GS: Improve state load determinism

CLUT wasn't force reloaded, so if the game didn't write to TEX0 before
its first draw after loading a state, you got either the CLUT before the
state was loaded, or random/uninitialized garbage.

Do the same for reset as well, except zero it out in that case.

pcsx2/GS/GSClut.cpp

@@ -13,8 +13,6 @@
 GSClut::GSClut(GSLocalMemory* mem)
 	: m_mem(mem)
 {
-	static constexpr u32 CLUT_ALLOC_SIZE = 4096 * 2;
-
 	// 1k + 1k for mirrored area simulating wrapping memory
 	m_clut = static_cast<u16*>(_aligned_malloc(CLUT_ALLOC_SIZE, VECTOR_ALIGNMENT));
 	if (!m_clut)
@@ -94,9 +92,7 @@ GSClut::GSClut(GSLocalMemory* mem)
 
 GSClut::~GSClut()
 {
-	if (m_gpu_clut4)
 	delete m_gpu_clut4;
-	if (m_gpu_clut8)
 	delete m_gpu_clut8;
 
 	_aligned_free(m_clut);
@@ -130,6 +126,16 @@ void GSClut::SetNextCLUTTEX0(u64 TEX0)
 	m_write.next_tex0 = TEX0;
 }
 
+void GSClut::Reset()
+{
+	std::memset(m_CBP, 0, sizeof(m_CBP));
+	std::memset(m_clut, 0, CLUT_ALLOC_SIZE);
+	m_write = {};
+	m_write.dirty = 1;
+	m_read = {};
+	m_read.dirty = true;
+}
+
 bool GSClut::InvalidateRange(u32 start_block, u32 end_block, bool is_draw)
 {
 	if (m_write.dirty & 2)

pcsx2/GS/GSClut.h

@@ -13,16 +13,18 @@ class GSTexture;
 
 class alignas(32) GSClut final : public GSAlignedClass<32>
 {
+	static constexpr u32 CLUT_ALLOC_SIZE = 4096 * 2;
+
 	static const GSVector4i m_bm;
 	static const GSVector4i m_gm;
 	static const GSVector4i m_rm;
 
 	GSLocalMemory* m_mem;
 
-	u32 m_CBP[2];
+	u32 m_CBP[2] = {};
-	u16* m_clut;
+	u16* m_clut = nullptr;
-	u32* m_buff32;
+	u32* m_buff32 = nullptr;
-	u64* m_buff64;
+	u64* m_buff64 = nullptr;
 
 	struct alignas(32) WriteState
 	{
@@ -31,7 +33,7 @@ class alignas(32) GSClut final : public GSAlignedClass<32>
 		u8 dirty;
 		u64 next_tex0;
 		bool IsDirty(const GIFRegTEX0& TEX0, const GIFRegTEXCLUT& TEXCLUT);
-	} m_write;
+	} m_write = {};
 
 	struct alignas(32) ReadState
 	{
@@ -42,7 +44,7 @@ class alignas(32) GSClut final : public GSAlignedClass<32>
 		int amin, amax;
 		bool IsDirty(const GIFRegTEX0& TEX0);
 		bool IsDirty(const GIFRegTEX0& TEX0, const GIFRegTEXA& TEXA);
-	} m_read;
+	} m_read = {};
 
 	GSTexture* m_gpu_clut4 = nullptr;
 	GSTexture* m_gpu_clut8 = nullptr;
@@ -96,6 +98,7 @@ public:
 
 	__fi GSTexture* GetGPUTexture() const { return m_current_gpu_clut; }
 
+	void Reset();
 	bool InvalidateRange(u32 start_block, u32 end_block, bool is_draw = false);
 	u8 IsInvalid();
 	void ClearDrawInvalidity();

pcsx2/GS/GSState.cpp

@@ -140,6 +140,7 @@ void GSState::Reset(bool hardware_reset)
 	memset(&m_v, 0, sizeof(m_v));
 
 	m_env.Reset();
+	m_mem.m_clut.Reset();
 
 	PRIM = &m_env.PRIM;
 
@@ -181,6 +182,7 @@ void GSState::Reset(bool hardware_reset)
 	m_index.tail = 0;
 	m_scanmask_used = 0;
 	m_texflush_flag = false;
+	m_channel_shuffle = false;
 	m_dirty_gs_regs = 0;
 	m_backed_up_ctx = -1;
 
@@ -2708,6 +2710,7 @@ int GSState::Defrost(const freezeData* fd)
 
 	ReadState(&m_q, data);
 
+	m_prev_env = m_env;
 	PRIM = &m_env.PRIM;
 
 	UpdateContext();
@@ -2725,6 +2728,10 @@ int GSState::Defrost(const freezeData* fd)
 
 	UpdateScissor();
 
+	// Force CLUT to be reloaded.
+	m_mem.m_clut.Reset();
+	(PRIM->CTXT == 0) ? ApplyTEX0<0>(m_context->TEX0) : ApplyTEX0<1>(m_context->TEX0);
+
 	g_perfmon.SetFrame(5000);
 
 	ResetPCRTC();