50 			  unsigned int bytes, int nrounds)  in chacha_dosimd()  argument
55 		while (bytes >= CHACHA_BLOCK_SIZE * 8) {  in chacha_dosimd()
56 			chacha_8block_xor_avx512vl(state, dst, src, bytes,  in chacha_dosimd()
58 			bytes -= CHACHA_BLOCK_SIZE * 8;  in chacha_dosimd()
63 		if (bytes > CHACHA_BLOCK_SIZE * 4) {  in chacha_dosimd()
64 			chacha_8block_xor_avx512vl(state, dst, src, bytes,  in chacha_dosimd()
66 			state[12] += chacha_advance(bytes, 8);  in chacha_dosimd()
69 		if (bytes > CHACHA_BLOCK_SIZE * 2) {  in chacha_dosimd()
70 			chacha_4block_xor_avx512vl(state, dst, src, bytes,  in chacha_dosimd()
72 			state[12] += chacha_advance(bytes, 4);  in chacha_dosimd()
75 		if (bytes) {  in chacha_dosimd()
76 			chacha_2block_xor_avx512vl(state, dst, src, bytes,  in chacha_dosimd()
78 			state[12] += chacha_advance(bytes, 2);  in chacha_dosimd()
84 		while (bytes >= CHACHA_BLOCK_SIZE * 8) {  in chacha_dosimd()
85 			chacha_8block_xor_avx2(state, dst, src, bytes, nrounds);  in chacha_dosimd()
86 			bytes -= CHACHA_BLOCK_SIZE * 8;  in chacha_dosimd()
91 		if (bytes > CHACHA_BLOCK_SIZE * 4) {  in chacha_dosimd()
92 			chacha_8block_xor_avx2(state, dst, src, bytes, nrounds);  in chacha_dosimd()
93 			state[12] += chacha_advance(bytes, 8);  in chacha_dosimd()
96 		if (bytes > CHACHA_BLOCK_SIZE * 2) {  in chacha_dosimd()
97 			chacha_4block_xor_avx2(state, dst, src, bytes, nrounds);  in chacha_dosimd()
98 			state[12] += chacha_advance(bytes, 4);  in chacha_dosimd()
101 		if (bytes > CHACHA_BLOCK_SIZE) {  in chacha_dosimd()
102 			chacha_2block_xor_avx2(state, dst, src, bytes, nrounds);  in chacha_dosimd()
103 			state[12] += chacha_advance(bytes, 2);  in chacha_dosimd()
108 	while (bytes >= CHACHA_BLOCK_SIZE * 4) {  in chacha_dosimd()
109 		chacha_4block_xor_ssse3(state, dst, src, bytes, nrounds);  in chacha_dosimd()
110 		bytes -= CHACHA_BLOCK_SIZE * 4;  in chacha_dosimd()
115 	if (bytes > CHACHA_BLOCK_SIZE) {  in chacha_dosimd()
116 		chacha_4block_xor_ssse3(state, dst, src, bytes, nrounds);  in chacha_dosimd()
117 		state[12] += chacha_advance(bytes, 4);  in chacha_dosimd()
120 	if (bytes) {  in chacha_dosimd()
121 		chacha_block_xor_ssse3(state, dst, src, bytes, nrounds);  in chacha_dosimd()
130 	int next_yield = 4096; /* bytes until next FPU yield */  in chacha_simd_stream_xor()