[NHWC][asm igemm][GFX908][GFX90a] fix fwd kernel that have computatio…

…n error in small cyx case (#1137) * add gfx908 fwd kernel for fix small cyx * add gfx90a fwd kernel for fix small cyx * add ctest for the failed case
ROCm · Sep 6, 2021 · 10a787b · 10a787b
1 parent 4e7b1ab
commit 10a787b
Show file tree

Hide file tree

Showing 303 changed files with 1,925 additions and 12,552 deletions.
diff --git a/...c_fp16_bx0_ex0_bt128x128x32_wt32x32x8_ws1x1_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x2x1_1x4x1x64.s b/...c_fp16_bx0_ex0_bt128x128x32_wt32x32x8_ws1x1_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x2x1_1x4x1x64.s
@@ -23,47 +23,9 @@
  * SOFTWARE.
  *
  *******************************************************************************/
-; generated by igemm_codegen.py (0ffceb844f35ce8d878168528ba6941b666ff7c0)
+; generated by igemm_codegen.py (32a41f791dcf0139e95f217f3905939fbbae794c)
 ;
-.macro .mdiv_u32_ss s_quot s_numer s_magic s_shift s_tmp
-    s_mul_hi_u32 s[\s_tmp], s[\s_magic], s[\s_numer]
-    s_add_u32 s[\s_tmp], s[\s_tmp], s[\s_numer]
-    s_lshr_b32 s[\s_quot], s[\s_tmp], s[\s_shift]
-.endm
-
-.macro .mdiv_u32_rem_ss s_rem s_quot s_numer s_magic s_shift s_denom s_tmp
-    .mdiv_u32_ss \s_quot,\s_numer,\s_magic,\s_shift,\s_tmp
-    s_mul_i32 s[\s_tmp], s[\s_denom], s[\s_quot]
-    s_sub_u32 s[\s_rem], s[\s_numer], s[\s_tmp]
-.endm
-
-.macro .mdiv_u32_vs v_quot v_numer s_magic s_shift v_tmp
-    v_mul_hi_u32 v[\v_tmp], s[\s_magic], v[\v_numer]
-    v_add_u32 v[\v_tmp], v[\v_tmp], v[\v_numer]
-    v_lshrrev_b32 v[\v_quot], s[\s_shift], v[\v_tmp]
-.endm
-
-.macro .mdiv_u32_rem_vs v_rem v_quot v_numer s_magic s_shift s_denom v_tmp
-    .mdiv_u32_vs \v_quot,\v_numer,\s_magic,\s_shift,\v_tmp
-    v_mul_lo_u32 v[\v_tmp], s[\s_denom], v[\v_quot]
-    v_sub_u32 v[\v_rem], v[\v_numer], v[\v_tmp]
-.endm
-
-.macro .v_clear_acc_c a, num
-    _a = \a
-    .rept \num
-        v_accvgpr_write_b32 a[_a], 0
-        _a = _a + 1
-    .endr
-.endm
-
-.macro .v_clear_nc vid, num
-    _v = \vid
-    .rept \num
-        v_mov_b32 v[_v], 0
-        _v = _v + 1
-    .endr
-.endm
+.include "igemm_fwd_gtcx_nhwc_fp16_utils.inc"
 
 ;----------------------------------------------------------
 ; starting of kernel igemm_fwd_gtcx_nhwc_fp16_bx0_ex0_bt128x128x32_wt32x32x8_ws1x1_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x2x1_1x4x1x64

diff --git a/...6_bx0_ex0_bt128x128x32_wt32x32x8_ws1x1_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x2x1_1x4x1x64_gkgs.s b/...6_bx0_ex0_bt128x128x32_wt32x32x8_ws1x1_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x2x1_1x4x1x64_gkgs.s
@@ -23,47 +23,9 @@
  * SOFTWARE.
  *
  *******************************************************************************/
-; generated by igemm_codegen.py (0ffceb844f35ce8d878168528ba6941b666ff7c0)
+; generated by igemm_codegen.py (32a41f791dcf0139e95f217f3905939fbbae794c)
 ;
-.macro .mdiv_u32_ss s_quot s_numer s_magic s_shift s_tmp
-    s_mul_hi_u32 s[\s_tmp], s[\s_magic], s[\s_numer]
-    s_add_u32 s[\s_tmp], s[\s_tmp], s[\s_numer]
-    s_lshr_b32 s[\s_quot], s[\s_tmp], s[\s_shift]
-.endm
-
-.macro .mdiv_u32_rem_ss s_rem s_quot s_numer s_magic s_shift s_denom s_tmp
-    .mdiv_u32_ss \s_quot,\s_numer,\s_magic,\s_shift,\s_tmp
-    s_mul_i32 s[\s_tmp], s[\s_denom], s[\s_quot]
-    s_sub_u32 s[\s_rem], s[\s_numer], s[\s_tmp]
-.endm
-
-.macro .mdiv_u32_vs v_quot v_numer s_magic s_shift v_tmp
-    v_mul_hi_u32 v[\v_tmp], s[\s_magic], v[\v_numer]
-    v_add_u32 v[\v_tmp], v[\v_tmp], v[\v_numer]
-    v_lshrrev_b32 v[\v_quot], s[\s_shift], v[\v_tmp]
-.endm
-
-.macro .mdiv_u32_rem_vs v_rem v_quot v_numer s_magic s_shift s_denom v_tmp
-    .mdiv_u32_vs \v_quot,\v_numer,\s_magic,\s_shift,\v_tmp
-    v_mul_lo_u32 v[\v_tmp], s[\s_denom], v[\v_quot]
-    v_sub_u32 v[\v_rem], v[\v_numer], v[\v_tmp]
-.endm
-
-.macro .v_clear_acc_c a, num
-    _a = \a
-    .rept \num
-        v_accvgpr_write_b32 a[_a], 0
-        _a = _a + 1
-    .endr
-.endm
-
-.macro .v_clear_nc vid, num
-    _v = \vid
-    .rept \num
-        v_mov_b32 v[_v], 0
-        _v = _v + 1
-    .endr
-.endm
+.include "igemm_fwd_gtcx_nhwc_fp16_utils.inc"
 
 ;----------------------------------------------------------
 ; starting of kernel igemm_fwd_gtcx_nhwc_fp16_bx0_ex0_bt128x128x32_wt32x32x8_ws1x1_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x2x1_1x4x1x64_gkgs

diff --git a/...c_fp16_bx0_ex0_bt128x256x32_wt32x32x8_ws1x2_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x4x1_1x4x1x64.s b/...c_fp16_bx0_ex0_bt128x256x32_wt32x32x8_ws1x2_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x4x1_1x4x1x64.s
@@ -23,47 +23,9 @@
  * SOFTWARE.
  *
  *******************************************************************************/
-; generated by igemm_codegen.py (0ffceb844f35ce8d878168528ba6941b666ff7c0)
+; generated by igemm_codegen.py (32a41f791dcf0139e95f217f3905939fbbae794c)
 ;
-.macro .mdiv_u32_ss s_quot s_numer s_magic s_shift s_tmp
-    s_mul_hi_u32 s[\s_tmp], s[\s_magic], s[\s_numer]
-    s_add_u32 s[\s_tmp], s[\s_tmp], s[\s_numer]
-    s_lshr_b32 s[\s_quot], s[\s_tmp], s[\s_shift]
-.endm
-
-.macro .mdiv_u32_rem_ss s_rem s_quot s_numer s_magic s_shift s_denom s_tmp
-    .mdiv_u32_ss \s_quot,\s_numer,\s_magic,\s_shift,\s_tmp
-    s_mul_i32 s[\s_tmp], s[\s_denom], s[\s_quot]
-    s_sub_u32 s[\s_rem], s[\s_numer], s[\s_tmp]
-.endm
-
-.macro .mdiv_u32_vs v_quot v_numer s_magic s_shift v_tmp
-    v_mul_hi_u32 v[\v_tmp], s[\s_magic], v[\v_numer]
-    v_add_u32 v[\v_tmp], v[\v_tmp], v[\v_numer]
-    v_lshrrev_b32 v[\v_quot], s[\s_shift], v[\v_tmp]
-.endm
-
-.macro .mdiv_u32_rem_vs v_rem v_quot v_numer s_magic s_shift s_denom v_tmp
-    .mdiv_u32_vs \v_quot,\v_numer,\s_magic,\s_shift,\v_tmp
-    v_mul_lo_u32 v[\v_tmp], s[\s_denom], v[\v_quot]
-    v_sub_u32 v[\v_rem], v[\v_numer], v[\v_tmp]
-.endm
-
-.macro .v_clear_acc_c a, num
-    _a = \a
-    .rept \num
-        v_accvgpr_write_b32 a[_a], 0
-        _a = _a + 1
-    .endr
-.endm
-
-.macro .v_clear_nc vid, num
-    _v = \vid
-    .rept \num
-        v_mov_b32 v[_v], 0
-        _v = _v + 1
-    .endr
-.endm
+.include "igemm_fwd_gtcx_nhwc_fp16_utils.inc"
 
 ;----------------------------------------------------------
 ; starting of kernel igemm_fwd_gtcx_nhwc_fp16_bx0_ex0_bt128x256x32_wt32x32x8_ws1x2_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x4x1_1x4x1x64

diff --git a/...6_bx0_ex0_bt128x256x32_wt32x32x8_ws1x2_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x4x1_1x4x1x64_gkgs.s b/...6_bx0_ex0_bt128x256x32_wt32x32x8_ws1x2_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x4x1_1x4x1x64_gkgs.s
@@ -23,47 +23,9 @@
  * SOFTWARE.
  *
  *******************************************************************************/
-; generated by igemm_codegen.py (0ffceb844f35ce8d878168528ba6941b666ff7c0)
+; generated by igemm_codegen.py (32a41f791dcf0139e95f217f3905939fbbae794c)
 ;
-.macro .mdiv_u32_ss s_quot s_numer s_magic s_shift s_tmp
-    s_mul_hi_u32 s[\s_tmp], s[\s_magic], s[\s_numer]
-    s_add_u32 s[\s_tmp], s[\s_tmp], s[\s_numer]
-    s_lshr_b32 s[\s_quot], s[\s_tmp], s[\s_shift]
-.endm
-
-.macro .mdiv_u32_rem_ss s_rem s_quot s_numer s_magic s_shift s_denom s_tmp
-    .mdiv_u32_ss \s_quot,\s_numer,\s_magic,\s_shift,\s_tmp
-    s_mul_i32 s[\s_tmp], s[\s_denom], s[\s_quot]
-    s_sub_u32 s[\s_rem], s[\s_numer], s[\s_tmp]
-.endm
-
-.macro .mdiv_u32_vs v_quot v_numer s_magic s_shift v_tmp
-    v_mul_hi_u32 v[\v_tmp], s[\s_magic], v[\v_numer]
-    v_add_u32 v[\v_tmp], v[\v_tmp], v[\v_numer]
-    v_lshrrev_b32 v[\v_quot], s[\s_shift], v[\v_tmp]
-.endm
-
-.macro .mdiv_u32_rem_vs v_rem v_quot v_numer s_magic s_shift s_denom v_tmp
-    .mdiv_u32_vs \v_quot,\v_numer,\s_magic,\s_shift,\v_tmp
-    v_mul_lo_u32 v[\v_tmp], s[\s_denom], v[\v_quot]
-    v_sub_u32 v[\v_rem], v[\v_numer], v[\v_tmp]
-.endm
-
-.macro .v_clear_acc_c a, num
-    _a = \a
-    .rept \num
-        v_accvgpr_write_b32 a[_a], 0
-        _a = _a + 1
-    .endr
-.endm
-
-.macro .v_clear_nc vid, num
-    _v = \vid
-    .rept \num
-        v_mov_b32 v[_v], 0
-        _v = _v + 1
-    .endr
-.endm
+.include "igemm_fwd_gtcx_nhwc_fp16_utils.inc"
 
 ;----------------------------------------------------------
 ; starting of kernel igemm_fwd_gtcx_nhwc_fp16_bx0_ex0_bt128x256x32_wt32x32x8_ws1x2_wr2x2_ta1x8x2x1_1x4x1x64_tb1x8x4x1_1x4x1x64_gkgs

diff --git a/...wc_fp16_bx0_ex0_bt128x32x32_wt64x16x4_ws1x1_wr1x1_ta1x4x4x1_1x8x1x32_tb1x4x1x1_1x8x1x32.s b/...wc_fp16_bx0_ex0_bt128x32x32_wt64x16x4_ws1x1_wr1x1_ta1x4x4x1_1x8x1x32_tb1x4x1x1_1x8x1x32.s
@@ -23,47 +23,9 @@
  * SOFTWARE.
  *
  *******************************************************************************/
-; generated by igemm_codegen.py (0ffceb844f35ce8d878168528ba6941b666ff7c0)
+; generated by igemm_codegen.py (32a41f791dcf0139e95f217f3905939fbbae794c)
 ;
-.macro .mdiv_u32_ss s_quot s_numer s_magic s_shift s_tmp
-    s_mul_hi_u32 s[\s_tmp], s[\s_magic], s[\s_numer]
-    s_add_u32 s[\s_tmp], s[\s_tmp], s[\s_numer]
-    s_lshr_b32 s[\s_quot], s[\s_tmp], s[\s_shift]
-.endm
-
-.macro .mdiv_u32_rem_ss s_rem s_quot s_numer s_magic s_shift s_denom s_tmp
-    .mdiv_u32_ss \s_quot,\s_numer,\s_magic,\s_shift,\s_tmp
-    s_mul_i32 s[\s_tmp], s[\s_denom], s[\s_quot]
-    s_sub_u32 s[\s_rem], s[\s_numer], s[\s_tmp]
-.endm
-
-.macro .mdiv_u32_vs v_quot v_numer s_magic s_shift v_tmp
-    v_mul_hi_u32 v[\v_tmp], s[\s_magic], v[\v_numer]
-    v_add_u32 v[\v_tmp], v[\v_tmp], v[\v_numer]
-    v_lshrrev_b32 v[\v_quot], s[\s_shift], v[\v_tmp]
-.endm
-
-.macro .mdiv_u32_rem_vs v_rem v_quot v_numer s_magic s_shift s_denom v_tmp
-    .mdiv_u32_vs \v_quot,\v_numer,\s_magic,\s_shift,\v_tmp
-    v_mul_lo_u32 v[\v_tmp], s[\s_denom], v[\v_quot]
-    v_sub_u32 v[\v_rem], v[\v_numer], v[\v_tmp]
-.endm
-
-.macro .v_clear_acc_c a, num
-    _a = \a
-    .rept \num
-        v_accvgpr_write_b32 a[_a], 0
-        _a = _a + 1
-    .endr
-.endm
-
-.macro .v_clear_nc vid, num
-    _v = \vid
-    .rept \num
-        v_mov_b32 v[_v], 0
-        _v = _v + 1
-    .endr
-.endm
+.include "igemm_fwd_gtcx_nhwc_fp16_utils.inc"
 
 ;----------------------------------------------------------
 ; starting of kernel igemm_fwd_gtcx_nhwc_fp16_bx0_ex0_bt128x32x32_wt64x16x4_ws1x1_wr1x1_ta1x4x4x1_1x8x1x32_tb1x4x1x1_1x8x1x32

