modplug-svn-commit Mailing List for modplug (Page 14)

SourceForge Headquarters 225 Broadway Suite 1600 San Diego, CA 92101 +1 (858) 422-6466

Author: manx
Date: Thu Jun  5 14:16:41 2025
New Revision: 23350
URL: https://source.openmpt.org/browse/openmpt/?op=revision&rev=23350

Log:
[Fix] Reverb: Use alignment-correct intrinsics for 64bit SSE2 load and store. Even though the underlying instruction is the same in both cases, and can do unaligned load and store even for the aligned intrinsics, and even though the intrinsics are documented to support unaligned load and store, the C++ language conflicts with that definition because the pointer argument is __m128i, which has 16 byte alignment requirements. And GCC decides to be pedantic here. Code generation is identical for all 3 major compilers.

Modified:
   trunk/OpenMPT/sounddsp/Reverb.cpp

Modified: trunk/OpenMPT/sounddsp/Reverb.cpp
==============================================================================

--- trunk/OpenMPT/sounddsp/Reverb.cpp	Wed Jun  4 23:09:30 2025	(r23349)
+++ trunk/OpenMPT/sounddsp/Reverb.cpp	Thu Jun  5 14:16:41 2025	(r23350)
@@ -24,6 +24,7 @@
 #include <intrin.h>
 #endif
 #include <emmintrin.h>
+#include <immintrin.h>
 #endif
 
