[FFmpeg-devel] [PATCH 1/2] libavutil/libavfilter: opencl wrapper based on comments on 20130329

Stefano Sabatini stefasab at gmail.com
Sat Mar 30 03:20:12 CET 2013


On date Friday 2013-03-29 21:32:35 +0800, Wei Gao encoded:
> 

> From dc939475b0ac425604e1293b816bbaeb0e651b7f Mon Sep 17 00:00:00 2001
> From: highgod0401 <highgod0401 at gmail.com>
> Date: Fri, 29 Mar 2013 20:13:55 +0800
> Subject: [PATCH 1/2] opencl wrapper based on comments on 20130329
> 
> ---
>  configure          |   4 +
>  libavutil/Makefile |   3 +
>  libavutil/opencl.c | 708 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>  libavutil/opencl.h | 203 +++++++++++++++
>  4 files changed, 918 insertions(+)
>  create mode 100644 libavutil/opencl.c
>  create mode 100644 libavutil/opencl.h
> 
> diff --git a/configure b/configure
> index 8443db4..9c42a85 100755
> --- a/configure
> +++ b/configure
> @@ -233,6 +233,7 @@ External library support:
>    --enable-libxvid         enable Xvid encoding via xvidcore,
>                             native MPEG-4/Xvid encoder exists [no]
>    --enable-openal          enable OpenAL 1.1 capture support [no]
> +  --enable-opencl          enable OpenCL code
>    --enable-openssl         enable openssl [no]
>    --enable-x11grab         enable X11 grabbing [no]
>    --enable-zlib            enable zlib [autodetect]
> @@ -1178,6 +1179,7 @@ EXTERNAL_LIBRARY_LIST="
>      libxavs
>      libxvid
>      openal
> +    opencl
>      openssl
>      x11grab
>      zlib
> @@ -3982,6 +3984,7 @@ enabled openal     && { { for al_libs in "${OPENAL_LIBS}" "-lopenal" "-lOpenAL32
>                          die "ERROR: openal not found"; } &&
>                        { check_cpp_condition "AL/al.h" "defined(AL_VERSION_1_1)" ||
>                          die "ERROR: openal must be installed and version must be 1.1 or compatible"; }
> +enabled opencl     && require2 opencl CL/cl.h clEnqueueNDRangeKernel -lOpenCL
>  enabled openssl    && { check_lib openssl/ssl.h SSL_library_init -lssl -lcrypto ||
>                          check_lib openssl/ssl.h SSL_library_init -lssl32 -leay32 ||
>                          check_lib openssl/ssl.h SSL_library_init -lssl -lcrypto -lws2_32 -lgdi32 ||
> @@ -4350,6 +4353,7 @@ echo "network support           ${network-no}"
>  echo "threading support         ${thread_type-no}"
>  echo "safe bitstream reader     ${safe_bitstream_reader-no}"
>  echo "SDL support               ${sdl-no}"
> +echo "opencl enabled            ${opencl-no}"
>  echo "texi2html enabled         ${texi2html-no}"
>  echo "perl enabled              ${perl-no}"
>  echo "pod2man enabled           ${pod2man-no}"
> diff --git a/libavutil/Makefile b/libavutil/Makefile
> index 103ce5e..b520473 100644
> --- a/libavutil/Makefile
> +++ b/libavutil/Makefile
> @@ -52,6 +52,8 @@ HEADERS = adler32.h                                                     \
>  
>  HEADERS-$(CONFIG_LZO)                   += lzo.h
>  
> +HEADERS-$(CONFIG_OPENCL)                += opencl.h
> +
>  ARCH_HEADERS = bswap.h                                                  \
>                 intmath.h                                                \
>                 intreadwrite.h                                           \
> @@ -106,6 +108,7 @@ OBJS = adler32.o                                                        \
>         xtea.o                                                           \
>  
>  OBJS-$(CONFIG_LZO)                      += lzo.o
> +OBJS-$(CONFIG_OPENCL)                   += opencl.o
>  
>  OBJS += $(COMPAT_OBJS:%=../compat/%)
>  
> diff --git a/libavutil/opencl.c b/libavutil/opencl.c
> new file mode 100644
> index 0000000..6281b32
> --- /dev/null
> +++ b/libavutil/opencl.c
> @@ -0,0 +1,708 @@
> +/*
> + * Copyright (C) 2012 Peng Gao <peng at multicorewareinc.com>
> + * Copyright (C) 2012 Li   Cao <li at multicorewareinc.com>
> + * Copyright (C) 2012 Wei  Gao <weigao at multicorewareinc.com>
> + *
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU Lesser General Public
> + * License as published by the Free Software Foundation; either
> + * version 2.1 of the License, or (at your option) any later version.
> + *
> + * FFmpeg is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> + * Lesser General Public License for more details.
> + *
> + * You should have received a copy of the GNU Lesser General Public
> + * License along with FFmpeg; if not, write to the Free Software
> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
> + */
> +
> +#include "opencl.h"
> +#include "avstring.h"
> +#include "log.h"
> +#include "avassert.h"
> +
> +#if HAVE_PTHREADS
> +
> +#include <pthread.h>
> +static pthread_mutex_t atomic_opencl_lock = PTHREAD_MUTEX_INITIALIZER;
> +
> +#define LOCK_OPENCL pthread_mutex_lock(&atomic_opencl_lock);
> +#define UNLOCK_OPENCL pthread_mutex_unlock(&atomic_opencl_lock);
> +
> +#elif !HAVE_THREADS
> +#define LOCK_OPENCL
> +#define UNLOCK_OPENCL
> +#endif
> +
> +
> +#define MAX_KERNEL_NUM 500

> +#define MAX_REG_NAME_LEN  64
> +#define MAX_REG_NUM 200
> +
> +typedef struct KernelNode {
> +    char kernel_name[AV_OPENCL_MAX_KERNEL_NAME_SIZE];
> +    cl_kernel kernel;
> +    int kernel_ref;
> +} KernelNode;
> +
> +typedef struct GPUEnv {
> +    int opencl_is_inited;
> +    cl_platform_id platform;
> +    cl_device_type device_type;
> +    cl_context context;
> +    cl_device_id *device_ids;
> +    cl_device_id device_id;
> +    cl_command_queue command_queue;
> +    cl_program program;

> +    int register_count;
> +    char register_names[MAX_REG_NUM][MAX_REG_NAME_LEN + 1];
> +    const char *kernel_code[MAX_REG_NUM];
> +    av_opencl_kernel_function kernel_functions[MAX_REG_NUM];

"register_names" sounds pretty confusing, it reminds "CPU registers"
to me, which have a completely different meaning in this context, and
doesn't tell much about what it's actually representing.

I suggest:
#define MAX_KERNEL_CODE_NAME_LEN  64
#define MAX_KERNEL_CODE_NUM 200
...
char                      kernel_code_names[MAX_KERNEL_CODE_NUM][MAX_KERNEL_CODE_NAME_LEN + 1];
const char               *kernel_code      [MAX_KERNEL_CODE_NUM];
int                       kernel_code_count;
av_opencl_kernel_function kernel_functions [MAX_KERNEL_CODE_NUM];

> +    int kernel_count;
> +    KernelNode kernel_node[MAX_KERNEL_NUM];
> +    int is_user_created; // 1: the opencl env is created by user and use AVOpenCLExternalInfo to pass to ffmpeg ,0:created by opencl wrapper.
> +} GPUEnv;
> +
> +typedef struct OpenclUtils {
> +    const AVClass *class;
> +    int log_offset;
> +    void *log_ctx;
> +} OpenclUtils;
> +
> +static const AVClass openclutils_class = {"OPENCLUTILS", av_default_item_name,
> +                                                   NULL, LIBAVUTIL_VERSION_INT,
> +                                                   offsetof(OpenclUtils, log_offset),
> +                                                   offsetof(OpenclUtils, log_ctx)};
> +static OpenclUtils openclutils = {&openclutils_class};
> +static GPUEnv gpu_env;
> +
> +typedef struct OpenclErrorMsg {
> +    int err_code;
> +    const char *err_str;
> +} OpenclErrorMsg;
> +
> +static const OpenclErrorMsg opencl_err_msg[] = {
> +        {CL_DEVICE_NOT_FOUND,                               "DEVICE NOT FOUND"},
> +        {CL_DEVICE_NOT_AVAILABLE,                           "DEVICE NOT AVAILABLE"},
> +        {CL_COMPILER_NOT_AVAILABLE,                         "COMPILER NOT AVAILABLE"},
> +        {CL_MEM_OBJECT_ALLOCATION_FAILURE,                  "MEM OBJECT ALLOCATION FAILURE"},
> +        {CL_OUT_OF_RESOURCES,                               "OUT OF RESOURCES"},
> +        {CL_OUT_OF_HOST_MEMORY,                             "OUT OF HOST MEMORY"},
> +        {CL_PROFILING_INFO_NOT_AVAILABLE,                   "PROFILING INFO NOT AVAILABLE"},
> +        {CL_MEM_COPY_OVERLAP,                               "MEM COPY OVERLAP"},
> +        {CL_IMAGE_FORMAT_MISMATCH,                          "IMAGE FORMAT MISMATCH"},
> +        {CL_IMAGE_FORMAT_NOT_SUPPORTED,                     "IMAGE FORMAT NOT_SUPPORTED"},
> +        {CL_BUILD_PROGRAM_FAILURE,                          "BUILD PROGRAM FAILURE"},
> +        {CL_MAP_FAILURE,                                    "MAP FAILURE"},
> +        {CL_MISALIGNED_SUB_BUFFER_OFFSET,                   "MISALIGNED SUB BUFFER OFFSET"},
> +        {CL_EXEC_STATUS_ERROR_FOR_EVENTS_IN_WAIT_LIST,      "EXEC STATUS ERROR FOR EVENTS IN WAIT LIST"},
> +        {CL_COMPILE_PROGRAM_FAILURE,                        "COMPILE PROGRAM FAILURE"},
> +        {CL_LINKER_NOT_AVAILABLE,                           "LINKER NOT AVAILABLE"},
> +        {CL_LINK_PROGRAM_FAILURE,                           "LINK PROGRAM FAILURE"},
> +        {CL_DEVICE_PARTITION_FAILED,                        "DEVICE PARTITION FAILED"},
> +        {CL_KERNEL_ARG_INFO_NOT_AVAILABLE,                  "KERNEL ARG INFO NOT AVAILABLE"},
> +        {CL_INVALID_VALUE,                                  "INVALID VALUE"},
> +        {CL_INVALID_DEVICE_TYPE,                            "INVALID DEVICE TYPE"},
> +        {CL_INVALID_PLATFORM,                               "INVALID PLATFORM"},
> +        {CL_INVALID_DEVICE,                                 "INVALID DEVICE"},
> +        {CL_INVALID_CONTEXT,                                "INVALID CONTEXT"},
> +        {CL_INVALID_QUEUE_PROPERTIES,                       "INVALID QUEUE PROPERTIES"},
> +        {CL_INVALID_COMMAND_QUEUE,                          "INVALID COMMAND QUEUE"},
> +        {CL_INVALID_HOST_PTR,                               "INVALID HOST PTR"},
> +        {CL_INVALID_MEM_OBJECT,                             "INVALID MEM OBJECT"},
> +        {CL_INVALID_IMAGE_FORMAT_DESCRIPTOR,                "INVALID IMAGE FORMAT DESCRIPTOR"},
> +        {CL_INVALID_IMAGE_SIZE,                             "INVALID IMAGE SIZE"},
> +        {CL_INVALID_SAMPLER,                                "INVALID SAMPLER"},
> +        {CL_INVALID_BINARY,                                 "INVALID BINARY"},
> +        {CL_INVALID_BUILD_OPTIONS,                          "INVALID BUILD OPTIONS"},
> +        {CL_INVALID_PROGRAM,                                "INVALID PROGRAM"},
> +        {CL_INVALID_PROGRAM_EXECUTABLE,                     "INVALID PROGRAM EXECUTABLE"},
> +        {CL_INVALID_KERNEL_NAME,                            "INVALID KERNEL NAME"},
> +        {CL_INVALID_KERNEL_DEFINITION,                      "INVALID KERNEL DEFINITION"},
> +        {CL_INVALID_KERNEL,                                 "INVALID KERNEL"},
> +        {CL_INVALID_ARG_INDEX,                              "INVALID ARG INDEX"},
> +        {CL_INVALID_ARG_VALUE,                              "INVALID ARG VALUE"},
> +        {CL_INVALID_ARG_SIZE,                               "INVALID ARG_SIZE"},
> +        {CL_INVALID_KERNEL_ARGS,                            "INVALID KERNEL ARGS"},
> +        {CL_INVALID_WORK_DIMENSION,                         "INVALID WORK DIMENSION"},
> +        {CL_INVALID_WORK_GROUP_SIZE,                        "INVALID WORK GROUP SIZE"},
> +        {CL_INVALID_WORK_ITEM_SIZE,                         "INVALID WORK ITEM SIZE"},
> +        {CL_INVALID_GLOBAL_OFFSET,                          "INVALID GLOBAL OFFSET"},
> +        {CL_INVALID_EVENT_WAIT_LIST,                        "INVALID EVENT WAIT LIST"},
> +        {CL_INVALID_EVENT,                                  "INVALID EVENT"},
> +        {CL_INVALID_OPERATION,                              "INVALID OPERATION"},
> +        {CL_INVALID_GL_OBJECT,                              "INVALID GL OBJECT"},
> +        {CL_INVALID_BUFFER_SIZE,                            "INVALID BUFFER SIZE"},
> +        {CL_INVALID_MIP_LEVEL,                              "INVALID MIP LEVEL"},
> +        {CL_INVALID_GLOBAL_WORK_SIZE,                       "INVALID GLOBAL WORK SIZE"},
> +        {CL_INVALID_PROPERTY,                               "INVALID PROPERTY"},
> +        {CL_INVALID_IMAGE_DESCRIPTOR,                       "INVALID IMAGE DESCRIPTOR"},
> +        {CL_INVALID_COMPILER_OPTIONS,                       "INVALID COMPILER OPTIONS"},
> +        {CL_INVALID_LINKER_OPTIONS,                         "INVALID LINKER OPTIONS"},
> +        {CL_INVALID_DEVICE_PARTITION_COUNT,                 "INVALID DEVICE PARTITION COUNT"},
> +};
> +
> +static const char *opencl_errstr(cl_int status)
> +{
> +    int i;
> +    for (i = 0; i < sizeof(opencl_err_msg); i++) {
> +        if (opencl_err_msg[i].err_code == status)
> +            return opencl_err_msg[i].err_str;
> +    }
> +    return "unknown error";
> +}
> +

