| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 6x8-minmax-neonfma-dup-ld128.c | 170 const float32x4_t va5c3 = vdupq_lane_f32(vget_high_f32(va5), 1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld128() local
176 vacc5x0123 = vfmaq_f32(vacc5x0123, va5c3, vb0123c3); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld128()
182 vacc5x4567 = vfmaq_f32(vacc5x4567, va5c3, vb4567c3); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld128()
|
| D | 6x8-minmax-wasmsimd-arm-splat.c | 174 const v128_t va5c3 = wasm_v32x4_shuffle(va5, va5, 3, 3, 3, 3); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_splat() local
184 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c3, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_splat()
190 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c3, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_splat()
|
| D | 6x8-minmax-neon-dup-ld128.c | 170 const float32x4_t va5c3 = vdupq_lane_f32(vget_high_f32(va5), 1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld128() local
176 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c3, vb0123c3); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld128()
182 vacc5x4567 = vmlaq_f32(vacc5x4567, va5c3, vb4567c3); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld128()
|
| D | 6x8-minmax-wasmsimd-x86-splat.c | 172 const v128_t va5c3 = wasm_v32x4_shuffle(va5, va5, 3, 3, 3, 3); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_splat() local
182 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c3, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_splat()
188 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c3, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_splat()
|
| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 6x8-minmax-neon-dup-ld128.c | 198 const float32x4_t va5c3 = vdupq_lane_f32(vget_high_f32(va5), 1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld128() local
204 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c3, vb0123c3); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld128()
210 vacc5x4567 = vmlaq_f32(vacc5x4567, va5c3, vb4567c3); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld128()
|
| D | 6x8-minmax-neonfma-dup-ld128.c | 198 const float32x4_t va5c3 = vdupq_lane_f32(vget_high_f32(va5), 1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld128() local
204 vacc5x0123 = vfmaq_f32(vacc5x0123, va5c3, vb0123c3); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld128()
210 vacc5x4567 = vfmaq_f32(vacc5x4567, va5c3, vb4567c3); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld128()
|
| D | 6x8-minmax-wasmsimd-arm-splat.c | 202 const v128_t va5c3 = wasm_v32x4_shuffle(va5, va5, 3, 3, 3, 3); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_splat() local
212 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c3, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_splat()
218 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c3, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_splat()
|
| D | 6x8-minmax-wasmsimd-x86-splat.c | 200 const v128_t va5c3 = wasm_v32x4_shuffle(va5, va5, 3, 3, 3, 3); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_splat() local
210 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c3, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_splat()
216 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c3, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_splat()
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 6x8inc-minmax-neon-dup-ld128.c | 172 const float32x4_t va5c3 = vdupq_lane_f32(vget_high_f32(va5), 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld128() local
178 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c3, vb0123c3); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld128()
184 vacc5x4567 = vmlaq_f32(vacc5x4567, va5c3, vb4567c3); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld128()
|
| D | 6x8inc-minmax-neonfma-dup-ld128.c | 172 const float32x4_t va5c3 = vdupq_lane_f32(vget_high_f32(va5), 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld128() local
178 vacc5x0123 = vfmaq_f32(vacc5x0123, va5c3, vb0123c3); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld128()
184 vacc5x4567 = vfmaq_f32(vacc5x4567, va5c3, vb4567c3); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld128()
|
| D | 6x8inc-minmax-wasmsimd-x86-splat.c | 174 const v128_t va5c3 = wasm_v32x4_shuffle(va5, va5, 3, 3, 3, 3); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_splat() local
184 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c3, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_splat()
190 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c3, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_splat()
|
| D | 6x8inc-minmax-wasmsimd-arm-splat.c | 176 const v128_t va5c3 = wasm_v32x4_shuffle(va5, va5, 3, 3, 3, 3); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_arm_splat() local
186 vacc5x0123 = wasm_f32x4_add(vacc5x0123, wasm_f32x4_mul(va5c3, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_arm_splat()
192 vacc5x4567 = wasm_f32x4_add(vacc5x4567, wasm_f32x4_mul(va5c3, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_arm_splat()
|
| /external/XNNPACK/src/f16-gemm/gen/ |
| D | 6x16-minmax-neonfp16arith-ld64.c | 230 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64() local
237 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64()
243 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3); in xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64()
|
| D | 6x8-minmax-neonfp16arith-ld64.c | 178 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64() local
185 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64()
|
| D | 8x16-minmax-neonfp16arith-ld64.c | 282 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64() local
291 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
299 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3); in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
|
| D | 8x8-minmax-neonfp16arith-ld64.c | 214 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64() local
223 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64()
|
| /external/XNNPACK/src/f16-igemm/gen/ |
| D | 6x16-minmax-neonfp16arith-ld64.c | 256 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64() local
263 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64()
269 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3); in xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64()
|
| D | 6x8-minmax-neonfp16arith-ld64.c | 204 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64() local
211 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64()
|
| D | 8x16-minmax-neonfp16arith-ld64.c | 314 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64() local
323 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64()
331 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3); in xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64()
|
| D | 8x8-minmax-neonfp16arith-ld64.c | 246 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64() local
255 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64()
|
| /external/XNNPACK/src/f16-gemm/gen-inc/ |
| D | 6x16inc-minmax-neonfp16arith-ld64.c | 232 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_gemminc_minmax_ukernel_6x16__neonfp16arith_ld64() local
239 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_gemminc_minmax_ukernel_6x16__neonfp16arith_ld64()
245 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3); in xnn_f16_gemminc_minmax_ukernel_6x16__neonfp16arith_ld64()
|
| D | 6x8inc-minmax-neonfp16arith-ld64.c | 180 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_gemminc_minmax_ukernel_6x8__neonfp16arith_ld64() local
187 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_gemminc_minmax_ukernel_6x8__neonfp16arith_ld64()
|
| D | 8x16inc-minmax-neonfp16arith-ld64.c | 284 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64() local
293 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64()
301 vacc5x89ABCDEF = vfmaq_f16(vacc5x89ABCDEF, va5c3, vb89ABCDEFc3); in xnn_f16_gemminc_minmax_ukernel_8x16__neonfp16arith_ld64()
|
| D | 8x8inc-minmax-neonfp16arith-ld64.c | 216 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3); in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64() local
225 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3); in xnn_f16_gemminc_minmax_ukernel_8x8__neonfp16arith_ld64()
|