void NoiseLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const Dtype* bottom_data = bottom[0]->cpu_data();
Dtype* top_data = top[0]->mutable_cpu_data();
Dtype* rand_vec_data = rand_vec_.mutable_cpu_data();
const int count = bottom[0]->count();
// create gaussian noise and add to top, in-place/ or not the same
if (sigma_> 0) {
caffe_rng_gaussian(count, Dtype(0), sigma_, rand_vec_data);
} else if (bottom[0] == top[0]) {
} else {
caffe_set(count, Dtype(0), rand_vec_data);
}
// use copy not add
caffe_add(count, rand_vec_data, bottom_data, top_data);
}
void DropoutLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const Dtype* bottom_data = bottom[0]->cpu_data();
Dtype* top_data = top[0]->mutable_cpu_data();
Dtype* mask = rand_vec_->mutable_cpu_data();
const int count = rand_vec_->count();
if (this->phase_ == TRAIN) {
switch (drop_type_){
case DropoutParameter_DropType_BERNOULLI:
{
// Create random numbers
caffe_rng_bernoulli(count, Dtype(1. - threshold_), mask);
break;
}
case DropoutParameter_DropType_GAUSSIAN:
{
caffe_rng_gaussian(count, Dtype(mu_), Dtype(sigma_), mask);
// clip to be in [0,1]
for (int i = 0; i < rand_vec_->count(); ++i){
Dtype m = mask[i];
mask[i] = m > 1 ? 1 : (m < 0 ? 0 : m);
}
break;
}
case DropoutParameter_DropType_UNIFORM:
{
caffe_rng_uniform(count, Dtype(a_), Dtype(b_), mask);
break;
}
}
if (drop_batch_){
Dtype drop = mask[0];
caffe_copy(top[0]->count(), bottom_data, top_data);
caffe_scal(top[0]->count(), Dtype(scale_ * drop), top_data);
}
else{
vector<Blob<Dtype>*> scale_bottom(2, NULL);
scale_bottom[0] = bottom[0];
scale_bottom[1] = rand_vec_;
const vector<Blob<Dtype>*> scale_top(1, top[0]);
scale_layer_->Forward(scale_bottom, scale_top);
caffe_scal(top[0]->count(), scale_, top_data);
}
} else {
caffe_copy(bottom[0]->count(), bottom_data, top_data);
}
}
void RngGaussianFill(const Dtype mu, const Dtype sigma, void* cpu_data) {
Dtype* rng_data = static_cast<Dtype*>(cpu_data);
caffe_rng_gaussian(sample_size_, mu, sigma, rng_data);
}
void Device::rng_gaussian_double(const uint_tp n, const double mu,
const double sigma, vptr<double> r) {
vector<double> random(n); // NOLINT
caffe_rng_gaussian(n, mu, sigma, &random[0]);
this->memcpy(sizeof(double) * n, &random[0], vptr<void>(r));
}
void Device::rng_gaussian_half(const uint_tp n, const half_fp mu,
const half_fp sigma, vptr<half_fp> r) {
vector<half_fp> random(n); // NOLINT
caffe_rng_gaussian(n, mu, sigma, &random[0]);
this->memcpy(sizeof(half_fp) * n, &random[0], vptr<void>(r));
}
void Device::rng_gaussian_float(const uint_tp n, const float mu,
const float sigma, vptr<float> r) {
vector<float> random(n); // NOLINT
caffe_rng_gaussian(n, mu, sigma, &random[0]);
this->memcpy(sizeof(float) * n, &random[0], vptr<void>(r));
}
TYPED_TEST(LibDNNBlasTest, TestGemvComparativeCPUGPU) {
Device *dc = Caffe::GetDefaultDevice();
TypeParam eps = 0.0;
if (std::is_same<TypeParam, half_fp>::value) {
