s:%i: %s = %zu\n",pos,line,varname,var);
}
void dvartrace_impl(const char* pos,int line,const char* varname, d var){
indent_trace();
printf(annestr(TRACERNAME) ":%s:%i: %s = %0.5e\n",pos,line,varname,var);
}
void cvartrace_impl(const char* pos,int line,const char* varname, c var){
indent_trace();
printf(annestr(TRACERNAME) ":%s:%i: %s = %0.5e+%0.5ei\n",pos,line,varname,creal(var),cimag(var));
}
#endif

14
beamforming/c/ascee_tracer.h
View File

@ -8,6 +8,17 @@
#pragma once
#ifndef ASCEE_TRACER_H
#define ASCEE_TRACER_H
#include <string.h>
#include <stdio.h>
static inline void clearScreen() {
printf("\033c\n");
}
/**
* Indent the rule for tracing visibility.
*/
void indent_trace();
// Some console colors
#define RESET "\033[0m"
@ -28,9 +39,6 @@
#define BOLDCYAN "\033[1m\033[36m" /* Bold Cyan */
#define BOLDWHITE "\033[1m\033[37m" /* Bold White */
#include <assert.h>
#include <string.h>
// Not so interesting part
#define rawstr(x) #x
#define namestr(x) rawstr(x)

56
beamforming/c/fft.c
View File

@ -14,7 +14,7 @@
typedef struct Fft_s {
us nfft;
us nchannels;
Fftr* fftr;
vd fft_work;
} Fft;
Fft* Fft_alloc(const us nfft,const us nchannels) {
@ -35,19 +35,20 @@ Fft* Fft_alloc(const us nfft,const us nchannels) {
fft->nfft = nfft;
fft->nchannels = nchannels;
fft->fftr = Fftr_alloc(nfft);
if(!fft->fftr) {
a_free(fft);
WARN(ALLOCFAILED "fftr");
return NULL;
}
/* Initialize foreign fft lib */
fft->fft_work = vd_alloc(2*nfft+15);
npy_rffti(nfft,fft->fft_work.ptr);
check_overflow_vx(fft->fft_work);
/* print_vd(&fft->fft_work); */
feTRACE(15);
return fft;
}
void Fft_free(Fft* fft) {
fsTRACE(15);
Fftr_free(fft->fftr);
dbgassert(fft,NULLPTRDEREF);
vd_free(&fft->fft_work);
a_free(fft);
feTRACE(15);
}
@ -55,20 +56,51 @@ us Fft_nchannels(const Fft* fft) {return fft->nchannels;}
us Fft_nfft(const Fft* fft) {return fft->nfft;}
void Fft_fft(const Fft* fft,const dmat* timedata,cmat* result) {
fsTRACE(15);
dbgassert(timedata && result,NULLPTRDEREF);
dbgassert(fft && timedata && result,NULLPTRDEREF);
dbgassert(timedata->n_rows == fft->nfft,"Invalid size for time data rows."
" Should be equal to nfft");
dbgassert(timedata->n_cols == fft->nchannels,"Invalid size for time data cols."
" Should be equal to nchannels");
const us nfft = fft->nfft;
vd fft_result = vd_alloc(fft->nfft);
for(us col=0;col<fft->nchannels;col++) {
Fftr_fftr(fft->fftr,
getdmatval(timedata,0,col),
getcmatval(result,0,col));
vd timedata_col = dmat_column((dmat*) timedata,col);
vd_copy(&fft_result,&timedata_col);
vd_free(&timedata_col);
npy_rfftf(fft->nfft,fft_result.ptr,fft->fft_work.ptr);
/* Fftpack stores the data a bit strange, the resulting array
* has the DC value at 0,the first cosine at 1, the first sine
* at 2 etc. This needs to be shifted properly in the
* resulting matrix, as for the complex data, the imaginary
* part of the DC component equals zero. */
check_overflow_vx(fft_result);
check_overflow_vx(fft->fft_work);
vc resultcol = cmat_column(result,col);
*getvcval(&resultcol,0) = *getvdval(&fft_result,0);
memcpy((void*) getvcval(&resultcol,1),
(void*) getvdval(&fft_result,1),
(nfft-1)*sizeof(d));
/* Set imaginary part of Nyquist frequency to zero */
((d*) getvcval(&resultcol,nfft/2))[1] = 0;
check_overflow_xmat(*timedata);
/* Free up storage of the result column */
vc_free(&resultcol);
}
/* print_vd(&fft->fft_work); */
vd_free(&fft_result);
feTRACE(15);
}

