[FFmpeg-devel] [PATCH]opencl: automatically select the fastest opencl device

Stefano Sabatini stefasab at gmail.com
Sun Dec 1 21:22:52 CET 2013


On date Saturday 2013-11-30 22:55:51 -0600, Lenny Wang encoded:
> On Fri, Nov 29, 2013 at 7:39 AM, Stefano Sabatini <stefasab at gmail.com> wrote:
[...]
> Please find attached new patch based on Stefano's feedback, thanks.

> From: Lenny Wang <lwanghpc at gmail.com>
> Date: Sat, 30 Nov 2013 22:45:00 -0600
> Subject: [PATCH] cmdutils & opencl: add -opencl_bench option to test and show available OpenCL devices 
> ---
>  cmdutils.c                     | 173 +++++++++++++++++++++++++++++++++++++++++
>  cmdutils.h                     |   6 ++
>  cmdutils_common_opts.h         |   1 +
>  cmdutils_opencl_bench_kernel.h |  86 ++++++++++++++++++++
>  doc/fftools-common-opts.texi   |   4 +
>  doc/utils.texi                 |   4 +-
>  libavutil/opencl.c             |  41 ++++++++++
>  libavutil/opencl.h             |  22 ++++++
>  8 files changed, 335 insertions(+), 2 deletions(-)
> 
> diff --git a/cmdutils.c b/cmdutils.c
> index 46ade3f..407c36e 100644
> --- a/cmdutils.c
> +++ b/cmdutils.c
> @@ -60,6 +60,8 @@
>  #endif
>  #if CONFIG_OPENCL
>  #include "libavutil/opencl.h"
> +#include "libavutil/time.h"
> +#include "cmdutils_opencl_bench_kernel.h"
>  #endif
>  
>  
> @@ -986,6 +988,177 @@ int opt_timelimit(void *optctx, const char *opt, const char *arg)
>  }
>  
>  #if CONFIG_OPENCL
> +#define OCLCHECK(method, ... )                                                 \
> +do {                                                                           \
> +    status = method(__VA_ARGS__);                                              \
> +    if (status != CL_SUCCESS) {                                                \
> +        av_log(NULL, AV_LOG_ERROR, # method " error '%s'\n",                   \
> +               av_opencl_errstr(status));                                      \
> +        ret = AVERROR_EXTERNAL;                                                \
> +        goto end;                                                              \
> +    }                                                                          \
> +} while (0)
> +
> +#define CREATEBUF(out, flags, size)                                            \
> +do {                                                                           \
> +    out = clCreateBuffer(ext_opencl_env->context, flags, size, NULL, &status); \
> +    if (status != CL_SUCCESS) {                                                \
> +        av_log(NULL, AV_LOG_ERROR, "Could not create OpenCL buffer\n");        \
> +        ret = AVERROR_EXTERNAL;                                                \
> +        goto end;                                                              \
> +    }                                                                          \
> +} while (0)
> +

> +static void fill_rand_int(int *data, int n)
> +{
> +    int i;
> +    srand(av_gettime());
> +    for (i = 0; i < n; i++)
> +        data[i] = rand();
> +}
> +
> +#define OPENCL_NB_ITER 3
> +static int64_t run_opencl_bench(AVOpenCLExternalEnv *ext_opencl_env)
> +{
> +    int i, arg = 0, width = 1920, height = 1088;
> +    int64_t start, ret = 0;
> +    cl_int status;
> +    size_t kernel_len;
> +    char *inbuf;
> +    int *mask;
> +    int buf_size = width * height * sizeof(char);
> +    int mask_size = sizeof(uint32_t) * 128;
> +
> +    cl_mem cl_mask, cl_inbuf, cl_outbuf;
> +    cl_kernel kernel = NULL;
> +    cl_program program = NULL;
> +    size_t local_work_size_2d[2] = {16, 16};
> +    size_t global_work_size_2d[2] = {(size_t)width, (size_t)height};
> +

> +    if (!(inbuf = av_malloc(buf_size))) {
> +        av_log(NULL, AV_LOG_ERROR, "Could not allocate buffer\n");
> +        return AVERROR(ENOMEM);
> +    }
> +    if (!(mask = av_malloc(mask_size))) {

> +        av_log(NULL, AV_LOG_ERROR, "Could not allocate buffer\n");

mask

> +        return AVERROR(ENOMEM);
> +    }

nit: there is an extremely very unlikely case that inbuf succeeds and
mask alloc doesn't, so I suggest:

    if (!(inbuf = av_malloc(buf_size))) {
        av_log(NULL, AV_LOG_ERROR, "Could not allocate buffer\n");
        ret = AVERROR(ENOMEM);
        goto end;
    }
    if (!(mask = av_malloc(mask_size))) {
        av_log(NULL, AV_LOG_ERROR, "Could not allocate mask\n");
        ret = AVERROR(ENOMEM);
        goto end;
    }

Also you can probably skip a log in case an alloc() fails (rationale:
in these cases the relevant information is that a memory error
occurred, knowing when - especially for small allocs, is not that
important).

