From ae17d2007963f7956f437ce7ff9f7eefd8d5d2ad Mon Sep 17 00:00:00 2001 From: ling Date: Tue, 25 Aug 2020 21:00:13 +0800 Subject: [PATCH] [MS][LITE][Develop] Max Pooling optimize --- mindspore/lite/nnacl/fp32/pooling.c | 77 ++++++++++++++--------------- 1 file changed, 38 insertions(+), 39 deletions(-) diff --git a/mindspore/lite/nnacl/fp32/pooling.c b/mindspore/lite/nnacl/fp32/pooling.c index e533eec5e0e..a7d56281d34 100644 --- a/mindspore/lite/nnacl/fp32/pooling.c +++ b/mindspore/lite/nnacl/fp32/pooling.c @@ -130,12 +130,11 @@ void MaxPooling(const float *input_ptr, float *output_ptr, PoolingParameter *poo int out_plane = output_w * output_h; int out_tile_count = UP_DIV(out_plane, TILE_NUM); int thread_num = pooling_param->thread_num_; - int c4 = UP_DIV(channel, C4NUM); - // input channel is equal to output channel + int c4 = UP_DIV(channel, C4NUM); /* oc && ic */ for (int batch = 0; batch < output_batch; batch++) { - int in_batch_offset = batch * in_h * in_w * channel; - int out_batch_offset = batch * output_h * output_w * channel; + const float *src_b_ptr = input_ptr + batch * in_h * in_w * channel; + float *dst_b_ptr = output_ptr + batch * output_h * output_w * channel; for (int thread_id = task_id; thread_id < out_tile_count; thread_id += thread_num) { int cal_start_index = thread_id * TILE_NUM; int real_cal_num = (out_plane - cal_start_index) > TILE_NUM ? TILE_NUM : (out_plane - cal_start_index); @@ -145,10 +144,18 @@ void MaxPooling(const float *input_ptr, float *output_ptr, PoolingParameter *poo int out_h_index = index / output_w; int in_w_index = out_w_index * stride_w - pad_w; int in_h_index = out_h_index * stride_h - pad_h; - int out_plane_offset = out_batch_offset + index * channel; - for (int j = 0; j < c4 - 1; j++) { - int in_channel_offset = in_batch_offset + j * C4NUM; - int out_channel_offset = out_plane_offset + j * C4NUM; + + const float *src_plane_ptr = src_b_ptr; + float *dst_plane_ptr = dst_b_ptr + index * channel; + + int real_win_h_start = MSMAX(0, -in_h_index); + int real_win_h_end = MSMIN(win_h, in_h - in_h_index); + int resl_win_w_start = MSMAX(0, -in_w_index); + int real_win_w_end = MSMIN(win_w, in_w - in_w_index); + + for (int ci = 0; ci < c4 - 1; ci++) { + const float *src_c_ptr = src_plane_ptr + ci * C4NUM; + float *dst_c_ptr = dst_plane_ptr + ci * C4NUM; #ifdef ENABLE_NEON float32x4_t tmp_max = vdupq_n_f32(-FLT_MAX); #else @@ -157,51 +164,43 @@ void MaxPooling(const float *input_ptr, float *output_ptr, PoolingParameter *poo float tmp_max3 = -FLT_MAX; float tmp_max4 = -FLT_MAX; #endif - for (int h = 0; h < win_h; h++) { - for (int w = 0; w < win_w; w++) { - if ((in_h_index + h) < 0 || (in_h_index + h) >= in_h || (in_w_index + w) < 0 || - (in_w_index + w) >= in_w) { - continue; - } else { - int in_offset = in_channel_offset + ((in_h_index + h) * in_w + in_w_index + w) * channel; + + for (int kh = real_win_h_start; kh < real_win_h_end; kh++) { + for (int kw = resl_win_w_start; kw < real_win_w_end; kw++) { + const float *src_win_ptr = src_c_ptr + ((in_h_index + kh) * in_w + in_w_index + kw) * channel; #ifdef ENABLE_NEON - tmp_max = vmaxq_f32(tmp_max, vld1q_f32(input_ptr + in_offset)); + tmp_max = vmaxq_f32(tmp_max, vld1q_f32(src_win_ptr)); #else - tmp_max1 = fmax(tmp_max1, *(input_ptr + in_offset)); - tmp_max2 = fmax(tmp_max2, *(input_ptr + in_offset + 1)); - tmp_max3 = fmax(tmp_max3, *(input_ptr + in_offset + 2)); - tmp_max4 = fmax(tmp_max4, *(input_ptr + in_offset + 3)); + tmp_max1 = fmax(tmp_max1, src_win_ptr[0]); + tmp_max2 = fmax(tmp_max2, src_win_ptr[1]); + tmp_max3 = fmax(tmp_max3, src_win_ptr[2]); + tmp_max4 = fmax(tmp_max4, src_win_ptr[3]); #endif - } } // win_w loop } // win_h loop #ifdef ENABLE_NEON - vst1q_f32(output_ptr + out_channel_offset, tmp_max); + vst1q_f32(dst_c_ptr, tmp_max); #else - *(output_ptr + out_channel_offset) = tmp_max1; - *(output_ptr + out_channel_offset + 1) = tmp_max2; - *(output_ptr + out_channel_offset + 2) = tmp_max3; - *(output_ptr + out_channel_offset + 3) = tmp_max4; + dst_c_ptr[0] = tmp_max1; + dst_c_ptr[1] = tmp_max2; + dst_c_ptr[2] = tmp_max3; + dst_c_ptr[3] = tmp_max4; #endif } // ic4-1 loop int channel_s = (c4 - 1) * C4NUM; - for (int k = channel_s; k < channel; k++) { - int in_channel_offset = in_batch_offset + k; - int out_channel_offset = out_plane_offset + k; + for (int ci = channel_s; ci < channel; ci++) { + float *dst_c_ptr = dst_plane_ptr + ci; + const float *src_c_ptr = src_plane_ptr + ci; float tmp_max = -FLT_MAX; - for (int h = 0; h < win_h; h++) { - for (int w = 0; w < win_w; w++) { - if ((in_h_index + h) < 0 || (in_h_index + h) >= in_h || (in_w_index + w) < 0 || - (in_w_index + w) >= in_w) { - continue; - } else { - int in_offset = in_channel_offset + ((in_h_index + h) * in_w + in_w_index + w) * channel; - tmp_max = fmax(tmp_max, *(input_ptr + in_offset)); - } + + for (int kh = real_win_h_start; kh < real_win_h_end; kh++) { + for (int kw = resl_win_w_start; kw < real_win_w_end; kw++) { + const float *src_win_ptr = src_c_ptr + ((in_h_index + kh) * in_w + in_w_index + kw) * channel; + tmp_max = fmax(tmp_max, src_win_ptr[0]); } // win_w loop } // win_h loop - *(output_ptr + out_channel_offset) = tmp_max; + dst_c_ptr[0] = tmp_max; } // channel_res loop } // real_cal_num loop } // out_plane loop