> +int av_opencl_register_kernel_code(const char *register_name, const char *kernel_code)

register_name -> kernel_code_name?

> +{
> +    int ret = 0;
> +    int i;
> +    LOCK_OPENCL;
> +    if (gpu_env.register_count >= MAX_REG_NUM) {
> +        av_log(&openclutils, AV_LOG_ERROR,
> +         "Could not register with name '%s', maximum number of registered kernels %d already reached\n",
> +         register_name, MAX_REG_NUM);
> +        ret = AVERROR(EINVAL);
> +        goto end;
> +    }
> +    if (strlen(register_name) > MAX_REG_NAME_LEN) {
> +         av_log(&openclutils, AV_LOG_ERROR, "Registered name %s is too long\n", register_name);
> +         ret = AVERROR(EINVAL);
> +         goto end;
> +    }
> +    for (i = 0;i < gpu_env.register_count;i++) {
> +        if (av_strcasecmp(register_name, gpu_env.register_names[i]) == 0) {
> +            av_log(&openclutils, AV_LOG_ERROR, "Register name '%s' has been registered\n", register_name);
> +            ret = AVERROR(EINVAL);
> +            goto end;
> +        }
> +    }
> +    gpu_env.kernel_code[gpu_env.register_count] = kernel_code;
> +    av_strlcpy(gpu_env.register_names[gpu_env.register_count], register_name,
> +                        sizeof(gpu_env.register_names[gpu_env.register_count]));
> +    gpu_env.register_count++;
> +end:
> +    UNLOCK_OPENCL;
> +    return ret;
> +}
> +

> +int av_opencl_create_kernel(AVOpenCLKernelEnv *env, const char *kernel_name)
> +{
> +    cl_int status;
> +    int i, ret = 0;
> +    LOCK_OPENCL;

> +    if (strlen(kernel_name) > sizeof(env->kernel_name)) {

strlen(str) > size-1?

> +        av_log(&openclutils, AV_LOG_ERROR, "Created kernel name %s is too long\n", kernel_name);
> +        ret = AVERROR(EINVAL);
> +        goto end;
> +    }
> +    if (!env->kernel) {
> +        for (i = 0;i < gpu_env.kernel_count;i++) {
> +            if (av_strcasecmp(kernel_name, gpu_env.kernel_node[i].kernel_name) == 0) {
> +                env->kernel = gpu_env.kernel_node[i].kernel;
> +                gpu_env.kernel_node[i].kernel_ref++;
> +                break;
> +            }
> +        }
> +        if (!env->kernel) {
> +            env->kernel = clCreateKernel(gpu_env.program, kernel_name, &status);
> +            if (status != CL_SUCCESS) {
> +                av_log(&openclutils, AV_LOG_ERROR, "Could not create OpenCL kernel: %s\n", opencl_errstr(status));
> +                ret = AVERROR_EXTERNAL;
> +                goto end;
> +            }
> +            av_strlcpy(gpu_env.kernel_node[gpu_env.kernel_count].kernel_name, kernel_name,
> +                       sizeof(gpu_env.kernel_node[gpu_env.kernel_count].kernel_name));
> +            gpu_env.kernel_node[gpu_env.kernel_count].kernel= env->kernel;
> +            gpu_env.kernel_node[gpu_env.kernel_count].kernel_ref++;
> +            gpu_env.kernel_count++;
> +        }

> +        env->context       = gpu_env.context;
> +        env->command_queue = gpu_env.command_queue;
> +        env->program       = gpu_env.program;

What's the need to set these fields in env if you can always access the
global environment?

> +        av_strlcpy(env->kernel_name, kernel_name, sizeof(env->kernel_name));
> +    }
> +end:
> +    UNLOCK_OPENCL;
> +    return ret;
> +}
> +
> +void av_opencl_release_kernel(AVOpenCLKernelEnv *env)
> +{
> +    cl_int status;
> +    int i;
> +    LOCK_OPENCL
> +    if (!env->kernel)
> +        goto end;
> +    for (i = 0;i < gpu_env.kernel_count;i++) {
> +        if (gpu_env.kernel_node[i].kernel == env->kernel) {
> +            if (gpu_env.kernel_node[i].kernel_ref == 1) {
> +                status = clReleaseKernel(env->kernel);
> +                if (status != CL_SUCCESS) {
> +                    av_log(&openclutils, AV_LOG_ERROR, "Could not release kernel: %s\n",
> +                          opencl_errstr(status));
> +                }
> +                gpu_env.kernel_node[i].kernel = NULL;
> +                env->kernel = NULL;
> +                memset(gpu_env.kernel_node[i].kernel_name, 0, sizeof(gpu_env.kernel_node[i].kernel_name));
> +                gpu_env.kernel_count--;
> +            }
> +            gpu_env.kernel_node[i].kernel_ref--;
> +        }
> +    }
> +end:
> +    UNLOCK_OPENCL

I feel this kernel referencing code is a bit overdesigned and could be
avoided if you allow the creation of more kernels associated to the
same kernel function. On the other hand I don't know what are the
drawbacks of this approach (performance loss?).

