From patchwork Fri Sep 4 15:05:13 2020 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Xu Jun X-Patchwork-Id: 22094 Return-Path: X-Original-To: patchwork@ffaux-bg.ffmpeg.org Delivered-To: patchwork@ffaux-bg.ffmpeg.org Received: from ffbox0-bg.mplayerhq.hu (ffbox0-bg.ffmpeg.org [79.124.17.100]) by ffaux.localdomain (Postfix) with ESMTP id 64E5E44B55B for ; Fri, 4 Sep 2020 18:09:11 +0300 (EEST) Received: from [127.0.1.1] (localhost [127.0.0.1]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTP id 4627F68B411; Fri, 4 Sep 2020 18:09:11 +0300 (EEST) X-Original-To: ffmpeg-devel@ffmpeg.org Delivered-To: ffmpeg-devel@ffmpeg.org Received: from smtp181.sjtu.edu.cn (smtp181.sjtu.edu.cn [202.120.2.181]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTPS id 2743A6880A5 for ; Fri, 4 Sep 2020 18:09:05 +0300 (EEST) Received: from proxy02.sjtu.edu.cn (smtp188.sjtu.edu.cn [202.120.2.188]) by smtp181.sjtu.edu.cn (Postfix) with ESMTPS id 8DC001008CBC2 for ; Fri, 4 Sep 2020 23:09:02 +0800 (CST) Received: from localhost (localhost.localdomain [127.0.0.1]) by proxy02.sjtu.edu.cn (Postfix) with ESMTP id 8FCA221B26704; Fri, 4 Sep 2020 23:09:02 +0800 (CST) X-Virus-Scanned: amavisd-new at Received: from proxy02.sjtu.edu.cn ([127.0.0.1]) by localhost (proxy02.sjtu.edu.cn [127.0.0.1]) (amavisd-new, port 10026) with ESMTP id pVuaNjiyPvst; Fri, 4 Sep 2020 23:09:02 +0800 (CST) Received: from localhost.localdomain (unknown [202.120.39.204]) (Authenticated sender: xujunzz@sjtu.edu.cn) by proxy02.sjtu.edu.cn (Postfix) with ESMTPSA id A423021B26702; Fri, 4 Sep 2020 23:09:00 +0800 (CST) From: xujunzz@sjtu.edu.cn To: ffmpeg-devel@ffmpeg.org Date: Fri, 4 Sep 2020 23:05:13 +0800 Message-Id: <20200904150511.5789-2-xujunzz@sjtu.edu.cn> X-Mailer: git-send-email 2.28.0 In-Reply-To: <20200904150511.5789-1-xujunzz@sjtu.edu.cn> References: <20200904150511.5789-1-xujunzz@sjtu.edu.cn> MIME-Version: 1.0 Subject: [FFmpeg-devel] [PATCH v4 2/2] dnn_backend_native_layer_conv2d.c:Add mutithread function X-BeenThere: ffmpeg-devel@ffmpeg.org X-Mailman-Version: 2.1.20 Precedence: list List-Id: FFmpeg development discussions and patches List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Reply-To: FFmpeg development discussions and patches Cc: xujunzz@sjtu.edu.cn Errors-To: ffmpeg-devel-bounces@ffmpeg.org Sender: "ffmpeg-devel" From: Xu Jun Use pthread to multithread dnn_execute_layer_conv2d. Can be tested with command "./ffmpeg_g -i input.png -vf \ format=yuvj420p,dnn_processing=dnn_backend=native:model= \ espcn.model:input=x:output=y:options=conv2d_threads=23 \ -y sr_native.jpg -benchmark" before patch: utime=11.238s stime=0.005s rtime=11.248s after patch: utime=20.817s stime=0.047s rtime=1.051s on my 3900X 12c24t @4.2GHz About the increase of utime, it's because that CPU HyperThreading technology makes logical cores twice of physical cores while cpu's counting performance improves less than double. And utime sums all cpu's logical cores' runtime. As a result, using threads num near cpu's logical core's number will double utime, while reduce rtime less than half for HyperThreading CPUs. Signed-off-by: Xu Jun --- v2: add check for HAVE_PTHREAD_CANCEL and modify FATE test dnn-layer-conv2d-test.c v4: use extern to call dnn_native_class in dnn-layer-conv2d-test.c .../dnn/dnn_backend_native_layer_conv2d.c | 107 ++++++++++++++++-- tests/dnn/dnn-layer-conv2d-test.c | 14 ++- 2 files changed, 108 insertions(+), 13 deletions(-) diff --git a/libavfilter/dnn/dnn_backend_native_layer_conv2d.c b/libavfilter/dnn/dnn_backend_native_layer_conv2d.c index d079795bf8..4068a13ab4 100644 --- a/libavfilter/dnn/dnn_backend_native_layer_conv2d.c +++ b/libavfilter/dnn/dnn_backend_native_layer_conv2d.c @@ -19,10 +19,27 @@ */ #include "libavutil/avassert.h" +#include "libavutil/thread.h" +#include "libavutil/cpu.h" #include "dnn_backend_native_layer_conv2d.h" #define CLAMP_TO_EDGE(x, w) ((x) < 0 ? 