IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'upstream/master' into layer_test_common

Browse Source
This commit is contained in:
Efode, Irina 2021-10-21 23:53:54 +03:00
commit 410c703682
10 changed files with 363 additions and 325 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

416
inference-engine/src/gna_plugin/backend/make_pwl.cpp
View File

@ -4,6 +4,7 @@
#include <vector> #include <vector>
#include <iostream> #include <iostream>
#include <iomanip>
#include <cmath> #include <cmath>
#include "runtime/pwl.h" #include "runtime/pwl.h"
@ -58,7 +59,72 @@ static void insert_extra_pwl_segments(std::vector<gna_pwl_segment_t>& gna_pwl,
} }
} }
void make_gna_pwl(const DnnActivation fun, static void print_segments_header(const DnnActivation& fun) {
gnalog() << "=========================== " << intel_dnn_activation_name[fun] <<
" segments ===========================\n";
gnalog() << std::setw(12) << std::setfill(' ') << "x" << std::setw(12) << std::setfill(' ') <<
"y" << std::setw(12) << std::setfill(' ') << "slope" << std::endl;
}
static void print_segment(double x, double y, double slope) {
gnalog() << std::setw(12) << std::setfill(' ') << x << std::setw(12) << std::setfill(' ') <<
y << std::setw(12) << std::setfill(' ') << slope << std::endl;
}
static std::vector<gna_pwl_segment_t> create_multisegment_gna_pwl(const std::vector<pwl_t>& pwl,
double in_scale,
double out_scale,
double min_x_val,
double max_x_val,
double min_y_val,
double max_y_val,
bool fake_quantize,
bool add_last_seg) {
std::vector<gna_pwl_segment_t> gna_pwl;
int32_t xbase = static_cast<int32_t> (INT32_MIN & XBASEMASK); // zero out the 2 lsb
int16_t ybase = FLOAT_TO_INT16(min_y_val * out_scale);
int16_t slope = 0;
gna_pwl.push_back({xbase, ybase, slope});
print_segment(xbase / in_scale, min_y_val, slope);
if (!fake_quantize && min_x_val > INT32_MIN / in_scale) {
auto s = gna_slope(pwl[0].m, in_scale, out_scale);
slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
xbase = (static_cast<int32_t>(min_x_val * in_scale) & XBASEMASK) | s.slope_scale_index;
ybase = FLOAT_TO_INT16(min_y_val * out_scale);
gna_pwl.push_back({xbase, ybase, slope});
print_segment(min_x_val, min_y_val, pwl[0].m);
}
for (uint32_t i = 0; i < pwl.size(); ++i) {
if (!fake_quantize && (pwl[i].alpha <= min_x_val ||
pwl[i].alpha <= INT32_MIN / in_scale ||
pwl[i].alpha >= max_x_val)) {
continue;
}
auto s = gna_slope(pwl[i].m, in_scale, out_scale);
xbase = ((static_cast<int32_t> (in_scale * pwl[i].alpha)) & XBASEMASK) | s.slope_scale_index;
ybase = FLOAT_TO_INT16(pwl[i].beta * out_scale);
slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl.push_back({xbase, ybase, slope});
print_segment(pwl[i].alpha, pwl[i].beta, pwl[i].m);
}
if (!fake_quantize && add_last_seg) {
// insert extra segment for xvalues > u_bound
xbase = static_cast<int32_t>(max_x_val * in_scale) & XBASEMASK;
ybase = FLOAT_TO_INT16(max_y_val * out_scale);
slope = 0;
gna_pwl.push_back({xbase, ybase, slope});
print_segment(max_x_val, max_y_val, slope);
}
return gna_pwl;
}
void make_gna_pwl(const DnnActivation& fun,
const std::vector<pwl_t>& pwl, const std::vector<pwl_t>& pwl,
const double l_bound, const double l_bound,
const double u_bound, const double u_bound,
@ -73,199 +139,56 @@ void make_gna_pwl(const DnnActivation fun,
gnalog() << "make_gna_pwl\n"; gnalog() << "make_gna_pwl\n";
gnalog() << " in_scale " << in_scale << "\n"; gnalog() << " in_scale " << in_scale << "\n";
gnalog() << " out_scale " << out_scale << "\n"; gnalog() << " out_scale " << out_scale << "\n";
print_segments_header(fun);
switch (fun) { switch (fun) {
case kActSigmoid: case kActSigmoid:
case kActTanh: case kActTanh:
case kActSoftSign: { case kActSoftSign: {
auto n_segments = static_cast<int32_t> (pwl_size) + 1;
gna_pwl.resize(n_segments);
// insert extra segment for x values < l_bound // insert extra segment for x values < l_bound
gna_pwl[0].xBase = static_cast<int32_t> (INT32_MIN & XBASEMASK); // zero out the 2 lsb double min_x_val;
double min_y_val;
if (fun == kActSigmoid) { if (fun == kActSigmoid) {
gnalog() << "=========================== Sigmoid Segments ===========================\n"; min_y_val = fun.fqParams.set ? pwl[0].beta : 0;
auto minVal = (fun.fqParams.set && *fun.fqParams.input_low > 0) ? FLOAT_TO_INT16(*fun.fqParams.input_low * out_scale) : 0; min_x_val = -pwl[0].b / pwl[0].m;
gna_pwl[0].yBase = gna_pwl[1].yBase = minVal;
gna_pwl[1].xBase = (static_cast<int32_t> (in_scale * (-pwl[0].b / pwl[0].m))) & XBASEMASK;
} else if (fun == kActTanh) { } else if (fun == kActTanh) {
gnalog() << "=========================== Tanh Segments ===========================\n"; min_y_val = fun.fqParams.set ? pwl[0].beta : -1.0;
auto minVal = (fun.fqParams.set && *fun.fqParams.input_low > -1) ? FLOAT_TO_INT16(*fun.fqParams.input_low * out_scale) : min_x_val = (-1.0 - pwl[0].b) / pwl[0].m;
static_cast<int16_t>(-1.0 * out_scale);
gna_pwl[0].yBase = gna_pwl[1].yBase = minVal;
gna_pwl[1].xBase = (static_cast<int32_t> (in_scale * (-1.0 - pwl[0].b) / pwl[0].m)) & XBASEMASK;
} else { } else {
gnalog() << "=========================== SoftSign Segments ===========================\n"; min_y_val = fun.fqParams.set ? pwl[0].beta : -1.0;
auto minVal = (fun.fqParams.set && *fun.fqParams.input_low > -1) ? FLOAT_TO_INT16(*fun.fqParams.input_low * out_scale) : min_x_val = (-1.0 - pwl[0].b) / pwl[0].m;
static_cast<int16_t>(-1.0 * out_scale);
gna_pwl[0].yBase = gna_pwl[1].yBase = minVal;
gna_pwl[1].xBase = (static_cast<int32_t> (in_scale * (-1.0 - pwl[0].b) / pwl[0].m)) & XBASEMASK;
} }
gna_pwl[0].slope = 0; double max_y_val = fun.fqParams.set ? pwl.back().beta : 1.0;
double max_x_val = fun.srcFQParams.set ? u_bound : (1.0 - pwl[pwl_size - 2].b) / pwl[pwl_size - 2].m;
gnalog() << (gna_pwl[0].xBase) / in_scale gna_pwl = create_multisegment_gna_pwl(pwl, in_scale, out_scale, min_x_val, max_x_val, min_y_val, max_y_val,
<< " " << (gna_pwl[0].yBase) / out_scale fun.fqParams.set, true);
<< " " << 0.0
<< "\n";
s = gna_slope(pwl[0].m, in_scale, out_scale);
gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index;
gnalog() << (gna_pwl[1].xBase/in_scale)
<< " " << (gna_pwl[1].yBase) / out_scale
<< " " << pwl[0].m
<< "\n";
for (uint32_t i = 1; i < pwl_size - 1; ++i) {
s = gna_slope(pwl[i].m, in_scale, out_scale);
gna_pwl[i + 1].xBase = (static_cast<int32_t> (in_scale * pwl[i].alpha)) & XBASEMASK;
gna_pwl[i + 1].yBase = FLOAT_TO_INT16(pwl[i].beta * out_scale);
gna_pwl[i + 1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[i + 1].xBase = gna_pwl[i + 1].xBase | s.slope_scale_index;
gnalog() << (pwl[i].alpha)
<< " " << pwl[i].beta
<< " " << pwl[i].m
<< "\n";
}
// insert extra segment for xvalues > u_bound
auto maxVal = (fun.fqParams.set && *fun.fqParams.input_high <= 1) ? *fun.fqParams.input_high : 1.0;
gna_pwl[n_segments - 1].xBase =
((uint32_t) (in_scale * (1.0 - pwl[pwl_size - 2].b) / pwl[pwl_size - 2].m)) & XBASEMASK;
gna_pwl[n_segments - 1].yBase = FLOAT_TO_INT16(maxVal * out_scale);
gna_pwl[n_segments - 1].slope = 0;
gnalog() << (gna_pwl[n_segments - 1].xBase / in_scale)
<< " " << 1.0
<< " " << 0.0
<< "\n";
break; break;
} }
case kActExp: { case kActExp: {
auto n_segments = static_cast<int32_t> (pwl_size) + 1; double min_x_val = -pwl[0].b / pwl[0].m;
gna_pwl.resize(n_segments); double max_x_val = (y_max/out_scale - pwl[pwl_size - 2].b) / pwl[pwl_size - 2].m;
// insert extra segment for x values < l_bound double min_y_val = fun.fqParams.set ? pwl[0].beta : 0;
gna_pwl[0].xBase = static_cast<int32_t> (INT32_MIN & XBASEMASK); // zero out the 2 lsb double max_y_val = fun.fqParams.set ? pwl.front().beta : y_max / out_scale;
gnalog() << "=========================== Exp Segments ===========================\n"; gna_pwl = create_multisegment_gna_pwl(pwl, in_scale, out_scale, min_x_val, max_x_val, min_y_val, max_y_val,
gna_pwl[0].yBase = gna_pwl[1].yBase = 0; fun.fqParams.set, true);
gna_pwl[1].xBase = (static_cast<int32_t> (in_scale * (-pwl[0].b / pwl[0].m))) & XBASEMASK;
gna_pwl[0].slope = 0;
gnalog() << (gna_pwl[0].xBase) / in_scale
<< " " << (gna_pwl[0].yBase) / out_scale
<< " " << 0.0
<< "\n";
s = gna_slope(pwl[0].m, in_scale, out_scale);
gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index;
gnalog() << ((int32_t)(gna_pwl[1].xBase & XBASEMASK) / in_scale)
<< " " << (gna_pwl[1].yBase) / out_scale
<< " " << pwl[0].m
<< "\n";
for (uint32_t i = 1; i < pwl_size - 1; ++i) {
s = gna_slope(pwl[i].m, in_scale, out_scale);
gna_pwl[i + 1].xBase = (static_cast<int32_t> (in_scale * pwl[i].alpha)) & XBASEMASK;
gna_pwl[i + 1].yBase = FLOAT_TO_INT16(pwl[i].beta * out_scale);
gna_pwl[i + 1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[i + 1].xBase = gna_pwl[i + 1].xBase | s.slope_scale_index;
gnalog() << (pwl[i].alpha)
<< " " << pwl[i].beta
<< " " << pwl[i].m
<< "\n";
}
// insert extra segment for xvalues > u_bound
gna_pwl[n_segments - 1].xBase =
((uint32_t)(in_scale * (y_max/out_scale - pwl[pwl_size - 2].b) / pwl[pwl_size - 2].m)) & XBASEMASK;
gna_pwl[n_segments - 1].yBase = y_max;
gna_pwl[n_segments - 1].slope = 0;
gnalog() << (gna_pwl[n_segments - 1].xBase / in_scale)
<< " " << 1.0
<< " " << 0.0
<< "\n";
break; break;
} }
case kActLog: { case kActLog: {
auto n_segments = static_cast<int32_t> (pwl_size); double min_x_val = 1 + ~XBASEMASK;
gna_pwl.resize(n_segments); double max_x_val = INT32_MAX / in_scale;
// insert extra segment for x values < l_bound double min_y_val = y_min / out_scale;
gna_pwl[0].xBase = static_cast<int32_t> (INT32_MIN & XBASEMASK); // zero out the 2 lsb double max_y_val = y_max / out_scale;
gnalog() << "=========================== Log Segments ===========================\n"; gna_pwl = create_multisegment_gna_pwl(pwl, in_scale, out_scale, min_x_val, max_x_val, min_y_val, max_y_val,
gna_pwl[0].yBase = gna_pwl[1].yBase = y_min; fun.fqParams.set, false);
gna_pwl[1].xBase = (static_cast<int32_t> (1 + ~XBASEMASK)); // smallest representable value
gna_pwl[0].slope = 0;
gnalog() << gna_pwl[0].xBase / in_scale
<< " " << (gna_pwl[0].yBase) / out_scale
<< " " << 0.0
<< "\n";
s = gna_slope(pwl[0].m, in_scale, out_scale);
gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index;
gnalog() << ((int32_t)(gna_pwl[1].xBase & XBASEMASK) / in_scale)
<< " " << (gna_pwl[1].yBase) / out_scale
<< " " << pwl[0].m
<< "\n";
for (uint32_t i = 1; i < pwl_size - 1; ++i) {
s = gna_slope(pwl[i].m, in_scale, out_scale);
gna_pwl[i + 1].xBase = (static_cast<int32_t> (in_scale * pwl[i].alpha)) & XBASEMASK;
gna_pwl[i + 1].yBase = FLOAT_TO_INT16(pwl[i].beta * out_scale);
gna_pwl[i + 1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[i + 1].xBase = gna_pwl[i + 1].xBase | s.slope_scale_index;
gnalog() << (pwl[i].alpha)
<< " " << pwl[i].beta
<< " " << pwl[i].m
<< "\n";
}
break; break;
} }
case kActNegLog: case kActNegLog:
case kActNegHalfLog: { case kActNegHalfLog: {
auto n_segments = static_cast<int32_t> (pwl_size); double min_x_val = 1 + ~XBASEMASK;
gna_pwl.resize(n_segments); double max_x_val = INT32_MAX / in_scale;
// insert extra segment for x values < l_bound double min_y_val = y_max / out_scale;
gna_pwl[0].xBase = static_cast<int32_t> (INT32_MIN & XBASEMASK); // zero out the 2 lsb double max_y_val = y_min / out_scale;
if (fun == kActNegHalfLog) gna_pwl = create_multisegment_gna_pwl(pwl, in_scale, out_scale, min_x_val, max_x_val, min_y_val, max_y_val,
gnalog() << "=========================== NegHalfLog Segments ===========================\n"; fun.fqParams.set, false);
else
gnalog() << "=========================== NegLog Segments ===========================\n";
gna_pwl[0].yBase = gna_pwl[1].yBase = y_max;
gna_pwl[1].xBase = (static_cast<int32_t> (1 + ~XBASEMASK)); // smallest representable value
gna_pwl[0].slope = 0;
gnalog() << gna_pwl[0].xBase / in_scale
<< " " << (gna_pwl[0].yBase) / out_scale
<< " " << 0.0
<< "\n";
s = gna_slope(pwl[0].m, in_scale, out_scale);
gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index;
gnalog() << ((int32_t)(gna_pwl[1].xBase & XBASEMASK) / in_scale)
<< " " << (gna_pwl[1].yBase) / out_scale
<< " " << pwl[0].m
<< "\n";
for (uint32_t i = 1; i < pwl_size - 1; ++i) {
s = gna_slope(pwl[i].m, in_scale, out_scale);
gna_pwl[i + 1].xBase = (static_cast<int32_t> (in_scale * pwl[i].alpha)) & XBASEMASK;
gna_pwl[i + 1].yBase = FLOAT_TO_INT16(pwl[i].beta * out_scale);
gna_pwl[i + 1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[i + 1].xBase = gna_pwl[i + 1].xBase | s.slope_scale_index;
gnalog() << (pwl[i].alpha)
<< " " << pwl[i].beta
<< " " << pwl[i].m
<< "\n";
}
break; break;
} }
case kActRelu: case kActRelu:
@ -273,10 +196,6 @@ void make_gna_pwl(const DnnActivation fun,
auto n_segments = 2; auto n_segments = 2;
gna_pwl.resize(n_segments); gna_pwl.resize(n_segments);
if (fun == kActRelu)
gnalog() << "=========================== ReLU Segments ===========================\n";
else
gnalog() << "=========================== LeakyReLU Segments ======================\n";
int32_t x_lower = INT32_MIN; int32_t x_lower = INT32_MIN;
int32_t x_upper = INT32_MAX; int32_t x_upper = INT32_MAX;
int32_t y_lower = y_min; int32_t y_lower = y_min;
@ -297,19 +216,16 @@ void make_gna_pwl(const DnnActivation fun,
gna_pwl[0].xBase = (x_lower & XBASEMASK) | s.slope_scale_index; // zero out the 2 lsb gna_pwl[0].xBase = (x_lower & XBASEMASK) | s.slope_scale_index; // zero out the 2 lsb
