fp16 optimizations (#985)

WickedEngine/shaders/ColorSpaceUtility.hlsli

@@ -34,41 +34,25 @@
 // are--the sRGB curve needs to be removed before involving the colors in linear mathematics such
 // as physically based lighting.
 
-float3 ApplySRGBCurve( float3 x )
-{
-    // Approximately pow(x, 1.0 / 2.2)
-    return select(x < 0.0031308, 12.92 * x, 1.055 * pow(x, 1.0 / 2.4) - 0.055);
-}
+// Note: modified for Wicked Engine to use macros, for better half precision mapping
 
-float3 RemoveSRGBCurve( float3 x )
-{
-    // Approximately pow(x, 2.2)
-	return select(x < 0.04045, x / 12.92, pow((x + 0.055) / 1.055, 2.4));
-}
+// Approximately pow(x, 1.0 / 2.2)
+#define ApplySRGBCurve( x ) select(x < 0.0031308, 12.92 * x, 1.055 * pow(x, 1.0 / 2.4) - 0.055)
+
+// Approximately pow(x, 2.2)
+#define RemoveSRGBCurve( x ) select(x < 0.04045, x / 12.92, pow((x + 0.055) / 1.055, 2.4))
 
 // These functions avoid pow() to efficiently approximate sRGB with an error < 0.4%.
-float3 ApplySRGBCurve_Fast( float3 x )
-{
-	return select(x < 0.0031308, 12.92 * x, 1.13005 * sqrt(x - 0.00228) - 0.13448 * x + 0.005719);
-}
+#define ApplySRGBCurve_Fast( x ) select(x < 0.0031308, 12.92 * x, 1.13005 * sqrt(x - 0.00228) - 0.13448 * x + 0.005719)
 
-float3 RemoveSRGBCurve_Fast( float3 x )
-{
-	return select(x < 0.04045, x / 12.92, -7.43605 * x - 31.24297 * sqrt(-0.53792 * x + 1.279924) + 35.34864);
-}
+#define RemoveSRGBCurve_Fast( x ) select(x < 0.04045, x / 12.92, -7.43605 * x - 31.24297 * sqrt(-0.53792 * x + 1.279924) + 35.34864)
 
 // The OETF recommended for content shown on HDTVs.  This "gamma ramp" may increase contrast as
 // appropriate for viewing in a dark environment.  Always use this curve with Limited RGB as it is
 // used in conjunction with HDTVs.
-float3 ApplyREC709Curve( float3 x )
-{
-	return select(x < 0.0181, 4.5 * x, 1.0993 * pow(x, 0.45) - 0.0993);
-}
+#define ApplyREC709Curve( x ) select(x < 0.0181, 4.5 * x, 1.0993 * pow(x, 0.45) - 0.0993)
 
-float3 RemoveREC709Curve( float3 x )
-{
-	return select(x < 0.08145, x / 4.5, pow((x + 0.0993) / 1.0993, 1.0 / 0.45));
-}
+#define RemoveREC709Curve( x ) select(x < 0.08145, x / 4.5, pow((x + 0.0993) / 1.0993, 1.0 / 0.45))
 
 // This is the new HDR transfer function, also called "PQ" for perceptual quantizer.  Note that REC2084
 // does not also refer to a color space.  REC2084 is typically used with the REC2020 color space.

WickedEngine/shaders/ShaderInterop_DDGI.h

@@ -230,7 +230,7 @@ half3 ddgi_sample_irradiance(float3 P, half3 N)
 			half dist_to_probe = length(probe_to_point);
 
 			//float2 temp = textureLod(depth_texture, tex_coord, 0.0f).rg;
-			half2 temp = bindless_textures[GetScene().ddgi.depth_texture].SampleLevel(sampler_linear_clamp, tex_coord, 0).xy;
+			half2 temp = bindless_textures_half4[GetScene().ddgi.depth_texture].SampleLevel(sampler_linear_clamp, tex_coord, 0).xy;
 			half mean = temp.x;
 			half variance = abs(sqr(temp.x) - temp.y);
 
@@ -254,7 +254,7 @@ half3 ddgi_sample_irradiance(float3 P, half3 N)
 		float2 tex_coord = ddgi_probe_color_uv(probe_grid_coord, irradiance_dir);
 
 		//float3 probe_irradiance = textureLod(irradiance_texture, tex_coord, 0.0f).rgb;
-		half3 probe_irradiance = bindless_textures[GetScene().ddgi.color_texture].SampleLevel(sampler_linear_clamp, tex_coord, 0).rgb;
+		half3 probe_irradiance = bindless_textures_half4[GetScene().ddgi.color_texture].SampleLevel(sampler_linear_clamp, tex_coord, 0).rgb;
 
 		// A tiny bit of light is really visible due to log perception, so
 		// crush tiny weights but keep the curve continuous. This must be done

WickedEngine/shaders/ShaderInterop_Font.h

@@ -26,12 +26,8 @@ struct FontConstants
 	int texture_index;
 	int padding0;
 
-	float4 color;
-
-	float softness;
-	float bolden;
-	uint flags;
-	float hdr_scaling;
+	uint2 color; // packed half4
+	uint2 softness_bolden_hdrscaling; // packed half3 | uint16 flags
 
 	float4x4 transform;
 };

WickedEngine/shaders/ShaderInterop_Renderer.h

@@ -195,12 +195,12 @@ struct alignas(16) ShaderTextureSlot
 	{
 		return f16tof32((uvset_aniso_lodclamp >> 16u) & 0xFFFF);
 	}
-	Texture2D GetTexture()
+	Texture2D<half4> GetTexture()
 	{
-		return bindless_textures[UniformTextureSlot(texture_descriptor)];
+		return bindless_textures_half4[UniformTextureSlot(texture_descriptor)];
 	}
-	float4 SampleVirtual(
-		in Texture2D tex,
+	half4 SampleVirtual(
+		in Texture2D<half4> tex,
 		in SamplerState sam,
 		in float2 uv,
 		in Texture2D<uint4> residency_map,
@@ -235,7 +235,7 @@ struct alignas(16) ShaderTextureSlot
 		const float clamped_lod = virtual_lod < max_nonpacked_lod ? max(virtual_lod, residency.z) : virtual_lod;
 
 		// Mip - more detailed:
-		float4 value0;
+		half4 value0;
 		{
 			uint lod0 = uint(clamped_lod);
 			const uint packed_mip_idx = packed_mips ? uint(virtual_lod - max_nonpacked_lod - 1) : 0;
@@ -249,7 +249,7 @@ struct alignas(16) ShaderTextureSlot
 		}
 
 		// Mip - less detailed:
-		float4 value1;
+		half4 value1;
 		{
 			uint lod1 = uint(clamped_lod + 1);
 			packed_mips = uint(lod1) > max_nonpacked_lod;
@@ -265,11 +265,11 @@ struct alignas(16) ShaderTextureSlot
 			value1 = tex.SampleLevel(sam, atlas_uv, 0);
 		}
 
-		return lerp(value0, value1, frac(clamped_lod)); // custom trilinear filtering
+		return lerp(value0, value1, (half)frac(clamped_lod)); // custom trilinear filtering
 	}
-	float4 Sample(in SamplerState sam, in float4 uvsets)
+	half4 Sample(in SamplerState sam, in float4 uvsets)
 	{
-		Texture2D tex = GetTexture();
+		Texture2D<half4> tex = GetTexture();
 		float2 uv = GetUVSet() == 0 ? uvsets.xy : uvsets.zw;
 
 #ifndef DISABLE_SVT
@@ -288,9 +288,9 @@ struct alignas(16) ShaderTextureSlot
 		return tex.Sample(sam, uv);
 	}
 
-	float4 SampleLevel(in SamplerState sam, in float4 uvsets, in float lod)
+	half4 SampleLevel(in SamplerState sam, in float4 uvsets, in float lod)
 	{
-		Texture2D tex = GetTexture();
+		Texture2D<half4> tex = GetTexture();
 		float2 uv = GetUVSet() == 0 ? uvsets.xy : uvsets.zw;
 
 #ifndef DISABLE_SVT
@@ -308,9 +308,9 @@ struct alignas(16) ShaderTextureSlot
 		return tex.SampleLevel(sam, uv, lod);
 	}
 
-	float4 SampleBias(in SamplerState sam, in float4 uvsets, in float bias)
+	half4 SampleBias(in SamplerState sam, in float4 uvsets, in float bias)
 	{
-		Texture2D tex = GetTexture();
+		Texture2D<half4> tex = GetTexture();
 		float2 uv = GetUVSet() == 0 ? uvsets.xy : uvsets.zw;
 
 #ifndef DISABLE_SVT
@@ -330,9 +330,9 @@ struct alignas(16) ShaderTextureSlot
 		return tex.SampleBias(sam, uv, bias);
 	}
 
-	float4 SampleGrad(in SamplerState sam, in float4 uvsets, in float4 uvsets_dx, in float4 uvsets_dy)
+	half4 SampleGrad(in SamplerState sam, in float4 uvsets, in float4 uvsets_dx, in float4 uvsets_dy)
 	{
-		Texture2D tex = GetTexture();
+		Texture2D<half4> tex = GetTexture();
 		float2 uv = GetUVSet() == 0 ? uvsets.xy : uvsets.zw;
 		float2 uv_dx = GetUVSet() == 0 ? uvsets_dx.xy : uvsets_dx.zw;
 		float2 uv_dy = GetUVSet() == 0 ? uvsets_dy.xy : uvsets_dy.zw;

WickedEngine/shaders/fontPS.hlsl

@@ -8,21 +8,27 @@ struct VertextoPixel
 	float2 bary : TEXCOORD1;
 };
 
-float4 main(VertextoPixel input) : SV_TARGET
+half4 main(VertextoPixel input) : SV_TARGET
 {
-	Texture2D tex = bindless_textures[font.texture_index];
+	Texture2D<half4> tex = bindless_textures_half4[font.texture_index];
 	half value = tex.SampleLevel(sampler_linear_clamp, input.uv, 0).r;
-	half4 color = font.color;
+	half4 color = unpack_half4(font.color);
+
+	const half3 softness_bolden_hdrscaling = unpack_half3(font.softness_bolden_hdrscaling);
+	const half softness = softness_bolden_hdrscaling.x;
+	const half bolden = softness_bolden_hdrscaling.y;
+	const half hdr_scaling = softness_bolden_hdrscaling.z;
+	const min16uint flags = font.softness_bolden_hdrscaling.y >> 16u;
 
 	[branch]
-	if (font.flags & FONT_FLAG_SDF_RENDERING)
+	if (flags & FONT_FLAG_SDF_RENDERING)
 	{
 		float2 bary_fw = fwidth(input.bary);
-		float w = max(bary_fw.x, bary_fw.y); // screen coverage dependency
+		half w = max(bary_fw.x, bary_fw.y); // screen coverage dependency
 		w = max(w, 1.0 / 255.0); // min softness to avoid pixelated hard edge in magnification
-		w += font.softness;
+		w += softness;
 		w = saturate(w);
-		half mid = lerp(SDF::onedge_value_unorm, 0, font.bolden);
+		half mid = lerp((half)SDF::onedge_value_unorm, 0, bolden);
 		color.a *= smoothstep(saturate(mid - w), saturate(mid + w), value);
 	}
 	else
@@ -31,7 +37,7 @@ float4 main(VertextoPixel input) : SV_TARGET
 	}
 
 	[branch]
-	if (font.flags & FONT_FLAG_OUTPUT_COLOR_SPACE_HDR10_ST2084)
+	if (flags & FONT_FLAG_OUTPUT_COLOR_SPACE_HDR10_ST2084)
 	{
 		// https://github.com/microsoft/DirectX-Graphics-Samples/blob/master/Samples/Desktop/D3D12HDR/src/presentPS.hlsl
 		const half referenceWhiteNits = 80.0;
@@ -42,10 +48,10 @@ float4 main(VertextoPixel input) : SV_TARGET
 		// Apply the ST.2084 curve to the result.
 		color.rgb = ApplyREC2084Curve(color.rgb * hdrScalar);
 	}
-	else if (font.flags & FONT_FLAG_OUTPUT_COLOR_SPACE_LINEAR)
+	else if (flags & FONT_FLAG_OUTPUT_COLOR_SPACE_LINEAR)
 	{
 		color.rgb = RemoveSRGBCurve_Fast(color.rgb);
-		color.rgb *= font.hdr_scaling;
+		color.rgb *= hdr_scaling;
 	}
 
 	return color;