Home
last modified time | relevance | path

Searched refs:vscale (Results 1 – 25 of 1596) sorted by relevance

12345678910>>...64

/external/ComputeLibrary/src/core/NEON/
DSVEAsymm.h71 const auto vscale = svdup_n_f32(scale); in svdequantize_z() local
73 …_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlb_u16(qv))), voffset)), vscale), in svdequantize_z()
74 …_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlb_u16(qv))), voffset)), vscale), in svdequantize_z()
75 …_s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlb_u32(svmovlt_u16(qv))), voffset)), vscale), in svdequantize_z()
76 …s32_z(pg, svsub_s32_z(pg, svreinterpret_s32_u32(svmovlt_u32(svmovlt_u16(qv))), voffset)), vscale)); in svdequantize_z()
105 const auto vscale = svdup_n_f32(scale); in svdequantize_z() local
107 …mul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlb_s16(qv)), voffset)), vscale), in svdequantize_z()
108 …mul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlb_s16(qv)), voffset)), vscale), in svdequantize_z()
109 …mul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlb_s32(svmovlt_s16(qv)), voffset)), vscale), in svdequantize_z()
110 …ul_f32_z(pg, svcvt_f32_s32_z(pg, svsub_s32_z(pg, svmovlt_s32(svmovlt_s16(qv)), voffset)), vscale)); in svdequantize_z()
[all …]
DNEAsymm.h420 const float32x4_t vscale = vdupq_n_f32(scale); in vdequantize() local
424 …f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(qv)))), voffset)), vscale), in vdequantize()
425 …32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(qv)))), voffset)), vscale), in vdequantize()
443 const float32x4_t vscale = vdupq_n_f32(scale); in vdequantize() local
447 … vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_low_s16(vmovl_s8(qv))), voffset)), vscale), in vdequantize()
448 … vmulq_f32(vcvtq_f32_s32(vsubq_s32(vmovl_s16(vget_high_s16(vmovl_s8(qv))), voffset)), vscale), in vdequantize()
466 const float32x4_t vscale = vdupq_n_f32(scale); in vdequantize() local
470 …_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(qv))))), voffset)), vscale), in vdequantize()
471 …s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(qv))))), voffset)), vscale), in vdequantize()
472 …s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(qv))))), voffset)), vscale), in vdequantize()
[all …]
/external/XNNPACK/src/qu8-f32-vcvt/gen/
Dvcvt-wasmsimd-x32.c31 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd.scale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32() local
62 vy0123 = wasm_f32x4_mul(vy0123, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
63 vy4567 = wasm_f32x4_mul(vy4567, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
64 vy89AB = wasm_f32x4_mul(vy89AB, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
65 vyCDEF = wasm_f32x4_mul(vyCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
66 vyGHIJ = wasm_f32x4_mul(vyGHIJ, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
67 vyKLMN = wasm_f32x4_mul(vyKLMN, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
68 vyOPQR = wasm_f32x4_mul(vyOPQR, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
69 vySTUV = wasm_f32x4_mul(vySTUV, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
92 vy_lo = wasm_f32x4_mul(vy_lo, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
[all …]
Dvcvt-neon-x32.c31 const float32x4_t vscale = vld1q_dup_f32(&params->neon.scale); in xnn_qu8_f32_vcvt_ukernel__neon_x32() local
61 vy0123 = vmulq_f32(vy0123, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x32()
62 vy4567 = vmulq_f32(vy4567, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x32()
63 vy89AB = vmulq_f32(vy89AB, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x32()
64 vyCDEF = vmulq_f32(vyCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x32()
65 vyGHIJ = vmulq_f32(vyGHIJ, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x32()
66 vyKLMN = vmulq_f32(vyKLMN, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x32()
67 vyOPQR = vmulq_f32(vyOPQR, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x32()
68 vySTUV = vmulq_f32(vySTUV, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x32()
90 vy_lo = vmulq_f32(vy_lo, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x32()
[all …]
Dvcvt-sse2-x32.c32 const __m128 vscale = _mm_load_ps(params->sse2.scale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32() local
65 vy0123 = _mm_mul_ps(vy0123, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
66 vy4567 = _mm_mul_ps(vy4567, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
67 vy89AB = _mm_mul_ps(vy89AB, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
68 vyCDEF = _mm_mul_ps(vyCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
69 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
70 vyKLMN = _mm_mul_ps(vyKLMN, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
71 vyOPQR = _mm_mul_ps(vyOPQR, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
72 vySTUV = _mm_mul_ps(vySTUV, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
95 vy_lo = _mm_mul_ps(vy_lo, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
[all …]
Dvcvt-wasmsimd-x24.c31 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd.scale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24() local
56 vy0123 = wasm_f32x4_mul(vy0123, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
57 vy4567 = wasm_f32x4_mul(vy4567, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
58 vy89AB = wasm_f32x4_mul(vy89AB, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
59 vyCDEF = wasm_f32x4_mul(vyCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
60 vyGHIJ = wasm_f32x4_mul(vyGHIJ, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
61 vyKLMN = wasm_f32x4_mul(vyKLMN, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
82 vy_lo = wasm_f32x4_mul(vy_lo, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
83 vy_hi = wasm_f32x4_mul(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
99 vy = wasm_f32x4_mul(vy, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
[all …]
Dvcvt-sse41-x32.c31 const __m128 vscale = _mm_load_ps(params->sse4.scale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32() local
61 vy0123 = _mm_mul_ps(vy0123, vscale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32()
62 vy4567 = _mm_mul_ps(vy4567, vscale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32()
63 vy89AB = _mm_mul_ps(vy89AB, vscale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32()
64 vyCDEF = _mm_mul_ps(vyCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32()
65 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32()
66 vyKLMN = _mm_mul_ps(vyKLMN, vscale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32()
67 vyOPQR = _mm_mul_ps(vyOPQR, vscale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32()
68 vySTUV = _mm_mul_ps(vySTUV, vscale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32()
86 vy = _mm_mul_ps(vy, vscale); in xnn_qu8_f32_vcvt_ukernel__sse41_x32()
[all …]
Dvcvt-sse2-x24.c32 const __m128 vscale = _mm_load_ps(params->sse2.scale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24() local
59 vy0123 = _mm_mul_ps(vy0123, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
60 vy4567 = _mm_mul_ps(vy4567, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
61 vy89AB = _mm_mul_ps(vy89AB, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
62 vyCDEF = _mm_mul_ps(vyCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
63 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
64 vyKLMN = _mm_mul_ps(vyKLMN, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
85 vy_lo = _mm_mul_ps(vy_lo, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
86 vy_hi = _mm_mul_ps(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
101 vy = _mm_mul_ps(vy, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
[all …]
Dvcvt-neon-x24.c31 const float32x4_t vscale = vld1q_dup_f32(&params->neon.scale); in xnn_qu8_f32_vcvt_ukernel__neon_x24() local
55 vy0123 = vmulq_f32(vy0123, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x24()
56 vy4567 = vmulq_f32(vy4567, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x24()
57 vy89AB = vmulq_f32(vy89AB, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x24()
58 vyCDEF = vmulq_f32(vyCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x24()
59 vyGHIJ = vmulq_f32(vyGHIJ, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x24()
60 vyKLMN = vmulq_f32(vyKLMN, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x24()
80 vy_lo = vmulq_f32(vy_lo, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x24()
81 vy_hi = vmulq_f32(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x24()
98 vy = vmulq_f32(vy, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x24()
[all …]
Dvcvt-wasmsimd-x16.c31 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd.scale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16() local
50 vy0123 = wasm_f32x4_mul(vy0123, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
51 vy4567 = wasm_f32x4_mul(vy4567, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
52 vy89AB = wasm_f32x4_mul(vy89AB, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
53 vyCDEF = wasm_f32x4_mul(vyCDEF, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
72 vy_lo = wasm_f32x4_mul(vy_lo, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
73 vy_hi = wasm_f32x4_mul(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
89 vy = wasm_f32x4_mul(vy, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
95 vy = wasm_f32x4_mul(vy, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
/external/XNNPACK/src/qs8-f32-vcvt/gen/
Dvcvt-wasmsimd-x32.c31 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd.scale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32() local
62 vy0123 = wasm_f32x4_mul(vy0123, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
63 vy4567 = wasm_f32x4_mul(vy4567, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
64 vy89AB = wasm_f32x4_mul(vy89AB, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
65 vyCDEF = wasm_f32x4_mul(vyCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
66 vyGHIJ = wasm_f32x4_mul(vyGHIJ, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
67 vyKLMN = wasm_f32x4_mul(vyKLMN, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
68 vyOPQR = wasm_f32x4_mul(vyOPQR, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
69 vySTUV = wasm_f32x4_mul(vySTUV, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
92 vy_lo = wasm_f32x4_mul(vy_lo, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
[all …]
Dvcvt-neon-x32.c31 const float32x4_t vscale = vld1q_dup_f32(&params->neon.scale); in xnn_qs8_f32_vcvt_ukernel__neon_x32() local
61 vy0123 = vmulq_f32(vy0123, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x32()
62 vy4567 = vmulq_f32(vy4567, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x32()
63 vy89AB = vmulq_f32(vy89AB, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x32()
64 vyCDEF = vmulq_f32(vyCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x32()
65 vyGHIJ = vmulq_f32(vyGHIJ, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x32()
66 vyKLMN = vmulq_f32(vyKLMN, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x32()
67 vyOPQR = vmulq_f32(vyOPQR, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x32()
68 vySTUV = vmulq_f32(vySTUV, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x32()
90 vy_lo = vmulq_f32(vy_lo, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x32()
[all …]
Dvcvt-wasmsimd-x24.c31 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd.scale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24() local
56 vy0123 = wasm_f32x4_mul(vy0123, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
57 vy4567 = wasm_f32x4_mul(vy4567, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
58 vy89AB = wasm_f32x4_mul(vy89AB, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
59 vyCDEF = wasm_f32x4_mul(vyCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
60 vyGHIJ = wasm_f32x4_mul(vyGHIJ, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
61 vyKLMN = wasm_f32x4_mul(vyKLMN, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
82 vy_lo = wasm_f32x4_mul(vy_lo, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
83 vy_hi = wasm_f32x4_mul(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
99 vy = wasm_f32x4_mul(vy, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
[all …]
Dvcvt-sse2-x32.c33 const __m128 vscale = _mm_load_ps(params->sse2.scale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32() local
70 vy0123 = _mm_mul_ps(vy0123, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
71 vy4567 = _mm_mul_ps(vy4567, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
72 vy89AB = _mm_mul_ps(vy89AB, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
73 vyCDEF = _mm_mul_ps(vyCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
74 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
75 vyKLMN = _mm_mul_ps(vyKLMN, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
76 vyOPQR = _mm_mul_ps(vyOPQR, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
77 vySTUV = _mm_mul_ps(vySTUV, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
101 vy_lo = _mm_mul_ps(vy_lo, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
[all …]
Dvcvt-sse41-x32.c31 const __m128 vscale = _mm_load_ps(params->sse4.scale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32() local
61 vy0123 = _mm_mul_ps(vy0123, vscale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32()
62 vy4567 = _mm_mul_ps(vy4567, vscale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32()
63 vy89AB = _mm_mul_ps(vy89AB, vscale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32()
64 vyCDEF = _mm_mul_ps(vyCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32()
65 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32()
66 vyKLMN = _mm_mul_ps(vyKLMN, vscale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32()
67 vyOPQR = _mm_mul_ps(vyOPQR, vscale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32()
68 vySTUV = _mm_mul_ps(vySTUV, vscale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32()
86 vy = _mm_mul_ps(vy, vscale); in xnn_qs8_f32_vcvt_ukernel__sse41_x32()
[all …]
Dvcvt-neon-x24.c31 const float32x4_t vscale = vld1q_dup_f32(&params->neon.scale); in xnn_qs8_f32_vcvt_ukernel__neon_x24() local
55 vy0123 = vmulq_f32(vy0123, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x24()
56 vy4567 = vmulq_f32(vy4567, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x24()
57 vy89AB = vmulq_f32(vy89AB, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x24()
58 vyCDEF = vmulq_f32(vyCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x24()
59 vyGHIJ = vmulq_f32(vyGHIJ, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x24()
60 vyKLMN = vmulq_f32(vyKLMN, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x24()
80 vy_lo = vmulq_f32(vy_lo, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x24()
81 vy_hi = vmulq_f32(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x24()
98 vy = vmulq_f32(vy, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x24()
[all …]
Dvcvt-wasmsimd-x16.c31 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd.scale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16() local
50 vy0123 = wasm_f32x4_mul(vy0123, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
51 vy4567 = wasm_f32x4_mul(vy4567, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
52 vy89AB = wasm_f32x4_mul(vy89AB, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
53 vyCDEF = wasm_f32x4_mul(vyCDEF, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
72 vy_lo = wasm_f32x4_mul(vy_lo, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
73 vy_hi = wasm_f32x4_mul(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
89 vy = wasm_f32x4_mul(vy, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
95 vy = wasm_f32x4_mul(vy, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
/external/XNNPACK/src/f32-qs8-vcvt/gen/
Dvcvt-neonv8-x32.c30 const float32x4_t vscale = vld1q_dup_f32(&params->neonv8.scale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32() local
44 vx0123 = vmulq_f32(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
45 vx4567 = vmulq_f32(vx4567, vscale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
46 vx89AB = vmulq_f32(vx89AB, vscale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
47 vxCDEF = vmulq_f32(vxCDEF, vscale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
48 vxGHIJ = vmulq_f32(vxGHIJ, vscale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
49 vxKLMN = vmulq_f32(vxKLMN, vscale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
50 vxOPQR = vmulq_f32(vxOPQR, vscale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
51 vxSTUV = vmulq_f32(vxSTUV, vscale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
88 vx_lo = vmulq_f32(vx_lo, vscale); in xnn_f32_qs8_vcvt_ukernel__neonv8_x32()
[all …]
Dvcvt-sse41-x32.c30 const __m128 vscale = _mm_load_ps(params->sse4.scale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32() local
46 vx0123 = _mm_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32()
47 vx4567 = _mm_mul_ps(vx4567, vscale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32()
48 vx89AB = _mm_mul_ps(vx89AB, vscale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32()
49 vxCDEF = _mm_mul_ps(vxCDEF, vscale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32()
50 vxGHIJ = _mm_mul_ps(vxGHIJ, vscale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32()
51 vxKLMN = _mm_mul_ps(vxKLMN, vscale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32()
52 vxOPQR = _mm_mul_ps(vxOPQR, vscale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32()
53 vxSTUV = _mm_mul_ps(vxSTUV, vscale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32()
99 vx_lo = _mm_mul_ps(vx_lo, vscale); in xnn_f32_qs8_vcvt_ukernel__sse41_x32()
[all …]
Dvcvt-sse2-x32.c30 const __m128 vscale = _mm_load_ps(params->sse2.scale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32() local
46 vx0123 = _mm_mul_ps(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
47 vx4567 = _mm_mul_ps(vx4567, vscale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
48 vx89AB = _mm_mul_ps(vx89AB, vscale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
49 vxCDEF = _mm_mul_ps(vxCDEF, vscale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
50 vxGHIJ = _mm_mul_ps(vxGHIJ, vscale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
51 vxKLMN = _mm_mul_ps(vxKLMN, vscale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
52 vxOPQR = _mm_mul_ps(vxOPQR, vscale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
53 vxSTUV = _mm_mul_ps(vxSTUV, vscale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
101 vx_lo = _mm_mul_ps(vx_lo, vscale); in xnn_f32_qs8_vcvt_ukernel__sse2_x32()
[all …]
Dvcvt-wasmsimd-cvt-x32.c30 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd_cvt.scale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32() local
45 vx0123 = wasm_f32x4_mul(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
