diff --git a/src/layer/crop.cpp b/src/layer/crop.cpp index 05ab8ab78..7982aa99a 100644 --- a/src/layer/crop.cpp +++ b/src/layer/crop.cpp @@ -20,6 +20,14 @@ DEFINE_LAYER_CREATOR(Crop) Crop::Crop() { + one_blob_only = false; + support_inplace = false; + support_vulkan = true; + +#if NCNN_VULKAN + pipeline_crop = 0; + pipeline_crop_pack4 = 0; +#endif // NCNN_VULKAN } int Crop::load_param(const ParamDict& pd) @@ -92,4 +100,182 @@ int Crop::forward(const std::vector& bottom_blobs, std::vector& top_bl return 0; } +#if NCNN_VULKAN +int Crop::create_pipeline() +{ + std::vector specializations(3); + specializations[0].i = woffset; + specializations[1].i = hoffset; + specializations[2].i = coffset; + + // pack1 + { + pipeline_crop = new Pipeline(vkdev); + pipeline_crop->set_optimal_local_size_xyz(); + pipeline_crop->create("crop", specializations, 2, 10); + } + + // pack4 + { + pipeline_crop_pack4 = new Pipeline(vkdev); + pipeline_crop_pack4->set_optimal_local_size_xyz(); + pipeline_crop_pack4->create("crop_pack4", specializations, 2, 10); + } + + return 0; +} + +int Crop::destroy_pipeline() +{ + delete pipeline_crop; + pipeline_crop = 0; + + delete pipeline_crop_pack4; + pipeline_crop_pack4 = 0; + + return 0; +} + +int Crop::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const +{ + int w = bottom_blob.w; + int h = bottom_blob.h; + int channels = bottom_blob.c; + size_t elemsize = bottom_blob.elemsize; + int packing = bottom_blob.packing; + + // TODO vec and image crop + int dims = bottom_blob.dims; + + int _outw = outw == -233 ? w - woffset : outw; + int _outh = outh == -233 ? h - hoffset : outh; + int _outc = outc == -233 ? channels * packing - coffset : outc; + + int out_packing = _outc % 4 == 0 ? 4 : 1; + size_t out_elemsize = elemsize / packing * out_packing; + + top_blob.create(_outw, _outh, _outc / out_packing, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator); + if (top_blob.empty()) + return -100; + +// fprintf(stderr, "Crop::forward %p %p\n", bottom_blob.buffer(), top_blob.buffer()); + + std::vector bindings(2); + bindings[0] = bottom_blob; + bindings[1] = top_blob; + + std::vector constants(10); + constants[0].i = bottom_blob.dims; + constants[1].i = bottom_blob.w; + constants[2].i = bottom_blob.h; + constants[3].i = bottom_blob.c; + constants[4].i = bottom_blob.cstep; + constants[5].i = top_blob.dims; + constants[6].i = top_blob.w; + constants[7].i = top_blob.h; + constants[8].i = top_blob.c; + constants[9].i = top_blob.cstep; + + const Pipeline* pipeline = 0; + if (packing == 1 && out_packing == 1) + { + pipeline = pipeline_crop; + } + else if (packing == 4 && out_packing == 4) + { + pipeline = pipeline_crop_pack4; + } + else if (packing == 1 && out_packing == 4) + { + // TODO + return -1; + } + else if (packing == 4 && out_packing == 1) + { + // TODO + return -1; + } + + // record + cmd.record_prepare_compute_barrier(bottom_blob); + cmd.record_prepare_compute_barrier(top_blob); + cmd.record_pipeline(pipeline, bindings, constants, top_blob); + + return 0; +} + +int Crop::forward(const std::vector& bottom_blobs, std::vector& top_blobs, VkCompute& cmd, const Option& opt) const +{ + const VkMat& bottom_blob = bottom_blobs[0]; + const VkMat& reference_blob = bottom_blobs[1]; + + int w = bottom_blob.w; + int h = bottom_blob.h; + int channels = bottom_blob.c; + size_t elemsize = bottom_blob.elemsize; + int packing = bottom_blob.packing; + + // TODO vec and image crop + int dims = bottom_blob.dims; + + int _outw = reference_blob.w; + int _outh = reference_blob.h; + int _outc = reference_blob.dims == 3 ? reference_blob.c : channels; + + int out_packing = _outc % 4 == 0 ? 4 : 1; + size_t out_elemsize = elemsize / packing * out_packing; + + VkMat& top_blob = top_blobs[0]; + + top_blob.create(_outw, _outh, _outc / out_packing, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator); + if (top_blob.empty()) + return -100; + +// fprintf(stderr, "Crop::forward %p %p\n", bottom_blob.buffer(), top_blob.buffer()); + + std::vector bindings(2); + bindings[0] = bottom_blob; + bindings[1] = top_blob; + + std::vector constants(10); + constants[0].i = bottom_blob.dims; + constants[1].i = bottom_blob.w; + constants[2].i = bottom_blob.h; + constants[3].i = bottom_blob.c; + constants[4].i = bottom_blob.cstep; + constants[5].i = top_blob.dims; + constants[6].i = top_blob.w; + constants[7].i = top_blob.h; + constants[8].i = top_blob.c; + constants[9].i = top_blob.cstep; + + const Pipeline* pipeline = 0; + if (packing == 1 && out_packing == 1) + { + pipeline = pipeline_crop; + } + else if (packing == 4 && out_packing == 4) + { + pipeline = pipeline_crop_pack4; + } + else if (packing == 1 && out_packing == 4) + { + // TODO + return -1; + } + else if (packing == 4 && out_packing == 1) + { + // TODO + return -1; + } + + // record + cmd.record_prepare_compute_barrier(bottom_blob); + cmd.record_prepare_compute_barrier(top_blob); + cmd.record_pipeline(pipeline, bindings, constants, top_blob); + + return 0; +} +#endif // NCNN_VULKAN + } // namespace ncnn diff --git a/src/layer/crop.h b/src/layer/crop.h index ee88ee076..781f9eaff 100644 --- a/src/layer/crop.h +++ b/src/layer/crop.h @@ -30,6 +30,15 @@ public: virtual int forward(const std::vector& bottom_blobs, std::vector& top_blobs, const Option& opt) const; +#if NCNN_VULKAN + virtual int create_pipeline(); + virtual int destroy_pipeline(); + + virtual int forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const; + + virtual int forward(const std::vector& bottom_blobs, std::vector& top_blobs, VkCompute& cmd, const Option& opt) const; +#endif // NCNN_VULKAN + public: int woffset; int hoffset; @@ -37,6 +46,11 @@ public: int outw; int outh; int outc; + +#if NCNN_VULKAN + Pipeline* pipeline_crop; + Pipeline* pipeline_crop_pack4; +#endif // NCNN_VULKAN }; } // namespace ncnn diff --git a/src/layer/shader/crop.comp b/src/layer/shader/crop.comp new file mode 100644 index 000000000..9e292b3dd --- /dev/null +++ b/src/layer/shader/crop.comp @@ -0,0 +1,59 @@ +// Tencent is pleased to support the open source community by making ncnn available. +// +// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved. +// +// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except +// in compliance with the License. You may obtain a copy of the License at +// +// https://opensource.org/licenses/BSD-3-Clause +// +// Unless required by applicable law or agreed to in writing, software distributed +// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR +// CONDITIONS OF ANY KIND, either express or implied. See the License for the +// specific language governing permissions and limitations under the License. + +#version 450 + +layout (constant_id = 0) const int woffset = 0; +layout (constant_id = 1) const int hoffset = 0; +layout (constant_id = 2) const int coffset = 0; + +layout (local_size_x_id = 233) in; +layout (local_size_y_id = 234) in; +layout (local_size_z_id = 235) in; + +layout (binding = 0) readonly buffer bottom_blob { float bottom_blob_data[]; }; +layout (binding = 1) writeonly buffer top_blob { float top_blob_data[]; }; + +layout (push_constant) uniform parameter +{ + int dims; + int w; + int h; + int c; + int cstep; + + int outdims; + int outw; + int outh; + int outc; + int outcstep; +} p; + +void main() +{ + int gx = int(gl_GlobalInvocationID.x); + int gy = int(gl_GlobalInvocationID.y); + int gz = int(gl_GlobalInvocationID.z); + + if (gx >= p.outw || gy >= p.outh || gz >= p.outc) + return; + + int x = gx + woffset; + int y = gy + hoffset; + int z = gz + coffset; + + int v_offset = z * p.cstep + y * p.w + x; + + top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset]; +} diff --git a/src/layer/shader/crop_pack4.comp b/src/layer/shader/crop_pack4.comp new file mode 100644 index 000000000..e7c0c1604 --- /dev/null +++ b/src/layer/shader/crop_pack4.comp @@ -0,0 +1,59 @@ +// Tencent is pleased to support the open source community by making ncnn available. +// +// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved. +// +// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except +// in compliance with the License. You may obtain a copy of the License at +// +// https://opensource.org/licenses/BSD-3-Clause +// +// Unless required by applicable law or agreed to in writing, software distributed +// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR +// CONDITIONS OF ANY KIND, either express or implied. See the License for the +// specific language governing permissions and limitations under the License. + +#version 450 + +layout (constant_id = 0) const int woffset = 0; +layout (constant_id = 1) const int hoffset = 0; +layout (constant_id = 2) const int coffset = 0; + +layout (local_size_x_id = 233) in; +layout (local_size_y_id = 234) in; +layout (local_size_z_id = 235) in; + +layout (binding = 0) readonly buffer bottom_blob { vec4 bottom_blob_data[]; }; +layout (binding = 1) writeonly buffer top_blob { vec4 top_blob_data[]; }; + +layout (push_constant) uniform parameter +{ + int dims; + int w; + int h; + int c; + int cstep; + + int outdims; + int outw; + int outh; + int outc; + int outcstep; +} p; + +void main() +{ + int gx = int(gl_GlobalInvocationID.x); + int gy = int(gl_GlobalInvocationID.y); + int gz = int(gl_GlobalInvocationID.z); + + if (gx >= p.outw || gy >= p.outh || gz >= p.outc) + return; + + int x = gx + woffset; + int y = gy + hoffset; + int z = gz + coffset / 4; + + int v_offset = z * p.cstep + y * p.w + x; + + top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset]; +}