1676 lines · plain
1// RUN: mlir-opt --split-input-file --tosa-infer-shapes --allow-unregistered-dialect %s | FileCheck %s2 3// CHECK-LABEL: @test_return4func.func @test_return(%arg0 : tensor<4xf32>) -> tensor<*xf32> {5 // CHECK: [[LOG:%.+]] = tosa.log %arg0 : (tensor<4xf32>) -> tensor<4xf32>6 // CHECK: tensor.cast [[LOG]] : tensor<4xf32> to tensor<*xf32>7 %0 = tosa.log %arg0 : (tensor<4xf32>) -> tensor<*xf32>8 return %0 : tensor<*xf32>9}10 11// -----12 13// CHECK-LABEL: @test_multiple14func.func @test_multiple(%arg0 : tensor<4xf32>, %arg1 : tensor<1xf32>, %arg2 : tensor<1xf32>) -> tensor<*xf32> {15 // CHECK: [[ADD:%.+]] = tosa.add %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>16 %0 = tosa.add %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>17 18 // CHECK: [[LOG:%.+]] = tosa.log %0 : (tensor<4xf32>) -> tensor<4xf32>19 %1 = tosa.log %0 : (tensor<*xf32>) -> tensor<*xf32>20 21 // CHECK: [[SUB:%.+]] = tosa.sub %0, %arg2 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>22 %2 = tosa.sub %0, %arg2 : (tensor<*xf32>, tensor<1xf32>) -> tensor<*xf32>23 return %0 : tensor<*xf32>24}25 26// -----27 28// CHECK-LABEL: @test_unary_f3229func.func @test_unary_f32(%arg0 : tensor<4xf32>) -> () {30 // CHECK: tosa.abs %arg0 : (tensor<4xf32>) -> tensor<4xf32>31 %0 = tosa.abs %arg0 : (tensor<4xf32>) -> tensor<*xf32>32 33 // CHECK: tosa.ceil %arg0 : (tensor<4xf32>) -> tensor<4xf32>34 %1 = tosa.ceil %arg0 : (tensor<4xf32>) -> tensor<*xf32>35 36 // CHECK: tosa.clamp %arg0 {{.+}} : (tensor<4xf32>) -> tensor<4xf32>37 %2 = tosa.clamp %arg0 { min_val = 0.0 : f32, max_val = 10.0 : f32 } : (tensor<4xf32>) -> tensor<*xf32>38 39 // CHECK: tosa.exp %arg0 : (tensor<4xf32>) -> tensor<4xf32>40 %3 = tosa.exp %arg0 : (tensor<4xf32>) -> tensor<*xf32>41 42 // CHECK: tosa.floor %arg0 : (tensor<4xf32>) -> tensor<4xf32>43 %4 = tosa.floor %arg0 : (tensor<4xf32>) -> tensor<*xf32>44 45 // CHECK: tosa.log %arg0 : (tensor<4xf32>) -> tensor<4xf32>46 %5 = tosa.log %arg0 : (tensor<4xf32>) -> tensor<*xf32>47 48 %in_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>49 %out_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>50 // CHECK: tosa.negate %arg0, {{.+}} : (tensor<4xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<4xf32>51 %6 = tosa.negate %arg0, %in_zp, %out_zp : (tensor<4xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<*xf32>52 53 // CHECK: tosa.reciprocal %arg0 : (tensor<4xf32>) -> tensor<4xf32>54 %7 = tosa.reciprocal %arg0 : (tensor<4xf32>) -> tensor<*xf32>55 56 // CHECK: tosa.reverse %arg0 {axis = 0 : i32} : (tensor<4xf32>) -> tensor<4xf32>57 %8 = tosa.reverse %arg0 { axis = 0 : i32 } : (tensor<4xf32>) -> tensor<?xf32>58 59 // CHECK: tosa.rsqrt %arg0 : (tensor<4xf32>) -> tensor<4xf32>60 %9 = tosa.rsqrt %arg0 : (tensor<4xf32>) -> tensor<*xf32>61 62 // CHECK: tosa.tanh %arg0 : (tensor<4xf32>) -> tensor<4xf32>63 %10 = tosa.tanh %arg0 : (tensor<4xf32>) -> tensor<*xf32>64 65 // CHECK: tosa.sigmoid %arg0 : (tensor<4xf32>) -> tensor<4xf32>66 %11 = tosa.sigmoid %arg0 : (tensor<4xf32>) -> tensor<*xf32>67 68 // CHECK: tosa.cast %arg0 : (tensor<4xf32>) -> tensor<4xi32>69 %12 = tosa.cast %arg0 : (tensor<4xf32>) -> tensor<*xi32>70 71 // CHECK: tosa.erf %arg0 : (tensor<4xf32>) -> tensor<4xf32>72 %13 = tosa.erf %arg0 : (tensor<4xf32>) -> tensor<*xf32>73 return74}75 76// -----77 78// CHECK-LABEL: @test_unary_i3279func.func @test_unary_i32(%arg0 : tensor<4xi32>, %arg1 : tensor<2xi8>) -> () {80 // CHECK: tosa.abs %arg0 : (tensor<4xi32>) -> tensor<4xi32>81 %0 = tosa.abs %arg0 : (tensor<4xi32>) -> tensor<*xi32>82 83 // CHECK: tosa.bitwise_not %arg0 : (tensor<4xi32>) -> tensor<4xi32>84 %1 = tosa.bitwise_not %arg0 : (tensor<4xi32>) -> tensor<*xi32>85 86 // CHECK: tosa.clamp %arg0 {{.+}} : (tensor<4xi32>) -> tensor<4xi32>87 %2 = tosa.clamp %arg0 { max_val = 10 : i32, min_val = 0 : i32} : (tensor<4xi32>) -> tensor<*xi32>88 89 // CHECK: tosa.clz %arg0 : (tensor<4xi32>) -> tensor<4xi32>90 %3 = tosa.clz %arg0 : (tensor<4xi32>) -> tensor<*xi32>91 92 %in_zp = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>93 %out_zp = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>94 // CHECK: tosa.negate %arg0, {{.+}} : (tensor<4xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<4xi32>95 %4 = tosa.negate %arg0, %in_zp, %out_zp : (tensor<4xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<*xi32>96 97 // CHECK: tosa.reverse %arg0 {axis = 0 : i32} : (tensor<4xi32>) -> tensor<4xi32>98 %5 = tosa.reverse %arg0 { axis = 0 : i32 } : (tensor<4xi32>) -> tensor<?xi32>99 100 // CHECK-DAG: %[[MULT:.+]] = "tosa.const"() <{values = dense<[42, 43]> : tensor<2xi16>}> : () -> tensor<2xi16>101 // CHECK-DAG: %[[SHIFT:.+]] = "tosa.const"() <{values = dense<[14, 15]> : tensor<2xi8>}> : () -> tensor<2xi8>102 // CHECK-DAG: %[[INPUTZP:.+]] = "tosa.const"() <{values = dense<43> : tensor<1xi8>}> : () -> tensor<1xi8>103 // CHECK-DAG: %[[OUTPUTZP:.+]] = "tosa.const"() <{values = dense<52> : tensor<1xi8>}> : () -> tensor<1xi8>104 // CHECK: tosa.rescale %arg1, %[[MULT]], %[[SHIFT]], %[[INPUTZP]], %[[OUTPUTZP]] {{.+}} : (tensor<2xi8>, tensor<2xi16>, tensor<2xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<2xi8>105 %multiplier = "tosa.const"() {values = dense<[42, 43]> : tensor<2xi16>} : () -> tensor<2xi16>106 %shift = "tosa.const"() {values = dense<[14, 15]> : tensor<2xi8>} : () -> tensor<2xi8>107 %input_zp = "tosa.const"() {values = dense<43> : tensor<1xi8>} : () -> tensor<1xi8>108 %output_zp = "tosa.const"() {values = dense<52> : tensor<1xi8>} : () -> tensor<1xi8>109 %6 = tosa.rescale %arg1, %multiplier, %shift, %input_zp, %output_zp {scale32 = false, rounding_mode = SINGLE_ROUND, per_channel = true, input_unsigned = true, output_unsigned = true} : (tensor<2xi8>, tensor<2xi16>, tensor<2xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<2xi8>110 111 // CHECK: tosa.identity %arg0 : (tensor<4xi32>) -> tensor<4xi32>112 %7 = tosa.identity %arg0 : (tensor<4xi32>) -> tensor<?xi32>113 return114}115 116// -----117 118// CHECK-LABEL: @test_binary_scalar_f32119func.func @test_binary_scalar_f32(%arg0 : tensor<4xf32>, %arg1 : tensor<1xf32>) -> () {120 // CHECK: tosa.add %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>121 %0 = tosa.add %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>122 123 // CHECK: tosa.maximum %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>124 %1 = tosa.maximum %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>125 126 // CHECK: tosa.minimum %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>127 %2 = tosa.minimum %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>128 129 %3 = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>130 // CHECK: tosa.mul %arg0, %arg1, %3 : (tensor<4xf32>, tensor<1xf32>, tensor<1xi8>) -> tensor<4xf32>131 %4 = tosa.mul %arg0, %arg1, %3 : (tensor<4xf32>, tensor<1xf32>, tensor<1xi8>) -> tensor<*xf32>132 133 // CHECK: tosa.pow %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>134 %5 = tosa.pow %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>135 136 // CHECK: tosa.sub %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>137 %6 = tosa.sub %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>138 139 // CHECK: tosa.equal %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xi1>140 %7 = tosa.equal %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xi1>141 142 // CHECK: tosa.greater %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xi1>143 %8 = tosa.greater %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xi1>144 145 // CHECK: tosa.greater_equal %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xi1>146 %9 = tosa.greater_equal %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xi1>147 148 return149}150 151// -----152 153// CHECK-LABEL: @test_binary_broadcast_f32154func.func @test_binary_broadcast_f32(%arg0 : tensor<4xf32>, %arg1 : tensor<1xf32>) -> () {155 // CHECK: tosa.add %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>156 %0 = tosa.add %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>157 158 // CHECK: tosa.maximum %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>159 %1 = tosa.maximum %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>160 161 // CHECK: tosa.minimum %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>162 %2 = tosa.minimum %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>163 164 %3 = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>165 // CHECK: tosa.mul %arg0, %arg1, %3 : (tensor<4xf32>, tensor<1xf32>, tensor<1xi8>) -> tensor<4xf32>166 %4 = tosa.mul %arg0, %arg1, %3 : (tensor<4xf32>, tensor<1xf32>, tensor<1xi8>) -> tensor<*xf32>167 168 // CHECK: tosa.pow %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>169 %5 = tosa.pow %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>170 171 // CHECK: tosa.sub %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xf32>172 %6 = tosa.sub %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xf32>173 174 // CHECK: tosa.equal %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xi1>175 %7 = tosa.equal %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xi1>176 177 // CHECK: tosa.greater %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xi1>178 %8 = tosa.greater %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xi1>179 180 // CHECK: tosa.greater_equal %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<4xi1>181 %9 = tosa.greater_equal %arg0, %arg1 : (tensor<4xf32>, tensor<1xf32>) -> tensor<*xi1>182 183 return184}185 186// -----187 188// CHECK-LABEL: @test_binary_i32189func.func @test_binary_i32(%arg0 : tensor<4xi32>, %arg1 : tensor<1xi32>) -> () {190 // CHECK: tosa.add %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>191 %0 = tosa.add %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>192 193 // CHECK: tosa.bitwise_and %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>194 %1 = tosa.bitwise_and %arg0, %arg1: (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>195 196 // CHECK: tosa.bitwise_or %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>197 %2 = tosa.bitwise_or %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>198 199 // CHECK: tosa.bitwise_xor %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>200 %3 = tosa.bitwise_xor %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>201 202 // CHECK: tosa.equal %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi1>203 %4 = tosa.equal %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi1>204 205 // CHECK: tosa.greater %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi1>206 %5 = tosa.greater %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi1>207 208 // CHECK: tosa.greater_equal %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi1>209 %6 = tosa.greater_equal %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi1>210 211 // CHECK: tosa.logical_left_shift %arg0, %arg1 {shift = 0 : i32} : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>212 %7 = tosa.logical_left_shift %arg0, %arg1 { shift = 0 : i32 }: (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>213 214 // CHECK: tosa.logical_right_shift %arg0, %arg1 {shift = 0 : i32} : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>215 %8 = tosa.logical_right_shift %arg0, %arg1 { shift = 0 : i32 }: (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>216 217 // CHECK: tosa.maximum %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>218 %9 = tosa.maximum %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>219 220 // CHECK: tosa.minimum %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>221 %10 = tosa.minimum %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>222 223 // CHECK: tosa.mul %arg0, %arg1, %{{.*}} : (tensor<4xi32>, tensor<1xi32>, tensor<1xi8>) -> tensor<4xi32>224 %shift = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>225 %11 = tosa.mul %arg0, %arg1, %shift : (tensor<4xi32>, tensor<1xi32>, tensor<1xi8>) -> tensor<*xi32>226 227 // CHECK: tosa.pow %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>228 %13 = tosa.pow %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>229 230 // CHECK: tosa.sub %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>231 %14 = tosa.sub %arg0, %arg1 : (tensor<4xi32>, tensor<1xi32>) -> tensor<*xi32>232 233 return234}235 236// -----237 238// CHECK-LABEL: @test_binary_i1239func.func @test_binary_i1(%arg0 : tensor<4xi1>, %arg1 : tensor<1xi1>) -> () {240 // CHECK: tosa.logical_and %arg0, %arg1 : (tensor<4xi1>, tensor<1xi1>) -> tensor<4xi1>241 %0 = tosa.logical_and %arg0, %arg1 : (tensor<4xi1>, tensor<1xi1>) -> tensor<*xi1>242 243 // CHECK: tosa.logical_or %arg0, %arg1 : (tensor<4xi1>, tensor<1xi1>) -> tensor<4xi1>244 %1 = tosa.logical_or %arg0, %arg1 : (tensor<4xi1>, tensor<1xi1>) -> tensor<*xi1>245 246 // CHECK: tosa.logical_xor %arg0, %arg1 : (tensor<4xi1>, tensor<1xi1>) -> tensor<4xi1>247 %2 = tosa.logical_xor %arg0, %arg1 : (tensor<4xi1>, tensor<1xi1>) -> tensor<*xi1>248 249 return250}251 252// -----253 254// CHECK-LABEL: @test_select_i32255func.func @test_select_i32(%arg0 : tensor<4xi1>, %arg1 : tensor<1xi32>, %arg2 : tensor<4xi32>) -> () {256 // CHECK: tosa.select %arg0, %arg1, %arg2 : (tensor<4xi1>, tensor<1xi32>, tensor<4xi32>) -> tensor<4xi32>257 %0 = tosa.select %arg0, %arg1, %arg2 : (tensor<4xi1>, tensor<1xi32>, tensor<4xi32>) -> tensor<*xi32>258 259 return260}261 262// -----263 264// CHECK-LABEL: @test_static_argmax265func.func @test_static_argmax(%arg0 : tensor<2x3xi32>) -> () {266 // CHECK: tosa.argmax %arg0 {axis = 0 : i32} : (tensor<2x3xi32>) -> tensor<3xi32>267 %0 = tosa.argmax %arg0 {axis = 0 : i32} : (tensor<2x3xi32>) -> tensor<?xi32>268 269 // CHECK: tosa.argmax %arg0 {axis = 1 : i32} : (tensor<2x3xi32>) -> tensor<2xi32>270 %1 = tosa.argmax %arg0 {axis = 1 : i32} : (tensor<2x3xi32>) -> tensor<?xi32>271 return272}273 274// -----275 276// CHECK-LABEL: @test_dynamic_argmax277func.func @test_dynamic_argmax(%arg0 : tensor<2x?xi32>) -> () {278 // CHECK: tosa.argmax %arg0 {axis = 0 : i32} : (tensor<2x?xi32>) -> tensor<?xi32>279 %0 = tosa.argmax %arg0 {axis = 0 : i32} : (tensor<2x?xi32>) -> tensor<?xi32>280 281 // CHECK: tosa.argmax %arg0 {axis = 1 : i32} : (tensor<2x?xi32>) -> tensor<2xi32>282 %1 = tosa.argmax %arg0 {axis = 1 : i32} : (tensor<2x?xi32>) -> tensor<?xi32>283 return284}285 286// -----287 288// CHECK-LABEL: @test_static_matmul289func.func @test_static_matmul(%arg0 : tensor<2x3x4xi32>, %arg1 : tensor<2x4x5xi32>) -> () {290 // CHECK tosa.matmul %arg0, %arg1, %0, %1 : (tensor<2x3x4xi32>, tensor<2x4x5xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<2x3x5xi32>291 %0 = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>292 %1 = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>293 %2 = tosa.matmul %arg0, %arg1, %0, %1 : (tensor<2x3x4xi32>, tensor<2x4x5xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<?x?x?xi32>294 295 return296}297 298// -----299 300// CHECK-LABEL: @test_dynamic_lhs_matmul301func.func @test_dynamic_lhs_matmul(%arg0 : tensor<?x?x?xi32>, %arg1 : tensor<2x4x5xi32>) -> () {302 // CHECK: tosa.matmul %arg0, %arg1, %0, %1 : (tensor<?x?x?xi32>, tensor<2x4x5xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<2x?x5xi32>303 %0 = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>304 %1 = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>305 %2 = tosa.matmul %arg0, %arg1, %0, %1 : (tensor<?x?x?xi32>, tensor<2x4x5xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<?x?x?xi32>306 307 return308}309 310// -----311 312// CHECK-LABEL: @test_dynamic_rhs_matmul313func.func @test_dynamic_rhs_matmul(%arg0 : tensor<2x3x4xi32>, %arg1 : tensor<?x?x?xi32>) -> () {314 // CHECK: tosa.matmul %arg0, %arg1, %0, %1 : (tensor<2x3x4xi32>, tensor<?x?x?xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<2x3x?xi32>315 %0 = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>316 %1 = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>317 %2 = tosa.matmul %arg0, %arg1, %0, %1 : (tensor<2x3x4xi32>, tensor<?x?x?xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<?x?x?xi32>318 319 return320}321 322// -----323 324// CHECK-LABEL: @test_dynamic_mixed_matmul325func.func @test_dynamic_mixed_matmul(%arg0 : tensor<?x3x?xi32>, %arg1 : tensor<?x?x5xi32>) -> () {326 // CHECK: tosa.matmul %arg0, %arg1, %0, %1 : (tensor<?x3x?xi32>, tensor<?x?x5xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<?x3x5xi32>327 %0 = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>328 %1 = "tosa.const"() <{values = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>329 %2 = tosa.matmul %arg0, %arg1, %0, %1 : (tensor<?x3x?xi32>, tensor<?x?x5xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<?x?x?xi32>330 331 return332}333 334// -----335 336// CHECK-LABEL: @test_unranked_zero_points_matmul337func.func @test_unranked_zero_points_matmul(%arg0: tensor<1x2x3xf32>, %arg1: tensor<1x3x4xf32>, %zero_point: tensor<1xf32>) -> tensor<1x2x4xf32> {338 // CHECK: %[[ZP:.*]] = tosa.cast %arg2 : (tensor<1xf32>) -> tensor<1xf32>339 %zero_point_unranked = "tosa.cast"(%zero_point) : (tensor<1xf32>) -> tensor<*xf32>340 // CHECK: tosa.matmul %arg0, %arg1, %[[ZP]], %[[ZP]] : (tensor<1x2x3xf32>, tensor<1x3x4xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x2x4xf32>341 %0 = tosa.matmul %arg0, %arg1, %zero_point_unranked, %zero_point_unranked : (tensor<1x2x3xf32>, tensor<1x3x4xf32>, tensor<*xf32>, tensor<*xf32>) -> tensor<1x2x4xf32>342 return %0 : tensor<1x2x4xf32>343}344 345// -----346 347// CHECK-LABEL: @test_accepts_unranked_scalar_tensor348func.func @test_accepts_unranked_scalar_tensor(%arg0: tensor<1x2x2xf32>, %arg1: tensor<1xf32>) -> tensor<*xf32> {349 // CHECK: %[[ZP:.*]] = tosa.cast %arg1 : (tensor<1xf32>) -> tensor<1xf32>350 %0 = tosa.cast %arg1 : (tensor<1xf32>) -> tensor<*xf32>351 // CHECK: %[[SHAPE:.*]] = tosa.const_shape352 %1 = tosa.const_shape {values = dense<[0, 0, 0, 1, 0, 1]> : tensor<6xindex>} : () -> !tosa.shape<6>353 // CHECK: tosa.pad %arg0, %[[SHAPE]], %[[ZP]] : (tensor<1x2x2xf32>, !tosa.shape<6>, tensor<1xf32>) -> tensor<1x3x3xf32>354 %2 = tosa.pad %arg0, %1, %0 : (tensor<1x2x2xf32>, !tosa.shape<6>, tensor<*xf32>) -> tensor<*xf32>355 return %2 : tensor<*xf32>356}357 358// -----359 360// CHECK-LABEL: @test_unranked_scalar_i8_tensor361func.func @test_unranked_scalar_i8_tensor(%arg0: tensor<4xi32>, %arg1: tensor<4xi32>, %arg2: tensor<1xi8>) -> tensor<4xi32> {362 // CHECK: %[[SHIFT:.*]] = tosa.cast %arg2 : (tensor<1xi8>) -> tensor<1xi8>363 %shift = tosa.cast %arg2 : (tensor<1xi8>) -> tensor<*xi8>364 // CHECK: tosa.mul %arg0, %arg1, %[[SHIFT]] : (tensor<4xi32>, tensor<4xi32>, tensor<1xi8>) -> tensor<4xi32>365 %0 = tosa.mul %arg0, %arg1, %shift : (tensor<4xi32>, tensor<4xi32>, tensor<*xi8>) -> tensor<4xi32>366 return %0 : tensor<4xi32>367}368 369// -----370 371// CHECK-LABEL: @test_table_static372func.func @test_table_static(%arg0 : tensor<4x5xi16>, %arg1 : tensor<513xi16>) -> () {373 // CHECK:tosa.table %arg0, %arg1 : (tensor<4x5xi16>, tensor<513xi16>) -> tensor<4x5xi16>374 %0 = tosa.table %arg0, %arg1 : (tensor<4x5xi16>, tensor<513xi16>) -> tensor<?x?xi16>375 return376}377 378// -----379 380// CHECK-LABEL: @test_table_dynamic381func.func @test_table_dynamic(%arg0 : tensor<4x?xi16>, %arg1 : tensor<513xi16>) -> () {382 // CHECK:tosa.table %arg0, %arg1 : (tensor<4x?xi16>, tensor<513xi16>) -> tensor<4x?xi16>383 %0 = tosa.table %arg0, %arg1 : (tensor<4x?xi16>, tensor<513xi16>) -> tensor<?x?xi16>384 return385}386 387// -----388 389// CHECK-LABEL: @test_static_reshape390func.func @test_static_reshape(%arg0 : tensor<4x4xi32>) -> () {391 // CHECK: %[[CONST3:.+]] = tosa.const_shape {values = dense<16> : tensor<1xindex>} : () -> !tosa.shape<1>392 %3 = tosa.const_shape {values = dense<16> : tensor<1xindex>} : () -> !tosa.shape<1>393 // CHECK: tosa.reshape %arg0, %[[CONST3]] : (tensor<4x4xi32>, !tosa.shape<1>) -> tensor<16xi32>394 %0 = tosa.reshape %arg0, %3 : (tensor<4x4xi32>, !tosa.shape<1>) -> tensor<16xi32>395 396 // CHECK: %[[CONST4:.+]] = tosa.const_shape {values = dense<-1> : tensor<1xindex>} : () -> !tosa.shape<1>397 // CHECK: tosa.reshape %arg0, %[[CONST4]] : (tensor<4x4xi32>, !tosa.shape<1>) -> tensor<16xi32>398 %4 = tosa.const_shape {values = dense<-1> : tensor<1xindex>} : () -> !tosa.shape<1>399 %1 = tosa.reshape %arg0, %4 : (tensor<4x4xi32>, !tosa.shape<1>) -> tensor<16xi32>400 401 // CHECK: %[[CONST5:.