renderer: improve perf for clear (avx2) and rect

1 files changed, 46 insertions, 17 deletions

diff --git a/src/renderer/RSoftwareBackend.cpp b/src/renderer/RSoftwareBackend.cpp
index 2606553..754a30f 100644
--- a/src/renderer/RSoftwareBackend.cpp
+++ b/src/renderer/RSoftwareBackend.cpp
@@ -8,6 +8,7 @@
 #include <algorithm>
 #include <cstdlib>
 #include <cstdio>
+#include <immintrin.h>
 
 
 RSoftwareBackend::RSoftwareBackend(Renderer& renderer)
@@ -17,6 +18,8 @@ RSoftwareBackend::RSoftwareBackend(Renderer& renderer)
     m_canvas.gshift = 8;
     m_canvas.bshift = 16;
     m_canvas.ashift = 24;
+    m_canvas.w = 0;
+    m_canvas.h = 0;
     m_canvas.pixels = nullptr;
 
 
@@ -77,16 +80,30 @@ RSoftwareBackend::Draw()
 void
 RSoftwareBackend::Resize(int32_t w, int32_t h)
 {
-    size_t realloc_size = (size_t)(w * h) * sizeof(m_canvas.pixels[0]);
-    void *realloc_data = realloc(m_canvas.pixels, realloc_size);
-    if (!realloc_data) {
-        printf("could not resize offscreen buffer\n");
+    if ((m_canvas.w == w && m_canvas.h == h)) {
         return;
     }
 
+    size_t alignment = 32;
+    size_t new_size = static_cast<size_t>(w) * static_cast<size_t>(h) * sizeof(m_canvas.pixels[0]);
+    if (new_size == 0) {
+        _mm_free(m_canvas.pixels);
+        m_canvas.pixels = nullptr;
+        m_canvas.w = 0;
+        m_canvas.h = 0;
+        return;
+    }
+
+    uint32_t* new_pixels = (uint32_t*)_mm_malloc(new_size, alignment);
+    if (!new_pixels) {
+        printf("_mm_malloc failed for resizing canvas\n");
+        return;
+    }
+
+    _mm_free(m_canvas.pixels);
     m_canvas.w = w;
     m_canvas.h = h;
-    m_canvas.pixels = (uint32_t*)realloc_data;
+    m_canvas.pixels = new_pixels;
 
     glViewport(0, 0, w, h);
 }
@@ -109,14 +126,24 @@ RSoftwareBackend::DrawClear()
     uint32_t gshift = m_canvas.gshift;
     uint32_t bshift = m_canvas.bshift;
 
-    for (int32_t y = 0; y < m_canvas.h; y++) {
-        for (int32_t x = 0; x < m_canvas.w; x++) {
-            uint32_t r = (uint32_t)(color.r * 255.0f);
-            uint32_t g = (uint32_t)(color.g * 255.0f);
-            uint32_t b = (uint32_t)(color.b * 255.0f);
-            uint32_t val = r << rshift | g << gshift | b << bshift;
-            m_canvas.pixels[y*m_canvas.w + x] = val;
-        }
+    uint32_t r = static_cast<uint32_t>(color.r * 255.0f);
+    uint32_t g = static_cast<uint32_t>(color.g * 255.0f);
+    uint32_t b = static_cast<uint32_t>(color.b * 255.0f);
+    uint32_t val = r << rshift | g << gshift | b << bshift;
+
+    size_t pixel_count = size_t(m_canvas.w) * size_t(m_canvas.h);
+    size_t chunk_count = pixel_count / 8;
+    size_t chunk_rest = pixel_count % 8;
+    uint32_t* pixels = m_canvas.pixels;
+
+    __m256i vec_val = _mm256_set1_epi32((int32_t)val);
+    for (size_t i = 0; i < chunk_count; i++) {
+        _mm256_storeu_si256(reinterpret_cast<__m256i*>(pixels), vec_val);
+        pixels += 8;
+    }
+
+    for (size_t i = 0; i < chunk_rest; i++) {
+        *pixels++ = val;
     }
 }
 
@@ -144,13 +171,15 @@ RSoftwareBackend::DrawRectangle(REntity_Rectangle& entity)
     uint32_t rshift = m_canvas.rshift;
     uint32_t gshift = m_canvas.gshift;
     uint32_t bshift = m_canvas.bshift;
+
+    uint32_t r = (uint32_t)(entity.color.r * 255.0f);
+    uint32_t g = (uint32_t)(entity.color.g * 255.0f);
+    uint32_t b = (uint32_t)(entity.color.b * 255.0f);
+    uint32_t val = r << rshift | g << gshift | b << bshift;
+
     for (int32_t y = ymin; y <= ymax; ++y) {
         uint32_t *pixel = m_canvas.pixels + y * m_canvas.w + xmin;
         for (int32_t x = xmin; x <= xmax; ++x) {
-            uint32_t r = (uint32_t)(entity.color.r * 255.0f);
-            uint32_t g = (uint32_t)(entity.color.g * 255.0f);
-            uint32_t b = (uint32_t)(entity.color.b * 255.0f);
-            uint32_t val = r << rshift | g << gshift | b << bshift;
             *pixel++ = val;
         }
     }