Changeset 134867 in webkit

Timestamp:

Nov 15, 2012 5:19:50 PM (11 years ago)

Author:

commit-queue@webkit.org

Message:

Optimize the multiply-add in Biquad.cpp::process
​https://bugs.webkit.org/show_bug.cgi?id=75528

Patch by Xingnan Wang <​xingnan.wang@intel.com> on 2012-11-15
Reviewed by Brent Fulgham.

Pipeline the multiply-add with SSE2 intrinsics.
Get ~45% performance improvement for the function.

• platform/audio/Biquad.cpp:

(WebCore::Biquad::process):

Location:

trunk/Source/WebCore

Files:

• ChangeLog (modified) (1 diff)
• platform/audio/Biquad.cpp (modified) (3 diffs)

trunk/Source/WebCore/ChangeLog

@@ +1 @@
+2012-11-15  Xingnan Wang  <xingnan.wang@intel.com>
+
+        Optimize the multiply-add in Biquad.cpp::process
+        https://bugs.webkit.org/show_bug.cgi?id=75528
+
+        Reviewed by Brent Fulgham.
+
+        Pipeline the multiply-add with SSE2 intrinsics.
+        Get ~45% performance improvement for the function.
+
+        * platform/audio/Biquad.cpp:
+        (WebCore::Biquad::process):
+
 2012-11-15  Alec Flett  <alecflett@chromium.org>
 

trunk/Source/WebCore/platform/audio/Biquad.cpp

@@ -42 +42 @@
 #endif
 
+#ifdef __SSE2__
+#include <emmintrin.h>
+#endif
+
 namespace WebCore {
 
@@ -97 +101 @@
     double a2 = m_a2;
 
+    // Optimize the hot multiply-add by pipelining with SSE2 intrinsics.
+#ifdef __SSE2__
+    __m128d mm0 = _mm_set_pd(x1, x2); // mm0 = (x1, x2)
+    __m128d mm1 = _mm_set_pd(y2, static_cast<double>(*sourceP)); // mm1 = (y2, x)
+    __m128d mm2 = _mm_set_pd(b1, b2); // mm2 = (b1, b2)
+    __m128d mm3 = _mm_set_pd(-a2, b0); // mm3 = (-a2, b0)
+    __m128d mm4 = _mm_set_sd(y1); // mm4 = y1, only use low part of mm4.
+    __m128d mm5;
+    __m128d mm6;
+    __m128 mm7; // Only use low part of mm7.
+    __m128d mma1 = _mm_set_sd(-a1); // mma1 = -a1, only use low part of mma1.
+
+    while (n) {
+        sourceP++;
+        mm6 = mm1; // mm6 = (y2, x)
+        mm1 = _mm_shuffle_pd(mm1, mm4, 0); // mm1 = (y1, x)
+        mm5 = _mm_mul_pd(mm2, mm0); // mm5 = (x1 * b1, x2 * b2)
+        mm6 = _mm_mul_pd(mm3, mm6); // mm6 = (-y2 * a2, x * b0)
+        mm0 = _mm_shuffle_pd(mm0, mm1, 1); // mm0 = (x, x1)
+        mm4 = _mm_mul_sd(mm4, mma1); // mm4 = -y1 * a1
+        mm5 = _mm_add_pd(mm5, mm6); // mm5 = (x1 * b1 - y2 * a2, x2 * b2 + x * b0)
+        n--;
+        mm6 = mm5; // mm6 = (x1 * b1 - y2 * a2, x2 * b2 + x * b0)
+        mm7 = _mm_load_ss(sourceP); // mm7 = *sourceP, load next x value.
+        mm1 = _mm_cvtss_sd(mm1, mm7); // mm1 = (y1, x)
+        mm5 = _mm_add_sd(mm4, mm5); // mm5 = x2 * b2 + x * b0 - y1 * a1, only care low part of mm5 here.
+        mm6 = _mm_shuffle_pd(mm6, mm6, 1); // mm6 = (x2 * b2 + x * b0, x1 * b1 - y2 * a2)
+        mm5 = _mm_add_sd(mm5, mm6); // mm5 = x * b0 + x1 * b1 + x2 * b2 - y1 * a1 - y2 * a2 = y, only care low part of mm5.
+        mm7 = _mm_cvtsd_ss(mm7, mm5); // mm7 = static_cast<float>(y)
+        _mm_store_ss(destP, mm7); // Store y to destP
+        mm4 = mm5; // mm4 = y
+        destP++;
+    }
+    _mm_storeh_pd(&x1, mm0);
+    _mm_storel_pd(&x2, mm0);
+    _mm_storel_pd(&y1, mm4);
+    _mm_storeh_pd(&y2, mm1);
+#else
     while (n--) {
-        // FIXME: this can be optimized by pipelining the multiply adds...
         float x = *sourceP++;
         float y = b0*x + b1*x1 + b2*x2 - a1*y1 - a2*y2;
@@ -110 +151 @@
         y1 = y;
     }
+#endif // __SSE2__
 
     // Local variables back to member. Flush denormals here so we