Combine OpenGL uniforms into batches

res/gamedata/shaders/gl/deffer_detail_s_flat.vs

@@ -1,18 +1,15 @@
 #include "common.h"
 #include "iostructs\v_detail.h"
 
-uniform float4 		consts; // {1/quant,1/quant,diffusescale,ambient}
-//uniform float4 		array	[200] : register(c12);
-//tbuffer DetailsData
-//{
-	uniform float4 		array[61*4];
-//}
+#include "shared\ubo_render_dump.h"
+
+uniform float4 		array[61*4];
 
 v2p_flat 	_main (v_detail v)
 {
 	v2p_flat 		O;
 	// index
-	int 	i 	= int(v.misc.w);
+	int 	i 	= gl_InstanceID*4;
 	float4  m0 	= array[i+0];
 	float4  m1 	= array[i+1];
 	float4  m2 	= array[i+2];

res/gamedata/shaders/gl/deffer_detail_w_flat.vs

@@ -1,20 +1,15 @@
 #include "common.h"
 #include "iostructs\v_detail.h"
 
-uniform float4 		consts; // {1/quant,1/quant,diffusescale,ambient}
-uniform float4 		wave; 	// cx,cy,cz,tm
-uniform float4 		dir2D; 
-//uniform float4 		array	[200] : register(c12);
-//tbuffer DetailsData
-//{
-	uniform float4 		array[61*4];
-//}
+#include "shared\ubo_render_dump.h"
+
+uniform float4 		array[61*4];
 
 v2p_flat 	_main (v_detail v)
 {
 	v2p_flat 		O;
 	// index
-	int 	i 	= int(v.misc.w);
+	int 	i 	= gl_InstanceID*4;
 	float4  m0 	= array[i+0];
 	float4  m1 	= array[i+1];
 	float4  m2 	= array[i+2];
@@ -33,7 +28,7 @@ v2p_flat 	_main (v_detail v)
 	float 	H 	= pos.y - base;			// height of vertex (scaled)
 	float 	frac 	= v.misc.z*consts.x;		// fractional
 	float 	inten 	= H * dp;
-	float2 	result	= calc_xz_wave	(dir2D.xz*inten,frac);
+	float2 	result	= calc_xz_wave	(wind.xz*inten,frac);
 	pos		= float4(pos.x+result.x, pos.y, pos.z+result.y, 1);
 
 	// Normal in world coords

res/gamedata/shaders/gl/shared/ubo_render_dump.h

@@ -0,0 +1,15 @@
+#ifndef OPENXRAY_UBO_RENDER_DUMP_H
+#define OPENXRAY_UBO_RENDER_DUMP_H
+
+layout(std140) uniform RenderDumpUBO {
+    float4x4 xform;
+    float4x4 formView;
+    float4 consts;
+    float4 scale;
+    float4 bias;
+    float4 wind;
+    float4 wave;
+    float4 sun;
+};
+
+#endif

src/Layers/xrRender/DetailManager.h

@@ -70,9 +70,9 @@ class ECORE_API CDetailManager
 
     struct SlotPart
     { //
-        u32 id; // ID модельки
-        SlotItemVec items; // список кустиков
-        SlotItemVec r_items[3]; // список кустиков for render
+        u32 id; // ID models
+        SlotItemVec items; // list of bushes
+        SlotItemVec r_items[3]; // list of bushes for render
     };
 
     enum SlotType : u32
@@ -185,19 +185,25 @@ class ECORE_API CDetailManager
     VertexStagingBuffer hw_VB;
     IndexStagingBuffer hw_IB;
 
+#ifndef USE_OGL
     ref_constant hwc_consts;
     ref_constant hwc_wave;
     ref_constant hwc_wind;
-    ref_constant hwc_array;
     ref_constant hwc_s_consts;
     ref_constant hwc_s_xform;
+#endif
+    ref_constant hwc_array;
     ref_constant hwc_s_array;
     void hw_Load();
     void hw_Load_Geom();
     void hw_Load_Shaders();
     void hw_Unload();
     void hw_Render(CBackend& cmd_list);
+#ifdef USE_OGL
+    void hw_Render_dump(CBackend& cmd_list, u32 var_id, u32 lod_id);
+#else
     void hw_Render_dump(CBackend& cmd_list, const Fvector4& consts, const Fvector4& wave, const Fvector4& wind, u32 var_id, u32 lod_id);
+#endif
 
