prelu shader

src/layer/prelu.cpp

@@ -22,6 +22,12 @@ PReLU::PReLU()
 {
     one_blob_only = true;
     support_inplace = true;
+    support_vulkan = true;
+
+#if NCNN_VULKAN
+    pipeline_prelu = 0;
+    pipeline_prelu_pack4 = 0;
+#endif // NCNN_VULKAN
 }
 
 int PReLU::load_param(const ParamDict& pd)
@@ -115,4 +121,84 @@ int PReLU::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
     return 0;
 }
 
+#if NCNN_VULKAN
+int PReLU::upload_model(VkTransfer& cmd)
+{
+    if (num_slope == 1)
+    {
+        // dup4 for pack4
+        Mat slope_data4(4);
+        slope_data4.fill(slope_data[0]);
+        cmd.record_upload(slope_data4, slope_data_gpu);
+    }
+    else
+    {
+        cmd.record_upload(slope_data, slope_data_gpu);
+    }
+
+    return 0;
+}
+
+int PReLU::create_pipeline()
+{
+    std::vector<vk_specialization_type> specializations(1);
+    specializations[0].i = num_slope;
+
+    // pack1
+    if (num_slope == 1 || num_slope % 4 != 0)
+    {
+        pipeline_prelu = new Pipeline(vkdev);
+        pipeline_prelu->set_optimal_local_size_xyz(8, 8, num_slope);
+        pipeline_prelu->create("prelu", specializations, 2, 5);
+    }
+
+    // pack4
+    if (num_slope == 1 || num_slope % 4 == 0)
+    {
+        pipeline_prelu_pack4 = new Pipeline(vkdev);
+        pipeline_prelu_pack4->set_optimal_local_size_xyz(8, 8, num_slope / 4);
+        pipeline_prelu_pack4->create("prelu_pack4", specializations, 2, 5);
+    }
+
+    return 0;
+}
+
+int PReLU::destroy_pipeline()
+{
+    delete pipeline_prelu;
+    pipeline_prelu = 0;
+
+    delete pipeline_prelu_pack4;
+    pipeline_prelu_pack4 = 0;
+
+    return 0;
+}
+
+int PReLU::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const
+{
+    int packing = bottom_top_blob.packing;
+
+//     fprintf(stderr, "PReLU::forward_inplace %p\n", bottom_top_blob.buffer());
+
+    std::vector<VkMat> bindings(2);
+    bindings[0] = bottom_top_blob;
+    bindings[1] = slope_data_gpu;
+
+    std::vector<vk_constant_type> constants(5);
+    constants[0].i = bottom_top_blob.dims;
+    constants[1].i = bottom_top_blob.w;
+    constants[2].i = bottom_top_blob.h;
+    constants[3].i = bottom_top_blob.c;
+    constants[4].i = bottom_top_blob.cstep;
+
+    const Pipeline* pipeline = packing == 4 ? pipeline_prelu_pack4 : pipeline_prelu;
+
+    // record
+    cmd.record_prepare_compute_barrier(bottom_top_blob);
+    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+
+    return 0;
+}
+#endif // NCNN_VULKAN
+
 } // namespace ncnn

src/layer/prelu.h

@@ -30,9 +30,24 @@ public:
 
     virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
 
+#if NCNN_VULKAN
+    virtual int upload_model(VkTransfer& cmd);
+
+    virtual int create_pipeline();
+    virtual int destroy_pipeline();
+
+    virtual int forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& opt) const;
+#endif // NCNN_VULKAN
+
 public:
     int num_slope;
     Mat slope_data;
+
+#if NCNN_VULKAN
+    VkMat slope_data_gpu;
+    Pipeline* pipeline_prelu;
+    Pipeline* pipeline_prelu_pack4;
+#endif // NCNN_VULKAN
 };
 
 } // namespace ncnn

src/layer/shader/prelu.comp

@@ -0,0 +1,86 @@
+// 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 num_slope = 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) buffer bottom_top_blob { float bottom_top_blob_data[]; };
+layout (binding = 1) readonly buffer slope_blob { float slope_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    if (p.dims == 1)
+    {
+        float v = bottom_top_blob_data[gx];
+
+        float slope = num_slope > 1 ? slope_blob_data[gx] : slope_blob_data[0];
+
+        v = v < 0 ? v * slope : v;
+
+        bottom_top_blob_data[gx] = v;
+
+        return;
+    }
+
+    if (p.dims == 2)
+    {
+        const int gi = gy * p.w + gx;
+
+        float v = bottom_top_blob_data[gi];
+
+        float slope = num_slope > 1 ? slope_blob_data[gy] : slope_blob_data[0];
+
+        v = v < 0 ? v * slope : v;
+
+        bottom_top_blob_data[gi] = v;
+
+        return;
+    }
+
+    if (p.dims == 3)
+    {
+        const int gi = gz * p.cstep + gy * p.w + gx;
+
+        float v = bottom_top_blob_data[gi];
+
+        float slope = num_slope > 1 ? slope_blob_data[gz] : slope_blob_data[0];
+
+        v = v < 0 ? v * slope : v;
+
+        bottom_top_blob_data[gi] = v;
+
+        return;
+    }
+}

src/layer/shader/prelu_pack4.comp

@@ -0,0 +1,86 @@
+// 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 num_slope = 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) buffer bottom_top_blob { vec4 bottom_top_blob_data[]; };
+layout (binding = 1) readonly buffer slope_blob { vec4 slope_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    if (p.dims == 1)
+    {
+        vec4 v = bottom_top_blob_data[gx];
+
+        vec4 slope = num_slope > 1 ? slope_blob_data[gx] : slope_blob_data[0];
+
+        v = mix(v, v * slope, lessThan(v, vec4(0.0)));
+
+        bottom_top_blob_data[gx] = v;
+
+        return;
+    }
+
+    if (p.dims == 2)
+    {
+        const int gi = gy * p.w + gx;
+
+        vec4 v = bottom_top_blob_data[gi];
+
+        vec4 slope = num_slope > 1 ? slope_blob_data[gy] : slope_blob_data[0];
+
+        v = mix(v, v * slope, lessThan(v, vec4(0.0)));
+
+        bottom_top_blob_data[gi] = v;
+
+        return;
+    }
+
+    if (p.dims == 3)
+    {
+        const int gi = gz * p.cstep + gy * p.w + gx;
+
+        vec4 v = bottom_top_blob_data[gi];
+
+        vec4 slope = num_slope > 1 ? slope_blob_data[gz] : slope_blob_data[0];
+
+        v = mix(v, v * slope, lessThan(v, vec4(0.0)));
+
+        bottom_top_blob_data[gi] = v;
+
+        return;
+    }
+}