fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-17 20:00:20 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

agx: Coalesce i2i16 and u2u16

Browse source 
Extract out the code for unpack_64_2x32_split_x and use it for other integer
downcasts too to coalesce out a move. Pointless, but I wanted to have a little
RA fun after getting stencil export working.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20365>
This commit is contained in:
Alyssa Rosenzweig 2022-12-02 11:36:42 -05:00 • committed by Marge Bot
parent 58d02e4f59
commit ef23bbfdbd
1 changed files with 29 additions and 12 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

41
src/asahi/compiler/agx_compile.c
View file

@ -202,6 +202,23 @@ agx_vec2(agx_builder *b, agx_index s0, agx_index s1)
return dst;
}
/*
* Extract the lower or upper N-bits from a (2*N)-bit quantity. We use a split
* without null destinations to let us CSE (and coalesce) the splits when both x
* and y are split.
*/
static agx_instr *
agx_subdivide_to(agx_builder *b, agx_index dst, agx_index s0, unsigned comp)
{
assert((s0.size == (dst.size + 1)) && "only 2x subdivide handled");
assert((comp == 0 || comp == 1) && "too many components");
agx_instr *split = agx_split(b, 2, s0);
split->dest[comp] = dst;
split->dest[1 - comp] = agx_temp(b->shader, dst.size);
return split;
}
static void
agx_block_add_successor(agx_block *block, agx_block *successor)
{
@ -869,7 +886,6 @@ agx_emit_alu(agx_builder *b, nir_alu_instr *instr)
UNOP(fddy_fine, dfdy);
UNOP(mov, mov);
UNOP(u2u16, mov);
UNOP(u2u32, mov);
UNOP(bitfield_reverse, bitrev);
UNOP(bit_count, popcount);
@ -943,10 +959,18 @@ agx_emit_alu(agx_builder *b, nir_alu_instr *instr)
return agx_asr_to(b, dst, ishl16, agx_immediate(8));
} else {
assert (s0.size == AGX_SIZE_32 && "other conversions lowered");
return agx_iadd_to(b, dst, s0, agx_zero(), 0);
return agx_subdivide_to(b, dst, s0, 0);
}
}
case nir_op_u2u16:
{
if (s0.size == AGX_SIZE_32)
return agx_subdivide_to(b, dst, s0, 0);
else
return agx_mov_to(b, dst, s0);
}
case nir_op_iadd_sat:
{
agx_instr *I = agx_iadd_to(b, dst, s0, s1, 0);
@ -1039,18 +1063,11 @@ agx_emit_alu(agx_builder *b, nir_alu_instr *instr)
return agx_emit_collect_to(b, dst, 2, idx);
}
/* Split a 64-bit word into 32-bit parts. Do not use null destinations to
* let us CSE (and coalesce) the splits when both x and y are split.
*/
case nir_op_unpack_64_2x32_split_x:
return agx_subdivide_to(b, dst, s0, 0);
case nir_op_unpack_64_2x32_split_y:
{
agx_instr *split = agx_split(b, 2, s0);
unsigned comp = instr->op == nir_op_unpack_64_2x32_split_y ? 1 : 0;
split->dest[comp] = dst;
split->dest[1 - comp] = agx_temp(b->shader, dst.size);
return split;
}
return agx_subdivide_to(b, dst, s0, 1);
case nir_op_vec2:
case nir_op_vec3: