diff --git a/include/algos.hpp b/include/algos.hpp
index d6c9f0c..ba3b8d0 100644
--- a/include/algos.hpp
+++ b/include/algos.hpp
@@ -24,7 +24,7 @@ void bgr2gray_opencv(const cv::Mat& src, cv::Mat& dst);
 void boxFilter_halide(uint16_t* src, uint16_t* dst, int height, int width);
 void boxFilter_opencv(const cv::Mat& src, cv::Mat& dst);
 void ascii_art_ref(const uint8_t* src, uint8_t* dst, int height, int width);
-void ascii_art_halide(uint8_t* src, uint8_t* dst, int input_height, int input_width);
+void ascii_art_halide(uint8_t* src, float* dst, int input_height, int input_width);
 
 void julia_ref(uint8_t* dst, int height, int width);
 void halide_julia(uint8_t* dst, int height, int width);
diff --git a/perf/perf_ascii.cpp b/perf/perf_ascii.cpp
index a614801..4647944 100644
--- a/perf/perf_ascii.cpp
+++ b/perf/perf_ascii.cpp
@@ -28,10 +28,10 @@ PERF_TEST(ascii_art, halide) {
     const int height = 1080;
     const std::string grey_scale = "$@B%8&WM#*OAHKDPQWMRZO0QLCJUYXVJFT/|()1{}[]?-_+~<>i!lI;:,^`'.  ";
 
-    Mat src(height, width, CV_8UC1), dst(height/ry, width/rx, CV_8U);
+    Mat src(height, width, CV_8UC1), dst(height/ry, width/rx, CV_32F);
     randu(src, 0, 256);
     PERF_SAMPLE_BEGIN()
-        ascii_art_halide(src.ptr<uint8_t>(), dst.ptr<uint8_t>(), src.rows, src.cols);
+        ascii_art_halide(src.ptr<uint8_t>(), dst.ptr<float>(), src.rows, src.cols);
     PERF_SAMPLE_END()
 
     SANITY_CHECK_NOTHING();
diff --git a/src/ASCII-ART.cpp b/src/ASCII-ART.cpp
index ad090f3..6e7470c 100644
--- a/src/ASCII-ART.cpp
+++ b/src/ASCII-ART.cpp
@@ -22,13 +22,13 @@ static const int norm_h = 3;
 static const int rx = 15;
 static const int ry = 19;
 
-void ascii_art_halide(uint8_t* src, uint8_t* dst, int input_height, int input_width) {
+void ascii_art_halide(uint8_t* src, float* dst, int input_height, int input_width) {
     int output_width = input_width / rx;
     int output_height = input_height / ry;
- 
-    
+
+
     Buffer<uint8_t> input(src, {input_width, input_height});
-    Buffer<uint8_t> output(dst, {input_width / rx, input_height / ry});
+    Buffer<float> output(dst, {input_width / rx, input_height / ry});
 #ifdef __riscv
     ascii_art(input, output);
 #else
@@ -38,17 +38,24 @@ void ascii_art_halide(uint8_t* src, uint8_t* dst, int input_height, int input_wi
 
         Var x("x"), y("y");
         RDom r(0, rx, 0, ry);
-        Expr s = sum(cast<uint32_t>(input(x*rx + r.x, y*ry + r.y)));
-    
+
+        Func casted;
+        casted(x, y) = cast<float>(input(x, y));
+
+        Expr s = sum(casted(x*rx + r.x, y*ry + r.y));
+
         //s = Halide::clamp(s/(rx*ry),0,255);
-        ascii(x, y) = cast<uint8_t>(s/(rx*ry));
-        // ascii.realize(output);
+        ascii(x, y) = s/(rx*ry);
+
+        casted.compute_root();
+        ascii.vectorize(x, 4);
+
         // Compile
         Target target;
         target.os = Target::OS::Linux;
         target.arch = Target::Arch::RISCV;
         target.bits = 64;
-        //target.vector_bits = factor * sizeof(uint8_t) * 8;
+        target.vector_bits = 128;
 
         // Tested XuanTie C906 has 128-bit vector unit
         CV_Assert(target.vector_bits <= 128);
@@ -76,11 +83,11 @@ void ascii_art_ref(const uint8_t* src, uint8_t* dst, int input_height, int input
             float lum = 0;
             for(int j = 0; j < ry; ++j){
                 for(int i = 0; i < rx; ++i){
-                    lum += src[ (y * ry + j) * input_width + x * rx +i];   
-                }   
+                    lum += src[ (y * ry + j) * input_width + x * rx +i];
+                }
             }
             dst[y * output_width + x] = static_cast<uint8_t>(lum/(rx*ry));
             lum = 0;
-        }   
+        }
     }
 }
diff --git a/test/test_ascii.cpp b/test/test_ascii.cpp
index 6af04ef..b29a16e 100644
--- a/test/test_ascii.cpp
+++ b/test/test_ascii.cpp
@@ -10,12 +10,10 @@ TEST(ascii_art, halide){
     const std::string grey_scale = "$@B%8&WM#*OAHKDPQWMRZO0QLCJUYXVJFT/|()1{}[]?-_+~<>i!lI;:,^`'.  ";
 
     cv::Mat src = imread("cat.jpeg", cv::IMREAD_GRAYSCALE);
-    cv::Mat dst(src.rows/ry, src.cols/rx, CV_8U),
+    cv::Mat dst(src.rows/ry, src.cols/rx, CV_32F),
     render(src.rows, src.cols, CV_8U, cv::Scalar(255));
 
-    ascii_art_halide(src.ptr<uint8_t>(), dst.ptr<uint8_t>(), src.rows, src.cols);
-    char *s;
-    s = (char*)dst.ptr<uint8_t>();
+    ascii_art_halide(src.ptr<uint8_t>(), dst.ptr<float>(), src.rows, src.cols);
     std::vector<std::pair<float, char>> lums={};
     std::vector<char> symbols(256);
     float lum_min = 255;
@@ -42,14 +40,16 @@ TEST(ascii_art, halide){
     for(int i = 0; i < dst.rows; i++)
         for(int j = 0; j < dst.cols; j++){
 
-            uint8_t lum = dst.at<uint8_t>(i, j);
-            int index = (static_cast<float>(lum)/255)*(lums.size()-1);
+            float lum = dst.at<float>(i, j);
+            int index = (min(lum, 255.0f)/255.0f)*(lums.size()-1);
+            std::cout << lum << std::endl;
+            CV_Assert(index < lums.size());
             cv::Mat roi =  render.colRange(j*rx, (j+1)*rx).rowRange(i*ry, (i+1)*ry);
             cv::putText(roi, std::string(1, grey_scale[index]), cv::Point(1, ry-1),
                 cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0), 1, cv::LINE_AA);
         }
 
-    imwrite("res_cat_ascii_halide.png", render);
-    imwrite("src_cat_ascii_halide.png", src);
-    ASSERT_EQ(true, true);
+    // imwrite("res_cat_ascii_halide.png", render);
+    // imwrite("src_cat_ascii_halide.png", src);
+    // ASSERT_EQ(true, true);
 }