Skip to content

Commit

Permalink
Add benchmarks results for 9aa82c6fb90e1dcd6e7f60626255d597ef0fdea1
Browse files Browse the repository at this point in the history
  • Loading branch information
github-actions committed Mar 13, 2024
1 parent 6468bd6 commit 07a1e92
Show file tree
Hide file tree
Showing 5 changed files with 895 additions and 5 deletions.
68 changes: 67 additions & 1 deletion analyze.json
View file
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
{
"lastUpdate": 1710160575440,
"lastUpdate": 1710360181784,
"repoUrl": "https://github.com/luau-lang/luau",
"entries": {
"luau-analyze": [
Expand Down Expand Up @@ -11416,6 +11416,72 @@
"extra": "luau-analyze"
}
]
},
{
"commit": {
"author": {
"email": "[email protected]",
"name": "Arseny Kapoulkine",
"username": "zeux"
},
"committer": {
"email": "[email protected]",
"name": "GitHub",
"username": "web-flow"
},
"distinct": true,
"id": "9aa82c6fb90e1dcd6e7f60626255d597ef0fdea1",
"message": "CodeGen: Improve lowering of NUM_TO_VEC on A64 for constants (#1194)\n\nWhen the input is a constant, we use a fairly inefficient sequence of\r\nfmov+fcvt+dup or, when the double isn't encodable in fmov,\r\nadr+ldr+fcvt+dup.\r\n\r\nInstead, we can use the same lowering as X64 when the input is a\r\nconstant, and load the vector from memory. However, if the constant is\r\nencodable via fmov, we can use a vector fmov instead (which is just one\r\ninstruction and doesn't need constant space).\r\n\r\nFortunately the bit encoding of fmov for 32-bit floating point numbers\r\nmatches that of 64-bit: the decoding algorithm is a little different\r\nbecause it expands into a larger exponent, but the values are\r\ncompatible, so if a double can be encoded into a scalar fmov with a\r\ngiven abcdefgh pattern, the same pattern should encode the same float;\r\ndue to the very limited number of mantissa and exponent bits, all values\r\nthat are encodable are also exact in both 32-bit and 64-bit floats.\r\n\r\nThis strategy is ~same as what gcc uses. For complex vectors, we\r\npreviously used 4 instructions and 8 bytes of constant storage, and now\r\nwe use 2 instructions and 16 bytes of constant storage, so the memory\r\nfootprint is the same; for simple vectors we just need 1 instruction (4\r\nbytes).\r\n\r\nclang lowers vector constants a little differently, opting to synthesize\r\na 64-bit integer using 4 instructions (mov/movk) and then move it to the\r\nvector register - this requires 5 instructions and 20 bytes, vs ours/gcc\r\n2 instructions and 8+16=24 bytes. I tried a simpler version of this that\r\nwould be more compact - synthesize a 32-bit integer constant with\r\nmov+movk, and move it to vector register via dup.4s - but this was a\r\nlittle slower on M2, so for now we prefer the slightly larger version as\r\nit's not a regression vs current implementation.\r\n\r\nOn the vector approximation benchmark we get:\r\n\r\n- Before this PR (flag=false): ~7.85 ns/op\r\n- After this PR (flag=true): ~7.74 ns/op\r\n- After this PR, with 0.125 instead of 0.123 in the benchmark code (to\r\nuse fmov): ~7.52 ns/op\r\n- Not part of this PR, but the mov/dup strategy described above: ~8.00\r\nns/op",
"timestamp": "2024-03-13T12:56:11-07:00",
"tree_id": "b46afdd603a2f3bd60b9cac918c2ddc0faf0d668",
"url": "https://github.com/luau-lang/luau/commit/9aa82c6fb90e1dcd6e7f60626255d597ef0fdea1"
},
"date": 1710360181780,
"tool": "benchmarkluau",
"benches": [
{
"name": "map-nonstrict",
"value": 4.78128,
"unit": "4ms",
"range": "±0%",
"extra": "luau-analyze"
},
{
"name": "map-strict",
"value": 5.84051,
"unit": "5ms",
"range": "±0%",
"extra": "luau-analyze"
},
{
"name": "map-dcr",
"value": 51.0637,
"unit": "ms",
"range": "±0%",
"extra": "luau-analyze"
},
{
"name": "regex-nonstrict",
"value": 7.7506,
"unit": "7ms",
"range": "±0%",
"extra": "luau-analyze"
},
{
"name": "regex-strict",
"value": 9.96327,
"unit": "9ms",
"range": "±0%",
"extra": "luau-analyze"
},
{
"name": "regex-dcr",
"value": 115.89,
"unit": "ms",
"range": "±0%",
"extra": "luau-analyze"
}
]
}
]
}
Expand Down
250 changes: 249 additions & 1 deletion bench-codegen.json
View file
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
{
"lastUpdate": 1710160575122,
"lastUpdate": 1710360181462,
"repoUrl": "https://github.com/luau-lang/luau",
"entries": {
"callgrind codegen": [
Expand Down Expand Up @@ -29668,6 +29668,254 @@
"extra": "luau-codegen"
}
]
},
{
"commit": {
"author": {
"email": "[email protected]",
"name": "Arseny Kapoulkine",
"username": "zeux"
},
"committer": {
"email": "[email protected]",
"name": "GitHub",
"username": "web-flow"
},
"distinct": true,
"id": "9aa82c6fb90e1dcd6e7f60626255d597ef0fdea1",
"message": "CodeGen: Improve lowering of NUM_TO_VEC on A64 for constants (#1194)\n\nWhen the input is a constant, we use a fairly inefficient sequence of\r\nfmov+fcvt+dup or, when the double isn't encodable in fmov,\r\nadr+ldr+fcvt+dup.\r\n\r\nInstead, we can use the same lowering as X64 when the input is a\r\nconstant, and load the vector from memory. However, if the constant is\r\nencodable via fmov, we can use a vector fmov instead (which is just one\r\ninstruction and doesn't need constant space).\r\n\r\nFortunately the bit encoding of fmov for 32-bit floating point numbers\r\nmatches that of 64-bit: the decoding algorithm is a little different\r\nbecause it expands into a larger exponent, but the values are\r\ncompatible, so if a double can be encoded into a scalar fmov with a\r\ngiven abcdefgh pattern, the same pattern should encode the same float;\r\ndue to the very limited number of mantissa and exponent bits, all values\r\nthat are encodable are also exact in both 32-bit and 64-bit floats.\r\n\r\nThis strategy is ~same as what gcc uses. For complex vectors, we\r\npreviously used 4 instructions and 8 bytes of constant storage, and now\r\nwe use 2 instructions and 16 bytes of constant storage, so the memory\r\nfootprint is the same; for simple vectors we just need 1 instruction (4\r\nbytes).\r\n\r\nclang lowers vector constants a little differently, opting to synthesize\r\na 64-bit integer using 4 instructions (mov/movk) and then move it to the\r\nvector register - this requires 5 instructions and 20 bytes, vs ours/gcc\r\n2 instructions and 8+16=24 bytes. I tried a simpler version of this that\r\nwould be more compact - synthesize a 32-bit integer constant with\r\nmov+movk, and move it to vector register via dup.4s - but this was a\r\nlittle slower on M2, so for now we prefer the slightly larger version as\r\nit's not a regression vs current implementation.\r\n\r\nOn the vector approximation benchmark we get:\r\n\r\n- Before this PR (flag=false): ~7.85 ns/op\r\n- After this PR (flag=true): ~7.74 ns/op\r\n- After this PR, with 0.125 instead of 0.123 in the benchmark code (to\r\nuse fmov): ~7.52 ns/op\r\n- Not part of this PR, but the mov/dup strategy described above: ~8.00\r\nns/op",
"timestamp": "2024-03-13T12:56:11-07:00",
"tree_id": "b46afdd603a2f3bd60b9cac918c2ddc0faf0d668",
"url": "https://github.com/luau-lang/luau/commit/9aa82c6fb90e1dcd6e7f60626255d597ef0fdea1"
},
"date": 1710360181456,
"tool": "benchmarkluau",
"benches": [
{
"name": "base64",
"value": 13.385,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "chess",
"value": 52.018,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "life",
"value": 23.356,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "matrixmult",
"value": 9.336,
"unit": "9ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "mesh-normal-scalar",
"value": 13,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "pcmmix",
"value": 1.38,
"unit": "1ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "qsort",
"value": 41.508,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "sha256",
"value": 4.525,
"unit": "4ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "ack",
"value": 40.021,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "binary-trees",
"value": 20.853,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "fannkuchen-redux",
"value": 3.878,
"unit": "3ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "fixpoint-fact",
"value": 49.032,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "heapsort",
"value": 7.701,
"unit": "7ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "mandel",
"value": 40.471,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "n-body",
"value": 9.707,
"unit": "9ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "qt",
"value": 24.955,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "queen",
"value": 0.805,
"unit": "0ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "scimark",
"value": 24.643,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "spectral-norm",
"value": 2.444,
"unit": "2ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "sieve",
"value": 82.952,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "3d-cube",
"value": 3.736,
"unit": "3ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "3d-morph",
"value": 3.744,
"unit": "3ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "3d-raytrace",
"value": 3.304,
"unit": "3ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "controlflow-recursive",
"value": 3.463,
"unit": "3ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "crypto-aes",
"value": 7.228,
"unit": "7ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "fannkuch",
"value": 6.068,
"unit": "6ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "math-cordic",
"value": 3.768,
"unit": "3ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "math-partial-sums",
"value": 1.872,
"unit": "1ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "n-body-oop",
"value": 13.714,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "tictactoe",
"value": 62.961,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "trig",
"value": 6.618,
"unit": "6ms",
"range": "±0.000%",
"extra": "luau-codegen"
},
{
"name": "voxelgen",
"value": 27.559,
"unit": "ms",
"range": "±0.000%",
"extra": "luau-codegen"
}
]
}
]
}
Expand Down
Loading

0 comments on commit 07a1e92

Please sign in to comment.