blut shader improvements, use bilinear sampling for 2x fewer samples

Author: mmozeiko

src/render/d3d11/generated/render_d3d11.meta.h

@@ -205,10 +205,14 @@ str8_lit_comp(
 "cbuffer Globals : register(b0)\n"
 "{\n"
 "  float4 rect;\n"
-"  float2 viewport_size;\n"
-"  float blur_size;\n"
-"  float is_vertical;\n"
 "  float4 corner_radii_px;\n"
+"  float2 direction;\n"
+"  float2 viewport_size;\n"
+"  uint blur_count;\n"
+"}\n"
+"\n"
+"cbuffer Kernel : register(b1)\n"
+"{\n"
 "  float4 kernel[32];\n"
 "}\n"
 "\n"
@@ -221,7 +225,7 @@ str8_lit_comp(
 "{\n"
 "  float4 position        : SV_POSITION;\n"
 "  float2 texcoord        : TEX;\n"
-"  float2 cornercoord     : CRN;\n"
+"  float2 sdf_sample_pos  : SDF;\n"
 "  float corner_radius    : RAD;\n"
 "};\n"
 "\n"
@@ -238,12 +242,12 @@ str8_lit_comp(
 "Vertex2Pixel\n"
 "vs_main(CPU2Vertex c2v)\n"
 "{\n"
-"  float4 vertex_positions__scrn[] =\n"
+"  float2 vertex_positions__scrn[] =\n"
 "  {\n"
-"    float4(rect.x, rect.w, 0, 1) * float4(1, -1, 1, 1) + float4(0, viewport_size.y, 0, 0),\n"
-"    float4(rect.x, rect.y, 0, 1) * float4(1, -1, 1, 1) + float4(0, viewport_size.y, 0, 0),\n"
-"    float4(rect.z, rect.w, 0, 1) * float4(1, -1, 1, 1) + float4(0, viewport_size.y, 0, 0),\n"
-"    float4(rect.z, rect.y, 0, 1) * float4(1, -1, 1, 1) + float4(0, viewport_size.y, 0, 0),\n"
+"    rect.xw,\n"
+"    rect.xy,\n"
+"    rect.zw,\n"
+"    rect.zy,\n"
 "  };\n"
 "  float corner_radii__px[] =\n"
 "  {\n"
@@ -252,22 +256,20 @@ str8_lit_comp(
 "    corner_radii_px.w,\n"
 "    corner_radii_px.z,\n"
 "  };\n"
-"  float2 cornercoords__pct[] =\n"
-"  {\n"
-"    float2(0, 1),\n"
-"    float2(0, 0),\n"
-"    float2(1, 1),\n"
-"    float2(1, 0),\n"
-"  };\n"
-"  float4 vertex_position__scrn = vertex_positions__scrn[c2v.vertex_id];\n"
-"  float4 vertex_position__clip = float4(2*vertex_position__scrn.x/viewport_size.x - 1,\n"
-"                                        2*vertex_position__scrn.y/viewport_size.y - 1,\n"
-"                                        0, 1);\n"
+"  float2 cornercoords__pct = float2(\n"
+"    (c2v.vertex_id >> 1) ? 1.f : 0.f,\n"
+"    (c2v.vertex_id & 1)  ? 0.f : 1.f);\n"
+"\n"
+"  float2 vertex_position__pct = vertex_positions__scrn[c2v.vertex_id] / viewport_size;\n"
+"  float2 vertex_position__scr = 2.f * vertex_position__pct - 1.f;\n"
+"\n"
+"  float2 rect_half_size = float2((rect.z-rect.x)/2, (rect.w-rect.y)/2);\n"
+"\n"
 "  Vertex2Pixel v2p;\n"
 "  {\n"
-"    v2p.position = vertex_position__clip;\n"
