caiyu/nextpnr
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fixes for building on windows

Browse Source
This commit is contained in:
Justin Rajewski 2023-10-26 10:56:19 -06:00 committed by myrtle
parent 95f0a19391
commit 7cac0249a1
1 changed files with 777 additions and 928 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

349
3rdparty/oourafft/fftsg2d.cc vendored
View File

@ -366,13 +366,11 @@ macro definitions
. .
*/ */
#include <stdio.h> #include <stdio.h>
#include <stdlib.h> #include <stdlib.h>
#define fft2d_alloc_error_check(p) \ #define fft2d_alloc_error_check(p) \
{ \ { \
if((p) == NULL) { \ if ((p) == NULL) { \
fprintf(stderr, "fft2d memory allocation error\n"); \ fprintf(stderr, "fft2d memory allocation error\n"); \
exit(1); \ exit(1); \
} \ } \
@ -390,14 +388,14 @@ macro definitions
#define fft2d_thread_t pthread_t #define fft2d_thread_t pthread_t
#define fft2d_thread_create(thp, func, argp) \ #define fft2d_thread_create(thp, func, argp) \
{ \ { \
if(pthread_create(thp, NULL, func, (void *)(argp)) != 0) { \ if (pthread_create(thp, NULL, func, (void *)(argp)) != 0) { \
fprintf(stderr, "fft2d thread error\n"); \ fprintf(stderr, "fft2d thread error\n"); \
exit(1); \ exit(1); \
} \ } \
} }
#define fft2d_thread_wait(th) \ #define fft2d_thread_wait(th) \
{ \ { \
if(pthread_join(th, NULL) != 0) { \ if (pthread_join(th, NULL) != 0) { \
fprintf(stderr, "fft2d thread error\n"); \ fprintf(stderr, "fft2d thread error\n"); \
exit(1); \ exit(1); \
} \ } \
@ -418,9 +416,8 @@ macro definitions
#define fft2d_thread_create(thp, func, argp) \ #define fft2d_thread_create(thp, func, argp) \
{ \ { \
DWORD thid; \ DWORD thid; \
*(thp) = CreateThread(NULL, 0, (LPTHREAD_START_ROUTINE)(func), \ *(thp) = CreateThread(NULL, 0, (LPTHREAD_START_ROUTINE)(func), (LPVOID)(argp), 0, &thid); \
(LPVOID)(argp), 0, &thid); \ if (*(thp) == 0) { \
if(*(thp) == 0) { \
fprintf(stderr, "fft2d thread error\n"); \ fprintf(stderr, "fft2d thread error\n"); \
exit(1); \ exit(1); \
} \ } \
@ -436,19 +433,28 @@ macro definitions
NEXTPNR_NAMESPACE_BEGIN NEXTPNR_NAMESPACE_BEGIN
void ddct(int n, int isgn, float *a, int *ip, float *w);
void ddst(int n, int isgn, float *a, int *ip, float *w);
void cdft(int n, int isgn, float *a, int *ip, float *w);
void *cdft2d_th(void *p);
void *ddxt2d0_th(void *p);
void *ddxt2d_th(void *p);
void rdft(int n, int isgn, float *a, int *ip, float *w);
void makewt(int nw, int *ip, float *w);
void cdft2d_sub(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w);
void makect(int nc, int *ip, float *c);
void rdft2d_sub(int n1, int isgn, float **a);
void ddxt2d_sub(int n1, int n2, int ics, int isgn, float **a, float *t, int *ip, float *w);
void cdft2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
{
void makewt(int nw, int* ip, float* w);
void cdft(int n, int isgn, float* a, int* ip, float* w);
void cdft2d_sub(
int n1, int n2, int isgn, float** a, float* t, int* ip, float* w);
#ifdef USE_FFT2D_THREADS #ifdef USE_FFT2D_THREADS
void xdft2d0_subth( void xdft2d0_subth(int n1, int n2, int icr, int isgn, float **a, int *ip, float *w);
int n1, int n2, int icr, int isgn, float** a, int* ip, float* w); void cdft2d_subth(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w);
void cdft2d_subth( void ddxt2d0_subth(int n1, int n2, int ics, int isgn, float **a, int *ip, float *w);
int n1, int n2, int isgn, float** a, float* t, int* ip, float* w); void ddxt2d_subth(int n1, int n2, int ics, int isgn, float **a, float *t, int *ip, float *w);
#endif /* USE_FFT2D_THREADS */ #endif /* USE_FFT2D_THREADS */
void cdft2d(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w)
{
int n, itnull, nthread, nt, i; int n, itnull, nthread, nt, i;
n = n1 << 1; n = n1 << 1;
@ -471,11 +477,11 @@ void cdft2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} else if (n2 < 4 * nthread) { } else if (n2 < 4 * nthread) {
nt >>= 2; nt >>= 2;
} }
t = (float*) malloc(sizeof(float) * nt); t = (float *)malloc(sizeof(float) * nt);
fft2d_alloc_error_check(t); fft2d_alloc_error_check(t);
} }
#ifdef USE_FFT2D_THREADS #ifdef USE_FFT2D_THREADS
if ((float) n1 * n2 >= (float) FFT2D_THREADS_BEGIN_N) { if ((float)n1 * n2 >= (float)FFT2D_THREADS_BEGIN_N) {
xdft2d0_subth(n1, n2, 0, isgn, a, ip, w); xdft2d0_subth(n1, n2, 0, isgn, a, ip, w);
cdft2d_subth(n1, n2, isgn, a, t, ip, w); cdft2d_subth(n1, n2, isgn, a, t, ip, w);
} else } else
@ -491,20 +497,8 @@ void cdft2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} }
} }
void rdft2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w) void rdft2d(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w)
{ {
void makewt(int nw, int* ip, float* w);
void makect(int nc, int* ip, float* c);
void rdft(int n, int isgn, float* a, int* ip, float* w);
void cdft2d_sub(
int n1, int n2, int isgn, float** a, float* t, int* ip, float* w);
void rdft2d_sub(int n1, int isgn, float** a);
#ifdef USE_FFT2D_THREADS
void xdft2d0_subth(
int n1, int n2, int icr, int isgn, float** a, int* ip, float* w);
void cdft2d_subth(
int n1, int n2, int isgn, float** a, float* t, int* ip, float* w);
#endif /* USE_FFT2D_THREADS */
int n, nw, nc, itnull, nthread, nt, i; int n, nw, nc, itnull, nthread, nt, i;
n = n1 << 1; n = n1 << 1;
@ -534,11 +528,11 @@ void rdft2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} else if (n2 < 4 * nthread) { } else if (n2 < 4 * nthread) {
nt >>= 2; nt >>= 2;
} }
t = (float*) malloc(sizeof(float) * nt); t = (float *)malloc(sizeof(float) * nt);
fft2d_alloc_error_check(t); fft2d_alloc_error_check(t);
} }
#ifdef USE_FFT2D_THREADS #ifdef USE_FFT2D_THREADS
if ((float) n1 * n2 >= (float) FFT2D_THREADS_BEGIN_N) { if ((float)n1 * n2 >= (float)FFT2D_THREADS_BEGIN_N) {
