On 04/06/2016 04:42 PM, Matteo Bruni wrote:
+static void regstore_set_double(struct d3dx_regstore *rs, unsigned int table, unsigned int offset, double v) +{
- BYTE *p;
- unsigned int reg_idx;
- p = (BYTE *)rs->tables[table] + table_info[table].component_size * offset;
- switch (table_info[table].type)
- {
case PRES_VT_FLOAT : *(float *)p = v; break;case PRES_VT_DOUBLE: *(double *)p = v; break;case PRES_VT_INT : *(int *)p = roundf(v); break;Same here. Also it's not clear that rounding away from 0 is correct, this would need tests (although IIRC preshaders generated by the MS compiler explicitly compute integral values for integer constants, that would make this hard to test and mostly irrelevant in practice). So please try to test for rounding behavior with 0.5 / -0.5 / 1.5, if nothing else to confirm that my memory serves me right. BTW unless testing proves otherwise lrint() is probably the best option here since it generates much nicer code (at least when building with gcc for 32 bits).
In fact the main test now does not have integer constants. Preshader code I've seen generated by MS compiler so far indeed does not use any rounding functions, doing some math with fract, neg and min/max instead. I already have some further patches involving tests of the newly added operations which do not overwrite the blob but modify the opcode in place. I used the same mini framework to test the other edge case which is currently not in the main test: 0 * INF (BTW the results are funny, preshaders exhibit IEEE FP behaviour, like in my implementation, while d3d9 shaders have 0 * INF == 0). I suppose it will be easier to send and less code changes to review if we consider the whole updated test blob with all the planned changes at once afterwards, can we go this way?