merge develop

include/ck_tile/ops/smoothquant/kernel/smoothquant_kernel.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/common.hpp"
+
+namespace ck_tile {
+
+// host side args
+struct SmoothquantHostArgs
+{
+    const void* p_x;      // [m ,n], input, fp16/bf16
+    const void* p_xscale; // [1, n], input, columnwise scale, fp32
+
+    void* p_yscale; // [m, 1], output, rowwise quant scale (amax / 127) of (p_x * p_xscale)
+    void* p_qy;     // [m, n], output, p_x * p_xscale / p_yscale
+
+    index_t m;
+    index_t n;
+    index_t stride; // row_stride
+};
+
+// TODO: Extract some type to wrapper class
+template <typename Pipeline_>
+struct Smoothquant
+{
+    using Pipeline = remove_cvref_t<Pipeline_>;
+    using Problem  = typename Pipeline::Problem;
+
+    using XDataType       = remove_cvref_t<typename Problem::XDataType>;
+    using XScaleDataType  = remove_cvref_t<typename Problem::XScaleDataType>;
+    using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>;
+    using YScaleDataType  = remove_cvref_t<typename Problem::YScaleDataType>;
+    using QYDataType      = remove_cvref_t<typename Problem::QYDataType>;
+
+    static constexpr index_t Block_M = Problem::BlockShape::Block_M;
+    static constexpr index_t Block_N = Problem::BlockShape::Block_N;
+    static constexpr bool kPadM      = false; // always no need to pad along M
+    static constexpr bool kPadN      = Problem::kPadN;
+    static constexpr bool kTwoPass   = Problem::kTwoPass;
+
+    static constexpr index_t ThreadPerWarp_N = Problem::BlockShape::ThreadPerWarp_N;
+    static constexpr index_t Vector_N        = Problem::BlockShape::Vector_N;
+    static constexpr index_t Repeat_N        = Problem::BlockShape::Repeat_N;
+
+    static constexpr auto I0 = number<0>{};
+    static constexpr auto I1 = number<1>{};
+
+    struct Kargs
+    {
+        const void* p_x;
+        const void* p_xscale;
+
+        void* p_yscale;
+        void* p_qy;
+
+        index_t m;
+        index_t n;
+        index_t stride; // row_stride
+    };
+    using Hargs = SmoothquantHostArgs;
+
+    CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs)
+    {
+        return Kargs{
+            hargs.p_x, hargs.p_xscale, hargs.p_yscale, hargs.p_qy, hargs.m, hargs.n, hargs.stride};
+    }
+
+    CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
+    {
+        return dim3(integer_divide_ceil(hargs.m, Block_M));
+    }
+
+    CK_TILE_HOST static constexpr auto BlockSize() { return Problem::BlockShape::BlockSize; }
+
+    // clang-format off
+    template <typename T> struct t2s;
+    template <> struct t2s<float> { static constexpr const char * name = "fp32"; };
+    template <> struct t2s<ck_tile::fp16_t> { static constexpr const char * name = "fp16"; };
+    template <> struct t2s<ck_tile::bf16_t> { static constexpr const char * name = "bf16"; };
+    template <> struct t2s<ck_tile::fp8_t> { static constexpr const char * name = "fp8"; };
+    template <> struct t2s<ck_tile::bf8_t> { static constexpr const char * name = "bf8"; };
+    // clang-format on
+
+    // in byte
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize() { return Pipeline::GetSmemSize(); }
+
+    CK_TILE_HOST static std::string GetName()
+    {
+        // clang-format off
+        using S_ = typename Problem::BlockShape;
+        auto surfix = [&] () {
+            std::string n;
+            if (kPadN) n += "_pn";
+            if (kTwoPass) n += "_2p";
+            return n; }();
+
+        #define _SS_  std::string
+        #define _TS_  std::to_string
+        return _SS_("smoothquant_fwd_") + _SS_(t2s<XDataType>::name) + "_" +
+             _TS_(S_::Block_M) + "x" + _TS_(S_::Block_N) + "_" + _TS_(S_::WarpPerBlock_M) + "x" + _TS_(S_::WarpPerBlock_N) + "_" +
+             _TS_(S_::Warp_M) + "x" + _TS_(S_::Warp_N) + "_" + _TS_(S_::Vector_M) + "x" + _TS_(S_::Vector_N) + "_" +
+             _SS_(Pipeline::name) + surfix;
+        #undef _SS_
+        #undef _TS_
+        // clang-format on
+    }
+
+    CK_TILE_DEVICE void operator()(Kargs kargs) const
+    {
+        const auto iM = get_block_id() * Block_M;
+
+        const auto x_window = [&]() {
+            const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
+                static_cast<const XDataType*>(kargs.p_x),
+                make_tuple(kargs.m, kargs.n),
+                make_tuple(kargs.stride, 1),
+                number<Vector_N>{},
+                number<1>{});
+
+            const auto tmp2_ = pad_tensor_view(
+                tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<kPadM, kPadN>{});
+            return make_tile_window(
+                tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
+        }();
+
+        const auto xscale_window = [&]() {
+            const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
+                static_cast<const XScaleDataType*>(kargs.p_xscale),
+                make_tuple(kargs.n),
