105 lines · plain
1// RUN: mlir-opt %s --sparse-gpu-codegen="num-threads=0" | FileCheck %s2 3#trait_sampled_dense_dense = {4 indexing_maps = [5 affine_map<(i,j,k) -> (i,k)>, // A6 affine_map<(i,j,k) -> (k,j)>, // B7 affine_map<(i,j,k) -> (i,j)> // S (out)8 ],9 iterator_types = ["parallel", "parallel", "reduction"],10 doc = "X(i,j) += S(i,j) SUM_k A(i,k) B(k,j)"11}12 13#trait_vec_op = {14 indexing_maps = [15 affine_map<(i,j) -> (i,j)>, // a (in)16 affine_map<(i,j) -> (i,j)>, // b (in)17 affine_map<(i,j) -> (i,j)> // x (out)18 ],19 iterator_types = ["parallel", "parallel"]20}21 22#CSR = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : dense, d1 : compressed) }>23 24// CHECK-LABEL: func.func @sparse_sampled_dd(25// CHECK-SAME: %[[VAL_0:.*]]: tensor<8x8xf64, #sparse{{[0-9]*}}>,26// CHECK-SAME: %[[VAL_1:.*]]: tensor<8x8xf64>,27// CHECK-SAME: %[[VAL_2:.*]]: tensor<8x8xf64>) -> tensor<8x8xf64, #sparse{{[0-9]*}}> {28// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 8 : index29// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 0 : index30// CHECK: %[[VAL_5:.*]] = sparse_tensor.number_of_entries %[[VAL_0]] : tensor<8x8xf64, #sparse{{[0-9]*}}>31// CHECK: %[[VAL_6:.*]] = bufferization.to_buffer %[[VAL_1]] : tensor<8x8xf64> to memref<8x8xf64>32// CHECK: %[[VAL_7:.*]] = gpu.wait async33// CHECK: %[[VAL_8:.*]], %[[VAL_9:.*]] = gpu.alloc async {{\[}}%[[VAL_7]]] () : memref<8x8xf64>34// CHECK: %[[VAL_10:.*]] = gpu.memcpy async {{\[}}%[[VAL_9]]] %[[VAL_8]], %[[VAL_6]] : memref<8x8xf64>, memref<8x8xf64>35// CHECK: %[[VAL_11:.*]] = bufferization.to_buffer %[[VAL_2]] : tensor<8x8xf64> to memref<8x8xf64>36// CHECK: %[[VAL_12:.*]] = gpu.wait async37// CHECK: %[[VAL_13:.*]], %[[VAL_14:.*]] = gpu.alloc async {{\[}}%[[VAL_12]]] () : memref<8x8xf64>38// CHECK: %[[VAL_15:.*]] = gpu.memcpy async {{\[}}%[[VAL_14]]] %[[VAL_13]], %[[VAL_11]] : memref<8x8xf64>, memref<8x8xf64>39// CHECK: %[[VAL_16:.*]] = sparse_tensor.positions %[[VAL_0]] {level = 1 : index} : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xindex>40// CHECK: %[[VAL_17:.*]] = sparse_tensor.coordinates %[[VAL_0]] {level = 1 : index} : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xindex>41// CHECK: %[[VAL_18:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xf64>42// CHECK: %[[VAL_19:.*]] = gpu.wait async43// CHECK: %[[VAL_20:.*]] = memref.dim %[[VAL_16]], %[[VAL_4]] : memref<?xindex>44// CHECK: %[[VAL_21:.*]], %[[VAL_22:.*]] = gpu.alloc async {{\[}}%[[VAL_19]]] (%[[VAL_20]]) : memref<?xindex>45// CHECK: %[[VAL_23:.*]] = gpu.memcpy async {{\[}}%[[VAL_22]]] %[[VAL_21]], %[[VAL_16]] : memref<?xindex>, memref<?xindex>46// CHECK: %[[VAL_24:.*]] = gpu.wait async47// CHECK: %[[VAL_25:.*]] = memref.dim %[[VAL_17]], %[[VAL_4]] : memref<?xindex>48// CHECK: %[[VAL_26:.*]], %[[VAL_27:.*]] = gpu.alloc async {{\[}}%[[VAL_24]]] (%[[VAL_25]]) : memref<?xindex>49// CHECK: %[[VAL_28:.*]] = gpu.memcpy async {{\[}}%[[VAL_27]]] %[[VAL_26]], %[[VAL_17]] : memref<?xindex>, memref<?xindex>50// CHECK: %[[VAL_29:.*]] = gpu.wait async51// CHECK: %[[VAL_30:.*]] = memref.dim %[[VAL_18]], %[[VAL_4]] : memref<?xf64>52// CHECK: %[[VAL_31:.*]], %[[VAL_32:.*]] = gpu.alloc async {{\[}}%[[VAL_29]]] (%[[VAL_30]]) : memref<?xf64>53// CHECK: %[[VAL_33:.