From 640a0372d50ee56a6877d458f456fed7940f8580 Mon Sep 17 00:00:00 2001
From: nihuini <nihuini@tencent.com>
Date: Wed, 24 Jul 2019 18:40:52 +0800
Subject: [PATCH] pooling pack4 arm neon

---
 src/layer/arm/pooling_arm.cpp | 281 +++++++++++++++++++++++++++++++++-
 1 file changed, 276 insertions(+), 5 deletions(-)

diff --git a/src/layer/arm/pooling_arm.cpp b/src/layer/arm/pooling_arm.cpp
index 3cd2cd21d..25c716052 100644
--- a/src/layer/arm/pooling_arm.cpp
+++ b/src/layer/arm/pooling_arm.cpp
@@ -31,6 +31,282 @@ int Pooling_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
     // max value in NxN window
     // avg value in NxN window
 
+    int w = bottom_blob.w;
+    int h = bottom_blob.h;
+    int channels = bottom_blob.c;
+    size_t elemsize = bottom_blob.elemsize;
+    int elempack = bottom_blob.elempack;
+
+    if (opt.use_packing_layout)
+    {
+
+//     fprintf(stderr, "Pooling     input %d x %d  pad = %d %d %d %d  ksize=%d %d  stride=%d %d\n", w, h, pad_left, pad_right, pad_top, pad_bottom, kernel_w, kernel_h, stride_w, stride_h);
+
+    if (elempack == 4)
+    {
+        if (global_pooling)
+        {
+            top_blob.create(channels, elemsize, elempack, opt.blob_allocator);
+            if (top_blob.empty())
+                return -100;
+
+            int size = w * h;
+
+            if (pooling_type == PoolMethod_MAX)
+            {
+                #pragma omp parallel for num_threads(opt.num_threads)
+                for (int q=0; q<channels; q++)
+                {
+                    const float* ptr = bottom_blob.channel(q);
+
+                    float32x4_t _max = vld1q_f32(ptr);
+                    for (int i=0; i<size; i++)
+                    {
+                        float32x4_t _val = vld1q_f32(ptr);
+                        _max = vmaxq_f32(_max, _val);
+                        ptr += 4;
+                    }
+
+                    float* outptr = top_blob;
+                    vst1q_f32(outptr + q * 4, _max);
+                }
+            }
+            else if (pooling_type == PoolMethod_AVE)
+            {
+                #pragma omp parallel for num_threads(opt.num_threads)
+                for (int q=0; q<channels; q++)
+                {
+                    const float* ptr = bottom_blob.channel(q);
+
+                    float32x4_t _sum = vdupq_n_f32(0.f);
+                    for (int i=0; i<size; i++)
+                    {
+                        float32x4_t _val = vld1q_f32(ptr);
+                        _sum = vaddq_f32(_sum, _val);
+                        ptr += 4;
+                    }
+
+                    float32x4_t _inv_size = vdupq_n_f32(1.f / size);
+                    float32x4_t _avg = vmulq_f32(_sum, _inv_size);
+
+                    float* outptr = top_blob;
+                    vst1q_f32(outptr + q * 4, _avg);
+                }
+            }
+
+            return 0;
+        }
+
+        Mat bottom_blob_bordered = bottom_blob;
+
+        float pad_value = 0.f;
+        if (pooling_type == PoolMethod_MAX)
+        {
+            pad_value = -FLT_MAX;
+        }
+        else if (pooling_type == PoolMethod_AVE)
+        {
+            pad_value = 0.f;
+        }
+
+        int wtailpad = 0;
+        int htailpad = 0;
+
+        if (pad_mode == 0) // full padding
+        {
+            int wtail = (w + pad_left + pad_right - kernel_w) % stride_w;
+            int htail = (h + pad_top + pad_bottom - kernel_h) % stride_h;
+
+            if (wtail != 0)
+                wtailpad = stride_w - wtail;
+            if (htail != 0)
+                htailpad = stride_h - htail;
+
+            copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom + htailpad, pad_left, pad_right + wtailpad, BORDER_CONSTANT, pad_value, opt.workspace_allocator, opt.num_threads);
+            if (bottom_blob_bordered.empty())
+                return -100;
+
+            w = bottom_blob_bordered.w;
+            h = bottom_blob_bordered.h;
+        }
+        else if (pad_mode == 1) // valid padding
+        {
+            copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, pad_value, opt.workspace_allocator, opt.num_threads);
+            if (bottom_blob_bordered.empty())
+                return -100;
+
+            w = bottom_blob_bordered.w;
+            h = bottom_blob_bordered.h;
+        }
+        else if (pad_mode == 2) // tensorflow padding=SAME
+        {
+            int wpad = kernel_w + (w - 1) / stride_w * stride_w - w;
+            int hpad = kernel_h + (h - 1) / stride_h * stride_h - h;
+            if (wpad > 0 || hpad > 0)
+            {
+                copy_make_border(bottom_blob, bottom_blob_bordered, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, BORDER_CONSTANT, pad_value, opt.workspace_allocator, opt.num_threads);
+                if (bottom_blob_bordered.empty())
+                    return -100;
+            }
+
+            w = bottom_blob_bordered.w;
+            h = bottom_blob_bordered.h;
+        }
+
+        int outw = (w - kernel_w) / stride_w + 1;
+        int outh = (h - kernel_h) / stride_h + 1;
+
+        top_blob.create(outw, outh, channels, elemsize, elempack, opt.blob_allocator);
+        if (top_blob.empty())
+            return -100;
+
+        const int maxk = kernel_w * kernel_h;
+
+        // kernel offsets
+        std::vector<int> _space_ofs(maxk);
+        int* space_ofs = &_space_ofs[0];
+        {
+            int p1 = 0;
+            int p2 = 0;
+            int gap = w - kernel_w;