-"    v2p.texcoord = float2(vertex_position__scrn.x/viewport_size.x, 1 - vertex_position__scrn.y/viewport_size.y);\n"
-"    v2p.cornercoord = cornercoords__pct[c2v.vertex_id];\n"
+"    v2p.position = float4(vertex_position__scr.x, -vertex_position__scr.y, 0.f, 1.f);\n"
+"    v2p.texcoord = vertex_position__pct;\n"
+"    v2p.sdf_sample_pos = (2.f * cornercoords__pct - 1.f) * rect_half_size;\n"
 "    v2p.corner_radius = corner_radii__px[c2v.vertex_id];\n"
 "  }\n"
 "  return v2p;\n"
@@ -279,26 +281,27 @@ str8_lit_comp(
 "ps_main(Vertex2Pixel v2p) : SV_TARGET\n"
 "{\n"
 "  // rjf: blend weighted texture samples into color\n"
-"  float4 color = stage_t2d.Sample(stage_sampler, v2p.texcoord) * kernel[0].x;\n"
+"  float4 color = kernel[0].x * stage_t2d.Sample(stage_sampler, v2p.texcoord);\n"
 "  color.a = kernel[0].x;\n"
-"  for(float i = 1; i < blur_size; i += 1)\n"
+"\n"
+"  for(uint i = 1; i < blur_count; i += 1)\n"
 "  {\n"
-"    float weight = ((float[4])kernel[uint(i)/4])[uint(i)%4];\n"
-"    float4 min_sample = stage_t2d.Sample(stage_sampler, v2p.texcoord - float2(!is_vertical*i/viewport_size.x, is_vertical*i/viewport_size.y));\n"
-"    float4 max_sample = stage_t2d.Sample(stage_sampler, v2p.texcoord + float2(!is_vertical*i/viewport_size.x, is_vertical*i/viewport_size.y));\n"
-"    min_sample.a = 1;\n"
-"    max_sample.a = 1;\n"
-"    color += min_sample*weight;\n"
-"    color += max_sample*weight;\n"
+"    float weight = kernel[i].x;\n"
+"    float offset = kernel[i].y;\n"
+"    float4 min_sample = stage_t2d.Sample(stage_sampler, v2p.texcoord - offset * direction);\n"
+"    float4 max_sample = stage_t2d.Sample(stage_sampler, v2p.texcoord + offset * direction);\n"
+"    min_sample.a = 1.f;\n"
+"    max_sample.a = 1.f;\n"
+"    color += min_sample * weight;\n"
+"    color += max_sample * weight;\n"
 "  }\n"
 "  \n"
 "  // rjf: determine SDF sample position\n"
 "  float2 rect_half_size = float2((rect.z-rect.x)/2, (rect.w-rect.y)/2);\n"
-"  float2 sdf_sample_pos = float2((2*v2p.cornercoord.x-1)*rect_half_size.x,\n"
-"                                 (2*v2p.cornercoord.y-1)*rect_half_size.y);\n"
+"  float2 sdf_sample_pos = v2p.sdf_sample_pos;\n"
 "  \n"
 "  // rjf: sample for corners\n"
-"  float corner_sdf_s = rect_sdf(sdf_sample_pos, rect_half_size - float2(2.f, 2.f), v2p.corner_radius);\n"
+"  float corner_sdf_s = rect_sdf(sdf_sample_pos, rect_half_size - 2.f, v2p.corner_radius);\n"
 "  float corner_sdf_t = 1-smoothstep(0, 2, corner_sdf_s);\n"
 "  \n"
 "  // rjf: weight output color by sdf\n"

src/render/d3d11/render_d3d11.cpp

@@ -1183,99 +1183,140 @@ r_window_submit(OS_Handle window, R_Handle window_equip, R_PassList *passes)
         case R_PassKind_Blur:
         {
           R_PassParams_Blur *params = pass->params_blur;
-          ID3D11SamplerState *sampler   = r_d3d11_state->samplers[R_Tex2DSampleKind_Nearest];
+          ID3D11SamplerState *sampler   = r_d3d11_state->samplers[R_Tex2DSampleKind_Linear];
           ID3D11VertexShader *vshad     = r_d3d11_state->vshads[R_D3D11_VShadKind_Blur];
           ID3D11PixelShader *pshad      = r_d3d11_state->pshads[R_D3D11_PShadKind_Blur];
           ID3D11Buffer *uniforms_buffer = r_d3d11_state->uniform_type_kind_buffers[R_D3D11_VShadKind_Blur];
-          
-          //- rjf: perform blur on each axis
-          ID3D11RenderTargetView *rtvs[Axis2_COUNT] =
-          {
-            wnd->stage_scratch_color_rtv,
-            wnd->stage_color_rtv,
-          };
-          ID3D11ShaderResourceView *srvs[Axis2_COUNT] =
+
+          // rjf: setup output merger
+          d_ctx->OMSetDepthStencilState(r_d3d11_state->noop_depth_stencil, 0);
+          d_ctx->OMSetBlendState(r_d3d11_state->main_blend_state, 0, 0xffffffff);
+
+          // rjf: set up viewport
+          Vec2S32 resolution = wnd->last_resolution;
+          D3D11_VIEWPORT viewport = { 0.0f, 0.0f, (F32)resolution.x, (F32)resolution.y, 0.0f, 1.0f };
+          d_ctx->RSSetViewports(1, &viewport);
+          d_ctx->RSSetState(r_d3d11_state->main_rasterizer);
+
+          // rjf: setup input assembly
+          d_ctx->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+          d_ctx->IASetInputLayout(0);
+
+          // rjf: setup shaders
+          d_ctx->VSSetShader(vshad, 0, 0);
+          d_ctx->VSSetConstantBuffers(0, 1, &uniforms_buffer);
+          d_ctx->PSSetShader(pshad, 0, 0);
+          d_ctx->PSSetSamplers(0, 1, &sampler);
+
+          // rjf: setup scissor rect
           {
-            wnd->stage_color_srv,
-            wnd->stage_scratch_color_srv,
-          };
-          for(Axis2 axis = (Axis2)0; axis < Axis2_COUNT; axis = (Axis2)(axis+1))
+              D3D11_RECT rect = { 0 };
+              rect.left = 0;
+              rect.right = (LONG)wnd->last_resolution.x;
+              rect.top = 0;
+              rect.bottom = (LONG)wnd->last_resolution.y;
+              d_ctx->RSSetScissorRects(1, &rect);
+          }
+
+          // rjf: set up uniforms
+          R_D3D11_Uniforms_Blur uniforms = { 0 };
           {
-            // rjf: setup output merger
-            d_ctx->OMSetRenderTargets(1, &rtvs[axis], 0);
-            d_ctx->OMSetDepthStencilState(r_d3d11_state->noop_depth_stencil, 0);
-            d_ctx->OMSetBlendState(r_d3d11_state->main_blend_state, 0, 0xffffffff);
-            
-            // rjf: set up viewport
-            Vec2S32 resolution = wnd->last_resolution;
-            D3D11_VIEWPORT viewport = { 0.0f, 0.0f, (F32)resolution.x, (F32)resolution.y, 0.0f, 1.0f };
-            d_ctx->RSSetViewports(1, &viewport);
-            d_ctx->RSSetState(r_d3d11_state->main_rasterizer);
-            
-            // rjf: setup input assembly
-            d_ctx->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
-            d_ctx->IASetInputLayout(0);
-            
-            // rjf: set up uniforms
+            F32 weights[ArrayCount(uniforms.kernel)*2] = {0};
+
+            F32 blur_size = Min(params->blur_size, ArrayCount(weights));
+            U64 blur_count = (U64)round_f32(blur_size);
+
+            F32 stdev = (blur_size-1.f)/2.f;
+            F32 one_over_root_2pi_stdev2 = 1/sqrt_f32(2*pi32*stdev*stdev);
+            F32 euler32 = 2.718281828459045f;
+
+            weights[0] = 1.f;
+            if(stdev > 0.f)
             {