 #endif // NO_REVERB
@@ -35,17 +36,6 @@
 #ifndef NO_REVERB
 
 
-#if defined(MPT_WANT_ARCH_INTRINSICS_X86_SSE2) && defined(MPT_ARCH_INTRINSICS_X86_SSE2)
-// Load two 32-bit values
-static MPT_FORCEINLINE __m128i Load64SSE(const int32 *x) { return _mm_loadl_epi64(reinterpret_cast<const __m128i *>(x)); }
-// Load four 16-bit values
-static MPT_FORCEINLINE __m128i Load64SSE(const LR16 (&x)[2]) { return _mm_loadl_epi64(&reinterpret_cast<const __m128i &>(x)); }
-// Store two 32-bit or four 16-bit values from register
-static MPT_FORCEINLINE void Store64SSE(int32 *dst, __m128i src) { return _mm_storel_epi64(reinterpret_cast<__m128i *>(dst), src); }
-static MPT_FORCEINLINE void Store64SSE(LR16 (&dst)[2], __m128i src) { return _mm_storel_epi64(&reinterpret_cast<__m128i &>(dst), src); }
-#endif
-
-
 CReverb::CReverb()
 {
 	// Reverb mix buffers
@@ -597,24 +587,24 @@
 #if defined(MPT_WANT_ARCH_INTRINSICS_X86_SSE2) && defined(MPT_ARCH_INTRINSICS_X86_SSE2)
 	if(CPU::HasFeatureSet(CPU::feature::sse2) && CPU::HasModesEnabled(CPU::mode::xmm128sse))
 	{
-		__m128i nDCRRvb_Y1 = Load64SSE(gnDCRRvb_Y1);
-		__m128i nDCRRvb_X1 = Load64SSE(gnDCRRvb_X1);
+		__m128i nDCRRvb_Y1 = _mm_loadu_si64(gnDCRRvb_Y1);
+		__m128i nDCRRvb_X1 = _mm_loadu_si64(gnDCRRvb_X1);
 		__m128i in = _mm_set1_epi32(0);
 		while(nSamples--)
 		{
-			in = Load64SSE(pRvb);
+			in = _mm_loadu_si64(pRvb);
 			pRvb += 2;
 			// x(n-1) - x(n)
 			__m128i diff = _mm_sub_epi32(nDCRRvb_X1, in);
 			nDCRRvb_X1 = _mm_add_epi32(nDCRRvb_Y1, _mm_sub_epi32(_mm_srai_epi32(diff, DCR_AMOUNT + 1), diff));
-			__m128i out = _mm_add_epi32(Load64SSE(pDry), nDCRRvb_X1);
+			__m128i out = _mm_add_epi32(_mm_loadu_si64(pDry), nDCRRvb_X1);
 			nDCRRvb_Y1 = _mm_sub_epi32(nDCRRvb_X1, _mm_srai_epi32(nDCRRvb_X1, DCR_AMOUNT));
 			nDCRRvb_X1 = in;
-			Store64SSE(pDry, out);
+			_mm_storeu_si64(pDry, out);
 			pDry += 2;
 		}
-		Store64SSE(gnDCRRvb_X1, in);
-		Store64SSE(gnDCRRvb_Y1, nDCRRvb_Y1);
+		_mm_storeu_si64(gnDCRRvb_X1, in);
+		_mm_storeu_si64(gnDCRRvb_Y1, nDCRRvb_Y1);
 		return;
 	}
 #endif
@@ -659,20 +649,20 @@
 #if defined(MPT_WANT_ARCH_INTRINSICS_X86_SSE2) && defined(MPT_ARCH_INTRINSICS_X86_SSE2)
 	if(CPU::HasFeatureSet(CPU::feature::sse2) && CPU::HasModesEnabled(CPU::mode::xmm128sse))
 	{
-		__m128i nDCRRvb_Y1 = Load64SSE(gnDCRRvb_Y1);
-		__m128i nDCRRvb_X1 = Load64SSE(gnDCRRvb_X1);
+		__m128i nDCRRvb_Y1 = _mm_loadu_si64(gnDCRRvb_Y1);
+		__m128i nDCRRvb_X1 = _mm_loadu_si64(gnDCRRvb_X1);
 		while(nSamples--)
 		{
-			__m128i in = Load64SSE(pBuffer);
+			__m128i in = _mm_loadu_si64(pBuffer);
 			__m128i diff = _mm_sub_epi32(nDCRRvb_X1, in);
 			__m128i out = _mm_add_epi32(nDCRRvb_Y1, _mm_sub_epi32(_mm_srai_epi32(diff, DCR_AMOUNT + 1), diff));
-			Store64SSE(pBuffer, out);
+			_mm_storeu_si64(pBuffer, out);
 			pBuffer += 2;
 			nDCRRvb_Y1 = _mm_sub_epi32(out, _mm_srai_epi32(out, DCR_AMOUNT));
 			nDCRRvb_X1 = in;
 		}
-		Store64SSE(gnDCRRvb_X1, nDCRRvb_X1);
-		Store64SSE(gnDCRRvb_Y1, nDCRRvb_Y1);
+		_mm_storeu_si64(gnDCRRvb_X1, nDCRRvb_X1);
+		_mm_storeu_si64(gnDCRRvb_Y1, nDCRRvb_Y1);
 		return;
 	}
 #endif
@@ -729,7 +719,7 @@
 		__m128i preDifCoeffs = _mm_cvtsi32_si128(pPreDelay->nPreDifCoeffs.lr);
 		while(nSamples--)
 		{
-			__m128i in32 = Load64SSE(pIn);					// 16-bit unsaturated reverb input [  r  |  l  ]
+			__m128i in32 = _mm_loadu_si64(pIn);					// 16-bit unsaturated reverb input [  r  |  l  ]
 			__m128i inSat = _mm_packs_epi32(in32, in32);	// [ r | l | r | l ] (16-bit saturated)
 			pIn += 2;
 			// Low-pass
@@ -806,10 +796,10 @@
 		__m128i delayPos = _mm_set_epi16(GETDELAY(7), GETDELAY(6), GETDELAY(5), GETDELAY(4), GETDELAY(3), GETDELAY(2), GETDELAY(1), GETDELAY(0));
 #undef GETDELAY
 		delayPos = _mm_sub_epi16(_mm_set1_epi16(static_cast<int16>(pPreDelay->nDelayPos - 1)), delayPos);
-		__m128i gain12 = _mm_unpacklo_epi64(Load64SSE(pPreDelay->Reflections[0].Gains), Load64SSE(pPreDelay->Reflections[1].Gains));
-		__m128i gain34 = _mm_unpacklo_epi64(Load64SSE(pPreDelay->Reflections[2].Gains), Load64SSE(pPreDelay->Reflections[3].Gains));
-		__m128i gain56 = _mm_unpacklo_epi64(Load64SSE(pPreDelay->Reflections[4].Gains), Load64SSE(pPreDelay->Reflections[5].Gains));
-		__m128i gain78 = _mm_unpacklo_epi64(Load64SSE(pPreDelay->Reflections[6].Gains), Load64SSE(pPreDelay->Reflections[7].Gains));
+		__m128i gain12 = _mm_unpacklo_epi64(_mm_loadu_si64(pPreDelay->Reflections[0].Gains), _mm_loadu_si64(pPreDelay->Reflections[1].Gains));
+		__m128i gain34 = _mm_unpacklo_epi64(_mm_loadu_si64(pPreDelay->Reflections[2].Gains), _mm_loadu_si64(pPreDelay->Reflections[3].Gains));
+		__m128i gain56 = _mm_unpacklo_epi64(_mm_loadu_si64(pPreDelay->Reflections[4].Gains), _mm_loadu_si64(pPreDelay->Reflections[5].Gains));
+		__m128i gain78 = _mm_unpacklo_epi64(_mm_loadu_si64(pPreDelay->Reflections[6].Gains), _mm_loadu_si64(pPreDelay->Reflections[7].Gains));
 		// For 28-bit final output: 16+15-3 = 28
 		__m128i refGain = _mm_srai_epi32(_mm_set_epi32(0, 0, pPreDelay->ReflectionsGain.c.r, pPreDelay->ReflectionsGain.c.l), 3);
 		__m128i delayInc = _mm_set1_epi16(1), delayMask = _mm_set1_epi16(SNDMIX_REFLECTIONS_DELAY_MASK);
@@ -836,7 +826,7 @@
 
 			__m128i out = _mm_madd_epi16(_mm_unpacklo_epi16(refOut, refOut), refGain);	// Apply reflections gain
 			// At this, point, this is the only output of the reverb
-			Store64SSE(pOut, out);
+			_mm_storeu_si64(pOut, out);
 			pOut += 2;
 		}
 		return;
@@ -892,10 +882,10 @@
 	if(CPU::HasFeatureSet(CPU::feature::sse2) && CPU::HasModesEnabled(CPU::mode::xmm128sse))
 	{
 		int delayPos = pReverb->nDelayPos & RVBDLY_MASK;
-		__m128i rvbOutGains = Load64SSE(pReverb->RvbOutGains);
-		__m128i difCoeffs = Load64SSE(pReverb->nDifCoeffs);
-		__m128i decayLP = Load64SSE(pReverb->nDecayLP);
-		__m128i lpHistory = Load64SSE(pReverb->LPHistory);
+		__m128i rvbOutGains = _mm_loadu_si64(pReverb->RvbOutGains);
+		__m128i difCoeffs = _mm_loadu_si64(pReverb->nDifCoeffs);
+		__m128i decayLP = _mm_loadu_si64(pReverb->nDecayLP);
+		__m128i lpHistory = _mm_loadu_si64(pReverb->LPHistory);
 		while(nSamples--)
 		{
 			__m128i refIn = _mm_cvtsi32_si128(pRefOut->lr);	// 16-bit stereo input
@@ -918,7 +908,7 @@
 			lpHistory = _mm_adds_epi16(_mm_adds_epi16(lpDecay, lpDecay), delay2);	// Low-passed decay
 
 			// Apply decay gain
-			__m128i histDecay = _mm_srai_epi32(_mm_madd_epi16(Load64SSE(pReverb->nDecayDC), lpHistory), 15);
+			__m128i histDecay = _mm_srai_epi32(_mm_madd_epi16(_mm_loadu_si64(pReverb->nDecayDC), lpHistory), 15);
 			__m128i histDecayPacked = _mm_shuffle_epi32(_mm_packs_epi32(histDecay, histDecay), _MM_SHUFFLE(2, 0, 2, 0));
 			__m128i histDecayIn = _mm_adds_epi16(histDecayPacked, _mm_srai_epi16(_mm_unpacklo_epi32(refIn, refIn), 2));
 			__m128i histDecayInDiff = _mm_subs_epi16(histDecayIn, _mm_mulhi_epi16(_mm_cvtsi32_si128(diffusion1), difCoeffs));
@@ -934,23 +924,23 @@
 				_mm_cvtsi32_si128(pReverb->Delay1[DELAY_OFFSET(RVBDLY1L_LEN)].lr),
 				_mm_cvtsi32_si128(pReverb->Delay1[DELAY_OFFSET(RVBDLY1R_LEN)].lr));
 
-			__m128i delay1Gains = _mm_srai_epi32(_mm_madd_epi16(delay1, Load64SSE(pReverb->Dif2InGains)), 15);
+			__m128i delay1Gains = _mm_srai_epi32(_mm_madd_epi16(delay1, _mm_loadu_si64(pReverb->Dif2InGains)), 15);
 			__m128i delay1GainsSat = _mm_shuffle_epi32(_mm_packs_epi32(delay1Gains, delay1Gains), _MM_SHUFFLE(2, 0, 2, 0));
 			__m128i histDelay1 = _mm_subs_epi16(_mm_adds_epi16(histDecayInDelay, delay1), delay1GainsSat);	// accumulate with reverb output
 			__m128i diff2out = _mm_subs_epi16(delay1GainsSat, _mm_mulhi_epi16(_mm_cvtsi32_si128(diffusion2), difCoeffs));
 			__m128i diff2outCoeffs = _mm_mulhi_epi16(difCoeffs, diff2out);
 			pReverb->Diffusion2[delayPos].lr = _mm_cvtsi128_si32(diff2out);
 
-			__m128i mixOut = Load64SSE(pMixOut);
+			__m128i mixOut = _mm_loadu_si64(pMixOut);
 			__m128i delay2out = _mm_adds_epi16(diff2outCoeffs, _mm_cvtsi32_si128(diffusion2));
 			pReverb->Delay2[delayPos].lr = _mm_cvtsi128_si32(delay2out);
 			delayPos = (delayPos + 1) & RVBDLY_MASK;
 			// Accumulate with reverb output
 			__m128i out = _mm_add_epi32(_mm_madd_epi16(_mm_adds_epi16(histDelay1, delay2out), rvbOutGains), mixOut);
-			Store64SSE(pMixOut, out);
+			_mm_storeu_si64(pMixOut, out);
 			pMixOut += 2;
 		}
-		Store64SSE(pReverb->LPHistory, lpHistory);
+		_mm_storeu_si64(pReverb->LPHistory, lpHistory);
 		pReverb->nDelayPos = delayPos;
 		return;
 	}




2006	Jan	Feb	Mar (1)	Apr (1)	May	Jun (1)	Jul	Aug (10)	Sep	Oct	Nov	Dec (3)
2007	Jan (1)	Feb (2)	Mar (3)	Apr (2)	May (10)	Jun (2)	Jul (1)	Aug	Sep	Oct	Nov (3)	Dec
2008	Jan (6)	Feb (4)	Mar (5)	Apr (2)	May (1)	Jun (1)	Jul (4)	Aug (6)	Sep (2)	Oct (9)	Nov (1)	Dec (4)
2009	Jan (9)	Feb (2)	Mar (2)	Apr (2)	May (6)	Jun (18)	Jul (33)	Aug (39)	Sep (33)	Oct (24)	Nov (23)	Dec (22)
2010	Jan (29)	Feb (32)	Mar (51)	Apr (17)	May (31)	Jun (21)	Jul (32)	Aug (28)	Sep (35)	Oct (27)	Nov (11)	Dec (13)
2011	Jan (14)	Feb (13)	Mar (27)	Apr (27)	May (28)	Jun (20)	Jul (43)	Aug (52)	Sep (66)	Oct (61)	Nov (11)	Dec (8)
2012	Jan (20)	Feb (30)	Mar (38)	Apr (21)	May (33)	Jun (21)	Jul (25)	Aug (9)	Sep (24)	Oct (42)	Nov (27)	Dec (41)
2013	Jan (20)	Feb (35)	Mar (156)	Apr (298)	May (258)	Jun (201)	Jul (105)	Aug (60)	Sep (193)	Oct (245)	Nov (280)	Dec (194)
2014	Jan (63)	Feb (202)	Mar (200)	Apr (23)	May (53)	Jun (105)	Jul (18)	Aug (26)	Sep (110)	Oct (187)	Nov (97)	Dec (74)
2015	Jan (45)	Feb (55)	Mar (116)	Apr (116)	May (193)	Jun (164)	Jul (50)	Aug (111)	Sep (98)	Oct (71)	Nov (103)	Dec (63)
2016	Jan (33)	Feb (101)	Mar (182)	Apr (139)	May (140)	Jun (103)	Jul (165)	Aug (286)	Sep (208)	Oct (127)	Nov (97)	Dec (54)
2017	Jan (64)	Feb (335)	Mar (202)	Apr (212)	May (139)	Jun (127)	Jul (294)	Aug (154)	Sep (170)	Oct (152)	Nov (156)	Dec (62)
2018	Jan (168)	Feb (237)	Mar (196)	Apr (174)	May (174)	Jun (161)	Jul (127)	Aug (88)	Sep (149)	Oct (66)	Nov (52)	Dec (135)
2019	Jan (146)	Feb (126)	Mar (104)	Apr (58)	May (60)	Jun (28)	Jul (197)	Aug (129)	Sep (141)	Oct (148)	Nov (63)	Dec (100)
2020	Jan (74)	Feb (37)	Mar (59)	Apr (154)	May (194)	Jun (133)	Jul (313)	Aug (197)	Sep (49)	Oct (162)	Nov (143)	Dec (57)
2021	Jan (120)	Feb (107)	Mar (314)	Apr (157)	May (524)	Jun (169)	Jul (72)	Aug (133)	Sep (135)	Oct (146)	Nov (198)	Dec (325)
2022	Jan (409)	Feb (249)	Mar (138)	Apr (95)	May (102)	Jun (221)	Jul (66)	Aug (120)	Sep (192)	Oct (131)	Nov (53)	Dec (171)
2023	Jan (357)	Feb (82)	Mar (168)	Apr (218)	May (196)	Jun (86)	Jul (115)	Aug (49)	Sep (190)	Oct (102)	Nov (45)	Dec (76)
2024	Jan (86)	Feb (50)	Mar (324)	Apr (209)	May (197)	Jun (232)	Jul (194)	Aug (247)	Sep (219)	Oct (266)	Nov (328)	Dec (304)
2025	Jan (191)	Feb (115)	Mar (137)	Apr (32)	May (126)	Jun (403)	Jul	Aug	Sep	Oct	Nov	Dec

modplug-svn-commit Mailing List for modplug (Page 14)

A music tracker software for Windows.

modplug-svn-commit — OpenMPT subversion commit notifications