#4079: Remove dprints from op kernels

tenstorrent · Dec 6, 2023 · 1ed70ec · 1ed70ec
1 parent 5e82e3c
commit 1ed70ec
Show file tree

Hide file tree

Showing 48 changed files with 82 additions and 96 deletions.
diff --git a/tt_eager/tt_dnn/kernels/compute/reduce_h.cpp b/tt_eager/tt_dnn/kernels/compute/reduce_h.cpp
@@ -4,8 +4,6 @@
 
 #include <cstdint>
 
-#include "debug/dprint.h"
-
 #include "compute_kernel_api/reduce.h"
 
 namespace NAMESPACE {

diff --git a/tt_eager/tt_dnn/kernels/compute/reduce_hw.cpp b/tt_eager/tt_dnn/kernels/compute/reduce_hw.cpp
@@ -4,8 +4,6 @@
 
 #include <cstdint>
 
-#include "debug/dprint.h"
-
 #include "compute_kernel_api/reduce.h"
 
 namespace NAMESPACE {

diff --git a/tt_eager/tt_dnn/kernels/compute/reduce_w.cpp b/tt_eager/tt_dnn/kernels/compute/reduce_w.cpp
@@ -4,8 +4,6 @@
 
 #include <cstdint>
 
-#include "debug/dprint.h"
-
 #include "compute_kernel_api/reduce.h"
 
 namespace NAMESPACE {

diff --git a/tt_eager/tt_dnn/kernels/dataflow/reader_binary_dtx.cpp b/tt_eager/tt_dnn/kernels/dataflow/reader_binary_dtx.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+
 inline void noc_async_read_from_dram_to_l1(uint32_t dram_addr, uint32_t dram_noc_x, uint32_t dram_noc_y, uint32_t l1_dest_addr, uint32_t read_size) {
     uint64_t src_noc_addr = get_noc_addr(dram_noc_x, dram_noc_y, dram_addr);
     noc_async_read(src_noc_addr, l1_dest_addr, read_size);

diff --git a/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_8bank.cpp b/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_8bank.cpp
@@ -5,7 +5,7 @@
 #include <stdint.h>
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 void kernel_main() {
     // same arg indices as in reader_binary_diff_lenghts for compat

diff --git a/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_8bank_output_tiles_partitioned.cpp b/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_8bank_output_tiles_partitioned.cpp
@@ -5,8 +5,6 @@
 #include <stdint.h>
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
-
 void kernel_main() {
     // same arg indices as in reader_binary_diff_lenghts for compat
     uint32_t src0_addr            = get_arg_val<uint32_t>(0);

diff --git a/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core.cpp b/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core.cpp
@@ -5,7 +5,7 @@
 #include <stdint.h>
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 #ifdef FUSE_BIAS
     #include "tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_bias.hpp"

diff --git a/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_bias.hpp b/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_bias.hpp
@@ -5,7 +5,7 @@
 #include <stdint.h>
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 template <bool bias_in_dram>

diff --git a/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_tilize_untilize.cpp b/tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_tilize_untilize.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 #ifdef FUSE_BIAS
     #include "tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_bias.hpp"

diff --git a/tt_eager/tt_dnn/kernels/dataflow/reader_tm_tile_layout_split_two_chunks.cpp b/tt_eager/tt_dnn/kernels/dataflow/reader_tm_tile_layout_split_two_chunks.cpp
@@ -8,7 +8,7 @@
 
 //#define DEBUG
 #ifdef DEBUG
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 #endif
 
 void kernel_main() {

diff --git a/tt_eager/tt_dnn/kernels/dataflow/reshape_interleaved.cpp b/tt_eager/tt_dnn/kernels/dataflow/reshape_interleaved.cpp
@@ -5,7 +5,7 @@
 #include <stdint.h>
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 using uint32_t = std::uint32_t;
 

diff --git a/tt_eager/tt_dnn/kernels/dataflow/writer_bmm_single_core_tiled.cpp b/tt_eager/tt_dnn/kernels/dataflow/writer_bmm_single_core_tiled.cpp
@@ -3,7 +3,7 @@
 // SPDX-License-Identifier: Apache-2.0
 
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 void kernel_main() {
     // This writer is for output tensor in tile format

diff --git a/tt_eager/tt_dnn/kernels/dataflow/writer_unary_stick_layout_8bank_blocks.cpp b/tt_eager/tt_dnn/kernels/dataflow/writer_unary_stick_layout_8bank_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 template <bool DRAM>
 inline void write_tiles_in_block(uint32_t cb_id_out0,

diff --git a/...op_library/conv/kernels/reader_conv1x1_activations_fast_for_col_major_conv_out_blocks.cpp b/...op_library/conv/kernels/reader_conv1x1_activations_fast_for_col_major_conv_out_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 inline void pad_l1_buffer_with_zeroes(uint32_t l1_addr, uint32_t pad_size_bytes) {
     volatile std::uint32_t* dst = reinterpret_cast<volatile std::uint32_t*>(l1_addr);

diff --git a/tt_eager/tt_dnn/op_library/conv/kernels/reader_conv_activations.cpp b/tt_eager/tt_dnn/op_library/conv/kernels/reader_conv_activations.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 inline void pad_l1_buffer_with_zeroes(uint32_t l1_addr, uint32_t pad_size_bytes) {
     volatile std::uint32_t* dst = reinterpret_cast<volatile std::uint32_t*>(l1_addr);

diff --git a/...op_library/conv/kernels/reader_conv_activations_2d_mcast_padded_with_halo_3x3_weights.cpp b/...op_library/conv/kernels/reader_conv_activations_2d_mcast_padded_with_halo_3x3_weights.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 FORCE_INLINE
 void read_channels(uint32_t& l1_write_addr_act, const uint32_t act_l1_read_addr, const uint32_t reader_channel_idx,

diff --git a/...brary/conv/kernels/reader_conv_activations_act_block_w_equals_channels_X_filter_width.cpp b/...brary/conv/kernels/reader_conv_activations_act_block_w_equals_channels_X_filter_width.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 inline void pad_l1_buffer_with_zeroes(uint32_t l1_addr, uint32_t pad_size_bytes) {
     volatile std::uint32_t* dst = reinterpret_cast<volatile std::uint32_t*>(l1_addr);

diff --git a/tt_eager/tt_dnn/op_library/conv/kernels/reader_conv_activations_fast.cpp b/tt_eager/tt_dnn/op_library/conv/kernels/reader_conv_activations_fast.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 inline void pad_l1_buffer_with_zeroes(uint32_t l1_addr, uint32_t pad_size_bytes) {
     volatile std::uint32_t* dst = reinterpret_cast<volatile std::uint32_t*>(l1_addr);

diff --git a/...nn/op_library/conv/kernels/reader_conv_activations_fast_for_col_major_conv_out_blocks.cpp b/...nn/op_library/conv/kernels/reader_conv_activations_fast_for_col_major_conv_out_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 inline void pad_l1_buffer_with_zeroes(uint32_t l1_addr, uint32_t pad_size_bytes) {
     volatile std::uint32_t* dst = reinterpret_cast<volatile std::uint32_t*>(l1_addr);

diff --git a/tt_eager/tt_dnn/op_library/conv/kernels/reader_conv_activations_fast_resnet50_first_conv.cpp b/tt_eager/tt_dnn/op_library/conv/kernels/reader_conv_activations_fast_resnet50_first_conv.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 void kernel_main() {
     uint32_t i = 0;

diff --git a/...ager/tt_dnn/op_library/conv/kernels/reader_conv_activations_fast_without_conv_padding.cpp b/...ager/tt_dnn/op_library/conv/kernels/reader_conv_activations_fast_without_conv_padding.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 inline void pad_l1_buffer_with_zeroes(uint32_t l1_addr, uint32_t pad_size_bytes) {
     volatile std::uint32_t* dst = reinterpret_cast<volatile std::uint32_t*>(l1_addr);

