SPU2: Multi-isa resampling functions

pcsx2/CMakeLists.txt

@@ -280,6 +280,10 @@ set(pcsx2SPU2Sources
 	SPU2/Wavedump_wav.cpp
 )
 
+set(pcsx2SPU2SourcesUnshared
+	SPU2/ReverbResample.cpp
+)
+
 # SPU2 headers
 set(pcsx2SPU2Headers
 	SPU2/Debug.h
@@ -762,7 +766,7 @@ if(DISABLE_ADVANCE_SIMD)
 	# Note: ld64 (macOS's linker) does not act the same way when presented with .a files, unless linked with `-force_load` (cmake WHOLE_ARCHIVE).
 	set(is_first_isa "1")
 	foreach(isa "sse4" "avx" "avx2")
-		add_library(GS-${isa} STATIC ${pcsx2GSSourcesUnshared} ${pcsx2IPUSourcesUnshared})
+		add_library(GS-${isa} STATIC ${pcsx2GSSourcesUnshared} ${pcsx2IPUSourcesUnshared} ${pcsx2SPU2SourcesUnshared})
 		target_link_libraries(GS-${isa} PRIVATE PCSX2_FLAGS)
 		target_compile_definitions(GS-${isa} PRIVATE MULTI_ISA_UNSHARED_COMPILATION=isa_${isa} MULTI_ISA_IS_FIRST=${is_first_isa} ${pcsx2_defs_${isa}})
 		target_compile_options(GS-${isa} PRIVATE ${compile_options_${isa}})
@@ -778,6 +782,7 @@ if(DISABLE_ADVANCE_SIMD)
 else()
 	list(APPEND pcsx2GSSources ${pcsx2GSSourcesUnshared})
 	list(APPEND pcsx2IPUSources ${pcsx2IPUSourcesUnshared})
+	list(APPEND pcsx2SPU2Sources ${pcsx2SPU2SourcesUnshared})
 endif()
 
 # DebugTools sources

pcsx2/SPU2/Reverb.cpp

@@ -19,7 +19,6 @@
 
 #include <array>
 
-
 void V_Core::AnalyzeReverbPreset()
 {
 	Console.WriteLn("Reverb Parameter Update for Core %d:", Index);
@@ -55,193 +54,6 @@ void V_Core::AnalyzeReverbPreset()
 	Console.WriteLn("----------------------------------------------------------");
 }
 
-static constexpr u32 NUM_TAPS = 39;
-// 39 tap filter, the 0's could be optimized out
-static constexpr std::array<s16, 48> filter_down_coefs alignas(32) = {
-	-1,
-	0,
-	2,
-	0,
-	-10,
-	0,
-	35,
-	0,
-	-103,
-	0,
-	266,
-	0,
-	-616,
-	0,
-	1332,
-	0,
-	-2960,
-	0,
-	10246,
-	16384,
-	10246,
-	0,
-	-2960,
-	0,
-	1332,
-	0,
-	-616,
-	0,
-	266,
-	0,
-	-103,
-	0,
-	35,
-	0,
-	-10,
-	0,
-	2,
-	0,
-	-1,
-};
-
-static constexpr std::array<s16, 48> make_up_coefs()
-{
-	std::array<s16, 48> ret = {};
-
-	for (u32 i = 0; i < NUM_TAPS; i++)
-	{
-		ret[i] = static_cast<s16>(std::clamp<s32>(filter_down_coefs[i] * 2, INT16_MIN, INT16_MAX));
-	}
-
-	return ret;
-}
-
-static constexpr std::array<s16, 48> filter_up_coefs alignas(32) = make_up_coefs();
-
-s32 __forceinline V_Core::ReverbDownsample(bool right)
-{
-	int index = (RevbSampleBufPos - NUM_TAPS) & 63;
-
-#if _M_SSE >= 0x501
-	auto c = GSVector8i::load<true>(&filter_down_coefs[0]);
-	auto s = GSVector8i::load<false>(&RevbDownBuf[right][index]);
-	auto acc = s.mul16hrs(c);
-
-	c = GSVector8i::load<true>(&filter_down_coefs[16]);
-	s = GSVector8i::load<false>(&RevbDownBuf[right][index + 16]);
-	acc = acc.adds16(s.mul16hrs(c));
-
-	c = GSVector8i::load<true>(&filter_down_coefs[32]);
-	s = GSVector8i::load<false>(&RevbDownBuf[right][index + 32]);
-	acc = acc.adds16(s.mul16hrs(c));
-
-	acc = acc.adds16(acc.ba());
-
-	acc = acc.hadds16(acc);
-	acc = acc.hadds16(acc);
-	acc = acc.hadds16(acc);
-#else
-	auto c = GSVector4i::load<true>(&filter_down_coefs[0]);
-	auto s = GSVector4i::load<false>(&RevbDownBuf[right][index]);
-	auto acc = s.mul16hrs(c);
-
-	c = GSVector4i::load<true>(&filter_down_coefs[8]);
-	s = GSVector4i::load<false>(&RevbDownBuf[right][index + 8]);
-	acc = acc.adds16(s.mul16hrs(c));
-
-	c = GSVector4i::load<true>(&filter_down_coefs[16]);
-	s = GSVector4i::load<false>(&RevbDownBuf[right][index + 16]);
-	acc = acc.adds16(s.mul16hrs(c));
-
-	c = GSVector4i::load<true>(&filter_down_coefs[24]);
-	s = GSVector4i::load<false>(&RevbDownBuf[right][index + 24]);
-	acc = acc.adds16(s.mul16hrs(c));
-
-	c = GSVector4i::load<true>(&filter_down_coefs[32]);
-	s = GSVector4i::load<false>(&RevbDownBuf[right][index + 32]);
-	acc = acc.adds16(s.mul16hrs(c));
-
-	acc = acc.hadds16(acc);
-	acc = acc.hadds16(acc);
-	acc = acc.hadds16(acc);
-#endif
-
-	return acc.I16[0];
-}
-
-StereoOut32 __forceinline V_Core::ReverbUpsample()
-{
-	int index = (RevbSampleBufPos - NUM_TAPS) & 63;
-
-#if _M_SSE >= 0x501
-	auto c = GSVector8i::load<true>(&filter_up_coefs[0]);
