Skip to content
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

Jump to bottom

Feature: Circular Padding #638

Draft
wants to merge 3 commits into
base: master
Choose a base branch
from
Draft

Feature: Circular Padding #638

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6e3e76e
Implement Circular Padding and Test Validation
gartia Dec 6, 2023
9395e55
Small fix in test function
gartia Dec 6, 2023
6fddbed
Rename circularpadding->pad-circular & Test
gartia Dec 7, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
3 changes: 3 additions & 0 deletions build.zig
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@ const builtin = @import("builtin");
// zig build run_test-mul-mat2
// zig build run_test-opt
// zig build run_test-vec1
// zig build run_test-pad-circular
// zig build run_test0
// zig build run_test1
// zig build run_test2
Expand Down Expand Up @@ -95,6 +96,7 @@ pub fn build(b: *std.build.Builder) void {
"test1",
"test2",
"test3",
"test-pad-circular",
} else .{
// "test-blas0",
// "test-grad0",
Expand All @@ -110,6 +112,7 @@ pub fn build(b: *std.build.Builder) void {
"test1",
"test2",
"test3",
"test-pad-circular",
};
inline for (tests) |name| {
const exe = b.addExecutable(.{
Expand Down
9 changes: 9 additions & 0 deletions include/ggml/ggml.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -450,6 +450,8 @@ extern "C" {
GGML_OP_CROSS_ENTROPY_LOSS,
GGML_OP_CROSS_ENTROPY_LOSS_BACK,

GGML_OP_PAD_CIRCULAR,

GGML_OP_COUNT,
};

Expand Down Expand Up @@ -1778,6 +1780,13 @@ extern "C" {
struct ggml_tensor * b,
struct ggml_tensor * c);

// Circular Padding

GGML_API struct ggml_tensor * ggml_pad_circular(
struct ggml_context * ctx,
struct ggml_tensor * a,
int padding);

//
// automatic differentiation
//
Expand Down
205 changes: 203 additions & 2 deletions src/ggml.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1648,9 +1648,11 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {

"CROSS_ENTROPY_LOSS",
"CROSS_ENTROPY_LOSS_BACK",

"PAD_CIRCULAR"
};

static_assert(GGML_OP_COUNT == 70, "GGML_OP_COUNT != 70");
static_assert(GGML_OP_COUNT == 71, "GGML_OP_COUNT != 71");

static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
Expand Down Expand Up @@ -1732,9 +1734,11 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {

"cross_entropy_loss(x,y)",
"cross_entropy_loss_back(x,y)",

"pad_circular(x)"
};

static_assert(GGML_OP_COUNT == 70, "GGML_OP_COUNT != 70");
static_assert(GGML_OP_COUNT == 71, "GGML_OP_COUNT != 71");

static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");

Expand Down Expand Up @@ -6303,6 +6307,44 @@ struct ggml_tensor * ggml_cross_entropy_loss_back(
return result;
}

//GGML_PAD_CIRCULAR

static struct ggml_tensor * ggml_pad_circular_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
int padding) {
bool is_node = false;

if (a->grad) {
GGML_ASSERT(false);
is_node = true;
}

int new_height = a->ne[0] + 2 * padding;
int new_width = a->ne[1] + 2 * padding;

struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
new_height,
new_width,
a->ne[2], a->ne[3]);

result->op = GGML_OP_PAD_CIRCULAR;
result->op_params[0] = padding;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
result->src[1] = NULL;

return result;
}

struct ggml_tensor * ggml_pad_circular(
struct ggml_context * ctx,
struct ggml_tensor * a,
int padding) {
return ggml_pad_circular_impl(ctx, a, padding);
}


////////////////////////////////////////////////////////////////////////////////

void ggml_set_param(
Expand Down Expand Up @@ -13834,6 +13876,152 @@ static void ggml_compute_forward_cross_entropy_loss_back(
}
}

// ggml_compute_pad_circular

static void ggml_compute_forward_pad_circular_f16(
const struct ggml_compute_params * params,
const struct ggml_tensor * src,
struct ggml_tensor * dst) {

if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
return;
}
const int padding = dst->op_params[0];
const int ith = params->ith;
const int nth = params->nth;

ggml_fp16_t * src_data = (ggml_fp16_t *) src->data;
ggml_fp16_t * dst_data = (ggml_fp16_t *) dst->data;

const int64_t orig_height = src->ne[0];
const int64_t orig_width = src->ne[1];
const int64_t new_height = orig_height + 2 * padding;
const int64_t new_width = orig_width + 2 * padding;

int64_t total_elements = new_height * new_width;
int64_t start_index = ith * (total_elements / nth);
int64_t end_index = (ith + 1) == nth ? total_elements : (ith + 1) * (total_elements / nth);

for (int64_t idx = start_index; idx < end_index; ++idx) {
int64_t i = idx / new_width;
int64_t j = idx % new_width;

int64_t orig_i = (i - padding + orig_height) % orig_height;
int64_t orig_j = (j - padding + orig_width) % orig_width;

dst_data[i * new_width + j] = src_data[orig_i * orig_width + orig_j];
}
}

static void ggml_compute_forward_pad_circular_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src,
struct ggml_tensor * dst) {

if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
return;
}

const int padding = dst->op_params[0];
const int ith = params->ith;
const int nth = params->nth;

const int64_t orig_height = src->ne[0];
const int64_t orig_width = src->ne[1];
const int64_t new_height = orig_height + 2 * padding;
const int64_t new_width = orig_width + 2 * padding;

float * src_data = (float *) src->data;
float * dst_data = (float *) dst->data;

int64_t total_elements = new_height * new_width;
int64_t start_index = ith * (total_elements / nth);
int64_t end_index = (ith + 1) == nth ? total_elements : (ith + 1) * (total_elements / nth);

for (int64_t idx = start_index; idx < end_index; ++idx) {
int64_t i = idx / new_width;
int64_t j = idx % new_width;

int64_t orig_i = (i - padding + orig_height) % orig_height;
int64_t orig_j = (j - padding + orig_width) % orig_width;

dst_data[i * new_width + j] = src_data[orig_i * orig_width + orig_j];
}
}

// TODO: Make sure this works i am not familiar with quantization.
static void ggml_compute_forward_pad_circular_q_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src,
struct ggml_tensor * dst) {

if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
return;
}

const int padding = dst->op_params[0];
const int ith = params->ith;
const int nth = params->nth;

const int64_t orig_height = src->ne[0];
const int64_t orig_width = src->ne[1];
const int64_t new_height = orig_height + 2 * padding;
const int64_t new_width = orig_width + 2 * padding;

float * src_data = (float *) src->data;
float * dst_data = (float *) dst->data;

int64_t total_elements = new_height * new_width;
int64_t start_index = ith * (total_elements / nth);
int64_t end_index = (ith + 1) == nth ? total_elements : (ith + 1) * (total_elements / nth);

for (int64_t idx = start_index; idx < end_index; ++idx) {
int64_t i = idx / new_width;
int64_t j = idx % new_width;

int64_t orig_i = (i - padding + orig_height) % orig_height;
int64_t orig_j = (j - padding + orig_width) % orig_width;

dst_data[i * new_width + j] = src_data[orig_i * orig_width + orig_j];
}

}

static void ggml_compute_forward_pad_circular(
const struct ggml_compute_params * params,
const struct ggml_tensor * src,
struct ggml_tensor * dst) {

int padding = dst->op_params[0];

switch (src->type) {
case GGML_TYPE_F32:
ggml_compute_forward_pad_circular_f32(params, src, dst);
break;
case GGML_TYPE_F16:
ggml_compute_forward_pad_circular_f16(params, src, dst);
break;
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
{
ggml_compute_forward_pad_circular_q_f32(params, src,dst);
} break;
default:
{
GGML_ASSERT(false);
} break;
}
}


/////////////////////////////////

static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
Expand Down Expand Up @@ -14155,6 +14343,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
ggml_compute_forward_cross_entropy_loss_back(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
}
break;
case GGML_OP_PAD_CIRCULAR:
{
ggml_compute_forward_pad_circular(params, tensor->src[0], tensor);
}
break;
case GGML_OP_NONE:
{
// nop
Expand Down Expand Up @@ -15220,6 +15413,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
{
GGML_ASSERT(false); // not supported
} break;
case GGML_OP_PAD_CIRCULAR:
{
GGML_ASSERT(false); // TODO: not implemented
} break;
case GGML_OP_NONE:
{
// nop
Expand Down Expand Up @@ -15848,6 +16045,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
{
n_tasks = n_threads;
} break;
case GGML_OP_PAD_CIRCULAR:
{
n_tasks=n_threads;
};
case GGML_OP_NONE:
{
n_tasks = 1;
Expand Down
9 changes: 9 additions & 0 deletions tests/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -404,3 +404,12 @@ add_executable(${TEST_TARGET} ${TEST_TARGET}.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml)
add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}>)
set_property(TEST ${TEST_TARGET} PROPERTY ENVIRONMENT "LLVM_PROFILE_FILE=${TEST_TARGET}.profraw")

#
# test-pad-circular

set(TEST_TARGET test-pad-circular)
add_executable(${TEST_TARGET} ${TEST_TARGET}.cpp)
target_link_libraries(${TEST_TARGET} PRIVATE ggml)
add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}>)
set_property(TEST ${TEST_TARGET} PROPERTY ENVIRONMENT "LLVM_PROFILE_FILE=${TEST_TARGET}.profraw")