eps = EPS_HALF;
}
if (std::is_same<TypeParam, float>::value) {
eps = EPS_FLOAT;
}
if (std::is_same<TypeParam, double>::value) {
eps = EPS_DOUBLE;
}
std::random_device rdev;
std::mt19937 rngen(rdev());
std::uniform_int_distribution<int_tp> dimsRand(1, 256);
std::uniform_int_distribution<int_tp> boolRand(0, 1);
std::uniform_int_distribution<int_tp> factorRand(-25, 25);
for (int_tp testIdx = 0; testIdx < 25; ++testIdx) {
int_tp M = dimsRand(rngen);
int_tp N = dimsRand(rngen);
CBLAS_TRANSPOSE trans_A = boolRand(rngen) ? CblasTrans : CblasNoTrans;
bool has_alpha = boolRand(rngen);
TypeParam alpha_val = factorRand(rngen) / 100.0;
bool has_beta = boolRand(rngen);
TypeParam beta_val = factorRand(rngen) / 100.0;
vector<int_tp> A_shape(4, 1);
vector<int_tp> x_shape(4, 1);
vector<int_tp> y_shape(4, 1);
A_shape[2] = M;
A_shape[3] = N;
x_shape[3] = trans_A == CblasTrans ? M : N;
y_shape[3] = trans_A == CblasTrans ? N : M;
Blob<TypeParam> A(A_shape, Caffe::GetDefaultDevice());
Blob<TypeParam> x(x_shape, Caffe::GetDefaultDevice());
Blob<TypeParam> y_GPU(y_shape, Caffe::GetDefaultDevice());
Blob<TypeParam> y_CPU(y_shape, Caffe::GetDefaultDevice());
caffe_rng_gaussian(M * N, (TypeParam)0.0, (TypeParam)0.25,
A.mutable_cpu_data());
caffe_rng_gaussian(trans_A == CblasTrans ? M : N, (TypeParam)0.0,
(TypeParam)0.25, x.mutable_cpu_data());
caffe_rng_gaussian(trans_A == CblasTrans ? N : M, (TypeParam)0.0,
(TypeParam)0.25, y_CPU.mutable_cpu_data());
caffe_copy(trans_A == CblasTrans ? N : M, y_CPU.cpu_data(),
y_GPU.mutable_cpu_data());
std::cout << "==== Test Case " << testIdx << " ====" << std::endl;
std::cout << "M: " << M << " N: " << N << std::endl;
std::cout << "alpha: " << (has_alpha ? alpha_val : (TypeParam)1.0) << " "
<< "beta: " << (has_beta ? beta_val : (TypeParam)0.0)
<< std::endl;
std::cout << "trans A: " << (trans_A == CblasTrans) << std::endl;
dc->GetLibDNNBlas<TypeParam, TypeParam>()->gemv(
trans_A,
M, N,
has_alpha ? alpha_val: (TypeParam)1.,
A.gpu_data(), x.gpu_data(),
has_beta ? beta_val : (TypeParam)0.,
y_GPU.mutable_gpu_data());
caffe_gemv<TypeParam>(
trans_A,
M, N,
has_alpha ? alpha_val: (TypeParam)1.,
A.cpu_data(), x.cpu_data(),
has_beta ? beta_val : (TypeParam)0.,
y_CPU.mutable_cpu_data());
for (int_tp i = 0; i < (trans_A == CblasTrans ? N : M); ++i) {
EXPECT_NEAR(y_CPU.cpu_data()[i], y_GPU.cpu_data()[i], eps);
// One error is enough to abort
if (fabs(y_CPU.cpu_data()[i] - y_GPU.cpu_data()[i]) >= eps) {
break;
}
}
}
}
TYPED_TEST(QuantBlasTest, TestAxpbyComparativeFloatQuant) {
typedef typename TypeParam::Dtype Dtype;
// Expect at most 5% error
float percentile_eps = 0.05;
std::random_device rdev;
std::mt19937 rngen(rdev());
// Need to test > 64 dimension
std::uniform_int_distribution<int_tp> dimsRand(1, 256);
std::uniform_int_distribution<int_tp> boolRand(0, 1);
std::uniform_int_distribution<int_tp> factorRand(-25, 25);
