/aosp_15_r20/external/XNNPACK/src/s8-maxpool/ |
H A D | 4p3x-minmax-neon-c16.c | 55 const int8x16_t vi2 = vld1q_s8(i2); i2 += 16; in xnn_s8_maxpool_minmax_ukernel_4p3x__neon_c16() local 69 const int8x16_t vi2 = vld1q_s8(i2); in xnn_s8_maxpool_minmax_ukernel_4p3x__neon_c16() local 116 const int8x16_t vi2 = vld1q_s8(i2); i2 += 16; in xnn_s8_maxpool_minmax_ukernel_4p3x__neon_c16() local 130 const int8x16_t vi2 = vld1q_s8(i2); in xnn_s8_maxpool_minmax_ukernel_4p3x__neon_c16() local
|
H A D | 9p8x-minmax-sse41-c16.c | 82 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16; in xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16() local 106 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16() local 191 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16; in xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16() local 216 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16() local
|
H A D | 9p8x-minmax-neon-c16.c | 80 const int8x16_t vi2 = vld1q_s8(i2); i2 += 16; in xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16() local 104 const int8x16_t vi2 = vld1q_s8(i2); in xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16() local 186 const int8x16_t vi2 = vld1q_s8(i2); i2 += 16; in xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16() local 210 const int8x16_t vi2 = vld1q_s8(i2); in xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16() local
|
H A D | 9p8x-minmax-wasmsimd-c16.c | 83 const v128_t vi2 = wasm_v128_load(i2); in xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local 114 const v128_t vi2 = wasm_v128_load(i2); in xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local 202 const v128_t vi2 = wasm_v128_load(i2); in xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local 233 const v128_t vi2 = wasm_v128_load(i2); in xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local
|
H A D | 9p8x-minmax-sse2-c16.c | 85 const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16() local 117 const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16() local 205 const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16() local 237 const __m128i vi2 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i2), vbias); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
|
/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/q8avgpool/ |
H A D | mp8x9p8q-neon.c | 66 const uint8x8_t vi2 = vld1_u8(i2); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local 119 const uint8x8_t vi2 = vreinterpret_u8_u64( in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local 172 const uint8x8_t vi2 = vld1_u8(i2); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local 222 const uint8x8_t vi2 = vreinterpret_u8_u64( in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local 294 const uint8x8_t vi2 = vld1_u8(i2); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local 371 const uint8x8_t vi2 = vreinterpret_u8_u64( in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
|
H A D | mp8x9p8q-sse2.c | 55 const __m128i vi2 = _mm_loadl_epi64((const __m128i*)i2); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local 117 const __m128i vi2 = in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local 179 const __m128i vi2 = _mm_loadl_epi64((const __m128i*)i2); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local 237 const __m128i vi2 = in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local 317 const __m128i vi2 = _mm_loadl_epi64((const __m128i*)i2); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local 396 const __m128i vi2 = in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
|
/aosp_15_r20/external/XNNPACK/src/f32-gavgpool/ |
H A D | 7p7x-minmax-neon-c4.c | 41 const float32x4_t vi2 = vld1q_f32(i2); i2 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local 72 const float32x4_t vi2 = vld1q_f32(i2); i2 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local 126 const float32x4_t vi2 = vld1q_f32(i2); i2 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local 154 const float32x4_t vi2 = vld1q_f32(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
|
H A D | 7p7x-minmax-wasmsimd-x86-c4.c | 43 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local 82 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local 143 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local 178 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
|
H A D | 7p7x-minmax-wasmsimd-arm-c4.c | 43 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local 82 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local 143 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local 178 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
|
H A D | 7p7x-minmax-sse-c4.c | 43 const __m128 vi2 = _mm_loadu_ps(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local 81 const __m128 vi2 = _mm_loadu_ps(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local 142 const __m128 vi2 = _mm_loadu_ps(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local 177 const __m128 vi2 = _mm_loadu_ps(i2); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
|
/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/q8gavgpool/ |
H A D | mp8x7p7q-neon.c | 46 const uint8x8_t vi2 = vld1_u8(i2); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local 89 const uint8x8_t vi2 = vld1_u8(i2); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local 168 const uint8x8_t vi2 = vld1_u8(i2); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local 242 const uint8x8_t vi2 = in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local
|
