gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fixed bug

Browse Source
This commit is contained in:
Philippe Tillet
2019-07-05 15:07:20 -07:00
parent 88ebdddf3d
commit c666f71fd6
3 changed files with 31 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
examples/python/tensorflow/run.py
View File

@@ -61,30 +61,30 @@ def run_shift():
np.random.seed(2)
a = tf.placeholder(tf.float32, shape=[C, H, W, B])
b = tf.placeholder(tf.float32, shape=[C, F])
hshift_h = np.random.randint(- (R//2), R//2 + 1, size=C, dtype=np.int32)
hshift_w = np.random.randint(- (S//2), R//2 + 1, size=C, dtype=np.int32)
#hshift_h = np.random.randint(- (R//2), R//2 + 1, size=C, dtype=np.int32)
#hshift_w = np.random.randint(- (S//2), R//2 + 1, size=C, dtype=np.int32)
hshift_h = -1*np.ones(C, dtype=np.int32)
hshift_w = -1*np.ones(C, dtype=np.int32)
print(hshift_h)
print(hshift_w)
#hshift_h = np.ones(C, dtype=np.int32)
#hshift_w = np.ones(C, dtype=np.int32)
c = module.shift_conv(a, b, shift_h=tf.make_tensor_proto(hshift_h), shift_w=tf.make_tensor_proto(hshift_w))
c = tf.math.reduce_sum(c)
# Reference
ha = np.random.rand(C, H, W, B)
hb = np.random.rand(C, F)
ha = np.ones((C, H, W, B), dtype=np.float32)
hb = np.ones((C, F), dtype=np.float32)
#ha = np.ones((C, H, W, B), dtype=np.int32)
#hb = np.ones((C, F), dtype=np.int32)
sess = tf.InteractiveSession()
grads = tf.test.compute_gradient([a, b], [(C, H, W, B), (C, F)], c, (C, H, W, B),
grads = tf.test.compute_gradient([a, b], [(C, H, W, B), (C, F)], c, (1,),
extra_feed_dict={a: ha, b: hb})
dx_t, dx_n = grads[0]
dw_t, dw_n = grads[1]
print(dx_t)
print(dx_n)
#print(np.max(dw_t - dw_n))
#print(dw_t - dw_n)
#np.savetxt('diff.dat', dw_t - dw_n, fmt='%2.4f')
#np.savetxt('theoretical.dat', dw_t, fmt='%2.4f')
#np.savetxt('numerical.dat', dw_n, fmt='%2.4f')
print(np.max(dx_t - dx_n))
print(np.max(np.abs(dw_t - dw_n)))
print(np.max(np.abs(dx_t - dx_n)))
np.savetxt('diff.dat', dx_t - dx_n, fmt='%2.4f')
np.savetxt('theoretical.dat', dx_t, fmt='%2.4f')
np.savetxt('numerical.dat', dx_n, fmt='%2.4f')

39
lib/dnn/shift.cpp
View File

@@ -83,12 +83,13 @@ void shift::build_deltas() {
h_deltas_[TK_ + c] = offset(c + TK_) - offset(c);
}
if(ty_ == BPROP){
for(unsigned c = 0; c < C_; c++)
for(unsigned c = 0; c < C_; c++){
h_deltas_[c] = shift_h_[c]*ld_c_[1] + shift_w_[c]*ld_c_[2];
}
}
if(ty_ == WGRAD){
for(unsigned c = 0; c < C_; c++)
h_deltas_[c] = -shift_h_[c]*ld_b_[1] + -shift_w_[c]*ld_b_[2];
h_deltas_[c] = shift_h_[c]*ld_b_[1] + shift_w_[c]*ld_b_[2];
}
}
@@ -202,12 +203,12 @@ if(ty_ == FPROP){
int32 rah[TM] = rahc % AH;
__constant__ int32* pd[TK] = delta + rka;
multiple_of(4) int32 d[TK] = *pd;
int1 maskh[TM] = (rah >= pad_h) && (rah < (AH - pad_h));
int1 maskw[TM] = (raw >= pad_w) && (raw < (AW - pad_w));
int1 mask[TM, TK] = maskh[:, newaxis] && maskw[:, newaxis];
int1 interiorh[TM] = (rah >= pad_h) && (rah < (AH - pad_h));
int1 interiorw[TM] = (raw >= pad_w) && (raw < (AW - pad_w));
int1 interior[TM, TK] = interiorh[:, newaxis] && interiorw[:, newaxis];
int32 inc_true[TM, TK] = d[newaxis, :];
int32 inc_false[TM, TK] = rka[newaxis, :] * lda;
int32 inc[TM, TK] = mask ? inc_true : inc_false;)";
int32 inc[TM, TK] = interior ? inc_true : inc_false;)";
}
if(ty_ == WGRAD){
os << R"(
@@ -225,10 +226,10 @@ if(ty_ == WGRAD){
int32 rbw[TK] = rbwhc % AW;
int32 rbhc[TK] = rbwhc / AW;
int32 rbh[TK] = rbhc % AH;
int1 maskh[TK] = (rbh >= pad_h) && (rbh < (AH - pad_h));
int1 maskw[TK] = (rbw >= pad_w) && (rbw < (AW - pad_w));
int1 mask[TK, TN] = maskh[:, newaxis] && maskw[:, newaxis];
int32 inc[TK, TN] = mask ? shift : 0;
int1 interiorh[TK] = (rbh >= pad_h) && (rbh < (AH - pad_h));
int1 interiorw[TK] = (rbw >= pad_w) && (rbw < (AW - pad_w));
int1 interior[TK, TN] = interiorh[:, newaxis] && interiorw[:, newaxis];
int32 inc[TK, TN] = interior ? shift : 0;
)" << b_ty_ << R"(* shifted_pb[TK, TN] = pb + inc;
)" << b_ty_ << R"( b[TK, TN] = *shifted_pb;)";
}
@@ -247,7 +248,7 @@ if(ty_ == FPROP){
d = *pd;
inc_true = d[newaxis, :];
inc_false = TK * lda;
inc = mask ? inc_true : inc_false;
inc = interior ? inc_true : inc_false;
pa = pa + inc;
@checka a = *pa;)";
}
@@ -264,10 +265,10 @@ if(ty_ == WGRAD){
rbw = rbwhc % AW;
rbhc = rbwhc / AW;
rbh = rbhc % AH;
maskh = (rbh >= pad_h) && (rbh < (AH - pad_h));
maskw = (rbw >= pad_w) && (rbw < (AW - pad_w));
mask = maskh[:, newaxis] && maskw[:, newaxis];
inc = mask ? 0 : shift;
interiorh = (rbh >= pad_h) && (rbh < (AH - pad_h));
interiorw = (rbw >= pad_w) && (rbw < (AW - pad_w));
interior = interiorh[:, newaxis] && interiorw[:, newaxis];
inc = interior ? shift : 0;
shifted_pb = pb + inc;
@checkb b = *shifted_pb;)";
}
@@ -290,12 +291,12 @@ if(ty_ == BPROP){
int32 rcw[TM] = rcwhc % AW;
int32 rchc[TM] = rcwhc / AW;
int32 rch[TM] = rchc % AH;
int1 maskh[TM] = (rch >= pad_h) && (rch < (AH - pad_h));
int1 maskw[TM] = (rcw >= pad_w) && (rcw < (AW - pad_w));
int1 interior[TM, TN] = maskh[:, newaxis] && maskw[:, newaxis];
int1 interiorh[TM] = (rch >= pad_h) && (rch < (AH - pad_h));
int1 interiorw[TM] = (rcw >= pad_w) && (rcw < (AW - pad_w));
int1 interior[TM, TN] = interiorh[:, newaxis] && interiorw[:, newaxis];
__constant__ int32* pd[TN] = delta + ryc;
fp32* shift_pc[TM, TN] = pc + (*pd)[newaxis, :];
pc = interior ? pc : shift_pc;
pc = interior ? shift_pc : pc;
@checkc __atomic_add(pc, C);
)";
}

2
lib/lang/node.cpp
View File

@@ -53,8 +53,6 @@ void node::implicit_cast(ir::builder &builder, ir::value *&lhs, ir::value *&rhs,
ir::type *right_ty = rhs->get_type()->get_scalar_ty();
// One operand is pointer
if(left_ty->is_pointer_ty() || right_ty->is_pointer_ty()){
if(right_ty->is_pointer_ty())
std::swap(lhs, rhs);
is_ptr = true;
}
// One operand is double