ir3: Make FS tex prefetch optimization optional

a610 and friends seem not to have tex prefetch.

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20991>

diff --git a/src/freedreno/common/freedreno_dev_info.h b/src/freedreno/common/freedreno_dev_info.h
index f05c130..e5d0ae0 100644 (file)
--- a/src/freedreno/common/freedreno_dev_info.h
+++ b/src/freedreno/common/freedreno_dev_info.h
@@ -72,6 +72,8 @@ struct fd_dev_info {
 
          bool has_hw_multiview;
 
+         bool has_fs_tex_prefetch;
+
          /* Whether the PC_MULTIVIEW_MASK register exists. */
          bool supports_multiview_mask;
 

diff --git a/src/freedreno/common/freedreno_devices.py b/src/freedreno/common/freedreno_devices.py
index cfe9108..1e09cc3 100644 (file)
--- a/src/freedreno/common/freedreno_devices.py
+++ b/src/freedreno/common/freedreno_devices.py
@@ -153,6 +153,7 @@ class A6xxGPUInfo(GPUInfo):
 
         self.a6xx.has_gmem_fast_clear = True
         self.a6xx.has_hw_multiview = True
+        self.a6xx.has_fs_tex_prefetch = True
 
         self.a6xx.sysmem_per_ccu_cache_size = 64 * 1024
         self.a6xx.gmem_ccu_color_cache_fraction = CCUColorCacheFraction.QUARTER.value
@@ -248,6 +249,7 @@ a6xx_gen1 = dict(
 a6xx_gen1_low = {**a6xx_gen1, **dict(
         has_gmem_fast_clear = False,
         has_hw_multiview = False,
+        has_fs_tex_prefetch = False,
         sysmem_per_ccu_cache_size = 8 * 1024,
         gmem_ccu_color_cache_fraction = CCUColorCacheFraction.HALF.value,
         vs_max_inputs_count = 16,

diff --git a/src/freedreno/ir3/ir3_compiler.c b/src/freedreno/ir3/ir3_compiler.c
index 272a308..37d2bec 100644 (file)
--- a/src/freedreno/ir3/ir3_compiler.c
+++ b/src/freedreno/ir3/ir3_compiler.c
@@ -204,6 +204,8 @@ ir3_compiler_create(struct fd_device *dev, const struct fd_dev_id *dev_id,
       compiler->shared_consts_base_offset = 504;
       compiler->shared_consts_size = 8;
       compiler->geom_shared_consts_size_quirk = 16;
+
+      compiler->has_fs_tex_prefetch = dev_info->a6xx.has_fs_tex_prefetch;
    } else {
       compiler->max_const_pipeline = 512;
       compiler->max_const_geom = 512;

diff --git a/src/freedreno/ir3/ir3_compiler.h b/src/freedreno/ir3/ir3_compiler.h
index 77360a8..1730fa4 100644 (file)
--- a/src/freedreno/ir3/ir3_compiler.h
+++ b/src/freedreno/ir3/ir3_compiler.h
@@ -237,6 +237,8 @@ struct ir3_compiler {
     * TODO: Keep an eye on this for next gens.
     */
    uint64_t geom_shared_consts_size_quirk;
+
+   bool has_fs_tex_prefetch;
 };
 
 void ir3_compiler_destroy(struct ir3_compiler *compiler);

diff --git a/src/freedreno/ir3/ir3_context.c b/src/freedreno/ir3/ir3_context.c
index e58bedd..6f8f738 100644 (file)
--- a/src/freedreno/ir3/ir3_context.c
+++ b/src/freedreno/ir3/ir3_context.c
@@ -123,7 +123,7 @@ ir3_context_init(struct ir3_compiler *compiler, struct ir3_shader *shader,
    /* Enable the texture pre-fetch feature only a4xx onwards.  But
     * only enable it on generations that have been tested:
     */
-   if ((so->type == MESA_SHADER_FRAGMENT) && (compiler->gen >= 6))
+   if ((so->type == MESA_SHADER_FRAGMENT) && compiler->has_fs_tex_prefetch)
       NIR_PASS_V(ctx->s, ir3_nir_lower_tex_prefetch);
 
    NIR_PASS(progress, ctx->s, nir_lower_phis_to_scalar, true);