248 lines · plain
1// RUN: mlir-opt -verify-diagnostics %s | mlir-opt | FileCheck %s2 3// CHECK-LABEL: func @avx512_mask_rndscale4func.func @avx512_mask_rndscale(%a: vector<16xf32>, %b: vector<8xf64>, %i32: i32, %i16: i16, %i8: i8)5 -> (vector<16xf32>, vector<8xf64>)6{7 // CHECK: x86vector.avx512.mask.rndscale {{.*}}: vector<16xf32>8 %0 = x86vector.avx512.mask.rndscale %a, %i32, %a, %i16, %i32 : vector<16xf32>9 // CHECK: x86vector.avx512.mask.rndscale {{.*}}: vector<8xf64>10 %1 = x86vector.avx512.mask.rndscale %b, %i32, %b, %i8, %i32 : vector<8xf64>11 return %0, %1: vector<16xf32>, vector<8xf64>12}13 14// CHECK-LABEL: func @avx512_scalef15func.func @avx512_scalef(%a: vector<16xf32>, %b: vector<8xf64>, %i32: i32, %i16: i16, %i8: i8)16 -> (vector<16xf32>, vector<8xf64>)17{18 // CHECK: x86vector.avx512.mask.scalef {{.*}}: vector<16xf32>19 %0 = x86vector.avx512.mask.scalef %a, %a, %a, %i16, %i32: vector<16xf32>20 // CHECK: x86vector.avx512.mask.scalef {{.*}}: vector<8xf64>21 %1 = x86vector.avx512.mask.scalef %b, %b, %b, %i8, %i32 : vector<8xf64>22 return %0, %1: vector<16xf32>, vector<8xf64>23}24 25// CHECK-LABEL: func @avx512_vp2intersect26func.func @avx512_vp2intersect(%a: vector<16xi32>, %b: vector<8xi64>)27 -> (vector<16xi1>, vector<16xi1>, vector<8xi1>, vector<8xi1>)28{29 // CHECK: x86vector.avx512.vp2intersect {{.*}} : vector<16xi32>30 %0, %1 = x86vector.avx512.vp2intersect %a, %a : vector<16xi32>31 // CHECK: x86vector.avx512.vp2intersect {{.*}} : vector<8xi64>32 %2, %3 = x86vector.avx512.vp2intersect %b, %b : vector<8xi64>33 return %0, %1, %2, %3 : vector<16xi1>, vector<16xi1>, vector<8xi1>, vector<8xi1>34}35 36// CHECK-LABEL: func @avx512_mask_compress37func.func @avx512_mask_compress(%k1: vector<16xi1>, %a1: vector<16xf32>,38 %k2: vector<8xi1>, %a2: vector<8xi64>)39 -> (vector<16xf32>, vector<16xf32>, vector<8xi64>)40{41 // CHECK: x86vector.avx512.mask.compress {{.*}} : vector<16xf32>42 %0 = x86vector.avx512.mask.compress %k1, %a1 : vector<16xf32>43 // CHECK: x86vector.avx512.mask.compress {{.*}} : vector<16xf32>44 %1 = x86vector.avx512.mask.compress %k1, %a1 {constant_src = dense<5.0> : vector<16xf32>} : vector<16xf32>45 // CHECK: x86vector.avx512.mask.compress {{.*}} : vector<8xi64>46 %2 = x86vector.avx512.mask.compress %k2, %a2, %a2 : vector<8xi64>, vector<8xi64>47 return %0, %1, %2 : vector<16xf32>, vector<16xf32>, vector<8xi64>48}49 50// CHECK-LABEL: func @avx512bf16_dot_12851func.func @avx512bf16_dot_128(%src: vector<4xf32>, %a: vector<8xbf16>,52 %b: vector<8xbf16>) -> (vector<4xf32>)53{54 // CHECK: x86vector.avx512.dot {{.*}} : vector<8xbf16> -> vector<4xf32>55 %0 = x86vector.avx512.dot %src, %a, %b : vector<8xbf16> -> vector<4xf32>56 return %0 : vector<4xf32>57}58 59// CHECK-LABEL: func @avx512bf16_dot_25660func.func @avx512bf16_dot_256(%src: vector<8xf32>, %a: vector<16xbf16>,61 %b: vector<16xbf16>) -> (vector<8xf32>)62{63 // CHECK: x86vector.avx512.dot {{.*}} : vector<16xbf16> -> vector<8xf32>64 %0 = x86vector.avx512.dot %src, %a, %b : vector<16xbf16> -> vector<8xf32>65 return %0 : vector<8xf32>66}67 68// CHECK-LABEL: func @avx512bf16_dot_51269func.func @avx512bf16_dot_512(%src: vector<16xf32>, %a: vector<32xbf16>,70 %b: vector<32xbf16>) -> (vector<16xf32>)71{72 // CHECK: x86vector.avx512.dot {{.