15
beamforming/c/mq.c
View File

@ -14,11 +14,12 @@
#include "mq.h"
#include <pthread.h>
#ifdef __linux__
/* #ifdef linux */
#define _GNU_SOURCE
#include <sys/types.h>
#include <unistd.h>
#include <sys/syscall.h>
#endif
/* #endif */
typedef struct {
void* job_ptr;
@ -252,13 +253,9 @@ void* JobQueue_assign(JobQueue* jq) {
TRACE(16,"JobQueue_assign: found ready job. Assigned to:");
#ifdef ASCEE_DEBUG
#ifdef __linux__
pid_t tid = syscall(SYS_gettid);
iVARTRACE(16,tid);
#endif // __linux__
#endif // ASCEE_DEBUG
pthread_t thisthread = pthread_self();
iVARTRACE(16,thisthread);
#endif
/* print_job_queue(jq); */
/* Find a job from the queue, assign it and return it */

72
beamforming/c/ps.c
View File

@ -15,12 +15,8 @@
typedef struct PowerSpectra_s {
vd window;
d win_pow; /**< The power of the window */
Fft* fft; /**< Fft routines storage */
cmat fft_work; /**< Work area for FFt's */
dmat timedata_work; /**< Work area for timedata */
vc j_vec_conj; /**< Work area for conjugate of j */
} PowerSpectra;
PowerSpectra* PowerSpectra_alloc(const us nfft,
@ -53,11 +49,9 @@ PowerSpectra* PowerSpectra_alloc(const us nfft,
/* Allocate vectors and matrices */
ps->window = vd_alloc(nfft);
ps->fft_work = cmat_alloc(nfft/2+1,nchannels);
ps->timedata_work= dmat_alloc(nfft,nchannels);
ps->j_vec_conj = vc_alloc(nfft/2+1);
rv = window_create(wt,&ps->window,&ps->win_pow);
check_overflow_vx(ps->window);
if(rv!=0) {
WARN("Error creating window function, continuing anyway");
}
@ -67,14 +61,9 @@ PowerSpectra* PowerSpectra_alloc(const us nfft,
void PowerSpectra_free(PowerSpectra* ps) {
fsTRACE(15);
Fft_free(ps->fft);
vd_free(&ps->window);
cmat_free(&ps->fft_work);
dmat_free(&ps->timedata_work);
vc_free(&ps->j_vec_conj);
a_free(ps);
feTRACE(15);
}
@ -93,8 +82,6 @@ void PowerSpectra_compute(const PowerSpectra* ps,
const d win_pow = ps->win_pow;
dVARTRACE(15,win_pow);
us i,j;
/* Sanity checks for the matrices */
dbgassert(timedata->n_cols == nchannels,"timedata n_cols "
"should be equal to nchannels");
@ -111,25 +98,27 @@ void PowerSpectra_compute(const PowerSpectra* ps,
/* Multiply time data with the window and store result in
* timedata_work. */
dmat timedata_work = ps->timedata_work;
dmat timedata_work = dmat_alloc(nfft,nchannels);
for(us i=0;i<nchannels;i++) {
vd column = dmat_column((dmat*) timedata,i);
vd column_work = dmat_column(&timedata_work,i);
for(i=0;i<nchannels;i++) {
d_elem_prod_d(getdmatval(&timedata_work,0,i), /* Result */
getdmatval(timedata,0,i),
ps->window.data,
nfft);
vd_elem_prod(&column_work,&column,&ps->window);
vd_free(&column);
vd_free(&column_work);
}
check_overflow_xmat(timedata_work);
/* print_dmat(&timedata_work); */
/* Compute fft of the time data */
cmat fft_work = ps->fft_work;
cmat fft_work = cmat_alloc(nfft/2+1,nchannels);
Fft_fft(ps->fft,
&timedata_work,
&fft_work);
dmat_free(&timedata_work);
TRACE(15,"fft done");
/* Scale fft such that power is easily comxputed */
@ -137,40 +126,53 @@ void PowerSpectra_compute(const PowerSpectra* ps,
cmat_scale(&fft_work,scale_fac);
TRACE(15,"scale done");
for(i=0;i< nchannels;i++) {
for(us i=0;i< nchannels;i++) {
/* Multiply DC term with 1/sqrt(2) */
*getcmatval(&fft_work,0,i) *= 1/d_sqrt(2.)+0*I;
/* Multiply Nyquist term with 1/sqrt(2) */
*getcmatval(&fft_work,nfft/2,i) *= 1/d_sqrt(2.)+0*I;
}
check_overflow_xmat(fft_work);
/* print_cmat(&fft_work); */
TRACE(15,"Nyquist and DC correction done");
vc j_vec_conj = ps->j_vec_conj;
vc j_vec_conj = vc_alloc(nfft/2+1);
/* Compute Cross-power spectra and store result */
for(i =0; i<nchannels;i++) {
for(j=0;j<nchannels;j++) {
iVARTRACE(15,i);
iVARTRACE(15,j);
for(us i =0; i<nchannels;i++) {
for(us j=0;j<nchannels;j++) {
/* The indices here are important. This is also how it
* is documented */
vc res = cmat_column(result,i+j*nchannels);
TRACE(15,"SFSG");
check_overflow_vx(res);
vc i_vec = cmat_column(&fft_work,i);
vc j_vec = cmat_column(&fft_work,j);
TRACE(15,"SFSG");
check_overflow_xmat(fft_work);
/* Compute the conjugate of spectra j */
vc_conj_vc(&j_vec_conj,&j_vec);
TRACE(15,"SFSG");
vc_conj(&j_vec_conj,&j_vec);
check_overflow_xmat(fft_work);
/* Compute the element-wise product of the two vectors and
* store the result as the result */
vc_elem_prod(&res,&i_vec,&j_vec_conj);
TRACE(15,"SFSG");
vc_hadamard(&res,&i_vec,&j_vec_conj);
vc_free(&i_vec);
vc_free(&j_vec);
vc_free(&res);
}
}
check_overflow_xmat(*result);
check_overflow_xmat(*timedata);
cmat_free(&fft_work);
vc_free(&j_vec_conj);
feTRACE(15);
}

11
beamforming/c/types.h
View File

@ -9,15 +9,18 @@
#pragma once
#ifndef TYPES_H
#define TYPES_H
// Branch prediction performance improvement
#if !defined(likely) && defined(__GNUC__) && !defined(ASCEE_DEBUG)
#include <stddef.h>
// // Branch prediction performance improvement
#if !defined(likely)
#if defined(__GNUC__) && !defined(ASCEE_DEBUG)
#define likely(x) __builtin_expect(!!(x), 1)
#define unlikely(x) __builtin_expect(!!(x), 0)
#else
#define likely(x) (x)
#define unlikely(x) (x)
#endif
#endif // if defined(__GNUC__) && !defined(ASCEE_DEBUG)
#endif // !defined(likely)
/// We often use boolean values
#include <stdbool.h> // true, false

30
beamforming/c/window.c
View File

@ -8,7 +8,7 @@
#include "window.h"
#include "signals.h"
#include <stdlib.h>
/**
* Compute the Hann window
*
@ -53,9 +53,12 @@ static d rectangle(us i,us N) {
dbgassert(i<N,"Invalid index for window function Hann");
return 1.0;
}
static d bartlett(us n,us N) {
dbgassert(n<N,"Invalid index for window function Bartlett");
return 1 - d_abs(2*(n - (N-1)/2.)/(N-1));
}
int window_create(const WindowType wintype,vd* result,d* win_pow) {
fsTRACE(15);
us nfft = result->size;
d (*win_fun)(us,us);
switch (wintype) {
@ -67,21 +70,26 @@ int window_create(const WindowType wintype,vd* result,d* win_pow) {
win_fun = hamming;
break;
}
case Blackman: {
win_fun = blackman;
break;
}
case Rectangular: {
win_fun = rectangle;
break;
}
case Bartlett: {
win_fun = bartlett;
break;
}
case Blackman: {
win_fun = blackman;
break;
}
default:
WARN("Unkown window function");
return FAILURE;
DBGWARN("BUG: Unknown window function");
abort();
break;
}
us index;
for(index=0;index<nfft;index++) {
/* Compute the window function value */
d val = win_fun(index,nfft);
@ -89,10 +97,10 @@ int window_create(const WindowType wintype,vd* result,d* win_pow) {
setvecval(result,index,val);
}
/* Store window power in result */
*win_pow = signal_power(result);
/* Store window power in result */
feTRACE(15);
return SUCCESS;
}

