permute shader

src/layer/permute.cpp

@@ -22,6 +22,12 @@ Permute::Permute()
 {
     one_blob_only = true;
     support_inplace = false;
+    support_vulkan = true;
+
+#if NCNN_VULKAN
+    pipeline_permute = 0;
+    pipeline_permute_pack4to1 = 0;
+#endif // NCNN_VULKAN
 }
 
 int Permute::load_param(const ParamDict& pd)
@@ -196,4 +202,157 @@ int Permute::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) c
     return 0;
 }
 
+#if NCNN_VULKAN
+int Permute::create_pipeline()
+{
+    std::vector<vk_specialization_type> specializations(1);
+    specializations[0].i = order_type;
+
+    pipeline_permute = new Pipeline(vkdev);
+    pipeline_permute->set_optimal_local_size_xyz();
+    pipeline_permute->create("permute", specializations, 2, 10);
+
+    // pack4
+    {
+        pipeline_permute_pack4to1 = new Pipeline(vkdev);
+        pipeline_permute_pack4to1->set_optimal_local_size_xyz();
+        pipeline_permute_pack4to1->create("permute_pack4to1", specializations, 2, 10);
+    }
+
+    return 0;
+}
+
+int Permute::destroy_pipeline()
+{
+    delete pipeline_permute;
+    pipeline_permute = 0;
+
+    delete pipeline_permute_pack4to1;
+    pipeline_permute_pack4to1 = 0;
+
+    return 0;
+}
+
+int Permute::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 dims = bottom_blob.dims;
+
+    int out_packing = 1;
+    size_t out_elemsize = elemsize / packing;
+
+    if (dims == 2)
+    {
+        // order_type
+        // 0 = w h
+        // 1 = h w
+
+        h = h * packing;
+
+        if (order_type == 0)
+        {
+            top_blob.create(w, h, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator);
+            if (top_blob.empty())
+                return -100;
+        }
+        else if (order_type == 1)
+        {
+            top_blob.create(h, w, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator);
+            if (top_blob.empty())
+                return -100;
+        }
+    }
+    else if (dims == 3)
+    {
+        // order_type
+        // 0 = w h c
+        // 1 = h w c
+        // 2 = w c h
+        // 3 = c w h
+        // 4 = h c w
+        // 5 = c h w
+
+        channels = channels * packing;
+
+        if (order_type == 0)
+        {
+            top_blob.create(w, h, channels, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator);
+            if (top_blob.empty())
+                return -100;
+        }
+        else if (order_type == 1)
+        {
+            top_blob.create(h, w, channels, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator);
+            if (top_blob.empty())
+                return -100;
+        }
+        else if (order_type == 2)
+        {
+            top_blob.create(w, channels, h, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator);
+            if (top_blob.empty())
+                return -100;
+        }
+        else if (order_type == 3)
+        {
+            top_blob.create(channels, w, h, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator);
+            if (top_blob.empty())
+                return -100;
+        }
+        else if (order_type == 4)
+        {
+            top_blob.create(h, channels, w, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator);
+            if (top_blob.empty())
+                return -100;
+        }
+        else if (order_type == 5)
+        {
+            top_blob.create(channels, h, w, out_elemsize, out_packing, opt.blob_vkallocator, opt.staging_vkallocator);
+            if (top_blob.empty())
+                return -100;
+        }
+    }
+
+    fprintf(stderr, "%d %d %d\n", top_blob.w, top_blob.h, top_blob.c);
+
+//     fprintf(stderr, "Permute::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;
+
+    // record
+    cmd.record_prepare_compute_barrier(bottom_blob);
+    cmd.record_prepare_compute_barrier(top_blob);
+
+    if (packing == 1)
+    {
+        cmd.record_pipeline(pipeline_permute, bindings, constants, top_blob);
+    }
+
+    if (packing == 4)
+    {
+        cmd.record_pipeline(pipeline_permute_pack4to1, bindings, constants, bottom_blob);
+    }
+
+    return 0;
+}
+#endif // NCNN_VULKAN
+
 } // namespace ncnn

src/layer/permute.h

@@ -28,8 +28,20 @@ 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
+
 public:
     int order_type;
+
+#if NCNN_VULKAN
+    Pipeline* pipeline_permute;
+    Pipeline* pipeline_permute_pack4to1;
+#endif // NCNN_VULKAN
 };
 
 } // namespace ncnn

src/layer/shader/packing_4to1.comp

@@ -47,9 +47,9 @@ void main()
 
     ivec4 z4 = ivec4(gz * 4) + ivec4(0, 1, 2, 3);
 
