/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/3643/include/ck_tile/ops/gemm_quant/pipeline/gemm_group_quant_utils.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/3643/include/ck_tile/ops/gemm_quant/pipeline/gemm_group_quant_utils.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/3643/include/ck_tile/ops/gemm_quant/pipeline/gemm_group_quant_utils.hpp Source File
Go to the documentation of this file.
 // Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 // SPDX-License-Identifier: MIT
  
 #pragma once
  
 #include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
  
 namespace ck_tile {
  
 template <typename Problem, typename DataType, index_t YPerTile, index_t XPerTile>
 CK_TILE_HOST_DEVICE static constexpr auto GetABQGlobalVectorLoadSize()
 {
     using I1                 = number<1>;
     constexpr index_t NWarps = Problem::BlockGemmShape::BlockWarps::at(I1{});
  
     constexpr index_t BlockSize = Problem::kBlockSize;
  
     // Data is replicated across warps along NWarps, so we divide BlockSize by NWarps
     constexpr index_t elements_per_thread = (YPerTile * XPerTile) / (BlockSize / NWarps);
     constexpr index_t PackedSize = ck_tile::numeric_traits<remove_cvref_t<DataType>>::PackedSize;
  
     // Define vector load candidates in descending order of priority
     constexpr std::array<index_t, 5> candidates{
         PackedSize * 32 / sizeof(DataType),
         PackedSize * 16 / sizeof(DataType),
         PackedSize * 8 / sizeof(DataType),
         PackedSize * 4 / sizeof(DataType),
         PackedSize * 2 / sizeof(DataType),
     };
  
     for(const auto vec_size : candidates)
     {
         if(vec_size <= 0 || XPerTile % vec_size != 0 || elements_per_thread % vec_size != 0)
             continue;
         bool is_valid = (vec_size > 0) && (XPerTile % vec_size == 0) &&
                         (elements_per_thread % vec_size == 0) && vec_size != candidates[4];
         if(is_valid)
         {
             return vec_size;
         }
     }
     return PackedSize; // Absolute fallback
 }
  
 // AQ holds groupquant scale data for A. Data is loaded from DRAM and partitioned across
 // threads. Post mfma scales are shuffled across threads in the warp and applied to
 // accum registers.
 template <typename BlockGemmShape,
           typename WarpGemm,
           index_t BlockSize,
           index_t YPerTile,
           index_t XPerTile,
           index_t KPerBlockAQ,
           index_t VecSize,
           bool PreshuffleQuant>
 struct tile_distribution_encoding_pattern_aq : public tile_distribution_encoding_pattern
 {
     static_assert(XPerTile % VecSize == 0, "XPerTile must be a multiple of VecSize!");
     static constexpr index_t warp_size = get_warp_size();
     static constexpr index_t num_warps = BlockSize / get_warp_size();
  
     static constexpr index_t MWarps = BlockGemmShape::BlockWarps::at(number<0>{});
     static constexpr index_t NWarps = BlockGemmShape::BlockWarps::at(number<1>{});
     static constexpr index_t KWarps = BlockGemmShape::BlockWarps::at(number<2>{});
  
     static constexpr index_t MIterPerWarp = BlockGemmShape::kM / (MWarps * WarpGemm::kM);
  
     static_assert(num_warps == MWarps * NWarps * KWarps);
  
     // KWarps > 1 isn't supported
     static_assert(KWarps == 1);
  
     CK_TILE_HOST_DEVICE static constexpr auto make_2d_static_tile_distribution()
     {
         if constexpr(PreshuffleQuant)
         {
             // # of elements per thread
             static_assert(XPerTile >= warp_size && XPerTile % warp_size == 0);
             constexpr index_t X1 = warp_size;
             constexpr index_t X0 = XPerTile / warp_size;
  
             constexpr index_t Y1 = MWarps;
             constexpr index_t Y0 = YPerTile / Y1;
             return make_static_tile_distribution(
                 tile_distribution_encoding<sequence<NWarps>,
                                            tuple<sequence<Y0, Y1>, sequence<X0, X1>>,
                                            tuple<sequence<1, 0>, sequence<2>>,
                                            tuple<sequence<1, 0>, sequence<1>>,
                                            sequence<1, 2>,
                                            sequence<0, 0>>{});
         }
         else
         {
             // # of elements per thread
             constexpr index_t X = XPerTile;
  