46 vx4567 = wasm_f32x4_mul(vx4567, vscale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
47 vx89AB = wasm_f32x4_mul(vx89AB, vscale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
48 vxCDEF = wasm_f32x4_mul(vxCDEF, vscale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
49 vxGHIJ = wasm_f32x4_mul(vxGHIJ, vscale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
50 vxKLMN = wasm_f32x4_mul(vxKLMN, vscale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
51 vxOPQR = wasm_f32x4_mul(vxOPQR, vscale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
52 vxSTUV = wasm_f32x4_mul(vxSTUV, vscale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
100 vx_lo = wasm_f32x4_mul(vx_lo, vscale); in xnn_f32_qs8_vcvt_ukernel__wasmsimd_cvt_x32()
[all …]
Dvcvt-neon-x32.c30 const float32x4_t vscale = vld1q_dup_f32(&params->neon.scale); in xnn_f32_qs8_vcvt_ukernel__neon_x32() local
45 vx0123 = vmulq_f32(vx0123, vscale); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
46 vx4567 = vmulq_f32(vx4567, vscale); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
47 vx89AB = vmulq_f32(vx89AB, vscale); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
48 vxCDEF = vmulq_f32(vxCDEF, vscale); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
49 vxGHIJ = vmulq_f32(vxGHIJ, vscale); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
50 vxKLMN = vmulq_f32(vxKLMN, vscale); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
51 vxOPQR = vmulq_f32(vxOPQR, vscale); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
52 vxSTUV = vmulq_f32(vxSTUV, vscale); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
93 vx_lo = vmulq_f32(vx_lo, vscale); in xnn_f32_qs8_vcvt_ukernel__neon_x32()
[all …]
/external/XNNPACK/src/f32-qu8-vcvt/gen/
Dvcvt-neonv8-x32.c30 const float32x4_t vscale = vld1q_dup_f32(&params->neonv8.scale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32() local
44 vx0123 = vmulq_f32(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
45 vx4567 = vmulq_f32(vx4567, vscale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
46 vx89AB = vmulq_f32(vx89AB, vscale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
47 vxCDEF = vmulq_f32(vxCDEF, vscale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
48 vxGHIJ = vmulq_f32(vxGHIJ, vscale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
49 vxKLMN = vmulq_f32(vxKLMN, vscale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
50 vxOPQR = vmulq_f32(vxOPQR, vscale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
51 vxSTUV = vmulq_f32(vxSTUV, vscale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
88 vx_lo = vmulq_f32(vx_lo, vscale); in xnn_f32_qu8_vcvt_ukernel__neonv8_x32()
[all …]
Dvcvt-sse2-x32.c30 const __m128 vscale = _mm_load_ps(params->sse2.scale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32() local
46 vx0123 = _mm_mul_ps(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32()
47 vx4567 = _mm_mul_ps(vx4567, vscale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32()
48 vx89AB = _mm_mul_ps(vx89AB, vscale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32()
49 vxCDEF = _mm_mul_ps(vxCDEF, vscale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32()
50 vxGHIJ = _mm_mul_ps(vxGHIJ, vscale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32()
51 vxKLMN = _mm_mul_ps(vxKLMN, vscale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32()
52 vxOPQR = _mm_mul_ps(vxOPQR, vscale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32()
53 vxSTUV = _mm_mul_ps(vxSTUV, vscale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32()
99 vx_lo = _mm_mul_ps(vx_lo, vscale); in xnn_f32_qu8_vcvt_ukernel__sse2_x32()
[all …]
Dvcvt-wasmsimd-cvt-x32.c30 const v128_t vscale = wasm_v128_load64_splat(params->wasmsimd_cvt.scale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32() local
45 vx0123 = wasm_f32x4_mul(vx0123, vscale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
46 vx4567 = wasm_f32x4_mul(vx4567, vscale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
47 vx89AB = wasm_f32x4_mul(vx89AB, vscale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
48 vxCDEF = wasm_f32x4_mul(vxCDEF, vscale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
49 vxGHIJ = wasm_f32x4_mul(vxGHIJ, vscale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
50 vxKLMN = wasm_f32x4_mul(vxKLMN, vscale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
51 vxOPQR = wasm_f32x4_mul(vxOPQR, vscale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
52 vxSTUV = wasm_f32x4_mul(vxSTUV, vscale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
100 vx_lo = wasm_f32x4_mul(vx_lo, vscale); in xnn_f32_qu8_vcvt_ukernel__wasmsimd_cvt_x32()
[all …]

12345678910>>...64