[Mesa-dev] [PATCH 07/12] nir: Allow [iu]mul_high on non-32-bit types

[Ian Romanick]

On 10/09/2018 10:17 AM, Jason Ekstrand wrote:
> On Tue, Oct 9, 2018 at 11:32 AM Ian Romanick <idr at freedesktop.org
> <mailto:idr at freedesktop.org>> wrote:
> 
>     On 10/08/2018 02:14 PM, Jason Ekstrand wrote:
>     > On Mon, Oct 8, 2018 at 3:46 PM Ian Romanick <idr at freedesktop.org
>     <mailto:idr at freedesktop.org>
>     > <mailto:idr at freedesktop.org <mailto:idr at freedesktop.org>>> wrote:
>     >
>     >     On 10/05/2018 09:10 PM, Jason Ekstrand wrote:
>     >     > ---
>     >     >  src/compiler/nir/nir_constant_expressions.py |  1 +
>     >     >  src/compiler/nir/nir_opcodes.py              | 43
>     >     ++++++++++++++++++--
>     >     >  2 files changed, 40 insertions(+), 4 deletions(-)
>     >     >
>     >     > diff --git a/src/compiler/nir/nir_constant_expressions.py
>     >     b/src/compiler/nir/nir_constant_expressions.py
>     >     > index 118af9f7818..afc0739e8b2 100644
>     >     > --- a/src/compiler/nir/nir_constant_expressions.py
>     >     > +++ b/src/compiler/nir/nir_constant_expressions.py
>     >     > @@ -79,6 +79,7 @@ template = """\
>     >     >  #include <math.h>
>     >     >  #include "util/rounding.h" /* for _mesa_roundeven */
>     >     >  #include "util/half_float.h"
>     >     > +#include "util/bigmath.h"
>     >     >  #include "nir_constant_expressions.h"
>     >     > 
>     >     >  /**
>     >     > diff --git a/src/compiler/nir/nir_opcodes.py
>     >     b/src/compiler/nir/nir_opcodes.py
>     >     > index 4ef4ecc6f22..209f0c5509b 100644
>     >     > --- a/src/compiler/nir/nir_opcodes.py
>     >     > +++ b/src/compiler/nir/nir_opcodes.py
>     >     > @@ -443,12 +443,47 @@ binop("isub", tint, "", "src0 - src1")
>     >     >  binop("fmul", tfloat, commutative + associative, "src0 * src1")
>     >     >  # low 32-bits of signed/unsigned integer multiply
>     >     >  binop("imul", tint, commutative + associative, "src0 * src1")
>     >     > +
>     >     >  # high 32-bits of signed integer multiply
>     >     > -binop("imul_high", tint32, commutative,
>     >     > -      "(int32_t)(((int64_t) src0 * (int64_t) src1) >> 32)")
>     >     > +binop("imul_high", tint, commutative, """
>     >
>     >     This will enable imul_high for all integer types (ditto for
>     umul_high
>     >     below).  A later patch adds lowering for 64-bit integer type. 
>     Will the
>     >     backend do the right thing for [iu]mul_high of 16- or 8-bit types?
>     >
>     >
>     > That's a good question.  Looks like lower_integer_multiplication
>     in the
>     > back-end will do nothing whatsoever, and we'll emit an illegal opcode
>     > which will probably hang the GPU.  For 8 and 16, it's easy enough to
>     > lower to a couple of conversions, a N*2-bit multiply, and a
>     shift.  It's
>     > also not obvious where the cut-off point for the optimization is. 
>     > Certainly, it's better in 64-bits than doing the division algorithm in
>     > the shader and I think it's better for 32 but maybe not in 8 and 16? 
>     > I'm not sure.  I'm pretty sure my 32-bit benchmark gave positive
>     results
>     > (about 40-50% faster) but it was very noisy.
>     >
>     > I don't think anything allows 8 and 16-bit arithmetic right now. 
>     Still,
>     > should probably fix it...
> 
>     Hm... if an extension adds GL or Vulkan support for 16-bit arithmetic, I
>     doubt it would add [iu]mul_high (e.g., GL_AMD_gpu_shader_int16).  I'd
>     expect every GPU would support a 16x16=32 multiplier.  Would it be
>     better to restrict this instruction to 32- and 64-bit and implement the
>     integer division optimization differently for 8- and 16-bit sources?
> 
> 
> We don't currently have a mechanism in NIR to restrict an instruction to
> a particular set of bit-widths.  Maybe we should consider adding that?

Oh.  I thought it could already do that.  There must be some subtlety of
tint vs. tint32 in this context that I don't understand.  Can you
restrict it to one size only vs. any size?

> As far as the optimization goes, I see a few different options;
> 
>  1) Add a few lines in opt_algebraic to do the lowering.  We could even
> make it unconditional for now.
>  2) Add a couple tiny helpers that switch on bit-width in opt_idiv_const
> and do the lowering on-the-fly.
>  3) Just disable the pass for 8 and 16-bit for now.  It's questionable
> whether it's actually worth emitting 4-6 instructions for a simple
> division at those bit sizes anyway.
> 
> I think my order of preference is 3 > 1 > 2.  Especially when you
> consider that we have no good way to test the pass for anything other
> than 32 and 64-bit right now.

Yeah... I didn't intend to open a can of worms. :)  Option 3 sounds fine
to me.

> --Jason