[Mesa-dev] [PATCH 2/2] nir: Lower ldexp to arithmetic.
Ian Romanick
idr at freedesktop.org
Wed Jan 20 11:04:32 PST 2016
On 01/19/2016 10:14 PM, Kenneth Graunke wrote:
> This is a port of Matt's GLSL IR lowering pass to NIR. It's required
> because we translate SPIR-V directly to NIR, bypassing GLSL IR.
>
> I haven't introduced a lower_ldexp flag, as I believe all current NIR
> consumers would set the flag. i965 wants this, vc4 doesn't implement
> this feature, and st_glsl_to_tgsi currently lowers ldexp
> unconditionally anyway.
>
> Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
> ---
> src/glsl/nir/nir_opt_algebraic.py | 63 +++++++++++++++++++++++++++++++++++++++
> 1 file changed, 63 insertions(+)
>
> diff --git a/src/glsl/nir/nir_opt_algebraic.py b/src/glsl/nir/nir_opt_algebraic.py
> index 7745b76..0976ed6 100644
> --- a/src/glsl/nir/nir_opt_algebraic.py
> +++ b/src/glsl/nir/nir_opt_algebraic.py
> @@ -1,4 +1,5 @@
> #! /usr/bin/env python
> +# -*- encoding: utf-8 -*-
> #
> # Copyright (C) 2014 Intel Corporation
> #
> @@ -265,6 +266,68 @@ for op in ['flt', 'fge', 'feq', 'fne',
> ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))),
> ]
>
> +def ldexp_to_arith(x, exp):
> + """
> + Translates
> + ldexp x exp
> + into
> +
> + extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
> + resulting_biased_exp = extracted_biased_exp + exp;
> +
> + if (resulting_biased_exp < 1) {
> + return copysign(0.0, x);
> + }
> +
> + return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
> + lshift(i2u(resulting_biased_exp), exp_shift));
> +
> + which we can't actually implement as such, since NIR doesn't have
> + vectorized if-statements. We actually implement it without branches
> + using conditional-select:
> +
> + extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
> + resulting_biased_exp = extracted_biased_exp + exp;
> +
> + is_not_zero_or_underflow = gequal(resulting_biased_exp, 1);
> + x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
> + resulting_biased_exp = csel(is_not_zero_or_underflow,
> + resulting_biased_exp, 0);
> +
> + return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
> + lshift(i2u(resulting_biased_exp), exp_shift));
Interesting. That's not the way I would have predicted the branches
would get lowered, but that does match the code in
lower_instructions.cpp. I was expecting
extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
resulting_biased_exp = extracted_biased_exp + exp;
is_not_zero_or_underflow = gequal(resulting_biased_exp, 1);
result = bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
lshift(i2u(resulting_biased_exp), exp_shift));
return csel(is_not_zero_or_underflow, result, copysign(0.0f, x));
Probably a question for Matt... why was one picked over the other?
> + """
> +
> + sign_mask = 0x80000000
> + exp_shift = 23
> + exp_width = 8
> +
> + # Extract the biased exponent from <x>.
> + extracted_biased_exp = ('ushr', ('iabs', x), exp_shift)
> + resulting_biased_exp = ('iadd', extracted_biased_exp, exp)
> +
> + # Test if result is ±0.0, subnormal, or underflow by checking if the
> + # resulting biased exponent would be less than 0x1. If so, the result is
> + # 0.0 with the sign of x. (Actually, invert the conditions so that
> + # immediate values are the second arguments, which is better for i965)
> + zero_sign_x = ('iand', x, sign_mask)
> +
> + is_not_zero_or_underflow = ('ige', resulting_biased_exp, 0x1)
> +
> + # We could test for overflows by checking if the resulting biased exponent
> + # would be greater than 0xFE. Turns out we don't need to because the GLSL
> + # spec says:
> + #
> + # "If this product is too large to be represented in the
> + # floating-point type, the result is undefined."
> +
> + return ('bitfield_insert',
> + ('bcsel', is_not_zero_or_underflow, x, zero_sign_x),
> + ('bcsel', is_not_zero_or_underflow, resulting_biased_exp, 0),
> + exp_shift, exp_width)
> +
> +optimizations += [(('ldexp', 'x', 'exp'), ldexp_to_arith('x', 'exp'))]
> +
> # This section contains "late" optimizations that should be run after the
> # regular optimizations have finished. Optimizations should go here if
> # they help code generation but do not necessarily produce code that is
>
More information about the mesa-dev
mailing list