hyperray 0.0.1__py3-none-any.whl

hyperdrone/_vendor/rl-tools/include/rl_tools/nn/layers/dynamic_conv2d/operations_generic.h

@@ -0,0 +1,536 @@
+#include "../../../version.h"
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_NN_LAYERS_DYNAMIC_CONV2D_OPERATIONS_GENERIC_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_NN_LAYERS_DYNAMIC_CONV2D_OPERATIONS_GENERIC_H
+
+#include "../../../containers/tensor/tensor.h"
+#include "layer.h"
+#include "../../../rl_tools.h"
+
+RL_TOOLS_NAMESPACE_WRAPPER_START
+namespace rl_tools{
+    // ======================== malloc / free ========================
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::dynamic_conv2d::LayerForward<SPEC>& layer) {}
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::dynamic_conv2d::LayerForward<SPEC>& layer) {}
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::dynamic_conv2d::LayerBackward<SPEC>& layer) {
+        malloc(device, layer.pre_activations);
+    }
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::dynamic_conv2d::LayerBackward<SPEC>& layer) {
+        free(device, layer.pre_activations);
+    }
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SPEC>& layer) {
+        malloc(device, static_cast<nn::layers::dynamic_conv2d::LayerBackward<SPEC>&>(layer));
+        malloc(device, layer.output);
+    }
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SPEC>& layer) {
+        free(device, layer.output);
+        free(device, static_cast<nn::layers::dynamic_conv2d::LayerBackward<SPEC>&>(layer));
+    }
+    template<typename DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::dynamic_conv2d::State&) {}
+    template<typename DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::dynamic_conv2d::State&) {}
+    template<typename DEVICE, typename BUFFER_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void malloc(DEVICE& device, nn::layers::dynamic_conv2d::Buffer<BUFFER_SPEC>& buffer) {
+        malloc(device, buffer.d_data_acc);
+        malloc(device, buffer.d_kernel_weights_acc);
+    }
+    template<typename DEVICE, typename BUFFER_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void free(DEVICE& device, nn::layers::dynamic_conv2d::Buffer<BUFFER_SPEC>& buffer) {
+        free(device, buffer.d_data_acc);
+        free(device, buffer.d_kernel_weights_acc);
+    }
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy(SOURCE_DEVICE&, TARGET_DEVICE&, nn::layers::dynamic_conv2d::State&, nn::layers::dynamic_conv2d::State&) {}
+    template<typename SPEC, typename DEVICE, typename RNG, typename MODE>
+    RL_TOOLS_FUNCTION_PLACEMENT void reset(DEVICE&, const nn::layers::dynamic_conv2d::LayerForward<SPEC>&, nn::layers::dynamic_conv2d::State&, RNG&, Mode<MODE>) {}
+
+    // ======================== init_weights (no-op) ========================
+    template<typename DEVICE, typename SPEC, typename RNG>
+    RL_TOOLS_FUNCTION_PLACEMENT void init_weights(DEVICE& device, nn::layers::dynamic_conv2d::LayerForward<SPEC>& layer, RNG& rng) {}
+
+    // ======================== evaluate ========================
+#ifndef RL_TOOLS_NN_DISABLE_GENERIC_FORWARD_BACKWARD
+    template<typename DEVICE, typename LAYER_SPEC, typename DATA_SPEC, typename KERNEL_WEIGHTS_SPEC, typename OUTPUT_SPEC, typename BUFFER_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void evaluate(DEVICE& device, const nn::layers::dynamic_conv2d::LayerForward<LAYER_SPEC>& layer, const Tensor<DATA_SPEC>& data, const Tensor<KERNEL_WEIGHTS_SPEC>& kernel_weights, Tensor<OUTPUT_SPEC>& output, nn::layers::dynamic_conv2d::Buffer<BUFFER_SPEC>&, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}) {
+        static_assert(nn::layers::dynamic_conv2d::check_data_output<LAYER_SPEC, DATA_SPEC, OUTPUT_SPEC>);
+        using TI = typename DEVICE::index_t;
+        using T = typename OUTPUT_SPEC::T;
+        using ACCUMULATOR_TYPE = typename LAYER_SPEC::TYPE_POLICY::template GET<numeric_types::categories::Accumulator>;
+        constexpr TI BATCH_SIZE = LAYER_SPEC::INTERNAL_BATCH_SIZE;
+        using INTERNAL_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, LAYER_SPEC::INPUT_CHANNELS>;
+        auto data_4d = view_memory<INTERNAL_DATA_SHAPE>(device, data);
+        if constexpr(LAYER_SPEC::IS_FULL_CONV){
+            constexpr TI OC = LAYER_SPEC::OUTPUT_CHANNELS;
+            constexpr TI IC = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_KERNEL_WEIGHTS_SHAPE = tensor::Shape<TI, BATCH_SIZE, OC, IC, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            using INTERNAL_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, OC>;
+            auto kw_5d = view_memory<INTERNAL_KERNEL_WEIGHTS_SHAPE>(device, kernel_weights);
+            auto output_4d = view_memory<INTERNAL_OUTPUT_SHAPE>(device, output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c_out = 0; c_out < OC; c_out++){
+                            ACCUMULATOR_TYPE acc = 0;
+                            for(TI c_in = 0; c_in < IC; c_in++){
