From 1edcf73a73d2007afc61289245763f5cf0c29e10 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 04 Dec 2014 07:20:29 +0000
Subject: [PATCH] Detection good, split up col images
---
src/opencl.c | 108 ++++++++++++++++++++++++++++++++++++++++++++++++-----
1 files changed, 97 insertions(+), 11 deletions(-)
diff --git a/src/opencl.c b/src/opencl.c
index 8f9edd3..981067a 100644
--- a/src/opencl.c
+++ b/src/opencl.c
@@ -1,18 +1,26 @@
#ifdef GPU
-#include "opencl.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <unistd.h>
+#ifdef CLBLAS
+#include <clBLAS.h>
+#endif
+
+#include "opencl.h"
+#include "utils.h"
+#include "activations.h"
cl_info cl = {0};
void check_error(cl_info info)
{
+ // clFinish(cl.queue);
if (info.error != CL_SUCCESS) {
printf("\n Error number %d", info.error);
+ abort();
exit(1);
}
}
@@ -27,13 +35,62 @@
// Fetch the Platform and Device IDs; we only want one.
cl_device_id devices[MAX_DEVICES];
info.error=clGetPlatformIDs(1, &info.platform, &num_platforms);
+
+ printf("=== %d OpenCL platform(s) found: ===\n", num_platforms);
+ char buffer[10240];
+ clGetPlatformInfo(info.platform, CL_PLATFORM_PROFILE, 10240, buffer, NULL);
+ printf(" PROFILE = %s\n", buffer);
+ clGetPlatformInfo(info.platform, CL_PLATFORM_VERSION, 10240, buffer, NULL);
+ printf(" VERSION = %s\n", buffer);
+ clGetPlatformInfo(info.platform, CL_PLATFORM_NAME, 10240, buffer, NULL);
+ printf(" NAME = %s\n", buffer);
+ clGetPlatformInfo(info.platform, CL_PLATFORM_VENDOR, 10240, buffer, NULL);
+ printf(" VENDOR = %s\n", buffer);
+ clGetPlatformInfo(info.platform, CL_PLATFORM_EXTENSIONS, 10240, buffer, NULL);
+ printf(" EXTENSIONS = %s\n", buffer);
+
check_error(info);
info.error=clGetDeviceIDs(info.platform, CL_DEVICE_TYPE_ALL, MAX_DEVICES, devices, &num_devices);
if(num_devices > MAX_DEVICES) num_devices = MAX_DEVICES;
+ printf("=== %d OpenCL device(s) found on platform:\n", num_devices);
+ int i;
+ for (i=0; i<num_devices; i++)
+ {
+ char buffer[10240];
+ cl_uint buf_uint;
+ cl_ulong buf_ulong;
+ printf(" -- %d --\n", i);
+ clGetDeviceInfo(devices[i], CL_DEVICE_NAME, sizeof(buffer), buffer, NULL);
+ printf(" DEVICE_NAME = %s\n", buffer);
+ clGetDeviceInfo(devices[i], CL_DEVICE_VENDOR, sizeof(buffer), buffer, NULL);
+ printf(" DEVICE_VENDOR = %s\n", buffer);
+ clGetDeviceInfo(devices[i], CL_DEVICE_VERSION, sizeof(buffer), buffer, NULL);
+ printf(" DEVICE_VERSION = %s\n", buffer);
+ clGetDeviceInfo(devices[i], CL_DRIVER_VERSION, sizeof(buffer), buffer, NULL);
+ printf(" DRIVER_VERSION = %s\n", buffer);
+ clGetDeviceInfo(devices[i], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(buf_uint), &buf_uint, NULL);
+ printf(" DEVICE_MAX_COMPUTE_UNITS = %u\n", (unsigned int)buf_uint);
+ clGetDeviceInfo(devices[i], CL_DEVICE_MAX_CLOCK_FREQUENCY, sizeof(buf_uint), &buf_uint, NULL);
+ printf(" DEVICE_MAX_CLOCK_FREQUENCY = %u\n", (unsigned int)buf_uint);
+ clGetDeviceInfo(devices[i], CL_DEVICE_GLOBAL_MEM_SIZE, sizeof(buf_ulong), &buf_ulong, NULL);
+ printf(" DEVICE_GLOBAL_MEM_SIZE = %llu\n", (unsigned long long)buf_ulong);
+ clGetDeviceInfo(devices[i], CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(buf_ulong), &buf_ulong, NULL);
+ printf(" DEVICE_MAX_MEM_ALLOC_SIZE = %llu\n", (unsigned long long)buf_ulong);
+ clGetDeviceInfo(devices[i], CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(buf_ulong), &buf_ulong, NULL);
+ printf(" DEVICE_MAX_WORK_GROUP_SIZE = %llu\n", (unsigned long long)buf_ulong);
+ cl_uint items;
+ clGetDeviceInfo( devices[i], CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS, sizeof(cl_uint),
+ &items, NULL);
+ printf(" DEVICE_MAX_WORK_ITEM_DIMENSIONS = %u\n", (unsigned int)items);
