88 lines · plain
1// RUN: mlir-translate -mlir-to-llvmir %s | FileCheck %s2 3// CHECK-LABEL: @convert_f32x2_to_f16x2_rn4llvm.func @convert_f32x2_to_f16x2_rn(%srcA : f32, %srcB : f32) {5 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rn(float %{{.*}}, float %{{.*}})6 %res1 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rn>} : vector<2xf16>7 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rn.satfinite(float %{{.*}}, float %{{.*}})8 %res2 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rn>, sat = #nvvm.sat_mode<satfinite>} : vector<2xf16>9 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rn.relu(float %{{.*}}, float %{{.*}})10 %res3 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rn>, relu = true} : vector<2xf16>11 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rn.relu.satfinite(float %{{.*}}, float %{{.*}})12 %res4 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rn>, relu = true, sat = #nvvm.sat_mode<satfinite>} : vector<2xf16>13 14 llvm.return15}16 17// CHECK-LABEL: @convert_f32x2_to_f16x2_rz18llvm.func @convert_f32x2_to_f16x2_rz(%srcA : f32, %srcB : f32) {19 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rz(float %{{.*}}, float %{{.*}})20 %res1 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rz>} : vector<2xf16>21 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rz.satfinite(float %{{.*}}, float %{{.*}})22 %res2 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rz>, sat = #nvvm.sat_mode<satfinite>} : vector<2xf16>23 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rz.relu(float %{{.*}}, float %{{.*}})24 %res3 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rz>, relu = true} : vector<2xf16>25 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rz.relu.satfinite(float %{{.*}}, float %{{.*}})26 %res4 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rz>, relu = true, sat = #nvvm.sat_mode<satfinite>} : vector<2xf16>27 28 llvm.return29}30 31// CHECK-LABEL: @convert_f32x2_to_f16x2_rs_stochastic32llvm.func @convert_f32x2_to_f16x2_rs_stochastic(%srcA : f32, %srcB : f32, %rbits : i32) {33 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rs(float %{{.*}}, float %{{.*}}, i32 %{{.*}})34 %res1 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB, %rbits {rnd = #nvvm.fp_rnd_mode<rs>} : vector<2xf16>35 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rs.relu(float %{{.*}}, float %{{.*}}, i32 %{{.*}})36 %res2 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB, %rbits {relu = true, rnd = #nvvm.fp_rnd_mode<rs>} : vector<2xf16>37 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rs.satfinite(float %{{.*}}, float %{{.*}}, i32 %{{.*}})38 %res3 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB, %rbits {rnd = #nvvm.fp_rnd_mode<rs>, sat = #nvvm.sat_mode<satfinite>} : vector<2xf16>39 // CHECK: %{{.*}} = call <2 x half> @llvm.nvvm.ff2f16x2.rs.relu.satfinite(float %{{.*}}, float %{{.*}}, i32 %{{.*}})40 %res4 = nvvm.convert.f32x2.to.f16x2 %srcA, %srcB, %rbits {relu = true, rnd = #nvvm.fp_rnd_mode<rs>, sat = #nvvm.sat_mode<satfinite>} : vector<2xf16>41 42 llvm.return43}44 45// -----46 47// CHECK-LABEL: @convert_f32x2_to_bf16x2_rn48llvm.func @convert_f32x2_to_bf16x2_rn(%srcA : f32, %srcB : f32) {49 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rn(float %{{.*}}, float %{{.*}})50 %res1 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rn>} : vector<2xbf16>51 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rn.satfinite(float %{{.*}}, float %{{.*}})52 %res2 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rn>, sat = #nvvm.sat_mode<satfinite>} : vector<2xbf16>53 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rn.relu(float %{{.*}}, float %{{.*}})54 %res3 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rn>, relu = true} : vector<2xbf16>55 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rn.relu.satfinite(float %{{.*}}, float %{{.*}})56 %res4 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rn>, relu = true, sat = #nvvm.sat_mode<satfinite>} : vector<2xbf16>57 58 llvm.return59}60 61// CHECK-LABEL: @convert_f32x2_to_bf16x2_rz62llvm.func @convert_f32x2_to_bf16x2_rz(%srcA : f32, %srcB : f32) {63 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rz(float %{{.*}}, float %{{.*}})64 %res1 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rz>} : vector<2xbf16>65 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rz.satfinite(float %{{.*}}, float %{{.*}})66 %res2 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rz>, sat = #nvvm.sat_mode<satfinite>} : vector<2xbf16>67 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rz.relu(float %{{.*}}, float %{{.*}})68 %res3 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rz>, relu = true} : vector<2xbf16>69 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rz.relu.satfinite(float %{{.*}}, float %{{.*}})70 %res4 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB {rnd = #nvvm.fp_rnd_mode<rz>, relu = true, sat = #nvvm.sat_mode<satfinite>} : vector<2xbf16>71 72 llvm.return73}74 75// CHECK-LABEL: @convert_f32x2_to_bf16x2_rs_stochastic76llvm.func @convert_f32x2_to_bf16x2_rs_stochastic(%srcA : f32, %srcB : f32, %rbits : i32) {77 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rs(float %{{.*}}, float %{{.*}}, i32 %{{.*}})78 %res1 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB, %rbits {rnd = #nvvm.fp_rnd_mode<rs>} : vector<2xbf16>79 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rs.relu(float %{{.*}}, float %{{.*}}, i32 %{{.*}})80 %res2 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB, %rbits {relu = true, rnd = #nvvm.fp_rnd_mode<rs>} : vector<2xbf16>81 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rs.satfinite(float %{{.*}}, float %{{.*}}, i32 %{{.*}})82 %res3 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB, %rbits {rnd = #nvvm.fp_rnd_mode<rs>, sat = #nvvm.sat_mode<satfinite>} : vector<2xbf16>83 // CHECK: %{{.*}} = call <2 x bfloat> @llvm.nvvm.ff2bf16x2.rs.relu.satfinite(float %{{.*}}, float %{{.*}}, i32 %{{.*}})84 %res4 = nvvm.convert.f32x2.to.bf16x2 %srcA, %srcB, %rbits {relu = true, rnd = #nvvm.fp_rnd_mode<rs>, sat = #nvvm.sat_mode<satfinite>} : vector<2xbf16>85 86 llvm.return87}88