From f02db165f4000d6576686b0e38a9f6085f3045b1 Mon Sep 17 00:00:00 2001
From: Admin AWS GPU <admin@aws.gpu>
Date: Thu, 29 Feb 2024 09:47:04 +0000
Subject: [PATCH 1/3] Fix bugs

---
 field/Cargo.toml          |  2 +-
 plonky2/src/fri/oracle.rs | 16 +++++++++-------
 2 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/field/Cargo.toml b/field/Cargo.toml
index 2eeafc0454..a0f377107b 100644
--- a/field/Cargo.toml
+++ b/field/Cargo.toml
@@ -32,7 +32,7 @@ proc-macro2 = "1"
 quote = "1"
 
 [features]
-default = ["no_cuda"]
+default = []
 cuda = ["cryptography_cuda"]
 precompile = []
 no_cuda = ["cryptography_cuda/no_cuda"]
\ No newline at end of file
diff --git a/plonky2/src/fri/oracle.rs b/plonky2/src/fri/oracle.rs
index 927128c8d8..479d490b4e 100644
--- a/plonky2/src/fri/oracle.rs
+++ b/plonky2/src/fri/oracle.rs
@@ -141,11 +141,7 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
     ) -> Self {
         let degree = polynomials[0].len();
         let log_n = log2_strict(degree) + rate_bits;
-        let num_gpus: usize = std::env::var("NUM_OF_GPUS")
-            .expect("NUM_OF_GPUS should be set")
-            .parse()
-            .unwrap();
-        
+
         #[cfg(feature = "cuda")]
         if(log_n > 10 && polynomials.len() > 0){
             let lde_values = Self::from_coeffs_gpu(
@@ -159,6 +155,12 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
                 degree
             );
 
+            let num_gpus: usize = std::env::var("NUM_OF_GPUS")
+                .expect("NUM_OF_GPUS should be set")
+                .parse()
+                .unwrap();
+        
+
             if num_gpus != 1 {
                 let mut leaves = timed!(timing, "transpose LDEs", transpose(&lde_values));
                 reverse_index_bits_in_place(&mut leaves);
@@ -314,12 +316,12 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
                             id as i32, 
                             device_transpose_data.as_mut_ptr(), 
                             device_output_data.as_mut_ptr(), 
-                            log2_strict(degree), 
+                            log_n, 
                             cfg_trans
                         );
 
                         let start = std::time::Instant::now();
-                        let nums: Vec<usize> = (0..poly_chunk.len()).collect();
+                        let nums: Vec<usize> = (0..(1<<log_n)).collect();
                         let r = nums
                             .par_iter()
                             .map(|i| {

From d1dd98e46b233378ffabc7e0989ce729b033c37a Mon Sep 17 00:00:00 2001
From: Admin AWS GPU <admin@aws.gpu>
Date: Thu, 29 Feb 2024 10:03:19 +0000
Subject: [PATCH 2/3] Add timing for transpose

---
 plonky2/src/fri/oracle.rs | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/plonky2/src/fri/oracle.rs b/plonky2/src/fri/oracle.rs
index 479d490b4e..ae657eebb4 100644
--- a/plonky2/src/fri/oracle.rs
+++ b/plonky2/src/fri/oracle.rs
@@ -312,6 +312,8 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
                         cfg_trans.batches = per_device_batch as u32;
                         cfg_trans.are_inputs_on_device = true;
                         cfg_trans.are_outputs_on_device = true;
+
+                        let start = std::time::Instant::now();
                         transpose_rev_batch(
                             id as i32, 
                             device_transpose_data.as_mut_ptr(), 
@@ -320,6 +322,8 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
                             cfg_trans
                         );
 
+                        println!("real transpose_rev_batch elapsed: {:?}", start.elapsed());
+
                         let start = std::time::Instant::now();
                         let nums: Vec<usize> = (0..(1<<log_n)).collect();
                         let r = nums

From e93d686bc03bf7577cae9ef30219e4d9043a2566 Mon Sep 17 00:00:00 2001
From: Admin AWS GPU <admin@aws.gpu>
Date: Thu, 29 Feb 2024 11:03:24 +0000
Subject: [PATCH 3/3] Update Cargo.toml

---
 field/Cargo.toml | 2 +-
 rust-toolchain   | 1 -
 2 files changed, 1 insertion(+), 2 deletions(-)
 delete mode 100644 rust-toolchain

diff --git a/field/Cargo.toml b/field/Cargo.toml
index a0f377107b..2eeafc0454 100644
--- a/field/Cargo.toml
+++ b/field/Cargo.toml
@@ -32,7 +32,7 @@ proc-macro2 = "1"
 quote = "1"
 
 [features]
-default = []
+default = ["no_cuda"]
 cuda = ["cryptography_cuda"]
 precompile = []
 no_cuda = ["cryptography_cuda/no_cuda"]
\ No newline at end of file
diff --git a/rust-toolchain b/rust-toolchain
deleted file mode 100644
index 07ade694b1..0000000000
--- a/rust-toolchain
+++ /dev/null
@@ -1 +0,0 @@
-nightly
\ No newline at end of file