Home
last modified time | relevance | path

Searched refs:vq31prod4567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup8x9-minmax-wasmsimd-mul16.c186 const v128_t vq31prod4567 = wasm_v32x4_shuffle(vprod45, vprod67, 1, 3, 5, 7); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() local
190 …v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
195 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod4567, vshift), wasm_i32x4_gt(vrem4567, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
302 const v128_t vq31prod4567 = wasm_v32x4_shuffle(vprod45, vprod67, 1, 3, 5, 7); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() local
306 …v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
311 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod4567, vshift), wasm_i32x4_gt(vrem4567, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16()
Dup16x9-minmax-wasmsimd-mul16.c243 const v128_t vq31prod4567 = wasm_v32x4_shuffle(vprod45, vprod67, 1, 3, 5, 7); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() local
249 …v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
256 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod4567, vshift), wasm_i32x4_gt(vrem4567, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
378 const v128_t vq31prod4567 = wasm_v32x4_shuffle(vprod45, vprod67, 1, 3, 5, 7); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() local
382 …v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
387 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod4567, vshift), wasm_i32x4_gt(vrem4567, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16()
Dup8x9-minmax-sse41-mul16.c222 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
228 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
235 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
382 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
388 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
395 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
Dup8x9-minmax-sse2-mul16.c243 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
249 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
256 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
423 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
429 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
436 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
Dup8x9-minmax-ssse3-mul16.c243 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
249 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
256 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
423 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
429 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
436 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
Dup24x9-minmax-wasmsimd-mul16.c300 const v128_t vq31prod4567 = wasm_v32x4_shuffle(vprod45, vprod67, 1, 3, 5, 7); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() local
308 …v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
317 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod4567, vshift), wasm_i32x4_gt(vrem4567, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
444 const v128_t vq31prod4567 = wasm_v32x4_shuffle(vprod45, vprod67, 1, 3, 5, 7); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() local
448 …v128_t vrem4567 = wasm_i32x4_add(wasm_v128_and(vq31prod4567, vremainder_mask), wasm_i32x4_shr(vq31… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
453 …vacc4567 = wasm_i32x4_sub(wasm_i32x4_shr(vq31prod4567, vshift), wasm_i32x4_gt(vrem4567, vthreshold… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16()
Dup16x9-minmax-sse41-mul16.c306 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
314 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
325 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
490 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
496 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
503 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
Dup16x9-minmax-ssse3-mul16.c343 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
351 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
362 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
547 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
553 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
560 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c343 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
351 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
362 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
547 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
553 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
560 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup24x9-minmax-sse41-mul16.c390 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
400 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
415 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
588 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
594 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
601 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
Dup24x9-minmax-ssse3-mul16.c443 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
453 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
468 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
661 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
667 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
674 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c443 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
453 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
468 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
661 const __m128i vq31prod4567 = _mm_shuffle_epi32(vq31prod4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
667 …_mm_add_epi32(_mm_and_si128(vq31prod4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), v… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
674 …_mm_sub_epi32(_mm_sra_epi32(vq31prod4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold)… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()