std::uniform_real_distribution<float> valRand(-2.0, 2.0);
for (int_tp testIdx = 0; testIdx < 25; ++testIdx) {
int_tp N = dimsRand(rngen);
bool has_alpha = boolRand(rngen);
bool has_beta = has_alpha ? boolRand(rngen) : true;
bool alpha_with_quant = boolRand(rngen) && has_alpha;
bool beta_with_quant = boolRand(rngen) && has_beta;
float alpha_val;
float beta_val;
if (has_alpha) {
alpha_val = alpha_with_quant ? valRand(rngen) : float(1.0);
} else {
alpha_val = 0.0;
}
if (has_beta) {
beta_val = beta_with_quant ? valRand(rngen) : float(1.0);
} else {
beta_val = 0.0;
}
vector<int_tp> x_shape(1, 1);
vector<int_tp> y_shape(1, 1);
x_shape[0] = N;
y_shape[0] = N;
Blob<float> x(x_shape, Caffe::GetDefaultDevice());
Blob<float> y(y_shape, Caffe::GetDefaultDevice());
Blob<float> y_result(y_shape, Caffe::GetDefaultDevice());
Blob<Dtype> x_quant(x_shape, Caffe::GetDefaultDevice());
Blob<Dtype> y_quant(y_shape, Caffe::GetDefaultDevice());
Blob<float> y_unquant(y_shape, Caffe::GetDefaultDevice());
caffe_rng_gaussian(N, (float)0.0, (float)0.5, x.mutable_cpu_data());
caffe_rng_gaussian(N, (float)0.0, (float)0.5, y.mutable_cpu_data());
caffe_copy(N, y.cpu_data(), y_result.mutable_cpu_data());
QuantizerParameter qpm_x;
QuantizerParameter qpm_y;
QuantizerParameter qpm_alpha;
QuantizerParameter qpm_beta;
qpm_x.set_mode(CAFFE_QUANT_OBSERVE);
qpm_y.set_mode(CAFFE_QUANT_OBSERVE);
qpm_alpha.set_mode(CAFFE_QUANT_OBSERVE);
qpm_beta.set_mode(CAFFE_QUANT_OBSERVE);
Quantizer<float, Dtype> xq(qpm_x);
Quantizer<float, Dtype> yq(qpm_y);
Quantizer<float, Dtype> alphaq(qpm_alpha);
Quantizer<float, Dtype> betaq(qpm_beta);
// Normal GEMM
caffe_axpby<float>(N, alpha_val, x.cpu_data(), beta_val,
y_result.mutable_cpu_data());
// Observe all values that will be relevant for quantization
xq.ObserveIn_cpu(N, x.cpu_data());
yq.ObserveIn_cpu(N, y.cpu_data());
yq.ObserveIn_cpu(N, y_result.cpu_data());
alphaq.ObserveIn_cpu(1, &alpha_val);
betaq.ObserveIn_cpu(1, &beta_val);
// Apply observed values to the quantizer
xq.update();
yq.update();
alphaq.update();
betaq.update();
// Quantize A, B and C
xq.Forward_cpu(N, x.cpu_data(), x_quant.mutable_cpu_data());
yq.Forward_cpu(N, y.cpu_data(), y_quant.mutable_cpu_data());
Dtype alpha_val_quant = has_alpha;
Dtype beta_val_quant = has_beta;
// Quantize alpha
if (alpha_with_quant) {
alphaq.Forward_cpu(1, &alpha_val, &alpha_val_quant);
}
// Quantize beta
if (beta_with_quant) {
betaq.Forward_cpu(1, &beta_val, &beta_val_quant);
}
if (Caffe::mode() == Caffe::Brew::CPU) {
// TODO: Not implemented yet
return;
/*caffe_axpby<Dtype>(N, alpha_val_quant, x_quant.cpu_data(),
beta_val_quant, y_quant.mutable_cpu_data(),
alpha_with_quant ? &(alphaq.out_quantizer_values()) : nullptr,
&(xq.out_quantizer_values()),
beta_with_quant ? &(betaq.out_quantizer_values()) : nullptr,
&(yq.out_quantizer_values()));*/
} else {
Caffe::GetDefaultDevice()->template axpby<Dtype>(N,
alpha_val_quant, x_quant.gpu_data(),
