#16812: Reordering cbs in reduce_init_delta (#16981)

### Ticket
[Link to Github
Issue](#16812)

### Problem description
In all compute kernel APIs, the ordering of the circular buffer
parameters are input cbs first, then output cbs. However, in the
reduce_init_delta function in the reduce.h kernel, the output cb
parameter comes before those of input cbs. This may cause confusion (and
it has in some files) as the expected ordering is inputs first and
outputs later, following all other kernel APIs.

### What's changed
The parameters were reordered such that the input cbs come before the
output cb. Call chains were updated accordingly.

### Checklist
- [x] [Post commit CI
passes](https://github.com/tenstorrent/tt-metal/actions/runs/12917899606)
- [x] [Blackhole Post
commit](https://github.com/tenstorrent/tt-metal/actions/runs/12917901314)
(if applicable)
- [ ] Model regression CI testing passes (if applicable)
- [ ] Device performance regression CI testing passes (if applicable)
- [ ] **(For models and ops writers)** Full [new
models](https://github.com/tenstorrent/tt-metal/actions/workflows/full-new-models-suite.yaml)
tests passes
- [ ] New/Existing tests provide coverage for changes

tests/tt_metal/tt_metal/test_kernels/compute/layernorm.cpp

@@ -98,7 +98,7 @@ void MAIN {
          */
         ACQ();
         cb_reserve_back(cb_ex, 1 * onetile);
-        reduce_init_delta<false>(cb_ex, cb_x, cb_scaler);
+        reduce_init_delta<false>(cb_x, cb_scaler, cb_ex);
         for (uint32_t wt = 0; wt < Wt; wt += blk) {
             cb_wait_front(cb_x, wt + blk);
             for (uint32_t j = 0; j < blk; j++) {
@@ -154,7 +154,7 @@ void MAIN {
          * TODO(AP): can save space here by reusing CB
          */
         cb_reserve_back(cb_ex2, 1);
-        reduce_init_delta<false>(cb_ex2, cb_xmm2, cb_scaler);
+        reduce_init_delta<false>(cb_xmm2, cb_scaler, cb_ex2);
         ACQ();
         cb_wait_front(cb_xmm2, Wt);
         // cb_wait_front(cb_xmm, Wt);

tests/tt_metal/tt_metal/test_kernels/compute/max_pool.cpp

@@ -77,7 +77,7 @@ inline void reduce_h(
     uint32_t out_cb_id) {
     cb_wait_front(in_cb_id, in_ntiles_hwc * out_nelems);
     cb_reserve_back(out_cb_id, out_ntiles_c * out_nelems);
-    reduce_init_delta<false, PoolType::MAX, ReduceDim::REDUCE_COL>(out_cb_id, in_cb_id, in_scalar_cb_id);
+    reduce_init_delta<false, PoolType::MAX, ReduceDim::REDUCE_COL>(in_cb_id, in_scalar_cb_id, out_cb_id);
     uint32_t base_tile_id = 0;
     for (uint32_t c_i = 0; c_i < in_ntiles_c * out_nelems; ++c_i) {
         // add to accumulator all the in_ntiles_hw in a column of tiles

tests/tt_metal/tt_metal/test_kernels/compute/max_pool_multi_core.cpp

@@ -77,7 +77,7 @@ inline void reduce_h(
     uint32_t out_cb_id) {
     cb_wait_front(in_cb_id, in_ntiles_hwc * out_nelems);
     cb_reserve_back(out_cb_id, out_ntiles_c * out_nelems);
-    reduce_init_delta<false, PoolType::MAX, ReduceDim::REDUCE_COL>(out_cb_id, in_cb_id, in_scalar_cb_id);
+    reduce_init_delta<false, PoolType::MAX, ReduceDim::REDUCE_COL>(in_cb_id, in_scalar_cb_id, out_cb_id);
     uint32_t base_tile_id = 0;
     for (uint32_t c_i = 0; c_i < in_ntiles_c * out_nelems; ++c_i) {
         // add to accumulator all the in_ntiles_hw in a column of tiles

