feat(RISC-V): Add RVV-optimized implementation for memcopy64 by anthony-zy · Pull Request #213 · google/snappy

anthony-zy · 2025-09-12T03:29:25Z

This PR add RVV to optimize Memcopy64 function , improving compression speed by ~46%.

Optimize Snappy 1.2.2 performance

Added RVV support for RISC-V in Snappy, optimizing Memcopy64 by leveraging RVV vector load/store instructions (e.g., vle8_v_u8m2, vse8_v_u8m2) to reduce memory copy overhead and improve decompression performance. lzbench 2.1 tests onsilesia.tar (GCC 13.2.1, 64-bit Linux) show:

Compressor	Compress	Decompress	Size	Ratio
Snappy 1.2.2 (Before)	64.3 MB/s	186 MB/s	101403263	47.84%
Snappy 1.2.2 (After)	64.4 MB/s	272 MB/s	101403263	47.84%

Snappy 1.2.2 unittest ([ PASSED ] 21 tests.)

Running main() from gmock_main.cc
[==========] Running 21 tests from 3 test suites.
[----------] Global test environment set-up.
[----------] 1 test from CorruptedTest
[ RUN ] CorruptedTest.VerifyCorrupted
Crazy decompression lengths not checked on 64-bit build
[ OK ] CorruptedTest.VerifyCorrupted (5 ms)
[----------] 1 test from CorruptedTest (5 ms total)

[----------] 17 tests from Snappy
[ RUN ] Snappy.SimpleTests
[ OK ] Snappy.SimpleTests (22 ms)
[ RUN ] Snappy.AppendSelfPatternExtensionEdgeCases
[ OK ] Snappy.AppendSelfPatternExtensionEdgeCases (3 ms)
[ RUN ] Snappy.AppendSelfPatternExtensionEdgeCasesExhaustive
[ OK ] Snappy.AppendSelfPatternExtensionEdgeCasesExhaustive (4830 ms)
[ RUN ] Snappy.MaxBlowup
[ OK ] Snappy.MaxBlowup (11 ms)
[ DISABLED ] Snappy.DISABLED_MoreThan4GB
[ RUN ] Snappy.RandomData
[ OK ] Snappy.RandomData (20311 ms)
[ RUN ] Snappy.FourByteOffset
[ OK ] Snappy.FourByteOffset (1 ms)
[ RUN ] Snappy.IOVecSourceEdgeCases
[ OK ] Snappy.IOVecSourceEdgeCases (0 ms)
[ RUN ] Snappy.IOVecSinkEdgeCases
[ OK ] Snappy.IOVecSinkEdgeCases (0 ms)
[ RUN ] Snappy.IOVecLiteralOverflow
[ OK ] Snappy.IOVecLiteralOverflow (0 ms)
[ RUN ] Snappy.IOVecCopyOverflow
[ OK ] Snappy.IOVecCopyOverflow (0 ms)
[ RUN ] Snappy.ReadPastEndOfBuffer
[ OK ] Snappy.ReadPastEndOfBuffer (0 ms)
[ RUN ] Snappy.ZeroOffsetCopy
[ OK ] Snappy.ZeroOffsetCopy (0 ms)
[ RUN ] Snappy.ZeroOffsetCopyValidation
[ OK ] Snappy.ZeroOffsetCopyValidation (0 ms)
[ RUN ] Snappy.FindMatchLength
[ OK ] Snappy.FindMatchLength (0 ms)
[ RUN ] Snappy.FindMatchLengthRandom
[ OK ] Snappy.FindMatchLengthRandom (1145 ms)
[ RUN ] Snappy.VerifyCharTable
[ OK ] Snappy.VerifyCharTable (0 ms)
[ RUN ] Snappy.TestBenchmarkFiles
[ OK ] Snappy.TestBenchmarkFiles (406 ms)
[----------] 17 tests from Snappy (26734 ms total)

[----------] 3 tests from SnappyCorruption
[ RUN ] SnappyCorruption.TruncatedVarint
[ OK ] SnappyCorruption.TruncatedVarint (0 ms)
[ RUN ] SnappyCorruption.UnterminatedVarint
[ OK ] SnappyCorruption.UnterminatedVarint (0 ms)
[ RUN ] SnappyCorruption.OverflowingVarint
[ OK ] SnappyCorruption.OverflowingVarint (0 ms)
[----------] 3 tests from SnappyCorruption (0 ms total)

[----------] Global test environment tear-down
[==========] 21 tests from 3 test suites ran. (26740 ms total)
[ PASSED ] 21 tests.

