Home
last modified time | relevance | path

Searched refs:vy_hi (Results 1 – 25 of 49) sorted by relevance

12

/external/XNNPACK/src/qs8-f32-vcvt/gen/
Dvcvt-wasmsimd-x8.c38 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x8() local
41 vy_hi = wasm_f32x4_convert_i32x4(vy_hi); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x8()
44 vy_hi = wasm_f32x4_mul(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x8()
47 wasm_v128_store(y + 4, vy_hi); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x8()
Dvcvt-sse2-x8.c42 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x8() local
45 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); in xnn_qs8_f32_vcvt_ukernel__sse2_x8()
48 vy_hi = _mm_mul_ps(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x8()
51 _mm_storeu_ps(y + 4, vy_hi); in xnn_qs8_f32_vcvt_ukernel__sse2_x8()
Dvcvt-wasmsimd-x16.c67 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16() local
70 vy_hi = wasm_f32x4_convert_i32x4(vy_hi); 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()
76 wasm_v128_store(y + 4, vy_hi); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
Dvcvt-wasmsimd-x24.c77 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24() local
80 vy_hi = wasm_f32x4_convert_i32x4(vy_hi); 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()
86 wasm_v128_store(y + 4, vy_hi); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
Dvcvt-sse2-x16.c74 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x16() local
77 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); in xnn_qs8_f32_vcvt_ukernel__sse2_x16()
80 vy_hi = _mm_mul_ps(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x16()
83 _mm_storeu_ps(y + 4, vy_hi); in xnn_qs8_f32_vcvt_ukernel__sse2_x16()
Dvcvt-wasmsimd-x32.c87 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32() local
90 vy_hi = wasm_f32x4_convert_i32x4(vy_hi); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
93 vy_hi = wasm_f32x4_mul(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
96 wasm_v128_store(y + 4, vy_hi); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
Dvcvt-sse2-x24.c85 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x24() local
88 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); in xnn_qs8_f32_vcvt_ukernel__sse2_x24()
91 vy_hi = _mm_mul_ps(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x24()
94 _mm_storeu_ps(y + 4, vy_hi); in xnn_qs8_f32_vcvt_ukernel__sse2_x24()
Dvcvt-sse2-x32.c96 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x32() local
99 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
102 vy_hi = _mm_mul_ps(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
105 _mm_storeu_ps(y + 4, vy_hi); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
Dvcvt-neon-x8.c41 float32x4_t vy_hi = vcvtq_f32_s32(vwx_hi); in xnn_qs8_f32_vcvt_ukernel__neon_x8() local
44 vy_hi = vmulq_f32(vy_hi, vscale); in xnn_qs8_f32_vcvt_ukernel__neon_x8()
47 vst1q_f32(y, vy_hi); y += 4; in xnn_qs8_f32_vcvt_ukernel__neon_x8()
/external/XNNPACK/src/qu8-f32-vcvt/gen/
Dvcvt-wasmsimd-x8.c38 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x8() local
41 vy_hi = wasm_f32x4_convert_i32x4(vy_hi); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x8()
44 vy_hi = wasm_f32x4_mul(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x8()
47 wasm_v128_store(y + 4, vy_hi); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x8()
Dvcvt-sse2-x8.c40 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); in xnn_qu8_f32_vcvt_ukernel__sse2_x8() local
43 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); in xnn_qu8_f32_vcvt_ukernel__sse2_x8()
46 vy_hi = _mm_mul_ps(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x8()
49 _mm_storeu_ps(y + 4, vy_hi); in xnn_qu8_f32_vcvt_ukernel__sse2_x8()
Dvcvt-wasmsimd-x16.c67 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16() local
70 vy_hi = wasm_f32x4_convert_i32x4(vy_hi); 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()
76 wasm_v128_store(y + 4, vy_hi); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
Dvcvt-wasmsimd-x24.c77 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24() local
80 vy_hi = wasm_f32x4_convert_i32x4(vy_hi); 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()
86 wasm_v128_store(y + 4, vy_hi); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
Dvcvt-sse2-x16.c70 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); in xnn_qu8_f32_vcvt_ukernel__sse2_x16() local
73 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); in xnn_qu8_f32_vcvt_ukernel__sse2_x16()
76 vy_hi = _mm_mul_ps(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x16()
79 _mm_storeu_ps(y + 4, vy_hi); in xnn_qu8_f32_vcvt_ukernel__sse2_x16()
Dvcvt-wasmsimd-x32.c87 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32() local
90 vy_hi = wasm_f32x4_convert_i32x4(vy_hi); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
93 vy_hi = wasm_f32x4_mul(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
96 wasm_v128_store(y + 4, vy_hi); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
Dvcvt-sse2-x24.c80 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); in xnn_qu8_f32_vcvt_ukernel__sse2_x24() local
83 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); 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()
89 _mm_storeu_ps(y + 4, vy_hi); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
Dvcvt-sse2-x32.c90 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); in xnn_qu8_f32_vcvt_ukernel__sse2_x32() local
93 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
96 vy_hi = _mm_mul_ps(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
99 _mm_storeu_ps(y + 4, vy_hi); in xnn_qu8_f32_vcvt_ukernel__sse2_x32()
Dvcvt-neon-x8.c41 float32x4_t vy_hi = vcvtq_f32_s32(vwx_hi); in xnn_qu8_f32_vcvt_ukernel__neon_x8() local
44 vy_hi = vmulq_f32(vy_hi, vscale); in xnn_qu8_f32_vcvt_ukernel__neon_x8()
47 vst1q_f32(y, vy_hi); y += 4; in xnn_qu8_f32_vcvt_ukernel__neon_x8()
/external/XNNPACK/src/math/
Dcvt-f32-qs8-neon.c33 int32x4_t vy_hi = vreinterpretq_s32_f32(vx_hi); in xnn_math_f32_qs8_cvt__neon() local
36 vy_hi = vqsubq_s32(vy_hi, vimagic); in xnn_math_f32_qs8_cvt__neon()
38 const int16x8_t vy = vcombine_s16(vqmovn_s32(vy_lo), vqmovn_s32(vy_hi)); in xnn_math_f32_qs8_cvt__neon()
Dcvt-f32-qu8-neon.c33 int32x4_t vy_hi = vreinterpretq_s32_f32(vx_hi); in xnn_math_f32_qu8_cvt__neon() local
36 vy_hi = vqsubq_s32(vy_hi, vimagic); in xnn_math_f32_qu8_cvt__neon()
38 const int16x8_t vy = vcombine_s16(vqmovn_s32(vy_lo), vqmovn_s32(vy_hi)); in xnn_math_f32_qu8_cvt__neon()
/external/XNNPACK/src/qs8-f32-vcvt/
Dwasmsimd.c.in64 v128_t vy_hi = wasm_i32x4_extend_high_i16x8(vx); variable
67 vy_hi = wasm_f32x4_convert_i32x4(vy_hi);
70 vy_hi = wasm_f32x4_mul(vy_hi, vscale);
73 wasm_v128_store(y + 4, vy_hi);
Dsse2.c.in74 __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); variable
77 vy_hi = _mm_sub_ps(vy_hi, vmagic_bias);
80 vy_hi = _mm_mul_ps(vy_hi, vscale);
83 _mm_storeu_ps(y + 4, vy_hi);
/external/XNNPACK/src/f32-qs8-vcvt/gen/
Dvcvt-sse41-x8.c47 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); in xnn_f32_qs8_vcvt_ukernel__sse41_x8() local
49 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); in xnn_f32_qs8_vcvt_ukernel__sse41_x8()
69 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); in xnn_f32_qs8_vcvt_ukernel__sse41_x8() local
71 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); in xnn_f32_qs8_vcvt_ukernel__sse41_x8()
Dvcvt-sse2-x8.c47 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); in xnn_f32_qs8_vcvt_ukernel__sse2_x8() local
49 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); in xnn_f32_qs8_vcvt_ukernel__sse2_x8()
69 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); in xnn_f32_qs8_vcvt_ukernel__sse2_x8() local
71 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); in xnn_f32_qs8_vcvt_ukernel__sse2_x8()
/external/XNNPACK/src/f32-qu8-vcvt/gen/
Dvcvt-sse2-x8.c47 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); in xnn_f32_qu8_vcvt_ukernel__sse2_x8() local
49 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); in xnn_f32_qu8_vcvt_ukernel__sse2_x8()
69 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi); in xnn_f32_qu8_vcvt_ukernel__sse2_x8() local
71 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi); in xnn_f32_qu8_vcvt_ukernel__sse2_x8()

12