/aosp_15_r20/external/XNNPACK/src/f32-ibilinear/gen/ |
H A D | neonfma-c8.c | 62 const float32x4_t vb0123 = vfmaq_f32(vbl0123, vbd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c8() local 67 const float32x4_t vb0123 = vfmaq_lane_f32(vbl0123, vbd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c8() local 97 const float32x4_t vb0123 = vfmaq_f32(vbl0123, vbd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c8() local 100 const float32x4_t vb0123 = vfmaq_lane_f32(vbl0123, vbd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c8() local 125 const float32x4_t vb0123 = vfmaq_f32(vbl0123, vbd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c8() local 128 const float32x4_t vb0123 = vfmaq_lane_f32(vbl0123, vbd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c8() local
|
H A D | neonfma-c4.c | 56 const float32x4_t vb0123 = vfmaq_f32(vbl0123, vbd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c4() local 59 const float32x4_t vb0123 = vfmaq_lane_f32(vbl0123, vbd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c4() local 83 const float32x4_t vb0123 = vfmaq_f32(vbl0123, vbd0123, valphah); in xnn_f32_ibilinear_ukernel__neonfma_c4() local 86 const float32x4_t vb0123 = vfmaq_lane_f32(vbl0123, vbd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neonfma_c4() local
|
H A D | neon-c8.c | 57 const float32x4_t vb0123 = vmlaq_lane_f32(vbl0123, vbd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neon_c8() local 80 const float32x4_t vb0123 = vmlaq_lane_f32(vbl0123, vbd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neon_c8() local 99 const float32x4_t vb0123 = vmlaq_lane_f32(vbl0123, vbd0123, valphahv, 0); in xnn_f32_ibilinear_ukernel__neon_c8() local
|
H A D | sse-c8.c | 64 const __m128 vb0123 = _mm_add_ps(vbl0123, _mm_mul_ps(vbd0123, valphah)); in xnn_f32_ibilinear_ukernel__sse_c8() local 92 const __m128 vb0123 = _mm_add_ps(vbl0123, _mm_mul_ps(vbd0123, valphah)); in xnn_f32_ibilinear_ukernel__sse_c8() local 111 const __m128 vb0123 = _mm_add_ps(vbl0123, _mm_mul_ps(vbd0123, valphah)); in xnn_f32_ibilinear_ukernel__sse_c8() local
|
/aosp_15_r20/external/XNNPACK/src/f32-vbinary/gen/ |
H A D | vsqrdiff-sse-x8.c | 38 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vsqrdiff_ukernel__sse_x8() local 57 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vsqrdiff_ukernel__sse_x8() local 67 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vsqrdiff_ukernel__sse_x8() local
|
H A D | vmin-sse-x8.c | 38 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmin_ukernel__sse_x8() local 55 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmin_ukernel__sse_x8() local 64 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmin_ukernel__sse_x8() local
|
H A D | vmax-sse-x8.c | 38 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmax_ukernel__sse_x8() local 55 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmax_ukernel__sse_x8() local 64 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmax_ukernel__sse_x8() local
|
H A D | vsqrdiff-neon-x8.c | 34 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vsqrdiff_ukernel__neon_x8() local 50 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vsqrdiff_ukernel__neon_x8() local 58 const float32x4_t vb0123 = vld1q_f32(b); in xnn_f32_vsqrdiff_ukernel__neon_x8() local
|
H A D | vmax-neon-x8.c | 34 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vmax_ukernel__neon_x8() local 48 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vmax_ukernel__neon_x8() local 55 const float32x4_t vb0123 = vld1q_f32(b); in xnn_f32_vmax_ukernel__neon_x8() local
|
H A D | vmin-neon-x8.c | 34 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vmin_ukernel__neon_x8() local 48 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vmin_ukernel__neon_x8() local 55 const float32x4_t vb0123 = vld1q_f32(b); in xnn_f32_vmin_ukernel__neon_x8() local
|
H A D | vadd-minmax-sse-x8.c | 40 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vadd_minmax_ukernel__sse_x8() local 62 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vadd_minmax_ukernel__sse_x8() local 73 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vadd_minmax_ukernel__sse_x8() local
|
H A D | vmul-minmax-sse-x8.c | 40 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmul_minmax_ukernel__sse_x8() local 62 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmul_minmax_ukernel__sse_x8() local 73 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmul_minmax_ukernel__sse_x8() local
|
H A D | vsub-minmax-sse-x8.c | 40 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vsub_minmax_ukernel__sse_x8() local 62 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vsub_minmax_ukernel__sse_x8() local 73 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vsub_minmax_ukernel__sse_x8() local
|
H A D | vdiv-minmax-sse-x8.c | 40 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vdiv_minmax_ukernel__sse_x8() local 62 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vdiv_minmax_ukernel__sse_x8() local 73 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vdiv_minmax_ukernel__sse_x8() local
|
H A D | vsub-minmax-neon-x8.c | 36 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vsub_minmax_ukernel__neon_x8() local 55 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vsub_minmax_ukernel__neon_x8() local 64 const float32x4_t vb0123 = vld1q_f32(b); in xnn_f32_vsub_minmax_ukernel__neon_x8() local
|
H A D | vadd-minmax-neon-x8.c | 36 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vadd_minmax_ukernel__neon_x8() local 55 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vadd_minmax_ukernel__neon_x8() local 64 const float32x4_t vb0123 = vld1q_f32(b); in xnn_f32_vadd_minmax_ukernel__neon_x8() local
|
H A D | vdiv-minmax-neon-x8.c | 36 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vdiv_minmax_ukernel__neon_x8() local 55 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vdiv_minmax_ukernel__neon_x8() local 64 const float32x4_t vb0123 = vld1q_f32(b); in xnn_f32_vdiv_minmax_ukernel__neon_x8() local
|
H A D | vmul-minmax-neon-x8.c | 36 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vmul_minmax_ukernel__neon_x8() local 55 const float32x4_t vb0123 = vld1q_f32(b); b += 4; in xnn_f32_vmul_minmax_ukernel__neon_x8() local 64 const float32x4_t vb0123 = vld1q_f32(b); in xnn_f32_vmul_minmax_ukernel__neon_x8() local
|
H A D | vmax-sse-x4.c | 37 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmax_ukernel__sse_x4() local 49 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmax_ukernel__sse_x4() local
|
H A D | vmin-sse-x4.c | 37 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmin_ukernel__sse_x4() local 49 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vmin_ukernel__sse_x4() local
|
H A D | vsqrdiff-sse-x4.c | 37 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vsqrdiff_ukernel__sse_x4() local 50 const __m128 vb0123 = _mm_loadu_ps(b); in xnn_f32_vsqrdiff_ukernel__sse_x4() local
|
/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/sgemm/ |
H A D | 5x8-neon.c | 68 const float32x4_t vb0123 = vld1q_f32(w); in pytorch_sgemm_ukernel_5x8__neon() local 99 const float32x4_t vb0123 = vld1q_f32(w); in pytorch_sgemm_ukernel_5x8__neon() local 136 const float32x4_t vb0123 = vld1q_f32(w); in pytorch_sgemm_ukernel_5x8__neon() local
|
H A D | 6x8-neon.c | 76 const float32x4_t vb0123 = vld1q_f32(w); in pytorch_sgemm_ukernel_6x8__neon() local 111 const float32x4_t vb0123 = vld1q_f32(w); in pytorch_sgemm_ukernel_6x8__neon() local 153 const float32x4_t vb0123 = vld1q_f32(w); in pytorch_sgemm_ukernel_6x8__neon() local
|
/aosp_15_r20/external/XNNPACK/src/f32-vbinary/ |
H A D | vop-sse.c.in | 83 const __m128 vb0123 = _mm_loadu_ps(b); variable 97 const __m128 vb0123 = _mm_loadu_ps(b); variable
|
H A D | vop-neon.c.in | 72 const float32x4_t vb0123 = vld1q_f32(b); b += 4; variable 84 const float32x4_t vb0123 = vld1q_f32(b); variable
|