From d85775fbcd92f46acd5a64c9edeba7d83d49ecac Mon Sep 17 00:00:00 2001
From: nihui <shuizhuyuanluo@126.com>
Date: Sun, 3 Mar 2019 14:20:05 +0800
Subject: [PATCH] fix softmax axis order on 3-dim, fix caffe reshape
 conversion, regenerate ssd param

---
 benchmark/mobilenet_ssd.param                 |  2 +-
 benchmark/squeezenet_ssd.param                |  2 +-
 src/layer/concat.cpp                          | 18 +++----
 src/layer/flatten.cpp                         | 11 +++-
 src/layer/shader/softmax_div_sum.comp         | 14 +++--
 src/layer/shader/softmax_div_sum_pack4.comp   | 14 +++--
 src/layer/shader/softmax_exp_sub_max.comp     | 14 +++--
 .../shader/softmax_exp_sub_max_pack4.comp     | 14 +++--
 src/layer/shader/softmax_reduce_max.comp      | 46 ++++++++++++----
 .../shader/softmax_reduce_max_pack4.comp      | 50 ++++++++++++++----
 src/layer/shader/softmax_reduce_sum.comp      | 48 +++++++++++++----
 .../shader/softmax_reduce_sum_pack4.comp      | 52 +++++++++++++++----
 src/layer/softmax.cpp                         | 44 ++++++++--------
 tools/caffe/caffe2ncnn.cpp                    |  6 +--
 14 files changed, 240 insertions(+), 95 deletions(-)

diff --git a/benchmark/mobilenet_ssd.param b/benchmark/mobilenet_ssd.param
index c1362b7dc..5bda98c28 100644
--- a/benchmark/mobilenet_ssd.param
+++ b/benchmark/mobilenet_ssd.param
@@ -123,7 +123,7 @@ PriorBox         conv17_2_mbox_priorbox 2 1 conv17_2_conv17_2/relu_splitncnn_0 d
 Concat           mbox_loc         6 1 conv11_mbox_loc_flat conv13_mbox_loc_flat conv14_2_mbox_loc_flat conv15_2_mbox_loc_flat conv16_2_mbox_loc_flat conv17_2_mbox_loc_flat mbox_loc 0=0
 Concat           mbox_conf        6 1 conv11_mbox_conf_flat conv13_mbox_conf_flat conv14_2_mbox_conf_flat conv15_2_mbox_conf_flat conv16_2_mbox_conf_flat conv17_2_mbox_conf_flat mbox_conf 0=0
 Concat           mbox_priorbox    6 1 conv11_mbox_priorbox conv13_mbox_priorbox conv14_2_mbox_priorbox conv15_2_mbox_priorbox conv16_2_mbox_priorbox conv17_2_mbox_priorbox mbox_priorbox 0=1
-Reshape          mbox_conf_reshape 1 1 mbox_conf mbox_conf_reshape 0=21 1=-1 2=0 3=0
+Reshape          mbox_conf_reshape 1 1 mbox_conf mbox_conf_reshape 0=21 1=-1 2=-233 3=0
 Softmax          mbox_conf_softmax 1 1 mbox_conf_reshape mbox_conf_softmax 0=1
 Flatten          mbox_conf_flatten 1 1 mbox_conf_softmax mbox_conf_flatten
 DetectionOutput  detection_out    3 1 mbox_loc mbox_conf_flatten mbox_priorbox detection_out 0=21 1=0.450000 2=100 3=100 4=0.250000
diff --git a/benchmark/squeezenet_ssd.param b/benchmark/squeezenet_ssd.param
index f4c4c6657..cb54f96ca 100644
--- a/benchmark/squeezenet_ssd.param
+++ b/benchmark/squeezenet_ssd.param
@@ -175,7 +175,7 @@ PriorBox         conv13_2_mbox_priorbox 2 1 conv13_2_conv13_2/relu_splitncnn_0 d
 Concat           mbox_loc         6 1 fire5_mbox_loc_flat fire9_mbox_loc_flat fire10_mbox_loc_flat fire11_mbox_loc_flat conv12_2_mbox_loc_flat conv13_2_mbox_loc_flat mbox_loc 0=0
 Concat           mbox_conf        6 1 fire5_mbox_conf_flat fire9_mbox_conf_flat fire10_mbox_conf_flat fire11_mbox_conf_flat conv12_2_mbox_conf_flat conv13_2_mbox_conf_flat mbox_conf 0=0
 Concat           mbox_priorbox    6 1 fire5_mbox_priorbox fire9_mbox_priorbox fire10_mbox_priorbox fire11_mbox_priorbox conv12_2_mbox_priorbox conv13_2_mbox_priorbox mbox_priorbox 0=1
-Reshape          mbox_conf_reshape 1 1 mbox_conf mbox_conf_reshape 0=21 1=-1 2=0 3=0
+Reshape          mbox_conf_reshape 1 1 mbox_conf mbox_conf_reshape 0=21 1=-1 2=-233 3=0
 Softmax          mbox_conf_softmax 1 1 mbox_conf_reshape mbox_conf_softmax 0=1
 Flatten          mbox_conf_flatten 1 1 mbox_conf_softmax mbox_conf_flatten
 DetectionOutput  detection_out    3 1 mbox_loc mbox_conf_flatten mbox_priorbox detection_out 0=21 1=0.450000 2=100 3=100 4=0.250000
diff --git a/src/layer/concat.cpp b/src/layer/concat.cpp
index b6f74f796..dc6c0f06d 100644
--- a/src/layer/concat.cpp
+++ b/src/layer/concat.cpp
@@ -370,15 +370,15 @@ int Concat::forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>&
             constants[10].i = woffset;
 