+                make_tuple(1),
+                number<Vector_N>{},
+                number<1>{});
+
+            const auto tmp2_ =
+                pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<kPadN>{});
+
+            return make_tile_window(tmp2_, make_tuple(number<Block_N>{}), {0});
+        }();
+
+        auto yscale_window = [&]() {
+            const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
+                static_cast<YScaleDataType*>(kargs.p_yscale),
+                make_tuple(kargs.m),
+                make_tuple(1),
+                number<1>{});
+
+            const auto tmp2_ =
+                pad_tensor_view(tmp_, make_tuple(number<Block_M>{}), sequence<kPadM>{});
+
+            return make_tile_window(tmp2_, make_tuple(number<Block_M>{}), {iM});
+        }();
+
+        auto qy_window = [&]() {
+            auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
+                static_cast<QYDataType*>(kargs.p_qy),
+                make_tuple(kargs.m, kargs.n),
+                make_tuple(kargs.stride, 1),
+                number<Vector_N>{},
+                number<1>{});
+
+            auto tmp2_ = pad_tensor_view(
+                tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<kPadM, kPadN>{});
+            return make_tile_window(
+                tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
+        }();
+
+        __shared__ char smem[GetSmemSize()];
+
+        Pipeline{}(x_window, xscale_window, yscale_window, qy_window, kargs.n, smem);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_default_policy.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/reduce/block/block_reduce2d_problem.hpp"
+#include "ck_tile/ops/reduce/block/block_reduce2d.hpp"
+
+namespace ck_tile {
+
+struct SmoothquantPipelineDefaultPolicy
+{
+    template <typename Problem>
+    CK_TILE_DEVICE static constexpr auto MakeXBlockTileDistribution()
+    {
+        using S = typename Problem::BlockShape;
+
+        return make_static_tile_distribution(
+            tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<S::Repeat_M, S::WarpPerBlock_M, S::ThreadPerWarp_M, S::Vector_M>,
+                      sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
+                tuple<sequence<1, 2>, sequence<1, 2>>,
+                tuple<sequence<1, 1>, sequence<2, 2>>,
+                sequence<1, 1, 2, 2>,
+                sequence<0, 3, 0, 3>>{});
+    }
+
+    template <typename Problem>
+    CK_TILE_DEVICE static constexpr auto MakeXScaleBlockTileDistribution()
+    {
+        using S = typename Problem::BlockShape;
+
+        return make_static_tile_distribution(
+            tile_distribution_encoding<
+                sequence<S::WarpPerBlock_M, S::ThreadPerWarp_M>,
+                tuple<sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
+                tuple<sequence<0, 1>, sequence<0, 1>>,
+                tuple<sequence<0, 1>, sequence<1, 2>>,
+                sequence<1, 1>,
+                sequence<0, 3>>{});
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2d()
+    {
+        using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
+                                        typename Problem::ComputeDataType,
+                                        typename Problem::BlockShape>;
+        return BlockReduce2d<P_>{};
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dSync()
+    {
+        using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
+                                        typename Problem::ComputeDataType,
+                                        typename Problem::BlockShape>;
+        return BlockReduce2dSync<P_>{};
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dCrossWarpSync()
+    {
+        using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
+                                        typename Problem::ComputeDataType,
+                                        typename Problem::BlockShape>;
+        return BlockReduce2dCrossWarpSync<P_>{};
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        if constexpr(Problem::kNeedCrossWarpSync)
+        {
+            using P_ = BlockReduce2dProblem<typename Problem::XDataType,
+                                            typename Problem::ComputeDataType,
+                                            typename Problem::BlockShape>;
+
+            using block_reduce2d = BlockReduce2d<P_>;
+            using x_block_tile =
+                decltype(make_static_distributed_tensor<typename Problem::XDataType>(
+                    MakeXBlockTileDistribution<Problem>()));
+            using y_block_tile = decltype(block_reduce2d::template MakeYBlockTile<x_block_tile>());
+
+            return GetBlockReduce2dCrossWarpSync<Problem>().template GetSmemSize<y_block_tile>();
+        }
+        else
+        {
+            return 1; // zero size arrays are an extension
+        }
+    }
+};
+} // namespace ck_tile

