Padding for attention: bmm+scale+softmax+bmm kernel (#385)

* add padding algo for bmm+scale+softmax+bmm. Version for verification

* remove verification code

* remove comments

* add padded bmm scale softmax bmm example

* format

* refactor

* add comments for usages of padding bmm+scale+softmax+bmm
Co-authored-by: Chao Liu <lc.roy86@gmail.com>

example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt

 add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
 add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+add_example_executable(example_padded_batched_gemm_scale_softmax_gemm_xdl_fp16 padded_batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)

example/32_batched_gemm_scale_softmax_gemm/padded_batched_gemm_scale_softmax_gemm_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+// When using padded DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle kernel, 2 specs should be set:
+// 1. GemmSpecialization should be set to MNPadding(or NPadding in future)
+// 2. Acc0ElementOp should be set to ScaleAndResetNaNToMinusInfinity
+// Otherwise, wrong result may be produced.
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::ScaleAndResetNaNToMinusInfinity;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    MNPadding,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    64,          // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    2,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<16, 16, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M             = 1020;
+    ck::index_t N             = 1020;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA       = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideC       = -1;
+    ck::index_t BatchStrideA  = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideC  = -1;
+    float alpha               = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 9)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+    }
+    else if(argc == 18)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+
+        StrideA  = std::stoi(argv[9]);
+        StrideB0 = std::stoi(argv[10]);
+        StrideB1 = std::stoi(argv[11]);
+        StrideC  = std::stoi(argv[12]);
+
+        BatchStrideA  = std::stoi(argv[13]);
+        BatchStrideB0 = std::stoi(argv[14]);
+        BatchStrideB1 = std::stoi(argv[15]);
+        BatchStrideC  = std::stoi(argv[16]);
+
+        alpha = std::stof(argv[17]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
+               "BatchStrideB0, BatchStrideB1, BatchStrideC\n");
+        printf("arg17: scale (alpha)\n");
+        exit(0);
+    }
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
+                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    auto argument =
+        gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+                          static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+                          M,
+                          N,
+                          K,
+                          O,
+                          BatchCount,
+                          StrideA,
+                          StrideB0,
+                          StrideB1,
+                          StrideC,
+                          BatchStrideA,
+                          BatchStrideB0,
+                          BatchStrideB1,
+                          BatchStrideC,
+                          a_element_op,
+                          b0_element_op,
+                          acc0_element_op,
+                          b1_element_op,
+                          c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+    if(do_verification)
+    {
+        // Output of Gemm0 is input A of Gemm1
+        Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData) ? 0 : 1;
+    }
+
+    return 0;
+}

include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp

@@ -111,6 +111,15 @@ __global__ void
 // Computes C = A * B0 * B1
 //              ^^^^^^ (Acc0)
 //              ^^^^^^^^^^^ (Acc1)
+
+// When using NPadding as GemmSpecialization, AccElementwiseOperation should be set to
+// ScaleAndResetNaNToMinusInfinity.
+// if !isNan(AccElement)
+//     AccElement *= scale
+// else
+//     AccElement = -INFINITY
+// Otherwise, result may be wrong.
+
 template <typename ALayout,
           typename BLayout, // B0Layout
           typename B1Layout,

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -97,6 +97,22 @@ struct Scale
     float scale_;
 };
 
+struct ScaleAndResetNaNToMinusInfinity
+{
+    __host__ __device__ ScaleAndResetNaNToMinusInfinity(float scale) : scale_(scale) {}
+
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        y = ck::math::isnan(x) ? -ck::NumericLimits<float>::Infinity() : scale_ * x;
+    };
+
+    float scale_;
+};
+
 struct UnaryDivide
 {
     __host__ __device__ UnaryDivide(const int32_t divider = 1) : divider_(divider) {}

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp

@@ -349,9 +349,13 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
                                const Block2CTileMap& block_2_ctile_map)
     {
         const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
-            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+            p_a_grid,
+            a_grid_desc_ak0_m_ak1.GetElementSpaceSize(),
+            NumericLimits<FloatAB>::QuietNaN());
         const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
-            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+            p_b_grid,
+            b_grid_desc_bk0_n_bk1.GetElementSpaceSize(),
+            NumericLimits<FloatAB>::QuietNaN());
         const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
         auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(

include/ck/utility/data_type.hpp

@@ -1023,6 +1023,8 @@ struct NumericLimits
     {
         return std::numeric_limits<T>::quiet_NaN();
     }
+
+    __host__ __device__ static constexpr T Infinity() { return std::numeric_limits<T>::infinity(); }
 };
 
 template <>