Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

perf(dsv4): compute four output rows per pass in coli_fp4_matvec_rows16_order (#1314)

Brianf59199e2bb77
diff --git a/c/deepseek_v4.c b/c/deepseek_v4.c
index 0b548d24..944922ee 100644
--- a/c/deepseek_v4.c
+++ b/c/deepseek_v4.c
@@ -16056,19 +16056,52 @@ void coli_fp4_matvec_rows16_order(float *y, const uint8_t *q4,
}
#else
#pragma omp parallel for schedule(static)
- for (int o = 0; o < O; o++) {
- const uint8_t *w = q4 + (int64_t)o * rb;
- const uint8_t *scl = e8s + (int64_t)o * ng;
- float sum = 0.0f;
+ for (int o = 0; o < O; o += 4) {
+ int o1 = o + 1 < O ? o + 1 : o;
+ int o2 = o + 2 < O ? o + 2 : o;
+ int o3 = o + 3 < O ? o + 3 : o;
+ const uint8_t *w0 = q4 + (int64_t)o * rb;
+ const uint8_t *w1 = q4 + (int64_t)o1 * rb;
+ const uint8_t *w2 = q4 + (int64_t)o2 * rb;
+ const uint8_t *w3 = q4 + (int64_t)o3 * rb;
+ const uint8_t *scl0 = e8s + (int64_t)o * ng;
+ const uint8_t *scl1 = e8s + (int64_t)o1 * ng;
+ const uint8_t *scl2 = e8s + (int64_t)o2 * ng;
+ const uint8_t *scl3 = e8s + (int64_t)o3 * ng;
+ float sum0 = 0.0f, sum1 = 0.0f, sum2 = 0.0f, sum3 = 0.0f;
+ /* Interleaving independent rows changes no row's c=0..I-1 direct
+ * fold: each accumulator receives the same rounded term sequence. */
for (int c = 0; c < I; c++) {
- uint8_t byte = w[c >> 1];
- float wv = coli_e2m1_decode((c & 1)
- ? (uint8_t)(byte >> 4) : (uint8_t)(byte & 0xF));
- float t = x[c] * wv;
- t = t * e8lut[scl[c / 32]];
- sum = sum + t;
+ float xc = x[c];
+ uint8_t byte0 = w0[c >> 1];
+ uint8_t byte1 = w1[c >> 1];
+ uint8_t byte2 = w2[c >> 1];
+ uint8_t byte3 = w3[c >> 1];
+ float wv0 = coli_e2m1_decode((c & 1)
+ ? (uint8_t)(byte0 >> 4) : (uint8_t)(byte0 & 0xF));
+ float wv1 = coli_e2m1_decode((c & 1)
+ ? (uint8_t)(byte1 >> 4) : (uint8_t)(byte1 & 0xF));
+ float wv2 = coli_e2m1_decode((c & 1)
+ ? (uint8_t)(byte2 >> 4) : (uint8_t)(byte2 & 0xF));
+ float wv3 = coli_e2m1_decode((c & 1)
+ ? (uint8_t)(byte3 >> 4) : (uint8_t)(byte3 & 0xF));
+ float t0 = xc * wv0;
+ float t1 = xc * wv1;
+ float t2 = xc * wv2;
+ float t3 = xc * wv3;
+ t0 = t0 * e8lut[scl0[c / 32]];
+ t1 = t1 * e8lut[scl1[c / 32]];
+ t2 = t2 * e8lut[scl2[c / 32]];
+ t3 = t3 * e8lut[scl3[c / 32]];
+ sum0 = sum0 + t0;
+ sum1 = sum1 + t1;
+ sum2 = sum2 + t2;
+ sum3 = sum3 + t3;
}
- y[o] = sum;
+ y[o] = sum0;
+ if (o1 != o) y[o1] = sum1;
+ if (o2 != o) y[o2] = sum2;
+ if (o3 != o) y[o3] = sum3;
}
#endif
}