Example
storylm0_tp_vocab.kina
let tp_logits = fn(
residual: tensor<F32, [1, 4, 4], @host>,
lm_head: tensor<F32, [4, 8], @host>
) -> tensor<F32, [1, 4, 8], @host> {
let lm0 = lm_head.slice(axis=1, start=0, stop=4) in
let lm1 = lm_head.slice(axis=1, start=4, stop=8) in
let lm0_b = lm0.broadcast([1, 4, 4]) in
let lm1_b = lm1.broadcast([1, 4, 4]) in
let z0 = residual @ lm0_b in
let z1 = residual @ lm1_b in
let logits = z0.concat(axis=2, z1) in
logits
} in
let storylm0_tp_vocab = fn(
token_ids: tensor<I64, [1, 4], @host>,
token_embedding: tensor<F32, [8, 4], @host>,
causal_mask: tensor<F32, [1, 1, 4, 4], @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, 4, 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 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, 4, 2, 2]).transpose([0, 2, 1, 3]) in
let k_t = k_flat.reshape([1, 4, 2, 2]).transpose([0, 2, 3, 1]) in
let v = v_flat.reshape([1, 4, 2, 2]).transpose([0, 2, 1, 3]) in
let scores = q @ k_t in
let scale_broadcast = attention_scale.broadcast([1, 2, 4, 4]) in
let mask_broadcast = causal_mask.broadcast([1, 2, 4, 4]) 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, 4, 4]) in
let projected = merged @ wo_broadcast in
let residual = x + projected in
let logits = tp_logits(residual, lm_head) in
logits
} in
let token_ids = tensor<I64, [1, 4], @host> in
let token_embedding = tensor<F32, [8, 4], @host> in
let causal_mask = tensor<F32, [1, 1, 4, 4], @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_tp_vocab(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<1x4x4xf32>, %126: tensor<4x8xf32>) -> tensor<1x4x8xf32> attributes {llvm.emit_c_interface} {
%127 = tensor.extract_slice %126[0, 0][4, 4][1, 1] : tensor<4x8xf32> to tensor<4x4xf32>
%128 = tensor.extract_slice %126[0, 4][4, 4][1, 1] : tensor<4x8xf32> to tensor<4x4xf32>
%129 = tensor.empty() : tensor<1x4x4xf32>
%130 = 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(%127 : tensor<4x4xf32>) outs(%129 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%131 = tensor.empty() : tensor<1x4x4xf32>
%132 = 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(%128 : tensor<4x4xf32>) outs(%131 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%133 = arith.constant dense<0.000000e+00> : tensor<1x4x4xf32>
%134 = linalg.batch_matmul ins(%125, %130 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%133 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
%135 = arith.constant dense<0.000000e+00> : tensor<1x4x4xf32>
%136 = linalg.batch_matmul ins(%125, %132 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%135 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
%137 = tensor.empty() : tensor<1x4x8xf32>
%138 = tensor.insert_slice %134 into %137[0, 0, 0] [1, 4, 4] [1, 1, 1] : tensor<1x4x4xf32> into tensor<1x4x8xf32>
%139 = tensor.insert_slice %136 into %138[0, 0, 4] [1, 4, 4] [1, 1, 1] : tensor<1x4x4xf32> into tensor<1x4x8xf32>
return %139 : tensor<1x4x8xf32>
}
func.func @fn_val_140(%141: tensor<1x4xi64>, %142: tensor<8x4xf32>, %143: tensor<1x1x4x4xf32>, %144: tensor<1x1x1x1xf32>, %145: tensor<4x4xf32>, %146: tensor<4x4xf32>, %147: tensor<4x4xf32>, %148: tensor<4x4xf32>, %149: tensor<4x8xf32>) -> tensor<1x4x8xf32> attributes {llvm.emit_c_interface} {
%150 = arith.constant 0 : i64
%151 = arith.constant 8 : i64
%152 = tensor.empty() : tensor<1x4xi64>
%153 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
} ins(%141 : tensor<1x4xi64>) outs(%152 : tensor<1x4xi64>) {
^bb0(%candidate: i64, %_: i64):
%154 = arith.cmpi sge, %candidate, %150 : i64
%155 = arith.cmpi slt, %candidate, %151 : i64
%156 = arith.andi %154, %155 : i1
cf.assert %156, "gather index out of bounds"
linalg.yield %candidate : i64
} -> tensor<1x4xi64>
%157 = tensor.empty() : tensor<1x4x4xf32>
%158 = 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(%153 : tensor<1x4xi64>) outs(%157 : tensor<1x4x4xf32>) {
^bb0(%idx: i64, %_: f32):
%159 = arith.index_cast %idx : i64 to index
%160 = linalg.index 2 : index
%161 = tensor.extract %142[%159, %160] : tensor<8x4xf32>
linalg.yield %161 : f32
} -> tensor<1x4x4xf32>
%162 = tensor.empty() : tensor<1x4x4xf32>
%163 = 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(%145 : tensor<4x4xf32>) outs(%162 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%164 = tensor.empty() : tensor<1x4x4xf32>
%165 = 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(%146 : tensor<4x4xf32>) outs(%164 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%166 = tensor.empty() : tensor<1x4x4xf32>
%167 = 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(%147 : tensor<4x4xf32>) outs(%166 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%168 = tensor.empty() : tensor<1x4x4xf32>
%169 = 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(%148 : tensor<4x4xf32>) outs(%168 : tensor<1x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x4x4xf32>
%170 = arith.constant dense<0.000000e+00> : tensor<1x4x4xf32>
%171 = linalg.batch_matmul ins(%158, %163 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%170 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
%172 = arith.constant dense<0.000000e+00> : tensor<1x4x4xf32>
%173 = linalg.batch_matmul ins(%158, %165 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%172 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
%174 = arith.constant dense<0.000000e+00> : tensor<1x4x4xf32>
%175 = linalg.batch_matmul ins(%158, %167 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%174 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
%176 = arith.constant dense<[1, 4, 2, 2]> : tensor<4xi32>
%177 = tensor.reshape %171(%176) : (tensor<1x4x4xf32>, tensor<4xi32>) -> tensor<1x4x2x2xf32>
%178 = tensor.empty() : tensor<1x2x4x2xf32>
%179 = linalg.transpose ins(%177 : tensor<1x4x2x2xf32>) outs(%178 : tensor<1x2x4x2xf32>) permutation = [0, 2, 1, 3]
%180 = arith.constant dense<[1, 4, 2, 2]> : tensor<4xi32>
%181 = tensor.reshape %173(%180) : (tensor<1x4x4xf32>, tensor<4xi32>) -> tensor<1x4x2x2xf32>
%182 = tensor.empty() : tensor<1x2x2x4xf32>
%183 = linalg.transpose ins(%181 : tensor<1x4x2x2xf32>) outs(%182 : tensor<1x2x2x4xf32>) permutation = [0, 2, 3, 1]
%184 = arith.constant dense<[1, 4, 2, 2]> : tensor<4xi32>
%185 = tensor.reshape %175(%184) : (tensor<1x4x4xf32>, tensor<4xi32>) -> tensor<1x4x2x2xf32>
%186 = tensor.empty() : tensor<1x2x4x2xf32>
%187 = linalg.transpose ins(%185 : tensor<1x4x2x2xf32>) outs(%186 : tensor<1x2x4x2xf32>) permutation = [0, 2, 1, 3]
%188 = tensor.collapse_shape %179 [[0, 1], [2], [3]] : tensor<1x2x4x2xf32> into tensor<2x4x2xf32>
%189 = tensor.collapse_shape %183 [[0, 1], [2], [3]] : tensor<1x2x2x4xf32> into tensor<2x2x4xf32>
%190 = arith.constant dense<0.000000e+00> : tensor<2x4x4xf32>
%191 = linalg.batch_matmul ins(%188, %189 : tensor<2x4x2xf32>, tensor<2x2x4xf32>) outs(%190 : tensor<2x4x4xf32>) -> tensor<2x4x4xf32>
%192 = tensor.expand_shape %191 [[0, 1], [2], [3]] : tensor<2x4x4xf32> into tensor<1x2x4x4xf32>
%193 = tensor.empty() : tensor<1x2x4x4xf32>
%194 = 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(%144 : tensor<1x1x1x1xf32>) outs(%193 : tensor<1x2x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x4x4xf32>
%195 = tensor.empty() : tensor<1x2x4x4xf32>
%196 = 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(%143 : tensor<1x1x4x4xf32>) outs(%195 : tensor<1x2x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x4x4xf32>
%197 = tensor.empty() : tensor<1x2x4x4xf32>
%198 = linalg.mul ins(%192, %194 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%197 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
%199 = tensor.empty() : tensor<1x2x4x4xf32>
%200 = linalg.add ins(%198, %196 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%199 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
%203 = arith.constant 0xFF800000 : f32
%201 = tensor.empty() : tensor<1x2x4x1xf32>
%202 = linalg.fill ins(%203 : f32) outs(%201 : tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32>
%204 = 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(%200 : tensor<1x2x4x4xf32>) outs(%202 : tensor<1x2x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%205 = arith.maximumf %a0, %a1 : f32
linalg.yield %205 : f32
} -> tensor<1x2x4x1xf32>
%206 = tensor.empty() : tensor<1x2x4x4xf32>
%207 = 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(%204 : tensor<1x2x4x1xf32>) outs(%206 : tensor<1x2x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x4x4xf32>
%208 = tensor.empty() : tensor<1x2x4x4xf32>
%209 = linalg.sub ins(%200, %207 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%208 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
%210 = tensor.empty() : tensor<1x2x4x4xf32>
%211 = linalg.exp ins(%209 : tensor<1x2x4x4xf32>) outs(%210 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
%214 = arith.constant 0xFF800000 : f32
%212 = tensor.empty() : tensor<1x2x4x1xf32>
%213 = linalg.fill ins(%214 : f32) outs(%212 : tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32>
%215 = 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(%200 : tensor<1x2x4x4xf32>) outs(%213 : tensor<1x2x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%216 = arith.maximumf %a0, %a1 : f32
linalg.yield %216 : f32
} -> tensor<1x2x4x1xf32>
%217 = tensor.empty() : tensor<1x2x4x4xf32>
%218 = 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(%215 : tensor<1x2x4x1xf32>) outs(%217 : tensor<1x2x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x4x4xf32>
%219 = tensor.empty() : tensor<1x2x4x4xf32>
%220 = linalg.sub ins(%200, %218 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%219 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
%221 = tensor.empty() : tensor<1x2x4x4xf32>
%222 = linalg.exp ins(%220 : tensor<1x2x4x4xf32>) outs(%221 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
%225 = arith.constant 0.000000e+00 : f32
%223 = tensor.empty() : tensor<1x2x4x1xf32>
%224 = linalg.fill ins(%225 : f32) outs(%223 : tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32>
%226 = 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(%222 : tensor<1x2x4x4xf32>) outs(%224 : tensor<1x2x4x1xf32>) {
^bb0(%a0: f32, %a1: f32):
%227 = arith.addf %a0, %a1 : f32
linalg.yield %227 : f32
} -> tensor<1x2x4x1xf32>
%228 = tensor.empty() : tensor<1x2x4x4xf32>
%229 = 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(%226 : tensor<1x2x4x1xf32>) outs(%228 : tensor<1x2x4x4xf32>) {
^bb0(%a0: f32, %a1: f32):
linalg.yield %a0 : f32
} -> tensor<1x2x4x4xf32>
%230 = tensor.empty() : tensor<1x2x4x4xf32>
%231 = linalg.div ins(%211, %229 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%230 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
%232 = tensor.collapse_shape %231 [[0, 1], [2], [3]] : tensor<1x2x4x4xf32> into tensor<2x4x4xf32>
%233 = tensor.collapse_shape %187 [[0, 1], [2], [3]] : tensor<1x2x4x2xf32> into tensor<2x4x2xf32>
%234 = arith.constant dense<0.000000e+00> : tensor<2x4x2xf32>
%235 = linalg.batch_matmul ins(%232, %233 : tensor<2x4x4xf32>, tensor<2x4x2xf32>) outs(%234 : tensor<2x4x2xf32>) -> tensor<2x4x2xf32>
%236 = tensor.expand_shape %235 [[0, 1], [2], [3]] : tensor<2x4x2xf32> into tensor<1x2x4x2xf32>
%237 = tensor.empty() : tensor<1x4x2x2xf32>
%238 = linalg.transpose ins(%236 : tensor<1x2x4x2xf32>) outs(%237 : tensor<1x4x2x2xf32>) permutation = [0, 2, 1, 3]
%239 = arith.constant dense<[1, 4, 4]> : tensor<3xi32>
%240 = tensor.reshape %238(%239) : (tensor<1x4x2x2xf32>, tensor<3xi32>) -> tensor<1x4x4xf32>
%241 = arith.constant dense<0.000000e+00> : tensor<1x4x4xf32>
%242 = linalg.batch_matmul ins(%240, %169 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%241 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
%243 = tensor.empty() : tensor<1x4x4xf32>
%244 = linalg.add ins(%158, %242 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%243 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
%245 = func.call @fn_val_124(%244, %149) : (tensor<1x4x4xf32>, tensor<4x8xf32>) -> tensor<1x4x8xf32>
return %245 : tensor<1x4x8xf32>
}
func.func @main() -> tensor<1x4x8xf32> {
%246 = arith.constant dense<0> : tensor<1x4xi64>
%247 = arith.constant dense<0.000000e+00> : tensor<8x4xf32>
%248 = arith.constant dense<0.000000e+00> : tensor<1x1x4x4xf32>
%249 = arith.constant dense<0.000000e+00> : tensor<1x1x1x1xf32>
%250 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
%251 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
%252 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
%253 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
%254 = arith.constant dense<0.000000e+00> : tensor<4x8xf32>
%255 = func.call @fn_val_140(%246, %247, %248, %249, %250, %251, %252, %253, %254) : (tensor<1x4xi64>, tensor<8x4xf32>, tensor<1x1x4x4xf32>, tensor<1x1x1x1xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x8xf32>) -> tensor<1x4x8xf32>
return %255 : tensor<1x4x8xf32>
}
}