Back out "Improve large-buffer crc32c performance by 50-85% [3/3]"

Summary: Original commit changeset: 5ee59d858cec Original Phabricator Diff: D52679940 Reviewed By: udippant Differential Revision: D52915808 fbshipit-source-id: 9aecf413abcd017467511ca16289fba019f71283
facebook · Jan 19, 2024 · b930cba · b930cba
1 parent 4a2f1aa
commit b930cba
Show file tree

Hide file tree

Showing 3 changed files with 4 additions and 118 deletions.
diff --git a/folly/hash/Checksum.cpp b/folly/hash/Checksum.cpp
@@ -22,8 +22,6 @@
 #include <boost/crc.hpp>
 
 #include <folly/CpuId.h>
-#include <folly/external/fast-crc32/avx512_crc32c_v8s3x4.h>
-#include <folly/external/fast-crc32/sse_crc32c_v8s3x3.h>
 #include <folly/hash/detail/ChecksumDetail.h>
 
 #if FOLLY_SSE_PREREQ(4, 2)
@@ -74,11 +72,6 @@ bool crc32c_hw_supported() {
   return id.sse42();
 }
 
-bool crc32c_hw_supported_avx512() {
-  static folly::CpuId id;
-  return id.avx512vl();
-}
-
 bool crc32_hw_supported() {
   static folly::CpuId id;
   return id.sse42();
@@ -97,10 +90,6 @@ bool crc32c_hw_supported() {
   return false;
 }
 
-bool crc32c_hw_supported_avx512() {
-  return false;
-}
-
 bool crc32_hw_supported() {
   return false;
 }
@@ -143,16 +132,8 @@ uint32_t crc32_sw(
 } // namespace detail
 
 uint32_t crc32c(const uint8_t* data, size_t nbytes, uint32_t startingChecksum) {
-  if (detail::crc32c_hw_supported_avx512()) {
-    if (nbytes <= 4096) {
-      return detail::crc32c_hw(data, nbytes, startingChecksum);
-    }
-    return detail::avx512_crc32c_v8s3x4(data, nbytes, startingChecksum);
-  } else if (detail::crc32c_hw_supported()) {
-    if (nbytes <= 4096) {
-      return detail::crc32c_hw(data, nbytes, startingChecksum);
-    }
-    return detail::sse_crc32c_v8s3x3(data, nbytes, startingChecksum);
+  if (detail::crc32c_hw_supported()) {
+    return detail::crc32c_hw(data, nbytes, startingChecksum);
   } else {
     return detail::crc32c_sw(data, nbytes, startingChecksum);
   }

diff --git a/folly/hash/detail/ChecksumDetail.h b/folly/hash/detail/ChecksumDetail.h
@@ -43,17 +43,11 @@ uint32_t crc32c_hw(
     const uint8_t* data, size_t nbytes, uint32_t startingChecksum = ~0U);
 
 /**
- * Check whether a SSE4.2 hardware-accelerated CRC-32C implementation is
+ * Check whether a hardware-accelerated CRC-32C implementation is
  * supported on the current CPU.
  */
 bool crc32c_hw_supported();
 
-/**
- * Check whether an AVX512VL hardware-accelerated CRC-32C implementation is
- * supported on the current CPU.
- */
-bool crc32c_hw_supported_avx512();
-
 /**
  * Compute a CRC-32C checksum of a buffer using a portable,
  * software-only implementation.

diff --git a/folly/hash/test/ChecksumTest.cpp b/folly/hash/test/ChecksumTest.cpp
@@ -20,8 +20,6 @@
 
 #include <folly/Benchmark.h>
 #include <folly/Random.h>
-#include <folly/external/fast-crc32/avx512_crc32c_v8s3x4.h>
-#include <folly/external/fast-crc32/sse_crc32c_v8s3x3.h>
 #include <folly/hash/Hash.h>
 #include <folly/hash/detail/ChecksumDetail.h>
 #include <folly/portability/GFlags.h>
@@ -146,93 +144,6 @@ TEST(Checksum, crc32cContinuationHardware) {
   }
 }
 
