merge develop

codegen/include/ck/host/types.hpp

@@ -66,6 +66,20 @@ enum class GemmType
 };
 std::string ToString(GemmType gt);
 
+enum class LoopScheduler
+{
+    Default,
+    Interwave,
+};
+std::string ToString(LoopScheduler ls);
+
+enum class PipelineVersion
+{
+    v1,
+    v2
+};
+std::string ToString(PipelineVersion pv);
+
 struct TensorDesc
 {
     DataType element;
@@ -84,6 +98,7 @@ const std::string S = SequenceStr({xs...});
 
 constexpr const char* PassThrough = "ck::tensor_operation::element_wise::PassThrough";
 constexpr const char* Bilinear    = "ck::tensor_operation::element_wise::Bilinear";
+constexpr const char* Scale       = "ck::tensor_operation::element_wise::Scale";
 
 } // namespace host
 } // namespace ck

codegen/src/device_batched_gemm_softmax_gemm.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/host/device_batched_gemm_softmax_gemm/problem.hpp"
+#include "ck/host/device_batched_gemm_softmax_gemm/operation.hpp"
+#include "ck/host/utils.hpp"
+#include <algorithm>
+
+namespace ck {
+namespace host {
+namespace device_batched_gemm_softmax_gemm {
+
+// return the relevant device op file based on the operation
+std::string Problem::GetIncludeHeader() const
+{
+    return "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp";
+}
+
+// returns templated instances when provided with a problem specification
+std::vector<Solution> Problem::GetSolutions(const std::string& arch,
+                                            const std::string& prologue,
+                                            const std::string& epilogue) const
+{
+    if(get_xdlop_archs().count(arch) == 0)
+        return {};
+    auto ops = ck::host::device_batched_gemm_softmax_gemm::Operation_Xdl_CShuffle::CreateOperations(
+        *this, prologue, epilogue); // obtains vector of instances
+    std::vector<Solution> result;
+    std::transform(ops.begin(), ops.end(), std::back_inserter(result), [&](const auto& op) {
+        return op.ToSolution(); // template instance with correct values
+    });
+    return result;
+}
+
+} // namespace device_batched_gemm_softmax_gemm
+} // namespace host
+} // namespace ck

codegen/src/device_gemm_multiple_d_operation_xdl_cshuffle.cpp

@@ -62,6 +62,12 @@ void Operation_Xdl_CShuffle::update_epilogue(const std::string& epi)
 // accounts for all possible combinations of Row/Col major
 static Layout ToLayout(bool Trans) { return Trans ? Layout::Column : Layout::Row; }
 
+// clang-format off
+// DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1,
+
+// DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+// clang-format on
+
 // Hard-code tuning parameters in modularized fashion, string them together into a vector of
 // instances
 std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
@@ -83,6 +89,8 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
   {   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,        1},
   {   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,        1},
   {   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,        1},
+//  Irregular tile
+  {    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,        1},
         // clang-format on
     };
 
@@ -100,6 +108,8 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
   {    S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1},
   {    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1},
   {    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1},
+//  Irregular tile
+  {    S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1},
         // clang-format on
     };
 
@@ -109,15 +119,17 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
 //   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|
 // Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |
 //                |               |               |               |               |               |          |
+  {    S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1},
+  {    S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1},
+  {    S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1},
+  {    S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1},
+  {    S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1},
+  {    S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1},
+  {    S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1},
+  {    S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1},
+//  Irregular tile
+  {    S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1},
         // clang-format on
-        {S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1},
-        {S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1},
-        {S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1},
-        {S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1},
-        {S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1},
-        {S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1},
-        {S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1},
-        {S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1},
     };
 
     std::vector<operation::BlockTransferDesc> b_block_descriptions_rowmajor = {
@@ -134,6 +146,8 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
   {    S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1},
   {    S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1},
   {    S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1},
+//  Irregular tile
+  {    S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1},
         // clang-format on
     };
 