> +    fill_rand_int((int*)inbuf, buf_size/4);
> +    fill_rand_int(mask, mask_size/4);
> +
> +    CREATEBUF(cl_mask, CL_MEM_READ_ONLY, mask_size);
> +    CREATEBUF(cl_inbuf, CL_MEM_READ_ONLY, buf_size);
> +    CREATEBUF(cl_outbuf, CL_MEM_READ_WRITE, buf_size);
> +
> +    kernel_len = strlen(ocl_bench_source);
> +    program = clCreateProgramWithSource(ext_opencl_env->context, 1, &ocl_bench_source,
> +                                        &kernel_len, &status);
> +    if (status != CL_SUCCESS || !program) {
> +        av_log(NULL, AV_LOG_ERROR, "OpenCL unable to create benchmark program\n");
> +        ret = AVERROR_EXTERNAL;
> +        goto end;
> +    }
> +    status = clBuildProgram(program, 1, &(ext_opencl_env->device_id), NULL, NULL, NULL);
> +    if (status != CL_SUCCESS) {
> +        av_log(NULL, AV_LOG_ERROR, "OpenCL unable to build benchmark program\n");
> +        ret = AVERROR_EXTERNAL;
> +        goto end;
> +    }
> +    kernel = clCreateKernel(program, "unsharp_bench", &status);
> +    if (status != CL_SUCCESS) {
> +        av_log(NULL, AV_LOG_ERROR, "OpenCL unable to create benchmark kernel\n");
> +        ret = AVERROR_EXTERNAL;
> +        goto end;
> +    }
> +
> +    OCLCHECK(clEnqueueWriteBuffer, ext_opencl_env->command_queue, cl_inbuf, CL_TRUE, 0,
> +             buf_size, inbuf, 0, NULL, NULL);
> +    OCLCHECK(clEnqueueWriteBuffer, ext_opencl_env->command_queue, cl_mask, CL_TRUE, 0,
> +             mask_size, mask, 0, NULL, NULL);
> +    OCLCHECK(clSetKernelArg, kernel, arg++, sizeof(cl_mem), &cl_inbuf);
> +    OCLCHECK(clSetKernelArg, kernel, arg++, sizeof(cl_mem), &cl_outbuf);
> +    OCLCHECK(clSetKernelArg, kernel, arg++, sizeof(cl_mem), &cl_mask);
> +    OCLCHECK(clSetKernelArg, kernel, arg++, sizeof(cl_int), &width);
> +    OCLCHECK(clSetKernelArg, kernel, arg++, sizeof(cl_int), &height);

> +    /* warm up */

Nit: this comment is probably more confusing than clarifying

> +    OCLCHECK(clEnqueueNDRangeKernel, ext_opencl_env->command_queue, kernel, 2, NULL,
> +             global_work_size_2d, local_work_size_2d, 0, NULL, NULL);
> +    clFinish(ext_opencl_env->command_queue);

Any reason why this is called outside the loop? This could be
(briefly) explained in a comment.

> +    start = av_gettime();
> +    for (i = 0; i < OPENCL_NB_ITER; i++)
> +	    OCLCHECK(clEnqueueNDRangeKernel, ext_opencl_env->command_queue, kernel, 2, NULL,
> +                 global_work_size_2d, local_work_size_2d, 0, NULL, NULL);
> +	clFinish(ext_opencl_env->command_queue);
> +    ret = (av_gettime() - start)/OPENCL_NB_ITER;
> +end:
> +    if (kernel)
> +        clReleaseKernel(kernel);
> +    if (program)
> +        clReleaseProgram(program);
> +    if (cl_inbuf)
> +        clReleaseMemObject(cl_inbuf);
> +    if (cl_outbuf)
> +        clReleaseMemObject(cl_outbuf);
> +    if (cl_mask)
> +        clReleaseMemObject(cl_mask);
> +    av_free(inbuf);
> +    av_free(mask);
> +    return ret;
> +}
> +
> +static int compare_ocl_device_desc(const void *a, const void *b)
> +{
> +    return ((AVOpenCLDeviceBenchmark*)a)->runtime - ((AVOpenCLDeviceBenchmark*)b)->runtime;
> +}
> +
> +int test_opencl(void *optctx, const char *opt, const char *arg)
> +{
> +    int i, j, nb_devices = 0, count = 0;
> +    int64_t score = 0;
> +    AVOpenCLDeviceList *device_list;
> +    AVOpenCLDeviceNode *device_node = NULL;
> +    AVOpenCLDeviceBenchmark *devices = NULL;
> +    cl_platform_id platform;
> +
> +    av_opencl_get_device_list(&device_list);
> +    for (i = 0; i < device_list->platform_num; i++)
> +        nb_devices += device_list->platform_node[i]->device_num;
> +    if (!nb_devices) {
> +        av_log(NULL, AV_LOG_ERROR, "No OpenCL device detected!\n");
> +        return AVERROR(EINVAL);
> +    }
> +    if (!(devices = av_malloc(sizeof(AVOpenCLDeviceBenchmark) * nb_devices))) {
> +        av_log(NULL, AV_LOG_ERROR, "Could not allocate buffer\n");
> +        return AVERROR(ENOMEM);
> +    }
> +
> +    for (i = 0; i < device_list->platform_num; i++) {
> +        for (j = 0; j < device_list->platform_node[i]->device_num; j++) {
> +            device_node = device_list->platform_node[i]->device_node[j];
> +            platform = device_list->platform_node[i]->platform_id;
> +            score = av_opencl_benchmark(device_node, platform, run_opencl_bench);
> +            if (score > 0) {
> +                devices[count].platform_idx = i;
> +                devices[count].device_idx = j;
> +                devices[count].runtime = score;
> +                strcpy(devices[count].device_name, device_node->device_name);
> +                count++;
> +            }
> +        }
> +    }

> +    qsort(devices, count, sizeof(AVOpenCLDeviceBenchmark), compare_ocl_device_desc);
> +    for (i = 0; i < count; i++)
> +        av_log(NULL, AV_LOG_WARNING,
> +               "platform_idx=%d\tdevice_idx=%d\tdevice_name=%s\truntime=%"PRId64"\n",
> +               devices[i].platform_idx, devices[i].device_idx,
> +               devices[i].device_name, devices[i].runtime);

This should be printed to stdout, so that the user can easily parse it
(also this is consistent with other similar options: in general the
expected textual output of the program should be sent to stdout, see
for example -pix_fmts and other cmdutils print utilities).

> +
> +    av_opencl_free_device_list(&device_list);
> +    av_free(devices);
> +    return 0;
> +}
> +
>  int opt_opencl(void *optctx, const char *opt, const char *arg)
>  {
>      char *key, *value;
> diff --git a/cmdutils.h b/cmdutils.h
> index b814961..5f47120 100644
> --- a/cmdutils.h
> +++ b/cmdutils.h
> @@ -101,6 +101,12 @@ int opt_codec_debug(void *optctx, const char *opt, const char *arg);
>  int opt_opencl(void *optctx, const char *opt, const char *arg);
>  
>  /**
> + * Perform benchmark on all OpenCL devices and print results.
> + * This option processing function does not utilize the arguments.
> + */
> +int test_opencl(void *optctx, const char *opt, const char *arg);

nit: opt_opencl_bench() -> consistent with other opt handlers and with
the option name