beta_val_quant, y_quant.mutable_gpu_data(),
alpha_with_quant ? &(alphaq.out_quantizer_values()) : nullptr,
&(xq.out_quantizer_values()),
beta_with_quant ? &(betaq.out_quantizer_values()) : nullptr,
&(yq.out_quantizer_values()));
}
yq.Backward_cpu(N, y_quant.cpu_data(), y_unquant.mutable_cpu_data());
const QuantizerValues cqv = yq.in_quantizer_values();
float eps = std::max(std::abs(cqv.get_max<float>()),
std::abs(cqv.get_min<float>())) * percentile_eps;
for (int_tp i = 0; i < N; ++i) {
EXPECT_NEAR(y_unquant.cpu_data()[i], y_result.cpu_data()[i], eps);
// One error is enough to abort
if (fabs(y_unquant.cpu_data()[i] - y_result.cpu_data()[i]) >= eps) {
break;
}
}
}
}
TYPED_TEST(QuantBlasTest, TestGemmComparativeFloatQuant) {
typedef typename TypeParam::Dtype Dtype;
// Expect at most 5% error
float percentile_eps = 0.05;
std::random_device rdev;
std::mt19937 rngen(rdev());
// Need to test > 64 dimension
std::uniform_int_distribution<int_tp> dimsRand(1, 256);
std::uniform_int_distribution<int_tp> boolRand(0, 1);
std::uniform_int_distribution<int_tp> factorRand(-25, 25);
std::uniform_real_distribution<float> valRand(-2.0, 2.0);
for (int_tp testIdx = 0; testIdx < 25; ++testIdx) {
int_tp M = dimsRand(rngen);
int_tp N = dimsRand(rngen);
int_tp K = dimsRand(rngen);
CBLAS_TRANSPOSE trans_A = boolRand(rngen) ? CblasTrans : CblasNoTrans;
CBLAS_TRANSPOSE trans_B = boolRand(rngen) ? CblasTrans : CblasNoTrans;
bool has_alpha = boolRand(rngen);
bool has_beta = has_alpha ? boolRand(rngen) : true;
bool alpha_with_quant = boolRand(rngen) && has_alpha;
bool beta_with_quant = boolRand(rngen) && has_beta;
float alpha_val;
float beta_val;
if (has_alpha) {
alpha_val = alpha_with_quant ? valRand(rngen) : float(1.0);
} else {
alpha_val = 0.0;
}
if (has_beta) {
beta_val = beta_with_quant ? valRand(rngen) : float(1.0);
} else {
beta_val = 0.0;
}
vector<int_tp> A_shape(4, 1);
vector<int_tp> B_shape(4, 1);
vector<int_tp> C_shape(4, 1);
A_shape[2] = M;
A_shape[3] = K;
B_shape[2] = K;
B_shape[3] = N;
C_shape[2] = M;
C_shape[3] = N;
Blob<float> A(A_shape, Caffe::GetDefaultDevice());
Blob<float> B(B_shape, Caffe::GetDefaultDevice());
Blob<float> C(C_shape, Caffe::GetDefaultDevice());
Blob<float> C_result(C_shape, Caffe::GetDefaultDevice());
Blob<Dtype> A_quant(A_shape, Caffe::GetDefaultDevice());
Blob<Dtype> B_quant(B_shape, Caffe::GetDefaultDevice());
Blob<Dtype> C_quant(C_shape, Caffe::GetDefaultDevice());
Blob<float> C_unquant(C_shape, Caffe::GetDefaultDevice());
caffe_rng_gaussian(M * K, (float)0.0, (float)0.5,
A.mutable_cpu_data());
caffe_rng_gaussian(K * N, (float)0.0, (float)0.5,
B.mutable_cpu_data());
caffe_rng_gaussian(M * N, (float)0.0, (float)0.5,
C.mutable_cpu_data());
caffe_copy(M * N, C.cpu_data(), C_result.mutable_cpu_data());
QuantizerParameter qpm_a;
QuantizerParameter qpm_b;
QuantizerParameter qpm_c;
QuantizerParameter qpm_alpha;
QuantizerParameter qpm_beta;
