Skip to content

Fixed issues introduced by the new broadcast operations #51

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
diaconuccalin wants to merge 9 commits into
base: main
Choose a base branch
from
Open

Fixed issues introduced by the new broadcast operations #51

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
0fab1bd
Fixed crashes in mhsa fp16 due to new transposition. Test not passing…
diaconuccalin May 13, 2025
11dbbf6
Fixed mhsa fp32 kernel issues
diaconuccalin May 13, 2025
c9deb96
Fixed problems introduced into the residual test
diaconuccalin May 13, 2025
f4c94f4
Fixed issues introduced in rnn kernel
diaconuccalin May 13, 2025
95b8428
Made default test demo model for tiny-vit
diaconuccalin May 13, 2025
31cdf06
Added necessary "USE_BIAS = 0" to the MHSA fp16 test
diaconuccalin May 14, 2025
4e4a7d0
Fixed mhsa fp16 test issues (memory allocation and error check to hexa)
diaconuccalin May 14, 2025
c457faa
Add model shape inferencing
diaconuccalin May 26, 2025
569b08a
Revert "Add model shape inferencing"
diaconuccalin May 26, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
261 changes: 146 additions & 115 deletions lib/sources/pulp_mhsa_fp16.c
View file
Open in desktop

Large diffs are not rendered by default.

179 changes: 92 additions & 87 deletions lib/sources/pulp_mhsa_fp32.c
View file
Open in desktop

Large diffs are not rendered by default.

23 changes: 5 additions & 18 deletions lib/sources/pulp_residual_fp16.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,20 +41,14 @@ void pulp_residualconn_fp16_fw(void *SkipConn_args_fp16) {
return;
}

int dims[] = {out->dim};

struct vect_sum_args_fp16 args_sum;

args_sum.op_1 = skip->data;
args_sum.op_2 = lout->data;
args_sum.dest = out->data;
args_sum.size = out->dim;

args_sum.op_1_dims = dims;
args_sum.op_2_dims = dims;

args_sum.op_1_dims_len = 1;
args_sum.op_2_dims_len = 1;

pi_cl_team_fork(NUM_CORES, array_broadcast_sum_fp16, &args_sum);
pi_cl_team_fork(NUM_CORES, vect_sum_fp16, &args_sum);
}


Expand All @@ -77,21 +71,14 @@ void pulp_sumnode_fp16_bw(void *SkipConn_args_fp16) {
return;
}

int dims[] = {skip->dim};

struct vect_sum_args_fp16 args_sum;

args_sum.op_1 = out->diff;
args_sum.op_2 = skip->diff;
args_sum.dest = skip->diff;
args_sum.size = skip->dim;

args_sum.op_1_dims = dims;
args_sum.op2_dims = dims;

args_sum.op_1_dims_len = 1;
args_sum.op_2_dims_len = 1;

pi_cl_team_fork(NUM_CORES, array_broadcast_sum_fp16, &args_sum);
pi_cl_team_fork(NUM_CORES, vect_sum_fp16, &args_sum);
}
}

Expand Down
11 changes: 6 additions & 5 deletions lib/sources/pulp_rnn_fp32.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -244,13 +244,13 @@ void pulp_rnn_fp32_bw_cl(void *Rnn_args) {

// Calculate gradient for State Weights
// Transpose State
int dims[] = {N, M};
int t_axes[] = {1, 0};
dims[0] = N;
dims[1] = M;

struct transp_args transp_args2;

transp_args2.matrix = hiddState;
transp_args2.transp_matrix = temp;
transp_args2.in_matrix = hiddState;
transp_args2.out_matrix = temp;
transp_args2.dim = dims;
transp_args2.transposed_axes = t_axes;
transp_args2.n_dim = 2;
Expand Down Expand Up @@ -301,7 +301,8 @@ void pulp_rnn_fp32_bw_cl(void *Rnn_args) {

// Calculate the Gradient of the Input
// Transpose Input Weights
dims = {K, M};
dims[0] = K;
dims[1] = M;

struct transp_args transp_args3;

Expand Down
31 changes: 30 additions & 1 deletion tests/test_matmul/utils/GM.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -127,12 +127,41 @@
B.transpose(0, 1)
else:
C = torch.mm(input=A, mat2=B, out=C)
elif (data_type == 'bf16'):
# Matrices to be multiplied
A = torch.Tensor(in_size, mid_size).to(torch.bfloat16)
if transp == '1':
B = torch.Tensor(out_size, mid_size).to(torch.bfloat16)
else:
B = torch.Tensor(mid_size, out_size).to(torch.bfloat16)
C = torch.Tensor(in_size, out_size).to(torch.bfloat16)

A = torch.div(torch.randn(in_size, mid_size), divider).to(torch.bfloat16)
for i in range(A.shape[0]):
for j in range(A.shape[1]):
A[i][j] += (i+j+0.1)/divider

if transp == '1':
B = torch.zeros(out_size, mid_size).to(torch.bfloat16)
else:
B = torch.zeros(mid_size, out_size).to(torch.bfloat16)

for i in range(B.shape[0]):
for j in range(B.shape[1]):
B[i][j] = i*j+0.1

if transp == '1':
C = torch.mm(input=A, mat2=B.transpose(0, 1), out=C)
B.transpose(0, 1)
else:
C = torch.mm(input=A, mat2=B, out=C)

else : # Error message
print('Invalid data type selection!!')
exit()


if data_type == 'bf16':
data_type = 'fp16'

# Print data and create data header file
f = open('net_args.h', "w")
Expand Down
8 changes: 4 additions & 4 deletions tests/test_mhsa_fp16/Makefile
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,10 +1,10 @@
APP = mhsa_fp16

# User settings
IN_H?=196 # Sequence Length
IN_W?=160 # Token Size
N_HEADS?=5
ATT_DIM?=160 #Hidden dimension
IN_H?=20 # Sequence Length
IN_W?=40 # Token Size
N_HEADS?=2
ATT_DIM?=40 #Hidden dimension

IN_CH?=1
OUT_CH?=1
Expand Down
4 changes: 2 additions & 2 deletions tests/test_mhsa_fp16/net.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,8 +34,8 @@
#define Tker_l0 (Tin_l0*Tout_l0)

// Tensor checksum definition
#define CHECK_TOLERANCE 0.001
#define ERROR_TOLERANCE 0.001
#define CHECK_TOLERANCE 0x00000021
#define ERROR_TOLERANCE 0x00000001

// PULP DEFINES
#define STACK_SIZE 4096
Expand Down
Loading