+]] = tosa.const_shape {values = dense<[2, -1]> : tensor<2xindex>} : () -> !tosa.shape<2>402 // CHECK: tosa.reshape %arg0, %[[CONST5]] : (tensor<4x4xi32>, !tosa.shape<2>) -> tensor<2x8xi32>403 %5 = tosa.const_shape {values = dense<[2, -1]> : tensor<2xindex>} : () -> !tosa.shape<2>404 %2 = tosa.reshape %arg0, %5 : (tensor<4x4xi32>, !tosa.shape<2>) -> tensor<2x8xi32>405 406 return407}408 409// -----410 411// CHECK-LABEL: @test_dynamic_reshape412func.func @test_dynamic_reshape(%arg0 : tensor<4x?xi32>) -> () {413 // CHECK: %0 = tosa.const_shape {values = dense<16> : tensor<1xindex>} : () -> !tosa.shape<1>414 %0 = tosa.const_shape {values = dense<16> : tensor<1xindex>} : () -> !tosa.shape<1>415 // CHECK: %1 = tosa.reshape %arg0, %0 : (tensor<4x?xi32>, !tosa.shape<1>) -> tensor<16xi32>416 %1 = tosa.reshape %arg0, %0 : (tensor<4x?xi32>, !tosa.shape<1>) -> tensor<?xi32>417 418 // CHECK: %2 = tosa.const_shape {values = dense<-1> : tensor<1xindex>} : () -> !tosa.shape<1>419 %2 = tosa.const_shape {values = dense<-1> : tensor<1xindex>} : () -> !tosa.shape<1>420 // CHECK: %3 = tosa.reshape %arg0, %2 : (tensor<4x?xi32>, !tosa.shape<1>) -> tensor<?xi32>421 %3 = tosa.reshape %arg0, %2 : (tensor<4x?xi32>, !tosa.shape<1>) -> tensor<?xi32>422 423 // CHECK: %4 = tosa.const_shape {values = dense<[2, -1]> : tensor<2xindex>} : () -> !tosa.shape<2>424 %4 = tosa.const_shape {values = dense<[2, -1]> : tensor<2xindex>} : () -> !tosa.shape<2>425 // CHECK: %5 = tosa.reshape %arg0, %4 : (tensor<4x?xi32>, !tosa.shape<2>) -> tensor<2x?xi32>426 %5 = tosa.reshape %arg0, %4 : (tensor<4x?xi32>, !tosa.shape<2>) -> tensor<?x?xi32>427 428 return429}430 431// -----432 433// CHECK: @test_reduce_binary434func.func @test_reduce_binary(%arg0 : tensor<2x3x?x?xi1>) -> () {435 // CHECK: tosa.reduce_all %arg0 {axis = 0 : i32} : (tensor<2x3x?x?xi1>) -> tensor<1x3x?x?xi1>436 %0 = tosa.reduce_all %arg0 {axis = 0 : i32} : (tensor<2x3x?x?xi1>) -> tensor<?x?x?x?xi1>437 438 // CHECK: tosa.reduce_all %arg0 {axis = 1 : i32} : (tensor<2x3x?x?xi1>) -> tensor<2x1x?x?xi1>439 %1 = tosa.reduce_all %arg0 {axis = 1 : i32} : (tensor<2x3x?x?xi1>) -> tensor<?x?x?x?xi1>440 441 // CHECK: tosa.reduce_all %arg0 {axis = 2 : i32} : (tensor<2x3x?x?xi1>) -> tensor<2x3x1x?xi1>442 %2 = tosa.reduce_all %arg0 {axis = 2 : i32} : (tensor<2x3x?x?xi1>) -> tensor<?x?x?x?xi1>443 444 // CHECK: tosa.reduce_all %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xi1>) -> tensor<2x3x?x1xi1>445 %3 = tosa.reduce_all %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xi1>) -> tensor<?x?x?x?xi1>446 447 // CHECK: tosa.reduce_any %arg0 {axis = 0 : i32} : (tensor<2x3x?x?xi1>) -> tensor<1x3x?x?xi1>448 %4 = tosa.reduce_any %arg0 {axis = 0 : i32} : (tensor<2x3x?x?xi1>) -> tensor<?x?x?x?xi1>449 450 return451}452 453// -----454 455// CHECK: @test_reduce_float456func.func @test_reduce_float(%arg0 : tensor<2x3x?x?xf32>) -> () {457 // CHECK: tosa.reduce_sum %arg0 {axis = 0 : i32} : (tensor<2x3x?x?xf32>) -> tensor<1x3x?x?xf32>458 %0 = tosa.reduce_sum %arg0 {axis = 0 : i32} : (tensor<2x3x?x?xf32>) -> tensor<?x?x?x?xf32>459 460 // CHECK: tosa.reduce_sum %arg0 {axis = 1 : i32} : (tensor<2x3x?x?xf32>) -> tensor<2x1x?x?xf32>461 %1 = tosa.reduce_sum %arg0 {axis = 1 : i32} : (tensor<2x3x?x?xf32>) -> tensor<?x?x?x?xf32>462 463 // CHECK: tosa.reduce_sum %arg0 {axis = 2 : i32} : (tensor<2x3x?x?xf32>) -> tensor<2x3x1x?xf32>464 %2 = tosa.reduce_sum %arg0 {axis = 2 : i32} : (tensor<2x3x?x?xf32>) -> tensor<?x?x?x?xf32>465 466 // CHECK: tosa.reduce_sum %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xf32>) -> tensor<2x3x?x1xf32>467 %3 = tosa.reduce_sum %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xf32>) -> tensor<?x?x?x?xf32>468 469 // CHECK: tosa.reduce_max %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xf32>) -> tensor<2x3x?x1xf32>470 %4 = tosa.reduce_max %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xf32>) -> tensor<?x?x?x?xf32>471 472 // CHECK: tosa.reduce_min %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xf32>) -> tensor<2x3x?x1xf32>473 %5 = tosa.reduce_min %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xf32>) -> tensor<?x?x?x?xf32>474 475 // CHECK: tosa.reduce_product %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xf32>) -> tensor<2x3x?x1xf32>476 %6 = tosa.reduce_product %arg0 {axis = 3 : i32} : (tensor<2x3x?x?xf32>) -> tensor<?x?x?x?xf32>477 478 return479}480 481// -----482 483// CHECK-LABEL: @test_concat484func.func @test_concat(%arg0 : tensor<1x2xf32>, %arg1 : tensor<2x2xf32>) -> () {485 // CHECK: tosa.concat %arg0, %arg1 {axis = 0 : i32} : (tensor<1x2xf32>, tensor<2x2xf32>) -> tensor<3x2xf32>486 %0 = tosa.concat %arg0, %arg1 {axis = 0 : i32} : (tensor<1x2xf32>, tensor<2x2xf32>) -> tensor<?x?xf32>487 488 return489}490 491// -----492 493// CHECK-LABEL: @test_concat_dynamic494func.func @test_concat_dynamic(%arg0 : tensor<1x2xf32>, %arg1 : tensor<2x?xf32>) -> () {495 // CHECK: tosa.concat %arg0, %arg1 {axis = 0 : i32} : (tensor<1x2xf32>, tensor<2x?xf32>) -> tensor<3x2xf32>496 %0 = tosa.concat %arg0, %arg1 {axis = 0 : i32} : (tensor<1x2xf32>, tensor<2x?xf32>) -> tensor<?x?xf32>497 498 return499}500 501// -----502 503// CHECK-LABEL: @test_concat_dynamic_axis504func.func @test_concat_dynamic_axis(%arg0 : tensor<?x2xf32>, %arg1 : tensor<2x2xf32>) -> () {505 // CHECK: tosa.concat %arg0, %arg1 {axis = 0 : i32} : (tensor<?x2xf32>, tensor<2x2xf32>) -> tensor<?x2xf32>506 %0 = tosa.concat %arg0, %arg1 {axis = 0 : i32} : (tensor<?x2xf32>, tensor<2x2xf32>) -> tensor<?x?xf32>507 508 return509}510 511// -----512 513// CHECK-LABEL: @test_concat_axis_1514func.func @test_concat_axis_1(%arg0 : tensor<2x1xf32>, %arg1 : tensor<2x2xf32>) -> () {515 // CHECK: tosa.concat %arg0, %arg1 {axis = 1 : i32} : (tensor<2x1xf32>, tensor<2x2xf32>) -> tensor<2x3xf32>516 %0 = tosa.concat %arg0, %arg1 {axis = 1 : i32} : (tensor<2x1xf32>, tensor<2x2xf32>) -> tensor<?x?xf32>517 518 return519}520 521 522// -----523 524// CHECK-LABEL:@test_padding_dynamic_input525func.func @test_padding_dynamic_input(%arg0 : tensor<1x?xf32>) -> () {526 %0 = tosa.const_shape { values = dense<[1, 2, 3, 4]> : tensor<4xindex> } : () -> !tosa.shape<4>527 %1 = "tosa.const"() {values = dense<3.14> : tensor<1xf32>} : () -> tensor<1xf32>528 // CHECK: tosa.pad %arg0, %0, %1 : (tensor<1x?xf32>, !tosa.shape<4>, tensor<1xf32>) -> tensor<4x?xf32>529 %2 = tosa.pad %arg0, %0, %1 : (tensor<1x?xf32>, !tosa.shape<4>, tensor<1xf32>) -> tensor<?x?xf32>530 return531}532 533// -----534 535// CHECK-LABEL: @test_padding_simple536func.func @test_padding_simple(%arg0 : tensor<1x2xf32>) -> () {537 %0 = tosa.const_shape { values = dense<[1, 2, 3, 4]> : tensor<4xindex> } : () -> !tosa.shape<4>538 %1 = "tosa.const"() {values = dense<3.14> : tensor<1xf32>} : () -> tensor<1xf32>539 // CHECK: tosa.pad %arg0, %0, %1 : (tensor<1x2xf32>, !tosa.shape<4>, tensor<1xf32>) -> tensor<4x9xf32>540 %2 = tosa.pad %arg0, %0, %1 : (tensor<1x2xf32>, !tosa.shape<4>, tensor<1xf32>) -> tensor<?x?xf32>541 return542}543 544// -----545 546// CHECK-LABEL: @test_slice547func.func @test_slice(%arg0 : tensor<?xi32>) -> () {548 // CHECK: %0 = tosa.const_shape {values = dense<1> : tensor<1xindex>}549 // CHECK: %1 = tosa.const_shape {values = dense<2> : tensor<1xindex>}550 // CHECK: %2 = tosa.slice %arg0, %0, %1 : (tensor<?xi32>, !tosa.shape<1>, !tosa.shape<1>) -> tensor<2xi32>551 %0 = tosa.const_shape {values = dense<1> : tensor<1xindex>} : () -> !tosa.shape<1>552 %1 = tosa.const_shape {values = dense<2> : tensor<1xindex>} : () -> !tosa.shape<1>553 %2= tosa.slice %arg0, %0, %1 : (tensor<?xi32>, !tosa.shape<1>, !tosa.shape<1>) -> tensor<?xi32>554 return555}556 557// -----558 559// CHECK-LABEL: @test_slice_size_minus_one560func.func @test_slice_size_minus_one(%arg0 : tensor<?x8x8x8xi32>) -> () {561 // CHECK: %[[START:.+]] = tosa.const_shape562 // CHECK: %[[SIZE:.+]] = tosa.const_shape563 // CHECK: %[[VAL:.+]] = tosa.slice %arg0, %[[START]], %[[SIZE]] : (tensor<?x8x8x8xi32>, !tosa.shape<4>, !tosa.shape<4>) -> tensor<?x7x?x?xi32>564 // this checks following565 // dim 0: size=-1, input dim=? => inferred output dim is ?566 // dim 1: size=-1 => inferred output dim is input_dim - start567 // dim 2: size=-1, start=-1 => inferred output dim is ?568 // dim 3: size=-1, start=8 => inferred output dim is ? because start is out of bound569 %start = tosa.const_shape {values = dense<[0, 1, -1, 8]> : tensor<4xindex>} : () -> !tosa.shape<4>570 %size = tosa.const_shape {values = dense<[-1, -1, -1, -1]> : tensor<4xindex>} : () -> !tosa.shape<4>571 %2= tosa.slice %arg0, %start, %size : (tensor<?x8x8x8xi32>, !tosa.shape<4>, !tosa.shape<4>) -> tensor<?x?x?x?xi32>572 return573}574 575// -----576 577// CHECK-LABEL: @test_slice_size_out_of_bound578func.func @test_slice_size_out_of_bound(%arg0 : tensor<8x8x8x?xi32>) -> () {579 // CHECK: %[[START:.+]] = tosa.const_shape580 // CHECK: %[[SIZE:.+]] = tosa.const_shape581 // CHECK: %[[VAL:.+]] = tosa.slice %arg0, %[[START]], %[[SIZE]] : (tensor<8x8x8x?xi32>, !tosa.shape<4>, !tosa.shape<4>) -> tensor<?x?x?x4xi32>582 // this checks following583 // dim 0: size=0 => inferred output dim is ?584 // dim 1: size=-2 => inferred output dim is ?585 // dim 3: start+size out of bound because size too big: inferred output dim is ?586 // dim 4: size=4, input dim=? => inferred output dim is 4587 %start = tosa.const_shape {values = dense<[0, 0, 0, 0]> : tensor<4xindex>} : () -> !tosa.shape<4>588 %size = tosa.const_shape {values = dense<[0, -2, 9, 4]> : tensor<4xindex>} : () -> !tosa.shape<4>589 %2= tosa.slice %arg0, %start, %size : (tensor<8x8x8x?xi32>, !tosa.shape<4>, !tosa.shape<4>) -> tensor<?x?x?x?xi32>590 return591}592 593// -----594 595// CHECK-LABEL: @test_slice_start_out_of_bound596func.func @test_slice_start_out_of_bound(%arg0 : tensor<8x8x8x?xi32>) -> () {597 // CHECK: %[[START:.+]] = tosa.const_shape598 // CHECK: %[[SIZE:.+]] = tosa.const_shape599 // CHECK: %[[VAL:.