single attention增加汇编

.gitignore

@@ -33,3 +33,4 @@ token
 /example/Android/LLMAssistant/local.properties
 /test/cmmlu/results/
 /models/
+/localtest/

src/devices/cpu/cpudevice.cpp

@@ -241,15 +241,44 @@ namespace fastllm {
                     qk[j] = -10000;
                     continue;
                 }
-                float sum = 0.0f;
-                for (int l = 0; l < q2; l++) {
-                    sum += qd[i * q2 + l] * kd[j * q2 + l];
+                float now = 0.0f;
+                int l = 0;
+#ifdef __aarch64__
+                float32x4_t sum = {0, 0, 0, 0};
+                for (; l + 3 < q2; l += 4) {
+                    sum = vaddq_f32(sum, vmulq_f32(vld1q_f32(qd + i * q2 + l),
+                                                   vld1q_f32(kd + j * q2 + l)));
+                }
+                now += sum[0] + sum[1] + sum[2] + sum[3];
+#elif defined(__AVX__)
+                __m256 vsum = _mm256_set1_ps(0.0f);
+                for (; l + 7 < q2; l += 8) {
+                    __m256 vx = _mm256_loadu_ps((const float *) (qd + i * q2 + l));
+                    __m256 vy = _mm256_loadu_ps((const float *) (kd + j * q2 + l));
+                    vsum = _mm256_add_ps(vsum, _mm256_mul_ps(vx, vy));
+                }
+                now += Floatsum(vsum);
+#endif
+                for (; l < q2; l++) {
+                    now += qd[i * q2 + l] * kd[j * q2 + l];
                 }
-                qk[j] = sum * scale;
-                maxValue = std::max(maxValue, sum * scale);
+                qk[j] = now * scale;
+                maxValue = std::max(maxValue, now * scale);
             }
-            for (int j = 0; j < k1; j++) {
+
+            int j = 0;
+#ifdef __aarch64__
+            float32x4_t vmax = vdupq_n_f32(maxValue);
+            for (; j + 3 < k1; j += 4) {
+                vst1q_f32(temp + j, exp_ps(vsubq_f32(vld1q_f32(qk + j), vmax)));
+            }
+#endif
+            for (; j < k1; j++) {
                 temp[j] = expf(qk[j] - maxValue);
+            }
+
+            sum = 0.0f;
+            for (int j = 0; j < k1; j++) {
                 sum += temp[j];
             }
             sum = std::max(sum, 0.1f);