From 9d7806cb43f475b079e59c02d891175ce85bb1a5 Mon Sep 17 00:00:00 2001
From: D-AIRY <admin@ds-servers.com>
Date: Sat, 28 Nov 2020 14:02:45 +0300
Subject: [PATCH] Reduced GPU-side vertex size
---
source/anim/AnimatedModelProvider.cpp | 16 +++++++++
source/anim/AnimatedModelShared.cpp | 40 +++++++++++++++++++----
source/anim/AnimatedModelShared.h | 2 +-
source/anim/DynamicModelProvider.cpp | 14 +++++++-
source/anim/DynamicModelShared.cpp | 28 +++++++++++++---
source/anim/DynamicModelShared.h | 2 +-
source/xcommon/resource/IXResourceModel.h | 20 ++++++++++++
7 files changed, 107 insertions(+), 15 deletions(-)
diff --git a/source/anim/AnimatedModelProvider.cpp b/source/anim/AnimatedModelProvider.cpp
index 014167dcd..a3728936d 100644
--- a/source/anim/AnimatedModelProvider.cpp
+++ b/source/anim/AnimatedModelProvider.cpp
@@ -30,6 +30,8 @@ void CAnimatedModelProvider::setDevice(IGXDevice *pDevice)
{
m_pRenderContext = pDevice;
+#if 0
+ // 72 bytes
GXVertexElement layoutDynamicEx[] =
{
{0, 0, GXDECLTYPE_FLOAT3, GXDECLUSAGE_POSITION},
@@ -41,6 +43,20 @@ void CAnimatedModelProvider::setDevice(IGXDevice *pDevice)
{0, 60, GXDECLTYPE_FLOAT4, GXDECLUSAGE_BLENDWEIGHT},
GX_DECL_END()
};
+#endif
+
+ // 52 bytes
+ GXVertexElement layoutDynamicEx[] =
+ {
+ {0, 0, GXDECLTYPE_FLOAT3, GXDECLUSAGE_POSITION},
+ {0, 12, GXDECLTYPE_FLOAT2, GXDECLUSAGE_TEXCOORD},
+ {0, 20, GXDECLTYPE_SHORT4N, GXDECLUSAGE_NORMAL},
+ {0, 28, GXDECLTYPE_SHORT4N, GXDECLUSAGE_TANGENT},
+ {0, 36, GXDECLTYPE_SHORT4N, GXDECLUSAGE_BINORMAL},
+ {0, 44, GXDECLTYPE_UBYTE4, GXDECLUSAGE_BLENDINDICES},
+ {0, 48, GXDECLTYPE_UBYTE4N, GXDECLUSAGE_BLENDWEIGHT},
+ GX_DECL_END()
+ };
m_pVertexDeclaration = getDevice()->createVertexDeclaration(layoutDynamicEx);
diff --git a/source/anim/AnimatedModelShared.cpp b/source/anim/AnimatedModelShared.cpp
index 2fdf6207a..44edd18cd 100644
--- a/source/anim/AnimatedModelShared.cpp
+++ b/source/anim/AnimatedModelShared.cpp
@@ -459,11 +459,11 @@ bool CAnimatedModelShared::init(UINT uResourceCount, IXResourceModelAnimated **p
}
UINT **ppIndices = new UINT*[m_uLodCount];
- XResourceModelAnimatedVertex **ppVertices = new XResourceModelAnimatedVertex*[m_uLodCount];
+ XResourceModelAnimatedVertexGPU **ppVertices = new XResourceModelAnimatedVertexGPU*[m_uLodCount];
for(UINT i = 0; i < m_uLodCount; ++i)
{
ppIndices[i] = new UINT[aLodIndexCount[i]];
- ppVertices[i] = new XResourceModelAnimatedVertex[aLodVertexCount[i]];
+ ppVertices[i] = new XResourceModelAnimatedVertexGPU[aLodVertexCount[i]];
}
@@ -494,9 +494,35 @@ bool CAnimatedModelShared::init(UINT uResourceCount, IXResourceModelAnimated **p
{
// append subset to lod (group by materialId)
- //memcpy(&ppIndices[uLod][subset.uStartIndex], pSubset->pIndices, sizeof(UINT) * pSubset->iIndexCount);
- memcpy(&ppVertices[uLod][subset.uStartVertex], pSubset->pVertices, sizeof(XResourceModelAnimatedVertex) * pSubset->iVertexCount);
+ //memcpy(&ppVertices[uLod][subset.uStartVertex], pSubset->pVertices, sizeof(XResourceModelAnimatedVertex) * pSubset->iVertexCount);
+ for(UINT k = 0; k < pSubset->iVertexCount; ++k)
+ {
+#define TO_SHORT(v) ((short)((v) * 32767.0f))
+ auto &dst = (ppVertices[uLod] + subset.uStartVertex)[k];
+ auto &src = pSubset->pVertices[k];
+ dst.vPos = src.vPos;
+ dst.vTex = src.vTex;
+ dst.vNorm[0] = TO_SHORT(src.vNorm.x);
+ dst.vNorm[1] = TO_SHORT(src.vNorm.y);
+ dst.vNorm[2] = TO_SHORT(src.vNorm.z);
+ dst.vTangent[0] = TO_SHORT(src.vTangent.x);
+ dst.vTangent[1] = TO_SHORT(src.vTangent.y);
+ dst.vTangent[2] = TO_SHORT(src.vTangent.z);
+ dst.vBinorm[0] = TO_SHORT(src.vBinorm.x);
+ dst.vBinorm[1] = TO_SHORT(src.vBinorm.y);
+ dst.vBinorm[2] = TO_SHORT(src.vBinorm.z);
+ for(UINT k = 0; k < 4; ++k)
+ {
+ dst.u8BoneIndices[k] = src.u8BoneIndices[k];
+ }
+ dst.vBoneWeights[0] = (byte)(src.vBoneWeights.x * 255.0f);
+ dst.vBoneWeights[1] = (byte)(src.vBoneWeights.y * 255.0f);
+ dst.vBoneWeights[2] = (byte)(src.vBoneWeights.z * 255.0f);
+ dst.vBoneWeights[3] = (byte)(src.vBoneWeights.w * 255.0f);
+#undef TO_SHORT
+ }
+ //memcpy(&ppIndices[uLod][subset.uStartIndex], pSubset->pIndices, sizeof(UINT) * pSubset->iIndexCount);
UINT uIndexDelta = aLodVertexCount[uLod];
for(UINT j = 0; j < pSubset->iIndexCount; ++j)
{
@@ -506,7 +532,7 @@ bool CAnimatedModelShared::init(UINT uResourceCount, IXResourceModelAnimated **p
//@TODO: optimize that!!!
for(UINT j = 0; j < pSubset->iVertexCount; ++j)
{
- XResourceModelAnimatedVertex &vtx = ppVertices[uLod][subset.uStartVertex + j];
+ XResourceModelAnimatedVertexGPU &vtx = ppVertices[uLod][subset.uStartVertex + j];
for(UINT k = 0; k < 4; ++k)
{
vtx.u8BoneIndices[k] = getBoneId(aSubsets[i].pResource->getBoneName(vtx.u8BoneIndices[k]));
@@ -546,7 +572,7 @@ bool CAnimatedModelShared::init(UINT uResourceCount, IXResourceModelAnimated **p
if(m_pProvider->getCore()->isOnMainThread())
{
m_ppIndexBuffer[i] = m_pDevice->createIndexBuffer(sizeof(UINT) * aLodIndexCount[i], GXBUFFER_USAGE_STATIC, GXIT_UINT32, ppIndices[i]);
- IGXVertexBuffer *pVertexBuffer = m_pDevice->createVertexBuffer(sizeof(XResourceModelAnimatedVertex) * aLodVertexCount[i], GXBUFFER_USAGE_STATIC, ppVertices[i]);
+ IGXVertexBuffer *pVertexBuffer = m_pDevice->createVertexBuffer(sizeof(XResourceModelAnimatedVertexGPU) * aLodVertexCount[i], GXBUFFER_USAGE_STATIC, ppVertices[i]);
m_ppRenderBuffer[i] = m_pDevice->createRenderBuffer(1, &pVertexBuffer, m_pProvider->getVertexDeclaration());
mem_release(pVertexBuffer);
}
@@ -680,7 +706,7 @@ void CAnimatedModelShared::initGPUresources()
for(UINT i = 0; i < m_uLodCount; ++i)
{
m_ppIndexBuffer[i] = m_pDevice->createIndexBuffer(sizeof(UINT) * m_puTempTotalIndices[i], GXBUFFER_USAGE_STATIC, GXIT_UINT32, m_ppTempIndices[i]);
- IGXVertexBuffer *pVertexBuffer = m_pDevice->createVertexBuffer(sizeof(XResourceModelAnimatedVertex) * m_puTempTotalVertices[i], GXBUFFER_USAGE_STATIC, m_ppTempVertices[i]);
+ IGXVertexBuffer *pVertexBuffer = m_pDevice->createVertexBuffer(sizeof(XResourceModelAnimatedVertexGPU) * m_puTempTotalVertices[i], GXBUFFER_USAGE_STATIC, m_ppTempVertices[i]);
m_ppRenderBuffer[i] = m_pDevice->createRenderBuffer(1, &pVertexBuffer, m_pProvider->getVertexDeclaration());
mem_release(pVertexBuffer);
diff --git a/source/anim/AnimatedModelShared.h b/source/anim/AnimatedModelShared.h
index f8e84fee2..9c67a1404 100644
--- a/source/anim/AnimatedModelShared.h
+++ b/source/anim/AnimatedModelShared.h
@@ -72,7 +72,7 @@ protected:
IGXIndexBuffer **m_ppIndexBuffer = NULL;
UINT **m_ppTempIndices = NULL;
- XResourceModelAnimatedVertex **m_ppTempVertices = NULL;
+ XResourceModelAnimatedVertexGPU **m_ppTempVertices = NULL;
UINT *m_puTempTotalIndices = 0;
UINT *m_puTempTotalVertices = 0;
diff --git a/source/anim/DynamicModelProvider.cpp b/source/anim/DynamicModelProvider.cpp
index 2d3fe536b..98c6715f2 100644
--- a/source/anim/DynamicModelProvider.cpp
+++ b/source/anim/DynamicModelProvider.cpp
@@ -209,7 +209,8 @@ void CDynamicModelProvider::setDevice(IGXDevice *pDevice)
{
m_pRenderContext = pDevice;
-
+#if 0
+ // 56 bytes
GXVertexElement layoutStaticEx[] =
{
{0, 0, GXDECLTYPE_FLOAT3, GXDECLUSAGE_POSITION},
@@ -219,6 +220,17 @@ void CDynamicModelProvider::setDevice(IGXDevice *pDevice)
{0, 44, GXDECLTYPE_FLOAT3, GXDECLUSAGE_BINORMAL},
GX_DECL_END()
};
+#endif
+ // 40 bytes/ 28?
