Bugfixes, cleanup

GPU/Common/FragmentShaderGenerator.cpp

@@ -399,6 +399,13 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 			WRITE(p, "uniform float u_mipBias;\n");
 		}
 
+		if (fsMinmaxDiscard) {
+			*uniformMask |= DIRTY_RASTER_OFFSET;  // they're updated together
+			WRITE(p, "uniform vec2 u_minZmaxZ;\n");
+		}
+
+		// Varyings
+
 		WRITE(p, "%s %s lowp vec4 v_color0;\n", shading, compat.varying_fs);
 		if (lmode) {
 			WRITE(p, "%s %s lowp vec3 v_color1;\n", shading, compat.varying_fs);

GPU/Common/ShaderUniforms.cpp

@@ -6,6 +6,7 @@
 #include "Common/Data/Convert/SmallDataConvert.h"
 #include "Common/Math/lin/matrix4x4.h"
 #include "Common/Math/math_util.h"
+#include "Common/Math/CrossSIMD.h"
 #include "Common/Math/lin/vec3.h"
 #include "Common/TimeUtil.h"
 #include "GPU/GPUState.h"
@@ -87,8 +88,8 @@ void BaseUpdateUniforms(UB_VS_FS_Base *ub, uint64_t dirtyUniforms, bool useBuffe
 	if (dirtyUniforms & DIRTY_RASTER_OFFSET) {
 		ub->rasterOffset[0] = gstate.getOffsetX();
 		ub->rasterOffset[1] = gstate.getOffsetY();
-		ub->minZmaxZ[0] = gstate.getDepthRangeMin();
-		ub->minZmaxZ[1] = gstate.getDepthRangeMax();
+		ub->minZmaxZ[0] = (float)gstate.getDepthRangeMin();
+		ub->minZmaxZ[1] = (float)gstate.getDepthRangeMax();
 
 		// test sine wave
 		// ub->minZmaxZ[0] = (sin(time_now_d()) * 0.5f + 0.5f) * 65536.0;
@@ -96,17 +97,17 @@ void BaseUpdateUniforms(UB_VS_FS_Base *ub, uint64_t dirtyUniforms, bool useBuffe
 
 	if (dirtyUniforms & DIRTY_VIEWPORT_UNIFORMS) {
 		// TODO: This should be a couple of SIMD instructions.
-		ub->vpScale[0] = gstate.getViewportXScale();
-		ub->vpScale[1] = gstate.getViewportYScale();
-		ub->vpScale[2] = gstate.getViewportZScale();
+		Vec4F32 vpScale = Vec4F32::LoadF24x3_DontCare(&gstate.viewportxscale);
+		Vec4F32 vpOffset = Vec4F32::LoadF24x3_DontCare(&gstate.viewportxcenter);
+		vpScale.Store(ub->vpScale);
+		vpOffset.Store(ub->vpOffset);
 		ub->NaN = std::numeric_limits<float>::quiet_NaN();  // Used in the shader for range culling.
-		ub->vpOffset[0] = gstate.getViewportXCenter();
-		ub->vpOffset[1] = gstate.getViewportYCenter();
-		ub->vpOffset[2] = gstate.getViewportZCenter();
 	}
 
 	// Transform
 	if (dirtyUniforms & DIRTY_WORLDMATRIX) {
+		// TODO: We could change the shader to directly read these "malformed" matrices, but we'd
+		// be doing the matrix multiplication manually.
 		ConvertMatrix4x3To3x4Transposed(ub->world, gstate.worldMatrix);
 	}
 	if (dirtyUniforms & DIRTY_VIEWMATRIX) {

GPU/D3D11/DrawEngineD3D11.cpp

@@ -297,7 +297,7 @@ void DrawEngineD3D11::Flush() {
 	useHWTransform = CheckBoundingDepths(useHWTransform);
 
 	if (useHWTransform != lastUseHwTransform_) {
-		gstate_c.Dirty(DIRTY_VERTEXSHADER_STATE | DIRTY_RASTER_STATE);
+		gstate_c.Dirty(DIRTY_VERTEXSHADER_STATE | DIRTY_FRAGMENTSHADER_STATE | DIRTY_RASTER_STATE);
 		lastUseHwTransform_ = useHWTransform;
 	}
 

