Merge pull request #98327 from clayjohn/revert-shader-float

Revert "Fix GPUParticles are not rendered for older AMD GPUs with OpenGL+Angle"

4 files changed, 41 insertions, 19 deletions

diff --git a/drivers/gles3/shader_gles3.cpp b/drivers/gles3/shader_gles3.cpp
index b73debf04a..83f74865f6 100644
--- a/drivers/gles3/shader_gles3.cpp
+++ b/drivers/gles3/shader_gles3.cpp
@@ -168,6 +168,10 @@ void ShaderGLES3::_build_variant_code(StringBuilder &builder, uint32_t p_variant
 		builder.append("#version 300 es\n");
 	}
 
+	if (GLES3::Config::get_singleton()->polyfill_half2float) {
+		builder.append("#define USE_HALF2FLOAT\n");
+	}
+
 	for (int i = 0; i < specialization_count; i++) {
 		if (p_specialization & (uint64_t(1) << uint64_t(i))) {
 			builder.append("#define " + String(specializations[i].name) + "\n");
diff --git a/drivers/gles3/shaders/stdlib_inc.glsl b/drivers/gles3/shaders/stdlib_inc.glsl
index f88c218506..004db9e5d0 100644
--- a/drivers/gles3/shaders/stdlib_inc.glsl
+++ b/drivers/gles3/shaders/stdlib_inc.glsl
@@ -1,25 +1,28 @@
-
 // Compatibility renames. These are exposed with the "godot_" prefix
 // to work around two distinct Adreno bugs:
 // 1. Some Adreno devices expose ES310 functions in ES300 shaders.
 //    Internally, we must use the "godot_" prefix, but user shaders
 //    will be mapped automatically.
 // 2. Adreno 3XX devices have poor implementations of the other packing
-//    functions, so we just use our own everywhere to keep it simple.
+//    functions, so we just use our own there to keep it simple.
 
+#ifdef USE_HALF2FLOAT
 // Floating point pack/unpack functions are part of the GLSL ES 300 specification used by web and mobile.
+// It appears to be safe to expose these on mobile, but when running through ANGLE this appears to break.
 uint float2half(uint f) {
-	uint b = f + uint(0x00001000);
-	uint e = (b & uint(0x7F800000)) >> 23;
-	uint m = b & uint(0x007FFFFF);
-	return (b & uint(0x80000000)) >> uint(16) | uint(e > uint(112)) * ((((e - uint(112)) << uint(10)) & uint(0x7C00)) | m >> uint(13)) | (uint(e < uint(113)) & uint(e > uint(101))) * ((((uint(0x007FF000) + m) >> (uint(125) - e)) + uint(1)) >> uint(1)) | uint(e > uint(143)) * uint(0x7FFF);
+	uint e = f & uint(0x7f800000);
+	if (e <= uint(0x38000000)) {
+		return uint(0);
+	} else {
+		return ((f >> uint(16)) & uint(0x8000)) |
+				(((e - uint(0x38000000)) >> uint(13)) & uint(0x7c00)) |
+				((f >> uint(13)) & uint(0x03ff));
+	}
 }
 
 uint half2float(uint h) {
-	uint e = (h & uint(0x7C00)) >> uint(10);
-	uint m = (h & uint(0x03FF)) << uint(13);
-	uint v = m >> uint(23);
-	return (h & uint(0x8000)) << uint(16) | uint(e != uint(0)) * ((e + uint(112)) << uint(23) | m) | (uint(e == uint(0)) & uint(m != uint(0))) * ((v - uint(37)) << uint(23) | ((m << (uint(150) - v)) & uint(0x007FE000)));
+	uint h_e = h & uint(0x7c00);
+	return ((h & uint(0x8000)) << uint(16)) | uint((h_e >> uint(10)) != uint(0)) * (((h_e + uint(0x1c000)) << uint(13)) | ((h & uint(0x03ff)) << uint(13)));
 }
 
 uint godot_packHalf2x16(vec2 v) {
@@ -50,6 +53,17 @@ vec2 godot_unpackSnorm2x16(uint p) {
 	return clamp((v - 32767.0) * vec2(0.00003051851), vec2(-1.0), vec2(1.0));
 }
 
+#define packHalf2x16 godot_packHalf2x16
+#define unpackHalf2x16 godot_unpackHalf2x16
+#define packUnorm2x16 godot_packUnorm2x16
+#define unpackUnorm2x16 godot_unpackUnorm2x16
+#define packSnorm2x16 godot_packSnorm2x16
+#define unpackSnorm2x16 godot_unpackSnorm2x16
+
+#endif // USE_HALF2FLOAT
+
+// Always expose these as they are ES310 functions and not available in ES300 or GLSL 330.
+
 uint godot_packUnorm4x8(vec4 v) {
 	uvec4 uv = uvec4(round(clamp(v, vec4(0.0), vec4(1.0)) * 255.0));
 	return uv.x | (uv.y << uint(8)) | (uv.z << uint(16)) | (uv.w << uint(24));
@@ -73,9 +87,3 @@ vec4 godot_unpackSnorm4x8(uint p) {
 #define unpackUnorm4x8 godot_unpackUnorm4x8
 #define packSnorm4x8 godot_packSnorm4x8
 #define unpackSnorm4x8 godot_unpackSnorm4x8
-#define packHalf2x16 godot_packHalf2x16
-#define unpackHalf2x16 godot_unpackHalf2x16
-#define packUnorm2x16 godot_packUnorm2x16
-#define unpackUnorm2x16 godot_unpackUnorm2x16
-#define packSnorm2x16 godot_packSnorm2x16
-#define unpackSnorm2x16 godot_unpackSnorm2x16
diff --git a/drivers/gles3/storage/config.cpp b/drivers/gles3/storage/config.cpp
index 209b7dd7d2..07dd5f58c1 100644
--- a/drivers/gles3/storage/config.cpp
+++ b/drivers/gles3/storage/config.cpp
@@ -231,6 +231,13 @@ Config::Config() {
 	} else if (rendering_device_name == "PowerVR Rogue GE8320") {
 		disable_transform_feedback_shader_cache = true;
 	}
+
+	if (OS::get_singleton()->get_current_rendering_driver_name() == "opengl3_angle") {
+		polyfill_half2float = false;
+	}
+#ifdef WEB_ENABLED
+	polyfill_half2float = false;
+#endif
 }
 
 Config::~Config() {
diff --git a/drivers/gles3/storage/config.h b/drivers/gles3/storage/config.h
index 06d88429be..51f6714320 100644
--- a/drivers/gles3/storage/config.h
+++ b/drivers/gles3/storage/config.h
@@ -95,13 +95,16 @@ public:
 	bool multiview_supported = false;
 	bool external_texture_supported = false;
 
-	// Adreno 3XX compatibility
-	bool disable_particles_workaround = false; // set to 'true' to disable 'GPUParticles'
+	// Adreno 3XX compatibility.
+	bool disable_particles_workaround = false; // Set to 'true' to disable 'GPUParticles'.
 	bool flip_xy_workaround = false;
 
-	// PowerVR GE 8320 workaround
+	// PowerVR GE 8320 workaround.
 	bool disable_transform_feedback_shader_cache = false;
 
+	// ANGLE shader workaround.
+	bool polyfill_half2float = true;
+
 #ifdef ANDROID_ENABLED
 	PFNGLFRAMEBUFFERTEXTUREMULTIVIEWOVRPROC eglFramebufferTextureMultiviewOVR = nullptr;
 	PFNGLTEXSTORAGE3DMULTISAMPLEPROC eglTexStorage3DMultisample = nullptr;