SIMD-accelerated numeric operations for Go, optimized for ARM64 (NEON and SVE).
Hand-tuned ARM64 assembly for NEON and SVE instruction setsAutomatic CPU detection - detects Graviton3/4, uses optimal thresholdsThreshold-based dispatch - uses scalar for small arrays, SIMD for largeScalar fallbacks for non-ARM64 platformsZero allocations in hot pathsFuzz tested for correctness
go get github.com/axiomhq/simd-go package main
import (
"fmt"
simd "github.com/axiomhq/simd-go"
)
func main () {
vals := []float64 {1.0 , 2.0 , 3.0 , 4.0 , 5.0 }
sum := simd .SumFloat64 (vals )
min := simd .MinFloat64 (vals )
max := simd .MaxFloat64 (vals )
fmt .Printf ("Sum: %v, Min: %v, Max: %v\n " , sum , min , max )
// Output: Sum: 15, Min: 1, Max: 5
}
Function
Description
SumFloat64(vals []float64) float64Sum of all values
MinFloat64(vals []float64) float64Minimum value
MaxFloat64(vals []float64) float64Maximum value
DotProductFloat64(a, b []float64) float64Dot product of two vectors
Function
Description
SumFloat32(vals []float32) float32Sum of all values
MinFloat32(vals []float32) float32Minimum value
MaxFloat32(vals []float32) float32Maximum value
DotProductFloat32(a, b []float32) float32Dot product of two vectors
Function
Description
SumInt64(vals []int64) int64Sum of all values
MinInt64(vals []int64) int64Minimum value
MaxInt64(vals []int64) int64Maximum value
DotProductInt64(a, b []int64) int64Dot product of two vectors
SumSqInt64(vals []int64) int64Sum of squares (Σv²)
AnyAbsGreaterThan(vals []int64, threshold int64) boolCheck if any |v| > threshold
Function
Description
SumInt32(vals []int32) int64Sum of all values (returns int64 to avoid overflow)
MinInt32(vals []int32) int32Minimum value
MaxInt32(vals []int32) int32Maximum value
DotProductInt32(a, b []int32) int64Dot product of two vectors (returns int64 to avoid overflow)
SumSqInt32(vals []int32) int64Sum of squares (returns int64 to avoid overflow)
AnyAbsGreaterThanInt32(vals []int32, threshold int32) boolCheck if any |v| > threshold
Function
Description
SumInt16(vals []int16) int64Sum of all values (returns int64 to avoid overflow)
MinInt16(vals []int16) int16Minimum value
MaxInt16(vals []int16) int16Maximum value
DotProductInt16(a, b []int16) int64Dot product of two vectors (returns int64 to avoid overflow)
SumSqInt16(vals []int16) int64Sum of squares (returns int64 to avoid overflow)
AnyAbsGreaterThanInt16(vals []int16, threshold int16) boolCheck if any |v| > threshold
Function
Description
HasSVE() boolReturns true if CPU supports SVE
HasNEON() boolReturns true if CPU supports NEON
IsARM64() boolReturns true if running on ARM64
CPUName() stringReturns detected CPU name (e.g., "AWS Graviton4 (Neoverse-V2)")
All benchmarks run with n=10,000 elements.
Type
Operation
Scalar
NEON
Speedup
Float64
Sum
2.83µs (28.3 GB/s)
551ns (145.3 GB/s)
5.1x
Float64
Min
11.8µs (6.8 GB/s)
556ns (143.9 GB/s)
21.2x
Float64
Max
11.9µs (6.7 GB/s)
561ns (142.7 GB/s)
21.1x
Float64
DotProduct
2.96µs (54.1 GB/s)
1.04µs (154.4 GB/s)
2.9x
Float32
Sum
3.02µs (13.3 GB/s)
289ns (138.5 GB/s)
10.4x
Float32
Min
11.7µs (3.4 GB/s)
285ns (140.2 GB/s)
41.1x
Float32
Max
11.8µs (3.4 GB/s)
289ns (138.4 GB/s)
40.8x
Float32
DotProduct
3.01µs (26.6 GB/s)
520ns (153.9 GB/s)
5.8x
Int64
Sum
2.97µs (26.9 GB/s)
567ns (141.0 GB/s)
5.2x
Int64
Min
3.00µs (26.6 GB/s)
806ns (99.3 GB/s)
3.7x
Int64
Max
2.97µs (26.9 GB/s)
823ns (97.2 GB/s)
3.6x
Int64
SumSq
2.98µs (26.9 GB/s)
1.55µs (51.6 GB/s)
1.9x
Int32
Sum
2.96µs (13.5 GB/s)
376ns (106.3 GB/s)
7.9x
Int32
Min
3.00µs (13.3 GB/s)
292ns (137.1 GB/s)
10.3x
Int32
Max
3.03µs (13.2 GB/s)
285ns (140.3 GB/s)
10.6x
Int32
DotProduct
2.97µs (27.0 GB/s)
748ns (107.0 GB/s)
4.0x
Int32
SumSq
2.96µs (13.5 GB/s)
743ns (53.8 GB/s)
4.0x
Int16
Sum
2.97µs (6.7 GB/s)
284ns (70.5 GB/s)
10.5x
Int16
Min
2.97µs (6.7 GB/s)
148ns (134.9 GB/s)
20.0x
Int16
Max
2.98µs (6.7 GB/s)
149ns (134.2 GB/s)
20.0x
Int16
DotProduct
2.98µs (13.4 GB/s)
560ns (71.4 GB/s)
5.3x
Int16
SumSq
2.96µs (6.8 GB/s)
560ns (35.7 GB/s)
5.3x
AWS Graviton3 (Neoverse-V1, SVE 256-bit)
Type
Operation
Scalar
NEON
SVE
Best
Float64
Sum
3.87µs (20.7 GB/s)
1.37µs (58.2 GB/s)
1.26µs (63.7 GB/s)
3.1x SVE
Float64
Min
15.4µs (5.2 GB/s)
1.56µs (51.3 GB/s)
1.61µs (49.7 GB/s)
9.9x NEON
Float64
Max
15.3µs (5.2 GB/s)
1.56µs (51.3 GB/s)
1.61µs (49.7 GB/s)
9.8x NEON
Float64
DotProduct
3.87µs (41.4 GB/s)
1.96µs (81.5 GB/s)
1.76µs (91.0 GB/s)
2.2x SVE
Float32
Sum
3.87µs (10.3 GB/s)
550ns (72.7 GB/s)
391ns (102.3 GB/s)
9.9x SVE
Float32
Min
15.4µs (2.6 GB/s)
570ns (70.2 GB/s)
574ns (69.7 GB/s)
27.0x NEON
Float32
Max
15.3µs (2.6 GB/s)
586ns (68.5 GB/s)
574ns (69.7 GB/s)
26.7x SVE
Float32
DotProduct
3.87µs (20.7 GB/s)
780ns (102.6 GB/s)
702ns (114.0 GB/s)
5.5x SVE
Int64
Sum
3.87µs (20.7 GB/s)
1.38µs (58.1 GB/s)
1.24µs (64.3 GB/s)
3.1x SVE
Int64
Min
3.91µs (20.5 GB/s)
1.99µs (40.2 GB/s)
1.61µs (49.6 GB/s)
2.4x SVE
Int64
Max
3.86µs (20.7 GB/s)
2.00µs (40.1 GB/s)
1.61µs (49.7 GB/s)
2.4x SVE
Int64
DotProduct
3.87µs (41.4 GB/s)
—
2.34µs (68.2 GB/s)
1.6x SVE
Int64
SumSq
3.87µs (20.7 GB/s)
3.18µs (25.2 GB/s)
2.19µs (36.6 GB/s)
1.8x SVE
Int32
Sum
3.86µs (10.4 GB/s)
765ns (52.3 GB/s)
1.27µs (31.5 GB/s)
5.1x NEON
Int32
Min
3.87µs (10.3 GB/s)
564ns (71.0 GB/s)
392ns (102.1 GB/s)
9.9x SVE
Int32
Max
3.91µs (10.2 GB/s)
564ns (71.0 GB/s)
393ns (101.9 GB/s)
9.9x SVE
Int32
DotProduct
3.87µs (20.7 GB/s)
1.76µs (45.4 GB/s)
2.76µs (29.0 GB/s)
2.2x NEON
Int32
SumSq
3.86µs (10.4 GB/s)
1.43µs (28.0 GB/s)
2.11µs (19.0 GB/s)
2.7x NEON
Int16
Sum
3.86µs (5.2 GB/s)
542ns (36.9 GB/s)
1.51µs (13.2 GB/s)
7.1x NEON
Int16
Min
3.87µs (5.2 GB/s)
287ns (69.8 GB/s)
201ns (99.3 GB/s)
19.2x SVE
Int16
Max
3.86µs (5.2 GB/s)
287ns (69.8 GB/s)
201ns (99.3 GB/s)
19.2x SVE
Int16
DotProduct
3.87µs (10.3 GB/s)
1.25µs (31.9 GB/s)
2.12µs (18.9 GB/s)
3.1x NEON
Int16
SumSq
3.86µs (5.2 GB/s)
1.12µs (17.8 GB/s)
1.87µs (10.7 GB/s)
3.4x NEON
AWS Graviton4 (Neoverse-V2, SVE2 128-bit)
Type
Operation
Scalar
NEON
SVE
SVE2
Best
Float64
Sum
3.59µs (22.3 GB/s)
1.00µs (79.7 GB/s)
1.00µs (79.6 GB/s)
—
3.6x NEON
Float64
Min
14.3µs (5.6 GB/s)
1.22µs (65.7 GB/s)
1.23µs (65.1 GB/s)
—
11.7x NEON
Float64
Max
14.3µs (5.6 GB/s)
1.22µs (65.6 GB/s)
1.23µs (65.0 GB/s)
—
11.7x NEON
Float64
DotProduct
4.06µs (40.0 GB/s)
1.71µs (93.6 GB/s)
1.66µs (96.6 GB/s)
—
2.5x SVE
Float32
Sum
3.90µs (10.4 GB/s)
461ns (86.8 GB/s)
444ns (90.1 GB/s)
—
8.8x SVE
Float32
Min
14.3µs (2.8 GB/s)
563ns (71.1 GB/s)
565ns (70.8 GB/s)
—
25.4x NEON
Float32
Max
14.2µs (2.8 GB/s)
563ns (71.1 GB/s)
565ns (70.8 GB/s)
—
25.3x NEON
Float32
DotProduct
3.59µs (22.3 GB/s)
759ns (105.4 GB/s)
770ns (103.9 GB/s)
—
4.7x NEON
Int64
Sum
4.06µs (20.0 GB/s)
1.00µs (79.9 GB/s)
1.00µs (79.8 GB/s)
—
4.1x NEON
Int64
Min
3.90µs (20.8 GB/s)
1.58µs (50.7 GB/s)
1.23µs (65.2 GB/s)
—
3.2x SVE
Int64
Max
3.58µs (22.3 GB/s)
1.51µs (53.2 GB/s)
1.23µs (65.1 GB/s)
—
2.9x SVE
Int64
DotProduct
3.59µs (44.6 GB/s)
—
1.96µs (81.6 GB/s)
—
1.8x SVE
Int64
SumSq
3.59µs (22.3 GB/s)
1.97µs (40.6 GB/s)
1.80µs (44.4 GB/s)
—
2.0x SVE
Int32
Sum
3.59µs (11.2 GB/s)
612ns (65.4 GB/s)
1.02µs (39.2 GB/s)
657ns (60.9 GB/s)
5.9x NEON
Int32
Min
4.19µs (9.7 GB/s)
558ns (71.7 GB/s)
444ns (90.2 GB/s)
—
9.5x SVE
Int32
Max
3.65µs (11.0 GB/s)
558ns (71.7 GB/s)
445ns (89.8 GB/s)
—
8.2x SVE
Int32
DotProduct
3.59µs (22.3 GB/s)
1.38µs (57.9 GB/s)
2.17µs (36.9 GB/s)
1.35µs (59.5 GB/s)
2.7x SVE2
Int32
SumSq
3.59µs (11.2 GB/s)
1.12µs (35.6 GB/s)
1.90µs (21.1 GB/s)
1.12µs (35.6 GB/s)
3.2x NEON
Int16
Sum
3.78µs (5.3 GB/s)
408ns (49.1 GB/s)
1.26µs (15.9 GB/s)
339ns (59.0 GB/s)
11.1x SVE2
Int16
Min
3.90µs (5.2 GB/s)
344ns (58.2 GB/s)
224ns (89.2 GB/s)
—
17.4x SVE
Int16
Max
3.90µs (5.2 GB/s)
342ns (58.5 GB/s)
224ns (89.2 GB/s)
—
17.4x SVE
Int16
DotProduct
3.90µs (10.4 GB/s)
965ns (41.5 GB/s)
2.01µs (19.9 GB/s)
746ns (53.7 GB/s)
5.2x SVE2
Int16
SumSq
3.59µs (5.6 GB/s)
896ns (22.3 GB/s)
1.75µs (11.4 GB/s)
587ns (34.1 GB/s)
6.1x SVE2
DotProductInt64 has no NEON implementation (NEON lacks 64-bit integer multiply)
Graviton4 uses 128-bit SVE vectors; SVE2 provides additional instructions for better Int16/Int32 performance
Apple M3 has no SVE support
go test -bench=. -benchmem -count=5 ./... | tee bench.txt
benchstat bench.txt Use benchstat for comparison:
benchstat -row /fn,/n -col /impl bench.txt ┌─────────────────────────────────────────────────────────┐
│ Public API (simd.go) │
│ SumFloat64, MinFloat64, MaxFloat64, DotProductFloat64 │
│ SumInt64, MinInt64, MaxInt64, DotProductInt64, ... │
└─────────────────────────────┬───────────────────────────┘
│
┌───────────────┴───────────────┐
│ │
┌─────────▼─────────┐ ┌──────────▼──────────┐
│ impl_arm64.go │ │ impl_stub.go │
│ (ARM64 dispatch) │ │ (other platforms) │
│ │ │ │
│ 1. Detect CPU │ │ → scalar.go │
│ (G3/G4/other) │ │ │
│ 2. Check size │ └─────────────────────┘
│ vs threshold │
│ 3. Dispatch to: │
│ scalar/NEON/ │
│ SVE │
└─────────┬─────────┘
│
┌─────────┼─────────┐
│ │ │
┌───▼───┐ ┌───▼───┐ ┌───▼───┐
│Scalar │ │ NEON │ │ SVE │
│ (.go) │ │ (.s) │ │ (.s) │
└───────┘ └───────┘ └───────┘
SumSqInt64 can overflow if input values are too large. Use AnyAbsGreaterThan to check:
const sqrtMaxInt64 = 3037000499
if simd .AnyAbsGreaterThan (vals , sqrtMaxInt64 ) {
// Handle potential overflow - use float64 or arbitrary precision
}
sum := simd .SumSqInt64 (vals )MIT
Contributions are welcome! Please feel free to submit a Pull Request.