diff options
author | Jonathan Marek <[email protected]> | 2019-06-20 23:05:13 -0400 |
---|---|---|
committer | Jonathan Marek <[email protected]> | 2019-06-26 14:56:28 -0400 |
commit | 0b5a483baaefa9a7e39c76607d8f0f435aa46315 (patch) | |
tree | 3ac4fb2b451e76fd82fd74bad160ab18ff3b8ca0 /src | |
parent | 10688257bd7cc1b4241f7b069a4ee9bc3338be47 (diff) |
nir: opt_vectorize: combine different constant sources
We can vectorize instructions with different constant sources by creating
a new load_const and using that.
Signed-off-by: Jonathan Marek <[email protected]>
Reviewed-by: Alyssa Rosenzweig <[email protected]>
Diffstat (limited to 'src')
-rw-r--r-- | src/compiler/nir/nir_opt_vectorize.c | 27 |
1 files changed, 25 insertions, 2 deletions
diff --git a/src/compiler/nir/nir_opt_vectorize.c b/src/compiler/nir/nir_opt_vectorize.c index 275202b0d85..dab7c2e1dad 100644 --- a/src/compiler/nir/nir_opt_vectorize.c +++ b/src/compiler/nir/nir_opt_vectorize.c @@ -33,8 +33,9 @@ static uint32_t hash_src(uint32_t hash, const nir_src *src) { assert(src->is_ssa); + void *hash_data = nir_src_is_const(*src) ? NULL : src->ssa; - return HASH(hash, src->ssa); + return HASH(hash, hash_data); } static uint32_t @@ -79,7 +80,8 @@ srcs_equal(const nir_src *src1, const nir_src *src2) assert(src1->is_ssa); assert(src2->is_ssa); - return src1->ssa == src2->ssa; + return src1->ssa == src2->ssa || + nir_src_is_const(*src1) == nir_src_is_const(*src2); } static bool @@ -185,6 +187,27 @@ instr_try_combine(nir_instr *instr1, nir_instr *instr2) new_alu->dest.write_mask = (1 << total_components) - 1; for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) { + /* handle constant merging case */ + if (alu1->src[i].src.ssa != alu2->src[i].src.ssa) { + nir_const_value *c1 = nir_src_as_const_value(alu1->src[i].src); + nir_const_value *c2 = nir_src_as_const_value(alu2->src[i].src); + assert(c1 && c2); + nir_const_value value[4]; + unsigned bit_size = alu1->src[i].src.ssa->bit_size; + + for (unsigned j = 0; j < total_components; j++) { + value[j].u64 = j < alu1_components ? + c1[alu1->src[i].swizzle[j]].u64 : + c2[alu2->src[i].swizzle[j - alu1_components]].u64; + } + nir_ssa_def *def = nir_build_imm(&b, total_components, bit_size, value); + + new_alu->src[i].src = nir_src_for_ssa(def); + for (unsigned j = 0; j < total_components; j++) + new_alu->src[i].swizzle[j] = j; + continue; + } + new_alu->src[i].src = alu1->src[i].src; for (unsigned j = 0; j < alu1_components; j++) |