Mesa (master): spirv,nir: Add a better vector_insert

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: c478f8ad6cf57425ffdae56a7c18b62b27985ea4
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=c478f8ad6cf57425ffdae56a7c18b62b27985ea4

Author: Jason Ekstrand <jason at jlekstrand.net>
Date:   Wed Apr  8 11:14:40 2020 -0500

spirv,nir: Add a better vector_insert

The old one in spirv_to_nir was besel'ing the whole vector for every
component.  If we think about this as a vector operation, we can do it
way more efficiently.

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira at intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4495>

---

 src/compiler/spirv/spirv_to_nir.c | 23 +++++++++++------------
 1 file changed, 11 insertions(+), 12 deletions(-)

diff --git a/src/compiler/spirv/spirv_to_nir.c b/src/compiler/spirv/spirv_to_nir.c
index e0e8fdba27c..249b09b56ff 100644
--- a/src/compiler/spirv/spirv_to_nir.c
+++ b/src/compiler/spirv/spirv_to_nir.c
@@ -3341,12 +3341,6 @@ vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
    return &vec->dest.dest.ssa;
 }
 
-static nir_ssa_def *
-nir_ieq_imm(nir_builder *b, nir_ssa_def *x, uint64_t i)
-{
-   return nir_ieq(b, x, nir_imm_intN_t(b, i, x->bit_size));
-}
-
 nir_ssa_def *
 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
                            nir_ssa_def *index)
@@ -3358,12 +3352,17 @@ nir_ssa_def *
 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
                           nir_ssa_def *insert, nir_ssa_def *index)
 {
-   nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
-   for (unsigned i = 1; i < src->num_components; i++)
-      dest = nir_bcsel(&b->nb, nir_ieq_imm(&b->nb, index, i),
-                       vtn_vector_insert(b, src, insert, i), dest);
-
-   return dest;
+   nir_const_value per_comp_idx_const[NIR_MAX_VEC_COMPONENTS];
+   for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++)
+      per_comp_idx_const[i] = nir_const_value_for_int(i, index->bit_size);
+   nir_ssa_def *per_comp_idx =
+      nir_build_imm(&b->nb, src->num_components,
+                    index->bit_size, per_comp_idx_const);
+
+   /* nir_builder will automatically splat out scalars to vectors so an insert
+    * is as simple as "if I'm the channel, replace me with the scalar."
+    */
+   return nir_bcsel(&b->nb, nir_ieq(&b->nb, index, per_comp_idx), insert, src);
 }
 
 static nir_ssa_def *