Modified: trunk/Source/WebCore/platform/audio/Biquad.cpp (110743 => 110744)
--- trunk/Source/WebCore/platform/audio/Biquad.cpp 2012-03-14 20:59:53 UTC (rev 110743)
+++ trunk/Source/WebCore/platform/audio/Biquad.cpp 2012-03-14 21:01:37 UTC (rev 110744)
@@ -96,6 +96,60 @@
double a1 = m_a1;
double a2 = m_a2;
+// Optimize the hot multiply-add by pipelining with SSE2 instructions.
+#ifdef __SSE2__
+ double na1 = -a1;
+ double na2 = -a2;
+
+ __asm__(
+ "movl %4, %%edx\n\t" // move sourceP to edx
+ "movl %5, %%ecx\n\t" // move destP to ecx
+ "movl %6, %%eax\n\t" // move n to eax
+ "testl %%eax, %%eax\n\t"
+ "je .LabelEnd\n\t"
+ "movss (%%edx), %%xmm7\n\t" // load x to xmm7[63:0]
+ "cvtss2sd %%xmm7, %%xmm1\n\t" // convert x from float to double
+ "movlpd %1, %%xmm0\n\t" // move x2 to xmm0[63:0]
+ "movlpd %9, %%xmm2\n\t" // move b2 to xmm2[63:0]
+ "movlpd %7, %%xmm3\n\t" // move b0 to xmm3[63:0]
+ "movhpd %0, %%xmm0\n\t" // move x1 to xmm0[127:64] ----> (x1 x2)
+ "movhpd %3, %%xmm1\n\t" // move y2 to xmm1[127:64] ----> (y2 x )
+ "movhpd %8, %%xmm2\n\t" // move b1 to xmm2[127:64] ----> (b1 b2)
+ "movhpd %11, %%xmm3\n\t" // move a2 to xmm3[127:64] ----> (a2 b0)
+ "movlpd %2, %%xmm4\n\t" // move y1 to xmm4[63:0]
+ ".LabelLoop:\n\t"
+ "addl $4, %%edx\n\t" // sourceP++
+ "movapd %%xmm0, %%xmm5\n\t" // copy (x1 x2)
+ "movapd %%xmm1, %%xmm6\n\t" // copy (y2 x )
+ "shufpd $0, %%xmm4, %%xmm1\n\t" // y2=y1
+ "mulpd %%xmm2, %%xmm5\n\t" // (x1*b1 x2*b2)
+ "mulpd %%xmm3, %%xmm6\n\t" // (y2*a2 x *b0)
+ "shufpd $1, %%xmm1, %%xmm0\n\t" // x2=x1 x1=x
+ "mulsd %10, %%xmm4\n\t" // a1*y1
+ "addpd %%xmm6, %%xmm5\n\t" // (x1*b1+y2*a2 x2*b2+x*b0)
+ "subl $1, %%eax\n\t" // n--
+ "movapd %%xmm5, %%xmm6\n\t"
+ "movss (%%edx), %%xmm7\n\t" // load x
+ "cvtss2sd %%xmm7, %%xmm1\n\t" // cvt x from float to double x = new x
+ "addsd %%xmm4, %%xmm5\n\t" // a1*y1 + (x2*b2+x*b0)
+ "shufpd $1, %%xmm6, %%xmm6\n\t" // (x1*b1+y2*a2 x2*b2+x*b0) -> (x2*b2+x*b0 x1*b1+y2*a2)
+ "addsd %%xmm6, %%xmm5\n\t" // y
+ "cvtsd2ss %%xmm5, %%xmm7\n\t"
+ "movss %%xmm7, (%%ecx)\n\t" // y -> *destP
+ "movapd %%xmm5, %%xmm4\n\t" // y1 = y
+ "addl $4, %%ecx\n\t" // destP++
+ "testl %%eax, %%eax\n\t"
+ "jne .LabelLoop\n\t" // while()
+ "movhpd %%xmm0, %0\n\t"
+ "movlpd %%xmm0, %1\n\t"
+ "movlpd %%xmm4, %2\n\t"
+ "movhpd %%xmm1, %3\n\t"
+ ".LabelEnd:\n\t"
+ :"+m"(x1), "+m"(x2), "+m"(y1), "+m"(y2)
+ :"m"(sourceP), "m"(destP), "m"(n), "m"(b0), "m"(b1), "m"(b2), "m"(na1), "m"(na2)
+ :"eax", "edx", "ecx", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+ );
+#else
while (n--) {
// FIXME: this can be optimized by pipelining the multiply adds...
float x = *sourceP++;
@@ -109,6 +163,7 @@
y2 = y1;
y1 = y;
}
+#endif
// Local variables back to member. Flush denormals here so we
// don't slow down the inner loop above.
