| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7x-minmax-fp32-sse41-c8.c | 94 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8() local
97 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
100 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
103 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
147 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8() local
150 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
153 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
156 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
|
| D | 7x-minmax-fp32-sse2-c8.c | 108 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8() local
111 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
114 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
117 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
177 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8() local
180 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
183 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
186 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
|
| D | 7x-minmax-fp32-sse41-c16.c | 111 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16() local
116 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
121 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
126 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
173 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16() local
176 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
179 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
182 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
|
| /external/XNNPACK/src/qu8-gavgpool/gen/ |
| D | 7x-minmax-fp32-sse41-c8.c | 95 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8() local
98 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
101 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
104 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
148 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8() local
151 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
154 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
157 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8()
|
| D | 7x-minmax-fp32-sse2-c8.c | 108 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8() local
111 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
114 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
117 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
176 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8() local
179 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
182 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
185 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8()
|
| D | 7x-minmax-fp32-sse41-c16.c | 112 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16() local
117 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
122 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
127 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
174 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16() local
177 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
180 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
183 vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16()
|
| /external/XNNPACK/src/qu8-vmulc/gen/ |
| D | minmax-fp32-sse41-mul16-ld64-x8.c | 49 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
52 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
55 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
83 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
86 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
89 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
|
| D | minmax-fp32-neonv8-ld64-x8.c | 43 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
46 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
49 vacc4567 = vcvtnq_s32_f32(vfpacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
81 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
84 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
87 vacc4567 = vcvtnq_s32_f32(vfpacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
|
| D | minmax-fp32-avx-mul16-ld64-x8.c | 49 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
52 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
55 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
83 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
86 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
89 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
|
| D | minmax-fp32-neon-ld64-x8.c | 43 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
46 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
49 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
83 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
86 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
89 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
|
| D | minmax-fp32-sse2-mul16-ld64-x8.c | 51 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
54 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
57 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
87 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
90 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
93 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
|
| /external/XNNPACK/src/qs8-vmulc/gen/ |
| D | minmax-fp32-sse41-mul16-ld64-x8.c | 49 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
52 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
55 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
83 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
86 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
89 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
|
| D | minmax-fp32-neonv8-ld64-x8.c | 43 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
46 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
49 vacc4567 = vcvtnq_s32_f32(vfpacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
81 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
84 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
87 vacc4567 = vcvtnq_s32_f32(vfpacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8()
|
| D | minmax-fp32-sse2-mul16-ld64-x8.c | 50 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
53 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
56 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
85 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
88 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
91 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
|
| D | minmax-fp32-avx-mul16-ld64-x8.c | 49 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
52 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
55 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
83 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
86 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
89 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x8()
|
| D | minmax-fp32-neon-ld64-x8.c | 43 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
46 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
49 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
83 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
86 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
89 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8()
|
| /external/XNNPACK/src/qs8-vmul/gen/ |
| D | minmax-fp32-neonv8-ld64-x8.c | 43 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
46 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
49 vacc4567 = vcvtnq_s32_f32(vfpacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
83 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
86 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
89 vacc4567 = vcvtnq_s32_f32(vfpacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
|
| D | minmax-fp32-sse41-mul16-ld64-x8.c | 50 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
53 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
56 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
86 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
89 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
92 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
|
| D | minmax-fp32-avx-mul16-ld64-x8.c | 50 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
53 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
56 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
86 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
89 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
92 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
|
| D | minmax-fp32-sse2-mul16-ld64-x8.c | 52 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
55 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
58 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
90 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
93 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
96 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
|
| D | minmax-fp32-neon-ld64-x8.c | 43 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
46 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
49 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias)); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
85 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
88 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
91 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias)); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8()
|
| /external/XNNPACK/src/qu8-vmul/gen/ |
| D | minmax-fp32-neonv8-ld64-x8.c | 43 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
46 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
49 vacc4567 = vcvtnq_s32_f32(vfpacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
83 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
86 vfpacc4567 = vmulq_f32(vfpacc4567, vscale); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
89 vacc4567 = vcvtnq_s32_f32(vfpacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8()
|
| D | minmax-fp32-sse41-mul16-ld64-x8.c | 50 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
53 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
56 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
86 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8() local
89 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
92 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x8()
|
| D | minmax-fp32-avx-mul16-ld64-x8.c | 50 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
53 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
56 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
86 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8() local
89 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
92 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x8()
|
| D | minmax-fp32-sse2-mul16-ld64-x8.c | 53 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
56 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
59 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
92 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8() local
95 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
98 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8()
|