merge from develop and revisison for pr#881

example/13_pool2d_fwd/CMakeLists.txt

-add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
-add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
-
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
+endif()

example/13_pool2d_fwd/pool2d_fwd_common.hpp

@@ -39,17 +39,18 @@ bool pool_test(bool do_verification,
                ck::index_t Wi,
                ck::index_t window_stride_h,
                ck::index_t window_stride_w,
+               ck::index_t window_dilation_h,
+               ck::index_t window_dilation_w,
                ck::index_t in_left_pad_h,
                ck::index_t in_left_pad_w,
                ck::index_t in_right_pad_h,
                ck::index_t in_right_pad_w)
 {
     using DevicePoolFwdInstance =
-        ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
-            InDataType,      // InDataType
-            OutDataType,     // OutDataType
-            IndexDataType,   // IndexDataType
-            ComputeDataType, // ComputeDataType
+        ck::tensor_operation::device::DevicePool2dFwd_NHWC_NHWC<InDataType,
+                                                                OutDataType,
+                                                                IndexDataType,
+                                                                ComputeDataType,
                                                                 ReduceOpId,
                                                                 OutputIndex,
                                                                 64, // BlockSize
@@ -57,13 +58,16 @@ bool pool_test(bool do_verification,
                                                                 1,  // ReduceKThreadClusterSize
                                                                 4,  // ReduceMThreadSliceSize
                                                                 1,  // ReduceKThreadSliceSize
-            4>; // InSrcOutDstVectorSize
+                                                                1>; // InSrcOutDstVectorSize
 
-    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
-    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
 
     const std::vector<ck::index_t> window_spatial_lengths{Y, X};
     const std::vector<ck::index_t> window_strides{window_stride_h, window_stride_w};
+    const std::vector<ck::index_t> window_dilations{window_dilation_h, window_dilation_w};
     const std::vector<ck::index_t> input_left_pads{in_left_pad_h, in_left_pad_w};
     const std::vector<ck::index_t> input_right_pads{in_right_pad_h, in_right_pad_w};
 
@@ -123,6 +127,7 @@ bool pool_test(bool 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});
@@ -144,8 +149,8 @@ bool pool_test(bool do_verification,
 
     float gb_per_sec = num_btype / 1.E6 / ave_time;
 
-    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
-              << std::endl;
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+              << " GB / s " << std::endl;
 
     bool pass = true;
 
@@ -169,6 +174,7 @@ bool pool_test(bool do_verification,
                                                              out_indices_n_c_ho_wo_host,
                                                              window_spatial_lengths,
                                                              window_strides,
+                                                             window_dilations,
                                                              input_left_pads,
                                                              input_right_pads);
 

example/13_pool2d_fwd/pool2d_fwd_fp16.cpp

@@ -42,6 +42,8 @@ int main(int argc, char* argv[])
     ck::index_t Wi                = 71;
     ck::index_t window_stride_h   = 2;
     ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
     ck::index_t in_left_pad_h     = 1;
     ck::index_t in_left_pad_w     = 1;
     ck::index_t in_right_pad_h    = 1;
@@ -59,7 +61,7 @@ int main(int argc, char* argv[])
         init_method     = std::stoi(argv[2]);
         time_kernel     = static_cast<bool>(std::stoi(argv[3]));
     }
-    else if(argc == 16)
+    else if(argc == 18)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
@@ -73,17 +75,19 @@ int main(int argc, char* argv[])
         Wi                = std::stoi(argv[9]);
         window_stride_h   = std::stoi(argv[10]);
         window_stride_w   = std::stoi(argv[11]);
-        in_left_pad_h   = std::stoi(argv[12]);
-        in_left_pad_w   = std::stoi(argv[13]);
-        in_right_pad_h  = std::stoi(argv[14]);
-        in_right_pad_w  = std::stoi(argv[15]);
+        window_dilation_h = std::stoi(argv[12]);
+        window_dilation_w = std::stoi(argv[13]);
+        in_left_pad_h     = std::stoi(argv[14]);
+        in_left_pad_w     = std::stoi(argv[15]);
+        in_right_pad_h    = std::stoi(argv[16]);
+        in_right_pad_w    = std::stoi(argv[17]);
     }
     else
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
         printf("arg3: time kernel (0=no, 1=yes)\n");
-        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
+        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
                "RightPx\n");
         exit(0);
     }
@@ -107,6 +111,8 @@ int main(int argc, char* argv[])
                                        Wi,
                                        window_stride_h,
                                        window_stride_w,
+                                       window_dilation_h,
+                                       window_dilation_w,
                                        in_left_pad_h,
                                        in_left_pad_w,
                                        in_right_pad_h,

