OSS MPZCH CUDA kernel in FBGEMM

fbgemm_gpu/FbgemmGpu.cmake

@@ -55,6 +55,11 @@ if(NOT FBGEMM_BUILD_VARIANT STREQUAL BUILD_VARIANT_CPU)
     src/metric_ops/metric_ops_host.cpp
     src/input_combine_ops/input_combine_gpu.cpp)
 
+  if(NOT FBGEMM_BUILD_VARIANT STREQUAL BUILD_VARIANT_ROCM)
+    list(APPEND fbgemm_gpu_sources_cpu_static
+      src/faster_hash_ops/faster_hash.cpp)
+  endif()
+
   if(NVML_LIB_PATH OR FBGEMM_BUILD_VARIANT STREQUAL BUILD_VARIANT_ROCM)
     message(STATUS "Adding merge_pooled_embeddings sources")
     list(APPEND fbgemm_gpu_sources_cpu_static
@@ -122,6 +127,11 @@ if(NOT FBGEMM_BUILD_VARIANT STREQUAL BUILD_VARIANT_CPU)
       src/sparse_ops/sparse_reorder_batched_ad.cu
       src/sparse_ops/sparse_segment_sum_csr.cu
       src/sparse_ops/sparse_zipf.cu)
+
+  if(NOT FBGEMM_BUILD_VARIANT STREQUAL BUILD_VARIANT_ROCM)
+    list(APPEND fbgemm_gpu_sources_gpu_static
+      src/faster_hash_ops/faster_hash.cu)
+  endif()
 endif()
 
 

fbgemm_gpu/include/fbgemm_gpu/faster_hash_ops/common_utils.cuh

@@ -0,0 +1,137 @@
+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (C) 2016 ExplosionAI GmbH, 2014-2015 Matthew Honnibal, 2016 spaCy
+ * GmbH
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ *
+ */
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <ATen/ATen.h>
+
+#define AT_DISPATCH_INTEGER_TYPES(TYPE, NAME, HINT, ...)                      \
+  AT_DISPATCH_SWITCH(                                                         \
+      TYPE,                                                                   \
+      NAME,                                                                   \
+      AT_PRIVATE_CASE_TYPE_USING_HINT(at::ScalarType::Int, HINT, __VA_ARGS__) \
+          AT_PRIVATE_CASE_TYPE_USING_HINT(                                    \
+              at::ScalarType::Long, HINT, __VA_ARGS__))
+
+namespace fbgemm_gpu {
+
+#if defined(TORBOREC_CUDA)
+#define TORBOREC_INLINE __device__ __host__ __inline__
+#else
+#define TORBOREC_INLINE inline
+#endif
+
+// NOLINTNEXTLINE:
+TORBOREC_INLINE uint64_t
+murmur_hash3_2x64(const uint64_t x, const uint64_t y, const uint64_t seed) {
+  const uint64_t c1 = 0x87c37b91114253d5;
+  const uint64_t c2 = 0x4cf5ad432745937f;
+
+  uint64_t h1 = seed;
+  uint64_t h2 = seed;
+
+  // First 64-bit block
+  uint64_t k1 = x;
+  k1 *= c1;
+  k1 = (k1 << 31) | (k1 >> (64 - 31));
+  k1 *= c2;
+  h1 ^= k1;
+  h1 = (h1 << 27) | (h1 >> (64 - 27));
+  h1 += h2;
+  h1 = h1 * 5 + 0x52dce729;
+
+  // Second 64-bit block
+  uint64_t k2 = y;
+  k2 *= c2;
+  k2 = (k2 << 33) | (k2 >> (64 - 33));
+  k2 *= c1;
+  h2 ^= k2;
+  h2 = (h2 << 31) | (h2 >> (64 - 31));
+  h2 += h1;
+  h2 = h2 * 5 + 0x38495ab5;
+
+  // Finalization
+  h1 ^= 16;
+  h2 ^= 16;
+  h1 += h2;
+  h2 += h1;
+  h1 ^= h1 >> 33;
+  h1 *= 0xff51afd7ed558ccd;
+  h1 ^= h1 >> 33;
+  h1 *= 0xc4ceb9fe1a85ec53;
+  h1 ^= h1 >> 33;
+  h2 ^= h2 >> 33;
+  h2 *= 0xff51afd7ed558ccd;
+  h2 ^= h2 >> 33;
+  h2 *= 0xc4ceb9fe1a85ec53;
+  h2 ^= h2 >> 33;
+  h1 += h2;
+  h2 += h1;
+
+  return h1 ^ h2;
+}
+
+// NOLINTNEXTLINE:
+template <bool CIRCULAR_PROBE>
+TORBOREC_INLINE int64_t next_output_index(
+    int64_t output_index,
+    int64_t modulo,
+    int64_t& /* max_probe_local */) {
+  static_assert(CIRCULAR_PROBE);
+  return (output_index + 1) % modulo;
+}
+
+// NOLINTNEXTLINE:
+template <>
+TORBOREC_INLINE int64_t next_output_index<false>(
+    int64_t output_index,
+    int64_t modulo,
+    int64_t& max_probe_local) {
+  output_index = (output_index + 1) % modulo;
+  if (output_index == 0) {
+    // circular, using max_probe_local to control exit.
+    max_probe_local = 0;
+  }
+  return output_index;
+}
+
+TORBOREC_INLINE bool is_eviction_enabled(
+    bool readonly,
+    int eviction_threshold,
+    int eviction_policy) {
+  return !readonly && (eviction_threshold > 0 || eviction_policy > 0);
+}
+
+#undef TORBOREC_INLINE
+
+} // namespace fbgemm_gpu

