Example
storylm0_rollout.kina
let greedy_row = fn(
logits: tensor<F32, [1, 1, 8], @host>
) -> tensor<I64, [1, 1], @host> {
let tokens = logits.reduce(op=argmax, axis=2) in
tokens
} in
let sample_row = fn(
logits: tensor<F32, [1, 1, 8], @host>,
u: tensor<F32, [1, 1, 8], @host>,
temperature: tensor<F32, [1, 1], @host>
) -> tensor<I64, [1, 1], @host> {
let log_u = log(u) in
let neg_log_u = -log_u in
let log_neg_log_u = log(neg_log_u) in
let gumbel = -log_neg_log_u in
let t_b = temperature.broadcast([1, 1, 8]) in
let scaled = logits / t_b in
let scores = scaled + gumbel in
let tokens = scores.reduce(op=argmax, axis=2) in
tokens
} in
let storylm0_rollout = fn(
token_ids: tensor<I64, [1, 36], @host>,
token_embedding: tensor<F32, [8, 4], @host>,
causal_mask: tensor<F32, [1, 1, 36, 36], @host>,
attention_scale: tensor<F32, [1, 1, 1, 1], @host>,
wq: tensor<F32, [4, 4], @host>,
wk: tensor<F32, [4, 4], @host>,
wv: tensor<F32, [4, 4], @host>,
wo: tensor<F32, [4, 4], @host>,
lm_head: tensor<F32, [4, 8], @host>
) -> tensor<F32, [1, 36, 8], @host> {
let x = token_embedding.gather(token_ids) in
let wq_broadcast = wq.broadcast([1, 4, 4]) in
let wk_broadcast = wk.broadcast([1, 4, 4]) in
let wv_broadcast = wv.broadcast([1, 4, 4]) in
let wo_broadcast = wo.broadcast([1, 4, 4]) in
let lm_head_broadcast = lm_head.broadcast([1, 4, 8]) in
let q_flat = x @ wq_broadcast in
let k_flat = x @ wk_broadcast in
let v_flat = x @ wv_broadcast in
let q = q_flat.reshape([1, 36, 2, 2]).transpose([0, 2, 1, 3]) in
let k_t = k_flat.reshape([1, 36, 2, 2]).transpose([0, 2, 3, 1]) in
let v = v_flat.reshape([1, 36, 2, 2]).transpose([0, 2, 1, 3]) in
let scores = q @ k_t in
let scale_broadcast = attention_scale.broadcast([1, 2, 36, 36]) in
let mask_broadcast = causal_mask.broadcast([1, 2, 36, 36]) in
let masked_scores = (scores * scale_broadcast) + mask_broadcast in
let probs = masked_scores.softmax(axis=3) in
let context = probs @ v in
let merged = context.transpose([0, 2, 1, 3]).reshape([1, 36, 4]) in
let projected = merged @ wo_broadcast in
let residual = x + projected in
let logits = residual @ lm_head_broadcast in
logits
} in
let token_ids = tensor<I64, [1, 36], @host> in
let token_embedding = tensor<F32, [8, 4], @host> in
let causal_mask = tensor<F32, [1, 1, 36, 36], @host> in
let attention_scale = tensor<F32, [1, 1, 1, 1], @host> in
let wq = tensor<F32, [4, 4], @host> in
let wk = tensor<F32, [4, 4], @host> in
let wv = tensor<F32, [4, 4], @host> in
let wo = tensor<F32, [4, 4], @host> in
let lm_head = tensor<F32, [4, 8], @host> in
storylm0_rollout(token_ids, token_embedding, causal_mask, attention_scale, wq, wk, wv, wo, lm_head)
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<1x1x8xf32>) -> tensor<1x1xi64> attributes {llvm.emit_c_interface} {
%126 = arith.constant 0xFF800000 : f32
%127 = arith.constant 0 : i64
%128 = tensor.empty() : tensor<1x1xf32>
%129 = linalg.fill ins(%126 : f32) outs(%128 : tensor<1x1xf32>) -> tensor<1x1xf32>
%130 = tensor.empty() : tensor<1x1xi64>
%131 = linalg.fill ins(%127 : i64) outs(%130 : tensor<1x1xi64>) -> tensor<1x1xi64>
%132, %133 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%125 : tensor<1x1x8xf32>) outs(%129, %131 : tensor<1x1xf32>, tensor<1x1xi64>) {
^bb0(%a0: f32, %a1: f32, %a2: i64):
%134 = arith.cmpf ogt, %a0, %a1 : f32
%135 = arith.select %134, %a0, %a1 : f32
%136 = linalg.index 2 : index
%137 = arith.index_cast %136 : index to i64
%138 = arith.select %134, %137, %a2 : i64
linalg.yield %135, %138 : f32, i64
} -> (tensor<1x1xf32>, tensor<1x1xi64>)
return %133 : tensor<1x1xi64>
}
func.func @fn_val_139(%140: tensor<1x1x8xf32>, %141: tensor<1x1x8xf32>, %142: tensor<1x1xf32>) -> tensor<1x1xi64> attributes {llvm.emit_c_interface} {
%143 = tensor.empty() : tensor<1x1x8xf32>
%144 = linalg.log ins(%141 : tensor<1x1x8xf32>) outs(%143 : tensor<1x1x8xf32>) -> tensor<1x1x8xf32>
%145 = tensor.empty() : tensor<1x1x8xf32>
%146 = linalg.negf ins(%144 : tensor<1x1x8xf32>) outs(%145 : tensor<1x1x8xf32>) -> tensor<1x1x8xf32>
%147 = tensor.empty() : tensor<1x1x8xf32>
%148 = linalg.log ins(%146 : tensor<1x1x8xf32>) outs(%147 : tensor<1x1x8xf32>) -> tensor<1x1x8xf32>
%149 = tensor.empty() : tensor<1x1x8xf32>
%150 = linalg.negf ins(%148 : tensor<1x1x8xf32>) outs(%149 : tensor<1x1x8xf32>) -> tensor<1x1x8xf32>
%151 = tensor.empty() : tensor<1x1x8xf32>
%152 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (0, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%142 : tensor<1x1xf32>) outs(%151 : tensor<1x1x8xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x1x8xf32>
%153 = tensor.empty() : tensor<1x1x8xf32>
%154 = linalg.div ins(%140, %152 : tensor<1x1x8xf32>, tensor<1x1x8xf32>) outs(%153 : tensor<1x1x8xf32>) -> tensor<1x1x8xf32>
%155 = tensor.empty() : tensor<1x1x8xf32>
%156 = linalg.add ins(%154, %150 : tensor<1x1x8xf32>, tensor<1x1x8xf32>) outs(%155 : tensor<1x1x8xf32>) -> tensor<1x1x8xf32>
%157 = arith.constant 0xFF800000 : f32
%158 = arith.constant 0 : i64
%159 = tensor.empty() : tensor<1x1xf32>
%160 = linalg.fill ins(%157 : f32) outs(%159 : tensor<1x1xf32>) -> tensor<1x1xf32>
%161 = tensor.empty() : tensor<1x1xi64>
%162 = linalg.fill ins(%158 : i64) outs(%161 : tensor<1x1xi64>) -> tensor<1x1xi64>
%163, %164 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>],
iterator_types = ["parallel", "parallel", "reduction"]
} ins(%156 : tensor<1x1x8xf32>) outs(%160, %162 : tensor<1x1xf32>, tensor<1x1xi64>) {
^bb0(%a0: f32, %a1: f32, %a2: i64):
%165 = arith.cmpf ogt, %a0, %a1 : f32
%166 = arith.select %165, %a0, %a1 : f32
%167 = linalg.index 2 : index
%168 = arith.index_cast %167 : index to i64
%169 = arith.select %165, %168, %a2 : i64
linalg.yield %166, %169 : f32, i64
} -> (tensor<1x1xf32>, tensor<1x1xi64>)
return %164 : tensor<1x1xi64>
}
func.func @fn_val_170(%171: tensor<1x36xi64>, %172: tensor<8x4xf32>, %173: tensor<1x1x36x36xf32>, %174: tensor<1x1x1x1xf32>, %175: tensor<4x4xf32>, %176: tensor<4x4xf32>, %177: tensor<4x4xf32>, %178: tensor<4x4xf32>, %179: tensor<4x8xf32>) -> tensor<1x36x8xf32> attributes {llvm.emit_c_interface} {
%180 = arith.constant 0 : i64
%181 = arith.constant 8 : i64
%182 = tensor.empty() : tensor<1x36xi64>
%183 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
} ins(%171 : tensor<1x36xi64>) outs(%182 : tensor<1x36xi64>) {
^bb0(%candidate: i64, %_: i64):
%184 = arith.cmpi sge, %candidate, %180 : i64
%185 = arith.cmpi slt, %candidate, %181 : i64
%186 = arith.andi %184, %185 : i1
cf.assert %186, "gather index out of bounds"
linalg.yield %candidate : i64
} -> tensor<1x36xi64>
%187 = tensor.empty() : tensor<1x36x4xf32>
%188 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%183 : tensor<1x36xi64>) outs(%187 : tensor<1x36x4xf32>) {
^bb0(%idx: i64, %_: f32):
%189 = arith.index_cast %idx : i64 to index
%190 = linalg.index 2 : index
%191 = tensor.extract %172[%189, %190] : tensor<8x4xf32>
linalg.yield %191 : f32
} -> tensor<1x36x4xf32>
%192 = tensor.empty() : tensor<1x4x4xf32>
%193 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%175 : tensor<4x4xf32>) outs(%192 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%194 = tensor.empty() : tensor<1x4x4xf32>
%195 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%176 : tensor<4x4xf32>) outs(%194 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%196 = tensor.empty() : tensor<1x4x4xf32>
%197 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%177 : tensor<4x4xf32>) outs(%196 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%198 = tensor.empty() : tensor<1x4x4xf32>
%199 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%178 : tensor<4x4xf32>) outs(%198 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%200 = tensor.empty() : tensor<1x4x8xf32>
%201 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]
} ins(%179 : tensor<4x8xf32>) outs(%200 : tensor<1x4x8xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x8xf32>
%202 = arith.constant dense<0.000000e+00> : tensor<1x36x4xf32>
%203 = linalg.batch_matmul ins(%188, %193 : tensor<1x36x4xf32>, tensor<1x4x4xf32>) outs(%202 : tensor<1x36x4xf32>) -> tensor<1x36x4xf32>
%204 = arith.constant dense<0.000000e+00> : tensor<1x36x4xf32>
%205 = linalg.batch_matmul ins(%188, %195 : tensor<1x36x4xf32>, tensor<1x4x4xf32>) outs(%204 : tensor<1x36x4xf32>) -> tensor<1x36x4xf32>
%206 = arith.constant dense<0.000000e+00> : tensor<1x36x4xf32>
%207 = linalg.batch_matmul ins(%188, %197 : tensor<1x36x4xf32>, tensor<1x4x4xf32>) outs(%206 : tensor<1x36x4xf32>) -> tensor<1x36x4xf32>
%208 = arith.constant dense<[1, 36, 2, 2]> : tensor<4xi32>
%209 = tensor.reshape %203(%208) : (tensor<1x36x4xf32>, tensor<4xi32>) -> tensor<1x36x2x2xf32>
%210 = tensor.empty() : tensor<1x2x36x2xf32>
%211 = linalg.transpose ins(%209 : tensor<1x36x2x2xf32>) outs(%210 : tensor<1x2x36x2xf32>) permutation = [0, 2, 1, 3]
%212 = arith.constant dense<[1, 36, 2, 2]> : tensor<4xi32>
%213 = tensor.reshape %205(%212) : (tensor<1x36x4xf32>, tensor<4xi32>) -> tensor<1x36x2x2xf32>
%214 = tensor.empty() : tensor<1x2x2x36xf32>
%215 = linalg.transpose ins(%213 : tensor<1x36x2x2xf32>) outs(%214 : tensor<1x2x2x36xf32>) permutation = [0, 2, 3, 1]
%216 = arith.constant dense<[1, 36, 2, 2]> : tensor<4xi32>
%217 = tensor.reshape %207(%216) : (tensor<1x36x4xf32>, tensor<4xi32>) -> tensor<1x36x2x2xf32>
%218 = tensor.empty() : tensor<1x2x36x2xf32>
%219 = linalg.transpose ins(%217 : tensor<1x36x2x2xf32>) outs(%218 : tensor<1x2x36x2xf32>) permutation = [0, 2, 1, 3]
%220 = tensor.collapse_shape %211 [[0, 1], [2], [3]] : tensor<1x2x36x2xf32> into tensor<2x36x2xf32>
%221 = tensor.collapse_shape %215 [[0, 1], [2], [3]] : tensor<1x2x2x36xf32> into tensor<2x2x36xf32>
%222 = arith.constant dense<0.000000e+00> : tensor<2x36x36xf32>
%223 = linalg.batch_matmul ins(%220, %221 : tensor<2x36x2xf32>, tensor<2x2x36xf32>) outs(%222 : tensor<2x36x36xf32>) -> tensor<2x36x36xf32>
%224 = tensor.expand_shape %223 [[0, 1], [2], [3]] : tensor<2x36x36xf32> into tensor<1x2x36x36xf32>
%225 = tensor.empty() : tensor<1x2x36x36xf32>
%226 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, 0, 0, 0)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
