Example
ad_nll.kina
let nll = fn(logits: tensor<F32, [1, 4, 8], @host>) -> tensor<F32, [1, 4], @host> {
let targets = tensor<I64, [1, 4], @host> in
let log_probs = logits.log_softmax(axis=2) in
let picked = log_probs.take_along_axis(targets, axis=2) in
let loss = -picked in
loss
} in
let logits = tensor<F32, [1, 4, 8], @host> in
handle nll(logits) with reverseMode
module {
func.func @fn_val_0(%1: tensor<1x32x64xf32>, %2: tensor<1x1x64xf32>) -> tensor<1x32x64xf32> attributes {llvm.emit_c_interface} {
%3 = tensor.empty() : tensor<1x32x64xf32>
%4 = linalg.mul ins(%1, %1 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%3 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
%7 = arith.constant 0.000000e+00 : f32
%5 = tensor.empty() : tensor<1x32x1xf32>
%6 = linalg.fill ins(%7 : f32) outs(%5 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
%8 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%4 : tensor<1x32x64xf32>) outs(%6 : tensor<1x32x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%9 = arith.addf %a0, %a1 : f32
linalg.yield %9 : f32
} -> tensor<1x32x1xf32>
%10 = arith.constant dense<0.000000e+00> : tensor<1x1x1xf32>
%11 = tensor.empty() : tensor<1x32x1xf32>
%12 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, 0, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%10 : tensor<1x1x1xf32>) outs(%11 : tensor<1x32x1xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x32x1xf32>
%13 = tensor.empty() : tensor<1x32x1xf32>
%14 = linalg.mul ins(%8, %12 : tensor<1x32x1xf32>, tensor<1x32x1xf32>) outs(%13 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
%15 = arith.constant dense<0.000000e+00> : tensor<1x1x1xf32>
%16 = tensor.empty() : tensor<1x32x1xf32>
%17 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, 0, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%15 : tensor<1x1x1xf32>) outs(%16 : tensor<1x32x1xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x32x1xf32>
%18 = tensor.empty() : tensor<1x32x1xf32>
%19 = linalg.add ins(%14, %17 : tensor<1x32x1xf32>, tensor<1x32x1xf32>) outs(%18 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
%23 = tensor.empty() : tensor<1x32x1xf32>
%24 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%19 : tensor<1x32x1xf32>) outs(%23 : tensor<1x32x1xf32>) {
^bb0(%x: f32, %_: f32):
%20 = math.sqrt %x : f32
%21 = arith.constant 1.000000e+00 : f32
%22 = arith.divf %21, %20 : f32
linalg.yield %22 : f32
} -> tensor<1x32x1xf32>
%25 = tensor.empty() : tensor<1x32x64xf32>
%26 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%24 : tensor<1x32x1xf32>) outs(%25 : tensor<1x32x64xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x32x64xf32>
%27 = tensor.empty() : tensor<1x32x64xf32>
%28 = linalg.mul ins(%1, %26 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%27 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
%29 = tensor.empty() : tensor<1x32x64xf32>
%30 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, 0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%2 : tensor<1x1x64xf32>) outs(%29 : tensor<1x32x64xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x32x64xf32>
%31 = tensor.empty() : tensor<1x32x64xf32>
%32 = linalg.mul ins(%28, %30 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%31 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
return %32 : tensor<1x32x64xf32>
}
func.func @fn_val_33(%34: tensor<1x32x64xf32>, %35: tensor<1x64x128xf32>, %36: tensor<1x128x64xf32>) -> tensor<1x32x64xf32> attributes {llvm.emit_c_interface} {