fbgemm_gpu/include/fbgemm_gpu/faster_hash_ops/faster_hash_ops.cuh

@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <ATen/ATen.h>
+
+/// @defgroup faster-hash-ops CUDA Operators
+/// The following are CUDA Operators
+
+namespace fbgemm_gpu {
+
+using at::Tensor;
+
+///@ingroup faster-hash-ops
+///
+/// CUDA implementation of zero collision hash
+///
+/// @param output the output tensor that will be modified in place
+/// @param evict_slots the slots that will be evicted
+/// @param input the input tensor
+/// @param identities the identity tensor
+/// @param max_probe the maximum number of probes
+/// @param circular_probe whether to use circular probe
+/// @param cur_hour the current hour
+/// @param readonly whether to use readonly mode
+/// @param support_evict whether to support evict
+/// @param local_sizes the local sizes tensor
+/// @param offsets the offsets tensor
+/// @param hash_identity whether to hash the identity
+/// @param metadata the metadata tensor
+/// @param disable_fallback whether to disable fallback
+/// @param input_metadata the input metadata tensor
+/// @param eviction_threshold the eviction threshold
+/// @param eviction_policy the eviction policy
+/// @param opt_in_prob the opt-in probability
+/// @param num_reserved_slots the number of reserved slots
+/// @param opt_in_rands the opt-in randoms tensor
+///
+/// @return None
+template <typename TInput, typename TIdentity>
+void _zero_collision_hash_cuda(
+    Tensor& output,
+    Tensor& evict_slots,
+    const Tensor& input,
+    Tensor& identities,
+    int64_t max_probe,
+    bool circular_probe,
+    int64_t cur_hour,
+    bool readonly,
+    bool support_evict,
+    const std::optional<Tensor>& local_sizes,
+    const std::optional<Tensor>& offsets,
+    int32_t hash_identity,
+    const std::optional<Tensor>& metadata,
+    bool disable_fallback,
+    const std::optional<Tensor>& input_metadata,
+    int64_t eviction_threshold,
+    int64_t eviction_policy,
+    int64_t opt_in_prob,
+    int64_t num_reserved_slots,
+    const std::optional<Tensor>& opt_in_rands);
+
+///@ingroup faster-hash-ops
+///
+/// CUDA implementation of murmurhash3
+///
+/// @param input the input tensor
+/// @param y the y value
+/// @param seed the seed value
+
+/// @return the output tensor
+Tensor murmur_hash3_cuda(const Tensor& input, int64_t y, int64_t seed);
+
+} // namespace fbgemm_gpu