diff --git a/paddle/fluid/operators/detection/density_prior_box_op.h b/paddle/fluid/operators/detection/density_prior_box_op.h index 3591681fc3f6951dfc8d73e8edce38180b771eaf..42137215e21af1a529563ecc995a54d610120beb 100644 --- a/paddle/fluid/operators/detection/density_prior_box_op.h +++ b/paddle/fluid/operators/detection/density_prior_box_op.h @@ -72,7 +72,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { #ifdef PADDLE_WITH_MKLML #pragma omp parallel for #endif - for (int i = 0; i < fixed_ratios.size(); i++) { + for (size_t i = 0; i < fixed_ratios.size(); i++) { sqrt_fixed_ratios.push_back(sqrt(fixed_ratios[i])); } @@ -115,11 +115,10 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { } } if (clip) { - platform::Transform trans; - ClipFunctor clip_func; - trans(ctx.template device_context(), - boxes->data(), boxes->data() + boxes->numel(), - boxes->data(), clip_func); + T* dt = boxes->data(); + std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T { + return std::min(std::max(v, 0.), 1.); + }); } framework::Tensor var_t; var_t.mutable_data( @@ -141,7 +140,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel { #pragma omp parallel for collapse(2) #endif for (int i = 0; i < box_num; ++i) { - for (int j = 0; j < variances.size(); ++j) { + for (size_t j = 0; j < variances.size(); ++j) { e_vars(i, j) = variances[j]; } } diff --git a/paddle/fluid/operators/detection/prior_box_op.h b/paddle/fluid/operators/detection/prior_box_op.h index 4e226abbb51c271502f0ca5419d488643b5a1a82..f84405664596ebe25983e5acbbb82bfc18c38124 100644 --- a/paddle/fluid/operators/detection/prior_box_op.h +++ b/paddle/fluid/operators/detection/prior_box_op.h @@ -46,13 +46,6 @@ inline void ExpandAspectRatios(const std::vector& input_aspect_ratior, } } -template -struct ClipFunctor { - HOSTDEVICE inline T operator()(T in) const { - return std::min(std::max(in, 0.), 1.); - } -}; - template class PriorBoxOpKernel : public framework::OpKernel { public: @@ -101,31 +94,30 @@ class PriorBoxOpKernel : public framework::OpKernel { boxes->mutable_data(ctx.GetPlace()); vars->mutable_data(ctx.GetPlace()); - auto e_boxes = framework::EigenTensor::From(*boxes); + T* b_t = boxes->data(); for (int h = 0; h < feature_height; ++h) { for (int w = 0; w < feature_width; ++w) { T center_x = (w + offset) * step_width; T center_y = (h + offset) * step_height; T box_width, box_height; - int idx = 0; for (size_t s = 0; s < min_sizes.size(); ++s) { auto min_size = min_sizes[s]; if (min_max_aspect_ratios_order) { box_width = box_height = min_size / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; if (max_sizes.size() > 0) { auto max_size = max_sizes[s]; // square prior with size sqrt(minSize * maxSize) box_width = box_height = sqrt(min_size * max_size) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } // priors with different aspect ratios for (size_t r = 0; r < aspect_ratios.size(); ++r) { @@ -135,11 +127,11 @@ class PriorBoxOpKernel : public framework::OpKernel { } box_width = min_size * sqrt(ar) / 2.; box_height = min_size / sqrt(ar) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } } else { // priors with different aspect ratios @@ -147,21 +139,21 @@ class PriorBoxOpKernel : public framework::OpKernel { float ar = aspect_ratios[r]; box_width = min_size * sqrt(ar) / 2.; box_height = min_size / sqrt(ar) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } if (max_sizes.size() > 0) { auto max_size = max_sizes[s]; // square prior with size sqrt(minSize * maxSize) box_width = box_height = sqrt(min_size * max_size) / 2.; - e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; - e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; - e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; - e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; - idx++; + b_t[0] = (center_x - box_width) / img_width; + b_t[1] = (center_y - box_height) / img_height; + b_t[2] = (center_x + box_width) / img_width; + b_t[3] = (center_y + box_height) / img_height; + b_t += 4; } } } @@ -169,11 +161,10 @@ class PriorBoxOpKernel : public framework::OpKernel { } if (clip) { - platform::Transform trans; - ClipFunctor clip_func; - trans(ctx.template device_context(), - boxes->data(), boxes->data() + boxes->numel(), - boxes->data(), clip_func); + T* dt = boxes->data(); + std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T { + return std::min(std::max(v, 0.), 1.); + }); } framework::Tensor var_t;