Add a denorm test fix (#603)

* Add type_convert implementations for bf16

* Add the fix for conv_fwd

* Add the fix for conv_bwd_data

* Add the fix for conv_bwd_weight

* Format

* Format

* Another format

* Add a macro to use workaround on MI200 only

* Format

---------
Co-authored-by: Rosty Geyyer <rosty.geyyer@amd.com>
Co-authored-by: zjing14 <zhangjing14@gmail.com>

example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 

include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

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

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

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

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -92,6 +92,17 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
     using GridwiseGemmPipe = remove_cvref_t<decltype(
         GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
+    // denorm test fix, required to work around fp16 mfma issue
+    // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
+    // when mfma if fixed, remove this section and update
+    // ABDataTypeAdjusted -> ABDataType throughout this file
+#if defined(__gfx90a__)
+    using ABDataTypeAdjusted =
+        conditional_t<is_same_v<ABDataType, ck::half_t>, ck::bhalf_t, ABDataType>;
+#else
+    using ABDataTypeAdjusted = ABDataType;
+#endif
+
     __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
     {
         // A matrix in LDS memory, dst of blockwise copy
@@ -397,7 +408,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
                                                 ABlockTransferThreadClusterLengths_AK0_M_AK1,
                                                 ABlockTransferThreadClusterArrangeOrder,
                                                 ABDataType,
-                                                ABDataType,
+                                                ABDataTypeAdjusted,
                                                 decltype(a_grid_desc_ak0_m_ak1),
                                                 decltype(a_block_desc_ak0_m_ak1),
                                                 ABlockTransferSrcAccessOrder,
@@ -428,7 +439,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
                                                 BBlockTransferThreadClusterLengths_BK0_N_BK1,
                                                 BBlockTransferThreadClusterArrangeOrder,
                                                 ABDataType,
-                                                ABDataType,
+                                                ABDataTypeAdjusted,
                                                 decltype(b_grid_desc_bk0_n_bk1),
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
@@ -458,11 +469,11 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
         // sanity check
         constexpr index_t KPack =
             math::max(math::lcm(AK1, BK1),
-                      MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+                      MfmaSelector<ABDataTypeAdjusted, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            ABDataType,
+            ABDataTypeAdjusted,
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -480,10 +491,11 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
             a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+            static_cast<ABDataTypeAdjusted*>(p_shared),
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
+            static_cast<ABDataTypeAdjusted*>(p_shared) + a_block_space_size_aligned,
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

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

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -166,15 +166,12 @@ __global__ void
                                       const CBlockClusterAdaptor c_block_cluster_adaptor)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
-    constexpr index_t shared_block_size =
-        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
-
-    __shared__ FloatAB p_shared_block[shared_block_size];
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
                                                   p_b_grid,
                                                   p_c_grid,
-                                                  p_shared_block,
+                                                  p_shared,
                                                   a_b_k0_m_k1_grid_desc,
                                                   b_b_k0_n_k1_grid_desc,
                                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
@@ -264,6 +261,16 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
     using GridwiseGemmPipe = remove_cvref_t<decltype(
         GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
+    // denorm test fix, required to work around fp16 mfma issue
+    // 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 defined(__gfx90a__)
+    using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
+#else
+    using FloatABAdjusted = FloatAB;
+#endif
+
     // M0/M1/M1Padding
     static constexpr auto M1PerBlock = Number<ABlockLdsM1PerBlock>{};
     static constexpr auto M0PerBlock = Number<ABlockLdsM0PerBlock>{};
@@ -605,7 +612,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
     __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
                                const FloatAB* __restrict__ p_b_grid,
                                FloatC* __restrict__ p_c_grid,
-                               FloatAB* __restrict__ p_shared_block,
+                               void* __restrict__ p_shared,
                                const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
                                const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
                                const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
@@ -666,7 +673,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
                                                 ABlockTransferThreadClusterLengths_K0_M_K1,
                                                 ABlockTransferThreadClusterArrangeOrder,
                                                 FloatAB,
-                                                FloatAB,
+                                                FloatABAdjusted,
                                                 decltype(a_b_k0_m_k1_grid_desc),
                                                 decltype(a_b_k0_m_k1_block_desc),
                                                 ABlockTransferSrcAccessOrder,
@@ -696,7 +703,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
                                                 BBlockTransferThreadClusterLengths_K0_N_K1,
                                                 BBlockTransferThreadClusterArrangeOrder,
                                                 FloatAB,
-                                                FloatAB,
+                                                FloatABAdjusted,
                                                 decltype(b_b_k0_n_k1_grid_desc),
                                                 decltype(b_b_k0_n_k1_block_desc),
                                                 BBlockTransferSrcAccessOrder,
@@ -725,11 +732,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
         // sanity check
 
         constexpr index_t KPack =
-            math::max(K1, MfmaSelector<FloatAB, MPerXDL, NPerXDL>::selected_mfma.k_per_blk);
+            math::max(K1, MfmaSelector<FloatABAdjusted, MPerXDL, NPerXDL>::selected_mfma.k_per_blk);
 
         auto blockwise_gemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
-                                                                FloatAB,
+                                                                FloatABAdjusted,
                                                                 FloatAcc,
                                                                 decltype(a_k0_m_k1_block_desc),
                                                                 decltype(b_k0_n_k1_block_desc),
@@ -745,16 +752,15 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
         constexpr auto a_block_space_size =
             math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
 
-        FloatAB* p_a_block = p_shared_block;
-        FloatAB* p_b_block = p_shared_block + a_block_space_size;
-
         constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
         constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
+            static_cast<FloatABAdjusted*>(p_shared), a_k0_m_k1_block_desc.GetElementSpaceSize());
+
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
+            static_cast<FloatABAdjusted*>(p_shared) + a_block_space_size,
+            b_k0_n_k1_block_desc.GetElementSpaceSize());
 
         // gridwise GEMM pipeline
         const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
@@ -798,8 +804,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
             constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
                 GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
 
-            void* p_shared = static_cast<void*>(p_shared_block);
-
             auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
                 static_cast<FloatC*>(p_shared),
                 c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());

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

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -131,6 +131,16 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
     using GridwiseGemmPipe = remove_cvref_t<decltype(
         GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
+    // denorm test fix, required to work around fp16 mfma issue
+    // 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 defined(__gfx90a__)
+    using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
+#else
+    using FloatABAdjusted = FloatAB;
+#endif
+
     __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
     {
         constexpr auto max_lds_align = K1;
@@ -281,7 +291,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
 
         using BlockwiseGemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
-                                                                FloatAB,
+                                                                FloatABAdjusted,
                                                                 FloatAcc,
                                                                 decltype(a_block_desc_k0_m_k1),
                                                                 decltype(b_block_desc_k0_n_k1),
@@ -367,7 +377,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
                                                 ABlockTransferThreadClusterLengths_K0_M_K1,
                                                 ABlockTransferThreadClusterArrangeOrder,
                                                 FloatAB,
-                                                FloatAB,
+                                                FloatABAdjusted,
                                                 decltype(a_grid_desc_k0_m_k1),
                                                 decltype(a_block_desc_k0_m_k1),
                                                 ABlockTransferSrcAccessOrder,
@@ -398,7 +408,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
                                                 BBlockTransferThreadClusterLengths_K0_N_K1,
                                                 BBlockTransferThreadClusterArrangeOrder,
                                                 FloatAB,
-                                                FloatAB,
+                                                FloatABAdjusted,
                                                 decltype(b_grid_desc_k0_n_k1),
                                                 decltype(b_block_desc_k0_n_k1),
                                                 BBlockTransferSrcAccessOrder,
@@ -428,7 +438,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
         // sanity check
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            FloatAB,
+            FloatABAdjusted,
             FloatAcc,
             decltype(a_block_desc_k0_m_k1),
             decltype(b_block_desc_k0_n_k1),
@@ -446,10 +456,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
             math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+            static_cast<FloatABAdjusted*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            static_cast<FloatABAdjusted*>(p_shared) + a_block_space_size_aligned,
             b_block_desc_k0_n_k1.GetElementSpaceSize());
 
         constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck/utility/data_type.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -1008,6 +1008,60 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float
     return uint16_t(u.int32 >> 16);
 }
 
+// convert bfp16 to fp16 via fp32
+template <>
+inline __host__ __device__ constexpr half_t type_convert<half_t, bhalf_t>(bhalf_t x)
+{
+    float x_fp32 = type_convert<float>(x);
+
+    return static_cast<half_t>(x_fp32);
+}
+
+// convert fp16 to bfp16 via fp32
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, half_t>(half_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return type_convert<bhalf_t>(x_fp32);
+}
+
+// convert bfp16 to int32 via fp32
+template <>
+inline __host__ __device__ constexpr int32_t type_convert<int32_t, bhalf_t>(bhalf_t x)
+{
+    float x_fp32 = type_convert<float>(x);
+
+    return static_cast<int32_t>(x_fp32);
+}
+
+// convert int32 to bfp16 via fp32
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int32_t>(int32_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return type_convert<bhalf_t>(x_fp32);
+}
+
+// convert bfp16 to int8 via fp32
+template <>
+inline __host__ __device__ constexpr int8_t type_convert<int8_t, bhalf_t>(bhalf_t x)
+{
+    float x_fp32 = type_convert<float>(x);
+
+    return static_cast<int8_t>(x_fp32);
+}
+
+// convert int8 to bfp16 via fp32
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return type_convert<bhalf_t>(x_fp32);
+}
+
 template <typename T>
 struct NumericLimits
 {