| /external/XNNPACK/src/f32-spmm/gen/ |
| D | 32x2-minmax-neonfma.c | 63 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 164 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 241 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 300 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local
|
| D | 32x1-minmax-wasmsimd-arm.c | 55 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local 121 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local
|
| D | 32x1-minmax-wasmsimd-x86-pipelined.c | 43 v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local 135 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
|
| D | 32x1-minmax-neon.c | 55 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local 124 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local
|
| D | 32x1-minmax-neonfma-pipelined.c | 43 float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local 135 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local
|
| D | 32x1-minmax-wasmsimd-arm-pipelined-x2.c | 43 v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local 175 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local
|
| D | 32x1-minmax-wasmsimd-x86.c | 55 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local 121 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
|
| D | 32x1-minmax-neonfma.c | 55 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local 124 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local
|
| D | 32x1-minmax-wasmsimd-arm-pipelined.c | 43 v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local 135 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local
|
| D | 32x1-minmax-wasmsimd-x86-pipelined-x2.c | 43 v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local 175 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local
|
| D | 32x1-minmax-sse.c | 55 const __m128 vi89AB = _mm_loadu_ps(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local 121 const __m128 vi89AB = _mm_loadu_ps(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local
|
| D | 32x1-minmax-neon-pipelined.c | 43 float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local 135 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local
|
| D | 32x4-minmax-neonfma.c | 79 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 246 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 331 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local 424 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local
|
| D | 12x2-minmax-neonfma.c | 53 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local 103 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local
|
| D | 16x2-minmax-neonfma.c | 55 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 115 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local
|
| D | 32x1-minmax-wasmsimd-arm-x2.c | 123 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x2() local 189 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_x2() local
|
| D | 32x1-minmax-neon-x2.c | 126 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() local 195 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() local
|
| D | 32x1-minmax-neonfma-x2.c | 126 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_x2() local 195 const float32x4_t vi89AB = vld1q_f32(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_x2() local
|
| D | 32x1-minmax-wasmsimd-x86-x2.c | 123 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() local 189 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_x2() local
|
| /external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| D | sse2-rr2-p5-x12.c | 45 const __m128 vi89AB = _mm_loadu_ps(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x12() local
|
| D | neonfma-rr1-p5-x12.c | 43 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x12() local
|
| D | neon-rr2-p5-x12-acc2.c | 45 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_p5_x12_acc2() local
|
| D | wasmsimd-rr2-p5-x12.c | 45 const v128_t vi89AB = wasm_v128_load(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x12() local
|
| D | neonfma-rr1-p5-x12-acc2.c | 44 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x12_acc2() local
|
| D | neonfma-rr1-p5-x12-acc3.c | 45 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_p5_x12_acc3() local
|