Home
last modified time | relevance | path

Searched refs:vacc1x13 (Results 1 – 25 of 38) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x4c8-xw-minmax-wasmsimd.c108 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__wasmsimd() local
111 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__wasmsimd()
D2x4c8-minmax-wasmsimd-ld128.c112 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld128() local
115 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld128()
D2x4c8-minmax-wasmsimd-ld64.c108 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld64() local
111 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld64()
D2x4c8-xw-minmax-sse2.c93 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2() local
96 …1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x… in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2()
D2x4c8-minmax-sse2-ld128.c95 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128() local
98 …1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128()
D2x4c8-minmax-sse2-ld64.c97 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64() local
100 …1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64()
D2x16c8-minmax-neon-mull-padal.c79 int32x4_t vacc1x13 = vacc0x13; in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
159 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
189 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
250 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c79 int32x4_t vacc1x13 = vacc0x13; in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
205 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
298 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
328 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
389 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D3x4c8-xw-minmax-wasmsimd.c132 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__wasmsimd() local
137 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__wasmsimd()
D3x4c8-minmax-wasmsimd-ld128.c136 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld128() local
141 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld128()
D3x4c8-minmax-wasmsimd-ld64.c132 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64() local
137 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64()
D2x16c16-minmax-neon-mlal-padal.c79 int32x4_t vacc1x13 = vacc0x13; in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
189 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
221 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
282 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x4c8-minmax-sse2-ld64.c114 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64() local
119 …1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
D3x4c8-minmax-sse2-ld128.c112 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128() local
117 …1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
D3x4c8-xw-minmax-sse2.c110 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2() local
115 …1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
/external/XNNPACK/src/qs8-igemm/gen/
D2x4c8-minmax-wasmsimd-ld128.c127 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld128() local
130 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld128()
D2x4c8-minmax-wasmsimd-ld64.c123 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld64() local
126 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld64()
D2x4c8-minmax-sse2-ld128.c110 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128() local
113 …1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128()
D2x4c8-minmax-sse2-ld64.c112 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld64() local
115 …1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld64()
D2x16c8-minmax-neon-mull-padal.c80 int32x4_t vacc1x13 = vacc0x13; in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
172 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
205 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
266 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c80 int32x4_t vacc1x13 = vacc0x13; in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
218 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
311 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
344 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
405 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D3x4c8-minmax-wasmsimd-ld128.c153 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld128() local
158 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld128()
D3x4c8-minmax-wasmsimd-ld64.c149 …const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x… in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64() local
154 …= wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, … in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64()
D3x4c8-minmax-sse2-ld128.c129 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128() local
134 …1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
D2x16c16-minmax-neon-mlal-padal.c80 int32x4_t vacc1x13 = vacc0x13; in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
202 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
237 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
298 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()

12