[frontend] Implement SiLU fusion with mul + sigmoid pattern

[GuoningHuang]

Fuse tosa.mul + tosa.sigmoid into a linalg-based SiLU op to enable efficient IR lowering and further backend optimization.

[linuxlonelyeagle]

Please rewrite the title as [frontend] Implement SiLU fusion with mul + sigmoid pattern

[linuxlonelyeagle]

add endline

[linuxlonelyeagle]

In my view, this is a completely useless check. You should also capture the SSA value.

[linuxlonelyeagle]

Keep the PR description concise—ideally under 50 characters. You should explain what you did and why you did it. This description will appear in the git log.

[GuoningHuang]

@linuxlonelyeagle Hi, thanks a lot for your suggestions!

Title: I'll update it to "[frontend] Implement SiLU fusion with mul + sigmoid pattern" for clarity and log consistency.

Newline: I'll add the missing newline at the end of the file.

Test check: I’ll revise the test in tests/Python/test_silu.py to also capture the SSA value as you suggested.

PR description: I'll shorten the description to under 50 characters and make sure it succinctly states “what” was done and “why”.

[linuxlonelyeagle]

# CHECK-LABEL: func @forward
#       CHECK:   %{{.*}} = tensor.empty() : tensor<4x4xf32>
#       CHECK:   %[[RES:.*]] = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel", "parallel"]} ins(%arg0 : tensor<4x4xf32>) outs(%0 : tensor<4x4xf32>) {
....
#       CHECK:   return %[[RES]] : tensor<4x4xf32>
#       CHECK: }

[linuxlonelyeagle]

The above is an example I provided.

• Eliminate unnecessary content from checks.
• Ensure the format of the check is sufficiently elegant.
• I believe you should also capture the SSA values within the generic op block.

[GuoningHuang]

is this ok?

# CHECK-LABEL: func.func @forward
# CHECK: %[[EMPTY:.*]] = tensor.empty() : tensor<4x4xf32>
# CHECK: %[[RES:.*]] = linalg.generic {.*} ins(%arg0 : tensor<4x4xf32>) outs(%[[EMPTY]] : tensor<4x4xf32>) {
# CHECK:   ^bb0(%in: f32, %out: f32):
# CHECK:     %[[NEG:.*]] = arith.negf %in : f32
# CHECK:     %[[EXP:.*]] = math.exp %[[NEG]] : f32
# CHECK:     %[[ONE:.*]] = arith.constant 1.000000e+00 : f32
# CHECK:     %[[ADD:.*]] = arith.addf %[[EXP]], %[[ONE]] : f32
# CHECK:     %[[DIV:.*]] = arith.divf %in, %[[ADD]] : f32
# CHECK:     linalg.yield %[[DIV]] : f32
# CHECK: } -> tensor<4x4xf32>
# CHECK: return %[[RES]] : tensor<4x4xf32>

[linuxlonelyeagle]

Take a close look at the example I provided: the colons in CHECK-LABEL and CHECK are aligned.

[GuoningHuang]

OK, thank you for pointing that out! Is this version ok? If so, I'll update my PR accordingly:

# CHECK-LABEL: func.func @forward
#       CHECK:   %[[EMPTY:.*]] = tensor.empty() : tensor<4x4xf32>
#       CHECK:   %[[RES:.*]] = linalg.generic {.*} ins(%arg0 : tensor<4x4xf32>) outs(%[[EMPTY]] : tensor<4x4xf32>) {
#       CHECK:   ^bb0(%in: f32, %out: f32):
#       CHECK:       %[[NEG:.*]] = arith.negf %in : f32 
#       CHECK:       %[[EXP:.*]] = math.exp %[[NEG]] : f32
#       CHECK:       %[[ONE:.*]] = arith.constant 1.000000e+00 : f32
#       CHECK:       %[[ADD:.*]] = arith.addf %[[EXP]], %[[ONE]] : f32
#       CHECK:       %[[DIV:.*]] = arith.divf %in, %[[ADD]] : f32
#       CHECK:       linalg.yield %[[DIV]] : f32
#       CHECK:    } -> tensor<4x4xf32>
#       CHECK: return %[[RES]] : tensor<4x4xf32>

[linuxlonelyeagle]

This is good.

[GuoningHuang]

ok，I have update my PR accordingly.

[zhanghb97]

@GuoningHuang, please check the CI error.

[GuoningHuang]

@zhanghb97 CI error has been fixed. Thanks for pointing it out!