Rework params packing.

cpp/render.hpp

@@ -85,22 +85,15 @@ enum : rk_uint {
 
 typedef rk_ushort rk_vertex_index;
 typedef rk_ushort rk_mesh_index;
+typedef rk_uint rk_vertex_input;
+typedef rk_uint rk_vertex_output;
+typedef rk_vertex_input rk_param_input;
 
 struct rk_mesh {
     rk_uint base_index;
     rk_uint ntriangles;
 };
 
-// param input types must be size compatible with an array of rk_param_input
-typedef rk_uint rk_param_input;
-
-template<typename _input_type>
-void rk_param_get_input_size(unsigned & _size, unsigned & _len) {
-    static_assert((sizeof(_input_type) % sizeof(rk_param_input)) == 0);
-    _size = sizeof(_input_type);
-    _len = sizeof(_input_type) / sizeof(rk_param_input);
-}
-
 RK_EXPORT void rk_render_initialize(
     rk_bool debug);
 

cpp/render/render_opengles.cpp

@@ -14,6 +14,7 @@
 // along with this program.  If not, see <http://www.gnu.org/licenses/>.
 
 #include "render_opengles.hpp"
+#include "vertex_format_opengles.hpp"
 #include "../display/display_glx.hpp"
 #include "../utils/cmp_memcpy.hpp"
 #include <cstdlib>
@@ -330,13 +331,13 @@ rk_vertices_t rk_create_vertices(
     for (rk_vertex_format const * f = format; *f; ++f, ++format_size) {
         switch (*f & RK_VERTEX_FORMAT_MASK) {
             case RK_VERTEX_FORMAT_VEC3_FLOAT:
-                vertex_size += sizeof(rk_vec3_float);
+                vertex_size += rk_vec3_float::get_output_size();
                 break;
             case RK_VERTEX_FORMAT_VEC3_INT10:
-                vertex_size += sizeof(rk_vec3_int10);
+                vertex_size += rk_vec3_int10::get_output_size();
                 break;
             case RK_VERTEX_FORMAT_VEC3_UINT10:
-                vertex_size += sizeof(rk_vec3_uint10);
+                vertex_size += rk_vec3_uint10::get_output_size();
                 break;
             default:
                 rk_printf("rk_create_vertices(): invalid vertex format.");
@@ -420,143 +421,6 @@ static void rk_buckets_alloc(
     }
 }
 
-#define rk_pack_short_norm(_f) (static_cast<rk_int>((_f) * 32767.f) & 65536)
-// #define rk_pack_short_norm(_f) (((static_cast<rk_int>((_f) * 32767.f) - 1) / 2) & 65535)
-
-#define rk_pack_int10_norm(_f) (static_cast<rk_int>((_f) * 511.f) & 1023)
-// #define rk_pack_int10_norm(_f) (((static_cast<rk_int>((_f) * 1023.f) - 1) / 2) & 1023)
-
-static void rk_pack_vec3_float(
-    unsigned const count,
-    rk_instance_index const * const __restrict indices,
-    rk_param_output * __restrict _dst,
-    rk_param_input const * const __restrict _src) {
-    rk_instance_index const * const last_index = indices + count;
-    rk_vec3_float * __restrict dst = reinterpret_cast<rk_vec3_float *>(_dst);
-    rk_vec3_float const * const __restrict src = reinterpret_cast<rk_vec3_float const *>(_src);
-    for (rk_instance_index const * __restrict index = indices; index < last_index; ++index, ++dst) {
-        *dst = src[*index];
-    }
-}
-
-static void rk_pack_vec3_short(
-    unsigned const count,
-    rk_instance_index const * const __restrict indices,
-    rk_param_output * __restrict _dst,
-    rk_param_input const * const __restrict _src) {
-    rk_instance_index const * const last_index = indices + count;
-    rk_vec3_short * __restrict dst = reinterpret_cast<rk_vec3_short *>(_dst);
-    rk_vec3_float const * const __restrict src = reinterpret_cast<rk_vec3_float const *>(_src);
-    for (rk_instance_index const * __restrict index = indices; index < last_index; ++index, ++dst) {
-        rk_vec3_float const & input = src[*index];
-        dst->x = static_cast<rk_short>(input.x);
-        dst->y = static_cast<rk_short>(input.y);
-        dst->z = static_cast<rk_short>(input.z);
-        dst->pad = 0;
-    }
-}
-
-static void rk_pack_vec3_short_norm(
-    unsigned const count,
-    rk_instance_index const * const __restrict indices,
-    rk_param_output * __restrict _dst,
-    rk_param_input const * const __restrict _src) {
-    rk_instance_index const * const last_index = indices + count;
-    rk_vec3_short * __restrict dst = reinterpret_cast<rk_vec3_short *>(_dst);
-    rk_vec3_float const * const __restrict src = reinterpret_cast<rk_vec3_float const *>(_src);
-    #define _convert(s) rk_pack_short_norm(s)
-    for (rk_instance_index const * __restrict index = indices; index < last_index; ++index, ++dst) {
-        rk_vec3_float const & input = src[*index];
-        dst->x = _convert(input.x);
-        dst->y = _convert(input.y);
-        dst->z = _convert(input.z);
-        dst->pad = 0;
-    }
-    #undef _convert
-}
-
-static void rk_pack_vec3_int10(
-    unsigned const count,
-    rk_instance_index const * const __restrict indices,
-    rk_param_output * __restrict _dst,
-    rk_param_input const * const __restrict _src) {
-    rk_instance_index const * const last_index = indices + count;
-    rk_vec3_int10 * __restrict dst = reinterpret_cast<rk_vec3_int10 *>(_dst);
-    rk_vec3_float const * const __restrict src = reinterpret_cast<rk_vec3_float const *>(_src);
-    #define _convert(s) (static_cast<rk_int>((s)) & 1023)
-    for (rk_instance_index const * __restrict index = indices; index < last_index; ++index, ++dst) {
-        rk_vec3_float const & input = src[*index];
-        *dst = _convert(input.x) | (_convert(input.y) << 10) | (_convert(input.z) << 20);
-    }
-    #undef _convert
-}
-
-static void rk_pack_vec3_int10_norm(
-    unsigned const count,
-    rk_instance_index const * const __restrict indices,
-    rk_param_output * __restrict _dst,
-    rk_param_input const * const __restrict _src) {
-    rk_instance_index const * const last_index = indices + count;
-    rk_vec3_int10 * __restrict dst = reinterpret_cast<rk_vec3_int10 *>(_dst);
-    rk_vec3_float const * const __restrict src = reinterpret_cast<rk_vec3_float const *>(_src);
-    #define _convert(s) rk_pack_int10_norm(s)
-    for (rk_instance_index const * __restrict index = indices; index < last_index; ++index, ++dst) {
-        rk_vec3_float const & input = src[*index];
-        *dst = _convert(input.x) | (_convert(input.y) << 10) | (_convert(input.z) << 20);
-    }
-    #undef _convert
-}
-
-static void rk_pack_mat3_float(
-    unsigned const count,
