diff --git a/driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded.hpp b/driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded.hpp
index 965f84b612..e550581fc4 100644
--- a/driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded.hpp
+++ b/driver/include/device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded.hpp
@@ -51,8 +51,8 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded(InDesc,
     wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
     out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
 
-#if 1
-    // each thread hold 64 data
+#if 0
+    // BlockSize = 256, each thread hold 64 data
     constexpr index_t BlockSize = 256;
 
     constexpr index_t BPerBlock = 16;
@@ -86,6 +86,43 @@ void device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_padded(InDesc,
     using WeiBlockCopySrcAccessOrder            = Sequence<1, 0>; // [K, E]
     using WeiBlockCopyDstAccessOrder            = Sequence<0, 1>; // [E, K]
 
+    constexpr index_t WeiBlockCopySrcDataPerRead_E  = 4;
+    constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
+#elif 1
+    // BlockSize = 64, each thread hold 64 data
+    constexpr index_t BlockSize = 64;
+
+    constexpr index_t BPerBlock = 8;
+    constexpr index_t KPerBlock = 64;
+    constexpr index_t EPerBlock = 8;
+
+    constexpr index_t GemmNRepeat = 2;
+
+    constexpr index_t GemmMPerThreadSubC = 4;
+    constexpr index_t GemmNPerThreadSubC = 4;
+    constexpr index_t GemmMLevel0Cluster = 4;
+    constexpr index_t GemmNLevel0Cluster = 4;
+    constexpr index_t GemmMLevel1Cluster = 2;
+    constexpr index_t GemmNLevel1Cluster = 2;
+    constexpr index_t GemmKPerThreadLoop = 1;
+    constexpr index_t GemmDataPerReadA   = 4;
+    constexpr index_t GemmDataPerReadB   = 4;
+
+    using InBlockCopySubLengths_E_N1_B_N2      = Sequence<1, 2, 1, 4>;
+    using InBlockCopyClusterLengths_E_N1_B_N2  = Sequence<8, 1, 8, 1>;
+    using InBlockCopyThreadClusterArrangeOrder = Sequence<0, 1, 3, 2>; // [E, N1, N2, B]
+    using InBlockCopySrcAccessOrder            = Sequence<0, 2, 1, 3>; // [E, B, N1, N2]
+    using InBlockCopyDstAccessOrder            = Sequence<0, 1, 2, 3>; // [E, N1, B, N2]
+
+    constexpr index_t InBlockCopySrcDataPerRead_B   = 1;
+    constexpr index_t InBlockCopyDstDataPerWrite_N2 = 4;
+
+    using WeiBlockCopySubLengths_E_K            = Sequence<4, 2>;
+    using WeiBlockCopyClusterLengths_E_K        = Sequence<2, 32>;
+    using WeiBlockCopyThreadClusterArrangeOrder = Sequence<1, 0>; // [K, E]
+    using WeiBlockCopySrcAccessOrder            = Sequence<1, 0>; // [K, E]
+    using WeiBlockCopyDstAccessOrder            = Sequence<0, 1>; // [E, K]
+
     constexpr index_t WeiBlockCopySrcDataPerRead_E  = 4;
     constexpr index_t WeiBlockCopyDstDataPerWrite_K = 1;
 #endif
diff --git a/driver/src/driver.cpp b/driver/src/driver.cpp
index 8ae1b3eb35..94c59c478c 100644
--- a/driver/src/driver.cpp
+++ b/driver/src/driver.cpp
@@ -311,6 +311,51 @@ int main(int argc, char* argv[])
 
     using LeftPads  = Sequence<0, 0>;
     using RightPads = Sequence<0, 0>;
+#elif 0
+    // 5x5 filter, 2x2 pad, 7x7 input
+    constexpr index_t N  = 128;
+    constexpr index_t C  = 48;
+    constexpr index_t HI = 7;
+    constexpr index_t WI = 7;
+    constexpr index_t K  = 128;
+    constexpr index_t Y  = 5;
+    constexpr index_t X  = 5;
+
+    using ConvStrides   = Sequence<1, 1>;
+    using ConvDilations = Sequence<1, 1>;
+
+    using LeftPads  = Sequence<2, 2>;
+    using RightPads = Sequence<2, 2>;
+#elif 0
+    // 7x1 filter, 3x0 pad, 17x17 input
+    constexpr index_t N  = 128;
+    constexpr index_t C  = 128;
+    constexpr index_t HI = 17;
+    constexpr index_t WI = 17;
+    constexpr index_t K  = 128;
+    constexpr index_t Y  = 7;
+    constexpr index_t X  = 1;
+
+    using ConvStrides   = Sequence<1, 1>;
+    using ConvDilations = Sequence<1, 1>;
+
+    using LeftPads  = Sequence<3, 0>;
+    using RightPads = Sequence<3, 0>;
+#elif 1
+    // 1x7 filter, 0x3 pad, 17x17 input
+    constexpr index_t N  = 128;
+    constexpr index_t C  = 128;
+    constexpr index_t HI = 17;
+    constexpr index_t WI = 17;
+    constexpr index_t K  = 128;
+    constexpr index_t Y  = 1;
+    constexpr index_t X  = 7;
+
+    using ConvStrides   = Sequence<1, 1>;
+    using ConvDilations = Sequence<1, 1>;
+
+    using LeftPads  = Sequence<0, 3>;
+    using RightPads = Sequence<0, 3>;
 #endif
 
     auto in_nchw_desc  = make_ConstantTensorDescriptor_packed(Sequence<N, C, HI, WI>{});
diff --git a/script/compile-hip.sh b/script/compile-hip.sh
index 51d1bed797..5a61bc1387 100755
--- a/script/compile-hip.sh
+++ b/script/compile-hip.sh
@@ -1,8 +1,8 @@
 #!/bin/bash
  export KMDUMPISA=1
  export KMDUMPLLVM=1
- export KMOPTLLC="-mattr=+enable-ds128"
-#export KMOPTLLC="-mattr=+enable-ds128 -amdgpu-enable-global-sgpr-addr"
+#export KMOPTLLC="-mattr=+enable-ds128"
+ export KMOPTLLC="-mattr=+enable-ds128 -amdgpu-enable-global-sgpr-addr"
 
 make -j driver
 /opt/rocm/hcc/bin/llvm-objdump -mcpu=gfx906 -source -line-numbers driver/dump-gfx906.isabin > driver/dump-gfx906.isabin.asm