freedreno/ir3: respect tex prefetch limits

Refactor a bit the limit checking in the bindless case, and add tex/samp
limit checking for the non-bindless case, to ensure we do not try to
prefetch textures which cannot be encoded in the # of bits available.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5431>

diff --git a/src/freedreno/ir3/ir3_compiler_nir.c b/src/freedreno/ir3/ir3_compiler_nir.c
index ff699a8925a7233fc77718e47cf75ac4855318be..19633077118d130534e2f94c2ae49737079c966e 100644 (file)
--- a/src/freedreno/ir3/ir3_compiler_nir.c
+++ b/src/freedreno/ir3/ir3_compiler_nir.c
@@ -3446,10 +3446,16 @@ collect_tex_prefetches(struct ir3_context *ctx, struct ir3 *ir)
                                fetch->dst = instr->regs[0]->num;
                                fetch->src = instr->prefetch.input_offset;
 
+                               /* These are the limits on a5xx/a6xx, we might need to
+                                * revisit if SP_FS_PREFETCH[n] changes on later gens:
+                                */
+                               assert(fetch->dst <= 0x3f);
+                               assert(fetch->tex_id <= 0x1f);
+                               assert(fetch->samp_id < 0xf);
+
                                ctx->so->total_in =
                                        MAX2(ctx->so->total_in, instr->prefetch.input_offset + 2);
 
-                               /* Disable half precision until supported. */
                                fetch->half_precision = !!(instr->regs[0]->flags & IR3_REG_HALF);
 
                                /* Remove the prefetch placeholder instruction: */

diff --git a/src/freedreno/ir3/ir3_nir_lower_tex_prefetch.c b/src/freedreno/ir3/ir3_nir_lower_tex_prefetch.c
index 543b40bfb6c00a318d076c9ce64d9f5b3edd89a1..38474749955e1d47cd9223621c8b5e21a17d5f4a 100644 (file)
--- a/src/freedreno/ir3/ir3_nir_lower_tex_prefetch.c
+++ b/src/freedreno/ir3/ir3_nir_lower_tex_prefetch.c
@@ -110,6 +110,46 @@ has_src(nir_tex_instr *tex, nir_tex_src_type type)
        return nir_tex_instr_src_index(tex, type) > 0;
 }
 
+static bool
+ok_bindless_src(nir_tex_instr *tex, nir_tex_src_type type)
+{
+       int idx = nir_tex_instr_src_index(tex, type);
+       assert(idx >= 0);
+       nir_intrinsic_instr *bindless = ir3_bindless_resource(tex->src[idx].src);
+
+       /* TODO from SP_FS_BINDLESS_PREFETCH[n] it looks like this limit should
+        * be 1<<8 ?
+        */
+       return nir_src_is_const(bindless->src[0]) &&
+                       (nir_src_as_uint(bindless->src[0]) < (1 << 16));
+}
+
+/**
+ * Check that we will be able to encode the tex/samp parameters
+ * successfully.  These limits are based on the layout of
+ * SP_FS_PREFETCH[n] and SP_FS_BINDLESS_PREFETCH[n], so at some
+ * point (if those regs changes) they may become generation
+ * specific.
+ */
+static bool
+ok_tex_samp(nir_tex_instr *tex)
+{
+       if (has_src(tex, nir_tex_src_texture_handle)) {
+               /* bindless case: */
+
+               assert(has_src(tex, nir_tex_src_sampler_handle));
+
+               return ok_bindless_src(tex, nir_tex_src_texture_handle) &&
+                               ok_bindless_src(tex, nir_tex_src_sampler_handle);
+       } else {
+               assert(!has_src(tex, nir_tex_src_texture_offset));
+               assert(!has_src(tex, nir_tex_src_sampler_offset));
+
+               return (tex->texture_index <= 0x1f) &&
+                               (tex->sampler_index <= 0xf);
+       }
+}
+
 static bool
 lower_tex_prefetch_block(nir_block *block)
 {
@@ -135,30 +175,14 @@ lower_tex_prefetch_block(nir_block *block)
                                has_src(tex, nir_tex_src_sampler_offset))
                        continue;
 
-               /* Disallow indirect or large bindless handles */
-               int idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_handle);
-               if (idx >= 0) {
-                       nir_intrinsic_instr *bindless =
-                               ir3_bindless_resource(tex->src[idx].src);
-                       if (!nir_src_is_const(bindless->src[0]) ||
-                               nir_src_as_uint(bindless->src[0]) >= (1 << 16))
-                               continue;
-               }
-
-               idx = nir_tex_instr_src_index(tex, nir_tex_src_sampler_handle);
-               if (idx >= 0) {
-                       nir_intrinsic_instr *bindless =
-                               ir3_bindless_resource(tex->src[idx].src);
-                       if (!nir_src_is_const(bindless->src[0]) ||
-                               nir_src_as_uint(bindless->src[0]) >= (1 << 16))
-                               continue;
-               }
-
                /* only prefetch for simple 2d tex fetch case */
                if (tex->sampler_dim != GLSL_SAMPLER_DIM_2D || tex->is_array)
                        continue;
 
-               idx = nir_tex_instr_src_index(tex, nir_tex_src_coord);
+               if (!ok_tex_samp(tex))
+                       continue;
+
+               int idx = nir_tex_instr_src_index(tex, nir_tex_src_coord);
                /* First source should be the sampling coordinate. */
                nir_tex_src *coord = &tex->src[idx];
                debug_assert(coord->src.is_ssa);