if (isgn < 0) { if (isgn < 0) {
rdft2d_sub(n1, isgn, a); rdft2d_sub(n1, isgn, a);
cdft2d_subth(n1, n2, isgn, a, t, ip, w); cdft2d_subth(n1, n2, isgn, a, t, ip, w);
@ -568,7 +562,7 @@ void rdft2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} }
} }
void rdft2dsort(int n1, int n2, int isgn, float** a) void rdft2dsort(int n1, int n2, int isgn, float **a)
{ {
int n1h, i; int n1h, i;
float x, y; float x, y;
@ -601,32 +595,8 @@ void rdft2dsort(int n1, int n2, int isgn, float** a)
} }
} }
void ddcst2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w) void ddcst2d(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w)
{ {
void makewt(int nw, int* ip, float* w);
void makect(int nc, int* ip, float* c);
void ddct(int n, int isgn, float* a, int* ip, float* w);
void ddst(int n, int isgn, float* a, int* ip, float* w);
void ddxt2d_sub(int n1,
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w);
#ifdef USE_FFT2D_THREADS
void ddxt2d0_subth(
int n1, int n2, int ics, int isgn, float** a, int* ip, float* w);
void ddxt2d_subth(int n1,
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w);
#endif /* USE_FFT2D_THREADS */
int n, nw, nc, itnull, nthread, nt, i; int n, nw, nc, itnull, nthread, nt, i;
n = n1; n = n1;
@ -656,11 +626,11 @@ void ddcst2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} else if (n2 < 2 * nthread) { } else if (n2 < 2 * nthread) {
nt >>= 2; nt >>= 2;
} }
t = (float*) malloc(sizeof(float) * nt); t = (float *)malloc(sizeof(float) * nt);
fft2d_alloc_error_check(t); fft2d_alloc_error_check(t);
} }
#ifdef USE_FFT2D_THREADS #ifdef USE_FFT2D_THREADS
if ((float) n1 * n2 >= (float) FFT2D_THREADS_BEGIN_N) { if ((float)n1 * n2 >= (float)FFT2D_THREADS_BEGIN_N) {
ddxt2d0_subth(n1, n2, 1, isgn, a, ip, w); ddxt2d0_subth(n1, n2, 1, isgn, a, ip, w);
ddxt2d_subth(n1, n2, 0, isgn, a, t, ip, w); ddxt2d_subth(n1, n2, 0, isgn, a, t, ip, w);
} else } else
@ -676,32 +646,8 @@ void ddcst2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} }
} }
void ddsct2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w) void ddsct2d(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w)
{ {
void makewt(int nw, int* ip, float* w);
void makect(int nc, int* ip, float* c);
void ddct(int n, int isgn, float* a, int* ip, float* w);
void ddst(int n, int isgn, float* a, int* ip, float* w);
void ddxt2d_sub(int n1,
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w);
#ifdef USE_FFT2D_THREADS
void ddxt2d0_subth(
int n1, int n2, int ics, int isgn, float** a, int* ip, float* w);
void ddxt2d_subth(int n1,
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w);
#endif /* USE_FFT2D_THREADS */
int n, nw, nc, itnull, nthread, nt, i; int n, nw, nc, itnull, nthread, nt, i;
n = n1; n = n1;
