warp-lang 1.6.1__py3-none-win_amd64.whl → 1.7.0__py3-none-win_amd64.whl

warp/native/cutlass_gemm.cu

@@ -1,373 +0,0 @@
-/** Copyright (c) 2022 NVIDIA CORPORATION.  All rights reserved.
- * NVIDIA CORPORATION and its licensors retain all intellectual property
- * and proprietary rights in and to this software, related documentation
- * and any modifications thereto.  Any use, reproduction, disclosure or
- * distribution of this software and related documentation without an express
- * license agreement from NVIDIA CORPORATION is strictly prohibited.
- */
-
-#include "builtin.h"
-#include "temp_buffer.h"
-#include "cuda_util.h"
-
-#include "cutlass/cutlass.h"
-#include "cutlass/gemm/device/gemm_universal.h"
-#include "cutlass/util/device_memory.h"
-
-#define F16_STR "<f2"
-#define F32_STR "<f4"
-#define F64_STR "<f8"
-
-namespace wp {
-
-template <typename Gemm>
-bool run_gemm(int m, int n, int k, int batch_count, const void* a, const void* b, const void* c, void* d, float alpha, float beta) {
-    //
-    // Initialize arguments
-    //
-    typename Gemm::EpilogueOutputOp::Params epilogue_params(
-        (typename Gemm::EpilogueOutputOp::ElementCompute)alpha,
-        (typename Gemm::EpilogueOutputOp::ElementCompute)beta);
-
-    typename Gemm::Arguments arguments{
-        batch_count == 1 ? cutlass::gemm::GemmUniversalMode::kGemm : cutlass::gemm::GemmUniversalMode::kBatched ,
-        cutlass::gemm::GemmCoord{m, n, k}, // Problem size
-        batch_count,
-        epilogue_params,
-        a, b, c, d,
-        int64_t(m * k), int64_t(k * n), int64_t(m * n), int64_t(m * n), // Batch strides
-        Gemm::LayoutA::packed({m, k}).stride(0), Gemm::LayoutB::packed({k, n}).stride(0), n, n
-    };
-
-    Gemm gemm;
-    size_t workspace_size = Gemm::get_workspace_size(arguments);
-    ScopedTemporary<> workspace(WP_CURRENT_CONTEXT, workspace_size);
-    cudaStream_t stream = static_cast<cudaStream_t>(cuda_stream_get_current());
-    cutlass::Status status = gemm.initialize(arguments, workspace.buffer(), stream);
-
-    if (status != cutlass::Status::kSuccess) {
-        cudaError_t error = cudaGetLastError();
-        std::cerr << "Error initializing GEMM: " << cudaGetErrorString(error) << "\n";
-        return false;
-    }
-
-    //
-    // Run the GEMM
-    //
-
-    status = gemm(stream);
-    if (status != cutlass::Status::kSuccess) {
-        cudaError_t error = cudaGetLastError();
-        std::cerr << "Runtime error: " << cudaGetErrorString(error) << "\n";
-        return false;
-    }
-
-    return true;
-}
-
-template <
-    int ComputeCapability,
-    typename Element_,
-    typename LayoutA,
-    typename LayoutB
->
-struct DefaultGemmConfig;
-
-//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-// Partial specialization for SM80 F64 Tensor Cores
-template <typename LayoutA, typename LayoutB>
-struct DefaultGemmConfig<80, double, LayoutA, LayoutB> {
-    using Gemm = cutlass::gemm::device::GemmUniversal<
-        double, LayoutA,                                                // ElementA and LayoutA
-        double, LayoutB,                                                // ElementB and LayoutB
-        double, cutlass::layout::RowMajor,                              // ElementC and LayoutC
-        double,                                                         // ElementAccumulator
-        cutlass::arch::OpClassTensorOp,                                 // Operation type
-        cutlass::arch::Sm80,                                            // Architecture
-        cutlass::gemm::GemmShape<128, 128, 16>,                         // ThreadblockShape
-        cutlass::gemm::GemmShape<32, 64, 16>,                           // WarpShape
-        cutlass::gemm::GemmShape<8, 8, 4>,                              // Instruction Shape
-        cutlass::epilogue::thread::LinearCombination<                   // Epilogue
-            double,
-            1,
-            double,
-            double>,
-        cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,   // Swizzling
-        3                                                               // Stages
-    >;
-};
-
-// Partial specialization for SM80 F32 Tensor Cores
-template <typename LayoutA, typename LayoutB>
-struct DefaultGemmConfig<80, float, LayoutA, LayoutB> {
-    using Gemm = cutlass::gemm::device::GemmUniversal<
-        float, LayoutA,                                                 // ElementA and LayoutA
-        float, LayoutB,                                                 // ElementB and LayoutB
-        float, cutlass::layout::RowMajor,                               // ElementC and LayoutC
-        float,                                                          // ElementAccumulator
-        cutlass::arch::OpClassTensorOp,                                 // Operation type
-        cutlass::arch::Sm80,                                            // Architecture
-        cutlass::gemm::GemmShape<256, 128, 16>,                         // ThreadblockShape
-        cutlass::gemm::GemmShape<64, 64, 16>,                           // WarpShape
-        cutlass::gemm::GemmShape<16, 8, 8>,                             // Instruction Shape
-        cutlass::epilogue::thread::LinearCombination<                   // Epilogue
-            float,
-            128 / cutlass::sizeof_bits<float>::value,
-            float,
-            float>,
-        cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,   // Swizzling
-        3,                                                              // Stages
-        4, 4,                                                           // AlignmentA and AlignmentB
-        cutlass::arch::OpMultiplyAddFastF32                             // Math mode -- use 3xTF32
-    >;
-};
-
-// Partial specialization for SM80 F16 Tensor Cores
-template <typename LayoutA, typename LayoutB>
-struct DefaultGemmConfig<80, cutlass::half_t, LayoutA, LayoutB> {
-    using Gemm = cutlass::gemm::device::GemmUniversal<
-        cutlass::half_t, LayoutA,                                       // ElementA and LayoutA
-        cutlass::half_t, LayoutB,                                       // ElementB and LayoutB
-        cutlass::half_t, cutlass::layout::RowMajor,                     // ElementC and LayoutC
-        cutlass::half_t,                                                // ElementAccumulator
-        cutlass::arch::OpClassTensorOp,                                 // Operation type
-        cutlass::arch::Sm80,                                            // Architecture
-        cutlass::gemm::GemmShape<256, 128, 32>,                         // ThreadblockShape
-        cutlass::gemm::GemmShape<64, 64, 32>,                           // WarpShape
-        cutlass::gemm::GemmShape<16, 8, 16>,                            // Instruction Shape
-        cutlass::epilogue::thread::LinearCombination<                   // Epilogue
-            cutlass::half_t,
-            128 / cutlass::sizeof_bits<cutlass::half_t>::value,
-            cutlass::half_t,
-            cutlass::half_t>,
-        cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,   // Swizzling
-        3                                                               // Stages
-    >;
-};
-
-//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-// Partial specialization for SM75 F16 Tensor Cores
-template <typename LayoutA, typename LayoutB>
-struct DefaultGemmConfig<75, cutlass::half_t, LayoutA, LayoutB> {
-    using Gemm = cutlass::gemm::device::GemmUniversal<
-        cutlass::half_t, LayoutA,                                       // ElementA and LayoutA
-        cutlass::half_t, LayoutB,                                       // ElementB and LayoutB
-        cutlass::half_t, cutlass::layout::RowMajor,                     // ElementC and LayoutC
-        cutlass::half_t,                                                // ElementAccumulator
-        cutlass::arch::OpClassTensorOp,                                 // Operation type
-        cutlass::arch::Sm75,                                            // Architecture
-        cutlass::gemm::GemmShape<256, 128, 32>,                         // ThreadblockShape
-        cutlass::gemm::GemmShape<64, 64, 32>,                           // WarpShape
-        cutlass::gemm::GemmShape<16, 8, 8>,                             // Instruction Shape
-        cutlass::epilogue::thread::LinearCombination<                   // Epilogue
-            cutlass::half_t,
-            128 / cutlass::sizeof_bits<cutlass::half_t>::value,
-            cutlass::half_t,
-            cutlass::half_t>,
-        cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,   // Swizzling
-        2                                                               // Stages
-    >;
-};
-
-//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-// Partial specialization for SM70 F16 Tensor Cores
-template <typename LayoutA, typename LayoutB>
-struct DefaultGemmConfig<70, cutlass::half_t, LayoutA, LayoutB> {
-    using Gemm = cutlass::gemm::device::GemmUniversal<
-        cutlass::half_t, LayoutA,                                       // ElementA and LayoutA
-        cutlass::half_t, LayoutB,                                       // ElementB and LayoutB
-        cutlass::half_t, cutlass::layout::RowMajor,                     // ElementC and LayoutC
-        cutlass::half_t,                                                // ElementAccumulator
-        cutlass::arch::OpClassTensorOp,                                 // Operation type
-        cutlass::arch::Sm70,                                            // Architecture
-        cutlass::gemm::GemmShape<256, 128, 32>,                         // ThreadblockShape
-        cutlass::gemm::GemmShape<64, 64, 32>,                           // WarpShape
-        cutlass::gemm::GemmShape<8, 8, 4>,                              // Instruction Shape
-        cutlass::epilogue::thread::LinearCombination<                   // Epilogue
-            cutlass::half_t,
-            128 / cutlass::sizeof_bits<cutlass::half_t>::value,
-            cutlass::half_t,
-            cutlass::half_t>,
-        cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,   // Swizzling
-        2                                                               // Stages
-    >;
-};
-
-//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-// Partial specialization for SM50 SIMT
-template <typename Element, typename LayoutA, typename LayoutB>
-struct DefaultGemmConfig<50, Element, LayoutA, LayoutB> {
-    using Gemm = cutlass::gemm::device::GemmUniversal<
-        Element, LayoutA,                                               // ElementA and LayoutA
-        Element, LayoutB,                                               // ElementB and LayoutB
-        Element, cutlass::layout::RowMajor,                             // ElementC and LayoutC
-        Element,                                                        // ElementAccumulator
-        cutlass::arch::OpClassSimt,                                     // Operation type
-        cutlass::arch::Sm50,                                            // Architecture
-        cutlass::gemm::GemmShape<128, 128, 8>,                          // ThreadblockShape
-        cutlass::gemm::GemmShape<32, 64, 8>,                            // WarpShape
-        cutlass::gemm::GemmShape<1, 1, 1>,                              // Instruction Shape
-        cutlass::epilogue::thread::LinearCombination<                   // Epilogue
-            Element,
-            1,
-            Element,
-            Element>,
-        cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>,   // Swizzling
-        2                                                               // Stages
-    >;
-};
-
-//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-
-extern "C" {
-
-WP_API
-bool cutlass_gemm(
-                  void* context, int compute_capability,
-                  int m, int n, int k,
-                  const char* datatype_str,
-                  const void* a, const void* b, const void* c, void* d,
-                  float alpha, float beta,