include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_one_pass.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
+#include <string>
+#include <type_traits>
+
+namespace ck_tile {
+
+template <typename Problem_, typename Policy_ = SmoothquantPipelineDefaultPolicy>
+struct SmoothquantPipelineOnePass
+{
+    using Problem = ck_tile::remove_cvref_t<Problem_>;
+    using Policy  = ck_tile::remove_cvref_t<Policy_>;
+
+    using XDataType       = ck_tile::remove_cvref_t<typename Problem::XDataType>;
+    using XScaleDataType  = ck_tile::remove_cvref_t<typename Problem::XScaleDataType>;
+    using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
+    using QYDataType      = ck_tile::remove_cvref_t<typename Problem::QYDataType>;
+    using YScaleDataType  = ck_tile::remove_cvref_t<typename Problem::YScaleDataType>;
+
+    static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
+    static constexpr bool kPadM              = false; // TODO - BlockSmoothquantProblem::kPadM
+    static constexpr bool kPadN              = Problem::kPadN;
+
+    static constexpr const char* name = []() {
+        if constexpr(kNeedCrossWarpSync)
+            return "bpr_op"; // block per row
+        else
+            return "wpr_op"; // warp per row
+    }();
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        return Policy::template GetSmemSize<Problem>();
+    }
+
+    template <typename XWindow, typename XScaleWindow, typename QYWindow, typename YScaleWindow>
+    CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
+                                   const XScaleWindow& xscale_window_,
+                                   YScaleWindow& yscale_window,
+                                   QYWindow& qy_window,
+                                   ck_tile::index_t,
+                                   void* smem) const
+    {
+        auto x_window =
+            make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
+        auto xscale_window = make_tile_window(
+            xscale_window_, Policy::template MakeXScaleBlockTileDistribution<Problem>());
+
+        auto reduce_absmax_func  = ReduceOp::AbsMax{};
+        auto reduce_max_func     = ReduceOp::Max{};
+        auto block_reduce2d      = Policy::template GetBlockReduce2d<Problem>();
+        auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
+        auto block_reduce2d_cross_warp_sync =
+            Policy::template GetBlockReduce2dCrossWarpSync<Problem>();
+
+        const auto x      = load_tile(x_window);
+        const auto xscale = load_tile(xscale_window);
+        auto y            = tile_elementwise_in(
+            [&](const auto& a, const auto& b) {
+                return type_convert<ComputeDataType>(a) * type_convert<ComputeDataType>(b);
+            },
+            x,
+            xscale);
+
+        // compute absmax, cross-lane->cross-warp
+        auto absmax = block_reduce2d(
+            y, reduce_absmax_func.GetIdentityValue<ComputeDataType>(), reduce_absmax_func);
+        block_reduce2d_sync(absmax, reduce_max_func);
+        block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
+
+        // ex: yscale = absmax / 127 if int8
+        auto yscale = tile_elementwise_in(
+            [&](const auto& v_) {
+                return v_ / type_convert<ComputeDataType>(numeric<QYDataType>::max());
+            },
+            absmax);
+        store_tile(yscale_window, cast_tile<YScaleDataType>(yscale));
+
+        // quantize y to qy
+        auto qy = make_static_distributed_tensor<QYDataType>(y.get_tile_distribution());
+        sweep_tile(qy, [&](auto idx) {
+            constexpr auto i_idx = make_tuple(idx[number<0>{}]);
+            auto qy_             = y[idx] / yscale[i_idx];
+            qy(idx)              = saturates<QYDataType>{}(qy_);
+        });
+        store_tile(qy_window, qy);
+    }
+};
+} // namespace ck_tile

include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_problem.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core/utility/type_traits.hpp"
+
+namespace ck_tile {
+
+// Y = X * XScale, QY = RowwiseDynamicQuant(Y) = SaturateCast(Y / YScale)
+template <typename XDataType_,
+          typename XScaleDataType_,
+          typename ComputeDataType_,
+          typename YScaleDataType_,
+          typename QYDataType_,
+          typename BlockShape_,
+          bool kPadN_,
+          bool kTwoPass_>
+struct SmoothquantPipelineProblem
+{
+    using XDataType       = remove_cvref_t<XDataType_>;
+    using XScaleDataType  = remove_cvref_t<XScaleDataType_>;
+    using ComputeDataType = remove_cvref_t<ComputeDataType_>;
+    using YScaleDataType  = remove_cvref_t<YScaleDataType_>;
+    using QYDataType      = remove_cvref_t<QYDataType_>;
+    using BlockShape      = remove_cvref_t<BlockShape_>;
+
+    static constexpr bool kNeedCrossLaneSync = BlockShape::ThreadPerWarp_N > 1;
+    static constexpr bool kNeedCrossWarpSync = BlockShape::WarpPerBlock_N > 1;
+
+    static constexpr bool kPadN    = kPadN_;
+    static constexpr bool kTwoPass = kTwoPass_;
+};
+
+} // namespace ck_tile

