nir/lower_tex: Add a way to lower TXS(non-0-LOD) instructions

The V3D driver has an open-coded solution for this, and we need the
same thing for Panfrost, so let's add a generic way to lower TXS(LOD)
into max(TXS(0) >> LOD, 1).

Changes in v2:
* Use == 0 instead of !
* Rework the minification logic as suggested by Jason
* Assign cursor pos at the beginning of the function
* Patch the LOD just after retrieving the old value

Signed-off-by: Boris Brezillon <[email protected]>
Reviewed-by: Alyssa Rosenzweig <[email protected]>

2 files changed, 52 insertions, 0 deletions

diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 4270df56511..8972b4af748 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -3427,6 +3427,12 @@ typedef struct nir_lower_tex_options {
    bool lower_txd_clamp_if_sampler_index_not_lt_16;
 
    /**
+    * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
+    * 0-lod followed by a nir_ishr.
+    */
+   bool lower_txs_lod;
+
+   /**
     * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
     * mixed-up tg4 locations.
     */
diff --git a/src/compiler/nir/nir_lower_tex.c b/src/compiler/nir/nir_lower_tex.c
index ace8600d4bb..8a36edfbc5e 100644
--- a/src/compiler/nir/nir_lower_tex.c
+++ b/src/compiler/nir/nir_lower_tex.c
@@ -983,6 +983,47 @@ lower_tg4_offsets(nir_builder *b, nir_tex_instr *tex)
 }
 
 static bool
+nir_lower_txs_lod(nir_builder *b, nir_tex_instr *tex)
+{
+   int lod_idx = nir_tex_instr_src_index(tex, nir_tex_src_lod);
+   if (lod_idx < 0 ||
+       (nir_src_is_const(tex->src[lod_idx].src) &&
+        nir_src_as_int(tex->src[lod_idx].src) == 0))
+      return false;
+
+   unsigned dest_size = nir_tex_instr_dest_size(tex);
+
+   b->cursor = nir_before_instr(&tex->instr);
+   nir_ssa_def *lod = nir_ssa_for_src(b, tex->src[lod_idx].src, 1);
+
+   /* Replace the non-0-LOD in the initial TXS operation by a 0-LOD. */
+   nir_instr_rewrite_src(&tex->instr, &tex->src[lod_idx].src,
+                         nir_src_for_ssa(nir_imm_int(b, 0)));
+
+   /* TXS(LOD) = max(TXS(0) >> LOD, 1) */
+   b->cursor = nir_after_instr(&tex->instr);
+   nir_ssa_def *minified = nir_imax(b, nir_ushr(b, &tex->dest.ssa, lod),
+                                    nir_imm_int(b, 1));
+
+   /* Make sure the component encoding the array size (if any) is not
+    * minified.
+    */
+   if (tex->is_array) {
+      nir_ssa_def *comp[3];
+
+      for (unsigned i = 0; i < dest_size - 1; i++)
+         comp[i] = nir_channel(b, minified, i);
+
+      comp[dest_size - 1] = nir_channel(b, &tex->dest.ssa, dest_size - 1);
+      minified = nir_vec(b, comp, dest_size);
+   }
+
+   nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, nir_src_for_ssa(minified),
+                                  minified->parent_instr);
+   return true;
+}
+
+static bool
 nir_lower_tex_block(nir_block *block, nir_builder *b,
                     const nir_lower_tex_options *options)
 {
@@ -1135,6 +1176,11 @@ nir_lower_tex_block(nir_block *block, nir_builder *b,
          continue;
       }
 
+      if (options->lower_txs_lod && tex->op == nir_texop_txs) {
+         progress |= nir_lower_txs_lod(b, tex);
+         continue;
+      }
+
       /* has to happen after all the other lowerings as the original tg4 gets
        * replaced by 4 tg4 instructions.
        */