使用 GtkGLArea 的 OpenCL / OpenGL Interop
OpenCL / OpenGL Interop using GtkGLArea
我知道我似乎只在需要时才去拜访，但我还有一个问题。
在我的 的基础上，我成功地 运行 了一个基本的 OpenGL 程序。我现在正在添加 OpenCL 互操作和一个简单的 CL 脚本，每次渲染时都会将三角形缩小一点。
我得到的只是一个空白屏幕。注释掉获取和释放 GL objects 的部分允许渲染像以前一样工作。只是未能集成 OpenCL 组件。
opencl.h 只是一个带有辅助函数的 header，我用它来注册所有 CL object 并在最后通过单个函数调用释放它们。
编辑：2015 年 6 月 12 日
取消注释显示更多信息的 /*fprintf(stderr, "ERROR: " x " failed %d\n", cl_stat);*/ 行：
ERROR: Set cl kernel arg failed -38
ERROR: Acquiring GL objects failed -5
根据 this list，OpenCL 错误代码 -38 表示它是无效内存 object，然后在尝试 re-acquire GL objects 时出现资源不足错误.
这里是main.c
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <glib.h>
#include <gdk/gdkx.h>
#include <epoxy/glx.h>
#include <epoxy/gl.h>
#include <gtk/gtk.h>
#include <CL/opencl.h>
#include "cl_utils.h"
#define IGNORE_VAR(type, identifier) \
{ \
type IGNORED_VARIABLE_abcd = identifier; \
identifier = IGNORED_VARIABLE_abcd; \
}
#define CL_ASSERT(x) \
/*if(print_info) \
printf(x "...\n"); */\
if(cl_stat != CL_SUCCESS) \
{ \
/*fprintf(stderr, "ERROR: " x " failed %d\n", cl_stat);*/ \
goto exception; \
}
const GLchar *vert_src = "\n" \
"#version 330 \n" \
"#extension GL_ARB_explicit_attrib_location: enable \n" \
" \n" \
"layout(location = 0) in vec2 in_position; \n" \
" \n" \
"void main() \n" \
"{ \n" \
" gl_Position = vec4(in_position, 0.0, 1.0); \n" \
"} \n";
const GLchar *frag_src = "\n" \
"void main (void) \n" \
"{ \n" \
" gl_FragColor = vec4(1.0, 1.0, 1.0, 1.0); \n" \
"} \n";
const char *cl_src = "\n" \
"typedef struct Point{ \n" \
" float x; \n" \
" float y; \n" \
"} Point; \n" \
" \n" \
"__kernel void cl_func(__global Point* point) \n" \
"{ \n" \
" const int i = get_global_id(0); \n" \
" const float d = 0.99; \n" \
" \n" \
" if(i>=3) \n" \
" return; \n" \
" \n" \
" point[i].x = point[i].x * d; \n" \
" point[i].y = point[i].y * d; \n" \
"} \n";
struct cl
{
clu_object_stack* stack;
cl_platform_id* platform;
cl_uint num_platforms;
cl_device_id* device;
cl_uint num_devices;
cl_context context;
cl_context_properties properties[7];
cl_command_queue queue;
cl_program program;
cl_kernel kernel;
cl_mem buffer;
} cl;
struct cl cl;
GLuint gl_vao, gl_buffer, gl_program;
cl_int init_cl(GtkGLArea *area)
{
cl_int cl_stat;
cl.stack = cluCreateObjectStack(44);
cl_stat = clGetPlatformIDs(0, NULL, &cl.num_platforms);
CL_ASSERT("Got number of platforms");
cl.platform = malloc((size_t)cl.num_platforms * sizeof(cl_platform_id));
cl_stat = clGetPlatformIDs(cl.num_platforms, cl.platform, NULL);
CL_ASSERT("Got platforms");
cl_stat = clGetDeviceIDs(cl.platform[0], CL_DEVICE_TYPE_GPU, 0, NULL, &cl.num_devices);
CL_ASSERT("Got number of devices");
printf("Number of GPU devices: %d\n", cl.num_devices);
if(cl.num_devices == 0)
{
fprintf(stderr, "Num devices cannot be 0\n");
goto exception;
}
cl.device = malloc((size_t)cl.num_devices * sizeof(cl_device_id));
cl_stat = clGetDeviceIDs(cl.platform[0], CL_DEVICE_TYPE_GPU, cl.num_devices, cl.device, NULL);
CL_ASSERT("Got devices");
if(cl.device == NULL)
{
fprintf(stderr, "Devices list is NULL\n");
goto exception;
}
gtk_gl_area_make_current (area);
cl.properties[0] = CL_GL_CONTEXT_KHR;
cl.properties[1] = (cl_context_properties) glXGetCurrentContext();
cl.properties[2] = CL_GLX_DISPLAY_KHR;
cl.properties[3] = (cl_context_properties) glXGetCurrentDisplay();
cl.properties[4] = CL_CONTEXT_PLATFORM;
cl.properties[5] = (cl_context_properties) cl.platform[0];
cl.properties[6] = 0;
cl.context = cluCreateContext(cl.stack, cl.properties, cl.num_devices, cl.device, NULL, NULL, &cl_stat);
CL_ASSERT("Created cl context");
cl.queue = cluCreateCommandQueue(cl.stack, cl.context, cl.device[0], 0, &cl_stat);
CL_ASSERT("Created command queue");
cl.buffer = cluCreateFromGLBuffer(cl.stack, cl.context, CL_MEM_WRITE_ONLY, gl_buffer, NULL);
CL_ASSERT("Created cl memory object from gl buffer");
cl.program = cluCreateProgramWithSource(cl.stack, cl.context, 1, (const char **)&cl_src, NULL, &cl_stat);
CL_ASSERT("Created cl program object");
cl_stat = clBuildProgram(cl.program, cl.num_devices, cl.device, NULL, NULL, NULL);
if(cl_stat != CL_SUCCESS)
{
size_t ret_size;
clGetProgramBuildInfo(cl.program, cl.device[0], CL_PROGRAM_BUILD_LOG, 0, NULL, &ret_size);
char e_str[ret_size];
clGetProgramBuildInfo(cl.program, cl.device[0], CL_PROGRAM_BUILD_LOG, ret_size, e_str, &ret_size);
printf("%s\n", e_str);
}
CL_ASSERT("Built cl program object");
cl.kernel = cluCreateKernel(cl.stack, cl.program, "cl_func", &cl_stat);
CL_ASSERT("Created cl kernel object");
return 0;
exception:
return 1;
}
static gboolean realise(GtkGLArea *area, GdkGLContext *context);
static gboolean render(GtkGLArea *area, GdkGLContext *context);
int main(int argc, char** argv)
{
gtk_init(&argc, &argv);
GtkWidget *window = gtk_window_new(GTK_WINDOW_TOPLEVEL),
*gl_area = gtk_gl_area_new();
g_signal_connect(window, "delete-event", G_CALLBACK(gtk_main_quit), NULL);
g_signal_connect(gl_area, "realize", G_CALLBACK(realise), NULL);
g_signal_connect(gl_area, "render", G_CALLBACK(render), NULL);
gtk_container_add(GTK_CONTAINER(window), gl_area);
gtk_widget_show_all(window);
gtk_main();
cluFreeObjectStack(cl.stack);
free(cl.platform);
free(cl.device);
return 0;
}
static gboolean realise(GtkGLArea *area, GdkGLContext *context)
{
IGNORE_VAR(GdkGLContext*, context);
gtk_gl_area_make_current(GTK_GL_AREA(area));
if (gtk_gl_area_get_error (GTK_GL_AREA(area)) != NULL)
{
printf("Failed to initialiize buffers\n");
return FALSE;
}
GLfloat verts[] =
{
+0.0f, +1.0f,
-1.0f, -1.0f,
+1.0f, -1.0f,
};
GLuint frag_shader, vert_shader;
frag_shader = glCreateShader(GL_FRAGMENT_SHADER);
vert_shader = glCreateShader(GL_VERTEX_SHADER);
glShaderSource(frag_shader, 1, &frag_src, NULL);
glShaderSource(vert_shader, 1, &vert_src, NULL);
glCompileShader(frag_shader);
glCompileShader(vert_shader);
gl_program = glCreateProgram();
glAttachShader(gl_program, frag_shader);
glAttachShader(gl_program, vert_shader);
glLinkProgram(gl_program);
glGenBuffers(1, &gl_buffer);
glBindBuffer(GL_ARRAY_BUFFER, gl_buffer);
glBufferData(GL_ARRAY_BUFFER, sizeof(verts), verts, GL_DYNAMIC_DRAW);
glGenVertexArrays(1, &gl_vao);
glBindVertexArray(gl_vao);
glEnableVertexAttribArray(0);
glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 0, (void*)0);
glBindVertexArray(0);
//glDeleteBuffers(1, &gl_buffer);
if(init_cl(area))
return FALSE;
return TRUE;
}
static gboolean render(GtkGLArea *area, GdkGLContext *context)
{
IGNORE_VAR(GdkGLContext*, context);
IGNORE_VAR(GtkGLArea*, area);
cl_int cl_stat;
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
glClearColor(0.0, 0.0, 0.0, 1.0);
glUseProgram(gl_program);
glBindVertexArray(gl_vao);
glDrawArrays(GL_TRIANGLES, 0, 3);
glBindVertexArray (0);
glUseProgram (0);
glFinish();
cl_stat = clEnqueueAcquireGLObjects(cl.queue, 1, &cl.buffer, 0, NULL, NULL);
CL_ASSERT("Acquiring GL objects");
cl_stat = clSetKernelArg(cl.kernel, 0, sizeof(cl_mem), (const void*)cl.buffer);
CL_ASSERT("Set cl kernel arg");
//size_t g_sz[1] = { 32 };
//cl_stat = clEnqueueNDRangeKernel(cl.queue, cl.kernel, 1, NULL, g_sz, NULL, 0, NULL, NULL);
cl_stat = clEnqueueTask(cl.queue, cl.kernel, 0, NULL, NULL);
CL_ASSERT("Executing cl kernel");
cl_stat = clEnqueueReleaseGLObjects(cl.queue, 1, &cl.buffer, 0, NULL, NULL);
CL_ASSERT("Releasing GL buffer");
clFinish(cl.queue);
return TRUE;
exception:
return FALSE;
}
SConstruct
import os
CFLAGS = '--std=c11 -g -Wall'
ENV = {'PATH':os.environ['PATH']}
env = Environment(CFLAGS=CFLAGS, ENV=ENV)
if os.name is 'posix':
env['ENV']['TERM'] = os.environ['TERM']
env.ParseConfig('pkg-config --cflags --libs gtk+-3.0')
env.Append(LIBS = ['epoxy', 'GL', 'OpenCL'])
env.Program(target='gl', source=['main.c', 'cl_utils.c'])
# vim: set filetype=python:
cl_utils.c
#include <stdlib.h>
#include <stdio.h>
#include "cl_utils.h"
#define FUNC __func__
#define CL_ASSERT(object, label) \
if(!object) \
{ \
fprintf(stderr, "CL_ERROR: failed to create CL %s object: %d\n", #object, *cl_stat); \
goto label; \
} \
else \
{ \
printf("CL_INFO: created cl_%s object\n", #object); \
}
#define PLATFORM_INFO(_) \
_(PROFILE) \
_(VERSION) \
_(NAME) \
_(VENDOR) \
_(EXTENSIONS)
const char* space = " ";
#define TO_STRING(TYPE, ENUM, ...) [ENUM] = #TYPE,
const char* CLU_TYPE_STRING[] = { CLU_TABLE(TO_STRING) };
#undef TO_STRING
#define TO_STRING(x) [CL_PLATFORM_##x] = #x,
const char* platform_info_string[] = { PLATFORM_INFO(TO_STRING) };
#undef TO_STRING