-    rk_instance_index const * const __restrict indices,
-    rk_param_output * __restrict _dst,
-    rk_param_input const * const __restrict _src) {
-    rk_instance_index const * const last_index = indices + count;
-    rk_mat3_float * __restrict dst = reinterpret_cast<rk_mat3_float *>(_dst);
-    rk_mat3_float const * const __restrict src = reinterpret_cast<rk_mat3_float const *>(_src);
-    for (rk_instance_index const * __restrict index = indices; index < last_index; ++index, ++dst) {
-        *dst = src[*index];
-    }
-    #undef _convert
-}
-
-static void rk_pack_mat3_int10(
-    unsigned const count,
-    rk_instance_index const * const __restrict indices,
-    rk_param_output * __restrict _dst,
-    rk_param_input const * const __restrict _src) {
-    rk_instance_index const * const last_index = indices + count;
-    rk_mat3_int10 * __restrict dst = reinterpret_cast<rk_mat3_int10 *>(_dst);
-    rk_mat3_float const * const __restrict src = reinterpret_cast<rk_mat3_float const *>(_src);
-    #define _convert(s) (static_cast<rk_int>((s)) & 1023)
-    for (rk_instance_index const * __restrict index = indices; index < last_index; ++index, ++dst) {
-        rk_mat3_float const & input = src[*index];
-        dst->x = _convert(input.x.x) | (_convert(input.x.y) << 10) | (_convert(input.x.z) << 20);
-        dst->y = _convert(input.y.x) | (_convert(input.y.y) << 10) | (_convert(input.y.z) << 20);
-        dst->z = _convert(input.z.x) | (_convert(input.z.y) << 10) | (_convert(input.z.z) << 20);
-    }
-    #undef _convert
-}
-
-static void rk_pack_mat3_int10_norm(
-    unsigned const count,
-    rk_instance_index const * const __restrict indices,
-    rk_param_output * __restrict _dst,
-    rk_param_input const * const __restrict _src) {
-    rk_instance_index const * const last_index = indices + count;
-    rk_mat3_int10 * __restrict dst = reinterpret_cast<rk_mat3_int10 *>(_dst);
-    rk_mat3_float const * const __restrict src = reinterpret_cast<rk_mat3_float const *>(_src);
-    #define _convert(s) rk_pack_int10_norm(s)
-    for (rk_instance_index const * __restrict index = indices; index < last_index; ++index, ++dst) {
-        rk_mat3_float const & input = src[*index];
-        dst->x = _convert(input.x.x) | (_convert(input.x.y) << 10) | (_convert(input.x.z) << 20);
-        dst->y = _convert(input.y.x) | (_convert(input.y.y) << 10) | (_convert(input.y.z) << 20);
-        dst->z = _convert(input.z.x) | (_convert(input.z.y) << 10) | (_convert(input.z.z) << 20);
-    }
-    #undef _convert
-}
-
 rk_batch_t rk_create_batch(
     rk_vertices_t _vertices,
     rk_uint max_size,
@@ -570,13 +434,13 @@ rk_batch_t rk_create_batch(
     for (rk_vertex_format const * f = vertices->format; *f; ++f) {
         switch (*f & RK_VERTEX_FORMAT_MASK) {
             case RK_VERTEX_FORMAT_VEC3_FLOAT:
-                vertex_size += sizeof(rk_vec3_float);
+                vertex_size += rk_vec3_float::get_output_size();
                 break;
             case RK_VERTEX_FORMAT_VEC3_INT10:
-                vertex_size += sizeof(rk_vec3_int10);
+                vertex_size += rk_vec3_int10::get_output_size();
                 break;
             case RK_VERTEX_FORMAT_VEC3_UINT10:
-                vertex_size += sizeof(rk_vec3_uint10);
+                vertex_size += rk_vec3_uint10::get_output_size();
                 break;
         }
     }
@@ -586,19 +450,19 @@ rk_batch_t rk_create_batch(
         for (rk_param_format const * f = params_format; *f; ++f, ++nparams) {
             switch (*f & RK_PARAM_FORMAT_MASK) {
                 case RK_PARAM_FORMAT_VEC3_FLOAT:
-                    params_size += sizeof(rk_vec3_float);
+                    params_size += rk_vec3_float::get_output_size();
                     break;
                 case RK_PARAM_FORMAT_VEC3_SHORT:
-                    params_size += sizeof(rk_vec3_short);
+                    params_size += rk_vec3_short::get_output_size();
                     break;
                 case RK_PARAM_FORMAT_VEC3_INT10:
-                    params_size += sizeof(rk_vec3_int10);
+                    params_size += rk_vec3_int10::get_output_size();
                     break;