@@ -151,6 +165,8 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
   {    S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1},
   {    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1},
   {    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1},
+//  Irregular tile
+  {    S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1},
         // clang-format on
     };
 
@@ -167,6 +183,7 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
   {          1,           1},
   {          1,           1},
   {          1,           1},
+  {          1,           1},
   {          1,           1},
         // clang-format on
     };
@@ -185,6 +202,8 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
   {              S<1, 16, 1, 8>,               8},
   {              S<1, 32, 1, 8>,               8},
   {              S<1, 32, 1, 8>,               8},
+//  Irregular tile
+  {              S<1, 16, 1, 4>,               1},
         // clang-format on
     };
 
@@ -199,6 +218,14 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
     assert(tile_descriptions.size() == cshuffle_descriptions.size());
     assert(tile_descriptions.size() == c_block_descriptions.size());
 
+    const std::vector<std::tuple<LoopScheduler, PipelineVersion>> scheduler_pipeline_descriptions =
+        {
+            {LoopScheduler::Default, PipelineVersion::v1},
+            {LoopScheduler::Interwave, PipelineVersion::v1},
+            {LoopScheduler::Default, PipelineVersion::v2},
+        };
+    for(auto [loop_scheduler, pipeline_version] : scheduler_pipeline_descriptions)
+    {
         // Put all values together into a single operation > store into the result vector
         for(std::size_t i = 0; i < tile_descriptions.size(); i++)
         {
@@ -223,10 +250,13 @@ std::vector<Operation_Xdl_CShuffle> Operation_Xdl_CShuffle::CreateOperations(
                                                 x.tile_desc.m_per_block,
                                                 x.tile_desc.n_per_block,
                                                 x.tile_desc.k_per_block);
+            x.loop_scheduler      = loop_scheduler;
+            x.pipeline_version    = pipeline_version;
             x.update_prologue(prologue);
             x.update_epilogue(epilogue);
             result.push_back(x);
         }
+    }
     return result;
 }
 
@@ -263,7 +293,7 @@ static const char* const DeviceGemmMultipleD_Xdl_CShuffleTemplate =
     "${BBlockTransferSrcScalarPerVector}, ${BBlockTransferDstScalarPerVector_BK1}, "
     "${BBlockLdsExtraN}, ${CShuffleMXdlPerWavePerShuffle}, ${CShuffleNXdlPerWavePerShuffle}, "
     "${CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock}, "
-    "${CDEBlockTransferScalarPerVector_NPerBlock}>";
+    "${CDEBlockTransferScalarPerVector_NPerBlock}, ${LoopScheduler}, ${PipelineVersion}>";
 
 // use hardcoded instances from vector of operations to substitute values into instance template
 Solution Operation_Xdl_CShuffle::ToSolution() const
@@ -336,6 +366,8 @@ Solution Operation_Xdl_CShuffle::ToSolution() const
          this->c_block_transfer.cluster_lengths_m_block_m_wave_m_per_Xdl_n_block_n_wave_n_per_Xdl},
         {"CDEBlockTransferScalarPerVector_NPerBlock",
          std::to_string(this->c_block_transfer.scalar_per_vector_n_wave_n_per_Xdl)},
+        {"LoopScheduler", ToString(this->loop_scheduler)},
+        {"PipelineVersion", ToString(this->pipeline_version)},
     };
 
     return Solution{InterpolateString(DeviceGemmMultipleD_Xdl_CShuffleTemplate, values),

codegen/src/headers.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include "ck/host/headers.hpp"
 #include "ck_headers.hpp"
 

codegen/src/types.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include "ck/host/types.hpp"
 #include "ck/host/stringutils.hpp"
 #include <algorithm>
@@ -56,6 +59,26 @@ std::string ToString(GemmType gt)
     throw std::runtime_error("Incorrect gemm type");
 }
 
