X-Git-Url: https://git.stderr.nl/gitweb?p=matthijs%2Fprojects%2Fmontium-fft.git;a=blobdiff_plain;f=FFT.mc;fp=FFT.mc;h=3ad081668c51b058cece7d041f772a5c415ad94c;hp=330b7f5be67632df7d362477e7651dc8492ac31c;hb=a62dbd706365d44bdcac9f8ac4c7c6dd05484642;hpb=57eb86fc79d7056a4b91438104cc1fe1eaf86c59 diff --git a/FFT.mc b/FFT.mc index 330b7f5..3ad0816 100644 --- a/FFT.mc +++ b/FFT.mc @@ -16,8 +16,10 @@ INLINE struct bf_out butterfly(struct bf_in in) { /* ALU 0 & 1 */ /* im(W) * im(b) */ aluexp Wixbi = west(fmul(rd1(in.W_im), rb1(in.b_im))); + /* re(W * b) = re(W) * re(b) - im(W) * im(b) */ aluexp Wxbr = ssub_acc(fmul(rc1(in.W_re), ra1(in.b_re)), Wixbi); + /* re(out_a) = re(a) + re(W * b) */ out.a_re = p0o0(sadd_bf(rb1(in.a_re), Wxbr)); @@ -147,10 +149,10 @@ INLINE void init_output_addresses_regular(struct mems m, bool stage_odd, bool se set_offset(m.output_b_re, 0-2); set_offset(m.output_b_im, 0-2); } else { - set_offset(m.output_a_re, 1-2); - set_offset(m.output_a_im, 1-2); - set_offset(m.output_b_re, 0-2); - set_offset(m.output_b_im, 0-2); + set_offset(m.output_a_re, 0-2); + set_offset(m.output_a_im, 0-2); + set_offset(m.output_b_re, 1-2); + set_offset(m.output_b_im, 1-2); } } @@ -176,10 +178,10 @@ INLINE void do_half_regular_stage(struct mems m, bool stage_odd, bool second_hal struct bf_in in = read_input_regular(m, EVEN_CYCLE, stage_odd); struct bf_out out = butterfly(in); - /* Now, do a single stage. That means N_t / 2 cycles. Since we do 2 + /* Now, do half a single stage. That means N_t / 4 cycles. Since we do 2 * cycles on every iteration, plus one before and after the loop, - * we will loop N_t / 4 - 1 times. */ - init_loop(LC2, (PARAM_N_t / 4) - 1); + * we will loop N_t / 8 - 1 times. */ + init_loop(LC2, (PARAM_N_t / 8) - 1); do { /* Write outputs of previous cycle */ write_output_regular(m, out, second_half); @@ -243,17 +245,22 @@ INLINE struct mems init_mem_mapping(bool stage_odd){ void run() { do { freeze(); } while (gpi(0) == 0); struct mems m; - - m = init_mem_mapping(EVEN_STAGE); - init_input_addresses_regular(m, EVEN_STAGE); - /* do_half_regular_stage will init output addresses */ - next_cycle(); - do_half_regular_stage(m, EVEN_STAGE, FIRST_HALF); - do_half_regular_stage(m, EVEN_STAGE, SECOND_HALF); - next_cycle(); - init_input_addresses_regular(m, ODD_STAGE); - m = init_mem_mapping(ODD_STAGE); - next_cycle(); - do_half_regular_stage(m, ODD_STAGE, FIRST_HALF); - do_half_regular_stage(m, ODD_STAGE, SECOND_HALF); + + /* We need to do n_t regular stages. Since we do two stages each + * iteration, we'll do n_t / 2 iterations. */ + init_loop(LC1, (PARAM_n_t / 2)); + do { + m = init_mem_mapping(EVEN_STAGE); + init_input_addresses_regular(m, EVEN_STAGE); + /* do_half_regular_stage will init output addresses */ + next_cycle(); + do_half_regular_stage(m, EVEN_STAGE, FIRST_HALF); + do_half_regular_stage(m, EVEN_STAGE, SECOND_HALF); + next_cycle(); + init_input_addresses_regular(m, ODD_STAGE); + m = init_mem_mapping(ODD_STAGE); + next_cycle(); + do_half_regular_stage(m, ODD_STAGE, FIRST_HALF); + do_half_regular_stage(m, ODD_STAGE, SECOND_HALF); + } while (loop_next(LC1)); }