Home
last modified time | relevance | path

Searched refs:vx89ABCDEF (Results 1 – 25 of 56) sorted by relevance

123

/external/XNNPACK/src/qs8-f32-vcvt/gen/
Dvcvt-sse2-x16.c37 __m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (x + 8)); in xnn_qs8_f32_vcvt_ukernel__sse2_x16() local
41 vx89ABCDEF = _mm_xor_si128(vx89ABCDEF, vsign_mask); in xnn_qs8_f32_vcvt_ukernel__sse2_x16()
44 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero); in xnn_qs8_f32_vcvt_ukernel__sse2_x16()
48 __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x16()
49 __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x16()
Dvcvt-sse2-x24.c37 __m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (x + 8)); in xnn_qs8_f32_vcvt_ukernel__sse2_x24() local
42 vx89ABCDEF = _mm_xor_si128(vx89ABCDEF, vsign_mask); in xnn_qs8_f32_vcvt_ukernel__sse2_x24()
46 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero); in xnn_qs8_f32_vcvt_ukernel__sse2_x24()
51 __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x24()
52 __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x24()
Dvcvt-sse2-x32.c37 __m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (x + 8)); in xnn_qs8_f32_vcvt_ukernel__sse2_x32() local
43 vx89ABCDEF = _mm_xor_si128(vx89ABCDEF, vsign_mask); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
48 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
54 __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
55 __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qs8_f32_vcvt_ukernel__sse2_x32()
Dvcvt-wasmsimd-x16.c34 v128_t vx89ABCDEF = wasm_i16x8_load8x8(x + 8); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16() local
38 vx89ABCDEF = wasm_i16x8_add(vx89ABCDEF, vminus_zero_point); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
42 v128_t vy89AB = wasm_i32x4_extend_low_i16x8(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
43 v128_t vyCDEF = wasm_i32x4_extend_high_i16x8(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x16()
Dvcvt-wasmsimd-x24.c34 v128_t vx89ABCDEF = wasm_i16x8_load8x8(x + 8); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24() local
39 vx89ABCDEF = wasm_i16x8_add(vx89ABCDEF, vminus_zero_point); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
44 v128_t vy89AB = wasm_i32x4_extend_low_i16x8(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
45 v128_t vyCDEF = wasm_i32x4_extend_high_i16x8(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x24()
Dvcvt-avx2-x16.c34 __m256i vx89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (x + 8))); in xnn_qs8_f32_vcvt_ukernel__avx2_x16() local
38 vx89ABCDEF = _mm256_add_epi32(vx89ABCDEF, vminus_zero_point); in xnn_qs8_f32_vcvt_ukernel__avx2_x16()
41 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__avx2_x16()
Dvcvt-wasmsimd-x32.c34 v128_t vx89ABCDEF = wasm_i16x8_load8x8(x + 8); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32() local
40 vx89ABCDEF = wasm_i16x8_add(vx89ABCDEF, vminus_zero_point); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
46 v128_t vy89AB = wasm_i32x4_extend_low_i16x8(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
47 v128_t vyCDEF = wasm_i32x4_extend_high_i16x8(vx89ABCDEF); in xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32()
/external/XNNPACK/src/f32-qu8-vcvt/gen/
Dvcvt-avx-x16.c37 __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8); in xnn_f32_qu8_vcvt_ukernel__avx_x16() local
41 vx89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vscale); in xnn_f32_qu8_vcvt_ukernel__avx_x16()
44 vx89ABCDEF = _mm256_min_ps(vx89ABCDEF, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx_x16()
47 const __m256i vacc89ABCDEF = _mm256_cvtps_epi32(vx89ABCDEF); in xnn_f32_qu8_vcvt_ukernel__avx_x16()
Dvcvt-avx-x24.c37 __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8); in xnn_f32_qu8_vcvt_ukernel__avx_x24() local
42 vx89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vscale); in xnn_f32_qu8_vcvt_ukernel__avx_x24()
46 vx89ABCDEF = _mm256_min_ps(vx89ABCDEF, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx_x24()
50 const __m256i vacc89ABCDEF = _mm256_cvtps_epi32(vx89ABCDEF); in xnn_f32_qu8_vcvt_ukernel__avx_x24()
Dvcvt-avx-x32.c37 __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8); in xnn_f32_qu8_vcvt_ukernel__avx_x32() local
43 vx89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vscale); in xnn_f32_qu8_vcvt_ukernel__avx_x32()
48 vx89ABCDEF = _mm256_min_ps(vx89ABCDEF, voutput_max_less_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx_x32()
53 const __m256i vacc89ABCDEF = _mm256_cvtps_epi32(vx89ABCDEF); in xnn_f32_qu8_vcvt_ukernel__avx_x32()
/external/XNNPACK/src/f32-qs8-vcvt/gen/
Dvcvt-avx-x16.c37 __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8); in xnn_f32_qs8_vcvt_ukernel__avx_x16() local
41 vx89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vscale); in xnn_f32_qs8_vcvt_ukernel__avx_x16()
44 vx89ABCDEF = _mm256_min_ps(vx89ABCDEF, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx_x16()
47 const __m256i vacc89ABCDEF = _mm256_cvtps_epi32(vx89ABCDEF); in xnn_f32_qs8_vcvt_ukernel__avx_x16()
Dvcvt-avx-x24.c37 __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8); in xnn_f32_qs8_vcvt_ukernel__avx_x24() local
42 vx89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vscale); in xnn_f32_qs8_vcvt_ukernel__avx_x24()
46 vx89ABCDEF = _mm256_min_ps(vx89ABCDEF, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx_x24()
50 const __m256i vacc89ABCDEF = _mm256_cvtps_epi32(vx89ABCDEF); in xnn_f32_qs8_vcvt_ukernel__avx_x24()
Dvcvt-avx-x32.c37 __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8); in xnn_f32_qs8_vcvt_ukernel__avx_x32() local
43 vx89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vscale); in xnn_f32_qs8_vcvt_ukernel__avx_x32()
48 vx89ABCDEF = _mm256_min_ps(vx89ABCDEF, voutput_max_less_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx_x32()
53 const __m256i vacc89ABCDEF = _mm256_cvtps_epi32(vx89ABCDEF); in xnn_f32_qs8_vcvt_ukernel__avx_x32()
/external/XNNPACK/src/qu8-f32-vcvt/gen/
Dvcvt-wasmsimd-x16.c34 v128_t vx89ABCDEF = wasm_u16x8_load8x8(x + 8); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16() local
38 vx89ABCDEF = wasm_i16x8_add(vx89ABCDEF, vminus_zero_point); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
42 v128_t vy89AB = wasm_i32x4_extend_low_i16x8(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
43 v128_t vyCDEF = wasm_i32x4_extend_high_i16x8(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x16()
Dvcvt-wasmsimd-x24.c34 v128_t vx89ABCDEF = wasm_u16x8_load8x8(x + 8); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24() local
39 vx89ABCDEF = wasm_i16x8_add(vx89ABCDEF, vminus_zero_point); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
44 v128_t vy89AB = wasm_i32x4_extend_low_i16x8(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
45 v128_t vyCDEF = wasm_i32x4_extend_high_i16x8(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x24()
Dvcvt-sse2-x16.c36 __m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (x + 8)); in xnn_qu8_f32_vcvt_ukernel__sse2_x16() local
41 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero); in xnn_qu8_f32_vcvt_ukernel__sse2_x16()
45 __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qu8_f32_vcvt_ukernel__sse2_x16()
46 __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qu8_f32_vcvt_ukernel__sse2_x16()
Dvcvt-avx2-x16.c34 __m256i vx89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (x + 8))); in xnn_qu8_f32_vcvt_ukernel__avx2_x16() local
38 vx89ABCDEF = _mm256_add_epi32(vx89ABCDEF, vminus_zero_point); in xnn_qu8_f32_vcvt_ukernel__avx2_x16()
41 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx2_x16()
Dvcvt-wasmsimd-x32.c34 v128_t vx89ABCDEF = wasm_u16x8_load8x8(x + 8); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32() local
40 vx89ABCDEF = wasm_i16x8_add(vx89ABCDEF, vminus_zero_point); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
46 v128_t vy89AB = wasm_i32x4_extend_low_i16x8(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
47 v128_t vyCDEF = wasm_i32x4_extend_high_i16x8(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32()
Dvcvt-sse2-x24.c36 __m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (x + 8)); in xnn_qu8_f32_vcvt_ukernel__sse2_x24() local
42 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
47 __m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
48 __m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp)); in xnn_qu8_f32_vcvt_ukernel__sse2_x24()
Dvcvt-avx2-x24.c34 __m256i vx89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (x + 8))); in xnn_qu8_f32_vcvt_ukernel__avx2_x24() local
39 vx89ABCDEF = _mm256_add_epi32(vx89ABCDEF, vminus_zero_point); in xnn_qu8_f32_vcvt_ukernel__avx2_x24()
43 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF); in xnn_qu8_f32_vcvt_ukernel__avx2_x24()
/external/XNNPACK/src/f16-vhswish/gen/
Dvhswish-neonfp16arith-x16.c37 float16x8_t vx89ABCDEF = vld1q_f16(x); x += 8; in xnn_f16_vhswish_ukernel__neonfp16arith_x16() local
41 float16x8_t vacc89ABCDEF = vaddq_f16(vx89ABCDEF, vthree); in xnn_f16_vhswish_ukernel__neonfp16arith_x16()
42 vx89ABCDEF = vmulq_f16(vx89ABCDEF, vsixth); in xnn_f16_vhswish_ukernel__neonfp16arith_x16()
51 vacc89ABCDEF = vmulq_f16(vacc89ABCDEF, vx89ABCDEF); in xnn_f16_vhswish_ukernel__neonfp16arith_x16()
Dvhswish-f16c-x16.c38 __m256 vx89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (x + 8))); in xnn_f16_vhswish_ukernel__f16c_x16() local
43 __m128i vacc89ABCDEF = _mm256_cvtps_ph(_mm256_add_ps(vx89ABCDEF, vthree), _MM_FROUND_NO_EXC); in xnn_f16_vhswish_ukernel__f16c_x16()
44 …vx89ABCDEF = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vx89ABCDEF, vsixth), _MM_FROUND_NO_EXC)… in xnn_f16_vhswish_ukernel__f16c_x16()
53 …vacc89ABCDEF = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc89ABCDEF), vx89ABCDEF), _MM_FROUN… in xnn_f16_vhswish_ukernel__f16c_x16()
/external/XNNPACK/src/f32-vlrelu/gen/
Dvlrelu-avx-x16.c30 const __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8); in xnn_f32_vlrelu_ukernel__avx_x16() local
34 __m256 vacc89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vslope); in xnn_f32_vlrelu_ukernel__avx_x16()
37 vacc89ABCDEF = _mm256_blendv_ps(vx89ABCDEF, vacc89ABCDEF, vx89ABCDEF); in xnn_f32_vlrelu_ukernel__avx_x16()
/external/XNNPACK/src/f16-vlrelu/gen/
Dvlrelu-neonfp16arith-x16.c32 const float16x8_t vx89ABCDEF = vld1q_f16(i); i += 8; in xnn_f16_vlrelu_ukernel__neonfp16arith_x16() local
36 float16x8_t vacc89ABCDEF = vmulq_f16(vx89ABCDEF, vslope); in xnn_f16_vlrelu_ukernel__neonfp16arith_x16()
37 const uint16x8_t vmask89ABCDEF = vcltq_s16(vreinterpretq_s16_f16(vx89ABCDEF), vmovq_n_s16(0)); in xnn_f16_vlrelu_ukernel__neonfp16arith_x16()
40 vacc89ABCDEF = vbslq_f16(vmask89ABCDEF, vacc89ABCDEF, vx89ABCDEF); in xnn_f16_vlrelu_ukernel__neonfp16arith_x16()
Dvlrelu-f16c-x16.c33 const __m256 vx89ABCDEF = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 8))); in xnn_f16_vlrelu_ukernel__f16c_x16() local
37 __m256 vacc89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vslope); in xnn_f16_vlrelu_ukernel__f16c_x16()
40 vacc89ABCDEF = _mm256_blendv_ps(vx89ABCDEF, vacc89ABCDEF, vx89ABCDEF); in xnn_f16_vlrelu_ukernel__f16c_x16()

123