                 case RK_PARAM_FORMAT_MAT3_FLOAT:
-                    params_size += sizeof(rk_mat3_float);
+                    params_size += rk_mat3_float::get_output_size();
                     break;
                 case RK_PARAM_FORMAT_MAT3_INT10:
-                    params_size += sizeof(rk_mat3_int10);
+                    params_size += rk_mat3_int10::get_output_size();
                     break;
                 default:
                     rk_printf("rk_create_batch(): invalid param format.");
@@ -667,19 +531,19 @@ rk_batch_t rk_create_batch(
                 glEnableVertexAttribArray(attrib);
                 glVertexAttribFormat(attrib, 3, GL_FLOAT, GL_FALSE, offset);
                 glVertexAttribBinding(attrib++, binding);
-                offset += sizeof(rk_vec3_float);
+                offset += rk_vec3_float::get_output_size();
                 break;
             case RK_VERTEX_FORMAT_VEC3_INT10:
                 glEnableVertexAttribArray(attrib);
                 glVertexAttribFormat(attrib, 4, GL_INT_2_10_10_10_REV, norm, offset);
                 glVertexAttribBinding(attrib++, binding);
-                offset += sizeof(rk_vec3_int10);
+                offset += rk_vec3_int10::get_output_size();
                 break;
             case RK_VERTEX_FORMAT_VEC3_UINT10:
                 glEnableVertexAttribArray(attrib);
                 glVertexAttribFormat(attrib, 4, GL_UNSIGNED_INT_2_10_10_10_REV, norm, offset);
                 glVertexAttribBinding(attrib++, binding);
-                offset += sizeof(rk_vec3_uint10);
+                offset += rk_vec3_uint10::get_output_size();
                 break;
         }
     }
@@ -694,64 +558,66 @@ rk_batch_t rk_create_batch(
             param->offset = offset;
             switch (*f & RK_PARAM_FORMAT_MASK) {
                 case RK_PARAM_FORMAT_VEC3_FLOAT:
-                    rk_param_get_input_size<rk_vec3>(param->src_size, param->src_len);
+                    param->src_size = rk_vec3_float::get_input_size();
-                    rk_param_get_output_size<rk_vec3_float>(param->dst_size, param->dst_len);
+                    param->dst_size = rk_vec3_float::get_output_size();
-                    param->packer = rk_pack_vec3_float;
+                    param->packer = rk_vec3_float::param_packer;
                     glBindVertexBuffer(binding, batch->params_buffer, param->offset, param->dst_size);
                     glEnableVertexAttribArray(attrib);
                     glVertexAttribFormat(attrib, 3, GL_FLOAT, GL_FALSE, 0);
                     glVertexAttribBinding(attrib++, binding);
                     break;
                 case RK_PARAM_FORMAT_VEC3_SHORT:
-                    rk_param_get_input_size<rk_vec3>(param->src_size, param->src_len);
+                    param->src_size = rk_vec3_short::get_input_size();
-                    rk_param_get_output_size<rk_vec3_short>(param->dst_size, param->dst_len);
+                    param->dst_size = rk_vec3_short::get_output_size();
-                    param->packer = norm ? rk_pack_vec3_short_norm : rk_pack_vec3_short;
+                    param->packer = norm ? rk_vec3_short_norm::param_packer : rk_vec3_short::param_packer;
                     glBindVertexBuffer(binding, batch->params_buffer, param->offset, param->dst_size);
                     glEnableVertexAttribArray(attrib);
                     glVertexAttribFormat(attrib, 3, GL_SHORT, norm, 0);
                     glVertexAttribBinding(attrib++, binding);
                     break;
                 case RK_PARAM_FORMAT_VEC3_INT10:
-                    rk_param_get_input_size<rk_vec3>(param->src_size, param->src_len);
+                    param->src_size = rk_vec3_int10::get_input_size();
-                    rk_param_get_output_size<rk_vec3_int10>(param->dst_size, param->dst_len);
+                    param->dst_size = rk_vec3_int10::get_output_size();
-                    param->packer = norm ? rk_pack_vec3_int10_norm : rk_pack_vec3_int10;
+                    param->packer = norm ? rk_vec3_int10_norm::param_packer : rk_vec3_int10::param_packer;