diff --git a/...r/tt_dnn/op_library/conv/kernels/reader_conv_activations_padded_with_halo_3x3_weights.cpp b/...r/tt_dnn/op_library/conv/kernels/reader_conv_activations_padded_with_halo_3x3_weights.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 void kernel_main() {

diff --git a/...t_dnn/op_library/conv/kernels/writer_and_reader_weights_resnet50_first_conv_tiled_out.cpp b/...t_dnn/op_library/conv/kernels/writer_and_reader_weights_resnet50_first_conv_tiled_out.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 #ifdef FUSE_BIAS
     #include "tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_bias.hpp"

diff --git a/...nn/op_library/conv/kernels/writer_and_reader_weights_resnet50_first_conv_untilize_out.cpp b/...nn/op_library/conv/kernels/writer_and_reader_weights_resnet50_first_conv_untilize_out.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 FORCE_INLINE void read_weight_blocks_inner_h_dim(uint32_t cb_id_weight,
                         uint32_t num_blocks_weight_h,

diff --git a/...y/conv/kernels/writer_tiled_out_1d_mcast_receiver_conv_weights_tiled_col_to_rm_blocks.cpp b/...y/conv/kernels/writer_tiled_out_1d_mcast_receiver_conv_weights_tiled_col_to_rm_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 void kernel_main() {

diff --git a/...ary/conv/kernels/writer_tiled_out_1d_mcast_sender_conv_weights_tiled_col_to_rm_blocks.cpp b/...ary/conv/kernels/writer_tiled_out_1d_mcast_sender_conv_weights_tiled_col_to_rm_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 void kernel_main() {

diff --git a/...y/conv/kernels/writer_tiled_out_2d_mcast_receiver_conv_weights_tiled_col_to_rm_blocks.cpp b/...y/conv/kernels/writer_tiled_out_2d_mcast_receiver_conv_weights_tiled_col_to_rm_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 void kernel_main() {

diff --git a/...ary/conv/kernels/writer_tiled_out_2d_mcast_sender_conv_weights_tiled_col_to_rm_blocks.cpp b/...ary/conv/kernels/writer_tiled_out_2d_mcast_sender_conv_weights_tiled_col_to_rm_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 void kernel_main() {

diff --git a/...rary/conv/kernels/writer_tiled_out_mcast_receiver_conv_weights_tiled_col_to_rm_blocks.cpp b/...rary/conv/kernels/writer_tiled_out_mcast_receiver_conv_weights_tiled_col_to_rm_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 void kernel_main() {

diff --git a/...tiled_out_mcast_receiver_conv_weights_tiled_col_to_rm_blocks_num_blocks_weight_h_eq_1.cpp b/...tiled_out_mcast_receiver_conv_weights_tiled_col_to_rm_blocks_num_blocks_weight_h_eq_1.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 void kernel_main() {

diff --git a/...ibrary/conv/kernels/writer_tiled_out_mcast_sender_conv_weights_tiled_col_to_rm_blocks.cpp b/...ibrary/conv/kernels/writer_tiled_out_mcast_sender_conv_weights_tiled_col_to_rm_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 void kernel_main() {

diff --git a/...r_tiled_out_mcast_sender_conv_weights_tiled_col_to_rm_blocks_num_blocks_weight_h_eq_1.cpp b/...r_tiled_out_mcast_sender_conv_weights_tiled_col_to_rm_blocks_num_blocks_weight_h_eq_1.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 
 void kernel_main() {

diff --git a/tt_eager/tt_dnn/op_library/conv/kernels/writer_tiled_out_reader_conv_weights_tiled.cpp b/tt_eager/tt_dnn/op_library/conv/kernels/writer_tiled_out_reader_conv_weights_tiled.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 #ifdef FUSE_BIAS
     #include "tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_bias.hpp"