const char* clGetPlatformInfoString(cl_platform_info platform_info)
{
return platform_info_string[platform_info];
}
cl_int infoPlatforms(cl_platform_id* platforms, cl_int num_platforms, cl_platform_info* params, cl_int num_params)
{
cl_int cl_stat;
size_t buffer_size = 10, buffer_ret_size = 0;
char* buffer = (char*)malloc(buffer_size);
for(cl_int i = 0; i < num_platforms; i++)
{
for(cl_int j = 0; j < num_params; j++)
{
cl_stat = clGetPlatformInfo(platforms[i], params[j], 0, NULL, &buffer_ret_size);
if(cl_stat != CL_SUCCESS)
{
fprintf(stderr, "ERROR: clGetPlatformInfo failed\n");
goto end;
}
if(buffer_ret_size > buffer_size)
{
void* tmp = NULL;
buffer_size = buffer_ret_size;
tmp = realloc(buffer, buffer_size);
if(tmp == NULL)
{
fprintf(stderr, "ERROR: Could not realloc memory\n");
perror("ERROR: ");
goto end;
}
else
buffer = (char*)tmp;
}
cl_stat = clGetPlatformInfo(platforms[i], params[j], buffer_size, buffer, &buffer_ret_size);
if(cl_stat != CL_SUCCESS)
{
fprintf(stderr, "ERROR: clGetPlatformInfo failed\n");
goto end;
}
printf("%s: %s\n", clGetPlatformInfoString(params[j]), buffer);
}
printf("\n");
}
end:
if(cl_stat != CL_SUCCESS)
{
printf("SENT TO ERROR HANDLER!\n");
printf("CL_ERROR: %d\n", cl_stat);
}
free(buffer);
if(cl_stat != CL_SUCCESS)
return 1;
return 0;
}
clu_object_stack* cluCreateObjectStack(int init_len)
{
clu_object_stack* stack = (clu_object_stack*)malloc(sizeof(clu_object_stack));
stack->length = init_len;
stack->num_objects = 0;
stack->list = (clu_object*)malloc(sizeof(clu_object) * init_len);
return stack;
}
void cluFreeObjectStack(clu_object_stack* stack)
{
cl_int ret = CL_SUCCESS;
printf("Freeing stack: %p\n", stack);
int i;
for(i=stack->num_objects-1; i>=0; i--)
{
switch(stack->list[i].type)
{
#define TO_FREE(cl_type, ENUM, element, cluRelease) \
case ENUM: \
ret = cluRelease(stack->list[i].element); \
/*printf("Releasing " #cl_type "\n");*/ \
break;
CLU_TABLE(TO_FREE)
#undef TO_FREE
default:
printf("ERROR: Invalid or unsupported object type\n");
break;
}
if(ret != CL_SUCCESS)
{
printf("Failed to release %s\n", CLU_TYPE_STRING[stack->list[i].type]);
}
}
free(stack->list);
free(stack);
}
int cluAssignToObjectGroup(struct clu_object_stack* stack, void* obj, cl_type type)
{
if(stack->num_objects >= stack->length)
{
printf("Stack Error\n");
return -1;
}
stack->list[stack->num_objects].type = type;
switch(type)
{
#define TO_ASSIGN(cl_type, ENUM, element, ...) \
case ENUM: \
stack->list[stack->num_objects].element=*(cl_type*)obj; \
/*printf("Assigning " #cl_type "\n");*/ \
break;
CLU_TABLE(TO_ASSIGN)
#undef TO_ASSIGN
default:
printf("cluAssignToObjectGroup Failed\n");
break;
}
stack->num_objects++;
return 0;
}
cl_context cluCreateContext(
clu_object_stack *stack,
cl_context_properties *properties,
cl_uint num_devices,
const cl_device_id *devices,
pfn_notify func,
void *user_data,
cl_int* cl_stat)
{
cl_context context= clCreateContext(properties, num_devices, devices, func, user_data, cl_stat);
if(context == 0)
printf("Created ZERO value cl context\n");
CL_ASSERT(context, error_ret);
cluAssignToObjectGroup(stack, &context, CLU_CONTEXT);
error_ret:
if(devices == NULL)
fprintf(stderr, "%2sdevices cannot be NULL\n", space);
if(num_devices == 0)
fprintf(stderr, "%2snum_devices cannot be zero\n", space);
if((func == NULL) && (user_data != NULL))
fprintf(stderr, "%2spfn_notify cannot be NULL when user_data is not NULL\n", space);
return context;
}
cl_command_queue cluCreateCommandQueue(
clu_object_stack *stack,
cl_context context,
cl_device_id device,
cl_command_queue_properties properties,
cl_int *cl_stat)
{
cl_command_queue queue = clCreateCommandQueue(context, device, properties, cl_stat);
CL_ASSERT(queue, error_ret);
cluAssignToObjectGroup(stack, &queue, CLU_COMMAND_QUEUE);
error_ret:
return queue;
}
cl_mem cluCreateBuffer(
clu_object_stack* stack,
cl_context context,
cl_mem_flags flags,
size_t size,
void *host_ptr,
cl_int *cl_stat)
{
cl_mem buffer = clCreateBuffer(context, flags, size, host_ptr, cl_stat);
CL_ASSERT(buffer, error_ret);
cluAssignToObjectGroup(stack, &buffer, CLU_MEM_OBJECT);
error_ret:
return buffer;
}
cl_program cluCreateProgramWithSource(
clu_object_stack* stack,
cl_context context,
