/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x4-minmax-wasmsimd-arm.c | 95 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local 137 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local 165 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local
|
D | up8x4-minmax-wasmsimd-x86.c | 95 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local 137 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local 165 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local
|
D | up8x4-minmax-sse.c | 95 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local 136 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local 164 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local
|
D | up8x4-wasmsimd.c | 93 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local 132 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local 159 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local
|
D | up8x4-minmax-neon.c | 88 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local 119 const float32x4_t vk3x0123 = vld1q_f32(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local 145 const float32x4_t vk3x0123 = vld1q_f32(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local
|
D | up8x4-minmax-sse-acc2.c | 95 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local 139 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local 169 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
|
D | up8x4-minmax-neonfma.c | 88 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local 119 const float32x4_t vk3x0123 = vld1q_f32(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local 145 const float32x4_t vk3x0123 = vld1q_f32(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local
|
D | up8x4-minmax-wasmsimd-arm-acc2.c | 95 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local 140 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local 170 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
|
D | up8x4-minmax-wasmsimd-x86-acc2.c | 95 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local 140 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local 170 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
|
D | up8x4-minmax-neonfma-acc2.c | 88 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local 122 const float32x4_t vk3x0123 = vld1q_f32(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local 150 const float32x4_t vk3x0123 = vld1q_f32(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
|
D | up8x4-minmax-neon-acc2.c | 88 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local 122 const float32x4_t vk3x0123 = vld1q_f32(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local 150 const float32x4_t vk3x0123 = vld1q_f32(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
|
D | up4x4-wasmsimd.c | 82 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up4x4__wasmsimd() local 109 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_ukernel_up4x4__wasmsimd() local
|
D | up4x4-minmax-wasmsimd-arm.c | 84 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm() local 113 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm() local
|
D | up4x4-minmax-wasmsimd-x86.c | 84 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86() local 113 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86() local
|
D | up4x4-minmax-sse.c | 84 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse() local 112 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse() local
|
D | up4x4-minmax-neon.c | 77 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon() local 103 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon() local
|
D | up4x4-minmax-sse-acc2.c | 84 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2() local 114 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2() local
|
D | up4x4-minmax-neonfma.c | 77 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma() local 103 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma() local
|
D | up4x4-minmax-wasmsimd-arm-acc2.c | 84 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm_acc2() local 115 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm_acc2() local
|
D | up4x4-minmax-wasmsimd-x86-acc2.c | 84 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86_acc2() local 115 const v128_t vk3x0123 = wasm_v128_load(w + 16); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86_acc2() local
|
D | up4x4-minmax-neonfma-acc2.c | 77 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2() local 105 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2() local
|
D | up4x4-minmax-neon-acc2.c | 77 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2() local 105 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2() local
|
D | up8x9-wasmsimd.c | 118 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 202 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 259 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
|
D | up8x9-minmax-wasmsimd-x86-acc2.c | 120 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 210 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local 270 const v128_t vk3x0123 = wasm_v128_load(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
|
D | up8x9-minmax-neon-acc2.c | 113 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 182 const float32x4_t vk3x0123 = vld1q_f32(w + 28); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local 230 const float32x4_t vk3x0123 = vld1q_f32(w + 32); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
|