diff --git a/npu/audio_exec.cpp b/npu/audio_exec.cpp index 0bf21e223..a98df279c 100644 --- a/npu/audio_exec.cpp +++ b/npu/audio_exec.cpp @@ -263,6 +263,59 @@ void load_init_i64(const OnnxGraphInfo& g, const OnnxInit& t, std::vector t, con float* r0 = r.data.data() + bi * size_t(M * N) + size_t(i0) * size_t(N); if (i1 - i0 == 4) { const float* a0 = ap + size_t(i0) * size_t(K); - for (int64_t k = 0; k < K; ++k) { - const float v0 = a0[size_t(k)], v1 = a0[size_t(K) + size_t(k)]; - const float v2 = a0[size_t(2 * K) + size_t(k)], v3 = a0[size_t(3 * K) + size_t(k)]; - if ((v0 == 0.0f) && (v1 == 0.0f) && (v2 == 0.0f) && (v3 == 0.0f)) continue; - const float* br = bpp + size_t(k) * size_t(N); - for (int64_t j = 0; j < N; ++j) { - const float bv = br[j]; - r0[j] += v0 * bv; - r0[N + j] += v1 * bv; - r0[2 * N + j] += v2 * bv; - r0[3 * N + j] += v3 * bv; - } - } + host_gemm_4rows()(a0, bpp, r0, K, N); // engine#1783: baseline or AVX2+FMA, picked once } else { for (int64_t i = i0; i < i1; ++i) { float* rr = r0 + size_t(i - i0) * size_t(N);