diff --git a/...16_bx0_ex0_bt128x32x32_wt64x16x4_ws1x1_wr1x1_ta1x4x4x1_1x8x1x32_tb1x4x1x1_1x8x1x32_gkgs.s b/...16_bx0_ex0_bt128x32x32_wt64x16x4_ws1x1_wr1x1_ta1x4x4x1_1x8x1x32_tb1x4x1x1_1x8x1x32_gkgs.s
@@ -23,47 +23,9 @@
  * SOFTWARE.
  *
  *******************************************************************************/
-; generated by igemm_codegen.py (0ffceb844f35ce8d878168528ba6941b666ff7c0)
+; generated by igemm_codegen.py (32a41f791dcf0139e95f217f3905939fbbae794c)
 ;
-.macro .mdiv_u32_ss s_quot s_numer s_magic s_shift s_tmp
-    s_mul_hi_u32 s[\s_tmp], s[\s_magic], s[\s_numer]
-    s_add_u32 s[\s_tmp], s[\s_tmp], s[\s_numer]
-    s_lshr_b32 s[\s_quot], s[\s_tmp], s[\s_shift]
-.endm
-
-.macro .mdiv_u32_rem_ss s_rem s_quot s_numer s_magic s_shift s_denom s_tmp
-    .mdiv_u32_ss \s_quot,\s_numer,\s_magic,\s_shift,\s_tmp
-    s_mul_i32 s[\s_tmp], s[\s_denom], s[\s_quot]
-    s_sub_u32 s[\s_rem], s[\s_numer], s[\s_tmp]
-.endm
-
-.macro .mdiv_u32_vs v_quot v_numer s_magic s_shift v_tmp
-    v_mul_hi_u32 v[\v_tmp], s[\s_magic], v[\v_numer]
-    v_add_u32 v[\v_tmp], v[\v_tmp], v[\v_numer]
-    v_lshrrev_b32 v[\v_quot], s[\s_shift], v[\v_tmp]
-.endm
-
-.macro .mdiv_u32_rem_vs v_rem v_quot v_numer s_magic s_shift s_denom v_tmp
-    .mdiv_u32_vs \v_quot,\v_numer,\s_magic,\s_shift,\v_tmp
-    v_mul_lo_u32 v[\v_tmp], s[\s_denom], v[\v_quot]
-    v_sub_u32 v[\v_rem], v[\v_numer], v[\v_tmp]
-.endm
-
-.macro .v_clear_acc_c a, num
-    _a = \a
-    .rept \num
-        v_accvgpr_write_b32 a[_a], 0
-        _a = _a + 1
-    .endr
-.endm
-
-.macro .v_clear_nc vid, num
-    _v = \vid
-    .rept \num
-        v_mov_b32 v[_v], 0
-        _v = _v + 1
-    .endr
-.endm
+.include "igemm_fwd_gtcx_nhwc_fp16_utils.inc"
 
 ;----------------------------------------------------------
 ; starting of kernel igemm_fwd_gtcx_nhwc_fp16_bx0_ex0_bt128x32x32_wt64x16x4_ws1x1_wr1x1_ta1x4x4x1_1x8x1x32_tb1x4x1x1_1x8x1x32_gkgs

diff --git a/...16_bx0_ex0_bt128x64x32_wt32x32x8_ws1x1_wr1x2_ta1x16x1x1_1x2x4x32_tb1x8x1x1_1x4x1x64_pta.s b/...16_bx0_ex0_bt128x64x32_wt32x32x8_ws1x1_wr1x2_ta1x16x1x1_1x2x4x32_tb1x8x1x1_1x4x1x64_pta.s
@@ -23,47 +23,9 @@
  * SOFTWARE.
  *
  *******************************************************************************/
-; generated by igemm_codegen.py (0ffceb844f35ce8d878168528ba6941b666ff7c0)
+; generated by igemm_codegen.py (32a41f791dcf0139e95f217f3905939fbbae794c)
 ;
-.macro .mdiv_u32_ss s_quot s_numer s_magic s_shift s_tmp
-    s_mul_hi_u32 s[\s_tmp], s[\s_magic], s[\s_numer]
-    s_add_u32 s[\s_tmp], s[\s_tmp], s[\s_numer]
-    s_lshr_b32 s[\s_quot], s[\s_tmp], s[\s_shift]
-.endm
-
-.macro .mdiv_u32_rem_ss s_rem s_quot s_numer s_magic s_shift s_denom s_tmp
-    .mdiv_u32_ss \s_quot,\s_numer,\s_magic,\s_shift,\s_tmp
-    s_mul_i32 s[\s_tmp], s[\s_denom], s[\s_quot]
-    s_sub_u32 s[\s_rem], s[\s_numer], s[\s_tmp]
-.endm
-
-.macro .mdiv_u32_vs v_quot v_numer s_magic s_shift v_tmp
-    v_mul_hi_u32 v[\v_tmp], s[\s_magic], v[\v_numer]
-    v_add_u32 v[\v_tmp], v[\v_tmp], v[\v_numer]
-    v_lshrrev_b32 v[\v_quot], s[\s_shift], v[\v_tmp]
-.endm
-
-.macro .mdiv_u32_rem_vs v_rem v_quot v_numer s_magic s_shift s_denom v_tmp
-    .mdiv_u32_vs \v_quot,\v_numer,\s_magic,\s_shift,\v_tmp
-    v_mul_lo_u32 v[\v_tmp], s[\s_denom], v[\v_quot]
-    v_sub_u32 v[\v_rem], v[\v_numer], v[\v_tmp]
-.endm
-
-.macro .v_clear_acc_c a, num
-    _a = \a
-    .rept \num
-        v_accvgpr_write_b32 a[_a], 0
-        _a = _a + 1
-    .endr
-.endm
-
-.macro .v_clear_nc vid, num
-    _v = \vid
-    .rept \num
-        v_mov_b32 v[_v], 0
-        _v = _v + 1
-    .endr
-.endm
+.include "igemm_fwd_gtcx_nhwc_fp16_utils.inc"
 
