Merge remote-tracking branch 'origin/develop' into aosewski/hotloop

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

@@ -74,8 +74,7 @@ __global__ void
                                           const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
                                           const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

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

@@ -60,8 +60,7 @@ __global__ void
             const index_t batch_count,
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -108,7 +107,7 @@ __global__ void
     ignore = block_2_ctile_map;
     ignore = batch_count;
     ignore = compute_base_ptr_of_batch;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 // Computes C = A * B0 * B1

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

@@ -83,8 +83,7 @@ __global__ void
                                 const Block2ETileMap block_2_etile_map)
 {
 
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

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

@@ -68,8 +68,7 @@ __global__ void
             const index_t batch_count,
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);

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

@@ -59,8 +59,7 @@ __global__ void
             const ComputeBasePrtOfBatch compute_base_ptr_of_batch_,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

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

@@ -67,8 +67,7 @@ __global__ void
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
             const C0MatrixMask c0_matrix_mask)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -127,7 +126,7 @@ __global__ void
     ignore = batch_count;
     ignore = compute_base_ptr_of_batch;
     ignore = c0_matrix_mask;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 // Computes C = A * B0 * B1

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

@@ -62,8 +62,7 @@ __global__ void
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
             const C0MatrixMask c0_matrix_mask)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -112,7 +111,7 @@ __global__ void
     ignore = batch_count;
     ignore = compute_base_ptr_of_batch;
     ignore = c0_matrix_mask;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 // Computes C = A * B0 * B1

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

@@ -52,8 +52,7 @@ __global__ void
 #endif
         kernel_batched_gemm_xdlops_v2r3(const typename DeviceOp::Argument karg)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / karg.Batch);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

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

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#include "ck/library/utility/numeric.hpp"
 #include "ck/tensor_description/tensor_descriptor.hpp"
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
 #include "ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp"

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

@@ -55,8 +55,7 @@ __global__ void
                 e_grid_desc_mblock_mperblock_nblock_nperblock,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_as_grid,

include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp

@@ -55,8 +55,7 @@ __global__ void
             const CElementwiseOperation c_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / num_batches);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
@@ -97,7 +96,7 @@ __global__ void
     ignore = b_element_op;
     ignore = c_element_op;
     ignore = block_2_ctile_map;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 // specialization for #D conv: in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]

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

@@ -50,9 +50,8 @@ __global__ void
             const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||         \
-    defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__) || \
-    defined(__gfx12__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx9__) || \
+    defined(__gfx103__) || defined(__gfx11__) || defined(__gfx12__))
 
     constexpr index_t shared_block_size =
         GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);

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

@@ -63,8 +63,7 @@ __global__ void
             const Block2ETileMap block_2_etile_map,
             index_t NRaw)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemmWelford::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemmWelford::template Run<HasMainKBlockLoop>(

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

@@ -60,8 +60,7 @@ __global__ void
             const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,

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

@@ -52,8 +52,7 @@ __global__ void
                                                 e_grid_desc_mblock_mperblock_nblock_nperblock,
                                             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,

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

@@ -205,7 +205,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
 
             const auto b2c_map = DefaultBlock2CTileMap{};
             index_t gdx, gdy, gdz;
-            std::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
+            ck::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
             const auto K0Padded    = karg.K0Padded;
 
             const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0Padded);

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

@@ -183,7 +183,7 @@ struct DeviceGemmXdlSplitKCShuffle_LdsDirectLoad : public DeviceGemmSplitK<ALayo
 
             const auto b2c_map = DefaultBlock2CTileMap{};
             index_t gdx, gdy, gdz;
-            std::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
+            ck::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
             const auto K0Padded    = karg.K0Padded;
 
             const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0Padded);

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

@@ -47,8 +47,7 @@ __global__ void
                 e_grid_desc_mblock_mperblock_nblock_nperblock,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,

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

@@ -37,8 +37,7 @@ __global__ void
             const BElementwiseOperation b_element_op,
             const CDEElementwiseOperation cde_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp

@@ -87,8 +87,7 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const ComputePtrOffsetOfN compute_ptr_offset_of_n)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     // offset base pointer for each work-group
     const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.z);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y);