YOU HAVE 1 DISABLED TEST
[ PASSED ] 21 tests.

…peed For RISC-V[skip ci]

…into add_rvv_support

anthony-zy · 2025-09-16T01:51:44Z

@danilak-G
I've revised the code, reducing branches and optimizing preprocessor directives for easier review. Could you check the current version when possible?
Also, in Snappy's maintenance mode, what's the review timeline and what changes are typically accepted? Thanks!

anthony-zy · 2025-10-09T03:10:09Z

@danilak-G I've revised the code, reducing branches and optimizing preprocessor directives for easier review. Could you check the current version when possible? Also, in Snappy's maintenance mode, what's the review timeline and what changes are typically accepted? Thanks!

Hi @danilak-G,
Just a quick follow-up on this PR. I've made some revisions for easier review. Any chance you could take a look when you have a moment?
Thanks!"

danilak-G · 2025-10-16T17:56:24Z

Overall LGTM, thanks for fixing many things and now it looks greate, some small comments

danilak-G · 2025-10-16T17:55:04Z

snappy.cc

+  size_t remaining_bytes = size;
+  //Loop as long as there are bytes remaining to be copied.
+  while (remaining_bytes > 0) {
+    //Set vector configuration: e8 (8-bit elements), m2 (LMUL=2).


Can you fix the indentation to 4 spaces?

danilak-G · 2025-10-16T17:55:15Z

snappy.cc

+  // RVV acceleration available on RISC-V when compiled with -march=rv64gcv
+#elif defined(__riscv) && SNAPPY_HAVE_RVV
+  // Cast pointers to the type we will operate on.
+  unsigned char* dst_ptr = (unsigned char*)dst;


Use reinterpret_cast, please

danilak-G · 2025-10-16T17:55:41Z

snappy.cc

+  vuint8m2_t vec = VLE8_V_U8M2(src_ptr, vl);
+    //Store data to the current destination pointer.
+  VSE8_V_U8M2(dst_ptr, vec, vl);
+    //Update pointers and the remaining count.


Set the space after // for all comments

danilak-G · 2025-10-16T17:55:53Z

snappy.cc

-                 static_cast<const uint8_t*>(src) + kShortMemCopy,
-                 64 - kShortMemCopy);
-  }
+                  static_cast<const uint8_t*>(src) + kShortMemCopy,


Extra not-needed space added

…into add_rvv_support

…ny-zy/snappy into add_rvv_optmized_memcopy64

anthony-zy · 2025-10-17T06:53:57Z

Overall LGTM, thanks for fixing many things and now it looks greate, some small comments

Fixed, thanks for the review!

danilak-G · 2025-10-17T09:36:39Z

snappy.cc

    std::memmove(dst + kShortMemCopy,
                 static_cast<const uint8_t*>(src) + kShortMemCopy,
                 64 - kShortMemCopy);
-  }


This closing bracket should not be deleted

Sure—bracket restored. Thanks for catching that!

danilak-G · 2025-10-17T09:36:47Z

snappy.cc

 #else
  std::memmove(dst, src, kShortMemCopy);
-  // Profiling shows that nearly all copies are short.
+    //Profiling shows that nearly all copies are short.


A space here as well

danilak-G · 2025-10-17T09:36:57Z

snappy.cc

+    size_t vl = VSETVL_E8M2(remaining_bytes);
+    // Load data from the current source pointer.
+    vuint8m2_t vec = VLE8_V_U8M2(src_ptr, vl);
+      // Store data to the current destination pointer.