*]] = gpu.memcpy async {{\[}}%[[VAL_32]]] %[[VAL_31]], %[[VAL_18]] : memref<?xf64>, memref<?xf64>54// CHECK: gpu.wait {{\[}}%[[VAL_10]], %[[VAL_15]], %[[VAL_23]], %[[VAL_28]], %[[VAL_33]]]55// CHECK: %[[VAL_34:.*]] = gpu.wait async56// CHECK: %[[VAL_37:.*]], %[[VAL_38:.*]] = gpu.create_dn_tensor async {{\[}}%[[VAL_34]]] %[[VAL_8]], %[[VAL_3]], %[[VAL_3]] : index, index into memref<8x8xf64>57// CHECK: %[[VAL_39:.*]], %[[VAL_40:.*]] = gpu.create_dn_tensor async {{\[}}%[[VAL_38]]] %[[VAL_13]], %[[VAL_3]], %[[VAL_3]] : index, index into memref<8x8xf64>58// CHECK: %[[VAL_41:.*]], %[[VAL_42:.*]] = gpu.create_csr async {{\[}}%[[VAL_40]]] %[[VAL_3]], %[[VAL_3]], %[[VAL_5]], %[[VAL_21]], %[[VAL_26]], %[[VAL_31]] : memref<?xindex>, memref<?xindex>, memref<?xf64>59// CHECK: %[[VAL_43:.*]], %[[VAL_44:.*]] = gpu.sddmm_buffer_size async {{\[}}%[[VAL_42]]] %[[VAL_37]], %[[VAL_39]], %[[VAL_41]] into f6460// CHECK: %[[VAL_45:.*]], %[[VAL_46:.*]] = gpu.alloc async {{\[}}%[[VAL_44]]] (%[[VAL_43]]) : memref<?xi8>61// CHECK: %[[VAL_47:.*]] = gpu.sddmm async {{\[}}%[[VAL_46]]] %[[VAL_37]], %[[VAL_39]], %[[VAL_41]], %[[VAL_45]] : memref<?xi8> into f6462// CHECK: %[[VAL_48:.*]] = gpu.destroy_dn_tensor async {{\[}}%[[VAL_47]]] %[[VAL_37]]63// CHECK: %[[VAL_49:.*]] = gpu.destroy_dn_tensor async {{\[}}%[[VAL_48]]] %[[VAL_39]]64// CHECK: %[[VAL_50:.*]] = gpu.destroy_sp_mat async {{\[}}%[[VAL_49]]] %[[VAL_41]]65// CHECK: %[[VAL_52:.*]] = gpu.dealloc async {{\[}}%[[VAL_50]]] %[[VAL_45]] : memref<?xi8>66// CHECK: %[[VAL_53:.*]] = gpu.dealloc async {{\[}}%[[VAL_52]]] %[[VAL_8]] : memref<8x8xf64>67// CHECK: %[[VAL_54:.*]] = gpu.dealloc async {{\[}}%[[VAL_53]]] %[[VAL_13]] : memref<8x8xf64>68// CHECK: %[[VAL_55:.*]] = gpu.dealloc async {{\[}}%[[VAL_54]]] %[[VAL_21]] : memref<?xindex>69// CHECK: %[[VAL_56:.*]] = gpu.dealloc async {{\[}}%[[VAL_55]]] %[[VAL_26]] : memref<?xindex>70// CHECK: %[[VAL_57:.*]] = gpu.memcpy async {{\[}}%[[VAL_56]]] %[[VAL_18]], %[[VAL_31]] : memref<?xf64>, memref<?xf64>71// CHECK: %[[VAL_58:.*]] = gpu.dealloc async {{\[}}%[[VAL_57]]] %[[VAL_31]] : memref<?xf64>72// CHECK: gpu.wait {{\[}}%[[VAL_58]]]73// CHECK: %[[VAL_59:.*]] = sparse_tensor.load %[[VAL_0]] : tensor<8x8xf64, #sparse{{[0-9]*}}>74// CHECK: return %[[VAL_59]] : tensor<8x8xf64, #sparse{{[0-9]*}}>75// CHECK: }76//77// A kernel that computes a direct sampled matrix matrix multiplication78// (with sparse result).79// Compute SDDMM C = C\spy AB80//81func.func @sparse_sampled_dd(%argS: tensor<8x8xf64, #CSR>,82 %argA: tensor<8x8xf64>,83 %argB: tensor<8x8xf64>) -> tensor<8x8xf64, #CSR> {84 %result = linalg.generic #trait_sampled_dense_dense85 ins(%argA, %argB: tensor<8x8xf64>, tensor<8x8xf64>)86 outs(%argS: tensor<8x8xf64, #CSR>) {87 ^bb(%a: f64, %b: f64, %s: f64):88 %f0 = arith.constant 0.0 : f6489 %u = sparse_tensor.unary %s : f64 to f6490 present={91 ^bb0(%p: f64):92 %mul = arith.mulf %a, %b : f6493 sparse_tensor.yield %mul : f6494 }95 absent={}96 %r = sparse_tensor.reduce %s, %u, %f0 : f64 {97 ^bb0(%p: f64, %q: f64):98 %add = arith.addf %p, %q : f6499 sparse_tensor.yield %add : f64100 }101 linalg.yield %r : f64102 } -> tensor<8x8xf64, #CSR>103 return %result : tensor<8x8xf64, #CSR>104}105