             const Pipeline* pipeline = 0;
-            if (packing == 1 && out_packing == 1)
+            if (bottom_blob.packing == 1 && out_packing == 1)
             {
                 pipeline = pipeline_concat;
             }
-            else if (packing == 4 && out_packing == 4)
+            else if (bottom_blob.packing == 4 && out_packing == 4)
             {
                 pipeline = pipeline_concat_pack4;
             }
-            else if (packing == 4 && out_packing == 1)
+            else if (bottom_blob.packing == 4 && out_packing == 1)
             {
                 pipeline = pipeline_concat_pack4to1;
             }
@@ -450,15 +450,15 @@ int Concat::forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>&
             constants[10].i = hoffset;
 
             const Pipeline* pipeline = 0;
-            if (packing == 1 && out_packing == 1)
+            if (bottom_blob.packing == 1 && out_packing == 1)
             {
                 pipeline = pipeline_concat;
             }
-            else if (packing == 4 && out_packing == 4)
+            else if (bottom_blob.packing == 4 && out_packing == 4)
             {
                 pipeline = pipeline_concat_pack4;
             }
-            else if (packing == 4 && out_packing == 1)
+            else if (bottom_blob.packing == 4 && out_packing == 1)
             {
                 pipeline = pipeline_concat_pack4to1;
             }
@@ -587,15 +587,15 @@ int Concat::forward(const std::vector<VkMat>& bottom_blobs, std::vector<VkMat>&
             constants[10].i = coffset;
 
             const Pipeline* pipeline = 0;
-            if (packing == 1 && out_packing == 1)
+            if (bottom_blob.packing == 1 && out_packing == 1)
             {
                 pipeline = pipeline_concat;
             }
-            else if (packing == 4 && out_packing == 4)
+            else if (bottom_blob.packing == 4 && out_packing == 4)
             {
                 pipeline = pipeline_concat_pack4;
             }
-            else if (packing == 4 && out_packing == 1)
+            else if (bottom_blob.packing == 4 && out_packing == 1)
             {
                 pipeline = pipeline_concat_pack4to1;
             }
diff --git a/src/layer/flatten.cpp b/src/layer/flatten.cpp
index b0e30cc4e..17baaad3b 100644
--- a/src/layer/flatten.cpp
+++ b/src/layer/flatten.cpp
@@ -83,12 +83,21 @@ int Flatten::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd,
 {
     int dims = bottom_blob.dims;
 
-    if (dims == 1 || dims == 2)
+    if (dims == 1)
     {
         top_blob = bottom_blob;
         return 0;
     }
 
+    if (dims == 2)
+    {
+        top_blob = bottom_blob;
+        top_blob.dims = 1;
+        top_blob.w = bottom_blob.w * bottom_blob.h;
+        top_blob.h = 1;
+        return 0;
+    }
+
     int w = bottom_blob.w;
     int h = bottom_blob.h;
     int channels = bottom_blob.c;
diff --git a/src/layer/shader/softmax_div_sum.comp b/src/layer/shader/softmax_div_sum.comp
index b1ffbbd03..48949a786 100644
--- a/src/layer/shader/softmax_div_sum.comp
+++ b/src/layer/shader/softmax_div_sum.comp
@@ -55,32 +55,36 @@ void main()
 
     if (p.dims == 2 && axis == 0)
     {
-        // FIXME TODO
+        int gi = gy * p.w + gx;
+        bottom_top_blob_data[gi] /= sum_workspace_data[gx];
         return;
     }
 
     if (p.dims == 2 && axis == 1)
     {
-        // FIXME TODO
+        int gi = gy * p.w + gx;
+        bottom_top_blob_data[gi] /= sum_workspace_data[gy];
         return;
     }
 
     if (p.dims == 3 && axis == 0)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        bottom_top_blob_data[gi] /= sum_workspace_data[ gy * p.w + gx ];
+        bottom_top_blob_data[gi] /= sum_workspace_data[gy * p.w + gx];
         return;
     }
 
     if (p.dims == 3 && axis == 1)
     {
-        // FIXME TODO
+        int gi = gz * p.cstep + gy * p.w + gx;
+        bottom_top_blob_data[gi] /= sum_workspace_data[gz * p.w + gx];
         return;
     }
 
     if (p.dims == 3 && axis == 2)
     {
-        // FIXME TODO
+        int gi = gz * p.cstep + gy * p.w + gx;
+        bottom_top_blob_data[gi] /= sum_workspace_data[gz * p.h + gy];
         return;
     }
 }
diff --git a/src/layer/shader/softmax_div_sum_pack4.comp b/src/layer/shader/softmax_div_sum_pack4.comp
index 1c1e10f57..702927808 100644
--- a/src/layer/shader/softmax_div_sum_pack4.comp
+++ b/src/layer/shader/softmax_div_sum_pack4.comp
@@ -55,32 +55,36 @@ void main()
 
     if (p.dims == 2 && axis == 0)
     {
-        // FIXME TODO
+        int gi = gy * p.w + gx;
+        bottom_top_blob_data[gi] /= sum_workspace_data[gx];
         return;
     }
 
     if (p.dims == 2 && axis == 1)
     {
-        // FIXME TODO
+        int gi = gy * p.w + gx;
+        bottom_top_blob_data[gi] /= sum_workspace_data[gy];
         return;
     }
 
     if (p.dims == 3 && axis == 0)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        bottom_top_blob_data[gi] /= sum_workspace_data[ gy * p.w + gx ];
+        bottom_top_blob_data[gi] /= sum_workspace_data[gy * p.w + gx];
         return;
     }
 
     if (p.dims == 3 && axis == 1)
     {
-        // FIXME TODO
+        int gi = gz * p.cstep + gy * p.w + gx;
+        bottom_top_blob_data[gi] /= sum_workspace_data[gz * p.w + gx];
         return;
     }
 
     if (p.dims == 3 && axis == 2)
     {
-        // FIXME TODO
+        int gi = gz * p.cstep + gy * p.w + gx;
+        bottom_top_blob_data[gi] /= sum_workspace_data[gz * p.h + gy];
         return;
     }
 }
diff --git a/src/layer/shader/softmax_exp_sub_max.comp b/src/layer/shader/softmax_exp_sub_max.comp
index 0c8ff8bb4..838c4e84d 100644
--- a/src/layer/shader/softmax_exp_sub_max.comp
+++ b/src/layer/shader/softmax_exp_sub_max.comp
@@ -55,32 +55,36 @@ void main()
 
     if (p.dims == 2 && axis == 0)
     {
-        // FIXME TODO
+        int gi = gy * p.w + gx;
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gx]);
         return;
     }
 
     if (p.dims == 2 && axis == 1)
     {
-        // FIXME TODO
+        int gi = gy * p.w + gx;
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gy]);
         return;
     }
 
     if (p.dims == 3 && axis == 0)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[ gy * p.w + gx ]);
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gy * p.w + gx]);
         return;
     }
 
     if (p.dims == 3 && axis == 1)
     {
-        // FIXME TODO
+        int gi = gz * p.cstep + gy * p.w + gx;
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gz * p.w + gx]);
         return;
     }
 
     if (p.dims == 3 && axis == 2)
     {
-        // FIXME TODO
+        int gi = gz * p.cstep + gy * p.w + gx;
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gz * p.h + gy]);
         return;
     }
 }
diff --git a/src/layer/shader/softmax_exp_sub_max_pack4.comp b/src/layer/shader/softmax_exp_sub_max_pack4.comp
index b16e8bedb..99af1498a 100644
--- a/src/layer/shader/softmax_exp_sub_max_pack4.comp
+++ b/src/layer/shader/softmax_exp_sub_max_pack4.comp
@@ -55,32 +55,36 @@ void main()
 
     if (p.dims == 2 && axis == 0)
     {
-        // FIXME TODO
+        int gi = gy * p.w + gx;
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gx]);
         return;
     }
 
     if (p.dims == 2 && axis == 1)
     {
-        // FIXME TODO
+        int gi = gy * p.w + gx;
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gy]);
         return;
     }
 
     if (p.dims == 3 && axis == 0)
     {
         int gi = gz * p.cstep + gy * p.w + gx;
-        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[ gy * p.w + gx ]);
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gy * p.w + gx]);
         return;
     }
 
     if (p.dims == 3 && axis == 1)
     {
-        // FIXME TODO
+        int gi = gz * p.cstep + gy * p.w + gx;
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gz * p.w + gx]);
         return;
     }
 
     if (p.dims == 3 && axis == 2)
     {
-        // FIXME TODO
+        int gi = gz * p.cstep + gy * p.w + gx;
+        bottom_top_blob_data[gi] = exp(bottom_top_blob_data[gi] - max_workspace_data[gz * p.h + gy]);
         return;
     }
 }
diff --git a/src/layer/shader/softmax_reduce_max.comp b/src/layer/shader/softmax_reduce_max.comp
index b32a547a2..0e9e7563c 100644
--- a/src/layer/shader/softmax_reduce_max.comp
+++ b/src/layer/shader/softmax_reduce_max.comp
@@ -44,13 +44,13 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= p.w || gy >= p.h || gz >= p.c)
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
         return;
 
     if (p.dims == 1) // axis == 0
     {
         float max_value = -99999999;
-        for (int i=0; i<p.w; i++)
+        for (int i = 0; i < p.w; i++)
         {
             max_value = max(max_value, bottom_top_blob_data[i]);
         }
@@ -61,38 +61,66 @@ void main()
 
     if (p.dims == 2 && axis == 0)
     {
-        // FIXME TODO
+        float max_value = -99999999;
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = i * p.w + gx;
+            max_value = max(max_value, bottom_top_blob_data[v_offset]);
+        }
+        max_workspace_data[gx] = max_value;
+
         return;
     }
 
     if (p.dims == 2 && axis == 1)
     {
-        // FIXME TODO
+        float max_value = -99999999;
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gx * p.w + i;
+            max_value = max(max_value, bottom_top_blob_data[v_offset]);
+        }
+        max_workspace_data[gx] = max_value;
+
         return;
     }
 
     if (p.dims == 3 && axis == 0)
     {
         float max_value = -99999999;
-        for (int z = 0; z < p.c; z++)
+        for (int i = 0; i < p.c; i++)
         {
-            int v_offset = z * p.cstep + gy * p.w + gx;
+            int v_offset = i * p.cstep + gy * p.w + gx;
             max_value = max(max_value, bottom_top_blob_data[v_offset]);
         }
-        max_workspace_data[ gy * p.w + gx ] = max_value;
+        max_workspace_data[gy * p.w + gx] = max_value;
 
