| 49 | template<class AType, class BType, class CType, class AccumType, |
| 50 | int MTile, int NTile, int KTile> |
| 51 | __global__ void reference_gpu_gemm_kernel(GemmParams<AType, BType, CType, AccumType> params) { |
| 52 | static_assert(MTile > 0); |
| 53 | static_assert(NTile > 0); |
| 54 | static_assert(KTile > 0); |
| 55 | int m = (blockIdx.y * blockDim.y + threadIdx.y) * MTile; |
| 56 | int n = (blockIdx.x * blockDim.x + threadIdx.x) * NTile; |
| 57 | AccumType accum[MTile * NTile] = {AccumType(0)}; |
| 58 | for (int k = 0; k < params.K; k += KTile) { |
| 59 | for (int mi = 0; mi < MTile; ++mi) { |
| 60 | for (int ki = 0; ki < KTile; ++ki) { |
| 61 | for (int ni = 0; ni < NTile; ++ni) { |
| 62 | AType aval = k + ki < params.K ? params.A[(m + mi) % params.M * params.K + k + ki] : AType(0); |
| 63 | BType bval = k + ki < params.K ? params.B[(n + ni) % params.N * params.K + k + ki] : BType(0); |
| 64 | accum[mi * NTile + ni] += AccumType(aval) * AccumType(bval); |
| 65 | } |
| 66 | } |
| 67 | } |
| 68 | } |
| 69 | for (int mi = 0; mi < MTile; ++mi) { |
| 70 | for (int ni = 0; ni < NTile; ++ni) { |
| 71 | if ((m + mi < params.M) && (n + ni < params.N)) { |
| 72 | params.C[(m + mi) * params.N + n + ni] = (CType)(accum[mi * NTile + ni] * params.alpha + params.beta); |
| 73 | } |
| 74 | } |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | template<class AType, class BType, class CType, class AccumType> |
| 79 | void reference_gpu_gemm(GemmParams<AType, BType, CType, AccumType> params) { |