Merge branch 'develop' into amd-develop

CHANGELOG.md

@@ -11,7 +11,7 @@ None
 None
 
 ### Additions
-* Introduced wrapper sublibrary (limited functionality). (#1071, #1098, #1108, #1126)
+* Introduced wrapper sublibrary (limited functionality). (#1071, #1098, #1108, #1126, #1139)
 
 ### Changes
 None

Jenkinsfile

@@ -135,6 +135,7 @@ def buildDocker(install_prefix){
             echo "Building image: ${image_name}"
             retimage = docker.build("${image_name}", dockerArgs + ' .')
             retimage.push()
+            sh 'docker images -q -f dangling=true | xargs --no-run-if-empty docker rmi'
         }
         else{
             echo "Checking for image: ${image_name}"

docs/sphinx/requirements.in

-rocm-docs-core==0.33.0
+rocm-docs-core==0.34.0
 sphinxcontrib-bibtex==2.6.2

docs/sphinx/requirements.txt

@@ -113,7 +113,7 @@ requests==2.31.0
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core==0.33.0
+rocm-docs-core==0.34.0
     # via -r requirements.in
 six==1.16.0
     # via

docs/wrapper.rst

@@ -89,3 +89,4 @@ Operations
 -------------------------------------
 
 .. doxygenfile:: copy.hpp
+.. doxygenfile:: gemm.hpp

example/35_splitK_gemm/CMakeLists.txt

@@ -10,6 +10,9 @@ foreach(gpu IN LISTS GPU_TARGETS)
    add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
    add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
 
+   add_example_executable(example_splitK_gemm_xdl_lds_direct_load_fp16 splitK_gemm_xdl_lds_direct_load_fp16.cpp)
+   add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_lds_direct_load_fp16)
+
    add_example_executable(example_splitK_gemm_xdl_bf16 splitK_gemm_xdl_bf16.cpp)
    add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bf16)
 

example/35_splitK_gemm/splitK_gemm_xdl_lds_direct_load_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+#define DIRECT_LOAD 1
+
+#if DIRECT_LOAD
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp"
+#else
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#endif
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using CDataType   = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+#if DIRECT_LOAD
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle_LdsDirectLoad
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|         ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|         BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|          ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar| AddExtraM|          ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |     |     | Wave| Wave| Lengths_KBatch_K0_M_K1|               |               |      PerVector|          | Lengths_KBatch_K0_N_K1|               |              |      PerVector|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |         |      |      |      |      |    |     |     |     |     |                       |               |               |               |          |                       |               |              |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        2,   128,    32,    16,     4,  16,   16,   16,    1,    1,         S<1, 2, 8, 8>,  S<0, 2, 1, 3>,              3,              2,      true,         S<1, 2, 8, 8>,  S<0, 2, 1, 3>,             3,              2,      true,           1,           1,                   S<1, 32, 1, 4>,               4>;
+// clang-format on
+
+#else
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+using DeviceGemmInstance          = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#endif
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }

include/ck/ck.hpp

@@ -44,16 +44,30 @@
 #define CK_USE_WAVES_PER_EU 0
 #endif
 
+// define general macros for various architectures
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+#define __gfx94__
+#endif
+#if defined(__gfx1010__) || defined(__gfx1011__) || defined(__gfx1012__)
+#define __gfx101__
+#endif
+#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \
+    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__)
+#define __gfx103__
+#endif
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__)
+#define __gfx11__
+#endif
+
 // buffer resource
 #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__) || defined(__gfx940__) || defined(__gfx941__) ||                          \
-    defined(__gfx942__) // for GPU code
+    defined(__gfx90a__) || defined(__gfx94__)
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
-#elif defined(__gfx1030__) // for GPU code
+#elif defined(__gfx103__)
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
-#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
+#elif defined(__gfx11__)
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
 #endif
 