%37 = arith.constant dense<0.000000e+00> : tensor<1x32x128xf32>
%38 = linalg.batch_matmul ins(%34, %35 : tensor<1x32x64xf32>, tensor<1x64x128xf32>) outs(%37 : tensor<1x32x128xf32>) -> tensor<1x32x128xf32>
%44 = tensor.empty() : tensor<1x32x128xf32>
%45 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%38 : tensor<1x32x128xf32>) outs(%44 : tensor<1x32x128xf32>) {
^bb0(%a0: f32, %a1: f32):
%39 = arith.constant 1.000000e+00 : f32
%40 = arith.negf %a0 : f32
%41 = math.exp %40 : f32
%42 = arith.addf %39, %41 : f32
%43 = arith.divf %39, %42 : f32
linalg.yield %43 : f32
} -> tensor<1x32x128xf32>
%46 = tensor.empty() : tensor<1x32x128xf32>
%47 = linalg.mul ins(%38, %45 : tensor<1x32x128xf32>, tensor<1x32x128xf32>) outs(%46 : tensor<1x32x128xf32>) -> tensor<1x32x128xf32>
%48 = arith.constant dense<0.000000e+00> : tensor<1x32x64xf32>
%49 = linalg.batch_matmul ins(%47, %36 : tensor<1x32x128xf32>, tensor<1x128x64xf32>) outs(%48 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
return %49 : tensor<1x32x64xf32>
}
func.func @fn_val_50(%51: tensor<1x32x64xf32>, %52: tensor<1x64x64xf32>, %53: tensor<1x64x64xf32>, %54: tensor<1x64x64xf32>, %55: tensor<1x64x64xf32>) -> tensor<1x32x64xf32> attributes {llvm.emit_c_interface} {
%56 = arith.constant dense<0.000000e+00> : tensor<1x32x64xf32>
%57 = linalg.batch_matmul ins(%51, %52 : tensor<1x32x64xf32>, tensor<1x64x64xf32>) outs(%56 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
%58 = arith.constant dense<0.000000e+00> : tensor<1x32x64xf32>
%59 = linalg.batch_matmul ins(%51, %53 : tensor<1x32x64xf32>, tensor<1x64x64xf32>) outs(%58 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
%60 = arith.constant dense<0.000000e+00> : tensor<1x32x64xf32>
%61 = linalg.batch_matmul ins(%51, %54 : tensor<1x32x64xf32>, tensor<1x64x64xf32>) outs(%60 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
%62 = tensor.empty() : tensor<1x64x32xf32>
%63 = linalg.transpose ins(%59 : tensor<1x32x64xf32>) outs(%62 : tensor<1x64x32xf32>) permutation = [0, 2, 1]
%64 = arith.constant dense<0.000000e+00> : tensor<1x32x32xf32>
%65 = linalg.batch_matmul ins(%57, %63 : tensor<1x32x64xf32>, tensor<1x64x32xf32>) outs(%64 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
%66 = arith.constant dense<0.000000e+00> : tensor<1x1x1xf32>
%67 = tensor.empty() : tensor<1x32x32xf32>
%68 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, 0, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%66 : tensor<1x1x1xf32>) outs(%67 : tensor<1x32x32xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x32x32xf32>
%69 = tensor.empty() : tensor<1x32x32xf32>
%70 = linalg.mul ins(%65, %68 : tensor<1x32x32xf32>, tensor<1x32x32xf32>) outs(%69 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
%73 = arith.constant 0xFF800000 : f32
%71 = tensor.empty() : tensor<1x32x1xf32>
%72 = linalg.fill ins(%73 : f32) outs(%71 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
%74 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%70 : tensor<1x32x32xf32>) outs(%72 : tensor<1x32x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%75 = arith.maximumf %a0, %a1 : f32
linalg.yield %75 : f32
} -> tensor<1x32x1xf32>
%76 = tensor.empty() : tensor<1x32x32xf32>
%77 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%74 : tensor<1x32x1xf32>) outs(%76 : tensor<1x32x32xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x32x32xf32>
%78 = tensor.empty() : tensor<1x32x32xf32>
%79 = linalg.sub ins(%70, %77 : tensor<1x32x32xf32>, tensor<1x32x32xf32>) outs(%78 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