cl_uint count,
const char **strings,
const size_t *lengths,
cl_int *cl_stat)
{
cl_program program = clCreateProgramWithSource(context, count, strings, lengths, cl_stat);
CL_ASSERT(program, error_ret);
cluAssignToObjectGroup(stack, &program, CLU_PROGRAM);
error_ret:
return program;
}
cl_kernel cluCreateKernel(
clu_object_stack* stack,
cl_program program,
const char *kernel_name,
cl_int *cl_stat)
{
cl_kernel kernel = clCreateKernel(program, kernel_name, cl_stat);
CL_ASSERT(kernel, error_ret);
cluAssignToObjectGroup(stack, &kernel, CLU_KERNEL);
error_ret:
return kernel;
}
cl_mem cluCreateFromGLBuffer(
clu_object_stack* stack,
cl_context context,
cl_mem_flags flags,
GLuint bufobj,
cl_int* cl_stat)
{
cl_mem gl_buffer = clCreateFromGLBuffer(context, flags, bufobj, cl_stat);
CL_ASSERT(gl_buffer, error_ret);
cluAssignToObjectGroup(stack, &gl_buffer, CLU_MEM_OBJECT);
error_ret:
return gl_buffer;
}
cl_utils.h
#ifndef __CL_UTILS_H__
#define __CL_UTILS_H__
#ifdef __cplusplus
extern "C" {
#endif /* C++ */
#include <stdarg.h>
#if defined(__APPLE__) || defined(MACOSX)
#include <OpenCL/opencl.h>
#else
#include <CL/opencl.h>
#endif
#include <epoxy/gl.h>
#define CLU_TABLE(_) \
_(cl_context, CLU_CONTEXT, context, clReleaseContext) \
_(cl_command_queue, CLU_COMMAND_QUEUE, queue, clReleaseCommandQueue) \
_(cl_mem,CLU_MEM_OBJECT, mem_object, clReleaseMemObject) \
_(cl_program, CLU_PROGRAM, program, clReleaseProgram) \
_(cl_kernel, CLU_KERNEL, kernel, clReleaseKernel)
#define TO_ENUM(cl_type, ENUM, ...) ENUM,
typedef enum cl_type{
CLU_TABLE(TO_ENUM)
} cl_type;
#undef TO_ENUM
typedef void (*pfn_notify)(const char *errinfo, const void *private_info, size_t cb, void *user_data);
//extern const char* CLU_TYPE_STRING[];
typedef struct clu_object clu_object;
struct clu_object{
cl_type type;
union{
cl_context context;
cl_command_queue queue;
cl_mem mem_object;
cl_program program;
cl_kernel kernel;
};
};
typedef struct clu_object_stack{
struct clu_object* list;
int num_objects;
int length;
} clu_object_stack;
const char* clGetPlatformInfoString(cl_platform_info platform_info);
int infoPlatforms(cl_platform_id* platforms, cl_int num_platforms, cl_platform_info* params, cl_int num_params);
clu_object_stack* cluCreateObjectStack(int init_len);
void cluFreeObjectStack(clu_object_stack* stack);
int cluAssignToObjectGroup(struct clu_object_stack* stack, void* obj, cl_type type);
cl_context cluCreateContext(
clu_object_stack *stack,
cl_context_properties *properties,
cl_uint num_devices,
const cl_device_id *devices,
pfn_notify func,
void *user_data,
cl_int* cl_stat);
cl_command_queue cluCreateCommandQueue(
clu_object_stack* stack,
cl_context context,
cl_device_id device,
cl_command_queue_properties properties,
cl_int *cl_stat);
cl_mem cluCreateBuffer(
clu_object_stack* stack,
cl_context context,
cl_mem_flags flags,
size_t size,
void *host_ptr,
cl_int * cl_stat);
cl_program cluCreateProgramWithSource(
clu_object_stack* stack,
cl_context context,
cl_uint count,
const char **strings,
const size_t *lengths,
cl_int *cl_stat);
cl_kernel cluCreateKernel(
clu_object_stack* stack,
cl_program program,
const char *kernel_name,
cl_int *cl_stat);
cl_mem cluCreateFromGLBuffer(
clu_object_stack* stack,
cl_context context,
cl_mem_flags flags,
GLuint bufobj,
cl_int* cl_stat);
#ifdef __cplusplus
}
#endif /* C++ */
#endif /* __CL_UTILS_H__ */
要启用 CL-GL 互操作性，必须按特定顺序进行设置：
1. Create OpenGL context
2. Create OpenCL context
3. Create OpenGL buffers
4. Start OpenGL rendering
一个可能的问题是您在 OpenCL 上下文之前（在调用 init_cl 之前）创建了 OpenGL 缓冲区。
找到问题了。 clSetKernelArg() 的最后一个参数需要一个指向 mem 对象的指针，我忘了在前面添加 & 运算符。
所以这个：
cl_stat = clSetKernelArg(cl.kernel, 0, sizeof(cl_mem), (const void*)cl.buffer);
变成这样：
cl_stat = clSetKernelArg(cl.kernel, 0, sizeof(cl_mem), (const void*)&cl.buffer);
很简单。
我知道我似乎只在需要时才去拜访，但我还有一个问题。
在我的
我得到的只是一个空白屏幕。注释掉获取和释放 GL objects 的部分允许渲染像以前一样工作。只是未能集成 OpenCL 组件。
opencl.h 只是一个带有辅助函数的 header，我用它来注册所有 CL object 并在最后通过单个函数调用释放它们。
编辑：2015 年 6 月 12 日
取消注释显示更多信息的 /*fprintf(stderr, "ERROR: " x " failed %d\n", cl_stat);*/ 行：
ERROR: Set cl kernel arg failed -38
ERROR: Acquiring GL objects failed -5
根据 this list，OpenCL 错误代码 -38 表示它是无效内存 object，然后在尝试 re-acquire GL objects 时出现资源不足错误.
这里是main.c
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <glib.h>
#include <gdk/gdkx.h>
#include <epoxy/glx.h>
#include <epoxy/gl.h>
#include <gtk/gtk.h>
#include <CL/opencl.h>
#include "cl_utils.h"
#define IGNORE_VAR(type, identifier) \
{ \
type IGNORED_VARIABLE_abcd = identifier; \
identifier = IGNORED_VARIABLE_abcd; \
}
#define CL_ASSERT(x) \
/*if(print_info) \
printf(x "...\n"); */\
if(cl_stat != CL_SUCCESS) \
{ \
/*fprintf(stderr, "ERROR: " x " failed %d\n", cl_stat);*/ \
goto exception; \
}
const GLchar *vert_src = "\n" \
"#version 330 \n" \
"#extension GL_ARB_explicit_attrib_location: enable \n" \
" \n" \
"layout(location = 0) in vec2 in_position; \n" \
" \n" \
"void main() \n" \
"{ \n" \
" gl_Position = vec4(in_position, 0.0, 1.0); \n" \
"} \n";
const GLchar *frag_src = "\n" \
"void main (void) \n" \
"{ \n" \
" gl_FragColor = vec4(1.0, 1.0, 1.0, 1.0); \n" \
"} \n";
const char *cl_src = "\n" \
"typedef struct Point{ \n" \
" float x; \n" \
" float y; \n" \
"} Point; \n" \
" \n" \
"__kernel void cl_func(__global Point* point) \n" \
"{ \n" \
" const int i = get_global_id(0); \n" \
" const float d = 0.99; \n" \
" \n" \
" if(i>=3) \n" \
" return; \n" \
" \n" \
" point[i].x = point[i].x * d; \n" \
" point[i].y = point[i].y * d; \n" \
"} \n";
struct cl
{
clu_object_stack* stack;
cl_platform_id* platform;
cl_uint num_platforms;
cl_device_id* device;
cl_uint num_devices;
cl_context context;
cl_context_properties properties[7];
cl_command_queue queue;
cl_program program;
cl_kernel kernel;
cl_mem buffer;
} cl;
struct cl cl;
GLuint gl_vao, gl_buffer, gl_program;
cl_int init_cl(GtkGLArea *area)
{
cl_int cl_stat;
cl.stack = cluCreateObjectStack(44);
cl_stat = clGetPlatformIDs(0, NULL, &cl.num_platforms);
CL_ASSERT("Got number of platforms");
cl.platform = malloc((size_t)cl.num_platforms * sizeof(cl_platform_id));
cl_stat = clGetPlatformIDs(cl.num_platforms, cl.platform, NULL);
CL_ASSERT("Got platforms");
cl_stat = clGetDeviceIDs(cl.platform[0], CL_DEVICE_TYPE_GPU, 0, NULL, &cl.num_devices);
CL_ASSERT("Got number of devices");
printf("Number of GPU devices: %d\n", cl.num_devices);
if(cl.num_devices == 0)
{
fprintf(stderr, "Num devices cannot be 0\n");
goto exception;
}
cl.device = malloc((size_t)cl.num_devices * sizeof(cl_device_id));
cl_stat = clGetDeviceIDs(cl.platform[0], CL_DEVICE_TYPE_GPU, cl.num_devices, cl.device, NULL);
CL_ASSERT("Got devices");
if(cl.device == NULL)
{
fprintf(stderr, "Devices list is NULL\n");
goto exception;
}
gtk_gl_area_make_current (area);
cl.properties[0] = CL_GL_CONTEXT_KHR;
cl.properties[1] = (cl_context_properties) glXGetCurrentContext();
cl.properties[2] = CL_GLX_DISPLAY_KHR;
cl.properties[3] = (cl_context_properties) glXGetCurrentDisplay();
cl.properties[4] = CL_CONTEXT_PLATFORM;
cl.properties[5] = (cl_context_properties) cl.platform[0];
cl.properties[6] = 0;
cl.context = cluCreateContext(cl.stack, cl.properties, cl.num_devices, cl.device, NULL, NULL, &cl_stat);
CL_ASSERT("Created cl context");
cl.queue = cluCreateCommandQueue(cl.stack, cl.context, cl.device[0], 0, &cl_stat);
CL_ASSERT("Created command queue");
cl.buffer = cluCreateFromGLBuffer(cl.stack, cl.context, CL_MEM_WRITE_ONLY, gl_buffer, NULL);
CL_ASSERT("Created cl memory object from gl buffer");
cl.program = cluCreateProgramWithSource(cl.stack, cl.context, 1, (const char **)&cl_src, NULL, &cl_stat);
CL_ASSERT("Created cl program object");
cl_stat = clBuildProgram(cl.program, cl.num_devices, cl.device, NULL, NULL, NULL);
if(cl_stat != CL_SUCCESS)
{
size_t ret_size;
clGetProgramBuildInfo(cl.program, cl.device[0], CL_PROGRAM_BUILD_LOG, 0, NULL, &ret_size);
char e_str[ret_size];
clGetProgramBuildInfo(cl.program, cl.device[0], CL_PROGRAM_BUILD_LOG, ret_size, e_str, &ret_size);
printf("%s\n", e_str);
}
CL_ASSERT("Built cl program object");
cl.kernel = cluCreateKernel(cl.stack, cl.program, "cl_func", &cl_stat);
CL_ASSERT("Created cl kernel object");
return 0;
exception:
return 1;
}
static gboolean realise(GtkGLArea *area, GdkGLContext *context);
static gboolean render(GtkGLArea *area, GdkGLContext *context);
int main(int argc, char** argv)
{
gtk_init(&argc, &argv);
GtkWidget *window = gtk_window_new(GTK_WINDOW_TOPLEVEL),
*gl_area = gtk_gl_area_new();
g_signal_connect(window, "delete-event", G_CALLBACK(gtk_main_quit), NULL);
g_signal_connect(gl_area, "realize", G_CALLBACK(realise), NULL);
g_signal_connect(gl_area, "render", G_CALLBACK(render), NULL);
gtk_container_add(GTK_CONTAINER(window), gl_area);
gtk_widget_show_all(window);
gtk_main();
cluFreeObjectStack(cl.stack);
free(cl.platform);
free(cl.device);
return 0;
}
static gboolean realise(GtkGLArea *area, GdkGLContext *context)
{
IGNORE_VAR(GdkGLContext*, context);
gtk_gl_area_make_current(GTK_GL_AREA(area));
if (gtk_gl_area_get_error (GTK_GL_AREA(area)) != NULL)
{
printf("Failed to initialiize buffers\n");
return FALSE;
}
GLfloat verts[] =
{
+0.0f, +1.0f,
-1.0f, -1.0f,
+1.0f, -1.0f,
};
GLuint frag_shader, vert_shader;
frag_shader = glCreateShader(GL_FRAGMENT_SHADER);
vert_shader = glCreateShader(GL_VERTEX_SHADER);
glShaderSource(frag_shader, 1, &frag_src, NULL);
glShaderSource(vert_shader, 1, &vert_src, NULL);
glCompileShader(frag_shader);
glCompileShader(vert_shader);
gl_program = glCreateProgram();
glAttachShader(gl_program, frag_shader);
glAttachShader(gl_program, vert_shader);
glLinkProgram(gl_program);
glGenBuffers(1, &gl_buffer);
glBindBuffer(GL_ARRAY_BUFFER, gl_buffer);
glBufferData(GL_ARRAY_BUFFER, sizeof(verts), verts, GL_DYNAMIC_DRAW);
glGenVertexArrays(1, &gl_vao);
glBindVertexArray(gl_vao);
glEnableVertexAttribArray(0);
glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 0, (void*)0);
glBindVertexArray(0);
//glDeleteBuffers(1, &gl_buffer);
if(init_cl(area))
return FALSE;
return TRUE;
}
static gboolean render(GtkGLArea *area, GdkGLContext *context)
{
IGNORE_VAR(GdkGLContext*, context);
IGNORE_VAR(GtkGLArea*, area);
cl_int cl_stat;
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
glClearColor(0.0, 0.0, 0.0, 1.0);
glUseProgram(gl_program);
glBindVertexArray(gl_vao);
glDrawArrays(GL_TRIANGLES, 0, 3);
glBindVertexArray (0);
glUseProgram (0);
glFinish();
cl_stat = clEnqueueAcquireGLObjects(cl.queue, 1, &cl.buffer, 0, NULL, NULL);
CL_ASSERT("Acquiring GL objects");
cl_stat = clSetKernelArg(cl.kernel, 0, sizeof(cl_mem), (const void*)cl.buffer);
CL_ASSERT("Set cl kernel arg");
//size_t g_sz[1] = { 32 };
//cl_stat = clEnqueueNDRangeKernel(cl.queue, cl.kernel, 1, NULL, g_sz, NULL, 0, NULL, NULL);
cl_stat = clEnqueueTask(cl.queue, cl.kernel, 0, NULL, NULL);
CL_ASSERT("Executing cl kernel");
cl_stat = clEnqueueReleaseGLObjects(cl.queue, 1, &cl.buffer, 0, NULL, NULL);
CL_ASSERT("Releasing GL buffer");
clFinish(cl.queue);
return TRUE;
exception:
return FALSE;
}
SConstruct
import os
CFLAGS = '--std=c11 -g -Wall'
ENV = {'PATH':os.environ['PATH']}
env = Environment(CFLAGS=CFLAGS, ENV=ENV)
if os.name is 'posix':
env['ENV']['TERM'] = os.environ['TERM']
env.ParseConfig('pkg-config --cflags --libs gtk+-3.0')
env.Append(LIBS = ['epoxy', 'GL', 'OpenCL'])
env.Program(target='gl', source=['main.c', 'cl_utils.c'])
# vim: set filetype=python:
cl_utils.c
#include <stdlib.h>
#include <stdio.h>
#include "cl_utils.h"
#define FUNC __func__
#define CL_ASSERT(object, label) \
if(!object) \
{ \
fprintf(stderr, "CL_ERROR: failed to create CL %s object: %d\n", #object, *cl_stat); \
goto label; \
} \
else \
{ \
printf("CL_INFO: created cl_%s object\n", #object); \
}
#define PLATFORM_INFO(_) \
_(PROFILE) \
_(VERSION) \
_(NAME) \
_(VENDOR) \
_(EXTENSIONS)
const char* space = " ";
#define TO_STRING(TYPE, ENUM, ...) [ENUM] = #TYPE,
const char* CLU_TYPE_STRING[] = { CLU_TABLE(TO_STRING) };
#undef TO_STRING
#define TO_STRING(x) [CL_PLATFORM_##x] = #x,
const char* platform_info_string[] = { PLATFORM_INFO(TO_STRING) };
#undef TO_STRING
const char* clGetPlatformInfoString(cl_platform_info platform_info)
{
