86 lines · plain
1// RUN: mlir-opt %s --linalg-generalize-named-ops \2// RUN: --pre-sparsification-rewrite \3// RUN: --sparse-reinterpret-map \4// RUN: --sparsification="parallelization-strategy=dense-outer-loop" \5// RUN: --sparse-gpu-codegen | FileCheck %s6 7#CSR = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : dense, d1 : compressed) }>8 9//10// Compute matrix matrix C = AB11//12// CHECK-LABEL: gpu.module @sparse_kernels13// CHECK-LABEL: gpu.func @kernel0(14// CHECK-SAME: %[[VAL_0:.*0]]: index,15// CHECK-SAME: %[[VAL_1:.*1]]: index,16// CHECK-SAME: %[[VAL_2:.*2]]: memref<?xindex>,17// CHECK-SAME: %[[VAL_3:.*3]]: memref<?xindex>,18// CHECK-SAME: %[[VAL_4:.*4]]: memref<?xf64>,19// CHECK-SAME: %[[VAL_5:.*5]]: memref<?x?xf64>,20// CHECK-SAME: %[[VAL_6:.*6]]: memref<?x?xf64>) kernel {21// CHECK-DAG: %[[VAL_7:.*]] = arith.constant 1 : index22// CHECK-DAG: %[[VAL_8:.*]] = arith.constant 0 : index23// CHECK: %[[VAL_9:.*]] = gpu.block_id x24// CHECK: %[[VAL_10:.*]] = gpu.block_dim x25// CHECK: %[[VAL_11:.*]] = gpu.thread_id x26// CHECK: %[[VAL_12:.*]] = gpu.grid_dim x27// CHECK: %[[VAL_13:.*]] = arith.muli %[[VAL_9]], %[[VAL_10]] : index28// CHECK: %[[VAL_14:.*]] = arith.addi %[[VAL_13]], %[[VAL_11]] : index29// CHECK: %[[VAL_15:.*]] = arith.muli %[[VAL_10]], %[[VAL_12]] : index30// CHECK: scf.for %[[VAL_16:.*]] = %[[VAL_14]] to %[[VAL_1]] step %[[VAL_15]] {31// CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_2]]{{\[}}%[[VAL_16]]] : memref<?xindex>32// CHECK: %[[VAL_18:.*]] = arith.addi %[[VAL_16]], %[[VAL_7]] : index33// CHECK: %[[VAL_19:.*]] = memref.load %[[VAL_2]]{{\[}}%[[VAL_18]]] : memref<?xindex>34// CHECK: scf.for %[[VAL_20:.*]] = %[[VAL_17]] to %[[VAL_19]] step %[[VAL_7]] {35// CHECK: %[[VAL_21:.*]] = memref.load %[[VAL_3]]{{\[}}%[[VAL_20]]] : memref<?xindex>36// CHECK: %[[VAL_22:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_20]]] : memref<?xf64>37// CHECK: scf.for %[[VAL_23:.*]] = %[[VAL_8]] to %[[VAL_0]] step %[[VAL_7]] {38// CHECK: %[[VAL_24:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_16]], %[[VAL_23]]] : memref<?x?xf64>39// CHECK: %[[VAL_25:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_21]], %[[VAL_23]]] : memref<?x?xf64>40// CHECK: %[[VAL_26:.*]] = arith.mulf %[[VAL_22]], %[[VAL_25]] : f6441// CHECK: %[[VAL_27:.*]] = arith.addf %[[VAL_24]], %[[VAL_26]] : f6442// CHECK: memref.store %[[VAL_27]], %[[VAL_5]]{{\[}}%[[VAL_16]], %[[VAL_23]]] : memref<?x?xf64>43// CHECK: } {"Emitted from" = "linalg.generic"}44// CHECK: } {"Emitted from" = "linalg.generic"}45// CHECK: }46// CHECK: gpu.return47// CHECK: }48//49//50// CHECK-LABEL: func.func @matmul51// CHECK: gpu.wait async52// CHECK: gpu.alloc async53// CHECK: %[[S0:.*]] = gpu.memcpy async54// CHECK: gpu.wait async55// CHECK: gpu.alloc async56// CHECK: %[[S1:.*]] = gpu.memcpy async57// CHECK: gpu.wait async58// CHECK: gpu.alloc async59// CHECK: %[[S2:.*]] = gpu.memcpy async60// CHECK: gpu.wait async61// CHECK: gpu.alloc async62// CHECK: %[[S3:.*]] = gpu.memcpy async63// CHECK: gpu.wait async64// CHECK: gpu.alloc async65// CHECK: %[[S4:.*]] = gpu.memcpy async66// CHECK: gpu.wait [%[[S0]], %[[S1]], %[[S2]], %[[S3]], %[[S4]]67// CHECK: %[[T0:.*]] = gpu.launch_func async @sparse_kernels::@kernel0 blocks68// CHECK: %[[M0:.*]] = gpu.memcpy async [%[[T0]]]69// CHECK: %[[M1:.*]] = gpu.dealloc async [%[[M0]]]70// CHECK: %[[M2:.*]] = gpu.wait async71// CHECK: %[[M3:.*]] = gpu.dealloc async [%[[M2]]]72// CHECK: %[[M4:.*]] = gpu.wait async73// CHECK: %[[M5:.*]] = gpu.dealloc async [%[[M4]]]74// CHECK: %[[M6:.*]] = gpu.wait async75// CHECK: %[[M7:.*]] = gpu.dealloc async [%[[M6]]]76// CHECK: %[[M8:.*]] = gpu.wait async77// CHECK: %[[M9:.*]] = gpu.dealloc async [%[[M8]]]78// CHECK: gpu.wait [%[[M1]], %[[M3]], %[[M5]], %[[M7]], %[[M9]]79//80func.func @matmul(%A: tensor<?x?xf64, #CSR>, %B: tensor<?x?xf64>, %C_in: tensor<?x?xf64>) -> tensor<?x?xf64> {81 %C_out = linalg.matmul82 ins(%A, %B: tensor<?x?xf64, #CSR>, tensor<?x?xf64>)83 outs(%C_in: tensor<?x?xf64>) -> tensor<?x?xf64>84 return %C_out : tensor<?x?xf64>85}86