Home
last modified time | relevance | path

Searched refs:va4c3 (Results 1 – 25 of 34) sorted by relevance

12

/external/XNNPACK/src/f32-gemm/gen/
D5x8-relu-wasmsimd-splat.c152 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemm_relu_ukernel_5x8__wasmsimd_splat() local
161 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemm_relu_ukernel_5x8__wasmsimd_splat()
166 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemm_relu_ukernel_5x8__wasmsimd_splat()
D5x8-wasmsimd-splat.c152 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemm_ukernel_5x8__wasmsimd_splat() local
161 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemm_ukernel_5x8__wasmsimd_splat()
166 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemm_ukernel_5x8__wasmsimd_splat()
D5x8-minmax-wasmsimd-arm-splat.c154 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_splat() local
163 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_splat()
168 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_splat()
D5x8-minmax-wasmsimd-x86-splat.c152 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_splat() local
161 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_splat()
166 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_splat()
D6x8-minmax-neonfma-dup-ld128.c169 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld128() local
175 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld128()
181 vacc4x4567 = vfmaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld128()
D6x8-minmax-wasmsimd-arm-splat.c173 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_splat() local
183 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_splat()
189 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_splat()
D6x8-minmax-neon-dup-ld128.c169 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld128() local
175 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld128()
181 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld128()
D6x8-minmax-wasmsimd-x86-splat.c171 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_splat() local
181 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_splat()
187 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_splat()
/external/XNNPACK/src/f32-igemm/gen/
D5x8-wasmsimd-splat.c177 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_igemm_ukernel_5x8__wasmsimd_splat() local
186 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_igemm_ukernel_5x8__wasmsimd_splat()
191 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_igemm_ukernel_5x8__wasmsimd_splat()
D5x8-relu-wasmsimd-splat.c177 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_igemm_relu_ukernel_5x8__wasmsimd_splat() local
186 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_igemm_relu_ukernel_5x8__wasmsimd_splat()
191 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_igemm_relu_ukernel_5x8__wasmsimd_splat()
D5x8-minmax-wasmsimd-x86-splat.c177 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_splat() local
186 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_splat()
191 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_x86_splat()
D5x8-minmax-wasmsimd-arm-splat.c179 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_arm_splat() local
188 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_arm_splat()
193 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_arm_splat()
D6x8-minmax-neon-dup-ld128.c197 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld128() local
203 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld128()
209 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld128()
D6x8-minmax-neonfma-dup-ld128.c197 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld128() local
203 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld128()
209 vacc4x4567 = vfmaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld128()
D6x8-minmax-wasmsimd-arm-splat.c201 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_splat() local
211 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_splat()
217 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_arm_splat()
D6x8-minmax-wasmsimd-x86-splat.c199 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_splat() local
209 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_splat()
215 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_igemm_minmax_ukernel_6x8__wasmsimd_x86_splat()
/external/XNNPACK/src/f32-gemm/gen-inc/
D5x8inc-minmax-wasmsimd-arm-splat.c156 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_arm_splat() local
165 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_arm_splat()
170 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_arm_splat()
D5x8inc-minmax-wasmsimd-x86-splat.c154 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_splat() local
163 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_splat()
168 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_splat()
D6x8inc-minmax-neon-dup-ld128.c171 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld128() local
177 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld128()
183 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld128()
D6x8inc-minmax-neonfma-dup-ld128.c171 const float32x4_t va4c3 = vdupq_lane_f32(vget_high_f32(va4), 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld128() local
177 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c3, vb0123c3); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld128()
183 vacc4x4567 = vfmaq_f32(vacc4x4567, va4c3, vb4567c3); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld128()
D6x8inc-minmax-wasmsimd-x86-splat.c173 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_splat() local
183 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_splat()
189 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_splat()
D6x8inc-minmax-wasmsimd-arm-splat.c175 const v128_t va4c3 = wasm_v32x4_shuffle(va4, va4, 3, 3, 3, 3); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_arm_splat() local
185 vacc4x0123 = wasm_f32x4_add(vacc4x0123, wasm_f32x4_mul(va4c3, vb0123c3)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_arm_splat()
191 vacc4x4567 = wasm_f32x4_add(vacc4x4567, wasm_f32x4_mul(va4c3, vb4567c3)); in xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_arm_splat()
/external/XNNPACK/src/f16-gemm/gen/
D6x16-minmax-neonfp16arith-ld64.c229 const float16x8_t va4c3 = vdupq_lane_f16(va4, 3); in xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64() local
236 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3); in xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64()
242 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c3, vb89ABCDEFc3); in xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64()
/external/XNNPACK/src/f16-igemm/gen/
D6x16-minmax-neonfp16arith-ld64.c255 const float16x8_t va4c3 = vdupq_lane_f16(va4, 3); in xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64() local
262 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3); in xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64()
268 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c3, vb89ABCDEFc3); in xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64()
/external/XNNPACK/src/f16-gemm/gen-inc/
D6x16inc-minmax-neonfp16arith-ld64.c231 const float16x8_t va4c3 = vdupq_lane_f16(va4, 3); in xnn_f16_gemminc_minmax_ukernel_6x16__neonfp16arith_ld64() local
238 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3); in xnn_f16_gemminc_minmax_ukernel_6x16__neonfp16arith_ld64()
244 vacc4x89ABCDEF = vfmaq_f16(vacc4x89ABCDEF, va4c3, vb89ABCDEFc3); in xnn_f16_gemminc_minmax_ukernel_6x16__neonfp16arith_ld64()

12