bend-ml-tensor-gpu

Releases (9)

• 0.1.5.1 indexed
• 0.1.5.0 unindexed
• 0.1.4.0 unindexed
• 0.1.3.0 unindexed
• 0.1.2.1 unindexed
• 0.1.2.0 unindexed
• 0.1.1.1 unindexed
• 0.1.1.0 unindexed
• 0.1.0.0 unindexed

Definitions (147)

main.bend

• def root4_new
• def go4
• def go4_new
• def run4
• def run4_new
• def blk_len
• def blocks_acc
• def precise_c
• def mm_precise3
• def mm_precise2
• def matmul_precise
• def rng_word
• def rng_f32
• def rng_mixmlbend-generative's mix32: xor-shift 16, multiply, xor-shift 13, multiply, xor-shift 16, over U32 that wraps
• def rng_hashthe hash of index i of counter s
• def rng_uniformthe uniform of index i of counter s, in (0, 1]: the hash's low 24 bits plus one, over 2^24 (exact)
• def rng_ln_mln m for m in [sqrt(1/2), sqrt(2)], from s = (m - 1) / (m + 1): ln m = 2 (s + s^3/3 + s^5/5 + s^7/7 + s^9/9)
• def rng_ln_wln x for a positive normal x: x = m * 2^ex with m in [sqrt(1/2), sqrt(2)) read from the bits (a mantissa above sqrt(2)'s, 0x3504F3, is halved and the exponent raised by one); ex * ln 2 in two parts, the first exact
• def rng_ln
• def rng_sinsin x and cos x for x in [0, pi/4], Taylor to x^9 and x^10 (the next terms are below 2^-28)
• def rng_cos
• type SCthe result of rng_sc: sin and cos of one angle
• def rng_sc_qsin and cos of a quarter turn's first eighth (x) or second (pi/2 - x, the two swapped), then turned by q quarters
• def rng_sc_x
• def rng_sc_wsin and cos of 2 pi v / 2^24 (v < 2^24): the quarter q is v's top 2 bits, w the rest; past the eighth (2^21) the angle is taken from the quarter's end (2^22 - w, exact), so the polynomial only sees [0, pi/4]
• def rng_sc
• def rng_pair
• def rng_gauss2the Gaussians of cells 2p and 2p + 1 of counter s (Box-Muller).
• def rng_pick
• def rng_gaussthe Gaussian of cell i alone (fill_gauss's cell i)
• type FEthe counter, the first index and the number of cells, the same for every leaf
• def fu_go
• def fu_leafleaf t writes the cells t, t + l, ..
• def fg_ifwrites cell i when it is below n (the cell before the first of an odd i0, which wraps to 2^32 - 1, and the cell after the last are not written)
• def fg_put
• def fg_gopair j of the slice is the global pair q + j (q = i0 / 2), cells 2j - h and 2j - h + 1 of the output (h = i0 mod 2)
• def fg_pairsthe number of global pairs that cells i0 ..
• def fg_leafleaf t writes the pairs t, t + l, ..
• def r1_out
• def fill_uniform_at
• def fill_gauss_at
• def fill_uniform
• def fill_gauss
• def uniform_mat
• def gauss_mat
• def uniform_mat_at
• def gauss_mat_at