                     glBindVertexBuffer(binding, batch->params_buffer, param->offset, param->dst_size);
                     glEnableVertexAttribArray(attrib);
                     glVertexAttribFormat(attrib, 4, GL_INT_2_10_10_10_REV, norm, 0);
                     glVertexAttribBinding(attrib++, binding);
                     break;
                 case RK_PARAM_FORMAT_MAT3_FLOAT:
-                    rk_param_get_input_size<rk_mat3>(param->src_size, param->src_len);
+                    param->src_size = rk_mat3_float::get_input_size();
-                    rk_param_get_output_size<rk_mat3_float>(param->dst_size, param->dst_len);
+                    param->dst_size = rk_mat3_float::get_output_size();
-                    param->packer = rk_pack_mat3_float;
+                    param->packer = rk_mat3_float::param_packer;
                     glBindVertexBuffer(binding, batch->params_buffer, param->offset, param->dst_size);
                     glEnableVertexAttribArray(attrib);
-                    glVertexAttribFormat(attrib, 3, GL_FLOAT, GL_FALSE, offsetof(rk_mat3_float, x));
+                    glVertexAttribFormat(attrib, 3, GL_FLOAT, GL_FALSE, rk_mat3_float::get_output_offset(0));
                     glVertexAttribBinding(attrib++, binding);
                     glEnableVertexAttribArray(attrib);
-                    glVertexAttribFormat(attrib, 3, GL_FLOAT, GL_FALSE, offsetof(rk_mat3_float, y));
+                    glVertexAttribFormat(attrib, 3, GL_FLOAT, GL_FALSE, rk_mat3_float::get_output_offset(1));
                     glVertexAttribBinding(attrib++, binding);
                     glEnableVertexAttribArray(attrib);
-                    glVertexAttribFormat(attrib, 3, GL_FLOAT, GL_FALSE, offsetof(rk_mat3_float, z));
+                    glVertexAttribFormat(attrib, 3, GL_FLOAT, GL_FALSE, rk_mat3_float::get_output_offset(2));
                     glVertexAttribBinding(attrib++, binding);
                     break;
                 case RK_PARAM_FORMAT_MAT3_INT10:
-                    rk_param_get_input_size<rk_mat3>(param->src_size, param->src_len);
+                    param->src_size = rk_mat3_int10::get_input_size();
-                    rk_param_get_output_size<rk_mat3_int10>(param->dst_size, param->dst_len);
+                    param->dst_size = rk_mat3_int10::get_output_size();
-                    param->packer = norm ? rk_pack_mat3_int10_norm : rk_pack_mat3_int10;
+                    param->packer = norm ? rk_mat3_int10_norm::param_packer : rk_mat3_int10::param_packer;
                     glBindVertexBuffer(binding, batch->params_buffer, param->offset, param->dst_size);
                     glEnableVertexAttribArray(attrib);
-                    glVertexAttribFormat(attrib, 4, GL_INT_2_10_10_10_REV, norm, offsetof(rk_mat3_int10, x));
+                    glVertexAttribFormat(attrib, 4, GL_INT_2_10_10_10_REV, norm, rk_mat3_int10::get_output_offset(0));
                     glVertexAttribBinding(attrib++, binding);
                     glEnableVertexAttribArray(attrib);
-                    glVertexAttribFormat(attrib, 4, GL_INT_2_10_10_10_REV, norm, offsetof(rk_mat3_int10, y));
+                    glVertexAttribFormat(attrib, 4, GL_INT_2_10_10_10_REV, norm, rk_mat3_int10::get_output_offset(1));
                     glVertexAttribBinding(attrib++, binding);
                     glEnableVertexAttribArray(attrib);
-                    glVertexAttribFormat(attrib, 4, GL_INT_2_10_10_10_REV, norm, offsetof(rk_mat3_int10, z));
+                    glVertexAttribFormat(attrib, 4, GL_INT_2_10_10_10_REV, norm, rk_mat3_int10::get_output_offset(2));
                     glVertexAttribBinding(attrib++, binding);
                     break;
             }
             glVertexBindingDivisor(binding, 1);
+            param->src_len = param->src_size / sizeof(rk_param_output);
+            param->dst_len = param->dst_size / sizeof(rk_param_output);
             param->source = new rk_param_input[max_size * param->src_len];
             memset(param->source, 0xFF, max_size * param->src_size);
             offset += max_size * param->dst_size;
@@ -762,6 +628,7 @@ rk_batch_t rk_create_batch(
     return reinterpret_cast<rk_batch_t>(batch);
 }
 
