update buffer addressing

composable_kernel/include/utility/config.amd.hpp.in

@@ -29,8 +29,8 @@
 #endif
 
 // only gfx908 support native floating point atomic add
-#ifndef CK_USE_AMD_BUFFER_ATOMIC_ADD
-#define CK_USE_AMD_BUFFER_ATOMIC_ADD 0
+#ifndef CK_USE_AMD_BUFFER_ATOMIC_FADD
+#define CK_USE_AMD_BUFFER_ATOMIC_FADD 0
 #endif
 
 // AMD XDLOPS
@@ -52,8 +52,8 @@
 #endif
 
 // experimental implementation
-#ifndef CK_EXPERIMENTAL_AMD_BUFFER_ADDRESSING_USE_OFFSET_TRICK
-#define CK_EXPERIMENTAL_AMD_BUFFER_ADDRESSING_USE_OFFSET_TRICK 1
+#ifndef CK_EXPERIMENTAL_USE_BUFFER_ADDRESS_OOB_CHECK
+#define CK_EXPERIMENTAL_USE_BUFFER_ADDRESS_OOB_CHECK 1
 #endif
 
 #ifndef CK_EXPERIMENTAL_BLOCKWISE_GEMM_USE_PIPELINE
@@ -73,10 +73,6 @@
 #ifndef CK_WORKAROUND_SWDEV_229564
 #define CK_WORKAROUND_SWDEV_229564 1
 #endif
-// workaround for buffer load/store fp16/bfp16 intrinsic bug
-#ifndef CK_WORKAROUND_SWDEV_231101
-#define CK_WORKAROUND_SWDEV_231101 1
-#endif
 
 namespace ck {
 

driver/include/device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp

@@ -187,7 +187,7 @@ void device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc,
     constexpr index_t GemmBBlockCopyDstDataPerWrite_GemmN = 4;
 
     constexpr index_t GemmCThreadCopyDstDataPerWrite_GemmN1 = 4;
-#elif 0
+#elif 1
     // cdata = 64, BlockSize = 256, 128x128x16
     constexpr index_t BlockSize = 256;
 

driver/src/conv_bwd_data_driver.cpp

@@ -51,10 +51,10 @@ int main(int argc, char* argv[])
 #elif 1
     // 3x3, 28x28
     constexpr index_t N  = 128;
-    constexpr index_t C  = 256;
+    constexpr index_t C  = 128;
     constexpr index_t HI = 28;
     constexpr index_t WI = 28;
-    constexpr index_t K  = 1024;
+    constexpr index_t K  = 128;
     constexpr index_t Y  = 3;
     constexpr index_t X  = 3;
 
@@ -245,7 +245,7 @@ int main(int argc, char* argv[])
     device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw
 #elif 0
     device_convolution_backward_data_implicit_gemm_v1r2_nchw_kcyx_nkhw
-#elif 0
+#elif 1
     device_convolution_backward_data_implicit_gemm_v4r1_nchw_kcyx_nkhw
 #elif 1
     device_convolution_backward_data_implicit_gemm_v5r1_nhwc_kyxc_nhwk

driver/src/conv_driver.cpp

@@ -201,7 +201,7 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 0>;
     using RightPads = Sequence<0, 0>;
-#elif 1
+#elif 0
     // 3x3, 35x35, stride 2
     constexpr index_t N  = 128;
     constexpr index_t C  = 288;
@@ -339,7 +339,7 @@ int main(int argc, char* argv[])
 #elif 1
     // 3x3, 28x28
     constexpr index_t N  = 128;
-    constexpr index_t C  = 192;
+    constexpr index_t C  = 128;
     constexpr index_t HI = 28;
     constexpr index_t WI = 28;
     constexpr index_t K  = 128;
@@ -561,7 +561,7 @@ int main(int argc, char* argv[])
                                                          LeftPads{},
                                                          RightPads{},
                                                          nrepeat);
-#elif 0
+#elif 1
     device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(in_nchw_desc,
                                                          in_nchw,
                                                          wei_kcyx_desc,