return platform_info_string[platform_info];
}
cl_int infoPlatforms(cl_platform_id* platforms, cl_int num_platforms, cl_platform_info* params, cl_int num_params)
{
cl_int cl_stat;
size_t buffer_size = 10, buffer_ret_size = 0;
char* buffer = (char*)malloc(buffer_size);
for(cl_int i = 0; i < num_platforms; i++)
{
for(cl_int j = 0; j < num_params; j++)
{
cl_stat = clGetPlatformInfo(platforms[i], params[j], 0, NULL, &buffer_ret_size);
if(cl_stat != CL_SUCCESS)
{
fprintf(stderr, "ERROR: clGetPlatformInfo failed\n");
goto end;
}
if(buffer_ret_size > buffer_size)
{
void* tmp = NULL;
buffer_size = buffer_ret_size;
tmp = realloc(buffer, buffer_size);
if(tmp == NULL)
{
fprintf(stderr, "ERROR: Could not realloc memory\n");
perror("ERROR: ");
goto end;
}
else
buffer = (char*)tmp;
}
cl_stat = clGetPlatformInfo(platforms[i], params[j], buffer_size, buffer, &buffer_ret_size);
if(cl_stat != CL_SUCCESS)
{
fprintf(stderr, "ERROR: clGetPlatformInfo failed\n");
goto end;
}
printf("%s: %s\n", clGetPlatformInfoString(params[j]), buffer);
}
printf("\n");
}
end:
if(cl_stat != CL_SUCCESS)
{
printf("SENT TO ERROR HANDLER!\n");
printf("CL_ERROR: %d\n", cl_stat);
}
free(buffer);
if(cl_stat != CL_SUCCESS)
return 1;
return 0;
}
clu_object_stack* cluCreateObjectStack(int init_len)
{
clu_object_stack* stack = (clu_object_stack*)malloc(sizeof(clu_object_stack));
stack->length = init_len;
stack->num_objects = 0;
stack->list = (clu_object*)malloc(sizeof(clu_object) * init_len);
return stack;
}
void cluFreeObjectStack(clu_object_stack* stack)
{
cl_int ret = CL_SUCCESS;
printf("Freeing stack: %p\n", stack);
int i;
for(i=stack->num_objects-1; i>=0; i--)
{
switch(stack->list[i].type)
{
#define TO_FREE(cl_type, ENUM, element, cluRelease) \
case ENUM: \
ret = cluRelease(stack->list[i].element); \
/*printf("Releasing " #cl_type "\n");*/ \
break;
CLU_TABLE(TO_FREE)
#undef TO_FREE
default:
printf("ERROR: Invalid or unsupported object type\n");
break;
}
if(ret != CL_SUCCESS)
{
printf("Failed to release %s\n", CLU_TYPE_STRING[stack->list[i].type]);
}
}
free(stack->list);
free(stack);
}
int cluAssignToObjectGroup(struct clu_object_stack* stack, void* obj, cl_type type)
{
if(stack->num_objects >= stack->length)
{
printf("Stack Error\n");
return -1;
}
stack->list[stack->num_objects].type = type;
switch(type)
{
#define TO_ASSIGN(cl_type, ENUM, element, ...) \
case ENUM: \
stack->list[stack->num_objects].element=*(cl_type*)obj; \
/*printf("Assigning " #cl_type "\n");*/ \
break;
CLU_TABLE(TO_ASSIGN)
#undef TO_ASSIGN
default:
printf("cluAssignToObjectGroup Failed\n");
break;
}
stack->num_objects++;
return 0;
}
cl_context cluCreateContext(
clu_object_stack *stack,
cl_context_properties *properties,
cl_uint num_devices,
const cl_device_id *devices,
pfn_notify func,
void *user_data,
cl_int* cl_stat)
{
cl_context context= clCreateContext(properties, num_devices, devices, func, user_data, cl_stat);
if(context == 0)
printf("Created ZERO value cl context\n");
CL_ASSERT(context, error_ret);
cluAssignToObjectGroup(stack, &context, CLU_CONTEXT);
error_ret:
if(devices == NULL)
fprintf(stderr, "%2sdevices cannot be NULL\n", space);
if(num_devices == 0)
fprintf(stderr, "%2snum_devices cannot be zero\n", space);
if((func == NULL) && (user_data != NULL))
fprintf(stderr, "%2spfn_notify cannot be NULL when user_data is not NULL\n", space);
return context;
}
cl_command_queue cluCreateCommandQueue(
clu_object_stack *stack,
cl_context context,
cl_device_id device,
cl_command_queue_properties properties,
cl_int *cl_stat)
{
cl_command_queue queue = clCreateCommandQueue(context, device, properties, cl_stat);
CL_ASSERT(queue, error_ret);
cluAssignToObjectGroup(stack, &queue, CLU_COMMAND_QUEUE);
error_ret:
return queue;
}
cl_mem cluCreateBuffer(
clu_object_stack* stack,
cl_context context,
cl_mem_flags flags,
size_t size,
void *host_ptr,
cl_int *cl_stat)
{
cl_mem buffer = clCreateBuffer(context, flags, size, host_ptr, cl_stat);
CL_ASSERT(buffer, error_ret);
