#4686: fix fp32 arg for GS path in llk

tenstorrent · Feb 16, 2024 · a8d1067 · a8d1067
1 parent c5d507a
commit a8d1067
Show file tree

Hide file tree

Showing 4 changed files with 6 additions and 6 deletions.
diff --git a/docs/source/tt_metal/apis/kernel_apis/compute/move_copy_tile.rst b/docs/source/tt_metal/apis/kernel_apis/compute/move_copy_tile.rst
@@ -2,6 +2,6 @@ move_copy_tile
 ==============
 
 
-.. doxygenfunction:: copy_tile_to_dst_init_short_with_dt(uint32_t old_cbid, uint32_t new_cbid, bool transpose = false) {
-.. doxygenfunction:: copy_tile_to_dst_init_short(uint32_t cbid = 0, bool transpose = false)
+.. doxygenfunction:: copy_tile_to_dst_init_short_with_dt(uint32_t old_cbid, uint32_t new_cbid) {
+.. doxygenfunction:: copy_tile_to_dst_init_short()
 .. doxygenfunction:: copy_tile_init()
diff --git a/tt_eager/tt_dnn/op_library/bmm/kernels/compute/bmm_large_block_zm_fused_bias_activation.cpp b/tt_eager/tt_dnn/op_library/bmm/kernels/compute/bmm_large_block_zm_fused_bias_activation.cpp
@@ -274,7 +274,7 @@ void MAIN {
                 PACK(( llk_init_packer_dest_offset_registers<SyncHalf,DstTileFaceLayout::ColMajor,false>()  ));
             #endif
             // reconfigure unpacker df for src B
-            unpack_reconfig_data_format(in0_cb_id, in1_cb_id);
+            unpack_reconfig_data_format(in1_cb_id, in0_cb_id);
         }
         #endif
     }

diff --git a/tt_metal/hw/ckernels/grayskull/metal/llk_api/llk_math_unary_datacopy_api.h b/tt_metal/hw/ckernels/grayskull/metal/llk_api/llk_math_unary_datacopy_api.h
@@ -20,15 +20,15 @@ inline void llk_math_eltwise_unary_datacopy(uint dst_index) {
     _llk_math_eltwise_unary_datacopy_<type, src_b_bcast_type, Dst, is_fp32_dest_acc_en>(dst_index);
 }
 
-template <DataCopyType type, BroadcastType src_b_bcast_type = BroadcastType::NONE, DstSync Dst = DstSync::SyncFull>
+template <DataCopyType type, BroadcastType src_b_bcast_type = BroadcastType::NONE, DstSync Dst = DstSync::SyncFull, bool is_fp32_dest_acc_en = false>
 inline void llk_math_eltwise_unary_datacopy_block(uint start_dst_index, uint ntiles) {
 
     for (uint32_t dst_index = start_dst_index; dst_index < start_dst_index + ntiles; dst_index++) {
         llk_math_eltwise_unary_datacopy<type, src_b_bcast_type, Dst>(dst_index);
     }
 }
 
-template <DataCopyType type, BroadcastType src_b_bcast_type = BroadcastType::NONE>
+template <DataCopyType type, BroadcastType src_b_bcast_type = BroadcastType::NONE, bool is_fp32_dest_acc_en = false>
 // within_face_16x16_transpose is used by unpacker, math does not transpose
 inline void llk_math_eltwise_unary_datacopy_init(
     const std::uint32_t transpose_of_faces = 0 /*unused*/,

diff --git a/tt_metal/hw/ckernels/grayskull/metal/llk_api/llk_pack_api.h b/tt_metal/hw/ckernels/grayskull/metal/llk_api/llk_pack_api.h
@@ -247,7 +247,7 @@ inline void llk_pack_reduce_config_v2(uint32_t output) {
     }
 }
 
-template <bool out_of_order_output = false, DstSync Dst = SyncFull, bool untilize = false>
+template <bool out_of_order_output = false, DstSync Dst = SyncFull, bool untilize = false, bool is_fp32_dest_acc_en = false /*unused*/>
 inline void llk_matmul_pack(std::uint32_t start_tile_index, std::uint32_t output, uint32_t ntiles, std::uint32_t output_tile_index = 0) {
     std::uint8_t output_id = get_output_id(output);
     const std::uint8_t OUTPUT_BASE_ID = (std::uint8_t) get_output_base_id();