| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 6x2-minmax-neon-lane-ld64.c | 86 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64() local 111 const float32x2_t va4 = vld1_dup_f32(a4); a4 += 1; in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64() local
|
| D | 5x8-minmax-neon-lane-ld64.c | 84 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() local 118 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() local
|
| D | 5x8-minmax-neonfma-lane-ld64.c | 84 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64() local 118 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1; in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64() local
|
| D | 5x8s4-wasmrelaxedsimd-fma.c | 88 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local 179 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local
|
| D | 6x2-minmax-neonfma-lane-ld64.c | 86 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64() local 141 const float32x2_t va4 = vld1_dup_f32(a4); a4 += 1; in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64() local
|
| D | 5x8s4-wasmsimd.c | 88 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd() local 179 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_ukernel_5x8s4__wasmsimd() local
|
| D | 5x8s4-minmax-sse.c | 88 __m128 va4 = _mm_loadu_ps(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse() local 179 __m128 va4 = _mm_loadu_ps(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__sse() local
|
| D | 6x8-minmax-neonfma-lane-ld64.c | 92 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() local 131 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() local
|
| D | 5x8s4-minmax-wasmrelaxedsimd.c | 90 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd() local 181 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd() local
|
| D | 5x8s4-minmax-wasmsimd-arm.c | 90 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm() local 181 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm() local
|
| D | 5x8s4-relu-wasmsimd.c | 88 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd() local 179 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_relu_ukernel_5x8s4__wasmsimd() local
|
| D | 5x8s4-minmax-wasmsimd-x86.c | 90 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86() local 181 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86() local
|
| D | 5x8s4-minmax-wasmrelaxedsimd-fma.c | 90 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local 181 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local
|
| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 6x2-minmax-neon-lane-ld64.c | 114 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64() local 139 const float32x2_t va4 = vld1_dup_f32(a4); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64() local
|
| D | 5x8s4-wasmrelaxedsimd-fma.c | 113 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local 204 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_ukernel_5x8s4__wasmrelaxedsimd_fma() local
|
| D | 5x8s4-relu-wasmsimd.c | 113 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd() local 204 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmsimd() local
|
| D | 6x2-minmax-neonfma-lane-ld64.c | 114 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64() local 169 const float32x2_t va4 = vld1_dup_f32(a4); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64() local
|
| D | 5x8s4-wasmsimd.c | 113 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd() local 204 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_ukernel_5x8s4__wasmsimd() local
|
| D | 5x8s4-relu-wasmrelaxedsimd-fma.c | 113 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma() local 204 v128_t va4 = wasm_v128_load(a4); in xnn_f32_igemm_relu_ukernel_5x8s4__wasmrelaxedsimd_fma() local
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 5x8inc-minmax-neonfma-lane-ld64.c | 86 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64() local 120 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1; in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64() local
|
| D | 5x8inc-minmax-neon-lane-ld64.c | 86 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64() local 120 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1; in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64() local
|
| D | 5x8s4inc-minmax-wasmsimd-x86.c | 92 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86() local 183 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86() local
|
| D | 5x8s4inc-minmax-wasmrelaxedsimd-fma.c | 92 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local 183 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma() local
|
| D | 6x8inc-minmax-neonfma-lane-ld64.c | 94 const float32x2_t va4 = vld1_f32(a4); a4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() local 133 const float32x4_t va4 = vld1q_dup_f32(a4); a4 += 1; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() local
|
| D | 5x8s4inc-minmax-wasmsimd-arm.c | 92 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm() local 183 v128_t va4 = wasm_v128_load(a4); in xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm() local
|