diff --git a/...n/op_library/conv/kernels/writer_tiled_out_reader_conv_weights_tiled_col_to_rm_blocks.cpp b/...n/op_library/conv/kernels/writer_tiled_out_reader_conv_weights_tiled_col_to_rm_blocks.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 #ifdef FUSE_BIAS
     #include "tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_bias.hpp"

diff --git a/...s/writer_tiled_out_reader_conv_weights_tiled_col_to_rm_blocks_read_weight_slices_once.cpp b/...s/writer_tiled_out_reader_conv_weights_tiled_col_to_rm_blocks_read_weight_slices_once.cpp
@@ -4,7 +4,7 @@
 
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 #ifdef FUSE_BIAS
     #include "tt_eager/tt_dnn/kernels/dataflow/reader_bmm_single_core_bias.hpp"

diff --git a/...nv/kernels/writer_unary_stick_8bank_blocks_reader_weight_tile_with_pow2_addr_gen_fast.cpp b/...nv/kernels/writer_unary_stick_8bank_blocks_reader_weight_tile_with_pow2_addr_gen_fast.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 FORCE_INLINE void read_weight_blocks_inner_h_dim(uint32_t cb_id_weight,
                         uint32_t num_blocks_weight_h,

diff --git a/...library/conv/kernels/writer_unary_stick_layout_8bank_blocks_reader_weight_tile_layout.cpp b/...library/conv/kernels/writer_unary_stick_layout_8bank_blocks_reader_weight_tile_layout.cpp
@@ -4,7 +4,7 @@
 
 #include <stdint.h>
 #include "dataflow_api.h"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 // TODO: FORCE INLINE
 inline void read_weight_blocks_inner_h_dim(uint32_t cb_id_weight,

diff --git a/tt_eager/tt_dnn/op_library/layernorm/kernels/dataflow/writer_unary_sharded_ln_rm_gb.cpp b/tt_eager/tt_dnn/op_library/layernorm/kernels/dataflow/writer_unary_sharded_ln_rm_gb.cpp
@@ -5,7 +5,7 @@
 #include <stdint.h>
 #include "dataflow_api.h"
 #include "hostdevcommon/common_values.hpp"
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 FORCE_INLINE void generate_bcast_scaler_c() {
     constexpr uint32_t cb_in_4 = tt::CB::c_in4;

diff --git a/tt_eager/tt_dnn/op_library/pool/kernels/dataflow/reader_max_pool_2d_multi_core.cpp b/tt_eager/tt_dnn/op_library/pool/kernels/dataflow/reader_max_pool_2d_multi_core.cpp
@@ -5,7 +5,7 @@
 #include <cstdint>
 #include "dataflow_api.h"
 
-#include "debug/dprint.h"
+// #include "debug/dprint.h"
 
 // SliceRange srr = SliceRange{ .h0 = 0, .h1 = 1, .hs = 8, .w0 = 0, .w1 = 32, .ws = 1 };
 // SliceRange srt = SliceRange{ .h0 = 0, .h1 = 4, .hs = 1, .w0 = 0, .w1 = 8, .ws = 1 };
@@ -127,8 +127,8 @@ void kernel_main() {
 
     // NOTE: batch is folded in
 
-    DPRINT << "NOC coords 0: " << (uint) my_x[0] << "," << (uint) my_y[0] << ENDL();
-    DPRINT << "NOC coords 1: " << (uint) my_x[1] << "," << (uint) my_y[1] << ENDL();
+    // DPRINT << "NOC coords 0: " << (uint) my_x[0] << "," << (uint) my_y[0] << ENDL();
+    // DPRINT << "NOC coords 1: " << (uint) my_x[1] << "," << (uint) my_y[1] << ENDL();
 
     uint32_t core_out_w_i_start = get_arg_val<int32_t>(38);
     uint32_t core_out_h_i_start = get_arg_val<int32_t>(39);