+                                for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                    for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                        TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                        TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                        if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                           iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                            TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                            TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                            acc += (ACCUMULATOR_TYPE)get(device, kw_5d, bi, c_out, c_in, kh, kw) * (ACCUMULATOR_TYPE)get(device, data_4d, bi, ih, iw, c_in);
+                                        }
+                                    }
+                                }
+                            }
+                            set(device, output_4d, (T)activation<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>(acc), bi, oh, ow, c_out);
+                        }
+                    }
+                }
+            }
+        } else {
+            constexpr TI C = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_KERNEL_WEIGHTS_SHAPE = tensor::Shape<TI, BATCH_SIZE, C, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            using INTERNAL_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, C>;
+            auto kw_4d = view_memory<INTERNAL_KERNEL_WEIGHTS_SHAPE>(device, kernel_weights);
+            auto output_4d = view_memory<INTERNAL_OUTPUT_SHAPE>(device, output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c = 0; c < C; c++){
+                            ACCUMULATOR_TYPE acc = 0;
+                            for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                    TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                    TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                    if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                       iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                        TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                        TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                        acc += (ACCUMULATOR_TYPE)get(device, kw_4d, bi, c, kh, kw) * (ACCUMULATOR_TYPE)get(device, data_4d, bi, ih, iw, c);
+                                    }
+                                }
+                            }
+                            set(device, output_4d, (T)activation<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>(acc), bi, oh, ow, c);
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    // ======================== forward (LayerBackward — stores pre-activations) ========================
+    template<typename DEVICE, typename LAYER_SPEC, typename DATA_SPEC, typename KERNEL_WEIGHTS_SPEC, typename OUTPUT_SPEC, typename BUFFER_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void forward(DEVICE& device, nn::layers::dynamic_conv2d::LayerBackward<LAYER_SPEC>& layer, const Tensor<DATA_SPEC>& data, const Tensor<KERNEL_WEIGHTS_SPEC>& kernel_weights, Tensor<OUTPUT_SPEC>& output, nn::layers::dynamic_conv2d::Buffer<BUFFER_SPEC>& buffer, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        static_assert(nn::layers::dynamic_conv2d::check_data_output<LAYER_SPEC, DATA_SPEC, OUTPUT_SPEC>);
+        using TI = typename DEVICE::index_t;
+        using T = typename OUTPUT_SPEC::T;
+        using ACCUMULATOR_TYPE = typename LAYER_SPEC::TYPE_POLICY::template GET<numeric_types::categories::Accumulator>;
+        constexpr TI BATCH_SIZE = LAYER_SPEC::INTERNAL_BATCH_SIZE;
+        using INTERNAL_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, LAYER_SPEC::INPUT_CHANNELS>;
+        auto data_4d = view_memory<INTERNAL_DATA_SHAPE>(device, data);
+        if constexpr(LAYER_SPEC::IS_FULL_CONV){
+            constexpr TI OC = LAYER_SPEC::OUTPUT_CHANNELS;
+            constexpr TI IC = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_KERNEL_WEIGHTS_SHAPE = tensor::Shape<TI, BATCH_SIZE, OC, IC, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            using INTERNAL_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, OC>;
+            auto kw_5d = view_memory<INTERNAL_KERNEL_WEIGHTS_SHAPE>(device, kernel_weights);
+            auto output_4d = view_memory<INTERNAL_OUTPUT_SHAPE>(device, output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c_out = 0; c_out < OC; c_out++){
+                            ACCUMULATOR_TYPE acc = 0;
+                            for(TI c_in = 0; c_in < IC; c_in++){
+                                for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                    for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                        TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                        TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                        if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                           iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                            TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                            TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                            acc += (ACCUMULATOR_TYPE)get(device, kw_5d, bi, c_out, c_in, kh, kw) * (ACCUMULATOR_TYPE)get(device, data_4d, bi, ih, iw, c_in);