+[[RK_HOT, RK_FAST]]
 static void rk_sort_batch(
     rk_batch const & batch) {
     rk_bucket const * const last_bucket = rk_buckets + batch.vertices->nmeshes;
@@ -798,6 +665,7 @@ static void rk_sort_batch(
     batch.state = RK_BATCH_STATE_SORTED;
 }
 
+[[RK_HOT, RK_FAST]]
 static void rk_pack_batch(
     rk_batch const & batch) {
     if (batch.nparams) {
@@ -821,6 +689,7 @@ static void rk_pack_batch(
     batch.state = RK_BATCH_STATE_PACKED;
 }
 
+[[RK_HOT, RK_FAST]]
 void rk_fill_batch(
     rk_batch_t _batch,
     rk_uint count,

cpp/render/render_opengles.hpp

@@ -21,6 +21,10 @@
 #include <GLES3/gl3ext.h>
 #include <GLES3/gl3platform.h>
 
+static_assert(sizeof(rk_vertex_output) == 4);
+
+typedef rk_vertex_output rk_param_output;
+
 struct rk_shader {
     GLuint vertex;
     GLuint fragment;
@@ -58,53 +62,11 @@ struct rk_command {
     GLuint base_instance;
 };
 
-// param output types must be size compatible with an array of rk_param_output
-typedef rk_uint rk_param_output;
-
-template<typename _output_type>
-void rk_param_get_output_size(unsigned & _size, unsigned & _len) {
-    static_assert((sizeof(_output_type) % sizeof(rk_param_output)) == 0);
-    _size = sizeof(_output_type);
-    _len = sizeof(_output_type) / sizeof(rk_param_output);
-}
-
-struct rk_vec3_float {
-    float x;
-    float y;
-    float z;
-};
-
-static_assert(sizeof(rk_vec3_float) == sizeof(rk_vec3));
-
-struct rk_vec3_short {
-    rk_short x;
-    rk_short y;
-    rk_short z;
-    rk_short pad;
-};
-
-typedef rk_int rk_vec3_int10;
-typedef rk_uint rk_vec3_uint10;
-
-struct rk_mat3_float {
-    rk_vec3_float x;
-    rk_vec3_float y;
-    rk_vec3_float z;
-};
-
-static_assert(sizeof(rk_mat3_float) == sizeof(rk_mat3));
-
-struct rk_mat3_int10 {
-    rk_vec3_int10 x;
-    rk_vec3_int10 y;
-    rk_vec3_int10 z;
-};
-
 typedef void (*rk_packer)(
     unsigned const, // count
-    rk_instance_index const * const, // indices
+    rk_instance_index const * const __restrict, // indices
-    rk_param_output *, // dst
+    rk_param_output * __restrict, // dst
-    rk_param_input const * const); // src
+    rk_param_input const * const __restrict); // src
 
 struct rk_parameter {
     mutable bool dirty;

cpp/render/vertex_format_opengles.hpp

@@ -0,0 +1,230 @@
+// Copyright (C) 2023 RozK
+//
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU Affero General Public License for more details.
+//
+// You should have received a copy of the GNU Affero General Public License
+// along with this program.  If not, see <http://www.gnu.org/licenses/>.
+
+#ifndef RK_ENGINE_VERTEX_FORMAT_H
+#define RK_ENGINE_VERTEX_FORMAT_H
+
+#include "render_opengles.hpp"
+#include <limits>
+
+namespace rk_vertex {
+
+#pragma pack(push, 1)
+
+template<typename _type>
+struct alignas(alignof(_type)) rk_input {
+    _type input;
+};
+
+template<typename _type, typename _input, bool _signed, bool _normalized>
+struct alignas(alignof(_type)) rk_output {
+    _type output;
+
+    [[RK_FAST]]
+    inline void convert(
+        rk_input<_input> const & __restrict src) {
+        output = static_cast<_type>(src.input);
+    }
+};
+
+template<typename _type, bool _signed>
+struct alignas(alignof(_type)) rk_output<_type, rk_float, _signed, true> {
+    _type output;
+
+    [[RK_FAST]]
+    inline void convert(
+        rk_input<rk_float> const & __restrict src) {
+        enum : _type { max = std::numeric_limits<_type>::max() };
+        output = static_cast<_type>(src.input * static_cast<float>(max));
+    }
+};
+
+#pragma pack(4)
+
+template<typename _input, unsigned _cols>
+struct alignas(4) rk_input_row {
+    rk_input<_input> input_col[_cols];
+};
+
+template<typename _output, typename _input, unsigned _cols, bool _signed, bool _normalized>
+struct alignas(4) rk_output_row {
+    rk_output<_output, _input, _signed, _normalized> output_col[_cols];
+
+    [[RK_FAST, RK_FLATTEN, RK_UNROLLED]]
+    inline void convert(
+        rk_input_row<_input, _cols> const & __restrict src) {
+        for (unsigned col = 0; col < _cols; ++col) {
+            output_col[col].convert(src.input_col[col]);
+        }
+    }
+};
+
+template<typename _input, unsigned _cols, bool _signed, bool _normalized>
+struct alignas(4) rk_output_row<rk_packed<_signed, _cols>, _input, _cols, _signed, _normalized> {
+    rk_output<rk_packed<_signed, _cols>, _input, _signed, _normalized> output_cols;
+
+    [[RK_FAST, RK_FLATTEN]]
+    inline void convert(
+        rk_input_row<_input, _cols> const & __restrict src) {
+        output_cols.convert(src);
+    }
+};
+
+template<typename _input, unsigned _cols, unsigned _rows>
+struct alignas(4) rk_input_format {
+    rk_input_row<_input, _cols> input_row[_rows];
+};
+
+template<typename _output, typename _input, unsigned _cols, unsigned _rows, bool _signed, bool _normalized>
+struct alignas(4) rk_output_format {
+    rk_output_row<_output, _input, _cols, _signed, _normalized> output_row[_rows];
+
+    [[RK_FAST, RK_FLATTEN, RK_UNROLLED]]
+    inline void convert(
+        rk_input_format<_input, _cols, _rows> const & __restrict src) {
+        for (unsigned row = 0; row < _rows; ++row) {
+            output_row[row].convert(src.input_row[row]);
+        }
+    }
+};
+
+template<typename _input, bool _signed, bool _normalized>
+struct alignas(alignof(rk_packed<_signed, 3>)) rk_output<rk_packed<_signed, 3>, _input, _signed, _normalized> {
+    rk_packed<_signed, 3> output;
+
+    [[RK_FAST]]
+    inline void convert(
+        rk_input_row<_input, 3> const & __restrict src) {
+        typedef typename rk_packed<_signed, 3>::type packed_type;
+        output.packed =
+            ((static_cast<packed_type>(src.input_col[0].input) & 1023)) |
+            ((static_cast<packed_type>(src.input_col[1].input) & 1023) << 10) |
+            ((static_cast<packed_type>(src.input_col[2].input) & 1023) << 20);
+    }
+};
+
+template<typename _input, bool _signed, bool _normalized>
+struct alignas(alignof(rk_packed<_signed, 4>)) rk_output<rk_packed<_signed, 4>, _input, _signed, _normalized> {
+    rk_packed<_signed, 4> output;
+
+    [[RK_FAST]]
+    inline void convert(
+        rk_input_row<_input, 4> const & __restrict src) {
+        typedef typename rk_packed<_signed, 4>::type packed_type;
+        output.packed =
+            ((static_cast<packed_type>(src.input_col[0].input) & 1023)) |
+            ((static_cast<packed_type>(src.input_col[1].input) & 1023) << 10) |
+            ((static_cast<packed_type>(src.input_col[2].input) & 1023) << 20) |
+            ((static_cast<packed_type>(src.input_col[3].input) &    3) << 30);
