/external/XNNPACK/src/qu8-gavgpool/gen/ |
D | 7p7x-minmax-fp32-wasmsimd-c16.c | 84 const v128_t vaccCDEF = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_high_u16x8(vacc89ABCDEF)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c16() local 145 v128_t vaccCDEF = wasm_v128_load(b + 12); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c16() local 232 v128_t vaccCDEF = wasm_v128_load(buffer + 12); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c16() local
|
D | 7p7x-minmax-fp32-neonv8-c16.c | 77 …const int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget… in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local 127 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local 203 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
|
D | 7p7x-minmax-fp32-sse41-c16.c | 86 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c16() local 153 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c16() local 240 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c16() local
|
D | 7p7x-minmax-rndnu-neon-c16.c | 76 …const int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget… in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local 126 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local 204 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
|
D | 7p7x-minmax-fp32-neon-c16.c | 76 …const int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget… in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local 126 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local 203 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
|
D | 7p7x-minmax-fp32-sse2-c16.c | 106 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local 192 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local 298 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local
|
D | 7p7x-minmax-fp32-wasmsimd-c24.c | 97 const v128_t vaccCDEF = wasm_i32x4_add(vinit_bias, wasm_u32x4_extend_high_u16x8(vacc89ABCDEF)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local 211 v128_t vaccCDEF = wasm_v128_load(b + 12); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local 356 v128_t vaccCDEF = wasm_v128_load(buffer + 12); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local
|
D | 7p7x-minmax-fp32-neonv8-c24.c | 90 …const int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget… in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 182 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 305 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
|
D | 7p7x-minmax-fp32-sse41-c24.c | 99 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 224 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 369 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local
|
D | 7p7x-minmax-rndnu-neon-c24.c | 89 …const int32x4_t vaccCDEF = vreinterpretq_s32_u32(vaddw_u16(vreinterpretq_u32_s32(vinit_bias), vget… in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 181 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 306 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
|
/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-fp32-wasmsimd-c16.c | 84 const v128_t vaccCDEF = wasm_i32x4_add(vinit_bias, wasm_i32x4_extend_high_i16x8(vacc89ABCDEF)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c16() local 145 v128_t vaccCDEF = wasm_v128_load(b + 12); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c16() local 232 v128_t vaccCDEF = wasm_v128_load(buffer + 12); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c16() local
|
D | 7p7x-minmax-rndnu-neon-c16.c | 76 const int32x4_t vaccCDEF = vaddw_s16(vinit_bias, vget_high_s16(vsum89ABCDEF)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local 126 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local 204 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c16() local
|
D | 7p7x-minmax-fp32-sse41-c16.c | 85 __m128i vaccCDEF = _mm_srai_epi32(_mm_unpackhi_epi16(vacc89ABCDEF, vacc89ABCDEF), 16); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c16() local 151 __m128i vaccCDEF = _mm_srai_epi32(_mm_unpackhi_epi16(vacc89ABCDEF, vacc89ABCDEF), 16); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c16() local 237 __m128i vaccCDEF = _mm_srai_epi32(_mm_unpackhi_epi16(vacc89ABCDEF, vacc89ABCDEF), 16); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c16() local
|
D | 7p7x-minmax-fp32-neonv8-c16.c | 77 const int32x4_t vaccCDEF = vaddw_s16(vinit_bias, vget_high_s16(vsum89ABCDEF)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local 127 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local 203 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c16() local
|
D | 7p7x-minmax-fp32-neon-c16.c | 76 const int32x4_t vaccCDEF = vaddw_s16(vinit_bias, vget_high_s16(vsum89ABCDEF)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local 126 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local 203 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c16() local
|
D | 7p7x-minmax-fp32-wasmsimd-c24.c | 97 const v128_t vaccCDEF = wasm_i32x4_add(vinit_bias, wasm_i32x4_extend_high_i16x8(vacc89ABCDEF)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local 211 v128_t vaccCDEF = wasm_v128_load(b + 12); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local 356 v128_t vaccCDEF = wasm_v128_load(buffer + 12); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local
|
D | 7p7x-minmax-fp32-sse41-c24.c | 98 __m128i vaccCDEF = _mm_srai_epi32(_mm_unpackhi_epi16(vacc89ABCDEF, vacc89ABCDEF), 16); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 222 __m128i vaccCDEF = _mm_srai_epi32(_mm_unpackhi_epi16(vacc89ABCDEF, vacc89ABCDEF), 16); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 366 __m128i vaccCDEF = _mm_srai_epi32(_mm_unpackhi_epi16(vacc89ABCDEF, vacc89ABCDEF), 16); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local
|
D | 7p7x-minmax-fp32-neonv8-c24.c | 90 const int32x4_t vaccCDEF = vaddw_s16(vinit_bias, vget_high_s16(vsum89ABCDEF)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 182 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 305 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
|
D | 7p7x-minmax-fp32-sse2-c16.c | 107 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vsgnacc89ABCDEF); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local 195 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vsgnacc89ABCDEF); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local 303 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vsgnacc89ABCDEF); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c16() local
|
D | 7p7x-minmax-rndnu-neon-c24.c | 89 const int32x4_t vaccCDEF = vaddw_s16(vinit_bias, vget_high_s16(vsum89ABCDEF)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 181 int32x4_t vaccCDEF = vld1q_s32(b + 12); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 306 int32x4_t vaccCDEF = vld1q_s32(buffer); buffer += 4; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
|
/external/XNNPACK/src/f32-vrelu/gen/ |
D | vrelu-wasmsimd-x16.c | 35 v128_t vaccCDEF = wasm_v128_load(x + 12); in xnn_f32_vrelu_ukernel__wasmsimd_x16() local
|
/external/XNNPACK/src/f32-vhswish/gen/ |
D | vhswish-wasmsimd-x16.c | 45 v128_t vaccCDEF = wasm_f32x4_add(vxCDEF, vthree); in xnn_f32_vhswish_ukernel__wasmsimd_x16() local
|
D | vhswish-neon-x16.c | 44 float32x4_t vaccCDEF = vaddq_f32(vxCDEF, vthree); in xnn_f32_vhswish_ukernel__neon_x16() local
|
/external/XNNPACK/src/qu8-vaddc/gen/ |
D | minmax-wasmsimd-x16.c | 41 …v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_u32x4_extend_high_u16x8(va89ABCDEF), v… in xnn_qu8_vaddc_minmax_ukernel__wasmsimd_x16() local
|
/external/XNNPACK/src/qs8-vaddc/gen/ |
D | minmax-wasmsimd-x16.c | 41 …v128_t vaccCDEF = wasm_i32x4_add(vbias, wasm_i32x4_mul(wasm_i32x4_extend_high_i16x8(va89ABCDEF), v… in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16() local
|