[libc] Perform bitfield zero initialization wave-parallel

Summary:
We need to set the bitfield memory to zero because the system does not
guarantee zeroed out memory. Even if fresh pages are zero, the system
allows re-use so we would need a `kfd` level API to skip this step.

Because we can't this patch updates the logic to perform the zero
initialization wave-parallel. This reduces the amount of time it takes
to allocate a fresh by up to a tenth.

This has the unfortunate side effect that the control flow is more
convoluted and we waste some extra registers, but it's worth it to
reduce the slab allocation latency.

Author: jhuber6

libc/src/__support/GPU/allocator.cpp

@@ -157,10 +157,19 @@ struct Slab {
     Header *header = reinterpret_cast<Header *>(memory);
     header->chunk_size = chunk_size;
     header->global_index = global_index;
+  }
 
-    // This memset is expensive and likely not necessary for the current 'kfd'
-    // driver. Until zeroed pages are exposed by the API we must be careful.
-    __builtin_memset(get_bitfield(), 0, bitfield_bytes(chunk_size));
+  // Set the necessary bitfield bytes to zero in parallel using many lanes. This
+  // must be called before the bitfield can be accessed safely, memory is not
+  // guaranteed to be zero initialized in the current implementation.
+  void initialize(uint64_t uniform) {
+    uint64_t mask = gpu::get_lane_mask();
+    uint32_t *bitfield = get_bitfield();
+    uint32_t workers = cpp::popcount(uniform);
+    uint32_t words = (bitfield_bytes(get_chunk_size()) + sizeof(uint32_t) - 1) /
+                     sizeof(uint32_t);
+    for (uint32_t i = impl::lane_count(mask & uniform); i < words; i += workers)
+      bitfield[i] = 0;
   }
 
   // Get the number of chunks that can theoretically fit inside this slab.
@@ -354,14 +363,7 @@ struct GuardPtr {
       void *raw = impl::rpc_allocate(sizeof(Slab));
       if (!raw)
         return nullptr;
-      Slab *mem = new (raw) Slab(cpp::forward<Args>(args)...);
-
-      cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
-      ptr.store(mem, cpp::MemoryOrder::RELAXED);
-      cpp::atomic_thread_fence(cpp::MemoryOrder::ACQUIRE);
-      if (!ref.acquire(n, count))
-        ref.reset(n, count);
-      return mem;
+      return new (raw) Slab(cpp::forward<Args>(args)...);
     }
 
     if (!expected || expected == reinterpret_cast<Slab *>(SENTINEL))
@@ -374,6 +376,16 @@ struct GuardPtr {
     return ptr.load(cpp::MemoryOrder::RELAXED);
   }
 
+  // Finalize the associated memory and signal that it is ready to use by
+  // resetting the counter.
+  void finalize(Slab *mem, uint32_t n, uint64_t &count) {
+    cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
+    ptr.store(mem, cpp::MemoryOrder::RELAXED);
+    cpp::atomic_thread_fence(cpp::MemoryOrder::ACQUIRE);
+    if (!ref.acquire(n, count))
+      ref.reset(n, count);
+  }
+
 public:
   // Attempt to lock access to the pointer, potentially creating it if empty.
   // The uniform mask represents which lanes share the same pointer. For each
@@ -392,6 +404,14 @@ struct GuardPtr {
     if (!result)
       return nullptr;
 
+    // We defer storing the newly allocated slab until now so that we can use
+    // multiple lanes to initialize it and release it for use.
+    if (count == cpp::numeric_limits<uint64_t>::max()) {
+      result->initialize(uniform);
+      if (gpu::get_lane_id() == uint32_t(cpp::countr_zero(uniform)))
+        finalize(result, cpp::popcount(uniform), count);
+    }
+
     if (count != cpp::numeric_limits<uint64_t>::max())
       count = count - cpp::popcount(uniform) + impl::lane_count(uniform) + 1;