/aosp_15_r20/external/XNNPACK/src/f32-gavgpool/ |
H A D | 7p7x-minmax-neon-c4.c | 47 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local 79 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local 133 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local 161 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
|
H A D | 7p7x-minmax-wasmsimd-x86-c4.c | 54 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local 94 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local 156 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local 185 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
|
H A D | 7p7x-minmax-wasmsimd-arm-c4.c | 54 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local 94 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local 156 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local 185 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
|
H A D | 7p7x-minmax-sse-c4.c | 54 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local 93 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local 155 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local 184 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
|
H A D | 7p7x-minmax-wasm-c1.c | 45 const float vsum01 = vi0 + vi1; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1() local 78 const float vsum01 = vi0 + vi1; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1() local 132 const float vsum01 = vi0 + vi1; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1() local
|
H A D | 7p7x-minmax-scalar-c1.c | 45 const float vsum01 = vi0 + vi1; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1() local 78 const float vsum01 = vi0 + vi1; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1() local 132 const float vsum01 = vi0 + vi1; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
|
/aosp_15_r20/external/XNNPACK/src/f32-avgpool/ |
H A D | 9p8x-minmax-wasmsimd-arm-c4.c | 103 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 180 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 280 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 309 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
|
H A D | 9p8x-minmax-wasmsimd-x86-c4.c | 103 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 180 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 280 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 309 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
|
H A D | 9p8x-minmax-sse-c4.c | 103 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local 179 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local 278 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local 307 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
|
H A D | 9p8x-minmax-neon-c4.c | 94 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local 162 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local 253 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local 281 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
|
H A D | 9p8x-minmax-scalar-c1.c | 94 const float vsum01 = vi0 + vi1; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local 163 const float vsum01 = vi0 + vi1; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local 253 const float vsum01 = vi0 + vi1; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local
|
H A D | 9p8x-minmax-wasm-c1.c | 94 const float vsum01 = vi0 + vi1; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local 163 const float vsum01 = vi0 + vi1; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local 253 const float vsum01 = vi0 + vi1; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local
|
/aosp_15_r20/external/XNNPACK/src/f16-avgpool/ |
H A D | 9p8x-minmax-neonfp16arith-c8.c | 94 const float16x8_t vsum01 = vaddq_f16(vi0, vi1); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local 162 const float16x8_t vsum01 = vaddq_f16(vi0, vi1); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local 253 const float16x8_t vsum01 = vaddq_f16(vi0, vi1); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local 281 const float16x8_t vsum01 = vaddq_f16(vi0, vi1); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
|
H A D | 9p8x-minmax-f16c-c8.c | 105 …const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local 182 …const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local 283 …const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local 312 …const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local
|
/aosp_15_r20/external/XNNPACK/src/f32-pavgpool/ |
H A D | 9p8x-minmax-wasmsimd-x86-c4.c | 103 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 180 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 283 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 312 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
|
H A D | 9p8x-minmax-wasmsimd-arm-c4.c | 103 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 180 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 283 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 312 const v128_t vsum01 = wasm_f32x4_add(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
|
H A D | 9p8x-minmax-sse-c4.c | 103 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local 179 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local 281 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local 310 const __m128 vsum01 = _mm_add_ps(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
|
H A D | 9p8x-minmax-neon-c4.c | 94 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local 162 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local 254 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local 282 const float32x4_t vsum01 = vaddq_f32(vi0, vi1); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
|
H A D | 9p8x-minmax-wasm-c1.c | 94 const float vsum01 = vi0 + vi1; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local 163 const float vsum01 = vi0 + vi1; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local 255 const float vsum01 = vi0 + vi1; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
|
H A D | 9p8x-minmax-scalar-c1.c | 94 const float vsum01 = vi0 + vi1; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local 163 const float vsum01 = vi0 + vi1; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local 255 const float vsum01 = vi0 + vi1; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
|
/aosp_15_r20/external/XNNPACK/src/f16-pavgpool/ |
H A D | 9p8x-minmax-neonfp16arith-c8.c | 94 const float16x8_t vsum01 = vaddq_f16(vi0, vi1); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local 162 const float16x8_t vsum01 = vaddq_f16(vi0, vi1); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local 254 const float16x8_t vsum01 = vaddq_f16(vi0, vi1); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local 282 const float16x8_t vsum01 = vaddq_f16(vi0, vi1); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
|
H A D | 9p8x-minmax-avx2-c8.c | 105 …const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local 182 …const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local 285 …const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local 314 …const __m256 vsum01 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi0, vi1), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local
|
/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/q8avgpool/ |
H A D | mp8x9p8q-neon.c | 187 const uint16x8_t vsum01 = vaddl_u8(vi0, vi1); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local 237 const uint16x8_t vsum01 = vaddl_u8(vi0, vi1); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local 311 const int16x8_t vsum01 = vreinterpretq_s16_u16(vaddl_u8(vi0, vi1)); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local 387 const int16x8_t vsum01 = vreinterpretq_s16_u16(vaddl_u8(vi0, vi1)); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
|
H A D | mp8x9p8q-sse2.c | 203 const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local 261 const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local 342 const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local 420 const __m128i vsum01 = _mm_add_epi16(vxi0, vxi1); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
|
/aosp_15_r20/external/XNNPACK/src/qu8-avgpool/ |
H A D | 9p8x-minmax-scalar-c1.c | 101 const uint32_t vsum01 = vi0 + vi1; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() local 172 const uint32_t vsum01 = vi0 + vi1; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() local 264 const uint32_t vsum01 = vi0 + vi1; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() local
|