fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-05 05:18:08 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

agx: fix fmin/fmax with (-0, 0) pair

Browse source 
We need additional lowering to handle negzero properly. fixes float_controls2
fails but strictly the bug was already present!

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29742>
This commit is contained in:
Alyssa Rosenzweig 2024-06-11 20:01:13 -04:00
parent ab21d179d6
commit 88cdcd8f72
1 changed files with 35 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

45
src/asahi/compiler/agx_compile.c
View file

@ -1666,6 +1666,37 @@ is_conversion_to_8bit(nir_op op)
}
}
static agx_instr *
agx_fminmax_to(agx_builder *b, agx_index dst, agx_index s0, agx_index s1,
nir_alu_instr *alu)
{
bool fmax = alu->op == nir_op_fmax;
enum agx_fcond fcond = fmax ? AGX_FCOND_GTN : AGX_FCOND_LTN;
enum agx_icond icond = fmax ? AGX_ICOND_SGT : AGX_ICOND_SLT;
/* Calculate min/max with the appropriate hardware instruction */
agx_index tmp = agx_fcmpsel(b, s0, s1, s0, s1, fcond);
/* The hardware ltn/gtn modes implement IEEE 754 NaN rules. Unfortunately,
* they do not respect signed zeros, causing conformance fails. To
* workaround, when both operands are zero, we instead use integer min/max to
* get the appropriate sign on the result. We detect that case by checking
* for float equality of the inputs: signed zero is the only case where the
* floats are equal but the integer values are not, so this works for all
* cases. It's annoying that we need 3 SCIB instructions instead of 1, but
* hopefully apps don't ask for signed zero preserve if they don't need it.
*/
if (nir_is_float_control_signed_zero_preserve(alu->fp_fast_math,
alu->def.bit_size)) {
agx_index iminmax = agx_icmpsel(b, s0, s1, s0, s1, icond);
tmp = agx_fcmpsel(b, s0, s1, iminmax, tmp, AGX_FCOND_EQ);
}
/* Flush denorms, as cmpsel will not. */
return agx_fadd_to(b, dst, tmp, agx_negzero());
}
static agx_instr *
agx_emit_alu(agx_builder *b, nir_alu_instr *instr)
{
@ -1773,16 +1804,10 @@ agx_emit_alu(agx_builder *b, nir_alu_instr *instr)
case nir_op_fneg:
return agx_fmov_to(b, dst, agx_neg(s0));
case nir_op_fmin: {
agx_index tmp = agx_fcmpsel(b, s0, s1, s0, s1, AGX_FCOND_LTN);
/* flush denorms */
return agx_fadd_to(b, dst, tmp, agx_negzero());
}
case nir_op_fmax: {
agx_index tmp = agx_fcmpsel(b, s0, s1, s0, s1, AGX_FCOND_GTN);
/* flush denorms */
return agx_fadd_to(b, dst, tmp, agx_negzero());
}
case nir_op_fmin:
case nir_op_fmax:
return agx_fminmax_to(b, dst, s0, s1, instr);
case nir_op_imin:
return agx_icmpsel_to(b, dst, s0, s1, s0, s1, AGX_ICOND_SLT);
case nir_op_imax: