Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
提交 5ba523e7 编写于 作者: S smilejames 提交者： GitHub
浏览文件
操作

Merge branch 'develop' into develop

上级 63250a65 2808898a
显示空白变更内容
内联 并排
Showing with 662 addition and 481 deletion
README.md
浏览文件 @ 5ba523e7
......@@ -28,13 +28,13 @@ Paddle-Moible是PaddlePaddle组织下的项目，是一个致力于嵌入式平
|mobilenet arm v7|1线程|2线程|4线程|
|------------|----|-----|-----|
|麒麟960(ms)|110.586|72.474|49.833|
|麒麟960(ms)|110.586|70.897|47.474|
|||||
|mobilenetssd arm v7|1线程|2线程|4线程|
|麒麟960(ms)|224.464|142.544|96.068|
|麒麟960(ms)|222.124|138.952|90.856|
|||||
|googlenet(v1) arm v7|1线程|2线程|4线程|
|麒麟960(ms)|348.018|242.689|169.998|
|麒麟960(ms)|348.018|240.304|169.998|
|||||
|squeezenet arm v7|1线程|2线程|4线程|
|麒麟960(ms)|84.685|56.544|38.833|
......
src/operators/math/depthwise_conv_3x3.cpp
浏览文件 @ 5ba523e7
......@@ -613,7 +613,6 @@ void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, const Tensor *filter,
}
int m;
for (m = 1; m < output_width - 4; m += 4) {
float *output_ptr = output_data + m;
float32x4_t in0, in1, in2, in3, tmp0, tmp1, tmp2, tmp3, out0;
......@@ -637,7 +636,8 @@ void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, const Tensor *filter,
}
vst1q_f32(output_ptr, out0);
}
for (m = 1; (m + 3) < output_width - 1; m = m + 4) {
for (m = 1; (m + 3) < output_width - 1; m += 4) {
}
for (int j = m; j < output_width - 1; j++) {
output_data[j] = input_data[j - 1] * w10 + input_data[j] * w11 +
......@@ -652,7 +652,7 @@ void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, const Tensor *filter,
}
}
for (m = 1; (m + 3) < output_width - 1; m = m + 4) {
for (m = 1; m < output_width - 4; m += 4) {
float *output_ptr =
output_data + (output_height - 1) * output_width + m;
......@@ -807,71 +807,60 @@ void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, const Tensor *filter,
output_data[0] = w11 * input_data[0] + w12 * input_data[1] +
w21 * input_data[l] + w22 * input_data[l + 1];
output_data[l - 1] = w10 * input_data[l - 2] + w11 * input_data[l - 1] +
w20 * input_data[2 * l - 2] +
w21 * input_data[2 * l - 1];
output_data[l - 1] = w10 * input_data[l - 2] + w11 * input_data[l -
1] + w20 * input_data[2 * l - 2] + w21 * input_data[2 * l - 1];
output_data[(l - 1) * l] =
w01 * input_data[(l - 2) * l] + w02 * input_data[(l - 2) * l + 1] +
w11 * input_data[(l - 1) * l] + w12 * input_data[(l - 1) * l + 1];
w01 * input_data[(l - 2) * l] + w02 * input_data[(l - 2) * l +
1] + w11 * input_data[(l - 1) * l] + w12 * input_data[(l - 1) * l + 1];
output_data[l * l - 1] = w00 * input_data[(l - 2) * (l + 1)] +
w01 * input_data[(l - 2) * (l + 1) + 1] +
w10 * input_data[l * l - 2] +
w11 * input_data[l * l - 1];
output_data[0] = output_data[0] * newscale_data[j] + newbias_data[j];
output_data[l - 1] =
output_data[l - 1] * newscale_data[j] + newbias_data[j];
output_data[(l - 1) * l] =
output_data[0] = output_data[0] * newscale_data[j] +
newbias_data[j]; output_data[l - 1] = output_data[l - 1] *
newscale_data[j] + newbias_data[j]; output_data[(l - 1) * l] =
output_data[(l - 1) * l] * newscale_data[j] + newbias_data[j];
output_data[l * l - 1] =
output_data[l * l - 1] * newscale_data[j] + newbias_data[j];
if (if_relu) {
output_data[0] = output_data[0] < 0 ? 0 : output_data[0];
output_data[l - 1] = output_data[l - 1] < 0 ? 0 : output_data[l - 1];
output_data[(l - 1) * l] =
output_data[(l - 1) * l] < 0 ? 0 : output_data[(l - 1) * l];
output_data[l * l - 1] =
output_data[l * l - 1] < 0 ? 0 : output_data[l * l - 1];
output_data[l - 1] = output_data[l - 1] < 0 ? 0 : output_data[l -