             constexpr index_t YR = 1;
             constexpr index_t Y0 = MIterPerWarp ? MIterPerWarp : 1;
             constexpr index_t Y1 = MWarps;
             constexpr index_t Y2 = WarpGemm::kM;
             static_assert(Y2 >= WarpGemm::kM,
                           "Scales for all rows must be available within the warp.");
             static_assert(Y0 * Y1 * Y2 == YPerTile, "Y0, Y1, Y2 must cover the blocktile along Y.");
             return make_static_tile_distribution(
                 tile_distribution_encoding<sequence<NWarps, YR>,
                                            tuple<sequence<Y0, Y1, Y2>, sequence<X>>,
                                            tuple<sequence<1, 0>, sequence<0, 1>>,
                                            tuple<sequence<1, 0>, sequence<1, 2>>,
                                            sequence<1, 2>,
                                            sequence<0, 0>>{});
         }
     }
     CK_TILE_HOST_DEVICE static constexpr auto make_2d_static_tile_distribution_transposed()
     {
  
         constexpr index_t Y0 = YPerTile;
         constexpr index_t X0 = 1;
         constexpr index_t X1 = MIterPerWarp ? MIterPerWarp : 1;
         constexpr index_t X2 = MWarps;
         constexpr index_t X3 = WarpGemm::kM;
  
         static_assert(X3 >= WarpGemm::kM, "Scales for all rows must be available within the warp.");
         static_assert(X0 * X1 * X2 * X3 == XPerTile,
                       "X0, X1, X2, X3 must cover the blocktile along X.");
  
         return make_static_tile_distribution(
             tile_distribution_encoding<sequence<NWarps>,
                                        tuple<sequence<Y0>, sequence<X0, X1, X2, X3>>,
                                        tuple<sequence<2, 0>, sequence<2, 2>>,
                                        tuple<sequence<2, 0>, sequence<0, 3>>,
                                        sequence<2, 1>,
                                        sequence<1, 0>>{});
     }
 };
  
 template <typename BlockGemmShape,
           typename WarpGemm,
           index_t BlockSize,
           index_t YPerTile,
           index_t XPerTile,
           index_t VecSize>
 struct tile_distribution_encoding_pattern_aq_transposed_c
     : public tile_distribution_encoding_pattern
 {
     // TODO: make pattern where below condition does not need to hold - GGemmMultiDSplitk!
     static_assert(XPerTile % VecSize == 0, "XPerTile must be a multiple of VecSize!");
     static constexpr index_t warp_size = get_warp_size();
     static constexpr index_t num_warps = BlockSize / get_warp_size();
  
     static constexpr index_t MWarps = BlockGemmShape::BlockWarps::at(number<0>{});
     static constexpr index_t NWarps = BlockGemmShape::BlockWarps::at(number<1>{});
     static constexpr index_t KWarps = BlockGemmShape::BlockWarps::at(number<2>{});
  
     static constexpr index_t MIterPerWarp = BlockGemmShape::kM / (MWarps * WarpGemm::kM);
  
     static_assert(num_warps == MWarps * NWarps * KWarps);
  
     // KWarps > 1 isn't supported
     static_assert(KWarps == 1);
  
     // # of elements per thread
     static constexpr index_t X  = XPerTile;
     static constexpr index_t XR = 2;
  
     // Number of iters per warp
     // MIters are indexed using (Y0, Y1)
     static constexpr index_t Y0 = MIterPerWarp;
  
     // # of warps in Y dim
     static constexpr index_t Y1 = MWarps;
  
     static constexpr index_t Y2 = WarpGemm::kM;
  
     static_assert(Y0 * Y1 * Y2 == YPerTile, "Y0, Y1, Y2 must cover the blocktile along Y.");
  
     CK_TILE_HOST_DEVICE static constexpr auto make_2d_static_tile_distribution()
     {
         return make_static_tile_distribution(
             tile_distribution_encoding<sequence<NWarps, XR>,
                                        tuple<sequence<Y0, Y1, Y2>, sequence<X>>,
                                        tuple<sequence<1, 0>, sequence<0, 1>>,
                                        tuple<sequence<1, 0>, sequence<1, 2>>,
                                        sequence<1, 2>,
                                        sequence<0, 0>>{});
     }
 };
  
