asahi: implement rba2 semantics for vbo

Different APIs have different robustness requirements for VBOs. Add a knob to
select the desired robustness so we can implement rba2 in honeykrisp.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29742>

src/asahi/lib/agx_linker.h

@@ -61,6 +61,9 @@ struct agx_vs_prolog_key {
 
    /* If !hw and the draw call is indexed, the index size */
    uint8_t sw_index_size_B;
+
+   /* Robustness settings for the vertex fetch */
+   struct agx_robustness robustness;
 };
 
 struct agx_fs_prolog_key {

src/asahi/lib/agx_nir_lower_vbo.c

@@ -11,6 +11,11 @@
 #include "util/u_math.h"
 #include "shader_enums.h"
 
+struct ctx {
+   struct agx_attribute *attribs;
+   struct agx_robustness rs;
+};
+
 static bool
 is_rgb10_a2(const struct util_format_description *desc)
 {
@@ -109,7 +114,8 @@ pass(struct nir_builder *b, nir_intrinsic_instr *intr, void *data)
    if (intr->intrinsic != nir_intrinsic_load_input)
       return false;
 
-   struct agx_attribute *attribs = data;
+   struct ctx *ctx = data;
+   struct agx_attribute *attribs = ctx->attribs;
    b->cursor = nir_instr_remove(&intr->instr);
 
    nir_src *offset_src = nir_get_io_offset_src(intr);
@@ -190,14 +196,12 @@ pass(struct nir_builder *b, nir_intrinsic_instr *intr, void *data)
     * before the load. That is faster than the 4 cmpsel required after the load,
     * and it avoids waiting on the load which should help prolog performance.
     *
-    * TODO: Plumb through soft fault information to skip this.
+    * TODO: Optimize.
     *
-    * TODO: Add a knob for robustBufferAccess2 semantics.
+    * TODO: We always clamp to handle null descriptors. Maybe optimize?
     */
-   bool robust = true;
-   if (robust) {
-      el = nir_umin(b, el, bounds);
-   }
+   nir_def *oob = nir_ult(b, bounds, el);
+   el = nir_bcsel(b, oob, nir_imm_int(b, 0), el);
 
    nir_def *base = nir_load_vbo_base_agx(b, buf_handle);
 
@@ -228,6 +232,12 @@ pass(struct nir_builder *b, nir_intrinsic_instr *intr, void *data)
       b, interchange_comps, interchange_register_size, base, stride_offset_el,
       .format = interchange_format, .base = shift);
 
+   /* TODO: Optimize per above */
+   if (ctx->rs.level >= AGX_ROBUSTNESS_D3D) {
+      nir_def *zero = nir_imm_zero(b, memory->num_components, memory->bit_size);
+      memory = nir_bcsel(b, oob, zero, memory);
+   }
+
    unsigned dest_size = intr->def.bit_size;
 
    /* Unpack but do not convert non-native non-array formats */
@@ -290,9 +300,12 @@ pass(struct nir_builder *b, nir_intrinsic_instr *intr, void *data)
 }
 
 bool
-agx_nir_lower_vbo(nir_shader *shader, struct agx_attribute *attribs)
+agx_nir_lower_vbo(nir_shader *shader, struct agx_attribute *attribs,
+                  struct agx_robustness robustness)
 {
    assert(shader->info.stage == MESA_SHADER_VERTEX);
+
+   struct ctx ctx = {.attribs = attribs, .rs = robustness};
    return nir_shader_intrinsics_pass(
-      shader, pass, nir_metadata_block_index | nir_metadata_dominance, attribs);
+      shader, pass, nir_metadata_block_index | nir_metadata_dominance, &ctx);
 }

src/asahi/lib/agx_nir_lower_vbo.h

@@ -33,7 +33,30 @@ struct agx_attribute {
    bool instanced : 1;
 };
 
-bool agx_nir_lower_vbo(nir_shader *shader, struct agx_attribute *attribs);
+enum agx_robustness_level {
+   /* No robustness */
+   AGX_ROBUSTNESS_DISABLED,
+
+   /* Invalid load/store must not fault, but undefined value/effect */
+   AGX_ROBUSTNESS_GLES,
+
+   /* Invalid load/store access something from the array (or 0) */
+   AGX_ROBUSTNESS_GL,
+
+   /* Invalid loads return 0 and invalid stores are dropped */
+   AGX_ROBUSTNESS_D3D,
+};
+
+struct agx_robustness {
+   enum agx_robustness_level level;
+
+   /* Whether hardware "soft fault" is enabled. */
+   bool soft_fault;
+};
+
+bool agx_nir_lower_vbo(nir_shader *shader, struct agx_attribute *attribs,
+                       struct agx_robustness rs);
+
 bool agx_vbo_supports_format(enum pipe_format format);
 
 #ifdef __cplusplus

src/asahi/lib/agx_nir_prolog_epilog.c

@@ -56,7 +56,8 @@ agx_nir_lower_poly_stipple(nir_shader *s)
 }
 
 static bool
-lower_vbo(nir_shader *s, const struct agx_velem_key *key)
+lower_vbo(nir_shader *s, const struct agx_velem_key *key,
+          const struct agx_robustness rs)
 {
    struct agx_attribute out[AGX_MAX_VBUFS];
 
@@ -69,7 +70,7 @@ lower_vbo(nir_shader *s, const struct agx_velem_key *key)
       };
    }
 
-   return agx_nir_lower_vbo(s, out);
+   return agx_nir_lower_vbo(s, out, rs);
 }
 
 static int
@@ -166,7 +167,7 @@ agx_nir_vs_prolog(nir_builder *b, const void *key_)
    nir_export_agx(b, nir_load_instance_id(b), .base = 6 * 2);
 
    /* Now lower the resulting program using the key */
-   lower_vbo(b->shader, key->attribs);
+   lower_vbo(b->shader, key->attribs, key->robustness);
 
    if (!key->hw) {
       agx_nir_lower_index_buffer(b->shader, key->sw_index_size_B, false);

src/gallium/drivers/asahi/agx_state.c

@@ -2213,6 +2213,14 @@ agx_update_vs(struct agx_context *ctx, unsigned index_size_B)
    struct agx_fast_link_key link_key = {
       .prolog.vs.hw = key.hw,
       .prolog.vs.sw_index_size_B = key.hw ? 0 : index_size_B,
+
+      /* TODO: We could optimize this */
+      .prolog.vs.robustness =
+         {
+            .level = AGX_ROBUSTNESS_GL,
+            .soft_fault = false,
+         },
+
       .main = ctx->vs,
    };