fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-08 11:18:08 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

i915/corm: multi-component ALU dest folding in vec construction

Browse source 
Generalize the scalar ALU dest fold to handle multi-component results.
When a vec source covers contiguous channels with identity swizzle and
all uses of the source come from this vec, patch the ALU instruction
to write directly into the vec's dest register with the appropriate
channel mask.

This eliminates redundant MOVs for patterns like
  vec4(%result.x, %result.y, %result.z, %other)
where %result is a vec3 ALU output — the ALU instruction now writes
directly to the output register's .xyz channels.

shader-db (I915_FS=nir): 233/403 compiled, 3328 alu
shader-db (I915_FS=both): nir won 233 (26 identical, 1 tied, 203 better, 3 only),
  54 TGSI, 116 neither

Assisted-by: Claude
This commit is contained in:
Adam Jackson 2026-05-07 09:46:52 -04:00
parent 879cf1bd74
commit 595b9850e0
1 changed files with 28 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

35
src/gallium/drivers/i915/i915_fpc_nir.c
View file

@ -550,27 +550,48 @@ emit_alu(struct nir_to_i915 *c, nir_alu_instr *alu)
neg_sel[i] = (srcs[i] >> UREG_CHANNEL_X_NEGATE_SHIFT) & 0x1;
}
/* Single-component ALU dest folding: if a vec source is a single-use
* scalar ALU result in a temp, patch that instruction to write directly
* into our dest with the right channel mask.
/* ALU dest folding: if a vec source is a single-use ALU result in a
* temp with identity swizzle, patch that instruction to write
* directly into our dest with the right channel mask.
*/
for (unsigned i = 0; i < n; i++) {
nir_def *src_def = alu->src[i].src.ssa;
uint32_t *prev_csr = c->def_csr[src_def->index];
if (!prev_csr || !list_is_singular(&src_def->uses))
if (!prev_csr)
continue;
if (GET_UREG_TYPE(srcs[i]) != REG_TYPE_R)
continue;
if (src_def->num_components != 1)
unsigned nc = src_def->num_components;
if (i + nc > n)
continue;
bool identity = true;
for (unsigned j = 0; j < nc && identity; j++)
identity = (j == 0 || alu->src[i + j].src.ssa == src_def) &&
(alu->src[i + j].swizzle[0] == j);
if (!identity)
continue;
bool all_from_this_vec = true;
nir_foreach_use(use, src_def) {
if (nir_src_use_instr(use) != &alu->instr) {
all_from_this_vec = false;
break;
}
}
if (!all_from_this_vec)
continue;
uint32_t fold_mask = 0;
for (unsigned j = 0; j < nc; j++)
fold_mask |= chan_mask[i + j];
prev_csr[0] = (prev_csr[0] & ~(A0_DEST_CHANNEL_ALL |
(0x1ff << A0_DEST_NR_SHIFT))) |
A0_DEST(dest) | chan_mask[i];
A0_DEST(dest) | fold_mask;
i915_release_temp(p, GET_UREG_NR(srcs[i]));
c->ureg_map[src_def->index] = dest;
emitted[i] = true;
for (unsigned j = 0; j < nc; j++)
emitted[i + j] = true;
}
/* Process real-register sources first, folding in any ZERO/ONE