fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-22 11:20:11 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

nir/lower_doubles: preserve NaN when asked to do so

Browse source 
v2: avoid generating unnecessary bcsel (Caio)

Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28702>
This commit is contained in:
Iván Briano 2024-02-12 16:40:25 -08:00
parent c6c52113a0
commit 53e130e333
1 changed files with 25 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

31
src/compiler/nir/nir_lower_double_ops.c
View file

@ -101,6 +101,19 @@ get_signed_zero(nir_builder *b, nir_def *src)
return zero; return zero;
} }
static nir_def *
preserve_nan(nir_builder *b, nir_def *src, nir_def *res)
{
uint32_t exec_mode = b->shader->info.float_controls_execution_mode;
if (nir_is_float_control_nan_preserve(exec_mode, 64)) {
nir_def *is_nan = nir_fneu(b, src, src);
return nir_bcsel(b, is_nan, src, res);
}
return res;
}
/* /*
* Generates the correctly-signed infinity if the source was zero, and flushes * Generates the correctly-signed infinity if the source was zero, and flushes
* the result to 0 if the source was infinity or the calculated exponent was * the result to 0 if the source was infinity or the calculated exponent was
@ -111,12 +124,14 @@ static nir_def *
fix_inv_result(nir_builder *b, nir_def *res, nir_def *src, fix_inv_result(nir_builder *b, nir_def *res, nir_def *src,
nir_def *exp) nir_def *exp)
{ {
/* If the exponent is too small or the original input was infinity/NaN, /* If the exponent is too small or the original input was infinity,
* force the result to 0 (flush denorms) to avoid the work of handling * force the result to 0 (flush denorms) to avoid the work of handling
* denorms properly. * denorms properly. If we are asked to preserve NaN, do so, otherwise
* we return the flushed result for it.
*/ */
res = nir_bcsel(b, nir_ior(b, nir_ile_imm(b, exp, 0), nir_feq_imm(b, nir_fabs(b, src), INFINITY)), res = nir_bcsel(b, nir_ior(b, nir_ile_imm(b, exp, 0), nir_feq_imm(b, nir_fabs(b, src), INFINITY)),
get_signed_zero(b, src), res); get_signed_zero(b, src), res);
res = preserve_nan(b, src, res);
/* If the original input was 0, generate the correctly-signed infinity */ /* If the original input was 0, generate the correctly-signed infinity */
res = nir_bcsel(b, nir_fneu_imm(b, src, 0.0f), res = nir_bcsel(b, nir_fneu_imm(b, src, 0.0f),
@ -301,16 +316,15 @@ lower_sqrt_rsq(nir_builder *b, nir_def *src, bool sqrt)
res = nir_ffma(b, y_1, r_1, y_1); res = nir_ffma(b, y_1, r_1, y_1);
} }
uint32_t exec_mode = b->shader->info.float_controls_execution_mode;
if (sqrt) { if (sqrt) {
/* Here, the special cases we need to handle are /* Here, the special cases we need to handle are
* 0 -> 0 and * 0 -> 0 and
* +inf -> +inf * +inf -> +inf
* NaN -> NaN
*/ */
const bool preserve_denorms =
b->shader->info.float_controls_execution_mode &
FLOAT_CONTROLS_DENORM_PRESERVE_FP64;
nir_def *src_flushed = src; nir_def *src_flushed = src;
if (!preserve_denorms) { if (!nir_is_denorm_preserve(exec_mode, 64)) {
src_flushed = nir_bcsel(b, src_flushed = nir_bcsel(b,
nir_flt_imm(b, nir_fabs(b, src), DBL_MIN), nir_flt_imm(b, nir_fabs(b, src), DBL_MIN),
get_signed_zero(b, src), get_signed_zero(b, src),
@ -318,10 +332,15 @@ lower_sqrt_rsq(nir_builder *b, nir_def *src, bool sqrt)
} }
res = nir_bcsel(b, nir_ior(b, nir_feq_imm(b, src_flushed, 0.0), nir_feq_imm(b, src, INFINITY)), res = nir_bcsel(b, nir_ior(b, nir_feq_imm(b, src_flushed, 0.0), nir_feq_imm(b, src, INFINITY)),
src_flushed, res); src_flushed, res);
res = preserve_nan(b, src, res);
} else { } else {
res = fix_inv_result(b, res, src, new_exp); res = fix_inv_result(b, res, src, new_exp);
} }
if (nir_is_float_control_nan_preserve(exec_mode, 64))
res = nir_bcsel(b, nir_feq_imm(b, src, -INFINITY),
nir_imm_double(b, NAN), res);
return res; return res;
} }