%80 = tensor.empty() : tensor<1x32x32xf32>
%81 = linalg.exp ins(%79 : tensor<1x32x32xf32>) outs(%80 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
%84 = arith.constant 0xFF800000 : f32
%82 = tensor.empty() : tensor<1x32x1xf32>
%83 = linalg.fill ins(%84 : f32) outs(%82 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
%85 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%70 : tensor<1x32x32xf32>) outs(%83 : tensor<1x32x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%86 = arith.maximumf %a0, %a1 : f32
linalg.yield %86 : f32
} -> tensor<1x32x1xf32>
%87 = tensor.empty() : tensor<1x32x32xf32>
%88 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%85 : tensor<1x32x1xf32>) outs(%87 : tensor<1x32x32xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x32x32xf32>
%89 = tensor.empty() : tensor<1x32x32xf32>
%90 = linalg.sub ins(%70, %88 : tensor<1x32x32xf32>, tensor<1x32x32xf32>) outs(%89 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
%91 = tensor.empty() : tensor<1x32x32xf32>
%92 = linalg.exp ins(%90 : tensor<1x32x32xf32>) outs(%91 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
%95 = arith.constant 0.000000e+00 : f32
%93 = tensor.empty() : tensor<1x32x1xf32>
%94 = linalg.fill ins(%95 : f32) outs(%93 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
%96 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%92 : tensor<1x32x32xf32>) outs(%94 : tensor<1x32x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%97 = arith.addf %a0, %a1 : f32
linalg.yield %97 : f32
} -> tensor<1x32x1xf32>
%98 = tensor.empty() : tensor<1x32x32xf32>
%99 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%96 : tensor<1x32x1xf32>) outs(%98 : tensor<1x32x32xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x32x32xf32>
%100 = tensor.empty() : tensor<1x32x32xf32>
%101 = linalg.div ins(%81, %99 : tensor<1x32x32xf32>, tensor<1x32x32xf32>) outs(%100 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
%102 = arith.constant dense<0.000000e+00> : tensor<1x32x64xf32>
%103 = linalg.batch_matmul ins(%101, %61 : tensor<1x32x32xf32>, tensor<1x32x64xf32>) outs(%102 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
%104 = arith.constant dense<0.000000e+00> : tensor<1x32x64xf32>
%105 = linalg.batch_matmul ins(%103, %55 : tensor<1x32x64xf32>, tensor<1x64x64xf32>) outs(%104 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
return %105 : tensor<1x32x64xf32>
}
func.func @fn_val_106(%107: tensor<1x32x64xf32>, %108: tensor<1x1x64xf32>, %109: tensor<1x64x64xf32>, %110: tensor<1x64x64xf32>, %111: tensor<1x64x64xf32>, %112: tensor<1x64x64xf32>, %113: tensor<1x1x64xf32>, %114: tensor<1x64x128xf32>, %115: tensor<1x128x64xf32>) -> tensor<1x32x64xf32> attributes {llvm.emit_c_interface} {
%116 = func.call @fn_val_0(%107, %108) : (tensor<1x32x64xf32>, tensor<1x1x64xf32>) -> tensor<1x32x64xf32>
%117 = func.call @fn_val_50(%116, %109, %110, %111, %112) : (tensor<1x32x64xf32>, tensor<1x64x64xf32>, tensor<1x64x64xf32>, tensor<1x64x64xf32>, tensor<1x64x64xf32>) -> tensor<1x32x64xf32>
%118 = tensor.empty() : tensor<1x32x64xf32>
%119 = linalg.add ins(%107, %117 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%118 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
%120 = func.call @fn_val_0(%119, %113) : (tensor<1x32x64xf32>, tensor<1x1x64xf32>) -> tensor<1x32x64xf32>
%121 = func.call @fn_val_33(%120, %114, %115) : (tensor<1x32x64xf32>, tensor<1x64x128xf32>, tensor<1x128x64xf32>) -> tensor<1x32x64xf32>