include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_two_pass.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
+#include <string>
+#include <type_traits>
+
+namespace ck_tile {
+
+template <typename Problem_, typename Policy_ = SmoothquantPipelineDefaultPolicy>
+struct SmoothquantPipelineTwoPass
+{
+    using Problem = ck_tile::remove_cvref_t<Problem_>;
+    using Policy  = ck_tile::remove_cvref_t<Policy_>;
+
+    using XDataType       = ck_tile::remove_cvref_t<typename Problem::XDataType>;
+    using XScaleDataType  = ck_tile::remove_cvref_t<typename Problem::XScaleDataType>;
+    using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
+    using QYDataType      = ck_tile::remove_cvref_t<typename Problem::QYDataType>;
+    using YScaleDataType  = ck_tile::remove_cvref_t<typename Problem::YScaleDataType>;
+
+    static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
+    static constexpr bool kPadM              = false; // TODO - BlockSmoothquantProblem::kPadM
+    static constexpr bool kPadN              = Problem::kPadN;
+
+    static constexpr const char* name = []() {
+        if constexpr(kNeedCrossWarpSync)
+            return "bpr_tp"; // block per row
+        else
+            return "wpr_tp"; // warp per row
+    }();
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        return Policy::template GetSmemSize<Problem>();
+    }
+
+    template <typename XWindow, typename XScaleWindow, typename QYWindow, typename YScaleWindow>
+    CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
+                                   const XScaleWindow& xscale_window_,
+                                   YScaleWindow& yscale_window,
+                                   QYWindow& qy_window,
+                                   ck_tile::index_t row_size,
+                                   void* smem) const
+    {
+        auto x_window =
+            make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
+        auto xscale_window = make_tile_window(
+            xscale_window_, Policy::template MakeXScaleBlockTileDistribution<Problem>());
+
+        static constexpr index_t Block_N = Problem::BlockShape::Block_N;
+        index_t num_n_tile_iteration =
+            __builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
+
+        auto reduce_absmax_func  = ReduceOp::AbsMax{};
+        auto reduce_max_func     = ReduceOp::Max{};
+        auto block_reduce2d      = Policy::template GetBlockReduce2d<Problem>();
+        auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
+        auto block_reduce2d_cross_warp_sync =
+            Policy::template GetBlockReduce2dCrossWarpSync<Problem>();
+
+        using XTensorType = decltype(cast_tile<ComputeDataType>(load_tile(x_window)));
+        auto absmax       = block_reduce2d.template MakeYBlockTile<XTensorType>();
+        set_tile(absmax, reduce_absmax_func.GetIdentityValue<ComputeDataType>());
+
+        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
+        {
+            const auto x      = load_tile(x_window);
+            const auto xscale = load_tile(xscale_window);
+            const auto y      = tile_elementwise_in(
+                [&](const auto& a, const auto& b) {
+                    return type_convert<ComputeDataType>(a) * type_convert<ComputeDataType>(b);
+                },
+                x,
+                xscale);
+
+            block_reduce2d(y, absmax, reduce_absmax_func);
+
+            move_tile_window(x_window, {0, Block_N});
+            move_tile_window(xscale_window, {Block_N});
+        }
+
+        // compute absmax, cross-lane->cross-warp
+        block_reduce2d_sync(absmax, reduce_max_func);
+        block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
+
+        // ex: yscale = absmax / 127 if int8
+        auto yscale = tile_elementwise_in(
+            [&](const auto& v_) {
+                return v_ / type_convert<ComputeDataType>(numeric<QYDataType>::max());
+            },
+            absmax);
+        store_tile(yscale_window, cast_tile<YScaleDataType>(yscale));
+
+        // reverse read x to reuse cache
+        ck_tile::index_t stride_to_right_most_window =
+            row_size % Block_N == 0 ? row_size - Block_N : row_size - row_size % Block_N;
+
+        move_tile_window(x_window, {0, -Block_N});
+        move_tile_window(xscale_window, {-Block_N});
+        move_tile_window(qy_window, {0, stride_to_right_most_window});
+
+        // recompute y and quantize y to qy
+        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
+        {
+            const auto x      = load_tile(x_window);
+            const auto xscale = load_tile(xscale_window);
+            const auto y      = tile_elementwise_in(
+                [&](const auto& a, const auto& b) {
+                    return type_convert<ComputeDataType>(a) * type_convert<ComputeDataType>(b);
+                },
+                x,
+                xscale);
+
+            auto qy = make_static_distributed_tensor<QYDataType>(y.get_tile_distribution());
+            sweep_tile(qy, [&](auto idx) {
+                constexpr auto i_idx = make_tuple(idx[number<0>{}]);
+                auto qy_             = y[idx] / yscale[i_idx];
+                qy(idx)              = saturates<QYDataType>{}(qy_);
+            });
+            store_tile(qy_window, qy);
+
+            move_tile_window(x_window, {0, -Block_N});
+            move_tile_window(xscale_window, {0, -Block_N});
+            move_tile_window(qy_window, {0, -Block_N});
+        }
+    }
+};
+} // namespace ck_tile

include/ck_tile/ops/softmax.hpp

@@ -5,4 +5,5 @@
 
 #include "ck_tile/ops/softmax/block/block_softmax_2d.hpp"
 #include "ck_tile/ops/softmax/block/block_softmax_2d_problem.hpp"
+#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/topk.hpp