@ -731,11 +677,11 @@ void ddsct2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} else if (n2 < 2 * nthread) { } else if (n2 < 2 * nthread) {
nt >>= 2; nt >>= 2;
} }
t = (float*) malloc(sizeof(float) * nt); t = (float *)malloc(sizeof(float) * nt);
fft2d_alloc_error_check(t); fft2d_alloc_error_check(t);
} }
#ifdef USE_FFT2D_THREADS #ifdef USE_FFT2D_THREADS
if ((float) n1 * n2 >= (float) FFT2D_THREADS_BEGIN_N) { if ((float)n1 * n2 >= (float)FFT2D_THREADS_BEGIN_N) {
ddxt2d0_subth(n1, n2, 0, isgn, a, ip, w); ddxt2d0_subth(n1, n2, 0, isgn, a, ip, w);
ddxt2d_subth(n1, n2, 1, isgn, a, t, ip, w); ddxt2d_subth(n1, n2, 1, isgn, a, t, ip, w);
} else } else
@ -751,31 +697,8 @@ void ddsct2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} }
} }
void ddct2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w) void ddct2d(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w)
{ {
void makewt(int nw, int* ip, float* w);
void makect(int nc, int* ip, float* c);
void ddct(int n, int isgn, float* a, int* ip, float* w);
void ddxt2d_sub(int n1,
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w);
#ifdef USE_FFT2D_THREADS
void ddxt2d0_subth(
int n1, int n2, int ics, int isgn, float** a, int* ip, float* w);
void ddxt2d_subth(int n1,
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w);
#endif /* USE_FFT2D_THREADS */
int n, nw, nc, itnull, nthread, nt, i; int n, nw, nc, itnull, nthread, nt, i;
n = n1; n = n1;
@ -805,11 +728,11 @@ void ddct2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} else if (n2 < 2 * nthread) { } else if (n2 < 2 * nthread) {
nt >>= 2; nt >>= 2;
} }
t = (float*) malloc(sizeof(float) * nt); t = (float *)malloc(sizeof(float) * nt);
fft2d_alloc_error_check(t); fft2d_alloc_error_check(t);
} }
#ifdef USE_FFT2D_THREADS #ifdef USE_FFT2D_THREADS
if ((float) n1 * n2 >= (float) FFT2D_THREADS_BEGIN_N) { if ((float)n1 * n2 >= (float)FFT2D_THREADS_BEGIN_N) {
ddxt2d0_subth(n1, n2, 0, isgn, a, ip, w); ddxt2d0_subth(n1, n2, 0, isgn, a, ip, w);
ddxt2d_subth(n1, n2, 0, isgn, a, t, ip, w); ddxt2d_subth(n1, n2, 0, isgn, a, t, ip, w);
} else } else
@ -825,31 +748,8 @@ void ddct2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} }
} }
void ddst2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w) void ddst2d(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w)
{ {
void makewt(int nw, int* ip, float* w);
void makect(int nc, int* ip, float* c);
void ddst(int n, int isgn, float* a, int* ip, float* w);
void ddxt2d_sub(int n1,
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w);
#ifdef USE_FFT2D_THREADS
void ddxt2d0_subth(
int n1, int n2, int ics, int isgn, float** a, int* ip, float* w);
void ddxt2d_subth(int n1,
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w);
#endif /* USE_FFT2D_THREADS */
int n, nw, nc, itnull, nthread, nt, i; int n, nw, nc, itnull, nthread, nt, i;
n = n1; n = n1;
