in c_reference/src/quantized_mbconv.c [264:520]
void q7xq15_q15_mbconv_block(const Q7_T* const input,
const Q15_T* const filter1, const Q15_T* const BN1W, const Q15_T* const BN1B,
const Q15_T* const filter2, const Q15_T* const BN2W, const Q15_T* const BN2B,
const Q15_T* const filter3, const Q15_T* const BN3W, const Q15_T* const BN3B,
Q15_T* const output, Q15_T* const convBuffer1, Q15_T* const convBuffer2,
ITER_T N, ITER_T H, ITER_T W, ITER_T CIn, ITER_T CTemp, ITER_T HF, ITER_T WF,
ITER_T COut, ITER_T HOut, ITER_T WOut, S_ITER_T HPadU, S_ITER_T HPadD,
S_ITER_T WPadL, S_ITER_T WPadR, ITER_T HStride, ITER_T WStride, Q31_T limit1,
Q31_T limit2, SCALE_T shrU1, SCALE_T shrX1, SCALE_T shrU2, SCALE_T shrX2,
SCALE_T shrU3, SCALE_T shrW3, SCALE_T shlU1, SCALE_T shlX1, SCALE_T shlU2,
SCALE_T shlX2, SCALE_T shlU3, SCALE_T shlW3) {
S_ITER_T HOffsetFL = (HF - 1) >> 1;
S_ITER_T WOffsetFL = (WF - 1) >> 1;
S_ITER_T HOffsetFR = HF >> 1;
S_ITER_T WOffsetFR = WF >> 1;
S_ITER_T HOffsetL = HOffsetFL - HPadU;
S_ITER_T WOffsetL = WOffsetFL - WPadL;
S_ITER_T HOffsetR = HOffsetFR - HPadD;
S_ITER_T WOffsetR = WOffsetFR - WPadR;
ITER_T HOffsetIn = W * CIn;
ITER_T NOffsetIn = H * HOffsetIn;
ITER_T HOffsetC1 = W * CTemp;
ITER_T GOffsetF = HF * WF;
ITER_T HOffsetOut = WOut * COut;
ITER_T NOffsetOut = HOut * HOffsetOut;
Q31_T sum;
for (ITER_T n = 0; n < N; n++) {
ITER_T NIndexIn = n * NOffsetIn;
ITER_T NIndexOut = n * NOffsetOut;
ITER_T margin = 0;
if ((S_ITER_T)HF - HPadU - (S_ITER_T)HStride > 0) {
margin = (ITER_T)((S_ITER_T)HF - HPadU - (S_ITER_T)HStride);
}
for (ITER_T i = 0; i < margin; i++) {
ITER_T HIndexIn = i * HOffsetIn + NIndexIn;
ITER_T HIndexC1 = i * HOffsetC1;
for (ITER_T j = 0; j < W; j++) {
ITER_T WIndexIn = j * CIn + HIndexIn;
Q15_T* convBuffer1_offset = ((Q15_T*)convBuffer1) + j * CTemp + HIndexC1;
for (ITER_T k = 0; k < CTemp; k++) {
sum = 0;
Q7_T* input_offset = (Q7_T*)input + WIndexIn;
Q15_T* filter1_offset = (Q15_T*)filter1 + k;
ITER_T in_channels = CIn;
#ifdef LOOP_UNROLL
ITER_T len_unroll = in_channels >> 2;
in_channels = in_channels % 4;
while (len_unroll--) {
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
}
#endif
while (in_channels--) {
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
}
#ifdef SHIFT
Q31_T x = (((Q31_T)(((sum << shlU1) >> shrU1) + BN1B[k])) *
((Q31_T)BN1W[k]));
#else
Q31_T x = (((Q31_T)((sum * shlU1) / shrU1 + BN1B[k])) *
((Q31_T)BN1W[k]));
#endif
x = q31_relu(x, limit1);