qpm_a.set_mode(CAFFE_QUANT_OBSERVE);
qpm_b.set_mode(CAFFE_QUANT_OBSERVE);
qpm_c.set_mode(CAFFE_QUANT_OBSERVE);
qpm_alpha.set_mode(CAFFE_QUANT_OBSERVE);
qpm_beta.set_mode(CAFFE_QUANT_OBSERVE);
Quantizer<float, Dtype> aq(qpm_a);
Quantizer<float, Dtype> bq(qpm_b);
Quantizer<float, Dtype> cq(qpm_c);
Quantizer<float, Dtype> alphaq(qpm_alpha);
Quantizer<float, Dtype> betaq(qpm_beta);
// Normal GEMM
caffe_gemm<float>(
trans_A, trans_B,
M, N, K,
alpha_val,
A.cpu_data(), B.cpu_data(),
beta_val,
C_result.mutable_cpu_data());
// Observe all values that will be relevant for quantization
aq.ObserveIn_cpu(M * K, A.cpu_data());
bq.ObserveIn_cpu(K * N, B.cpu_data());
cq.ObserveIn_cpu(M * N, C.cpu_data());
cq.ObserveIn_cpu(M * N, C_result.cpu_data());
alphaq.ObserveIn_cpu(1, &alpha_val);
betaq.ObserveIn_cpu(1, &beta_val);
// Apply observed values to the quantizer
aq.update();
bq.update();
cq.update();
alphaq.update();
betaq.update();
// Quantize A, B and C
aq.Forward_cpu(M * K, A.cpu_data(), A_quant.mutable_cpu_data());
bq.Forward_cpu(K * N, B.cpu_data(), B_quant.mutable_cpu_data());
cq.Forward_cpu(M * N, C.cpu_data(), C_quant.mutable_cpu_data());
Dtype alpha_val_quant = has_alpha;
Dtype beta_val_quant = has_beta;
// Quantize alpha
if (alpha_with_quant) {
alphaq.Forward_cpu(1, &alpha_val, &alpha_val_quant);
}
// Quantize beta
if (beta_with_quant) {
betaq.Forward_cpu(1, &beta_val, &beta_val_quant);
}
/*
std::cout << "C max:" << cq.in_quantizer_values().max << std::endl;
std::cout << "C min:" << cq.in_quantizer_values().min << std::endl;
std::cout << "C zero:" << cq.in_quantizer_values().zero << std::endl;
std::cout << "C scale:" << cq.in_quantizer_values().scale << std::endl;
std::cout << "C max:" << cq.out_quantizer_values().max << std::endl;
std::cout << "C min:" << cq.out_quantizer_values().min << std::endl;
std::cout << "C zero:" << cq.out_quantizer_values().zero << std::endl;
std::cout << "C scale:" << cq.out_quantizer_values().scale << std::endl;
*/
if (Caffe::mode() == Caffe::Brew::CPU) {
caffe_gemm<Dtype>(
trans_A, trans_B,
M, N, K,
alpha_val_quant,
A_quant.cpu_data(), B_quant.cpu_data(),
beta_val_quant,
C_quant.mutable_cpu_data(),
alpha_with_quant ? &(alphaq.out_quantizer_values()) : nullptr,
&(aq.out_quantizer_values()),
&(bq.out_quantizer_values()),
beta_with_quant ? &(betaq.out_quantizer_values()) : nullptr,
&(cq.out_quantizer_values()));
} else {
Caffe::GetDefaultDevice()->template gemm<Dtype>(trans_A, trans_B,
M, N, K,
alpha_val_quant,
A_quant.gpu_data(), B_quant.gpu_data(),
beta_val_quant,
C_quant.mutable_gpu_data(),
alpha_with_quant ? &(alphaq.out_quantizer_values()) : nullptr,
&(aq.out_quantizer_values()),
&(bq.out_quantizer_values()),
beta_with_quant ? &(betaq.out_quantizer_values()) : nullptr,
&(cq.out_quantizer_values()));
}
cq.Backward_cpu(M * N, C_quant.cpu_data(), C_unquant.mutable_cpu_data());
// print_matrix(A_quant.cpu_data(), M, K);