*}} : vector<32xbf16> -> vector<16xf32>73 %0 = x86vector.avx512.dot %src, %a, %b : vector<32xbf16> -> vector<16xf32>74 return %0 : vector<16xf32>75}76 77// CHECK-LABEL: func @avx512bf16_cvt_packed_f32_to_bf16_25678func.func @avx512bf16_cvt_packed_f32_to_bf16_256(79 %a: vector<8xf32>) -> (vector<8xbf16>)80{81 // CHECK: x86vector.avx512.cvt.packed.f32_to_bf16 {{.*}} :82 // CHECK-SAME: vector<8xf32> -> vector<8xbf16>83 %0 = x86vector.avx512.cvt.packed.f32_to_bf16 %a : vector<8xf32> -> vector<8xbf16>84 return %0 : vector<8xbf16>85}86 87// CHECK-LABEL: func @avx512bf16_cvt_packed_f32_to_bf16_51288func.func @avx512bf16_cvt_packed_f32_to_bf16_512(89 %a: vector<16xf32>) -> (vector<16xbf16>)90{91 // CHECK: x86vector.avx512.cvt.packed.f32_to_bf16 {{.*}} :92 // CHECK-SAME: vector<16xf32> -> vector<16xbf16>93 %0 = x86vector.avx512.cvt.packed.f32_to_bf16 %a : vector<16xf32> -> vector<16xbf16>94 return %0 : vector<16xbf16>95}96 97// CHECK-LABEL: func @avxbf16_cvt_packed_even_indexed_bf16_to_f32_12898func.func @avxbf16_cvt_packed_even_indexed_bf16_to_f32_128(99 %a: memref<8xbf16>) -> vector<4xf32>100{101 // CHECK: x86vector.avx.cvt.packed.even.indexed_to_f32 {{.*}} :102 // CHECK-SAME: memref<8xbf16> -> vector<4xf32>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: x86vector.avx.cvt.packed.even.indexed_to_f32 {{.*}} :112 // CHECK-SAME: memref<16xbf16> -> vector<8xf32>113 %0 = x86vector.avx.cvt.packed.even.indexed_to_f32 %a : memref<16xbf16> -> vector<8xf32>114 return %0 : vector<8xf32>115}116 117// CHECK-LABEL: func @avxbf16_cvt_packed_odd_indexed_bf16_to_f32_128118func.func @avxbf16_cvt_packed_odd_indexed_bf16_to_f32_128(119 %a: memref<8xbf16>) -> vector<4xf32>120{121 // CHECK: x86vector.avx.cvt.packed.odd.indexed_to_f32 {{.*}} :122 // CHECK-SAME: memref<8xbf16> -> vector<4xf32>123 %0 = x86vector.avx.cvt.packed.odd.indexed_to_f32 %a : memref<8xbf16> -> vector<4xf32>124 return %0 : vector<4xf32>125}126 127// CHECK-LABEL: func @avxbf16_cvt_packed_odd_indexed_bf16_to_f32_256128func.func @avxbf16_cvt_packed_odd_indexed_bf16_to_f32_256(129 %a: memref<16xbf16>) -> vector<8xf32>130{131 // CHECK: x86vector.avx.cvt.packed.odd.indexed_to_f32 {{.*}} :132 // CHECK-SAME: memref<16xbf16> -> vector<8xf32>133 %0 = x86vector.avx.cvt.packed.odd.indexed_to_f32 %a : memref<16xbf16> -> vector<8xf32>134 return %0 : vector<8xf32>135}136 137// CHECK-LABEL: func @avxbf16_bcst_bf16_to_f32_128138func.func @avxbf16_bcst_bf16_to_f32_128(139 %a: memref<1xbf16>) -> vector<4xf32>140{141 // CHECK: x86vector.avx.bcst_to_f32.packed {{.*}} :142 // CHECK-SAME: memref<1xbf16> -> vector<4xf32>143 %0 = x86vector.avx.bcst_to_f32.packed %a : memref<1xbf16> -> vector<4xf32>144 return %0 : vector<4xf32>145}146 147// CHECK-LABEL: func @avxbf16_bcst_bf16_to_f32_256148func.func @avxbf16_bcst_bf16_to_f32_256(149 %a: memref<1xbf16>) -> vector<8xf32>150{151 // CHECK: x86vector.avx.bcst_to_f32.packed {{.*}} :152 // CHECK-SAME: memref<1xbf16> -> vector<8xf32>153 %0 = x86vector.avx.bcst_to_f32.packed %a : memref<1xbf16> -> vector<8xf32>154 return %0 : vector<8xf32>155}156 157// CHECK-LABEL: func @avxf16_cvt_packed_even_indexed_f16_to_f32_128158func.func @avxf16_cvt_packed_even_indexed_f16_to_f32_128(159 %a: memref<8xf16>) -> vector<4xf32>160{161 // CHECK: x86vector.avx.cvt.packed.even.indexed_to_f32 {{.