89     uint32_t x2 = x;  in OneU4()  local
95                 convert_float4(py0[x2]) * coeff[2] +  in OneU4()
101                 convert_float4(py1[x2]) * coeff[7] +  in OneU4()
107                 convert_float4(py2[x2]) * coeff[12] +  in OneU4()
113                 convert_float4(py3[x2]) * coeff[17] +  in OneU4()
119                 convert_float4(py4[x2]) * coeff[22] +  in OneU4()
132     uint32_t x2 = x;  in OneU2()  local
138                 convert_float2(py0[x2]) * coeff[2] +  in OneU2()
144                 convert_float2(py1[x2]) * coeff[7] +  in OneU2()
150                 convert_float2(py2[x2]) * coeff[12] +  in OneU2()
156                 convert_float2(py3[x2]) * coeff[17] +  in OneU2()
162                 convert_float2(py4[x2]) * coeff[22] +  in OneU2()
175     uint32_t x2 = x;  in OneU1()  local
181                (float)(py0[x2]) * coeff[2] +  in OneU1()
187                (float)(py1[x2]) * coeff[7] +  in OneU1()
193                (float)(py2[x2]) * coeff[12] +  in OneU1()
199                (float)(py3[x2]) * coeff[17] +  in OneU1()
205                (float)(py4[x2]) * coeff[22] +  in OneU1()
218     uint32_t x2 = x;  in OneF4()  local
224                 py0[x2] * coeff[2] +  in OneF4()
230                 py1[x2] * coeff[7] +  in OneF4()
236                 py2[x2] * coeff[12] +  in OneF4()
242                 py3[x2] * coeff[17] +  in OneF4()
248                 py4[x2] * coeff[22] +  in OneF4()
260     uint32_t x2 = x;  in OneF2()  local
266                 py0[x2] * coeff[2] +  in OneF2()
272                 py1[x2] * coeff[7] +  in OneF2()
278                 py2[x2] * coeff[12] +  in OneF2()
284                 py3[x2] * coeff[17] +  in OneF2()
290                 py4[x2] * coeff[22] +  in OneF2()
302     uint32_t x2 = x;  in OneF1()  local
308                py0[x2] * coeff[2] +  in OneF1()
314                py1[x2] * coeff[7] +  in OneF1()
320                py2[x2] * coeff[12] +  in OneF1()
326                py3[x2] * coeff[17] +  in OneF1()
332                py4[x2] * coeff[22] +  in OneF1()
368     uint32_t x2 = xend;  in kernelU4()  local
370     while((x1 < x2) && (x1 < 2)) {  in kernelU4()
378     if (gArchUseSIMD &&((x1 + 6) < x2)) {  in kernelU4()
380         uint32_t len = (x2 - x1 - 3) >> 2;  in kernelU4()
388     if(gArchUseSIMD && ((x1 + 3) < x2)) {  in kernelU4()
389         uint32_t len = (x2 - x1 - 3) >> 1;  in kernelU4()
396     while(x1 < x2) {  in kernelU4()
428     uint32_t x2 = xend;  in kernelU2()  local
430     while((x1 < x2) && (x1 < 2)) {  in kernelU2()
437     if((x1 + 3) < x2) {  in kernelU2()
438         uint32_t len = (x2 - x1 - 3) >> 1;  in kernelU2()
445     while(x1 < x2) {  in kernelU2()
477     uint32_t x2 = xend;  in kernelU1()  local
479     while((x1 < x2) && (x1 < 2)) {  in kernelU1()
486     if((x1 + 3) < x2) {  in kernelU1()
487         uint32_t len = (x2 - x1 - 3) >> 1;  in kernelU1()
494     while(x1 < x2) {  in kernelU1()
526     uint32_t x2 = xend;  in kernelF4()  local
528     while((x1 < x2) && (x1 < 2)) {  in kernelF4()
535     if((x1 + 3) < x2) {  in kernelF4()
536         uint32_t len = (x2 - x1 - 3) >> 1;  in kernelF4()
543     while(x1 < x2) {  in kernelF4()
575     uint32_t x2 = xend;  in kernelF2()  local
577     while((x1 < x2) && (x1 < 2)) {  in kernelF2()
584     if((x1 + 3) < x2) {  in kernelF2()
585         uint32_t len = (x2 - x1 - 3) >> 1;  in kernelF2()
592     while(x1 < x2) {  in kernelF2()
624     uint32_t x2 = xend;  in kernelF1()  local
626     while((x1 < x2) && (x1 < 2)) {  in kernelF1()
633     if((x1 + 3) < x2) {  in kernelF1()
634         uint32_t len = (x2 - x1 - 3) >> 1;  in kernelF1()
641     while(x1 < x2) {  in kernelF1()