On Thu, Oct 1, 2015 at 8:36 AM, Daniel Scharrer <[email protected]> wrote: > The result of POW for a negative base is undefined. Even when the result > is multiplied by zero (which is the case here whenever the base is > negative), the Inf and NaNs can propagate past that. > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91342 > Signed-off-by: Daniel Scharrer <[email protected]> > Cc: "10.6 11.0" <[email protected]>
Reviewed-by: Ilia Mirkin <[email protected]> This also fixes the trace in question on nouveau (tested with a GK208). Nice find! I did some digging in the ARB_vertex_program spec, and it explicitly says that LG2(negative) is undefined, and that POW can be expressed as EX2 + LG2, so the compilers are well within their right to compile the program as they have been. Let me know if you need me to push this out for you. -ilia > --- > src/mesa/main/ffvertex_prog.c | 17 ++++++++++++++--- > 1 file changed, 14 insertions(+), 3 deletions(-) > > diff --git a/src/mesa/main/ffvertex_prog.c b/src/mesa/main/ffvertex_prog.c > index 95b428d..a6183b4 100644 > --- a/src/mesa/main/ffvertex_prog.c > +++ b/src/mesa/main/ffvertex_prog.c > @@ -293,9 +293,10 @@ struct ureg { > GLuint file:4; > GLint idx:9; /* relative addressing may be negative */ > /* sizeof(idx) should == sizeof(prog_src_reg::Index) */ > + GLuint abs:1; > GLuint negate:1; > GLuint swz:12; > - GLuint pad:6; > + GLuint pad:5; > }; > > > @@ -324,6 +325,7 @@ static const struct ureg undef = { > 0, > 0, > 0, > + 0, > 0 > }; > > @@ -342,6 +344,7 @@ static struct ureg make_ureg(GLuint file, GLint idx) > struct ureg reg; > reg.file = file; > reg.idx = idx; > + reg.abs = 0; > reg.negate = 0; > reg.swz = SWIZZLE_NOOP; > reg.pad = 0; > @@ -350,6 +353,14 @@ static struct ureg make_ureg(GLuint file, GLint idx) > > > > +static struct ureg absolute( struct ureg reg ) > +{ > + reg.abs = 1; > + reg.negate = 0; > + return reg; > +} > + > + > static struct ureg negate( struct ureg reg ) > { > reg.negate ^= 1; > @@ -526,8 +537,8 @@ static void emit_arg( struct prog_src_register *src, > src->File = reg.file; > src->Index = reg.idx; > src->Swizzle = reg.swz; > + src->Abs = reg.abs; > src->Negate = reg.negate ? NEGATE_XYZW : NEGATE_NONE; > - src->Abs = 0; > src->RelAddr = 0; > /* Check that bitfield sizes aren't exceeded */ > assert(src->Index == reg.idx); > @@ -953,7 +964,7 @@ static struct ureg calculate_light_attenuation( struct > tnl_program *p, > > emit_op2(p, OPCODE_DP3, spot, 0, negate(VPpli), spot_dir_norm); > emit_op2(p, OPCODE_SLT, slt, 0, swizzle1(spot_dir_norm,W), spot); > - emit_op2(p, OPCODE_POW, spot, 0, spot, swizzle1(attenuation, W)); > + emit_op2(p, OPCODE_POW, spot, 0, absolute(spot), swizzle1(attenuation, > W)); > emit_op2(p, OPCODE_MUL, att, 0, slt, spot); > > release_temp(p, spot); > -- > 2.6.0 > > _______________________________________________ > mesa-dev mailing list > [email protected] > http://lists.freedesktop.org/mailman/listinfo/mesa-dev _______________________________________________ mesa-dev mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/mesa-dev