1]; output_data[(l - 1) * l] = output_data[(l - 1) * l] < 0 ? 0 :
output_data[(l - 1) * l]; output_data[l * l - 1] = output_data[l * l - 1]
< 0 ? 0 : output_data[l * l - 1];
}
for (int i = 1; i < l - 1; ++i) {
output_data[i * l] =
w01 * input_data[i * l - l] + w02 * input_data[i * l - l + 1] +
w11 * input_data[i * l] + w12 * input_data[i * l + 1] +
w21 * input_data[i * l + l] + w22 * input_data[i * l + l + 1];
output_data[i * l + l - 1] = w00 * input_data[i * l + l - 1 - l - 1] +
w01 * input_data[i * l + l - 1 - l] +
w10 * input_data[i * l + l - 1 - 1] +
w11 * input_data[i * l + l - 1] +
w20 * input_data[i * l + l - 1 + l - 1] +
w21 * input_data[i * l + l - 1 + l];
output_data[i * l] =
output_data[i * l] * newscale_data[j] + newbias_data[j];
output_data[i * l + l - 1] =
output_data[i * l + l - 1] * newscale_data[j] + newbias_data[j];
w01 * input_data[i * l - l] + w02 * input_data[i * l - l + 1]
+ w11 * input_data[i * l] + w12 * input_data[i * l + 1] + w21 *
input_data[i * l + l] + w22 * input_data[i * l + l + 1]; output_data[i *
l + l - 1] = w00 * input_data[i * l + l - 1 - l - 1] + w01 * input_data[i
* l + l - 1 - l] + w10 * input_data[i * l + l - 1 - 1] + w11 *
input_data[i * l + l - 1] + w20 * input_data[i * l + l - 1 + l - 1] + w21
* input_data[i * l + l - 1 + l]; output_data[i * l] = output_data[i * l]
* newscale_data[j] + newbias_data[j]; output_data[i * l + l - 1] =
output_data[i * l + l - 1] * newscale_data[j] +
newbias_data[j];
if (if_relu) {
output_data[i * l] = output_data[i * l] < 0 ? 0 : output_data[i *
l]; output_data[i * l + l - 1] =
output_data[i * l + l - 1] < 0 ? 0 : output_data[i * l + l - 1];
output_data[i * l] = output_data[i * l] < 0 ? 0 : output_data[i
* l]; output_data[i * l + l - 1] = output_data[i * l + l - 1] < 0 ? 0 :
output_data[i * l + l - 1];
}
}
// top 1 row and bottom 1 row
const float *input_tmp = input_data;
float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1, tmp2,
tmp3, tmp4, tmp5, out0;
in0 = vld1q_f32(input_tmp);
in2 = vld1q_f32(input_tmp + l);
const float *input_tmp_end = input_tmp + (l - 2) * l;
in4 = vld1q_f32(input_tmp_end);
in6 = vld1q_f32(input_tmp_end + l);
int c_mid = l_mid;
auto output_ptr = output_data + 1;
for (; c_mid > 3; c_mid -= 4) {
in1 = vld1q_f32(input_tmp + 4);
in3 = vld1q_f32(input_tmp + l + 4);
float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1,
tmp2, tmp3, tmp4, tmp5, out0; in0 = vld1q_f32(input_tmp); in2 =
vld1q_f32(input_tmp + l); const float *input_tmp_end = input_tmp + (l -
2) * l; in4 = vld1q_f32(input_tmp_end); in6 = vld1q_f32(input_tmp_end +
l); int c_mid = l_mid; auto output_ptr = output_data + 1; for (; c_mid >
3; c_mid -= 4) { in1 = vld1q_f32(input_tmp + 4); in3 =
vld1q_f32(input_tmp + l + 4);
tmp0 = vextq_f32(in0, in1, 1);
tmp1 = vextq_f32(in0, in1, 2);
......@@ -1068,6 +1057,7 @@ void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, const Tensor *filter,
}
}
*/
#endif
}
......@@ -1482,230 +1472,421 @@ void DepthwiseConvAddBNRelu3x3s2p1v2(const Tensor *input, const Tensor *filter,
const float *newscale_data = new_scale->data<float>();
const float *newbias_data = new_bias->data<float>();
float32x4_t vnewbias = vdupq_n_f32(0.0);
float32x4_t vnewscale = vdupq_n_f32(1.0);
const int in_h = static_cast<int>(input->dims()[2]);
const int in_w = static_cast<int>(input->dims()[3]);
const int out_h = static_cast<int>(output->dims()[2]);
const int out_w = static_cast<int>(output->dims()[3]);
const int out_l = out_h;
const int in_l = in_h;