@@ -5,4 +5,5 @@
 
 #include "ck_tile/ops/topk/block/block_topk_stream_2d.hpp"
 #include "ck_tile/ops/topk/block/block_topk_stream_2d_problem.hpp"
+#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/topk_softmax.hpp

@@ -7,4 +7,5 @@
 #include "ck_tile/ops/topk_softmax/pipeline/topk_softmax_warp_per_row_pipeline.hpp"
 #include "ck_tile/ops/topk_softmax/pipeline/topk_softmax_warp_per_row_policy.hpp"
 #include "ck_tile/ops/topk_softmax/pipeline/topk_softmax_warp_per_row_problem.hpp"
+#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/welford.hpp

@@ -6,4 +6,5 @@
 #include "ck_tile/ops/welford/block/block_welford.hpp"
 #include "ck_tile/ops/welford/block/block_welford_problem.hpp"
 #include "ck_tile/ops/welford/thread/thread_welford.hpp"
+#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/welford/block/block_welford.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck_tile/remod.py

+from datetime import datetime
 import pathlib
 from pathlib import Path
 import subprocess
@@ -8,8 +9,8 @@ NS = 'ck_tile'
 OPS = 'ops'
 OPS_COMMON = 'common' # common header will be duplicated into ops/* other module
 
-HEADER_COMMON = """// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.\n
+HEADER_COMMON = f"""// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-{datetime.now().year}, Advanced Micro Devices, Inc. All rights reserved.\n
 """
 
 # aa/bb/cc/file.hpp -> (aa, bb, cc, file.hpp)

library/src/tensor_operation_instance/gpu/CMakeLists.txt

@@ -67,6 +67,21 @@ function(add_instance_library INSTANCE_NAME)
          list(REMOVE_ITEM ARGN "${source}")
     endif()
     endforeach()
+    # Do not build gemm_universal_f8 or gemm_multiply_multiply_f8 for any targets except gfx94
+    if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH)
+    foreach(source IN LISTS ARGN)
+    if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "gemm_multiply_multiply_xdl_f8")
+         message("removing gemm_multiply_multiply_f8 instance ${source} ")
+         list(REMOVE_ITEM ARGN "${source}")
+    endif()
+    endforeach()
+    foreach(source IN LISTS ARGN)
+    if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "gemm_xdl_universal" AND source MATCHES "_f8_")
+         message("removing gemm_universal_f8 instance ${source} ")
+         list(REMOVE_ITEM ARGN "${source}")
+    endif()
+    endforeach()
+    endif()
     #only continue if there are some source files left on the list
     if(ARGN)
         set(INST_OBJ)
@@ -74,11 +89,20 @@ function(add_instance_library INSTANCE_NAME)
             set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
             if(source MATCHES "_xdl")
                 list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
-            elseif(ARGN MATCHES "_wmma")
+            elseif(source MATCHES "_wmma")
                 list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
-            elseif(ARGN MATCHES "mha")
+            elseif(source MATCHES "mha")
                 list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
             endif()
+            #only build the fp8 gemm instances for gfx908/90a if the build argument is set
+            if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH)
+                if(source MATCHES "gemm_xdl_universal" AND source MATCHES "f8")
+                    list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
+                endif()
+                if(source MATCHES "gemm_multiply_multiply_f8")
+                    list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
+                endif()
+            endif()
             set(offload_targets)
             foreach(target IN LISTS INST_TARGETS)
                     string(APPEND offload_targets "--offload-arch=${target} ")
@@ -108,7 +132,7 @@ function(add_instance_library INSTANCE_NAME)
 
         # flags to compress the library
         if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
-            message("Adding --offload-compress flag for ${INSTANCE_NAME}")
+            #message("Adding --offload-compress flag for ${INSTANCE_NAME}")
             target_compile_options(${INSTANCE_NAME} PRIVATE --offload-compress)
         endif()
 

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp

@@ -36,12 +36,12 @@ static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
 
 template <GemmSpecialization GemmSpec>
 using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances = std::tuple<
-    // clang-format off
+// clang-format off
         //################################| ALayout| BLayout|         DsLayout| ELayout|AData| BData|          DsData| EData| AccData| Cshuffle|           A|           B|              C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
         //################################|        |        |                 |        | Type|  Type|            Type|  Type|    Type|     Type| Elementwise| Elementwise|    Elementwise|Specialization|  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|                               Pipeline|                     Pipeline|
         //################################|        |        |                 |        |     |      |                |      |        |         |   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|                              Scheduler|                     Verision|
         //################################|        |        |                 |        |     |      |                |      |        |         |            |            |               |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
-        
+#if defined(__gfx94__) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
         // Compute friendly
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   256,   256,    64,  16,  16,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,   128,   128,  16,  16,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
@@ -58,17 +58,18 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances = std
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,    64,    128, 16,  16,  32,   32,    2,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,   128,    128, 16,  16,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    64,    128, 16,  16,  32,   32,    1,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
+#endif
     // clang-format on
     >;
 
 template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
 using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_instances = std::tuple<
-    // clang-format off
+// clang-format off
         //################################| ALayout| BLayout|         DsLayout| ELayout|AData| BData|          DsData| EData| AccData| Cshuffle|           A|           B|               C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|    Block-wiseGemm|               Block-wiseGemm|
         //################################|        |        |                 |        | Type|  Type|            Type|  Type|    Type|     Type| Elementwise| Elementwise|     Elementwise|Specialization|  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|          Pipeline|                     Pipeline|
         //################################|        |        |                 |        |     |      |                |      |        |         |   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|         Scheduler|                     Verision|
         //################################|        |        |                 |        |     |      |                |      |        |         |            |            |                |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                  |                             |
-
+#if defined(__gfx94__) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
         // Latency friendly 
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,   128,    32,   16,    128, 16,  16,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,      S<2, 2, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,    64,    16,   16,    128, 16,  16,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 4>,      S<4, 4, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
@@ -90,6 +91,7 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_instances = std:
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,   128,    32,  128,    128, 16,  16,  32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,     S<8, 8, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,   256,    16,  256,    128, 16,  16,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,   256,    32,  256,    128, 16,  16,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<8, 8, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>
+#endif
     // clang-format on
     >;
 } // namespace instance

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn.hpp

@@ -62,12 +62,12 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_instances = std::tuple<
 
 template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
 using device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_instances = std::tuple<
-    // clang-format off
+// clang-format off
         //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle|           A|           B|           C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|    Block-wiseGemm|               Block-wiseGemm|
         //#########################|        |        |        | Type|  Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  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|          Pipeline|                     Pipeline|
         //#########################|        |        |        |     |      |      |        |         |   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|         Scheduler|                     Verision|
         //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                  |                             |
-
+#if defined(__gfx94__) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
         // Latency friendly 
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    128, 16,   4,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 4, 1>,     S<0, 2, 1>,    S<0, 2, 1>,               1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    128, 16,   4,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 2, 1>,     S<0, 2, 1>,    S<0, 2, 1>,               1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
@@ -90,6 +90,7 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_instances = std::tuple<
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   64,    128, 16,   4,  16,   16,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 4, 1>,     S<0, 2, 1>,    S<0, 2, 1>,               1,             16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,  128,    128, 16,   8,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 8, 1>,     S<0, 2, 1>,    S<0, 2, 1>,               1,             16,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    128,  8,   8,  16,   16,    1,    4,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,    S<0, 2, 1>,               1,             16,              8,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>
+#endif
     // clang-format on
     >;
 } // namespace instance

