From fd8574a845df9fa9848675f3aec8eb22bb4b4e83 Mon Sep 17 00:00:00 2001 From: Stenzek Date: Sun, 4 Oct 2026 16:17:40 +1000 Subject: [PATCH] SPU: Pack reverb FIR coefficients into 16-bit lanes Use three packed multiply-adds for the 20-tap upsampling FIR and widen the shared coefficients for downsampling. Preserve the final shift and clamp. The absolute coefficient sum is 31142, keeping every s32 sum within range. Validated output, persistent state, and RAM against the original across 8,192 frames using both SSE4.1 and SSE2; compiled with clang-cl /O2. --- src/core/spu.cpp | 39 ++++++++++++++++++++------------------- 1 file changed, 20 insertions(+), 19 deletions(-) diff --git a/src/core/spu.cpp b/src/core/spu.cpp index b08824121..5c7717e84 100644 --- a/src/core/spu.cpp +++ b/src/core/spu.cpp @@ -2279,7 +2279,7 @@ void SPU::ProcessReverb(s32 left_in, s32 right_in, s32* left_out, s32* right_out // // Zeros have been removed since the result is always zero, therefore the multiply is redundant. - alignas(VECTOR_ALIGNMENT) static constexpr std::array resample_coeff = { + alignas(VECTOR_ALIGNMENT) static constexpr std::array resample_coeff = { -0x0001, 0x0002, -0x000A, 0x0023, -0x0067, 0x010A, -0x0268, 0x0534, -0x0B90, 0x2806, 0x2806, -0x0B90, 0x0534, -0x0268, 0x010A, -0x0067, 0x0023, -0x000A, 0x0002, -0x0001}; @@ -2310,16 +2310,21 @@ void SPU::ProcessReverb(s32 left_in, s32 right_in, s32* left_out, s32* right_out { const s16* src = &s_state.reverb_downsample_buffer[channel][(s_state.reverb_resample_buffer_position - 38) & 0x3F]; - GSVector4i acc = - GSVector4i::load(&resample_coeff[0]).mul32l(GSVector4i::load(&src[0]).sll32(16).sra32(16)); - acc = acc.add32( - GSVector4i::load(&resample_coeff[4]).mul32l(GSVector4i::load(&src[8]).sll32(16).sra32(16))); - acc = acc.add32( - GSVector4i::load(&resample_coeff[8]).mul32l(GSVector4i::load(&src[16]).sll32(16).sra32(16))); - acc = acc.add32( - GSVector4i::load(&resample_coeff[12]).mul32l(GSVector4i::load(&src[24]).sll32(16).sra32(16))); - acc = acc.add32( - GSVector4i::load(&resample_coeff[16]).mul32l(GSVector4i::load(&src[32]).sll32(16).sra32(16))); + GSVector4i acc = GSVector4i::loadl(&resample_coeff[0]) + .s16to32() + .mul32l(GSVector4i::load(&src[0]).sll32(16).sra32(16)); + acc = acc.add32(GSVector4i::loadl(&resample_coeff[4]) + .s16to32() + .mul32l(GSVector4i::load(&src[8]).sll32(16).sra32(16))); + acc = acc.add32(GSVector4i::loadl(&resample_coeff[8]) + .s16to32() + .mul32l(GSVector4i::load(&src[16]).sll32(16).sra32(16))); + acc = acc.add32(GSVector4i::loadl(&resample_coeff[12]) + .s16to32() + .mul32l(GSVector4i::load(&src[24]).sll32(16).sra32(16))); + acc = acc.add32(GSVector4i::loadl(&resample_coeff[16]) + .s16to32() + .mul32l(GSVector4i::load(&src[32]).sll32(16).sra32(16))); // Horizontal reduction, middle 0x4000. Moved here so we don't need another 4 elements above. downsampled[channel] = Clamp16((acc.addv_s32() + (0x4000 * src[19])) >> 15); @@ -2398,14 +2403,10 @@ void SPU::ProcessReverb(s32 left_in, s32 right_in, s32* left_out, s32* right_out const s16* src = &s_state.reverb_upsample_buffer[channel][((s_state.reverb_resample_buffer_position >> 1) - 19) & 0x1F]; - GSVector4i srcs = GSVector4i::load(&src[0]); - GSVector4i acc = GSVector4i::load(&resample_coeff[0]).mul32l(srcs.s16to32()); - acc = acc.add32(GSVector4i::load(&resample_coeff[4]).mul32l(srcs.uph64().s16to32())); - srcs = GSVector4i::load(&src[8]); - acc = acc.add32(GSVector4i::load(&resample_coeff[8]).mul32l(srcs.s16to32())); - acc = acc.add32(GSVector4i::load(&resample_coeff[12]).mul32l(srcs.uph64().s16to32())); - srcs = GSVector4i::loadl(&src[16]); - acc = acc.add32(GSVector4i::load(&resample_coeff[16]).mul32l(srcs.s16to32())); + // The absolute coefficient sum is 31142, so s32 accumulation cannot overflow. + GSVector4i acc = GSVector4i::load(&src[0]).madd_s16(GSVector4i::load(&resample_coeff[0])); + acc = acc.add32(GSVector4i::load(&src[8]).madd_s16(GSVector4i::load(&resample_coeff[8]))); + acc = acc.add32(GSVector4i::loadl(&src[16]).madd_s16(GSVector4i::loadl(&resample_coeff[16]))); out[channel] = std::clamp(acc.addv_s32() >> 14, -32768, 32767); }