     // get unpacked slot
     DetailSlot& QueryDB(int sx, int sz);
@@ -225,5 +231,16 @@ class ECORE_API CDetailManager
 
     CDetailManager();
     virtual ~CDetailManager();
+
+#ifdef USE_OGL
+private:
+    // TODO move to global registry ?
+    GLuint renderDumpBlockBinding = 3;
+    ShaderInstanceData renderDumpData[3];
+
+    UniformBufferObject renderDumpUBO = {
+        BUFFER_DEBUG_NAME("DetailManagerUBO") GL_NONE, GL_DYNAMIC_DRAW, sizeof(ShaderInstanceData)*3
+    };
+#endif
 };
 } // namespace xray::render::RENDER_NAMESPACE

src/Layers/xrRender/R_Backend.h

@@ -21,6 +21,10 @@
 #include "Layers/xrRenderDX11/StateManager/dx11ShaderResourceStateCache.h"
 #include "Layers/xrRenderDX11/StateManager/dx11StateCache.h"
 #endif
+#ifdef USE_OGL
+#include <glm/glm.hpp>
+#include "gl_backend_extension.h"
+#endif
 
 #include "FVF.h"
 
@@ -60,6 +64,9 @@ struct R_statistics
 #pragma warning(push)
 #pragma warning(disable:4324)
 class ECORE_API CBackend
+#ifdef USE_OGL
+    : public CBackendExtension
+#endif
 {
 public:
     enum
@@ -103,7 +110,6 @@ class ECORE_API CBackend
 #else
 #   error No graphics API selected or enabled!
 #endif
-
     // Vertices/Indices/etc
     SDeclaration* decl;
     VertexBufferHandle vb;
@@ -404,6 +410,7 @@ class ECORE_API CBackend
 public:
 #if defined(USE_OGL)
     ICF bool is_TessEnabled() { return false; }
+    void RenderInstanced(D3DPRIMITIVETYPE T, u32 baseV, u32 startV, u32 countV, u32 startI, u32 PC, GLsizei instanceCount);
 #elif defined(USE_DX11)
     ICF bool is_TessEnabled();
 #else

src/Layers/xrRender/R_Backend_Runtime.cpp

@@ -470,6 +470,10 @@ void CBackend::OnDeviceCreate()
 
     // invalidate caching
     Invalidate();
+
+#ifdef USE_OGL
+    initializeUniforms();
+#endif
 }
 
 void CBackend::OnDeviceDestroy()

src/Layers/xrRender/SH_Atomic.h

@@ -31,6 +31,7 @@ struct ECORE_API SVS : public xr_resource_named
     ID3DVertexShader* sh;
 #elif defined(USE_OGL)
     GLuint sh;
+    uint32_t bindingSlots = 0;
 #else
 #   error No graphics API selected or enabled!
 #endif

src/Layers/xrRender/SkeletonX.cpp

@@ -73,16 +73,30 @@ void CSkeletonX::_Render(CBackend& cmd_list, ref_geom& hGeom, u32 vCount, u32 iO
         // transfer matrices
         ref_constant array = cmd_list.get_c(s_bones_array_const);
         u32 count = RMS_bonecount;
+#ifdef USE_OGL
+        xr_vector<glm::vec4> uniformBuffer;
+        uniformBuffer.reserve(count*3);
+#endif
+
         for (u32 mid = 0; mid < count; mid++)
         {
             Fmatrix& M = Parent->LL_GetTransform_R(u16(mid));
+#ifdef USE_OGL
+            uniformBuffer.emplace_back(M._11, M._21, M._31, M._41);
+            uniformBuffer.emplace_back(M._12, M._22, M._32, M._42);
+            uniformBuffer.emplace_back(M._13, M._23, M._33, M._43);
+#else
             u32 id = mid * 3;
             cmd_list.set_ca(&*array, id + 0, M._11, M._21, M._31, M._41);
             cmd_list.set_ca(&*array, id + 1, M._12, M._22, M._32, M._42);
             cmd_list.set_ca(&*array, id + 2, M._13, M._23, M._33, M._43);
+#endif
         }
 
         // render
+#ifdef USE_OGL
+        cmd_list.setUniforms(array->vs.program, array->vs.location, uniformBuffer);
+#endif
         cmd_list.set_Geometry(hGeom);
         cmd_list.Render(D3DPT_TRIANGLELIST, 0, 0, vCount, iOffset, pCount);
         if (RM_SKINNING_1B == RenderMode || RM_SKINNING_1B_HQ == RenderMode)