         return;
     }
 
     if (p.dims == 3 && axis == 1)
     {
-        // FIXME TODO
+        float max_value = -99999999;
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = gy * p.cstep + i * p.w + gx;
+            max_value = max(max_value, bottom_top_blob_data[v_offset]);
+        }
+        max_workspace_data[gy * p.w + gx] = max_value;
+
         return;
     }
 
     if (p.dims == 3 && axis == 2)
     {
-        // FIXME TODO
+        float max_value = -99999999;
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gy * p.cstep + gx * p.w + i;
+            max_value = max(max_value, bottom_top_blob_data[v_offset]);
+        }
+        max_workspace_data[gy * p.h + gx] = max_value;
+
         return;
     }
 }
diff --git a/src/layer/shader/softmax_reduce_max_pack4.comp b/src/layer/shader/softmax_reduce_max_pack4.comp
index 48e5b9179..fda3e56dc 100644
--- a/src/layer/shader/softmax_reduce_max_pack4.comp
+++ b/src/layer/shader/softmax_reduce_max_pack4.comp
@@ -44,13 +44,13 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= p.w || gy >= p.h || gz >= p.c)
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
         return;
 
     if (p.dims == 1) // axis == 0
     {
         vec4 max_value = vec4(-99999999);
-        for (int i=0; i<p.w; i++)
+        for (int i = 0; i < p.w; i++)
         {
             max_value = max(max_value, bottom_top_blob_data[i]);
         }
@@ -62,39 +62,71 @@ void main()
 
     if (p.dims == 2 && axis == 0)
     {
-        // FIXME TODO
+        vec4 max_value = vec4(-99999999);
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = i * p.w + gx;
+            max_value = max(max_value, bottom_top_blob_data[v_offset]);
+        }
+        vec2 max2 = max(max_value.rg, max_value.ba);
+        max_workspace_data[gx] = max(max2.r, max2.g);
+
         return;
     }
 
     if (p.dims == 2 && axis == 1)
     {
-        // FIXME TODO
+        vec4 max_value = vec4(-99999999);
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gx * p.w + i;
+            max_value = max(max_value, bottom_top_blob_data[v_offset]);
+        }
+        vec2 max2 = max(max_value.rg, max_value.ba);
+        max_workspace_data[gx] = max(max2.r, max2.g);
+
         return;
     }
 
     if (p.dims == 3 && axis == 0)
     {
         vec4 max_value = vec4(-99999999);
-        for (int z = 0; z < p.c; z++)
+        for (int i = 0; i < p.c; i++)
         {
-            int v_offset = z * p.cstep + gy * p.w + gx;
+            int v_offset = i * p.cstep + gy * p.w + gx;
             max_value = max(max_value, bottom_top_blob_data[v_offset]);
         }
         vec2 max2 = max(max_value.rg, max_value.ba);
-        max_workspace_data[ gy * p.w + gx ] = max(max2.r, max2.g);
+        max_workspace_data[gy * p.w + gx] = max(max2.r, max2.g);
 
         return;
     }
 
     if (p.dims == 3 && axis == 1)
     {
-        // FIXME TODO
+        vec4 max_value = vec4(-99999999);
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = gy * p.cstep + i * p.w + gx;
+            max_value = max(max_value, bottom_top_blob_data[v_offset]);
+        }
+        vec2 max2 = max(max_value.rg, max_value.ba);
+        max_workspace_data[gy * p.w + gx] = max(max2.r, max2.g);
+
         return;
     }
 
     if (p.dims == 3 && axis == 2)
     {
-        // FIXME TODO
+        vec4 max_value = vec4(-99999999);
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gy * p.cstep + gx * p.w + i;
+            max_value = max(max_value, bottom_top_blob_data[v_offset]);
+        }
+        vec2 max2 = max(max_value.rg, max_value.ba);
+        max_workspace_data[gy * p.h + gx] = max(max2.r, max2.g);
+
         return;
     }
 }
diff --git a/src/layer/shader/softmax_reduce_sum.comp b/src/layer/shader/softmax_reduce_sum.comp
index 9e6886dcb..77fc8de5a 100644
--- a/src/layer/shader/softmax_reduce_sum.comp
+++ b/src/layer/shader/softmax_reduce_sum.comp
@@ -44,13 +44,13 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= p.w || gy >= p.h || gz >= p.c)
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
         return;
 
     if (p.dims == 1) // axis == 0
     {
         float sum_value = 0.f;
-        for (int i=0; i<p.w; i++)
+        for (int i = 0; i < p.w; i++)
         {
             sum_value += bottom_top_blob_data[i];
         }
@@ -61,38 +61,66 @@ void main()
 
     if (p.dims == 2 && axis == 0)
     {
-        // FIXME TODO
+        float sum_value = 0.f;
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = i * p.w + gx;
+            sum_value += bottom_top_blob_data[v_offset];
+        }
+        sum_workspace_data[gx] = sum_value;
+
         return;
     }
 
     if (p.dims == 2 && axis == 1)
     {
-        // FIXME TODO
+        float sum_value = 0.f;
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gx * p.w + i;
+            sum_value += bottom_top_blob_data[v_offset];
+        }
+        sum_workspace_data[gx] = sum_value;
+
         return;
     }
 
     if (p.dims == 3 && axis == 0)
     {
-        float sum_value = 0;
-        for (int z = 0; z < p.c; z++)
+        float sum_value = 0.f;
+        for (int i = 0; i < p.c; i++)
         {
-            int v_offset = z * p.cstep + gy * p.w + gx;
+            int v_offset = i * p.cstep + gy * p.w + gx;
             sum_value += bottom_top_blob_data[v_offset];
         }
-        sum_workspace_data[ gy * p.w + gx ] = sum_value;
+        sum_workspace_data[gy * p.w + gx] = sum_value;
 
         return;
     }
 
     if (p.dims == 3 && axis == 1)
     {
-        // FIXME TODO
+        float sum_value = 0.f;
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = gy * p.cstep + i * p.w + gx;
+            sum_value += bottom_top_blob_data[v_offset];
+        }
+        sum_workspace_data[gy * p.w + gx] = sum_value;
+
         return;
     }
 
     if (p.dims == 3 && axis == 2)
     {
-        // FIXME TODO
+        float sum_value = 0.f;
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gy * p.cstep + gx * p.w + i;
+            sum_value += bottom_top_blob_data[v_offset];
+        }
+        sum_workspace_data[gy * p.h + gx] = sum_value;
+
         return;
     }
 }
diff --git a/src/layer/shader/softmax_reduce_sum_pack4.comp b/src/layer/shader/softmax_reduce_sum_pack4.comp
index 8f5aaff11..c084010d2 100644
--- a/src/layer/shader/softmax_reduce_sum_pack4.comp
+++ b/src/layer/shader/softmax_reduce_sum_pack4.comp
@@ -44,13 +44,13 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= p.w || gy >= p.h || gz >= p.c)
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
         return;
 
     if (p.dims == 1) // axis == 0
     {
-        vec4 sum_value = vec4(0.0);
-        for (int i=0; i<p.w; i++)
+        vec4 sum_value = vec4(0.f);
+        for (int i = 0; i < p.w; i++)
         {
             sum_value += bottom_top_blob_data[i];
         }
@@ -62,22 +62,38 @@ void main()
 
     if (p.dims == 2 && axis == 0)
     {
-        // FIXME TODO
+        vec4 sum_value = vec4(0.f);
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = i * p.w + gx;
+            sum_value += bottom_top_blob_data[v_offset];
+        }
+        vec2 sum2 = sum_value.rg + sum_value.ba;
+        sum_workspace_data[gx] = sum2.r + sum2.g;
+
         return;
     }
 
     if (p.dims == 2 && axis == 1)
     {
-        // FIXME TODO
+        vec4 sum_value = vec4(0.f);
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gx * p.w + i;
+            sum_value += bottom_top_blob_data[v_offset];
+        }
+        vec2 sum2 = sum_value.rg + sum_value.ba;
+        sum_workspace_data[gx] = sum2.r + sum2.g;
+
         return;
     }
 