+]] = tosa.slice %arg0, %[[START]], %[[SIZE]] : (tensor<8x8x8x?xi32>, !tosa.shape<4>, !tosa.shape<4>) -> tensor<?x?x?x4xi32>600 // this checks following601 // dim 0: start=-1 => inferred output dim is ?602 // dim 1: start=8 => inferred output dim is ?603 // dim 2: start+size out of bound: inferred output dim is ?604 // dim 3: start=8000000, size=4, input dim=? => inferred output dim is 4605 %start = tosa.const_shape {values = dense<[-1, 8, 6, 8000000]> : tensor<4xindex>} : () -> !tosa.shape<4>606 %size = tosa.const_shape {values = dense<[1, 1, 3, 4]> : tensor<4xindex>} : () -> !tosa.shape<4>607 %2= tosa.slice %arg0, %start, %size : (tensor<8x8x8x?xi32>, !tosa.shape<4>, !tosa.shape<4>) -> tensor<?x?x?x?xi32>608 return609}610 611// -----612 613// CHECK-LABEL: @test_slice_dynamic614func.func @test_slice_dynamic(%arg0 : tensor<10x?x2xf32>) -> () {615 // CHECK: %0 = tosa.const_shape {values = dense<[1, 0, 0]> : tensor<3xindex>}616 // CHECK: %1 = tosa.const_shape {values = dense<[7, -1, 1]> : tensor<3xindex>}617 // CHECK: %2 = tosa.slice %arg0, %0, %1 : (tensor<10x?x2xf32>, !tosa.shape<3>, !tosa.shape<3>) -> tensor<7x?x1xf32>618 %0 = tosa.const_shape {values = dense<[1, 0, 0]> : tensor<3xindex>} : () -> !tosa.shape<3>619 %1 = tosa.const_shape {values = dense<[7, -1, 1]> : tensor<3xindex>} : () -> !tosa.shape<3>620 %2= tosa.slice %arg0, %0, %1 : (tensor<10x?x2xf32>, !tosa.shape<3>, !tosa.shape<3>) -> tensor<?x?x?xf32>621 return622}623 624// -----625 626// CHECK-LABEL: @test_tile627func.func @test_tile(%arg0 : tensor<2x3x?xi32>) -> () {628 // CHECK: %[[CST:.*]] = tosa.const_shape {values = dense<[2, 1, 5]> : tensor<3xindex>} : () -> !tosa.shape<3>629 // CHECK: tosa.tile %arg0, %[[CST]] : (tensor<2x3x?xi32>, !tosa.shape<3>) -> tensor<4x3x?xi32>630 %cst = tosa.const_shape {values = dense<[2, 1, 5]> : tensor<3xindex>} : () -> !tosa.shape<3>631 %0 = tosa.tile %arg0, %cst : (tensor<2x3x?xi32>, !tosa.shape<3>) -> tensor<?x?x?xi32>632 return633}634 635// -----636 637// CHECK-LABEL: @test_tile_unknown_multiples638func.func @test_tile_unknown_multiples(%arg0 : tensor<2x3x?xi32>) -> () {639 // CHECK: %[[CST:.*]] = tosa.const_shape {values = dense<[2, -1, 5]> : tensor<3xindex>} : () -> !tosa.shape<3>640 // CHECK: tosa.tile %arg0, %[[CST]] : (tensor<2x3x?xi32>, !tosa.shape<3>) -> tensor<4x?x?xi32>641 %cst = tosa.const_shape {values = dense<[2, -1, 5]> : tensor<3xindex>} : () -> !tosa.shape<3>642 %0 = tosa.tile %arg0, %cst : (tensor<2x3x?xi32>, !tosa.shape<3>) -> tensor<?x?x?xi32>643 return644}645 646// -----647 648// CHECK-LABEL: @test_transpose_static649func.func @test_transpose_static(%arg0 : tensor<3x4x5xi32>) -> () {650 // CHECK: tosa.transpose %arg0 {perms = array<i32: 2, 1, 0>} : (tensor<3x4x5xi32>) -> tensor<5x4x3xi32>651 %1 = tosa.transpose %arg0 { perms = array<i32: 2, 1, 0> }: (tensor<3x4x5xi32>) -> tensor<?x?x?xi32>652 return653}654 655// -----656 657// CHECK-LABEL: @gather_static658func.func @gather_static(%arg0 : tensor<3x4x5xi32>, %arg1 : tensor<3x6xi32>) {659 // CHECK: tosa.gather %arg0, %arg1 : (tensor<3x4x5xi32>, tensor<3x6xi32>) -> tensor<3x6x5xi32>660 %0 = tosa.gather %arg0, %arg1 : (tensor<3x4x5xi32>, tensor<3x6xi32>) -> tensor<?x?x?xi32>661 return662}663 664// -----665 666// CHECK-LABEL: @gather_dynamic_values667func.func @gather_dynamic_values(%arg0 : tensor<?x?x?xi32>, %arg1 : tensor<3x6xi32>) {668 // CHECK: tosa.gather %arg0, %arg1 : (tensor<?x?x?xi32>, tensor<3x6xi32>) -> tensor<3x6x?xi32>669 %0 = tosa.gather %arg0, %arg1 : (tensor<?x?x?xi32>, tensor<3x6xi32>) -> tensor<?x?x?xi32>670 return671}672 673// -----674 675// CHECK-LABEL: @gather_dynamic_indices676func.func @gather_dynamic_indices(%arg0 : tensor<3x4x5xi32>, %arg1 : tensor<?x?xi32>) {677 // CHECK: tosa.gather %arg0, %arg1 : (tensor<3x4x5xi32>, tensor<?x?xi32>) -> tensor<3x?x5xi32>678 %0 = tosa.gather %arg0, %arg1 : (tensor<3x4x5xi32>, tensor<?x?xi32>) -> tensor<?x?x?xi32>679 return680}681 682// -----683 684// CHECK-LABEL: @gather_minimum_info685func.func @gather_minimum_info(%arg0 : tensor<3x?x5xi32>, %arg1 : tensor<?x6xi32>) {686 // CHECK: tosa.gather %arg0, %arg1 : (tensor<3x?x5xi32>, tensor<?x6xi32>) -> tensor<3x6x5xi32>687 %0 = tosa.gather %arg0, %arg1 : (tensor<3x?x5xi32>, tensor<?x6xi32>) -> tensor<?x?x?xi32>688 return689}690 691// -----692 693// CHECK-LABEL: @scatter_static694func.func @scatter_static(%arg0 : tensor<3x8x5xi32>, %arg1 : tensor<3x6xi32>, %arg2 : tensor<3x6x5xi32>) {695 // CHECK: tosa.scatter %arg0, %arg1, %arg2 : (tensor<3x8x5xi32>, tensor<3x6xi32>, tensor<3x6x5xi32>) -> tensor<3x8x5xi32>696 %0 = tosa.scatter %arg0, %arg1, %arg2 : (tensor<3x8x5xi32>, tensor<3x6xi32>, tensor<3x6x5xi32>) -> tensor<?x?x?xi32>697 return698}699 700// -----701 702// CHECK-LABEL: @scatter_static_values703func.func @scatter_static_values(%arg0 : tensor<3x4x5xi32>, %arg1 : tensor<?x?xi32>, %arg2 : tensor<?x?x?xi32>) {704 // CHECK: tosa.scatter %arg0, %arg1, %arg2 : (tensor<3x4x5xi32>, tensor<?x?xi32>, tensor<?x?x?xi32>) -> tensor<3x4x5xi32>705 %0 = tosa.scatter %arg0, %arg1, %arg2 : (tensor<3x4x5xi32>, tensor<?x?xi32>, tensor<?x?x?xi32>) -> tensor<?x?x?xi32>706 return707}708 709// -----710 711// CHECK-LABEL: @scatter_static_indices712func.func @scatter_static_indices(%arg0 : tensor<?x?x?xi32>, %arg1 : tensor<3x6xi32>, %arg2 : tensor<?x?x?xi32>) {713 // CHECK: tosa.scatter %arg0, %arg1, %arg2 : (tensor<?x?x?xi32>, tensor<3x6xi32>, tensor<?x?x?xi32>) -> tensor<3x?x?xi32>714 %0 = tosa.scatter %arg0, %arg1, %arg2 : (tensor<?x?x?xi32>, tensor<3x6xi32>, tensor<?x?x?xi32>) -> tensor<?x?x?xi32>715 return716}717 718// -----719 720// CHECK-LABEL: @scatter_static_input721func.func @scatter_static_input(%arg0 : tensor<?x?x?xi32>, %arg1 : tensor<?x?xi32>, %arg2 : tensor<3x6x5xi32>) {722 // CHECK: tosa.scatter %arg0, %arg1, %arg2 : (tensor<?x?x?xi32>, tensor<?x?xi32>, tensor<3x6x5xi32>) -> tensor<3x?x5xi32>723 %0 = tosa.scatter %arg0, %arg1, %arg2 : (tensor<?x?x?xi32>, tensor<?x?xi32>, tensor<3x6x5xi32>) -> tensor<?x?x?xi32>724 return725}726 727// -----728 729// CHECK-LABEL: @scatter_minimum_static730func.func @scatter_minimum_static(%arg0 : tensor<?x4x?xi32>, %arg1 : tensor<3x?xi32>, %arg2 : tensor<?x?x5xi32>) {731 // CHECK: tosa.scatter %arg0, %arg1, %arg2 : (tensor<?x4x?xi32>, tensor<3x?xi32>, tensor<?x?x5xi32>) -> tensor<3x4x5xi32>732 %0 = tosa.scatter %arg0, %arg1, %arg2 : (tensor<?x4x?xi32>, tensor<3x?xi32>, tensor<?x?x5xi32>) -> tensor<?x?x?xi32>733 return734}735 736// -----737 738// CHECK-LABEL: @test_pool_static739func.func @test_pool_static(%arg0: tensor<3x5x6x7xf32>) {740 %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>741 %output_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>742 743 // CHECK: -> tensor<3x2x4x7xf32>744 %0 = tosa.avg_pool2d %arg0, %input_zp, %output_zp {acc_type = f32, kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>745 746 // CHECK: -> tensor<3x2x4x7xf32>747 %1 = tosa.max_pool2d %arg0 {kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>) -> tensor<?x?x?x?xf32>748 return749}750 751// -----752 753// CHECK-LABEL: @conv2d_static754func.func @conv2d_static(%input: tensor<2x8x9x3xf32>, %weights: tensor<5x3x6x3xf32>, %bias: tensor<5xf32>, %input_zp: tensor<1xf32>, %weight_zp: tensor<1xf32>) -> () {755 // CHECK: -> tensor<2x6x4x5xf32>756 %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1>} : (tensor<2x8x9x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>757 return758}759 760// -----761 762// CHECK-LABEL: @conv2d_dynamic_input763func.func @conv2d_dynamic_input(%input: tensor<?x?x?x?xf32>, %weights: tensor<5x3x6x3xf32>, %bias: tensor<5xf32>, %input_zp: tensor<1xf32>, %weight_zp: tensor<1xf32>) -> () {764 // CHECK: -> tensor<?x?x?x5xf32>765 %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>766 return767}768 769// -----770 771// CHECK-LABEL: @test_pool_dynamic_input772func.func @test_pool_dynamic_input(%arg0: tensor<?x?x?x?xf32>) {773 %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>774 %output_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>775 776 // CHECK: -> tensor<?x?x?x?xf32>777 %0 = tosa.avg_pool2d %arg0, %input_zp, %output_zp {acc_type = f32, kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>778 779 // CHECK: -> tensor<?x?x?x?xf32>780 %1 = tosa.max_pool2d %arg0 {kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>781 return782}783 784// -----785 786// CHECK-LABEL: @test_pool_padded787func.func @test_pool_padded(%arg0: tensor<3x5x6x7xf32>) {788 %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>789 %output_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>790 791 // CHECK: -> tensor<3x7x5x7xf32>792 %0 = tosa.avg_pool2d %arg0, %input_zp, %output_zp {acc_type = f32, kernel = array<i64: 4, 3>, pad = array<i64: 3, 2, 1, 0>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>793 794 // CHECK: -> tensor<3x7x5x7xf32>795 %1 = tosa.max_pool2d %arg0 {kernel = array<i64: 4, 3>, pad = array<i64: 3, 2, 1, 0>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>) -> tensor<?x?x?x?xf32>796 return797}798 799// -----800 801// CHECK-LABEL: @conv2d_dynamic_weight802func.func @conv2d_dynamic_weight(%input: tensor<2x8x9x3xf32>, %weights: tensor<?x?x?x?xf32>, %bias: tensor<5xf32>, %input_zp: tensor<1xf32>, %weight_zp: tensor<1xf32>) -> () {803 // CHECK: -> tensor<2x?x?x5xf32>804 %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1>} : (tensor<2x8x9x3xf32>, tensor<?x?x?x?xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>805 return806}807 808// -----809 810// CHECK-LABEL: @conv2d_dynamic_bias811func.func @conv2d_dynamic_bias(%input: tensor<2x8x9x3xf32>, %weights: tensor<5x3x6x3xf32>, %bias: tensor<?xf32>, %input_zp: tensor<1xf32>, %weight_zp: tensor<1xf32>) -> () {812 // CHECK: -> tensor<2x6x4x5xf32>813 %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1>} : (tensor<2x8x9x3xf32>, tensor<5x3x6x3xf32>, tensor<?xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>814 return815}816 817// -----818 819// CHECK-LABEL: @test_pool_stride820func.func @test_pool_stride(%arg0: tensor<3x14x12x7xf32>) {821 %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>822 %output_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>823 824 // CHECK: -> tensor<3x6x4x7xf32>825 %0 = tosa.avg_pool2d %arg0, %input_zp, %output_zp {acc_type = f32, kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 3>} : (tensor<3x14x12x7xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>826 827 // CHECK: -> tensor<3x6x4x7xf32>828 %1 = tosa.max_pool2d %arg0 {kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 3>} : (tensor<3x14x12x7xf32>) -> tensor<?x?x?x?xf32>829 return830}831 832// -----833 834// CHECK-LABEL: @conv2d_padded835func.func @conv2d_padded(%input: tensor<2x8x9x3xf32>, %weights: tensor<5x3x6x3xf32>, %bias: tensor<5xf32>, %input_zp: tensor<1xf32>, %weight_zp: tensor<1xf32>) -> () {836 // CHECK: -> tensor<2x9x11x5xf32>837 %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 1, 2, 3, 4>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1>} : (tensor<2x8x9x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>838 return839}840 841// -----842 843// CHECK-LABEL: @conv2d_dilated844func.func @conv2d_dilated(%input: tensor<2x12x14x3xf32>, %weights: tensor<5x3x6x3xf32>, %bias: tensor<5xf32>, %input_zp: tensor<1xf32>, %weight_zp: tensor<1xf32>) -> () {845 // CHECK: -> tensor<2x6x4x5xf32>846 %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 3, 2>} : (tensor<2x12x14x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>847 return848}849 850// -----851 852// CHECK-LABEL: @conv2d_strided853func.func @conv2d_strided(%input: tensor<1x13x15x1xf32>, %weights: tensor<1x1x1x1xf32>, %bias: tensor<1xf32>, %input_zp: tensor<1xf32>, %weight_zp: tensor<1xf32>) -> () {854 // CHECK: -> tensor<1x5x8x1xf32>855 %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 3, 2>, dilation = array<i64: 1, 1>} : (tensor<1x13x15x1xf32>, tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?xf32>856 return857}858 859// -----860 861// CHECK-LABEL: @conv3d_static862func.func @conv3d_static(%input: tensor<2x8x9x10x3xf32>, %weights: tensor<5x3x6x4x3xf32>, %bias: tensor<5xf32>, %input_zp: tensor<1xf32>, %weight_zp: tensor<1xf32>) -> () {863 // CHECK: -> tensor<2x6x4x7x5xf32>864 %0 = tosa.conv3d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, dilation = array<i64: 1, 1, 1>, pad = array<i64: 0, 0, 0, 0, 0, 0>, stride = array<i64: 1, 1, 1>} : (tensor<2x8x9x10x3xf32>, tensor<5x3x6x4x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?x?xf32>865 return866}867 868// -----869 870// CHECK-LABEL: @conv3d_dynamic_input871func.func @conv3d_dynamic_input(%arg0: tensor<?x?x?x?x?xf32>, %arg1: tensor<5x3x6x4x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {872 // CHECK: -> tensor<?x?x?x?x5xf32>873 %0 = tosa.conv3d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1, 1>, pad = array<i64: 0, 0, 0, 0, 0, 0>, stride = array<i64: 1, 1, 1>} : (tensor<?x?x?x?x?xf32>, tensor<5x3x6x4x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?x?xf32>874 return875}876 877// -----878 879// CHECK-LABEL: @conv3d_dynamic_weight880func.func @conv3d_dynamic_weight(%arg0: tensor<2x8x9x10x3xf32>, %arg1: tensor<?x?x?x?x?xf32>, %arg2: tensor<7xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {881 // CHECK: -> tensor<2x?x?x?x7xf32>882 %0 = tosa.conv3d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1, 1>, pad = array<i64: 0, 0, 0, 0, 0, 0>, stride = array<i64: 1, 1, 1>} : (tensor<2x8x9x10x3xf32>, tensor<?x?x?x?x?xf32>, tensor<7xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?x?xf32>883 return884}885 886// -----887 888// CHECK-LABEL: @conv3d_dynamic_bias889func.func @conv3d_dynamic_bias(%arg0: tensor<2x8x9x10x3xf32>, %arg1: tensor<5x3x6x4x3xf32>, %arg2: tensor<?xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {890 // CHECK: -> tensor<2x6x4x7x5xf32>891 %0 = tosa.conv3d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1, 1>, pad = array<i64: 0, 0, 0, 0, 0, 0>, stride = array<i64: 1, 1, 1>} : (tensor<2x8x9x10x3xf32>, tensor<5x3x6x4x3xf32>, tensor<?xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?x?xf32>892 return893}894 895// -----896 897// CHECK-LABEL: @conv3d_padded898func.func @conv3d_padded(%arg0: tensor<2x8x9x10x3xf32>, %arg1: tensor<5x3x6x4x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {899 // CHECK: -> tensor<2x9x11x18x5xf32>900 %0 = tosa.conv3d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1, 1>, pad = array<i64: 1, 2, 3, 4, 5, 6>, stride = array<i64: 1, 1, 1>} : (tensor<2x8x9x10x3xf32>, tensor<5x3x6x4x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?x?xf32>901 return902}903 904// -----905 906// CHECK-LABEL: @conv3d_dilated907func.func @conv3d_dilated(%arg0: tensor<2x12x14x16x3xf32>, %arg1: tensor<5x3x6x2x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {908 // CHECK: -> tensor<2x6x4x12x5xf32>909 %0 = tosa.conv3d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 3, 2, 4>, pad = array<i64: 0, 0, 0, 0, 0, 0>, stride = array<i64: 1, 1, 1>} : (tensor<2x12x14x16x3xf32>, tensor<5x3x6x2x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?x?xf32>910 return911}912 913// -----914 915// CHECK-LABEL: @conv3d_strided916func.func @conv3d_strided(%arg0: tensor<1x13x17x17x1xf32>, %arg1: tensor<1x1x1x1x1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {917 // CHECK: -> tensor<1x5x9x5x1xf32>918 %0 = tosa.conv3d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1, 1>, pad = array<i64: 0, 0, 0, 0, 0, 0>, stride = array<i64: 3, 2, 4>} : (tensor<1x13x17x17x1xf32>, tensor<1x1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x?x?xf32>919 return920}921 922// -----923 924// CHECK-LABEL: @depthwise_conv2d_static925func.func @depthwise_conv2d_static(%arg0: tensor<2x8x9x3xf32>, %arg1: tensor<3x6x3x5xf32>, %arg2: tensor<15xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {926 // CHECK: -> tensor<2x6x4x15xf32>927 %0 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x8x9x3xf32>, tensor<3x6x3x5xf32>, tensor<15xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x6x4x15xf32>928 return929}930 931// -----932 933// CHECK-LABEL: @depthwise_conv2d_dynamic_input934func.func @depthwise_conv2d_dynamic_input(%arg0: tensor<?x?x?x?xf32>, %arg1: tensor<3x6x3x5xf32>, %arg2: tensor<15xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {935 // CHECK: -> tensor<?x?x?x15xf32>936 %0 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>, tensor<3x6x3x5xf32>, tensor<15xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x15xf32>937 return938}939 940// -----941 942// CHECK-LABEL: @depthwise_conv2d_dynamic_weight943func.func @depthwise_conv2d_dynamic_weight(%arg0: tensor<2x8x9x3xf32>, %arg1: tensor<?x?x?x?xf32>, %arg2: tensor<15xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {944 // CHECK: -> tensor<2x?x?x15xf32>945 %0 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x8x9x3xf32>, tensor<?x?x?x?xf32>, tensor<15xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x15xf32>946 return947}948 949// -----950 951// CHECK-LABEL: @depthwise_conv2d_dynamic_bias952func.func @depthwise_conv2d_dynamic_bias(%arg0: tensor<2x8x9x3xf32>, %arg1: tensor<3x6x3x5xf32>, %arg2: tensor<?xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {953 // CHECK: -> tensor<2x6x4x15xf32>954 %0 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x8x9x3xf32>, tensor<3x6x3x5xf32>, tensor<?xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x6x4x15xf32>955 return956}957 958// -----959 960// CHECK-LABEL: @depthwise_conv2d_padded961func.func @depthwise_conv2d_padded(%arg0: tensor<2x8x9x3xf32>, %arg1: tensor<3x6x3x5xf32>, %arg2: tensor<15xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {962 // CHECK: -> tensor<2x9x11x15xf32>963 %0 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 1, 2, 3, 4>, stride = array<i64: 1, 1>} : (tensor<2x8x9x3xf32>, tensor<3x6x3x5xf32>, tensor<15xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x9x11x15xf32>964 return965}966 967// -----968 969// CHECK-LABEL: @depthwise_conv2d_dilated970func.func @depthwise_conv2d_dilated(%arg0: tensor<2x12x14x3xf32>, %arg1: tensor<3x6x3x5xf32>, %arg2: tensor<15xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {971 // CHECK: -> tensor<2x6x4x15xf32>972 %0 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 3, 2>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x12x14x3xf32>, tensor<3x6x3x5xf32>, tensor<15xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x6x4x15xf32>973 return974}975 976// -----977 978// CHECK-LABEL: @depthwise_conv2d_strided979func.func @depthwise_conv2d_strided(%arg0: tensor<1x13x15x1xf32>, %arg1: tensor<1x1x1x1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {980 // CHECK: -> tensor<1x5x8x1xf32>981 %0 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 3, 2>} : (tensor<1x13x15x1xf32>, tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x5x8x1xf32>982 return983}984 985// -----986 987// CHECK-LABEL: @transpose_conv2d_out_shape988func.func @transpose_conv2d_out_shape(%arg0: tensor<2x?x?x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {989 // CHECK: -> tensor<2x8x9x5xf32>990 %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x?x?x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x8x9x5xf32>991 return992}993 994// -----995 996// CHECK-LABEL: @transpose_conv2d_static997func.func @transpose_conv2d_static(%arg0: tensor<2x16x14x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {998 // CHECK: -> tensor<2x18x19x5xf32>999 %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x16x14x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x5xf32>1000 return1001}1002 1003// -----1004 1005// CHECK-LABEL: @transpose_conv2d_static_strided1006func.func @transpose_conv2d_static_strided(%arg0: tensor<2x16x14x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {1007 // CHECK: -> tensor<2x33x45x5xf32>1008 %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 3>} : (tensor<2x16x14x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x5xf32>1009 return1010}1011 1012// -----1013 1014// CHECK-LABEL: @transpose_conv2d_dynamic_input1015func.func @transpose_conv2d_dynamic_input(%arg0: tensor<?x?x?x?xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {1016 // CHECK: -> tensor<?x?x?x5xf32>1017 %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x5xf32>1018 return1019}1020 1021// -----1022 1023// CHECK-LABEL: @transpose_conv2d_dynamic_weights1024func.func @transpose_conv2d_dynamic_weights(%arg0: tensor<2x6x4x3xf32>, %arg1: tensor<?x?x?x?xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {1025 // CHECK: -> tensor<2x?x?x5xf32>1026 %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x6x4x3xf32>, tensor<?x?x?x?xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x5xf32>1027 return1028}1029 1030// -----1031 1032// CHECK-LABEL: @transpose_conv2d_dynamic_bias1033func.func @transpose_conv2d_dynamic_bias(%arg0: tensor<2x6x4x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<?xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {1034 // CHECK: -> tensor<2x8x9x5xf32>1035 %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x6x4x3xf32>, tensor<5x3x6x3xf32>, tensor<?xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x8x9x5xf32>1036 return1037}1038 1039// -----1040 1041// CHECK-LABEL: @transpose_conv2d_padded1042func.func @transpose_conv2d_padded(%arg0: tensor<2x9x11x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {1043 // CHECK: -> tensor<2x12x19x5xf32>1044 %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 1, 0, 3, 0>, stride = array<i64: 1, 1>} : (tensor<2x9x11x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x12x19x5xf32>1045 return1046}1047 1048// -----1049 1050// CHECK-LABEL: @transpose_conv2d_strided1051func.func @transpose_conv2d_strided(%arg0: tensor<1x5x7x1xf32>, %arg1: tensor<1x1x1x1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {1052 // CHECK: -> tensor<1x13x13x1xf32>1053 %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 3, 2>} : (tensor<1x5x7x1xf32>, tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x13x13x1xf32>1054 return1055}1056 1057// -----1058 1059// CHECK-LABEL: @transpose_conv2d_dynamic_out_channels1060func.func @transpose_conv2d_dynamic_out_channels(%arg0: tensor<2x1x1x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {1061 // CHECK: -> tensor<2x3x6x5xf32>1062 %0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x1x1x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x3x6x?xf32>1063 return1064}1065 1066// -----1067 1068// CHECK-LABEL: @resize_int_horizontal1069func.func @resize_int_horizontal(%arg0: tensor<1x15x13x1xi8>) {1070 %scale = tosa.const_shape { values = dense<[11, 7, 89, 6]> : tensor<4xindex> } : () -> !tosa.shape<4>1071 %offset = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1072 %border = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1073 // CHECK: -> tensor<1x23x179x1xi8>1074 %0 = tosa.resize %arg0, %scale, %offset, %border {mode = NEAREST_NEIGHBOR} : (tensor<1x15x13x1xi8>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<?x?x?x?xi8>1075 return1076}1077 1078// -----1079 1080// CHECK-LABEL: @resize_int_vertical1081func.func @resize_int_vertical(%arg0: tensor<1x49x42x1xi16>) {1082 %scale = tosa.const_shape { values = dense<[37, 16, 219, 41]> : tensor<4xindex> } : () -> !tosa.shape<4>1083 %offset = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1084 %border = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1085 // CHECK: -> tensor<1x112x220x1xi16>1086 %0 = tosa.resize %arg0, %scale, %offset, %border {mode = NEAREST_NEIGHBOR} : (tensor<1x49x42x1xi16>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<?x?x?x?xi16>1087 return1088}1089 1090// -----1091 1092// CHECK-LABEL: @resize_int_power_of_two_upscale1093func.func @resize_int_power_of_two_upscale(%arg0: tensor<1x23x19x1xi8>) {1094 %scale = tosa.const_shape { values = dense<[16, 1, 16, 1]> : tensor<4xindex> } : () -> !tosa.shape<4>1095 %offset = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1096 %border = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1097 // CHECK: -> tensor<1x353x289x1xi32>1098 %0 = tosa.resize %arg0, %scale, %offset, %border {mode = BILINEAR} : (tensor<1x23x19x1xi8>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<?x?x?x?xi32>1099 return1100}1101 1102// -----1103 1104// CHECK-LABEL: @resize_int_power_of_two_upscale_offsetted1105func.func @resize_int_power_of_two_upscale_offsetted(%arg0: tensor<1x41x26x1xi16>) {1106 %scale = tosa.const_shape { values = dense<[16, 2, 16, 2]> : tensor<4xindex> } : () -> !tosa.shape<4>1107 %offset = tosa.const_shape { values = dense<[-7, -7]> : tensor<2xindex> } : () -> !tosa.shape<2>1108 %border = tosa.const_shape { values = dense<[7, 7]> : tensor<2xindex> } : () -> !tosa.shape<2>1109 // CHECK: -> tensor<1x328x208x1xi48>1110 %0 = tosa.resize %arg0, %scale, %offset, %border {mode = BILINEAR} : (tensor<1x41x26x1xi16>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<?x?x?x?xi48>1111 return1112}1113 1114// -----1115// CHECK-LABEL: @resize_fp_horizontal1116func.func @resize_fp_horizontal(%arg0: tensor<1x50x48x1xf32>) {1117 %scale = tosa.const_shape { values = dense<[15, 7, 84, 47]> : tensor<4xindex> } : () -> !tosa.shape<4>1118 %offset = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1119 %border = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1120 // CHECK: -> tensor<1x106x85x1xf32>1121 %0 = tosa.resize %arg0, %scale, %offset, %border {mode = BILINEAR} : (tensor<1x50x48x1xf32>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<?x?x?x?xf32>1122 return1123}1124 1125// -----1126// CHECK-LABEL: @resize_fp_vertical1127func.func @resize_fp_vertical(%arg0: tensor<1x50x48x1xf32>) {1128 %scale = tosa.const_shape { values = dense<[127, 49, 12, 47]> : tensor<4xindex> } : () -> !tosa.shape<4>1129 %offset = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1130 %border = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1131 // CHECK: -> tensor<1x128x13x1xf32>1132 %0 = tosa.resize %arg0, %scale, %offset, %border {mode = NEAREST_NEIGHBOR} : (tensor<1x50x48x1xf32>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<?x?x?x?xf32>1133 return1134}1135 1136// -----1137 1138// CHECK-LABEL: @resize_fp_power_of_two_upscale1139func.func @resize_fp_power_of_two_upscale(%arg0: tensor<1x23x23x1xf32>) {1140 %scale = tosa.const_shape { values = dense<[4, 1, 4, 1]> : tensor<4xindex> } : () -> !tosa.shape<4>1141 %offset = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1142 %border = tosa.const_shape { values = dense<0> : tensor<2xindex> } : () -> !tosa.shape<2>1143 // CHECK: -> tensor<1x89x89x1xf32>1144 %0 = tosa.resize %arg0, %scale, %offset, %border {mode = BILINEAR} : (tensor<1x23x23x1xf32>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<?x?x?x?xf32>1145 return1146}1147 1148// -----1149 1150// CHECK-LABEL: @resize_fp_power_of_two_upscale_offsetted1151func.func @resize_fp_power_of_two_upscale_offsetted(%arg0: tensor<1x50x48x1xf32>) {1152 %scale = tosa.const_shape { values = dense<[64, 2, 64, 2]> : tensor<4xindex> } : () -> !tosa.shape<4>1153 %offset = tosa.const_shape { values = dense<[-31, -31]> : tensor<2xindex> } : () -> !tosa.shape<2>1154 %border = tosa.const_shape { values = dense<[31, 31]> : tensor<2xindex> } : () -> !tosa.shape<2>1155 // CHECK: -> tensor<1x1600x1536x1xf32>1156 %0 = tosa.resize %arg0, %scale, %offset, %border {mode = NEAREST_NEIGHBOR} : (tensor<1x50x48x1xf32>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<?x?x?x?xf32>1157 return1158}1159 1160// -----1161 1162// CHECK-LABEL: @resize_negative_output_dim1163func.func @resize_negative_output_dim(%arg0: tensor<1x3x1x1xi8>) {1164 %scale = tosa.const_shape { values = dense<[1, 3, 1, 1]> : tensor<4xindex> } : () -> !tosa.shape<4>1165 %offset = tosa.const_shape { values = dense<[6, 1]> : tensor<2xindex> } : () -> !tosa.shape<2>1166 %border = tosa.const_shape { values = dense<[-15, 0]> : tensor<2xindex> } : () -> !tosa.shape<2>1167 // expected-error@+1 {{calculated output height and width must be non-negative, got height = -5, width = 0}}1168 %0 = tosa.resize %arg0, %scale, %offset, %border {mode = NEAREST_NEIGHBOR} : (tensor<1x3x1x1xi8>, !tosa.shape<4>, !tosa.shape<2>, !tosa.shape<2>) -> tensor<*xi8>1169 return1170}1171 1172// -----1173 1174// CHECK-LABEL: @if_test_simple1175func.func @if_test_simple(%arg0 : tensor<f32>, %arg1 : tensor<f32>, %arg2 : tensor<i1>) -> () {1176 %a = tosa.log %arg0 : (tensor<f32>) -> tensor<f32>1177 %b = tosa.log %arg1 : (tensor<f32>) -> tensor<f32>1178 1179 // CHECK: tosa.cond_if1180 // CHECK: -> tensor<f32>1181 %0 = tosa.cond_if %arg2 : tensor<i1> -> tensor<f32> {1182 tosa.yield %a : tensor<f32>1183 } else {1184 tosa.yield %b : tensor<f32>1185 }1186 return1187}1188 1189// -----1190 1191// CHECK-LABEL: @if_test_dynamic1192func.func @if_test_dynamic(%arg0 : tensor<2xf32>, %arg1 : tensor<3xf32>, %arg2 : tensor<i1>) -> () {1193 // CHECK: tosa.cond_if1194 // CHECK: -> tensor<?xf32>1195 %0 = tosa.cond_if %arg2 : tensor<i1> -> tensor<?xf32> {1196 tosa.yield %arg0 : tensor<2xf32>1197 } else {1198 tosa.yield %arg1 : tensor<3xf32>1199 }1200 return1201}1202 1203// -----1204 1205// CHECK-LABEL: @if_test_unranked1206func.func @if_test_unranked(%arg0 : tensor<f32>, %arg1 : tensor<3xf32>, %arg2 : tensor<i1>) -> () {1207 // CHECK: tosa.cond_if1208 // CHECK: -> tensor<*xf32>1209 %0 = tosa.cond_if %arg2 : tensor<i1> -> tensor<*xf32> {1210 tosa.yield %arg0 : tensor<f32>1211 } else {1212 tosa.yield %arg1 : tensor<3xf32>1213 }1214 return1215}1216 1217// -----1218 1219// CHECK-LABEL: @if_test_propagate1220func.func @if_test_propagate(%arg0 : tensor<f32>, %arg1 : tensor<f32>, %arg2 : tensor<i1>) -> () {1221 // CHECK: tosa.cond_if1222 // CHECK: -> tensor<f32>1223 %0 = tosa.cond_if %arg2 : tensor<i1> -> tensor<f32> {1224 %1 = tosa.add %arg0, %arg1 : (tensor<f32>, tensor<f32>) -> tensor<f32>1225 tosa.yield %1 : tensor<f32>1226 } else {1227 %1 = tosa.sub %arg0, %arg1 : (tensor<f32>, tensor<f32>) -> tensor<f32>1228 tosa.yield %1 : tensor<f32>1229 }1230 return1231}1232 1233// -----1234 1235// CHECK-LABEL: @while_test1236func.func @while_test(%arg0 : tensor<i32>) -> (tensor<*xi32>) {1237 // CHECK: tosa.add1238 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i32>1239 %0 = tosa.add %arg0, %arg0 : (tensor<i32>, tensor<i32>) -> tensor<*xi32>1240 1241 // CHECK: tosa.while_loop1242 // CHECK-SAME: (tensor<i32>) -> tensor<i32>1243 %1 = tosa.while_loop (%arg1 = %0) : (tensor<*xi32>) -> tensor<*xi32> {1244 %2 = "tosa.const"() <{values = dense<3> : tensor<i32>}> : () -> tensor<i32>1245 1246 // CHECK: tosa.greater_equal1247 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i1>1248 %3 = tosa.greater_equal %2, %arg1 : (tensor<i32>, tensor<*xi32>) -> tensor<*xi1>1249 1250 // CHECK: tosa.yield1251 // CHECK-SAME: tensor<i1>1252 tosa.yield %3 : tensor<*xi1>1253 1254 } do {1255 1256 // CHECK: ^bb01257 // CHECK-SAME: tensor<i32>1258 ^bb0(%arg1: tensor<*xi32>):1259 %2 = "tosa.const"() <{values = dense<1> : tensor<i32>}> : () -> tensor<i32>1260 1261 // CHECK: tosa.add1262 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i32>1263 %3 = tosa.add %arg1, %2 : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>1264 1265 // CHECK: tosa.yield1266 // CHECK-SAME: tensor<i32>1267 tosa.yield %3 : tensor<*xi32>1268 }1269 1270 // CHECK: tensor.cast1271 return %1 : tensor<*xi32>1272}1273 1274// -----1275 1276// CHECK-LABEL: @while_test1277func.func @while_test(%arg0 : tensor<i32>, %arg1 : tensor<1xi32>) -> () {1278 // CHECK: tosa.while_loop1279 // CHECK-SAME: (tensor<i32>, tensor<1xi32>) -> (tensor<i32>, tensor<?xi32>)1280 %0:2 = tosa.while_loop (%arg2 = %arg0, %arg3 = %arg1) : (tensor<i32>, tensor<1xi32>) -> (tensor<i32>, tensor<?xi32>) {1281 %1 = "tosa.const"() <{values = dense<3> : tensor<i32>}> : () -> tensor<i32>1282 // CHECK: tosa.greater_equal1283 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i1>1284 %2 = tosa.greater_equal %1, %arg2 : (tensor<i32>, tensor<i32>) -> tensor<i1>1285 1286 // CHECK: tosa.yield1287 // CHECK-SAME: tensor<i1>1288 tosa.yield %2 : tensor<i1>1289 } do {1290 1291 // CHECK: ^bb01292 // CHECK-SAME: tensor<i32>1293 // CHECK-SAME: tensor<?xi32>1294 ^bb0(%arg2: tensor<i32>, %arg3: tensor<?xi32>):1295 %1 = "tosa.const"() <{values = dense<1> : tensor<i32>}> : () -> tensor<i32>1296 1297 // CHECK: tosa.add1298 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i32>1299 %2 = tosa.add %arg2, %1 : (tensor<i32>, tensor<i32>) -> tensor<i32>1300 1301 // CHECK: tosa.concat1302 // CHECK-SAME: (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>1303 %3 = tosa.concat %arg3, %arg3 {axis = 0 : i32} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>1304 1305 // CHECK: tosa.yield1306 // CHECK-SAME: tensor<i32>1307 // CHECK-SAME: tensor<?xi32>1308 tosa.yield %2, %3 : tensor<i32>, tensor<?xi32>1309 }1310 return1311}1312 1313// -----1314 1315// This test locks down a fix for a crash in the type inference process.1316// The relevant pattern is a while loop whose body contains a TOSA operation which is1317// consumed by a non-inferrable user in the same body.1318// Previously, this would trigger a crash due to how types are cached and then1319// reapplied to the operations in the loops body.1320 1321// CHECK-LABEL: @while_dont_crash1322func.func @while_dont_crash(%arg0 : tensor<i32>) -> (tensor<*xi32>) {1323 %0 = tosa.add %arg0, %arg0 : (tensor<i32>, tensor<i32>) -> tensor<*xi32>1324 // CHECK: tosa.while_loop1325 // CHECK-SAME: (tensor<i32>) -> tensor<i32>1326 %1 = tosa.while_loop (%arg1 = %0) : (tensor<*xi32>) -> tensor<*xi32> {1327 %2 = "tosa.const"() <{values = dense<3> : tensor<i32>}> : () -> tensor<i32>1328 // CHECK: tosa.greater_equal1329 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i1>1330 %3 = tosa.greater_equal %2, %arg1 : (tensor<i32>, tensor<*xi32>) -> tensor<*xi1>1331 tosa.yield %3 : tensor<*xi1>1332 } do {1333 // CHECK: ^bb01334 // CHECK-SAME: tensor<i32>1335 ^bb0(%arg1: tensor<*xi32>):1336 // CHECK: tosa.add1337 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i32>1338 %3 = tosa.add %arg1, %arg1 : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>1339 // CHECK: %[[CAST:.+]] = tensor.cast %{{.*}} : tensor<i32> to tensor<*xi32>1340 // CHECK: "use"(%[[CAST]]) : (tensor<*xi32>) -> ()1341 "use"(%3) : (tensor<*xi32>) -> ()1342 tosa.yield %3 : tensor<*xi32>1343 }1344 // CHECK: tensor.cast1345 return %1 : tensor<*xi32>1346}1347 1348// -----1349 1350// This test locks down a fix for a crash in the type inference process.1351// The relevant pattern is a while loop whose body contains a TOSA operation which is1352// consumed by a non-inferrable user in the same body.1353 1354// CHECK-LABEL: @while_dont_crash_nested1355func.func @while_dont_crash_nested(%arg0 : tensor<i32>) -> (tensor<*xi32>) {1356 %0 = tosa.add %arg0, %arg0 : (tensor<i32>, tensor<i32>) -> tensor<*xi32>1357 // CHECK: tosa.while_loop1358 // CHECK-SAME: (tensor<i32>) -> tensor<i32>1359 %1 = tosa.while_loop (%arg1 = %0) : (tensor<*xi32>) -> tensor<*xi32> {1360 %2 = "tosa.const"() <{values = dense<3> : tensor<i32>}> : () -> tensor<i32>1361 // CHECK: tosa.greater_equal1362 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i1>1363 %3 = tosa.greater_equal %2, %arg1 : (tensor<i32>, tensor<*xi32>) -> tensor<*xi1>1364 // CHECK: tosa.yield1365 // CHECK-SAME: tensor<i1>1366 tosa.yield %3 : tensor<*xi1>1367 } do {1368 // CHECK: ^bb01369 // CHECK-SAME: tensor<i32>1370 ^bb0(%arg1: tensor<*xi32>):1371 // CHECK: tosa.while_loop1372 // CHECK-SAME: (tensor<i32>) -> tensor<i32>1373 %1 = tosa.while_loop (%arg2 = %arg1) : (tensor<*xi32>) -> tensor<*xi32> {1374 %2 = "tosa.const"() <{values = dense<3> : tensor<i32>}> : () -> tensor<i32>1375 // CHECK: tosa.greater_equal1376 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i1>1377 %4 = tosa.greater_equal %2, %arg2 : (tensor<i32>, tensor<*xi32>) -> tensor<*xi1>1378 // CHECK: tosa.yield1379 // CHECK-SAME: tensor<i1>1380 tosa.yield %4 : tensor<*xi1>1381 } do {1382 // CHECK: ^bb01383 // CHECK-SAME: tensor<i32>1384 ^bb0(%arg2: tensor<*xi32>):1385 // CHECK: tosa.add1386 // CHECK-SAME: (tensor<i32>, tensor<i32>) -> tensor<i32>1387 %4 = tosa.add %arg2, %arg2 : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>1388 // CHECK: %[[CAST:.+]] = tensor.cast %{{.*}} : tensor<i32> to tensor<*xi32>1389 // CHECK: "use"(%[[CAST]]) : (tensor<*xi32>) -> ()1390 "use"(%4) : (tensor<*xi32>) -> ()1391 // CHECK: tosa.yield1392 // CHECK-SAME: tensor<i32>1393 tosa.yield %4 : tensor<*xi32>1394 }1395 // CHECK: tosa.yield1396 // CHECK-SAME: tensor<i32>1397 tosa.yield %1 : tensor<*xi32>1398 }1399 1400 // CHECK: tensor.cast1401 return %1 : tensor<*xi32>1402}1403 1404// -----1405 1406// CHECK-LABEL: @test_static_rfft2d1407func.func @test_static_rfft2d(%arg0: tensor<5x2x8xf32>) -> () {1408 // CHECK: -> (tensor<5x2x5xf32>, tensor<5x2x5xf32>)1409 %output_real, %output_imag = tosa.rfft2d %arg0 : (tensor<5x2x8xf32>) -> (tensor<?x?x?xf32>, tensor<?x?x?xf32>)1410 return1411}1412 1413// -----1414 1415// CHECK-LABEL: @test_dynamic_batch_rfft2d1416func.func @test_dynamic_batch_rfft2d(%arg0 : tensor<?x2x4xf32>) -> () {1417 // CHECK: -> (tensor<?x2x3xf32>, tensor<?x2x3xf32>)1418 %output_real, %output_imag = tosa.rfft2d %arg0 : (tensor<?x2x4xf32>) -> (tensor<?x?x?xf32>, tensor<?x?x?xf32>)1419 return1420}1421 1422// -----1423 1424// CHECK-LABEL: @test_dynamic_width_rfft2d1425func.func @test_dynamic_width_rfft2d(%arg0 : tensor<5x2x?xf32>) -> () {1426 // CHECK: -> (tensor<5x2x?xf32>, tensor<5x2x?xf32>)1427 %output_real, %output_imag = tosa.rfft2d %arg0 : (tensor<5x2x?xf32>) -> (tensor<?x?x?xf32>, tensor<?x?x?xf32>)1428 return1429}1430 1431// -----1432 1433// CHECK-LABEL: @test_static_fft2d1434func.func @test_static_fft2d(%arg0: tensor<1x4x8xf32>, %arg1: tensor<1x4x8xf32>) -> (tensor<1x4x8xf32>, tensor<1x4x8xf32>) {1435 // CHECK: -> (tensor<1x4x8xf32>, tensor<1x4x8xf32>)1436 %output_real, %output_imag = tosa.fft2d %arg0, %arg1 {inverse = false} : (tensor<1x4x8xf32>, tensor<1x4x8xf32>) -> (tensor<1x4x8xf32>, tensor<1x4x8xf32>)1437 return %output_real, %output_imag : tensor<1x4x8xf32>, tensor<1x4x8xf32>1438}1439 1440// -----1441 1442// CHECK-LABEL: @test_dynamic_batch_fft2d1443func.func @test_dynamic_batch_fft2d(%arg0: tensor<?x4x8xf32>, %arg1: tensor<?x4x8xf32>) -> (tensor<?x4x8xf32>, tensor<?x4x8xf32>) {1444 // CHECK: -> (tensor<?x4x8xf32>, tensor<?x4x8xf32>)1445 %output_real, %output_imag = tosa.fft2d %arg0, %arg1 {inverse = false} : (tensor<?x4x8xf32>, tensor<?x4x8xf32>) -> (tensor<?x4x8xf32>, tensor<?x4x8xf32>)1446 return %output_real, %output_imag : tensor<?x4x8xf32>, tensor<?x4x8xf32>1447}1448 1449// -----1450 1451// CHECK-LABEL: @test_unranked_equal1452func.func @test_unranked_equal(%arg0 : tensor<*xf32>, %arg1 : tensor<f32>) -> () {1453 // CHECK: tosa.equal %arg0, %arg1 : (tensor<*xf32>, tensor<f32>) -> tensor<*xi1>1454 %0 = tosa.equal %arg0, %arg1 : (tensor<*xf32>, tensor<f32>) -> tensor<*xi1>1455 return1456}1457 1458// -----1459 1460// CHECK-LABEL: test_non_tosa_consumer_shape1461func.func @test_non_tosa_consumer_shape(%arg0: tensor<4x4xf32>) -> !shape.shape {1462 // CHECK: tosa.log %arg0 : (tensor<4x4xf32>) -> tensor<4x4xf32>1463 %0 = tosa.log %arg0 : (tensor<4x4xf32>) -> tensor<*xf32>1464 %1 = shape.shape_of %0 : tensor<*xf32> -> !shape.shape1465 return %1 : !shape.shape1466}1467 1468// -----1469 1470// CHECK-LABEL: test_non_tosa_consumer_shape1471func.func @test_non_tosa_consumer_shape2(%arg0: tensor<4x4xf32>) -> tensor<?xindex> {1472 // CHECK: tosa.log %arg0 : (tensor<4x4xf32>) -> tensor<4x4xf32>1473 %0 = tosa.log %arg0 : (tensor<4x4xf32>) -> tensor<*xf32>1474 %1 = shape.shape_of %0 : tensor<*xf32> -> tensor<?xindex>1475 return %1 : tensor<?xindex>1476}1477 1478// -----1479 1480// CHECK-LABEL: test_non_tosa_consumer_extract1481func.func @test_non_tosa_consumer_extract(%arg0: tensor<4x4xf32>, %arg1: index) -> f32 {1482 // CHECK: tosa.log %arg0 : (tensor<4x4xf32>) -> tensor<4x4xf32>1483 %0 = tosa.log %arg0 : (tensor<4x4xf32>) -> tensor<?x?xf32>1484 %1 = tensor.extract %0[%arg1, %arg1] : tensor<?x?xf32>1485 return %1 : f321486}1487 1488// -----1489 1490// CHECK-LABEL: test_non_tosa_consumer_still_propagates1491func.func @test_non_tosa_consumer_still_propagates(%arg0: tensor<1x1x8xf32>, %arg1: tensor<1x8x1xf32>) -> tensor<?x?xf32> {1492 // CHECK: tosa.matmul %arg0, %arg1, %0, %1 : (tensor<1x1x8xf32>, tensor<1x8x1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x1x1xf32>1493 %0 = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>1494 %1 = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>1495 %2 = tosa.matmul %arg0, %arg1, %0, %1 : (tensor<1x1x8xf32>, tensor<1x8x1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x1x1xf32>1496 %3 = arith.constant dense<[1, 1]> : tensor<2xindex>1497 %4 = tensor.reshape %2(%3) : (tensor<?x1x1xf32>, tensor<2xindex>) -> tensor<?x?xf32>1498 return %4 : tensor<?x?xf32>1499}1500 1501// -----1502 1503// CHECK-LABEL: test_tosa_use_def_chain1504func.func @test_tosa_use_def_chain(%arg0: tensor<1x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) -> tensor<?x16x16x16xf32> {1505 // CHECK: [[CONV:%.