const int inhxw = in_h * in_w;
const int outhxw = out_h * out_w;
const int if_pad = in_l - 1 == (out_l - 1) * 2 ? 1 : 0;
const int batch_size = static_cast<int>(input->dims()[0]);
const int c = static_cast<int>(input->dims()[1]);
const float *input_row_ptr;
float *output_row_ptr;
const int input_channel = static_cast<int>(input->dims()[1]);
const int w_times = (out_w - 2) / 3;
const int input_height = static_cast<int>(input->dims()[2]);
const int input_width = static_cast<int>(input->dims()[3]);
const int output_height = static_cast<int>(output->dims()[2]);
const int output_width = static_cast<int>(output->dims()[3]);
const int inhxw = input_height * input_width;
const int outhxw = output_height * output_width;
float32x4x2_t input_buff_mid{}, input_buff_bottom[w_times + 1];
float32x4_t elewise_res0, elewise_res1, elewise_res2, res3;
int out2in_mid;
float32x4_t vnewbias = vdupq_n_f32(0.0);
float32x4_t vnewscale = vdupq_n_f32(1.0);
float32x4_t zero = vdupq_n_f32(0.0);
for (int b = batch_size; b > 0; --b) {
const float *filter_data_tmp = filter_data;
for (int j = 0; j < c; ++j) {
auto output_data_tmp = output_data + j * out_h * out_w;
auto input_data_tmp = input_data + j * in_h * in_w;
auto input_const = input_data_tmp;
vnewbias = vdupq_n_f32(newbias_data[j]);
vnewscale = vdupq_n_f32(newscale_data[j]);
float w00 = filter_data_tmp[0];
float w01 = filter_data_tmp[1];
float w02 = filter_data_tmp[2];
float w10 = filter_data_tmp[3];
float w11 = filter_data_tmp[4];
float w12 = filter_data_tmp[5];
float w20 = filter_data_tmp[6];
float w21 = filter_data_tmp[7];
float w22 = filter_data_tmp[8];
int h_mid = 0;
for (; h_mid < out_h - 1; h_mid++) {
input_row_ptr = input_data_tmp + 1 + h_mid * 2 * in_w;
output_row_ptr = output_data_tmp + 1 + h_mid * out_w;
for (int b = 0; b < batch_size; b++) {
filter_data = filter->data<float>();
for (int c = 0; c < input_channel; c++) {
vnewbias = vdupq_n_f32(newbias_data[c]);
vnewscale = vdupq_n_f32(newscale_data[c]);
for (int w4 = 0; w4 < w_times + 1; w4++) {
if (h_mid == 0) {
elewise_res1 = zero;
elewise_res0 = zero;
elewise_res2 = zero;
} else {
elewise_res1 = vmulq_n_f32(input_buff_bottom[w4].val[1], w01);
elewise_res0 = vmulq_n_f32(input_buff_bottom[w4].val[0], w00);
elewise_res2 = vmulq_n_f32(input_buff_bottom[w4].val[0], w02);
}
input_buff_mid = vld2q_f32(input_row_ptr);
input_buff_bottom[w4] = vld2q_f32(input_row_ptr + in_w);
float w00 = filter_data[0];
float w01 = filter_data[1];
float w02 = filter_data[2];
float w10 = filter_data[3];
float w11 = filter_data[4];
float w12 = filter_data[5];
float w20 = filter_data[6];
float w21 = filter_data[7];
float w22 = filter_data[8];
elewise_res1 = vmlaq_n_f32(elewise_res1, input_buff_mid.val[1], w11);
elewise_res0 = vmlaq_n_f32(elewise_res0, input_buff_mid.val[0], w10);
elewise_res2 = vmlaq_n_f32(elewise_res2, input_buff_mid.val[0], w12);
int m;
for (m = 1; m < output_width - 2; m = m + 3) {
float *output_ptr = output_data + m;
float32x4x2_t input_buff_mid{}, input_buff_bottom{};
float32x4_t in0, in1, in2, in3, tmp0, tmp1, tmp2, tmp3, out0;
input_buff_mid = vld2q_f32(input_data + (2 * m - 1));
input_buff_bottom = vld2q_f32(input_data + input_width + (2 * m - 1));
elewise_res1 =
vmlaq_n_f32(elewise_res1, input_buff_bottom[w4].val[1], w21);
elewise_res0 =
vmlaq_n_f32(elewise_res0, input_buff_bottom[w4].val[0], w20);
elewise_res2 =
vmlaq_n_f32(elewise_res2, input_buff_bottom[w4].val[0], w22);
in0 = input_buff_mid.val[0];
tmp0 = input_buff_mid.val[1];
tmp1 = vextq_f32(in0, zero, 1);
