/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x9-wasmsimd.c | 136 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 214 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 267 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
|
D | up4x9-wasmsimd.c | 119 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local 172 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
|
D | up8x9-minmax-sse.c | 138 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 218 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 272 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
|
D | up8x9-minmax-neon.c | 127 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 187 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local 233 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
|
D | up8x9-minmax-sse-acc2.c | 138 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 221 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local 277 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
|
D | up8x9-minmax-neonfma.c | 127 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 187 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local 233 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
|
D | up8x9-minmax-wasmsimd-x86-acc2.c | 138 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 222 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 278 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
|
D | up8x9-minmax-wasmsimd-arm-acc2.c | 138 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 222 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 278 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
|
D | up8x9-minmax-wasmsimd-x86.c | 138 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 219 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local 273 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
|
D | up8x9-minmax-neonfma-acc2.c | 127 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 190 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local 238 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
|
D | up8x9-minmax-wasmsimd-arm.c | 138 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 219 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local 273 const v128_t vk5x0123 = wasm_v128_load(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
|
D | up8x9-minmax-neon-acc2.c | 127 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 190 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 238 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
|
D | up4x9-minmax-wasmsimd-arm.c | 121 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local 176 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
|
D | up4x9-minmax-wasmsimd-x86-acc2.c | 121 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local 178 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
|
D | up4x9-minmax-sse.c | 121 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local 175 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
|
D | up4x9-minmax-neon-acc2.c | 110 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local 158 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
|
D | up4x9-minmax-neonfma.c | 110 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local 156 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
|
D | up4x9-minmax-sse-acc2.c | 121 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local 177 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
|
D | up4x9-minmax-wasmsimd-x86.c | 121 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local 176 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
|
D | up4x9-minmax-neonfma-acc2.c | 110 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local 158 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
|
D | up4x9-minmax-neon.c | 110 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local 156 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
|
D | up4x9-minmax-wasmsimd-arm-acc2.c | 121 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local 178 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
|
D | up4x25-wasmsimd.c | 199 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local 348 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
|
D | up4x25-minmax-wasmsimd-arm-acc2.c | 201 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local 354 const v128_t vk5x0123 = wasm_v128_load(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
|
D | up4x25-minmax-sse.c | 201 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local 351 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
|