H A D | mp8x7p7q-sse2.c | 48 const __m128i vi2 = _mm_loadl_epi64((const __m128i*)i2); in pytorch_q8gavgpool_ukernel_mp8x7p7q__sse2() local 102 const __m128i vi2 = _mm_loadl_epi64((const __m128i*)i2); in pytorch_q8gavgpool_ukernel_mp8x7p7q__sse2() local 178 const __m128i vi2 = _mm_loadl_epi64((const __m128i*)i2); in pytorch_q8gavgpool_ukernel_mp8x7p7q__sse2() local 254 const __m128i vi2 = in pytorch_q8gavgpool_ukernel_mp8x7p7q__sse2() local
|
/aosp_15_r20/external/XNNPACK/src/u8-maxpool/ |
H A D | 9p8x-minmax-neon-c16.c | 83 const uint8x16_t vi2 = vld1q_u8(i2); i2 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local 106 const uint8x16_t vi2 = vld1q_u8(i2); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local 187 const uint8x16_t vi2 = vld1q_u8(i2); i2 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local 210 const uint8x16_t vi2 = vld1q_u8(i2); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local
|
H A D | 9p8x-minmax-sse2-c16.c | 85 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local 109 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local 194 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); i2 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local 219 const __m128i vi2 = _mm_loadu_si128((const __m128i*) i2); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
|
H A D | 9p8x-minmax-wasmsimd-c16.c | 83 const v128_t vi2 = wasm_v128_load(i2); in xnn_u8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local 114 const v128_t vi2 = wasm_v128_load(i2); in xnn_u8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local 202 const v128_t vi2 = wasm_v128_load(i2); in xnn_u8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local 233 const v128_t vi2 = wasm_v128_load(i2); in xnn_u8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local
|
/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/u8maxpool/ |
H A D | 16x9p8q-sse2.c | 76 const __m128i vi2 = _mm_loadu_si128((const __m128i*)i2); in pytorch_u8maxpool_ukernel_16x9p8q__sse2() local 122 const __m128i vi2 = _mm_loadu_si128((const __m128i*)i2); in pytorch_u8maxpool_ukernel_16x9p8q__sse2() local 184 const __m128i vi2 = _mm_loadu_si128((const __m128i*)i2); in pytorch_u8maxpool_ukernel_16x9p8q__sse2() local 228 const __m128i vi2 = _mm_loadu_si128((const __m128i*)i2); in pytorch_u8maxpool_ukernel_16x9p8q__sse2() local
|
H A D | 16x9p8q-neon.c | 73 const uint8x16_t vi2 = vld1q_u8(i2); in pytorch_u8maxpool_ukernel_16x9p8q__neon() local 119 const uint8x16_t vi2 = vld1q_u8(i2); in pytorch_u8maxpool_ukernel_16x9p8q__neon() local 181 const uint8x16_t vi2 = vld1q_u8(i2); in pytorch_u8maxpool_ukernel_16x9p8q__neon() local 225 const uint8x16_t vi2 = vld1q_u8(i2); in pytorch_u8maxpool_ukernel_16x9p8q__neon() local
|
/aosp_15_r20/external/XNNPACK/src/f32-maxpool/ |
H A D | 9p8x-minmax-wasmsimd-arm-c4.c | 82 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 115 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 197 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 227 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
|
H A D | 9p8x-minmax-neon-c4.c | 80 const float32x4_t vi2 = vld1q_f32(i2); i2 += 4; in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local 103 const float32x4_t vi2 = vld1q_f32(i2); i2 += 4; in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local 176 const float32x4_t vi2 = vld1q_f32(i2); i2 += 4; in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local 199 const float32x4_t vi2 = vld1q_f32(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local
|
H A D | 9p8x-minmax-sse-c4.c | 82 const __m128 vi2 = _mm_loadu_ps(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local 115 const __m128 vi2 = _mm_loadu_ps(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local 198 const __m128 vi2 = _mm_loadu_ps(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local 228 const __m128 vi2 = _mm_loadu_ps(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local
|
H A D | 9p8x-minmax-wasmsimd-x86-c4.c | 82 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 118 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 203 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 236 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
|
/aosp_15_r20/external/XNNPACK/src/f16-maxpool/ |
H A D | 9p8x-minmax-neonfp16arith-c8.c | 80 const float16x8_t vi2 = vld1q_f16(i2); i2 += 8; in xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8() local 103 const float16x8_t vi2 = vld1q_f16(i2); i2 += 8; in xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8() local 180 const float16x8_t vi2 = vld1q_f16(i2); i2 += 8; in xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8() local 203 const float16x8_t vi2 = vld1q_f16(i2); in xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
|
H A D | 9p8x-minmax-f16c-c8.c | 83 const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_maxpool_minmax_ukernel_9p8x__f16c_c8() local 116 const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_maxpool_minmax_ukernel_9p8x__f16c_c8() local 205 const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_maxpool_minmax_ukernel_9p8x__f16c_c8() local 235 const __m256 vi2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_maxpool_minmax_ukernel_9p8x__f16c_c8() local
|
/aosp_15_r20/external/XNNPACK/src/f32-avgpool/ |
H A D | 9p8x-minmax-wasmsimd-arm-c4.c | 88 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 166 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 265 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 301 const v128_t vi2 = wasm_v128_load(i2); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
|