 // TODO:: might need to update
 template <typename BlockGemmShape,
           typename WarpGemm,
           index_t BlockSize,
           index_t KPerTile,
           index_t NPerTile,
           index_t NPerQ,
           index_t KPerQ,
           typename BQLayout    = tensor_layout::gemm::ColumnMajor,
           bool PreshuffleQuant = false>
 struct tile_distribution_encoding_pattern_bq : public tile_distribution_encoding_pattern
 {
     static constexpr index_t warp_size = get_warp_size();
     static constexpr index_t num_warps = BlockSize / get_warp_size();
  
     static constexpr index_t MWarps = BlockGemmShape::BlockWarps::at(number<0>{});
     static constexpr index_t NWarps = BlockGemmShape::BlockWarps::at(number<1>{});
     static constexpr index_t KWarps = BlockGemmShape::BlockWarps::at(number<2>{});
  
     static constexpr index_t NIterPerWarp = BlockGemmShape::kN / (NWarps * WarpGemm::kN);
  
     static_assert(num_warps == MWarps * NWarps * KWarps);
     static_assert(KWarps == 1);
  
     CK_TILE_HOST_DEVICE static constexpr auto make_2d_static_tile_distribution()
     {
         // Preshuffle only supported for ColumnMajor currently
         static_assert(!(PreshuffleQuant && std::is_same_v<BQLayout, tensor_layout::gemm::RowMajor>),
                       "PreshuffleQuant only supported for ColumnMajor BQLayout");
  
         if constexpr(PreshuffleQuant)
         {
             // =============================================================================
             // PRE-SHUFFLED BQ SCALE TILE DISTRIBUTION
             // =============================================================================
             // For pre-shuffled quantization, the BQ scale tensor has been reorganized
             // (pre-shuffled) to optimize memory access patterns during dequantization.
             //
             // Tile Dimensions:
             //   - K-axis (Y in encoding): Corresponds to the K-dimension iteration
             //   - N-axis (X in encoding): Flattened scale index combining N and K groups
             //
             // The encoding distributes work across threads such that each thread loads
             // the correct pre-shuffled scale for its corresponding B-matrix elements.
             // =============================================================================
             if constexpr(NPerQ <= WarpGemm::kN)
             {
                 // =========================================================================
                 // CASE 1: Fine-grained Quantization (NPerQ <= WarpGemm::kN)
                 // =========================================================================
                 // Multiple quantization scales exist within a single warp's N-dimension.
                 // Each warp processes multiple scales: WarpGemm::kN / NPerQ scales per warp.
                 //
                 // Example: NPerQ=8, WarpGemm::kN=16, KPerQ=128, BlockGemmShape::kK=256
                 //          → 2 scales per warp in N, 2 K-groups per block
  
                 // N1: Number of K-dimension quantization groups per block,
                 //      Each K-group of KPerQ elements shares the same scale.
                 // N0: Number of scales per warp in N-dimension, Since NPerQ
                 //      <= WarpGemm::kN, each warp handles multiple scales.
                 // N2: Elements per thread
                 // NR1: Elements sharing the same scale in N-dimension
                 // NR0: Interleave factor to ensure full warp utilization
                 // K1: Number of warps distributed along this dimension
                 // K0: Iterations per warp to cover the K-tile
                 // KR: No replication in K-dimension
                 constexpr auto N1  = BlockGemmShape::kK / KPerQ;
                 constexpr auto N0  = WarpGemm::kN / NPerQ;
                 constexpr auto N2  = 1;
                 constexpr auto NR1 = NPerQ;
                 constexpr auto NR0 =
                     (warp_size <= (N0 * N1 * N2 * NR1)) ? 1 : warp_size / (N0 * N1 * N2 * NR1);
                 constexpr auto K1 = NWarps;
                 constexpr auto K0 = KPerTile / K1;
                 constexpr auto KR = 1;
  
                 return make_static_tile_distribution(
                     tile_distribution_encoding<sequence<MWarps, NR0, NR1, KR>,
                                                tuple<sequence<K0, K1>, sequence<N0, N1, N2>>,
                                                tuple<sequence<0, 1>, sequence<0, 2, 0, 2, 0>>,
                                                tuple<sequence<0, 1>, sequence<1, 0, 2, 1, 3>>,
                                                sequence<1, 2>,
                                                sequence<0, 2>>{});
             }
             else if constexpr(NPerQ < WarpGemm::kN * NWarps)
             {
                 // =========================================================================
                 // CASE 2: Medium-grained Quantization (WarpGemm::kN < NPerQ < WarpGemm::kN *
                 // NWarps)
                 // =========================================================================
                 // Each warp handles exactly one quantization scale in N-dimension.
                 // Some warps share the same scale (KR > 1 creates warp grouping).
                 //
                 // Example: NPerQ=32, WarpGemm::kN=16, NWarps=4
                 //          → KR=2 (2 warps share same scale), K1=2 (2 unique scale groups)
  