+]] = tosa.conv2d %arg0, %arg1, %arg21506 // CHECK: (tensor<1x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x32x32x16xf32>1507 %0 = tosa.conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 1, 1, 1, 1>, stride = array<i64: 1, 1>} : (tensor<1x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x32x32x16xf32>1508 // CHECK: tosa.max_pool2d [[CONV]]1509 // CHECK: (tensor<1x32x32x16xf32>) -> tensor<1x16x16x16xf32>1510 %1 = tosa.max_pool2d %0 {kernel = array<i64: 2, 2>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 2>} : (tensor<?x32x32x16xf32>) -> tensor<?x16x16x16xf32>1511 return %1 : tensor<?x16x16x16xf32>1512}1513 1514// -----1515 1516// This test locks two bug fixes manifested in the code below.1517//1518// 1. Context1519//1520// When shape propagation hits an operation that does not support shape1521// inference (here 'tensor.expand_shape'), it must revert the currently1522// inferred shape of its consumers back to the originally expected input1523// type to avoid potential op verification errors. This type reversal is1524// done through an additional 'tensor.cast' op.1525//1526//1527// 2. Preserving list of non-inferrable consumers1528//1529// When multiple non-inferrable consumers of a shape-inferred value are found1530// (here, the 2 occurrences of 'tensor.expand_shape' consuming the output of1531// 'tosa.cast'), their input argument ('%0') must be altered to consume the1532// output the new 'tensor.cast' op. While these replacements occur, the use list1533// of the producer ('tosa.cast') is also implicitly altered, invalidating any1534// iterators associated with it. It is therefore necessary to create a copy of1535// this use list ahead of time. Before this bug fix, the second1536// 'tensor.expand_shape' op below was not updated correctly.1537//1538// 3. Guaranteeing def-use order1539//1540// When emitting the 'tensor.cast' op, it is important to guarantee that its1541// output value is defined before all of its consumers (here, both of the1542// 'tensor.expand_shape' ops. In a previous version of the code, this insertion1543// occurred right before the first encountered consumer. Since use lists are1544// saved in reverse order, the 'tensor.cast' op was inserted before the second1545// 'tensor.expand_shape' op, leading to a def-use order violation when the1546// first 'tensor.expand_shape' op was later updated. The current implementation1547// sets the insertion point right after the producer of the last shape-inferred1548// value (here 'tosa.cast'), which guarantees correct def-use order for all1549// future operand updates.1550 1551// CHECK-LABEL: test_multiple_non_inferrable_consumers1552// CHECK-SAME: %[[ARG:.*]]: tensor<1x2x8xf32>1553func.func @test_multiple_non_inferrable_consumers(%arg0: tensor<1x2x8xf32>) {1554 // CHECK: %[[TOSA_CAST:.*]] = tosa.cast %[[ARG]] : (tensor<1x2x8xf32>) -> tensor<1x2x8xf32>1555 // CHECK: %[[TENSOR_CAST:.*]] = tensor.cast %[[TOSA_CAST]] : tensor<1x2x8xf32> to tensor<?x2x8xf32>1556 %0 = tosa.cast %arg0 : (tensor<1x2x8xf32>) -> tensor<?x2x8xf32>1557 1558 %c0 = arith.constant 0 : index1559 %dim = tensor.dim %0, %c0 : tensor<?x2x8xf32>1560 1561 // CHECK: tensor.expand_shape %[[TENSOR_CAST]]1562 // CHECK: tensor.expand_shape %[[TENSOR_CAST]]1563 %expanded_0 = tensor.expand_shape %0 [[0], [1, 2], [3]] output_shape [%dim, 1, 4, 8] : tensor<?x2x8xf32> into tensor<?x1x2x8xf32>1564 %expanded_1 = tensor.expand_shape %0 [[0], [1, 2], [3]] output_shape [%dim, 1, 4, 8] : tensor<?x2x8xf32> into tensor<?x1x2x8xf32>1565 return1566}1567 1568// -----1569// CHECK-LABEL: test_mul_scalar1570func.func @test_mul_scalar(%arg0: tensor<f32>, %arg1: tensor<f32>) -> tensor<*xf32> {1571 // CHECK: %[[SHIFT:.*]] = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>1572 // CHECK: tosa.mul %arg0, %arg1, %[[SHIFT]] : (tensor<f32>, tensor<f32>, tensor<1xi8>) -> tensor<f32>1573 %shift = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>1574 %0 = tosa.mul %arg0, %arg1, %shift : (tensor<f32>, tensor<f32>, tensor<1xi8>) -> tensor<*xf32>1575 return %0 : tensor<*xf32>1576}1577 1578// -----1579 1580// CHECK-LABEL: test_matmul_t_block_scaled_static1581func.func @test_matmul_t_block_scaled_static(%arg0: tensor<4x8x32xf8E4M3FN>, %arg1: tensor<4x8x1xf8E8M0FNU>, %arg2: tensor<1x16x32xf8E4M3FN>, %arg3: tensor<1x16x1xf8E8M0FNU>) -> tensor<?x?x?xf32> {1582 // CHECK: -> tensor<4x8x16xf32>1583 %0 = tosa.matmul_t_block_scaled %arg0, %arg1, %arg2, %arg3 {block_size = #tosa.block_size<BLOCK_SIZE_32>} : (tensor<4x8x32xf8E4M3FN>, tensor<4x8x1xf8E8M0FNU>, tensor<1x16x32xf8E4M3FN>, tensor<1x16x1xf8E8M0FNU>) -> tensor<?x?x?xf32>1584 return %0 : tensor<?x?x?xf32>1585}1586 1587// -----1588 1589// CHECK-LABEL: test_matmul_t_block_scaled_unranked_a_data1590func.func @test_matmul_t_block_scaled_unranked_a_data(%arg0: tensor<*xf8E4M3FN>, %arg1: tensor<4x8x1xf8E8M0FNU>, %arg2: tensor<4x16x32xf8E4M3FN>, %arg3: tensor<4x16x1xf8E8M0FNU>) -> tensor<?x?x?xf32> {1591 // CHECK: -> tensor<4x8x16xf32>1592 %0 = tosa.matmul_t_block_scaled %arg0, %arg1, %arg2, %arg3 {block_size = #tosa.block_size<BLOCK_SIZE_32>} : (tensor<*xf8E4M3FN>, tensor<4x8x1xf8E8M0FNU>, tensor<4x16x32xf8E4M3FN>, tensor<4x16x1xf8E8M0FNU>) -> tensor<?x?x?xf32>1593 return %0 : tensor<?x?x?xf32>1594}1595 1596// -----1597 1598// CHECK-LABEL: test_matmul_t_block_scaled_unranked_b_data_and_scale1599func.func @test_matmul_t_block_scaled_unranked_b_data_and_scale(%arg0: tensor<4x8x32xf8E4M3FN>, %arg1: tensor<4x8x1xf8E8M0FNU>, %arg2: tensor<*xf8E4M3FN>, %arg3: tensor<*xf8E8M0FNU>) -> tensor<?x?x?xf32> {1600 // CHECK: -> tensor<4x8x?xf32>1601 %0 = tosa.matmul_t_block_scaled %arg0, %arg1, %arg2, %arg3 {block_size = #tosa.block_size<BLOCK_SIZE_32>} : (tensor<4x8x32xf8E4M3FN>, tensor<4x8x1xf8E8M0FNU>, tensor<*xf8E4M3FN>, tensor<*xf8E8M0FNU>) -> tensor<?x?x?xf32>1602 return %0 : tensor<?x?x?xf32>1603}1604 1605// -----1606 1607// CHECK-LABEL: test_matmul_t_block_scaled_unranked_all1608func.func @test_matmul_t_block_scaled_unranked_all(%arg0: tensor<*xf8E4M3FN>, %arg1: tensor<*xf8E8M0FNU>, %arg2: tensor<*xf8E4M3FN>, %arg3: tensor<*xf8E8M0FNU>) -> tensor<?x?x?xf32> {1609 // CHECK: -> tensor<?x?x?xf32>1610 %0 = tosa.matmul_t_block_scaled %arg0, %arg1, %arg2, %arg3 {block_size = #tosa.block_size<BLOCK_SIZE_32>} : (tensor<*xf8E4M3FN>, tensor<*xf8E8M0FNU>, tensor<*xf8E4M3FN>, tensor<*xf8E8M0FNU>) -> tensor<?x?x?xf32>1611 return %0 : tensor<?x?x?xf32>1612}1613 1614// -----1615 1616// CHECK-LABEL: test_matmul_t_block_scaled_broadcast_b_data1617func.func @test_matmul_t_block_scaled_broadcast_b_data(%arg0: tensor<*xf8E4M3FN>, %arg1: tensor<*xf8E8M0FNU>, %arg2: tensor<1x4x32xf8E4M3FN>, %arg3: tensor<1x4x1xf8E8M0FNU>) -> tensor<?x?x?xf32> {1618 // CHECK: -> tensor<?x?x4xf32>1619 %0 = tosa.matmul_t_block_scaled %arg0, %arg1, %arg2, %arg3 {block_size = #tosa.block_size<BLOCK_SIZE_32>} : (tensor<*xf8E4M3FN>, tensor<*xf8E8M0FNU>, tensor<1x4x32xf8E4M3FN>, tensor<1x4x1xf8E8M0FNU>) -> tensor<?x?x?xf32>1620 return %0 : tensor<?x?x?xf32>1621}1622 1623// -----1624 1625// CHECK-LABEL: test_matmul_t_block_scaled_broadcast_b_scale1626func.func @test_matmul_t_block_scaled_broadcast_b_scale(%arg0: tensor<*xf8E4M3FN>, %arg1: tensor<*xf8E8M0FNU>, %arg2: tensor<*xf8E4M3FN>, %arg3: tensor<1x4x1xf8E8M0FNU>) -> tensor<?x?x?xf32> {1627 // CHECK: -> tensor<?x?x4xf32>1628 %0 = tosa.matmul_t_block_scaled %arg0, %arg1, %arg2, %arg3 {block_size = #tosa.block_size<BLOCK_SIZE_32>} : (tensor<*xf8E4M3FN>, tensor<*xf8E8M0FNU>, tensor<*xf8E4M3FN>, tensor<1x4x1xf8E8M0FNU>) -> tensor<?x?x?xf32>1629 return %0 : tensor<?x?x?xf32>1630}1631 1632// -----1633 1634// CHECK-LABEL: test_cast_from_block_scaled_static1635func.func @test_cast_from_block_scaled_static(%arg0: tensor<4x32xf4E2M1FN>, %arg1: tensor<4x1xf8E8M0FNU>) -> tensor<*xf32> {1636 // CHECK: -> tensor<4x32xf32>1637 %0 = tosa.cast_from_block_scaled %arg0, %arg1 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32} : (tensor<4x32xf4E2M1FN>, tensor<4x1xf8E8M0FNU>) -> tensor<*xf32>1638 return %0 : tensor<*xf32>1639}1640 1641// -----1642 1643// CHECK-LABEL: test_cast_from_block_scaled_unranked_input_scale1644func.func @test_cast_from_block_scaled_unranked_input_scale(%arg0: tensor<4x32xf4E2M1FN>, %arg1: tensor<*xf8E8M0FNU>) -> tensor<*xf32> {1645 // CHECK: -> tensor<4x32xf32>1646 %0 = tosa.cast_from_block_scaled %arg0, %arg1 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32} : (tensor<4x32xf4E2M1FN>, tensor<*xf8E8M0FNU>) -> tensor<*xf32>1647 return %0 : tensor<*xf32>1648}1649 1650// -----1651 1652// CHECK-LABEL: test_cast_to_block_scaled_static1653func.func @test_cast_to_block_scaled_static(%arg0: tensor<4x32xf32>) -> (tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>) {1654 // CHECK: -> (tensor<4x32xf4E2M1FN>, tensor<4x1xf8E8M0FNU>)1655 %0:2 = tosa.cast_to_block_scaled %arg0 {block_size = #tosa.block_size<BLOCK_SIZE_32>} : (tensor<4x32xf32>) -> (tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>)1656 return %0#0, %0#1 : tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>1657}1658 1659// -----1660 1661// CHECK-LABEL: test_cast_to_block_scaled_unranked1662func.func @test_cast_to_block_scaled_unranked(%arg0: tensor<*xf32>) -> (tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>) {1663 // CHECK: -> (tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>)1664 %0:2 = tosa.cast_to_block_scaled %arg0 {block_size = #tosa.block_size<BLOCK_SIZE_32>} : (tensor<*xf32>) -> (tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>)1665 return %0#0, %0#1 : tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>1666}1667 1668// -----1669 1670// CHECK-LABEL: test_cast_to_block_scaled_dynamic_scales1671func.func @test_cast_to_block_scaled_dynamic_scales(%arg0: tensor<4x?xf32>) -> (tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>) {1672 // CHECK: -> (tensor<4x?xf4E2M1FN>, tensor<4x?xf8E8M0FNU>)1673 %0:2 = tosa.cast_to_block_scaled %arg0 {block_size = #tosa.block_size<BLOCK_SIZE_32>} : (tensor<4x?xf32>) -> (tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>)1674 return %0#0, %0#1 : tensor<*xf4E2M1FN>, tensor<*xf8E8M0FNU>1675}1676