-    ivec4 v_offset = z4 * p.cstep + ivec4(gy * p.w + gx);
+    ivec4 v_offset = z4 * p.outcstep + ivec4(gy * p.outw + gx);
 
-    vec4 v = bottom_blob_data[gz * p.outcstep + gy * p.outw + gx];
+    vec4 v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
 
     top_blob_data[v_offset.r] = v.r;
     top_blob_data[v_offset.g] = v.g;

src/layer/shader/permute.comp

@@ -0,0 +1,104 @@
+// 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 order_type = 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 v_offset;
+
+    if (p.dims == 2)
+    {
+        // order_type
+        // 0 = w h
+        // 1 = h w
+
+        if (order_type == 0)
+        {
+            v_offset = gy * p.w + gx;
+        }
+        if (order_type == 1)
+        {
+            v_offset = gx * p.w + gy;
+        }
+    }
+    else if (p.dims == 3)
+    {
+        // order_type
+        // 0 = w h c
+        // 1 = h w c
+        // 2 = w c h
+        // 3 = c w h
+        // 4 = h c w
+        // 5 = c h w
+
+        if (order_type == 0)
+        {
+            v_offset = gz * p.cstep + gy * p.w + gx;
+        }
+        if (order_type == 1)
+        {
+            v_offset = gz * p.cstep + gx * p.w + gy;
+        }
+        if (order_type == 2)
+        {
+            v_offset = gy * p.cstep + gz * p.w + gx;
+        }
+        if (order_type == 3)
+        {
+            v_offset = gx * p.cstep + gz * p.w + gy;
+        }
+        if (order_type == 4)
+        {
+            v_offset = gy * p.cstep + gx * p.w + gz;
+        }
+        if (order_type == 5)
+        {
+            v_offset = gx * p.cstep + gy * p.w + gz;
+        }
+    }
+
+    top_blob_data[gz * p.outcstep + gy * p.outw + gx] = bottom_blob_data[v_offset];
+}

src/layer/shader/permute_pack4to1.comp

@@ -0,0 +1,109 @@
+// 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 order_type = 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 { 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.w || gy >= p.h || gz >= p.c)
+        return;
+
+    ivec4 v_offset;
+
+    if (p.dims == 2)
+    {
+        // order_type
+        // 0 = w h
+        // 1 = h w
+
+        if (order_type == 0)
+        {
+            v_offset = ivec4((gy * 4) * p.outw + gx) + ivec4(0, 1, 2, 3) * p.outw;
+        }
+        if (order_type == 1)
+        {
+            v_offset = ivec4(gx * p.outw + gy) + ivec4(0, 1, 2, 3);
+        }
+    }
+    else if (p.dims == 3)
+    {
+        // order_type
+        // 0 = w h c
+        // 1 = h w c
+        // 2 = w c h
+        // 3 = c w h
+        // 4 = h c w
+        // 5 = c h w
+
+        if (order_type == 0)
+        {
+            v_offset = ivec4((gz * 4) * p.outcstep + gy * p.outw + gx) + ivec4(0, 1, 2, 3) * p.outcstep;
+        }
+        if (order_type == 1)
+        {
+            v_offset = ivec4((gz * 4) * p.outcstep + gx * p.outw + gy) + ivec4(0, 1, 2, 3) * p.outcstep;
+        }
+        if (order_type == 2)
+        {
+            v_offset = ivec4(gy * p.outcstep + (gz * 4) * p.outw + gx) + ivec4(0, 1, 2, 3) * p.outw;
+        }
+        if (order_type == 3)
+        {
+            v_offset = ivec4(gy * p.outcstep + gx * p.outw + gz * 4) + ivec4(0, 1, 2, 3);
+        }
+        if (order_type == 4)
+        {
+            v_offset = ivec4(gx * p.outcstep + (gz * 4) * p.outw + gy) + ivec4(0, 1, 2, 3) * p.outw;
+        }
+        if (order_type == 5)
+        {
+            v_offset = ivec4(gx * p.outcstep + gy * p.outw + gz * 4) + ivec4(0, 1, 2, 3);
+        }
+    }
+
+    vec4 v = bottom_blob_data[gz * p.cstep + gy * p.w + gx];
+
+    top_blob_data[v_offset.r] = v.r;
+    top_blob_data[v_offset.g] = v.g;
+    top_blob_data[v_offset.b] = v.b;
+    top_blob_data[v_offset.a] = v.a;
+}