experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Fixing a couple bugs in the xtensa_hifimini conv implementation:

Browse Source 
1. inner loop counters were not incremented correctly
  2. quantization multiplier scaling was wrong

PiperOrigin-RevId: 336117753
Change-Id: I65bae033097436ed1ed58a342ac28ac731f4296c
This commit is contained in:
A. Unique TensorFlower 2020-10-08 10:44:05 -07:00 committed by TensorFlower Gardener
parent 7ada484ae4
commit dc666bf0f4
1 changed files with 8 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
tensorflow/lite/micro/kernels/xtensa_hifimini/conv.cc
View File

@ -82,7 +82,6 @@ void ConvPerChannel(const ConvParams& params, const int32_t* output_multiplier,
const int input_height = input_shape.Dims(1); const int input_height = input_shape.Dims(1);
const int input_width = input_shape.Dims(2); const int input_width = input_shape.Dims(2);
const int input_depth = input_shape.Dims(3); const int input_depth = input_shape.Dims(3);
const int input_depth_iters = input_depth / 2;
const int filter_height = filter_shape.Dims(1); const int filter_height = filter_shape.Dims(1);
const int filter_width = filter_shape.Dims(2); const int filter_width = filter_shape.Dims(2);
@ -106,7 +105,7 @@ void ConvPerChannel(const ConvParams& params, const int32_t* output_multiplier,
ae_q56s acc_56 = AE_ZEROQ56(); ae_q56s acc_56 = AE_ZEROQ56();
for (int filter_y = 0; filter_y < filter_height; ++filter_y) { for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
for (int filter_x = 0; filter_x < filter_width; ++filter_x) { for (int filter_x = 0; filter_x < filter_width; filter_x += 2) {
const int in_x = in_x_origin + dilation_width_factor * filter_x; const int in_x = in_x_origin + dilation_width_factor * filter_x;
const int in_y = in_y_origin + dilation_height_factor * filter_y; const int in_y = in_y_origin + dilation_height_factor * filter_y;
const bool is_point_inside_image = const bool is_point_inside_image =
@ -119,10 +118,10 @@ void ConvPerChannel(const ConvParams& params, const int32_t* output_multiplier,
// with intrinsics: // with intrinsics:
int input_idx = int input_idx =
((batch * input_height + in_y) * input_width + in_x) * ((batch * input_height + in_y) * input_width + in_x) *
input_depth - input_depth * 2 -
2; 2;
const int8_t* input_vals_offset_ptr = input_data + input_idx; const int8_t* input_vals_offset_ptr = input_data + input_idx;
for (int i = 0; i < input_depth_iters; ++i) { for (int i = 0; i < input_depth; i += 2) {
// Load signed 2x 8bit values and right shift into 24bit // Load signed 2x 8bit values and right shift into 24bit
// alignment: // alignment:
ae_p24x2s input_vals_24x2; ae_p24x2s input_vals_24x2;
@ -139,7 +138,7 @@ void ConvPerChannel(const ConvParams& params, const int32_t* output_multiplier,
((out_channel * filter_height + filter_y) * filter_width + ((out_channel * filter_height + filter_y) * filter_width +
filter_x) * filter_x) *
filter_depth + filter_depth +
(i * 2) - 2; i - 2;
const int8_t* filter_vals_offset_ptr = const int8_t* filter_vals_offset_ptr =
filter_data + filter_idx; filter_data + filter_idx;
@ -171,9 +170,10 @@ void ConvPerChannel(const ConvParams& params, const int32_t* output_multiplier,
ae_p24x2s acc_24x2 = AE_TRUNCP24Q48(acc_56); ae_p24x2s acc_24x2 = AE_TRUNCP24Q48(acc_56);
// Apply quantized multiplier and accumulate result at 48bit // Apply quantized multiplier and accumulate result at 48bit
// alignment: // alignment. Convert the (unsigned) 32-bit multiplier down to a
acc_56 = ops::micro::xtensa::hifimini::MultiplyByQuantizedMultiplier( // 24-bit multiplier.
acc_24x2, output_multiplier[out_channel], acc_56 = micro::xtensa::hifimini::MultiplyByQuantizedMultiplier(
acc_24x2, output_multiplier[out_channel] >> 8,
output_shift[out_channel]); output_shift[out_channel]);
// Add output offset, cap activation, and assign to the output: // Add output offset, cap activation, and assign to the output: