Unroll and inline strings SSE and AVX

src/CMakeLists.txt

@@ -43,8 +43,11 @@ set (SFIZZ_HEADERS
     sfizz/effects/impl/ResonantArrayAVX.h
     sfizz/effects/impl/ResonantArraySSE.h
     sfizz/effects/impl/ResonantString.h
+    sfizz/effects/impl/ResonantStringInline.h
     sfizz/effects/impl/ResonantStringAVX.h
+    sfizz/effects/impl/ResonantStringAVXInline.h
     sfizz/effects/impl/ResonantStringSSE.h
+    sfizz/effects/impl/ResonantStringSSEInline.h
     sfizz/effects/Apan.h
     sfizz/effects/CommonLFO.h
     sfizz/effects/CommonLFO.hpp

src/sfizz/effects/impl/ResonantArray.cpp

@@ -6,6 +6,7 @@
 
 #include "ResonantArray.h"
 #include "ResonantString.h"
+#include "ResonantStringInline.h"
 #include "SIMDHelpers.h"
 
 namespace sfz {

src/sfizz/effects/impl/ResonantArrayAVX.cpp

@@ -5,6 +5,7 @@
 // If not, contact the sfizz maintainers at https://github.com/sfztools/sfizz
 
 #include "ResonantArrayAVX.h"
+#include "ResonantStringAVXInline.h"
 #include "Config.h"
 #include <cstring>
 
@@ -82,7 +83,29 @@ void ResonantArrayAVX::process(const float *inPtr, float *outPtr, unsigned numFr
     __m256* outputs8 = reinterpret_cast<__m256*>(_workBuffer.data());
     std::memset(outputs8, 0, numFrames * sizeof(__m256));
 
-    for (unsigned p = 0; p < numStringPacks; ++p) {
+    unsigned p = 0;
+    for (; p + 7 < numStringPacks; p += 8) {
+        ResonantStringAVX& rs = reinterpret_cast<ResonantStringAVX&>(stringPacks[p]);
+        for (unsigned i = 0; i < numFrames; ++i) {
+            __m256 o1 = rs.process(_mm256_broadcast_ss(&inPtr[i]));
+            __m256 o2 = rs.process(_mm256_broadcast_ss(&inPtr[i + 1]));
+            __m256 o3 = rs.process(_mm256_broadcast_ss(&inPtr[i + 2]));
+            __m256 o4 = rs.process(_mm256_broadcast_ss(&inPtr[i + 3]));
+            __m256 o5 = rs.process(_mm256_broadcast_ss(&inPtr[i + 4]));
+            __m256 o6 = rs.process(_mm256_broadcast_ss(&inPtr[i + 5]));
+            __m256 o7 = rs.process(_mm256_broadcast_ss(&inPtr[i + 6]));
+            __m256 o8 = rs.process(_mm256_broadcast_ss(&inPtr[i + 7]));
+            __m256 output8 = outputs8[i];
+            o1 = _mm256_add_ps(o1, o2);
+            o2 = _mm256_add_ps(o3, o4);
+            o3 = _mm256_add_ps(o5, o6);
+            o4 = _mm256_add_ps(o7, o8);
+            o1 = _mm256_add_ps(o1, o2);
+            o2 = _mm256_add_ps(o3, o4);
+            outputs8[i] = _mm256_add_ps(o1, _mm256_add_ps(output8, o2));
+        }
+    }
+    for (; p < numStringPacks; ++p) {
         ResonantStringAVX& rs = reinterpret_cast<ResonantStringAVX&>(stringPacks[p]);
         for (unsigned i = 0; i < numFrames; ++i)
             outputs8[i] = _mm256_add_ps(

src/sfizz/effects/impl/ResonantArraySSE.cpp

@@ -5,6 +5,7 @@
 // If not, contact the sfizz maintainers at https://github.com/sfztools/sfizz
 
 #include "ResonantArraySSE.h"
+#include "ResonantStringSSEInline.h"
 #include "Config.h"
 #include <cstring>
 
@@ -82,7 +83,29 @@ void ResonantArraySSE::process(const float *inPtr, float *outPtr, unsigned numFr
     __m128* outputs4 = reinterpret_cast<__m128*>(_workBuffer.data());
     std::memset(outputs4, 0, numFrames * sizeof(__m128));
 
-    for (unsigned p = 0; p < numStringPacks; ++p) {
+    unsigned p = 0;
+    for (; p + 7 < numStringPacks; p += 8) {
+        ResonantStringSSE& rs = reinterpret_cast<ResonantStringSSE&>(stringPacks[p]);
+        for (unsigned i = 0; i < numFrames; ++i) {
+            __m128 o1 = rs.process(_mm_load1_ps(&inPtr[i]));
+            __m128 o2 = rs.process(_mm_load1_ps(&inPtr[i + 1]));
+            __m128 o3 = rs.process(_mm_load1_ps(&inPtr[i + 2]));
+            __m128 o4 = rs.process(_mm_load1_ps(&inPtr[i + 3]));
+            __m128 o5 = rs.process(_mm_load1_ps(&inPtr[i + 4]));
+            __m128 o6 = rs.process(_mm_load1_ps(&inPtr[i + 5]));
+            __m128 o7 = rs.process(_mm_load1_ps(&inPtr[i + 6]));
+            __m128 o8 = rs.process(_mm_load1_ps(&inPtr[i + 7]));
+            __m128 output4 = outputs4[i];
+            o1 = _mm_add_ps(o1, o2);
+            o2 = _mm_add_ps(o3, o4);
+            o3 = _mm_add_ps(o5, o6);
+            o4 = _mm_add_ps(o7, o8);
+            o1 = _mm_add_ps(o1, o2);
+            o2 = _mm_add_ps(o3, o4);
+            outputs4[i] = _mm_add_ps(o1, _mm_add_ps(output4, o2));
+        }
+    }
+    for (; p < numStringPacks; ++p) {
         ResonantStringSSE& rs = stringPacks[p];
         for (unsigned i = 0; i < numFrames; ++i)
             outputs4[i] = _mm_add_ps(

src/sfizz/effects/impl/ResonantString.cpp

@@ -85,19 +85,5 @@ void ResonantString::setResonanceFrequency(float frequency, float bandwidth)
     fControl[17] = (fControl[10] + (4.0f - fControl[11]));
 }
 
-float ResonantString::process(float input)
-{
-    fRec0[0] = (fControl[1] * ((fControl[4] * fRec1[1]) + (fControl[5] * fRec0[1])));
-    float fTemp0 = input;
-    fRec2[0] = (fTemp0 - (fControl[15] * ((fControl[16] * fRec2[1]) + (fControl[17] * fRec2[2]))));
-    fRec1[0] = (((fControl[14] * fRec2[2]) + ((fControl[5] * fRec1[1]) + (fControl[13] * fRec2[0]))) - (fControl[4] * fRec0[1]));
-    float output = float((fControl[0] * fRec0[0]));
-    fRec0[1] = fRec0[0];
-    fRec2[2] = fRec2[1];
-    fRec2[1] = fRec2[0];
-    fRec1[1] = fRec1[0];
-    return output;
-}
-
 } // namespace sfz
 } // namespace fx

src/sfizz/effects/impl/ResonantStringAVX.cpp

@@ -110,20 +110,6 @@ void ResonantStringAVX::setResonanceFrequency(__m256 frequency, __m256 bandwidth
     fControl[17] = _mm256_add_ps(fControl[10], _mm256_sub_ps(_mm256_set1_ps(4.0f), fControl[11]));
 }
 
-__m256 ResonantStringAVX::process(__m256 input)
-{
-    fRec0[0] = _mm256_mul_ps(fControl[1], _mm256_add_ps(_mm256_mul_ps(fControl[4], fRec1[1]), _mm256_mul_ps(fControl[5], fRec0[1])));
-    __m256 fTemp0 = input;
-    fRec2[0] = _mm256_sub_ps(fTemp0, _mm256_mul_ps(fControl[15], _mm256_add_ps(_mm256_mul_ps(fControl[16], fRec2[1]), _mm256_mul_ps(fControl[17], fRec2[2]))));
-    fRec1[0] = _mm256_sub_ps(_mm256_add_ps(_mm256_mul_ps(fControl[14], fRec2[2]), _mm256_add_ps(_mm256_mul_ps(fControl[5], fRec1[1]), _mm256_mul_ps(fControl[13], fRec2[0]))),_mm256_mul_ps(fControl[4], fRec0[1]));
-    __m256 output = _mm256_mul_ps(fControl[0], fRec0[0]);
-    fRec0[1] = fRec0[0];
-    fRec2[2] = fRec2[1];
-    fRec2[1] = fRec2[0];
-    fRec1[1] = fRec1[0];
-    return output;
-}
-
 } // namespace sfz
 } // namespace fx
 #endif

src/sfizz/effects/impl/ResonantStringAVXInline.h

@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: BSD-2-Clause
+
+// This code is part of the sfizz library and is licensed under a BSD 2-clause
+// license. You should have receive a LICENSE.md file along with the code.
+// If not, contact the sfizz maintainers at https://github.com/sfztools/sfizz
+
+#pragma once
+#include "ResonantStringAVX.h"
+
+namespace sfz {
+namespace fx {
+
+inline __m256 ResonantStringAVX::process(__m256 input)
+{
+    fRec0[0] = _mm256_mul_ps(fControl[1], _mm256_add_ps(_mm256_mul_ps(fControl[4], fRec1[1]), _mm256_mul_ps(fControl[5], fRec0[1])));
+    __m256 fTemp0 = input;
+    fRec2[0] = _mm256_sub_ps(fTemp0, _mm256_mul_ps(fControl[15], _mm256_add_ps(_mm256_mul_ps(fControl[16], fRec2[1]), _mm256_mul_ps(fControl[17], fRec2[2]))));
+    fRec1[0] = _mm256_sub_ps(_mm256_add_ps(_mm256_mul_ps(fControl[14], fRec2[2]), _mm256_add_ps(_mm256_mul_ps(fControl[5], fRec1[1]), _mm256_mul_ps(fControl[13], fRec2[0]))),_mm256_mul_ps(fControl[4], fRec0[1]));
+    __m256 output = _mm256_mul_ps(fControl[0], fRec0[0]);
+    fRec0[1] = fRec0[0];
+    fRec2[2] = fRec2[1];
+    fRec2[1] = fRec2[0];
+    fRec1[1] = fRec1[0];
+    return output;
+}
+
+} // namespace sfz
+} // namespace fx

src/sfizz/effects/impl/ResonantStringInline.h

@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: BSD-2-Clause
+
+// This code is part of the sfizz library and is licensed under a BSD 2-clause
+// license. You should have receive a LICENSE.md file along with the code.
+// If not, contact the sfizz maintainers at https://github.com/sfztools/sfizz
+
+#include "ResonantString.h"
+
+namespace sfz {
+namespace fx {
+
+inline float ResonantString::process(float input)
+{
+    fRec0[0] = (fControl[1] * ((fControl[4] * fRec1[1]) + (fControl[5] * fRec0[1])));
+    float fTemp0 = input;
+    fRec2[0] = (fTemp0 - (fControl[15] * ((fControl[16] * fRec2[1]) + (fControl[17] * fRec2[2]))));
+    fRec1[0] = (((fControl[14] * fRec2[2]) + ((fControl[5] * fRec1[1]) + (fControl[13] * fRec2[0]))) - (fControl[4] * fRec0[1]));
+    float output = float((fControl[0] * fRec0[0]));
+    fRec0[1] = fRec0[0];
+    fRec2[2] = fRec2[1];
+    fRec2[1] = fRec2[0];
+    fRec1[1] = fRec1[0];
+    return output;
+}
+
+} // namespace sfz
+} // namespace fx

src/sfizz/effects/impl/ResonantStringSSE.cpp

@@ -110,20 +110,6 @@ void ResonantStringSSE::setResonanceFrequency(__m128 frequency, __m128 bandwidth
     fControl[17] = _mm_add_ps(fControl[10], _mm_sub_ps(_mm_set1_ps(4.0f), fControl[11]));
 }
 
-__m128 ResonantStringSSE::process(__m128 input)
-{
-    fRec0[0] = _mm_mul_ps(fControl[1], _mm_add_ps(_mm_mul_ps(fControl[4], fRec1[1]), _mm_mul_ps(fControl[5], fRec0[1])));
-    __m128 fTemp0 = input;
-    fRec2[0] = _mm_sub_ps(fTemp0, _mm_mul_ps(fControl[15], _mm_add_ps(_mm_mul_ps(fControl[16], fRec2[1]), _mm_mul_ps(fControl[17], fRec2[2]))));
-    fRec1[0] = _mm_sub_ps(_mm_add_ps(_mm_mul_ps(fControl[14], fRec2[2]), _mm_add_ps(_mm_mul_ps(fControl[5], fRec1[1]), _mm_mul_ps(fControl[13], fRec2[0]))),_mm_mul_ps(fControl[4], fRec0[1]));
-    __m128 output = _mm_mul_ps(fControl[0], fRec0[0]);
-    fRec0[1] = fRec0[0];
-    fRec2[2] = fRec2[1];
-    fRec2[1] = fRec2[0];
-    fRec1[1] = fRec1[0];
-    return output;
-}
-
 } // namespace sfz
 } // namespace fx
 #endif

src/sfizz/effects/impl/ResonantStringSSEInline.h

@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: BSD-2-Clause
+
+// This code is part of the sfizz library and is licensed under a BSD 2-clause
+// license. You should have receive a LICENSE.md file along with the code.
+// If not, contact the sfizz maintainers at https://github.com/sfztools/sfizz
+
+#pragma once
+#include "ResonantStringSSE.h"
+
+namespace sfz {
+namespace fx {
+
+inline __m128 ResonantStringSSE::process(__m128 input)
+{
+    fRec0[0] = _mm_mul_ps(fControl[1], _mm_add_ps(_mm_mul_ps(fControl[4], fRec1[1]), _mm_mul_ps(fControl[5], fRec0[1])));
+    __m128 fTemp0 = input;
+    fRec2[0] = _mm_sub_ps(fTemp0, _mm_mul_ps(fControl[15], _mm_add_ps(_mm_mul_ps(fControl[16], fRec2[1]), _mm_mul_ps(fControl[17], fRec2[2]))));
+    fRec1[0] = _mm_sub_ps(_mm_add_ps(_mm_mul_ps(fControl[14], fRec2[2]), _mm_add_ps(_mm_mul_ps(fControl[5], fRec1[1]), _mm_mul_ps(fControl[13], fRec2[0]))),_mm_mul_ps(fControl[4], fRec0[1]));
+    __m128 output = _mm_mul_ps(fControl[0], fRec0[0]);
+    fRec0[1] = fRec0[0];
+    fRec2[2] = fRec2[1];
+    fRec2[1] = fRec2[0];
+    fRec1[1] = fRec1[0];
+    return output;
+}
+
+} // namespace sfz
+} // namespace fx