| /external/tensorflow/tensorflow/contrib/lite/kernels/internal/optimized/ |
| H A D | depthwiseconv_float.h | 36 const float* filter_ptr, float* acc_buffer_ptr) {
40 filter[i] = vld1q_f32(filter_ptr + 4 * i);
97 const float* filter_ptr, float* acc_buffer_ptr) {
98 const float32x2_t filters = vld1_f32(filter_ptr);
180 const float* filter_ptr, float* acc_buffer_ptr) {
183 const float* local_filter_ptr = filter_ptr;
251 const float* filter_ptr, float* acc_buffer_ptr) {
254 const float* local_filter_ptr = filter_ptr;
321 const float* filter_ptr, float* acc_buffer_ptr) {
324 const float* local_filter_ptr = filter_ptr;
34 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
95 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
178 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
249 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
319 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
425 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
458 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
493 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
526 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
578 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
618 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
656 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
692 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
739 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
883 const float* filter_ptr = filter_base_ptr; local
[all...] |
| H A D | depthwiseconv_uint8.h | 36 int input_ptr_increment, const uint8* filter_ptr,
40 filter_u8.val[0] = vld1_u8(filter_ptr);
41 filter_u8.val[1] = vld1_u8(filter_ptr + 8);
83 int input_ptr_increment, const uint8* filter_ptr,
86 const uint8x8_t filter_u8 = vld1_u8(filter_ptr);
151 int input_ptr_increment, const uint8* filter_ptr,
154 const uint8x8_t filter_u8 = vld1_u8(filter_ptr);
221 int input_ptr_increment, const uint8* filter_ptr,
226 const uint8x8_t filter_u8 = vld1_u8(filter_ptr + 8 * i);
298 int input_ptr_increment, const uint8* filter_ptr,
34 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
81 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
149 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
219 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
296 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
362 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
476 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
536 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
631 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
701 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
786 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
872 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
937 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1022 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1076 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1110 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1154 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1210 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1259 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1291 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1354 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1418 Run(int num_output_pixels, int input_depth, int depth_multiplier, const uint8* input_ptr, int16 input_offset, int input_ptr_increment, const uint8* filter_ptr, int16 filter_offset, int32* acc_buffer_ptr) argument
1584 const uint8* filter_ptr = filter_base_ptr; local
[all...] |
| H A D | optimized_ops.h | 418 const uint8* filter_ptr = filter_data + in + (out + k) * input_size; local
419 filter_val_u8[k] = vld1q_u8(filter_ptr);
420 preload_l1_stream(filter_ptr + 64);
453 const uint8* filter_ptr = filter_data + in + (out + k) * input_size; local
454 filter_val_u8[k] = vld1_u8(filter_ptr);
|
| /external/tensorflow/tensorflow/core/kernels/ |
| H A D | dilation_ops_gpu.cu.cc | 40 const T* filter_ptr, int batch, int input_rows,
66 filter_ptr[d + depth * (w + filter_cols * h)];
80 const int32 nthreads, const T* input_ptr, const T* filter_ptr,
110 filter_ptr[d + depth * (w + filter_cols * h)];
129 const int32 nthreads, const T* input_ptr, const T* filter_ptr,
159 filter_ptr[d + depth * (w + filter_cols * h)];
