| /external/XNNPACK/src/f32-vmulcaddc/gen/ |
| D | c8-psimd-2x.c | 53 psimd_f32 vacc0x0123 = psimd_load_f32(i0); in xnn_f32_vmulcaddc_ukernel_c8__psimd_2x() local
90 psimd_f32 vacc0x0123 = psimd_load_f32(i0); in xnn_f32_vmulcaddc_ukernel_c8__psimd_2x() local
116 psimd_f32 vacc0x0123 = psimd_load_f32(i0); in xnn_f32_vmulcaddc_ukernel_c8__psimd_2x() local
|
| D | c8-sse-2x.c | 53 __m128 vacc0x0123 = _mm_loadu_ps(i0); in xnn_f32_vmulcaddc_ukernel_c8__sse_2x() local
95 __m128 vacc0x0123 = _mm_loadu_ps(i0); in xnn_f32_vmulcaddc_ukernel_c8__sse_2x() local
124 __m128 vacc0x0123 = _mm_loadu_ps(i0); in xnn_f32_vmulcaddc_ukernel_c8__sse_2x() local
|
| D | c8-neon-2x.c | 53 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_vmulcaddc_ukernel_c8__neon_2x() local
89 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_vmulcaddc_ukernel_c8__neon_2x() local
112 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 = (const float*) ((uintptr_t) i0 + c); in xnn_f32_vmulcaddc_ukernel_c8__neon_2x() local
|
| D | c8-neonfma-2x.c | 53 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_vmulcaddc_ukernel_c8__neonfma_2x() local
85 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_vmulcaddc_ukernel_c8__neonfma_2x() local
106 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 = (const float*) ((uintptr_t) i0 + c); in xnn_f32_vmulcaddc_ukernel_c8__neonfma_2x() local
|
| D | c4-psimd-2x.c | 52 psimd_f32 vacc0x0123 = psimd_load_f32(i0); in xnn_f32_vmulcaddc_ukernel_c4__psimd_2x() local
78 psimd_f32 vacc0x0123 = psimd_load_f32(i0); in xnn_f32_vmulcaddc_ukernel_c4__psimd_2x() local
|
| D | c4-sse-2x.c | 52 __m128 vacc0x0123 = _mm_loadu_ps(i0); in xnn_f32_vmulcaddc_ukernel_c4__sse_2x() local
81 __m128 vacc0x0123 = _mm_loadu_ps(i0); in xnn_f32_vmulcaddc_ukernel_c4__sse_2x() local
|
| D | c4-neon-2x.c | 52 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_vmulcaddc_ukernel_c4__neon_2x() local
75 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 = (const float*) ((uintptr_t) i0 + c); in xnn_f32_vmulcaddc_ukernel_c4__neon_2x() local
|
| D | c4-neonfma-2x.c | 52 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_vmulcaddc_ukernel_c4__neonfma_2x() local
73 float32x4_t vacc0x0123 = vld1q_f32(i0); i0 = (const float*) ((uintptr_t) i0 + c); in xnn_f32_vmulcaddc_ukernel_c4__neonfma_2x() local
|
| /external/XNNPACK/src/f32-prelu/gen/ |
| D | psimd-2x8.c | 61 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x8() local
97 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x8() local
123 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x8() local
|
| D | sse41-2x8.c | 66 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x8() local
100 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x8() local
126 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x8() local
|
| D | neon-2x8.c | 58 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x8() local
95 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x8() local
120 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x8() local
|
| D | sse2-2x8.c | 70 …__m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0… in xnn_f32_prelu_ukernel__sse2_2x8() local
106 …__m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0… in xnn_f32_prelu_ukernel__sse2_2x8() local
134 …__m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0… in xnn_f32_prelu_ukernel__sse2_2x8() local
|
| D | psimd-2x4.c | 58 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x4() local
84 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x4() local
|
| D | sse41-2x4.c | 61 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x4() local
87 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x4() local
|
| D | neon-2x4.c | 55 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x4() local
80 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x4() local
|
| D | sse2-2x4.c | 63 …__m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0… in xnn_f32_prelu_ukernel__sse2_2x4() local
91 …__m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0… in xnn_f32_prelu_ukernel__sse2_2x4() local
|
| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 1x8-avx-broadcast.c | 72 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567); in xnn_f32_gemm_ukernel_1x8__avx_broadcast() local
|
| D | 1x8-fma3-broadcast.c | 72 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567); in xnn_f32_gemm_ukernel_1x8__fma3_broadcast() local
|
| D | 1x8-psimd-loadsplat.c | 42 psimd_f32 vacc0x0123 = psimd_load_f32(w + 0); in xnn_f32_gemm_ukernel_1x8__psimd_loadsplat() local
|
| D | 1x8-sse-load1.c | 42 __m128 vacc0x0123 = _mm_load_ps(w + 0); in xnn_f32_gemm_ukernel_1x8__sse_load1() local
|
| D | 1x16-fma3-broadcast.c | 85 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567); in xnn_f32_gemm_ukernel_1x16__fma3_broadcast() local
|
| D | 1x16-avx-broadcast.c | 85 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567); in xnn_f32_gemm_ukernel_1x16__avx_broadcast() local
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 1x8-fma3-broadcast.c | 74 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567); in xnn_f32_gemminc_ukernel_1x8__fma3_broadcast() local
|
| D | 1x8-avx-broadcast.c | 74 __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567); in xnn_f32_gemminc_ukernel_1x8__avx_broadcast() local
|
| D | 1x8-psimd-loadsplat.c | 44 psimd_f32 vacc0x0123 = psimd_load_f32(acc + 0); in xnn_f32_gemminc_ukernel_1x8__psimd_loadsplat() local
|