238 lines · plain
1// RUN: mlir-opt %s -convert-vector-to-llvm="enable-x86vector" | mlir-opt | FileCheck %s2 3// CHECK-LABEL: func @avx512_mask_rndscale4func.func @avx512_mask_rndscale(5 %src: vector<16xf32>, %a: vector<16xf32>, %b: vector<8xf64>,6 %imm_i16: i16, %imm_i8: i8, %scale_k_i16: i16, %scale_k_i8: i8)7 -> (vector<16xf32>, vector<8xf64>, vector<16xf32>, vector<8xf64>)8{9 %rnd_k = arith.constant 15 : i3210 %rnd = arith.constant 42 : i3211 // CHECK: llvm.call_intrinsic "llvm.x86.avx512.mask.rndscale.ps.512"12 %0 = x86vector.avx512.mask.rndscale %src, %rnd_k, %a, %imm_i16, %rnd : vector<16xf32>13 // CHECK: llvm.call_intrinsic "llvm.x86.avx512.mask.rndscale.pd.512"14 %1 = x86vector.avx512.mask.rndscale %b, %rnd_k, %b, %imm_i8, %rnd : vector<8xf64>15 16 // CHECK: llvm.call_intrinsic "llvm.x86.avx512.mask.scalef.ps.512"17 %2 = x86vector.avx512.mask.scalef %a, %a, %a, %scale_k_i16, %rnd : vector<16xf32>18 // CHECK: llvm.call_intrinsic "llvm.x86.avx512.mask.scalef.pd.512"19 %3 = x86vector.avx512.mask.scalef %b, %b, %b, %scale_k_i8, %rnd : vector<8xf64>20 21 // Keep results alive.22 return %0, %1, %2, %3 : vector<16xf32>, vector<8xf64>, vector<16xf32>, vector<8xf64>23}24 25// CHECK-LABEL: func @avx512_mask_compress26func.func @avx512_mask_compress(27 %k1: vector<16xi1>, %a1: vector<16xf32>, %k2: vector<8xi1>, %a2: vector<8xi64>)28 -> (vector<16xf32>, vector<16xf32>, vector<8xi64>)29{30 // CHECK: llvm.mlir.constant(dense<0.000000e+00> : vector<16xf32>)31 // CHECK: llvm.call_intrinsic "llvm.x86.avx512.mask.compress"32 %0 = x86vector.avx512.mask.compress %k1, %a1 : vector<16xf32>33 // CHECK: llvm.mlir.constant(dense<5.000000e+00> : vector<16xf32>)34 // CHECK: llvm.call_intrinsic "llvm.x86.avx512.mask.compress"35 %1 = x86vector.avx512.mask.compress %k1, %a136 {constant_src = dense<5.0> : vector<16xf32>} : vector<16xf32>37 // CHECK: llvm.call_intrinsic "llvm.x86.avx512.mask.compress"38 %2 = x86vector.avx512.mask.compress %k2, %a2, %a2 : vector<8xi64>, vector<8xi64>39 return %0, %1, %2 : vector<16xf32>, vector<16xf32>, vector<8xi64>40}41 42// CHECK-LABEL: func @avx512_vp2intersect43func.func @avx512_vp2intersect(%a: vector<16xi32>, %b: vector<8xi64>)44 -> (vector<16xi1>, vector<16xi1>, vector<8xi1>, vector<8xi1>)45{46 // CHECK: llvm.call_intrinsic "llvm.x86.avx512.vp2intersect.d.512"47 %0, %1 = x86vector.avx512.vp2intersect %a, %a : vector<16xi32>48 // CHECK: llvm.call_intrinsic "llvm.x86.avx512.vp2intersect.q.512"49 %2, %3 = x86vector.avx512.vp2intersect %b, %b : vector<8xi64>50 return %0, %1, %2, %3 : vector<16xi1>, vector<16xi1>, vector<8xi1>, vector<8xi1>51}52 53// CHECK-LABEL: func @avx512bf16_dot_12854func.func @avx512bf16_dot_128(%src: vector<4xf32>, %a: vector<8xbf16>,55 %b: vector<8xbf16>) -> (vector<4xf32>)56{57 // CHECK: llvm.call_intrinsic "llvm.x86.avx512bf16.dpbf16ps.128"58 %0 = x86vector.avx512.dot %src, %a, %b : vector<8xbf16> -> vector<4xf32>59 return %0 : vector<4xf32>60}61 62// CHECK-LABEL: func @avx512bf16_dot_25663func.func @avx512bf16_dot_256(%src: vector<8xf32>, %a: vector<16xbf16>,64 %b: vector<16xbf16>) -> (vector<8xf32>)65{66 // CHECK: llvm.call_intrinsic "llvm.x86.avx512bf16.dpbf16ps.256"67 %0 = x86vector.avx512.dot %src, %a, %b : vector<16xbf16> -> vector<8xf32>68 return %0 : vector<8xf32>69}70 71// CHECK-LABEL: func @avx512bf16_dot_51272func.func @avx512bf16_dot_512(%src: vector<16xf32>, %a: vector<32xbf16>,73 %b: vector<32xbf16>) -> (vector<16xf32>)74{75 // CHECK: llvm.call_intrinsic "llvm.x86.avx512bf16.dpbf16ps.512"76 %0 = x86vector.avx512.dot %src, %a, %b : vector<32xbf16> -> vector<16xf32>77 return %0 : vector<16xf32>78}79 80// CHECK-LABEL: func @avx512bf16_cvt_packed_f32_to_bf16_25681func.func @avx512bf16_cvt_packed_f32_to_bf16_256(82 %a: vector<8xf32>) -> (vector<8xbf16>)83{84 // CHECK: llvm.call_intrinsic "llvm.x86.avx512bf16.cvtneps2bf16.256"85 %0 = x86vector.avx512.cvt.packed.f32_to_bf16 %a : vector<8xf32> -> vector<8xbf16>86 return %0 : vector<8xbf16>87}88 89// CHECK-LABEL: func @avx512bf16_cvt_packed_f32_to_bf16_51290func.func @avx512bf16_cvt_packed_f32_to_bf16_512(91 %a: vector<16xf32>) -> (vector<16xbf16>)92{93 // CHECK: llvm.call_intrinsic "llvm.x86.avx512bf16.cvtneps2bf16.512"94 %0 = x86vector.avx512.cvt.packed.f32_to_bf16 %a : vector<16xf32> -> vector<16xbf16>95 return %0 : vector<16xbf16>96}97 98// CHECK-LABEL: func @avxbf16_cvt_packed_even_indexed_bf16_to_f32_12899func.func @avxbf16_cvt_packed_even_indexed_bf16_to_f32_128(100 %a: memref<8xbf16>) -> vector<4xf32>101{102 // CHECK: llvm.call_intrinsic "llvm.x86.vcvtneebf162ps128"103 %0 = x86vector.avx.cvt.packed.even.indexed_to_f32 %a : memref<8xbf16> -> vector<4xf32>104 return %0 : vector<4xf32>105}106 107// CHECK-LABEL: func @avxbf16_cvt_packed_even_indexed_bf16_to_f32_256108func.func @avxbf16_cvt_packed_even_indexed_bf16_to_f32_256(109 %a: memref<16xbf16>) -> vector<8xf32>110{111 // CHECK: llvm.call_intrinsic "llvm.x86.vcvtneebf162ps256"112 %0 = x86vector.avx.cvt.packed.even.indexed_to_f32 %a : memref<16xbf16> -> vector<8xf32>113 return %0 : vector<8xf32>114}115 116// CHECK-LABEL: func @avxbf16_cvt_packed_odd_indexed_bf16_to_f32_128117func.func @avxbf16_cvt_packed_odd_indexed_bf16_to_f32_128(118 %a: memref<8xbf16>) -> vector<4xf32>119{120 // CHECK: llvm.call_intrinsic "llvm.x86.vcvtneobf162ps128"121 %0 = x86vector.avx.cvt.packed.odd.indexed_to_f32 %a : memref<8xbf16> -> vector<4xf32>122 return %0 : vector<4xf32>123}124 125// CHECK-LABEL: func @avxbf16_cvt_packed_odd_indexed_bf16_to_f32_256126func.func @avxbf16_cvt_packed_odd_indexed_bf16_to_f32_256(127 %a: memref<16xbf16>) -> vector<8xf32>128{129 // CHECK: llvm.call_intrinsic "llvm.x86.vcvtneobf162ps256"130 %0 = x86vector.avx.cvt.packed.odd.indexed_to_f32 %a : memref<16xbf16> -> vector<8xf32>131 return %0 : vector<8xf32>132}133 134// CHECK-LABEL: func @avxbf16_bsct_bf16_to_f32_packed_128135func.func @avxbf16_bsct_bf16_to_f32_packed_128(136 %a: memref<1xbf16>) -> vector<4xf32>137{138 // CHECK: llvm.call_intrinsic "llvm.x86.vbcstnebf162ps128"139 %0 = x86vector.avx.bcst_to_f32.packed %a : memref<1xbf16> -> vector<4xf32>140 return %0 : vector<4xf32>141}142 143// CHECK-LABEL: func @avxbf16_bsct_bf16_to_f32_packed_256144func.func @avxbf16_bsct_bf16_to_f32_packed_256(145 %a: memref<1xbf16>) -> vector<8xf32>146{147 // CHECK: llvm.call_intrinsic "llvm.x86.vbcstnebf162ps256"148 %0 = x86vector.avx.bcst_to_f32.packed %a : memref<1xbf16> -> vector<8xf32>149 return %0 : vector<8xf32>150}151 152// CHECK-LABEL: func @avxf16_cvt_packed_even_indexed_f16_to_f32_128153func.func @avxf16_cvt_packed_even_indexed_f16_to_f32_128(154 %a: memref<8xf16>) -> vector<4xf32>155{156 // CHECK: llvm.call_intrinsic "llvm.x86.vcvtneeph2ps128"157 %0 = x86vector.avx.cvt.packed.even.indexed_to_f32 %a : memref<8xf16> -> vector<4xf32>158 return %0 : vector<4xf32>159}160 161// CHECK-LABEL: func @avxf16_cvt_packed_even_indexed_f16_to_f32_256162func.func @avxf16_cvt_packed_even_indexed_f16_to_f32_256(163 %a: memref<16xf16>) -> vector<8xf32>164{165 // CHECK: llvm.call_intrinsic "llvm.x86.vcvtneeph2ps256"166 %0 = x86vector.avx.cvt.packed.even.indexed_to_f32 %a : memref<16xf16> -> vector<8xf32>167 return %0 : vector<8xf32>168}169 170// CHECK-LABEL: func @avxf16_cvt_packed_odd_indexed_f16_to_f32_128171func.func @avxf16_cvt_packed_odd_indexed_f16_to_f32_128(172 %a: memref<8xf16>) -> vector<4xf32>173{174 // CHECK: llvm.call_intrinsic "llvm.x86.vcvtneoph2ps128"175 %0 = x86vector.avx.cvt.packed.odd.indexed_to_f32 %a : memref<8xf16> -> vector<4xf32>176 return %0 : vector<4xf32>177}178 179// CHECK-LABEL: func @avxf16_cvt_packed_odd_indexed_f16_to_f32_256180func.func @avxf16_cvt_packed_odd_indexed_f16_to_f32_256(181 %a: memref<16xf16>) -> vector<8xf32>182{183 // CHECK: llvm.call_intrinsic "llvm.x86.vcvtneoph2ps256"184 %0 = x86vector.avx.cvt.packed.odd.indexed_to_f32 %a : memref<16xf16> -> vector<8xf32>185 return %0 : vector<8xf32>186}187 188// CHECK-LABEL: func @avxf16_bsct_f16_to_f32_packed_128189func.func @avxf16_bsct_f16_to_f32_packed_128(190 %a: memref<1xf16>) -> vector<4xf32>191{192 // CHECK: llvm.call_intrinsic "llvm.x86.vbcstnesh2ps128"193 %0 = x86vector.avx.bcst_to_f32.packed %a : memref<1xf16> -> vector<4xf32>194 return %0 : vector<4xf32>195}196 197// CHECK-LABEL: func @avxf16_bsct_f16_to_f32_packed_256198func.func @avxf16_bsct_f16_to_f32_packed_256(199 %a: memref<1xf16>) -> vector<8xf32>200{201 // CHECK: llvm.call_intrinsic "llvm.x86.vbcstnesh2ps256"202 %0 = x86vector.avx.bcst_to_f32.packed %a : memref<1xf16> -> vector<8xf32>203 return %0 : vector<8xf32>204}205 206// CHECK-LABEL: func @avx_rsqrt207func.func @avx_rsqrt(%a: vector<8xf32>) -> (vector<8xf32>)208{209 // CHECK: llvm.call_intrinsic "llvm.x86.avx.rsqrt.ps.256"210 %0 = x86vector.avx.rsqrt %a : vector<8xf32>211 return %0 : vector<8xf32>212}213 214// CHECK-LABEL: func @avx_dot215func.func @avx_dot(%a: vector<8xf32>, %b: vector<8xf32>) -> (vector<8xf32>)216{217 // CHECK: llvm.mlir.constant(-1 : i8)218 // CHECK: llvm.call_intrinsic "llvm.x86.avx.dp.ps.256"219 %0 = x86vector.avx.intr.dot %a, %b : vector<8xf32>220 return %0 : vector<8xf32>221}222 223// CHECK-LABEL: func @avx_dot_i8_128224func.func @avx_dot_i8_128(%w: vector<4xi32>, %a: vector<16xi8>,225 %b: vector<16xi8>) -> vector<4xi32> {226 // CHECK: llvm.call_intrinsic "llvm.x86.avx2.vpdpbssd.128"227 %0 = x86vector.avx.dot.i8 %w, %a, %b : vector<16xi8> -> vector<4xi32>228 return %0 : vector<4xi32>229}230 231// CHECK-LABEL: func @avx_dot_i8_256232func.func @avx_dot_i8_256(%w: vector<8xi32>, %a: vector<32xi8>,233 %b: vector<32xi8>) -> vector<8xi32> {234 // CHECK: llvm.call_intrinsic "llvm.x86.avx2.vpdpbssd.256"235 %0 = x86vector.avx.dot.i8 %w, %a, %b : vector<32xi8> -> vector<8xi32>236 return %0 : vector<8xi32>237}238