Skip to content

Rectangular kernel conv #32

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
patschmidt2 wants to merge 18 commits into
base: dev
Choose a base branch
from
Open

Rectangular kernel conv #32

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
3069fcd
Added defines for sigmoid and tanh activation
patschmidt2 Dec 15, 2022
2c13e80
Added setting of activation function MSB
patschmidt2 Jan 2, 2023
a9b8ea8
Added dummy versions of VGG16 and SSD300 networks
patschmidt2 Jan 26, 2023
488f779
added dummy version of Detection_Header of FFTRadNet network
Jan 27, 2023
483932f
added quantized weights and bias for Detection_Header; and .py file
Jan 31, 2023
a3d3c46
Added testcase for scaling matmul and transposed conv
patschmidt2 Feb 28, 2023
41f006a
Merge branch 'dev' of https://gitlab.itiv.kit.edu/ai-hardware-acceler…
patschmidt2 Feb 28, 2023
e149d5c
Merge of rectangular conv support from upstream
patschmidt2 Feb 28, 2023
4723b99
WIP of convolution with rectangular kernel
patschmidt2 Mar 5, 2023
2f38ccb
Fixed flatten_weights function
patschmidt2 Mar 5, 2023
f448bb6
Fixes for conv with rectangular conv
patschmidt2 Mar 9, 2023
bb27213
Fixed convolutions with weight transforms and added testcases for them
patschmidt2 Mar 13, 2023
a06f852
Added separate parameters for kernel dims
patschmidt2 Mar 14, 2023
47e6670
Added support for rectangular depthwise conv
patschmidt2 Mar 21, 2023
b825cc4
Merge branch 'ucb-bar:dev' into rectangular_kernel_conv
patschmidt2 Mar 21, 2023
053a886
Fixed conv with weight transform and rectangular kernel
patschmidt2 Mar 22, 2023
cda773f
Merge branch 'rectangular_kernel_conv' of https://github.com/patschmi…
patschmidt2 Mar 22, 2023
ae1581a
Removed FFTRadNet for pull request
patschmidt2 Mar 22, 2023
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
4 changes: 4 additions & 0 deletions bareMetalC/Makefile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,10 @@ tests = \
mvin_scale \
conv \
conv_rect \
conv_rect_kernel \
conv_rect_kernel_trans_weight_0132 \
conv_rect_kernel_trans_weight_1203 \
conv_rect_kernel_and_input \
conv_with_pool \
conv_with_rot180 \
conv_with_kernel_dilation \
Expand Down
50 changes: 26 additions & 24 deletions bareMetalC/conv.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,8 @@
#define IN_COL_DIM 224
#define IN_CHANNELS 3
#define OUT_CHANNELS 32
#define KERNEL_DIM 3
#define KERNEL_ROW_DIM 3
#define KERNEL_COL_DIM 3
#define PADDING 1
#define STRIDE 2

Expand All @@ -38,36 +39,37 @@
#endif

#define BATCH_SIZE 2
#define KERNEL_DIM 3
#define KERNEL_ROW_DIM 3
#define KERNEL_COL_DIM 3
#define PADDING 1
#define STRIDE 2

#endif

#define NO_BIAS false

#define OUT_ROW_DIM ((IN_ROW_DIM + 2*PADDING - KERNEL_DIM) / STRIDE + 1)
#define OUT_COL_DIM ((IN_COL_DIM + 2*PADDING - KERNEL_DIM) / STRIDE + 1)
#define PATCH_SIZE (KERNEL_DIM * KERNEL_DIM * IN_CHANNELS)
#define OUT_ROW_DIM ((IN_ROW_DIM + 2*PADDING - KERNEL_ROW_DIM) / STRIDE + 1)
#define OUT_COL_DIM ((IN_COL_DIM + 2*PADDING - KERNEL_COL_DIM) / STRIDE + 1)
#define PATCH_SIZE (KERNEL_ROW_DIM * KERNEL_COL_DIM * IN_CHANNELS)
#define N_PATCHES (BATCH_SIZE * OUT_ROW_DIM * OUT_COL_DIM)