+            for (int i = 0; i < kernel_h; i++)
+            {
+                for (int j = 0; j < kernel_w; j++)
+                {
+                    space_ofs[p1] = p2;
+                    p1++;
+                    p2++;
+                }
+                p2 += gap;
+            }
+        }
+
+        if (pooling_type == PoolMethod_MAX)
+        {
+            #pragma omp parallel for num_threads(opt.num_threads)
+            for (int q=0; q<channels; q++)
+            {
+                const Mat m = bottom_blob_bordered.channel(q);
+                float* outptr = top_blob.channel(q);
+
+                for (int i = 0; i < outh; i++)
+                {
+                    for (int j = 0; j < outw; j++)
+                    {
+                        const float* sptr = m.row(i*stride_h) + j*stride_w * 4;
+
+                        float32x4_t _max = vld1q_f32(sptr);
+
+                        for (int k = 0; k < maxk; k++)
+                        {
+                            float32x4_t _val = vld1q_f32( sptr + space_ofs[k] * 4 );
+                            _max = vmaxq_f32(_max, _val);
+                        }
+
+                        vst1q_f32(outptr + j * 4, _max);
+                    }
+
+                    outptr += outw * 4;
+                }
+            }
+        }
+        else if (pooling_type == PoolMethod_AVE)
+        {
+            #pragma omp parallel for num_threads(opt.num_threads)
+            for (int q=0; q<channels; q++)
+            {
+                const Mat m = bottom_blob_bordered.channel(q);
+                float* outptr = top_blob.channel(q);
+
+                float32x4_t _inv_maxk = vdupq_n_f32(1.f / maxk);
+
+                for (int i = 0; i < outh; i++)
+                {
+                    for (int j = 0; j < outw; j++)
+                    {
+                        const float* sptr = m.row(i*stride_h) + j*stride_w * 4;
+
+                        float32x4_t _sum = vdupq_n_f32(0.f);
+
+                        for (int k = 0; k < maxk; k++)
+                        {
+                            float32x4_t _val = vld1q_f32( sptr + space_ofs[k] * 4 );
+                            _sum = vaddq_f32(_sum, _val);
+                        }
+
+                        float32x4_t _avg = vmulq_f32(_sum, _inv_maxk);
+                        vst1q_f32(outptr + j * 4, _avg);
+                    }
+
+                    outptr += outw * 4;
+                }
+
+                // fix pad
+                if (pad_top != 0)
+                {
+                    const float scale = (float)kernel_h / (kernel_h - pad_top);
+                    float32x4_t _scale = vdupq_n_f32(scale);
+
+                    outptr = top_blob.channel(q).row(0);
+                    for (int i = 0; i < outw; i++)
+                    {
+                        float32x4_t _v = vld1q_f32(outptr);
+                        _v = vmulq_f32(_v, _scale);
+                        vst1q_f32(outptr, _v);
+                        outptr += 4;
+                    }
+                }
+                if (pad_bottom + htailpad != 0)
+                {
+                    const float scale = (float)kernel_h / (kernel_h - pad_bottom - htailpad);
+                    float32x4_t _scale = vdupq_n_f32(scale);
+
+                    outptr = top_blob.channel(q).row(outh - 1);
+                    for (int i = 0; i < outw; i++)
+                    {
+                        float32x4_t _v = vld1q_f32(outptr);
+                        _v = vmulq_f32(_v, _scale);
+                        vst1q_f32(outptr, _v);
+                        outptr += 4;
+                    }
+                }
+                if (pad_left != 0)
+                {
+                    const float scale = (float)kernel_w / (kernel_w - pad_left);
+                    float32x4_t _scale = vdupq_n_f32(scale);
+
+                    outptr = top_blob.channel(q);
+                    for (int i = 0; i < outh; i++)
+                    {
+                        float32x4_t _v = vld1q_f32(outptr);
+                        _v = vmulq_f32(_v, _scale);
+                        vst1q_f32(outptr, _v);
+                        outptr += outw * 4;
+                    }
+                }
+                if (pad_right + wtailpad != 0)
+                {
+                    const float scale = (float)kernel_w / (kernel_w - pad_right - wtailpad);
+                    float32x4_t _scale = vdupq_n_f32(scale);
+
+                    outptr = top_blob.channel(q);
+                    outptr += (outw - 1) * 4;
+                    for (int i = 0; i < outh; i++)
+                    {
+                        float32x4_t _v = vld1q_f32(outptr);
+                        _v = vmulq_f32(_v, _scale);
+                        vst1q_f32(outptr, _v);
+                        outptr += outw * 4;
+                    }
+                }
+            }
+        }
+
+        return 0;
+    }
+
+    } // opt.use_packing_layout
+
     if (kernel_w != kernel_h || stride_w != stride_h)
     {
         return Pooling::forward(bottom_blob, top_blob, opt);
@@ -49,11 +325,6 @@ int Pooling_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
         return Pooling::forward(bottom_blob, top_blob, opt);
     }
 
-    int w = bottom_blob.w;
-    int h = bottom_blob.h;
-    int channels = bottom_blob.c;
-    size_t elemsize = bottom_blob.elemsize;
-
     Mat bottom_blob_bordered = bottom_blob;
 
     float pad_value = 0.f;