Claude (noether)
97629e935c
Live instrumentation on panthor: WLS reuse verified working (644 REUSE/46 ALLOC) but the crash is dominated by the TLS *stack* alloc in emit_tls() at EndCommandBuffer — L-variant matmul shader spills 10x more registers (tls_size 544->5328 B/thread => 38MB->304MB stack), x~12 cmdbufs retained in the tls pool => ~3GB, over the 4GB priv heap. Remaining fix is structural (free tls-pool BOs post-submit and/or fix the x19 sparse-core overallocation). See docs/root-cause-final.md. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
75 lines
3.1 KiB
Diff
75 lines
3.1 KiB
Diff
--- a/src/panfrost/vulkan/csf/panvk_cmd_buffer.h
|
|
+++ src/panfrost/vulkan/csf/panvk_cmd_buffer.h
|
|
@@ -409,6 +409,11 @@
|
|
struct pan_ptr desc;
|
|
struct pan_tls_info info;
|
|
unsigned max_wg_count;
|
|
+
|
|
+ /* rocky-vulkan-llama: grow-only WLS buffer reused across dispatch
|
|
+ * commands (PANVK_WLS_REUSE) instead of one BO per dispatch. */
|
|
+ uint64_t wls_ptr;
|
|
+ unsigned wls_size;
|
|
};
|
|
|
|
struct panvk_cond_render_state {
|
|
--- a/src/panfrost/vulkan/csf/panvk_vX_cmd_dispatch.c
|
|
+++ src/panfrost/vulkan/csf/panvk_vX_cmd_dispatch.c
|
|
@@ -13,6 +13,7 @@
|
|
#include "genxml/gen_macros.h"
|
|
|
|
#include "panvk_buffer.h"
|
|
+#include <stdio.h>
|
|
#include "panvk_cmd_alloc.h"
|
|
#include "panvk_cmd_buffer.h"
|
|
#include "panvk_cmd_desc_state.h"
|
|
@@ -89,8 +90,14 @@
|
|
unsigned core_id_range;
|
|
pan_query_core_count(&phys_dev->kmod.dev->props, &core_id_range);
|
|
|
|
+ /* rocky-vulkan-llama: size WLS with the grid-INDEPENDENT upper bound
|
|
+ * on instances (max_instances_per_core), not the per-dispatch grid.
|
|
+ * Makes wls_total_size constant per pipeline so the grow-only reuse
|
|
+ * below collapses to ~1 BO per distinct shader wls_size for the whole
|
|
+ * command buffer instead of one-per-new-high-water. Mirrors the
|
|
+ * indirect path and the JM/gallium single-batch-WLS model. */
|
|
tlsinfo.wls.instances = pan_calc_wls_instances(
|
|
- &cs->cs.local_size, &phys_dev->kmod.dev->props, indirect ? NULL : dim);
|
|
+ &cs->cs.local_size, &phys_dev->kmod.dev->props, NULL);
|
|
|
|
unsigned wls_total_size = pan_calc_total_wls_size(
|
|
tlsinfo.wls.size, tlsinfo.wls.instances, core_id_range);
|
|
@@ -100,10 +107,29 @@
|
|
* instance count) might differ significantly between dispatch commands,
|
|
* rather than track a single maximum size, we might want to consider
|
|
* multiple allocations for different size buckets. */
|
|
- tlsinfo.wls.ptr =
|
|
- panvk_cmd_alloc_dev_mem(cmdbuf, tls, wls_total_size, 4096).gpu;
|
|
- if (!tlsinfo.wls.ptr)
|
|
- return 0;
|
|
+ /* rocky-vulkan-llama fix (addresses the TODO above): reuse a grow-only
|
|
+ * WLS buffer across dispatches. CSF allocated a fresh WLS BO per
|
|
+ * dispatch, exhausting the sub-4GB priv VA heap over a ~300-dispatch
|
|
+ * LLM prefill graph -> OOM at ub>64. Gated by PANVK_WLS_REUSE (A/B). */
|
|
+ const bool wls_reuse = getenv("PANVK_WLS_REUSE") != NULL;
|
|
+ bool wls_did_reuse = false;
|
|
+ if (wls_reuse && wls_total_size <= cmdbuf->state.tls.wls_size) {
|
|
+ tlsinfo.wls.ptr = cmdbuf->state.tls.wls_ptr;
|
|
+ wls_did_reuse = true;
|
|
+ } else {
|
|
+ tlsinfo.wls.ptr =
|
|
+ panvk_cmd_alloc_dev_mem(cmdbuf, tls, wls_total_size, 4096).gpu;
|
|
+ if (!tlsinfo.wls.ptr)
|
|
+ return 0;
|
|
+ if (wls_reuse) {
|
|
+ cmdbuf->state.tls.wls_ptr = tlsinfo.wls.ptr;
|
|
+ cmdbuf->state.tls.wls_size = wls_total_size;
|
|
+ }
|
|
+ }
|
|
+ if (getenv("PANVK_WLS_LOG"))
|
|
+ fprintf(stderr, "[panvk-wls] total_size=%u %s prev_max=%u\n",
|
|
+ wls_total_size, wls_did_reuse ? "REUSE" : "ALLOC",
|
|
+ cmdbuf->state.tls.wls_size);
|
|
}
|
|
|
|
cmdbuf->state.tls.info.tls.size =
|