-                  bool row_major_a, bool row_major_b,
-                  bool allow_tf32x3_arith,
-                  int batch_count) {
-
-    std::string datatype(datatype_str);
-
-    ContextGuard guard(context);
-
-    // Specializations for using Tensor Cores and A/B RowMajor/ColumnMajor designations
-    if (compute_capability == 80) {
-        if (datatype == F64_STR) {
-            if (row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<80, double, cutlass::layout::RowMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (!row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<80, double, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<80, double, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (!row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<80, double, cutlass::layout::ColumnMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            }
-        } else if (datatype == F32_STR && allow_tf32x3_arith) {
-            if (row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<80, float, cutlass::layout::RowMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (!row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<80, float, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<80, float, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (!row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<80, float, cutlass::layout::ColumnMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            }
-        } else if (datatype == F16_STR) {
-            if (row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<80, cutlass::half_t, cutlass::layout::RowMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (!row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<80, cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<80, cutlass::half_t, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (!row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<80, cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            }
-        }
-    } else if (compute_capability == 75) {
-        if (datatype == F16_STR) {
-            if (row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<75, cutlass::half_t, cutlass::layout::RowMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);    
-            } else if (!row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<75, cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);    
-            } else if (row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<75, cutlass::half_t, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);    
-            } else if (!row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<75, cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);    
-            }
-        }
-    } else if (compute_capability == 70) {
-        if (datatype == F16_STR) {
-            if (row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<70, cutlass::half_t, cutlass::layout::RowMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (!row_major_a && row_major_b) {
-                using Gemm = DefaultGemmConfig<70, cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<70, cutlass::half_t, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            } else if (!row_major_a && !row_major_b) {
-                using Gemm = DefaultGemmConfig<70, cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::layout::ColumnMajor>::Gemm;
-                return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-            }
-        }
-    }
-
-    // No Tensor Core capability available. Run a SIMT kernel
-    if (datatype == F64_STR) {
-        if (row_major_a && row_major_b) {
-            using Gemm = DefaultGemmConfig<50, double, cutlass::layout::RowMajor, cutlass::layout::RowMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        } else if (!row_major_a && row_major_b) {
-            using Gemm = DefaultGemmConfig<50, double, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        } else if (row_major_a && !row_major_b) {
-            using Gemm = DefaultGemmConfig<50, double, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        } else if (!row_major_a && !row_major_b) {
-            using Gemm = DefaultGemmConfig<50, double, cutlass::layout::ColumnMajor, cutlass::layout::ColumnMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        }
-    } else if (datatype == F32_STR) {
-        if (row_major_a && row_major_b) {
-            using Gemm = DefaultGemmConfig<50, float, cutlass::layout::RowMajor, cutlass::layout::RowMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        } else if (!row_major_a && row_major_b) {
-            using Gemm = DefaultGemmConfig<50, float, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        } else if (row_major_a && !row_major_b) {
-            using Gemm = DefaultGemmConfig<50, float, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        } else if (!row_major_a && !row_major_b) {
-            using Gemm = DefaultGemmConfig<50, float, cutlass::layout::ColumnMajor, cutlass::layout::ColumnMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        }
-    } else if (datatype == F16_STR) {
-        if (row_major_a && row_major_b) {
-            using Gemm = DefaultGemmConfig<50, cutlass::half_t, cutlass::layout::RowMajor, cutlass::layout::RowMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        } else if (!row_major_a && row_major_b) {
-            using Gemm = DefaultGemmConfig<50, cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        } else if (row_major_a && !row_major_b) {
-            using Gemm = DefaultGemmConfig<50, cutlass::half_t, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        } else if (!row_major_a && !row_major_b) {
-            using Gemm = DefaultGemmConfig<50, cutlass::half_t, cutlass::layout::ColumnMajor, cutlass::layout::ColumnMajor>::Gemm;
-            return run_gemm<Gemm>(m, n, k, batch_count, a, b, c, d, alpha, beta);
-        }
-    }
-
-    std::cerr << "Data type " << datatype << " is not currently supported." << std::endl;
-    return false;
-}
-
-}
-
-} // namespace wp