Merge branch 'amd-develop' into amd-master

example/44_conv2d_fwd_quantization/CMakeLists.txt

-add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
-add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
-add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)

include/ck/ck.hpp

@@ -31,20 +31,20 @@
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_BUFFER_RESOURCE_3RD_DWORD -1
 #elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
-    defined(__gfx90a__) // for GPU code
+    defined(__gfx90a__) || defined(__gfx940__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
 #elif defined(__gfx1030__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
 #elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
-#define CK_BUFFER_RESOURCE_3RD_DWORD 0x10020000
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
 #endif
 
 // FMA instruction
 #ifndef __HIP_DEVICE_COMPILE__                   // for host code, define nothing
 #elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
 #define CK_USE_AMD_V_MAC_F32
-#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx1030__) // for GPU code
+#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) || \
+    defined(__gfx940__) // for GPU code
 #define CK_USE_AMD_V_FMAC_F32
 #define CK_USE_AMD_V_DOT2_F32_F16
 #define CK_USE_AMD_V_DOT4_I32_I8
@@ -53,14 +53,18 @@
 // MFMA instruction
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_USE_AMD_MFMA
-#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code
+#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
 #define CK_USE_AMD_MFMA
 #endif
 
-#if defined(__gfx90a__)
+#if(defined(__gfx90a__) || defined(__gfx940__))
 #define CK_USE_AMD_MFMA_BF16_1K_OP
 #endif
 
+#if defined(__gfx940__)
+#define CK_USE_AMD_MFMA_GFX940
+#endif
+
 // WMMA instruction
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_USE_AMD_WMMA
@@ -80,13 +84,13 @@
 // buffer atomic add: floating point
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
-#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code
+#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
 #else // for GPU code
 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
 #endif
 
-#if defined(__gfx90a__) // for GPU code
+#if(defined(__gfx90a__) || defined(__gfx940__)) // for GPU code
 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
 #else
 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0
@@ -163,9 +167,16 @@
 // workaround: compiler not emiting reciprocal instruction frm __frcp_rn()
 #define CK_WORKAROUND_SWDEV_383542 1
 
+// workaround: compiler issue on gfx908
+#define CK_WORKAROUND_SWDEV_388832 1
 // flag to enable (1) or disable (0) the debugging output in some kernels
 #define DEBUG_LOG 0
 
+// denorm test fix, required to work around dissue
+#ifndef CK_WORKAROUND_DENORM_FIX
+#define CK_WORKAROUND_DENORM_FIX 0
+#endif
+
 namespace ck {
 
 enum struct InMemoryDataOperationEnum

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

@@ -47,7 +47,8 @@ __global__ void
                 e_grid_desc_mblock_mperblock_nblock_nperblock,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -416,7 +417,8 @@ struct DeviceGemm_Xdl_WaveletModel_CShuffle : public DeviceGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

@@ -134,7 +134,8 @@ __global__ void
             const Block2CTileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \
+    defined(__gfx90a__) || defined(__gfx908__))
     // offset base pointer for each work-group
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -314,9 +315,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
         const auto in_gemmm_gemmk_desc =
             matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
 
-        const auto M = in_gemmm_gemmk_desc.GetLength(I0);
-        const auto K = in_gemmm_gemmk_desc.GetLength(I1);
-
+        const auto M   = in_gemmm_gemmk_desc.GetLength(I0);
+        const auto K   = in_gemmm_gemmk_desc.GetLength(I1);
         const auto AK0 = K / K1;
 
         return transform_tensor_descriptor(
@@ -709,7 +709,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
         namespace ctc = tensor_layout::convolution;
 
         // check device
-        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
+             ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx908"))
         {
             return false;
         }
@@ -834,6 +835,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
         {
             return false;
         }
+
         // check Gridwise GEMM
         return GridwiseGemm::CheckValidity(
             arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.e_grid_desc_m_n_);

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/device_grouped_gemm.hpp

@@ -31,7 +31,7 @@ struct DeviceGroupedGemm : public BaseOperator
 {
     static constexpr index_t NumDTensor = DsDataType::Size();
 
-    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsisiten NumDTensor");
+    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsistent NumDTensor");
 
     virtual std::unique_ptr<BaseArgument>
     MakeArgumentPointer(std::vector<const void*>& p_a,

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

@@ -43,7 +43,8 @@ __global__ void
             const B1ElementwiseOperation b1_element_op,
             const CElementwiseOperation c_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();
@@ -678,7 +679,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

+#pragma once
+#include <iostream>
+#include <vector>
+
+#include "device_grouped_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemmSplitK : public DeviceGroupedGemm<ALayout,
+                                                          BLayout,
+                                                          DsLayout,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          DsDataType,
+                                                          EDataType,
+                                                          AElementwiseOperation,
+                                                          BElementwiseOperation,
+                                                          CElementwiseOperation>
+{
+    virtual void SetKBatchSize(BaseArgument* p_arg, index_t kbatch) const = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -56,7 +56,8 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t num_blocks_per_batch =
@@ -938,7 +939,8 @@ struct DeviceSplitKContractionMultipleD_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

@@ -56,7 +56,8 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t num_blocks_per_batch =
@@ -839,7 +840,8 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

@@ -74,7 +74,8 @@ __global__ void
                                           const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
                                           const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     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

@@ -60,7 +60,8 @@ __global__ void
             const index_t batch_count,
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -588,7 +589,8 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

@@ -83,7 +83,8 @@ __global__ void
                                 const Block2ETileMap block_2_etile_map)
 {
 
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     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);
@@ -579,7 +580,8 @@ struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

@@ -68,7 +68,8 @@ __global__ void
             const index_t batch_count,
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -804,7 +805,8 @@ struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

@@ -59,7 +59,8 @@ __global__ void
             const ComputeBasePrtOfBatch compute_base_ptr_of_batch_,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     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

@@ -67,7 +67,8 @@ __global__ void
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
             const C0MatrixMask c0_matrix_mask)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -714,7 +715,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
         arg.Print();
 #endif
 
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

@@ -62,7 +62,8 @@ __global__ void
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
             const C0MatrixMask c0_matrix_mask)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -612,7 +613,8 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

@@ -75,7 +75,8 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     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_contraction_multiple_d_xdl_cshuffle.hpp

@@ -52,7 +52,8 @@ __global__ void
                 e_grid_desc_mblock_mperblock_nblock_nperblock,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -581,7 +582,8 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

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

@@ -55,7 +55,8 @@ __global__ void
             const CElementwiseOperation c_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     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);