example/13_pool2d_fwd/pool2d_fwd_fp32.cpp

@@ -42,6 +42,8 @@ int main(int argc, char* argv[])
     ck::index_t Wi                = 71;
     ck::index_t window_stride_h   = 2;
     ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
     ck::index_t in_left_pad_h     = 1;
     ck::index_t in_left_pad_w     = 1;
     ck::index_t in_right_pad_h    = 1;
@@ -59,7 +61,7 @@ int main(int argc, char* argv[])
         init_method     = std::stoi(argv[2]);
         time_kernel     = static_cast<bool>(std::stoi(argv[3]));
     }
-    else if(argc == 16)
+    else if(argc == 18)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
@@ -73,17 +75,19 @@ int main(int argc, char* argv[])
         Wi                = std::stoi(argv[9]);
         window_stride_h   = std::stoi(argv[10]);
         window_stride_w   = std::stoi(argv[11]);
-        in_left_pad_h   = std::stoi(argv[12]);
-        in_left_pad_w   = std::stoi(argv[13]);
-        in_right_pad_h  = std::stoi(argv[14]);
-        in_right_pad_w  = std::stoi(argv[15]);
+        window_dilation_h = std::stoi(argv[12]);
+        window_dilation_w = std::stoi(argv[13]);
+        in_left_pad_h     = std::stoi(argv[14]);
+        in_left_pad_w     = std::stoi(argv[15]);
+        in_right_pad_h    = std::stoi(argv[16]);
+        in_right_pad_w    = std::stoi(argv[17]);
     }
     else
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
         printf("arg3: time kernel (0=no, 1=yes)\n");
-        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
+        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
                "RightPx\n");
         exit(0);
     }
@@ -107,6 +111,8 @@ int main(int argc, char* argv[])
                                        Wi,
                                        window_stride_h,
                                        window_stride_w,
+                                       window_dilation_h,
+                                       window_dilation_w,
                                        in_left_pad_h,
                                        in_left_pad_w,
                                        in_right_pad_h,

example/14_gemm_quantization/CMakeLists.txt

+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
 # dlops
-add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
+if(DL_KERNELS)
+  add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
+endif()
 
 # xdlops
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
@@ -11,3 +14,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
+endif()
\ No newline at end of file

example/15_grouped_gemm/CMakeLists.txt

 add_custom_target(example_grouped_gemm_xdl)
 
-add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
-add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
-add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
-add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
-add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
-add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
-
-
-add_dependencies(example_grouped_gemm_xdl
-                 example_grouped_gemm_xdl_fp32
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
+  add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
+  add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
+  add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16 grouped_gemm_xdl_fixed_nk_fp16.cpp)
+  add_example_executable(example_grouped_gemm_xdl_fixed_nk_bias_fp16 grouped_gemm_xdl_fixed_nk_bias_fp16.cpp)
+  add_dependencies(example_grouped_gemm_xdl 
                    example_grouped_gemm_xdl_fp16
-                 example_grouped_gemm_xdl_bfp16
-                 example_grouped_gemm_xdl_int8
                    example_grouped_gemm_multiple_d_dl_fp16
-                 example_grouped_gemm_xdl_splitk_fp16)
-
+                   example_grouped_gemm_xdl_splitk_fp16
+                   example_grouped_gemm_xdl_fixed_nk_fp16
+                   example_grouped_gemm_xdl_fixed_nk_bias_fp16)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bfp16)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
+endif()
 if(USE_BITINT_EXTENSION_INT4)
   add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
   add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)