                 // KR: Number of warps sharing the same scale
                 // K1: Number of distinct warp groups (unique scales)
                 // K0: Iterations to cover K-tile per warp group
                 // N1: K-dimension quantization groups
                 // N0: Scales per warp in N-dim (1 since NPerQ >= WarpGemm::kN)
                 // N2: Elements per thread
                 // NR1: Scale broadcast factor (full NPerQ)
                 // NR0: Remaining interleave factor
  
                 constexpr auto KR  = NPerQ / WarpGemm::kN;
                 constexpr auto K1  = NWarps / KR;
                 constexpr auto K0  = KPerTile / K1;
                 constexpr auto N1  = BlockGemmShape::kK / KPerQ;
                 constexpr auto N0  = 1;
                 constexpr auto N2  = 1;
                 constexpr auto NR1 = NPerQ;
                 constexpr auto NR0 =
                     (warp_size <= (N0 * N1 * N2 * NR1)) ? 1 : warp_size / (N0 * N1 * N2 * NR1);
  
                 return make_static_tile_distribution(
                     tile_distribution_encoding<sequence<MWarps, NR0, NR1, KR>,
                                                tuple<sequence<K0, K1>, sequence<N0, N1, N2>>,
                                                tuple<sequence<0, 1, 0>, sequence<0, 2, 0, 2>>,
                                                tuple<sequence<0, 1, 3>, sequence<1, 0, 2, 1>>,
                                                sequence<1, 2>,
                                                sequence<0, 2>>{});
             }
             else
             {
                 // =========================================================================
                 // CASE 3: Coarse-grained Quantization (NPerQ >= WarpGemm::kN * NWarps)
                 // =========================================================================
                 // The quantization group spans ALL warps in N-dimension.
                 // All warps share the same scale value for their N-tiles.
                 //
                 // Example: NPerQ=128, WarpGemm::kN=16, NWarps=4
                 //          → 128 >= 16*4=64, so all 4 warps use the same scale
  
                 // N1: K-dimension quantization groups
                 // N0: Minimal (1) since scale is shared across N
                 // N2: Elements per thread
                 // NR1: Fixed broadcast size
                 // NR0: Remaining interleave factor
  
                 constexpr auto N1  = BlockGemmShape::kK / KPerQ;
                 constexpr auto N0  = 1;
                 constexpr auto N2  = 1;
                 constexpr auto NR1 = 32;
                 constexpr auto NR0 =
                     (warp_size <= (N0 * N1 * N2 * NR1)) ? 1 : warp_size / (N0 * N1 * N2 * NR1);
                 return make_static_tile_distribution(
                     tile_distribution_encoding<sequence<MWarps, NWarps, NR0, NR1>,
                                                tuple<sequence<KPerTile>, sequence<N0, N1, N2>>,
                                                tuple<sequence<0, 0>, sequence<0, 2, 0, 2>>,
                                                tuple<sequence<0, 1>, sequence<2, 0, 3, 1>>,
                                                sequence<1, 2>,
                                                sequence<0, 2>>{});
             }
         }
         else
         {
             if constexpr(NPerQ < WarpGemm::kN)
             {
                 // Case 1: Fine-grained - multiple quantization scales within a single warp
                 // N dimension needs to be partitioned the same way regardless of layout
                 constexpr index_t NR = 1;                    // No N replication needed
                 constexpr index_t N0 = NIterPerWarp;         // Iterations per warp in N-dim
                 constexpr index_t N1 = NWarps;               // Number of warps in N-dim
                 constexpr index_t N2 = WarpGemm::kN / NPerQ; // Number of scales per warp
  
                 static_assert(N0 * N1 * N2 == NPerTile,
                               "N0, N1, N2 must cover the blocktile along N dimension.");
  