-              F32 stdev = (params->blur_size-1.f)/2.f;
-              F32 one_over_root_2pi_stdev2 = 1/sqrt_f32(2*pi32*stdev*stdev);
-              F32 euler32 = 2.718281828459045f;
-              R_D3D11_Uniforms_Blur uniforms = {0};
-              uniforms.viewport_size  = v2f32(resolution.x, resolution.y);
-              uniforms.rect           = params->rect;
-              uniforms.blur_size      = params->blur_size;
-              uniforms.is_vertical    = (F32)!!axis;
-              MemoryCopyArray(uniforms.corner_radii.v, params->corner_radii);
-              F32 kernel_x = 0;
-              uniforms.kernel[0].v[0] = 1.f;
-              if(stdev > 0.f)
-              {
-                for(U64 idx = 0; idx < ArrayCount(uniforms.kernel); idx += 1)
-                {
-                  for(U64 v_idx = 0; v_idx < ArrayCount(uniforms.kernel[idx].v); v_idx += 1)
-                  {
-                    uniforms.kernel[idx].v[v_idx] = one_over_root_2pi_stdev2*pow_f32(euler32, -kernel_x*kernel_x/(2.f*stdev*stdev)); 
-                    kernel_x += 1;
-                  }
-                }
-              }
-              if(uniforms.kernel[0].v[0] > 1.f)
+              for(U64 idx = 0; idx < blur_count; idx += 1)
               {
-                MemoryZeroArray(uniforms.kernel);
-                uniforms.kernel[0].v[0] = 1.f;
+                F32 kernel_x = (F32)idx;
+                weights[idx] = one_over_root_2pi_stdev2*pow_f32(euler32, -kernel_x*kernel_x/(2.f*stdev*stdev)); 
               }
-              D3D11_MAPPED_SUBRESOURCE sub_rsrc = {0};
-              r_d3d11_state->device_ctx->Map(uniforms_buffer, 0, D3D11_MAP_WRITE_DISCARD, 0, &sub_rsrc);
-              MemoryCopy((U8 *)sub_rsrc.pData, &uniforms, sizeof(uniforms));
-              r_d3d11_state->device_ctx->Unmap(uniforms_buffer, 0);
             }
-            
-            // rjf: setup shaders
-            d_ctx->VSSetShader(vshad, 0, 0);
-            d_ctx->VSSetConstantBuffers(0, 1, &uniforms_buffer);
-            d_ctx->PSSetShader(pshad, 0, 0);
-            d_ctx->PSSetConstantBuffers(0, 1, &uniforms_buffer);
-            d_ctx->PSSetShaderResources(0, 1, &srvs[axis]);
-            d_ctx->PSSetSamplers(0, 1, &sampler);
-            
-            // rjf: setup scissor rect
+            if(weights[0] > 1.f)
             {
-              D3D11_RECT rect = {0};
-              rect.left = 0;
-              rect.right = (LONG)wnd->last_resolution.x;
-              rect.top = 0;
-              rect.bottom = (LONG)wnd->last_resolution.y;
-              d_ctx->RSSetScissorRects(1, &rect);
+              MemoryZeroArray(weights);
+              weights[0] = 1.f;
             }
-            
-            // rjf: draw
-            d_ctx->Draw(4, 0);
-            
-            // rjf: unset srv
-            ID3D11ShaderResourceView *srv = 0;
-            d_ctx->PSSetShaderResources(0, 1, &srv);
+            else
+            {
+              // prepare weights & offsets for bilinear lookup
+              // blur filter wants to calculate w0*pixel[pos] + w1*pixel[pos+1] + ...
+              // with bilinear filter we can do this calulation by doing only w*sample(pos+t) = w*((1-t)*pixel[pos] + t*pixel[pos+1])
+              // we can see w0=w*(1-t) and w1=w*t
+              // thus w=w0+w1 and t=w1/w
+              for (U64 idx = 1; idx < blur_count; idx += 2)
+              {
+                  F32 w0 = weights[idx + 0];
+                  F32 w1 = weights[idx + 1];
+                  F32 w = w0 + w1;
+                  F32 t = w1 / w;
+
+                  // each kernel element is float2(weight, offset)
+                  // weights & offsets are adjusted for bilinear sampling