@@ -61,12 +75,12 @@
 #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__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) // for GPU code
+#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx103__) || \
+    defined(__gfx94__) // 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
-#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+#elif defined(__gfx11__)
 #define CK_USE_AMD_V_FMAC_F32
 #define CK_USE_AMD_V_DOT2_F32_F16
 #define CK_USE_AMD_V_DOT4_I32_I8_GFX11
@@ -75,23 +89,22 @@
 // MFMA instruction
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_USE_AMD_MFMA
-#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
-    defined(__gfx942__) // for GPU code
+#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) // for GPU code
 #define CK_USE_AMD_MFMA
 #endif
 
-#if(defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+#if(defined(__gfx90a__) || defined(__gfx94__))
 #define CK_USE_AMD_MFMA_BF16_1K_OP
 #endif
 
-#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+#if defined(__gfx94__)
 #define CK_USE_AMD_MFMA_GFX940
 #endif
 
 // WMMA instruction
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_USE_AMD_WMMA
-#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
+#elif defined(__gfx11__) // for GPU code
 #define CK_USE_AMD_WMMA
 #endif
 
@@ -107,15 +120,13 @@
 // 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__) || defined(__gfx940__) || defined(__gfx941__) || \
-    defined(__gfx942__) // for GPU code
+#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) // 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__) || defined(__gfx940__) || defined(__gfx941__) || \
-    defined(__gfx942__)) // for GPU code
+#if(defined(__gfx90a__) || defined(__gfx94__)) // for GPU code
 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
 #else
 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0

include/ck/host_utility/device_prop.hpp

@@ -65,4 +65,23 @@ inline bool is_lds_direct_load_supported()
            ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942";
 }
 
+inline bool is_navi1_supported()
+{
+    return ck::get_device_name() == "gfx1010" || ck::get_device_name() == "gfx1011" ||
+           ck::get_device_name() == "gfx1012";
+}
+
+inline bool is_navi2_supported()
+{
+    return ck::get_device_name() == "gfx1030" || ck::get_device_name() == "gfx1031" ||
+           ck::get_device_name() == "gfx1032" || ck::get_device_name() == "gfx1034" ||
+           ck::get_device_name() == "gfx1035" || ck::get_device_name() == "gfx1036";
+}
+
+inline bool is_navi3_supported()
+{
+    return ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102" || ck::get_device_name() == "gfx1103";
+}
+
 } // namespace ck

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

@@ -770,8 +770,7 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
-           ck::get_device_name() == "gfx1102")
+        if(ck::is_navi3_supported())
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {

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

@@ -57,7 +57,7 @@ __global__ void
             const Block2ETileMap block_2_etile_map)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t num_blocks_per_batch =

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

@@ -75,7 +75,7 @@ __global__ void
                                           const Block2ETileMap block_2_etile_map)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     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

@@ -61,7 +61,7 @@ __global__ void
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __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_multi_d_xdl.hpp

@@ -84,7 +84,7 @@ __global__ void
 {
 
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     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_dl.hpp

@@ -70,9 +70,8 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||             \
-    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
-    defined(__gfx1101__) || defined(__gfx1102__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
+    defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__))
 
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -648,11 +647,8 @@ struct DeviceBatchedGemmMultipleD_Dl : public DeviceBatchedGemmMultiD<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        // TODO: Enable for gfx90a after complier fix
-        if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx90a" ||
-           ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx1030" ||
-           ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx1100" ||
-           ck::get_device_name() == "gfx1101" || ck::get_device_name() == "gfx1102")
+        if(ck::get_device_name() == "gfx906" || ck::is_xdl_supported() ||
+           ck::is_navi2_supported() || ck::is_navi3_supported())
         {
             bool pass = true;
             pass      = pass && arg.K_ % K1 == 0;

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

@@ -69,7 +69,7 @@ __global__ void
             const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __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

@@ -60,7 +60,7 @@ __global__ void
             const Block2CTileMap block_2_ctile_map)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     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

@@ -68,7 +68,7 @@ __global__ void
             const C0MatrixMask c0_matrix_mask)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __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_softmax_gemm_xdl_cshuffle.hpp

@@ -63,7 +63,7 @@ __global__ void
             const C0MatrixMask c0_matrix_mask)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __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_xdl.hpp

@@ -53,7 +53,7 @@ __global__ void
         kernel_batched_gemm_xdlops_v2r3(const typename DeviceOp::Argument karg)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     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);