iterator_types = ["parallel", "parallel", "parallel", "parallel"]
} ins(%174 : tensor<1x1x1x1xf32>) outs(%225 : tensor<1x2x36x36xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x36x36xf32>
%227 = tensor.empty() : tensor<1x2x36x36xf32>
%228 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, 0, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
iterator_types = ["parallel", "parallel", "parallel", "parallel"]
} ins(%173 : tensor<1x1x36x36xf32>) outs(%227 : tensor<1x2x36x36xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x36x36xf32>
%229 = tensor.empty() : tensor<1x2x36x36xf32>
%230 = linalg.mul ins(%224, %226 : tensor<1x2x36x36xf32>, tensor<1x2x36x36xf32>) outs(%229 : tensor<1x2x36x36xf32>) -> tensor<1x2x36x36xf32>
%231 = tensor.empty() : tensor<1x2x36x36xf32>
%232 = linalg.add ins(%230, %228 : tensor<1x2x36x36xf32>, tensor<1x2x36x36xf32>) outs(%231 : tensor<1x2x36x36xf32>) -> tensor<1x2x36x36xf32>
%235 = arith.constant 0xFF800000 : f32
%233 = tensor.empty() : tensor<1x2x36x1xf32>
%234 = linalg.fill ins(%235 : f32) outs(%233 : tensor<1x2x36x1xf32>) -> tensor<1x2x36x1xf32>
%236 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>],
iterator_types = ["parallel", "parallel", "parallel", "reduction"]
} ins(%232 : tensor<1x2x36x36xf32>) outs(%234 : tensor<1x2x36x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%237 = arith.maximumf %a0, %a1 : f32
linalg.yield %237 : f32
} -> tensor<1x2x36x1xf32>
%238 = tensor.empty() : tensor<1x2x36x36xf32>
%239 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, d1, d2, 0)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
iterator_types = ["parallel", "parallel", "parallel", "parallel"]
} ins(%236 : tensor<1x2x36x1xf32>) outs(%238 : tensor<1x2x36x36xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x36x36xf32>
%240 = tensor.empty() : tensor<1x2x36x36xf32>
%241 = linalg.sub ins(%232, %239 : tensor<1x2x36x36xf32>, tensor<1x2x36x36xf32>) outs(%240 : tensor<1x2x36x36xf32>) -> tensor<1x2x36x36xf32>
%242 = tensor.empty() : tensor<1x2x36x36xf32>
%243 = linalg.exp ins(%241 : tensor<1x2x36x36xf32>) outs(%242 : tensor<1x2x36x36xf32>) -> tensor<1x2x36x36xf32>
%246 = arith.constant 0xFF800000 : f32
%244 = tensor.empty() : tensor<1x2x36x1xf32>
%245 = linalg.fill ins(%246 : f32) outs(%244 : tensor<1x2x36x1xf32>) -> tensor<1x2x36x1xf32>
%247 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>],
iterator_types = ["parallel", "parallel", "parallel", "reduction"]
} ins(%232 : tensor<1x2x36x36xf32>) outs(%245 : tensor<1x2x36x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%248 = arith.maximumf %a0, %a1 : f32
linalg.yield %248 : f32
} -> tensor<1x2x36x1xf32>
%249 = tensor.empty() : tensor<1x2x36x36xf32>
%250 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, d1, d2, 0)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
iterator_types = ["parallel", "parallel", "parallel", "parallel"]
} ins(%247 : tensor<1x2x36x1xf32>) outs(%249 : tensor<1x2x36x36xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x36x36xf32>
%251 = tensor.empty() : tensor<1x2x36x36xf32>
%252 = linalg.sub ins(%232, %250 : tensor<1x2x36x36xf32>, tensor<1x2x36x36xf32>) outs(%251 : tensor<1x2x36x36xf32>) -> tensor<1x2x36x36xf32>
%253 = tensor.empty() : tensor<1x2x36x36xf32>
%254 = linalg.exp ins(%252 : tensor<1x2x36x36xf32>) outs(%253 : tensor<1x2x36x36xf32>) -> tensor<1x2x36x36xf32>