0 : ((x) >= (w) ? (w - 1) : (x))) +//struct to pass parameters +typedef struct thread_common_param{ + DnnOperand *operands; + const int32_t *input_operand_indexes; + int32_t output_operand_index; + const void *parameters; + NativeContext *ctx; + int thread_num; +} thread_common_param; + +typedef struct thread_param{ + thread_common_param *thread_common_param; + int thread_index; +} thread_param; + int dnn_load_layer_conv2d(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num) { ConvolutionalParams *conv_params; @@ -88,17 +105,20 @@ int dnn_load_layer_conv2d(Layer *layer, AVIOContext *model_file_context, int fil return dnn_size; } -int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_indexes, - int32_t output_operand_index, const void *parameters, NativeContext *ctx) +static void * dnn_execute_layer_conv2d_thread(void *threadarg) { + //pass parameters + thread_param *thread_param = (struct thread_param *)threadarg; + thread_common_param *thread_common_param = thread_param->thread_common_param; + DnnOperand *operands = thread_common_param->operands; float *output; - int32_t input_operand_index = input_operand_indexes[0]; + int32_t input_operand_index = thread_common_param->input_operand_indexes[0]; int number = operands[input_operand_index].dims[0]; int height = operands[input_operand_index].dims[1]; int width = operands[input_operand_index].dims[2]; int channel = operands[input_operand_index].dims[3]; const float *input = operands[input_operand_index].data; - const ConvolutionalParams *conv_params = (const ConvolutionalParams *)parameters; + const ConvolutionalParams *conv_params = (const ConvolutionalParams *)(thread_common_param->parameters); int radius = conv_params->kernel_size >> 1; int src_linesize = width * conv_params->input_num; @@ -106,7 +126,11 @@ int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_ int filter_size = conv_params->kernel_size * filter_linesize; int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 * conv_params->dilation : 0; - DnnOperand *output_operand = &operands[output_operand_index]; + int thread_stride = (height - pad_size * 2) / thread_common_param->thread_num; + int thread_start = thread_stride * thread_param->thread_index + pad_size; + int thread_end = (thread_param->thread_index == thread_common_param->thread_num - 1) ? (height - pad_size) : (thread_start + thread_stride); + + DnnOperand *output_operand = &operands[thread_common_param->output_operand_index]; output_operand->dims[0] = number; output_operand->dims[1] = height - pad_size * 2; output_operand->dims[2] = width - pad_size * 2; @@ -114,19 +138,21 @@ int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_ output_operand->data_type = operands[input_operand_index].data_type; output_operand->length = calculate_operand_data_length(output_operand); if (output_operand->length <= 0) { - av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n"); - return DNN_ERROR; + av_log(thread_common_param->ctx, AV_LOG_ERROR, "The output data length overflow\n"); + return (void *)DNN_ERROR; } output_operand->data = av_realloc(output_operand->data, output_operand->length); if (!output_operand->data) { - av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n"); - return DNN_ERROR; + av_log(thread_common_param->ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n"); + return (void *)DNN_ERROR; } + output = output_operand->data; + output += (conv_params->output_num) * (width - 2 * pad_size) * (thread_start - pad_size); av_assert0(channel == conv_params->input_num); - for (int y = pad_size; y < height - pad_size; ++y) { + for (int y = thread_start; y < thread_end; ++y) { for (int x = pad_size; x < width - pad_size; ++x) { for (int n_filter = 0; n_filter < conv_params->output_num; ++n_filter) { if (conv_params->has_bias) @@ -174,5 +200,64 @@ int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_ output += conv_params->output_num; } } - return 0; + return (void *)0; +} + + +int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_indexes, + int32_t output_operand_index, const void *parameters, NativeContext *ctx) +{ + int thread_num = (ctx->options.conv2d_threads <= 0 || ctx->options.conv2d_threads > av_cpu_count()) + ? (av_cpu_count() + 1) : (ctx->options.conv2d_threads); +#if HAVE_PTHREAD_CANCEL + pthread_t *thread_id = av_malloc(thread_num * sizeof(pthread_t)); +#endif + thread_param **thread_param = av_malloc(thread_num * sizeof(*thread_param)); + void *res; + int error_flag = 0; + + //struct used to pass parameters + thread_common_param thread_common_param; + thread_common_param.operands = operands; + thread_common_param.input_operand_indexes = input_operand_indexes; + thread_common_param.output_operand_index = output_operand_index; + thread_common_param.parameters = parameters; + thread_common_param.ctx = ctx; +#if HAVE_PTHREAD_CANCEL + thread_common_param.thread_num = thread_num; + + //create threads + for (int i = 0; i < thread_num; i++){ + thread_param[i] = av_malloc(sizeof(thread_param)); + thread_param[i]->thread_common_param = &thread_common_param; + thread_param[i]->thread_index = i; + pthread_create(&thread_id[i], NULL, dnn_execute_layer_conv2d_thread, (void *)thread_param[i]); + } + + //join threads, res gets function return + for (int i = 0; i < thread_num; i++){ + pthread_join(thread_id[i], &res); + if ((int)res != 0) + error_flag = (int)res; + } + + //release memory + av_free(thread_id); + + for (int i = 0; i < thread_num; i++){ + av_free(thread_param[i]); + } +#else + thread_common_param.thread_num = 1; + thread_param[0] = av_malloc(sizeof(thread_param)); + thread_param[0]->thread_common_param = &thread_common_param; + thread_param[0]->thread_index = 0; + res = dnn_execute_layer_conv2d_thread((void *)thread_param[0]); + if ((int)res != 0) + error_flag = (int)res; + av_free(thread_param[0]); +#endif + + av_free(thread_param); + return error_flag; } diff --git a/tests/dnn/dnn-layer-conv2d-test.c b/tests/dnn/dnn-layer-conv2d-test.c index 836839cc64..378a05eafc 100644 --- a/tests/dnn/dnn-layer-conv2d-test.c +++ b/tests/dnn/dnn-layer-conv2d-test.c @@ -25,6 +25,8 @@ #define EPSON 0.00001 +extern const AVClass dnn_native_class; + static int test_with_same_dilate(void) { // the input data and expected data are generated with below python code. @@ -96,6 +98,10 @@ static int test_with_same_dilate(void) }; float bias[2] = { -1.6574852, -0.72915393 }; + NativeContext ctx; + ctx.class = &dnn_native_class; + ctx.options.conv2d_threads = 1; + params.activation = TANH; params.has_bias = 1; params.biases = bias; @@ -114,7 +120,7 @@ static int test_with_same_dilate(void) operands[1].data = NULL; input_indexes[0] = 0; - dnn_execute_layer_conv2d(operands, input_indexes, 1, ¶ms, NULL); + dnn_execute_layer_conv2d(operands, input_indexes, 1, ¶ms, &ctx); output = operands[1].data; for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) { @@ -196,6 +202,10 @@ static int test_with_valid(void) }; float bias[2] = { -0.4773722, -0.19620377 }; + NativeContext ctx; + ctx.class = &dnn_native_class; + ctx.options.conv2d_threads = 1; + params.activation = TANH; params.has_bias = 1; params.biases = bias; @@ -214,7 +224,7 @@ static int test_with_valid(void) operands[1].data = NULL; input_indexes[0] = 0; - dnn_execute_layer_conv2d(operands, input_indexes, 1, ¶ms, NULL); + dnn_execute_layer_conv2d(operands, input_indexes, 1, ¶ms, &ctx); output = operands[1].data; for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) {