| /aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/q8avgpool/ |
| H A D | mp8x9p8q-neon.c | 68 const uint8x8_t vi3 = vld1_u8(i3); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
121 const uint8x8_t vi3 = vreinterpret_u8_u64( in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
174 const uint8x8_t vi3 = vld1_u8(i3); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
224 const uint8x8_t vi3 = vreinterpret_u8_u64( in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
296 const uint8x8_t vi3 = vld1_u8(i3); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
373 const uint8x8_t vi3 = vreinterpret_u8_u64( in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
|
| H A D | mp8x9p8q-sse2.c | 57 const __m128i vi3 = _mm_loadl_epi64((const __m128i*)i3); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
119 const __m128i vi3 = in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
181 const __m128i vi3 = _mm_loadl_epi64((const __m128i*)i3); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
239 const __m128i vi3 = in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
319 const __m128i vi3 = _mm_loadl_epi64((const __m128i*)i3); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
398 const __m128i vi3 = in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-gavgpool/ |
| H A D | 7p7x-minmax-neon-c4.c | 42 const float32x4_t vi3 = vld1q_f32(i3); i3 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
73 const float32x4_t vi3 = vld1q_f32(i3); i3 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
127 const float32x4_t vi3 = vld1q_f32(i3); i3 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
155 const float32x4_t vi3 = vld1q_f32(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
|
| H A D | 7p7x-minmax-wasmsimd-x86-c4.c | 45 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
84 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
145 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
179 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
|
| H A D | 7p7x-minmax-wasmsimd-arm-c4.c | 45 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
84 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
145 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
179 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
|
| H A D | 7p7x-minmax-sse-c4.c | 45 const __m128 vi3 = _mm_loadu_ps(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
83 const __m128 vi3 = _mm_loadu_ps(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
144 const __m128 vi3 = _mm_loadu_ps(i3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
178 const __m128 vi3 = _mm_loadu_ps(i3); 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 | 48 const uint8x8_t vi3 = vld1_u8(i3); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local
91 const uint8x8_t vi3 = vld1_u8(i3); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local
170 const uint8x8_t vi3 = vld1_u8(i3); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local
244 const uint8x8_t vi3 = in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local
|
| H A D | mp8x7p7q-sse2.c | 50 const __m128i vi3 = _mm_loadl_epi64((const __m128i*)i3); in pytorch_q8gavgpool_ukernel_mp8x7p7q__sse2() local
104 const __m128i vi3 = _mm_loadl_epi64((const __m128i*)i3); in pytorch_q8gavgpool_ukernel_mp8x7p7q__sse2() local
180 const __m128i vi3 = _mm_loadl_epi64((const __m128i*)i3); in pytorch_q8gavgpool_ukernel_mp8x7p7q__sse2() local
256 const __m128i vi3 = in pytorch_q8gavgpool_ukernel_mp8x7p7q__sse2() local
|
| /aosp_15_r20/external/XNNPACK/src/u8-maxpool/ |
| H A D | 9p8x-minmax-neon-c16.c | 84 const uint8x16_t vi3 = vld1q_u8(i3); i3 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local
107 const uint8x16_t vi3 = vld1q_u8(i3); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local
188 const uint8x16_t vi3 = vld1q_u8(i3); i3 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local
211 const uint8x16_t vi3 = vld1q_u8(i3); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local
|
| H A D | 9p8x-minmax-sse2-c16.c | 86 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
110 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
195 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
220 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
|
| H A D | 9p8x-minmax-wasmsimd-c16.c | 85 const v128_t vi3 = wasm_v128_load(i3); in xnn_u8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local
115 const v128_t vi3 = wasm_v128_load(i3); in xnn_u8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local
204 const v128_t vi3 = wasm_v128_load(i3); in xnn_u8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local
234 const v128_t vi3 = wasm_v128_load(i3); 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 | 78 const __m128i vi3 = _mm_loadu_si128((const __m128i*)i3); in pytorch_u8maxpool_ukernel_16x9p8q__sse2() local
123 const __m128i vi3 = _mm_loadu_si128((const __m128i*)i3); in pytorch_u8maxpool_ukernel_16x9p8q__sse2() local
186 const __m128i vi3 = _mm_loadu_si128((const __m128i*)i3); in pytorch_u8maxpool_ukernel_16x9p8q__sse2() local
229 const __m128i vi3 = _mm_loadu_si128((const __m128i*)i3); in pytorch_u8maxpool_ukernel_16x9p8q__sse2() local
|