%257 = arith.constant 0.000000e+00 : f32
%255 = tensor.empty() : tensor<1x2x36x1xf32>
%256 = linalg.fill ins(%257 : f32) outs(%255 : tensor<1x2x36x1xf32>) -> tensor<1x2x36x1xf32>
%258 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>],
iterator_types = ["parallel", "parallel", "parallel", "reduction"]
} ins(%254 : tensor<1x2x36x36xf32>) outs(%256 : tensor<1x2x36x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%259 = arith.addf %a0, %a1 : f32
linalg.yield %259 : f32
} -> tensor<1x2x36x1xf32>
%260 = tensor.empty() : tensor<1x2x36x36xf32>
%261 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, d1, d2, 0)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
iterator_types = ["parallel", "parallel", "parallel", "parallel"]
} ins(%258 : tensor<1x2x36x1xf32>) outs(%260 : tensor<1x2x36x36xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x36x36xf32>
%262 = tensor.empty() : tensor<1x2x36x36xf32>
%263 = linalg.div ins(%243, %261 : tensor<1x2x36x36xf32>, tensor<1x2x36x36xf32>) outs(%262 : tensor<1x2x36x36xf32>) -> tensor<1x2x36x36xf32>
%264 = tensor.collapse_shape %263 [[0, 1], [2], [3]] : tensor<1x2x36x36xf32> into tensor<2x36x36xf32>
%265 = tensor.collapse_shape %219 [[0, 1], [2], [3]] : tensor<1x2x36x2xf32> into tensor<2x36x2xf32>
%266 = arith.constant dense<0.000000e+00> : tensor<2x36x2xf32>
%267 = linalg.batch_matmul ins(%264, %265 : tensor<2x36x36xf32>, tensor<2x36x2xf32>) outs(%266 : tensor<2x36x2xf32>) -> tensor<2x36x2xf32>
%268 = tensor.expand_shape %267 [[0, 1], [2], [3]] : tensor<2x36x2xf32> into tensor<1x2x36x2xf32>
%269 = tensor.empty() : tensor<1x36x2x2xf32>
%270 = linalg.transpose ins(%268 : tensor<1x2x36x2xf32>) outs(%269 : tensor<1x36x2x2xf32>) permutation = [0, 2, 1, 3]
%271 = arith.constant dense<[1, 36, 4]> : tensor<3xi32>
%272 = tensor.reshape %270(%271) : (tensor<1x36x2x2xf32>, tensor<3xi32>) -> tensor<1x36x4xf32>
%273 = arith.constant dense<0.000000e+00> : tensor<1x36x4xf32>
%274 = linalg.batch_matmul ins(%272, %199 : tensor<1x36x4xf32>, tensor<1x4x4xf32>) outs(%273 : tensor<1x36x4xf32>) -> tensor<1x36x4xf32>
%275 = tensor.empty() : tensor<1x36x4xf32>
%276 = linalg.add ins(%188, %274 : tensor<1x36x4xf32>, tensor<1x36x4xf32>) outs(%275 : tensor<1x36x4xf32>) -> tensor<1x36x4xf32>
%277 = arith.constant dense<0.000000e+00> : tensor<1x36x8xf32>
%278 = linalg.batch_matmul ins(%276, %201 : tensor<1x36x4xf32>, tensor<1x4x8xf32>) outs(%277 : tensor<1x36x8xf32>) -> tensor<1x36x8xf32>
return %278 : tensor<1x36x8xf32>
}
func.func @main() -> tensor<1x36x8xf32> {
%279 = arith.constant dense<0> : tensor<1x36xi64>
%280 = arith.constant dense<0.000000e+00> : tensor<8x4xf32>
%281 = arith.constant dense<0.000000e+00> : tensor<1x1x36x36xf32>
%282 = arith.constant dense<0.000000e+00> : tensor<1x1x1x1xf32>
%283 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
%284 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
%285 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
%286 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
%287 = arith.constant dense<0.000000e+00> : tensor<4x8xf32>
%288 = func.call @fn_val_170(%279, %280, %281, %282, %283, %284, %285, %286, %287) : (tensor<1x36xi64>, tensor<8x4xf32>, tensor<1x1x36x36xf32>, tensor<1x1x1x1xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x8xf32>) -> tensor<1x36x8xf32>
return %288 : tensor<1x36x8xf32>
}
}