GPU/GLES/DrawEngineGLES.cpp

@@ -257,7 +257,7 @@ void DrawEngineGLES::Flush() {
 	useHWTransform = CheckBoundingDepths(useHWTransform);
 
 	if (useHWTransform != lastUseHwTransform_) {
-		gstate_c.Dirty(DIRTY_VERTEXSHADER_STATE | DIRTY_RASTER_STATE);
+		gstate_c.Dirty(DIRTY_VERTEXSHADER_STATE | DIRTY_FRAGMENTSHADER_STATE | DIRTY_RASTER_STATE);
 		lastUseHwTransform_ = useHWTransform;
 	}
 

GPU/GPUCommonHW.cpp

@@ -585,6 +585,8 @@ u32 GPUCommonHW::CheckGPUFeatures() const {
 		features |= GPU_USE_BLEND_MINMAX;
 	}
 
+	// The next three (clipDistance, cullDistance, depthClamp) are the critical features that let us correctly emulate PSP behavior without fallbacks.
+	// This is except for the case of clipped polygons still reaching outside the guardband, where we still need to software-clip if detected.
 	if (draw_->GetDeviceCaps().maxClipDistances >= 3) {
 		features |= GPU_USE_CLIP_DISTANCE;
 	}
@@ -593,15 +595,15 @@ u32 GPUCommonHW::CheckGPUFeatures() const {
 		features |= GPU_USE_CULL_DISTANCE;
 	}
 
-	if (draw_->GetDeviceCaps().textureDepthSupported) {
-		features |= GPU_USE_DEPTH_TEXTURE;
-	}
-
 	if (draw_->GetDeviceCaps().depthClampSupported) {
 		// Some backends always do GPU_USE_ACCURATE_DEPTH, but it's required for depth clamp.
 		features |= GPU_USE_DEPTH_CLAMP;
 	}
 
+	if (draw_->GetDeviceCaps().textureDepthSupported) {
+		features |= GPU_USE_DEPTH_TEXTURE;
+	}
+
 	if (draw_->GetDeviceCaps().framebufferFetchSupported) {
 		features |= GPU_USE_FRAMEBUFFER_FETCH;
 		features |= GPU_USE_SHADER_BLENDING;   // doesn't matter if we are buffered or not here.

GPU/Math3D.h

@@ -1176,26 +1176,26 @@ inline void ConvertMatrix4x3To4x4Transposed(float *m4x4, const float *m4x3) {
 // 4567
 // 89AB
 // Don't see a way to SIMD that. Should be pretty fast anyway.
-inline void ConvertMatrix4x3To3x4Transposed(float *m4x4, const float *m4x3) {
+inline void ConvertMatrix4x3To3x4Transposed(float *m3x4, const float *m4x3) {
 #if PPSSPP_ARCH(ARM_NEON)
 	// vld3q is a perfect match here!
 	float32x4x3_t packed = vld3q_f32(m4x3);
-	vst1q_f32(m4x4, packed.val[0]);
-	vst1q_f32(m4x4 + 4, packed.val[1]);
-	vst1q_f32(m4x4 + 8, packed.val[2]);
+	vst1q_f32(m3x4, packed.val[0]);
+	vst1q_f32(m3x4 + 4, packed.val[1]);
+	vst1q_f32(m3x4 + 8, packed.val[2]);
 #else
-	m4x4[0] = m4x3[0];
-	m4x4[1] = m4x3[3];
-	m4x4[2] = m4x3[6];
-	m4x4[3] = m4x3[9];
-	m4x4[4] = m4x3[1];
-	m4x4[5] = m4x3[4];
-	m4x4[6] = m4x3[7];
-	m4x4[7] = m4x3[10];
-	m4x4[8] = m4x3[2];
-	m4x4[9] = m4x3[5];
-	m4x4[10] = m4x3[8];
-	m4x4[11] = m4x3[11];
+	m3x4[0] = m4x3[0];
+	m3x4[1] = m4x3[3];
+	m3x4[2] = m4x3[6];
+	m3x4[3] = m4x3[9];
+	m3x4[4] = m4x3[1];
+	m3x4[5] = m4x3[4];
+	m3x4[6] = m4x3[7];
+	m3x4[7] = m4x3[10];
+	m3x4[8] = m4x3[2];
+	m3x4[9] = m4x3[5];
+	m3x4[10] = m4x3[8];
+	m3x4[11] = m4x3[11];
 #endif
 }