// print_matrix(B_quant.cpu_data(), K, N);
// print_matrix(C_quant.cpu_data(), M, N);
// print_matrix(C_result.cpu_data(), M, N);
// print_matrix(C_unquant.cpu_data(), M, N);
const QuantizerValues cqv = cq.in_quantizer_values();
float eps = std::max(std::abs(cqv.get_max<float>()),
std::abs(cqv.get_min<float>())) * percentile_eps;
for (int_tp i = 0; i < M * N; ++i) {
EXPECT_NEAR(C_unquant.cpu_data()[i], C_result.cpu_data()[i], eps);
// One error is enough to abort
if (fabs(C_unquant.cpu_data()[i] - C_result.cpu_data()[i]) >= eps) {
break;
}
}
}
}
TYPED_TEST(QuantBlasTest, TestGemvComparativeFloatQuant) {
typedef typename TypeParam::Dtype Dtype;
// Expect at most 5% error
float percentile_eps = 0.05;
std::random_device rdev;
std::mt19937 rngen(rdev());
// Need to test > 64 dimension
std::uniform_int_distribution<int_tp> dimsRand(1, 256);
std::uniform_int_distribution<int_tp> boolRand(0, 1);
std::uniform_int_distribution<int_tp> factorRand(-25, 25);
std::uniform_real_distribution<float> valRand(-2.0, 2.0);
for (int_tp testIdx = 0; testIdx < 25; ++testIdx) {
int_tp M = dimsRand(rngen);
int_tp N = dimsRand(rngen);
CBLAS_TRANSPOSE trans_A = boolRand(rngen) ? CblasTrans : CblasNoTrans;
bool has_alpha = boolRand(rngen);
bool has_beta = has_alpha ? boolRand(rngen) : true;
bool alpha_with_quant = boolRand(rngen) && has_alpha;
bool beta_with_quant = boolRand(rngen) && has_beta;
float alpha_val;
float beta_val;
if (has_alpha) {
alpha_val = alpha_with_quant ? valRand(rngen) : float(1.0);
} else {
alpha_val = 0.0;
}
if (has_beta) {
beta_val = beta_with_quant ? valRand(rngen) : float(1.0);
} else {
beta_val = 0.0;
}
vector<int_tp> A_shape(4, 1);
vector<int_tp> x_shape(4, 1);
vector<int_tp> y_shape(4, 1);
A_shape[2] = M;
A_shape[3] = N;
x_shape[3] = trans_A == CblasTrans ? M : N;
y_shape[3] = trans_A == CblasTrans ? N : M;
Blob<float> A(A_shape, Caffe::GetDefaultDevice());
Blob<float> x(x_shape, Caffe::GetDefaultDevice());
Blob<float> y(y_shape, Caffe::GetDefaultDevice());
Blob<float> y_result(y_shape, Caffe::GetDefaultDevice());
Blob<Dtype> A_quant(A_shape, Caffe::GetDefaultDevice());
Blob<Dtype> x_quant(x_shape, Caffe::GetDefaultDevice());
Blob<Dtype> y_quant(y_shape, Caffe::GetDefaultDevice());
Blob<float> y_unquant(y_shape, Caffe::GetDefaultDevice());
caffe_rng_gaussian(M * N, (float)0.0, (float)0.5,
A.mutable_cpu_data());
caffe_rng_gaussian(trans_A == CblasTrans ? M : N,
(float)0.0, (float)0.5, x.mutable_cpu_data());
caffe_rng_gaussian(trans_A == CblasTrans ? N : M,
(float)0.0, (float)0.5, y.mutable_cpu_data());
caffe_copy(trans_A == CblasTrans ? N : M,
y.cpu_data(), y_result.mutable_cpu_data());
QuantizerParameter qpm_a;
QuantizerParameter qpm_x;
QuantizerParameter qpm_y;
QuantizerParameter qpm_alpha;
QuantizerParameter qpm_beta;
qpm_a.set_mode(CAFFE_QUANT_OBSERVE);
qpm_x.set_mode(CAFFE_QUANT_OBSERVE);
qpm_y.set_mode(CAFFE_QUANT_OBSERVE);
