From 5f5ce3a0c18185df00410bcf30c3a19fea1c3064 Mon Sep 17 00:00:00 2001
From: Yifan Zhao <yifanz16@illinois.edu>
Date: Tue, 5 Nov 2019 19:56:52 -0600
Subject: [PATCH] Add some debug prints
---
.../include/functional/reduce.cuh | 6 +--
.../tensor_runtime/src/img_tensor_runtime.cu | 12 +++---
.../tensor_runtime/src/img_tensor_utils.cpp | 39 ++++++++++---------
3 files changed, 29 insertions(+), 28 deletions(-)
diff --git a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/functional/reduce.cuh b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/functional/reduce.cuh
index bcd58f90bd..51d8c4da3f 100644
--- a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/functional/reduce.cuh
+++ b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/functional/reduce.cuh
@@ -146,7 +146,7 @@ __host__ Tensor *reduceDim(
// Calculate approximation parameters
if (skip_rate != 0.0f)
- INFO("Approximation happening...");
+ INFO("Approximation happening...\n");
size_t approx_row_size = (size_t)((1 - skip_rate) * row_size);
// If # of output entries is small, and row_size is enough for 16 threads,
@@ -154,7 +154,7 @@ __host__ Tensor *reduceDim(
// Remember if reducing dim in parallel, threads must be (16, 32).
if (num_rows < NThreads * MaxNBlocks && row_size >= AlongDimTh * 8) {
DEBUG(
- "Reducing in parallel, row size = %lu, actually using %lu", row_size,
+ "Reducing in parallel, row size = %lu, actually using %lu\n", row_size,
approx_row_size);
size_t grid_x = std::min(MaxBlocksPerDim, ceilDiv(num_irows, 32ul));
size_t grid_y = std::min(
@@ -166,7 +166,7 @@ __host__ Tensor *reduceDim(
num_irows, num_orows, row_size, approx_row_size);
} else {
DEBUG(
- "Reducing sequentially, row size = %lu, actually using %lu", row_size,
+ "Reducing sequentially, row size = %lu, actually using %lu\n", row_size,
approx_row_size);
// Reduce sequentially.
size_t threads = std::min(NThreads, num_irows);
diff --git a/llvm/projects/hpvm-tensor-rt/tensor_runtime/src/img_tensor_runtime.cu b/llvm/projects/hpvm-tensor-rt/tensor_runtime/src/img_tensor_runtime.cu
index 4abcc52c99..633c5a6655 100644
--- a/llvm/projects/hpvm-tensor-rt/tensor_runtime/src/img_tensor_runtime.cu
+++ b/llvm/projects/hpvm-tensor-rt/tensor_runtime/src/img_tensor_runtime.cu
@@ -18,7 +18,7 @@
void *tensorFft(void *input) {
// https://docs.nvidia.com/cuda/cufft/index.html#twod-complex-to-real-transforms
// Tensor checking
- INFO("FFT");
+ INFO("FFT\n");
auto *t_input = (Tensor *)input;
if (t_input->data_type != CUDNN_DATA_FLOAT)
throw std::runtime_error("Only float32 is supported");
@@ -54,7 +54,7 @@ void *tensorFft(void *input) {
}
void *tensorReduce(void *input, size_t axis, void *func) {
- INFO("Reduce");
+ INFO("Reduce\n");
auto *src = (Tensor *)input;
if (axis >= src->dims.num_dims)
throw std::runtime_error("Dimension out of range");
@@ -67,7 +67,7 @@ void *tensorReduce(void *input, size_t axis, void *func) {
void *tensorReductionSamplingReduce(
void *input, size_t axis, void *func, int skip_level) {
- INFO("Reduce with sampling");
+ INFO("Reduce with sampling\n");
auto *src = (Tensor *)input;
if (axis >= src->dims.num_dims)
throw std::runtime_error("Dimension out of range");
@@ -92,19 +92,19 @@ void *tensorProjectiveT(void *input, void *transformation) {
