Stores meshes into vertices.

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)

__init__.py

@@ -44,6 +44,10 @@ def _ushort_addr(x):
     assert x.typecode == 'H'
     return x.buffer_info()[0]
 
+def _uint_addr(x):
+    assert x.typecode == 'I'
+    return x.buffer_info()[0]
+
 def _float_addr(x):
     assert x.typecode == 'f'
     return x.buffer_info()[0]
@@ -371,10 +375,13 @@ _create_vertices.argtypes = (
     ctypes.c_uint, # nvertices
     ctypes.c_void_p, # vertices
     ctypes.c_uint, # nindices
-    ctypes.c_void_p) # indices
+    ctypes.c_void_p, # vertices
+    ctypes.c_uint, # nmeshes
+    ctypes.c_void_p) # meshes
 
-def create_vertices(format, nvertices, vertices, indices):
-    return _create_vertices(format, nvertices, _ubyte_addr(vertices), len(indices), _ushort_addr(indices))
+def create_vertices(format, nvertices, vertices, indices, meshes):
+    return _create_vertices(format,
+        nvertices, _ubyte_addr(vertices), len(indices), _ushort_addr(indices), len(meshes), _uint_addr(meshes))
 
 create_batch = _engine.rk_create_batch
 create_batch.restype = _handle
@@ -382,7 +389,6 @@ create_batch.errcheck = _check_handle
 create_batch.argtypes = (
     ctypes.c_void_p, # vertices
     ctypes.c_uint, # max_size
-    ctypes.c_uint, # max_meshes
     ctypes.c_char_p) # params_format
 
 clear_buffer = _engine.rk_clear_buffer

cpp/render.hpp

@@ -116,12 +116,13 @@ RK_EXPORT rk_vertices_t rk_create_vertices(
     rk_uint nvertices,
     rk_ubyte const * vertices,
     rk_uint nindices,
-    rk_ushort const * indices);
+    rk_ushort const * indices,
+    rk_uint nmeshes,
+    rk_mesh const * meshes);
 
 RK_EXPORT rk_batch_t rk_create_batch(
     rk_vertices_t vertices,
     rk_uint max_size,
-    rk_uint max_meshes,
     rk_param_format const * params_format);
 
 RK_EXPORT void rk_clear_buffer(

cpp/render/render_opengles.cpp

@@ -314,7 +314,9 @@ rk_vertices_t rk_create_vertices(
     rk_uint nvertices,
     rk_ubyte const * _vertices,
     rk_uint nindices,
-    rk_ushort const * indices) {
+    rk_ushort const * indices,
+    rk_uint nmeshes,
+    rk_mesh const * meshes) {
     if (!format || !nvertices || !_vertices || !nindices || !indices) {
         rk_printf("rk_create_vertices(): invalid params.");
         return RK_INVALID_HANDLE;
@@ -345,12 +347,17 @@ rk_vertices_t rk_create_vertices(
     rk_vertices * const vertices = new rk_vertices;
     vertices->nvertices = nvertices;
     vertices->nindices = nindices;
+    vertices->nmeshes = nmeshes;
     vertices->format = new rk_vertex_format[format_size + 1];
     memcpy(vertices->format, format, (format_size + 1) * sizeof(rk_vertex_format));
     vertices->vertices = new rk_ubyte[nvertices * vertex_size];
     memcpy(vertices->vertices, _vertices, nvertices * vertex_size);
     vertices->indices = new rk_ushort[nindices];
     memcpy(vertices->indices, indices, nindices * sizeof(rk_ushort));
+    vertices->meshes = new rk_mesh[nmeshes];
+    memcpy(vertices->meshes, meshes, nmeshes * sizeof(rk_mesh));
+    vertices->vertices_buffer = 0;
+    vertices->indices_buffer = 0;
     return reinterpret_cast<rk_vertices_t>(vertices);
 }
 
@@ -488,9 +495,8 @@ static void rk_pack_mat3_int10_norm(
 rk_batch_t rk_create_batch(
     rk_vertices_t _vertices,
     rk_uint max_size,
-    rk_uint max_meshes,
     rk_param_format const * params_format) {
-    rk_vertices const * const vertices = reinterpret_cast<rk_vertices const *>(_vertices);
+    rk_vertices * const vertices = reinterpret_cast<rk_vertices *>(_vertices);
     if (!vertices || !max_size || max_size > RK_BATCH_MAX_SIZE) {
         rk_printf("rk_create_batch(): invalid params.");
         return RK_INVALID_HANDLE;
@@ -542,12 +548,12 @@ rk_batch_t rk_create_batch(
     batch->ncommands = 0;
     batch->ninstances = 0;
     batch->max_size = max_size;
-    batch->max_meshes = max_meshes;
+    batch->max_meshes = vertices->nmeshes;
     batch->nparams = nparams;
     batch->flags = new rk_instance_flags[max_size];
     batch->meshes = new rk_mesh[max_size];
     batch->indices = new rk_ushort[max_size];
-    batch->commands = new rk_command[max_meshes];
+    batch->commands = new rk_command[batch->max_meshes];
     if (nparams) {
         batch->params = new rk_parameter[nparams];
     } else {
@@ -555,17 +561,23 @@ rk_batch_t rk_create_batch(
     }
     glGenVertexArrays(1, &batch->vertex_array);
     glBindVertexArray(batch->vertex_array);
-    glGenBuffers(1, &batch->vertices_buffer);
-    glBindBuffer(GL_ARRAY_BUFFER, batch->vertices_buffer);
-    glBufferData(GL_ARRAY_BUFFER, vertices->nvertices * vertex_size, vertices->vertices, GL_STATIC_DRAW);
-    glBindBuffer(GL_ARRAY_BUFFER, 0);
-    glGenBuffers(1, &batch->indices_buffer);
-    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, batch->indices_buffer);
-    glBufferData(GL_ELEMENT_ARRAY_BUFFER, vertices->nindices * sizeof(rk_ushort), vertices->indices, GL_STATIC_DRAW);
+    if (!vertices->vertices_buffer) {
+        glGenBuffers(1, &vertices->vertices_buffer);
+        glBindBuffer(GL_ARRAY_BUFFER, vertices->vertices_buffer);
+        glBufferData(GL_ARRAY_BUFFER, vertices->nvertices * vertex_size, vertices->vertices, GL_STATIC_DRAW);
+        glBindBuffer(GL_ARRAY_BUFFER, 0);
+    }
+    if (vertices->indices_buffer) {
+        glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vertices->indices_buffer);
+    } else {
+        glGenBuffers(1, &vertices->indices_buffer);
+        glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vertices->indices_buffer);
+        glBufferData(GL_ELEMENT_ARRAY_BUFFER, vertices->nindices * sizeof(rk_ushort), vertices->indices, GL_STATIC_DRAW);
+    }
     if (rk_MultiDrawElementsIndirect) {
         glGenBuffers(1, &batch->commands_buffer);
         glBindBuffer(GL_DRAW_INDIRECT_BUFFER, batch->commands_buffer);
-        glBufferData(GL_DRAW_INDIRECT_BUFFER, max_meshes * sizeof(rk_command), nullptr, GL_DYNAMIC_DRAW);
+        glBufferData(GL_DRAW_INDIRECT_BUFFER, batch->max_meshes * sizeof(rk_command), nullptr, GL_DYNAMIC_DRAW);
         glBindBuffer(GL_DRAW_INDIRECT_BUFFER, 0);
     }
     if (nparams) {
@@ -577,7 +589,7 @@ rk_batch_t rk_create_batch(
     unsigned binding = 0;
     unsigned attrib = 0;
     unsigned offset = 0;
-    glBindVertexBuffer(binding, batch->vertices_buffer, 0, vertex_size);
+    glBindVertexBuffer(binding, vertices->vertices_buffer, 0, vertex_size);
     for (rk_vertex_format const * f = vertices->format; *f; ++f) {
         GLboolean const norm = (*f & RK_VERTEX_FORMAT_NORMALIZE) != 0;
         switch (*f & RK_VERTEX_FORMAT_MASK) {
@@ -837,6 +849,32 @@ 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 +899,28 @@ 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 {
-        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));
+        return;
     }
+    batch->count = count;
+    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) {
@@ -1005,8 +1041,6 @@ void rk_destroy_batch(
             delete[] batch->params;
             glDeleteBuffers(1, &batch->params_buffer);
         }
-        glDeleteBuffers(1, &batch->indices_buffer);
-        glDeleteBuffers(1, &batch->vertices_buffer);
         glDeleteVertexArrays(1, &batch->vertex_array);
         delete batch;
     }
@@ -1029,6 +1063,12 @@ void rk_destroy_vertices(
         delete[] vertices->format;
         delete[] vertices->vertices;
         delete[] vertices->indices;
+        if (vertices->vertices_buffer) {
+            glDeleteBuffers(1, &vertices->vertices_buffer);
+        }
+        if (vertices->indices_buffer) {
+            glDeleteBuffers(1, &vertices->indices_buffer);
+        }
         delete vertices;
     }
 }

cpp/render/render_opengles.hpp

@@ -41,9 +41,13 @@ struct rk_triangles {
 struct rk_vertices {
     unsigned nvertices;
     unsigned nindices;
+    unsigned nmeshes;
     rk_vertex_format * format;
     rk_ubyte * vertices;
     rk_ushort * indices;
+    rk_mesh * meshes;
+    GLuint vertices_buffer;
+    GLuint indices_buffer;
 };
 
 struct rk_command {
@@ -119,8 +123,6 @@ struct rk_batch {
     rk_command * commands;
     rk_parameter * params;
     GLuint vertex_array;
-    GLuint vertices_buffer;
-    GLuint indices_buffer;
     GLuint commands_buffer;
     GLuint params_buffer;
 };