Merge remote-tracking branch 'origin/develop' into migx-jit-lib

include/ck/utility/amd_xdlops.hpp

@@ -297,6 +297,44 @@ struct intrin_mfma_i32_16x16x16i8<16, 16>
     }
 };
 
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_i32_32x32x16i8;
+
+template <>
+struct intrin_mfma_i32_32x32x16i8<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x8_t& reg_a, const int8x8_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x16_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_i32_32x32x16_i8(bit_cast<int64_t>(reg_a),
+                                                  bit_cast<int64_t>(reg_b),
+                                                  reg_c.template AsType<int32x16_t>()[Number<0>{}],
+                                                  0,
+                                                  0,
+                                                  0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_i32_16x16x32i8;
+
+template <>
+struct intrin_mfma_i32_16x16x32i8<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x8_t& reg_a, const int8x8_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x4_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_i32_16x16x32i8(bit_cast<int64_t>(reg_a),
+                                                 bit_cast<int64_t>(reg_b),
+                                                 reg_c.template AsType<int32x4_t>()[Number<0>{}],
+                                                 0,
+                                                 0,
+                                                 0);
+    }
+};
+
 template <index_t MPerWave, index_t NPerWave>
 struct intrin_mfma_f64_16x16x4f64;
 
@@ -306,7 +344,7 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
     template <class FloatC>
     __device__ static void Run(const double& reg_a, const double& reg_b, FloatC& reg_c)
     {
-#ifdef __gfx90a__
+#if defined(__gfx90a__) || defined(__gfx940__)
         reg_c.template AsType<double4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f64_16x16x4f64(
             reg_a, reg_b, reg_c.template AsType<double4_t>()[Number<0>{}], 0, 0, 0);
 #else

include/ck/utility/data_type.hpp

@@ -898,6 +898,8 @@ struct vector_type<T, 256>
     }
 };
 
+using int64_t = long;
+
 // fp64
 using double2_t = typename vector_type<double, 2>::type;
 using double4_t = typename vector_type<double, 4>::type;
@@ -974,37 +976,6 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float
         uint32_t int32;
     } u = {x};
 
-    // When the exponent bits are not all 1s, then the value is zero, normal,
-    // or subnormal. We round the bfloat16 mantissa up by adding 0x7FFF, plus
-    // 1 if the least significant bit of the bfloat16 mantissa is 1 (odd).
-    // This causes the bfloat16's mantissa to be incremented by 1 if the 16
-    // least significant bits of the float mantissa are greater than 0x8000,
-    // or if they are equal to 0x8000 and the least significant bit of the
-    // bfloat16 mantissa is 1 (odd). This causes it to be rounded to even when
-    // the lower 16 bits are exactly 0x8000. If the bfloat16 mantissa already
-    // has the value 0x7f, then incrementing it causes it to become 0x00 and
-    // the exponent is incremented by one, which is the next higher FP value
-    // to the unrounded bfloat16 value. When the bfloat16 value is subnormal
-    // with an exponent of 0x00 and a mantissa of 0x7f, it may be rounded up
-    // to a normal value with an exponent of 0x01 and a mantissa of 0x00.
-    // When the bfloat16 value has an exponent of 0xFE and a mantissa of 0x7F,
-    // incrementing it causes it to become an exponent of 0xFF and a mantissa
-    // of 0x00, which is Inf, the next higher value to the unrounded value.
-    bool flag0 = ~u.int32 & 0x7f800000;
-
-    // When all of the exponent bits are 1, the value is Inf or NaN.
-    // Inf is indicated by a zero mantissa. NaN is indicated by any nonzero
-    // mantissa bit. Quiet NaN is indicated by the most significant mantissa
-    // bit being 1. Signaling NaN is indicated by the most significant
-    // mantissa bit being 0 but some other bit(s) being 1. If any of the
-    // lower 16 bits of the mantissa are 1, we set the least significant bit
-    // of the bfloat16 mantissa, in order to preserve signaling NaN in case
-    // the bfloat16's mantissa bits are all 0.
-    bool flag1 = !flag0 && (u.int32 & 0xffff);
-
-    u.int32 += flag0 ? 0x7fff + ((u.int32 >> 16) & 1) : 0; // Round to nearest, round to even
-    u.int32 |= flag1 ? 0x10000 : 0x0;                      // Preserve signaling NaN
-
     return uint16_t(u.int32 >> 16);
 }
 
@@ -1062,6 +1033,63 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
     return type_convert<bhalf_t>(x_fp32);
 }
 
+// Declare a template function for bf16 conversion using RTN
+template <typename Y, typename X>
+__host__ __device__ constexpr Y bf16_convert_rtn(X x);
+
+// Convert fp32 to bf16 with RTN if higher precision is needed
+template <>
+inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, float>(float x)
+{
+    union
+    {
+        float fp32;
+        uint32_t int32;
+    } u = {x};
+
+    // When the exponent bits are not all 1s, then the value is zero, normal,
+    // or subnormal. We round the bfloat16 mantissa up by adding 0x7FFF, plus
+    // 1 if the least significant bit of the bfloat16 mantissa is 1 (odd).
+    // This causes the bfloat16's mantissa to be incremented by 1 if the 16
+    // least significant bits of the float mantissa are greater than 0x8000,
+    // or if they are equal to 0x8000 and the least significant bit of the
+    // bfloat16 mantissa is 1 (odd). This causes it to be rounded to even when
+    // the lower 16 bits are exactly 0x8000. If the bfloat16 mantissa already
+    // has the value 0x7f, then incrementing it causes it to become 0x00 and
+    // the exponent is incremented by one, which is the next higher FP value
+    // to the unrounded bfloat16 value. When the bfloat16 value is subnormal
+    // with an exponent of 0x00 and a mantissa of 0x7f, it may be rounded up
+    // to a normal value with an exponent of 0x01 and a mantissa of 0x00.
+    // When the bfloat16 value has an exponent of 0xFE and a mantissa of 0x7F,
+    // incrementing it causes it to become an exponent of 0xFF and a mantissa
+    // of 0x00, which is Inf, the next higher value to the unrounded value.
+    bool flag0 = ~u.int32 & 0x7f800000;
+
+    // When all of the exponent bits are 1, the value is Inf or NaN.
+    // Inf is indicated by a zero mantissa. NaN is indicated by any nonzero
+    // mantissa bit. Quiet NaN is indicated by the most significant mantissa
+    // bit being 1. Signaling NaN is indicated by the most significant
+    // mantissa bit being 0 but some other bit(s) being 1. If any of the
+    // lower 16 bits of the mantissa are 1, we set the least significant bit
+    // of the bfloat16 mantissa, in order to preserve signaling NaN in case
+    // the bfloat16's mantissa bits are all 0.
+    bool flag1 = !flag0 && (u.int32 & 0xffff);
+
+    u.int32 += flag0 ? 0x7fff + ((u.int32 >> 16) & 1) : 0; // Round to nearest, round to even
+    u.int32 |= flag1 ? 0x10000 : 0x0;                      // Preserve signaling NaN
+
+    return uint16_t(u.int32 >> 16);
+}
+
+// convert fp16 to bfp16 via fp32 with RTN if higher precision is needed
+template <>
+inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(half_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return bf16_convert_rtn<bhalf_t>(x_fp32);
+}
+
 template <typename T>
 struct NumericLimits
 {

library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
+struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_ms_ks,
+                 const Tensor<BDataType>& b_ns_ks,
+                 Tensor<CDataType>& c_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op)
+            : a_ms_ks_{a_ms_ks},
+              b_ns_ks_{b_ns_ks},
+              c_ms_ns_{c_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_ms_ks_;
+        const Tensor<BDataType>& b_ns_ks_;
+        Tensor<CDataType>& c_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_M2_N2_K2::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
+                const ck::index_t K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
+                const ck::index_t K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(ck::index_t k0 = 0; k0 < K0; ++k0)
+                {
+                    for(ck::index_t k1 = 0; k1 < K1; ++k1)
+                    {
+                        AccDataType v_a;
+                        AccDataType v_b;
+
+                        arg.a_element_op_(
+                            v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
+                        arg.b_element_op_(
+                            v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
+
+                        v_acc += v_a * v_b;
+                    }
+                }
+
+                arg.c_ms_ns_(m0, m1, n0, n1) = v_acc;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.c_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.c_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.c_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.c_ms_ns_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
+                             const Tensor<BDataType>& b_ns_ks,
+                             Tensor<CDataType>& c_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op)
+    {
+        return Argument{a_ms_ks, b_ns_ks, c_ms_ns, a_element_op, b_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M2_N2_K2"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp

@@ -6,6 +6,7 @@
 #include <iostream>
 #include <sstream>
 
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
 #include "ck/tensor_operation/gpu/device/device_base.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 
@@ -66,8 +67,26 @@ struct ReferenceGemm : public device::BaseOperator
                     ADataType v_a;
                     BDataType v_b;
 
-                    arg.a_element_op_(v_a, arg.a_m_k_(m, k));
-                    arg.b_element_op_(v_b, arg.b_k_n_(k, n));
+                    // use PassThrough instead of ConvertBF16RTN for reference calculation
+                    if constexpr(is_same_v<AElementwiseOperation,
+                                           ck::tensor_operation::element_wise::ConvertBF16RTN>)
+                    {
+                        ck::tensor_operation::element_wise::PassThrough{}(v_a, arg.a_m_k_(m, k));
+                    }
+                    else
+                    {
+                        arg.a_element_op_(v_a, arg.a_m_k_(m, k));
+                    }
+                    // same for B matrix
+                    if constexpr(is_same_v<BElementwiseOperation,
+                                           ck::tensor_operation::element_wise::ConvertBF16RTN>)
+                    {
+                        ck::tensor_operation::element_wise::PassThrough{}(v_b, arg.b_k_n_(k, n));
+                    }
+                    else
+                    {
+                        arg.b_element_op_(v_b, arg.b_k_n_(k, n));
+                    }
 
                     v_acc +=
                         ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);

library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp

@@ -68,6 +68,58 @@ void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
                                                   PassThrough,
                                                   PassThrough>>>& instances);
 
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
 template <typename ALayout,
           typename BLayout,
           typename ELayout,
@@ -109,11 +161,17 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                          is_same_v<ELayout, Row>)
             {
                 add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
+                    op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
                               is_same_v<ELayout, Row>)
             {
                 add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
+                    op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
                               is_same_v<ELayout, Row>)

library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt

@@ -3,4 +3,8 @@ add_instance_library(device_grouped_gemm_instance
    device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
    device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
    device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
 )

profiler/README.md

@@ -46,3 +46,33 @@ out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
 ....
 Best Perf: 1.42509 ms, 102.988 TFlops, 234.086 GB/s
 ```
+
+## Profile contraction kernels
+```bash
+#arg1: tensor operation (contraction_bilinear=CONTRACTION+Bilinear)
+#arg2: data type (0: fp32; 1: f64)\n"
+#arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
+#                     1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
+#                     2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
+#                     3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1])
+#arg4: verification (0: no; 1: yes)
+#arg5: initialization (0: no init; 1: integer value; 2: decimal value)
+#arg6: print tensor value (0: no; 1: yes)
+#arg7: time kernel (0: no, 1: yes)
+#arg8 and arg9: alpha and beta
+#arg10 to 15: M0, M1, N0, N1, K0, K1
+#arg16 to 31: Strides for A, B, D and E (skip for default)
+
+################                   op  datatype  layout  verify  init  log  time  alpha  beta  M0  M1  N0  N1  K0  K1
+./bin/ckProfiler contraction_bilinear         0       1       0     0    0     1    1.0   1.0 128 128 128 128 128 128
+```
+
+Result (MI100)
+```bash
+a_m_k: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1}
+b_k_n: dim 4, lengths {128, 128, 128, 128}, strides {128, 1, 2097152, 16384}
+d_m_n: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1}
+e_m_n: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1}
+....
+Best Perf: 211.405 ms, 41.6077 TFlops, 15.2372 GB/s
+```

profiler/src/CMakeLists.txt

@@ -30,6 +30,8 @@ set(PROFILER_SOURCES
     profile_batchnorm_bwd.cpp
     profile_batchnorm_infer.cpp
     profile_grouped_gemm_fastgelu.cpp
+    profile_contraction_bilinear.cpp
+    profile_contraction_scale.cpp
 )
 
 set(PROFILER_EXECUTABLE ckProfiler)
@@ -70,4 +72,6 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
 rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)