39 DilationKernel(const int32 nthreads, const T* input_ptr, const T* filter_ptr, int batch, int input_rows, int input_cols, int depth, int filter_rows, int filter_cols, int output_rows, int output_cols, int stride_rows, int stride_cols, int rate_rows, int rate_cols, int pad_top, int pad_left, T* output_ptr) argument
79 DilationBackpropInputKernel( const int32 nthreads, const T* input_ptr, const T* filter_ptr, const T* out_backprop_ptr, int batch, int input_rows, int input_cols, int depth, int filter_rows, int filter_cols, int output_rows, int output_cols, int stride_rows, int stride_cols, int rate_rows, int rate_cols, int pad_top, int pad_left, T* in_backprop_ptr) argument
128 DilationBackpropFilterKernel( const int32 nthreads, const T* input_ptr, const T* filter_ptr, const T* out_backprop_ptr, int batch, int input_rows, int input_cols, int depth, int filter_rows, int filter_cols, int output_rows, int output_cols, int stride_rows, int stride_cols, int rate_rows, int rate_cols, int pad_top, int pad_left, T* filter_backprop_ptr) argument
|
| H A D | conv_ops.cc | 179 auto filter_ptr = filter.template flat<float>().data(); local
182 functor::DeepConv2D<CPUDevice, float>()(ctx, args, input_ptr, filter_ptr,
246 auto filter_ptr = filter.template flat<float>().data(); local
250 ctx, desc, input_ptr, filter_ptr, output_ptr);
676 auto filter_ptr = local
723 input_desc, input_ptr, filter_desc, filter_ptr, conv_desc,
760 filter_ptr, conv_desc, output_desc,
|
| H A D | conv_ops_3d.cc | 362 auto filter_ptr = local
410 input_desc, input_ptr, filter_desc, filter_ptr, conv_desc,
445 filter_ptr, conv_desc, output_desc,
|
| H A D | depthwise_conv_op.cc | 400 auto filter_ptr = filter.template flat<T>().data(); variable
402 LaunchDepthwiseConvOp<Device, T>()(context, args, input_ptr, filter_ptr,
|
| H A D | conv_grad_input_ops.cc | 171 auto filter_ptr = kernel.data(); local
175 context, desc, input_ptr, filter_ptr, output_ptr);
926 auto filter_ptr = local
973 filter_desc, filter_ptr, output_desc, out_backprop_ptr,
1007 filter_desc, filter_ptr, output_desc, out_backprop_ptr, conv_desc,
|
| H A D | depthwise_conv_op_gpu.cu.cc | 275 const T* filter_ptr = filter_read_offset + shared_data;
279 filter_ptr -= kBlockDepth;
281 const T filter_value = *filter_ptr;
287 filter_ptr += kBlockDepth;
557 const T* filter_ptr = filter_read_offset + shared_data;
561 filter_ptr -= kBlockDepth;
563 const T filter_value = *filter_ptr;
569 filter_ptr += kBlockDepth;
|
| H A D | conv_grad_ops_3d.cc | 630 auto filter_ptr = variable
677 filter_desc, filter_ptr, output_desc, out_backprop_ptr,
713 filter_desc, filter_ptr, output_desc, out_backprop_ptr,
|
| H A D | conv_grad_filter_ops.cc | 169 auto filter_ptr = filter.data(); local
172 context, desc, input_ptr, filter_ptr, output_ptr);
|
| H A D | depthwise_conv_grad_op.cc | 575 auto filter_ptr = filter.template flat<T>().data(); variable
582 context, args, out_backprop_ptr, filter_ptr, in_backprop_ptr,
|
| /external/tensorflow/tensorflow/core/kernels/neon/ |
| H A D | depthwiseconv_float.h | 40 const float* filter_ptr, float* acc_buffer_ptr) {
44 filter[i] = vld1q_f32(filter_ptr + 4 * i);
101 const float* filter_ptr, float* acc_buffer_ptr) {
102 const float32x2_t filters = vld1_f32(filter_ptr);
184 const float* filter_ptr, float* acc_buffer_ptr) {
187 const float* local_filter_ptr = filter_ptr;
253 const float* filter_ptr, float* acc_buffer_ptr) {
256 const float* local_filter_ptr = filter_ptr;
320 const float* filter_ptr, float* acc_buffer_ptr) {
323 const float* local_filter_ptr = filter_ptr;
38 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
99 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
182 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
251 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
318 Run(int num_output_pixels, int input_depth, int depth_multiplier, const float* input_ptr, int input_ptr_increment, const float* filter_ptr, float* acc_buffer_ptr) argument
525 const float* filter_ptr = filter_base_ptr; local
[all...] |
| H A D | neon_depthwise_conv_op.cc | 123 const float* filter_ptr = filter.template flat<float>().data(); variable
136 input_ptr, input_neon_dims, filter_ptr, filter_neon_dims, bias_ptr,
|
| /external/tensorflow/tensorflow/contrib/fused_conv/kernels/ |
| H A D | fused_conv2d_bias_activation_op.cc | 504 auto filter_ptr = local
559 filter_desc, filter_ptr, conv_desc, side_input_ptr,
598 filter_ptr, conv_desc, side_input_ptr, side_input_scale,
|