res3 = vaddq_f32(vextq_f32(elewise_res2, zero, 1),
vaddq_f32(elewise_res0, elewise_res1));
res3 = vmlaq_f32(vnewbias, vnewscale, res3);
in2 = input_buff_bottom.val[0];
tmp2 = input_buff_bottom.val[1];
tmp3 = vextq_f32(in2, zero, 1);
out0 = vmulq_n_f32(in0, w10);
out0 = vmlaq_n_f32(out0, tmp0, w11);
out0 = vmlaq_n_f32(out0, tmp1, w12);
out0 = vmlaq_n_f32(out0, in2, w20);
out0 = vmlaq_n_f32(out0, tmp2, w21);
out0 = vmlaq_n_f32(out0, tmp3, w22);
out0 = vmlaq_f32(vnewbias, vnewscale, out0);
if (if_relu) {
res3 = vmaxq_f32(res3, zero);
out0 = vmaxq_f32(out0, zero);
}
vst1q_f32(output_row_ptr, res3);
input_row_ptr += 6;
output_row_ptr += 3;
vst1q_f32(output_ptr, out0);
}
for (m = 1; m < output_width - 2; m += 3) {
}
for (int j = m; j < output_width; j++) {
output_data[j] = input_data[2 * j - 1] * w10 + input_data[2 * j] * w11 +
input_data[2 * j + 1] * w12 +
input_data[2 * j - 1 + input_width] * w20 +
input_data[2 * j + input_width] * w21 +
input_data[2 * j + 1 + input_width] * w22;
output_data[j] = newscale_data[c] * output_data[j] + newbias_data[c];
if (if_relu) {
output_data[j] = output_data[j] < 0 ? 0 : output_data[j];
}
}
clock();
input_row_ptr = input_data_tmp + 1 + h_mid * 2 * in_w;
output_row_ptr = output_data_tmp + 1 + h_mid * out_w;
#pragma omp parallel for
for (int w4 = 0; w4 < w_times + 1; w4++) {
elewise_res1 = vmulq_n_f32(input_buff_bottom[w4].val[1], w01);
elewise_res0 = vmulq_n_f32(input_buff_bottom[w4].val[0], w00);
elewise_res2 = vmulq_n_f32(input_buff_bottom[w4].val[0], w02);
for (int i = 1; i < output_height; i += 1) {
for (int m = 1; m < output_width - 2; m += 3) {
float *output_ptr = output_data + i * output_width + m;
float32x4x2_t input_buff_top{}, input_buff_mid{}, input_buff_bottom{};
float32x4_t in0, in1, in2, in3, in4, in5, tmp0, tmp1, tmp2, tmp3,
tmp4, tmp5, out0;
input_buff_top =
vld2q_f32(input_data + (2 * i - 1) * input_width + (2 * m - 1));
input_buff_mid =
vld2q_f32(input_data + (2 * i) * input_width + (2 * m - 1));
input_buff_bottom =
vld2q_f32(input_data + (2 * i + 1) * input_width + (2 * m - 1));
input_buff_mid = vld2q_f32(input_row_ptr);
input_buff_bottom[w4] = vld2q_f32(input_row_ptr + in_w);
in0 = input_buff_top.val[0];
tmp0 = input_buff_top.val[1];
tmp1 = vextq_f32(in0, zero, 1);
elewise_res1 = vmlaq_n_f32(elewise_res1, input_buff_mid.val[1], w11);
elewise_res0 = vmlaq_n_f32(elewise_res0, input_buff_mid.val[0], w10);
elewise_res2 = vmlaq_n_f32(elewise_res2, input_buff_mid.val[0], w12);
in2 = input_buff_mid.val[0];
tmp2 = input_buff_mid.val[1];
tmp3 = vextq_f32(in2, zero, 1);
if (!if_pad) {
elewise_res1 =
vmlaq_n_f32(elewise_res1, input_buff_bottom[w4].val[1], w21);
elewise_res0 =
vmlaq_n_f32(elewise_res0, input_buff_bottom[w4].val[0], w20);
elewise_res2 =
vmlaq_n_f32(elewise_res2, input_buff_bottom[w4].val[0], w22);
}
res3 = vaddq_f32(vextq_f32(elewise_res2, zero, 1),
vaddq_f32(elewise_res0, elewise_res1));
res3 = vmlaq_f32(vnewbias, vnewscale, res3);
in4 = input_buff_bottom.val[0];
tmp4 = input_buff_bottom.val[1];
tmp5 = vextq_f32(in4, zero, 1);
out0 = vmulq_n_f32(in0, w00);
out0 = vmlaq_n_f32(out0, tmp0, w01);
out0 = vmlaq_n_f32(out0, tmp1, w02);
out0 = vmlaq_n_f32(out0, in2, w10);
out0 = vmlaq_n_f32(out0, tmp2, w11);
out0 = vmlaq_n_f32(out0, tmp3, w12);
out0 = vmlaq_n_f32(out0, in4, w20);
out0 = vmlaq_n_f32(out0, tmp4, w21);
out0 = vmlaq_n_f32(out0, tmp5, w22);
out0 = vmlaq_f32(vnewbias, vnewscale, out0);
if (if_relu) {
res3 = vmaxq_f32(res3, zero);
}
if ((w4 != w_times)) {
vst1q_f32(output_row_ptr, res3);
} else {
if (out_l - 2 - w_times * 3 == 1) {
vst1q_lane_f32(output_row_ptr, res3, 0);
} else if (out_l - 2 - w_times * 3 == 2) {
vst1q_lane_f32(output_row_ptr, res3, 0);
vst1q_lane_f32(output_row_ptr + 1, res3, 1);
out0 = vmaxq_f32(out0, zero);
}
vst1q_f32(output_ptr, out0);
}
input_row_ptr += 6;
output_row_ptr += 3;
int m;
for (m = 1; m < output_width - 2; m += 3) {
}
output_data_tmp[0] = input_const[0] * w11 + input_const[1] * w12 +
input_const[in_l] * w21 +
input_const[in_l + 1] * w22;
out2in_mid = (out_l - 1) * 2;
output_data_tmp[out_l - 1] =
w10 * input_const[out2in_mid - 1] + w11 * input_const[out2in_mid] +
w20 * input_const[out2in_mid + in_w - 1] +
w21 * input_const[out2in_mid + in_w] +
(1 - if_pad) * (w12 * input_const[out2in_mid + 1] +
w22 * input_const[out2in_mid + in_w + 1]);
out2in_mid = (out_l - 1) * 2 * in_w;
output_data_tmp[out_l * (out_l - 1)] =
w01 * input_const[out2in_mid - in_w] +
w02 * input_const[out2in_mid - in_w + 1] +
w11 * input_const[out2in_mid] + w12 * input_const[out2in_mid + 1] +
(1 - if_pad) * (w21 * input_const[out2in_mid + in_w] +
w22 * input_const[out2in_mid + in_w + 1]);
out2in_mid = (out_l - 1) * 2 * in_w + (out_l - 1) * 2;
output_data_tmp[out_l * out_l - 1] =
w00 * input_const[out2in_mid - in_w - 1] +
w01 * input_const[out2in_mid - in_w] +
w10 * input_const[out2in_mid - 1] + w11 * input_const[out2in_mid] +
(1 - if_pad) * (w20 * input_const[out2in_mid + in_w - 1] +
w21 * input_const[out2in_mid + in_w] +
w02 * input_const[out2in_mid - in_w + 1] +
w12 * input_const[out2in_mid + 1] +
w22 * input_const[out2in_mid + in_w + 1]);
output_data_tmp[0] =
output_data_tmp[0] * newscale_data[j] + newbias_data[j];
output_data_tmp[out_l - 1] =
output_data_tmp[out_l - 1] * newscale_data[j] + newbias_data[j];
output_data_tmp[out_l * (out_l - 1)] =
output_data_tmp[out_l * (out_l - 1)] * newscale_data[j] +
newbias_data[j];
output_data_tmp[out_l * out_l - 1] =
output_data_tmp[out_l * out_l - 1] * newscale_data[j] +
newbias_data[j];
for (int j = m; j < output_width; j++) {
output_data[i * output_width + j] =
input_data[(2 * i - 1) * input_width + 2 * j - 1] * w00 +
input_data[(2 * i - 1) * input_width + 2 * j] * w01 +
input_data[(2 * i - 1) * input_width + 2 * j + 1] * w02 +
input_data[(2 * i) * input_width + 2 * j - 1] * w10 +
input_data[(2 * i) * input_width + 2 * j] * w11 +
input_data[(2 * i) * input_width + 2 * j + 1] * w12 +
input_data[(2 * i + 1) * input_width + 2 * j - 1] * w20 +
input_data[(2 * i + 1) * input_width + 2 * j] * w21 +
input_data[(2 * i + 1) * input_width + 2 * j + 1] * w22;
output_data[i * output_width + j] =
newscale_data[c] * output_data[i * output_width + j] +
newbias_data[c];
if (if_relu) {
output_data_tmp[0] = output_data_tmp[0] < 0 ? 0 : output_data_tmp[0];
output_data_tmp[out_l - 1] =
output_data_tmp[out_l - 1] < 0 ? 0 : output_data_tmp[out_l - 1];
output_data_tmp[out_l * (out_l - 1)] =
output_data_tmp[out_l * (out_l - 1)] < 0
? 0
: output_data_tmp[out_l * (out_l - 1)];
output_data_tmp[out_l * out_l - 1] =
output_data_tmp[out_l * out_l - 1] < 0
output_data[i * output_width + j] =
output_data[i * output_width + j] < 0
? 0
: output_data_tmp[out_l * out_l - 1];
: output_data[i * output_width + j];
}
for (int i = 1; i < out_h - 1; i++) {
out2in_mid = i * 2 * in_w;
output_data_tmp[i * out_l] = w01 * input_const[out2in_mid - in_w] +
w02 * input_const[out2in_mid - in_w + 1] +
w11 * input_const[out2in_mid] +
w12 * input_const[out2in_mid + 1] +
w21 * input_const[out2in_mid + in_w] +
w22 * input_const[out2in_mid + in_w + 1];
}
}
output_data[0] = input_data[0] * w11 + input_data[1] * w12 +
input_data[input_height] * w21 +
input_data[input_height + 1] * w22;
out2in_mid = i * 2 * in_w + (out_l - 1) * 2;
output_data_tmp[i * out_l + out_l - 1] =
w00 * input_const[out2in_mid - in_w - 1] +
w01 * input_const[out2in_mid - in_w] +
w10 * input_const[out2in_mid - 1] + w11 * input_const[out2in_mid] +
w20 * input_const[out2in_mid + in_w - 1] +
w21 * input_const[out2in_mid + in_w] +
(1 - if_pad) * (w02 * input_const[out2in_mid - in_w + 1] +
w12 * input_const[out2in_mid + 1] +
w22 * input_const[out2in_mid + in_w + 1]);
output_data_tmp[i * out_l] =
output_data_tmp[i * out_l] * newscale_data[j] + newbias_data[j];
output_data_tmp[i * out_l + out_l - 1] =
output_data_tmp[i * out_l + out_l - 1] * newscale_data[j] +
newbias_data[j];
output_data[0] = newscale_data[c] * output_data[0] + newbias_data[c];
if (if_relu) {
output_data_tmp[i * out_l] =
output_data_tmp[i * out_l] < 0 ? 0 : output_data_tmp[i * out_l];
output_data_tmp[i * out_l + out_l - 1] =
output_data_tmp[i * out_l + out_l - 1] < 0
output_data[0] = output_data[0] < 0 ? 0 : output_data[0];
}
for (int i = 1; i < output_height; i++) {
output_data[i * output_width] =
input_data[(2 * i - 1) * input_width] * w01 +
input_data[(2 * i - 1) * input_width + 1] * w02 +
input_data[(2 * i) * input_width] * w11 +
input_data[(2 * i) * input_width + 1] * w12 +
input_data[(2 * i + 1) * input_width] * w21 +
input_data[(2 * i + 1) * input_width + 1] * w22;
output_data[i * output_width] =
newscale_data[c] * output_data[i * output_width] + newbias_data[c];
if (if_relu) {
output_data[i * output_width] = output_data[i * output_width] < 0
? 0
: output_data_tmp[i * out_l + out_l - 1];
: output_data[i * output_width];
}
}
filter_data_tmp += 9;
input_data = input_data + inhxw;
output_data = output_data + outhxw;
filter_data = filter_data + 9;
}
input_data += inhxw * c;
output_data += outhxw * c;
}
// const float *input_data = input->data<float>();
// const float *filter_data = filter->data<float>();
// float *output_data = output->data<float>();
// const float *newscale_data = new_scale->data<float>();
// const float *newbias_data = new_bias->data<float>();
//
// float32x4_t vnewbias = vdupq_n_f32(0.0);
// float32x4_t vnewscale = vdupq_n_f32(1.0);
//
// const int in_h = static_cast<int>(input->dims()[2]);
// const int in_w = static_cast<int>(input->dims()[3]);
// const int out_h = static_cast<int>(output->dims()[2]);
// const int out_w = static_cast<int>(output->dims()[3]);
// const int out_l = out_h;
// const int in_l = in_h;
// const int inhxw = in_h * in_w;
// const int outhxw = out_h * out_w;
// const int if_pad = in_l - 1 == (out_l - 1) * 2 ? 1 : 0;
// const int batch_size = static_cast<int>(input->dims()[0]);
// const int c = static_cast<int>(input->dims()[1]);
// const float *input_row_ptr;
// float *output_row_ptr;
//
// const int w_times = (out_w - 2) / 3;
//
// float32x4x2_t input_buff_mid{}, input_buff_bottom[w_times + 1];
// float32x4_t elewise_res0, elewise_res1, elewise_res2, res3;
// int out2in_mid;
// float32x4_t zero = vdupq_n_f32(0.0);
// for (int b = batch_size; b > 0; --b) {
// const float *filter_data_tmp = filter_data;
// for (int j = 0; j < c; ++j) {
// auto output_data_tmp = output_data + j * out_h * out_w;
// auto input_data_tmp = input_data + j * in_h * in_w;
// auto input_const = input_data_tmp;
//
// vnewbias = vdupq_n_f32(newbias_data[j]);
// vnewscale = vdupq_n_f32(newscale_data[j]);
//
// float w00 = filter_data_tmp[0];
// float w01 = filter_data_tmp[1];
// float w02 = filter_data_tmp[2];
// float w10 = filter_data_tmp[3];
// float w11 = filter_data_tmp[4];
// float w12 = filter_data_tmp[5];
// float w20 = filter_data_tmp[6];
// float w21 = filter_data_tmp[7];
// float w22 = filter_data_tmp[8];
//
// int h_mid = 0;
//
// for (; h_mid < out_h - 1; h_mid++) {
// input_row_ptr = input_data_tmp + 1 + h_mid * 2 * in_w;
// output_row_ptr = output_data_tmp + 1 + h_mid * out_w;
//
// for (int w4 = 0; w4 < w_times + 1; w4++) {
// if (h_mid == 0) {
// elewise_res1 = zero;
// elewise_res0 = zero;
// elewise_res2 = zero;
// } else {
// elewise_res1 = vmulq_n_f32(input_buff_bottom[w4].val[1], w01);
// elewise_res0 = vmulq_n_f32(input_buff_bottom[w4].val[0], w00);
// elewise_res2 = vmulq_n_f32(input_buff_bottom[w4].val[0], w02);
// }
// input_buff_mid = vld2q_f32(input_row_ptr);
// input_buff_bottom[w4] = vld2q_f32(input_row_ptr + in_w);
//
// elewise_res1 = vmlaq_n_f32(elewise_res1, input_buff_mid.val[1],
// w11); elewise_res0 = vmlaq_n_f32(elewise_res0,
// input_buff_mid.val[0], w10); elewise_res2 =
// vmlaq_n_f32(elewise_res2, input_buff_mid.val[0], w12);
//
// elewise_res1 =
// vmlaq_n_f32(elewise_res1, input_buff_bottom[w4].val[1],
// w21);
// elewise_res0 =
// vmlaq_n_f32(elewise_res0, input_buff_bottom[w4].val[0],
// w20);
// elewise_res2 =
// vmlaq_n_f32(elewise_res2, input_buff_bottom[w4].val[0],
// w22);
//
// res3 = vaddq_f32(vextq_f32(elewise_res2, zero, 1),
// vaddq_f32(elewise_res0, elewise_res1));
// res3 = vmlaq_f32(vnewbias, vnewscale, res3);
//
// if (if_relu) {
// res3 = vmaxq_f32(res3, zero);
// }
// vst1q_f32(output_row_ptr, res3);
//
// input_row_ptr += 6;
// output_row_ptr += 3;
// }
// }
// clock();
//
// input_row_ptr = input_data_tmp + 1 + h_mid * 2 * in_w;
// output_row_ptr = output_data_tmp + 1 + h_mid * out_w;
//
// for (int w4 = 0; w4 < w_times + 1; w4++) {
// elewise_res1 = vmulq_n_f32(input_buff_bottom[w4].val[1], w01);
// elewise_res0 = vmulq_n_f32(input_buff_bottom[w4].val[0], w00);
// elewise_res2 = vmulq_n_f32(input_buff_bottom[w4].val[0], w02);
//
// input_buff_mid = vld2q_f32(input_row_ptr);
// input_buff_bottom[w4] = vld2q_f32(input_row_ptr + in_w);
//
// elewise_res1 = vmlaq_n_f32(elewise_res1, input_buff_mid.val[1],
// w11); elewise_res0 = vmlaq_n_f32(elewise_res0,
// input_buff_mid.val[0], w10); elewise_res2 =
// vmlaq_n_f32(elewise_res2, input_buff_mid.val[0], w12);
//
// if (!if_pad) {
// elewise_res1 =
// vmlaq_n_f32(elewise_res1, input_buff_bottom[w4].val[1],
// w21);
// elewise_res0 =
// vmlaq_n_f32(elewise_res0, input_buff_bottom[w4].val[0],
// w20);
// elewise_res2 =
// vmlaq_n_f32(elewise_res2, input_buff_bottom[w4].val[0],
// w22);
// }
// res3 = vaddq_f32(vextq_f32(elewise_res2, zero, 1),
// vaddq_f32(elewise_res0, elewise_res1));
// res3 = vmlaq_f32(vnewbias, vnewscale, res3);
//
// if (if_relu) {
// res3 = vmaxq_f32(res3, zero);
// }
// if ((w4 != w_times)) {
// vst1q_f32(output_row_ptr, res3);
// } else {
// if (out_l - 2 - w_times * 3 == 1) {
// vst1q_lane_f32(output_row_ptr, res3, 0);
// } else if (out_l - 2 - w_times * 3 == 2) {
// vst1q_lane_f32(output_row_ptr, res3, 0);
// vst1q_lane_f32(output_row_ptr + 1, res3, 1);
// }
// }
// input_row_ptr += 6;
// output_row_ptr += 3;
// }
//
// output_data_tmp[0] = input_const[0] * w11 + input_const[1] * w12 +
// input_const[in_l] * w21 +
// input_const[in_l + 1] * w22;
//
// out2in_mid = (out_l - 1) * 2;
// output_data_tmp[out_l - 1] =
