pan/mdg: Lower PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK on Midgard

The register file on Midgard is not large enough to sustain 256 threads in a
threadgroup when all ISA-defined registers are used. As such, we want
to advertise the smallest MAX_THREADS_PER_BLOCK permissible by the spec to
avoid compiling shaders that will necessarily spill. The minimum-maximum in
OpenGL ES 3.1 is 128, so set that on Midgard.

6 compute shaders LOST in shader-db due to exceeding this new limit. These
shaders would fault if they were attempted to be executed.

Cc: mesa-stable
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19092>
(cherry picked from commit 9b19104a30)

.pick_status.json

@@ -40,7 +40,7 @@
         "description": "pan/mdg: Lower PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK on Midgard",
         "nominated": true,
         "nomination_type": 0,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": null
     },

src/gallium/drivers/panfrost/pan_screen.c

@@ -697,7 +697,15 @@ panfrost_get_compute_param(struct pipe_screen *pscreen, enum pipe_shader_ir ir_t
 		RET(((uint64_t []) { 256, 256, 256 }));
 
 	case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
-		RET((uint64_t []) { 256 });
+                /* On Bifrost and newer, all GPUs can support at least 256 threads
+                 * regardless of register usage, so we report 256.
+                 *
+                 * On Midgard, with maximum register usage, the maximum
+                 * thread count is only 64. We would like to report 64 here, but
+                 * the GLES3.1 spec minimum is 128, so we report 128 and limit
+                 * the register allocation of affected compute kernels.
+                 */
+		RET((uint64_t []) { dev->arch >= 6 ? 256 : 128 });
 
 	case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
 		RET((uint64_t []) { 1024*1024*512 /* Maybe get memory */ });