> +
> +/**
>   * Limit the execution time.
>   */
>  int opt_timelimit(void *optctx, const char *opt, const char *arg);
> diff --git a/cmdutils_common_opts.h b/cmdutils_common_opts.h
> index 3e3f0ac..276aa6e 100644
> --- a/cmdutils_common_opts.h
> +++ b/cmdutils_common_opts.h
> @@ -22,5 +22,6 @@
>      { "max_alloc"  , HAS_ARG,  {.func_arg = opt_max_alloc},     "set maximum size of a single allocated block", "bytes" },
>      { "cpuflags"   , HAS_ARG | OPT_EXPERT, { .func_arg = opt_cpuflags }, "force specific cpu flags", "flags" },
>  #if CONFIG_OPENCL
> +    { "opencl_bench", OPT_EXIT, {.func_arg = test_opencl},      "run benchmark across available OpenCL devices" },
>      { "opencl_options", HAS_ARG, {.func_arg = opt_opencl},      "set OpenCL environment options" },
>  #endif
> diff --git a/cmdutils_opencl_bench_kernel.h b/cmdutils_opencl_bench_kernel.h
> new file mode 100644
> index 0000000..ce6bad5
> --- /dev/null
> +++ b/cmdutils_opencl_bench_kernel.h
> @@ -0,0 +1,86 @@
> +/*
> + * Copyright (C) 2013 Lenny Wang
> + *
> + * 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
> + */
> +
> +#ifndef CMDUTILS_OPENCL_BENCH_KERNEL_H
> +#define CMDUTILS_OPENCL_BENCH_KERNEL_H
> +
> +#include "libavutil/opencl.h"
> +
> +const char *ocl_bench_source = AV_OPENCL_KERNEL(
> +inline unsigned char clip_uint8(int a)
> +{
> +    if (a & (~0xFF))
> +        return (-a)>>31;
> +    else
> +        return a;
> +}
> +
> +kernel void unsharp_bench(
> +                    global unsigned char *src,
> +                    global unsigned char *dst,
> +                    global int *mask,
> +                    int width,
> +                    int height)
> +{
> +    int i, j, local_idx, lc_idx, sum = 0;
> +    int2 thread_idx, block_idx, global_idx, lm_idx;
> +    thread_idx.x = get_local_id(0);
> +    thread_idx.y = get_local_id(1);
> +    block_idx.x = get_group_id(0);
> +    block_idx.y = get_group_id(1);
> +    global_idx.x = get_global_id(0);
> +    global_idx.y = get_global_id(1);
> +    local uchar data[32][32];
> +    local int lc[128];
> +
> +    for (i = 0; i <= 1; i++) {
> +        lm_idx.y = -8 + (block_idx.y + i) * 16 + thread_idx.y;
> +        lm_idx.y = lm_idx.y < 0 ? 0 : lm_idx.y;
> +        lm_idx.y = lm_idx.y >= height ? height - 1: lm_idx.y;
> +        for (j = 0; j <= 1; j++) {
> +            lm_idx.x = -8 + (block_idx.x + j) * 16 + thread_idx.x;
> +            lm_idx.x = lm_idx.x < 0 ? 0 : lm_idx.x;
> +            lm_idx.x = lm_idx.x >= width ? width - 1: lm_idx.x;
> +            data[i*16 + thread_idx.y][j*16 + thread_idx.x] = src[lm_idx.y*width + lm_idx.x];
> +        }
> +    }
> +    local_idx = thread_idx.y*16 + thread_idx.x;
> +    if (local_idx < 128)
> +        lc[local_idx] = mask[local_idx];
> +    barrier(CLK_LOCAL_MEM_FENCE);
> +
> +    \n#pragma unroll\n
> +    for (i = -4; i <= 4; i++) {
> +        lm_idx.y = 8 + i + thread_idx.y;
> +        \n#pragma unroll\n
> +        for (j = -4; j <= 4; j++) {
> +            lm_idx.x = 8 + j + thread_idx.x;
> +            lc_idx = (i + 4)*8 + j + 4;
> +            sum += (int)data[lm_idx.y][lm_idx.x] * lc[lc_idx];
> +        }
> +    }
> +    int temp = (int)data[thread_idx.y + 8][thread_idx.x + 8];
> +    int res = temp + (((temp - (int)((sum + 1<<15) >> 16))) >> 16);
> +    if (global_idx.x < width && global_idx.y < height)
> +        dst[global_idx.x + global_idx.y*width] = clip_uint8(res);
> +}
> +);
> +
> +#endif
> diff --git a/doc/fftools-common-opts.texi b/doc/fftools-common-opts.texi
> index 617af2f..d63b3d1 100644
> --- a/doc/fftools-common-opts.texi
> +++ b/doc/fftools-common-opts.texi
> @@ -250,6 +250,10 @@ Possible flags for this option are:
>  @end table
>  @end table
>  
> + at item -opencl_bench
> +Benchmark available OpenCL devices and output the results. This option
> +is only available when FFmpeg has been compiled with @code{--enable-opencl}.
> +
>  @item -opencl_options options (@emph{global})
>  Set OpenCL environment options. This option is only available when
>  FFmpeg has been compiled with @code{--enable-opencl}.
> diff --git a/doc/utils.texi b/doc/utils.texi
> index 9377139..3055662 100644
> --- a/doc/utils.texi
> +++ b/doc/utils.texi
> @@ -1051,13 +1051,13 @@ See reference "OpenCL Specification Version: 1.2 chapter 5.6.4".
>  Select the index of the platform to run OpenCL code.
>  
>  The specified index must be one of the indexes in the device list
> -which can be obtained with @code{av_opencl_get_device_list()}.
> +which can be obtained with @code{ffmpeg -opencl_bench} or @code{av_opencl_get_device_list()}.
>  
>  @item device_idx
>  Select the index of the device used to run OpenCL code.
>  
>  The specifed index must be one of the indexes in the device list which
> -can be obtained with @code{av_opencl_get_device_list()}.
> +can be obtained with @code{ffmpeg -opencl_bench} or @code{av_opencl_get_device_list()}.
>  
>  @end table
>  
> diff --git a/libavutil/opencl.c b/libavutil/opencl.c
> index ae4c476..1ac2d4e 100644
> --- a/libavutil/opencl.c
> +++ b/libavutil/opencl.c
> @@ -761,3 +761,44 @@ int av_opencl_buffer_read_image(uint8_t **dst_data, int *plane_size, int plane_n
>      }
>      return 0;
>  }
> +
> +int64_t av_opencl_benchmark(AVOpenCLDeviceNode *device_node, cl_platform_id platform,
> +                            int64_t (*benchmark)(AVOpenCLExternalEnv *ext_opencl_env))
> +{
> +    int64_t ret = 0;
> +    cl_int status;
> +    cl_context_properties cps[3];
> +    AVOpenCLExternalEnv *ext_opencl_env = NULL;
> +
> +    ext_opencl_env = av_opencl_alloc_external_env();
> +    ext_opencl_env->device_id = device_node->device_id;
> +    ext_opencl_env->device_type = device_node->device_type;
> +    av_log(&opencl_ctx, AV_LOG_VERBOSE, "Performing test on OpenCL device %s\n",
> +           device_node->device_name);
> +
> +    cps[0] = CL_CONTEXT_PLATFORM;
> +    cps[1] = (cl_context_properties)platform;
> +    cps[2] = 0;
> +    ext_opencl_env->context = clCreateContextFromType(cps, ext_opencl_env->device_type,
> +                                                      NULL, NULL, &status);
> +    if (status != CL_SUCCESS || !ext_opencl_env->context) {
> +        ret = AVERROR_EXTERNAL;
> +        goto end;
> +    }
> +    ext_opencl_env->command_queue = clCreateCommandQueue(ext_opencl_env->context,
> +                                                         ext_opencl_env->device_id, 0, &status);
> +    if (status != CL_SUCCESS || !ext_opencl_env->command_queue) {
> +        ret = AVERROR_EXTERNAL;
> +        goto end;
> +    }
> +    ret = benchmark(ext_opencl_env);

