#0: squash later

tt_eager/tt_dnn/op_library/conv/kernels/conv_bmm_tilize_col_major_out_blocks.cpp

@@ -15,7 +15,6 @@
 #endif
 
 #include "compute_kernel_api/eltwise_unary/sfpu_split_includes.h"
-#include "debug/dprint.h"
 
 #define DEBUG_PRINT 0
 
@@ -47,20 +46,16 @@ inline void tilize_in(
     uint32_t in_num_subblocks,
     uint32_t out_cb_id
 ) {
-    PACK(( DPRINT << "Start tilize" << ENDL()));
     tilize_init_short(in_cb_id, in_block_w);
     for (uint32_t in_subblock = 0; in_subblock < in_num_subblocks; ++in_subblock) {
         for (uint32_t h = 0; h < in_subblock_h; ++h) {
             cb_wait_front(in_cb_id, in_block_w);
-            PACK(( DPRINT << "Going to reserve tilize in0 cb" << in_block_w << ENDL()));
             cb_reserve_back(out_cb_id, in_block_w);
-            PACK(( DPRINT << "Reserved tilize in0 cb" << ENDL()));
             tilize_block(in_cb_id, in_block_w, out_cb_id);
             cb_push_back(out_cb_id, in_block_w);
             cb_pop_front(in_cb_id, in_block_w);
         }
     }
-    PACK(( DPRINT << "Done tilize" << ENDL()));
     tilize_uninit();
 } // tilize_in()
 
@@ -175,9 +170,6 @@ void MAIN {
     #endif
 
     mm_block_init(mm_in0_cb_id, in1_cb_id, out_cb_id);
-    //PACK(( DPRINT << "in0_num_blocks_h=" << in0_num_blocks_h));
-    //PACK(( DPRINT << "in0_num_subblocks=" << in0_num_subblocks));
-    //PACK(( DPRINT << "in0_subblock_h=" << in0_subblock_h));
     // in1 num blocks w is the outer loop. Output blocks are computed in col major order.
     for(uint32_t in1_block_w_i = 0; in1_block_w_i < in1_num_blocks_w; ++in1_block_w_i) {
 
@@ -306,8 +298,7 @@ void MAIN {
                     }
                 }
 
-                cb_pop_front(in0_cb_id, in0_block_num_tiles);
-                PACK(( DPRINT << "Popped from tilize cb " << in0_block_num_tiles << ENDL()));
+                cb_pop_front(mm_in0_cb_id, in0_block_num_tiles);
                 cb_pop_front(in1_cb_id, in1_block_num_tiles);
             } // for in0_num_blocks_w
             if constexpr(matmul_partials_cb == mm_out_cb_id) {

tt_eager/tt_dnn/op_library/conv/kernels/reader_conv_activations_2d_mcast_padded_with_halo_3x3_weights_v2.cpp

@@ -175,7 +175,6 @@ void kernel_main() {
     reader_idx = 0;
     for (uint32_t nbh = 0; nbh < act_num_blocks_h; nbh++) {
         cb_reserve_back(cb_id_act_row_major_bfloat16, act_block_num_tiles);
-        DPRINT << "reserved space for one act block" << ENDL();
         uint32_t l1_write_addr_act = get_write_ptr(cb_id_act_row_major_bfloat16);
 
         constexpr uint32_t stride_h_bytes = (conv_act_size_w+2) * conv_act_c_read_bytes;
@@ -192,10 +191,8 @@ void kernel_main() {
         // noc_async_read_inc_num_issued(num_issued_reads_per_block); // "false" on read
         noc_async_read_barrier();
         cb_push_back(cb_id_act_row_major_bfloat16, act_block_num_tiles);
-        //DPRINT << "pushed one act block" << ENDL();
         // compute tilizes and pops cb_id_act and pushes to tilized_in0_cb_id
         cb_wait_front(tilized_in0_cb_id, act_block_num_tiles);
-        //DPRINT << "Waited for tilized act block" << ENDL();
 
         // Round robin self-mcast and receive tilized act matrix in cb_id_act
         // Compute should function like regular mm
@@ -239,7 +236,6 @@ void kernel_main() {
             }
             cb_push_back(cb_id_act, act_block_num_tiles);
         } // act_w_num_outer
-        cb_pop_front(tilized_in0_cb_id, act_block_num_tiles); // passing
+        cb_pop_front(tilized_in0_cb_id, act_block_num_tiles);
     }
-    DPRINT << "done reader" << ENDL();
 }

tt_eager/tt_dnn/op_library/conv/multi_core_optimized_conv_sharded/optimized_conv_op_sharded_v2.cpp

@@ -552,10 +552,10 @@ operation::ProgramWithCallbacks multi_core_optimized_conv_sharded_v2_(const Tens
     if (fully_buffer_weights) {
         num_weight_cb_tiles *= window_outer;
     } else if (per_core_weight_matrix_width_ntiles < 8) {
-        num_weight_cb_tiles = num_weight_cb_tiles;
+        num_weight_cb_tiles = num_weight_cb_tiles * 2;
     }
     if (conv_act_size_c / conv_act_c_blocks < 256) {
-        num_act_cb_tiles = num_act_cb_tiles; // double buffered
+        num_act_cb_tiles = num_act_cb_tiles * 2; // double buffered
     }
     cout << "here" << endl;
     uint32_t writer_output_block_num_tiles = out_block_h_ntiles * weight_block_w_ntiles;