+                                        }
+                                    }
+                                }
+                            }
+                            set(device, layer.pre_activations, (T)acc, bi, oh, ow, c_out);
+                            set(device, output_4d, (T)activation<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>(acc), bi, oh, ow, c_out);
+                        }
+                    }
+                }
+            }
+        } else {
+            constexpr TI C = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_KERNEL_WEIGHTS_SHAPE = tensor::Shape<TI, BATCH_SIZE, C, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            using INTERNAL_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, C>;
+            auto kw_4d = view_memory<INTERNAL_KERNEL_WEIGHTS_SHAPE>(device, kernel_weights);
+            auto output_4d = view_memory<INTERNAL_OUTPUT_SHAPE>(device, output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c = 0; c < C; c++){
+                            ACCUMULATOR_TYPE acc = 0;
+                            for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                    TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                    TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                    if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                       iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                        TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                        TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                        acc += (ACCUMULATOR_TYPE)get(device, kw_4d, bi, c, kh, kw) * (ACCUMULATOR_TYPE)get(device, data_4d, bi, ih, iw, c);
+                                    }
+                                }
+                            }
+                            set(device, layer.pre_activations, (T)acc, bi, oh, ow, c);
+                            set(device, output_4d, (T)activation<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>(acc), bi, oh, ow, c);
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    template<typename DEVICE, typename LAYER_SPEC, typename DATA_SPEC, typename KERNEL_WEIGHTS_SPEC, typename BUFFER_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void forward(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<LAYER_SPEC>& layer, const Tensor<DATA_SPEC>& data, const Tensor<KERNEL_WEIGHTS_SPEC>& kernel_weights, nn::layers::dynamic_conv2d::Buffer<BUFFER_SPEC>& buffer, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}) {
+        forward(device, static_cast<nn::layers::dynamic_conv2d::LayerBackward<LAYER_SPEC>&>(layer), data, kernel_weights, layer.output, buffer, rng, mode);
+    }
+    template<typename DEVICE, typename LAYER_SPEC, typename DATA_SPEC, typename KERNEL_WEIGHTS_SPEC, typename OUTPUT_SPEC, typename BUFFER_SPEC, typename RNG, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void forward(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<LAYER_SPEC>& layer, const Tensor<DATA_SPEC>& data, const Tensor<KERNEL_WEIGHTS_SPEC>& kernel_weights, Tensor<OUTPUT_SPEC>& output, nn::layers::dynamic_conv2d::Buffer<BUFFER_SPEC>& buffer, RNG& rng, const Mode<MODE>& mode = Mode<mode::Default<>>{}) {
+        forward(device, layer, data, kernel_weights, buffer, rng, mode);
+        copy(device, device, layer.output, output);
+    }
+
+    // ======================== backward_input (d_data only) ========================
+    template<typename DEVICE, typename LAYER_SPEC, typename KERNEL_WEIGHTS_SPEC, typename D_OUTPUT_SPEC, typename D_DATA_SPEC, typename BUFFER_SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void backward_input(DEVICE& device, const nn::layers::dynamic_conv2d::LayerBackward<LAYER_SPEC>& layer, const Tensor<KERNEL_WEIGHTS_SPEC>& kernel_weights, const Tensor<D_OUTPUT_SPEC>& d_output, Tensor<D_DATA_SPEC>& d_data, nn::layers::dynamic_conv2d::Buffer<BUFFER_SPEC>& buffer, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        using TI = typename DEVICE::index_t;
+        using T = typename D_DATA_SPEC::T;
+        using ACCUMULATOR_TYPE = typename LAYER_SPEC::TYPE_POLICY::template GET<numeric_types::categories::Accumulator>;
+        constexpr TI BATCH_SIZE = LAYER_SPEC::INTERNAL_BATCH_SIZE;
+        set_all(device, buffer.d_data_acc, (ACCUMULATOR_TYPE)0);
+        if constexpr(LAYER_SPEC::IS_FULL_CONV){
+            constexpr TI OC = LAYER_SPEC::OUTPUT_CHANNELS;
+            constexpr TI IC = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_KERNEL_WEIGHTS_SHAPE = tensor::Shape<TI, BATCH_SIZE, OC, IC, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            using INTERNAL_D_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, OC>;