void conv(int batch_size, int in_channels,
int in_row_dim, int in_col_dim,
int out_channels, int kernel_dim,
int out_channels, int kernel_row_dim, int kernel_col_dim,
int out_row_dim, int out_col_dim,
int stride, int padding,
elem_t input[batch_size][in_row_dim][in_col_dim][in_channels],
elem_t weights[out_channels][kernel_dim][kernel_dim][in_channels],
elem_t weights[out_channels][kernel_row_dim][kernel_col_dim][in_channels],
acc_t bias[out_channels],
elem_t output[batch_size][out_row_dim][out_col_dim][out_channels]) {

#ifdef GEMMINI_ASSERTIONS
if (out_row_dim != (in_row_dim + 2 * padding - kernel_dim) / stride + 1) {
if (out_row_dim != (in_row_dim + 2 * padding - kernel_row_dim) / stride + 1) {
printf("conv out_row_dim is not correct\n");
exit(1);
}

if (out_col_dim != (in_col_dim + 2 * padding - kernel_dim) / stride + 1) {
if (out_col_dim != (in_col_dim + 2 * padding - kernel_col_dim) / stride + 1) {
printf("conv out_col_dim is not correct\n");
exit(1);
}
Expand All @@ -79,8 +81,8 @@ void conv(int batch_size, int in_channels,
for (int och = 0; och < out_channels; och++) {
acc_t result = bias[och];

for (int krow = 0; krow < kernel_dim; krow++) {
for (int kcol = 0; kcol < kernel_dim; kcol++) {
for (int krow = 0; krow < kernel_row_dim; krow++) {
for (int kcol = 0; kcol < kernel_col_dim; kcol++) {
for (int kch = 0; kch < in_channels; kch++) {
int irow = orow * stride + krow - padding;
int icol = ocol * stride + kcol - padding;
Expand All @@ -106,18 +108,18 @@ void conv(int batch_size, int in_channels,
}
}

void flatten_weights(int out_channels, int kernel_dim, int in_channels,
void flatten_weights(int out_channels, int kernel_row_dim, int kernel_col_dim, int in_channels,
int patch_size,
elem_t weights[out_channels][kernel_dim][kernel_dim][in_channels],
elem_t weights[out_channels][kernel_row_dim][kernel_col_dim][in_channels],
elem_t weights_mat[patch_size][out_channels]) {

assert(patch_size == kernel_dim * kernel_dim * in_channels);
assert(patch_size == kernel_row_dim * kernel_col_dim * in_channels);

for (int outc = 0; outc < out_channels; outc++) {
for (int krow = 0; krow < kernel_dim; krow++) {
for (int kcol = 0; kcol < kernel_dim; kcol++) {
for (int krow = 0; krow < kernel_row_dim; krow++) {
for (int kcol = 0; kcol < kernel_col_dim; kcol++) {
for (int inc = 0; inc < in_channels; inc++) {
int wmatrow = krow * kernel_dim * in_channels +
int wmatrow = krow * kernel_col_dim * in_channels +
kcol * in_channels +
inc;

Expand Down Expand Up @@ -182,7 +184,7 @@ int main() {
printf("Output dimensions (rows by columns): %u by %u\n\n", OUT_ROW_DIM, OUT_COL_DIM);

static elem_t input[BATCH_SIZE][IN_ROW_DIM][IN_COL_DIM][IN_CHANNELS];
static elem_t weights[OUT_CHANNELS][KERNEL_DIM][KERNEL_DIM][IN_CHANNELS];
static elem_t weights[OUT_CHANNELS][KERNEL_ROW_DIM][KERNEL_COL_DIM][IN_CHANNELS];
static acc_t bias[OUT_CHANNELS];
static elem_t output[BATCH_SIZE][OUT_ROW_DIM][OUT_COL_DIM][OUT_CHANNELS];

Expand All @@ -203,7 +205,7 @@ int main() {
#ifndef FAST
conv(BATCH_SIZE, IN_CHANNELS,
IN_ROW_DIM, IN_COL_DIM,
OUT_CHANNELS, KERNEL_DIM,
OUT_CHANNELS, KERNEL_ROW_DIM, KERNEL_COL_DIM,
OUT_ROW_DIM, OUT_COL_DIM,
STRIDE, PADDING,
input,
Expand All @@ -218,7 +220,7 @@ int main() {
static elem_t output_mat[N_PATCHES][OUT_CHANNELS];

printf("Flatten weights...\n");
flatten_weights(OUT_CHANNELS, KERNEL_DIM, IN_CHANNELS,
flatten_weights(OUT_CHANNELS, KERNEL_ROW_DIM, KERNEL_COL_DIM, IN_CHANNELS,
PATCH_SIZE,
weights,
weights_mat);
Expand All @@ -228,7 +230,7 @@ int main() {
tiled_conv_auto(
BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM, IN_CHANNELS,
OUT_CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
STRIDE, 1, 1, PADDING, KERNEL_DIM,
STRIDE, 1, 1, PADDING, KERNEL_ROW_DIM, KERNEL_COL_DIM,
false, false, false, false, false,

(elem_t*)input,
Expand Down Expand Up @@ -271,9 +273,9 @@ int main() {
printf("weights:\n");
for (int och = 0; och < OUT_CHANNELS; och++) {
printf("[");
for (int wrow = 0; wrow < KERNEL_DIM; wrow++) {
for (int wrow = 0; wrow < KERNEL_ROW_DIM; wrow++) {
printf("[");
for (int wcol = 0; wcol < KERNEL_DIM; wcol++) {
for (int wcol = 0; wcol < KERNEL_COL_DIM; wcol++) {
printf("[");
for (int ich = 0; ich < IN_CHANNELS; ich++) {
printf("%d,", weights[och][wrow][wcol][ich]);
Expand All @@ -287,7 +289,7 @@ int main() {
printf("\b\n\n");

printf("weights_mat:\n");
for (int wrow = 0; wrow < KERNEL_DIM * KERNEL_DIM * IN_CHANNELS; wrow++) {
for (int wrow = 0; wrow < KERNEL_ROW_DIM * KERNEL_COL_DIM * IN_CHANNELS; wrow++) {
printf("[");
for (int wcol = 0; wcol < OUT_CHANNELS; wcol++) {
printf("%d,", weights_mat[wrow][wcol]);
Expand Down
48 changes: 34 additions & 14 deletions bareMetalC/conv_dw.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,13 +8,23 @@
#endif
#include "include/gemmini_testutils.h"

#define RECT_KERNEL

#ifndef BAREMETAL

#define BATCH_SIZE 3
#define IN_ROW_DIM 112
#define IN_COL_DIM 112
#define CHANNELS 17
#define KERNEL_DIM 3

#ifdef RECT_KERNEL
#define KERNEL_ROW_DIM 4
#define KERNEL_COL_DIM 2
#else
#define KERNEL_ROW_DIM 3
#define KERNEL_COL_DIM 3
#endif

#define PADDING 1
#define STRIDE 2

Expand All @@ -30,21 +40,31 @@

#define IN_ROW_DIM 17
#define IN_COL_DIM 17
#define CHANNELS 15
#define CHANNELS 2

#endif

#define BATCH_SIZE 3
#define KERNEL_DIM 3

#ifdef RECT_KERNEL
#define KERNEL_ROW_DIM 4
#define KERNEL_COL_DIM 2
#else
#define KERNEL_ROW_DIM 3
#define KERNEL_COL_DIM 3
#endif

