save progress

include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp

@@ -129,7 +129,11 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
         return make_tuple(c_thread_m, c_thread_n);
     }
 
-    __host__ __device__ BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle()
+    using Tuple5 = decltype(CalculateAThreadOriginDataIndex());
+    __host__ __device__ BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle(
+        Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
+        Tuple4 b_origin = CalculateBThreadOriginDataIndex())
+        : a_thread_copy_(a_origin), b_thread_copy_(b_origin)
     {
         static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
                           BK0NK1BlockDesc::IsKnownAtCompileTime(),
@@ -299,8 +303,8 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
                                                          B_K1,
                                                          B_K1>;
 
-    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
-    BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
 };
 
 // block wise level pipe designed for inline asm

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp

@@ -13,7 +13,7 @@
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_wmma_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_wmma_cshuffle.hpp"
 #include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"

include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp

@@ -414,7 +414,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
         constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
 
         auto blockwise_gemm =
-            BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<BlockSize,
+            BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle<BlockSize,
                                                          FloatA,
                                                          FloatB,
                                                          FloatAcc,

include/ck/utility/amd_wmma.hpp

@@ -23,11 +23,11 @@ struct intrin_wmma_f32_16x16x16_f16_w32<16, 16>
     {
         // * Inline assembly need to elimate the duplicated data load, compiler won't help you
         // delete them.
-        amd_assembly_wmma_f32_16x16x16_f16_w32(
-            reg_a, reg_b, reg_c.template AsType<float8_t>()(Number<0>{}));
-        // reg_c.template AsType<float8_t>()(Number<0>{}) =
-        // __builtin_amdgcn_wmma_f32_16x16x16_f16_w32( reg_a, reg_b, reg_c.template
-        // AsType<float8_t>()[Number<0>{}]);
+        // amd_assembly_wmma_f32_16x16x16_f16_w32(
+            // reg_a, reg_b, reg_c.template AsType<float8_t>()(Number<0>{}));
+        reg_c.template AsType<float8_t>()(Number<0>{}) =
+        __builtin_amdgcn_wmma_f32_16x16x16_f16_w32( reg_a, reg_b, reg_c.template
+        AsType<float8_t>()[Number<0>{}]);
     }
 };