|
| 1 | +#include <cublas_v2.h> |
| 2 | + |
| 3 | +#include "src/turbomind/core/cuda_data_type.h" |
| 4 | +#include "src/turbomind/core/data_type.h" |
| 5 | + |
| 6 | +#include "src/turbomind/kernels/gemm/kernel.h" |
| 7 | +#include "src/turbomind/kernels/gemm/registry.h" |
| 8 | +#include "src/turbomind/kernels/gemm/types.h" |
| 9 | + |
| 10 | +namespace turbomind::gemm { |
| 11 | + |
| 12 | +class CublasKernel: public Kernel { |
| 13 | +public: |
| 14 | + explicit CublasKernel() |
| 15 | + { |
| 16 | + cublasCreate(&cublas_); |
| 17 | + if (0) { |
| 18 | + cublasSetMathMode(cublas_, CUBLAS_MATH_DISALLOW_REDUCED_PRECISION_REDUCTION); |
| 19 | + } |
| 20 | + desc_ = {}; |
| 21 | + desc_.backend = 1; |
| 22 | + name_ = GetName(); |
| 23 | + } |
| 24 | + |
| 25 | + int Launch(const Operation& operation, |
| 26 | + float alpha, |
| 27 | + const void* A, |
| 28 | + const MatrixLayout& Adesc, |
| 29 | + const void* U, |
| 30 | + const MatrixLayout& Udesc, |
| 31 | + const void* B, |
| 32 | + const MatrixLayout& Bdesc, |
| 33 | + const void* V, |
| 34 | + const MatrixLayout& Vdesc, |
| 35 | + float beta, |
| 36 | + const void* C, |
| 37 | + const MatrixLayout& Cdesc, |
| 38 | + void* D, |
| 39 | + const MatrixLayout& Ddesc, |
| 40 | + int swizzle, |
| 41 | + int splits, |
| 42 | + Workspace& workspace, |
| 43 | + cudaStream_t stream) override |
| 44 | + { |
| 45 | + cublasOperation_t transa = Adesc.order == kColMajor ? CUBLAS_OP_N : CUBLAS_OP_T; |
| 46 | + cublasOperation_t transb = Bdesc.order == kColMajor ? CUBLAS_OP_N : CUBLAS_OP_T; |
| 47 | + |
| 48 | + const int m = Adesc.rows; |
| 49 | + const int n = Bdesc.cols; |
| 50 | + const int k = Adesc.cols; |
| 51 | + |
| 52 | + TM_CHECK_EQ(Bdesc.rows, k); |
| 53 | + TM_CHECK_EQ(Ddesc.rows, m); |
| 54 | + TM_CHECK_EQ(Ddesc.cols, n); |
| 55 | + |
| 56 | + TM_CHECK(C == nullptr || C == D); |
| 57 | + |
| 58 | + if (stream_ != stream) { |
| 59 | + cublasSetStream(cublas_, stream); |
| 60 | + stream_ = stream; |
| 61 | + } |
| 62 | + |
| 63 | + if (workspace_ != workspace.partials || workspace_size_ != workspace.partials_size) { |
| 64 | + cublasSetWorkspace(cublas_, workspace.partials, workspace.partials_size); |
| 65 | + workspace_ = workspace.partials; |
| 66 | + workspace_size_ = workspace.partials_size; |
| 67 | + } |
| 68 | + |
| 69 | + auto ec = cublasGemmEx(cublas_, |
| 70 | + transa, |
| 71 | + transb, |
| 72 | + m, |
| 73 | + n, |
| 74 | + k, |
| 75 | + &alpha, |
| 76 | + A, |
| 77 | + to_cuda_dtype(Adesc.type), |
| 78 | + Adesc.ld, |
| 79 | + B, |
| 80 | + to_cuda_dtype(Bdesc.type), |
| 81 | + Bdesc.ld, |
| 82 | + &beta, |
| 83 | + D, |
| 84 | + to_cuda_dtype(Ddesc.type), |
| 85 | + Ddesc.ld, |
| 86 | + CUDA_R_32F, |
| 87 | + CUBLAS_GEMM_DEFAULT_TENSOR_OP); |
| 88 | + |
| 89 | + return ec == CUBLAS_STATUS_SUCCESS ? 0 : 1; |
| 90 | + } |
| 91 | + |
| 92 | + bool is_feasible(const GemmDesc& desc) const noexcept override |
| 93 | + { |
| 94 | + constexpr std::tuple flat3{Striding::kFlat, Striding::kFlat, Striding::kFlat}; |
| 95 | + |
| 96 | + if (std::tie(desc.striding_a, desc.striding_b, desc.striding_c) != flat3) { |
| 97 | + return false; |
| 98 | + } |
| 99 | + if (std::tie(desc.pack_a, desc.pack_b, desc.pack_u, desc.pack_v) != std::tuple{0, 0, 0, 0}) { |
| 100 | + return false; |
| 101 | + } |
| 102 | + if (desc.epilogue != Epilogue::kNone) { |
| 103 | + return false; |
| 104 | + } |
| 105 | + if (desc.num > 1) { |
| 106 | + return false; |
| 107 | + } |
| 108 | + if (desc.quant_a || desc.quant_b) { |
| 109 | + return false; |
| 110 | + } |
| 111 | + if (desc.sched) { |
| 112 | + return false; |
| 113 | + } |
| 114 | + if (desc.order_c != kColMajor) { |
| 115 | + return false; |
| 116 | + } |
| 117 | + if (desc.type_a != kHalf && desc.type_a != kBfloat16 && desc.type_a != kFloat) { |
| 118 | + return false; |
| 119 | + } |
| 120 | + if (desc.type_b != desc.type_a) { |
| 121 | + return false; |
| 122 | + } |
| 123 | + if (desc.type_c != desc.type_a && desc.type_c != kFloat) { |
| 124 | + return false; |
| 125 | + } |
| 126 | + return true; |
| 127 | + } |
| 128 | + |
| 129 | + int GetMaxSplits(const int4&, int64_t, size_t, size_t) const override |
| 130 | + { |
| 131 | + return 1; |
| 132 | + } |
| 133 | + |
| 134 | + int GetSwizzle(int m, int n, int k, int splits, int swizzle) const override |
| 135 | + { |
| 136 | + return 0; |
| 137 | + } |
| 138 | + |
| 139 | +private: |
| 140 | + cublasHandle_t cublas_{}; |
| 141 | + cudaStream_t stream_{}; |
| 142 | + void* workspace_{}; |
| 143 | + size_t workspace_size_{}; |
| 144 | +}; |
| 145 | + |
| 146 | +void Registry::cublas_float() |
| 147 | +{ |
| 148 | + Add(std::make_unique<CublasKernel>()); |
| 149 | +} |
| 150 | + |
| 151 | +} // namespace turbomind::gemm |
0 commit comments