gna_pwl[0].slope = FLOAT_TO_INT16(s.slope * s.slope_scale); gna_pwl[0].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gnalog() << (int32_t)(gna_pwl[0].xBase & XBASEMASK) / in_scale print_segment((int32_t)(gna_pwl[0].xBase & XBASEMASK) / in_scale,
<< " " << gna_pwl[0].yBase / out_scale gna_pwl[0].yBase / out_scale,
<< " " << (gna_pwl[0].slope * in_scale) / (out_scale*s.slope_scale) (gna_pwl[0].slope * in_scale) / (out_scale*s.slope_scale));
<< "\n";
gna_pwl[1].xBase = 0; gna_pwl[1].xBase = 0;
gna_pwl[1].yBase = 0; gna_pwl[1].yBase = 0;
s = gna_slope(1.0, in_scale, out_scale); s = gna_slope(1.0, in_scale, out_scale);
gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale); gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index; gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index;
gnalog() << 0.0 print_segment(0.0, 0.0, (gna_pwl[1].slope * in_scale) / (out_scale*s.slope_scale));
<< " " << 0.0
<< " " << (gna_pwl[1].slope * in_scale) / (out_scale*s.slope_scale)
<< "\n";
if (fun.fqParams.set) { // need a right segment if (fun.fqParams.set) { // need a right segment
gna_pwl.push_back({ gna_pwl.push_back({
@ -317,10 +233,7 @@ void make_gna_pwl(const DnnActivation fun,
y_upper, y_upper,
0 }); 0 });
gnalog() << (x_upper & XBASEMASK) / in_scale print_segment((x_upper & XBASEMASK) / in_scale, gna_pwl[n_segments].yBase / out_scale, 0.0);
<< " " << gna_pwl[n_segments].yBase / out_scale
<< " " << 0
<< "\n";
} }
break; break;
} }
@ -328,34 +241,28 @@ void make_gna_pwl(const DnnActivation fun,
auto n_segments = 3; auto n_segments = 3;
gna_pwl.resize(n_segments); gna_pwl.resize(n_segments);
gnalog() << "=========================== Sign Segments ===========================\n";
int32_t x_lower = INT32_MIN; int32_t x_lower = INT32_MIN;
int16_t y_lower = static_cast<int16_t>(-1.0 * out_scale); int16_t y_lower = static_cast<int16_t>(-1.0 * out_scale);
gna_pwl[0].yBase = y_lower; gna_pwl[0].yBase = y_lower;
gna_pwl[0].xBase = (x_lower & XBASEMASK); // zero out the 2 lsb gna_pwl[0].xBase = (x_lower & XBASEMASK); // zero out the 2 lsb
gna_pwl[0].slope = 0; gna_pwl[0].slope = 0;
gnalog() << gna_pwl[0].xBase / in_scale print_segment(gna_pwl[0].xBase / in_scale, gna_pwl[0].yBase / out_scale,
<< " " << gna_pwl[0].yBase / out_scale (gna_pwl[0].slope * in_scale) / (out_scale*s.slope_scale));
<< " " << (gna_pwl[0].slope * in_scale) / (out_scale*s.slope_scale)
<< "\n";
gna_pwl[1].xBase = -1; gna_pwl[1].xBase = -1;
gna_pwl[1].yBase = 0; gna_pwl[1].yBase = 0;
gna_pwl[1].slope = 0; gna_pwl[1].slope = 0;
gna_pwl[1].xBase = gna_pwl[1].xBase & XBASEMASK; gna_pwl[1].xBase = gna_pwl[1].xBase & XBASEMASK;
gnalog() << gna_pwl[1].xBase / in_scale print_segment(gna_pwl[1].xBase / in_scale, gna_pwl[1].yBase / out_scale,
<< " " << gna_pwl[1].yBase / out_scale (gna_pwl[1].slope * in_scale) / (out_scale*s.slope_scale));
<< " " << (gna_pwl[1].slope * in_scale) / (out_scale*s.slope_scale)
<< "\n";
gna_pwl[2].xBase = 1 + ~XBASEMASK; // smallest representable positive number gna_pwl[2].xBase = 1 + ~XBASEMASK; // smallest representable positive number
gna_pwl[2].yBase = static_cast<int16_t>(1.0 * out_scale); gna_pwl[2].yBase = static_cast<int16_t>(1.0 * out_scale);
s = gna_slope(1.0, in_scale, out_scale); s = gna_slope(1.0, in_scale, out_scale);
gna_pwl[2].slope = 0; gna_pwl[2].slope = 0;
gna_pwl[2].xBase = gna_pwl[2].xBase & XBASEMASK; gna_pwl[2].xBase = gna_pwl[2].xBase & XBASEMASK;
gnalog() << gna_pwl[2].xBase / in_scale print_segment(gna_pwl[2].xBase / in_scale, gna_pwl[2].yBase / out_scale,
<< " " << gna_pwl[2].yBase / out_scale (gna_pwl[2].slope * in_scale) / (out_scale*s.slope_scale));
<< " " << (gna_pwl[2].slope * in_scale) / (out_scale*s.slope_scale)
<< "\n";
break; break;
} }
case kActIdentity: case kActIdentity:
@ -373,7 +280,6 @@ void make_gna_pwl(const DnnActivation fun,
} }
auto n_segments = 2; auto n_segments = 2;
if (fun == kActKaldiLstmClipping) { if (fun == kActKaldiLstmClipping) {
gnalog() << "=========================== Clipping Segments ===========================\n";
if (x_lower < l_bound * in_scale) { if (x_lower < l_bound * in_scale) {
if (y_lower < l_bound * out_scale) { if (y_lower < l_bound * out_scale) {
x_lower = FLOAT_TO_INT32(l_bound * in_scale); x_lower = FLOAT_TO_INT32(l_bound * in_scale);
@ -391,42 +297,32 @@ void make_gna_pwl(const DnnActivation fun,
} }
} }
} else if (fun == kActIdentity) { } else if (fun == kActIdentity) {
gnalog() << "=========================== Identity Segments ===========================\n";
if (x_lower < y_lower * in_scale / out_scale) x_lower = FLOAT_TO_INT32(y_lower * in_scale / out_scale); if (x_lower < y_lower * in_scale / out_scale) x_lower = FLOAT_TO_INT32(y_lower * in_scale / out_scale);
if (x_upper > y_upper * in_scale / out_scale) x_upper = FLOAT_TO_INT32(y_upper * in_scale / out_scale); if (x_upper > y_upper * in_scale / out_scale) x_upper = FLOAT_TO_INT32(y_upper * in_scale / out_scale);
if (y_lower < x_lower * out_scale / in_scale) y_lower = FLOAT_TO_INT16(x_lower * out_scale / in_scale); if (y_lower < x_lower * out_scale / in_scale) y_lower = FLOAT_TO_INT16(x_lower * out_scale / in_scale);
if (y_upper > x_upper * out_scale / in_scale) y_upper = FLOAT_TO_INT16(x_upper * out_scale / in_scale); if (y_upper > x_upper * out_scale / in_scale) y_upper = FLOAT_TO_INT16(x_upper * out_scale / in_scale);
} else if (fun == kActFakeQuantize) {
gnalog() << "=========================== Fake Quantize Segments ===========================\n";
} }
gna_pwl.resize(n_segments); gna_pwl.resize(n_segments);
gna_pwl[0].xBase = INT32_MIN & XBASEMASK; // zero out the 2 lsb gna_pwl[0].xBase = INT32_MIN & XBASEMASK; // zero out the 2 lsb
gna_pwl[0].yBase = y_lower; gna_pwl[0].yBase = y_lower;
gna_pwl[0].slope = 0; gna_pwl[0].slope = 0;
gnalog() << gna_pwl[0].xBase / in_scale print_segment(gna_pwl[0].xBase / in_scale, gna_pwl[0].yBase / out_scale, 0.0);
<< " " << gna_pwl[0].yBase / out_scale
<< " " << 0
<< "\n";
gna_pwl[1].xBase = x_lower & XBASEMASK; // zero out the 2 lsb gna_pwl[1].xBase = x_lower & XBASEMASK; // zero out the 2 lsb
gna_pwl[1].yBase = y_lower; gna_pwl[1].yBase = y_lower;
s = gna_slope(1.0, in_scale, out_scale); s = gna_slope(1.0, in_scale, out_scale);
gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale); gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index; gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index;
gnalog() << (int32_t)(gna_pwl[1].xBase & XBASEMASK) / in_scale print_segment((int32_t)(gna_pwl[1].xBase & XBASEMASK) / in_scale, gna_pwl[1].yBase / out_scale, 1.0);
<< " " << gna_pwl[1].yBase / out_scale
<< " " << 1.0
<< "\n";
if (INT32_MAX > x_upper) { // need a right segment if (INT32_MAX > x_upper) { // need a right segment