+ size_t workitem_size[10];
+ clGetDeviceInfo(devices[i], CL_DEVICE_MAX_WORK_ITEM_SIZES, 10*sizeof(workitem_size), workitem_size, NULL);
+ printf(" DEVICE_MAX_WORK_ITEM_SIZES = %u / %u / %u \n", (unsigned int)workitem_size[0], (unsigned int)workitem_size[1], (unsigned int)workitem_size[2]);
+
+ }
int index = getpid()%num_devices;
+ index = 0;
printf("%d rand, %d devices, %d index\n", getpid(), num_devices, index);
- //info.device = devices[index];
- info.device = devices[1];
+ info.device = devices[index];
fprintf(stderr, "Found %d device(s)\n", num_devices);
check_error(info);
@@ -45,6 +102,10 @@
check_error(info);
info.queue = clCreateCommandQueue(info.context, info.device, 0, &info.error);
check_error(info);
+ #ifdef CLBLAS
+ info.error = clblasSetup();
+ #endif
+ check_error(info);
info.initialized = 1;
return info;
}
@@ -52,21 +113,22 @@
cl_program cl_fprog(char *filename, char *options, cl_info info)
{
size_t srcsize;
- char src[8192];
- memset(src, 0, 8192);
+ char src[64*1024];
+ memset(src, 0, 64*1024);
FILE *fil=fopen(filename,"r");
+ if(fil == 0) file_error(filename);
srcsize=fread(src, sizeof src, 1, fil);
fclose(fil);
const char *srcptr[]={src};
// Submit the source code of the example kernel to OpenCL
cl_program prog=clCreateProgramWithSource(info.context,1, srcptr, &srcsize, &info.error);
check_error(info);
- char build_c[4096];
+ char build_c[1024*64];
// and compile it (after this we could extract the compiled version)
info.error=clBuildProgram(prog, 0, 0, options, 0, 0);
if ( info.error != CL_SUCCESS ) {
fprintf(stderr, "Error Building Program: %d\n", info.error);
- clGetProgramBuildInfo( prog, info.device, CL_PROGRAM_BUILD_LOG, 4096, build_c, 0);
+ clGetProgramBuildInfo( prog, info.device, CL_PROGRAM_BUILD_LOG, 1024*64, build_c, 0);
fprintf(stderr, "Build Log for %s program:\n%s\n", filename, build_c);
}
check_error(info);
@@ -76,6 +138,7 @@
void cl_setup()
{
if(!cl.initialized){
+ printf("initializing\n");
cl = cl_init();
}
}
@@ -92,21 +155,31 @@
void cl_read_array(cl_mem mem, float *x, int n)
{
cl_setup();
- clEnqueueReadBuffer(cl.queue, mem, CL_TRUE, 0, sizeof(float)*n,x,0,0,0);
+ cl.error = clEnqueueReadBuffer(cl.queue, mem, CL_TRUE, 0, sizeof(float)*n,x,0,0,0);
check_error(cl);
}
+float cl_checksum(cl_mem mem, int n)
+{
+
+ float *x = calloc(n, sizeof(float));
+ cl_read_array(mem, x, n);
+ float sum = sum_array(x, n);
+ free(x);
+ return sum;
+}
+
void cl_write_array(cl_mem mem, float *x, int n)
{
cl_setup();
- clEnqueueWriteBuffer(cl.queue, mem, CL_TRUE, 0,sizeof(float)*n,x,0,0,0);
+ cl.error = clEnqueueWriteBuffer(cl.queue, mem, CL_TRUE, 0,sizeof(float)*n,x,0,0,0);
check_error(cl);
}
void cl_copy_array(cl_mem src, cl_mem dst, int n)
{
cl_setup();
- clEnqueueCopyBuffer(cl.queue, src, dst, 0, 0, sizeof(float)*n,0,0,0);
+ cl.error = clEnqueueCopyBuffer(cl.queue, src, dst, 0, 0, sizeof(float)*n,0,0,0);
check_error(cl);
}
@@ -115,10 +188,12 @@
cl_buffer_region r;
r.origin = offset*sizeof(float);
r.size = size*sizeof(float);
- cl_mem sub = clCreateSubBuffer(src, CL_MEM_USE_HOST_PTR, CL_BUFFER_CREATE_TYPE_REGION, &r, 0);
+ cl_mem sub = clCreateSubBuffer(src, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &r, &cl.error);
+ check_error(cl);
return sub;
}
+
cl_mem cl_make_array(float *x, int n)
{
cl_setup();
@@ -126,6 +201,17 @@
CL_MEM_READ_WRITE|CL_MEM_COPY_HOST_PTR,
sizeof(float)*n, x, &cl.error);
check_error(cl);
+ activate_array_ongpu(mem, n, LINEAR);
+ return mem;
+}
+
+cl_mem cl_make_int_array(int *x, int n)
+{
+ cl_setup();
+ cl_mem mem = clCreateBuffer(cl.context,
+ CL_MEM_READ_WRITE|CL_MEM_COPY_HOST_PTR,
+ sizeof(int)*n, x, &cl.error);
+ check_error(cl);
return mem;
}
--
Gitblit v1.10.0