AutoHotkey Community

Code: Select all

//calculate the sigmoid of the values of the matrix mIn and put them into mOut
//to use the same matrix for input and output you can set the input to the output
void sigmoid(double* mOut, double* mIn, unsigned int w, unsigned int h) {
	//using the tailor series to calculate e^x
	for (unsigned int i =0; i<w*h; i++) {
		double x = mIn[i];
		double result = x + 1;
		double dividend = x;
		double divisor = 1;
		for (double i=2; i<66; i++) { //we could increase precision especially for larger x by increasing the limit here
		//the effect is not major or neccessary though
			dividend *= x;
			divisor *= i;
			result += dividend/divisor;
		}
		mOut[i] = result/(result+1);
	}
}

//calculates the derivative of sigmoid of all values of mIn and puts them into mOut
//assumes that the sigmoid function has run over the input matrix
//to use the same matrix for input and output you can set the input to the output
void sigmoidDerivative(double* mOut, double* mIn, unsigned int w, unsigned int h) {
	for (unsigned int i=0; i<w*h; i++) {
		double var = mIn[i];
		mOut[i] = var * (1 - var);
	}
}

Code: Select all

mov	 ecx, DWORD PTR _w$[esp-4]
imul	 ecx, DWORD PTR _h$[esp-4]
test	 ecx, ecx
je	 SHORT $LN3@sigmoid
mov	 edx, DWORD PTR _mIn$[esp-4]
mov	 eax, DWORD PTR _mOut$[esp-4]
sub	 edx, eax
movsd	 xmm4, QWORD PTR __real@3ff0000000000000
movsd	 xmm7, QWORD PTR __real@4050800000000000
npad	 9
$LL4@sigmoid:
movsd	 xmm6, QWORD PTR [edx+eax]
movaps	 xmm5, xmm4
movsd	 xmm2, QWORD PTR __real@4000000000000000
movaps	 xmm1, xmm6
addsd	 xmm1, xmm4
movaps	 xmm3, xmm6
npad	 6
$LL7@sigmoid:
movaps	 xmm0, xmm2
mulsd	 xmm3, xmm6
mulsd	 xmm5, xmm0
addsd	 xmm2, xmm4
movaps	 xmm0, xmm3
comisd	 xmm7, xmm2
divsd	 xmm0, xmm5
addsd	 xmm1, xmm0
ja	 SHORT $LL7@sigmoid
movaps	 xmm0, xmm1
addsd	 xmm0, xmm4
divsd	 xmm1, xmm0
movsd	 QWORD PTR [eax], xmm1
add	 eax, 8
sub	 ecx, 1
jne	 SHORT $LL4@sigmoid
$LN3@sigmoid:
; Line 18
ret	 0

Code: Select all

mov	 ecx, DWORD PTR _w$[esp-4]
xor	 edx, edx
imul	 ecx, DWORD PTR _h$[esp-4]
push	 ebx
mov	 ebx, DWORD PTR _mOut$[esp]
push	 esi
mov	 esi, DWORD PTR _mIn$[esp+4]
mov	 DWORD PTR _w$[esp+4], ecx
test	 ecx, ecx
je	 $LN16@sigmoidDer
push	 ebp
push	 edi
cmp	 ecx, 8
jb	 $LN9@sigmoidDer
lea	 eax, DWORD PTR [esi-8]
lea	 eax, DWORD PTR [eax+ecx*8]
lea	 edi, DWORD PTR [ebx-8]
lea	 edi, DWORD PTR [edi+ecx*8]
cmp	 ebx, eax
ja	 SHORT $LN10@sigmoidDer
cmp	 edi, esi
jae	 $LN9@sigmoidDer
$LN10@sigmoidDer:
movaps	 xmm2, XMMWORD PTR __xmm@3ff00000000000003ff0000000000000
lea	 eax, DWORD PTR [ebx+16]
mov	 ebp, ecx
lea	 edi, DWORD PTR [esi+48]
mov	 ecx, esi
and	 ebp, -8			; fffffff8H
sub	 ecx, ebx
mov	 DWORD PTR tv1327[esp+12], ecx
mov	 ecx, DWORD PTR _w$[esp+12]
mov	 ebx, DWORD PTR tv1327[esp+12]
$LL4@sigmoidDer:
movups	 xmm1, XMMWORD PTR [edi-48]
add	 edx, 8
movaps	 xmm0, xmm2
subpd	 xmm0, xmm1
mulpd	 xmm0, xmm1
movups	 xmm1, XMMWORD PTR [ebx+eax]
movups	 XMMWORD PTR [eax-16], xmm0
movaps	 xmm0, xmm2
subpd	 xmm0, xmm1
mulpd	 xmm0, xmm1
movups	 xmm1, XMMWORD PTR [edi-16]
movups	 XMMWORD PTR [eax], xmm0
movaps	 xmm0, xmm2
subpd	 xmm0, xmm1
mulpd	 xmm0, xmm1
movups	 xmm1, XMMWORD PTR [edi]
add	 edi, 64			; 00000040H
movups	 XMMWORD PTR [eax+16], xmm0
movaps	 xmm0, xmm2
subpd	 xmm0, xmm1
mulpd	 xmm0, xmm1
movups	 XMMWORD PTR [eax+32], xmm0
add	 eax, 64			; 00000040H
cmp	 edx, ebp
jb	 SHORT $LL4@sigmoidDer
mov	 ebx, DWORD PTR _mOut$[esp+12]
$LN9@sigmoidDer:
cmp	 edx, ecx
jae	 $LN22@sigmoidDer
movsd	 xmm2, QWORD PTR __real@3ff0000000000000
mov	 eax, ecx
sub	 eax, edx
cmp	 eax, 4
jb	 $LC17@sigmoidDer
mov	 eax, esi
lea	 ebp, DWORD PTR [esi+24]
sub	 eax, ebx
lea	 edi, DWORD PTR [edx+1]
mov	 DWORD PTR tv1325[esp+12], eax
lea	 ebp, DWORD PTR [ebp+edx*8]
mov	 esi, DWORD PTR tv1325[esp+12]
lea	 edi, DWORD PTR [ebx+edi*8]
mov	 eax, ecx
sub	 eax, edx
sub	 eax, 4
shr	 eax, 2
inc	 eax
lea	 edx, DWORD PTR [edx+eax*4]
npad	 11
$LL18@sigmoidDer:
movaps	 xmm0, xmm2
subsd	 xmm0, QWORD PTR [ebp-24]
mulsd	 xmm0, QWORD PTR [ebp-24]
movsd	 QWORD PTR [edi-8], xmm0
movaps	 xmm0, xmm2
subsd	 xmm0, QWORD PTR [esi+edi]
mulsd	 xmm0, QWORD PTR [esi+edi]
movsd	 QWORD PTR [edi], xmm0
movaps	 xmm0, xmm2
subsd	 xmm0, QWORD PTR [ebp-8]
mulsd	 xmm0, QWORD PTR [ebp-8]
movsd	 QWORD PTR [edi+8], xmm0
movaps	 xmm0, xmm2
subsd	 xmm0, QWORD PTR [ebp]
mulsd	 xmm0, QWORD PTR [ebp]
add	 ebp, 32			; 00000020H
movsd	 QWORD PTR [edi+16], xmm0
add	 edi, 32			; 00000020H
sub	 eax, 1
jne	 SHORT $LL18@sigmoidDer
mov	 esi, DWORD PTR _mIn$[esp+12]
$LC17@sigmoidDer:
cmp	 edx, ecx
jae	 SHORT $LN22@sigmoidDer
sub	 esi, ebx
lea	 eax, DWORD PTR [ebx+edx*8]
sub	 ecx, edx
$LC8@sigmoidDer:
movaps	 xmm0, xmm2
subsd	 xmm0, QWORD PTR [esi+eax]
mulsd	 xmm0, QWORD PTR [esi+eax]
movsd	 QWORD PTR [eax], xmm0
add	 eax, 8
sub	 ecx, 1
jne	 SHORT $LC8@sigmoidDer
$LN22@sigmoidDer:
pop	 edi
pop	 ebp
$LN16@sigmoidDer:
pop	 esi
pop	 ebx
ret	 0