+ GXVertexElement layoutStaticEx[] =
+ {
+ {0, 0, GXDECLTYPE_FLOAT3, GXDECLUSAGE_POSITION},
+ {0, 12, GXDECLTYPE_FLOAT2, GXDECLUSAGE_TEXCOORD},
+ {0, 20, GXDECLTYPE_SHORT4N, GXDECLUSAGE_NORMAL},
+ {0, 28, GXDECLTYPE_SHORT4N, GXDECLUSAGE_TANGENT},
+ {0, 36, GXDECLTYPE_SHORT4N, GXDECLUSAGE_BINORMAL},
+ GX_DECL_END()
+ };
m_pVertexDeclaration = pDevice->createVertexDeclaration(layoutStaticEx);
diff --git a/source/anim/DynamicModelShared.cpp b/source/anim/DynamicModelShared.cpp
index 0216f5d03..f4444d798 100644
--- a/source/anim/DynamicModelShared.cpp
+++ b/source/anim/DynamicModelShared.cpp
@@ -138,7 +138,7 @@ bool CDynamicModelShared::init(IXResourceModelStatic *pResource)
if(!m_pProvider->getCore()->isOnMainThread())
{
m_ppTempIndices = new UINT*[uLodCount];
- m_ppTempVertices = new XResourceModelStaticVertex*[uLodCount];
+ m_ppTempVertices = new XResourceModelStaticVertexGPU*[uLodCount];
m_puTempTotalIndices = new UINT[uLodCount];
m_puTempTotalVertices = new UINT[uLodCount];
}
@@ -157,7 +157,7 @@ bool CDynamicModelShared::init(IXResourceModelStatic *pResource)
uTotalVertices += pSubset->iVertexCount;
}
- XResourceModelStaticVertex *pVertices = new XResourceModelStaticVertex[uTotalVertices];
+ XResourceModelStaticVertexGPU *pVertices = new XResourceModelStaticVertexGPU[uTotalVertices];
UINT *pIndices = new UINT[uTotalIndices];
subset_t subset;
@@ -174,7 +174,25 @@ bool CDynamicModelShared::init(IXResourceModelStatic *pResource)
subset.uIndexCount = pSubset->iIndexCount;
subset.uVertexCount = pSubset->iVertexCount;
- memcpy(pVertices + subset.uStartVertex, pSubset->pVertices, sizeof(XResourceModelStaticVertex) * pSubset->iVertexCount);
+ //memcpy(pVertices + subset.uStartVertex, pSubset->pVertices, sizeof(XResourceModelStaticVertex) * pSubset->iVertexCount);
+ for(UINT k = 0; k < pSubset->iVertexCount; ++k)
+ {
+#define TO_SHORT(v) ((short)((v) * 32767.0f))
+ auto &dst = (pVertices + subset.uStartVertex)[k];
+ auto &src = pSubset->pVertices[k];
+ dst.vPos = src.vPos;
+ dst.vTex = src.vTex;
+ dst.vNorm[0] = TO_SHORT(src.vNorm.x);
+ dst.vNorm[1] = TO_SHORT(src.vNorm.y);
+ dst.vNorm[2] = TO_SHORT(src.vNorm.z);
+ dst.vTangent[0] = TO_SHORT(src.vTangent.x);
+ dst.vTangent[1] = TO_SHORT(src.vTangent.y);
+ dst.vTangent[2] = TO_SHORT(src.vTangent.z);
+ dst.vBinorm[0] = TO_SHORT(src.vBinorm.x);
+ dst.vBinorm[1] = TO_SHORT(src.vBinorm.y);
+ dst.vBinorm[2] = TO_SHORT(src.vBinorm.z);
+#undef TO_SHORT
+ }
memcpy(pIndices + subset.uStartIndex, pSubset->pIndices, sizeof(UINT) * pSubset->iIndexCount);
m_aLods[i][uMaterial] = subset;
@@ -204,7 +222,7 @@ bool CDynamicModelShared::init(IXResourceModelStatic *pResource)
if(m_pProvider->getCore()->isOnMainThread())
{
m_ppIndexBuffer[i] = m_pDevice->createIndexBuffer(sizeof(UINT) * uTotalIndices, GXBUFFER_USAGE_STATIC, GXIT_UINT32, pIndices);
- IGXVertexBuffer *pVertexBuffer = m_pDevice->createVertexBuffer(sizeof(XResourceModelStaticVertex) * uTotalVertices, GXBUFFER_USAGE_STATIC, pVertices);
+ IGXVertexBuffer *pVertexBuffer = m_pDevice->createVertexBuffer(sizeof(XResourceModelStaticVertexGPU) * uTotalVertices, GXBUFFER_USAGE_STATIC, pVertices);
m_ppRenderBuffer[i] = m_pDevice->createRenderBuffer(1, &pVertexBuffer, m_pProvider->getVertexDeclaration());
mem_release(pVertexBuffer);
@@ -358,7 +376,7 @@ void CDynamicModelShared::initGPUresources()
for(UINT i = 0, l = m_aLods.size(); i < l; ++i)
{
m_ppIndexBuffer[i] = m_pDevice->createIndexBuffer(sizeof(UINT) * m_puTempTotalIndices[i], GXBUFFER_USAGE_STATIC, GXIT_UINT32, m_ppTempIndices[i]);
- IGXVertexBuffer *pVertexBuffer = m_pDevice->createVertexBuffer(sizeof(XResourceModelStaticVertex) * m_puTempTotalVertices[i], GXBUFFER_USAGE_STATIC, m_ppTempVertices[i]);
+ IGXVertexBuffer *pVertexBuffer = m_pDevice->createVertexBuffer(sizeof(XResourceModelStaticVertexGPU) * m_puTempTotalVertices[i], GXBUFFER_USAGE_STATIC, m_ppTempVertices[i]);
m_ppRenderBuffer[i] = m_pDevice->createRenderBuffer(1, &pVertexBuffer, m_pProvider->getVertexDeclaration());
mem_release(pVertexBuffer);
diff --git a/source/anim/DynamicModelShared.h b/source/anim/DynamicModelShared.h
index 0fba2c613..92ee37785 100644
--- a/source/anim/DynamicModelShared.h
+++ b/source/anim/DynamicModelShared.h
@@ -63,7 +63,7 @@ protected:
IGXIndexBuffer **m_ppIndexBuffer = NULL;
UINT **m_ppTempIndices = NULL;
- XResourceModelStaticVertex **m_ppTempVertices = NULL;
+ XResourceModelStaticVertexGPU **m_ppTempVertices = NULL;
UINT *m_puTempTotalIndices = 0;
UINT *m_puTempTotalVertices = 0;
diff --git a/source/xcommon/resource/IXResourceModel.h b/source/xcommon/resource/IXResourceModel.h
index 0ed5a3925..b8f83caa9 100644
--- a/source/xcommon/resource/IXResourceModel.h
+++ b/source/xcommon/resource/IXResourceModel.h
@@ -98,6 +98,15 @@ struct XResourceModelStaticVertex
float3_t vBinorm; /*!< Бинормаль */
};
+struct XResourceModelStaticVertexGPU
+{
+ float3_t vPos; /*!< Позиция */
+ float2_t vTex; /*!< Текстурные координаты */
+ short vNorm[4]; /*!< Нормаль */
+ short vTangent[4]; /*!< Тангент */
+ short vBinorm[4]; /*!< Бинормаль */
+};
+
struct XResourceModelStaticSubset
{
uint32_t iMaterialID = 0; //!< Идентификатор материала
@@ -118,6 +127,17 @@ struct XResourceModelAnimatedVertex
float4_t vBoneWeights; /*!< Веса костей */
};
+struct XResourceModelAnimatedVertexGPU
+{
+ float3_t vPos; /*!< Позиция */
+ float2_t vTex; /*!< Текстурные координаты */
+ short vNorm[4]; /*!< Нормаль */
+ short vTangent[4]; /*!< Тангент */
+ short vBinorm[4]; /*!< Бинормаль */
+ byte u8BoneIndices[4]; /*!< Индексы костей */
+ byte vBoneWeights[4]; /*!< Веса костей */
+};
+
struct XResourceModelAnimatedSubset
{
uint32_t iMaterialID = 0; //!< Идентификатор материала
--
GitLab