/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x9-wasmsimd.c | 127 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 208 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 263 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
|
D | up8x9-wasmrelaxedsimd-fma.c | 127 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma() local 208 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma() local 263 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma() local
|
D | up8x9-minmax-sse.c | 129 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 212 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 268 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
|
D | up8x9-minmax-neonfma-acc2.c | 120 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 186 const float32x4_t vk4x0123 = vld1q_f32(w + 36); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 234 const float32x4_t vk4x0123 = vld1q_f32(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
|
D | up8x9-minmax-wasmrelaxedsimd-acc2.c | 129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2() local 216 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2() local 274 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_acc2() local
|
D | up8x9-minmax-wasmrelaxedsimd-fma.c | 129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma() local 213 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma() local 269 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma() local
|
D | up8x9-minmax-wasmsimd-x86.c | 129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 213 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 269 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
|
D | up8x9-minmax-wasmsimd-x86-acc2.c | 129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 216 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 274 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
|
D | up8x9-minmax-neon.c | 120 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 183 const float32x4_t vk4x0123 = vld1q_f32(w + 36); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 229 const float32x4_t vk4x0123 = vld1q_f32(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
|
D | up8x9-minmax-wasmsimd-arm.c | 129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 213 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 269 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
|
D | up8x9-minmax-wasmrelaxedsimd.c | 129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd() local 213 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd() local 269 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd() local
|
D | up8x9-wasmsimd-acc2.c | 127 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2() local 211 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2() local 268 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd_acc2() local
|
D | up8x9-minmax-wasmsimd-arm-acc2.c | 129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 216 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 274 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
|
D | up8x9-minmax-neon-acc2.c | 120 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 186 const float32x4_t vk4x0123 = vld1q_f32(w + 36); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 234 const float32x4_t vk4x0123 = vld1q_f32(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
|
D | up8x9-minmax-wasmrelaxedsimd-fma-acc2.c | 129 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2() local 216 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2() local 274 const v128_t vk4x0123 = wasm_v128_load(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma_acc2() local
|
D | up8x9-minmax-sse-acc2.c | 129 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 215 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 273 const __m128 vk4x0123 = _mm_load_ps(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
|
D | up8x9-minmax-neonfma.c | 120 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 183 const float32x4_t vk4x0123 = vld1q_f32(w + 36); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 229 const float32x4_t vk4x0123 = vld1q_f32(w + 40); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
|
D | up4x9-wasmsimd.c | 113 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local 168 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
|
D | up4x9-wasmrelaxedsimd-fma.c | 113 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_ukernel_up4x9__wasmrelaxedsimd_fma() local 168 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_ukernel_up4x9__wasmrelaxedsimd_fma() local
|
D | up4x9-minmax-neonfma-acc2.c | 106 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local 154 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
|
D | up4x9-minmax-neonfma.c | 106 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local 152 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
|
D | up4x9-minmax-neon-acc2.c | 106 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local 154 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
|
D | up4x9-minmax-neon.c | 106 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local 152 const float32x4_t vk4x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
|
D | up4x9-minmax-sse.c | 115 const __m128 vk4x0123 = _mm_load_ps(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local 171 const __m128 vk4x0123 = _mm_load_ps(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
|
D | up4x9-minmax-wasmsimd-arm.c | 115 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local 172 const v128_t vk4x0123 = wasm_v128_load(w + 20); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
|