Skip to content

Fizzy 0.4.0
Compare
Choose a tag to compare
@axic axic released this 01 Sep 19:07
· 811 commits to master since this release
v0.4.0
bbdb5f8

This release introduces complete floating-point support.

With that in place, Fizzy passes all(*) of the official test suite (spectest 1.0):

  • 18896 of 18899 binary parser and execution tests,
  • 989 of 989 validation tests,
  • 477 skipped due to testing text format parser.

(*) The three failures are caused by, in our opinion, two bugs in the test cases, and the last one because we restrict memory to 256 Mb maximum during instantiation, while the test expects 4 Gb to succeed.

The addition of the new set of instructions causes changes to the execution times. Effects of some changes are truly surprising, e.g. 67fd4c8 causes over -10% performance regression on GCC 10 by adding two bytes of unused code; and bffe032, which adds two new instructions, makes binaries built with Clang 10 much faster. In the end we decided to start using Link-Time Optimized builds for benchmarking to mitigate some of these butterfly effects. See also README: Performance testing.

The performance regression of the execution time is +7% for GCC 10 and +2% for Clang 11. This is expected due to increased code size of the interpreter loop (increased iTLB-load-misses and LLC-load-misses have been observed). We plan to mitigate the impact of the change in the future by using [Profile-Guided Optimization].

The performance of Wasm binary parsing and instantiation remains unchanged.

Comparing Fizzy 0.3.0 to 0.4.0 (Intel Haswell CPU 4.0 GHz, GCC 10, LTO)
Benchmark                                  CPU Time [µs]       Old       New
----------------------------------------------------------------------------

fizzy/execute/blake2b/512_bytes_rounds_1         +0.0857        78        84
fizzy/execute/blake2b/512_bytes_rounds_16        +0.0967      1164      1276
fizzy/execute/ecpairing/onepoint                 +0.0734    395683    424732
fizzy/execute/keccak256/512_bytes_rounds_1       +0.0457        94        98
fizzy/execute/keccak256/512_bytes_rounds_16      +0.0492      1374      1442
fizzy/execute/memset/256_bytes                   +0.1102         6         7
fizzy/execute/memset/60000_bytes                 +0.1106      1390      1544
fizzy/execute/mul256_opt0/input0                 +0.0205        25        26
fizzy/execute/mul256_opt0/input1                 +0.0198        25        26
fizzy/execute/sha1/512_bytes_rounds_1            +0.0477        86        90
fizzy/execute/sha1/512_bytes_rounds_16           +0.0467      1196      1252
fizzy/execute/sha256/512_bytes_rounds_1          +0.0885        83        91
fizzy/execute/sha256/512_bytes_rounds_16         +0.0951      1142      1250
fizzy/execute/micro/eli_interpreter/halt         +0.1457         0         0
fizzy/execute/micro/eli_interpreter/exec105      +0.1711         4         5
fizzy/execute/micro/factorial/10                 +0.0643         0         0
fizzy/execute/micro/factorial/20                 +0.0213         1         1
fizzy/execute/micro/fibonacci/24                 +0.0518      7131      7500
fizzy/execute/micro/host_adler32/1               +0.0242         0         0
fizzy/execute/micro/host_adler32/100             +0.0054         3         3
fizzy/execute/micro/host_adler32/1000            -0.0126        30        29
fizzy/execute/micro/spinner/1                    +0.0860         0         0
fizzy/execute/micro/spinner/1000                 +0.1976         8        10

Added

  • f32.const, f64.const, f32.add, f64.add instructions. #424 #467
  • f32.sub, f64.sub instructions. #474
  • f32.mul, f64.mul instructions. #473
  • f32.div, f64.div instructions. #468
  • f32.sqrt, f64.sqrt instructions. #480
  • f32.copysign, f64.copysign instructions. #471
  • f32.abs, f64.abs instructions. #476
  • f32.neg, f64.neg instructions. #477
  • Floating-point comparison instructions: f32.eq, f32.ne, f32.lt, f32.gt, f32.le, f64.ge, f64.eq, f64.ne, f64.lt, f64.gt, f64.le, f64.ge. #449
  • f32.min, f32.max, f64.min, f64.max instructions. #472
  • f32.ceil, f32.floor, f32.trunc, f64.ceil, f64.floor, f64.trunc instructions. #481
  • f32.nearest, f64.nearest instructions. #486
  • Floating-point to integer truncation instructions: i32.trunc_f32_s, i32.trunc_f32_u, i32.trunc_f64_s, i32.trunc_f64_u, i64.trunc_f32_s, i64.trunc_f32_u,i64.trunc_f64_s, i64.trunc_f64_u. #447
  • Integer to floating-point conversion instructions: f32.convert_i32_s, f32.convert_i32_u, f32.convert_i64_s, f32.convert_i64_u, f64.convert_i32_s, f64.convert_i32_u, f64.convert_i64_s, f64.convert_i64_u. #455
  • f64.promote_f32 instruction. #457
  • f32.demote_f64 instruction. #458
  • Reinterpret instructions: i32.reinterpret_f32, i64.reinterpret_f64, f32.reinterpret_i32, f64.reinterpret_i64. #459
  • Floating-point memory instructions: f32.load, f64.load, f32.store, f64.store. #456
  • Floating-point globals support. #446
  • fizzy-spectests support for floating-point tests. #445 #460 #484
  • Comprehensive tests for all NaNs and all rounding directions. #475 #487 #488 #500
  • A single-precision floating-point benchmark approximating pi using Taylor's series. #482
  • A double-precision floating-point benchmark approximating pi using Ramanujan's algorithm. #483

Changed

  • Execution API now uses union-based Value type instead of uint64_t to represent values of all supported types. #419 #423 #448 #462 #464
  • fizzy-spectests output can be configured for various verbosity. #453
  • ExternalGlobal now includes type of value, which is checked against module definition for imported globals. #493
  • Mark execute() completely exception free (noexcept). #438
Assets 2
Loading