%122 = tensor.empty() : tensor<1x32x64xf32>
%123 = linalg.add ins(%119, %121 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%122 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
return %123 : tensor<1x32x64xf32>
}
func.func @fn_val_124(%125: tensor<1x4x8xf32>) -> tensor<1x4xf32> attributes {llvm.emit_c_interface} {
%126 = arith.constant dense<0> : tensor<1x4xi64>
%129 = arith.constant 0xFF800000 : f32
%127 = tensor.empty() : tensor<1x4x1xf32>
%128 = linalg.fill ins(%129 : f32) outs(%127 : tensor<1x4x1xf32>) -> tensor<1x4x1xf32>
%130 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%125 : tensor<1x4x8xf32>) outs(%128 : tensor<1x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%131 = arith.maximumf %a0, %a1 : f32
linalg.yield %131 : f32
} -> tensor<1x4x1xf32>
%132 = tensor.empty() : tensor<1x4x8xf32>
%133 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%130 : tensor<1x4x1xf32>) outs(%132 : tensor<1x4x8xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x8xf32>
%134 = tensor.empty() : tensor<1x4x8xf32>
%135 = linalg.sub ins(%125, %133 : tensor<1x4x8xf32>, tensor<1x4x8xf32>) outs(%134 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%138 = arith.constant 0xFF800000 : f32
%136 = tensor.empty() : tensor<1x4x1xf32>
%137 = linalg.fill ins(%138 : f32) outs(%136 : tensor<1x4x1xf32>) -> tensor<1x4x1xf32>
%139 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%125 : tensor<1x4x8xf32>) outs(%137 : tensor<1x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%140 = arith.maximumf %a0, %a1 : f32
linalg.yield %140 : f32
} -> tensor<1x4x1xf32>
%141 = tensor.empty() : tensor<1x4x8xf32>
%142 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%139 : tensor<1x4x1xf32>) outs(%141 : tensor<1x4x8xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x8xf32>
%143 = tensor.empty() : tensor<1x4x8xf32>
%144 = linalg.sub ins(%125, %142 : tensor<1x4x8xf32>, tensor<1x4x8xf32>) outs(%143 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%145 = tensor.empty() : tensor<1x4x8xf32>
%146 = linalg.exp ins(%144 : tensor<1x4x8xf32>) outs(%145 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%149 = arith.constant 0.000000e+00 : f32
%147 = tensor.empty() : tensor<1x4x1xf32>
%148 = linalg.fill ins(%149 : f32) outs(%147 : tensor<1x4x1xf32>) -> tensor<1x4x1xf32>
%150 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%146 : tensor<1x4x8xf32>) outs(%148 : tensor<1x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%151 = arith.addf %a0, %a1 : f32
linalg.yield %151 : f32
} -> tensor<1x4x1xf32>
%152 = tensor.empty() : tensor<1x4x8xf32>
%153 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%150 : tensor<1x4x1xf32>) outs(%152 : tensor<1x4x8xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x8xf32>
%154 = tensor.empty() : tensor<1x4x8xf32>
%155 = linalg.log ins(%153 : tensor<1x4x8xf32>) outs(%154 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%156 = tensor.empty() : tensor<1x4x8xf32>
%157 = linalg.sub ins(%135, %155 : tensor<1x4x8xf32>, tensor<1x4x8xf32>) outs(%156 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%158 = arith.constant 0 : i64
%159 = arith.constant 8 : i64
%160 = tensor.empty() : tensor<1x4xi64>
%161 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
} ins(%126 : tensor<1x4xi64>) outs(%160 : tensor<1x4xi64>) {
^bb0(%candidate: i64, %_: i64):
%162 = arith.cmpi sge, %candidate, %158 : i64
%163 = arith.cmpi slt, %candidate, %159 : i64
%164 = arith.andi %162, %163 : i1
cf.assert %164, "take_along_axis index out of bounds"
linalg.yield %candidate : i64
} -> tensor<1x4xi64>
%165 = tensor.empty() : tensor<1x4xf32>
%166 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
} ins(%161 : tensor<1x4xi64>) outs(%165 : tensor<1x4xf32>) {
^bb0(%idx: i64, %_: f32):
%167 = linalg.index 0 : index
%168 = linalg.index 1 : index
%169 = arith.index_cast %idx : i64 to index
%170 = tensor.extract %157[%167, %168, %169] : tensor<1x4x8xf32>
linalg.yield %170 : f32
} -> tensor<1x4xf32>
%171 = tensor.empty() : tensor<1x4xf32>
%172 = linalg.negf ins(%166 : tensor<1x4xf32>) outs(%171 : tensor<1x4xf32>) -> tensor<1x4xf32>
return %172 : tensor<1x4xf32>
}
func.func @main() -> (tensor<1x4xf32>, tensor<1x4x8xf32>) {
%173 = arith.constant dense<0.000000e+00> : tensor<1x4x8xf32>
%174 = arith.constant dense<0> : tensor<1x4xi64>
%177 = arith.constant 0xFF800000 : f32
%175 = tensor.empty() : tensor<1x4x1xf32>
%176 = linalg.fill ins(%177 : f32) outs(%175 : tensor<1x4x1xf32>) -> tensor<1x4x1xf32>
%178 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%173 : tensor<1x4x8xf32>) outs(%176 : tensor<1x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%179 = arith.maximumf %a0, %a1 : f32
linalg.yield %179 : f32
} -> tensor<1x4x1xf32>
%180 = tensor.empty() : tensor<1x4x8xf32>
%181 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%178 : tensor<1x4x1xf32>) outs(%180 : tensor<1x4x8xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x8xf32>
%182 = tensor.empty() : tensor<1x4x8xf32>
%183 = linalg.sub ins(%173, %181 : tensor<1x4x8xf32>, tensor<1x4x8xf32>) outs(%182 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%186 = arith.constant 0xFF800000 : f32
%184 = tensor.empty() : tensor<1x4x1xf32>
%185 = linalg.fill ins(%186 : f32) outs(%184 : tensor<1x4x1xf32>) -> tensor<1x4x1xf32>
%187 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%173 : tensor<1x4x8xf32>) outs(%185 : tensor<1x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%188 = arith.maximumf %a0, %a1 : f32
linalg.yield %188 : f32
} -> tensor<1x4x1xf32>
%189 = tensor.empty() : tensor<1x4x8xf32>
%190 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%187 : tensor<1x4x1xf32>) outs(%189 : tensor<1x4x8xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x8xf32>
%191 = tensor.empty() : tensor<1x4x8xf32>
%192 = linalg.sub ins(%173, %190 : tensor<1x4x8xf32>, tensor<1x4x8xf32>) outs(%191 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%193 = tensor.empty() : tensor<1x4x8xf32>
%194 = linalg.exp ins(%192 : tensor<1x4x8xf32>) outs(%193 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%197 = arith.constant 0.000000e+00 : f32
%195 = tensor.empty() : tensor<1x4x1xf32>
%196 = linalg.fill ins(%197 : f32) outs(%195 : tensor<1x4x1xf32>) -> tensor<1x4x1xf32>
%198 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%194 : tensor<1x4x8xf32>) outs(%196 : tensor<1x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%199 = arith.addf %a0, %a1 : f32
linalg.yield %199 : f32
} -> tensor<1x4x1xf32>
%200 = tensor.empty() : tensor<1x4x8xf32>
%201 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%198 : tensor<1x4x1xf32>) outs(%200 : tensor<1x4x8xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x8xf32>
%202 = tensor.empty() : tensor<1x4x8xf32>
%203 = linalg.log ins(%201 : tensor<1x4x8xf32>) outs(%202 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%204 = tensor.empty() : tensor<1x4x8xf32>
%205 = linalg.sub ins(%183, %203 : tensor<1x4x8xf32>, tensor<1x4x8xf32>) outs(%204 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%206 = arith.constant 0 : i64
%207 = arith.constant 8 : i64
%208 = tensor.empty() : tensor<1x4xi64>
%209 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
} ins(%174 : tensor<1x4xi64>) outs(%208 : tensor<1x4xi64>) {
^bb0(%candidate: i64, %_: i64):
%210 = arith.cmpi sge, %candidate, %206 : i64
%211 = arith.cmpi slt, %candidate, %207 : i64
%212 = arith.andi %210, %211 : i1
cf.assert %212, "take_along_axis index out of bounds"
linalg.yield %candidate : i64
} -> tensor<1x4xi64>
%213 = tensor.empty() : tensor<1x4xf32>
%214 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
} ins(%209 : tensor<1x4xi64>) outs(%213 : tensor<1x4xf32>) {
^bb0(%idx: i64, %_: f32):
%215 = linalg.index 0 : index
%216 = linalg.index 1 : index
%217 = arith.index_cast %idx : i64 to index
%218 = tensor.extract %205[%215, %216, %217] : tensor<1x4x8xf32>
linalg.yield %218 : f32
} -> tensor<1x4xf32>
%219 = tensor.empty() : tensor<1x4xf32>
%220 = linalg.negf ins(%214 : tensor<1x4xf32>) outs(%219 : tensor<1x4xf32>) -> tensor<1x4xf32>
%221 = arith.constant dense<1.000000e+00> : tensor<1x4xf32>
%222 = arith.constant dense<1.000000e+00> : tensor<1x4xf32>
%223 = tensor.empty() : tensor<1x4xf32>
%224 = linalg.negf ins(%222 : tensor<1x4xf32>) outs(%223 : tensor<1x4xf32>) -> tensor<1x4xf32>
%225 = arith.constant 0 : i64
%226 = arith.constant 8 : i64
%227 = tensor.empty() : tensor<1x4xi64>
%228 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
} ins(%174 : tensor<1x4xi64>) outs(%227 : tensor<1x4xi64>) {
^bb0(%candidate: i64, %_: i64):
%229 = arith.cmpi sge, %candidate, %225 : i64
%230 = arith.cmpi slt, %candidate, %226 : i64
%231 = arith.andi %229, %230 : i1
cf.assert %231, "scatter_along_axis index out of bounds"
linalg.yield %candidate : i64
} -> tensor<1x4xi64>
%232 = tensor.empty() : tensor<1x4x8xf32>
%233 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%228, %224 : tensor<1x4xi64>, tensor<1x4xf32>) outs(%232 : tensor<1x4x8xf32>) {
^bb0(%idx: i64, %upd: f32, %_: f32):
%234 = linalg.index 2 : index
%235 = arith.index_cast %234 : index to i64
%236 = arith.cmpi eq, %idx, %235 : i64
%237 = arith.constant 0.000000e+00 : f32
%238 = arith.select %236, %upd, %237 : f32
linalg.yield %238 : f32
} -> tensor<1x4x8xf32>
%239 = tensor.empty() : tensor<1x4x8xf32>
%240 = linalg.exp ins(%205 : tensor<1x4x8xf32>) outs(%239 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%243 = arith.constant 0.000000e+00 : f32
%241 = tensor.empty() : tensor<1x4x1xf32>
%242 = linalg.fill ins(%243 : f32) outs(%241 : tensor<1x4x1xf32>) -> tensor<1x4x1xf32>
%244 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%233 : tensor<1x4x8xf32>) outs(%242 : tensor<1x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%245 = arith.addf %a0, %a1 : f32
linalg.yield %245 : f32
} -> tensor<1x4x1xf32>
%246 = tensor.empty() : tensor<1x4x8xf32>
%247 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%244 : tensor<1x4x1xf32>) outs(%246 : tensor<1x4x8xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x8xf32>
%248 = tensor.empty() : tensor<1x4x8xf32>
%249 = linalg.mul ins(%247, %240 : tensor<1x4x8xf32>, tensor<1x4x8xf32>) outs(%248 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
%250 = tensor.empty() : tensor<1x4x8xf32>
%251 = linalg.sub ins(%233, %249 : tensor<1x4x8xf32>, tensor<1x4x8xf32>) outs(%250 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
return %220, %251 : tensor<1x4xf32>, tensor<1x4x8xf32>
}
}