+    }
+};
+
+template<>
+struct alignas(alignof(rk_packed<true, 3>)) rk_output<rk_packed<true, 3>, rk_float, true, true> {
+    rk_packed<true, 3> output;
+
+    [[RK_FAST]]
+    inline void convert(
+        rk_input_row<rk_float, 3> const & __restrict src) {
+        output.packed =
+            ((static_cast<rk_int>(src.input_col[0].input * 511.f) & 1023)) |
+            ((static_cast<rk_int>(src.input_col[1].input * 511.f) & 1023) << 10) |
+            ((static_cast<rk_int>(src.input_col[2].input * 511.f) & 1023) << 20);
+    }
+};
+
+template<>
+struct alignas(alignof(rk_packed<true, 4>)) rk_output<rk_packed<true, 4>, rk_float, true, true> {
+    rk_packed<true, 4> output;
+
+    [[RK_FAST]]
+    inline void convert(
+        rk_input_row<rk_float, 4> const & __restrict src) {
+        output.packed =
+            ((static_cast<rk_uint>(src.input_col[0].input * 511.f) & 1023)) |
+            ((static_cast<rk_uint>(src.input_col[1].input * 511.f) & 1023) << 10) |
+            ((static_cast<rk_uint>(src.input_col[2].input * 511.f) & 1023) << 20) |
+            ((static_cast<rk_uint>(src.input_col[3].input)         &    3) << 30);
+    }
+};
+
+#pragma pack(pop)
+
+template<typename _output, typename _input, unsigned _cols, unsigned _rows, bool _signed, bool _normalized>
+struct rk_format {
+    typedef rk_input<_input> input;
+    typedef rk_output<_output, input, _signed, _normalized> output;
+
+    typedef rk_input_row<_input, _cols> input_row;
+    typedef rk_output_row<_output, _input, _cols, _signed, _normalized> output_row;
+
+    typedef rk_input_format<_input, _cols, _rows> input_format;
+    typedef rk_output_format<_output, _input, _cols, _rows, _signed, _normalized> output_format;
+
+    static_assert(sizeof(input) == sizeof(_input));
+    static_assert(sizeof(output) == sizeof(_output));
+    static_assert((sizeof(input_row) % sizeof(rk_vertex_input)) == 0);
+    static_assert((sizeof(output_row) % sizeof(rk_vertex_output)) == 0);
+    static_assert((sizeof(input_format) % sizeof(rk_vertex_input)) == 0);
+    static_assert((sizeof(output_format) % sizeof(rk_vertex_output)) == 0);
+
+    static unsigned get_input_size() {
+        return sizeof(input_format);
+    }
+
+    static unsigned get_output_size() {
+        return sizeof(output_format);
+    }
+
+    static unsigned get_output_offset(unsigned const index) {
+        return index * sizeof(output_row);
+    }
+
+    [[RK_FAST, RK_FLATTEN]]
+    inline static void convert(
+        output_format & __restrict dst,
+        input_format const & __restrict src) {
+        dst.convert(src);
+    }
+
+    [[RK_HOT, RK_FAST, RK_FLATTEN]]
+    static void param_packer(
+        unsigned const count,
+        rk_instance_index const * const __restrict indices,
+        rk_param_output * __restrict _dst,
+        rk_param_input const * const __restrict _src) {
+        rk_instance_index const * const last_index = indices + count;
+        output_format * __restrict dst = reinterpret_cast<output_format *>(_dst);
+        input_format const * const __restrict src = reinterpret_cast<input_format const *>(_src);
+        for (rk_instance_index const * __restrict index = indices; index < last_index; ++index, ++dst) {
+            dst->convert(src[*index]);
+        }
+    }
+};
+
+} // namepace rk_vertex
+
+typedef rk_vertex::rk_format<rk_float, rk_float,           3, 1, true, false> rk_vec3_float;
+typedef rk_vertex::rk_format<rk_short, rk_float,           3, 1, true, false> rk_vec3_short;
+typedef rk_vertex::rk_format<rk_short, rk_float,           3, 1, true, true>  rk_vec3_short_norm;
+typedef rk_vertex::rk_format<rk_packed<true, 3>, rk_float, 3, 1, true, false> rk_vec3_int10;
+typedef rk_vertex::rk_format<rk_packed<true, 3>, rk_float, 3, 1, true, true>  rk_vec3_int10_norm;
+typedef rk_vertex::rk_format<rk_packed<true, 3>, rk_float, 3, 1, true, false> rk_vec3_uint10;
+typedef rk_vertex::rk_format<rk_packed<true, 3>, rk_float, 3, 1, true, true>  rk_vec3_uint10_norm;
+typedef rk_vertex::rk_format<rk_float,           rk_float, 3, 3, true, false> rk_mat3_float;
+typedef rk_vertex::rk_format<rk_packed<true, 3>, rk_float, 3, 3, true, false> rk_mat3_int10;
+typedef rk_vertex::rk_format<rk_packed<true, 3>, rk_float, 3, 3, true, true>  rk_mat3_int10_norm;
+
+#endif // RK_ENGINE_VERTEX_FORMAT_H