+            auto kw_5d = view_memory<INTERNAL_KERNEL_WEIGHTS_SHAPE>(device, kernel_weights);
+            auto d_output_4d = view_memory<INTERNAL_D_OUTPUT_SHAPE>(device, d_output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c_out = 0; c_out < OC; c_out++){
+                            ACCUMULATOR_TYPE d_pre_act = d_activation_d_x<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>((ACCUMULATOR_TYPE)get(device, layer.pre_activations, bi, oh, ow, c_out)) * (ACCUMULATOR_TYPE)get(device, d_output_4d, bi, oh, ow, c_out);
+                            for(TI c_in = 0; c_in < IC; c_in++){
+                                for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                    for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                        TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                        TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                        if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                           iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                            TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                            TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                            increment(device, buffer.d_data_acc, (ACCUMULATOR_TYPE)get(device, kw_5d, bi, c_out, c_in, kh, kw) * d_pre_act, bi, ih, iw, c_in);
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            using INTERNAL_D_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, IC>;
+            auto d_data_4d = view_memory<INTERNAL_D_DATA_SHAPE>(device, d_data);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++) for(TI ih = 0; ih < LAYER_SPEC::INPUT_HEIGHT; ih++) for(TI iw = 0; iw < LAYER_SPEC::INPUT_WIDTH; iw++) for(TI c_in = 0; c_in < IC; c_in++)
+                set(device, d_data_4d, (T)get(device, buffer.d_data_acc, bi, ih, iw, c_in), bi, ih, iw, c_in);
+        } else {
+            constexpr TI C = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_KERNEL_WEIGHTS_SHAPE = tensor::Shape<TI, BATCH_SIZE, C, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            using INTERNAL_D_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, C>;
+            auto kw_4d = view_memory<INTERNAL_KERNEL_WEIGHTS_SHAPE>(device, kernel_weights);
+            auto d_output_4d = view_memory<INTERNAL_D_OUTPUT_SHAPE>(device, d_output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c = 0; c < C; c++){
+                            ACCUMULATOR_TYPE d_pre_act = d_activation_d_x<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>((ACCUMULATOR_TYPE)get(device, layer.pre_activations, bi, oh, ow, c)) * (ACCUMULATOR_TYPE)get(device, d_output_4d, bi, oh, ow, c);
+                            for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                    TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                    TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                    if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                       iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                        TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                        TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                        increment(device, buffer.d_data_acc, (ACCUMULATOR_TYPE)get(device, kw_4d, bi, c, kh, kw) * d_pre_act, bi, ih, iw, c);
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            using INTERNAL_D_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, C>;
+            auto d_data_4d = view_memory<INTERNAL_D_DATA_SHAPE>(device, d_data);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++) for(TI ih = 0; ih < LAYER_SPEC::INPUT_HEIGHT; ih++) for(TI iw = 0; iw < LAYER_SPEC::INPUT_WIDTH; iw++) for(TI c = 0; c < C; c++)
+                set(device, d_data_4d, (T)get(device, buffer.d_data_acc, bi, ih, iw, c), bi, ih, iw, c);
+        }
+    }
+
+    // ======================== backward_kernel_weights (d_kernel_weights only) ========================
+    template<typename DEVICE, typename LAYER_SPEC, typename DATA_SPEC, typename D_OUTPUT_SPEC, typename D_KERNEL_WEIGHTS_SPEC, typename BUFFER_SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void backward_kernel_weights(DEVICE& device, const nn::layers::dynamic_conv2d::LayerBackward<LAYER_SPEC>& layer, const Tensor<DATA_SPEC>& data, const Tensor<D_OUTPUT_SPEC>& d_output, Tensor<D_KERNEL_WEIGHTS_SPEC>& d_kernel_weights, nn::layers::dynamic_conv2d::Buffer<BUFFER_SPEC>& buffer, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        using TI = typename DEVICE::index_t;
+        using T = typename D_KERNEL_WEIGHTS_SPEC::T;
+        using ACCUMULATOR_TYPE = typename LAYER_SPEC::TYPE_POLICY::template GET<numeric_types::categories::Accumulator>;
+        constexpr TI BATCH_SIZE = LAYER_SPEC::INTERNAL_BATCH_SIZE;
+        set_all(device, buffer.d_kernel_weights_acc, (ACCUMULATOR_TYPE)0);
+        if constexpr(LAYER_SPEC::IS_FULL_CONV){
+            constexpr TI OC = LAYER_SPEC::OUTPUT_CHANNELS;
