Mercurial > hg > Game > Cerium

changeset 1731:dc7dd1eaf6de draft

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
add file
author Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
date Thu, 31 Oct 2013 21:07:30 +0900
parents 2e8931baef7f
children 6acaf7f6ad97
files example/fft/gpu/bitReverse.cc example/fft/gpu/bitReverse.h example/fft/gpu/bitReverse.o example/fft/gpu/butterfly.cc example/fft/gpu/butterfly.h example/fft/gpu/butterfly.o example/fft/gpu/gpu_task_init.o example/fft/gpu/highPassFilter.cc example/fft/gpu/highPassFilter.h example/fft/gpu/highPassFilter.o example/fft/gpu/norm.cc example/fft/gpu/norm.h example/fft/gpu/norm.o example/fft/gpu/spinFact.cc example/fft/gpu/spinFact.h example/fft/gpu/spinFact.o example/fft/gpu/task_init.cc example/fft/gpu/task_init.o example/fft/gpu/transpose.cc example/fft/gpu/transpose.h example/fft/gpu/transpose.o example/fft/ppe/task_init.cc
diffstat 22 files changed, 336 insertions(+), 0 deletions(-) [+]
line wrap: on
line diff
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/bitReverse.cc	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,34 @@
+#include "bitReverse.h"
+#ifdef __APPLE__
+#include <OpenCL/opencl.h>
+#else
+#include <CL/cl.h>
+#endif
+#include "Func.h"
+
+SchedDefineTask1(bitReverse,bitReverse);
+
+static int
+bitReverse(SchedTask* s, void* rbuf, void* wbuf)
+{
+    cl_float2* src = (cl_float2*)s->get_inputAddr(0);
+    cl_float2* dst = (cl_float2*)s->get_outputAddr(0);
+    
+    unsigned long gid = s->x; // (unsigned long)s->get_param(0);
+    unsigned long nid = s->y; // (unsigned long)s->get_param(1);
+    
+    unsigned long m = (unsigned long)s->get_param(3);
+    unsigned long n = (unsigned long)s->get_param(4);
+    unsigned int j = gid;
+
+    j = (j & 0x55555555) << 1 | (j & 0xAAAAAAAA) >> 1;
+    j = (j & 0x33333333) << 2 | (j & 0xCCCCCCCC) >> 2;
+    j = (j & 0x0F0F0F0F) << 4 | (j & 0xF0F0F0F0) >> 4;
+    j = (j & 0x00FF00FF) << 8 | (j & 0xFF00FF00) >> 8;
+    j = (j & 0x0000FFFF) << 16 | (j & 0xFFFF0000) >> 16;
+
+    j >>= (32-m);
+
+    dst[nid*n+j] = src[nid*n+gid];
+    return 0;
+}
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/bitReverse.h	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,8 @@
+#ifndef INCLUDED_TASK_BITREVERSE
+#define INCLUDED_TASK_BITREVERSE
+
+#ifndef INCLUDED_SCHED_TASK
+#include "SchedTask.h"
+#endif
+
+#endif
Binary file example/fft/gpu/bitReverse.o has changed
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/butterfly.cc	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,66 @@
+#include "butterfly.h"
+#ifdef __APPLE__
+#include <OpenCL/opencl.h>
+#else
+#include <CL/cl.h>
+#endif
+#include "Func.h"
+SchedDefineTask1(butterfly,butterfly);
+
+static int
+butterfly(SchedTask* s,void* rbuf,void* wbuf)
+{
+    cl_float2* x_in = (cl_float2*)s->get_inputAddr(0);
+    cl_float2* w = (cl_float2*)s->get_inputAddr(1);
+    
+    cl_float2* x_out = (cl_float2*)s->get_outputAddr(0);
+
+    unsigned long gid = s->x; // (unsigned long)s->get_param(0);
+    unsigned long nid = s->y; // (unsigned long)s->get_param(1);
+    
+    long n = (long)s->get_param(3);
+    unsigned long direction_flag = (unsigned long)s->get_param(4);
+    long iter = (long)s->get_param(5);
+
+    int butterflySize = 1 << (iter-1);
+    int butterflyGrpDist = 1 << iter;
+    int butterflyGrpNum = n >> iter;
+    int butterflyGrpBase = (gid >> (iter-1))*(butterflyGrpDist);
+    int butterflyGrpOffset = gid & (butterflySize-1);
+
+    int a = nid * n + butterflyGrpBase + butterflyGrpOffset;
+    int b = a + butterflySize;
+    
+    int l = butterflyGrpNum * butterflyGrpOffset;
+
+    cl_float2 xa, xb, xbxx, xbyy, wab, wayx, wbyx, resa, resb;
+
