adding tensor regression test.
This commit is contained in:
@@ -3,11 +3,7 @@
|
||||
|
||||
#define KERNEL_ARG_DEV_MEM_ADDR 0x7ffff000
|
||||
|
||||
#define FP_ENABLE
|
||||
|
||||
#ifdef FP_ENABLE
|
||||
#define TYPE float
|
||||
#else
|
||||
#ifndef TYPE
|
||||
#define TYPE int
|
||||
#endif
|
||||
|
||||
@@ -17,4 +13,4 @@ typedef struct {
|
||||
uint64_t dst_addr;
|
||||
} kernel_arg_t;
|
||||
|
||||
#endif
|
||||
#endif
|
||||
|
||||
@@ -5,14 +5,14 @@
|
||||
|
||||
void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
|
||||
uint32_t num_points = arg->num_points;
|
||||
TYPE* src_ptr = (TYPE*)arg->src_addr;
|
||||
TYPE* dst_ptr = (TYPE*)arg->dst_addr;
|
||||
auto src_ptr = (TYPE*)arg->src_addr;
|
||||
auto dst_ptr = (TYPE*)arg->dst_addr;
|
||||
|
||||
TYPE ref_value = src_ptr[task_id];
|
||||
auto ref_value = src_ptr[task_id];
|
||||
|
||||
uint32_t pos = 0;
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
TYPE cur_value = src_ptr[i];
|
||||
auto cur_value = src_ptr[i];
|
||||
pos += (cur_value < ref_value) || ((cur_value == ref_value) && (i < task_id));
|
||||
}
|
||||
dst_ptr[pos] = ref_value;
|
||||
|
||||
@@ -66,8 +66,8 @@ void gen_input_data(uint32_t num_points) {
|
||||
src_data.resize(num_points);
|
||||
|
||||
for (uint32_t i = 0; i < num_points; ++i) {
|
||||
float r = static_cast<float>(std::rand()) / RAND_MAX;
|
||||
TYPE value = r * num_points;
|
||||
auto r = static_cast<float>(std::rand()) / RAND_MAX;
|
||||
auto value = static_cast<TYPE>(r * num_points);
|
||||
src_data[i] = value;
|
||||
std::cout << std::dec << i << ": value=" << value << std::endl;
|
||||
}
|
||||
@@ -172,19 +172,16 @@ int main(int argc, char *argv[]) {
|
||||
{
|
||||
std::cout << "allocate staging buffer" << std::endl;
|
||||
uint32_t staging_buf_size = std::max<uint32_t>(src_buf_size,
|
||||
std::max<uint32_t>(dst_buf_size,
|
||||
sizeof(kernel_arg_t)));
|
||||
std::max<uint32_t>(dst_buf_size,
|
||||
sizeof(kernel_arg_t)));
|
||||
staging_buf.resize(staging_buf_size);
|
||||
}
|
||||
|
||||
// upload kernel argument
|
||||
{
|
||||
std::cout << "upload kernel argument" << std::endl;
|
||||
auto buf_ptr = staging_buf.data();
|
||||
memcpy(buf_ptr, &kernel_arg, sizeof(kernel_arg_t));
|
||||
RT_CHECK(vx_copy_to_dev(device, KERNEL_ARG_DEV_MEM_ADDR, staging_buf.data(), sizeof(kernel_arg_t)));
|
||||
}
|
||||
|
||||
std::cout << "upload kernel argument" << std::endl;
|
||||
memcpy(staging_buf.data(), &kernel_arg, sizeof(kernel_arg_t));
|
||||
RT_CHECK(vx_copy_to_dev(device, KERNEL_ARG_DEV_MEM_ADDR, staging_buf.data(), sizeof(kernel_arg_t)));
|
||||
|
||||
// upload source buffer
|
||||
{
|
||||
std::cout << "upload source buffer" << std::endl;
|
||||
|
||||
Reference in New Issue
Block a user