From 7970fcff2540b7678cc07ed269ec7fb860a4d2de Mon Sep 17 00:00:00 2001 From: Alok Hota Date: Wed, 16 May 2018 11:14:19 -0500 Subject: swr/rast: fix VCVTPD2PS generation for AVX512 Reviewed-By: George Kyriazis --- .../swr/rasterizer/jitter/functionpasses/lower_x86.cpp | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) (limited to 'src/gallium/drivers/swr') diff --git a/src/gallium/drivers/swr/rasterizer/jitter/functionpasses/lower_x86.cpp b/src/gallium/drivers/swr/rasterizer/jitter/functionpasses/lower_x86.cpp index 3caea67e169..e0296f6255f 100644 --- a/src/gallium/drivers/swr/rasterizer/jitter/functionpasses/lower_x86.cpp +++ b/src/gallium/drivers/swr/rasterizer/jitter/functionpasses/lower_x86.cpp @@ -265,8 +265,16 @@ namespace SwrJit // Assuming the intrinsics are consistent and place the src operand and mask last in the argument list. if (mTarget == AVX512) { - args.push_back(GetZeroVec(vecWidth, pElemTy)); - args.push_back(GetMask(vecWidth)); + if (pFunc->getName().equals("meta.intrinsic.VCVTPD2PS")) { + args.push_back(GetZeroVec(W256, pCallInst->getType()->getScalarType())); + args.push_back(GetMask(W256)); + // for AVX512 VCVTPD2PS, we also have to add rounding mode + args.push_back(B->C(_MM_FROUND_TO_NEAREST_INT | + _MM_FROUND_NO_EXC)); + } else { + args.push_back(GetZeroVec(vecWidth, pElemTy)); + args.push_back(GetMask(vecWidth)); + } } return B->CALLA(pIntrin, args); -- cgit v1.2.3