                 if constexpr(std::is_same_v<BQLayout, tensor_layout::gemm::ColumnMajor>)
                 {
                     // ColumnMajor: [(N0, N1, N2), K] - N on Y-axis, partition Y
                     return make_static_tile_distribution(
                         tile_distribution_encoding<sequence<MWarps, NR, NPerQ>,
                                                    tuple<sequence<N0, N1, N2>, sequence<KPerTile>>,
                                                    tuple<sequence<0, 1>, sequence<0, 1, 0>>,
                                                    tuple<sequence<0, 1>, sequence<1, 2, 2>>,
                                                    sequence<1, 2>,
                                                    sequence<0, 0>>{});
                 }
                 else
                 {
                     // RowMajor: [K, (N0, N1, N2)] - N on X-axis, partition X
                     return make_static_tile_distribution(
                         tile_distribution_encoding<sequence<MWarps, NR, NPerQ>,
                                                    tuple<sequence<KPerTile>, sequence<N0, N1, N2>>,
                                                    tuple<sequence<0, 2>, sequence<0, 2, 0>>,
                                                    tuple<sequence<0, 1>, sequence<1, 2, 2>>,
                                                    sequence<2, 1>,
                                                    sequence<0, 0>>{});
                 }
             }
             else if constexpr(NPerQ <= WarpGemm::kN * NWarps)
             {
                 // Case 2: Medium-grained - one quantization scale per warp
                 constexpr auto NR = NPerQ / WarpGemm::kN; // Scale replication factor
                 constexpr auto N1 = NWarps / NR;          // Warps per unique scale
                 constexpr auto N0 = NPerTile / N1;        // Iterations to cover N dimension
  
                 if constexpr(std::is_same_v<BQLayout, tensor_layout::gemm::ColumnMajor>)
                 {
                     // ColumnMajor: [(N0, N1), K] - N on Y-axis
                     return make_static_tile_distribution(
                         tile_distribution_encoding<sequence<MWarps, NR, get_warp_size()>,
                                                    tuple<sequence<N0, N1>, sequence<KPerTile>>,
                                                    tuple<sequence<0, 1, 0>, sequence<0>>,
                                                    tuple<sequence<0, 1, 1>, sequence<2>>,
                                                    sequence<1, 2>,
                                                    sequence<0, 0>>{});
                 }
                 else
                 {
                     // RowMajor: [K, (N0, N1)] - N on X-axis
                     return make_static_tile_distribution(
                         tile_distribution_encoding<sequence<MWarps, NR, get_warp_size()>,
                                                    tuple<sequence<KPerTile>, sequence<N0, N1>>,
                                                    tuple<sequence<0, 2, 0>, sequence<0>>,
                                                    tuple<sequence<0, 1, 1>, sequence<2>>,
                                                    sequence<2, 1>,
                                                    sequence<0, 0>>{});
                 }
             }
             else // NPerQ > WarpGemm::kN * NWarps
             {
                 // Case 3: Coarse-grained - quantization group spans all warps
                 // All warps in N-dimension share the same quantization scale
                 if constexpr(std::is_same_v<BQLayout, tensor_layout::gemm::ColumnMajor>)
                 {
                     // ColumnMajor: [N, K]
                     return make_static_tile_distribution(
                         tile_distribution_encoding<sequence<MWarps, NWarps, get_warp_size()>,
                                                    tuple<sequence<NPerTile>, sequence<KPerTile>>,
                                                    tuple<sequence<0, 0>, sequence<0>>,
                                                    tuple<sequence<0, 1>, sequence<2>>,
                                                    sequence<1, 2>,
                                                    sequence<0, 0>>{});
                 }
                 else
                 {
                     // RowMajor: [K, N]
                     return make_static_tile_distribution(
                         tile_distribution_encoding<sequence<MWarps, NWarps, get_warp_size()>,
                                                    tuple<sequence<KPerTile>, sequence<NPerTile>>,
                                                    tuple<sequence<0, 0>, sequence<0>>,
                                                    tuple<sequence<0, 1>, sequence<2>>,
                                                    sequence<2, 1>,
                                                    sequence<0, 0>>{});
                 }
             }
         }
     }
 };
  
 template <typename GroupSizes>
 struct QuantGroupShape
 {
     static constexpr index_t kM = GroupSizes::at(number<0>{});
     static constexpr index_t kN = GroupSizes::at(number<1>{});
     static constexpr index_t kK = GroupSizes::at(number<2>{});
  
     [[nodiscard]] CK_TILE_HOST static const std::string GetName()
     {
         return concat('_', "quant_group_shape", concat('x', kM, kN, kK));
     }
 };
  
 } // namespace ck_tile