Optimize FlashMask v3 performance (PaddlePaddle#75737)

* tune bwd tile size

* tune bwd tile size for seqlen <= 8192

* fix cuda 700 cause by incorrect bwd tile size

* set scheduler_needs_semaphore to true

* update fa submodule

* update fa submodule

* update fa submodule

* update fa submodule

Author: umiswing

paddle/phi/kernels/gpu/flash_attn_v3_grad_kernel.cu

@@ -1026,11 +1026,9 @@ void FlashMaskV2GradBaseKernel(
                 ? 64
                 : (is_causal && softcap || is_flashmask > 0.0 ? 96 : 128)
           : (head_size_rounded <= 128
-                 ? (is_flashmask && !is_causal)
-                       ? 64
-                       : (is_causal || is_local || is_flashmask || softcap > 0.0
-                              ? 64
-                              : 80)
+                 ? (is_causal || is_local || softcap > 0.0
+                        ? 64
+                        : 64)
                  : 64);
 
   int const kBlockM_sm80 = head_size_rounded <= 64 ? 128 : 64;
@@ -1040,7 +1038,7 @@ void FlashMaskV2GradBaseKernel(
                  : (arch == 86 || arch == 89 ? kBlockM_sm86 : kBlockM_sm80);
   int const kBlockN_sm90 =
       head_size_rounded <= 64 && (is_flashmask && !is_causal) ? 96
-      : head_size_rounded <= 128 ? (is_flashmask && !is_causal) ? 64 : 128
+      : head_size_rounded <= 128 ? (is_causal || is_local || softcap > 0.f || seqlen_q >= 8192 ? 128 : 64)
                                  : (head_size_rounded <= 192 ? 96 : 80);
   int const kBlockN_sm80 =
       head_size_rounded <= 128 ? 128 : (head_size_rounded <= 192 ? 80 : 64);

paddle/phi/kernels/gpu/flash_attn_v3_kernel.cu

@@ -1762,9 +1762,7 @@ void FlashMaskV2BaseKernel(
   const int params_arch =
       phi::dynload::flashmaskv2_fwd_params_get_arch(params_handle);
   bool const scheduler_needs_semaphore =
-      params_arch >= 90 ? (((params_is_causal || params_is_local) &&
-                            (params_num_splits == 1)) ||
-                           is_varlen)
+      params_arch >= 90 ? true
                         : ((params_is_causal && !is_varlen) ||
                            (is_varlen && params_num_splits > 1));
   if (scheduler_needs_semaphore || use_dynamic_split) {