-	auto l = GSVector8i::load<false>(&RevbUpBuf[0][index]);
-	auto r = GSVector8i::load<false>(&RevbUpBuf[1][index]);
-
-	auto lacc = l.mul16hrs(c);
-	auto racc = r.mul16hrs(c);
-
-	c = GSVector8i::load<true>(&filter_up_coefs[16]);
-	l = GSVector8i::load<false>(&RevbUpBuf[0][index + 16]);
-	r = GSVector8i::load<false>(&RevbUpBuf[1][index + 16]);
-	lacc = lacc.adds16(l.mul16hrs(c));
-	racc = racc.adds16(r.mul16hrs(c));
-
-	c = GSVector8i::load<true>(&filter_up_coefs[32]);
-	l = GSVector8i::load<false>(&RevbUpBuf[0][index + 32]);
-	r = GSVector8i::load<false>(&RevbUpBuf[1][index + 32]);
-	lacc = lacc.adds16(l.mul16hrs(c));
-	racc = racc.adds16(r.mul16hrs(c));
-
-	lacc = lacc.adds16(lacc.ba());
-	racc = racc.adds16(racc.ba());
-
-	lacc = lacc.hadds16(lacc);
-	lacc = lacc.hadds16(lacc);
-	lacc = lacc.hadds16(lacc);
-
-	racc = racc.hadds16(racc);
-	racc = racc.hadds16(racc);
-	racc = racc.hadds16(racc);
-#else
-	auto c = GSVector4i::load<true>(&filter_up_coefs[0]);
-	auto l = GSVector4i::load<false>(&RevbUpBuf[0][index]);
-	auto r = GSVector4i::load<false>(&RevbUpBuf[1][index]);
-
-	auto lacc = l.mul16hrs(c);
-	auto racc = r.mul16hrs(c);
-
-	c = GSVector4i::load<true>(&filter_up_coefs[8]);
-	l = GSVector4i::load<false>(&RevbUpBuf[0][index + 8]);
-	r = GSVector4i::load<false>(&RevbUpBuf[1][index + 8]);
-	lacc = lacc.adds16(l.mul16hrs(c));
-	racc = racc.adds16(r.mul16hrs(c));
-
-	c = GSVector4i::load<true>(&filter_up_coefs[16]);
-	l = GSVector4i::load<false>(&RevbUpBuf[0][index + 16]);
-	r = GSVector4i::load<false>(&RevbUpBuf[1][index + 16]);
-	lacc = lacc.adds16(l.mul16hrs(c));
-	racc = racc.adds16(r.mul16hrs(c));
-
-	c = GSVector4i::load<true>(&filter_up_coefs[24]);
-	l = GSVector4i::load<false>(&RevbUpBuf[0][index + 24]);
-	r = GSVector4i::load<false>(&RevbUpBuf[1][index + 24]);
-	lacc = lacc.adds16(l.mul16hrs(c));
-	racc = racc.adds16(r.mul16hrs(c));
-
-	c = GSVector4i::load<true>(&filter_up_coefs[32]);
-	l = GSVector4i::load<false>(&RevbUpBuf[0][index + 32]);
-	r = GSVector4i::load<false>(&RevbUpBuf[1][index + 32]);
-	lacc = lacc.adds16(l.mul16hrs(c));
-	racc = racc.adds16(r.mul16hrs(c));
-
-	lacc = lacc.hadds16(lacc);
-	lacc = lacc.hadds16(lacc);
-	lacc = lacc.hadds16(lacc);
-
-	racc = racc.hadds16(racc);
-	racc = racc.hadds16(racc);
-	racc = racc.hadds16(racc);
-#endif
-
-	return {lacc.I16[0], racc.I16[0]};
-}
-
 __forceinline s32 V_Core::RevbGetIndexer(s32 offset)
 {
 	u32 start = EffectsStartA & 0x3f'ffff;
@@ -325,7 +137,7 @@ StereoOut32 V_Core::DoReverb(const StereoOut32& Input)
 	s32 in, same, diff, apf1, apf2, out;
 
