/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x9-wasmsimd.c | 154 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 226 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 275 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
|
D | up8x9-minmax-wasmsimd-x86-acc2.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 234 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 286 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
|
D | up8x9-minmax-neon-acc2.c | 141 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 198 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 246 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
|
D | up8x9-minmax-neonfma.c | 141 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 195 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 241 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
|
D | up8x9-minmax-sse.c | 156 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 230 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 280 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
|
D | up8x9-minmax-sse-acc2.c | 156 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 233 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 285 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
|
D | up8x9-minmax-wasmsimd-arm.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 231 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 281 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
|
D | up8x9-minmax-neon.c | 141 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 195 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 241 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
|
D | up8x9-minmax-wasmsimd-arm-acc2.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 234 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 286 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
|
D | up8x9-minmax-neonfma-acc2.c | 141 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 198 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 246 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
|
D | up8x9-minmax-wasmsimd-x86.c | 156 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 231 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 281 const v128_t vk7x0123 = wasm_v128_load(w + 64); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
|
D | up4x9-wasmsimd.c | 131 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local 180 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
|
D | up4x9-minmax-wasmsimd-x86.c | 133 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local 184 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
|
D | up4x9-minmax-wasmsimd-arm.c | 133 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local 184 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
|
D | up4x9-minmax-sse-acc2.c | 133 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local 185 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
|
D | up4x9-minmax-sse.c | 133 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local 183 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
|
D | up4x9-minmax-neon-acc2.c | 118 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local 166 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
|
D | up4x9-minmax-neon.c | 118 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local 164 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
|
D | up4x9-minmax-neonfma.c | 118 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local 164 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
|
D | up4x9-minmax-wasmsimd-arm-acc2.c | 133 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local 186 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
|
D | up4x9-minmax-wasmsimd-x86-acc2.c | 133 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local 186 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
|
D | up4x9-minmax-neonfma-acc2.c | 118 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local 166 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
|
D | up4x25-wasmsimd.c | 211 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 356 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-x86.c | 213 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local 360 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
|
D | up4x25-minmax-wasmsimd-x86-acc2.c | 213 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local 362 const v128_t vk7x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
|