mirror of
https://github.com/python-pillow/Pillow.git
synced 2025-08-21 04:34:47 +03:00
SIMD Filter. a bit faster 5x5 filter
This commit is contained in:
Alexander
Alexander Karpinsky
parent
c07157fd4b
commit
a84ad68ca8
|
|
@ -155,7 +155,7 @@ ImagingFilter3x3(Imaging imOut, Imaging im, const float *kernel, float offset) {
|
||||||
|
|
||||||
out[0] = in0[0];
|
out[0] = in0[0];
|
||||||
x = 1;
|
x = 1;
|
||||||
for (; x < im->xsize-3; x += 3) {
|
for (; x < im->xsize-1-2; x += 3) {
|
||||||
__m128 ss;
|
__m128 ss;
|
||||||
__m128i ssi0, ssi1, ssi2;
|
__m128i ssi0, ssi1, ssi2;
|
||||||
|
|
||||||
|
|
@ -205,11 +205,9 @@ ImagingFilter3x3(Imaging imOut, Imaging im, const float *kernel, float offset) {
|
||||||
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[8]), pix21));
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[8]), pix21));
|
||||||
ssi2 = _mm_cvtps_epi32(ss);
|
ssi2 = _mm_cvtps_epi32(ss);
|
||||||
|
|
||||||
|
|
||||||
ssi0 = _mm_packs_epi32(ssi0, ssi1);
|
ssi0 = _mm_packs_epi32(ssi0, ssi1);
|
||||||
ssi1 = _mm_packs_epi32(ssi2, ssi2);
|
ssi1 = _mm_packs_epi32(ssi2, ssi2);
|
||||||
ssi0 = _mm_packus_epi16(ssi0, ssi1);
|
ssi0 = _mm_packus_epi16(ssi0, ssi1);
|
||||||
out[x] = _mm_cvtsi128_si32(ssi0);
|
|
||||||
_mm_storeu_si128((__m128i*) &out[x], ssi0);
|
_mm_storeu_si128((__m128i*) &out[x], ssi0);
|
||||||
}
|
}
|
||||||
for (; x < im->xsize-1; x++) {
|
for (; x < im->xsize-1; x++) {
|
||||||
|
|
@ -288,10 +286,209 @@ ImagingFilter5x5(Imaging imOut, Imaging im, const float *kernel, float offset) {
|
||||||
UINT32* in1 = (UINT32*) im->image[y+1];
|
UINT32* in1 = (UINT32*) im->image[y+1];
|
||||||
UINT32* in2 = (UINT32*) im->image[y+2];
|
UINT32* in2 = (UINT32*) im->image[y+2];
|
||||||
UINT32* out = (UINT32*) imOut->image[y];
|
UINT32* out = (UINT32*) imOut->image[y];
|
||||||
|
__m128 pix00 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in2)[0]));
|
||||||
|
__m128 pix01 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in2)[1]));
|
||||||
|
__m128 pix02 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in2)[2]));
|
||||||
|
__m128 pix03 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in2)[3]));
|
||||||
|
__m128 pix10 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in1)[0]));
|
||||||
|
__m128 pix11 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in1)[1]));
|
||||||
|
__m128 pix12 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in1)[2]));
|
||||||
|
__m128 pix13 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in1)[3]));
|
||||||
|
__m128 pix20 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[0]));
|
||||||
|
__m128 pix21 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[1]));
|
||||||
|
__m128 pix22 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[2]));
|
||||||
|
__m128 pix23 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[3]));
|
||||||
|
__m128 pix30 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_1)[0]));
|
||||||
|
__m128 pix31 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_1)[1]));
|
||||||
|
__m128 pix32 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_1)[2]));
|
||||||
|
__m128 pix33 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_1)[3]));
|
||||||
|
__m128 pix40 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_2)[0]));
|
||||||
|
__m128 pix41 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_2)[1]));
|
||||||
|
__m128 pix42 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_2)[2]));
|
||||||
|
__m128 pix43 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_2)[3]));
|
||||||
|
__m128 pix04, pix14, pix24, pix34, pix44;
|
||||||
|
|
||||||
out[0] = in0[0];
|
out[0] = in0[0];
|
||||||
out[1] = in0[1];
|
out[1] = in0[1];
|
||||||
for (x = 2; x < im->xsize-2; x++) {
|
x = 2;
|
||||||
|
for (; x < im->xsize-2-4; x += 5) {
|
||||||
|
__m128 ss;
|
||||||
|
__m128i ssi0, ssi1, ssi2, ssi3;
|
||||||
|
|
||||||
|
ss = _mm_set1_ps(offset);
|
||||||
|
pix04 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in2)[x+2]));
|
||||||
|
pix14 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in1)[x+2]));
|
||||||
|
pix24 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[x+2]));
|
||||||
|
pix34 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_1)[x+2]));
|
||||||
|
pix44 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_2)[x+2]));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[0]), pix00));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[5]), pix10));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[10]), pix20));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[15]), pix30));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[20]), pix40));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[1]), pix01));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[6]), pix11));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[11]), pix21));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[16]), pix31));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[21]), pix41));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[2]), pix02));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[7]), pix12));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[12]), pix22));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[17]), pix32));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[22]), pix42));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[3]), pix03));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[8]), pix13));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[13]), pix23));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[18]), pix33));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[23]), pix43));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[4]), pix04));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[9]), pix14));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[14]), pix24));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[19]), pix34));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[24]), pix44));
|
||||||
|
ssi0 = _mm_cvtps_epi32(ss);
|
||||||
|
|
||||||
|
ss = _mm_set1_ps(offset);
|
||||||
|
pix00 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in2)[x+3]));
|
||||||
|
pix10 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in1)[x+3]));
|
||||||
|
pix20 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[x+3]));
|
||||||
|
pix30 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_1)[x+3]));
|
||||||
|
pix40 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_2)[x+3]));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[0]), pix01));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[5]), pix11));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[10]), pix21));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[15]), pix31));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[20]), pix41));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[1]), pix02));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[6]), pix12));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[11]), pix22));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[16]), pix32));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[21]), pix42));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[2]), pix03));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[7]), pix13));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[12]), pix23));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[17]), pix33));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[22]), pix43));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[3]), pix04));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[8]), pix14));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[13]), pix24));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[18]), pix34));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[23]), pix44));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[4]), pix00));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[9]), pix10));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[14]), pix20));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[19]), pix30));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[24]), pix40));
|
||||||
|
ssi1 = _mm_cvtps_epi32(ss);
|
||||||
|
|
||||||
|
ss = _mm_set1_ps(offset);
|
||||||
|
pix01 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in2)[x+4]));
|
||||||
|
pix11 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in1)[x+4]));
|
||||||
|
pix21 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[x+4]));
|
||||||
|
pix31 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_1)[x+4]));
|
||||||
|
pix41 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_2)[x+4]));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[0]), pix02));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[5]), pix12));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[10]), pix22));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[15]), pix32));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[20]), pix42));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[1]), pix03));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[6]), pix13));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[11]), pix23));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[16]), pix33));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[21]), pix43));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[2]), pix04));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[7]), pix14));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[12]), pix24));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[17]), pix34));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[22]), pix44));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[3]), pix00));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[8]), pix10));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[13]), pix20));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[18]), pix30));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[23]), pix40));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[4]), pix01));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[9]), pix11));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[14]), pix21));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[19]), pix31));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[24]), pix41));
|
||||||
|
ssi2 = _mm_cvtps_epi32(ss);
|
||||||
|
|
||||||
|
ss = _mm_set1_ps(offset);
|
||||||
|
pix02 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in2)[x+5]));
|
||||||
|
pix12 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in1)[x+5]));
|
||||||
|
pix22 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[x+5]));
|
||||||
|
pix32 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_1)[x+5]));
|
||||||
|
pix42 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_2)[x+5]));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[0]), pix03));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[5]), pix13));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[10]), pix23));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[15]), pix33));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[20]), pix43));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[1]), pix04));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[6]), pix14));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[11]), pix24));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[16]), pix34));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[21]), pix44));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[2]), pix00));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[7]), pix10));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[12]), pix20));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[17]), pix30));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[22]), pix40));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[3]), pix01));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[8]), pix11));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[13]), pix21));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[18]), pix31));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[23]), pix41));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[4]), pix02));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[9]), pix12));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[14]), pix22));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[19]), pix32));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[24]), pix42));
|
||||||
|
ssi3 = _mm_cvtps_epi32(ss);
|
||||||
|
|
||||||
|
ssi0 = _mm_packs_epi32(ssi0, ssi1);
|
||||||
|
ssi1 = _mm_packs_epi32(ssi2, ssi3);
|
||||||
|
ssi0 = _mm_packus_epi16(ssi0, ssi1);
|
||||||
|
_mm_storeu_si128((__m128i*) &out[x], ssi0);
|
||||||
|
|
||||||
|
ss = _mm_set1_ps(offset);
|
||||||
|
pix03 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in2)[x+6]));
|
||||||
|
pix13 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in1)[x+6]));
|
||||||
|
pix23 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in0)[x+6]));
|
||||||
|
pix33 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_1)[x+6]));
|
||||||
|
pix43 = _mm_cvtepi32_ps(_mm_cvtepu8_epi32(*(__m128i *) &(in_2)[x+6]));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[0]), pix04));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[5]), pix14));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[10]), pix24));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[15]), pix34));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[20]), pix44));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[1]), pix00));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[6]), pix10));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[11]), pix20));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[16]), pix30));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[21]), pix40));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[2]), pix01));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[7]), pix11));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[12]), pix21));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[17]), pix31));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[22]), pix41));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[3]), pix02));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[8]), pix12));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[13]), pix22));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[18]), pix32));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[23]), pix42));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[4]), pix03));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[9]), pix13));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[14]), pix23));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[19]), pix33));
|
||||||
|
ss = _mm_add_ps(ss, _mm_mul_ps(_mm_set1_ps(kernel[24]), pix43));
|
||||||
|
ssi0 = _mm_cvtps_epi32(ss);
|
||||||
|
ssi0 = _mm_packs_epi32(ssi0, ssi0);
|
||||||
|
ssi0 = _mm_packus_epi16(ssi0, ssi0);
|
||||||
|
out[x+4] = _mm_cvtsi128_si32(ssi0);
|
||||||
|
}
|
||||||
|
for (; x < im->xsize-2; x++) {
|
||||||
__m128 ss = _mm_set1_ps(offset);
|
__m128 ss = _mm_set1_ps(offset);
|
||||||
__m128i ssi;
|
__m128i ssi;
|
||||||
|
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue
Block a user