}
void *tensorMap1(void *f, void *i) {
- INFO("Map1");
+ INFO("Map1\n");
auto *src = (Tensor *)i;
return mapGeneral<float, 1>(f, {src});
}
void *tensorMap2(void *f2, void *i1, void *i2) {
- INFO("Map2");
+ INFO("Map2\n");
auto *src1 = (Tensor *)i1, *src2 = (Tensor *)i2;
return mapGeneral<float, 2>(f2, {src1, src2});
}
void *tensorMap3(void *f3, void *i1, void *i2, void *i3) {
- INFO("Map3");
+ INFO("Map3\n");
auto *src1 = (Tensor *)i1, *src2 = (Tensor *)i2, *src3 = (Tensor *)i3;
return mapGeneral<float, 3>(f3, {src1, src2, src3});
}
diff --git a/llvm/projects/hpvm-tensor-rt/tensor_runtime/src/img_tensor_utils.cpp b/llvm/projects/hpvm-tensor-rt/tensor_runtime/src/img_tensor_utils.cpp
index 6d2477ab46..172229d532 100644
--- a/llvm/projects/hpvm-tensor-rt/tensor_runtime/src/img_tensor_utils.cpp
+++ b/llvm/projects/hpvm-tensor-rt/tensor_runtime/src/img_tensor_utils.cpp
@@ -22,7 +22,7 @@ static inline bool isRegFile(const char *path, dirent *dp) {
return false;
struct stat sb {};
if (lstat(path, &sb) == -1) {
- INFO("lstat failed for file %s", path);
+ INFO("lstat failed for file %s\n", path);
return false;
}
mode_t type = sb.st_mode & S_IFMT;
@@ -72,19 +72,19 @@ static inline float *uint8_to_float(const uint8_t *ui, size_t len) {
static Tensor *to_nhwc(Tensor *t) {
if (t->data_format == CUDNN_TENSOR_NHWC) {
- DEBUG("Tensor already in NHWC format, no conversion needed");
+ DEBUG("Tensor already in NHWC format, no conversion needed\n");
return t;
} else if (t->data_format != CUDNN_TENSOR_NCHW) {
throw std::runtime_error(
"Unknown tensor format: " + std::to_string(t->data_format));
} else {
- DEBUG("Converting to NHWC format");
+ DEBUG("Converting to NHWC format\n");
}
size_t *dim_arr = t->dims.dim_sizes;
size_t n = dim_arr[0], c = dim_arr[1], h = dim_arr[2], w = dim_arr[3];
auto *out_tensor =
- (Tensor *)create4DTensor(t->data_type, CUDNN_TENSOR_NHWC, n, c, h, w);
+ (Tensor *)create4DTensor(t->data_type, CUDNN_TENSOR_NHWC, n, h, w, c);
size_t nhwc_offset = 0;
size_t element_size = getTypeSize(t->data_type);
char *out_data = (char *)(out_tensor->host_data),
@@ -104,13 +104,13 @@ static Tensor *to_nhwc(Tensor *t) {
static Tensor *to_nchw(Tensor *t) {
if (t->data_format == CUDNN_TENSOR_NCHW) {
- DEBUG("Tensor already in NCHW format, no conversion needed");
+ DEBUG("Tensor already in NCHW format, no conversion needed\n");
return t;
} else if (t->data_format != CUDNN_TENSOR_NHWC) {
throw std::runtime_error(
"Unknown tensor format: " + std::to_string(t->data_format));
} else {
- DEBUG("Converting to NCHW format");
+ DEBUG("Converting to NCHW format\n");
}
size_t *dim_arr = t->dims.dim_sizes;
size_t n = dim_arr[0], h = dim_arr[1], w = dim_arr[2], c = dim_arr[3];
@@ -134,15 +134,15 @@ static Tensor *to_nchw(Tensor *t) {
}
Tensor *readDataSet(const char *path, size_t n_color) {
- INFO("Loading image dataset from path %s", path);
+ INFO("Loading image dataset from path %s\n", path);
auto *first_image = (Tensor *)loadAsImage(sample_file(path).c_str(), n_color);
std::vector<size_t> sizes = ::sizes(first_image);
delete first_image;
size_t h = sizes[2], w = sizes[3];
size_t count = count_file(path);
- DEBUG("Counted %d images in path.", count);
+ DEBUG("Loading shape: (%lu, %lu, %lu, %lu)\n", count, n_color, h, w);
