Dash — CPU Acceleration Gate Layer

EXECUTE — CLIENT REFERENCE ENGINE

The browser executes the SAME op contracts as POST /execute (scalar JS reference — real math, real timing, deterministic seeded tensors). The daemon dispatches the identical op through its AVX2/AVX512/NEON kernel and returns exec_path + time_ms. Path detection is daemon-side (is_x86_feature_detected!) — shown verbatim below.

LAST OP

—

TIME_MS (REAL)

—

THROUGHPUT

—

RESULT Σ

—

TENSOR OUT

—

// executions appear here — request JSON (verbatim POST /execute contract) + real results

QUANTIZATION — Q8_0 BLOCK FORMAT (port of quantize.rs)

f32 → Q8_0: blocks of 32, per-block scale d = amax/127, qs[j] = round(v·d⁻¹) as i8. Dequantize: v' = qs[j]·d. Real quantization error computed on the live tensor.

run QUANTIZE Q8_0 to quantize the live tensor

PATH SELECTION — simd.rs (verbatim detection cascade)

pub fn detect_hardware() -> SimdPath {
    if cfg!(target_arch = "x86_64") {
        if is_x86_feature_detected!("avx512f") { return SimdPath::Avx512; }
        if is_x86_feature_detected!("avx2")   { return SimdPath::Avx2;   }
    }
    if cfg!(target_arch = "aarch64") {
        if is_aarch64_feature_detected!("neon") { return SimdPath::Neon; }
    }
    SimdPath::Scalar
}

PATH	F32 LANES	TARGET	NOTE
AVX512	16	x86_64 server (PowerEdge class)	detected first — widest
AVX2	8	x86_64 desktop/server	primary path
NEON	4	ARM64 / Apple Silicon	embedded + mobile
SCALAR	1	any hardware	always works, never fastest

POST /execute — API CONTRACT (server/mod.rs verbatim)

ExecuteRequest {
    op: String,                      // "matmul" | "relu" | "silu" | "gelu" | "softmax" | "layer_norm" | "conv2d" | "embedding_lookup"
    a, b, t, input, weight: Option<Tensor>,  // Tensor { data: Vec<u8>, shape: Vec<usize>, dtype }
    axis, stride, padding: Option<usize>,
    indices: Option<Vec<usize>>, weights: Option<Tensor>,
} → ExecuteResponse { result: Tensor, exec_path: String, time_ms: f64 }

WHAT DASH IS NOT (README verbatim)

Not an LLM runtime (that is Reach) · Not a model manager (that is Reach) · Not a cloud service · Not a wrapper around Python ML libraries · Not dependent on CUDA, ROCm, or any GPU driver