/external/XNNPACK/src/f16-gavgpool/gen/ |
D | 7p7x-minmax-f16c-c24.c | 56 …__m128i vaccGHIJKLMN = _mm256_cvtps_ph(_mm256_add_ps(vi0xGHIJKLMN, vi1xGHIJKLMN), _MM_FROUND_NO_EX… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 127 __m128i vaccGHIJKLMN = _mm_loadu_si128((const __m128i*) (b + 16)); in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 237 … __m128i vaccGHIJKLMN = _mm_loadu_si128((const __m128i*) buffer); buffer = (uint16_t*) buffer + 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
|
D | 7p7x-minmax-neonfp16arith-c24.c | 55 float16x8_t vaccGHIJKLMN = vaddq_f16(vi0xGHIJKLMN, vi1xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 126 float16x8_t vaccGHIJKLMN = vld1q_f16(b + 16); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 234 float16x8_t vaccGHIJKLMN = vld1q_f16(buffer); buffer = (__fp16*) buffer + 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
|
D | 7p7x-minmax-f16c-c32.c | 58 …__m128i vaccGHIJKLMN = _mm256_cvtps_ph(_mm256_add_ps(vi0xGHIJKLMN, vi1xGHIJKLMN), _MM_FROUND_NO_EX… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 141 __m128i vaccGHIJKLMN = _mm_loadu_si128((const __m128i*) (b + 16)); in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 267 … __m128i vaccGHIJKLMN = _mm_loadu_si128((const __m128i*) buffer); buffer = (uint16_t*) buffer + 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
|
D | 7p7x-minmax-neonfp16arith-c32.c | 57 float16x8_t vaccGHIJKLMN = vaddq_f16(vi0xGHIJKLMN, vi1xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 140 float16x8_t vaccGHIJKLMN = vld1q_f16(b + 16); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 264 float16x8_t vaccGHIJKLMN = vld1q_f16(buffer); buffer = (__fp16*) buffer + 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
|
/external/XNNPACK/src/qu8-gavgpool/gen/ |
D | 7x-minmax-rndnu-neon-c24.c | 137 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c24() local 141 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c24() local
|
D | 7x-minmax-fp32-neon-c24.c | 143 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c24() local 147 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c24() local
|
D | 7x-minmax-fp32-neonv8-c24.c | 136 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() local 140 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() local
|
D | 7x-minmax-rndnu-neon-c32.c | 158 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32() local 163 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32() local
|
D | 7x-minmax-fp32-neonv8-c32.c | 157 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() local 162 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() local
|
D | 7p7x-minmax-fp32-wasmsimd-c24.c | 57 v128_t vaccGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local 171 v128_t vaccGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local 316 v128_t vaccGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local
|
D | 7p7x-minmax-fp32-sse41-c24.c | 58 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 183 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 328 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local
|
/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7x-minmax-rndnu-neon-c24.c | 137 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c24() local 141 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c24() local
|
D | 7x-minmax-fp32-neon-c24.c | 143 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c24() local 147 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c24() local
|
D | 7x-minmax-fp32-neonv8-c24.c | 136 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() local 140 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() local
|
D | 7x-minmax-rndnu-neon-c32.c | 158 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32() local 163 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c32() local
|
D | 7x-minmax-fp32-neonv8-c32.c | 157 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() local 162 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() local
|
D | 7p7x-minmax-fp32-wasmsimd-c24.c | 57 v128_t vaccGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local 171 v128_t vaccGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local 316 v128_t vaccGHIJKLMN = wasm_i16x8_add(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c24() local
|
D | 7p7x-minmax-fp32-sse41-c24.c | 58 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 182 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 326 __m128i vaccGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local
|
/external/XNNPACK/src/qs8-vaddc/gen/ |
D | minmax-avx2-mul32-ld64-x24.c | 42 __m256i vaccGHIJKLMN = _mm256_add_epi32(vbias, _mm256_mullo_epi32(vaGHIJKLMN, va_multiplier)); in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x24() local
|
D | minmax-avx2-mul32-ld64-x32.c | 43 __m256i vaccGHIJKLMN = _mm256_add_epi32(vbias, _mm256_mullo_epi32(vaGHIJKLMN, va_multiplier)); in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x32() local
|
/external/XNNPACK/src/f32-qs8-vcvt/gen/ |
D | vcvt-avx-x24.c | 51 const __m256i vaccGHIJKLMN = _mm256_cvtps_epi32(vxGHIJKLMN); in xnn_f32_qs8_vcvt_ukernel__avx_x24() local
|
D | vcvt-avx-x32.c | 54 const __m256i vaccGHIJKLMN = _mm256_cvtps_epi32(vxGHIJKLMN); in xnn_f32_qs8_vcvt_ukernel__avx_x32() local
|
/external/XNNPACK/src/f32-qu8-vcvt/gen/ |
D | vcvt-avx-x24.c | 51 const __m256i vaccGHIJKLMN = _mm256_cvtps_epi32(vxGHIJKLMN); in xnn_f32_qu8_vcvt_ukernel__avx_x24() local
|
D | vcvt-avx-x32.c | 54 const __m256i vaccGHIJKLMN = _mm256_cvtps_epi32(vxGHIJKLMN); in xnn_f32_qu8_vcvt_ukernel__avx_x32() local
|
/external/XNNPACK/src/qs8-vadd/gen/ |
D | minmax-avx2-mul32-ld64-x24.c | 45 __m256i vaccGHIJKLMN = _mm256_add_epi32(vbias, _mm256_mullo_epi32(vaGHIJKLMN, va_multiplier)); in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x24() local
|