auto *batch = (Tensor *)create4DTensor(
- CUDNN_DATA_FLOAT, CUDNN_TENSOR_NHWC, count, n_color, h, w);
+ CUDNN_DATA_FLOAT, CUDNN_TENSOR_NHWC, count, h, w, n_color);
size_t n_floats = n_color * h * w;
auto *base_data = (float *)batch->host_data;
auto dirp = opendir(path);
@@ -164,18 +164,18 @@ Tensor *readDataSet(const char *path, size_t n_color) {
(void)closedir(dirp);
auto *nchw_batch = to_nchw(batch);
delete batch;
- DEBUG("Loaded all images.");
+ DEBUG("Loaded all images.\n");
return nchw_batch;
}
void saveDataSet(
const char *path, const char *prefix, Tensor *batch) {
- INFO("Saving image dataset to path %s", path);
- DEBUG("Copying to CPU before printing");
+ INFO("Saving image dataset to path %s\n", path);
+ DEBUG("Copying to CPU before printing\n");
deviceToHostCopy(batch);
Tensor *converted_batch = batch;
if (batch->data_format == CUDNN_TENSOR_NCHW) {
- DEBUG("Copy-converting to NHWC format");
+ DEBUG("Copy-converting to NHWC format\n");
converted_batch = to_nhwc(batch);
}
std::vector<size_t> sizes = ::sizes(converted_batch);
@@ -200,14 +200,15 @@ void saveDataSet(
}
void *loadAsImage(const char *filename, size_t n_color) {
- INFO("Loading image from path=%s", filename);
+ INFO("Loading image from path=%s\n", filename);
int x, y, n; // x = width, y = height, n = # 8-bit components per pixel
uint8_t *data = stbi_load(filename, &x, &y, &n, n_color);
if (data == nullptr)
throw std::runtime_error("Image load failed");
float *converted = uint8_to_float(data, x * y * n);
+ DEBUG("Loading shape: (1, %lu, %lu, %lu)(NHWC)\n", y, x, n_color);
auto *image =
- (Tensor *)create4DTensor(CUDNN_DATA_FLOAT, CUDNN_TENSOR_NHWC, 1, n, y, x);
+ (Tensor *)create4DTensor(CUDNN_DATA_FLOAT, CUDNN_TENSOR_NHWC, 1, y, x, n);
std::memcpy(image->host_data, converted, x * y * n * sizeof(float));
auto *nchw_image = to_nchw(image);
stbi_image_free(data);
@@ -216,11 +217,11 @@ void *loadAsImage(const char *filename, size_t n_color) {
}
void saveToImage(const char *filename, Tensor *tensor) {
- INFO("Saving image data to path=%s", filename);
+ INFO("Saving image data to path=%s\n", filename);
deviceToHostCopy(tensor);
Tensor *converted_tensor = tensor;
if (tensor->data_format == CUDNN_TENSOR_NCHW) {
- DEBUG("Copy-converting to NHWC format");
+ DEBUG("Copy-converting to NHWC format\n");
converted_tensor = to_nhwc(tensor);
}
auto *hdr_data = (float *)converted_tensor->host_data;
@@ -236,7 +237,7 @@ void saveToImage(const char *filename, Tensor *tensor) {
void *createFilterFromData(
int data_type, void *data, size_t w, size_t h, size_t n_chan) {
- DEBUG("Creating filter from data");
+ DEBUG("Creating filter from data\n");
auto *tensor =
(Tensor *)create4DTensor(data_type, CUDNN_TENSOR_NCHW, 1, n_chan, h, w);
char *tensor_data;
@@ -265,7 +266,7 @@ std::vector<float> PSNR(void *gold_ptr, void *approx_ptr) {
size_t batch_dim = dim_sizes[0];
size_t image_size = dim_sizes[1] * dim_sizes[2] * dim_sizes[3];
float image_size_f = image_size;
- DEBUG("batch_dim = %lu, image_size = %lu", batch_dim, image_size);
+ DEBUG("batch_dim = %lu, image_size = %lu\n", batch_dim, image_size);
auto *image_size_tensor = (Tensor *)create4DTensor(
CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 1, 1, 1, 1
);
--
GitLab