Change flag to CK_GFX90A_DENORM_WORKAROUND (#1817)

* Change flag from CK_WORKAROUND_DENORM_FIX to CK_GFX90A_DENORM_WORKAROUND for more clarity. Also changed the definition macros to be more clear.

include/ck/ck.hpp

@@ -235,13 +235,18 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
 // workaround: compiler issue on gfx908
 #define CK_WORKAROUND_SWDEV_388832 1
 
-// denorm test fix, required to work around dissue
-#ifndef CK_WORKAROUND_DENORM_FIX
-#define CK_WORKAROUND_DENORM_FIX 0
+// denorm test fix, necessary for gfx90a
+#ifndef CK_GFX90A_DENORM_WORKAROUND
+#define CK_GFX90A_DENORM_WORKAROUND 0
+#endif // CK_GFX90A_DENORM_WORKAROUND
+// Enable only for gfx90a
+#if defined(__gfx90a__)
+#if CK_GFX90A_DENORM_WORKAROUND
+#define CK_GFX90A_DENORM_WORKAROUND 1
+#endif // CK_GFX90A_DENORM_WORKAROUND is set to 1
 #else
-// enable only for gfx90a
-#define CK_WORKAROUND_DENORM_FIX = CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
-#endif // CK_WORKAROUND_DENORM_FIX
+#define CK_GFX90A_DENORM_WORKAROUND 0
+#endif // gfx90a
 
 // set flag to 1 to build deprecated instances
 #define CK_BUILD_DEPRECATED 1

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

@@ -101,7 +101,7 @@ struct GridwiseGemmMultipleABD_xdl_cshuffle
     using GridwiseGemmPipe = remove_cvref_t<
         decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
-#if CK_WORKAROUND_DENORM_FIX
+#if CK_GFX90A_DENORM_WORKAROUND
     using AComputeDataType =
         conditional_t<is_same_v<AComputeDataType_, ck::half_t>, ck::bhalf_t, AComputeDataType_>;
     using BComputeDataType =

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

@@ -100,7 +100,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
     using GridwiseGemmPipe = remove_cvref_t<
         decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
-#if CK_WORKAROUND_DENORM_FIX
+#if CK_GFX90A_DENORM_WORKAROUND
     using AComputeDataType =
         conditional_t<is_same_v<AComputeDataType_, ck::half_t>, ck::bhalf_t, AComputeDataType_>;
     using BComputeDataType =

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

@@ -164,7 +164,7 @@ struct GridwiseGemmMultipleD_Xdl_CShuffle_LdsDirectLoad
     using GridwiseGemmPipe = remove_cvref_t<
         decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
-#if CK_WORKAROUND_DENORM_FIX
+#if CK_GFX90A_DENORM_WORKAROUND
     using AComputeDataType =
         conditional_t<is_same_v<AComputeDataType_, ck::half_t>, ck::bhalf_t, AComputeDataType_>;
 #else

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

@@ -271,7 +271,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
     // when mfma if fixed, remove this section and update
     // FloatAAdjusted -> ComputeTypeA, FloatBAdjusted -> ComputeTypeB,
     // throughout this file
-#if CK_WORKAROUND_DENORM_FIX
+#if CK_GFX90A_DENORM_WORKAROUND
     using FloatAAdjusted =
         conditional_t<is_same_v<ComputeTypeA, ck::half_t>, ck::bhalf_t, ComputeTypeA>;
     using FloatBAdjusted =

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp

@@ -254,7 +254,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
     // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
     // when mfma if fixed, remove this section and update
     // FloatABAdjusted -> FloatAB throughout this file
-#if CK_WORKAROUND_DENORM_FIX
+#if CK_GFX90A_DENORM_WORKAROUND
     using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
 #else
     using FloatABAdjusted = FloatAB;