[Mesa-dev] [PATCH 04/51] glsl: Print 16-bit constants

[Pohjolainen, Topi]

On Mon, Nov 27, 2017 at 08:41:42PM +0200, Pohjolainen, Topi wrote:
> On Mon, Nov 27, 2017 at 11:40:49AM -0800, Matt Turner wrote:
> > On 11/24, Topi Pohjolainen wrote:
> > > ---
> > > src/compiler/glsl/ir_print_visitor.cpp | 1 +
> > > 1 file changed, 1 insertion(+)
> > > 
> > > diff --git a/src/compiler/glsl/ir_print_visitor.cpp b/src/compiler/glsl/ir_print_visitor.cpp
> > > index ea14cdeb6c..ab9a35d73f 100644
> > > --- a/src/compiler/glsl/ir_print_visitor.cpp
> > > +++ b/src/compiler/glsl/ir_print_visitor.cpp
> > > @@ -482,6 +482,7 @@ void ir_print_visitor::visit(ir_constant *ir)
> > > 	 case GLSL_TYPE_UINT:  fprintf(f, "%u", ir->value.u[i]); break;
> > > 	 case GLSL_TYPE_INT:   fprintf(f, "%d", ir->value.i[i]); break;
> > > 	 case GLSL_TYPE_FLOAT:
> > > +	 case GLSL_TYPE_FLOAT16:
> > 
> > Are fp16 values stored as 32-bit floats in ir->value.f[]? That seems
> > surprising.
> 
> Right. This needs more thought I guess. In the GLSL lowering pass I'm only
> changing types at the moment and the conversion of constants from 32-bits
> to 16-bits later in GLSL -> NIR. See patch:
> 
> glsl: Add support for 16-bit float constants in nir-conversion

Okay, coming back to this. At the moment there is no separate storage for
16-bit float in "union ir_constant_data". I thought about adding it but then
couldn't see the added value as we might as well store 16-bit constants with
32-bit precision. They are originally 32-bit anyway - it is the lowering pass
that might change the type to 16-bit.

What do you think?