 #define MUL(x, y) ((x) * (y) >> 15)
-	in = MUL(R ? Revb.IN_COEF_R : Revb.IN_COEF_L, ReverbDownsample(R));
+	in = MUL(R ? Revb.IN_COEF_R : Revb.IN_COEF_L, ReverbDownsample(*this, R));
 
 	same = MUL(Revb.IIR_VOL, in + MUL(Revb.WALL_VOL, _spu2mem[same_src]) - _spu2mem[same_prv]) + _spu2mem[same_prv];
 	diff = MUL(Revb.IIR_VOL, in + MUL(Revb.WALL_VOL, _spu2mem[diff_src]) - _spu2mem[diff_prv]) + _spu2mem[diff_prv];
@@ -356,5 +168,5 @@ StereoOut32 V_Core::DoReverb(const StereoOut32& Input)
 
 	RevbSampleBufPos = (RevbSampleBufPos + 1) & 63;
 
-	return ReverbUpsample();
+	return ReverbUpsample(*this);
 }

pcsx2/SPU2/ReverbResample.cpp

@@ -0,0 +1,257 @@
+#include "GS/GSVector.h"
+#include "Global.h"
+
+MULTI_ISA_UNSHARED_START
+
+static constexpr u32 NUM_TAPS = 39;
+// 39 tap filter, the 0's could be optimized out
+static constexpr std::array<s16, 48> filter_down_coefs alignas(32) = {
+	-1,
+	0,
+	2,
+	0,
+	-10,
+	0,
+	35,
+	0,
+	-103,
+	0,
+	266,
+	0,
+	-616,
+	0,
+	1332,
+	0,
+	-2960,
+	0,
+	10246,
+	16384,
+	10246,
+	0,
+	-2960,
+	0,
+	1332,
+	0,
+	-616,
+	0,
+	266,
+	0,
+	-103,
+	0,
+	35,
+	0,
+	-10,
+	0,
+	2,
+	0,
+	-1,
+};
+
+static constexpr std::array<s16, 48> make_up_coefs()
+{
+	std::array<s16, 48> ret = {};
+
+	for (u32 i = 0; i < NUM_TAPS; i++)
+	{
+		ret[i] = static_cast<s16>(std::clamp<s32>(filter_down_coefs[i] * 2, INT16_MIN, INT16_MAX));
+	}
+
+	return ret;
+}
+
+static constexpr std::array<s16, 48> filter_up_coefs alignas(32) = make_up_coefs();
+
+s32 __forceinline ReverbDownsample_reference(V_Core& core, bool right)
+{
+	int index = (core.RevbSampleBufPos - NUM_TAPS) & 63;
+	s32 out = 0;
+
+	for (int i = 0; i < NUM_TAPS; i++)
+	{
+		out += core.RevbDownBuf[right][index + i] * filter_down_coefs[i];
+	}
+
+	out >>= 15;
+
+	return clamp_mix(out);
+}
+
+#if _M_SSE >= 0x501
+s32 __forceinline ReverbDownsample_avx(V_Core& core, bool right)
+{
+	int index = (core.RevbSampleBufPos - NUM_TAPS) & 63;
+
+	auto c = GSVector8i::load<true>(&filter_down_coefs[0]);
+	auto s = GSVector8i::load<false>(&core.RevbDownBuf[right][index]);
+	auto acc = s.mul16hrs(c);
+
+	c = GSVector8i::load<true>(&filter_down_coefs[16]);
+	s = GSVector8i::load<false>(&core.RevbDownBuf[right][index + 16]);
+	acc = acc.adds16(s.mul16hrs(c));
+
+	c = GSVector8i::load<true>(&filter_down_coefs[32]);
+	s = GSVector8i::load<false>(&core.RevbDownBuf[right][index + 32]);
+	acc = acc.adds16(s.mul16hrs(c));
+
+	acc = acc.adds16(acc.ba());
+
+	acc = acc.hadds16(acc);
+	acc = acc.hadds16(acc);
+	acc = acc.hadds16(acc);
+
+	return acc.I16[0];
+}
+#endif
+
+s32 __forceinline ReverbDownsample_sse(V_Core& core, bool right)
+{
+	int index = (core.RevbSampleBufPos - NUM_TAPS) & 63;
+
+	auto c = GSVector4i::load<true>(&filter_down_coefs[0]);
+	auto s = GSVector4i::load<false>(&core.RevbDownBuf[right][index]);
+	auto acc = s.mul16hrs(c);
+
+	c = GSVector4i::load<true>(&filter_down_coefs[8]);
+	s = GSVector4i::load<false>(&core.RevbDownBuf[right][index + 8]);
+	acc = acc.adds16(s.mul16hrs(c));
+
+	c = GSVector4i::load<true>(&filter_down_coefs[16]);
+	s = GSVector4i::load<false>(&core.RevbDownBuf[right][index + 16]);
+	acc = acc.adds16(s.mul16hrs(c));
+
+	c = GSVector4i::load<true>(&filter_down_coefs[24]);
+	s = GSVector4i::load<false>(&core.RevbDownBuf[right][index + 24]);
+	acc = acc.adds16(s.mul16hrs(c));
+
+	c = GSVector4i::load<true>(&filter_down_coefs[32]);
+	s = GSVector4i::load<false>(&core.RevbDownBuf[right][index + 32]);
+	acc = acc.adds16(s.mul16hrs(c));
+
+	acc = acc.hadds16(acc);
+	acc = acc.hadds16(acc);
+	acc = acc.hadds16(acc);
