Remove TempVgprCount to reduce the risk of regression

Author: linqun

include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma_cshuffle_v3.hpp

@@ -579,19 +579,17 @@ struct GridwiseGemm_wmma_cshuffle_v3
             NPerBlock * KPerBlock / NWave / WaveSize * sizeof(ComputeTypeB) / sizeof(uint32_t);
         // WMMA input is duplicated in GFX11
         constexpr index_t InputVgprCount = IsGfx11 ? BaseInputVgprCount * 2 : BaseInputVgprCount;
-        // VGPR used in buffer load and LDS store
-        constexpr index_t TempVgprCount = BaseInputVgprCount / 2;
         // VGPR used in Accumulator
         constexpr index_t AccVgprCount =
             MPerBlock * NPerBlock / BlockSize * sizeof(AccDataType) / sizeof(uint32_t);
 
         if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
         {
-            return InputVgprCount + TempVgprCount + AccVgprCount;
+            return InputVgprCount + AccVgprCount;
         }
         else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
         {
-            return InputVgprCount * 2 + TempVgprCount + AccVgprCount;
+            return InputVgprCount * 2 + AccVgprCount;
         }
         else
         {