/external/XNNPACK/src/f32-ibilinear-chw/gen/ |
D | neonfma-p4.c | 50 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p4() local 62 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p4() 102 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p4() local 106 const float32x2_t vldrd0 = vsub_f32(vblbr0, vtltr0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p4()
|
D | neon-p4.c | 50 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neon_p4() local 62 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neon_p4() 102 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neon_p4() local 106 const float32x2_t vldrd0 = vsub_f32(vblbr0, vtltr0); in xnn_f32_ibilinear_chw_ukernel__neon_p4()
|
D | neonfma-p8.c | 58 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8() local 80 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8() 138 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8() local 150 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8() 190 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8() local 194 const float32x2_t vldrd0 = vsub_f32(vblbr0, vtltr0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p8()
|
D | neon-p8.c | 58 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neon_p8() local 80 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neon_p8() 138 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neon_p8() local 150 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neon_p8() 190 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neon_p8() local 194 const float32x2_t vldrd0 = vsub_f32(vblbr0, vtltr0); in xnn_f32_ibilinear_chw_ukernel__neon_p8()
|
D | neonfma-p16.c | 76 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p16() local 118 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p16() 210 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p16() local 222 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neonfma_p16() 262 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p16() local 266 const float32x2_t vldrd0 = vsub_f32(vblbr0, vtltr0); in xnn_f32_ibilinear_chw_ukernel__neonfma_p16()
|
D | neon-p16.c | 76 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neon_p16() local 118 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neon_p16() 210 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neon_p16() local 222 const float32x4_t vblbr01 = vcombine_f32(vblbr0, vblbr1); in xnn_f32_ibilinear_chw_ukernel__neon_p16() 262 const float32x2_t vblbr0 = vld1_f32(ibl0); in xnn_f32_ibilinear_chw_ukernel__neon_p16() local 266 const float32x2_t vldrd0 = vsub_f32(vblbr0, vtltr0); in xnn_f32_ibilinear_chw_ukernel__neon_p16()
|
D | wasmsimd-p8.c | 60 const v128_t vblbr0 = wasm_v128_load64_splat(ibl0); in xnn_f32_ibilinear_chw_ukernel__wasmsimd_p8() local 82 const v128_t vblbr01 = wasm_f64x2_replace_lane(vblbr0, 1, vblbr1); in xnn_f32_ibilinear_chw_ukernel__wasmsimd_p8() 137 const v128_t vblbr0 = wasm_v128_load64_splat(ibl0); in xnn_f32_ibilinear_chw_ukernel__wasmsimd_p8() local 149 const v128_t vblbr01 = wasm_f64x2_replace_lane(vblbr0, 1, vblbr1); in xnn_f32_ibilinear_chw_ukernel__wasmsimd_p8()
|
D | sse-p8.c | 60 const __m128 vblbr0 = _mm_loadl_pi(_mm_undefined_ps(), (const __m64*) ibl0); in xnn_f32_ibilinear_chw_ukernel__sse_p8() local 74 const __m128 vblbr01 = _mm_loadh_pi(vblbr0, (const __m64*) ibl1); in xnn_f32_ibilinear_chw_ukernel__sse_p8() 129 const __m128 vblbr0 = _mm_loadl_pi(_mm_undefined_ps(), (const __m64*) ibl0); in xnn_f32_ibilinear_chw_ukernel__sse_p8() local 137 const __m128 vblbr01 = _mm_loadh_pi(vblbr0, (const __m64*) ibl1); in xnn_f32_ibilinear_chw_ukernel__sse_p8()
|
D | wasmsimd-p4.c | 51 const v128_t vblbr0 = wasm_v128_load64_splat(ibl0); in xnn_f32_ibilinear_chw_ukernel__wasmsimd_p4() local 63 const v128_t vblbr01 = wasm_f64x2_replace_lane(vblbr0, 1, vblbr1); in xnn_f32_ibilinear_chw_ukernel__wasmsimd_p4()
|
D | sse-p4.c | 51 const __m128 vblbr0 = _mm_loadl_pi(_mm_undefined_ps(), (const __m64*) ibl0); in xnn_f32_ibilinear_chw_ukernel__sse_p4() local 59 const __m128 vblbr01 = _mm_loadh_pi(vblbr0, (const __m64*) ibl1); in xnn_f32_ibilinear_chw_ukernel__sse_p4()
|
/external/XNNPACK/src/amalgam/ |
D | sse.c | 5019 const __m128 vblbr0 = _mm_loadl_pi(_mm_undefined_ps(), (const __m64*) ibl0); in xnn_f32_ibilinear_chw_ukernel__sse_p8() local 5033 const __m128 vblbr01 = _mm_loadh_pi(vblbr0, (const __m64*) ibl1); in xnn_f32_ibilinear_chw_ukernel__sse_p8() 5088 const __m128 vblbr0 = _mm_loadl_pi(_mm_undefined_ps(), (const __m64*) ibl0); in xnn_f32_ibilinear_chw_ukernel__sse_p8() local 5096 const __m128 vblbr01 = _mm_loadh_pi(vblbr0, (const __m64*) ibl1); in xnn_f32_ibilinear_chw_ukernel__sse_p8()
|