Home
last modified time | relevance | path

Searched defs:vsum23 (Results 1 – 25 of 67) sorted by relevance

123

/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/q8avgpool/
H A Dmp8x9p8q-neon.c82 const uint16x8_t vsum23 = vaddl_u8(vi2, vi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
135 const uint16x8_t vsum23 = vaddl_u8(vi2, vi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
188 const uint16x8_t vsum23 = vaddl_u8(vi2, vi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
238 const uint16x8_t vsum23 = vaddl_u8(vi2, vi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
312 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3)); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
388 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3)); in pytorch_q8avgpool_ukernel_mp8x9p8q__neon() local
H A Dmp8x9p8q-sse2.c81 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
143 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
204 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
262 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
343 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
421 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in pytorch_q8avgpool_ukernel_mp8x9p8q__sse2() local
/aosp_15_r20/external/XNNPACK/src/f32-gavgpool/
H A D7p7x-minmax-neon-c4.c48 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
80 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
134 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
162 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
H A D7p7x-minmax-wasmsimd-x86-c4.c55 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
95 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
157 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
186 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
H A D7p7x-minmax-wasmsimd-arm-c4.c55 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
95 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
157 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
186 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
H A D7p7x-minmax-sse-c4.c55 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
94 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
156 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
185 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
H A D7p7x-minmax-wasm-c1.c46 const float vsum23 = vi2 + vi3; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1() local
79 const float vsum23 = vi2 + vi3; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1() local
133 const float vsum23 = vi2 + vi3; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1() local
H A D7p7x-minmax-scalar-c1.c46 const float vsum23 = vi2 + vi3; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
79 const float vsum23 = vi2 + vi3; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
133 const float vsum23 = vi2 + vi3; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
/aosp_15_r20/external/pytorch/aten/src/ATen/native/quantized/cpu/qnnpack/src/q8gavgpool/
H A Dmp8x7p7q-neon.c59 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3)); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local
104 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3)); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local
185 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3)); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local
258 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3)); in pytorch_q8gavgpool_ukernel_mp8x7p7q__neon() local
/aosp_15_r20/external/XNNPACK/src/f32-avgpool/
H A D9p8x-minmax-wasmsimd-arm-c4.c104 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
181 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
281 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
310 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
H A D9p8x-minmax-wasmsimd-x86-c4.c104 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
181 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
281 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
310 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
H A D9p8x-minmax-sse-c4.c104 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
180 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
279 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
308 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
H A D9p8x-minmax-neon-c4.c95 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
163 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
254 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
282 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
/aosp_15_r20/external/XNNPACK/src/f16-avgpool/
H A D9p8x-minmax-neonfp16arith-c8.c95 const float16x8_t vsum23 = vaddq_f16(vi2, vi3); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
163 const float16x8_t vsum23 = vaddq_f16(vi2, vi3); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
254 const float16x8_t vsum23 = vaddq_f16(vi2, vi3); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
282 const float16x8_t vsum23 = vaddq_f16(vi2, vi3); in xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
H A D9p8x-minmax-f16c-c8.c106 …const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local
183 …const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local
284 …const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local
313 …const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC)); in xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8() local
/aosp_15_r20/external/XNNPACK/src/f32-pavgpool/
H A D9p8x-minmax-wasmsimd-x86-c4.c104 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
181 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
284 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
313 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
H A D9p8x-minmax-wasmsimd-arm-c4.c104 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
181 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
284 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
313 const v128_t vsum23 = wasm_f32x4_add(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
H A D9p8x-minmax-sse-c4.c104 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
180 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
282 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
311 const __m128 vsum23 = _mm_add_ps(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
H A D9p8x-minmax-neon-c4.c95 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
163 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
255 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
283 const float32x4_t vsum23 = vaddq_f32(vi2, vi3); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
H A D9p8x-minmax-wasm-c1.c95 const float vsum23 = vi2 + vi3; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
164 const float vsum23 = vi2 + vi3; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
256 const float vsum23 = vi2 + vi3; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
H A D9p8x-minmax-scalar-c1.c95 const float vsum23 = vi2 + vi3; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
164 const float vsum23 = vi2 + vi3; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
256 const float vsum23 = vi2 + vi3; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
/aosp_15_r20/external/XNNPACK/src/f16-pavgpool/
H A D9p8x-minmax-neonfp16arith-c8.c95 const float16x8_t vsum23 = vaddq_f16(vi2, vi3); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
163 const float16x8_t vsum23 = vaddq_f16(vi2, vi3); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
255 const float16x8_t vsum23 = vaddq_f16(vi2, vi3); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
283 const float16x8_t vsum23 = vaddq_f16(vi2, vi3); in xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8() local
H A D9p8x-minmax-avx2-c8.c106 …const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local
183 …const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local
286 …const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _MM_FROUND_NO_EXC)); in xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8() local
315 …const __m256 vsum23 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vi2, vi3), _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.c106 const uint16x8_t vsum23 = vaddl_u8(vi2, vi3); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local
179 const uint16x8_t vsum23 = vaddl_u8(vi2, vi3); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local
274 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3)); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local
348 const int16x8_t vsum23 = vreinterpretq_s16_u16(vaddl_u8(vi2, vi3)); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local
H A D9p8x-minmax-sse2-c8.c111 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local
194 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local
300 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local
375 const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local

123