+    xa = x_in[a];
+    xb = x_in[b];
+    xbxx.x = xbxx.y = xb.x;
+    xbyy.x = xbyy.y = xb.y;
+    
+    wab.x = w[l].x;
+    if(direction_flag == 0x80000000) {
+        wab.y = -w[l].y;
+    } else {
+        wab.y = w[l].y;
+    }
+
+    wayx.x = -wab.y;
+    wayx.y = wab.x;
+
+    wbyx.x = wab.y;
+    wbyx.y = -wab.x;
+
+    resa.x = xa.x + xbxx.x*wab.x + xbyy.x*wayx.x;
+    resa.y = xa.y + xbxx.y*wab.y + xbyy.y*wayx.y;
+
+    resb.x = xa.x - xbxx.x*wab.x + xbyy.x*wbyx.x;
+    resb.y = xa.y - xbxx.y*wab.y + xbyy.y*wbyx.y;
+
+    x_out[a] = resa;
+    x_out[b] = resb;
+
+    return 0;
+}
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/butterfly.h	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,8 @@
+#ifndef INCUDED_TASK_BUTTERFLY
+#define INCUDED_TASK_BUTTERFLY
+
+#ifndef INCLUDED_SCHED_TASK
+#include "SchedTask.h"
+#endif
+
+#endif
Binary file example/fft/gpu/butterfly.o has changed
Binary file example/fft/gpu/gpu_task_init.o has changed
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/highPassFilter.cc	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,55 @@
+#include "highPassFilter.h"
+#ifdef __APPLE__
+#include <OpenCL/opencl.h>
+#else
+#include <CL/cl.h>
+#endif
+#include "Func.h"
+SchedDefineTask1(highPassFilter,highPassFilter);
+
+static int
+highPassFilter(SchedTask* s,void* rbuf,void* wbuf)
+{
+    cl_float2* in = (cl_float2*)s->get_inputAddr(0);
+
+    cl_float2* image = (cl_float2*)s->get_outputAddr(0);
+
+    unsigned long xgid = s->x; // (unsigned long)s->get_param(0);
+    unsigned long ygid = s->y; // (unsigned long)s->get_param(1);
+
+    long n = (long)s->get_param(3);
+    long radius = (long)s->get_param(4);
+
+    cl_int2 n_2;
+    n_2.x = n_2.y = n>>1;
+    
+    cl_int2 mask;
+    mask.x = mask.y = n-1;
+
+    cl_int2 gid;
+    gid.x = (xgid + n_2.x) & mask.x;
+    gid.y = (ygid + n_2.y) & mask.y;
+
+    cl_int2 diff;
+    diff.x = n_2.x - gid.x;
+    diff.y = n_2.y - gid.y;
+    
+    cl_int2 diff2;
+    diff2.x = diff.x * diff.x;
+    diff2.y = diff.y * diff.y;
+
+    int dist2 = diff2.x + diff2.y;
+
+    cl_int2 window;
+
+    if (dist2 < radius*radius) {
+        window.x = window.y = (int)0L;
+    } else {
+        window.x = window.y = (int)-1L;
+    }
+
+    image[ygid*n+xgid].x = (float)((int)in[ygid*n+xgid].x & window.x);
+    image[ygid*n+xgid].y = (float)((int)in[ygid*n+xgid].y & window.y);
+
+    return 0;
+}
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/highPassFilter.h	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,8 @@
+#ifndef INCLUDED_TASK_HIGHPASSFILTER
+#define INCLUDED_TASK_HIGHPASSFILTER
+
+#ifndef INCLUDED_SCHED_TASK
+#include "SchedTask.h"
+#endif
+
+#endif
Binary file example/fft/gpu/highPassFilter.o has changed
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/norm.cc	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,26 @@
+#include "norm.h"
+#ifdef __APPLE__
+#include <OpenCL/opencl.h>
+#else
+#include <CL/cl.h>
+#endif
+#include "Func.h"
+SchedDefineTask1(norm,norm);
+
+static int
+norm(SchedTask* s, void* rbuf,void* wbuf)
+{
+    cl_float2* in_x = (cl_float2*)s->get_inputAddr(0);
+
+    cl_float2* out_x = (cl_float2*)s->get_outputAddr(0);
+
+    unsigned long gid = s->x; // (unsigned long)s->get_param(0);
+    unsigned long nid = s->y; //(unsigned long)s->get_param(1);
+
+    long n = (long)s->get_param(3);
+
+    out_x[nid*n+gid].x = in_x[nid*n+gid].x / (float)n;
+    out_x[nid*n+gid].y = in_x[nid*n+gid].y / (float)n;
+
+    return 0;
+}
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/norm.h	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,8 @@
+#ifndef INCLUDED_TASK_NORM
+#define INCLUDED_TASK_NORM
+
+#ifndef INCLUDED_SCHED_TASK
+#include "SchedTask.h"
+#endif
+
+#endif
Binary file example/fft/gpu/norm.o has changed
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/spinFact.cc	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,32 @@
+#include <math.h>
+#include "spinFact.h"
+#ifdef __APPLE__
+#include <OpenCL/opencl.h>
+#else
+#include <CL/cl.h>
+#endif
+#include "Func.h"
+
+#define PI 3.14159265358979323846
+#define PI_2 1.57079632679489661923
+
+SchedDefineTask1(spinFact,spinFact);
+
+static int
+spinFact(SchedTask* s,void* rbuf,void* wbuf)
+{
+    cl_float2* w = (cl_float2*)s->get_outputAddr(0);
+
+    unsigned long i = s->x; // (unsigned long)s->get_param(0);
+    
+    long n = (long)s->get_param(3);
+
+    cl_float2 angle;
+    angle.x = (float)(2*i*PI/(float)n);
+    angle.y = (float)((2*i*PI/(float)n) + PI_2);
+
+    w[i].x = cos(angle.x);
+    w[i].y = cos(angle.y);
+
+    return 0;
+}
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/spinFact.h	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,8 @@
+#ifndef INCLUDED_TASK_SPINFACT
+#define INCLUDED_TASK_SPINFACT
+
+#ifndef INCLUDED_SCHED_TASK
+#include "SchedTask.h"
+#endif
+
+#endif
Binary file example/fft/gpu/spinFact.o has changed
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/task_init.cc	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,20 @@
+#include "Func.h"
+#include "Scheduler.h"
+
+SchedExternTask(spinFact);
+SchedExternTask(bitReverse);
+SchedExternTask(norm);
+SchedExternTask(butterfly);
+SchedExternTask(transpose);
+SchedExternTask(highPassFilter);
+
+void
+task_init(void)
+{
+    SchedRegisterTask(SPIN_FACT,spinFact);
+    SchedRegisterTask(NORMALIZATION, norm);
+    SchedRegisterTask(BIT_REVERSE, bitReverse);
+    SchedRegisterTask(BUTTERFLY, butterfly);
+    SchedRegisterTask(TRANSPOSE, transpose);
+    SchedRegisterTask(HIGH_PASS_FILTER, highPassFilter);
+}
Binary file example/fft/gpu/task_init.o has changed
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/transpose.cc	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,29 @@
+#include "transpose.h"
+#ifdef __APPLE__
+#include <OpenCL/opencl.h>
+#else
+#include <CL/cl.h>
+#endif
+#include "Func.h"
+
+SchedDefineTask1(transpose,transpose);
+
+static int
+transpose(SchedTask* s,void* rbuf,void* wbuf)
+{
+    cl_float2* src = (cl_float2*)s->get_inputAddr(0);
+
+    cl_float2* dst = (cl_float2*)s->get_outputAddr(0);
+
+    unsigned long xgid = s->x; // (unsigned long)s->get_param(0);
+    unsigned long ygid = s->y; // (unsigned long)s->get_param(1);
+
+    long n = (long)s->get_param(3);
+
+    unsigned int iid = ygid * n + xgid;
+    unsigned int oid = xgid * n + ygid;
+
+    dst[oid] = src[iid];
+
+    return 0;
+}
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/gpu/transpose.h	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,8 @@
+#ifndef INCLUDED_TASK_TRANSPOSE
+#define INCLUDED_TASK_TRANSPOSE
+
+#ifndef INCLUDED_SCHED_TASK
+#include "SchedTask.h"
+#endif
+
+#endif
Binary file example/fft/gpu/transpose.o has changed
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/example/fft/ppe/task_init.cc	Thu Oct 31 21:07:30 2013 +0900
@@ -0,0 +1,26 @@
+#include "Func.h"
+#include "Scheduler.h"
+
+SchedExternTask(spinFact);
+SchedExternTask(bitReverse);
+SchedExternTask(norm);
+SchedExternTask(butterfly);
+SchedExternTask(transpose);
+SchedExternTask(highPassFilter);
+
+void
+task_init(void)
+{
+    SchedRegisterTask(SPIN_FACT,spinFact);
+    SchedRegisterTask(NORMALIZATION, norm);
+    SchedRegisterTask(BIT_REVERSE, bitReverse);
+    SchedRegisterTask(BUTTERFLY, butterfly);
+    SchedRegisterTask(TRANSPOSE, transpose);
+    SchedRegisterTask(HIGH_PASS_FILTER, highPassFilter);
+}
+
+void
+gpu_task_init()
+{
+
+}