 ;----------------------------------------------------------
 ; starting of kernel igemm_fwd_gtcx_nhwc_fp16_bx0_ex0_bt128x64x32_wt32x32x8_ws1x1_wr1x2_ta1x16x1x1_1x2x4x32_tb1x8x1x1_1x4x1x64_pta

diff --git a/...0_ex0_bt128x64x32_wt32x32x8_ws1x1_wr1x2_ta1x16x1x1_1x2x4x32_tb1x8x1x1_1x4x1x64_pta_gkgs.s b/...0_ex0_bt128x64x32_wt32x32x8_ws1x1_wr1x2_ta1x16x1x1_1x2x4x32_tb1x8x1x1_1x4x1x64_pta_gkgs.s
@@ -23,47 +23,9 @@
  * SOFTWARE.
  *
  *******************************************************************************/
-; generated by igemm_codegen.py (0ffceb844f35ce8d878168528ba6941b666ff7c0)
+; generated by igemm_codegen.py (32a41f791dcf0139e95f217f3905939fbbae794c)
 ;
-.macro .mdiv_u32_ss s_quot s_numer s_magic s_shift s_tmp
-    s_mul_hi_u32 s[\s_tmp], s[\s_magic], s[\s_numer]
-    s_add_u32 s[\s_tmp], s[\s_tmp], s[\s_numer]
-    s_lshr_b32 s[\s_quot], s[\s_tmp], s[\s_shift]
-.endm
-
-.macro .mdiv_u32_rem_ss s_rem s_quot s_numer s_magic s_shift s_denom s_tmp
-    .mdiv_u32_ss \s_quot,\s_numer,\s_magic,\s_shift,\s_tmp
-    s_mul_i32 s[\s_tmp], s[\s_denom], s[\s_quot]
-    s_sub_u32 s[\s_rem], s[\s_numer], s[\s_tmp]
-.endm
-
-.macro .mdiv_u32_vs v_quot v_numer s_magic s_shift v_tmp
-    v_mul_hi_u32 v[\v_tmp], s[\s_magic], v[\v_numer]
-    v_add_u32 v[\v_tmp], v[\v_tmp], v[\v_numer]
-    v_lshrrev_b32 v[\v_quot], s[\s_shift], v[\v_tmp]
-.endm
-
-.macro .mdiv_u32_rem_vs v_rem v_quot v_numer s_magic s_shift s_denom v_tmp
-    .mdiv_u32_vs \v_quot,\v_numer,\s_magic,\s_shift,\v_tmp
-    v_mul_lo_u32 v[\v_tmp], s[\s_denom], v[\v_quot]
-    v_sub_u32 v[\v_rem], v[\v_numer], v[\v_tmp]
-.endm
-
-.macro .v_clear_acc_c a, num
-    _a = \a
-    .rept \num
-        v_accvgpr_write_b32 a[_a], 0
-        _a = _a + 1
-    .endr
-.endm
-
-.macro .v_clear_nc vid, num
-    _v = \vid
-    .rept \num
-        v_mov_b32 v[_v], 0
-        _v = _v + 1
-    .endr
-.endm
+.include "igemm_fwd_gtcx_nhwc_fp16_utils.inc"
 
 ;----------------------------------------------------------
 ; starting of kernel igemm_fwd_gtcx_nhwc_fp16_bx0_ex0_bt128x64x32_wt32x32x8_ws1x1_wr1x2_ta1x16x1x1_1x2x4x32_tb1x8x1x1_1x4x1x64_pta_gkgs