Fix issues of tanh(x) when x=INF or -INF

[amdrexu]

We didn't check this special case. We just computed tanh(x) by sinh(x)/cosh(x). But when x=INF or -INF, the limit of tanh(x) is defined as follow:

lim(tanh(x)) = 1.0, x -> INF; lim(tanh(x)) = -1.0, x -> -INF

[amdvlk-admin]

Test summary for commit 50d32aa

CTS tests (Failed: 0/137949)

• Built with version 1.3.5.2
Ubuntu navi3x, Srdcvk
• Passed: 35353/68947 (51.3%)
• Failed: 0/68947 (0.0%)
• Not Supported: 33594/68947 (48.7%)
• Warnings: 0/68947 (0.0%)
Ubuntu navi2x, Srdcvk
• Passed: 35424/69002 (51.3%)
• Failed: 0/69002 (0.0%)
• Not Supported: 33578/69002 (48.7%)
• Warnings: 0/69002 (0.0%)

[nhaehnle]

I feel like I've had a discussion with @kezhaoAMD about this recently. But I think it was in an llpcfe context. Did you discuss with him? He had a solution using copysign which might be a little better.

[xazhangAMD]

I feel like I've had a discussion with @kezhaoAMD about this recently. But I think it was in an llpcfe context. Did you discuss with him? He had a solution using copysign which might be a little better.

Yes, llpcfe uses a copysign.

  Value *FPOne = ConstantFP::get(call->getType(), 1);
  Value *isInf = m_builder->CreateIsInf(val);
  Value *infResult = m_builder->CreateBinaryIntrinsic(Intrinsic::copysign, FPOne, val);
  return m_builder->CreateSelect(isInf, infResult, tanh);

[amdrexu]

I feel like I've had a discussion with @kezhaoAMD about this recently. But I think it was in an llpcfe context. Did you discuss with him? He had a solution using copysign which might be a little better.

Yes, llpcfe uses a copysign.

  Value *FPOne = ConstantFP::get(call->getType(), 1);
  Value *isInf = m_builder->CreateIsInf(val);
  Value *infResult = m_builder->CreateBinaryIntrinsic(Intrinsic::copysign, FPOne, val);
  return m_builder->CreateSelect(isInf, infResult, tanh);

Thank you. Yes, this is better and saves a v_cndmask. I will follow it.

[xazhangAMD]

I feel like I've had a discussion with @kezhaoAMD about this recently. But I think it was in an llpcfe context. Did you discuss with him? He had a solution using copysign which might be a little better.

Yes, llpcfe uses a copysign.

  Value *FPOne = ConstantFP::get(call->getType(), 1);
  Value *isInf = m_builder->CreateIsInf(val);
  Value *infResult = m_builder->CreateBinaryIntrinsic(Intrinsic::copysign, FPOne, val);
  return m_builder->CreateSelect(isInf, infResult, tanh);

Thank you. Yes, this is better and saves a v_cndmask. I will follow it.

I didn't add it into lgc because I was not sure whether GLSL spec requires this. I can drop the llpcfe approach when the lgc commit is merged.

[amdvlk-admin]

Test summary for commit db6e682

CTS tests (Failed: 0/137949)

• Built with version 1.3.5.2
Ubuntu navi3x, Srdcvk
• Passed: 35353/68947 (51.3%)
• Failed: 0/68947 (0.0%)
• Not Supported: 33594/68947 (48.7%)
• Warnings: 0/68947 (0.0%)
Ubuntu navi2x, Srdcvk
• Passed: 35424/69002 (51.3%)
• Failed: 0/69002 (0.0%)
• Not Supported: 33578/69002 (48.7%)
• Warnings: 0/69002 (0.0%)

[amdrexu]

I feel like I've had a discussion with @kezhaoAMD about this recently. But I think it was in an llpcfe context. Did you discuss with him? He had a solution using copysign which might be a little better.

Yes, llpcfe uses a copysign.

  Value *FPOne = ConstantFP::get(call->getType(), 1);
  Value *isInf = m_builder->CreateIsInf(val);
  Value *infResult = m_builder->CreateBinaryIntrinsic(Intrinsic::copysign, FPOne, val);
  return m_builder->CreateSelect(isInf, infResult, tanh);

Thank you. Yes, this is better and saves a v_cndmask. I will follow it.

I didn't add it into lgc because I was not sure whether GLSL spec requires this. I can drop the llpcfe approach when the lgc commit is merged.

Thank you. Actually, GLSL spec allows this but we didn't encounter such issues because there is no test.

[nhaehnle]

Thanks!