Align comments with code

…ny-zy/snappy into add_rvv_optmized_memcopy64

danilak-G · 2025-10-17T10:07:53Z

snappy.cc

+// Suppress -Wignored-attributes warning for __m128i in x86 SSE2 environment
+// warning: ignoring attributes on template argument 'snappy::internal::V128' {aka '__vector(2) long long int'} [-Wignored-attributes]
+// This occurs because __m128i has vector attributes (e.g., __attribute__((vector_size(16)))) that are ignored in template parameters.
+#ifdef __SSE2__


This is deleting LoadPatternAndReshuffleMask completely for non SSE2. SNAPPY_HAVE_VECTOR_BYTE_SHUFFLE should be correct, I guess

There is a warning when compiling here on x86.I reverted to the previous code.

…ny-zy/snappy into add_rvv_optmized_memcopy64

anthony-zy added 9 commits August 26, 2025 20:51

Use high-performance FindMatchLength to Optimize Snappy compression s…

758804a

…peed For RISC-V[skip ci]

add RVV support and optimized uncompress speed

1f6ba69

add RVV support and optmized uncompress speed

802ef73

Merge branch 'add_rvv_support' of https://github.com/anthony-zy/snappy …

0f989a2

…into add_rvv_support

Merge branch 'add_rvv_support' of https://github.com/anthony-zy/snappy …

eb8d19f

…into add_rvv_support

Merge branch 'add_rvv_support' of https://github.com/anthony-zy/snappy …

27d8915

…into add_rvv_support

Merge branch 'google:main' into add_rvv_support

4410326

Merge branch 'google:main' into add_rvv_support

c860cab

Merge branch 'add_rvv_support' of https://github.com/anthony-zy/snappy …

2b81365

…into add_rvv_support

danilak-G reviewed Oct 16, 2025

View reviewed changes

anthony-zy added 4 commits October 17, 2025 10:20

Merge branch 'add_rvv_support' of https://github.com/anthony-zy/snappy …

c19d705

…into add_rvv_support

Merge branch 'add_rvv_optmized_memcopy64' of https://github.com/antho…

77a61a4

…ny-zy/snappy into add_rvv_optmized_memcopy64

fix by some comments

9fd2b72

Merge branch 'add_rvv_optmized_memcopy64' of https://github.com/antho…

d457ac7

…ny-zy/snappy into add_rvv_optmized_memcopy64

anthony-zy requested a review from danilak-G October 17, 2025 09:02

danilak-G reviewed Oct 17, 2025

View reviewed changes

anthony-zy added 2 commits October 17, 2025 17:52

Merge branch 'add_rvv_optmized_memcopy64' of https://github.com/antho…

6ab5882

…ny-zy/snappy into add_rvv_optmized_memcopy64

Merge branch 'add_rvv_optmized_memcopy64' of https://github.com/antho…

c098f68

…ny-zy/snappy into add_rvv_optmized_memcopy64

danilak-G reviewed Oct 17, 2025

View reviewed changes

anthony-zy added 4 commits October 20, 2025 09:53

Merge branch 'add_rvv_optmized_memcopy64' of https://github.com/antho…

9cc1596

…ny-zy/snappy into add_rvv_optmized_memcopy64

Merge branch 'add_rvv_optmized_memcopy64' of https://github.com/antho…

b63569f

…ny-zy/snappy into add_rvv_optmized_memcopy64

Merge branch 'add_rvv_optmized_memcopy64' of https://github.com/antho…

2d6c8f1

…ny-zy/snappy into add_rvv_optmized_memcopy64

Merge branch 'add_rvv_optmized_memcopy64' of https://github.com/antho…

e92cb6a

…ny-zy/snappy into add_rvv_optmized_memcopy64

danilak-G merged commit 633cb9c into google:main Oct 20, 2025
1 check passed

Conversation

anthony-zy commented Sep 12, 2025

Optimize Snappy 1.2.2 performance

Snappy 1.2.2 unittest ([ PASSED ] 21 tests.)

Uh oh!

anthony-zy commented Sep 16, 2025

Uh oh!

anthony-zy commented Oct 9, 2025

Uh oh!

danilak-G commented Oct 16, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anthony-zy commented Oct 17, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants