Skip to content

Commit

Permalink
Auto generate script for header
Browse files Browse the repository at this point in the history
  • Loading branch information
@RahulSundarMCW
RahulSundarMCW committed Jan 9, 2025
1 parent ed8080a commit 3cc217a
Show file tree
Hide file tree
Showing 25 changed files with 3,465 additions and 438 deletions.
1 change: 0 additions & 1 deletion scripts/generate-tests.sh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,6 @@ tools/generate-gemm-test.py --ukernel f16-f32acc-gemm-minmax --output-test test/
tools/generate-gemm-test.py --ukernel f32-gemm --output-test test/f32-gemm.cc &
tools/generate-gemm-test.py --ukernel f32-gemm-relu --output-test test/f32-gemm-relu.cc &
tools/generate-gemm-test.py --ukernel f32-gemm-minmax --output-test test/f32-gemm-minmax.cc &
### TODO
tools/generate-gemm-test.py --ukernel f32-gemminc-minmax --output-test test/f32-gemminc-minmax.cc &
tools/generate-gemm-test.py --ukernel f32-gemm-goi-minmax --output-test test/f32-gemm-goi-minmax.cc &

Expand Down
65 changes: 35 additions & 30 deletions src/bf16-gemm/bf16-gemm-minmax.h
View file
Original file line number Diff line number Diff line change
@@ -1,33 +1,38 @@
// Copyright 2023 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
// Arguments are:
// XNN_GEMM_MINMAX(arch_flags, fn_name, k_block, is_pipelined, mr, nr, kr, sr, mr_packed, is_igemm, datatype, params_type, init_fn, pack_fn)

//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
// Arguments are:
// XNN_GEMM(arch_flags, fn_name, k_block, is_pipelined, mr, nr, kr, sr, mr_packed, is_igemm, datatype, params_type, init_fn, pack_fn)


#if XNN_ARCH_ARM || XNN_ARCH_ARM64
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_shland, 8, false, 1, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_shland, 8, false, 2, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_shland, 8, false, 3, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_shland, 8, false, 4, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland, 8, false, 5, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_zip, 8, false, 1, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip, 8, false, 2, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_zip, 8, false, 3, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_zip, 8, false, 4, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_zip, 8, false, 5, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x8c2__neonbf16_bfdot_lane_ld128, 8, false, 1, 8, 2, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x8c2__neonbf16_bfdot_lane_ld128, 8, false, 4, 8, 2, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x8c2__neonbf16_bfdot_lane_ld128, 8, false, 5, 8, 2, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_6x8c2__neonbf16_bfdot_lane_ld128, 8, false, 6, 8, 2, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonbf16_bfdot, 8, false, 1, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonbf16_bfdot, 8, false, 2, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonbf16_bfdot, 8, false, 3, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonbf16_bfdot, 8, false, 4, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfdot, 8, false, 5, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonbf16_bfmlal, 8, 1, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonbf16_bfmlal, 8, false, 2, 4, 8, 1, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonbf16_bfmlal, 8, false, 3, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonbf16_bfmlal, 8, false, 4, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfmlal, 8, false, 5, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_shland, 8, false, 1, 4, 8, 1, 1, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_shland, 8, false, 2, 4, 8, 1, 2, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_shland, 8, false, 3, 4, 8, 1, 3, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_shland, 8, false, 4, 4, 8, 1, 4, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland, 8, false, 5, 4, 8, 1, 5, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_zip, 8, false, 1, 4, 8, 1, 1, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip, 8, false, 2, 4, 8, 1, 2, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_zip, 8, false, 3, 4, 8, 1, 3, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_zip, 8, false, 4, 4, 8, 1, 4, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_fma, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_zip, 8, false, 5, 4, 8, 1, 5, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64

#if XNN_ENABLE_ARM_BF16 && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_1x8c2__neonbf16_bfdot_lane_ld128, 8, false, 1, 8, 2, 1, 1, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_4x8c2__neonbf16_bfdot_lane_ld128, 8, false, 4, 8, 2, 1, 4, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_5x8c2__neonbf16_bfdot_lane_ld128, 8, false, 5, 8, 2, 1, 5, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_6x8c2__neonbf16_bfdot_lane_ld128, 8, false, 6, 8, 2, 1, 6, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonbf16_bfdot, 8, false, 1, 4, 8, 1, 1, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonbf16_bfdot, 8, false, 2, 4, 8, 1, 2, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonbf16_bfdot, 8, false, 3, 4, 8, 1, 3, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonbf16_bfdot, 8, false, 4, 4, 8, 1, 4, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfdot, 8, false, 5, 4, 8, 1, 5, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonbf16_bfmlal, 8, false, 1, 4, 8, 1, 1, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonbf16_bfmlal, 8, false, 2, 4, 8, 1, 2, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonbf16_bfmlal, 8, false, 3, 4, 8, 1, 3, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonbf16_bfmlal, 8, false, 4, 4, 8, 1, 4, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(xnn_arch_arm_neon_bf16, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfmlal, 8, false, 5, 4, 8, 1, 5, false, xnn_bfloat16, struct xnn_bf16_default_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
#endif // XNN_ENABLE_ARM_BF16 && (XNN_ARCH_ARM || XNN_ARCH_ARM64)

33 changes: 33 additions & 0 deletions src/bf16-gemm/bf16-gemm-minmax1.h
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,33 @@
// Copyright 2023 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
// Arguments are:
// XNN_GEMM_MINMAX(arch_flags, fn_name, k_block, is_pipelined, mr, nr, kr, sr, mr_packed, is_igemm, datatype, params_type, init_fn, pack_fn)

#if XNN_ARCH_ARM || XNN_ARCH_ARM64
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_shland, 8, false, 1, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_shland, 8, false, 2, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_shland, 8, false, 3, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_shland, 8, false, 4, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_shland, 8, false, 5, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonfma_zip, 8, false, 1, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip, 8, false, 2, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonfma_zip, 8, false, 3, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonfma_zip, 8, false, 4, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonfma_zip, 8, false, 5, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x8c2__neonbf16_bfdot_lane_ld128, 8, false, 1, 8, 2, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x8c2__neonbf16_bfdot_lane_ld128, 8, false, 4, 8, 2, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x8c2__neonbf16_bfdot_lane_ld128, 8, false, 5, 8, 2, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_6x8c2__neonbf16_bfdot_lane_ld128, 8, false, 6, 8, 2, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonbf16_bfdot, 8, false, 1, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonbf16_bfdot, 8, false, 2, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonbf16_bfdot, 8, false, 3, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonbf16_bfdot, 8, false, 4, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfdot, 8, false, 5, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_1x4c8__neonbf16_bfmlal, 8, 1, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_2x4c8__neonbf16_bfmlal, 8, false, 2, 4, 8, 1, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_3x4c8__neonbf16_bfmlal, 8, false, 3, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_4x4c8__neonbf16_bfmlal, 8, false, 4, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
XNN_GEMM(XNN_ARCH_ARM, xnn_bf16_gemm_minmax_ukernel_5x4c8__neonbf16_bfmlal, 8, false, 5, 4, 8, 1, false, xnn_bf16_default_params, struct xnn_bf16_minmax_params, xnn_init_bf16_minmax_scalar_params, xnn_pack_f16_gemm_goi_w)
#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
Loading

0 comments on commit 3cc217a

Please sign in to comment.