AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

pan/bi: Fix and improve the !abs && !coarse case in bi_emit_derivative()

Browse Source 
The lane source passed to CLPER must be wrap-invariant, and we currently
violate this constraint in the !abs && !coarse case. So let's rework
the code to use a XOR modifier on the current lane instead, and adjust
the sign of the (righ - left) operation afterwards.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Lars-Ivar Hesselberg Simonsen <lars-ivar.simonsen@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34676>
This commit is contained in:
Boris Brezillon
2025-04-24 09:48:12 +02:00
committed by Marge Bot
parent 5ed79c2d2b
commit 59c307a3f3
1 changed files with 32 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/panfrost/compiler/bifrost_compile.c
+32 -16
View File
@@ -1881,33 +1881,49 @@ static void
bi_emit_derivative(bi_builder *b, bi_index dst, nir_intrinsic_instr *instr,
unsigned axis, bool coarse)
{
assert(axis == 1 || axis == 2);
bi_index left, right;
bi_index s0 = bi_src_index(&instr->src[0]);
unsigned sz = instr->def.bit_size;
/* If all uses are fabs, the sign of the derivative doesn't matter. This is
* inherently based on fine derivatives so we can't do it for coarse.
*/
if (nir_def_all_uses_ignore_sign_bit(&instr->def) && !coarse) {
if (coarse) {
left = bi_clper(b, s0, bi_imm_u8(0), BI_LANE_OP_NONE);
right = bi_clper(b, s0, bi_imm_u8(axis), BI_LANE_OP_NONE);
} else {
left = s0;
right = bi_clper(b, s0, bi_imm_u8(axis), BI_LANE_OP_XOR);
} else {
bi_index lane1, lane2;
if (coarse) {
lane1 = bi_imm_u32(0);
lane2 = bi_imm_u32(axis);
} else {
lane1 = bi_lshift_and_i32(b, bi_fau(BIR_FAU_LANE_ID, false),
bi_imm_u32(0x3 & ~axis), bi_imm_u8(0));
lane2 = bi_iadd_u32(b, lane1, bi_imm_u32(axis), false);
}
left = bi_clper(b, s0, bi_byte(lane1, 0), BI_LANE_OP_NONE);
right = bi_clper(b, s0, bi_byte(lane2, 0), BI_LANE_OP_NONE);
}
bi_fadd_to(b, sz, dst, right, bi_neg(left));
if (!coarse && !nir_def_all_uses_ignore_sign_bit(&instr->def)) {
/* If the user cares about the sign, we need to take into account the fact
* left/right (or top/bottom) might be inverted. Instead of using a couple
* CSEL, we just invert the sign bit with
*
* sign_bit = XOR(sign_bit, axis_bit(lane_id)).
*/
bi_index res = bi_fadd(b, sz, right, bi_neg(left));
bi_index lane = bi_fau(BIR_FAU_LANE_ID, false);
bi_index lane_shift = bi_imm_u8(sz - ffs(axis));
/* Clear the low bit before the shift if this is the Y-axis we want.
* We skip it on the X-axis, because the lshift-by-31 will get us a
* clean mask. */
if (axis == 2)
lane = bi_lshift_and_i32(b, lane, bi_imm_u32(2), bi_imm_u8(0));
if (sz == 16)
lane = bi_half(lane, false);
/* And here comes the final XOR on the sign bit. */
bi_lshift_xor_to(b, sz, dst, lane, res, lane_shift);
} else {
bi_fadd_to(b, sz, dst, right, bi_neg(left));
}
}
static enum bi_subgroup