6
beamforming/c/window.h
View File

@ -13,8 +13,10 @@
typedef enum {
Hann = 0,
Hamming = 1,
Blackman = 2,
Rectangular = 3,
Rectangular = 2,
Bartlett = 3,
Blackman = 4,
} WindowType;
/**

6
beamforming/config.pxi.in
View File

@ -38,6 +38,12 @@ cdef extern from "ascee_math.h":
us n_cols,
c* data)
cmat cmat_alloc(us n_rows,us n_cols)
dmat dmat_alloc(us n_rows,us n_cols)
void dmat_free(dmat*)
void cmat_free(cmat*)
void cmat_copy(cmat* to,cmat* from_)
cdef extern from "ascee_tracer.h":
void clearScreen()

54
beamforming/fft.pyx → beamforming/wrappers.pyx
View File

@ -1,6 +1,19 @@
include "config.pxi"
# setTracerLevel(0)
setTracerLevel(-5)
cdef extern from "cblas.h":
int openblas_get_num_threads()
void openblas_set_num_threads(int)
# If we touch this variable: we get segfaults when running from
# Spyder!
#openblas_set_num_threads(8)
# print("Number of threads: ",
# openblas_get_num_threads())
def cls():
clearScreen()
cls()
cdef extern from "fft.h":
ctypedef struct c_Fft "Fft"
@ -30,14 +43,16 @@ cdef class Fft:
assert timedata.shape[0] ==nfft
assert timedata.shape[1] == nchannels
result = np.empty((nfft//2+1,
nchannels),
dtype=NUMPY_COMPLEX_TYPE,order='F')
result = np.empty((nfft//2+1,nchannels),
dtype=NUMPY_COMPLEX_TYPE,
order='F')
# result[:,:] = np.nan+1j*np.nan
cdef c[::1,:] result_view = result
cdef cmat r = cmat_foreign(result.shape[0],
result.shape[1],
&result_view[0,0])
cdef dmat t = dmat_foreign(timedata.shape[0],
timedata.shape[1],
&timedata[0,0])
@ -54,8 +69,14 @@ cdef extern from "window.h":
ctypedef enum WindowType:
Hann = 0
Hamming = 1
Blackman = 2
Rectangular = 3
Rectangular = 2
Bartlett = 3
Blackman = 4
hann = 0
hamming = 1
rectangular = 2
bartlett = 3
blackman = 4
cdef extern from "ps.h":
ctypedef struct c_PowerSpectra "PowerSpectra"
@ -73,8 +94,8 @@ cdef class PowerSpectra:
cdef:
c_PowerSpectra* _ps
def __cinit__(self, us nfft,us nchannels):
self._ps = PowerSpectra_alloc(nfft,nchannels,Rectangular)
def __cinit__(self, us nfft,us nchannels,us window=rectangular):
self._ps = PowerSpectra_alloc(nfft,nchannels,<WindowType> window)
if self._ps == NULL:
raise RuntimeError('PowerSpectra allocation failed')
@ -85,6 +106,8 @@ cdef class PowerSpectra:
int rv
dmat td
cmat result_mat
td = dmat_foreign(nfft,
nchannels,
&timedata[0,0])
@ -99,12 +122,15 @@ cdef class PowerSpectra:
result = np.empty((nfft//2+1,nchannels,nchannels),
dtype = NUMPY_COMPLEX_TYPE,
order='F')
cdef c[::1,:,:] result_view = result
result_mat = cmat_foreign(nfft//2+1,
nchannels*nchannels,
&result_view[0,0,0])
PowerSpectra_compute(self._ps,&td,&result_mat)
dmat_free(&td)
@ -137,11 +163,14 @@ cdef class AvPowerSpectra:
c_AvPowerSpectra* aps
us nfft, nchannels
def __cinit__(self,us nfft,us nchannels,d overlap_percentage):
def __cinit__(self,us nfft,
us nchannels,
d overlap_percentage,
us window=rectangular):
self.aps = AvPowerSpectra_alloc(nfft,
nchannels,
overlap_percentage,
Rectangular)
<WindowType> window)
self.nchannels = nchannels
self.nfft = nfft
@ -196,3 +225,8 @@ cdef class AvPowerSpectra:
return result
def getFreq(self, d fs):
cdef d df = fs/self.nfft # frequency resolution
cdef us K = self.nfft//2+1 # number of frequency bins
return np.linspace(0, (K-1)*df, K)

124
fftpack/fftpack.h
View File

@ -1,111 +1,27 @@
// fftpack.h
//
// Author: J.A. de Jong - ASCEE
//
// Description:
// Header file for FFT routines of fftpack
//////////////////////////////////////////////////////////////////////
#pragma once
#ifndef FFTPACK_H
#define FFTPACK_H
#include "ascee_tracer.h"
#include "ascee_alloc.h"
#include "npy_fftpack.h"
/*
* Fortran has two kind of routines: functions and subroutines. A
Fortran function is a C function that returns a single value.
A * subroutine called from C is in fact a C function
which returns void.
* This file is part of tela the Tensor Language.
* Copyright (c) 1994-1995 Pekka Janhunen
*/
#if ASCEE_DOUBLE_PRECISION
// These are all subroutines
extern void dffti_ (int *nfft,d* wsave);
extern void dfftf_ (int *nfft,d* r,d* wsave);
#ifdef __cplusplus
extern "C" {
#endif
#define NPY_VISIBILITY_HIDDEN
#ifdef ASCEE_DOUBLE_PRECISION
#define Treal double
#else
extern void rffti_ (int *nfft,d* wsave);
extern void rfftf_ (int *nfft,d* r,d* wsave);
#endif // ASCEE_DOUBLE_PRECISION
#define Treal float
#endif
typedef struct Fftr_s {
d *work_area;
d* wsave_ptr;
int nfft;
} Fftr;
extern NPY_VISIBILITY_HIDDEN void npy_cfftf(int N, Treal data[], const Treal wrk[]);
extern NPY_VISIBILITY_HIDDEN void npy_cfftb(int N, Treal data[], const Treal wrk[]);
extern NPY_VISIBILITY_HIDDEN void npy_cffti(int N, Treal wrk[]);
/**
*
*
* @param nfft the length of the sequence to be transformed
extern NPY_VISIBILITY_HIDDEN void npy_rfftf(int N, Treal data[], const Treal wrk[]);
extern NPY_VISIBILITY_HIDDEN void npy_rfftb(int N, Treal data[], const Treal wrk[]);
extern NPY_VISIBILITY_HIDDEN void npy_rffti(int N, Treal wrk[]);
* @param wsave a work array which must be dimensioned at least
* 2*nfft+15. the same work array can be used for both rfftf and
* rfftb as long as n remains unchanged. different wsave arrays
* are required for different values of n. the contents of wsave
* must not be changed between calls of rfftf or rfftb.
*/
Fftr* Fftr_alloc(int nfft) {
fsTRACE(15);
Fftr* fftr = a_malloc(sizeof(fftr));
dbgassert(nfft>0,"Invalid nfft")
dbgassert(fftr,ALLOCFAILED "Fftr_alloc");
fftr->work_area = a_malloc(sizeof(d)*(3*nfft+15));
fftr->wsave_ptr = &fftr->work_area[nfft];
fftr->nfft = nfft;
#if ASCEE_DOUBLE_PRECISION
// dffti_(&nfft,fftr->wsave);
npy_rffti(nfft,fftr->wsave_ptr);
#else
// rffti_(&nfft,fftr->wsave);
#endif
feTRACE(15);
return fftr;
#ifdef __cplusplus
}
void Fftr_free(Fftr* fftr) {
dbgassert(fftr,NULLPTRDEREF "Fftr_free");
a_free(fftr->work_area);
a_free(fftr);
}
/**
*
*
* @param nfft
* @param wsave
* @param input
* @param work
* @param result
*/
static inline void Fftr_fftr(Fftr* fftr,d* input,c* result) {
// Copy contents of input to the work area
d_copy(fftr->work_area,input,fftr->nfft);
int nfft = fftr->nfft;
#if ASCEE_DOUBLE_PRECISION
// dfftf_(&nfft,fftr->work,fftr->wsave);
npy_rfftf(nfft,fftr->work_area,fftr->wsave_ptr);
#else
NOT TESTED
rfftf_(&nfft,fftr->work_area,fftr->wsave_ptr);
#endif
result[0] = fftr->work_area[0];
d_copy((d*) (&result[1]),&fftr->work_area[1],nfft);
// Not portable way of setting imaginary part to zero. Works with
// gcc, though.
__imag__ result[nfft/2] = 0;
}
#endif // FFTPACK_H
//////////////////////////////////////////////////////////////////////
#endif

29
fftpack/npy_fftpack.h
View File

@ -1,29 +0,0 @@
/*
* This file is part of tela the Tensor Language.
* Copyright (c) 1994-1995 Pekka Janhunen
*/
#ifdef __cplusplus
extern "C" {
#endif
#define NPY_VISIBILITY_HIDDEN
#define DOUBLE
#ifdef DOUBLE
#define Treal double
#else
#define Treal float
#endif
extern NPY_VISIBILITY_HIDDEN void npy_cfftf(int N, Treal data[], const Treal wrk[]);
extern NPY_VISIBILITY_HIDDEN void npy_cfftb(int N, Treal data[], const Treal wrk[]);
extern NPY_VISIBILITY_HIDDEN void npy_cffti(int N, Treal wrk[]);
extern NPY_VISIBILITY_HIDDEN void npy_rfftf(int N, Treal data[], const Treal wrk[]);
extern NPY_VISIBILITY_HIDDEN void npy_rfftb(int N, Treal data[], const Treal wrk[]);
extern NPY_VISIBILITY_HIDDEN void npy_rffti(int N, Treal wrk[]);
#ifdef __cplusplus
}
#endif

7
test/benchmark.py
View File

@ -1,12 +1,15 @@
#!/usr/bin/env python3
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
Created on Mon Jan 15 19:45:33 2018
@author: anne
"""
from pycallgraph import PyCallGraph
from pycallgraph.output import GraphvizOutput
import numpy as np
from beamforming.fft import Fft
from beamforming import Fft
nfft=2**17

2
test/fft_test.py
View File

@ -6,7 +6,7 @@ Created on Mon Jan 15 19:45:33 2018
@author: anne
"""
import numpy as np
from beamforming.fft import Fft
from beamforming import Fft
nfft=6
print('nfft:',nfft)

6
test/ps_test.py
View File

@ -6,9 +6,9 @@ Created on Mon Jan 15 19:45:33 2018
@author: anne
"""
import numpy as np
from beamforming.fft import Fft, PowerSpectra
nfft=2**4
from beamforming import Fft, PowerSpectra, cls
cls
nfft=2048
print('nfft:',nfft)
#print(nfft)
nchannels = 2

2
test/test_bf.c
View File

@ -35,7 +35,7 @@ int main() {
vc res2 = vc_alloc(3);
c_elem_prod_c(res2.data,vc1.data,vc2.data,3);
c_hadamard(res2.ptr,vc1.ptr,vc2.ptr,3);
print_vc(&res2);

26
test/test_math.c
View File

@ -9,25 +9,39 @@
#include "fft.h"
#include "ascee_math.h"
#include "ascee_tracer.h"
#include "ascee_alg.h"
int main() {
iVARTRACE(15,getTracerLevel());
vc a = vc_alloc(5);
vc_set(&a,2+3*I);
c a[5];
c_set(a,1-I,5);
c_conj_inplace(a.data,a.size);
print_vc(&a);
/* print_vc(&a); */
vc b = vc_alloc(5);
vc_set(&b,2);
c_conj_c(b.data,a.data,5);
printf("b:\n");
print_vc(&b);
vc_free(&a);
vc c1 = vc_alloc(5);
/* vc_set(&c1,10); */
/* c_add_to(c1.ptr,a.ptr,1,3); */
c_hadamard(c1.ptr,a,b.ptr,5);
printf("c1:\n");
print_vc(&c1);
vc_free(&b);
return 0;
}

5
test/test_workers.c
View File

@ -8,6 +8,7 @@
#include "worker.h"
#include "mq.h"
#include "ascee_tracer.h"
#include "ascee_assert.h"
#include <unistd.h>
static void* walloc(void*);
static int worker(void*,void*);
@ -22,7 +23,7 @@ int main() {
us njobs = 4;
JobQueue* jq = JobQueue_alloc(njobs);
assert(jq);
dbgassert(jq,NULLPTRDEREF);
Workers* w = Workers_create(njobs,
jq,
@ -30,7 +31,7 @@ int main() {
worker,
wfree,
(void*) 101);
assert(w);
dbgassert(jq,NULLPTRDEREF);
for(us i=0; i< njobs; i++) {
iVARTRACE(15,i);