#define PADDING 1
#define STRIDE 2

#endif



#define NO_BIAS false

#define OUT_ROW_DIM ((IN_ROW_DIM + 2*PADDING - KERNEL_DIM) / STRIDE + 1)
#define OUT_COL_DIM ((IN_COL_DIM + 2*PADDING - KERNEL_DIM) / STRIDE + 1)
#define OUT_ROW_DIM ((IN_ROW_DIM + 2*PADDING - KERNEL_ROW_DIM) / STRIDE + 1)
#define OUT_COL_DIM ((IN_COL_DIM + 2*PADDING - KERNEL_COL_DIM) / STRIDE + 1)

bool vec_is_equal(elem_t * a, elem_t * b, int len) {
for (int i = 0; i < len; i++)
Expand All @@ -57,11 +77,11 @@ void init_random(elem_t * buf, int len) {
elem_t i = 0;
for (elem_t * ptr = buf; ptr < buf + len; ptr++) {
// *ptr = (rand() % 32) - 16;
#ifdef FAST
//#ifdef FAST
*ptr = 1;
#else
*ptr = (rand() % 5) - 2;
#endif
//#else
// *ptr = (rand() % 5) - 2;
//#endif
}
}

Expand Down Expand Up @@ -99,7 +119,7 @@ int main() {
printf("Output dimensions: %u by %u\n\n", OUT_ROW_DIM, OUT_COL_DIM);

static elem_t input[BATCH_SIZE][IN_ROW_DIM][IN_COL_DIM][CHANNELS];
static elem_t weights[CHANNELS][KERNEL_DIM][KERNEL_DIM];
static elem_t weights[CHANNELS][KERNEL_ROW_DIM][KERNEL_COL_DIM];
static acc_t bias[CHANNELS];
static elem_t output[BATCH_SIZE][OUT_ROW_DIM][OUT_COL_DIM][CHANNELS];

Expand All @@ -120,7 +140,7 @@ int main() {
#ifndef FAST
tiled_conv_dw_auto(BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM,
CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
STRIDE, PADDING, KERNEL_DIM,
STRIDE, PADDING, KERNEL_ROW_DIM, KERNEL_COL_DIM,

(elem_t*)input,
(elem_t*)weights,
Expand All @@ -140,7 +160,7 @@ int main() {
uint64_t start_gemmini = read_cycles();
tiled_conv_dw_auto(BATCH_SIZE, IN_ROW_DIM, IN_COL_DIM,
CHANNELS, OUT_ROW_DIM, OUT_COL_DIM,
STRIDE, PADDING, KERNEL_DIM,
STRIDE, PADDING, KERNEL_ROW_DIM, KERNEL_COL_DIM,

(elem_t*)input,
(elem_t*)weights,
Expand Down Expand Up @@ -180,9 +200,9 @@ int main() {
printf("weights:\n");
for (int och = 0; och < CHANNELS; och++) {
printf("[");
for (int wrow = 0; wrow < KERNEL_DIM; wrow++) {
for (int wrow = 0; wrow < KERNEL_ROW_DIM; wrow++) {
printf("[");
for (int wcol = 0; wcol < KERNEL_DIM; wcol++) {
for (int wcol = 0; wcol < KERNEL_COL_DIM; wcol++) {
printf("%d,", weights[och][wrow][wcol]);
}
printf("\b],\n");
Expand Down
2 changes: 1 addition & 1 deletion bareMetalC/conv_dw_perf.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,7 +85,7 @@ int main (int argc, char * argv[]) {
uint64_t start_gemmini = read_cycles();

tiled_conv_dw_auto(BATCH_SIZE, IN_DIM, IN_DIM, CHANNELS, OUT_DIM, OUT_DIM,
STRIDE, PADDING, KERNEL_DIM,
STRIDE, PADDING, KERNEL_DIM, KERNEL_DIM,

(elem_t*)input,
(elem_t*)weights,
Expand Down
Loading