Commit b755f375 authored by aska-0096's avatar aska-0096
Browse files

add save_x=true instance

parent 35ba0864
......@@ -141,6 +141,17 @@ float add_rmsnorm2d_rdquant_fwd_b16_(add_rmsnorm2d_rdquant_fwd_traits t,
r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1, 8, 1, 1024, 1, true, false, false>>(s, a);
}
}
else{
if(t.save_x){
if (a.n % 8 == 0)
r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1, 2, 1, 512, 8, false, true, false>>(s, a);
else if (a.n % 4 == 0)
r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1, 4, 1, 512, 4, false, true, false>>(s, a);
else if (a.n % 2 == 0)
r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1, 4, 1, 1024, 2, false, true, false>>(s, a);
else
r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1, 8, 1, 1024, 1, false, true, false>>(s, a);
}
else{
if (a.n % 8 == 0)
r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1, 2, 1, 512, 8, false, false, false>>(s, a);
......@@ -152,6 +163,7 @@ float add_rmsnorm2d_rdquant_fwd_b16_(add_rmsnorm2d_rdquant_fwd_traits t,
r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1, 8, 1, 1024, 1, false, false, false>>(s, a);
}
}
}
else if(a.n > 8192) {
if (a.n % 8 == 0)
r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1, 2, 1, 512, 8, true, true, true>>(s, a);
......
......@@ -14,6 +14,10 @@ template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 512, 4, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 8, 1, 1024, 1, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 512, 8, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 512, 4, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 8, 1, 1024, 1, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1, 512, 8, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 512, 4, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2, false, false, false>>(const S&, A);
......@@ -26,6 +30,10 @@ template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 512, 4, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 8, 1, 1024, 1, true, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 512, 8, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 512, 4, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 8, 1, 1024, 1, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1, 512, 8, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 512, 4, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, false, false, false>>(const S&, A);
......
......@@ -18,6 +18,10 @@ template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 512, 4, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 1024, 1, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 2, 1, 512, 8, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 512, 4, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::int8_t, 1, 8, 1, 1024, 1, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 2, 1, 512, 8, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 512, 4, true, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, true, true, false>>(const S&, A);
......@@ -30,4 +34,8 @@ template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 512, 4, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 1024, 1, false, false, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 2, 1, 512, 8, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 512, 4, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2, false, true, false>>(const S&, A);
template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, ck_tile::fp8_t, 1, 8, 1, 1024, 1, false, true, false>>(const S&, A);
// clang-format on
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment