/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | psimd-p5-x8.c | 47 const psimd_f32 vi4567 = psimd_load_f32(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8() local 52 const psimd_f32 vx4567 = psimd_sub_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8()
|
D | psimd-p5-x8-acc2.c | 48 const psimd_f32 vi4567 = psimd_load_f32(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2() local 53 const psimd_f32 vx4567 = psimd_sub_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x8_acc2()
|
D | neon-p5-x8-acc2.c | 48 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() local 52 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2()
|
D | sse2-p5-x8.c | 47 const __m128 vi4567 = _mm_loadu_ps(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8() local 52 const __m128 vx4567 = _mm_sub_ps(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8()
|
D | neonfma-p5-x8-acc2.c | 47 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() local 51 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2()
|
D | neon-p5-x8.c | 47 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() local 51 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8()
|
D | sse2-p5-x8-acc2.c | 48 const __m128 vi4567 = _mm_loadu_ps(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2() local 53 const __m128 vx4567 = _mm_sub_ps(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2()
|
D | neonfma-p5-x8.c | 46 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() local 50 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8()
|
D | psimd-p5-x12.c | 47 const psimd_f32 vi4567 = psimd_load_f32(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() local 53 const psimd_f32 vx4567 = psimd_sub_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
|
D | psimd-p5-x12-acc2.c | 48 const psimd_f32 vi4567 = psimd_load_f32(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() local 54 const psimd_f32 vx4567 = psimd_sub_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
|
D | psimd-p5-x12-acc3.c | 49 const psimd_f32 vi4567 = psimd_load_f32(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() local 55 const psimd_f32 vx4567 = psimd_sub_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
|
D | sse2-p5-x12.c | 47 const __m128 vi4567 = _mm_loadu_ps(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() local 53 const __m128 vx4567 = _mm_sub_ps(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
|
D | neon-p5-x12.c | 47 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local 52 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
|
D | neonfma-p5-x12.c | 46 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local 51 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
|
D | neonfma-p5-x12-acc2.c | 47 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local 52 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
|
D | neon-p5-x12-acc2.c | 48 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() local 53 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
|
D | psimd-p5-x16-acc2.c | 48 const psimd_f32 vi4567 = psimd_load_f32(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2() local 55 const psimd_f32 vx4567 = psimd_sub_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
|
D | neon-lut64-p2-x8-acc2.c | 48 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2() local 52 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2()
|
D | neonfma-lut64-p2-x8-acc2.c | 47 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2() local 51 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2()
|
D | sse2-p5-x12-acc3.c | 49 const __m128 vi4567 = _mm_loadu_ps(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3() local 55 const __m128 vx4567 = _mm_sub_ps(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
|
D | neon-lut64-p2-x8.c | 47 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8() local 51 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8()
|
D | psimd-p5-x16.c | 47 const psimd_f32 vi4567 = psimd_load_f32(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16() local 54 const psimd_f32 vx4567 = psimd_sub_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
|
D | psimd-p5-x16-acc4.c | 50 const psimd_f32 vi4567 = psimd_load_f32(input + 4); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4() local 57 const psimd_f32 vx4567 = psimd_sub_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
|
D | neonfma-p5-x12-acc3.c | 48 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3() local 53 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
|
D | neonfma-lut64-p2-x8.c | 46 const float32x4_t vi4567 = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8() local 50 const float32x4_t vx4567 = vsubq_f32(vi4567, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8()
|