+            constexpr TI IC = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, IC>;
+            using INTERNAL_D_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, OC>;
+            auto data_4d = view_memory<INTERNAL_DATA_SHAPE>(device, data);
+            auto d_output_4d = view_memory<INTERNAL_D_OUTPUT_SHAPE>(device, d_output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c_out = 0; c_out < OC; c_out++){
+                            ACCUMULATOR_TYPE d_pre_act = d_activation_d_x<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>((ACCUMULATOR_TYPE)get(device, layer.pre_activations, bi, oh, ow, c_out)) * (ACCUMULATOR_TYPE)get(device, d_output_4d, bi, oh, ow, c_out);
+                            for(TI c_in = 0; c_in < IC; c_in++){
+                                for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                    for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                        TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                        TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                        if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                           iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                            TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                            TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                            increment(device, buffer.d_kernel_weights_acc, (ACCUMULATOR_TYPE)get(device, data_4d, bi, ih, iw, c_in) * d_pre_act, bi, c_out, c_in, kh, kw);
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            using INTERNAL_D_KW_SHAPE = tensor::Shape<TI, BATCH_SIZE, OC, IC, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            auto d_kw_5d = view_memory<INTERNAL_D_KW_SHAPE>(device, d_kernel_weights);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++) for(TI c_out = 0; c_out < OC; c_out++) for(TI c_in = 0; c_in < IC; c_in++) for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++) for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++)
+                set(device, d_kw_5d, (T)get(device, buffer.d_kernel_weights_acc, bi, c_out, c_in, kh, kw), bi, c_out, c_in, kh, kw);
+        } else {
+            constexpr TI C = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, C>;
+            using INTERNAL_D_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, C>;
+            auto data_4d = view_memory<INTERNAL_DATA_SHAPE>(device, data);
+            auto d_output_4d = view_memory<INTERNAL_D_OUTPUT_SHAPE>(device, d_output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c = 0; c < C; c++){
+                            ACCUMULATOR_TYPE d_pre_act = d_activation_d_x<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>((ACCUMULATOR_TYPE)get(device, layer.pre_activations, bi, oh, ow, c)) * (ACCUMULATOR_TYPE)get(device, d_output_4d, bi, oh, ow, c);
+                            for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                    TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                    TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                    if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                       iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                        TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                        TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                        increment(device, buffer.d_kernel_weights_acc, (ACCUMULATOR_TYPE)get(device, data_4d, bi, ih, iw, c) * d_pre_act, bi, c, kh, kw);
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            using INTERNAL_D_KW_SHAPE = tensor::Shape<TI, BATCH_SIZE, C, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            auto d_kw_4d = view_memory<INTERNAL_D_KW_SHAPE>(device, d_kernel_weights);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++) for(TI c = 0; c < C; c++) for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++) for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++)
+                set(device, d_kw_4d, (T)get(device, buffer.d_kernel_weights_acc, bi, c, kh, kw), bi, c, kh, kw);
+        }
+    }
+
+    // ======================== backward_full (d_data + d_kernel_weights, fused) ========================
+    template<typename DEVICE, typename LAYER_SPEC, typename DATA_SPEC, typename KERNEL_WEIGHTS_SPEC, typename D_OUTPUT_SPEC, typename D_DATA_SPEC, typename D_KERNEL_WEIGHTS_SPEC, typename BUFFER_SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT void backward_full(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<LAYER_SPEC>& layer, const Tensor<DATA_SPEC>& data, const Tensor<KERNEL_WEIGHTS_SPEC>& kernel_weights, Tensor<D_OUTPUT_SPEC>& d_output, Tensor<D_DATA_SPEC>& d_data, Tensor<D_KERNEL_WEIGHTS_SPEC>& d_kernel_weights, nn::layers::dynamic_conv2d::Buffer<BUFFER_SPEC>& buffer, const Mode<MODE>& mode = Mode<mode::Default<>>{}){
+        using TI = typename DEVICE::index_t;
+        using T_DATA = typename D_DATA_SPEC::T;
+        using T_KW = typename D_KERNEL_WEIGHTS_SPEC::T;
