Automatic batch freezing.

Makefile

@@ -21,7 +21,7 @@ cpp/math.cpp
 
 OUTPUTFILE = engine.so
 
-CXXFLAGS = -fpic -Wall -Werror -O2 -flto -fomit-frame-pointer -ffast-math -funroll-loops -fno-rtti -fno-exceptions
+CXXFLAGS = -Wall -Werror -O2 -march=x86-64 -mtune=generic -msse2 -ffast-math -fpic -flto -fno-rtti -fno-exceptions
 
 .PHONY: all
 all: clean $(OUTPUTFILE)

cpp/render/render_opengles.cpp

@@ -837,6 +837,32 @@ static void rk_pack_batch(
     batch.state = RK_BATCH_STATE_PACKED;
 }
 
+static bool rk_compare_replace(
+    void * __restrict _dst,
+    void const * __restrict _src,
+    unsigned const size) {
+    rk_ulong hash = 0;
+    rk_ulong * dst = reinterpret_cast<rk_ulong *>(_dst);
+    rk_ulong const * src = reinterpret_cast<rk_ulong const *>(_src);
+    unsigned count = size / sizeof(rk_ulong);
+    unsigned remain = (size - count * sizeof(rk_ulong));
+    if (count) {
+        do {
+            hash |= *dst ^ *src;
+            *dst++ = *src++;
+        } while(--count > 0);
+    }
+    rk_ubyte * rdst = reinterpret_cast<rk_ubyte *>(dst);
+    rk_ubyte const * rsrc = reinterpret_cast<rk_ubyte const *>(src);
+    if (remain) {
+        do {
+            hash |= *rdst ^ *rsrc;
+            *rdst++ = *rsrc++;
+        } while(--remain > 0);
+    }
+    return (hash != 0);
+}
+
 void rk_fill_batch(
     rk_batch_t _batch,
     rk_uint count,
@@ -861,30 +887,28 @@ void rk_fill_batch(
         }
     }
     bool const is_empty = (batch->state < RK_BATCH_STATE_FILLED);
+    bool const resized = (count != batch->count);
     bool const got_everything = (flags && meshes && got_all_params);
-    bool const need_sorting = (is_empty || flags || meshes || count != batch->count);
     if (is_empty && !got_everything) {
         rk_printf("rk_fill_batch(): cannot freeze and empty batch.");
+        return;
     } else if (count > batch->count && !got_everything) {
         rk_printf("rk_fill_batch(): cannot grow a frozen batch.");
-    } else {
+        return;
+    }
     batch->count = count;
-    }
-    if (flags) {
-        memcpy(batch->flags, flags, batch->count * sizeof(rk_instance_flags));
-    }
-    if (meshes) {
-        memcpy(batch->meshes, meshes, batch->count * sizeof(rk_mesh));
-    }
+    bool const cmp_flags =
+        (flags && rk_compare_replace(batch->flags, flags, batch->count * sizeof(rk_instance_flags)));
+    bool const cmp_meshes =
+        (meshes && rk_compare_replace(batch->meshes, meshes, batch->count * sizeof(rk_mesh)));
+    bool const need_sorting = (cmp_flags || cmp_meshes || resized);
     if (batch->nparams) {
         rk_parameter const * const last_param = batch->params + batch->nparams;
         if (got_any_params) {
             rk_ubyte const * const * src = params;
             for (rk_parameter const * dst = batch->params; dst < last_param; ++dst, ++src) {
-                dst->dirty = (need_sorting || *src);
-                if (*src) {
-                    memcpy(dst->source, *src, batch->count * dst->src_size);
-                }
+                dst->dirty =
+                    (*src && rk_compare_replace(dst->source, *src, batch->count * dst->src_size)) || need_sorting;
             }
         } else if (need_sorting) {
             for (rk_parameter const * dst = batch->params; dst < last_param; ++dst) {