cluAssignToObjectGroup(stack, &buffer, CLU_MEM_OBJECT);
error_ret:
return buffer;
}
cl_program cluCreateProgramWithSource(
clu_object_stack* stack,
cl_context context,
cl_uint count,
const char **strings,
const size_t *lengths,
cl_int *cl_stat)
{
cl_program program = clCreateProgramWithSource(context, count, strings, lengths, cl_stat);
CL_ASSERT(program, error_ret);
cluAssignToObjectGroup(stack, &program, CLU_PROGRAM);
error_ret:
return program;
}
cl_kernel cluCreateKernel(
clu_object_stack* stack,
cl_program program,
const char *kernel_name,
cl_int *cl_stat)
{
cl_kernel kernel = clCreateKernel(program, kernel_name, cl_stat);
CL_ASSERT(kernel, error_ret);
cluAssignToObjectGroup(stack, &kernel, CLU_KERNEL);
error_ret:
return kernel;
}
cl_mem cluCreateFromGLBuffer(
clu_object_stack* stack,
cl_context context,
cl_mem_flags flags,
GLuint bufobj,
cl_int* cl_stat)
{
cl_mem gl_buffer = clCreateFromGLBuffer(context, flags, bufobj, cl_stat);
CL_ASSERT(gl_buffer, error_ret);
cluAssignToObjectGroup(stack, &gl_buffer, CLU_MEM_OBJECT);
error_ret:
return gl_buffer;
}
cl_utils.h
#ifndef __CL_UTILS_H__
#define __CL_UTILS_H__
#ifdef __cplusplus
extern "C" {
#endif /* C++ */
#include <stdarg.h>
#if defined(__APPLE__) || defined(MACOSX)
#include <OpenCL/opencl.h>
#else
#include <CL/opencl.h>
#endif
#include <epoxy/gl.h>
#define CLU_TABLE(_) \
_(cl_context, CLU_CONTEXT, context, clReleaseContext) \
_(cl_command_queue, CLU_COMMAND_QUEUE, queue, clReleaseCommandQueue) \
_(cl_mem,CLU_MEM_OBJECT, mem_object, clReleaseMemObject) \
_(cl_program, CLU_PROGRAM, program, clReleaseProgram) \
_(cl_kernel, CLU_KERNEL, kernel, clReleaseKernel)
#define TO_ENUM(cl_type, ENUM, ...) ENUM,
typedef enum cl_type{
CLU_TABLE(TO_ENUM)
} cl_type;
#undef TO_ENUM
typedef void (*pfn_notify)(const char *errinfo, const void *private_info, size_t cb, void *user_data);
//extern const char* CLU_TYPE_STRING[];
typedef struct clu_object clu_object;
struct clu_object{
cl_type type;
union{
cl_context context;
cl_command_queue queue;
cl_mem mem_object;
cl_program program;
cl_kernel kernel;
};
};
typedef struct clu_object_stack{
struct clu_object* list;
int num_objects;
int length;
} clu_object_stack;
const char* clGetPlatformInfoString(cl_platform_info platform_info);
int infoPlatforms(cl_platform_id* platforms, cl_int num_platforms, cl_platform_info* params, cl_int num_params);
clu_object_stack* cluCreateObjectStack(int init_len);
void cluFreeObjectStack(clu_object_stack* stack);
int cluAssignToObjectGroup(struct clu_object_stack* stack, void* obj, cl_type type);
cl_context cluCreateContext(
clu_object_stack *stack,
cl_context_properties *properties,
cl_uint num_devices,
const cl_device_id *devices,
pfn_notify func,
void *user_data,
cl_int* cl_stat);
cl_command_queue cluCreateCommandQueue(
clu_object_stack* stack,
cl_context context,
cl_device_id device,
cl_command_queue_properties properties,
cl_int *cl_stat);
cl_mem cluCreateBuffer(
clu_object_stack* stack,
cl_context context,
cl_mem_flags flags,
size_t size,
void *host_ptr,
cl_int * cl_stat);
cl_program cluCreateProgramWithSource(
clu_object_stack* stack,
cl_context context,
cl_uint count,
const char **strings,
const size_t *lengths,
cl_int *cl_stat);
cl_kernel cluCreateKernel(
clu_object_stack* stack,
cl_program program,
const char *kernel_name,
cl_int *cl_stat);
cl_mem cluCreateFromGLBuffer(
clu_object_stack* stack,
cl_context context,
cl_mem_flags flags,
GLuint bufobj,
cl_int* cl_stat);
#ifdef __cplusplus
}
#endif /* C++ */
#endif /* __CL_UTILS_H__ */
要启用 CL-GL 互操作性，必须按特定顺序进行设置：
1. Create OpenGL context
2. Create OpenCL context
3. Create OpenGL buffers
4. Start OpenGL rendering
一个可能的问题是您在 OpenCL 上下文之前（在调用 init_cl 之前）创建了 OpenGL 缓冲区。
找到问题了。 clSetKernelArg() 的最后一个参数需要一个指向 mem 对象的指针，我忘了在前面添加 & 运算符。
所以这个：
cl_stat = clSetKernelArg(cl.kernel, 0, sizeof(cl_mem), (const void*)cl.buffer);
变成这样：
cl_stat = clSetKernelArg(cl.kernel, 0, sizeof(cl_mem), (const void*)&cl.buffer);
很简单。