-TEST(Checksum, crc32cHardwareSse42) {
-  if (folly::detail::crc32c_hw_supported()) {
-    testCRC32C(folly::detail::sse_crc32c_v8s3x3);
-  } else {
-    LOG(WARNING) << "skipping SSE4.2 hardware-accelerated CRC-32C tests"
-                 << " (not supported on this CPU)";
-  }
-}
-
-TEST(Checksum, crc32cHardwareEqSse42) {
-  if (folly::detail::crc32c_hw_supported()) {
-    for (size_t i = 0; i < 1000; i++) {
-      auto sw = folly::detail::crc32c_sw(buffer, i, 0);
-      auto hw = folly::detail::sse_crc32c_v8s3x3(buffer, i, 0);
-      ASSERT_EQ(sw, hw);
-    }
-  } else {
-    LOG(WARNING) << "skipping SSE4.2 hardware-accelerated CRC-32C tests"
-                 << " (not supported on this CPU)";
-  }
-}
-
-TEST(Checksum, crc32cContinuationHardwareSse42) {
-  if (folly::detail::crc32c_hw_supported()) {
-    testCRC32CContinuation(folly::detail::sse_crc32c_v8s3x3);
-  } else {
-    LOG(WARNING) << "skipping SSE4.2 hardware-accelerated CRC-32C tests"
-                 << " (not supported on this CPU)";
-  }
-}
-
-TEST(Checksum, crc32cHardwareAvx512) {
-  if (folly::detail::crc32c_hw_supported_avx512()) {
-    testCRC32C(folly::detail::avx512_crc32c_v8s3x4);
-  } else {
-    LOG(WARNING) << "skipping AVX512 hardware-accelerated CRC-32C tests"
-                 << " (not supported on this CPU)";
-  }
-}
-
-TEST(Checksum, crc32cHardwareEqAvx512) {
-  if (folly::detail::crc32c_hw_supported_avx512()) {
-    for (size_t i = 0; i < 1000; i++) {
-      auto sw = folly::detail::crc32c_sw(buffer, i, 0);
-      auto hw = folly::detail::avx512_crc32c_v8s3x4(buffer, i, 0);
-      ASSERT_EQ(sw, hw);
-    }
-  } else {
-    LOG(WARNING) << "skipping AVX512 hardware-accelerated CRC-32C tests"
-                 << " (not supported on this CPU)";
-  }
-}
-
-TEST(Checksum, crc32cContinuationHardwareAvx512) {
-  if (folly::detail::crc32c_hw_supported_avx512()) {
-    testCRC32CContinuation(folly::detail::avx512_crc32c_v8s3x4);
-  } else {
-    LOG(WARNING) << "skipping AVX512 hardware-accelerated CRC-32C tests"
-                 << " (not supported on this CPU)";
-  }
-}
-
-// Test on very large buffer inputs to attempt to sanity check 32-bit
-// overflow problems on 64-bit platforms.
-#ifdef __LP64__
-TEST(Checksum, crc32clargeBuffers) {
-  constexpr size_t kLargeBufSz = 5ull * 1024 * 1024 * 1024; // 5GiB
-  auto buf = std::make_unique<uint8_t[]>(kLargeBufSz); // 5GiB
-  auto* bufp = buf.get();
-  // Fill with non-zero pattern.
-  memset(bufp, 0x2e, kLargeBufSz);
-
-  constexpr uint32_t kCrc = 2860399007;
-
-  if (folly::detail::crc32c_hw_supported()) {
-    auto crcSse42 = folly::detail::sse_crc32c_v8s3x3(bufp, kLargeBufSz, ~0);
-    ASSERT_EQ(kCrc, crcSse42);
-    auto crcHw = folly::detail::crc32c_hw(bufp, kLargeBufSz, ~0);
-    ASSERT_EQ(kCrc, crcHw);
-  }
-  if (folly::detail::crc32c_hw_supported_avx512()) {
-    auto crcAvx = folly::detail::avx512_crc32c_v8s3x4(bufp, kLargeBufSz, ~0);
-    ASSERT_EQ(kCrc, crcAvx);
-  }
-}
-#endif
-
 TEST(Checksum, crc32cAutodetect) {
   testCRC32C(folly::crc32c);
 }
@@ -468,7 +379,7 @@ int main(int argc, char** argv) {
   // on which to compute checksums
   const uint8_t* src = buffer;
   uint64_t* dst = (uint64_t*)buffer;
-  const uint64_t* end = (const uint64_t*)(buffer + sizeof(buffer));
+  const uint64_t* end = (const uint64_t*)(buffer + BUFFER_SIZE);
   *dst++ = 0;
   while (dst < end) {
     *dst++ = folly::hash::fnv64_buf((const char*)src, sizeof(uint64_t));