gna_pwl.push_back({ gna_pwl.push_back({
static_cast<int32_t>(x_upper & XBASEMASK), // zero out the 2 lsb static_cast<int32_t>(x_upper & XBASEMASK), // zero out the 2 lsb
y_upper, y_upper,
0 }); 0 });
gnalog() << (x_upper & XBASEMASK) / in_scale print_segment((x_upper & XBASEMASK) / in_scale, gna_pwl[n_segments].yBase / out_scale, 0.0);
<< " " << gna_pwl[n_segments].yBase / out_scale
<< " " << 0
<< "\n";
} }
break; break;
} }
@ -440,7 +336,6 @@ void make_gna_pwl(const DnnActivation fun,
if (y_upper > x_upper * out_scale / in_scale) y_upper = FLOAT_TO_INT16(x_upper * out_scale / in_scale); if (y_upper > x_upper * out_scale / in_scale) y_upper = FLOAT_TO_INT16(x_upper * out_scale / in_scale);
if (x_upper > y_upper * in_scale / out_scale) x_upper = FLOAT_TO_INT32(y_upper * in_scale / out_scale); if (x_upper > y_upper * in_scale / out_scale) x_upper = FLOAT_TO_INT32(y_upper * in_scale / out_scale);
gnalog() << "=========================== Abs Segments ===========================\n";
if (y_upper == y_max) { // saturation at ends - need one more segment if (y_upper == y_max) { // saturation at ends - need one more segment
n_segments += 1; n_segments += 1;
gna_pwl.resize(n_segments); gna_pwl.resize(n_segments);
@ -457,19 +352,14 @@ void make_gna_pwl(const DnnActivation fun,
s = gna_slope(-1.0, in_scale, out_scale); s = gna_slope(-1.0, in_scale, out_scale);
gna_pwl[i].slope = FLOAT_TO_INT16(s.slope * s.slope_scale); gna_pwl[i].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[i].xBase = gna_pwl[i].xBase | s.slope_scale_index; gna_pwl[i].xBase = gna_pwl[i].xBase | s.slope_scale_index;
gnalog() << (int32_t)(gna_pwl[i].xBase & XBASEMASK) / in_scale print_segment((int32_t)(gna_pwl[i].xBase & XBASEMASK) / in_scale, gna_pwl[i].yBase / out_scale, -1.0);
<< " " << gna_pwl[i].yBase / out_scale
<< " " << -1.0
<< "\n";
gna_pwl[i + 1].xBase = 0; gna_pwl[i + 1].xBase = 0;
gna_pwl[i + 1].yBase = 0; gna_pwl[i + 1].yBase = 0;
s = gna_slope(1.0, in_scale, out_scale); s = gna_slope(1.0, in_scale, out_scale);
gna_pwl[i + 1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale); gna_pwl[i + 1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[i + 1].xBase = gna_pwl[i + 1].xBase | s.slope_scale_index; gna_pwl[i + 1].xBase = gna_pwl[i + 1].xBase | s.slope_scale_index;
gnalog() << (int32_t)(gna_pwl[i + 1].xBase & XBASEMASK) / in_scale print_segment((int32_t)(gna_pwl[i + 1].xBase & XBASEMASK) / in_scale, gna_pwl[i + 1].yBase / out_scale, 1.0);
<< " " << gna_pwl[i + 1].yBase / out_scale
<< " " << 1.0
<< "\n";
break; break;
} }
case kActPow: { case kActPow: {
@ -551,11 +441,7 @@ void make_gna_pwl(const DnnActivation fun,
gna_pwl[0].xBase = INT32_MIN & XBASEMASK; // zero out the 2 lsb gna_pwl[0].xBase = INT32_MIN & XBASEMASK; // zero out the 2 lsb
gna_pwl[0].yBase = y_lower; gna_pwl[0].yBase = y_lower;
gna_pwl[0].slope = 0; gna_pwl[0].slope = 0;
gnalog() << gna_pwl[0].xBase / in_scale print_segment(gna_pwl[0].xBase / in_scale, gna_pwl[0].yBase / out_scale, 0.0);
<< " " << gna_pwl[0].yBase / out_scale
<< " " << 0
<< "\n";
gna_pwl[1].xBase = x_lower & XBASEMASK; // zero out the 2 lsb gna_pwl[1].xBase = x_lower & XBASEMASK; // zero out the 2 lsb
gna_pwl[1].yBase = y_lower; gna_pwl[1].yBase = y_lower;
@ -563,73 +449,27 @@ void make_gna_pwl(const DnnActivation fun,
s = gna_slope(slope, in_scale, out_scale); s = gna_slope(slope, in_scale, out_scale);
gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale); gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index; gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index;
gnalog() << (int32_t)(gna_pwl[1].xBase & XBASEMASK) / in_scale print_segment((int32_t)(gna_pwl[1].xBase & XBASEMASK) / in_scale, gna_pwl[1].yBase / out_scale, 1.0);
<< " " << gna_pwl[1].yBase / out_scale
<< " " << 1.0
<< "\n";
if (INT32_MAX > x_upper) { // need a right segment if (INT32_MAX > x_upper) { // need a right segment
gna_pwl.push_back({ gna_pwl.push_back({
static_cast<int32_t>(x_upper & XBASEMASK), // zero out the 2 lsb static_cast<int32_t>(x_upper & XBASEMASK), // zero out the 2 lsb
y_upper, y_upper,
0 }); 0 });
gnalog() << (x_upper & XBASEMASK) / in_scale print_segment((x_upper & XBASEMASK) / in_scale, gna_pwl[2].yBase / out_scale, 0.0);
<< " " << gna_pwl[2].yBase / out_scale
<< " " << 0
<< "\n";
} }
} else { } else {
auto n_segments = static_cast<int32_t> (pwl_size) + 1; double min_x_val = -pwl[0].b / pwl[0].m;
gna_pwl.resize(n_segments); double max_x_val = (y_max/out_scale - pwl[pwl_size - 2].b) / pwl[pwl_size - 2].m;
// insert extra segment for x values < l_bound double min_y_val = fun.fqParams.set ? pwl[0].beta : 0;
gna_pwl[0].xBase = static_cast<int32_t> (INT32_MIN & XBASEMASK); // zero out the 2 lsb double max_y_val = fun.fqParams.set ? pwl.front().beta : y_max / out_scale;
gnalog() << "=========================== Exp Segments ===========================\n"; gna_pwl = create_multisegment_gna_pwl(pwl, in_scale, out_scale, min_x_val, max_x_val, min_y_val, max_y_val,
gna_pwl[0].yBase = gna_pwl[1].yBase = 0; fun.fqParams.set, true);
gna_pwl[1].xBase = (static_cast<int32_t> (in_scale * (-pwl[0].b / pwl[0].m))) & XBASEMASK;
gna_pwl[0].slope = 0;
gnalog() << (gna_pwl[0].xBase) / in_scale
<< " " << (gna_pwl[0].yBase) / out_scale
<< " " << 0.0
<< "\n";
s = gna_slope(pwl[0].m, in_scale, out_scale);
gna_pwl[1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[1].xBase = gna_pwl[1].xBase | s.slope_scale_index;
gnalog() << ((int32_t)(gna_pwl[1].xBase & XBASEMASK) / in_scale)
<< " " << (gna_pwl[1].yBase) / out_scale
<< " " << pwl[0].m
<< "\n";
for (uint32_t i = 1; i < pwl_size - 1; ++i) {
s = gna_slope(pwl[i].m, in_scale, out_scale);
gna_pwl[i + 1].xBase = (static_cast<int32_t> (in_scale * pwl[i].alpha)) & XBASEMASK;
gna_pwl[i + 1].yBase = FLOAT_TO_INT16(pwl[i].beta * out_scale);
gna_pwl[i + 1].slope = FLOAT_TO_INT16(s.slope * s.slope_scale);
gna_pwl[i + 1].xBase = gna_pwl[i + 1].xBase | s.slope_scale_index;
gnalog() << (pwl[i].alpha)
<< " " << pwl[i].beta
<< " " << pwl[i].m
<< "\n";
}
// insert extra segment for xvalues > u_bound
gna_pwl[n_segments - 1].xBase =
((uint32_t)(in_scale * (y_max / out_scale - pwl[pwl_size - 2].b) / pwl[pwl_size - 2].m)) & XBASEMASK;
gna_pwl[n_segments - 1].yBase = y_max;
gna_pwl[n_segments - 1].slope = 0;
gnalog() << (gna_pwl[n_segments - 1].xBase / in_scale)
<< " " << 1.0
<< " " << 0.0
<< "\n";
break; break;
} }
break; break;
} }
default: default:
gnalog() << "Unexpected function activation!\n";
THROW_GNA_EXCEPTION << "Unexpected function activation!" << fun; THROW_GNA_EXCEPTION << "Unexpected function activation!" << fun;
} }
insert_extra_pwl_segments(gna_pwl, y_min, y_max); insert_extra_pwl_segments(gna_pwl, y_min, y_max);

2
inference-engine/src/gna_plugin/backend/make_pwl.hpp
View File

@ -7,7 +7,7 @@
#include <vector> #include <vector>
#include "runtime/pwl.h" #include "runtime/pwl.h"
void make_gna_pwl(const DnnActivation fun, void make_gna_pwl(const DnnActivation& fun,
const std::vector<pwl_t>& pwl, const std::vector<pwl_t>& pwl,
const double l_bound, const double l_bound,
const double u_bound, const double u_bound,

167
inference-engine/tests/functional/plugin/gna/scale_factors_tests/eltwise_act_fq.cpp Normal file
View File

@ -0,0 +1,167 @@
// Copyright (C) 2021 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//
#include <vector>
#include <memory>
#include <tuple>
#include <vector>
#include <string>
#include <ie_core.hpp>
#include "common_test_utils/common_utils.hpp"
#include "functional_test_utils/plugin_cache.hpp"
#include "shared_test_classes/base/layer_test_utils.hpp"
#include "functional_test_utils/blob_utils.hpp"
#include "ngraph_functions/utils/ngraph_helpers.hpp"
#include "ngraph_functions/builders.hpp"
#include "ngraph_functions/pass/convert_prc.hpp"
static std::map<ngraph::helpers::ActivationTypes, std::string> activationNames = {
{ngraph::helpers::ActivationTypes::Sigmoid, "Sigmoid"},
{ngraph::helpers::ActivationTypes::Tanh, "Tanh"},
{ngraph::helpers::ActivationTypes::Relu, "Relu"},
{ngraph::helpers::ActivationTypes::Exp, "Exp"},
{ngraph::helpers::ActivationTypes::Log, "Log"},
{ngraph::helpers::ActivationTypes::Sign, "Sign"},
{ngraph::helpers::ActivationTypes::Abs, "Abs"}
};
typedef std::tuple<
InferenceEngine::Precision, // Network Precision
std::string, // Target Device
std::map<std::string, std::string>, // Configuration
std::pair<float, float>, // Input values
ngraph::helpers::ActivationTypes // Activation type
> eltwiseActFqParams;
namespace LayerTestsDefinitions {
class EltwiseActFqTest : public testing::WithParamInterface<eltwiseActFqParams>,
public LayerTestsUtils::LayerTestsCommon {
public:
static std::string getTestCaseName(testing::TestParamInfo<eltwiseActFqParams> obj) {
InferenceEngine::Precision netPrecision;
std::string targetDevice;
std::map<std::string, std::string> configuration;
std::pair<float, float> inputValues;
ngraph::helpers::ActivationTypes act;
std::tie(netPrecision, targetDevice, configuration, inputValues, act) = obj.param;
std::ostringstream result;
result << "netPRC=" << netPrecision.name() << "_";
result << "targetDevice=" << targetDevice << "_";
for (auto const& configItem : configuration) {
result << "_configItem=" << configItem.first << "_" << configItem.second;
}
result << "_range=(" << inputValues.first << ", " << inputValues.second << ")";
result << "_act=" << activationNames[act];
return result.str();
}
InferenceEngine::Blob::Ptr GenerateInput(const InferenceEngine::InputInfo& info) const override {
InferenceEngine::Blob::Ptr blob = make_blob_with_precision(info.getTensorDesc());
blob->allocate();
auto* rawBlobDataPtr = blob->buffer().as<float*>();
std::vector<float> values = CommonTestUtils::generate_float_numbers(blob->size(), inputDataMin, inputDataMax);
for (size_t i = 0; i < blob->size(); i++) {
rawBlobDataPtr[i] = values[i];
}
return blob;
}
protected:
void SetUp() override {
InferenceEngine::Precision netPrecision;
std::pair<float, float> inputValues;
ngraph::helpers::ActivationTypes act;
std::tie(netPrecision, targetDevice, configuration, inputValues, act) = this->GetParam();
std::tie(inputDataMin, inputDataMax) = inputValues;
if (act == ngraph::helpers::ActivationTypes::Log) {
// clamp not positive values
inputDataMin = 1.0e-3;
}
auto ngPrc = FuncTestUtils::PrecisionUtils::convertIE2nGraphPrc(netPrecision);
const ngraph::Shape shape = {1, 128};
auto params = ngraph::builder::makeParams(ngPrc, {shape});
auto lowNodeIn = ngraph::builder::makeConstant<float>(ngPrc, {1}, { 100 * inputDataMin });
auto highNodeIn = ngraph::builder::makeConstant<float>(ngPrc, {1}, { 100 * inputDataMax });
auto fqIn = std::make_shared<ngraph::opset8::FakeQuantize>(params[0], lowNodeIn, highNodeIn,
lowNodeIn, highNodeIn, levels16);
auto constant = ngraph::builder::makeConstant<float>(ngPrc, shape,
CommonTestUtils::generate_float_numbers(shape[1], inputDataMin, inputDataMax));
auto add = std::make_shared<ngraph::opset8::Add>(fqIn, constant);
auto lowNode = ngraph::builder::makeConstant<float>(ngPrc, {1}, { 2 * inputDataMin });
auto highNode = ngraph::builder::makeConstant<float>(ngPrc, {1}, { 2 * inputDataMax });
auto fq = std::make_shared<ngraph::opset8::FakeQuantize>(add, lowNode, highNode,
lowNode, highNode, levels32);
auto tanh = ngraph::builder::makeActivation(fq, ngPrc, act);
auto lowNodeOut = ngraph::builder::makeConstant<float>(ngPrc, {1}, { std::tanh(2 * inputDataMin) });
auto highNodeOut = ngraph::builder::makeConstant<float>(ngPrc, {1}, { std::tanh(2 * inputDataMax) });
auto fqOut = std::make_shared<ngraph::opset8::FakeQuantize>(tanh, lowNodeOut, highNodeOut,
lowNodeOut, highNodeOut, levels16);
ngraph::ResultVector results{std::make_shared<ngraph::opset8::Result>(fqOut)};
function = std::make_shared<ngraph::Function>(results, params, "TanhFq");
}
float inputDataMax = 1.0;
float inputDataMin = -1.0;
const size_t levels16 = std::numeric_limits<uint16_t>::max();
const size_t levels32 = std::numeric_limits<uint32_t>::max();
// to reproduce the problem with quite big distance between min int and min value from stats
const size_t sf_reducer = 100;
};
TEST_P(EltwiseActFqTest, CompareWithRefImpl) {
Run();
};
const std::vector<InferenceEngine::Precision> netPrecisions = {
InferenceEngine::Precision::FP32,
InferenceEngine::Precision::FP16
};
const std::vector<std::map<std::string, std::string>> configs = {
{
{"GNA_DEVICE_MODE", "GNA_SW_EXACT"},
}
};
const std::vector<std::pair<float, float>> inputValues = {
{-10.0, 10.0},
{-5.0, 5.0},
{-1.0, 1.0},
{-0.04, 0.04}
};
const std::vector<ngraph::helpers::ActivationTypes> activationTypes = {
ngraph::helpers::ActivationTypes::Sigmoid,
ngraph::helpers::ActivationTypes::Tanh,
ngraph::helpers::ActivationTypes::Relu,
ngraph::helpers::ActivationTypes::Exp,
ngraph::helpers::ActivationTypes::Log,
ngraph::helpers::ActivationTypes::Sign,
ngraph::helpers::ActivationTypes::Abs
};
INSTANTIATE_TEST_SUITE_P(smoke_base, EltwiseActFqTest,
::testing::Combine(
::testing::ValuesIn(netPrecisions),
::testing::Values(CommonTestUtils::DEVICE_GNA),
::testing::ValuesIn(configs),
::testing::ValuesIn(inputValues),
::testing::ValuesIn(activationTypes)),
EltwiseActFqTest::getTestCaseName);
} // namespace LayerTestsDefinitions

2
inference-engine/tests/functional/plugin/gna/shared_tests_instances/skip_tests_config.cpp
View File

@ -19,6 +19,8 @@ std::vector<std::string> disabledTestPatterns() {
// TODO: FIX BUG 32210 // TODO: FIX BUG 32210
R"(.*ActivationLayerTest.CompareWithRefs/(Sigmoid|Tanh|Exp|Log).*)", R"(.*ActivationLayerTest.CompareWithRefs/(Sigmoid|Tanh|Exp|Log).*)",
R"(.*ActivationFQSubgraph.*activation=(Exp|Log).*)", R"(.*ActivationFQSubgraph.*activation=(Exp|Log).*)",
// TODO: Issue 68586
R"(.*EltwiseActFqTest.*act=Log.*)",
// TODO: Issue 32542 // TODO: Issue 32542
R"(.*(EltwiseLayerTest).*eltwiseOpType=(Sum|Sub).*opType=SCALAR.*)", R"(.*(EltwiseLayerTest).*eltwiseOpType=(Sum|Sub).*opType=SCALAR.*)",
R"(.*(EltwiseLayerTest).*eltwiseOpType=Prod.*secondaryInputType=PARAMETER.*opType=SCALAR.*)", R"(.*(EltwiseLayerTest).*eltwiseOpType=Prod.*secondaryInputType=PARAMETER.*opType=SCALAR.*)",

6
ngraph/core/include/openvino/runtime/tensor.hpp
View File

@ -79,8 +79,8 @@ public:
* @param type Tensor element type * @param type Tensor element type
* @param shape Tensor shape * @param shape Tensor shape
* @param host_ptr Pointer to pre-allocated host memory * @param host_ptr Pointer to pre-allocated host memory
* @param strides Optional strides parameters in elements. Strides are supposed to be equal to shape if they are not * @param strides Optional strides parameters in bytes. Strides are supposed to be computed automatically based
* set * on shape and element size
*/ */
Tensor(const element::Type type, const Shape& shape, void* host_ptr, const Strides& strides = {}); Tensor(const element::Type type, const Shape& shape, void* host_ptr, const Strides& strides = {});
@ -124,7 +124,7 @@ public:
size_t get_byte_size() const; size_t get_byte_size() const;
/** /**
* @return Tensor's strides in elements * @return Tensor's strides in bytes
*/ */
Strides get_strides() const; Strides get_strides() const;

32
ngraph/core/src/runtime/ov_tensor.cpp
View File

@ -40,15 +40,26 @@ Tensor::Tensor(const element::Type element_type, const Shape& shape, const Alloc
_impl->allocate(); _impl->allocate();
} }
Tensor::Tensor(const element::Type element_type, const Shape& shape, void* host_ptr, const Strides& strides) { Tensor::Tensor(const element::Type element_type, const Shape& shape, void* host_ptr, const Strides& byte_strides) {
ie::SizeVector blk_order(shape.size()); ie::SizeVector blk_order(shape.size());
std::iota(blk_order.begin(), blk_order.end(), 0); std::iota(blk_order.begin(), blk_order.end(), 0);
ie::SizeVector dim_offset(shape.size(), 0); ie::SizeVector dim_offset(shape.size(), 0);
ie::SizeVector blk_strides; ie::SizeVector blk_strides;
if (strides.empty()) { if (byte_strides.empty()) {
blk_strides = ov::row_major_strides(shape); blk_strides = ov::row_major_strides(shape);
} else { } else {
blk_strides.assign(strides.begin(), strides.end()); blk_strides.resize(byte_strides.size());
std::transform(byte_strides.begin(),
byte_strides.end(),
blk_strides.begin(),
[&element_type](size_t byte_stride) {
OPENVINO_ASSERT(byte_stride % element_type.size() == 0,
"Limitation: Stride in bytes ",
byte_stride,
" should be divisible by size of element ",
element_type.size());
return byte_stride / element_type.size();
});
} }
try { try {
@ -93,7 +104,19 @@ Strides Tensor::get_strides() const {
OPENVINO_ASSERT(get_element_type().bitwidth() >= 8, OPENVINO_ASSERT(get_element_type().bitwidth() >= 8,
"Could not get strides for types with bitwidths less then 8 bit. Tensor type: ", "Could not get strides for types with bitwidths less then 8 bit. Tensor type: ",
get_element_type()); get_element_type());
OV_TENSOR_STATEMENT(return _impl->getTensorDesc().getBlockingDesc().getStrides()); OV_TENSOR_STATEMENT({
const auto& element_strides = _impl->getTensorDesc().getBlockingDesc().getStrides();
const size_t elem_size = get_element_type().size();
Strides byte_strides;
byte_strides.resize(element_strides.size());
std::transform(element_strides.begin(),
element_strides.end(),
byte_strides.begin(),
[&elem_size](size_t stride) {
return stride * elem_size;
});
return byte_strides;
});
} }
size_t Tensor::get_size() const { size_t Tensor::get_size() const {
@ -120,6 +143,7 @@ void* Tensor::data(const element::Type element_type) const {
", is not representable as pointer to ", ", is not representable as pointer to ",
element_type); element_type);
} }
// since we don't use byte offsets, we need to explicitly multiply by element_size
auto byte_offset = _impl->getTensorDesc().getBlockingDesc().getOffsetPadding() * get_element_type().size(); auto byte_offset = _impl->getTensorDesc().getBlockingDesc().getOffsetPadding() * get_element_type().size();
OPENVINO_ASSERT((get_element_type().bitwidth() >= 8) || (byte_offset == 0), OPENVINO_ASSERT((get_element_type().bitwidth() >= 8) || (byte_offset == 0),
"ROI access for types with bitwidths less then 8 bit is not implemented. Tensor type: ", "ROI access for types with bitwidths less then 8 bit is not implemented. Tensor type: ",

51
ngraph/test/ov_tensor_test.cpp
View File

@ -18,6 +18,13 @@
using OVTensorTest = ::testing::Test; using OVTensorTest = ::testing::Test;
inline ov::Strides byteStrides(const ov::Strides& strides, const ov::element::Type& type) {
ov::Strides byte_strides(strides.size());
for (size_t i = 0; i < strides.size(); ++i)
byte_strides[i] = strides[i] * type.size();
return byte_strides;
}
TEST_F(OVTensorTest, canCreateTensor) { TEST_F(OVTensorTest, canCreateTensor) {
ov::Shape shape = {4, 3, 2}; ov::Shape shape = {4, 3, 2};
ov::runtime::Tensor t{ov::element::f32, shape}; ov::runtime::Tensor t{ov::element::f32, shape};
@ -27,7 +34,7 @@ TEST_F(OVTensorTest, canCreateTensor) {
ASSERT_EQ(ov::element::f32, t.get_element_type()); ASSERT_EQ(ov::element::f32, t.get_element_type());
ASSERT_EQ(shape, t.get_shape()); ASSERT_EQ(shape, t.get_shape());
ASSERT_NE(shape, t.get_strides()); ASSERT_NE(shape, t.get_strides());
ASSERT_EQ(ov::Strides({6, 2, 1}), t.get_strides()); ASSERT_EQ(byteStrides(ov::Strides({6, 2, 1}), t.get_element_type()), t.get_strides());
ASSERT_EQ(ov::element::f32.size() * totalSize, t.get_byte_size()); ASSERT_EQ(ov::element::f32.size() * totalSize, t.get_byte_size());
ASSERT_THROW(t.data(ov::element::i64), ov::Exception); ASSERT_THROW(t.data(ov::element::i64), ov::Exception);
ASSERT_THROW(t.data<std::int32_t>(), ov::Exception); ASSERT_THROW(t.data<std::int32_t>(), ov::Exception);
@ -72,7 +79,7 @@ TEST_F(OVTensorTest, canAccessExternalData) {
ASSERT_EQ(data, t.data(ov::element::f32)); ASSERT_EQ(data, t.data(ov::element::f32));
ASSERT_EQ(data, ptr); ASSERT_EQ(data, ptr);
ASSERT_THROW(t.data<std::int16_t>(), ov::Exception); ASSERT_THROW(t.data<std::int16_t>(), ov::Exception);
ASSERT_EQ(ov::row_major_strides(shape), t.get_strides()); ASSERT_EQ(byteStrides(ov::row_major_strides(shape), t.get_element_type()), t.get_strides());
ASSERT_EQ(ov::shape_size(shape), t.get_size()); ASSERT_EQ(ov::shape_size(shape), t.get_size());
ASSERT_EQ(ov::shape_size(shape) * ov::element::f32.size(), t.get_byte_size()); ASSERT_EQ(ov::shape_size(shape) * ov::element::f32.size(), t.get_byte_size());
} }
@ -81,11 +88,11 @@ TEST_F(OVTensorTest, canAccessExternalData) {
TEST_F(OVTensorTest, canAccessExternalDataWithStrides) { TEST_F(OVTensorTest, canAccessExternalDataWithStrides) {
ov::Shape shape = {2, 3}; ov::Shape shape = {2, 3};
float data[] = {5.f, 6.f, 7.f, 0.f, 1.f, 42.f, 3.f, 0.f}; float data[] = {5.f, 6.f, 7.f, 0.f, 1.f, 42.f, 3.f, 0.f};
ov::runtime::Tensor t{ov::element::f32, shape, data, {4, 1}}; ov::runtime::Tensor t{ov::element::f32, shape, data, {16, 4}};
ASSERT_EQ(ov::Strides({4, 1}), t.get_strides()); ASSERT_EQ(ov::Strides({16, 4}), t.get_strides());
{ {
ASSERT_EQ((ov::Shape{2, 3}), t.get_shape()); ASSERT_EQ((ov::Shape{2, 3}), t.get_shape());
float* ptr = t.data<float>(); const float* ptr = t.data<const float>();
ASSERT_EQ(ptr[5], 42); ASSERT_EQ(ptr[5], 42);
} }
} }
@ -98,16 +105,23 @@ TEST_F(OVTensorTest, cannotCreateTensorWithExternalNullptr) {
TEST_F(OVTensorTest, cannotCreateTensorWithWrongStrides) { TEST_F(OVTensorTest, cannotCreateTensorWithWrongStrides) {
ov::Shape shape = {2, 3}; ov::Shape shape = {2, 3};
float data[] = {5.f, 6.f, 7.f, 0.f, 1.f, 42.f, 3.f, 0.f}; float data[] = {5.f, 6.f, 7.f, 0.f, 1.f, 42.f, 3.f, 0.f};
const auto el = ov::element::f32;
{ {
// strides.size() != shape.size() // strides.size() != shape.size()
EXPECT_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {6, 3, 1}), ov::Exception); EXPECT_THROW(ov::runtime::Tensor(el, shape, data, byteStrides({6, 3, 1}, el)), ov::Exception);
} }
{ {
// strides values are element-wise >= ov::row_major_strides(shape) values // strides values are element-wise >= ov::row_major_strides(shape) values
EXPECT_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {2, 1}), ov::Exception); EXPECT_THROW(ov::runtime::Tensor(el, shape, data, byteStrides({2, 1}, el)), ov::Exception);
EXPECT_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {3, 0}), ov::Exception); EXPECT_THROW(ov::runtime::Tensor(el, shape, data, byteStrides({3, 0}, el)), ov::Exception);
EXPECT_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {3, 2}), ov::Exception); EXPECT_THROW(ov::runtime::Tensor(el, shape, data, byteStrides({3, 2}, el)), ov::Exception);
EXPECT_NO_THROW(ov::runtime::Tensor(ov::element::f32, shape, data, {6, 2})); EXPECT_NO_THROW(ov::runtime::Tensor(el, shape, data, byteStrides({6, 2}, el)));
}
{
// strides are not divisible by elem_size
EXPECT_THROW(ov::runtime::Tensor(el, shape, data, {7, el.size()}), ov::Exception);
EXPECT_THROW(ov::runtime::Tensor(el, shape, data, {3, 0}), ov::Exception);
EXPECT_THROW(ov::runtime::Tensor(el, shape, data, {el.size(), 3}), ov::Exception);
} }
} }
@ -119,7 +133,7 @@ TEST_F(OVTensorTest, saveDimsAndSizeAfterMove) {
ASSERT_EQ(shape, new_tensor.get_shape()); ASSERT_EQ(shape, new_tensor.get_shape());
ASSERT_EQ(ov::element::f32, new_tensor.get_element_type()); ASSERT_EQ(ov::element::f32, new_tensor.get_element_type());
ASSERT_EQ(ov::row_major_strides(shape), new_tensor.get_strides()); ASSERT_EQ(byteStrides(ov::row_major_strides(shape), new_tensor.get_element_type()), new_tensor.get_strides());
ASSERT_THROW(t.get_size(), ov::Exception); ASSERT_THROW(t.get_size(), ov::Exception);
ASSERT_THROW(t.get_element_type(), ov::Exception); ASSERT_THROW(t.get_element_type(), ov::Exception);
@ -141,7 +155,7 @@ TEST_F(OVTensorTest, canSetShape) {
ASSERT_EQ(t.get_shape(), origShape); ASSERT_EQ(t.get_shape(), origShape);
ASSERT_NO_THROW(t.set_shape({4, 5, 6})); ASSERT_NO_THROW(t.set_shape({4, 5, 6}));
ASSERT_EQ(newShape, t.get_shape()); ASSERT_EQ(newShape, t.get_shape());
ASSERT_EQ(ov::row_major_strides(newShape), t.get_strides()); ASSERT_EQ(byteStrides(ov::row_major_strides(newShape), t.get_element_type()), t.get_strides());
ASSERT_NE(orig_data, t.data()); ASSERT_NE(orig_data, t.data());
// check that setShape for copy changes original Tensor // check that setShape for copy changes original Tensor
@ -180,7 +194,7 @@ TEST_F(OVTensorTest, makeRangeRoiTensor) {
ASSERT_EQ(roi_tensor.data<int32_t>() - t.data<int32_t>(), ref_offset_elems); ASSERT_EQ(roi_tensor.data<int32_t>() - t.data<int32_t>(), ref_offset_elems);
ASSERT_EQ(reinterpret_cast<uint8_t*>(roi_tensor.data()) - reinterpret_cast<uint8_t*>(t.data()), ref_offset_bytes); ASSERT_EQ(reinterpret_cast<uint8_t*>(roi_tensor.data()) - reinterpret_cast<uint8_t*>(t.data()), ref_offset_bytes);
ASSERT_EQ(roi_tensor.get_strides(), t.get_strides()); ASSERT_EQ(roi_tensor.get_strides(), t.get_strides());
ASSERT_EQ(ref_strides, roi_tensor.get_strides()); ASSERT_EQ(byteStrides(ref_strides, roi_tensor.get_element_type()), roi_tensor.get_strides());
ASSERT_EQ(roi_tensor.get_element_type(), t.get_element_type()); ASSERT_EQ(roi_tensor.get_element_type(), t.get_element_type());
} }
@ -218,14 +232,15 @@ TEST_F(OVTensorTest, readRangeRoiBlob) {
ov::runtime::Tensor roi_tensor{t, {0, 0, 2, 4}, {1, 3, 4, 8}}; ov::runtime::Tensor roi_tensor{t, {0, 0, 2, 4}, {1, 3, 4, 8}};
ASSERT_NE(false, static_cast<bool>(roi_tensor)); ASSERT_NE(false, static_cast<bool>(roi_tensor));
{ {
auto roi = roi_tensor.data<int32_t>(); const std::uint8_t* roi = reinterpret_cast<const std::uint8_t*>(roi_tensor.data());
ASSERT_NE(nullptr, roi); ASSERT_NE(nullptr, roi);
auto strides = roi_tensor.get_strides(); auto strides = roi_tensor.get_strides();
for (auto&& c : ngraph::CoordinateTransformBasic{roi_tensor.get_shape()}) { for (auto&& c : ngraph::CoordinateTransformBasic{roi_tensor.get_shape()}) {
auto actual = roi[c[3] * strides[3] + c[2] * strides[2] + c[1] * strides[1] + c[0] * strides[0]]; auto actual_addr = roi + c[3] * strides[3] + c[2] * strides[2] + c[1] * strides[1] + c[0] * strides[0];
auto expected = t.data<int32_t>()[(c[3] + 4) * strides[3] + (c[2] + 2) * strides[2] + auto expected_addr = t.data<int32_t>() + ((c[3] + 4) * strides[3] + (c[2] + 2) * strides[2] +
(c[1] + 0) * strides[1] + (c[0] + 0) * strides[0]]; (c[1] + 0) * strides[1] + (c[0] + 0) * strides[0]) /
ASSERT_EQ(expected, actual) << c; t.get_element_type().size();
ASSERT_EQ(actual_addr, reinterpret_cast<const std::uint8_t*>(expected_addr));
} }
} }
} }

8
runtime/bindings/python/src/pyopenvino/core/common.cpp
View File

@ -73,14 +73,6 @@ const std::map<py::str, ov::element::Type> dtype_to_ov_type = {
{"bool", ov::element::boolean}, {"bool", ov::element::boolean},
}; };
ov::Strides to_numpy_strides(const ov::Strides& strides, const ov::element::Type& ov_type) {
ov::Strides numpy_strides(strides.size());
std::transform(strides.begin(), strides.end(), numpy_strides.begin(), [&ov_type](size_t stride) {
return stride * ov_type.size();
});
return numpy_strides;
}
InferenceEngine::Layout get_layout_from_string(const std::string& layout) { InferenceEngine::Layout get_layout_from_string(const std::string& layout) {
return layout_str_to_enum.at(layout); return layout_str_to_enum.at(layout);
} }

2
runtime/bindings/python/src/pyopenvino/core/common.hpp
View File

@ -36,8 +36,6 @@ namespace Common
extern const std::map<ov::element::Type, py::dtype> ov_type_to_dtype; extern const std::map<ov::element::Type, py::dtype> ov_type_to_dtype;
extern const std::map<py::str, ov::element::Type> dtype_to_ov_type; extern const std::map<py::str, ov::element::Type> dtype_to_ov_type;
ov::Strides to_numpy_strides(const ov::Strides& strides, const ov::element::Type& ov_type);
InferenceEngine::Layout get_layout_from_string(const std::string& layout); InferenceEngine::Layout get_layout_from_string(const std::string& layout);
const std::string& get_layout_from_enum(const InferenceEngine::Layout& layout); const std::string& get_layout_from_enum(const InferenceEngine::Layout& layout);

2
runtime/bindings/python/src/pyopenvino/core/tensor.cpp
View File

@ -71,7 +71,7 @@ void regclass_Tensor(py::module m) {
cls.def_property_readonly("data", [](ov::runtime::Tensor& self) { cls.def_property_readonly("data", [](ov::runtime::Tensor& self) {
return py::array(Common::ov_type_to_dtype.at(self.get_element_type()), return py::array(Common::ov_type_to_dtype.at(self.get_element_type()),
self.get_shape(), self.get_shape(),
Common::to_numpy_strides(self.get_strides(), self.get_element_type()), self.get_strides(),
self.data(), self.data(),
py::cast(self)); py::cast(self));
}); });