Home
last modified time | relevance | path

Searched defs:vsum67 (Results 1 – 25 of 46) sorted by relevance

12

/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/q8avgpool/
H A Dmp8x9p8q-neon.c84 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
137 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
190 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
240 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
314 const int16x8_t vsum67 = vreinterpretq_s16_u16(vaddl_u8(vi6, vi7)); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
390 const int16x8_t vsum67 = vreinterpretq_s16_u16(vaddl_u8(vi6, vi7)); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
H A Dmp8x9p8q-sse2.c83 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
145 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
206 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
264 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
345 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
423 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
/aosp_15_r20/external/XNNPACK/src/f32-avgpool/
H A D9p8x-minmax-wasmsimd-arm-c4.c106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
283 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
312 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
H A D9p8x-minmax-wasmsimd-x86-c4.c106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
283 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
312 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
H A D9p8x-minmax-sse-c4.c106 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
182 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
281 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
310 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
H A D9p8x-minmax-neon-c4.c97 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
165 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
256 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
284 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
H A D9p8x-minmax-scalar-c1.c97 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local
166 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local
256 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local
H A D9p8x-minmax-wasm-c1.c97 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local
166 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local
256 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local
H A D9x-minmax-sse-c4.c130 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__sse_c4() local
158 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__sse_c4() local
H A D9x-minmax-wasmsimd-arm-c4.c130 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_arm_c4() local
159 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_arm_c4() local
H A D9x-minmax-wasmsimd-x86-c4.c130 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_x86_c4() local
159 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_x86_c4() local
/aosp_15_r20/external/XNNPACK/src/f16-avgpool/
H A D9p8x-minmax-neonfp16arith-c8.c97 const float16x8_t vsum67 = vaddq_f16(vi6, vi7); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
165 const float16x8_t vsum67 = vaddq_f16(vi6, vi7); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
256 const float16x8_t vsum67 = vaddq_f16(vi6, vi7); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
284 const float16x8_t vsum67 = vaddq_f16(vi6, vi7); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
H A D9p8x-minmax-f16c-c8.c108 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local
185 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local
286 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local
315 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local
H A D9x-minmax-f16c-c8.c132 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9x__f16c_c8() local
161 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9x__f16c_c8() local
/aosp_15_r20/external/XNNPACK/src/f32-pavgpool/
H A D9p8x-minmax-wasmsimd-x86-c4.c106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
286 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
315 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
H A D9p8x-minmax-wasmsimd-arm-c4.c106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
286 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
315 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
H A D9p8x-minmax-sse-c4.c106 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
182 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
284 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
313 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
H A D9p8x-minmax-neon-c4.c97 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
165 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
257 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
285 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
H A D9p8x-minmax-wasm-c1.c97 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
166 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
258 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
H A D9p8x-minmax-scalar-c1.c97 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
166 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
258 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
/aosp_15_r20/external/XNNPACK/src/f16-pavgpool/
H A D9p8x-minmax-neonfp16arith-c8.c97 const float16x8_t vsum67 = vaddq_f16(vi6, vi7); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
165 const float16x8_t vsum67 = vaddq_f16(vi6, vi7); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
257 const float16x8_t vsum67 = vaddq_f16(vi6, vi7); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
285 const float16x8_t vsum67 = vaddq_f16(vi6, vi7); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
H A D9p8x-minmax-avx2-c8.c108 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local
185 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local
288 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local
317 …const __m256 vsum67 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi6, vi7), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local
/aosp_15_r20/external/XNNPACK/src/qu8-avgpool/
H A D9p8x-minmax-neon-c8.c108 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local
181 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local
276 const int16x8_t vsum67 = vreinterpretq_s16_u16(vaddl_u8(vi6, vi7)); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local
350 const int16x8_t vsum67 = vreinterpretq_s16_u16(vaddl_u8(vi6, vi7)); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local
H A D9p8x-minmax-sse2-c8.c113 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local
196 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local
302 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local
377 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local
H A D9p8x-minmax-scalar-c1.c104 const uint32_t vsum67 = vi6 + vi7; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() local
175 const uint32_t vsum67 = vi6 + vi7; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() local
267 const uint32_t vsum67 = vi6 + vi7; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() local

12