// w10 * input_const[out2in_mid - 1] + w11 *
// input_const[out2in_mid] + w20 * input_const[out2in_mid + in_w -
// 1] + w21 * input_const[out2in_mid + in_w] + (1 - if_pad) * (w12
// * input_const[out2in_mid + 1] +
// w22 * input_const[out2in_mid + in_w + 1]);
//
// out2in_mid = (out_l - 1) * 2 * in_w;
//
// output_data_tmp[out_l * (out_l - 1)] =
// w01 * input_const[out2in_mid - in_w] +
// w02 * input_const[out2in_mid - in_w + 1] +
// w11 * input_const[out2in_mid] + w12 * input_const[out2in_mid +
// 1] + (1 - if_pad) * (w21 * input_const[out2in_mid + in_w] +
// w22 * input_const[out2in_mid + in_w + 1]);
// out2in_mid = (out_l - 1) * 2 * in_w + (out_l - 1) * 2;
//
// output_data_tmp[out_l * out_l - 1] =
// w00 * input_const[out2in_mid - in_w - 1] +
// w01 * input_const[out2in_mid - in_w] +
// w10 * input_const[out2in_mid - 1] + w11 *
// input_const[out2in_mid] + (1 - if_pad) * (w20 *
// input_const[out2in_mid + in_w - 1] +
// w21 * input_const[out2in_mid + in_w] +
// w02 * input_const[out2in_mid - in_w + 1] +
// w12 * input_const[out2in_mid + 1] +
// w22 * input_const[out2in_mid + in_w + 1]);
// output_data_tmp[0] =
// output_data_tmp[0] * newscale_data[j] + newbias_data[j];
// output_data_tmp[out_l - 1] =
// output_data_tmp[out_l - 1] * newscale_data[j] + newbias_data[j];
// output_data_tmp[out_l * (out_l - 1)] =
// output_data_tmp[out_l * (out_l - 1)] * newscale_data[j] +
// newbias_data[j];
// output_data_tmp[out_l * out_l - 1] =
// output_data_tmp[out_l * out_l - 1] * newscale_data[j] +
// newbias_data[j];
// if (if_relu) {
// output_data_tmp[0] = output_data_tmp[0] < 0 ? 0 :
// output_data_tmp[0]; output_data_tmp[out_l - 1] =
// output_data_tmp[out_l - 1] < 0 ? 0 : output_data_tmp[out_l -
// 1];
// output_data_tmp[out_l * (out_l - 1)] =
// output_data_tmp[out_l * (out_l - 1)] < 0
// ? 0
// : output_data_tmp[out_l * (out_l - 1)];
// output_data_tmp[out_l * out_l - 1] =
// output_data_tmp[out_l * out_l - 1] < 0
// ? 0
// : output_data_tmp[out_l * out_l - 1];
// }
// for (int i = 1; i < out_h - 1; i++) {
// out2in_mid = i * 2 * in_w;
// output_data_tmp[i * out_l] = w01 * input_const[out2in_mid - in_w]
// +
// w02 * input_const[out2in_mid - in_w +
// 1] + w11 * input_const[out2in_mid] +
// w12 * input_const[out2in_mid + 1] +
// w21 * input_const[out2in_mid + in_w]
// + w22 * input_const[out2in_mid + in_w
// + 1];
//
// out2in_mid = i * 2 * in_w + (out_l - 1) * 2;
// output_data_tmp[i * out_l + out_l - 1] =
// w00 * input_const[out2in_mid - in_w - 1] +
// w01 * input_const[out2in_mid - in_w] +
// w10 * input_const[out2in_mid - 1] + w11 *
// input_const[out2in_mid] + w20 * input_const[out2in_mid + in_w
// - 1] + w21 * input_const[out2in_mid + in_w] + (1 - if_pad) *
// (w02 * input_const[out2in_mid - in_w + 1] +
// w12 * input_const[out2in_mid + 1] +
// w22 * input_const[out2in_mid + in_w + 1]);
// output_data_tmp[i * out_l] =
// output_data_tmp[i * out_l] * newscale_data[j] +
// newbias_data[j];
// output_data_tmp[i * out_l + out_l - 1] =
// output_data_tmp[i * out_l + out_l - 1] * newscale_data[j] +
// newbias_data[j];
// if (if_relu) {
// output_data_tmp[i * out_l] =
// output_data_tmp[i * out_l] < 0 ? 0 : output_data_tmp[i *
// out_l];
// output_data_tmp[i * out_l + out_l - 1] =
// output_data_tmp[i * out_l + out_l - 1] < 0
// ? 0
// : output_data_tmp[i * out_l + out_l - 1];
// }
// }
// filter_data_tmp += 9;
// }
// input_data += inhxw * c;
// output_data += outhxw * c;
// }
#endif
}
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或