/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | psimd-p5-x12.c | 48 const psimd_f32 vi89AB = psimd_load_f32(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12() local 54 const psimd_f32 vx89AB = psimd_sub_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12()
|
D | psimd-p5-x12-acc2.c | 49 const psimd_f32 vi89AB = psimd_load_f32(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2() local 55 const psimd_f32 vx89AB = psimd_sub_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc2()
|
D | psimd-p5-x12-acc3.c | 50 const psimd_f32 vi89AB = psimd_load_f32(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3() local 56 const psimd_f32 vx89AB = psimd_sub_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x12_acc3()
|
D | sse2-p5-x12.c | 48 const __m128 vi89AB = _mm_loadu_ps(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12() local 54 const __m128 vx89AB = _mm_sub_ps(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12()
|
D | neon-p5-x12.c | 48 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local 53 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12()
|
D | neonfma-p5-x12.c | 47 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local 52 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12()
|
D | neonfma-p5-x12-acc2.c | 48 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local 53 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2()
|
D | neon-p5-x12-acc2.c | 49 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() local 54 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2()
|
D | psimd-p5-x16-acc2.c | 49 const psimd_f32 vi89AB = psimd_load_f32(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2() local 56 const psimd_f32 vx89AB = psimd_sub_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc2()
|
D | sse2-p5-x12-acc3.c | 50 const __m128 vi89AB = _mm_loadu_ps(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3() local 56 const __m128 vx89AB = _mm_sub_ps(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3()
|
D | psimd-p5-x16.c | 48 const psimd_f32 vi89AB = psimd_load_f32(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16() local 55 const psimd_f32 vx89AB = psimd_sub_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16()
|
D | psimd-p5-x16-acc4.c | 51 const psimd_f32 vi89AB = psimd_load_f32(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4() local 58 const psimd_f32 vx89AB = psimd_sub_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x16_acc4()
|
D | neonfma-p5-x12-acc3.c | 49 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3() local 54 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3()
|
D | neon-p5-x12-acc3.c | 50 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3() local 55 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3()
|
D | sse2-p5-x12-acc2.c | 49 const __m128 vi89AB = _mm_loadu_ps(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2() local 55 const __m128 vx89AB = _mm_sub_ps(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2()
|
D | neonfma-p5-x16-acc4.c | 50 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4() local 56 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4()
|
D | psimd-p5-x20-acc2.c | 49 const psimd_f32 vi89AB = psimd_load_f32(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2() local 57 const psimd_f32 vx89AB = psimd_sub_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__psimd_p5_x20_acc2()
|
D | neon-p5-x16-acc2.c | 49 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2() local 55 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2()
|
D | sse2-p5-x16.c | 48 const __m128 vi89AB = _mm_loadu_ps(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16() local 55 const __m128 vx89AB = _mm_sub_ps(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16()
|
D | neon-p5-x16.c | 48 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16() local 54 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16()
|
D | neonfma-p5-x16.c | 47 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16() local 53 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16()
|
D | sse2-p5-x16-acc4.c | 51 const __m128 vi89AB = _mm_loadu_ps(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4() local 58 const __m128 vx89AB = _mm_sub_ps(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4()
|
D | neonfma-p5-x16-acc2.c | 48 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2() local 54 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2()
|
D | sse2-p5-x16-acc2.c | 49 const __m128 vi89AB = _mm_loadu_ps(input + 8); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2() local 56 const __m128 vx89AB = _mm_sub_ps(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2()
|
D | neon-p5-x16-acc4.c | 51 const float32x4_t vi89AB = vld1q_f32(input); input += 4; in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4() local 57 const float32x4_t vx89AB = vsubq_f32(vi89AB, vi_max); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4()
|