| H A D | 16x9p8q-neon.c | 75 const uint8x16_t vi3 = vld1q_u8(i3); in pytorch_u8maxpool_ukernel_16x9p8q__neon() local
120 const uint8x16_t vi3 = vld1q_u8(i3); in pytorch_u8maxpool_ukernel_16x9p8q__neon() local
183 const uint8x16_t vi3 = vld1q_u8(i3); in pytorch_u8maxpool_ukernel_16x9p8q__neon() local
226 const uint8x16_t vi3 = vld1q_u8(i3); in pytorch_u8maxpool_ukernel_16x9p8q__neon() local
|
| /aosp_15_r20/external/XNNPACK/src/s8-maxpool/ |
| H A D | 9p8x-minmax-sse41-c16.c | 83 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16; in xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16() local
107 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16() local
192 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); i3 += 16; in xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16() local
217 const __m128i vi3 = _mm_loadu_si128((const __m128i*) i3); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16() local
|
| H A D | 9p8x-minmax-neon-c16.c | 81 const int8x16_t vi3 = vld1q_s8(i3); i3 += 16; in xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16() local
105 const int8x16_t vi3 = vld1q_s8(i3); in xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16() local
187 const int8x16_t vi3 = vld1q_s8(i3); i3 += 16; in xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16() local
211 const int8x16_t vi3 = vld1q_s8(i3); in xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16() local
|
| H A D | 9p8x-minmax-wasmsimd-c16.c | 85 const v128_t vi3 = wasm_v128_load(i3); in xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local
115 const v128_t vi3 = wasm_v128_load(i3); in xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local
204 const v128_t vi3 = wasm_v128_load(i3); in xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local
234 const v128_t vi3 = wasm_v128_load(i3); in xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16() local
|
| H A D | 9p8x-minmax-sse2-c16.c | 87 const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
118 const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
207 const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
238 const __m128i vi3 = _mm_xor_si128(_mm_loadu_si128((const __m128i*) i3), vbias); in xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-maxpool/ |
| H A D | 9p8x-minmax-wasmsimd-arm-c4.c | 84 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
117 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
199 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
228 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
|
| H A D | 9p8x-minmax-neon-c4.c | 81 const float32x4_t vi3 = vld1q_f32(i3); i3 += 4; in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local
104 const float32x4_t vi3 = vld1q_f32(i3); i3 += 4; in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local
177 const float32x4_t vi3 = vld1q_f32(i3); i3 += 4; in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local
200 const float32x4_t vi3 = vld1q_f32(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local
|
| H A D | 9p8x-minmax-sse-c4.c | 84 const __m128 vi3 = _mm_loadu_ps(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local
117 const __m128 vi3 = _mm_loadu_ps(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local
200 const __m128 vi3 = _mm_loadu_ps(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local
229 const __m128 vi3 = _mm_loadu_ps(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local
|
| H A D | 9p8x-minmax-wasmsimd-x86-c4.c | 84 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
120 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
205 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
237 const v128_t vi3 = wasm_v128_load(i3); 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 | 81 const float16x8_t vi3 = vld1q_f16(i3); i3 += 8; in xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
104 const float16x8_t vi3 = vld1q_f16(i3); i3 += 8; in xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
181 const float16x8_t vi3 = vld1q_f16(i3); i3 += 8; in xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
204 const float16x8_t vi3 = vld1q_f16(i3); in xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
|
| H A D | 9p8x-minmax-f16c-c8.c | 85 const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3)); in xnn_f16_maxpool_minmax_ukernel_9p8x__f16c_c8() local
118 const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3)); in xnn_f16_maxpool_minmax_ukernel_9p8x__f16c_c8() local
207 const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3)); in xnn_f16_maxpool_minmax_ukernel_9p8x__f16c_c8() local
236 const __m256 vi3 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3)); 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 | 90 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
168 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
267 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
302 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
|
| H A D | 9p8x-minmax-wasmsimd-x86-c4.c | 90 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
168 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
267 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
302 const v128_t vi3 = wasm_v128_load(i3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
|