| # Copyright 2019 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. |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_ld64 |
| k-block: 2 |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a57 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a57 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__aarch64_neonfma_cortex_a57 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__aarch64_neonfma_cortex_a75 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a57 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75 |
| k-block: 8 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_1x12__aarch64_neonfma_cortex_a53 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x12__aarch64_neonfma_cortex_a53 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_ld64 |
| k-block: 2 |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_pld_cortex_a75 |
| k-block: 4 |
| pipelined: true |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_ld64 |
| k-block: 2 |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7 |
| k-block: 2 |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x4__aarch32_vfp_ld64 |
| k-block: 2 |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_ld128 |
| k-block: 4 |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_ld64 |
| k-block: 2 |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_ld128 |
| k-block: 4 |
| assembly: true |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld128 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld128 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__neonfma_lane_ld64 |
| k-block: 2 |
| arch: |
| - aarch64 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64 |
| k-block: 2 |
| arch: |
| - aarch64 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld128 |
| k-block: 4 |
| arch: |
| - aarch64 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64 |
| k-block: 2 |
| arch: |
| - aarch64 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64 |
| k-block: 2 |
| arch: |
| - aarch64 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld128 |
| k-block: 4 |
| arch: |
| - aarch64 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__neon_dup_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__neon_dup_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__neon_dup_ld128 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld128 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__neonfma_dup_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld128 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64 |
| k-block: 2 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld128 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8s4__neon |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8s4__neon |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8s4__neon |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_8x8s4__neon |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8s4__neonfma |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8s4__neonfma |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8s4__neonfma |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_8x8s4__neonfma |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__sse_load1 |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8__sse_load1 |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__sse_load1 |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__sse_load1 |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__sse_dup |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8__sse_dup |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__sse_dup |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__sse_dup |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8s4__sse |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8s4__sse |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8s4__sse |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8s4__sse |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x2c4__sse |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__sse2_dup |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8__sse2_dup |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__avx_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__avx_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__avx_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__avx_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_7x8__avx_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_1x16__avx_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_3x16__avx_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x16__avx_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_5x16__avx_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_7x8__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_8x8__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_1x16__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_3x16__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x16__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_5x16__fma3_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_5x16s4__fma3_broadcast |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x16__avx512f_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_arm_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_arm_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_arm_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_x86_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_x86_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_x86_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_arm_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_arm_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_arm_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_x86_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_x86_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_x86_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_splat |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_arm |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8s4__wasmsimd_arm |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8s4__wasmsimd_arm |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8s4__wasmsimd_arm |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_3x8s4__wasmsimd_x86 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x8s4__wasmsimd_x86 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_6x8s4__wasmsimd_x86 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_arm |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_x86 |
| k-block: 4 |
| - name: xnn_f32_gemm_minmax_ukernel_1x4__wasm |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_2x4__wasm |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x4__wasm |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x2__wasm |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_1x4__scalar |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_2x4__scalar |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x4__scalar |
| k-block: 1 |
| - name: xnn_f32_gemm_minmax_ukernel_4x2__scalar |
| k-block: 1 |