+        using ACCUMULATOR_TYPE = typename LAYER_SPEC::TYPE_POLICY::template GET<numeric_types::categories::Accumulator>;
+        constexpr TI BATCH_SIZE = LAYER_SPEC::INTERNAL_BATCH_SIZE;
+        set_all(device, buffer.d_data_acc, (ACCUMULATOR_TYPE)0);
+        set_all(device, buffer.d_kernel_weights_acc, (ACCUMULATOR_TYPE)0);
+        if constexpr(LAYER_SPEC::IS_FULL_CONV){
+            constexpr TI OC = LAYER_SPEC::OUTPUT_CHANNELS;
+            constexpr TI IC = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, IC>;
+            using INTERNAL_KERNEL_WEIGHTS_SHAPE = tensor::Shape<TI, BATCH_SIZE, OC, IC, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            using INTERNAL_D_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, OC>;
+            auto data_4d = view_memory<INTERNAL_DATA_SHAPE>(device, data);
+            auto kw_5d = view_memory<INTERNAL_KERNEL_WEIGHTS_SHAPE>(device, kernel_weights);
+            auto d_output_4d = view_memory<INTERNAL_D_OUTPUT_SHAPE>(device, d_output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c_out = 0; c_out < OC; c_out++){
+                            ACCUMULATOR_TYPE d_pre_act = d_activation_d_x<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>((ACCUMULATOR_TYPE)get(device, layer.pre_activations, bi, oh, ow, c_out)) * (ACCUMULATOR_TYPE)get(device, d_output_4d, bi, oh, ow, c_out);
+                            for(TI c_in = 0; c_in < IC; c_in++){
+                                for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                    for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                        TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                        TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                        if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                           iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                            TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                            TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                            increment(device, buffer.d_data_acc, (ACCUMULATOR_TYPE)get(device, kw_5d, bi, c_out, c_in, kh, kw) * d_pre_act, bi, ih, iw, c_in);
+                                            increment(device, buffer.d_kernel_weights_acc, (ACCUMULATOR_TYPE)get(device, data_4d, bi, ih, iw, c_in) * d_pre_act, bi, c_out, c_in, kh, kw);
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            using INTERNAL_D_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, IC>;
+            using INTERNAL_D_KW_SHAPE = tensor::Shape<TI, BATCH_SIZE, OC, IC, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            auto d_data_4d = view_memory<INTERNAL_D_DATA_SHAPE>(device, d_data);
+            auto d_kw_5d = view_memory<INTERNAL_D_KW_SHAPE>(device, d_kernel_weights);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++) for(TI ih = 0; ih < LAYER_SPEC::INPUT_HEIGHT; ih++) for(TI iw = 0; iw < LAYER_SPEC::INPUT_WIDTH; iw++) for(TI c_in = 0; c_in < IC; c_in++)
+                set(device, d_data_4d, (T_DATA)get(device, buffer.d_data_acc, bi, ih, iw, c_in), bi, ih, iw, c_in);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++) for(TI c_out = 0; c_out < OC; c_out++) for(TI c_in = 0; c_in < IC; c_in++) for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++) for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++)
+                set(device, d_kw_5d, (T_KW)get(device, buffer.d_kernel_weights_acc, bi, c_out, c_in, kh, kw), bi, c_out, c_in, kh, kw);
+        } else {
+            constexpr TI C = LAYER_SPEC::INPUT_CHANNELS;
+            using INTERNAL_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, C>;
+            using INTERNAL_KERNEL_WEIGHTS_SHAPE = tensor::Shape<TI, BATCH_SIZE, C, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            using INTERNAL_D_OUTPUT_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::OUTPUT_HEIGHT, LAYER_SPEC::OUTPUT_WIDTH, C>;
+            auto data_4d = view_memory<INTERNAL_DATA_SHAPE>(device, data);
+            auto kw_4d = view_memory<INTERNAL_KERNEL_WEIGHTS_SHAPE>(device, kernel_weights);
+            auto d_output_4d = view_memory<INTERNAL_D_OUTPUT_SHAPE>(device, d_output);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++){
+                for(TI oh = 0; oh < LAYER_SPEC::OUTPUT_HEIGHT; oh++){
+                    for(TI ow = 0; ow < LAYER_SPEC::OUTPUT_WIDTH; ow++){
+                        for(TI c = 0; c < C; c++){
+                            ACCUMULATOR_TYPE d_pre_act = d_activation_d_x<typename DEVICE::SPEC::MATH, ACCUMULATOR_TYPE, LAYER_SPEC::ACTIVATION_FUNCTION>((ACCUMULATOR_TYPE)get(device, layer.pre_activations, bi, oh, ow, c)) * (ACCUMULATOR_TYPE)get(device, d_output_4d, bi, oh, ow, c);
+                            for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++){
+                                for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++){
+                                    TI ih_padded = oh * LAYER_SPEC::STRIDE_H + kh;
