*/
static void
-nir_alu_ssa_dest_init(nir_alu_instr *instr, unsigned num_components)
+nir_alu_ssa_dest_init(nir_alu_instr *instr, unsigned num_components,
+ unsigned bit_size)
{
- nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components, NULL);
+ nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components,
+ bit_size, NULL);
instr->dest.write_mask = (1 << num_components) - 1;
}
nir_ssa_def *last = NULL;
for (unsigned i = 0; i < num_components; i++) {
nir_alu_instr *chan = nir_alu_instr_create(builder->shader, chan_op);
- nir_alu_ssa_dest_init(chan, 1);
+ nir_alu_ssa_dest_init(chan, 1, instr->dest.dest.ssa.bit_size);
nir_alu_src_copy(&chan->src[0], &instr->src[0], chan);
chan->src[0].swizzle[0] = chan->src[0].swizzle[i];
if (nir_op_infos[chan_op].num_inputs > 1) {
nir_alu_src_copy(&chan->src[1], &instr->src[1], chan);
chan->src[1].swizzle[0] = chan->src[1].swizzle[i];
}
+ chan->exact = instr->exact;
nir_builder_instr_insert(builder, &chan->instr);
nir_instr_remove(&instr->instr);
}
-static void
+static bool
lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b)
{
unsigned num_src = nir_op_infos[instr->op].num_inputs;
assert(instr->dest.write_mask != 0);
b->cursor = nir_before_instr(&instr->instr);
+ b->exact = instr->exact;
#define LOWER_REDUCTION(name, chan, merge) \
case name##2: \
case name##3: \
case name##4: \
lower_reduction(instr, chan, merge, b); \
- return;
+ return true;
switch (instr->op) {
case nir_op_vec4:
/* We don't need to scalarize these ops, they're the ones generated to
* group up outputs into a value that can be SSAed.
*/
- return;
+ return false;
case nir_op_pack_half_2x16:
if (!b->shader->options->lower_pack_half_2x16)
- return;
+ return false;
nir_ssa_def *val =
nir_pack_half_2x16_split(b, nir_channel(b, instr->src[0].src.ssa,
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
nir_instr_remove(&instr->instr);
- return;
+ return true;
case nir_op_unpack_unorm_4x8:
case nir_op_unpack_snorm_4x8:
/* There is no scalar version of these ops, unless we were to break it
* down to bitshifts and math (which is definitely not intended).
*/
- return;
+ return false;
case nir_op_unpack_half_2x16: {
if (!b->shader->options->lower_unpack_half_2x16)
- return;
+ return false;
nir_ssa_def *comps[2];
comps[0] = nir_unpack_half_2x16_split_x(b, instr->src[0].src.ssa);
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec));
nir_instr_remove(&instr->instr);
- return;
+ return true;
+ }
+
+ case nir_op_pack_uvec2_to_uint: {
+ assert(b->shader->options->lower_pack_snorm_2x16 ||
+ b->shader->options->lower_pack_unorm_2x16);
+
+ nir_ssa_def *word =
+ nir_extract_u16(b, instr->src[0].src.ssa, nir_imm_int(b, 0));
+ nir_ssa_def *val =
+ nir_ior(b, nir_ishl(b, nir_channel(b, word, 1), nir_imm_int(b, 16)),
+ nir_channel(b, word, 0));
+
+ nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
+ nir_instr_remove(&instr->instr);
+ break;
+ }
+
+ case nir_op_pack_uvec4_to_uint: {
+ assert(b->shader->options->lower_pack_snorm_4x8 ||
+ b->shader->options->lower_pack_unorm_4x8);
+
+ nir_ssa_def *byte =
+ nir_extract_u8(b, instr->src[0].src.ssa, nir_imm_int(b, 0));
+ nir_ssa_def *val =
+ nir_ior(b, nir_ior(b, nir_ishl(b, nir_channel(b, byte, 3), nir_imm_int(b, 24)),
+ nir_ishl(b, nir_channel(b, byte, 2), nir_imm_int(b, 16))),
+ nir_ior(b, nir_ishl(b, nir_channel(b, byte, 1), nir_imm_int(b, 8)),
+ nir_channel(b, byte, 0)));
+
+ nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
+ nir_instr_remove(&instr->instr);
+ break;
}
case nir_op_fdph: {
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
nir_instr_remove(&instr->instr);
- return;
+ return true;
}
+ case nir_op_unpack_64_2x32:
+ return false;
+
LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd);
LOWER_REDUCTION(nir_op_ball_fequal, nir_op_feq, nir_op_iand);
LOWER_REDUCTION(nir_op_ball_iequal, nir_op_ieq, nir_op_iand);
}
if (instr->dest.dest.ssa.num_components == 1)
- return;
+ return false;
unsigned num_components = instr->dest.dest.ssa.num_components;
nir_ssa_def *comps[] = { NULL, NULL, NULL, NULL };
lower->src[i].swizzle[j] = instr->src[i].swizzle[src_chan];
}
- nir_alu_ssa_dest_init(lower, 1);
+ nir_alu_ssa_dest_init(lower, 1, instr->dest.dest.ssa.bit_size);
lower->dest.saturate = instr->dest.saturate;
comps[chan] = &lower->dest.dest.ssa;
+ lower->exact = instr->exact;
nir_builder_instr_insert(b, &lower->instr);
}
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec));
nir_instr_remove(&instr->instr);
-}
-
-static bool
-lower_alu_to_scalar_block(nir_block *block, void *builder)
-{
- nir_foreach_instr_safe(block, instr) {
- if (instr->type == nir_instr_type_alu)
- lower_alu_instr_scalar(nir_instr_as_alu(instr), builder);
- }
-
return true;
}
-static void
+static bool
nir_lower_alu_to_scalar_impl(nir_function_impl *impl)
{
nir_builder builder;
nir_builder_init(&builder, impl);
+ bool progress = false;
+
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type == nir_instr_type_alu) {
+ progress = lower_alu_instr_scalar(nir_instr_as_alu(instr),
+ &builder) || progress;
+ }
+ }
+ }
+
+ nir_metadata_preserve(impl, nir_metadata_block_index |
+ nir_metadata_dominance);
- nir_foreach_block(impl, lower_alu_to_scalar_block, &builder);
+ return progress;
}
-void
+bool
nir_lower_alu_to_scalar(nir_shader *shader)
{
- nir_foreach_function(shader, function) {
+ bool progress = false;
+
+ nir_foreach_function(function, shader) {
if (function->impl)
- nir_lower_alu_to_scalar_impl(function->impl);
+ progress = nir_lower_alu_to_scalar_impl(function->impl) || progress;
}
+
+ return progress;
}
projects / mesa.git / blobdiff