> +}
> +
> +static int init_opencl_env(GPUEnv *gpu_env, AVOpenCLExternalInfo *ext_opencl_info)
> +{
> +    size_t device_length;
> +    cl_int status;
> +    cl_uint num_platforms, num_devices;
> +    cl_platform_id *platform_ids = NULL;
> +    cl_context_properties cps[3];
> +    char platform_name[100];
> +    int i, ret = 0;
> +    cl_device_type device_type[] = {CL_DEVICE_TYPE_GPU, CL_DEVICE_TYPE_CPU, CL_DEVICE_TYPE_DEFAULT};
> +    if (ext_opencl_info) {
> +        if (gpu_env->is_user_created)
> +            return 0;
> +        gpu_env->platform        = ext_opencl_info->platform;
> +        gpu_env->is_user_created = 1;
> +        gpu_env->command_queue   = ext_opencl_info->command_queue;
> +        gpu_env->context         = ext_opencl_info->context;
> +        gpu_env->device_ids      = ext_opencl_info->device_ids;
> +        gpu_env->device_id       = ext_opencl_info->device_id;
> +        gpu_env->device_type     = ext_opencl_info->device_type;
> +    } else {
> +        if (!gpu_env->is_user_created) {
> +            status = clGetPlatformIDs(0, NULL, &num_platforms);
> +            if (status != CL_SUCCESS) {
> +                av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL platform ids: %s\n", opencl_errstr(status));
> +                return AVERROR_EXTERNAL;
> +            }
> +            if (num_platforms > 0) {
> +                platform_ids = av_mallocz(num_platforms * sizeof(cl_platform_id));
> +                if (!platform_ids) {
> +                    ret = AVERROR(ENOMEM);
> +                    goto end;
> +                }
> +                status = clGetPlatformIDs(num_platforms, platform_ids, NULL);
> +                if (status != CL_SUCCESS) {
> +                    av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL platform ids: %s\n", opencl_errstr(status));
> +                    ret = AVERROR_EXTERNAL;
> +                    goto end;
> +                }
> +                for (i = 0; i < num_platforms; i++) {
> +                    status = clGetPlatformInfo(platform_ids[i], CL_PLATFORM_VENDOR,
> +                                               sizeof(platform_name), platform_name,
> +                                               NULL);
> +
> +                    if (status != CL_SUCCESS) {
> +                        av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL platform info: %s\n", opencl_errstr(status));
> +                        ret = AVERROR_EXTERNAL;
> +                        goto end;
> +                    }
> +                    gpu_env->platform = platform_ids[i];
> +                    status = clGetDeviceIDs(gpu_env->platform, CL_DEVICE_TYPE_GPU,
> +                                            0, NULL, &num_devices);
> +                    if (status != CL_SUCCESS) {
> +                        av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL device number:%s\n", opencl_errstr(status));
> +                        ret = AVERROR_EXTERNAL;
> +                        goto end;
> +                    }
> +                    if (num_devices == 0) {
> +                        //find CPU device
> +                        status = clGetDeviceIDs(gpu_env->platform, CL_DEVICE_TYPE_CPU,
> +                                             0, NULL, &num_devices);
> +                    }
> +                    if (status != CL_SUCCESS) {
> +                        av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL device ids: %s\n", opencl_errstr(status));
> +                        ret = AVERROR_EXTERNAL;
> +                        goto end;
> +                    }
> +                    if (num_devices)
> +                       break;
> +
> +                }
> +            }
> +            if (!gpu_env->platform) {
> +                av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL platforms\n");
> +                ret = AVERROR_EXTERNAL;
> +                goto end;
> +            }
> +
> +           /*
> +                 * Use available platform.
> +                 */
> +            av_log(&openclutils, AV_LOG_VERBOSE, "Platform Name: %s\n", platform_name);
> +            cps[0] = CL_CONTEXT_PLATFORM;
> +            cps[1] = (cl_context_properties)gpu_env->platform;
> +            cps[2] = 0;
> +            /* Check for GPU. */
> +
> +            for (i = 0; i < sizeof(device_type); i++) {
> +                gpu_env->device_type = device_type[i];
> +                gpu_env->context     = clCreateContextFromType(cps, gpu_env->device_type,
> +                                                               NULL, NULL, &status);
> +                if (status != CL_SUCCESS) {
> +                    av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL context from device type: %s\n", opencl_errstr(status));
> +                    ret = AVERROR_EXTERNAL;
> +                    goto end;
> +                }
> +                if (gpu_env->context)
> +                    break;
> +            }
> +            if (!gpu_env->context) {
> +                av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL context from device type\n");
> +                ret = AVERROR_EXTERNAL;
> +                goto end;
> +            }
> +            /* Detect OpenCL devices. */
> +            /* First, get the size of device list data */
> +            status = clGetContextInfo(gpu_env->context, CL_CONTEXT_DEVICES,
> +                                      0, NULL, &device_length);
> +            if (status != CL_SUCCESS) {
> +                av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL device length: %s\n", opencl_errstr(status));
> +                ret = AVERROR_EXTERNAL;
> +                goto end;
> +            }
> +            if (device_length == 0) {
> +                av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL device length\n");
> +                ret = AVERROR_EXTERNAL;
> +                goto end;
> +            }
> +            /* Now allocate memory for device list based on the size we got earlier */
> +            gpu_env->device_ids = av_mallocz(device_length);
> +            if (!gpu_env->device_ids) {
> +                ret = AVERROR(ENOMEM);
> +                goto end;
> +            }
> +            /* Now, get the device list data */
> +            status = clGetContextInfo(gpu_env->context, CL_CONTEXT_DEVICES, device_length,
> +                                      gpu_env->device_ids, NULL);
> +            if (status != CL_SUCCESS) {
> +                av_log(&openclutils, AV_LOG_ERROR, "Could not get OpenCL context info: %s\n", opencl_errstr(status));
> +                ret = AVERROR_EXTERNAL;
> +                goto end;
> +            }
> +            /* Create OpenCL command queue. */
> +            gpu_env->command_queue = clCreateCommandQueue(gpu_env->context, gpu_env->device_ids[0],
> +                                                          0, &status);
> +            if (status != CL_SUCCESS) {
> +                av_log(&openclutils, AV_LOG_ERROR, "Could not create OpenCL command queue: %s\n", opencl_errstr(status));
> +                ret = AVERROR_EXTERNAL;
> +                goto end;
> +            }
> +        }
> +    }
> +end:
> +    av_free(platform_ids);
> +    return ret;
> +}
> +
> +
> +int av_opencl_register_kernel_function(const char *register_name, av_opencl_kernel_function function)
> +{
> +    int i, ret = 0;
> +    LOCK_OPENCL
> +    for (i = 0; i < gpu_env.register_count; i++) {
> +        if (av_strcasecmp(register_name, gpu_env.register_names[i]) == 0) {
> +            gpu_env.kernel_functions[i] = function;
> +            goto end;
> +        }
> +    }
> +    av_log(&openclutils, AV_LOG_ERROR, "Could not find a register name '%s', cannot register function\n", register_name);
> +    ret = AVERROR(EINVAL);
> +end:
> +    UNLOCK_OPENCL
> +    return ret;
> +}
> +
> +static int compile_kernel_file(GPUEnv *gpu_env, const char *build_option)
> +{
> +    cl_int status;
> +    char *temp, *source_str = NULL;
> +    size_t source_str_len = 0;
> +    int i, ret = 0;
> +
> +    if (gpu_env->program)
> +        return ret;
> +
> +    for (i = 0; i < gpu_env->register_count; i++) {
> +        source_str_len += strlen(gpu_env->kernel_code[i]);
> +    }
> +    source_str = av_mallocz(source_str_len + 1);
> +    if (!source_str) {
> +        return AVERROR(ENOMEM);
> +    }
> +    temp = source_str;
> +    for (i = 0; i < gpu_env->register_count; i++) {
> +        memcpy(temp, gpu_env->kernel_code[i], strlen(gpu_env->kernel_code[i]));
> +        temp += strlen(gpu_env->kernel_code[i]);
> +    }
> +    /* create a CL program using the kernel source */
> +    gpu_env->program = clCreateProgramWithSource(gpu_env->context, 1, (const char **)(&source_str),
> +                                                 &source_str_len, &status);
> +    if(status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not create OpenCL program with source code: %s\n",
> +               opencl_errstr(status));
> +        ret = AVERROR_EXTERNAL;
> +        goto end;
> +    }
> +    if (!gpu_env->program) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Created program is NULL\n");
> +        ret = AVERROR_EXTERNAL;
> +        goto end;
> +    }
> +    /* create a cl program executable for all the devices specified */
> +    if (!gpu_env->is_user_created)
> +        status = clBuildProgram(gpu_env->program, 1, gpu_env->device_ids,
> +                                build_option, NULL, NULL);
> +    else
> +        status = clBuildProgram(gpu_env->program, 1, &(gpu_env->device_id),
> +                                 build_option, NULL, NULL);
> +
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not compile OpenCL kernel: %s\n", opencl_errstr(status));
> +        ret = AVERROR_EXTERNAL;
> +        goto end;
> +    }
> +end:
> +    av_free(source_str);
> +    return ret;
> +}
> +
> +int av_opencl_run_kernel(const char *register_name, void **userdata)
> +{
> +    av_opencl_kernel_function function = NULL;
> +    int i, ret = 0;
> +    LOCK_OPENCL
> +    for (i = 0; i < gpu_env.register_count; i++) {
> +        if (av_strcasecmp(register_name, gpu_env.register_names[i]) == 0) {
> +            function = gpu_env.kernel_functions[i];
> +            break;
> +        }
> +    }
> +    if (!function) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not find kernel: %s\n", register_name);
> +        ret = AVERROR(EINVAL);
> +        goto end;
> +    }
> +    ret = (function(userdata));