     if (p.dims == 3 && axis == 0)
     {
-        vec4 sum_value = vec4(0.0);
-        for (int z = 0; z < p.c; z++)
+        vec4 sum_value = vec4(0.f);
+        for (int i = 0; i < p.c; i++)
         {
-            int v_offset = z * p.cstep + gy * p.w + gx;
+            int v_offset = i * p.cstep + gy * p.w + gx;
             sum_value += bottom_top_blob_data[v_offset];
         }
         vec2 sum2 = sum_value.rg + sum_value.ba;
@@ -88,13 +104,29 @@ void main()
 
     if (p.dims == 3 && axis == 1)
     {
-        // FIXME TODO
+        vec4 sum_value = vec4(0.f);
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = gy * p.cstep + i * p.w + gx;
+            sum_value += bottom_top_blob_data[v_offset];
+        }
+        vec2 sum2 = sum_value.rg + sum_value.ba;
+        sum_workspace_data[gy * p.w + gx] = sum2.r + sum2.g;
+
         return;
     }
 
     if (p.dims == 3 && axis == 2)
     {
-        // FIXME TODO
+        vec4 sum_value = vec4(0.f);
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gy * p.cstep + gx * p.w + i;
+            sum_value += bottom_top_blob_data[v_offset];
+        }
+        vec2 sum2 = sum_value.rg + sum_value.ba;
+        sum_workspace_data[gy * p.h + gx] = sum2.r + sum2.g;
+
         return;
     }
 }
diff --git a/src/layer/softmax.cpp b/src/layer/softmax.cpp
index 6fb768aa4..c99b231cf 100644
--- a/src/layer/softmax.cpp
+++ b/src/layer/softmax.cpp
@@ -278,7 +278,7 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
         int channels = bottom_top_blob.c;
 
         Mat max;
-        max.create(h, channels, elemsize, opt.workspace_allocator);
+        max.create(w, channels, elemsize, opt.workspace_allocator);
         if (max.empty())
             return -100;
         max.fill(-FLT_MAX);
@@ -290,13 +290,11 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 
             for (int i=0; i<h; i++)
             {
-                float max = -FLT_MAX;
                 for (int j=0; j<w; j++)
                 {
-                    max = std::max(max, ptr[j]);
+                    maxptr[j] = std::max(maxptr[j], ptr[j]);
                 }
 
-                maxptr[i] = max;
                 ptr += w;
             }
         }
@@ -309,10 +307,9 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 
             for (int i=0; i<h; i++)
             {
-                float max = maxptr[i];
                 for (int j=0; j<w; j++)
                 {
-                    ptr[j] = exp(ptr[j] - max);
+                    ptr[j] = exp(ptr[j] - maxptr[j]);
                 }
 
                 ptr += w;
@@ -320,7 +317,7 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
         }
 
         Mat sum;
-        sum.create(h, channels, elemsize, opt.workspace_allocator);
+        sum.create(w, channels, elemsize, opt.workspace_allocator);
         if (sum.empty())
             return -100;
         sum.fill(0.f);
@@ -332,13 +329,11 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 
             for (int i=0; i<h; i++)
             {
-                float sum = 0.f;
                 for (int j=0; j<w; j++)
                 {
-                    sum += ptr[j];
+                    sumptr[j] += ptr[j];
                 }
 
-                sumptr[i] = sum;
                 ptr += w;
             }
         }
@@ -351,10 +346,9 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 
             for (int i=0; i<h; i++)
             {
-                float sum = sumptr[i];
                 for (int j=0; j<w; j++)
                 {
-                    ptr[j] /= sum;
+                    ptr[j] /= sumptr[j];
                 }
 
                 ptr += w;
@@ -371,7 +365,7 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
         int channels = bottom_top_blob.c;
 
         Mat max;
-        max.create(w, channels, elemsize, opt.workspace_allocator);
+        max.create(h, channels, elemsize, opt.workspace_allocator);
         if (max.empty())
             return -100;
         max.fill(-FLT_MAX);
@@ -383,11 +377,13 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 
             for (int i=0; i<h; i++)
             {
+                float max = -FLT_MAX;
                 for (int j=0; j<w; j++)
                 {
-                    maxptr[j] = std::max(maxptr[j], ptr[j]);
+                    max = std::max(max, ptr[j]);
                 }
 
