workaround rocm-6.2 compiler issue (#1421)

example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py

@@ -427,12 +427,19 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm
 
                     pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
                     pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                else:
+                    if bias == "bias":
+                        # TODO: rocm 6.2 compiler problem if using qr_async for bias case
+                        pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
                     else:
                         pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
                         pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
                         pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
                         pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
-                    if receipt == 1:
+                    if receipt == 1 and bias != "bias":
                         pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
                         pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
         elif dtype in ['fp8', 'bf8']:

include/ck_tile/core/tensor/tile_window.hpp

@@ -393,7 +393,10 @@ struct tile_window_with_static_distribution
                     bottom_tensor_thread_coord,
                     bool_constant<oob_conditional_check>{},
                     pre_nop_);
-
+#if CK_TILE_WORKAROUND_ROCM_6_1_SCRATCH_MEMORY_ISSUE
+                asm volatile(
+                    ""); // this is starting from rocm-6.2, but same sympton, reuse this flag
+#endif
                 // move thread coordinate
                 if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
                 {

include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp

@@ -231,7 +231,9 @@ struct BlockFmhaPipelineQRKSVSAsync
         // TODO: we use async Copy for K, which is inline asm
         // a side effect is we have to use inline asm for q as well
         auto q = decltype(load_tile(q_dram_window)){};
-        set_tile(q, number<0>{}); // use per-dword clear to avoid scratch
+        // TODO: start from rocm-6.2, compiler will have problem if manually set clear of q.
+        // however, q would be cleared in the constructor of static distributed tensor
+        // set_tile(q, number<0>{}); // use per-dword clear to avoid scratch
         load_tile_raw(q, q_dram_window);
         __builtin_amdgcn_sched_barrier(0);