*}} :162 // CHECK-SAME: memref<8xf16> -> vector<4xf32>163 %0 = x86vector.avx.cvt.packed.even.indexed_to_f32 %a : memref<8xf16> -> vector<4xf32>164 return %0 : vector<4xf32>165}166 167// CHECK-LABEL: func @avxf16_cvt_packed_even_indexed_f16_to_f32_256168func.func @avxf16_cvt_packed_even_indexed_f16_to_f32_256(169 %a: memref<16xf16>) -> vector<8xf32>170{171 // CHECK: x86vector.avx.cvt.packed.even.indexed_to_f32 {{.*}} :172 // CHECK-SAME: memref<16xf16> -> vector<8xf32>173 %0 = x86vector.avx.cvt.packed.even.indexed_to_f32 %a : memref<16xf16> -> vector<8xf32>174 return %0 : vector<8xf32>175}176 177// CHECK-LABEL: func @avxf16_cvt_packed_odd_indexed_f16_to_f32_128178func.func @avxf16_cvt_packed_odd_indexed_f16_to_f32_128(179 %a: memref<8xf16>) -> vector<4xf32>180{181 // CHECK: x86vector.avx.cvt.packed.odd.indexed_to_f32 {{.*}} :182 // CHECK-SAME: memref<8xf16> -> vector<4xf32>183 %0 = x86vector.avx.cvt.packed.odd.indexed_to_f32 %a : memref<8xf16> -> vector<4xf32>184 return %0 : vector<4xf32>185}186 187// CHECK-LABEL: func @avxf16_cvt_packed_odd_indexed_f16_to_f32_256188func.func @avxf16_cvt_packed_odd_indexed_f16_to_f32_256(189 %a: memref<16xf16>) -> vector<8xf32>190{191 // CHECK: x86vector.avx.cvt.packed.odd.indexed_to_f32 {{.*}} :192 // CHECK-SAME: memref<16xf16> -> vector<8xf32>193 %0 = x86vector.avx.cvt.packed.odd.indexed_to_f32 %a : memref<16xf16> -> vector<8xf32>194 return %0 : vector<8xf32>195}196 197// CHECK-LABEL: func @avxf16_bcst_f16_to_f32_128198func.func @avxf16_bcst_f16_to_f32_128(199 %a: memref<1xf16>) -> vector<4xf32>200{201 // CHECK: x86vector.avx.bcst_to_f32.packed {{.*}} :202 // CHECK-SAME: memref<1xf16> -> vector<4xf32>203 %0 = x86vector.avx.bcst_to_f32.packed %a : memref<1xf16> -> vector<4xf32>204 return %0 : vector<4xf32>205}206 207// CHECK-LABEL: func @avxf16_bcst_f16_to_f32_256208func.func @avxf16_bcst_f16_to_f32_256(209 %a: memref<1xf16>) -> vector<8xf32>210{211 // CHECK: x86vector.avx.bcst_to_f32.packed {{.*}} :212 // CHECK-SAME: memref<1xf16> -> vector<8xf32>213 %0 = x86vector.avx.bcst_to_f32.packed %a : memref<1xf16> -> vector<8xf32>214 return %0 : vector<8xf32>215}216 217// CHECK-LABEL: func @avx_rsqrt218func.func @avx_rsqrt(%a: vector<8xf32>) -> (vector<8xf32>)219{220 // CHECK: x86vector.avx.rsqrt {{.*}} : vector<8xf32>221 %0 = x86vector.avx.rsqrt %a : vector<8xf32>222 return %0 : vector<8xf32>223}224 225// CHECK-LABEL: func @avx_dot226func.func @avx_dot(%a: vector<8xf32>, %b: vector<8xf32>) -> (vector<8xf32>)227{228 // CHECK: x86vector.avx.intr.dot {{.*}} : vector<8xf32>229 %0 = x86vector.avx.intr.dot %a, %b : vector<8xf32>230 return %0 : vector<8xf32>231}232 233// CHECK-LABEL: func @avx_dot_i8_128234func.func @avx_dot_i8_128(%w: vector<4xi32>, %a: vector<16xi8>,235 %b: vector<16xi8>) -> vector<4xi32> {236 // CHECK: x86vector.avx.dot.i8 {{.*}} : vector<16xi8> -> vector<4xi32>237 %0 = x86vector.avx.dot.i8 %w, %a, %b : vector<16xi8> -> vector<4xi32>238 return %0 : vector<4xi32>239}240 241// CHECK-LABEL: func @avx_dot_i8_256242func.func @avx_dot_i8_256(%w: vector<8xi32>, %a: vector<32xi8>,243 %b: vector<32xi8>) -> vector<8xi32> {244 // CHECK: x86vector.avx.dot.i8 {{.*}} : vector<32xi8> -> vector<8xi32>245 %0 = x86vector.avx.dot.i8 %w, %a, %b : vector<32xi8> -> vector<8xi32>246 return %0 : vector<8xi32>247}248