reorg shader

src/layer/reorg.cpp

@@ -22,6 +22,13 @@ Reorg::Reorg()
 {
     one_blob_only = true;
     support_inplace = false;
+    support_vulkan = true;
+
+#if NCNN_VULKAN
+    pipeline_reorg = 0;
+    pipeline_reorg_pack4 = 0;
+    pipeline_reorg_pack1to4 = 0;
+#endif // NCNN_VULKAN
 }
 
 int Reorg::load_param(const ParamDict& pd)
@@ -75,4 +82,105 @@ int Reorg::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) con
     return 0;
 }
 
+#if NCNN_VULKAN
+int Reorg::create_pipeline()
+{
+    std::vector<vk_specialization_type> specializations(1);
+    specializations[0].i = stride;
+
+    pipeline_reorg = new Pipeline(vkdev);
+    pipeline_reorg->set_optimal_local_size_xyz();
+    pipeline_reorg->create("reorg", specializations, 2, 10);
+
+    // pack4
+    {
+        pipeline_reorg_pack4 = new Pipeline(vkdev);
+        pipeline_reorg_pack4->set_optimal_local_size_xyz();
+        pipeline_reorg_pack4->create("reorg_pack4", specializations, 2, 10);
+    }
+
+    // pack1to4
+    {
+        pipeline_reorg_pack1to4 = new Pipeline(vkdev);
+        pipeline_reorg_pack1to4->set_optimal_local_size_xyz();
+        pipeline_reorg_pack1to4->create("reorg_pack1to4", specializations, 2, 10);
+    }
+
+    return 0;
+}
+
+int Reorg::destroy_pipeline()
+{
+    delete pipeline_reorg;
+    pipeline_reorg = 0;
+
+    delete pipeline_reorg_pack4;
+    pipeline_reorg_pack4 = 0;
+
+    delete pipeline_reorg_pack1to4;
+    pipeline_reorg_pack1to4 = 0;
+
+    return 0;
+}
+
+int Reorg::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;
+
+    int outw = w / stride;
+    int outh = h / stride;
+    int outc = channels * packing * stride * stride;
+
+    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, "Reorg::forward %p %p\n", bottom_blob.buffer(), top_blob.buffer());
+
+    std::vector<VkMat> bindings(2);
+    bindings[0] = bottom_blob;
+    bindings[1] = top_blob;
+
+    std::vector<vk_constant_type> 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_reorg;
+    }
+    else if (packing == 4) // assert out_packing == 4
+    {
+        pipeline = pipeline_reorg_pack4;
+    }
+    else if (packing == 1 && out_packing == 4)
+    {
+        pipeline = pipeline_reorg_pack1to4;
+    }
+
+    // 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

src/layer/reorg.h

@@ -28,8 +28,21 @@ public:
 
     virtual int forward(const Mat& bottom_blob, Mat& top_blob, 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;
+#endif // NCNN_VULKAN
+
 private:
     int stride;
+
+#if NCNN_VULKAN
+    Pipeline* pipeline_reorg;
+    Pipeline* pipeline_reorg_pack4;
+    Pipeline* pipeline_reorg_pack1to4;
+#endif // NCNN_VULKAN
 };
 
 } // namespace ncnn

src/layer/shader/reorg.comp

@@ -0,0 +1,60 @@
+// 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 stride = 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 z = gz / (stride * stride);
+    int zi = gz % (stride * stride);
+    int y = gy * stride + zi / stride;
+    int x = gx * stride + zi % stride;
+
+    int v_offset = z * p.cstep + y * p.w + x;
+
+    float v = bottom_blob_data[v_offset];
+
+    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+}

src/layer/shader/reorg_pack1to4.comp

@@ -0,0 +1,67 @@
+// 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 stride = 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 { 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;
+
+    ivec4 gz4 = ivec4(gz * 4) + ivec4(0, 1, 2, 3);
+
+    ivec4 z4 = gz4 / (stride * stride);
+    ivec4 zi4 = gz4 % (stride * stride);
+    ivec4 y4 = gy * stride + zi4 / stride;
+    ivec4 x4 = gx * stride + zi4 % stride;
+
+    ivec4 v_offset = z4 * p.cstep + y4 * p.w + x4;
+
+    vec4 v;
+
+    v.r = bottom_blob_data[v_offset.r];
+    v.g = bottom_blob_data[v_offset.g];
+    v.b = bottom_blob_data[v_offset.b];
+    v.a = bottom_blob_data[v_offset.a];
+
+    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+}

src/layer/shader/reorg_pack4.comp

@@ -0,0 +1,91 @@
+// 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 stride = 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;
+
+    ivec4 gz4 = ivec4(gz * 4) + ivec4(0, 1, 2, 3);
+
+    ivec4 z4 = gz4 / (stride * stride);
+    ivec4 zi4 = gz4 % (stride * stride);
+    ivec4 y4 = gy * stride + zi4 / stride;
+    ivec4 x4 = gx * stride + zi4 % stride;
+
+    ivec4 v_offset = (z4 / 4) * p.cstep + y4 * p.w + x4;
+
+    ivec4 lane4 = z4 % 4;
+
+    // v = v4[lane]
+    vec4 v;
+
+    vec4 v4;
+
+    v4 = bottom_blob_data[v_offset.r];
+    if (lane4.r == 0) v.r = v4.r;
+    else if (lane4.r == 1) v.r = v4.g;
+    else if (lane4.r == 2) v.r = v4.b;
+    else /* if (lane4.r == 3) */ v.r = v4.a;
+
+    v4 = bottom_blob_data[v_offset.g];
+    if (lane4.g == 0) v.g = v4.r;
+    else if (lane4.g == 1) v.g = v4.g;
+    else if (lane4.g == 2) v.g = v4.b;
+    else /* if (lane4.g == 3) */ v.g = v4.a;
+
+    v4 = bottom_blob_data[v_offset.b];
+    if (lane4.b == 0) v.b = v4.r;
+    else if (lane4.b == 1) v.b = v4.g;
+    else if (lane4.b == 2) v.b = v4.b;
+    else /* if (lane4.b == 3) */ v.b = v4.a;
+
+    v4 = bottom_blob_data[v_offset.a];
+    if (lane4.a == 0) v.a = v4.r;
+    else if (lane4.a == 1) v.a = v4.g;
+    else if (lane4.a == 2) v.a = v4.b;
+    else /* if (lane4.a == 3) */ v.a = v4.a;
+
+    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = v;
+}