#ifdef SHIFT
*convBuffer1_offset++ = ((x << shlX1) >> shrX1);
#else
*convBuffer1_offset++ = (x * shlX1) / shrX1;
#endif
}
}
}
ITER_T hout = 0;
for (S_ITER_T h = HOffsetL; h < (S_ITER_T)H - HOffsetR; hout++, h += (S_ITER_T)HStride) {
ITER_T HIndexOut = hout * HOffsetOut + NIndexOut;
for (ITER_T i = 0; i < HStride; i++) {
ITER_T iRed = (i + margin + hout * HStride) % HF;
ITER_T iFull = i + margin + hout * HStride;
ITER_T HIndexC1 = iRed * HOffsetC1;
ITER_T HIndexIn = iFull * HOffsetIn + NIndexIn;
if (iFull < H){
for (ITER_T j = 0; j < W; j++) {
ITER_T WIndexIn = j * CIn + HIndexIn;
Q15_T* convBuffer1_offset = ((Q15_T*)convBuffer1) + j * CTemp + HIndexC1;
for (ITER_T k = 0; k < CTemp; k++) {
sum = 0;
Q7_T* input_offset = (Q7_T*)input + WIndexIn;
Q15_T* filter1_offset = (Q15_T*)filter1 + k;
ITER_T in_channels = CIn;
#ifdef LOOP_UNROLL
ITER_T len_unroll = in_channels >> 2;
in_channels = in_channels % 4;
while (len_unroll--) {
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
}
#endif
while (in_channels--) {
sum += ((Q31_T)*input_offset++) * ((Q31_T)*filter1_offset);
filter1_offset += CTemp;
}
#ifdef SHIFT
Q31_T x = (((Q31_T)(((sum << shlU1) >> shrU1) + BN1B[k])) *
((Q31_T)BN1W[k]));
#else
Q31_T x = (((Q31_T)((sum * shlU1) / shrU1 + BN1B[k])) *
((Q31_T)BN1W[k]));
#endif
x = q31_relu(x, limit1);
#ifdef SHIFT
*convBuffer1_offset++ = ((x << shlX1) >> shrX1);
#else
*convBuffer1_offset++ = (x * shlX1) / shrX1;
#endif
}
}
} else {
Q15_T* convBuffer1_offset = (Q15_T*)convBuffer1 + HIndexC1;
for (ITER_T j = 0; j < W; j++) {
Q15_T* BN1B_offset = (Q15_T*)BN1B;
Q15_T* BN1W_offset = (Q15_T*)BN1W;
ITER_T temp_channels = CTemp;
#ifdef LOOP_UNROLL
ITER_T len_unroll = temp_channels >> 2;
temp_channels = temp_channels % 4;
while (len_unroll--) {
Q31_T w = q31_relu(((Q31_T)*BN1B_offset++) * ((Q31_T)*BN1W_offset++), limit1);
Q31_T x = q31_relu(((Q31_T)*BN1B_offset++) * ((Q31_T)*BN1W_offset++), limit1);
Q31_T y = q31_relu(((Q31_T)*BN1B_offset++) * ((Q31_T)*BN1W_offset++), limit1);
Q31_T z = q31_relu(((Q31_T)*BN1B_offset++) * ((Q31_T)*BN1W_offset++), limit1);
#ifdef SHIFT
*convBuffer1_offset++ = ((w << shlX1) >> shrX1);
*convBuffer1_offset++ = ((x << shlX1) >> shrX1);
*convBuffer1_offset++ = ((y << shlX1) >> shrX1);
*convBuffer1_offset++ = ((z << shlX1) >> shrX1);
#else
*convBuffer1_offset++ = ((w * shlX1) / shrX1);
*convBuffer1_offset++ = ((x * shlX1) / shrX1);
*convBuffer1_offset++ = ((y * shlX1) / shrX1);
*convBuffer1_offset++ = ((z * shlX1) / shrX1);
#endif
}
#endif
while (temp_channels--) {
Q15_T w = q15_relu(((Q15_T)*BN1B_offset++) * ((Q15_T)*BN1W_offset++), limit1);
#ifdef SHIFT
*convBuffer1_offset++ = ((w << shlX1) >> shrX1);
#else
*convBuffer1_offset++ = ((w * shlX1) / shrX1);
#endif
}
}
}
}
ITER_T wout = 0;
for (S_ITER_T w = WOffsetL; w < ((S_ITER_T)W) - WOffsetR; wout++, w += ((S_ITER_T)WStride)) {
Q15_T* output_offset = ((Q15_T*)output) + wout * COut + HIndexOut;
for (ITER_T g = 0; g < CTemp; g++) {
sum = 0;
ITER_T GIndexF = g * GOffsetF;
for (S_ITER_T hf = -HOffsetFL; hf <= HOffsetFR; hf++) {
S_ITER_T hindex = h + hf;
if ((hindex < 0) || (hindex >= (S_ITER_T)H)){
continue;
}
ITER_T HIndexC1 = (((ITER_T)hindex) % HF) * HOffsetC1 + g;
ITER_T HIndexF = ((ITER_T)(hf + HOffsetFL)) * WF + GIndexF;
for (S_ITER_T wf = -WOffsetFL; wf <= WOffsetFR; wf++) {
S_ITER_T windex = w + wf;
if ((windex < 0) || (windex >= (S_ITER_T)W)) {
continue;
} else {
sum += ((Q31_T)convBuffer1[HIndexC1 + ((ITER_T)windex) * CTemp]) *
((Q31_T)filter2[HIndexF + ((ITER_T)(wf + WOffsetFL))]);
}
}
}
#ifdef SHIFT
Q31_T x = (((Q31_T)(((sum << shlU2) >> shrU2) + BN2B[g])) *
((Q31_T)BN2W[g]));
#else
Q31_T x = (((Q31_T)((sum * shlU2) / shrU2 + BN2B[g])) *
((Q31_T)BN2W[g]));
#endif
x = q31_relu(x, limit2);
#ifdef SHIFT
convBuffer2[g] = ((x << shlX2) >> shrX2);
#else
convBuffer2[g] = (x * shlX2) / shrX2;
#endif
}
for (ITER_T i = 0; i < COut; i++) {
sum = 0;
Q15_T* convBuffer2_offset = (Q15_T*)convBuffer2;
Q15_T* filter3_offset = (Q15_T*)filter3 + i;
ITER_T temp_channels = CTemp;
#ifdef LOOP_UNROLL
ITER_T len_unroll = temp_channels >> 2;
temp_channels = temp_channels % 4;
while (len_unroll--) {
sum += ((Q31_T)*convBuffer2_offset++) * ((Q31_T)*filter3_offset);
filter3_offset += COut;
sum += ((Q31_T)*convBuffer2_offset++) * ((Q31_T)*filter3_offset);
filter3_offset += COut;
sum += ((Q31_T)*convBuffer2_offset++) * ((Q31_T)*filter3_offset);
filter3_offset += COut;
sum += ((Q31_T)*convBuffer2_offset++) * ((Q31_T)*filter3_offset);
filter3_offset += COut;
}
#endif
while (temp_channels--) {
sum += ((Q31_T)*convBuffer2_offset++) * ((Q31_T)*filter3_offset);
filter3_offset += COut;
}
#ifdef SHIFT
*output_offset++ = (((((Q31_T)(((sum << shlU3) >> shrU3) + BN3B[i])) *
((Q31_T) BN3W[i])) << shlW3) >> shrW3);
#else
*output_offset++ = ((((Q31_T)((sum * shlU3) / shrU3 + BN3B[i])) *
((Q31_T) BN3W[i])) * shlW3) / shrW3;
#endif
}
}
}
}
}