tests/tt_metal/tt_metal/test_kernels/compute/reduce_h.cpp

@@ -48,7 +48,7 @@ void MAIN {
 #ifndef SHORT_INIT
     reduce_init<at_start>(tt::CBIndex::c_0, tt::CBIndex::c_2, tt::CBIndex::c_16);
 #else
-    reduce_init_delta<at_start>(tt::CBIndex::c_16, tt::CBIndex::c_0, tt::CBIndex::c_2);
+    reduce_init_delta<at_start>(tt::CBIndex::c_0, tt::CBIndex::c_2, tt::CBIndex::c_16);
 #endif
 
     cb_wait_front(tt::CBIndex::c_2, 1);  // scaler tile from the reader

tests/tt_metal/tt_metal/test_kernels/compute/reduce_hw.cpp

@@ -48,7 +48,7 @@ void MAIN {
 #ifndef SHORT_INIT
     reduce_init<at_start>(tt::CBIndex::c_0, tt::CBIndex::c_2, tt::CBIndex::c_16);
 #else
-    reduce_init_delta<at_start>(tt::CBIndex::c_16, tt::CBIndex::c_0, tt::CBIndex::c_2);
+    reduce_init_delta<at_start>(tt::CBIndex::c_0, tt::CBIndex::c_2, tt::CBIndex::c_16);
 #endif
 
     cb_wait_front(tt::CBIndex::c_2, 1);  // scaler tile from the reader

tests/tt_metal/tt_metal/test_kernels/compute/reduce_w.cpp

@@ -48,7 +48,7 @@ void MAIN {
 #ifndef SHORT_INIT
     reduce_init<at_start>(tt::CBIndex::c_0, tt::CBIndex::c_2, tt::CBIndex::c_16);
 #else
-    reduce_init_delta<at_start>(tt::CBIndex::c_16, tt::CBIndex::c_0, tt::CBIndex::c_2);
+    reduce_init_delta<at_start>(tt::CBIndex::c_0, tt::CBIndex::c_2, tt::CBIndex::c_16);
 #endif
 
     cb_wait_front(tt::CBIndex::c_2, 1);  // scaler tile from the reader

tests/tt_metal/tt_metal/test_kernels/compute/rmsnorm.cpp

@@ -111,7 +111,7 @@ void MAIN {
          * compute E[(x)^2]
          */
         cb_reserve_back(cb_ex2, 1);
-        reduce_init_delta<false>(cb_ex2, cb_x2, cb_scaler);
+        reduce_init_delta<false>(cb_x2, cb_scaler, cb_ex2);
         ACQ();
         cb_wait_front(cb_x2, Wt);
         // cb_wait_front(cb_xmm, Wt);

tests/tt_metal/tt_metal/test_kernels/compute/softmax.cpp

@@ -125,7 +125,7 @@ void MAIN {
 
         ACQ();
         cb_reserve_back(cb_recipsumexps, onetile);
-        reduce_init_delta<false>(cb_recipsumexps, cb_exps, cb_bcast_scaler);
+        reduce_init_delta<false>(cb_exps, cb_bcast_scaler, cb_recipsumexps);
         for (uint32_t wt = 0; wt < Wt; wt++) {
             cb_wait_front(cb_exps, wt + 1);        // must be a cumulative wait for correctness
             constexpr uint32_t bcast_scaler0 = 0;  // 0th index from bcast_scaler CB

tt_metal/include/compute_kernel_api/reduce.h

@@ -37,7 +37,7 @@ ALWI void reduce_init_short(uint32_t icb, uint32_t icb_scaler, uint32_t ocb) {
 }
 
 template <bool at_start, PoolType reduce_type = REDUCE_OP, ReduceDim reduce_dim = REDUCE_DIM>
-ALWI void reduce_init_delta(uint32_t ocb, uint32_t icb0, uint32_t icb1) {
+ALWI void reduce_init_delta(uint32_t icb0, uint32_t icb1, uint32_t ocb) {
     // FIXME: API Update needed in compute kernel?
     UNPACK((llk_unpack_AB_reduce_init<reduce_dim>(icb0, icb1)));
 

ttnn/cpp/ttnn/deprecated/tt_dnn/kernels/compute/moreh_common.hpp

@@ -138,7 +138,7 @@ ALWI void reduce_init_delta_with_dt(uint32_t ocb = 16, uint32_t icb0 = 0, uint32
 #if defined FP32_DEST_ACC_EN
     reconfig_data_format(icb0, icb1);
 #endif
-    reduce_init_delta<at_start, reduce_type, reduce_dim>(ocb, icb0, icb1);
+    reduce_init_delta<at_start, reduce_type, reduce_dim>(icb0, icb1, ocb);
 }
 
 class ArgFetcher {

ttnn/cpp/ttnn/operations/moreh/moreh_clip_grad_norm/moreh_clip_grad_norm_step1/device/kernels/moreh_clip_grad_norm_step1_kernel.cpp

@@ -134,7 +134,7 @@ void MAIN {
     cb_wait_front(cb_xpowadd, onetile);
     cb_reserve_back(cb_y, onetile);
 
-    reduce_init_delta<false>(cb_y, cb_xpowadd, cb_one);
+    reduce_init_delta<false>(cb_xpowadd, cb_one, cb_y);
     reduce_tile(cb_xpowadd, cb_one, 0, 0, dst0);
     reduce_revert_delta(cb_y);
     tile_regs_commit();

ttnn/cpp/ttnn/operations/moreh/moreh_dot/device/kernels/moreh_dot.cpp

@@ -47,7 +47,7 @@ void MAIN {
         }
 
         cb_wait_front(tt::CBIndex::c_24, onetile);
-        reduce_init_delta<false>(tt::CBIndex::c_16, tt::CBIndex::c_24, tt::CBIndex::c_2);
+        reduce_init_delta<false>(tt::CBIndex::c_24, tt::CBIndex::c_2, tt::CBIndex::c_16);
         reduce_tile(tt::CBIndex::c_24, tt::CBIndex::c_2, 0, 0, 0);
         cb_pop_front(tt::CBIndex::c_24, onetile);
         reduce_revert_delta(tt::CBIndex::c_16);

ttnn/cpp/ttnn/operations/moreh/moreh_linear_backward/device/kernels/moreh_bias_backward_multi_core_h.cpp

@@ -102,7 +102,7 @@ void MAIN {
 #if defined FP32_DEST_ACC_EN
                 reconfig_data_format(cb_reduce, cb_scaler);
 #endif
-                reduce_init_delta<false>(cb_out0, cb_reduce, cb_scaler);
+                reduce_init_delta<false>(cb_reduce, cb_scaler, cb_out0);
                 reduce_tile(cb_reduce, cb_scaler, 0, 0, 0);
                 reduce_revert_delta(cb_out0);
 

ttnn/cpp/ttnn/operations/moreh/moreh_linear_backward/device/kernels/moreh_bias_backward_single_core_hw.cpp

@@ -101,7 +101,7 @@ void MAIN {
 #if defined FP32_DEST_ACC_EN
                 reconfig_data_format(cb_reduce, cb_scaler);
 #endif
-                reduce_init_delta<false>(cb_out0, cb_reduce, cb_scaler);
+                reduce_init_delta<false>(cb_reduce, cb_scaler, cb_out0);
                 reduce_tile((do_mask) ? (cb_intermed0) : (cb_in0), cb_scaler, 0, 0, 0);
                 reduce_revert_delta(cb_out0);
 

ttnn/cpp/ttnn/operations/moreh/moreh_sum/device/moreh_sum_h_impl_kernels/moreh_sum_h.cpp

@@ -47,7 +47,7 @@ void MAIN {
 #if defined FP32_DEST_ACC_EN
                     reconfig_data_format(cb_input, cb_scaler);
 #endif
-                    reduce_init_delta<false>(cb_accum_dst, cb_input, cb_scaler);
+                    reduce_init_delta<false>(cb_input, cb_scaler, cb_accum_dst);
                     reduce_tile(cb_input, cb_scaler, 0, 0, reduce_dst_idx);
                     reduce_revert_delta(cb_accum_dst);
 
@@ -104,7 +104,7 @@ void MAIN {
 #if defined FP32_DEST_ACC_EN
             reconfig_data_format(cb_input, cb_scaler);
 #endif
-            reduce_init_delta<false>(cb_out, cb_input, cb_scaler);
+            reduce_init_delta<false>(cb_input, cb_scaler, cb_out);
             reduce_tile(cb_input, cb_scaler, 0, 0, reduce_dst_idx);
             reduce_revert_delta(cb_out);
             tile_regs_commit();

ttnn/cpp/ttnn/operations/normalization/groupnorm/device/kernels/compute/groupnorm_sharded_v2.cpp

@@ -199,7 +199,7 @@ void MAIN {
 
             // Partial-E[x]
             index_h_offset = 0;
-            reduce_init_delta<false>(cb_ex_partial, cb_x, cb_scaler);
+            reduce_init_delta<false>(cb_x, cb_scaler, cb_ex_partial);
             cb_reserve_back(cb_ex_partial, 1);
             tile_regs_acquire();
             cb_wait_front(cb_scaler, 1);
@@ -219,7 +219,7 @@ void MAIN {
             reduce_revert_delta(cb_ex_partial);
 
             if constexpr (is_mcast_sender and num_cores_per_mcast_group > 1) {
-                reduce_init_delta<false>(cb_ex_global, cb_ex_external, cb_scaler_global);
+                reduce_init_delta<false>(cb_ex_external, cb_scaler_global, cb_ex_global);
                 cb_reserve_back(cb_ex_global, 1);
                 cb_reserve_back(cb_ex, 1);
                 tile_regs_acquire();
@@ -316,7 +316,7 @@ void MAIN {
 
             // Partial-Var(x)
             index_h_offset = 0;
-            reduce_init_delta<false>(cb_ex_partial, cb_xmm, cb_scaler);
+            reduce_init_delta<false>(cb_xmm, cb_scaler, cb_ex_partial);
             cb_reserve_back(cb_ex_partial, 1);
             tile_regs_acquire();
             cb_wait_front(cb_xmm, block_hw);
@@ -337,7 +337,7 @@ void MAIN {
             reduce_revert_delta(cb_ex_partial);
 
             if constexpr (is_mcast_sender and num_cores_per_mcast_group > 1) {
-                reduce_init_delta<false>(cb_ex_global, cb_ex_external, cb_scaler_global);
+                reduce_init_delta<false>(cb_ex_external, cb_scaler_global, cb_ex_global);
                 cb_reserve_back(cb_ex_global, 1);
                 cb_reserve_back(cb_ex, 1);
                 tile_regs_acquire();

ttnn/cpp/ttnn/operations/normalization/layernorm/device/kernels/compute/layernorm.cpp

@@ -122,7 +122,7 @@ void MAIN {
          */
         ACQ();
         cb_reserve_back(cb_ex, onetile);
-        reduce_init_delta<false>(cb_ex, cb_x, cb_scaler);
+        reduce_init_delta<false>(cb_x, cb_scaler, cb_ex);
         for (uint32_t wt = 0; wt < Wt; wt += blk) {
             cb_wait_front(cb_x, wt + blk);
             for (uint32_t j = 0; j < blk; j++) {
@@ -193,7 +193,7 @@ void MAIN {
             reconfig_data_format(cb_xmm2, cb_scaler);
         }
         cb_reserve_back(cb_ex2, 1);
-        reduce_init_delta<false>(cb_ex2, cb_xmm2, cb_scaler);
+        reduce_init_delta<false>(cb_xmm2, cb_scaler, cb_ex2);
         ACQ();
         cb_wait_front(cb_xmm2, Wt);
         // cb_wait_front(cb_xmm, Wt);

ttnn/cpp/ttnn/operations/normalization/layernorm/device/kernels/compute/layernorm_sharded.cpp

@@ -142,7 +142,7 @@ void MAIN {
 #ifndef RMSNORM
     // E[x],
     index_h_offset = 0;
-    reduce_init_delta<false>(cb_ex_partial, cb_in, cb_scaler);
+    reduce_init_delta<false>(cb_in, cb_scaler, cb_ex_partial);
     cb_wait_front(cb_scaler, 1);
     cb_reserve_back(cb_ex_partial, block_h);
     for (uint32_t i = 0; i < block_h; i++) {
@@ -163,7 +163,7 @@ void MAIN {
 
     // global reduce, cb_ex <-- cb_ex_external, cb_ex_partial
     if constexpr (is_allgather_worker) {
-        reduce_init_delta<false>(cb_ex, cb_ex_external, cb_scaler_global);
+        reduce_init_delta<false>(cb_ex_external, cb_scaler_global, cb_ex);
         cb_reserve_back(cb_ex, num_tiles_per_allgather_worker);
 
         for (uint32_t i = 0; i < num_tiles_per_allgather_worker; i++) {
@@ -262,7 +262,7 @@ void MAIN {
     cb_wait_front(cb_scaler, 1);
 #endif
     cb_reserve_back(cb_ex_partial2, block_h);
-    reduce_init_delta<false>(cb_ex_partial2, cb_xmm2, cb_scaler);
+    reduce_init_delta<false>(cb_xmm2, cb_scaler, cb_ex_partial2);
     index_h_offset = 0;
     for (uint32_t i = 0; i < block_h; i++) {
         tile_regs_acquire();
@@ -281,7 +281,7 @@ void MAIN {
 
     // global reduce, cb_ex <-- cb_ex_external, cb_ex_partial
     if constexpr (is_allgather_worker) {
-        reduce_init_delta<false>(cb_ex2, cb_ex_external2, cb_scaler_global);
+        reduce_init_delta<false>(cb_ex_external2, cb_scaler_global, cb_ex2);
         cb_reserve_back(cb_ex2, num_tiles_per_allgather_worker);
 
         for (uint32_t i = 0; i < num_tiles_per_allgather_worker; i++) {

ttnn/cpp/ttnn/operations/normalization/layernorm/device/kernels/compute/layernorm_sharded_post_allgather.cpp

@@ -108,7 +108,7 @@ void MAIN {
 #endif
 
             cb_wait_front(cb_scaler_global, 1);
-            reduce_init_delta<false>(cb_var, cb_stats, cb_scaler_global);
+            reduce_init_delta<false>(cb_stats, cb_scaler_global, cb_var);
             tile_regs_acquire();
             // striding over cb_stats, consisting [E(X), E(X^2)] from all the distributed devices in interleaved order
             for (uint32_t w = 0; w < stats_tiles * num_distributed_blocks; w++) {

ttnn/cpp/ttnn/operations/normalization/layernorm/device/kernels/compute/layernorm_sharded_pre_allgather.cpp

@@ -119,7 +119,7 @@ void MAIN {
 #endif
     // E[x],
     index_h_offset = 0;
-    reduce_init_delta<false>(cb_ex_partial2, cb_in0, cb_scaler);
+    reduce_init_delta<false>(cb_in0, cb_scaler, cb_ex_partial2);
 
     cb_reserve_back(cb_ex_partial2, block_h);
     for (uint32_t i = 0; i < block_h; i++) {
@@ -177,7 +177,7 @@ void MAIN {
 
     cb_reserve_back(cb_ex_partial2, block_h);  // RMS E(x2) #Layernorm //E(x) and E(x^2)
 
-    reduce_init_delta<false>(cb_ex_partial2, cb_x2, cb_scaler);
+    reduce_init_delta<false>(cb_x2, cb_scaler, cb_ex_partial2);
     index_h_offset = 0;
     for (uint32_t i = 0; i < block_h; i++) {
         tile_regs_acquire();
@@ -200,7 +200,7 @@ void MAIN {
         cb_wait_front(cb_scaler_global, 1);
         reconfig_data_format_srca(cb_x2, cb_ex_external2);
         reconfig_data_format_srcb(cb_scaler, cb_scaler_global);
-        reduce_init_delta<false>(cb_reduction_out, cb_ex_external2, cb_scaler_global);
+        reduce_init_delta<false>(cb_ex_external2, cb_scaler_global, cb_reduction_out);
         cb_reserve_back(cb_reduction_out, num_tiles_per_partial_result * num_tiles_per_allgather_worker);
 
         for (uint32_t i = 0; i < num_tiles_per_allgather_worker; i++) {  // loops over height

ttnn/cpp/ttnn/operations/normalization/layernorm_distributed/device/kernels/compute/layernorm_post_allgather.cpp

@@ -88,7 +88,7 @@ void MAIN {
          * cb_stats = [sum(x0**2), sum(x0), sum(x1**2), sum(x1), ...]
          * RMSNorm packs mean(x**2) into cb_var. Layernorm just uses cb_stats_reduced.
          */
-        reduce_init_delta<false>(cb_stats_reduced, cb_stats, cb_reduce);
+        reduce_init_delta<false>(cb_stats, cb_reduce, cb_stats_reduced);
         cb_wait_front(cb_stats, stats_tiles_cols);
         cb_reserve_back(cb_stats_reduced, stats_tile_stride);
 #ifdef RMSNORM

ttnn/cpp/ttnn/operations/normalization/layernorm_distributed/device/kernels/compute/layernorm_pre_allgather.cpp

@@ -68,7 +68,7 @@ void MAIN {
          */
         reconfig_data_format(cb_x2, cb_reduce);
         pack_reconfig_data_format(cb_out);
-        reduce_init_delta<false>(cb_out, cb_x2, cb_reduce);
+        reduce_init_delta<false>(cb_x2, cb_reduce, cb_out);
         cb_wait_front(cb_x2, Wt);
         cb_reserve_back(cb_out, onetile);
         ACQ();
@@ -89,7 +89,7 @@ void MAIN {
          */
         reconfig_data_format(cb_inp, cb_reduce);
         pack_reconfig_data_format(cb_out);
-        reduce_init_delta<false>(cb_out, cb_inp, cb_reduce);
+        reduce_init_delta<false>(cb_inp, cb_reduce, cb_out);
         cb_reserve_back(cb_out, onetile);
         ACQ();
         for (uint32_t wtr = 0; wtr < Wt; wtr++) {

ttnn/cpp/ttnn/operations/normalization/softmax/device/kernels/compute/softmax.cpp

@@ -33,7 +33,7 @@ void calc_numeric_stable(
     reconfig_data_format(cb_in, cb_bcast_scaler);
     cb_reserve_back(cb_max, 1);
     cb_wait_front(cb_bcast_scaler, 1);
-    reduce_init_delta<false, PoolType::MAX, ReduceDim::REDUCE_ROW>(cb_max, cb_in, cb_bcast_scaler);
+    reduce_init_delta<false, PoolType::MAX, ReduceDim::REDUCE_ROW>(cb_in, cb_bcast_scaler, cb_max);
     for (uint32_t wt = 0; wt < Wt; wt++) {
         cb_wait_front(cb_in, wt + 1);
         constexpr uint32_t bcast_scaler0 = 0;
@@ -256,7 +256,7 @@ void MAIN {
 
         ACQ();
         cb_reserve_back(cb_recipsumexps, onetile);
-        reduce_init_delta<false>(cb_recipsumexps, cb_exps, cb_bcast_scaler);
+        reduce_init_delta<false>(cb_exps, cb_bcast_scaler, cb_recipsumexps);
         for (uint32_t wt = 0; wt < Wt; wt++) {
             cb_wait_front(cb_exps, wt + 1);        // must be a cumulative wait for correctness
             constexpr uint32_t bcast_scaler0 = 0;  // 0th index from bcast_scaler CB

ttnn/cpp/ttnn/operations/normalization/softmax/device/kernels/compute/softmax_sharded.cpp

@@ -25,7 +25,7 @@ ALWI void calc_numeric_stable(uint32_t cb_in, uint32_t cb_bcast_scaler, uint32_t
     reconfig_data_format(cb_in, cb_bcast_scaler);
     pack_reconfig_data_format(cb_max);
     cb_reserve_back(cb_max, 1);
-    reduce_init_delta<false, PoolType::MAX, ReduceDim::REDUCE_ROW>(cb_max, cb_in, cb_bcast_scaler);
+    reduce_init_delta<false, PoolType::MAX, ReduceDim::REDUCE_ROW>(cb_in, cb_bcast_scaler, cb_max);
     cb_wait_front(cb_bcast_scaler, 1);
     for (uint32_t w = 0; w < block_w; w++) {
         constexpr uint32_t bcast_scaler0 = 0;
@@ -203,7 +203,7 @@ void MAIN {
 
         // sum(exp(x))
         ACQ();
-        reduce_init_delta<false>(cb_recipsumexps, cb_exps, cb_bcast_scaler);
+        reduce_init_delta<false>(cb_exps, cb_bcast_scaler, cb_recipsumexps);
         cb_wait_front(cb_exps, block_w);
         cb_wait_front(cb_bcast_scaler, 1);
         cb_reserve_back(cb_recipsumexps, 1);