Modify ckProfiler

library/include/ck/library/tensor_operation_instance/gpu/pool2d_fwd.hpp

@@ -17,8 +17,9 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-static constexpr auto InOutRank  = 4;
-static constexpr auto WindowRank = 2;
+// pool2d inherit pool3d
+static constexpr auto InOutRank  = 5;
+static constexpr auto WindowRank = 3;
 
 static constexpr auto MaxOp = ck::ReduceTensorOp::MAX;
 static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;

profiler/include/profiler/profile_pool2d_fwd_impl.hpp

@@ -31,15 +31,16 @@ bool profile_pool2d_fwd_impl(int do_verification,
                              std::vector<index_t> in_length, // NCHW
                              std::vector<index_t> window_spatial_lengths,
                              std::vector<index_t> window_strides,
+                             std::vector<index_t> window_dilations,
                              std::vector<index_t> input_left_pads,
                              std::vector<index_t> input_right_pads)
 {
-    constexpr index_t InOutRank  = 4;
-    constexpr index_t WindowRank = 2;
+    constexpr index_t InOutRank  = 5;
+    constexpr index_t WindowRank = 3;
 
     if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
-       window_strides.size() != WindowRank || input_left_pads.size() != WindowRank ||
-       input_right_pads.size() != WindowRank)
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
         return false;
 
     std::vector<index_t> out_length(InOutRank);
@@ -53,11 +54,13 @@ bool profile_pool2d_fwd_impl(int do_verification,
     // Calculate Ho, Wo
     for(int i = 2; i < InOutRank; ++i)
     {
-        auto pad1           = input_left_pads[i - 2];
-        auto pad2           = input_right_pads[i - 2];
-        auto windows_size   = window_spatial_lengths[i - 2];
-        auto windows_stride = window_strides[i - 2];
-        out_length[i]       = (in_length[i] + pad1 + pad2 - windows_size) / windows_stride + 1;
+        auto pad1             = input_left_pads[i - 2];
+        auto pad2             = input_right_pads[i - 2];
+        auto windows_size     = window_spatial_lengths[i - 2];
+        auto windows_stride   = window_strides[i - 2];
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
     }
 
     int Hi = in_length[2];
@@ -131,6 +134,7 @@ bool profile_pool2d_fwd_impl(int do_verification,
                                              out_indices_n_c_ho_wo_host,
                                              window_spatial_lengths,
                                              window_strides,
+                                             window_dilations,
                                              input_left_pads,
                                              input_right_pads);
         auto ref_invoker  = ref.MakeInvoker();
@@ -152,6 +156,7 @@ bool profile_pool2d_fwd_impl(int do_verification,
             {C * Ho * Wo, 1, Wo * C, C},
             {C * Ho * Wo, 1, Wo * C, C},
             window_strides,
+            window_dilations,
             input_left_pads,
             input_right_pads,
             {2, 3});

profiler/include/profiler/profile_pool3d_fwd_impl.hpp

@@ -31,6 +31,7 @@ bool profile_pool3d_fwd_impl(int do_verification,
                              std::vector<index_t> in_length, // NCDHW
                              std::vector<index_t> window_spatial_lengths,
                              std::vector<index_t> window_strides,
+                             std::vector<index_t> window_dilations,
                              std::vector<index_t> input_left_pads,
                              std::vector<index_t> input_right_pads)
 {
@@ -38,8 +39,8 @@ bool profile_pool3d_fwd_impl(int do_verification,
     constexpr index_t WindowRank = 3;
 
     if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
-       window_strides.size() != WindowRank || input_left_pads.size() != WindowRank ||
-       input_right_pads.size() != WindowRank)
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
         return false;
 
     std::vector<index_t> out_length(InOutRank);
@@ -53,11 +54,13 @@ bool profile_pool3d_fwd_impl(int do_verification,
     // Calculate Do, Ho, Wo
     for(int i = 2; i < InOutRank; ++i)
     {
-        auto pad1           = input_left_pads[i - 2];
-        auto pad2           = input_right_pads[i - 2];
-        auto windows_size   = window_spatial_lengths[i - 2];
-        auto windows_stride = window_strides[i - 2];
-        out_length[i]       = (in_length[i] + pad1 + pad2 - windows_size) / windows_stride + 1;
+        auto pad1             = input_left_pads[i - 2];
+        auto pad2             = input_right_pads[i - 2];
+        auto windows_size     = window_spatial_lengths[i - 2];
+        auto windows_stride   = window_strides[i - 2];
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i] = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
     }
 
     int Di = in_length[2];
@@ -136,6 +139,7 @@ bool profile_pool3d_fwd_impl(int do_verification,
                                              out_indices_n_c_do_ho_wo_host,
                                              window_spatial_lengths,
                                              window_strides,
+                                             window_dilations,
                                              input_left_pads,
                                              input_right_pads);
         auto ref_invoker  = ref.MakeInvoker();
@@ -157,6 +161,7 @@ bool profile_pool3d_fwd_impl(int do_verification,
             {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
             {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
             window_strides,
+            window_dilations,
             input_left_pads,
             input_right_pads,
             {2, 3, 4});

profiler/src/profile_avg_pool2d_fwd.cpp

@@ -13,8 +13,12 @@ using ck::index_t;
 
 struct avgPoolFwdArgParser
 {
-    std::unordered_map<std::string, std::vector<int>> long_opts = {
-        {"length", {}}, {"wsize", {}}, {"wstride", {}}, {"pad1", {}}, {"pad2", {}}};
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
+                                                                   {"wsize", {}},
+                                                                   {"wstride", {}},
+                                                                   {"wdilation", {}},
+                                                                   {"pad1", {}},
+                                                                   {"pad2", {}}};
 
     bool parse_opt(int argc, char* argv[], const std::string& key, int i)
     {
@@ -55,10 +59,11 @@ void print_help_avg_pool2d_fwd()
               << "--length: input tensor length for NDHW(e.g, --length 2 32 30 30) \n"
               << "--wsize: window size for YX (e.g, --wsize 2 2) \n"
               << "--wstride: window stride for HW (e.g, --wstride 2 2) \n"
+              << "--wdilation: window dilation for DHW (e.g, --wdilation 1 1) \n"
               << "--pad1: left side of padding in HW (e.g, --pad1 1 1) \n"
               << "--pad2: right side of padding in HW (e.g, --pad2 1 1) \n"
               << "eg: ckProfiler avg_pool2d_fwd 0 1 2 0 1 0 --length 2 32 30 30 --wsize 2 2 "
-                 "--wstride 2 2 --pad1 1 1 --pad2 1 1"
+                 "--wstride 2 2 --wdilation 1 1 --pad1 1 1 --pad2 1 1"
               << std::endl;
 }
 
@@ -73,15 +78,16 @@ int profile_avg_pool2d_fwd(int argc, char* argv[])
     std::vector<index_t> in_length = {2, 32, 30, 30};
     std::vector<index_t> wsize     = {2, 2};
     std::vector<index_t> wstride   = {2, 2};
+    std::vector<index_t> wdilation = {1, 1};
     std::vector<index_t> pad1      = {1, 1};
     std::vector<index_t> pad2      = {1, 1};
 
-    if(argc != 2 && argc != 25)
+    if(argc != 2 && argc != 26)
     {
         print_help_avg_pool2d_fwd();
         return 0;
     }
-    else if(argc == 25)
+    else if(argc == 26)
     {
         data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
         do_verification = std::stoi(argv[3]);
@@ -95,6 +101,7 @@ int profile_avg_pool2d_fwd(int argc, char* argv[])
         in_length = arg_parser.long_opts["length"];
         wsize     = arg_parser.long_opts["wsize"];
         wstride   = arg_parser.long_opts["wstride"];
+        wdilation = arg_parser.long_opts["wdilation"];
         pad1      = arg_parser.long_opts["pad1"];
         pad2      = arg_parser.long_opts["pad2"];
     }
@@ -114,6 +121,7 @@ int profile_avg_pool2d_fwd(int argc, char* argv[])
             in_length,
             wsize,
             wstride,
+            wdilation,
             pad1,
             pad2);
     }
@@ -127,6 +135,7 @@ int profile_avg_pool2d_fwd(int argc, char* argv[])
             in_length,
             wsize,
             wstride,
+            wdilation,
             pad1,
             pad2);
     }

profiler/src/profile_max_pool3d_fwd.cpp

@@ -13,8 +13,12 @@ using ck::index_t;
 
 struct maxPoolFwdArgParser
 {
-    std::unordered_map<std::string, std::vector<int>> long_opts = {
-        {"length", {}}, {"wsize", {}}, {"wstride", {}}, {"pad1", {}}, {"pad2", {}}};
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
+                                                                   {"wsize", {}},
+                                                                   {"wstride", {}},
+                                                                   {"wdilation", {}},
+                                                                   {"pad1", {}},
+                                                                   {"pad2", {}}};
 
     bool parse_opt(int argc, char* argv[], const std::string& key, int i)
     {
@@ -56,10 +60,11 @@ void print_help_max_pool3d_fwd()
               << "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
               << "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
               << "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
+              << "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
               << "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
               << "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
               << "eg: ckProfiler max_pool3d_fwd 0 1 2 0 1 0 --length 2 32 30 30 30 --wsize 2 2 2 "
-                 "--wstride 2 2 2 --pad1 1 1 1 --pad2 1 1 1"
+                 "--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
               << std::endl;
 }
 
@@ -75,15 +80,16 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
     std::vector<index_t> in_length = {2, 32, 30, 30, 30};
     std::vector<index_t> wsize     = {2, 2, 2};
     std::vector<index_t> wstride   = {2, 2, 2};
+    std::vector<index_t> wdilation = {1, 1, 1};
     std::vector<index_t> pad1      = {1, 1, 1};
     std::vector<index_t> pad2      = {1, 1, 1};
 
-    if(argc != 2 && argc != 30)
+    if(argc != 2 && argc != 31)
     {
         print_help_max_pool3d_fwd();
         return 0;
     }
-    else if(argc == 30)
+    else if(argc == 31)
     {
         data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
         do_verification = std::stoi(argv[3]);
@@ -98,6 +104,7 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
         in_length = arg_parser.long_opts["length"];
         wsize     = arg_parser.long_opts["wsize"];
         wstride   = arg_parser.long_opts["wstride"];
+        wdilation = arg_parser.long_opts["wdilation"];
         pad1      = arg_parser.long_opts["pad1"];
         pad2      = arg_parser.long_opts["pad2"];
     }
@@ -118,6 +125,7 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
                 in_length,
                 wsize,
                 wstride,
+                wdilation,
                 pad1,
                 pad2);
         else
@@ -129,6 +137,7 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
                 in_length,
                 wsize,
                 wstride,
+                wdilation,
                 pad1,
                 pad2);
     }
@@ -143,6 +152,7 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
                 in_length,
                 wsize,
                 wstride,
+                wdilation,
                 pad1,
                 pad2);
         else
@@ -154,6 +164,7 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
                 in_length,
                 wsize,
                 wstride,
+                wdilation,
                 pad1,
                 pad2);
     }