What about specifying the function when you run the kernel?

I mean something like:
int av_opencl_run_kernel(av_opencl_function fn, void **userdata);

This way you avoid the registration process, simplify the API, avoid
function look-up when running the kernel, and also increase
flexibility since you can pass different functions to the kernel.

Also I do not understand why do you need to use a register name (used
to register a kernel code) to associate a function to a kernel.

> +end:
> +    UNLOCK_OPENCL
> +    return ret;
> +}
> +
> +int av_opencl_init(AVDictionary *options, AVOpenCLExternalInfo *ext_opencl_info)
> +{
> +    int ret = 0;
> +    LOCK_OPENCL
> +    if (!gpu_env.opencl_is_inited) {
> +        /*initialize devices, context, command_queue*/
> +        AVDictionaryEntry *opt_entry = av_dict_get(options, "build_option", NULL, 0);
> +        ret = init_opencl_env(&gpu_env, ext_opencl_info);
> +        if (ret < 0)
> +            goto end;
> +        /*initialize program, kernel_name, kernel_count*/
> +        if (opt_entry)
> +            ret = compile_kernel_file(&gpu_env, opt_entry->value);
> +        else
> +            ret = compile_kernel_file(&gpu_env, NULL);
> +        if (ret < 0)
> +            goto end;
> +        av_assert1(gpu_env.register_count > 0);
> +        gpu_env.opencl_is_inited = 1;
> +    }
> +end:
> +    UNLOCK_OPENCL
> +    return ret;
> +}
> +
> +void av_opencl_uninit(void)
> +{
> +    cl_int status;
> +    int i;
> +    LOCK_OPENCL
> +    if (!gpu_env.opencl_is_inited)
> +        goto end;
> +    if (gpu_env.is_user_created)
> +        goto end;
> +    for (i = 0;i < gpu_env.kernel_count;i++) {
> +        if (gpu_env.kernel_node[i].kernel)
> +            goto end;
> +    }
> +
> +    if (gpu_env.program) {
> +        status = clReleaseProgram(gpu_env.program);
> +        if (status != CL_SUCCESS) {
> +            av_log(&openclutils, AV_LOG_ERROR, "Could not release OpenCL program: %s\n", opencl_errstr(status));
> +        }
> +        gpu_env.program = NULL;
> +    }
> +    if (gpu_env.command_queue) {
> +        status = clReleaseCommandQueue(gpu_env.command_queue);
> +        if (status != CL_SUCCESS) {
> +            av_log(&openclutils, AV_LOG_ERROR, "Could not release OpenCL command queue: %s\n", opencl_errstr(status));
> +        }
> +        gpu_env.command_queue = NULL;
> +    }
> +    if (gpu_env.context) {
> +        status = clReleaseContext(gpu_env.context);
> +        if (status != CL_SUCCESS) {
> +            av_log(&openclutils, AV_LOG_ERROR, "Could not release OpenCL context: %s\n", opencl_errstr(status));
> +        }
> +        gpu_env.context = NULL;
> +    }
> +    av_freep(&(gpu_env.device_ids));
> +    gpu_env.opencl_is_inited = 0;
> +end:
> +    UNLOCK_OPENCL
> +}
> +
> +int av_opencl_buffer_create(cl_mem *cl_buf, size_t cl_buf_size, int flags, void *host_ptr)
> +{
> +    cl_int status;
> +    *cl_buf = clCreateBuffer(gpu_env.context, flags, cl_buf_size, host_ptr, &status);
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not create OpenCL buffer: %s\n", opencl_errstr(status));
> +        return AVERROR_EXTERNAL;
> +    }
> +    return 0;
> +}
> +
> +void av_opencl_buffer_release(cl_mem cl_buf)
> +{
> +    cl_int status = 0;
> +    if (!cl_buf)
> +        return;
> +    status = clReleaseMemObject(cl_buf);
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not release OpenCL buffer: %s\n", opencl_errstr(status));
> +    }
> +}
> +
> +int av_opencl_buffer_write(cl_mem dst_cl_buf, uint8_t *src_buf, size_t buf_size)
> +{
> +    cl_int status;
> +    void *mapped = clEnqueueMapBuffer(gpu_env.command_queue, dst_cl_buf,
> +                                      CL_TRUE,CL_MAP_WRITE, 0, sizeof(uint8_t) * buf_size,
> +                                      0, NULL, NULL, &status);
> +
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not map OpenCL buffer: %s\n", opencl_errstr(status));
> +        return AVERROR_EXTERNAL;
> +    }
> +    memcpy(mapped, src_buf, buf_size);
> +
> +    status = clEnqueueUnmapMemObject(gpu_env.command_queue, dst_cl_buf, mapped, 0, NULL, NULL);
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not unmap OpenCL buffer: %s\n", opencl_errstr(status));
> +        return AVERROR_EXTERNAL;
> +    }
> +    return 0;
> +}
> +
> +int av_opencl_buffer_read(uint8_t *dst_buf, cl_mem src_cl_buf, size_t buf_size)
> +{
> +    cl_int status;
> +    void *mapped = clEnqueueMapBuffer(gpu_env.command_queue, src_cl_buf,
> +                                      CL_TRUE,CL_MAP_READ, 0, buf_size,
> +                                      0, NULL, NULL, &status);
> +
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not map OpenCL buffer: %s\n", opencl_errstr(status));
> +        return AVERROR_EXTERNAL;
> +    }
> +    memcpy(dst_buf, mapped, buf_size);
> +
> +    status = clEnqueueUnmapMemObject(gpu_env.command_queue, src_cl_buf, mapped, 0, NULL, NULL);
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not unmap OpenCL buffer: %s\n", opencl_errstr(status));
> +        return AVERROR_EXTERNAL;
> +    }
> +    return 0;
> +}
> +
> +int av_opencl_buffer_write_image(cl_mem dst_cl_buf, size_t cl_buffer_size,
> +                                        uint8_t **src_data, int *plane_size, int plane_num, int offset)
> +{
> +    int i, buffer_size = 0;
> +    uint8_t *temp;
> +    cl_int status;
> +    void *mapped;
> +    if ((unsigned int)plane_num > 8) {
> +        return AVERROR(EINVAL);
> +    }
> +    for (i = 0;i < plane_num;i++) {
> +        buffer_size += plane_size[i];
> +    }
> +    if (buffer_size > cl_buffer_size) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Cannot write image to OpenCL buffer: buffer too small\n");
> +        return AVERROR(EINVAL);
> +    }
> +    mapped = clEnqueueMapBuffer(gpu_env.command_queue, dst_cl_buf,
> +                                      CL_TRUE,CL_MAP_WRITE, 0, buffer_size + offset,
> +                                      0, NULL, NULL, &status);
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not map OpenCL buffer: %s\n", opencl_errstr(status));
> +        return AVERROR_EXTERNAL;
> +    }
> +    temp = mapped;
> +    temp += offset;
> +    for (i = 0; i < plane_num; i++) {
> +        memcpy(temp, src_data[i], plane_size[i]);
> +        temp += plane_size[i];
> +    }
> +    status = clEnqueueUnmapMemObject(gpu_env.command_queue, dst_cl_buf, mapped, 0, NULL, NULL);
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not unmap OpenCL buffer: %s\n", opencl_errstr(status));
> +        return AVERROR_EXTERNAL;
> +    }
> +    return 0;
> +}
> +
> +int av_opencl_buffer_read_image(uint8_t **dst_data, int *plane_size, int plane_num,
> +                                       cl_mem src_cl_buf, size_t cl_buffer_size)
> +{
> +    int i,buffer_size = 0,ret = 0;
> +    uint8_t *temp;
> +    void *mapped;
> +    cl_int status;
> +    if ((unsigned int)plane_num > 8) {
> +        return AVERROR(EINVAL);
> +    }
> +    for (i = 0;i < plane_num;i++) {
> +        buffer_size += plane_size[i];
> +    }
> +    if (buffer_size > cl_buffer_size) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Cannot write image to CPU buffer: OpenCL buffer too small\n");
> +        return AVERROR(EINVAL);
> +    }
> +    mapped = clEnqueueMapBuffer(gpu_env.command_queue, src_cl_buf,
> +                                      CL_TRUE,CL_MAP_READ, 0, buffer_size,
> +                                      0, NULL, NULL, &status);
> +
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not map OpenCL buffer: %s\n", opencl_errstr(status));
> +        return AVERROR_EXTERNAL;
> +    }
> +    temp = mapped;
> +    if (ret >= 0) {
> +        for (i = 0;i < plane_num;i++) {
> +            memcpy(dst_data[i], temp, plane_size[i]);
> +            temp += plane_size[i];
> +        }
> +    }
> +    status = clEnqueueUnmapMemObject(gpu_env.command_queue, src_cl_buf, mapped, 0, NULL, NULL);
> +    if (status != CL_SUCCESS) {
> +        av_log(&openclutils, AV_LOG_ERROR, "Could not unmap OpenCL buffer: %s\n", opencl_errstr(status));
> +        return AVERROR_EXTERNAL;
> +    }
> +    return 0;
> +}
> +
> diff --git a/libavutil/opencl.h b/libavutil/opencl.h
> new file mode 100644
> index 0000000..6bb42fa
> --- /dev/null
> +++ b/libavutil/opencl.h
> @@ -0,0 +1,203 @@
> +/*
> + * Copyright (C) 2012 Peng Gao <peng at multicorewareinc.com>
> + * Copyright (C) 2012 Li   Cao <li at multicorewareinc.com>
> + * Copyright (C) 2012 Wei  Gao <weigao at multicorewareinc.com>
> + *
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU Lesser General Public
> + * License as published by the Free Software Foundation; either
> + * version 2.1 of the License, or (at your option) any later version.
> + *
> + * FFmpeg is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> + * Lesser General Public License for more details.
> + *
> + * You should have received a copy of the GNU Lesser General Public
> + * License along with FFmpeg; if not, write to the Free Software
> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
> + */
> +
> +#include "config.h"
> +#include "dict.h"
> +
> +#ifndef LIBAVUTIL_OPENCLWRAPPER_H
> +#define LIBAVUTIL_OPENCLWRAPPER_H
> +
> +#include <CL/cl.h>
> +
> +#define AV_OPENCL_KERNEL( ... )# __VA_ARGS__
> +
> +#define AV_OPENCL_MAX_KERNEL_NAME_SIZE 150
> +
> +typedef struct AVOpenCLKernelEnv {
> +    cl_context context;
> +    cl_command_queue command_queue;
> +    cl_program program;
> +    cl_kernel kernel;
> +    char kernel_name[AV_OPENCL_MAX_KERNEL_NAME_SIZE];
> +} AVOpenCLKernelEnv;
> +
> +typedef struct AVOpenCLExternalInfo {
> +    cl_platform_id platform;
> +    cl_device_type device_type;
> +    cl_context context;
> +    cl_device_id *device_ids;
> +    cl_device_id  device_id;
> +    cl_command_queue command_queue;
> +    char *platform_name;
> +} AVOpenCLExternalInfo;
> +