+
+	return acc.I16[0];
+}
+
+s32 ReverbDownsample(V_Core& core, bool right)
+{
+#if _M_SSE >= 0x501
+	return ReverbDownsample_avx(core, right);
+#else
+	return ReverbDownsample_sse(core, right);
+#endif
+}
+
+StereoOut32 __forceinline ReverbUpsample_reference(V_Core& core)
+{
+	int index = (core.RevbSampleBufPos - NUM_TAPS) & 63;
+	s32 l = 0, r = 0;
+
+	for (int i = 0; i < NUM_TAPS; i++)
+	{
+		l += core.RevbUpBuf[0][index + i] * filter_up_coefs[i];
+		r += core.RevbUpBuf[1][index + i] * filter_up_coefs[i];
+	}
+
+	l >>= 15;
+	r >>= 15;
+
+	return {clamp_mix(l), clamp_mix(r)};
+}
+
+#if _M_SSE >= 0x501
+StereoOut32 __forceinline ReverbUpsample_avx(V_Core& core)
+{
+	int index = (core.RevbSampleBufPos - NUM_TAPS) & 63;
+
+	auto c = GSVector8i::load<true>(&filter_up_coefs[0]);
+	auto l = GSVector8i::load<false>(&core.RevbUpBuf[0][index]);
+	auto r = GSVector8i::load<false>(&core.RevbUpBuf[1][index]);
+
+	auto lacc = l.mul16hrs(c);
+	auto racc = r.mul16hrs(c);
+
+	c = GSVector8i::load<true>(&filter_up_coefs[16]);
+	l = GSVector8i::load<false>(&core.RevbUpBuf[0][index + 16]);
+	r = GSVector8i::load<false>(&core.RevbUpBuf[1][index + 16]);
+	lacc = lacc.adds16(l.mul16hrs(c));
+	racc = racc.adds16(r.mul16hrs(c));
+
+	c = GSVector8i::load<true>(&filter_up_coefs[32]);
+	l = GSVector8i::load<false>(&core.RevbUpBuf[0][index + 32]);
+	r = GSVector8i::load<false>(&core.RevbUpBuf[1][index + 32]);
+	lacc = lacc.adds16(l.mul16hrs(c));
+	racc = racc.adds16(r.mul16hrs(c));
+
+	lacc = lacc.adds16(lacc.ba());
+	racc = racc.adds16(racc.ba());
+
+	lacc = lacc.hadds16(lacc);
+	lacc = lacc.hadds16(lacc);
+	lacc = lacc.hadds16(lacc);
+
+	racc = racc.hadds16(racc);
+	racc = racc.hadds16(racc);
+	racc = racc.hadds16(racc);
+
+	return {lacc.I16[0], racc.I16[0]};
+}
+#endif
+
+StereoOut32 __forceinline ReverbUpsample_sse(V_Core& core)
+{
+	int index = (core.RevbSampleBufPos - NUM_TAPS) & 63;
+
+	auto c = GSVector4i::load<true>(&filter_up_coefs[0]);
+	auto l = GSVector4i::load<false>(&core.RevbUpBuf[0][index]);
+	auto r = GSVector4i::load<false>(&core.RevbUpBuf[1][index]);
+
+	auto lacc = l.mul16hrs(c);
+	auto racc = r.mul16hrs(c);
+
+	c = GSVector4i::load<true>(&filter_up_coefs[8]);
+	l = GSVector4i::load<false>(&core.RevbUpBuf[0][index + 8]);
+	r = GSVector4i::load<false>(&core.RevbUpBuf[1][index + 8]);
+	lacc = lacc.adds16(l.mul16hrs(c));
+	racc = racc.adds16(r.mul16hrs(c));
+
+	c = GSVector4i::load<true>(&filter_up_coefs[16]);
+	l = GSVector4i::load<false>(&core.RevbUpBuf[0][index + 16]);
+	r = GSVector4i::load<false>(&core.RevbUpBuf[1][index + 16]);
+	lacc = lacc.adds16(l.mul16hrs(c));
+	racc = racc.adds16(r.mul16hrs(c));
+
+	c = GSVector4i::load<true>(&filter_up_coefs[24]);
+	l = GSVector4i::load<false>(&core.RevbUpBuf[0][index + 24]);
+	r = GSVector4i::load<false>(&core.RevbUpBuf[1][index + 24]);
+	lacc = lacc.adds16(l.mul16hrs(c));
+	racc = racc.adds16(r.mul16hrs(c));
+
+	c = GSVector4i::load<true>(&filter_up_coefs[32]);
+	l = GSVector4i::load<false>(&core.RevbUpBuf[0][index + 32]);
+	r = GSVector4i::load<false>(&core.RevbUpBuf[1][index + 32]);
+	lacc = lacc.adds16(l.mul16hrs(c));
+	racc = racc.adds16(r.mul16hrs(c));
+
+	lacc = lacc.hadds16(lacc);
+	lacc = lacc.hadds16(lacc);
+	lacc = lacc.hadds16(lacc);
+
+	racc = racc.hadds16(racc);
+	racc = racc.hadds16(racc);
+	racc = racc.hadds16(racc);
+
+	return {lacc.I16[0], racc.I16[0]};
+}
+
+StereoOut32 ReverbUpsample(V_Core& core)
+{
+#if _M_SSE >= 0x501
+	return ReverbUpsample_avx(core);
+#else
+	return ReverbUpsample_sse(core);
+#endif
+}
+
+MULTI_ISA_UNSHARED_END

pcsx2/SPU2/defs.h

@@ -19,6 +19,8 @@
 #include "SPU2/SndOut.h"
 #include "SPU2/Global.h"
 
+#include "GS/MultiISA.h"
+
 #include <array>
 
 // --------------------------------------------------------------------------------------
@@ -486,9 +488,6 @@ struct V_Core
 	StereoOut32 DoReverb(const StereoOut32& Input);
 	s32 RevbGetIndexer(s32 offset);
 
-	s32 ReverbDownsample(bool right);
-	StereoOut32 ReverbUpsample();
-
 	StereoOut32 ReadInput();
 	StereoOut32 ReadInput_HiFi();
 
@@ -537,6 +536,14 @@ struct V_Core
 	void FinishDMAwrite();
 };
 
+MULTI_ISA_DEF(
+	StereoOut32 ReverbUpsample(V_Core& core);
+	s32 ReverbDownsample(V_Core& core, bool right);
+)
+
+extern StereoOut32 (*ReverbUpsample)(V_Core& core);
+extern s32 (*ReverbDownsample)(V_Core& core, bool right);
+
 extern V_Core Cores[2];
 extern V_SPDIF Spdif;
 

pcsx2/SPU2/spu2sys.cpp

@@ -46,6 +46,9 @@ int PlayMode;
 
 static bool has_to_call_irq[2] = { false, false };
 static bool has_to_call_irq_dma[2] = { false, false };
+StereoOut32 (*ReverbUpsample)(V_Core& core);
+s32 (*ReverbDownsample)(V_Core& core, bool right);
+
 
 static bool psxmode = false;
 
@@ -111,6 +114,9 @@ void V_Core::Init(int index)
 	if (SPU2::MsgToConsole())
 		SPU2::ConLog("* SPU2: Init SPU2 core %d \n", index);
 
+	ReverbDownsample = MULTI_ISA_SELECT(ReverbDownsample);
+	ReverbUpsample = MULTI_ISA_SELECT(ReverbUpsample);
+
 	//memset(this, 0, sizeof(V_Core));
 	// Explicitly initializing variables instead.
 	Mute = false;

pcsx2/pcsx2.vcxproj

@@ -264,6 +264,7 @@
     <ClCompile Include="SPU2\Mixer.cpp" />
     <ClCompile Include="SPU2\ReadInput.cpp" />
     <ClCompile Include="SPU2\Reverb.cpp" />
+    <ClCompile Include="SPU2\ReverbResample.cpp" />
     <ClCompile Include="SPU2\spu2.cpp" />
     <ClCompile Include="IPU\IPUdma.cpp" />
     <ClCompile Include="IPU\IPUdither.cpp" />
@@ -872,4 +873,4 @@
   <Import Condition="$(Configuration.Contains(Debug)) Or $(Configuration.Contains(Devel))" Project="$(SolutionDir)3rdparty\winpixeventruntime\WinPixEventRuntime.props" />
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets" />
-</Project>
+</Project>