| // RUN: iree-opt --pass-pipeline='builtin.module(iree-hal-assign-target-devices{targetBackends=rocm},iree-hal-transformation-pipeline{serialize-executables=false})' --iree-rocm-target-chip=gfx940 %s | FileCheck %s --check-prefix=MI300 |
| // RUN: iree-opt --pass-pipeline='builtin.module(iree-hal-assign-target-devices{targetBackends=rocm},iree-hal-transformation-pipeline{serialize-executables=false})' --iree-rocm-target-chip=gfx942 %s | FileCheck %s --check-prefix=MI300 |
| // RUN: iree-opt --pass-pipeline='builtin.module(iree-hal-assign-target-devices{targetBackends=rocm},iree-hal-transformation-pipeline{serialize-executables=false})' --iree-rocm-target-chip=gfx1100 %s | FileCheck %s --check-prefix=RDNA3 |
| |
| // MI300: mma_intrinsics = [#iree_gpu.mma_layout<MFMA_F16_16x16x16_F32>, #iree_gpu.mma_layout<MFMA_F16_32x32x8_F32>] |
| // RDNA3: mma_intrinsics = [#iree_gpu.mma_layout<WMMA_F16_16x16x16_F32>] |
| |
| stream.executable public @reduce_dispatch { |
| stream.executable.export @reduce_dispatch workgroups(%arg0: index) -> (index, index, index) { |
| %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0 |
| stream.return %x, %y, %z : index, index, index |
| } |
| builtin.module { |
| func.func @reduce_dispatch(%arg0_binding: !stream.binding, %arg1_binding: !stream.binding) { |
| %c0 = arith.constant 0 : index |
| %arg0 = stream.binding.subspan %arg0_binding[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<16xf32>> |
| %arg1 = stream.binding.subspan %arg1_binding[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<f32>> |
| %0 = tensor.empty() : tensor<f32> |
| %1 = flow.dispatch.tensor.load %arg0, offsets=[0], sizes=[16], strides=[1] : !flow.dispatch.tensor<readonly:tensor<16xf32>> -> tensor<16xf32> |
| %3 = linalg.generic {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> ()>], iterator_types = ["reduction"]} ins(%1 : tensor<16xf32>) outs(%0 : tensor<f32>) { |
| ^bb0(%arg2: f32, %arg3: f32): |
| %4 = arith.addf %arg2, %arg3 : f32 |
| linalg.yield %4 : f32 |
| } -> tensor<f32> |
| flow.dispatch.tensor.store %3, %arg1, offsets=[], sizes=[], strides=[] : tensor<f32> -> !flow.dispatch.tensor<writeonly:tensor<f32>> |
| return |
| } |
| } |
| } |