> +/**
> + * User defined function, used to set the input parameter in the kernel
> + * environment. This function launches kernel and copies data from GPU to
> + * CPU, or from CPU to GPU.
> + */
> +typedef int (* av_opencl_kernel_function) (void **userdata);


/**
 * User defined function, used to set the input parameters in the kernel
 * environment, and execute a routine defined in the kernel, and move
 * data from between CPU and GPU.
 */
typedef int (* av_opencl_kernel_function) (void **userdata);

> +
> +/**
> + * Register kernel code.
> +
> + *  The registered kernel code is stored in a global context, and compiled
> + *  in the runtime environment when av_opencl_init() is called.
> + *

> + * @param register_name   register name is a index, it is used to find kernel code which to be compiled
> + *                                      in GPU and user defined function which as a interface between CPU and GPU (av_opencl_kernel_function)

Same remark already done on "register_name", which I suggest to rename
to "kernel_code_name". Also docs could be improved, I suggest:

@param kernel_code_name  name assigned to the registered kernel code

> + * @param kernel_code    kernel code to be compiled in GPU

nit: GPU -> to be compiled in the OpenCL runtime environment

GPU sounds too much low level to me.

> + * @return  >=0 on success, a negative error code in case of failure
> + */
> +int av_opencl_register_kernel_code(const char *register_name, const char *kernel_code);
> +
> +
> +/**
> + * Initialize the run time OpenCL environment.
> + *
> + * This must be done after all the kernels have been registered with
> + * the av_opencl_register_kernel_code().
> + *

> + * Currently, the only accepted option is "build_options", used to set
> + * compile OpenCL registered kernels code, see reference "OpenCL
> + * Specification Version: 1.2 chapter 5.6.4".

Currently, the only accepted option is "build_options", used to set
options to compile registered kernels code. See reference "OpenCL
Specification Version: 1.2 chapter 5.6.4".

> + *
> + * @param options               dictionary of key/value options
> + * @param ext_opencl_info    external OpenCL environment, created by an
> + *                                       application program, ignored if set to NULL
> + * @return  >=0 on success, a negative error value on failure
> + */
> +int av_opencl_init(AVDictionary *options, AVOpenCLExternalInfo *ext_opencl_info);
> +

> +/**
> + * Register a function for running the kernel specified by the kernel name.

> + * @param register_name   register name is a index, it is used to find kernel code which to be compiled in GPU and user defined function which as a interface between CPU and GPU (av_opencl_kernel_function)

Same note as above about name and documentation. Also mismatch with
the first sentence.

> + * @param function         user defined function,should not be NULL, it is a interface between CPU and GPU which used to set the input parameter in the kernel environment
> + * @return  >=0 on success, a negative error value on failure
> + */
> +int av_opencl_register_kernel_function(const char *register_name, av_opencl_kernel_function function);

See my note in the implementation, I think this could be removed.

> +
> +/**
> + * Create kernel object on the specified OpenCL run in env.
> + *
> + * @param env                  it is used for user to get the kernel env created by av_opencl_init, user will use it to run kernel.

> + * @param kernel_name      kernel function name

@param kernel_name      name of a function defined in the OpenCL runtime environment

> + * @return >=0 on success, a negative error code on failure
> + */
> +int av_opencl_create_kernel(AVOpenCLKernelEnv *env, const char *kernel_name);
> +

[...]
-- 
FFmpeg = Frightening and Faithless Multipurpose Practical Extravagant Gem


More information about the ffmpeg-devel mailing list