+std::string ToString(LoopScheduler ls)
+{
+    switch(ls)
+    {
+    case LoopScheduler::Default: return "ck::LoopScheduler::Default";
+    case LoopScheduler::Interwave: return "ck::LoopScheduler::Interwave";
+    }
+    throw std::runtime_error("Incorrect LoopScheduler type");
+}
+
+std::string ToString(PipelineVersion pv)
+{
+    switch(pv)
+    {
+    case PipelineVersion::v1: return "ck::PipelineVersion::v1";
+    case PipelineVersion::v2: return "ck::PipelineVersion::v2";
+    }
+    throw std::runtime_error("Incorrect PipelineVersion type");
+}
+
 std::string SequenceStr(const std::vector<int>& v)
 {
     return "ck::Sequence<" +

codegen/test/gemm_multiple_d.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include "ck/host/device_gemm_multiple_d/problem.hpp"
 #include "ck/host/device_gemm_multiple_d/operation.hpp"
 #include "ck/host/headers.hpp"

codegen/test/grouped_conv_fwd_multiple_d_v1.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_op.hpp"
 #include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_problem.hpp"
 #include "ck/host/headers.hpp"

codegen/test/grouped_conv_fwd_multiple_d_v2.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_op.hpp"
 #include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_problem.hpp"
 #include "ck/host/headers.hpp"

codegen/test/grouped_conv_fwd_multiple_d_v3.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_op.hpp"
 #include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_problem.hpp"
 #include "ck/host/headers.hpp"

codegen/test/grouped_conv_fwd_multiple_d_v4.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_op.hpp"
 #include "ck/host/device_grouped_conv_fwd_multiple_d/conv_fwd_problem.hpp"
 #include "ck/host/headers.hpp"

codegen/test/include/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 #include <algorithm>
 #include <cmath>

codegen/test/rtc/include/rtc/compile_kernel.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #ifndef GUARD_HOST_TEST_RTC_INCLUDE_RTC_COMPILE_KERNEL
 #define GUARD_HOST_TEST_RTC_INCLUDE_RTC_COMPILE_KERNEL
 

codegen/test/rtc/include/rtc/hip.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #ifndef GUARD_HOST_TEST_RTC_INCLUDE_RTC_HIP
 #define GUARD_HOST_TEST_RTC_INCLUDE_RTC_HIP
 
 #include <hip/hip_runtime_api.h>
 #include <memory>
+#include <stdexcept>
 #include <string>
 #include <stdexcept>
 

codegen/test/rtc/include/rtc/kernel.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #ifndef GUARD_HOST_TEST_RTC_INCLUDE_RTC_KERNEL
 #define GUARD_HOST_TEST_RTC_INCLUDE_RTC_KERNEL
 

codegen/test/rtc/include/rtc/manage_ptr.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #ifndef GUARD_HOST_TEST_RTC_INCLUDE_RTC_MANAGE_POINTER
 #define GUARD_HOST_TEST_RTC_INCLUDE_RTC_MANAGE_POINTER
 

codegen/test/rtc/include/rtc/tmp_dir.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #ifndef GUARD_HOST_TEST_RTC_INCLUDE_RTC_TMP_DIR
 #define GUARD_HOST_TEST_RTC_INCLUDE_RTC_TMP_DIR
 

codegen/test/rtc/src/compile_kernel.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include <rtc/hip.hpp>
 #include <rtc/compile_kernel.hpp>
 #include <rtc/tmp_dir.hpp>

codegen/test/rtc/src/hip.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include <rtc/hip.hpp>
 #include <rtc/manage_ptr.hpp>
 #include <stdexcept>

codegen/test/rtc/src/kernel.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
 #include <rtc/kernel.hpp>
 #include <rtc/manage_ptr.hpp>
 #include <rtc/hip.hpp>
+#include <stdexcept>
 #include <cassert>
 
 // extern declare the function since hip/hip_ext.h header is broken