qpm_alpha.set_mode(CAFFE_QUANT_OBSERVE);
qpm_beta.set_mode(CAFFE_QUANT_OBSERVE);
Quantizer<float, Dtype> aq(qpm_a);
Quantizer<float, Dtype> xq(qpm_x);
Quantizer<float, Dtype> yq(qpm_y);
Quantizer<float, Dtype> alphaq(qpm_alpha);
Quantizer<float, Dtype> betaq(qpm_beta);
// Normal GEMM
caffe_gemv<float>(
trans_A,
M, N,
alpha_val,
A.cpu_data(), x.cpu_data(),
beta_val,
y_result.mutable_cpu_data());
// Observe all values that will be relevant for quantization
aq.ObserveIn_cpu(M * N, A.cpu_data());
xq.ObserveIn_cpu(trans_A == CblasTrans ? M : N, x.cpu_data());
yq.ObserveIn_cpu(trans_A == CblasTrans ? N : M, y.cpu_data());
yq.ObserveIn_cpu(trans_A == CblasTrans ? N : M, y_result.cpu_data());
alphaq.ObserveIn_cpu(1, &alpha_val);
betaq.ObserveIn_cpu(1, &beta_val);
// Apply observed values to the quantizer
aq.update();
xq.update();
yq.update();
alphaq.update();
betaq.update();
// Quantize A, B and C
aq.Forward_cpu(M * N, A.cpu_data(), A_quant.mutable_cpu_data());
xq.Forward_cpu(trans_A == CblasTrans ? M : N,
x.cpu_data(), x_quant.mutable_cpu_data());
yq.Forward_cpu(trans_A == CblasTrans ? N : M,
y.cpu_data(), y_quant.mutable_cpu_data());
Dtype alpha_val_quant = has_alpha;
Dtype beta_val_quant = has_beta;
// Quantize alpha
if (alpha_with_quant) {
alphaq.Forward_cpu(1, &alpha_val, &alpha_val_quant);
}
// Quantize beta
if (beta_with_quant) {
betaq.Forward_cpu(1, &beta_val, &beta_val_quant);
}
if (Caffe::mode() == Caffe::Brew::CPU) {
caffe_gemv<Dtype>(trans_A, M, N,
alpha_val_quant,
A_quant.cpu_data(), x_quant.cpu_data(),
beta_val_quant,
y_quant.mutable_cpu_data(),
alpha_with_quant ? &(alphaq.out_quantizer_values()) : nullptr,
&(aq.out_quantizer_values()),
&(xq.out_quantizer_values()),
beta_with_quant ? &(betaq.out_quantizer_values()) : nullptr,
&(yq.out_quantizer_values()));
} else {
Caffe::GetDefaultDevice()->template gemv<Dtype>(trans_A, M, N,
alpha_val_quant,
A_quant.gpu_data(), x_quant.gpu_data(),
beta_val_quant,
y_quant.mutable_gpu_data(),
alpha_with_quant ? &(alphaq.out_quantizer_values()) : nullptr,
&(aq.out_quantizer_values()),
&(xq.out_quantizer_values()),
beta_with_quant ? &(betaq.out_quantizer_values()) : nullptr,
&(yq.out_quantizer_values()));
}
yq.Backward_cpu(trans_A == CblasTrans ? N : M,
y_quant.cpu_data(), y_unquant.mutable_cpu_data());
// print_matrix(A_quant.cpu_data(), M, K);
// print_matrix(B_quant.cpu_data(), K, N);
// print_matrix(C_quant.cpu_data(), M, N);
// print_matrix(C_result.cpu_data(), M, N);
// print_matrix(C_unquant.cpu_data(), M, N);
const QuantizerValues cqv = yq.in_quantizer_values();
float eps = std::max(std::abs(cqv.get_max<float>()),
std::abs(cqv.get_min<float>())) * percentile_eps;
for (int_tp i = 0; i < (trans_A == CblasTrans ? N : M); ++i) {
EXPECT_NEAR(y_unquant.cpu_data()[i], y_result.cpu_data()[i], eps);
// One error is enough to abort
if (fabs(y_unquant.cpu_data()[i] - y_result.cpu_data()[i]) >= eps) {
break;
}
}
}
}