this repo has no description
alignment fixing for USE_SIMD - all memory allocations are rounded up to 16-byte boundaries. This fixes crashes where the sub-buffers for complex FFTs end up unaligned violating compiler expectations
+41
-22
+4
-2
kiss_fft.c
+4
-2
kiss_fft.c
···
332
332
* */
333
333
kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem )
334
334
{
335
+
KISS_FFT_ALIGN_CHECK(mem)
336
+
335
337
kiss_fft_cfg st=NULL;
336
-
size_t memneeded = sizeof(struct kiss_fft_state)
337
-
+ sizeof(kiss_fft_cpx)*(nfft-1); /* twiddle factors*/
338
+
size_t memneeded = KISS_FFT_ALIGN_SIZE_UP(sizeof(struct kiss_fft_state)
339
+
+ sizeof(kiss_fft_cpx)*(nfft-1)); /* twiddle factors*/
338
340
339
341
if ( lenmem==NULL ) {
340
342
st = ( kiss_fft_cfg)KISS_FFT_MALLOC( memneeded );
+4
kiss_fft.h
+4
kiss_fft.h
···
37
37
# define kiss_fft_scalar __m128
38
38
# ifndef KISS_FFT_MALLOC
39
39
# define KISS_FFT_MALLOC(nbytes) _mm_malloc(nbytes,16)
40
+
# define KISS_FFT_ALIGN_CHECK(ptr) assert((((uintptr_t) ptr) & 0xF) == 0);
41
+
# define KISS_FFT_ALIGN_SIZE_UP(size) ((size + 15UL) & ~0xFUL)
40
42
# endif
41
43
# ifndef KISS_FFT_FREE
42
44
# define KISS_FFT_FREE _mm_free
43
45
# endif
44
46
#else
47
+
# define KISS_FFT_ALIGN_CHECK(ptr)
48
+
# define KISS_FFT_ALIGN_SIZE_UP(size) (size)
45
49
# ifndef KISS_FFT_MALLOC
46
50
# define KISS_FFT_MALLOC malloc
47
51
# endif
+15
-12
tools/kiss_fftnd.c
+15
-12
tools/kiss_fftnd.c
···
19
19
20
20
kiss_fftnd_cfg kiss_fftnd_alloc(const int *dims,int ndims,int inverse_fft,void*mem,size_t*lenmem)
21
21
{
22
+
KISS_FFT_ALIGN_CHECK(mem)
23
+
22
24
kiss_fftnd_cfg st = NULL;
23
25
int i;
24
26
int dimprod=1;
25
-
size_t memneeded = sizeof(struct kiss_fftnd_state);
26
-
char * ptr;
27
+
size_t memneeded = KISS_FFT_ALIGN_SIZE_UP(sizeof(struct kiss_fftnd_state));
28
+
char * ptr = NULL;
27
29
28
30
for (i=0;i<ndims;++i) {
29
31
size_t sublen=0;
···
31
33
memneeded += sublen; /* st->states[i] */
32
34
dimprod *= dims[i];
33
35
}
34
-
memneeded += sizeof(int) * ndims;/* st->dims */
35
-
memneeded += sizeof(void*) * ndims;/* st->states */
36
-
memneeded += sizeof(kiss_fft_cpx) * dimprod; /* st->tmpbuf */
36
+
memneeded += KISS_FFT_ALIGN_SIZE_UP(sizeof(int) * ndims);/* st->dims */
37
+
memneeded += KISS_FFT_ALIGN_SIZE_UP(sizeof(void*) * ndims);/* st->states */
38
+
memneeded += KISS_FFT_ALIGN_SIZE_UP(sizeof(kiss_fft_cpx) * dimprod); /* st->tmpbuf */
37
39
38
40
if (lenmem == NULL) {/* allocate for the caller*/
39
-
st = (kiss_fftnd_cfg) malloc (memneeded);
41
+
ptr = (char *) malloc (memneeded);
40
42
} else { /* initialize supplied buffer if big enough */
41
43
if (*lenmem >= memneeded)
42
-
st = (kiss_fftnd_cfg) mem;
44
+
ptr = (char *) mem;
43
45
*lenmem = memneeded; /*tell caller how big struct is (or would be) */
44
46
}
45
-
if (!st)
47
+
if (!ptr)
46
48
return NULL; /*malloc failed or buffer too small */
47
49
50
+
st = (kiss_fftnd_cfg) ptr;
48
51
st->dimprod = dimprod;
49
52
st->ndims = ndims;
50
-
ptr=(char*)(st+1);
53
+
ptr += KISS_FFT_ALIGN_SIZE_UP(sizeof(struct kiss_fftnd_state));
51
54
52
55
st->states = (kiss_fft_cfg *)ptr;
53
-
ptr += sizeof(void*) * ndims;
56
+
ptr += KISS_FFT_ALIGN_SIZE_UP(sizeof(void*) * ndims);
54
57
55
58
st->dims = (int*)ptr;
56
-
ptr += sizeof(int) * ndims;
59
+
ptr += KISS_FFT_ALIGN_SIZE_UP(sizeof(int) * ndims);
57
60
58
61
st->tmpbuf = (kiss_fft_cpx*)ptr;
59
-
ptr += sizeof(kiss_fft_cpx) * dimprod;
62
+
ptr += KISS_FFT_ALIGN_SIZE_UP(sizeof(kiss_fft_cpx) * dimprod);
60
63
61
64
for (i=0;i<ndims;++i) {
62
65
size_t len;
+16
-8
tools/kiss_fftndr.c
+16
-8
tools/kiss_fftndr.c
···
29
29
30
30
kiss_fftndr_cfg kiss_fftndr_alloc(const int *dims,int ndims,int inverse_fft,void*mem,size_t*lenmem)
31
31
{
32
+
KISS_FFT_ALIGN_CHECK(mem)
33
+
32
34
kiss_fftndr_cfg st = NULL;
33
35
size_t nr=0 , nd=0,ntmp=0;
34
36
int dimReal = dims[ndims-1];
35
37
int dimOther = prod(dims,ndims-1);
36
38
size_t memneeded;
37
-
39
+
char * ptr = NULL;
40
+
38
41
(void)kiss_fftr_alloc(dimReal,inverse_fft,NULL,&nr);
39
42
(void)kiss_fftnd_alloc(dims,ndims-1,inverse_fft,NULL,&nd);
40
43
ntmp =
41
44
MAX( 2*dimOther , dimReal+2) * sizeof(kiss_fft_scalar) // freq buffer for one pass
42
45
+ dimOther*(dimReal+2) * sizeof(kiss_fft_scalar); // large enough to hold entire input in case of in-place
43
46
44
-
memneeded = sizeof( struct kiss_fftndr_state ) + nr + nd + ntmp;
47
+
memneeded = KISS_FFT_ALIGN_SIZE_UP(sizeof( struct kiss_fftndr_state )) + KISS_FFT_ALIGN_SIZE_UP(nr) + KISS_FFT_ALIGN_SIZE_UP(nd) + KISS_FFT_ALIGN_SIZE_UP(ntmp);
45
48
46
49
if (lenmem==NULL) {
47
-
st = (kiss_fftndr_cfg) malloc(memneeded);
50
+
ptr = (char*) malloc(memneeded);
48
51
}else{
49
52
if (*lenmem >= memneeded)
50
-
st = (kiss_fftndr_cfg)mem;
53
+
ptr = (char *)mem;
51
54
*lenmem = memneeded;
52
55
}
53
-
if (st==NULL)
56
+
if (ptr==NULL)
54
57
return NULL;
58
+
59
+
st = (kiss_fftndr_cfg) ptr;
55
60
memset( st , 0 , memneeded);
61
+
ptr += KISS_FFT_ALIGN_SIZE_UP(sizeof(struct kiss_fftndr_state));
56
62
57
63
st->dimReal = dimReal;
58
64
st->dimOther = dimOther;
59
-
st->cfg_r = kiss_fftr_alloc( dimReal,inverse_fft,st+1,&nr);
60
-
st->cfg_nd = kiss_fftnd_alloc(dims,ndims-1,inverse_fft, ((char*) st->cfg_r)+nr,&nd);
61
-
st->tmpbuf = (char*)st->cfg_nd + nd;
65
+
st->cfg_r = kiss_fftr_alloc( dimReal,inverse_fft,ptr,&nr);
66
+
ptr += KISS_FFT_ALIGN_SIZE_UP(nr);
67
+
st->cfg_nd = kiss_fftnd_alloc(dims,ndims-1,inverse_fft, ptr,&nd);
68
+
ptr += KISS_FFT_ALIGN_SIZE_UP(nd);
69
+
st->tmpbuf = ptr;
62
70
63
71
return st;
64
72
}