/external/XNNPACK/src/f32-spmm/gen/ |
D | 12x2-minmax-neonfma.c | 51 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local 101 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local 148 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local 187 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local 223 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local 252 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local
|
D | 16x2-minmax-neonfma.c | 53 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 113 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 165 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 204 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 240 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local 269 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local
|
D | 32x2-minmax-neonfma.c | 61 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 162 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 239 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 298 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 346 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 385 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 421 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local 450 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local
|
D | 8x2-minmax-neonfma.c | 49 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_8x2__neonfma() local 89 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_8x2__neonfma() local 129 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_8x2__neonfma() local 158 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_8x2__neonfma() local
|
D | 12x4-minmax-neonfma.c | 57 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local 133 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local 184 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local 241 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local 279 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local 318 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local
|
D | 16x4-minmax-neonfma.c | 61 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local 155 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local 211 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local 268 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local 306 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local 345 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local
|
D | 32x1-minmax-wasmsimd-arm-pipelined.c | 41 v128_t vi0123 = wasm_v128_load(input + 0); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local 133 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local 176 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local 208 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined() local
|
D | 32x1-minmax-neon.c | 53 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local 122 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local 164 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local 195 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neon() local
|
D | 32x1-minmax-wasmsimd-arm.c | 53 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local 119 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local 162 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local 194 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm() local
|
D | 32x1-minmax-neon-pipelined.c | 41 float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local 133 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local 178 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local 212 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() local
|
D | 32x1-minmax-wasmsimd-arm-pipelined-x2.c | 41 v128_t vi0123 = wasm_v128_load(input + 0); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local 173 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local 216 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local 248 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm_pipelined_x2() local
|
D | 32x1-minmax-neonfma.c | 53 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local 122 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local 164 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local 195 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() local
|
D | 32x1-minmax-wasmsimd-x86-pipelined.c | 41 v128_t vi0123 = wasm_v128_load(input + 0); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local 133 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local 176 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local 208 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined() local
|
D | 32x1-minmax-neonfma-pipelined.c | 41 float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local 133 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local 178 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local 212 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() local
|
D | 32x1-minmax-wasmsimd-x86.c | 53 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local 119 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local 162 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local 194 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86() local
|
D | 32x1-minmax-wasmsimd-x86-pipelined-x2.c | 41 v128_t vi0123 = wasm_v128_load(input + 0); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local 173 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local 216 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local 248 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86_pipelined_x2() local
|
D | 32x1-minmax-sse.c | 53 const __m128 vi0123 = _mm_loadu_ps(input); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local 119 const __m128 vi0123 = _mm_loadu_ps(input); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local 161 const __m128 vi0123 = _mm_loadu_ps(input); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local 192 const __m128 vi0123 = _mm_loadu_ps(input); in xnn_f32_spmm_minmax_ukernel_32x1__sse() local
|
D | 16x1-minmax-neon-pipelined.c | 41 float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined() local 102 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined() local 135 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x1__neon_pipelined() local
|
D | 16x1-minmax-neonfma.c | 49 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma() local 95 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma() local 126 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x1__neonfma() local
|
D | 16x1-minmax-wasmsimd-x86-pipelined.c | 41 v128_t vi0123 = wasm_v128_load(input + 0); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined() local 103 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined() local 135 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86_pipelined() local
|
D | 12x1-minmax-neonfma.c | 48 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma() local 89 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma() local 120 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x1__neonfma() local
|
D | 12x1-minmax-neon.c | 48 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x1__neon() local 89 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x1__neon() local 120 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_12x1__neon() local
|
D | 16x1-minmax-wasmsimd-arm-pipelined-x2.c | 41 v128_t vi0123 = wasm_v128_load(input + 0); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined_x2() local 127 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined_x2() local 159 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_arm_pipelined_x2() local
|
D | 16x1-minmax-wasmsimd-x86.c | 49 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86() local 93 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86() local 125 const v128_t vi0123 = wasm_v128_load(input); in xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86() local
|
D | 16x1-minmax-neon.c | 49 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x1__neon() local 95 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x1__neon() local 126 const float32x4_t vi0123 = vld1q_f32(input); in xnn_f32_spmm_minmax_ukernel_16x1__neon() local
|