1// RUN: mlir-opt --split-input-file %s \ 2// RUN: --pass-pipeline='builtin.module(func.func(convert-arith-to-amdgpu{chipset=gfx940 saturate-fp8-truncf=true}))' \ 3// RUN: | FileCheck %s 4 5// CHECK-LABEL: func.func @scalar_trunc 6// CHECK-SAME: ([[V:%.+]]: f16) 7// CHECK-DAG: [[CMin:%.+]] = arith.constant -5.734400e+04 : f16 8// CHECK-DAG: [[CMax:%.+]] = arith.constant 5.734400e+04 : f16 9// CHECK-DAG: [[CInf:%.+]] = arith.constant 0x7C00 : f16 10// CHECK-DAG: [[CNegInf:%.+]] = arith.constant 0xFC00 : f16 11// CHECK: [[ISINF:%.+]] = arith.cmpf oeq, [[V]], [[CInf]] 12// CHECK: [[ISNEGINF:%.+]] = arith.cmpf oeq, [[V]], [[CNegInf]] 13// CHECK: [[ISNAN:%.+]] = arith.cmpf uno, [[V]], [[V]] 14// CHECK: [[ISNONFINITE_1:%.+]] = arith.ori [[ISINF]], [[ISNEGINF]] 15// CHECK: [[ISNONFINITE:%.+]] = arith.ori [[ISNONFINITE_1]], [[ISNAN]] 16// CHECK: [[CLAMPEDBELOW:%.+]] = arith.maximumf [[V]], [[CMin]] 17// CHECK: [[CLAMPED:%.+]] = arith.minimumf [[CLAMPEDBELOW]], [[CMax]] 18// CHECK: [[SATURATED:%.+]] = arith.select [[ISNONFINITE]], [[V]], [[CLAMPED]] 19// CHECK: [[FLOAT:%.+]] = arith.extf [[SATURATED]] : f16 to f32 20// CHECK: [[TRUNCV:%.+]] = amdgpu.packed_trunc_2xfp8 [[FLOAT]], undef into undef[word 0] : f32 to vector<4xf8E5M2FNUZ> 21// CHECK: [[W:%.+]] = vector.extract [[TRUNCV]][0] : f8E5M2FNUZ from vector<4xf8E5M2FNUZ> 22// CHECK: return [[W]] : f8E5M2FNUZ 23func.func @scalar_trunc(%v: f16) -> f8E5M2FNUZ { 24 %w = arith.truncf %v : f16 to f8E5M2FNUZ 25 return %w : f8E5M2FNUZ 26} 27 28// No 0-D test because arith.truncf hasn't been extended to support it. 29 30// ----- 31 32// CHECK-LABEL: func.func @vector_trunc 33// CHECK-SAME: ([[V:%.+]]: vector<2xf32>) -> vector<2xf8E4M3FNUZ> { 34// CHECK-DAG: [[CMin:%.+]] = arith.constant dense<-2.400000e+02> : vector<2xf32> 35// CHECK-DAG: [[CMax:%.+]] = arith.constant dense<2.400000e+02> : vector<2xf32> 36// CHECK-DAG: [[CInf:%.+]] = arith.constant dense<0x7F800000> : vector<2xf32> 37// CHECK-DAG: [[CNegInf:%.+]] = arith.constant dense<0xFF800000> : vector<2xf32> 38// CHECK: [[ISINF:%.+]] = arith.cmpf oeq, [[V]], [[CInf]] 39// CHECK: [[ISNEGINF:%.+]] = arith.cmpf oeq, [[V]], [[CNegInf]] 40// CHECK: [[ISNAN:%.+]] = arith.cmpf uno, [[V]], [[V]] 41// CHECK: [[ISNONFINITE_1:%.+]] = arith.ori [[ISINF]], [[ISNEGINF]] 42// CHECK: [[ISNONFINITE:%.+]] = arith.ori [[ISNONFINITE_1]], [[ISNAN]] 43// CHECK: [[CLAMPEDBELOW:%.+]] = arith.maximumf [[V]], [[CMin]] 44// CHECK: [[CLAMPED:%.+]] = arith.minimumf [[CLAMPEDBELOW]], [[CMax]] 45// CHECK: [[SATURATED:%.+]] = arith.select [[ISNONFINITE]], [[V]], [[CLAMPED]] 46// CHECK: [[F0:%.+]] = vector.extract [[SATURATED]][0] 47// CHECK: [[F1:%.+]] = vector.extract [[SATURATED]][1] 48// CHECK: [[W0:%.+]] = amdgpu.packed_trunc_2xfp8 [[F0]], [[F1]] into undef[word 0] : f32 to vector<4xf8E4M3FNUZ> 49// CHECK: [[W:%.+]] = vector.extract_strided_slice [[W0]] {offsets = [0], sizes = [2], strides = [1]} : vector<4xf8E4M3FNUZ> to vector<2xf8E4M3FNUZ> 50// CHECK: return [[W]] : vector<2xf8E4M3FNUZ> 51func.func @vector_trunc_short(%v: vector<2xf32>) -> vector<2xf8E4M3FNUZ> { 52 %w = arith.truncf %v : vector<2xf32> to vector<2xf8E4M3FNUZ> 53 return %w : vector<2xf8E4M3FNUZ> 54} 55