[Mesa-dev] [PATCH] radv: flush f32->f16 conversion denormals to zero.

[Matt Arsenault]

> On Mar 16, 2017, at 20:02, Dave Airlie <airlied at gmail.com> wrote:
> 
> From: Dave Airlie <airlied at redhat.com>
> 
> SPIR-V defines the f32->f16 operation as flushing denormals to 0,
> this compares the class using amd class opcode.
> 
> Thanks to Matt Arsenault for figuring it out.
> 
> This fixes:
> dEQP-VK.spirv_assembly.instruction.compute.opquantize.flush_to_zero
> 
> Signed-off-by: Dave Airlie <airlied at redhat.com>
> ---
> src/amd/common/ac_nir_to_llvm.c |  9 ++++++++-
> src/amd/common/sid.h            | 13 +++++++++++++
> 2 files changed, 21 insertions(+), 1 deletion(-)
> 
> diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
> index 77e3a85..ac80677 100644
> --- a/src/amd/common/ac_nir_to_llvm.c
> +++ b/src/amd/common/ac_nir_to_llvm.c
> @@ -1027,11 +1027,18 @@ static LLVMValueRef emit_f2f16(struct nir_to_llvm_context *ctx,
> {
> 	LLVMValueRef result;
> 	LLVMValueRef cond;
> +	LLVMValueRef args[2];
> 	src0 = to_float(ctx, src0);
> 	result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
> -	result = ac_build_intrinsic(&ctx->ac, "llvm.canonicalize.f16", ctx->f16, &result, 1, AC_FUNC_ATTR_READNONE);
> +	LLVMValueRef mask = LLVMConstInt(ctx->i32, N_SUBNORMAL | P_SUBNORMAL, false);
> +	

I don’t think you need the canonicalize here. This will also only work on VI+ which supports f16 instructions

-Matt