+                  // zw elements are not used, a bit of waste but it allows for simpler shader code
+                  uniforms.kernel[(idx+1)/2] = v4f32(w, (F32)idx + t, 0, 0);
+              }
+              uniforms.kernel[0].x = weights[0];
+            }
+
+            // technically we need just direction be different
+            // but there are 256 bytes of usable space anyway for each constant buffer chunk
+
+            uniforms.passes[Axis2_X].viewport_size = v2f32(resolution.x, resolution.y);
+            uniforms.passes[Axis2_X].rect          = params->rect;
+            uniforms.passes[Axis2_X].direction     = v2f32(1.f / resolution.x, 0);
+            uniforms.passes[Axis2_X].blur_count    = 1 + blur_count / 2; // 2x smaller because of bilinear sampling
+            MemoryCopyArray(uniforms.passes[Axis2_X].corner_radii.v, params->corner_radii);
+
+            uniforms.passes[Axis2_Y].viewport_size = v2f32(resolution.x, resolution.y);
+            uniforms.passes[Axis2_Y].rect          = params->rect;
+            uniforms.passes[Axis2_Y].direction     = v2f32(0, 1.f / resolution.y);
+            uniforms.passes[Axis2_Y].blur_count    = 1 + blur_count / 2; // 2x smaller because of bilinear sampling
+            MemoryCopyArray(uniforms.passes[Axis2_Y].corner_radii.v, params->corner_radii);
+
+            D3D11_MAPPED_SUBRESOURCE sub_rsrc = {0};
+            r_d3d11_state->device_ctx->Map(uniforms_buffer, 0, D3D11_MAP_WRITE_DISCARD, 0, &sub_rsrc);
+            MemoryCopy((U8 *)sub_rsrc.pData, &uniforms, sizeof(uniforms));
+            r_d3d11_state->device_ctx->Unmap(uniforms_buffer, 0);
           }
+
+          ID3D11Buffer* uniforms_buffers[] = { uniforms_buffer, uniforms_buffer };
+
+          U32 uniform_offset[Axis2_COUNT][2] =
+          {
+              { 0 * sizeof(R_D3D11_Uniforms_BlurPass) / 16, OffsetOf(R_D3D11_Uniforms_Blur, kernel) / 16 },
+              { 1 * sizeof(R_D3D11_Uniforms_BlurPass) / 16, OffsetOf(R_D3D11_Uniforms_Blur, kernel) / 16 },
+          };
+
+          U32 uniform_count[Axis2_COUNT][2] =
+          {
+              { sizeof(R_D3D11_Uniforms_BlurPass) / 16, sizeof(uniforms.kernel) / 16 },
+              { sizeof(R_D3D11_Uniforms_BlurPass) / 16, sizeof(uniforms.kernel) / 16 },
+          };
+
+          // rjf: for unsetting srv
+          ID3D11ShaderResourceView* srv = 0;
+
+          // horizontal pass
+          d_ctx->OMSetRenderTargets(1, &wnd->stage_scratch_color_rtv, 0);
+          d_ctx->PSSetConstantBuffers1(0, ArrayCount(uniforms_buffers), uniforms_buffers, uniform_offset[Axis2_X], uniform_count[Axis2_X]);
+          d_ctx->PSSetShaderResources(0, 1, &wnd->stage_color_srv);
+          d_ctx->Draw(4, 0);
+          d_ctx->PSSetShaderResources(0, 1, &srv);
+
+          // vertical pass
+          d_ctx->OMSetRenderTargets(1, &wnd->stage_color_rtv, 0);
+          d_ctx->PSSetConstantBuffers1(0, ArrayCount(uniforms_buffers), uniforms_buffers, uniform_offset[Axis2_Y], uniform_count[Axis2_Y]);
+          d_ctx->PSSetShaderResources(0, 1, &wnd->stage_scratch_color_srv);
+          d_ctx->Draw(4, 0);
+          d_ctx->PSSetShaderResources(0, 1, &srv);
         }break;
 
 

src/render/d3d11/render_d3d11.h

@@ -32,13 +32,20 @@ struct R_D3D11_Uniforms_Rect
   Vec2F32 xform_scale;
 };
 
-struct R_D3D11_Uniforms_Blur
+struct R_D3D11_Uniforms_BlurPass
 {
   Rng2F32 rect;
-  Vec2F32 viewport_size;
-  F32 blur_size;
-  F32 is_vertical;
   Vec4F32 corner_radii;
+  Vec2F32 direction;
+  Vec2F32 viewport_size;
+  U32 blur_count;
+  U8 _padding0_[204];
+};
+StaticAssert(sizeof(R_D3D11_Uniforms_BlurPass) % 256 == 0, NotAligned); // constant count/offset must be aligned to 256 bytes
+
+struct R_D3D11_Uniforms_Blur
+{
+  R_D3D11_Uniforms_BlurPass passes[Axis2_COUNT];
   Vec4F32 kernel[32];
 };