Mesa (main): pan/bi: Align accesses with packed TLS

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 1a882ecdab06d3cf3786698d23eb015c2baa332a
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=1a882ecdab06d3cf3786698d23eb015c2baa332a

Author: Alyssa Rosenzweig <alyssa at collabora.com>
Date:   Thu Jun 16 19:14:58 2022 -0400

pan/bi: Align accesses with packed TLS

When lowering vars to scratch, we need to be careful with alignment on Valhall,
where packed TLS access must not straddle a 16-byte boundary. Fixes regressions
when enabling indirect access to temps on Valhall.

Fixes: 6761dbf8915 ("panfrost: Use packed TLS on Valhall")
Signed-off-by: Alyssa Rosenzweig <alyssa at collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>

---

 src/panfrost/bifrost/bifrost_compile.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/src/panfrost/bifrost/bifrost_compile.c b/src/panfrost/bifrost/bifrost_compile.c
index 4eea93f529a..15f5924cf31 100644
--- a/src/panfrost/bifrost/bifrost_compile.c
+++ b/src/panfrost/bifrost/bifrost_compile.c
@@ -4835,9 +4835,19 @@ bi_finalize_nir(nir_shader *nir, unsigned gpu_id, bool is_blend)
         /* Get rid of any global vars before we lower to scratch. */
         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 
+        /* Valhall introduces packed thread local storage, which improves cache
+         * locality of TLS access. However, access to packed TLS cannot
+         * straddle 16-byte boundaries. As such, when packed TLS is in use
+         * (currently unconditional for Valhall), we force vec4 alignment for
+         * scratch access.
+         */
+        bool packed_tls = (gpu_id >= 0x9000);
+
         /* Lower large arrays to scratch and small arrays to bcsel (TODO: tune
          * threshold, but not until addresses / csel is optimized better) */
         NIR_PASS_V(nir, nir_lower_vars_to_scratch, nir_var_function_temp, 16,
+                        packed_tls ?
+                        glsl_get_vec4_size_align_bytes :
                         glsl_get_natural_size_align_bytes);
         NIR_PASS_V(nir, nir_lower_indirect_derefs, nir_var_function_temp, ~0);