@ -879,11 +779,11 @@ void ddst2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
} else if (n2 < 2 * nthread) { } else if (n2 < 2 * nthread) {
nt >>= 2; nt >>= 2;
} }
t = (float*) malloc(sizeof(float) * nt); t = (float *)malloc(sizeof(float) * nt);
fft2d_alloc_error_check(t); fft2d_alloc_error_check(t);
} }
#ifdef USE_FFT2D_THREADS #ifdef USE_FFT2D_THREADS
if ((float) n1 * n2 >= (float) FFT2D_THREADS_BEGIN_N) { if ((float)n1 * n2 >= (float)FFT2D_THREADS_BEGIN_N) {
ddxt2d0_subth(n1, n2, 1, isgn, a, ip, w); ddxt2d0_subth(n1, n2, 1, isgn, a, ip, w);
ddxt2d_subth(n1, n2, 1, isgn, a, t, ip, w); ddxt2d_subth(n1, n2, 1, isgn, a, t, ip, w);
} else } else
@ -901,15 +801,8 @@ void ddst2d(int n1, int n2, int isgn, float** a, float* t, int* ip, float* w)
/* -------- child routines -------- */ /* -------- child routines -------- */
void cdft2d_sub(int n1, void cdft2d_sub(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w)
int n2,
int isgn,
float** a,
float* t,
int* ip,
float* w)
{ {
void cdft(int n, int isgn, float* a, int* ip, float* w);
int i, j; int i, j;
if (n2 > 4) { if (n2 > 4) {
@ -967,7 +860,7 @@ void cdft2d_sub(int n1,
} }
} }
void rdft2d_sub(int n1, int isgn, float** a) void rdft2d_sub(int n1, int isgn, float **a)
{ {
int n1h, i, j; int n1h, i, j;
float xi; float xi;
@ -994,17 +887,8 @@ void rdft2d_sub(int n1, int isgn, float** a)
} }
} }
void ddxt2d_sub(int n1, void ddxt2d_sub(int n1, int n2, int ics, int isgn, float **a, float *t, int *ip, float *w)
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w)
{ {
void ddct(int n, int isgn, float* a, int* ip, float* w);
void ddst(int n, int isgn, float* a, int* ip, float* w);
int i, j; int i, j;
if (n2 > 2) { if (n2 > 2) {
@ -1061,22 +945,16 @@ struct fft2d_arg_st
int n2; int n2;
int ic; int ic;
int isgn; int isgn;
float** a; float **a;
float* t; float *t;
int* ip; int *ip;
float* w; float *w;
}; };
typedef struct fft2d_arg_st fft2d_arg_t; typedef struct fft2d_arg_st fft2d_arg_t;
void xdft2d0_subth(int n1, void xdft2d0_subth(int n1, int n2, int icr, int isgn, float **a, int *ip, float *w)
int n2,
int icr,
int isgn,
float** a,
int* ip,
float* w)
{ {
void* xdft2d0_th(void* p); void *xdft2d0_th(void *p);
fft2d_thread_t th[FFT2D_MAX_THREADS]; fft2d_thread_t th[FFT2D_MAX_THREADS];
fft2d_arg_t ag[FFT2D_MAX_THREADS]; fft2d_arg_t ag[FFT2D_MAX_THREADS];
int nthread, i; int nthread, i;
@ -1102,15 +980,8 @@ void xdft2d0_subth(int n1,
} }
} }
void cdft2d_subth(int n1, void cdft2d_subth(int n1, int n2, int isgn, float **a, float *t, int *ip, float *w)
int n2,
int isgn,
float** a,
float* t,
int* ip,
float* w)
{ {
void* cdft2d_th(void* p);
fft2d_thread_t th[FFT2D_MAX_THREADS]; fft2d_thread_t th[FFT2D_MAX_THREADS];
fft2d_arg_t ag[FFT2D_MAX_THREADS]; fft2d_arg_t ag[FFT2D_MAX_THREADS];
int nthread, nt, i; int nthread, nt, i;
@ -1140,15 +1011,8 @@ void cdft2d_subth(int n1,
} }
} }
void ddxt2d0_subth(int n1, void ddxt2d0_subth(int n1, int n2, int ics, int isgn, float **a, int *ip, float *w)
int n2,
int ics,
int isgn,
float** a,
int* ip,
float* w)
{ {
void* ddxt2d0_th(void* p);
fft2d_thread_t th[FFT2D_MAX_THREADS]; fft2d_thread_t th[FFT2D_MAX_THREADS];
fft2d_arg_t ag[FFT2D_MAX_THREADS]; fft2d_arg_t ag[FFT2D_MAX_THREADS];
int nthread, i; int nthread, i;
@ -1174,16 +1038,8 @@ void ddxt2d0_subth(int n1,
} }
} }
void ddxt2d_subth(int n1, void ddxt2d_subth(int n1, int n2, int ics, int isgn, float **a, float *t, int *ip, float *w)
int n2,
int ics,
int isgn,
float** a,
float* t,
int* ip,
float* w)
{ {
void* ddxt2d_th(void* p);
fft2d_thread_t th[FFT2D_MAX_THREADS]; fft2d_thread_t th[FFT2D_MAX_THREADS];
fft2d_arg_t ag[FFT2D_MAX_THREADS]; fft2d_arg_t ag[FFT2D_MAX_THREADS];
int nthread, nt, i; int nthread, nt, i;
@ -1214,22 +1070,20 @@ void ddxt2d_subth(int n1,
} }
} }
void* xdft2d0_th(void* p) void *xdft2d0_th(void *p)
{ {
void cdft(int n, int isgn, float* a, int* ip, float* w);
void rdft(int n, int isgn, float* a, int* ip, float* w);
int nthread, n0, n1, n2, icr, isgn, *ip, i; int nthread, n0, n1, n2, icr, isgn, *ip, i;
float **a, *w; float **a, *w;
nthread = ((fft2d_arg_t*) p)->nthread; nthread = ((fft2d_arg_t *)p)->nthread;
n0 = ((fft2d_arg_t*) p)->n0; n0 = ((fft2d_arg_t *)p)->n0;
n1 = ((fft2d_arg_t*) p)->n1; n1 = ((fft2d_arg_t *)p)->n1;
n2 = ((fft2d_arg_t*) p)->n2; n2 = ((fft2d_arg_t *)p)->n2;
icr = ((fft2d_arg_t*) p)->ic; icr = ((fft2d_arg_t *)p)->ic;
isgn = ((fft2d_arg_t*) p)->isgn; isgn = ((fft2d_arg_t *)p)->isgn;
a = ((fft2d_arg_t*) p)->a; a = ((fft2d_arg_t *)p)->a;
ip = ((fft2d_arg_t*) p)->ip; ip = ((fft2d_arg_t *)p)->ip;
w = ((fft2d_arg_t*) p)->w; w = ((fft2d_arg_t *)p)->w;
if (icr == 0) { if (icr == 0) {
for (i = n0; i < n1; i += nthread) { for (i = n0; i < n1; i += nthread) {
cdft(n2, isgn, a[i], ip, w); cdft(n2, isgn, a[i], ip, w);
@ -1239,24 +1093,23 @@ void* xdft2d0_th(void* p)
rdft(n2, isgn, a[i], ip, w); rdft(n2, isgn, a[i], ip, w);
} }
} }
return (void*) 0; return (void *)0;
} }
void* cdft2d_th(void* p) void *cdft2d_th(void *p)
{ {
void cdft(int n, int isgn, float* a, int* ip, float* w);
int nthread, n0, n1, n2, isgn, *ip, i, j; int nthread, n0, n1, n2, isgn, *ip, i, j;
float **a, *t, *w; float **a, *t, *w;
nthread = ((fft2d_arg_t*) p)->nthread; nthread = ((fft2d_arg_t *)p)->nthread;
n0 = ((fft2d_arg_t*) p)->n0; n0 = ((fft2d_arg_t *)p)->n0;
n1 = ((fft2d_arg_t*) p)->n1; n1 = ((fft2d_arg_t *)p)->n1;
n2 = ((fft2d_arg_t*) p)->n2; n2 = ((fft2d_arg_t *)p)->n2;
isgn = ((fft2d_arg_t*) p)->isgn; isgn = ((fft2d_arg_t *)p)->isgn;
a = ((fft2d_arg_t*) p)->a; a = ((fft2d_arg_t *)p)->a;
t = ((fft2d_arg_t*) p)->t; t = ((fft2d_arg_t *)p)->t;
ip = ((fft2d_arg_t*) p)->ip; ip = ((fft2d_arg_t *)p)->ip;
w = ((fft2d_arg_t*) p)->w; w = ((fft2d_arg_t *)p)->w;
if (n2 > 4 * nthread) { if (n2 > 4 * nthread) {
for (j = 8 * n0; j < n2; j += 8 * nthread) { for (j = 8 * n0; j < n2; j += 8 * nthread) {
for (i = 0; i < n1; i++) { for (i = 0; i < n1; i++) {
@ -1310,25 +1163,23 @@ void* cdft2d_th(void* p)
a[i][2 * n0 + 1] = t[2 * i + 1]; a[i][2 * n0 + 1] = t[2 * i + 1];
} }
} }
return (void*) 0; return (void *)0;
} }
void* ddxt2d0_th(void* p) void *ddxt2d0_th(void *p)
{ {
void ddct(int n, int isgn, float* a, int* ip, float* w);
void ddst(int n, int isgn, float* a, int* ip, float* w);
int nthread, n0, n1, n2, ics, isgn, *ip, i; int nthread, n0, n1, n2, ics, isgn, *ip, i;
float **a, *w; float **a, *w;
nthread = ((fft2d_arg_t*) p)->nthread; nthread = ((fft2d_arg_t *)p)->nthread;
n0 = ((fft2d_arg_t*) p)->n0; n0 = ((fft2d_arg_t *)p)->n0;
n1 = ((fft2d_arg_t*) p)->n1; n1 = ((fft2d_arg_t *)p)->n1;
n2 = ((fft2d_arg_t*) p)->n2; n2 = ((fft2d_arg_t *)p)->n2;
ics = ((fft2d_arg_t*) p)->ic; ics = ((fft2d_arg_t *)p)->ic;
isgn = ((fft2d_arg_t*) p)->isgn; isgn = ((fft2d_arg_t *)p)->isgn;
a = ((fft2d_arg_t*) p)->a; a = ((fft2d_arg_t *)p)->a;
ip = ((fft2d_arg_t*) p)->ip; ip = ((fft2d_arg_t *)p)->ip;
w = ((fft2d_arg_t*) p)->w; w = ((fft2d_arg_t *)p)->w;
if (ics == 0) { if (ics == 0) {
for (i = n0; i < n1; i += nthread) { for (i = n0; i < n1; i += nthread) {
ddct(n2, isgn, a[i], ip, w); ddct(n2, isgn, a[i], ip, w);
@ -1338,26 +1189,24 @@ void* ddxt2d0_th(void* p)
ddst(n2, isgn, a[i], ip, w); ddst(n2, isgn, a[i], ip, w);
} }
} }
return (void*) 0; return (void *)0;
} }
void* ddxt2d_th(void* p) void *ddxt2d_th(void *p)
{ {
void ddct(int n, int isgn, float* a, int* ip, float* w);
void ddst(int n, int isgn, float* a, int* ip, float* w);
int nthread, n0, n1, n2, ics, isgn, *ip, i, j; int nthread, n0, n1, n2, ics, isgn, *ip, i, j;
float **a, *t, *w; float **a, *t, *w;
nthread = ((fft2d_arg_t*) p)->nthread; nthread = ((fft2d_arg_t *)p)->nthread;
n0 = ((fft2d_arg_t*) p)->n0; n0 = ((fft2d_arg_t *)p)->n0;
n1 = ((fft2d_arg_t*) p)->n1; n1 = ((fft2d_arg_t *)p)->n1;
n2 = ((fft2d_arg_t*) p)->n2; n2 = ((fft2d_arg_t *)p)->n2;
ics = ((fft2d_arg_t*) p)->ic; ics = ((fft2d_arg_t *)p)->ic;
isgn = ((fft2d_arg_t*) p)->isgn; isgn = ((fft2d_arg_t *)p)->isgn;
a = ((fft2d_arg_t*) p)->a; a = ((fft2d_arg_t *)p)->a;
t = ((fft2d_arg_t*) p)->t; t = ((fft2d_arg_t *)p)->t;
ip = ((fft2d_arg_t*) p)->ip; ip = ((fft2d_arg_t *)p)->ip;
w = ((fft2d_arg_t*) p)->w; w = ((fft2d_arg_t *)p)->w;
if (n2 > 2 * nthread) { if (n2 > 2 * nthread) {
for (j = 4 * n0; j < n2; j += 4 * nthread) { for (j = 4 * n0; j < n2; j += 4 * nthread) {
for (i = 0; i < n1; i++) { for (i = 0; i < n1; i++) {
@ -1413,7 +1262,7 @@ void* ddxt2d_th(void* p)
a[i][n0] = t[i]; a[i][n0] = t[i];
} }
} }
return (void*) 0; return (void *)0;
} }
#endif /* USE_FFT2D_THREADS */ #endif /* USE_FFT2D_THREADS */