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn.hpp

@@ -35,12 +35,12 @@ static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
 
 template <GemmSpecialization GemmSpec>
 using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_instances = std::tuple<
-    // clang-format off
+// clang-format off
         //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle|           A|           B|           C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
         //#########################|        |        |        | Type|  Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  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|                               Pipeline|                     Pipeline|
         //#########################|        |        |        |     |      |      |        |         |   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|                              Scheduler|                     Verision|
         //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
-        
+#if defined(__gfx94__) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
         // Compute friendly
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    64,  16,  16,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,   128,  16,  16,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
@@ -57,17 +57,18 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_instances = std::tuple<
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,    64,    128, 16,  16,  32,   32,    2,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
         // DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    64,   128,    128, 16,  16,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    64,    64,    128, 16,  16,  32,   32,    1,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
+#endif
     // clang-format on
     >;
 
 template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
 using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_instances = std::tuple<
-    // clang-format off
+// clang-format off
         //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle|           A|           B|           C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|    Block-wiseGemm|               Block-wiseGemm|
         //#########################|        |        |        | Type|  Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  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|          Pipeline|                     Pipeline|
         //#########################|        |        |        |     |      |      |        |         |   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|         Scheduler|                     Verision|
         //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                  |                             |
-
+#if defined(__gfx94__) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
         // Latency friendly 
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    128, 16,  16,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    128, 16,  16,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
@@ -97,6 +98,7 @@ using device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_instances = std::tuple<
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,  128,    128, 16,  16,  32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    128, 16,  16,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     F8,     F8,    BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    32,  256,    128, 16,  16,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,              8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>
+#endif
     // clang-format on
     >;
 } // namespace instance

profiler/src/profile_gemm_universal.cpp

@@ -101,7 +101,9 @@ int profile_gemm_universal(int argc, char* argv[])
     using F32  = float;
     using F16  = ck::half_t;
     using BF16 = ck::bhalf_t;
-    using F8   = ck::f8_t;
+#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
+    using F8 = ck::f8_t;
+#endif
 
     using Row = ck::tensor_layout::gemm::RowMajor;
     using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -162,6 +164,7 @@ int profile_gemm_universal(int argc, char* argv[])
     {
         return profile(F16{}, F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
     }
+#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
     else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
         return profile(F16{}, F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
@@ -178,6 +181,7 @@ int profile_gemm_universal(int argc, char* argv[])
     {
         return profile(F8{}, F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
     }
+#endif
     else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
         return profile(BF16{}, BF16{}, BF16{}, F32{}, BF16{}, Row{}, Row{}, Row{});
@@ -194,6 +198,7 @@ int profile_gemm_universal(int argc, char* argv[])
     {
         return profile(BF16{}, BF16{}, BF16{}, F32{}, BF16{}, Col{}, Row{}, Row{});
     }
+#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
     else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
         return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Row{}, Row{});
@@ -202,6 +207,7 @@ int profile_gemm_universal(int argc, char* argv[])
     {
         return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Col{}, Row{});
     }
+#endif
     else
     {
         std::cout << "this data_type & layout is not implemented" << std::endl;

test/gemm_universal/test_gemm_universal_xdl.cpp

@@ -56,7 +56,7 @@ class TestGemmUniversal_KM_NK
 using KernelTypes_MK_KN = ::testing::Types<
     //         ADataType, BDataType, ComputeDataType, CDataType
     std::tuple<      F16,       F16,             F16,     F16>,
-#if (defined CK_ENABLE_FP8)
+#if defined(CK_ENABLE_FP8) && defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
     std::tuple<      F16,        F8,             F16,     F16>,
     std::tuple<       F8,       F16,             F16,     F16>,
     std::tuple<       F8,        F8,              F8,    BF16>,
@@ -66,7 +66,7 @@ using KernelTypes_MK_KN = ::testing::Types<
 using KernelTypes_MK_NK = ::testing::Types<
     //         ADataType, BDataType, ComputeDataType, CDataType
     std::tuple<      F16,       F16,             F16,     F16>,
-#if (defined CK_ENABLE_FP8)
+#if defined(CK_ENABLE_FP8) && defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
     std::tuple<      F16,        F8,             F16,     F16>,
     std::tuple<       F8,       F16,             F16,     F16>,
     std::tuple<       F8,        F8,              F8,    BF16>,