+                maxptr[i] = max;
                 ptr += w;
             }
         }
@@ -400,9 +396,10 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 
             for (int i=0; i<h; i++)
             {
+                float max = maxptr[i];
                 for (int j=0; j<w; j++)
                 {
-                    ptr[j] = exp(ptr[j] - maxptr[j]);
+                    ptr[j] = exp(ptr[j] - max);
                 }
 
                 ptr += w;
@@ -410,7 +407,7 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
         }
 
         Mat sum;
-        sum.create(w, channels, elemsize, opt.workspace_allocator);
+        sum.create(h, channels, elemsize, opt.workspace_allocator);
         if (sum.empty())
             return -100;
         sum.fill(0.f);
@@ -422,11 +419,13 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 
             for (int i=0; i<h; i++)
             {
+                float sum = 0.f;
                 for (int j=0; j<w; j++)
                 {
-                    sumptr[j] += ptr[j];
+                    sum += ptr[j];
                 }
 
+                sumptr[i] = sum;
                 ptr += w;
             }
         }
@@ -439,9 +438,10 @@ int Softmax::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 
             for (int i=0; i<h; i++)
             {
+                float sum = sumptr[i];
                 for (int j=0; j<w; j++)
                 {
-                    ptr[j] /= sumptr[j];
+                    ptr[j] /= sum;
                 }
 
                 ptr += w;
@@ -558,13 +558,13 @@ int Softmax::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Optio
     }
     else if (dims == 3 && axis == 1)
     {
-        max_workspace.create(h, channels, 4u, opt.workspace_vkallocator, opt.staging_vkallocator);
-        sum_workspace.create(h, channels, 4u, opt.workspace_vkallocator, opt.staging_vkallocator);
+        max_workspace.create(w, channels, 4u, opt.workspace_vkallocator, opt.staging_vkallocator);
+        sum_workspace.create(w, channels, 4u, opt.workspace_vkallocator, opt.staging_vkallocator);
     }
     else if (dims == 3 && axis == 2)
     {
-        max_workspace.create(w, channels, 4u, opt.workspace_vkallocator, opt.staging_vkallocator);
-        sum_workspace.create(w, channels, 4u, opt.workspace_vkallocator, opt.staging_vkallocator);
+        max_workspace.create(h, channels, 4u, opt.workspace_vkallocator, opt.staging_vkallocator);
+        sum_workspace.create(h, channels, 4u, opt.workspace_vkallocator, opt.staging_vkallocator);
     }
 
 //     fprintf(stderr, "Softmax::forward_inplace %p\n", bottom_top_blob.buffer());
diff --git a/tools/caffe/caffe2ncnn.cpp b/tools/caffe/caffe2ncnn.cpp
index 5c1a2f58d..2451b5367 100644
--- a/tools/caffe/caffe2ncnn.cpp
+++ b/tools/caffe/caffe2ncnn.cpp
@@ -1482,7 +1482,7 @@ int main(int argc, char** argv)
             const caffe::ReorgParameter& reorg_param = layer.reorg_param();
             fprintf(pp, " 0=%d", reorg_param.stride());
         }
-        else if (layer.type() == "Reshape")// -1 1 512
+        else if (layer.type() == "Reshape")
         {
             const caffe::ReshapeParameter& reshape_param = layer.reshape_param();
             const caffe::BlobShape& bs = reshape_param.shape();
@@ -1492,11 +1492,11 @@ int main(int argc, char** argv)
             }
             else if (bs.dim_size() == 2)
             {
-                fprintf(pp, " 0=%ld 1=%ld 2=-233", bs.dim(1), bs.dim(0));
+                fprintf(pp, " 0=%ld 1=-233 2=-233", bs.dim(1));
             }
             else if (bs.dim_size() == 3)
             {
-                fprintf(pp, " 0=%ld 1=%ld 2=%ld", bs.dim(2), bs.dim(1), bs.dim(0));
+                fprintf(pp, " 0=%ld 1=%ld 2=-233", bs.dim(2), bs.dim(1));
             }
             else // bs.dim_size() == 4
             {