Enable UInt8 math (#27)

tenstorrent · Aug 2, 2024 · 545b798 · 545b798
1 parent 136f561
commit 545b798
Show file tree

Hide file tree

Showing 2 changed files with 13 additions and 2 deletions.
diff --git a/common/inc/cpack_common.h b/common/inc/cpack_common.h
@@ -248,6 +248,9 @@ namespace ckernel::packer
                                                       (pack_output_src_format == (uint)DataFormat::Int32) |
                                                       (pack_output_src_format == (uint)DataFormat::Float32) |
                                                       (is_fp32_dest_acc_en ? 1 : 0);
+      if (pack_dst_format == (uint)DataFormat::UInt8) {
+         dest_rd_ctrl.f.PCK_DEST_RD_CTRL_Read_unsigned = 1;
+      }
 
       //Round to 10 bit mantissa from fp32 dest
       if(is_fp32_dest_acc_en && (pack_output_src_format == (uint)DataFormat::Float16)) {
@@ -349,7 +352,7 @@ namespace ckernel::packer
 
       t6_mutex_release(mutex::REG_RMW);
 
-      set_packer_config<is_fp32_dest_acc_en>(pack_output_src_format, pack_output_dst_format, num_faces, partial_face);
+      set_packer_config<is_fp32_dest_acc_en>(pack_src_format, pack_dst_format, num_faces, partial_face);
 
       // PACK_COUNTERS_SEC0_pack_per_xy_plane = cfg_reg_array[3][0 +: 8];
       // PACK_COUNTERS_SEC0_pack_reads_per_xy_plane = cfg_reg_array[3][8 +: 8];

diff --git a/common/inc/cunpack_common.h b/common/inc/cunpack_common.h
@@ -241,10 +241,18 @@ namespace ckernel::unpacker
                                    ((uint)unpA_dst_format_masked == (uint)DataFormat::Int32) ||
                                    ((uint)unpB_dst_format_masked == (uint)DataFormat::Int32);
 
-      constexpr uint alu_format_mask = ALU_FORMAT_SPEC_REG0_SrcA_MASK | ALU_FORMAT_SPEC_REG1_SrcB_MASK;
+      constexpr uint alu_format_mask = ALU_FORMAT_SPEC_REG0_SrcA_MASK | ALU_FORMAT_SPEC_REG1_SrcB_MASK |
+                                       ALU_FORMAT_SPEC_REG0_SrcAUnsigned_MASK | ALU_FORMAT_SPEC_REG0_SrcBUnsigned_MASK;
       alu_payload.f.ALU_FORMAT_SPEC_REG0_SrcA = unpA_dst_format_masked;
       alu_payload.f.ALU_FORMAT_SPEC_REG1_SrcB = row_pool ? ((uint) DataFormat::Float16 | (exp_width<<2)) : unpB_dst_format_masked;
 
+      if ((uint)unpA_src_format == (uint)DataFormat::UInt8) {
+         alu_payload.f.ALU_FORMAT_SPEC_REG0_SrcAUnsigned = 1;
+      }
+      if ((uint)unpB_src_format == (uint)DataFormat::UInt8) {
+         alu_payload.f.ALU_FORMAT_SPEC_REG0_SrcBUnsigned = 1;
+      }
+
       // FP32 accumulation and SFPU to read dest as FP32
       // NOTE: This assumes these config fields are adjacent and in same register!!
       static_assert(ALU_ACC_CTRL_Fp32_enabled_ADDR32 == ALU_FORMAT_SPEC_REG0_SrcA_ADDR32);