Merge remote-tracking branch 'origin/develop' into ck_tile/layernorm_fusion

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1536_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 3, 4,   64, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 3, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 3, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 6, 1,  256, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n2048_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 1, 1,  256, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 8, 1,  256, 1,  true,  true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n256_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  4,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n3072_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 3, 1,  128, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 3, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 6, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 3, 1, 1024, 1,  true,  true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1,  256, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1, 1024, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1, 1024, 1,  true,  true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_tp_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1,  256, 8,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1,  256, 4,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1, 1024, 2,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1, 1024, 1,  true,  true, true>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n512_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  4,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  8,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n64_n128_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn    pd     x      3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n768_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  3,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  6,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1, 12,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1024_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                                rm  rn  tm  tn  vn  pd      x     3p
+#if 0
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  8,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 16,  4,  64, 1,  true , true, false>>(const S&, A);
+
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  1, 256, 4,  true , true, false>>(const S&, A);
+#endif
+
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 1, 2,  128, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 2,  128, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1536_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 4,   64, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 6, 1,  256, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n2048_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 1, 1,  256, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 8, 1,  256, 1,  true,  true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n256_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n3072_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  128, 8,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  256, 4,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 6, 1,  256, 2,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1, 1024, 1,  true, true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 8,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 4,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1, 1024, 2,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1, 1024, 1,  true, true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_tp_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 8,  true, true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 4,  true, true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1, 1024, 2,  true, true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1, 1024, 1,  true, true, true>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n512_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  8,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n64_n128_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n768_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  3,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  6,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 12,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_instance_common.hpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "add_rmsnorm2d_rdquant_fwd.hpp"
+#include <iostream>
+
+#pragma once
+
+using S = ck_tile::stream_config;
+using A = add_rmsnorm2d_rdquant_fwd_args;
+
+template <typename DataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kSaveInvRms_,
+          bool kTwoPass_>
+using trait_ = add_rmsnorm2d_rdquant_fwd_traits_<DataType_,
+                                                 Repeat_M_,
+                                                 Repeat_N_,
+                                                 ThreadPerBlock_M_,
+                                                 ThreadPerBlock_N_,
+                                                 Vector_N_,
+                                                 kPadN_,
+                                                 kSaveInvRms_,
+                                                 kTwoPass_>;
+
+template <typename Traits_>
+float add_rmsnorm2d_rdquant_fwd_(const S& s, A a)
+{
+    using DataType = typename Traits_::DataType;
+
+    using PipelineProblem = ck_tile::AddRmsnorm2dRdquantFwdPipelineProblem<
+        typename AddRmsnormRdquantTypeConfig<DataType>::ADataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::BDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::GammaDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::ComputeDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::XDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::YScaleDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::QYDataType,
+        typename Traits_::Shape,
+        Traits_::kPadN,
+        Traits_::kSaveX,
+        Traits_::kThreePass>;
+
+    using OnePassPipeline   = ck_tile::AddRmsnorm2dRdquantFwdPipelineOnePass<PipelineProblem>;
+    using ThreePassPipeline = ck_tile::AddRmsnorm2dRdquantFwdPipelineThreePass<PipelineProblem>;
+    using Pipeline = std::conditional_t<Traits_::kThreePass, ThreePassPipeline, OnePassPipeline>;
+
+    using Kernel = ck_tile::AddRmsnorm2dRdquantFwd<Pipeline>;
+
+    const dim3 grids                       = Kernel::GridSize(a);
+    constexpr dim3 blocks                  = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+
+    auto kargs = Kernel::MakeKargs(a);
+    if(s.log_level_ > 0)
+        std::cout << ", " << Kernel::GetName() << std::flush;
+
+    return ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+}