+                                    TI iw_padded = ow * LAYER_SPEC::STRIDE_W + kw;
+                                    if(ih_padded >= LAYER_SPEC::PADDING_H && ih_padded < LAYER_SPEC::INPUT_HEIGHT + LAYER_SPEC::PADDING_H &&
+                                       iw_padded >= LAYER_SPEC::PADDING_W && iw_padded < LAYER_SPEC::INPUT_WIDTH + LAYER_SPEC::PADDING_W){
+                                        TI ih = ih_padded - LAYER_SPEC::PADDING_H;
+                                        TI iw = iw_padded - LAYER_SPEC::PADDING_W;
+                                        increment(device, buffer.d_data_acc, (ACCUMULATOR_TYPE)get(device, kw_4d, bi, c, kh, kw) * d_pre_act, bi, ih, iw, c);
+                                        increment(device, buffer.d_kernel_weights_acc, (ACCUMULATOR_TYPE)get(device, data_4d, bi, ih, iw, c) * d_pre_act, bi, c, kh, kw);
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            using INTERNAL_D_DATA_SHAPE = tensor::Shape<TI, BATCH_SIZE, LAYER_SPEC::INPUT_HEIGHT, LAYER_SPEC::INPUT_WIDTH, C>;
+            using INTERNAL_D_KW_SHAPE = tensor::Shape<TI, BATCH_SIZE, C, LAYER_SPEC::KERNEL_HEIGHT, LAYER_SPEC::KERNEL_WIDTH>;
+            auto d_data_4d = view_memory<INTERNAL_D_DATA_SHAPE>(device, d_data);
+            auto d_kw_4d = view_memory<INTERNAL_D_KW_SHAPE>(device, d_kernel_weights);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++) for(TI ih = 0; ih < LAYER_SPEC::INPUT_HEIGHT; ih++) for(TI iw = 0; iw < LAYER_SPEC::INPUT_WIDTH; iw++) for(TI c = 0; c < C; c++)
+                set(device, d_data_4d, (T_DATA)get(device, buffer.d_data_acc, bi, ih, iw, c), bi, ih, iw, c);
+            for(TI bi = 0; bi < BATCH_SIZE; bi++) for(TI c = 0; c < C; c++) for(TI kh = 0; kh < LAYER_SPEC::KERNEL_HEIGHT; kh++) for(TI kw = 0; kw < LAYER_SPEC::KERNEL_WIDTH; kw++)
+                set(device, d_kw_4d, (T_KW)get(device, buffer.d_kernel_weights_acc, bi, c, kh, kw), bi, c, kh, kw);
+        }
+    }
+#endif
+
+    // ======================== zero_gradient / update / _reset_optimizer_state (no-ops) ========================
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void zero_gradient(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SPEC>& layer) {}
+    template<typename DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void add_gradient(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SOURCE_SPEC>& source, nn::layers::dynamic_conv2d::LayerGradient<TARGET_SPEC>& target) {}
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy_gradient(SOURCE_DEVICE& source_device, TARGET_DEVICE& target_device, const nn::layers::dynamic_conv2d::LayerGradient<SOURCE_SPEC>& source, nn::layers::dynamic_conv2d::LayerGradient<TARGET_SPEC>& target) {}
+    template<typename DEVICE, typename SPEC, typename OPTIMIZER>
+    RL_TOOLS_FUNCTION_PLACEMENT void update(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SPEC>& layer, OPTIMIZER& optimizer) {}
+    template<typename DEVICE, typename SPEC, typename OPTIMIZER>
+    RL_TOOLS_FUNCTION_PLACEMENT void _reset_optimizer_state(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SPEC>& layer, OPTIMIZER& optimizer) {}
+
+    // ======================== copy ========================
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy(SOURCE_DEVICE& source_device, TARGET_DEVICE& target_device, const nn::layers::dynamic_conv2d::LayerForward<SOURCE_SPEC>& source, nn::layers::dynamic_conv2d::LayerForward<TARGET_SPEC>& target) {
+        static_assert(nn::layers::dynamic_conv2d::check_spec_memory<SOURCE_SPEC, TARGET_SPEC>);
+    }
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy(SOURCE_DEVICE& source_device, TARGET_DEVICE& target_device, const nn::layers::dynamic_conv2d::LayerBackward<SOURCE_SPEC>& source, nn::layers::dynamic_conv2d::LayerBackward<TARGET_SPEC>& target) {
+        static_assert(nn::layers::dynamic_conv2d::check_spec_memory<SOURCE_SPEC, TARGET_SPEC>);
+        copy(source_device, target_device, source.pre_activations, target.pre_activations);
+    }
+    template<typename SOURCE_DEVICE, typename TARGET_DEVICE, typename SOURCE_SPEC, typename TARGET_SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void copy(SOURCE_DEVICE& source_device, TARGET_DEVICE& target_device, const nn::layers::dynamic_conv2d::LayerGradient<SOURCE_SPEC>& source, nn::layers::dynamic_conv2d::LayerGradient<TARGET_SPEC>& target) {
+        static_assert(nn::layers::dynamic_conv2d::check_spec_memory<SOURCE_SPEC, TARGET_SPEC>);
+        copy(source_device, target_device, static_cast<const nn::layers::dynamic_conv2d::LayerBackward<SOURCE_SPEC>&>(source), static_cast<nn::layers::dynamic_conv2d::LayerBackward<TARGET_SPEC>&>(target));
+        copy(source_device, target_device, source.output, target.output);
+    }
+
+    // ======================== abs_diff ========================
+    template<typename DEVICE, typename S1, typename S2>
+    RL_TOOLS_FUNCTION_PLACEMENT typename S1::TYPE_POLICY::DEFAULT abs_diff(DEVICE& device, const nn::layers::dynamic_conv2d::LayerForward<S1>&, const nn::layers::dynamic_conv2d::LayerForward<S2>&) { return 0; }
+    template<typename DEVICE, typename S1, typename S2>
+    RL_TOOLS_FUNCTION_PLACEMENT typename S1::TYPE_POLICY::DEFAULT abs_diff(DEVICE& device, const nn::layers::dynamic_conv2d::LayerBackward<S1>& l1, const nn::layers::dynamic_conv2d::LayerBackward<S2>& l2) {
+        return abs_diff(device, l1.pre_activations, l2.pre_activations);
+    }
+    template<typename DEVICE, typename S1, typename S2>
+    RL_TOOLS_FUNCTION_PLACEMENT typename S1::TYPE_POLICY::DEFAULT abs_diff(DEVICE& device, const nn::layers::dynamic_conv2d::LayerGradient<S1>& l1, const nn::layers::dynamic_conv2d::LayerGradient<S2>& l2) {
+        return abs_diff(device, static_cast<const nn::layers::dynamic_conv2d::LayerBackward<S1>&>(l1), static_cast<const nn::layers::dynamic_conv2d::LayerBackward<S2>&>(l2))
+             + abs_diff(device, l1.output, l2.output);
+    }
+    template<typename DEVICE>
+    RL_TOOLS_FUNCTION_PLACEMENT auto abs_diff(DEVICE& device, const nn::layers::dynamic_conv2d::State&, const nn::layers::dynamic_conv2d::State&) { return 0; }
+
+    // ======================== reset_forward_state ========================
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void reset_forward_state(DEVICE& device, nn::layers::dynamic_conv2d::LayerBackward<SPEC>& l) {
+        set_all(device, l.pre_activations, 0);
+    }
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT void reset_forward_state(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SPEC>& l) {
+        reset_forward_state(device, static_cast<nn::layers::dynamic_conv2d::LayerBackward<SPEC>&>(l));
+        set_all(device, l.output, 0);
+    }
+
+    // ======================== is_nan ========================
+    template<typename DEVICE, typename SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT bool is_nan(DEVICE&, const nn::layers::dynamic_conv2d::LayerForward<SPEC>&, const Mode<MODE>& = Mode<mode::Default<>>{}) { return false; }
+    template<typename DEVICE, typename SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT bool is_nan(DEVICE& device, const nn::layers::dynamic_conv2d::LayerBackward<SPEC>& l, const Mode<MODE>& mode = Mode<mode::Default<>>{}) {
+        return is_nan(device, l.pre_activations, mode);
+    }
+    template<typename DEVICE, typename SPEC, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT bool is_nan(DEVICE& device, const nn::layers::dynamic_conv2d::LayerGradient<SPEC>& l, const Mode<MODE>& mode = Mode<mode::Default<>>{}) {
+        return is_nan(device, static_cast<const nn::layers::dynamic_conv2d::LayerBackward<SPEC>&>(l), mode) || is_nan(device, l.output, mode);
+    }
+    template<typename DEVICE, typename MODE = mode::Default<>>
+    RL_TOOLS_FUNCTION_PLACEMENT bool is_nan(DEVICE&, nn::layers::dynamic_conv2d::State&, const Mode<MODE>& = Mode<mode::Default<>>{}) { return false; }
+
+    // ======================== output ========================
+    template<typename DEVICE, typename SPEC>
+    RL_TOOLS_FUNCTION_PLACEMENT auto output(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SPEC>& l){
+        return view_memory<typename SPEC::OUTPUT_SHAPE>(device, l.output);
+    }
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+#endif

hyperdrone/_vendor/rl-tools/include/rl_tools/nn/layers/dynamic_conv2d/persist.h

@@ -0,0 +1,32 @@
+#include "../../../version.h"
+#if (defined(RL_TOOLS_DISABLE_INCLUDE_GUARDS) || !defined(RL_TOOLS_NN_LAYERS_DYNAMIC_CONV2D_PERSIST_H)) && (RL_TOOLS_USE_THIS_VERSION == 1)
+#pragma once
+#define RL_TOOLS_NN_LAYERS_DYNAMIC_CONV2D_PERSIST_H
+#include "layer.h"
+RL_TOOLS_NAMESPACE_WRAPPER_START
+namespace rl_tools {
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    void save(DEVICE& device, nn::layers::dynamic_conv2d::LayerForward<SPEC>& layer, GROUP& group) {
+        set_attribute(device, group, "type", "dynamic_conv2d");
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    void save(DEVICE& device, nn::layers::dynamic_conv2d::LayerBackward<SPEC>& layer, GROUP& group) {
+        save(device, (nn::layers::dynamic_conv2d::LayerForward<SPEC>&)layer, group);
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    void save(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SPEC>& layer, GROUP& group) {
+        save(device, (nn::layers::dynamic_conv2d::LayerBackward<SPEC>&)layer, group);
+    }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::dynamic_conv2d::LayerForward<SPEC>& layer, GROUP& group) { return true; }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::dynamic_conv2d::LayerBackward<SPEC>& layer, GROUP& group) { return true; }
+    template<typename DEVICE, typename SPEC, typename GROUP>
+    bool load(DEVICE& device, nn::layers::dynamic_conv2d::LayerGradient<SPEC>& layer, GROUP& group) { return true; }
+    template<typename DEVICE, typename GROUP>
+    void save(DEVICE& device, nn::layers::dynamic_conv2d::State& state, GROUP& group) {}
+    template<typename DEVICE, typename GROUP>
+    bool load(DEVICE& device, nn::layers::dynamic_conv2d::State& state, GROUP& group) { return true; }
+}
+RL_TOOLS_NAMESPACE_WRAPPER_END
+#endif