/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/smoothquant/kernel/moe_smoothquant_kernel.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/smoothquant/kernel/moe_smoothquant_kernel.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck_tile/ops/smoothquant/kernel/moe_smoothquant_kernel.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2025, 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 MoeSmoothquantHostArgs
 {
     const void* p_x;        // [tokens ,hidden_size], input, fp16/bf16
     const void* p_smscale;  // [experts, hidden_size], input, columnwise scale, fp32
     const void* p_topk_ids; // [tokens, topk]
  
     void* p_yscale; // [topk * tokens,  1], output, rowwise quant scale
     void* p_qy;     // [topk * tokens, hidden_size], output
  
     index_t tokens;
     index_t hidden_size;
     index_t experts;
     index_t topk;
     index_t x_stride; // input x row stride
     index_t y_stride; // output y stride(stride for topk)
 };
  
 // TODO: Extract some type to wrapper class
 template <typename Pipeline_>
 struct MoeSmoothquant
 {
     using Pipeline = remove_cvref_t<Pipeline_>;
     using Problem  = typename Pipeline::Problem;
  
     using XDataType           = remove_cvref_t<typename Problem::XDataType>;
     using SmoothScaleDataType = remove_cvref_t<typename Problem::SmoothScaleDataType>;
     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 index_t kBlockSize      = Problem::BlockShape::BlockSize;
  
     static constexpr auto I0 = number<0>{};
     static constexpr auto I1 = number<1>{};
  
     static_assert(Problem::BlockShape::Repeat_M == 1);
  
     struct Kargs
     {
         const void* p_x;        // [tokens ,hidden_size], input, fp16/bf16
         const void* p_smscale;  // [experts, hidden_size], input, columnwise scale, fp32
         const void* p_topk_ids; // [tokens, topk]
  
         void* p_yscale; // [topk, tokens, 1], output, rowwise quant scale
         void* p_qy;     // [topk, tokens, hidden_size], output
  
         index_t tokens;
         index_t hidden_size;
         index_t experts;
         index_t topk;
         index_t x_stride; // input x row stride
         index_t y_stride; // output y stride(stride for topk)
     };
     using Hargs = MoeSmoothquantHostArgs;
  
     CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs)
     {
         return Kargs{hargs.p_x,
                      hargs.p_smscale,
                      hargs.p_topk_ids,
                      hargs.p_yscale,
                      hargs.p_qy,
                      hargs.tokens,
                      hargs.hidden_size,
                      hargs.experts,
                      hargs.topk,
                      hargs.x_stride,
                      hargs.y_stride};
     }
  
     CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
     {
         return dim3(hargs.topk, integer_divide_ceil(hargs.tokens, Block_M), 1);
     }
  
     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"; };
     template <> struct t2s<ck_tile::int8_t> { static constexpr const char * name = "i8"; };
     // 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_("moe_smoothquant_") + _SS_(t2s<XDataType>::name) + "_" +  _SS_(t2s<QYDataType>::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 index_t i_topk  = blockIdx.x;
         const index_t i_token = blockIdx.y * Block_M;
         const index_t i_token_in_thrd =
             __builtin_amdgcn_readfirstlane(threadIdx.x / Problem::BlockShape::ThreadPerBlock_N);
  
         const index_t i_expert = reinterpret_cast<const index_t*>(
             kargs.p_topk_ids)[(i_token + i_token_in_thrd) * kargs.topk + i_topk];
  
         // [tokens ,hidden_size]
         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.tokens, kargs.hidden_size),
                 make_tuple(kargs.x_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>{}), {i_token, 0});
         }();
  
         // [experts, hidden_size],
         const auto smscale_window = [&]() {
             const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
                 static_cast<const SmoothScaleDataType*>(kargs.p_smscale) +
                     i_expert * kargs.hidden_size,
                 make_tuple(kargs.hidden_size),
                 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});
         }();
  
         // [topk, tokens]
         auto yscale_window = [&]() {
             const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
                 static_cast<YScaleDataType*>(kargs.p_yscale) + i_topk * kargs.tokens,
                 make_tuple(kargs.tokens),
                 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>{}), {i_token});
         }();
  
         // [topk, tokens, hidden_size]
         auto qy_window = [&]() {
             auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
                 static_cast<QYDataType*>(kargs.p_qy) + i_topk * kargs.tokens * kargs.y_stride,
                 make_tuple(kargs.tokens, kargs.hidden_size),
                 make_tuple(kargs.y_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>{}), {i_token, 0});
         }();
  
         __shared__ char smem[GetSmemSize()];
  
         Pipeline{}(x_window, smscale_window, yscale_window, qy_window, kargs.hidden_size, smem);
     }
 };
  
 } // namespace ck_tile