wave8_transpose_16x4_pipe4()

FASTLED_FORCE_INLINE FL_IRAM FL_OPTIMIZE_FUNCTION void fl::detail::wave8_transpose_16x4_pipe4	(	const Wave8Byte	lane_waves_a[16],
		const Wave8Byte	lane_waves_b[16],
		const Wave8Byte	lane_waves_c[16],
		const Wave8Byte	lane_waves_d[16],
		u8	output_a[16 *sizeof(Wave8Byte)],
		u8	output_b[16 *sizeof(Wave8Byte)],
		u8	output_c[16 *sizeof(Wave8Byte)],
		u8	output_d[16 *sizeof(Wave8Byte)] )

Pipe4: transpose 16-lane × 4-byte-positions in one fused call.

Bit-identical to four sequential wave8_transpose_16 calls. Extends the pipe2 idea to 4 positions — empirically the peak of the curve on the in-order RV32 P4 core (#2548). pipe2 saved 26% over baseline, pipe3 36%, pipe4 41%, pipe6 regressed to 94% (register spill). Stays within the 32-GPR budget: 4 × 4 = 16 OR-tree accumulators + ~8 misc GPRs = ~24 live; pipe6 would push this past 32.

Definition at line 273 of file wave8.hpp.

                                                                     {
    for (int symbol_idx = 0; symbol_idx < 8; symbol_idx++) {
        u8 la[16];
        u8 lb[16];
        u8 lc[16];
        u8 ld[16];
        for (int lane = 0; lane < 16; lane++) {
            la[lane] = lane_waves_a[lane].symbols[symbol_idx].data;
            lb[lane] = lane_waves_b[lane].symbols[symbol_idx].data;
            lc[lane] = lane_waves_c[lane].symbols[symbol_idx].data;
            ld[lane] = lane_waves_d[lane].symbols[symbol_idx].data;
        }
        spread_transpose16_symbol(la, output_a + symbol_idx * 16);
        spread_transpose16_symbol(lb, output_b + symbol_idx * 16);
        spread_transpose16_symbol(lc, output_c + symbol_idx * 16);
        spread_transpose16_symbol(ld, output_d + symbol_idx * 16);
    }
}

References spread_transpose16_symbol().

Referenced by fl::wave8Transpose_16x4_pipe4().

Here is the call graph for this function:

Here is the caller graph for this function: