initial gfx1036 and gfx1103 support

- initial support for gfx1036 and gfx1103 as a build target
- updated also the gfx1010 configuration settings to be
  more similar in composable kernel and miopen

fixes: #101
fixes: #103

Signed-off-by: Mika Laitio <lamikr@gmail.com>

Author: lamikr

binfo/040_02_onnxruntime_deepspeed.binfo

@@ -14,7 +14,7 @@ BINFO_APP_PRE_CONFIG_CMD_ARRAY=(
 
 BINFO_APP_BUILD_CMD_ARRAY=(
     "cd ${BINFO_APP_SRC_DIR}"
-    "./build_rocm.sh \"${SEMICOLON_SEPARATED_GPU_TARGET_LIST_DEFAULT}\""
+    "./build_rocm.sh ${SEMICOLON_SEPARATED_GPU_TARGET_LIST_DEFAULT}"
 )
 
 BINFO_APP_INSTALL_CMD_ARRAY=(

binfo/user_config.sh

@@ -3,7 +3,7 @@
 # Function to select ROCM SDK build target GPUs
 func_build_cfg_user() {
     local message="Select ROCM SDK build target GPUs. Space to select, Enter to finish save, ESC to cancel."
-    local options="gfx906|gfx90a|gfx940|gfx1010|gfx1011|gfx1012|gfx1030|gfx1031|gfx1035|gfx1100|gfx1101|gfx1102|gfx1150|gfx1151"
+    local options="gfx906|gfx90a|gfx940|gfx1010|gfx1011|gfx1012|gfx1030|gfx1031|gfx1035|gfx1036|gfx1100|gfx1101|gfx1102|gfx1103|gfx1150|gfx1151"
     local script_path="./build/checkbox.sh"
 
     # Check if the script exists and is executable
@@ -14,4 +14,4 @@ func_build_cfg_user() {
 
     # Execute the checkbox script with the specified parameters
     "$script_path" --message="$message" --options="$options" --multiple
-}
+}

patches/rocm-6.1.2/DeepSpeed/0001-deepspeed-rocm-preconfig-and-build_install-scripts.patch

@@ -1,4 +1,4 @@
-From 108755dda9e284382bd0b63e5e351cba9f7abb16 Mon Sep 17 00:00:00 2001
+From 3d237f904f4f74618da216179f9279fd027b05e8 Mon Sep 17 00:00:00 2001
 From: Mika Laitio <lamikr@gmail.com>
 Date: Mon, 20 May 2024 22:36:23 -0700
 Subject: [PATCH 1/4] deepspeed rocm preconfig and build_install scripts

patches/rocm-6.1.2/DeepSpeed/0002-check-rocm-path-from-installed-pytorch-variables.patch

@@ -1,4 +1,4 @@
-From 40f44ac1ca3cc73851bfb7c3016276894022eb4f Mon Sep 17 00:00:00 2001
+From ff7f27c49a5ac7ac31f7ebc70e02c18f3aa0ac91 Mon Sep 17 00:00:00 2001
 From: Mika Laitio <lamikr@gmail.com>
 Date: Tue, 21 May 2024 07:57:53 -0700
 Subject: [PATCH 2/4] check rocm path from installed pytorch variables

patches/rocm-6.1.2/DeepSpeed/0003-allow-building-deepspeed-for-rocm-in-virtual-linux.patch

@@ -1,4 +1,4 @@
-From 59fe85c08a0900acaf07bd2aa506853af3fbd83e Mon Sep 17 00:00:00 2001
+From e423db20d2eb75934214c4487fb4193f3df389a9 Mon Sep 17 00:00:00 2001
 From: Mika Laitio <lamikr@gmail.com>
 Date: Wed, 26 Jun 2024 14:44:04 -0700
 Subject: [PATCH 3/4] allow building deepspeed for rocm in virtual linux
@@ -10,24 +10,26 @@ fixes: https://github.com/lamikr/rocm_sdk_builder/issues/75
 
 Signed-off-by: Mika Laitio <lamikr@gmail.com>
 ---
- build_rocm.sh         |  2 ++
+ build_rocm.sh         |  4 +++-
  op_builder/builder.py |  2 +-
  setup.py              | 18 ++++++++++--------
- 3 files changed, 13 insertions(+), 9 deletions(-)
+ 3 files changed, 14 insertions(+), 10 deletions(-)
 
 diff --git a/build_rocm.sh b/build_rocm.sh
-index f4a75855..3f3ce526 100755
+index f4a75855..b48a0ad9 100755
 --- a/build_rocm.sh
 +++ b/build_rocm.sh
-@@ -15,6 +15,8 @@ mkdir -p deepspeed/ops/spatial
+@@ -15,7 +15,9 @@ mkdir -p deepspeed/ops/spatial
  #export CFLAGS="-I/usr/include"
  #export LDFLAGS="-L/usr/lib64"
 
 +# needed by real accelerator.py to detect the cuda when build on virtual linux without access to real hardware
 +export DS_ACCELERATOR=cuda
  # install command will create wheel and install it. bdist_wheel comamnd would only create the wheel
- AMDGPU_TARGETS=${amd_target_gpu} DS_BUILD_AIO=0 DS_BUILD_FP_QUANTIZER=0 DS_BUILD_QUANTIZER=0 DS_BUILD_SPARSE_ATTN=0 DS_BUILD_RAGGED_DEVICE_OPS=0 DS_BUILD_CUTLASS_OPS=0 DS_BUILD_EVOFORMER_ATTN=0 DS_BUILD_OPS=1 python setup.py bdist_wheel
+-AMDGPU_TARGETS=${amd_target_gpu} DS_BUILD_AIO=0 DS_BUILD_FP_QUANTIZER=0 DS_BUILD_QUANTIZER=0 DS_BUILD_SPARSE_ATTN=0 DS_BUILD_RAGGED_DEVICE_OPS=0 DS_BUILD_CUTLASS_OPS=0 DS_BUILD_EVOFORMER_ATTN=0 DS_BUILD_OPS=1 python setup.py bdist_wheel
++AMDGPU_TARGETS="${amd_target_gpu}" DS_BUILD_AIO=0 DS_BUILD_FP_QUANTIZER=0 DS_BUILD_QUANTIZER=0 DS_BUILD_SPARSE_ATTN=0 DS_BUILD_RAGGED_DEVICE_OPS=0 DS_BUILD_CUTLASS_OPS=0 DS_BUILD_EVOFORMER_ATTN=0 DS_BUILD_OPS=1 python setup.py bdist_wheel
 
+ #DS_BUILD_UTILS=1 DS_BUILD_CPU_ADAGRAD=1 DS_BUILD_RANDOM_LTD=1 DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 DS_BUILD_FUSED_LAMB=1 DS_BUILD_CCL_COMM=1 python setup.py develop
 diff --git a/op_builder/builder.py b/op_builder/builder.py
 index a27b134c..4980a528 100644
 --- a/op_builder/builder.py

patches/rocm-6.1.2/DeepSpeed/0004-remove-linear_kernel-which-fails-on-rocm.patch

@@ -1,4 +1,4 @@
-From bae0b1c89880e23f8ec885eba5f18043726319a0 Mon Sep 17 00:00:00 2001
+From 44b7a213e104e7f7b9e6f9b7e63f5eb8ff7be88f Mon Sep 17 00:00:00 2001
 From: Mika Laitio <lamikr@gmail.com>
 Date: Tue, 21 May 2024 11:41:20 -0700
 Subject: [PATCH 4/4] remove linear_kernel which fails on rocm

patches/rocm-6.1.2/MIOpen/0001-Do-not-fail-on-install-for-missing-kdb.bz2-file.patch

@@ -1,7 +1,7 @@
-From 915a3c074ff1c3694566fd74f2e83003a1253104 Mon Sep 17 00:00:00 2001
+From 7b51b6d06842de4851b6c20b61dda2ead78e68ff Mon Sep 17 00:00:00 2001
 From: Mika Laitio <lamikr@pilppa.org>
 Date: Wed, 8 May 2024 13:48:21 -0700
-Subject: [PATCH 1/4] Do not fail on install for missing kdb.bz2 file
+Subject: [PATCH 1/5] Do not fail on install for missing kdb.bz2 file
 
 Do not fail if kdb.bz2 file for some GPU does not exist because
 these kdb files are not mandatory. Their function is to speed up the
@@ -43,5 +43,5 @@ index 32d9a2e5b..d6c2db704 100644
  endif()
 
 -- 
-2.41.1
+2.45.2
 

patches/rocm-6.1.2/MIOpen/0002-fix-libroctx64.so-linking-error.patch

@@ -1,7 +1,7 @@
-From ecb981cd1b66749186404fa76c56237c758953b5 Mon Sep 17 00:00:00 2001
+From fdc0e268b2edbb0a57af8e8820b388816af38471 Mon Sep 17 00:00:00 2001
 From: Mika Laitio <lamikr@pilppa.org>
 Date: Fri, 3 May 2024 14:15:09 -0700
-Subject: [PATCH 2/4] fix libroctx64.so linking error
+Subject: [PATCH 2/5] fix libroctx64.so linking error
 
 search the library and if found link it from there
 instead of expecting it to be in the ld library path
@@ -31,5 +31,5 @@ index 0741a6023..ae4405eed 100644
 
  ############################################################
 -- 
-2.41.1
+2.45.2
 

patches/rocm-6.1.2/MIOpen/0003-MIOpen-gfx1010-and-gfx1035-support.patch

@@ -1,7 +1,7 @@
-From 6578a68e3226e97716aad12d445632358f2a463e Mon Sep 17 00:00:00 2001
+From 4b65dfcb0208bfe1eb64c474fb97e31e2b8bf60f Mon Sep 17 00:00:00 2001
 From: Mika Laitio <lamikr@gmail.com>
 Date: Tue, 19 Dec 2023 15:13:46 -0800
-Subject: [PATCH 3/4] MIOpen gfx1010 and gfx1035 support
+Subject: [PATCH 3/5] MIOpen gfx1010 and gfx1035 support
 
 - todo: check gfx1010 specific parts
 
@@ -174,5 +174,5 @@ index 16ce78f04..2ec3eaf09 100644
                         "gfx1101",
                         "gfx1102"};
 -- 
-2.41.1
+2.45.2
 

patches/rocm-6.1.2/MIOpen/0004-improved-gfx1010-support.patch

@@ -1,7 +1,7 @@
-From 5e7803271cbbe475da352ab188f09b345006d9c0 Mon Sep 17 00:00:00 2001
+From 3c0552223df66774e8f3613826a6094939763cce Mon Sep 17 00:00:00 2001
 From: Mika Laitio <lamikr@pilppa.org>
 Date: Mon, 8 Jul 2024 21:44:10 +0300
-Subject: [PATCH 4/4] improved gfx1010 support
+Subject: [PATCH 4/5] improved gfx1010 support
 
 - allows running pytorch gpu benchmark
   on gfx1010/amd rx 5700
@@ -135,5 +135,5 @@ index bf02d4d55..c3fa2bd3a 100644
          {"Rembrandt", "gfx1035"},
      };
 -- 
-2.41.1
+2.45.2