105 lines · plain
1// RUN: mlir-opt %s --transform-interpreter --split-input-file | FileCheck %s2 3func.func @matmul(%lhs: memref<32x32xf32>, %rhs: memref<32x32xf32>, %out: memref<32x32xf32>) {4 %c8 = arith.constant 8 : index5 %c0 = arith.constant 0 : index6 %cst = arith.constant dense<0.000000e+00> : vector<16x16xf32>7 %c16 = arith.constant 16 : index8 %c32 = arith.constant 32 : index9 %cst_0 = arith.constant 0.000000e+00 : f3210 %3 = gpu.thread_id x11 %4 = gpu.thread_id y12 %5 = affine.apply affine_map<()[s0] -> (s0 * 16)>()[%4]13 %6 = affine.apply affine_map<()[s0] -> ((s0 floordiv 32) * 16)>()[%3]14 // CHECK: scf.for {{.*}} -> (vector<16x16xf32>) {15 // CHECK-COUNT-2: vector.transfer_read {{.*}} vector<16x8xf32>16 // CHECK-COUNT-2: vector.transfer_read {{.*}} vector<8x16xf32>17 // CHECK-COUNT-2: vector.contract {{.*}} vector<16x8xf32>, vector<8x16xf32> into vector<16x16xf32>18 // CHECK: scf.yield {{.*}} : vector<16x16xf32>19 // CHECK: }20 %7 = scf.for %arg0 = %c0 to %c32 step %c16 iter_args(%arg1 = %cst) -> (vector<16x16xf32>) {21 %10 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%5]22 %11 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%arg0]23 %12 = vector.transfer_read %lhs[%10, %11], %cst_0 {in_bounds = [true, true]} : memref<32x32xf32>, vector<16x16xf32>24 %16 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%6]25 %17 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%arg0]26 %18 = vector.transfer_read %rhs[%17, %16], %cst_0 {in_bounds = [true, true]} : memref<32x32xf32>, vector<16x16xf32>27 %22 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %12, %18, %arg1 : vector<16x16xf32>, vector<16x16xf32> into vector<16x16xf32>28 scf.yield %22 : vector<16x16xf32>29 }30 %8 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%5]31 %9 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%6]32 vector.transfer_write %7, %out[%8, %9] {in_bounds = [true, true]} : vector<16x16xf32>, memref<32x32xf32>33 return34}35 36module attributes {transform.with_named_sequence} {37 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {38 %func_op = transform.structured.match ops{["func.func"]} in %arg0 : (!transform.any_op) -> !transform.op<"func.func">39 transform.apply_patterns to %func_op {40 transform.apply_patterns.gpu.unroll_vectors_subgroup_mma [16, 16, 8]41 } : !transform.op<"func.func">42 transform.yield43 }44}45 46// -----47 48// CHECK-LABEL: func.func @gathered_matmul49func.func @gathered_matmul(%lhs: memref<32x32xf32>, %rhs: memref<32x32xf32>, %out: memref<32x32xf32>) {50 %c8 = arith.constant 8 : index51 %c0 = arith.constant 0 : index52 %cst = arith.constant dense<0.000000e+00> : vector<16x16xf32>53 %cst_mask = arith.constant dense<true> : vector<4x4xi1>54 %cst_pt = arith.constant dense<0.000000e+00> : vector<4x4xf32>55 %c16 = arith.constant 16 : index56 %c32 = arith.constant 32 : index57 %cst_0 = arith.constant 0.000000e+00 : f3258 %cst_1 = arith.constant dense<[0, 1, 2, 3]> : vector<4xindex>59 %cst_2 = arith.constant dense<1> : vector<4x4xindex>60 %alloc = memref.alloc() {alignment = 64 : i64} : memref<32x32xf32>61 %3 = gpu.thread_id x62 %4 = gpu.thread_id y63 %5 = affine.apply affine_map<()[s0] -> (s0 * 16)>()[%4]64 %6 = affine.apply affine_map<()[s0] -> ((s0 floordiv 32) * 16)>()[%3]65 // CHECK: scf.for {{.*}} -> (vector<16x16xf32>) {66 // CHECK: arith.addi {{.*}} : vector<4xindex>67 // CHECK: vector.gather {{.*}} : memref<32x32xf32>, vector<4x4xindex>, vector<4x4xi1>, vector<4x4xf32> into vector<4x4xf32>68 // CHECK-COUNT-8: vector.transfer_read {{.*}} vector<8x4xf32>69 // CHECK-COUNT-4: vector.transfer_read {{.*}} vector<4x16xf32>70 // CHECK-COUNT-8: vector.contract {{.*}} vector<8x4xf32>, vector<4x16xf32> into vector<8x16xf32>71 // CHECK: scf.yield {{.*}} : vector<16x16xf32>72 // CHECK: }73 %7 = scf.for %arg0 = %c0 to %c32 step %c16 iter_args(%arg1 = %cst) -> (vector<16x16xf32>) {74 %10 = vector.broadcast %arg0 : index to vector<4xindex>75 %11 = arith.addi %10, %cst_1 : vector<4xindex>76 %12 = vector.broadcast %11 : vector<4xindex> to vector<4x4xindex>77 %13 = arith.addi %12, %cst_2 : vector<4x4xindex>78 %14 = vector.gather %lhs[%c0, %c0] [%13], %cst_mask, %cst_pt : memref<32x32xf32>, vector<4x4xindex>, vector<4x4xi1>, vector<4x4xf32> into vector<4x4xf32>79 vector.transfer_write %14, %alloc[%c0, %c0] {in_bounds = [true, true]} : vector<4x4xf32>, memref<32x32xf32>80 gpu.barrier81 %15 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%5]82 %16 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%arg0]83 %17 = vector.transfer_read %alloc[%15, %16], %cst_0 {in_bounds = [true, true]} : memref<32x32xf32>, vector<16x16xf32>84 %18 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%6]85 %19 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%arg0]86 %20 = vector.transfer_read %rhs[%19, %18], %cst_0 {in_bounds = [true, true]} : memref<32x32xf32>, vector<16x16xf32>87 %21 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %17, %20, %arg1 : vector<16x16xf32>, vector<16x16xf32> into vector<16x16xf32>88 scf.yield %21 : vector<16x16xf32>89 }90 %8 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%5]91 %9 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%c0)[%6]92 vector.transfer_write %7, %out[%8, %9] {in_bounds = [true, true]} : vector<16x16xf32>, memref<32x32xf32>93 return94}95 96module attributes {transform.with_named_sequence} {97 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {98 %func_op = transform.structured.match ops{["func.func"]} in %arg0 : (!transform.any_op) -> !transform.op<"func.func">99 transform.apply_patterns to %func_op {100 transform.apply_patterns.gpu.unroll_vectors_subgroup_mma [8, 16, 4]101 } : !transform.op<"func.func">102 transform.yield103 }104}105