> +    if (ret < 0)
> +        av_log(&opencl_ctx, AV_LOG_ERROR, "Benchmark failed with this device!\n");

Nit: you can provide a bit more context by specifying the device name

> +end:
> +    if (ext_opencl_env->command_queue)
> +        clReleaseCommandQueue(ext_opencl_env->command_queue);
> +    if (ext_opencl_env->context)
> +        clReleaseContext(ext_opencl_env->context);
> +    av_opencl_free_external_env(&ext_opencl_env);
> +    return ret;
> +}
> diff --git a/libavutil/opencl.h b/libavutil/opencl.h
> index e4ecbf8..8a26e06 100644
> --- a/libavutil/opencl.h
> +++ b/libavutil/opencl.h
> @@ -85,6 +85,13 @@ typedef struct {
>      char *platform_name;
>  } AVOpenCLExternalEnv;
>  

> +typedef struct {
> +    int device_idx;
> +    int platform_idx;
> +    int64_t runtime;
> +    char device_name[AV_OPENCL_MAX_DEVICE_NAME_SIZE];
> +} AVOpenCLDeviceBenchmark;

This is not used by the API and should be probably moved to the
cmdutils level.

> +
>  /**
>   * Get OpenCL device list.
>   *
> @@ -310,4 +317,19 @@ void av_opencl_release_kernel(AVOpenCLKernelEnv *env);
>   */
>  void av_opencl_uninit(void);
>  
> +/**
> + * Benchmark an OpenCL device with a user defined callback function.  This function
> + * sets up an external OpenCL environment including context and command queue on
> + * the device then tears it down in the end.  The callback function should perform
> + * the rest of the work.
> + *
> + * @param device            pointer to the OpenCL device to be used
> + * @param platform          cl_platform_id handle to which the device belongs to
> + * @param benchmark         callback function to perform the benchmark, return a
> + *                          negative value in case of failure
> + * @return the score passed from the callback function, a negative value in case of failure
> + */
> +int64_t av_opencl_benchmark(AVOpenCLDeviceNode *device, cl_platform_id platform,
> +                            int64_t (*benchmark)(AVOpenCLExternalEnv *ext_opencl_env));
> +

Looks OK otherwise, thanks.
-- 
FFmpeg = Fundamentalist and Free Mysterious Problematic Elitarian Gadget


More information about the ffmpeg-devel mailing list